kern_synch.c revision 81397
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 81397 2001-08-10 06:37:05Z jhb $ 40 */ 41 42#include "opt_ktrace.h" 43 44#include <sys/param.h> 45#include <sys/systm.h> 46#include <sys/condvar.h> 47#include <sys/kernel.h> 48#include <sys/ktr.h> 49#include <sys/lock.h> 50#include <sys/mutex.h> 51#include <sys/proc.h> 52#include <sys/resourcevar.h> 53#include <sys/signalvar.h> 54#include <sys/smp.h> 55#include <sys/sx.h> 56#include <sys/sysctl.h> 57#include <sys/sysproto.h> 58#include <sys/vmmeter.h> 59#include <vm/vm.h> 60#include <vm/vm_extern.h> 61#ifdef KTRACE 62#include <sys/uio.h> 63#include <sys/ktrace.h> 64#endif 65 66#include <machine/cpu.h> 67 68static void sched_setup __P((void *dummy)); 69SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 70 71int hogticks; 72int lbolt; 73int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 74 75static struct callout schedcpu_callout; 76static struct callout roundrobin_callout; 77 78static void endtsleep __P((void *)); 79static void roundrobin __P((void *arg)); 80static void schedcpu __P((void *arg)); 81 82static int 83sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 84{ 85 int error, new_val; 86 87 new_val = sched_quantum * tick; 88 error = sysctl_handle_int(oidp, &new_val, 0, req); 89 if (error != 0 || req->newptr == NULL) 90 return (error); 91 if (new_val < tick) 92 return (EINVAL); 93 sched_quantum = new_val / tick; 94 hogticks = 2 * sched_quantum; 95 return (0); 96} 97 98SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 99 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 100 101/* 102 * Arrange to reschedule if necessary, taking the priorities and 103 * schedulers into account. 104 */ 105void 106maybe_resched(p) 107 struct proc *p; 108{ 109 110 mtx_assert(&sched_lock, MA_OWNED); 111 if (p->p_pri.pri_level < curproc->p_pri.pri_level) 112 need_resched(curproc); 113} 114 115int 116roundrobin_interval(void) 117{ 118 return (sched_quantum); 119} 120 121/* 122 * Force switch among equal priority processes every 100ms. 123 * We don't actually need to force a context switch of the current process. 124 * The act of firing the event triggers a context switch to softclock() and 125 * then switching back out again which is equivalent to a preemption, thus 126 * no further work is needed on the local CPU. 127 */ 128/* ARGSUSED */ 129static void 130roundrobin(arg) 131 void *arg; 132{ 133 134#ifdef SMP 135 mtx_lock_spin(&sched_lock); 136 forward_roundrobin(); 137 mtx_unlock_spin(&sched_lock); 138#endif 139 140 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 141} 142 143/* 144 * Constants for digital decay and forget: 145 * 90% of (p_estcpu) usage in 5 * loadav time 146 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 147 * Note that, as ps(1) mentions, this can let percentages 148 * total over 100% (I've seen 137.9% for 3 processes). 149 * 150 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 151 * 152 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 153 * That is, the system wants to compute a value of decay such 154 * that the following for loop: 155 * for (i = 0; i < (5 * loadavg); i++) 156 * p_estcpu *= decay; 157 * will compute 158 * p_estcpu *= 0.1; 159 * for all values of loadavg: 160 * 161 * Mathematically this loop can be expressed by saying: 162 * decay ** (5 * loadavg) ~= .1 163 * 164 * The system computes decay as: 165 * decay = (2 * loadavg) / (2 * loadavg + 1) 166 * 167 * We wish to prove that the system's computation of decay 168 * will always fulfill the equation: 169 * decay ** (5 * loadavg) ~= .1 170 * 171 * If we compute b as: 172 * b = 2 * loadavg 173 * then 174 * decay = b / (b + 1) 175 * 176 * We now need to prove two