kern_synch.c revision 83787
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 83787 2001-09-21 19:21:18Z jhb $ 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(kg) 111 struct ksegrp *kg; 112{ 113 114 mtx_assert(&sched_lock, MA_OWNED); 115 if (kg->kg_pri.pri_level < curthread->td_ksegrp->kg_pri.pri_level) 116 curthread->td_kse->ke_flags |= KEF_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 struct kse *ke; 249 register struct ksegrp *kg; 250 register int realstathz; 251 int awake; 252 253 realstathz = stathz ? stathz : hz; 254 sx_slock(&allproc_lock); 255 FOREACH_PROC_IN_SYSTEM(p) { 256 mtx_lock_spin(&sched_lock); 257 p->p_swtime++; 258 FOREACH_KSEGRP_IN_PROC(p, kg) { 259 awake = 0; 260 FOREACH_KSE_IN_GROUP(kg, ke) { 261 /* 262 * Increment time in/out of memory and sleep 263 * time (if sleeping). We ignore overflow; 264 * with 16-bit int's (remember them?) 265 * overflow takes 45 days. 266 */ 267 /* XXXKSE */ 268 /* if ((ke->ke_flags & KEF_ONRUNQ) == 0) */ 269 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) { 270 ke->ke_slptime++; 271 } else { 272 ke->ke_slptime = 0; 273 awake = 1; 274 } 275 276 /* 277 * pctcpu is only for ps? 278 * Do it per kse.. and add them up at the end? 279 * XXXKSE 280 */ 281 ke->ke_pctcpu = (ke->ke_pctcpu * ccpu) >> FSHIFT; 282 /* 283 * If the kse has been idle the entire second, 284 * stop recalculating its priority until 285 * it wakes up. 286 */ 287 if (ke->ke_slptime > 1) { 288 continue; 289 } 290 291#if (FSHIFT >= CCPU_SHIFT) 292 ke->ke_pctcpu += (realstathz == 100) ? 293 ((fixpt_t) ke->ke_cpticks) << 294 (FSHIFT - CCPU_SHIFT) : 295 100 * (((fixpt_t) ke->ke_cpticks) << 296 (FSHIFT - CCPU_SHIFT)) / realstathz; 297#else 298 ke->ke_pctcpu += ((FSCALE - ccpu) * 299 (ke->ke_cpticks * FSCALE / realstathz)) >> 300 FSHIFT; 301#endif 302 ke->ke_cpticks = 0; 303 } /* end of kse loop */ 304 if (awake == 0) { 305 kg->kg_slptime++; 306 } else { 307 kg->kg_slptime = 0; 308 } 309 kg->kg_estcpu = decay_cpu(loadfac, kg->kg_estcpu); 310 resetpriority(kg); 311 if (kg->kg_pri.pri_level >= PUSER && 312 (p->p_sflag & PS_INMEM)) { 313 int changedqueue = 314 ((kg->kg_pri.pri_level / RQ_PPQ) != 315 (kg->kg_pri.pri_user / RQ_PPQ)); 316 317 kg->kg_pri.pri_level = kg->kg_pri.pri_user; 318 FOREACH_KSE_IN_GROUP(kg, ke) { 319 if ((ke->ke_oncpu == NOCPU) && /* idle */ 320 (p->p_stat == SRUN) && /* XXXKSE */ 321 changedqueue) { 322 remrunqueue(ke->ke_thread); 323 setrunqueue(ke->ke_thread); 324 } 325 } 326 } 327 } /* end of ksegrp loop */ 328 mtx_unlock_spin(&sched_lock); 329 } /* end of process loop */ 330 sx_sunlock(&allproc_lock); 331 vmmeter(); 332 wakeup((caddr_t)&lbolt); 333 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 334} 335 336/* 337 * Recalculate the priority of a process after it has slept for a while. 