kern_synch.c revision 82711
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 82711 2001-09-01 03:54:09Z dillon $ 40 */ 41 42#include "opt_ddb.h" 43#include "opt_ktrace.h" 44 45#include <sys/param.h> 46#include <sys/systm.h> 47#include <sys/condvar.h> 48#include <sys/kernel.h> 49#include <sys/ktr.h> 50#include <sys/lock.h> 51#include <sys/mutex.h> 52#include <sys/proc.h> 53#include <sys/resourcevar.h> 54#include <sys/signalvar.h> 55#include <sys/smp.h> 56#include <sys/sx.h> 57#include <sys/sysctl.h> 58#include <sys/sysproto.h> 59#include <sys/vmmeter.h> 60#include <vm/vm.h> 61#include <vm/vm_extern.h> 62#ifdef DDB 63#include <ddb/ddb.h> 64#endif 65#ifdef KTRACE 66#include <sys/uio.h> 67#include <sys/ktrace.h> 68#endif 69 70#include <machine/cpu.h> 71 72static void sched_setup __P((void *dummy)); 73SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 74 75int hogticks; 76int lbolt; 77int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 78 79static struct callout schedcpu_callout; 80static struct callout roundrobin_callout; 81 82static void endtsleep __P((void *)); 83static void roundrobin __P((void *arg)); 84static void schedcpu __P((void *arg)); 85 86static int 87sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 88{ 89 int error, new_val; 90 91 new_val = sched_quantum * tick; 92 error = sysctl_handle_int(oidp, &new_val, 0, req); 93 if (error != 0 || req->newptr == NULL) 94 return (error); 95 if (new_val < tick) 96 return (EINVAL); 97 sched_quantum = new_val / tick; 98 hogticks = 2 * sched_quantum; 99 return (0); 100} 101 102SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 103 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 104 105/* 106 * Arrange to reschedule if necessary, taking the priorities and 107 * schedulers into account. 108 */ 109void 110maybe_resched(p) 111 struct proc *p; 112{ 113 114 mtx_assert(&sched_lock, MA_OWNED); 115 if (p->p_pri.pri_level < curproc->p_pri.pri_level) 116 curproc->p_sflag |= PS_NEEDRESCHED; 117} 118 119int 120roundrobin_interval(void) 121{ 122 return (sched_quantum); 123} 124 125/* 126 * Force switch among equal priority processes every 100ms. 127 * We don't actually need to force a context switch of the current process. 128 * The act of firing the event triggers a context switch to softclock() and 129 * then switching back out again which is equivalent to a preemption, thus 130 * no further work is needed on the local CPU. 131 */ 132/* ARGSUSED */ 133static void 134roundrobin(arg) 135 void *arg; 136{ 137 138#ifdef SMP 139 mtx_lock_spin(&sched_lock); 140 forward_roundrobin(); 141 mtx_unlock_spin(&sched_lock); 142#endif 143 144 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 145} 146 147/* 148 * Constants for digital decay and forget: 149 * 90% of (p_estcpu) usage in 5 * loadav time 150 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 151 * Note that, as ps(1) mentions, this can let percentages 152 * total over 100% (I've seen 137.9% for 3 processes). 153 * 154 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 155 * 156 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 157 * That is, the system wants to compute a value of decay such 158 * that the following for loop: 159 * for (i = 0; i < (5 * loadavg); i++) 160 * p_estcpu *= decay; 161 * will compute 162 * p_estcpu *= 0.1; 163 * for all values of loadavg: 164 * 165 * Mathematically this loop can be expressed by saying: 166 * decay ** (5 * loadavg) ~= .1 167 * 168 * The system computes decay as: 169 * decay = (2 * loadavg) / (2 * loadavg + 1) 170 * 171 * We wish to prove that the system's computation of decay 172 * will always fulfill the equation: 173 * decay ** (5 * loadavg) ~= .1 174 * 175 * If we compute b as: 176 * b = 2 * loadavg 177 * then 178 * decay = b / (b + 1) 179 * 180 * We now need to prove two things: 181 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 182 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 183 * 184 * Facts: 185 * For x close to zero, exp(x) =~ 1 + x, since 186 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 187 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 188 * For x close to zero, ln(1+x) =~ x, since 189 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 190 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 191 * ln(.1) =~ -2.30 192 * 193 * Proof of (1): 194 * Solve (factor)**(power) =~ .1 given power (5*loadav): 195 * solving for factor, 196 * ln(factor) =~ (-2.30/5*loadav), or 197 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 198 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 199 * 200 * Proof of (2): 201 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 202 * solving for power, 203 * power*ln(b/(b+1)) =~ -2.30, or 204 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 205 * 206 * Actual power values for the implemented algorithm are as follows: 207 * loadav: 1 2 3 4 208 * power: 5.68 10.32 14.94 19.55 209 */ 210 211/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 212#define loadfactor(loadav) (2 * (loadav)) 213#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 214 215/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 216static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 217SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 218 219/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 220static int fscale __unused = FSCALE; 221SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 222 223/* 224 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 225 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 226 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 227 * 228 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 229 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 230 * 231 * If you don't want to bother with the faster/more-accurate formula, you 232 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 233 * (more general) method of calculating the %age of CPU used by a process. 234 */ 235#define CCPU_SHIFT 11 236 237/* 238 * Recompute process priorities, every hz ticks. 239 * MP-safe, called without the Giant mutex. 240 */ 241/* ARGSUSED */ 242static void 243schedcpu(arg) 244 void *arg; 245{ 246 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 247 register struct proc *p; 248 register int realstathz; 249 250 realstathz = stathz ? stathz : hz; 251 sx_slock(&allproc_lock); 252 LIST_FOREACH(p, &allproc, p_list) { 253 /* 254 * Increment time in/out of memory and sleep time 255 * (if sleeping). We ignore overflow; with 16-bit int's 256 * (remember them?) overflow takes 45 days. 257 */ 258 mtx_lock_spin(&sched_lock); 259 p->p_swtime++; 260 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 261 p->p_slptime++; 262 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 263 /* 264 * If the process has slept the entire second, 265 * stop recalculating its priority until it wakes up. 266 */ 267 if (p->p_slptime > 1) { 268 mtx_unlock_spin(&sched_lock); 269 continue; 270 } 271 272 /* 273 * p_pctcpu is only for ps. 274 */ 275#if (FSHIFT >= CCPU_SHIFT) 276 p->p_pctcpu += (realstathz == 100)? 277 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 278 100 * (((fixpt_t) p->p_cpticks) 279 << (FSHIFT - CCPU_SHIFT)) / realstathz; 280#else 281 p->p_pctcpu += ((FSCALE - ccpu) * 282 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT; 283#endif 284 p->p_cpticks = 0; 285 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu); 286 resetpriority(p); 287 if (p->p_pri.pri_level >= PUSER) { 288 if (p->p_oncpu == NOCPU && /* idle */ 289 p->p_stat == SRUN && 290 (p->p_sflag & PS_INMEM) && 291 (p->p_pri.pri_level / RQ_PPQ) != 292 (p->p_pri.pri_user / RQ_PPQ)) { 293 remrunqueue(p); 294 p->p_pri.pri_level = p->p_pri.pri_user; 295 setrunqueue(p); 296 } else 297 p->p_pri.pri_level = p->p_pri.pri_user; 298 } 299 mtx_unlock_spin(&sched_lock); 300 } 301 sx_sunlock(&allproc_lock); 302 vmmeter(); 303 wakeup((caddr_t)&lbolt); 304 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 305} 306 307/* 308 * Recalculate the priority of a process after it has slept for a while. 