things: 177 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 178 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 179 * 180 * Facts: 181 * For x close to zero, exp(x) =~ 1 + x, since 182 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 183 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 184 * For x close to zero, ln(1+x) =~ x, since 185 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 186 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 187 * ln(.1) =~ -2.30 188 * 189 * Proof of (1): 190 * Solve (factor)**(power) =~ .1 given power (5*loadav): 191 * solving for factor, 192 * ln(factor) =~ (-2.30/5*loadav), or 193 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 194 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 195 * 196 * Proof of (2): 197 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 198 * solving for power, 199 * power*ln(b/(b+1)) =~ -2.30, or 200 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 201 * 202 * Actual power values for the implemented algorithm are as follows: 203 * loadav: 1 2 3 4 204 * power: 5.68 10.32 14.94 19.55 205 */ 206 207/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 208#define loadfactor(loadav) (2 * (loadav)) 209#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 210 211/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 212static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 213SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 214 215/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 216static int fscale __unused = FSCALE; 217SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 218 219/* 220 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 221 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 222 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 223 * 224 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 225 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 226 * 227 * If you don't want to bother with the faster/more-accurate formula, you 228 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 229 * (more general) method of calculating the %age of CPU used by a process. 230 */ 231#define CCPU_SHIFT 11 232 233/* 234 * Recompute process priorities, every hz ticks. 235 * MP-safe, called without the Giant mutex. 236 */ 237/* ARGSUSED */ 238static void 239schedcpu(arg) 240 void *arg; 241{ 242 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 243 register struct proc *p; 244 register int realstathz; 245 246 realstathz = stathz ? stathz : hz; 247 sx_slock(&allproc_lock); 248 LIST_FOREACH(p, &allproc, p_list) { 249 /* 250 * Increment time in/out of memory and sleep time 251 * (if sleeping). We ignore overflow; with 16-bit int's 252 * (remember them?) overflow takes 45 days. 253 */ 254 mtx_lock_spin(&sched_lock); 255 p->p_swtime++; 256 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 257 p->p_slptime++; 258 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 259 /* 260 * If the process has slept the entire second, 261 * stop recalculating its priority until it wakes up. 262 */ 263 if (p->p_slptime > 1) { 264 mtx_unlock_spin(&sched_lock); 265 continue; 266 } 267 268 /* 269 * p_pctcpu is only for ps. 270 */ 271#if (FSHIFT >= CCPU_SHIFT) 272 p->p_pctcpu += (realstathz == 100)? 