338 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 339 * least six times the loadfactor will decay p_estcpu to zero. 340 */ 341void 342updatepri(td) 343 register struct thread *td; 344{ 345 register struct ksegrp *kg; 346 register unsigned int newcpu; 347 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 348 349 if (td == NULL) 350 return; 351 kg = td->td_ksegrp; 352 newcpu = kg->kg_estcpu; 353 if (kg->kg_slptime > 5 * loadfac) 354 kg->kg_estcpu = 0; 355 else { 356 kg->kg_slptime--; /* the first time was done in schedcpu */ 357 while (newcpu && --kg->kg_slptime) 358 newcpu = decay_cpu(loadfac, newcpu); 359 kg->kg_estcpu = newcpu; 360 } 361 resetpriority(td->td_ksegrp); 362} 363 364/* 365 * We're only looking at 7 bits of the address; everything is 366 * aligned to 4, lots of things are aligned to greater powers 367 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 368 */ 369#define TABLESIZE 128 370static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE]; 371#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 372 373void 374sleepinit(void) 375{ 376 int i; 377 378 sched_quantum = hz/10; 379 hogticks = 2 * sched_quantum; 380 for (i = 0; i < TABLESIZE; i++) 381 TAILQ_INIT(&slpque[i]); 382} 383 384/* 385 * General sleep call. Suspends the current process until a wakeup is 386 * performed on the specified identifier. The process will then be made 387 * runnable with the specified priority. Sleeps at most timo/hz seconds 388 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 389 * before and after sleeping, else signals are not checked. Returns 0 if 390 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 391 * signal needs to be delivered, ERESTART is returned if the current system 392 * call should be restarted if possible, and EINTR is returned if the system 393 * call should be interrupted by the signal (return EINTR). 394 * 395 * The mutex argument is exited before the caller is suspended, and 396 * entered before msleep returns. If priority includes the PDROP 397 * flag the mutex is not entered before returning. 398 */ 399int 400msleep(ident, mtx, priority, wmesg, timo) 401 void *ident; 402 struct mtx *mtx; 403 int priority, timo; 404 const char *wmesg; 405{ 406 struct proc *p = curproc; 407 struct thread *td = curthread; 408 int sig, catch = priority & PCATCH; 409 int rval = 0; 410 WITNESS_SAVE_DECL(mtx); 411 412#ifdef KTRACE 413 if (p && KTRPOINT(p, KTR_CSW)) 414 ktrcsw(p->p_tracep, 1, 0); 415#endif 416 WITNESS_SLEEP(0, &mtx->mtx_object); 417 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 418 ("sleeping without a mutex")); 419 mtx_lock_spin(&sched_lock); 420 if (cold || panicstr) { 421 /* 422 * After a panic, or during autoconfiguration, 423 * just give interrupts a chance, then just return; 424 * don't run any other procs or panic below, 425 * in case this is the idle process and already asleep. 426 */ 427 if (mtx != NULL && priority & PDROP) 428 mtx_unlock_flags(mtx, MTX_NOSWITCH); 429 mtx_unlock_spin(&sched_lock); 430 return (0); 431 } 432 433 DROP_GIANT_NOSWITCH(); 434 435 if (mtx != NULL) { 436 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 437 WITNESS_SAVE(&mtx->mtx_object, mtx); 438 mtx_unlock_flags(mtx, MTX_NOSWITCH); 439 if (priority & PDROP) 440 mtx = NULL; 441 } 442 443 KASSERT(p != NULL, ("msleep1")); 444 KASSERT(ident != NULL && td->td_proc->p_stat == SRUN, ("msleep")); 445 446 td->td_wchan = ident; 447 td->td_wmesg = wmesg; 448 td->td_kse->ke_slptime = 0; /* XXXKSE */ 449 td->td_ksegrp->kg_slptime = 0; 450 td->td_ksegrp->kg_pri.pri_level = priority & PRIMASK; 451 CTR5(KTR_PROC, "msleep: thread %p (pid %d, %s) on %s (%p)", 452 td, p->p_pid, p->p_comm, wmesg, ident); 453 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], td, td_slpq); 454 if (timo) 455 callout_reset(&td->td_slpcallout, timo, endtsleep, td); 456 /* 457 * We put ourselves on the sleep queue and start our timeout 458 * before calling CURSIG, as we could stop there, and a wakeup 459 * or a SIGCONT (or both) could occur while we were stopped. 