309 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 310 * least six times the loadfactor will decay p_estcpu to zero. 311 */ 312void 313updatepri(p) 314 register struct proc *p; 315{ 316 register unsigned int newcpu = p->p_estcpu; 317 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 318 319 if (p->p_slptime > 5 * loadfac) 320 p->p_estcpu = 0; 321 else { 322 p->p_slptime--; /* the first time was done in schedcpu */ 323 while (newcpu && --p->p_slptime) 324 newcpu = decay_cpu(loadfac, newcpu); 325 p->p_estcpu = newcpu; 326 } 327 resetpriority(p); 328} 329 330/* 331 * We're only looking at 7 bits of the address; everything is 332 * aligned to 4, lots of things are aligned to greater powers 333 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 334 */ 335#define TABLESIZE 128 336static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE]; 337#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 338 339void 340sleepinit(void) 341{ 342 int i; 343 344 sched_quantum = hz/10; 345 hogticks = 2 * sched_quantum; 346 for (i = 0; i < TABLESIZE; i++) 347 TAILQ_INIT(&slpque[i]); 348} 349 350/* 351 * General sleep call. Suspends the current process until a wakeup is 352 * performed on the specified identifier. The process will then be made 353 * runnable with the specified priority. Sleeps at most timo/hz seconds 354 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 355 * before and after sleeping, else signals are not checked. Returns 0 if 356 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 357 * signal needs to be delivered, ERESTART is returned if the current system 358 * call should be restarted if possible, and EINTR is returned if the system 359 * call should be interrupted by the signal (return EINTR). 360 * 361 * The mutex argument is exited before the caller is suspended, and 362 * entered before msleep returns. If priority includes the PDROP 363 * flag the mutex is not entered before returning. 364 */ 365int 366msleep(ident, mtx, priority, wmesg, timo) 367 void *ident; 368 struct mtx *mtx; 369 int priority, timo; 370 const char *wmesg; 371{ 372 struct proc *p = curproc; 373 int sig, catch = priority & PCATCH; 374 int rval = 0; 375 WITNESS_SAVE_DECL(mtx); 376 377#ifdef KTRACE 378 if (p && KTRPOINT(p, KTR_CSW)) 379 ktrcsw(p->p_tracep, 1, 0); 380#endif 381 WITNESS_SLEEP(0, &mtx->mtx_object); 382 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 383 ("sleeping without a mutex")); 384 mtx_lock_spin(&sched_lock); 385 if (cold || panicstr) { 386 /* 387 * After a panic, or during autoconfiguration, 388 * just give interrupts a chance, then just return; 389 * don't run any other procs or panic below, 390 * in case this is the idle process and already asleep. 391 */ 392 if (mtx != NULL && priority & PDROP) 393 mtx_unlock_flags(mtx, MTX_NOSWITCH); 394 mtx_unlock_spin(&sched_lock); 395 return (0); 396 } 397 398 DROP_GIANT_NOSWITCH(); 399 400 if (mtx != NULL) { 401 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 402 WITNESS_SAVE(&mtx->mtx_object, mtx); 403 mtx_unlock_flags(mtx, MTX_NOSWITCH); 404 if (priority & PDROP) 405 mtx = NULL; 406 } 407 408 KASSERT(p != NULL, ("msleep1")); 409 KASSERT(ident != NULL && p->p_stat == SRUN, ("msleep")); 410 411 p->p_wchan = ident; 412 p->p_wmesg = wmesg; 413 p->p_slptime = 0; 414 p->p_pri.pri_level = priority & PRIMASK; 415 CTR5(KTR_PROC, "msleep: proc %p (pid %d, %s) on %s (%p)", p, p->p_pid, 416 p->p_comm, wmesg, ident); 417 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 418 if (timo) 419 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 420 /* 421 * We put ourselves on the sleep queue and start our timeout 422 * before calling CURSIG, as we could stop there, and a wakeup 423 * or a SIGCONT (or both) could occur while we were stopped. 