273 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 274 100 * (((fixpt_t) p->p_cpticks) 275 << (FSHIFT - CCPU_SHIFT)) / realstathz; 276#else 277 p->p_pctcpu += ((FSCALE - ccpu) * 278 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT; 279#endif 280 p->p_cpticks = 0; 281 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu); 282 resetpriority(p); 283 if (p->p_pri.pri_level >= PUSER) { 284 if (p->p_oncpu == NOCPU && /* idle */ 285 p->p_stat == SRUN && 286 (p->p_sflag & PS_INMEM) && 287 (p->p_pri.pri_level / RQ_PPQ) != 288 (p->p_pri.pri_user / RQ_PPQ)) { 289 remrunqueue(p); 290 p->p_pri.pri_level = p->p_pri.pri_user; 291 setrunqueue(p); 292 } else 293 p->p_pri.pri_level = p->p_pri.pri_user; 294 } 295 mtx_unlock_spin(&sched_lock); 296 } 297 sx_sunlock(&allproc_lock); 298 vmmeter(); 299 wakeup((caddr_t)&lbolt); 300 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 301} 302 303/* 304 * Recalculate the priority of a process after it has slept for a while. 305 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 306 * least six times the loadfactor will decay p_estcpu to zero. 307 */ 308void 309updatepri(p) 310 register struct proc *p; 311{ 312 register unsigned int newcpu = p->p_estcpu; 313 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 314 315 if (p->p_slptime > 5 * loadfac) 316 p->p_estcpu = 0; 317 else { 318 p->p_slptime--; /* the first time was done in schedcpu */ 319 while (newcpu && --p->p_slptime) 320 newcpu = decay_cpu(loadfac, newcpu); 321 p->p_estcpu = newcpu; 322 } 323 resetpriority(p); 324} 325 326/* 327 * We're only looking at 7 bits of the address; everything is 328 * aligned to 4, lots of things are aligned to greater powers 329 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 330 */ 331#define TABLESIZE 128 332static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE]; 333#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 334 335void 336sleepinit(void) 337{ 338 int i; 339 340 sched_quantum = hz/10; 341 hogticks = 2 * sched_quantum; 342 for (i = 0; i < TABLESIZE; i++) 343 TAILQ_INIT(&slpque[i]); 344} 345 346/* 347 * General sleep call. Suspends the current process until a wakeup is 348 * performed on the specified identifier. The process will then be made 349 * runnable with the specified priority. Sleeps at most timo/hz seconds 350 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 351 * before and after sleeping, else signals are not checked. Returns 0 if 352 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 353 * signal needs to be delivered, ERESTART is returned if the current system 354 * call should be restarted if possible, and EINTR is returned if the system 355 * call should be interrupted by the signal (return EINTR). 356 * 357 * The mutex argument is exited before the caller is suspended, and 358 * entered before msleep returns. If priority includes the PDROP 359 * flag the mutex is not entered before returning. 360 */ 361int 362msleep(ident, mtx, priority, wmesg, timo) 363 void *ident; 364 struct mtx *mtx; 365 int priority, timo; 366 const char *wmesg; 367{ 368 struct proc *p = curproc; 369 int sig, catch = priority & PCATCH; 370 int rval = 0; 371 WITNESS_SAVE_DECL(mtx); 372 373#ifdef KTRACE 374 if (p && KTRPOINT(p, KTR_CSW)) 375 ktrcsw(p->p_tracep, 1, 0); 376#endif 377 WITNESS_SLEEP(0, &mtx->mtx_object); 378 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 379 ("sleeping without a mutex")); 380 mtx_lock_spin(&sched_lock); 381 if (cold || panicstr) { 382 /* 383 * After a panic, or during autoconfiguration, 384 * just give interrupts a chance, then just return; 385 * don't run any other procs or panic below, 386 * in case this is the idle process and already asleep. 