460 * A SIGCONT would cause us to be marked as SSLEEP 461 * without resuming us, thus we must be ready for sleep 462 * when CURSIG is called. If the wakeup happens while we're 463 * stopped, td->td_wchan will be 0 upon return from CURSIG. 464 */ 465 if (catch) { 466 CTR3(KTR_PROC, "msleep caught: proc %p (pid %d, %s)", p, 467 p->p_pid, p->p_comm); 468 td->td_flags |= TDF_SINTR; 469 mtx_unlock_spin(&sched_lock); 470 PROC_LOCK(p); 471 sig = CURSIG(p); 472 mtx_lock_spin(&sched_lock); 473 PROC_UNLOCK_NOSWITCH(p); 474 if (sig != 0) { 475 if (td->td_wchan != NULL) 476 unsleep(td); 477 } else if (td->td_wchan == NULL) 478 catch = 0; 479 } else 480 sig = 0; 481 if (td->td_wchan != NULL) { 482 td->td_proc->p_stat = SSLEEP; 483 p->p_stats->p_ru.ru_nvcsw++; 484 mi_switch(); 485 } 486 CTR3(KTR_PROC, "msleep resume: proc %p (pid %d, %s)", td, p->p_pid, 487 p->p_comm); 488 KASSERT(td->td_proc->p_stat == SRUN, ("running but not SRUN")); 489 td->td_flags &= ~TDF_SINTR; 490 if (td->td_flags & TDF_TIMEOUT) { 491 td->td_flags &= ~TDF_TIMEOUT; 492 if (sig == 0) 493 rval = EWOULDBLOCK; 494 } else if (td->td_flags & TDF_TIMOFAIL) 495 td->td_flags &= ~TDF_TIMOFAIL; 496 else if (timo && callout_stop(&td->td_slpcallout) == 0) { 497 /* 498 * This isn't supposed to be pretty. If we are here, then 499 * the endtsleep() callout is currently executing on another 500 * CPU and is either spinning on the sched_lock or will be 501 * soon. If we don't synchronize here, there is a chance 502 * that this process may msleep() again before the callout 503 * has a chance to run and the callout may end up waking up 504 * the wrong msleep(). Yuck. 505 */ 506 td->td_flags |= TDF_TIMEOUT; 507 p->p_stats->p_ru.ru_nivcsw++; 508 mi_switch(); 509 } 510 mtx_unlock_spin(&sched_lock); 511 512 if (rval == 0 && catch) { 513 PROC_LOCK(p); 514 /* XXX: shouldn't we always be calling CURSIG() */ 515 if (sig != 0 || (sig = CURSIG(p))) { 516 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 517 rval = EINTR; 518 else 519 rval = ERESTART; 520 } 521 PROC_UNLOCK(p); 522 } 523 PICKUP_GIANT(); 524#ifdef KTRACE 525 mtx_lock(&Giant); 526 if (KTRPOINT(p, KTR_CSW)) 527 ktrcsw(p->p_tracep, 0, 0); 528 mtx_unlock(&Giant); 529#endif 530 if (mtx != NULL) { 531 mtx_lock(mtx); 532 WITNESS_RESTORE(&mtx->mtx_object, mtx); 533 } 534 return (rval); 535} 536 537/* 538 * Implement timeout for msleep() 539 * 540 * If process hasn't been awakened (wchan non-zero), 541 * set timeout flag and undo the sleep. If proc 542 * is stopped, just unsleep so it will remain stopped. 543 * MP-safe, called without the Giant mutex. 544 */ 545static void 546endtsleep(arg) 547 void *arg; 548{ 549 register struct thread *td = arg; 550 551 CTR3(KTR_PROC, "endtsleep: thread %p (pid %d, %s)", td, td->td_proc->p_pid, 552 td->td_proc->p_comm); 553 mtx_lock_spin(&sched_lock); 554 /* 555 * This is the other half of the synchronization with msleep() 556 * described above. If the PS_TIMEOUT flag is set, we lost the 557 * race and just need to put the process back on the runqueue. 