424 * A SIGCONT would cause us to be marked as SSLEEP 425 * without resuming us, thus we must be ready for sleep 426 * when CURSIG is called. If the wakeup happens while we're 427 * stopped, p->p_wchan will be 0 upon return from CURSIG. 428 */ 429 if (catch) { 430 CTR3(KTR_PROC, "msleep caught: proc %p (pid %d, %s)", p, 431 p->p_pid, p->p_comm); 432 p->p_sflag |= PS_SINTR; 433 mtx_unlock_spin(&sched_lock); 434 PROC_LOCK(p); 435 sig = CURSIG(p); 436 mtx_lock_spin(&sched_lock); 437 PROC_UNLOCK_NOSWITCH(p); 438 if (sig != 0) { 439 if (p->p_wchan != NULL) 440 unsleep(p); 441 } else if (p->p_wchan == NULL) 442 catch = 0; 443 } else 444 sig = 0; 445 if (p->p_wchan != NULL) { 446 p->p_stat = SSLEEP; 447 p->p_stats->p_ru.ru_nvcsw++; 448 mi_switch(); 449 } 450 CTR3(KTR_PROC, "msleep resume: proc %p (pid %d, %s)", p, p->p_pid, 451 p->p_comm); 452 KASSERT(p->p_stat == SRUN, ("running but not SRUN")); 453 p->p_sflag &= ~PS_SINTR; 454 if (p->p_sflag & PS_TIMEOUT) { 455 p->p_sflag &= ~PS_TIMEOUT; 456 if (sig == 0) 457 rval = EWOULDBLOCK; 458 } else if (p->p_sflag & PS_TIMOFAIL) 459 p->p_sflag &= ~PS_TIMOFAIL; 460 else if (timo && callout_stop(&p->p_slpcallout) == 0) { 461 /* 462 * This isn't supposed to be pretty. If we are here, then 463 * the endtsleep() callout is currently executing on another 464 * CPU and is either spinning on the sched_lock or will be 465 * soon. If we don't synchronize here, there is a chance 466 * that this process may msleep() again before the callout 467 * has a chance to run and the callout may end up waking up 468 * the wrong msleep(). Yuck. 469 */ 470 p->p_sflag |= PS_TIMEOUT; 471 p->p_stats->p_ru.ru_nivcsw++; 472 mi_switch(); 473 } 474 mtx_unlock_spin(&sched_lock); 475 476 if (rval == 0 && catch) { 477 PROC_LOCK(p); 478 /* XXX: shouldn't we always be calling CURSIG() */ 479 if (sig != 0 || (sig = CURSIG(p))) { 480 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 481 rval = EINTR; 482 else 483 rval = ERESTART; 484 } 485 PROC_UNLOCK(p); 486 } 487 PICKUP_GIANT(); 488#ifdef KTRACE 489 mtx_lock(&Giant); 490 if (KTRPOINT(p, KTR_CSW)) 491 ktrcsw(p->p_tracep, 0, 0); 492 mtx_unlock(&Giant); 493#endif 494 if (mtx != NULL) { 495 mtx_lock(mtx); 496 WITNESS_RESTORE(&mtx->mtx_object, mtx); 497 } 498 return (rval); 499} 500 501/* 502 * Implement timeout for msleep() 503 * 504 * If process hasn't been awakened (wchan non-zero), 505 * set timeout flag and undo the sleep. If proc 506 * is stopped, just unsleep so it will remain stopped. 507 * MP-safe, called without the Giant mutex. 508 */ 509static void 510endtsleep(arg) 511 void *arg; 512{ 513 register struct proc *p; 514 515 p = (struct proc *)arg; 516 CTR3(KTR_PROC, "endtsleep: proc %p (pid %d, %s)", p, p->p_pid, 517 p->p_comm); 518 mtx_lock_spin(&sched_lock); 519 /* 520 * This is the other half of the synchronization with msleep() 521 * described above. If the PS_TIMEOUT flag is set, we lost the 522 * race and just need to put the process back on the runqueue. 523 */ 524 if ((p->p_sflag & PS_TIMEOUT) != 0) { 525 p->p_sflag &= ~PS_TIMEOUT; 526 setrunqueue(p); 527 } else if (p->p_wchan != NULL) { 528 if (p->p_stat == SSLEEP) 529 setrunnable(p); 530 else 531 unsleep(p); 532 p->p_sflag |= PS_TIMEOUT; 533 } else 534 p->p_sflag |= PS_TIMOFAIL; 535 mtx_unlock_spin(&sched_lock); 536} 537 538/* 539 * Remove a process from its wait queue 540 */ 541void 542unsleep(p) 543 register struct proc *p; 544{ 545 546 mtx_lock_spin(&sched_lock); 547 if (p->p_wchan != NULL) { 548 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_slpq); 549 p->p_wchan = NULL; 550 } 551 mtx_unlock_spin(&sched_lock); 552} 553 554/* 555 * Make all processes sleeping on the specified identifier runnable. 