387 */ 388 if (mtx != NULL && priority & PDROP) 389 mtx_unlock_flags(mtx, MTX_NOSWITCH); 390 mtx_unlock_spin(&sched_lock); 391 return (0); 392 } 393 394 DROP_GIANT_NOSWITCH(); 395 396 if (mtx != NULL) { 397 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 398 WITNESS_SAVE(&mtx->mtx_object, mtx); 399 mtx_unlock_flags(mtx, MTX_NOSWITCH); 400 if (priority & PDROP) 401 mtx = NULL; 402 } 403 404 KASSERT(p != NULL, ("msleep1")); 405 KASSERT(ident != NULL && p->p_stat == SRUN, ("msleep")); 406 407 p->p_wchan = ident; 408 p->p_wmesg = wmesg; 409 p->p_slptime = 0; 410 p->p_pri.pri_level = priority & PRIMASK; 411 CTR5(KTR_PROC, "msleep: proc %p (pid %d, %s) on %s (%p)", p, p->p_pid, 412 p->p_comm, wmesg, ident); 413 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 414 if (timo) 415 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 416 /* 417 * We put ourselves on the sleep queue and start our timeout 418 * before calling CURSIG, as we could stop there, and a wakeup 419 * or a SIGCONT (or both) could occur while we were stopped. 420 * A SIGCONT would cause us to be marked as SSLEEP 421 * without resuming us, thus we must be ready for sleep 422 * when CURSIG is called. If the wakeup happens while we're 423 * stopped, p->p_wchan will be 0 upon return from CURSIG. 424 */ 425 if (catch) { 426 CTR3(KTR_PROC, "msleep caught: proc %p (pid %d, %s)", p, 427 p->p_pid, p->p_comm); 428 p->p_sflag |= PS_SINTR; 429 mtx_unlock_spin(&sched_lock); 430 PROC_LOCK(p); 431 sig = CURSIG(p); 432 mtx_lock_spin(&sched_lock); 433 PROC_UNLOCK_NOSWITCH(p); 434 if (sig != 0) { 435 if (p->p_wchan) 436 unsleep(p); 437 } else if (p->p_wchan == NULL) 438 catch = 0; 439 } else 440 sig = 0; 441 if (p->p_wchan != NULL) { 442 p->p_stat = SSLEEP; 443 p->p_stats->p_ru.ru_nvcsw++; 444 mi_switch(); 445 } 446 CTR3(KTR_PROC, "msleep resume: proc %p (pid %d, %s)", p, p->p_pid, 447 p->p_comm); 448 KASSERT(p->p_stat == SRUN, ("running but not SRUN")); 449 p->p_sflag &= ~PS_SINTR; 450 if (p->p_sflag & PS_TIMEOUT) { 451 p->p_sflag &= ~PS_TIMEOUT; 452 if (sig == 0) 453 rval = EWOULDBLOCK; 454 } else if (timo) 455 callout_stop(&p->p_slpcallout); 456 mtx_unlock_spin(&sched_lock); 457 458 if (rval == 0 && catch) { 459 PROC_LOCK(p); 460 /* XXX: shouldn't we always be calling CURSIG() */ 461 if (sig != 0 || (sig = CURSIG(p))) { 462 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 463 rval = EINTR; 464 else 465 rval = ERESTART; 466 } 467 PROC_UNLOCK(p); 468 } 469 PICKUP_GIANT(); 470#ifdef KTRACE 471 mtx_lock(&Giant); 472 if (KTRPOINT(p, KTR_CSW)) 473 ktrcsw(p->p_tracep, 0, 0); 474 mtx_unlock(&Giant); 475#endif 476 if (mtx != NULL) { 477 mtx_lock(mtx); 478 WITNESS_RESTORE(&mtx->mtx_object, mtx); 479 } 480 return (rval); 481} 482 483/* 484 * Implement timeout for msleep() 485 * 486 * If process hasn't been awakened (wchan non-zero), 487 * set timeout flag and undo the sleep. If proc 488 * is stopped, just unsleep so it will remain stopped. 489 * MP-safe, called without the Giant mutex. 490 */ 491static void 492endtsleep(arg) 493 void *arg; 494{ 495 register struct proc *p; 496 497 p = (struct proc *)arg; 498 CTR3(KTR_PROC, "endtsleep: proc %p (pid %d, %s)", p, p->p_pid, 499 p->p_comm); 500 mtx_lock_spin(&sched_lock); 501 if (p->p_wchan) { 502 if (p->p_stat == SSLEEP) 503 setrunnable(p); 504 else 505 unsleep(p); 506 p->p_sflag |= PS_TIMEOUT; 507 } 508 mtx_unlock_spin(&sched_lock); 509} 510 511/* 512 * Remove a process from its wait queue 513 */ 514void 515unsleep(p) 516 register struct proc *p; 517{ 518 519 mtx_lock_spin(&sched_lock); 520 if (p->p_wchan) { 521 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_slpq); 522 p->p_wchan = NULL; 523 } 524 mtx_unlock_spin(&sched_lock); 525} 526 527/* 528 * Make all processes sleeping on the specified identifier runnable. 