kern_synch.c revision 85227
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 85227 2001-10-20 13:10:43Z iedowse $ 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#ifdef DDB 61#include <ddb/ddb.h> 62#endif 63#ifdef KTRACE 64#include <sys/uio.h> 65#include <sys/ktrace.h> 66#endif 67 68#include <machine/cpu.h> 69 70static void sched_setup __P((void *dummy)); 71SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 72 73int hogticks; 74int lbolt; 75int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 76 77static struct callout schedcpu_callout; 78static struct callout roundrobin_callout; 79 80struct loadavg averunnable = 81 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 82/* 83 * Constants for averages over 1, 5, and 15 minutes 84 * when sampling at 5 second intervals. 85 */ 86static fixpt_t cexp[3] = { 87 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 88 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 89 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 90}; 91 92static void endtsleep __P((void *)); 93static void loadav __P((struct loadavg *)); 94static void roundrobin __P((void *arg)); 95static void schedcpu __P((void *arg)); 96 97static int 98sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 99{ 100 int error, new_val; 101 102 new_val = sched_quantum * tick; 103 error = sysctl_handle_int(oidp, &new_val, 0, req); 104 if (error != 0 || req->newptr == NULL) 105 return (error); 106 if (new_val < tick) 107 return (EINVAL); 108 sched_quantum = new_val / tick; 109 hogticks = 2 * sched_quantum; 110 return (0); 111} 112 113SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 114 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 115 116/* 117 * Arrange to reschedule if necessary, taking the priorities and 118 * schedulers into account. 119 */ 120void 121maybe_resched(kg) 122 struct ksegrp *kg; 123{ 124 125 mtx_assert(&sched_lock, MA_OWNED); 126 if (kg->kg_pri.pri_level < curthread->td_ksegrp->kg_pri.pri_level) 127 curthread->td_kse->ke_flags |= KEF_NEEDRESCHED; 128} 129 130int 131roundrobin_interval(void) 132{ 133 return (sched_quantum); 134} 135 136/* 137 * Force switch among equal priority processes every 100ms. 138 * We don't actually need to force a context switch of the current process. 139 * The act of firing the event triggers a context switch to softclock() and 140 * then switching back out again which is equivalent to a preemption, thus 141 * no further work is needed on the local CPU. 142 */ 143/* ARGSUSED */ 144static void 145roundrobin(arg) 146 void *arg; 147{ 148 149#ifdef SMP 150 mtx_lock_spin(&sched_lock); 151 forward_roundrobin(); 152 mtx_unlock_spin(&sched_lock); 153#endif 154 155 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 156} 157 158/* 159 * Constants for digital decay and forget: 160 * 90% of (p_estcpu) usage in 5 * loadav time 161 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 162 * Note that, as ps(1) mentions, this can let percentages 163 * total over 100% (I've seen 137.9% for 3 processes). 164 * 165 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 166 * 167 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 168 * That is, the system wants to compute a value of decay such 169 * that the following for loop: 170 * for (i = 0; i < (5 * loadavg); i++) 171 * p_estcpu *= decay; 172 * will compute 173 * p_estcpu *= 0.1; 174 * for all values of loadavg: 175 * 176 * Mathematically this loop can be expressed by saying: 177 * decay ** (5 * loadavg) ~= .1 178 * 179 * The system computes decay as: 180 * decay = (2 * loadavg) / (2 * loadavg + 1) 181 * 182 * We wish to prove that the system's computation of decay 183 * will always fulfill the equation: 184 * decay ** (5 * loadavg) ~= .1 185 * 186 * If we compute b as: 187 * b = 2 * loadavg 188 * then 189 * decay = b / (b + 1) 190 * 191 * We now need to prove two things: 192 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 193 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 194 * 195 * Facts: 196 * For x close to zero, exp(x) =~ 1 + x, since 197 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 198 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 199 * For x close to zero, ln(1+x) =~ x, since 200 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 201 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 202 * ln(.1) =~ -2.30 203 * 204 * Proof of (1): 205 * Solve (factor)**(power) =~ .1 given power (5*loadav): 206 * solving for factor, 207 * ln(factor) =~ (-2.30/5*loadav), or 208 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 209 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 210 * 211 * Proof