558 */ 559 if ((td->td_flags & TDF_TIMEOUT) != 0) { 560 td->td_flags &= ~TDF_TIMEOUT; 561 setrunqueue(td); 562 } else if (td->td_wchan != NULL) { 563 if (td->td_proc->p_stat == SSLEEP) /* XXXKSE */ 564 setrunnable(td); 565 else 566 unsleep(td); 567 td->td_flags |= TDF_TIMEOUT; 568 } else { 569 td->td_flags |= TDF_TIMOFAIL; 570 } 571 mtx_unlock_spin(&sched_lock); 572} 573 574/* 575 * Remove a process from its wait queue 576 */ 577void 578unsleep(struct thread *td) 579{ 580 581 mtx_lock_spin(&sched_lock); 582 if (td->td_wchan != NULL) { 583 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_slpq); 584 td->td_wchan = NULL; 585 } 586 mtx_unlock_spin(&sched_lock); 587} 588 589/* 590 * Make all processes sleeping on the specified identifier runnable. 591 */ 592void 593wakeup(ident) 594 register void *ident; 595{ 596 register struct slpquehead *qp; 597 register struct thread *td; 598 struct proc *p; 599 600 mtx_lock_spin(&sched_lock); 601 qp = &slpque[LOOKUP(ident)]; 602restart: 603 TAILQ_FOREACH(td, qp, td_slpq) { 604 p = td->td_proc; 605 if (td->td_wchan == ident) { 606 TAILQ_REMOVE(qp, td, td_slpq); 607 td->td_wchan = NULL; 608 if (td->td_proc->p_stat == SSLEEP) { 609 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 610 CTR3(KTR_PROC, "wakeup: thread %p (pid %d, %s)", 611 td, p->p_pid, p->p_comm); 612 if (td->td_ksegrp->kg_slptime > 1) 613 updatepri(td); 614 td->td_ksegrp->kg_slptime = 0; 615 td->td_kse->ke_slptime = 0; 616 td->td_proc->p_stat = SRUN; 617 if (p->p_sflag & PS_INMEM) { 618 setrunqueue(td); 619 maybe_resched(td->td_ksegrp); 620 } else { 621 p->p_sflag |= PS_SWAPINREQ; 622 wakeup((caddr_t)&proc0); 623 } 624 /* END INLINE EXPANSION */ 625 goto restart; 626 } 627 } 628 } 629 mtx_unlock_spin(&sched_lock); 630} 631 632/* 633 * Make a process sleeping on the specified identifier runnable. 634 * May wake more than one process if a target process is currently 635 * swapped out. 636 */ 637void 638wakeup_one(ident) 639 register void *ident; 640{ 641 register struct slpquehead *qp; 642 register struct thread *td; 643 register struct proc *p; 644 645 mtx_lock_spin(&sched_lock); 646 qp = &slpque[LOOKUP(ident)]; 647 648 TAILQ_FOREACH(td, qp, td_slpq) { 649 p = td->td_proc; 650 if (td->td_wchan == ident) { 651 TAILQ_REMOVE(qp, td, td_slpq); 652 td->td_wchan = NULL; 653 if (td->td_proc->p_stat == SSLEEP) { 654 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 655 CTR3(KTR_PROC, "wakeup1: proc %p (pid %d, %s)", 656 p, p->p_pid, p->p_comm); 657 if (td->td_ksegrp->kg_slptime > 1) 658 updatepri(td); 659 td->td_ksegrp->kg_slptime = 0; 660 td->td_kse->ke_slptime = 0; 661 td->td_proc->p_stat = SRUN; 662 if (p->p_sflag & PS_INMEM) { 663 setrunqueue(td); 664 maybe_resched(td->td_ksegrp); 665 break; 666 } else { 667 p->p_sflag |= PS_SWAPINREQ; 668 wakeup((caddr_t)&proc0); 669 } 670 /* END INLINE EXPANSION */ 671 } 672 } 673 } 674 mtx_unlock_spin(&sched_lock); 675} 676 677/* 678 * The machine independent parts of mi_switch(). 