529 */ 530void 531wakeup(ident) 532 register void *ident; 533{ 534 register struct slpquehead *qp; 535 register struct proc *p; 536 537 mtx_lock_spin(&sched_lock); 538 qp = &slpque[LOOKUP(ident)]; 539restart: 540 TAILQ_FOREACH(p, qp, p_slpq) { 541 if (p->p_wchan == ident) { 542 TAILQ_REMOVE(qp, p, p_slpq); 543 p->p_wchan = NULL; 544 if (p->p_stat == SSLEEP) { 545 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 546 CTR3(KTR_PROC, "wakeup: proc %p (pid %d, %s)", 547 p, p->p_pid, p->p_comm); 548 if (p->p_slptime > 1) 549 updatepri(p); 550 p->p_slptime = 0; 551 p->p_stat = SRUN; 552 if (p->p_sflag & PS_INMEM) { 553 setrunqueue(p); 554 maybe_resched(p); 555 } else { 556 p->p_sflag |= PS_SWAPINREQ; 557 wakeup((caddr_t)&proc0); 558 } 559 /* END INLINE EXPANSION */ 560 goto restart; 561 } 562 } 563 } 564 mtx_unlock_spin(&sched_lock); 565} 566 567/* 568 * Make a process sleeping on the specified identifier runnable. 569 * May wake more than one process if a target process is currently 570 * swapped out. 571 */ 572void 573wakeup_one(ident) 574 register void *ident; 575{ 576 register struct slpquehead *qp; 577 register struct proc *p; 578 579 mtx_lock_spin(&sched_lock); 580 qp = &slpque[LOOKUP(ident)]; 581 582 TAILQ_FOREACH(p, qp, p_slpq) { 583 if (p->p_wchan == ident) { 584 TAILQ_REMOVE(qp, p, p_slpq); 585 p->p_wchan = NULL; 586 if (p->p_stat == SSLEEP) { 587 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 588 CTR3(KTR_PROC, "wakeup1: proc %p (pid %d, %s)", 589 p, p->p_pid, p->p_comm); 590 if (p->p_slptime > 1) 591 updatepri(p); 592 p->p_slptime = 0; 593 p->p_stat = SRUN; 594 if (p->p_sflag & PS_INMEM) { 595 setrunqueue(p); 596 maybe_resched(p); 597 break; 598 } else { 599 p->p_sflag |= PS_SWAPINREQ; 600 wakeup((caddr_t)&proc0); 601 } 602 /* END INLINE EXPANSION */ 603 } 604 } 605 } 606 mtx_unlock_spin(&sched_lock); 607} 608 609/* 610 * The machine independent parts of mi_switch(). 611 */ 612void 613mi_switch() 614{ 615 struct timeval new_switchtime; 616 register struct proc *p = curproc; /* XXX */ 617#if 0 618 register struct rlimit *rlim; 619#endif 620 critical_t sched_crit; 621 u_int sched_nest; 622 623 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 624 625 /* 626 * Compute the amount of time during which the current 627 * process was running, and add that to its total so far. 628 */ 629 microuptime(&new_switchtime); 630 if (timevalcmp(&new_switchtime, PCPU_PTR(switchtime), <)) { 631#if 0 632 /* XXX: This doesn't play well with sched_lock right now. */ 633 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 634 PCPU_GET(switchtime.tv_sec), PCPU_GET(switchtime.tv_usec), 635 new_switchtime.tv_sec, new_switchtime.tv_usec); 636#endif 637 new_switchtime = PCPU_GET(switchtime); 638 } else { 639 p->p_runtime += (new_switchtime.tv_usec - PCPU_GET(switchtime.tv_usec)) + 640 (new_switchtime.tv_sec - PCPU_GET(switchtime.tv_sec)) * 641 (int64_t)1000000; 642 } 643 644#if 0 645 /* 646 * Check if the process exceeds its cpu resource allocation. 647 * If over max, kill it. 648 * 649 * XXX drop sched_lock, pickup Giant 650 */ 651 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 652 p->p_runtime > p->p_limit->p_cpulimit) { 653 rlim = &p->p_rlimit[RLIMIT_CPU]; 654 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 655 mtx_unlock_spin(&sched_lock); 656 PROC_LOCK(p); 657 killproc(p, "exceeded maximum CPU limit"); 658 mtx_lock_spin(&sched_lock); 659 PROC_UNLOCK_NOSWITCH(p); 660 } else { 661 mtx_unlock_spin(&sched_lock); 662 PROC_LOCK(p); 663 psignal(p, SIGXCPU); 664 mtx_lock_spin(&sched_lock); 665 PROC_UNLOCK_NOSWITCH(p); 666 if (rlim->rlim_cur < rlim->rlim_max) { 667 /* XXX: we should make a private copy */ 668 rlim->rlim_cur += 5; 669 } 670 } 671 } 672#endif 673 674 /* 675 * Pick a new current process and record its start time. 676 */ 677 cnt.v_swtch++; 678 PCPU_SET(switchtime, new_switchtime); 679 CTR3(KTR_PROC, "mi_switch: old proc %p (pid %d, %s)", p, p->p_pid, 680 p->p_comm); 681 sched_crit = sched_lock.mtx_savecrit; 682 sched_nest = sched_lock.mtx_recurse; 683 p->p_lastcpu = p->p_oncpu; 684 p->p_oncpu = NOCPU; 685 clear_resched(p); 686 cpu_switch(); 687 p->p_oncpu = PCPU_GET(cpuid); 688 sched_lock.mtx_savecrit = sched_crit; 689 sched_lock.mtx_recurse = sched_nest; 690 sched_lock.mtx_lock = (uintptr_t)p; 691 CTR3(KTR_PROC, "mi_switch: new proc %p (pid %d, %s)", p, p->p_pid, 692 p->p_comm); 693 if (PCPU_GET(switchtime.tv_sec) == 0) 694 microuptime(PCPU_PTR(switchtime)); 695 PCPU_SET(switchticks, ticks); 696} 697 698/* 699 * Change process state to be runnable, 700 * placing it on the run queue if it is in memory, 701 * and awakening the swapper if it isn't in memory. 