556 */ 557void 558wakeup(ident) 559 register void *ident; 560{ 561 register struct slpquehead *qp; 562 register struct proc *p; 563 564 mtx_lock_spin(&sched_lock); 565 qp = &slpque[LOOKUP(ident)]; 566restart: 567 TAILQ_FOREACH(p, qp, p_slpq) { 568 if (p->p_wchan == ident) { 569 TAILQ_REMOVE(qp, p, p_slpq); 570 p->p_wchan = NULL; 571 if (p->p_stat == SSLEEP) { 572 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 573 CTR3(KTR_PROC, "wakeup: proc %p (pid %d, %s)", 574 p, p->p_pid, p->p_comm); 575 if (p->p_slptime > 1) 576 updatepri(p); 577 p->p_slptime = 0; 578 p->p_stat = SRUN; 579 if (p->p_sflag & PS_INMEM) { 580 setrunqueue(p); 581 maybe_resched(p); 582 } else { 583 p->p_sflag |= PS_SWAPINREQ; 584 wakeup((caddr_t)&proc0); 585 } 586 /* END INLINE EXPANSION */ 587 goto restart; 588 } 589 } 590 } 591 mtx_unlock_spin(&sched_lock); 592} 593 594/* 595 * Make a process sleeping on the specified identifier runnable. 596 * May wake more than one process if a target process is currently 597 * swapped out. 598 */ 599void 600wakeup_one(ident) 601 register void *ident; 602{ 603 register struct slpquehead *qp; 604 register struct proc *p; 605 606 mtx_lock_spin(&sched_lock); 607 qp = &slpque[LOOKUP(ident)]; 608 609 TAILQ_FOREACH(p, qp, p_slpq) { 610 if (p->p_wchan == ident) { 611 TAILQ_REMOVE(qp, p, p_slpq); 612 p->p_wchan = NULL; 613 if (p->p_stat == SSLEEP) { 614 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 615 CTR3(KTR_PROC, "wakeup1: proc %p (pid %d, %s)", 616 p, p->p_pid, p->p_comm); 617 if (p->p_slptime > 1) 618 updatepri(p); 619 p->p_slptime = 0; 620 p->p_stat = SRUN; 621 if (p->p_sflag & PS_INMEM) { 622 setrunqueue(p); 623 maybe_resched(p); 624 break; 625 } else { 626 p->p_sflag |= PS_SWAPINREQ; 627 wakeup((caddr_t)&proc0); 628 } 629 /* END INLINE EXPANSION */ 630 } 631 } 632 } 633 mtx_unlock_spin(&sched_lock); 634} 635 636/* 637 * The machine independent parts of mi_switch(). 638 */ 639void 640mi_switch() 641{ 642 struct timeval new_switchtime; 643 register struct proc *p = curproc; /* XXX */ 644#if 0 645 register struct rlimit *rlim; 646#endif 647 critical_t sched_crit; 648 u_int sched_nest; 649 650 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 651 652 /* 653 * Compute the amount of time during which the current 654 * process was running, and add that to its total so far. 655 */ 656 microuptime(&new_switchtime); 657 if (timevalcmp(&new_switchtime, PCPU_PTR(switchtime), <)) { 658#if 0 659 /* XXX: This doesn't play well with sched_lock right now. */ 660 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 661 PCPU_GET(switchtime.tv_sec), PCPU_GET(switchtime.tv_usec), 662 new_switchtime.tv_sec, new_switchtime.tv_usec); 663#endif 664 new_switchtime = PCPU_GET(switchtime); 665 } else { 666 p->p_runtime += (new_switchtime.tv_usec - PCPU_GET(switchtime.tv_usec)) + 667 (new_switchtime.tv_sec - PCPU_GET(switchtime.tv_sec)) * 668 (int64_t)1000000; 669 } 670 671#ifdef DDB 672 /* 673 * Don't perform context switches from the debugger. 674 */ 675 if (db_active) { 676 mtx_unlock_spin(&sched_lock); 677 db_error("Context switches not allowed in the debugger."); 678 } 679#endif 680 681#if 0 682 /* 683 * Check if the process exceeds its cpu resource allocation. 684 * If over max, kill it. 685 * 686 * XXX drop sched_lock, pickup Giant 687 */ 688 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 689 p->p_runtime > p->p_limit->p_cpulimit) { 690 rlim = &p->p_rlimit[RLIMIT_CPU]; 691 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 692 mtx_unlock_spin(&sched_lock); 693 PROC_LOCK(p); 694 killproc(p, "exceeded maximum CPU limit"); 695 mtx_lock_spin(&sched_lock); 696 PROC_UNLOCK_NOSWITCH(p); 697 } else { 698 mtx_unlock_spin(&sched_lock); 699 PROC_LOCK(p); 700 psignal(p, SIGXCPU); 701 mtx_lock_spin(&sched_lock); 702 PROC_UNLOCK_NOSWITCH(p); 703 if (rlim->rlim_cur < rlim->rlim_max) { 704 /* XXX: we should make a private copy */ 705 rlim->rlim_cur += 5; 706 } 707 } 708 } 709#endif 710 711 /* 712 * Pick a new current process and record its start time. 713 */ 714 cnt.v_swtch++; 715 PCPU_SET(switchtime, new_switchtime); 716 CTR3(KTR_PROC, "mi_switch: old proc %p (pid %d, %s)", p, p->p_pid, 717 p->p_comm); 718 sched_crit = sched_lock.mtx_savecrit; 719 sched_nest = sched_lock.mtx_recurse; 720 p->p_lastcpu = p->p_oncpu; 721 p->p_oncpu = NOCPU; 722 p->p_sflag &= ~PS_NEEDRESCHED; 723 cpu_switch(); 724 p->p_oncpu = PCPU_GET(cpuid); 725 sched_lock.mtx_savecrit = sched_crit; 726 sched_lock.mtx_recurse = sched_nest; 727 sched_lock.mtx_lock = (uintptr_t)p; 728 CTR3(KTR_PROC, "mi_switch: new proc %p (pid %d, %s)", p, p->p_pid, 729 p->p_comm); 730 if (PCPU_GET(switchtime.tv_sec) == 0) 731 microuptime(PCPU_PTR(switchtime)); 732 PCPU_SET(switchticks, ticks); 733} 734 735/* 736 * Change process state to be runnable, 737 * placing it on the run queue if it is in memory, 738 * and awakening the swapper if it isn't in memory. 739 */ 740void 741setrunnable(p) 742 register struct proc *p; 743{ 744 745 mtx_lock_spin(&sched_lock); 746 switch (p->p_stat) { 747 case 0: 748 case SRUN: 749 case SZOMB: 750 case SWAIT: 751 default: 752 panic("setrunnable"); 753 case SSTOP: 754 case SSLEEP: /* e.g. when sending signals */ 755 if (p->p_sflag & PS_CVWAITQ) 756 cv_waitq_remove(p); 757 else 758 unsleep(p); 759 break; 760 761 case SIDL: 762 break; 763 } 764 p->p_stat = SRUN; 765 if (p->p_slptime > 1) 766 updatepri(p); 767 p->p_slptime = 0; 768 if ((p->p_sflag & PS_INMEM) == 0) { 769 p->p_sflag |= PS_SWAPINREQ; 770 wakeup((caddr_t)&proc0); 771 } else { 772 setrunqueue(p); 773 maybe_resched(p); 774 } 775 mtx_unlock_spin(&sched_lock); 776} 777 778/* 779 * Compute the priority of a process when running in user mode. 780 * Arrange to reschedule if the resulting priority is better 781 * than that of the current process. 782 */ 783void 784resetpriority(p) 785 register struct proc *p; 786{ 787 register unsigned int newpriority; 788 789 mtx_lock_spin(&sched_lock); 790 if (p->p_pri.pri_class == PRI_TIMESHARE) { 791 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 792 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 793 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 794 PRI_MAX_TIMESHARE); 795 p->p_pri.pri_user = newpriority; 796 } 797 maybe_resched(p); 798 mtx_unlock_spin(&sched_lock); 799} 800 801/* ARGSUSED */ 802static void 803sched_setup(dummy) 804 void *dummy; 805{ 806 807 callout_init(&schedcpu_callout, 1); 808 callout_init(&roundrobin_callout, 0); 809 810 /* Kick off timeout driven events by calling first time. */ 811 roundrobin(NULL); 812 schedcpu(NULL); 813} 814 815/* 816 * We adjust the priority of the current process. The priority of 817 * a process gets worse as it accumulates CPU time. The cpu usage 818 * estimator (p_estcpu) is increased here. resetpriority() will 819 * compute a different priority each time p_estcpu increases by 820 * INVERSE_ESTCPU_WEIGHT 821 * (until MAXPRI is reached). The cpu usage estimator ramps up 822 * quite quickly when the process is running (linearly), and decays 823 * away exponentially, at a rate which is proportionally slower when 824 * the system is busy. The basic principle is that the system will 825 * 90% forget that the process used a lot of CPU time in 5 * loadav 826 * seconds. This causes the system to favor processes which haven't 827 * run much recently, and to round-robin among other processes. 828 */ 829void 830schedclock(p) 831 struct proc *p; 832{ 833 834 p->p_cpticks++; 835 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 836 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 837 resetpriority(p); 838 if (p->p_pri.pri_level >= PUSER) 839 p->p_pri.pri_level = p->p_pri.pri_user; 840 } 841} 842 843/* 844 * General purpose yield system call 845 */ 846int 847yield(struct proc *p, struct yield_args *uap) 848{ 849 p->p_retval[0] = 0; 850 851 mtx_lock_spin(&sched_lock); 852 mtx_assert(&Giant, MA_NOTOWNED); 853#if 0 854 DROP_GIANT_NOSWITCH(); 855#endif 856 p->p_pri.pri_level = PRI_MAX_TIMESHARE; 857 setrunqueue(p); 858 p->p_stats->p_ru.ru_nvcsw++; 859 mi_switch(); 860 mtx_unlock_spin(&sched_lock); 861#if 0 862 PICKUP_GIANT(); 863#endif 864 865 return (0); 866} 867 868