679 */ 680void 681mi_switch() 682{ 683 struct timeval new_switchtime; 684 struct thread *td = curthread; /* XXX */ 685 register struct proc *p = td->td_proc; /* XXX */ 686#if 0 687 register struct rlimit *rlim; 688#endif 689 critical_t sched_crit; 690 u_int sched_nest; 691 692 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 693 694 /* 695 * Compute the amount of time during which the current 696 * process was running, and add that to its total so far. 697 */ 698 microuptime(&new_switchtime); 699 if (timevalcmp(&new_switchtime, PCPU_PTR(switchtime), <)) { 700#if 0 701 /* XXX: This doesn't play well with sched_lock right now. */ 702 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 703 PCPU_GET(switchtime.tv_sec), PCPU_GET(switchtime.tv_usec), 704 new_switchtime.tv_sec, new_switchtime.tv_usec); 705#endif 706 new_switchtime = PCPU_GET(switchtime); 707 } else { 708 p->p_runtime += (new_switchtime.tv_usec - PCPU_GET(switchtime.tv_usec)) + 709 (new_switchtime.tv_sec - PCPU_GET(switchtime.tv_sec)) * 710 (int64_t)1000000; 711 } 712 713#ifdef DDB 714 /* 715 * Don't perform context switches from the debugger. 716 */ 717 if (db_active) { 718 mtx_unlock_spin(&sched_lock); 719 db_error("Context switches not allowed in the debugger."); 720 } 721#endif 722 723#if 0 724 /* 725 * Check if the process exceeds its cpu resource allocation. 726 * If over max, kill it. 727 * 728 * XXX drop sched_lock, pickup Giant 729 */ 730 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 731 p->p_runtime > p->p_limit->p_cpulimit) { 732 rlim = &p->p_rlimit[RLIMIT_CPU]; 733 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 734 mtx_unlock_spin(&sched_lock); 735 PROC_LOCK(p); 736 killproc(p, "exceeded maximum CPU limit"); 737 mtx_lock_spin(&sched_lock); 738 PROC_UNLOCK_NOSWITCH(p); 739 } else { 740 mtx_unlock_spin(&sched_lock); 741 PROC_LOCK(p); 742 psignal(p, SIGXCPU); 743 mtx_lock_spin(&sched_lock); 744 PROC_UNLOCK_NOSWITCH(p); 745 if (rlim->rlim_cur < rlim->rlim_max) { 746 /* XXX: we should make a private copy */ 747 rlim->rlim_cur += 5; 748 } 749 } 750 } 751#endif 752 753 /* 754 * Pick a new current process and record its start time. 755 */ 756 cnt.v_swtch++; 757 PCPU_SET(switchtime, new_switchtime); 758 CTR3(KTR_PROC, "mi_switch: old proc %p (pid %d, %s)", p, p->p_pid, 759 p->p_comm); 760 sched_crit = sched_lock.mtx_savecrit; 761 sched_nest = sched_lock.mtx_recurse; 762 td->td_lastcpu = td->td_kse->ke_oncpu; 763 td->td_kse->ke_oncpu = NOCPU; 764 td->td_kse->ke_flags &= ~KEF_NEEDRESCHED; 765 cpu_switch(); 766 td->td_kse->ke_oncpu = PCPU_GET(cpuid); 767 sched_lock.mtx_savecrit = sched_crit; 768 sched_lock.mtx_recurse = sched_nest; 769 sched_lock.mtx_lock = (uintptr_t)td; 770 CTR3(KTR_PROC, "mi_switch: new proc %p (pid %d, %s)", p, p->p_pid, 771 p->p_comm); 772 if (PCPU_GET(switchtime.tv_sec) == 0) 773 microuptime(PCPU_PTR(switchtime)); 774 PCPU_SET(switchticks, ticks); 775} 776 777/* 778 * Change process state to be runnable, 779 * placing it on the run queue if it is in memory, 780 * and awakening the swapper if it isn't in memory. 