702 */ 703void 704setrunnable(p) 705 register struct proc *p; 706{ 707 708 mtx_lock_spin(&sched_lock); 709 switch (p->p_stat) { 710 case 0: 711 case SRUN: 712 case SZOMB: 713 case SWAIT: 714 default: 715 panic("setrunnable"); 716 case SSTOP: 717 case SSLEEP: /* e.g. when sending signals */ 718 if (p->p_sflag & PS_CVWAITQ) 719 cv_waitq_remove(p); 720 else 721 unsleep(p); 722 break; 723 724 case SIDL: 725 break; 726 } 727 p->p_stat = SRUN; 728 if (p->p_slptime > 1) 729 updatepri(p); 730 p->p_slptime = 0; 731 if ((p->p_sflag & PS_INMEM) == 0) { 732 p->p_sflag |= PS_SWAPINREQ; 733 wakeup((caddr_t)&proc0); 734 } else { 735 setrunqueue(p); 736 maybe_resched(p); 737 } 738 mtx_unlock_spin(&sched_lock); 739} 740 741/* 742 * Compute the priority of a process when running in user mode. 743 * Arrange to reschedule if the resulting priority is better 744 * than that of the current process. 745 */ 746void 747resetpriority(p) 748 register struct proc *p; 749{ 750 register unsigned int newpriority; 751 752 mtx_lock_spin(&sched_lock); 753 if (p->p_pri.pri_class == PRI_TIMESHARE) { 754 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 755 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 756 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 757 PRI_MAX_TIMESHARE); 758 p->p_pri.pri_user = newpriority; 759 } 760 maybe_resched(p); 761 mtx_unlock_spin(&sched_lock); 762} 763 764/* ARGSUSED */ 765static void 766sched_setup(dummy) 767 void *dummy; 768{ 769 770 callout_init(&schedcpu_callout, 1); 771 callout_init(&roundrobin_callout, 0); 772 773 /* Kick off timeout driven events by calling first time. */ 774 roundrobin(NULL); 775 schedcpu(NULL); 776} 777 778/* 779 * We adjust the priority of the current process. The priority of 780 * a process gets worse as it accumulates CPU time. The cpu usage 781 * estimator (p_estcpu) is increased here. resetpriority() will 782 * compute a different priority each time p_estcpu increases by 783 * INVERSE_ESTCPU_WEIGHT 784 * (until MAXPRI is reached). The cpu usage estimator ramps up 785 * quite quickly when the process is running (linearly), and decays 786 * away exponentially, at a rate which is proportionally slower when 787 * the system is busy. The basic principle is that the system will 788 * 90% forget that the process used a lot of CPU time in 5 * loadav 789 * seconds. This causes the system to favor processes which haven't 790 * run much recently, and to round-robin among other processes. 791 */ 792void 793schedclock(p) 794 struct proc *p; 795{ 796 797 p->p_cpticks++; 798 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 799 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 800 resetpriority(p); 801 if (p->p_pri.pri_level >= PUSER) 802 p->p_pri.pri_level = p->p_pri.pri_user; 803 } 804} 805 806/* 807 * General purpose yield system call 808 */ 809int 810yield(struct proc *p, struct yield_args *uap) 811{ 812 813 p->p_retval[0] = 0; 814 815 mtx_lock_spin(&sched_lock); 816 DROP_GIANT_NOSWITCH(); 817 p->p_pri.pri_level = PRI_MAX_TIMESHARE; 818 setrunqueue(p); 819 p->p_stats->p_ru.ru_nvcsw++; 820 mi_switch(); 821 mtx_unlock_spin(&sched_lock); 822 PICKUP_GIANT(); 823 824 return (0); 825} 826