781 */ 782void 783setrunnable(struct thread *td) 784{ 785 struct proc *p = td->td_proc; 786 787 mtx_lock_spin(&sched_lock); 788 switch (p->p_stat) { 789 case SZOMB: /* not a thread flag XXXKSE */ 790 panic("setrunnable(1)"); 791 } 792 switch (td->td_proc->p_stat) { 793 case 0: 794 case SRUN: 795 case SWAIT: 796 default: 797 panic("setrunnable(2)"); 798 case SSTOP: 799 case SSLEEP: /* e.g. when sending signals */ 800 if (td->td_flags & TDF_CVWAITQ) 801 cv_waitq_remove(td); 802 else 803 unsleep(td); 804 break; 805 806 case SIDL: 807 break; 808 } 809 td->td_proc->p_stat = SRUN; 810 if (td->td_ksegrp->kg_slptime > 1) 811 updatepri(td); 812 td->td_ksegrp->kg_slptime = 0; 813 td->td_kse->ke_slptime = 0; 814 if ((p->p_sflag & PS_INMEM) == 0) { 815 p->p_sflag |= PS_SWAPINREQ; 816 wakeup((caddr_t)&proc0); 817 } else { 818 setrunqueue(td); 819 maybe_resched(td->td_ksegrp); 820 } 821 mtx_unlock_spin(&sched_lock); 822} 823 824/* 825 * Compute the priority of a process when running in user mode. 826 * Arrange to reschedule if the resulting priority is better 827 * than that of the current process. 828 */ 829void 830resetpriority(kg) 831 register struct ksegrp *kg; 832{ 833 register unsigned int newpriority; 834 835 mtx_lock_spin(&sched_lock); 836 if (kg->kg_pri.pri_class == PRI_TIMESHARE) { 837 newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT + 838 NICE_WEIGHT * (kg->kg_nice - PRIO_MIN); 839 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 840 PRI_MAX_TIMESHARE); 841 kg->kg_pri.pri_user = newpriority; 842 } 843 maybe_resched(kg); 844 mtx_unlock_spin(&sched_lock); 845} 846 847/* ARGSUSED */ 848static void 849sched_setup(dummy) 850 void *dummy; 851{ 852 853 callout_init(&schedcpu_callout, 1); 854 callout_init(&roundrobin_callout, 0); 855 856 /* Kick off timeout driven events by calling first time. */ 857 roundrobin(NULL); 858 schedcpu(NULL); 859} 860 861/* 862 * We adjust the priority of the current process. The priority of 863 * a process gets worse as it accumulates CPU time. The cpu usage 864 * estimator (p_estcpu) is increased here. resetpriority() will 865 * compute a different priority each time p_estcpu increases by 866 * INVERSE_ESTCPU_WEIGHT 867 * (until MAXPRI is reached). The cpu usage estimator ramps up 868 * quite quickly when the process is running (linearly), and decays 869 * away exponentially, at a rate which is proportionally slower when 870 * the system is busy. The basic principle is that the system will 871 * 90% forget that the process used a lot of CPU time in 5 * loadav 872 * seconds. This causes the system to favor processes which haven't 873 * run much recently, and to round-robin among other processes. 874 */ 875void 876schedclock(td) 877 struct thread *td; 878{ 879 struct kse *ke = td->td_kse; 880 struct ksegrp *kg = td->td_ksegrp; 881 882 if (td) { 883 ke->ke_cpticks++; 884 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1); 885 if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 886 resetpriority(td->td_ksegrp); 887 if (kg->kg_pri.pri_level >= PUSER) 888 kg->kg_pri.pri_level = kg->kg_pri.pri_user; 889 } 890 } else { 891 panic("schedclock"); 892 } 893} 894 895/* 896 * General purpose yield system call 897 */ 898int 899yield(struct thread *td, struct yield_args *uap) 900{ 901 struct ksegrp *kg = td->td_ksegrp; 902 903 mtx_assert(&Giant, MA_NOTOWNED); 904 mtx_lock_spin(&sched_lock); 905 kg->kg_pri.pri_level = PRI_MAX_TIMESHARE; 906 setrunqueue(td); 907 kg->kg_proc->p_stats->p_ru.ru_nvcsw++; 908 mi_switch(); 909 mtx_unlock_spin(&sched_lock); 910 td->td_retval[0] = 0; 911 912 return (0); 913} 914 915