of (2): 212 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 213 * solving for power, 214 * power*ln(b/(b+1)) =~ -2.30, or 215 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 216 * 217 * Actual power values for the implemented algorithm are as follows: 218 * loadav: 1 2 3 4 219 * power: 5.68 10.32 14.94 19.55 220 */ 221 222/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 223#define loadfactor(loadav) (2 * (loadav)) 224#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 225 226/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 227static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 228SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 229 230/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 231static int fscale __unused = FSCALE; 232SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 233 234/* 235 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 236 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 237 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 238 * 239 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 240 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 241 * 242 * If you don't want to bother with the faster/more-accurate formula, you 243 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 244 * (more general) method of calculating the %age of CPU used by a process. 245 */ 246#define CCPU_SHIFT 11 247 248/* 249 * Recompute process priorities, every hz ticks. 250 * MP-safe, called without the Giant mutex. 251 */ 252/* ARGSUSED */ 253static void 254schedcpu(arg) 255 void *arg; 256{ 257 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 258 register struct proc *p; 259 register struct kse *ke; 260 register struct ksegrp *kg; 261 register int realstathz; 262 int awake; 263 264 realstathz = stathz ? stathz : hz; 265 sx_slock(&allproc_lock); 266 FOREACH_PROC_IN_SYSTEM(p) { 267 mtx_lock_spin(&sched_lock); 268 p->p_swtime++; 269 FOREACH_KSEGRP_IN_PROC(p, kg) { 270 awake = 0; 271 FOREACH_KSE_IN_GROUP(kg, ke) { 272 /* 273 * Increment time in/out of memory and sleep 274 * time (if sleeping). We ignore overflow; 275 * with 16-bit int's (remember them?) 276 * overflow takes 45 days. 277 */ 278 /* XXXKSE */ 279 /* if ((ke->ke_flags & KEF_ONRUNQ) == 0) */ 280 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) { 281 ke->ke_slptime++; 282 } else { 283 ke->ke_slptime = 0; 284 awake = 1; 285 } 286 287 /* 288 * pctcpu is only for ps? 289 * Do it per kse.. and add them up at the end? 290 * XXXKSE 291 */ 292 ke->ke_pctcpu = (ke->ke_pctcpu * ccpu) >> FSHIFT; 293 /* 294 * If the kse has been idle the entire second, 295 * stop recalculating its priority until 296 * it wakes up. 297 */ 298 if (ke->ke_slptime > 1) { 299 continue; 300 } 301 302#if (FSHIFT >= CCPU_SHIFT) 303 ke->ke_pctcpu += (realstathz == 100) ? 304 ((fixpt_t) ke->ke_cpticks) << 305 (FSHIFT - CCPU_SHIFT) : 306 100 * (((fixpt_t) ke->ke_cpticks) << 307 (FSHIFT - CCPU_SHIFT)) / realstathz; 308#else 309 ke->ke_pctcpu += ((FSCALE - ccpu) * 310 (ke->ke_cpticks * FSCALE / realstathz)) >> 311 FSHIFT; 312#endif 313 ke->ke_cpticks = 0; 314 } /* end of kse loop */ 315 if (awake == 0) { 316 kg->kg_slptime++; 317 } else { 318 kg->kg_slptime = 0; 319 } 320 kg->kg_estcpu = decay_cpu(loadfac, kg->kg_estcpu); 321 resetpriority(kg); 322 if (kg->kg_pri.pri_level >= PUSER && 323 (p->p_sflag & PS_INMEM)) { 324 int changedqueue = 325 ((kg->kg_pri.pri_level / RQ_PPQ) != 326 (kg->kg_pri.pri_user / RQ_PPQ)); 327 328 kg->kg_pri.pri_level = kg->kg_pri.pri_user; 329 FOREACH_KSE_IN_GROUP(kg, ke) { 330 if ((ke->ke_oncpu == NOCPU) && /* idle */ 331 (p->p_stat == SRUN) && /* XXXKSE */ 332 changedqueue) { 333 remrunqueue(ke->ke_thread); 334 setrunqueue(ke->ke_thread); 335 } 336 } 337 } 338 } /* end of ksegrp loop */ 339 mtx_unlock_spin(&sched_lock); 340 } /* end of process loop */ 341 sx_sunlock(&allproc_lock); 342 if (time_second % 5 == 0) 343 loadav(&averunnable); 344 wakeup((caddr_t)&lbolt); 345 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 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 */ 353void 354updatepri(td) 355 register struct thread *td; 356{ 357 register struct ksegrp *kg; 358 register unsigned int newcpu; 359 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 360 361 if (td == NULL) 362 return; 363 kg = td->td_ksegrp; 364 newcpu = kg->kg_estcpu; 365 if (kg->kg_slptime > 5 * loadfac) 366 kg->kg_estcpu = 0; 367 else { 368 kg->kg_slptime--; /* the first time was done in schedcpu */ 369 while (newcpu && --kg->kg_slptime) 370 newcpu = decay_cpu(loadfac, newcpu); 371 kg->kg_estcpu = newcpu; 372 } 373 resetpriority(td->td_ksegrp); 374} 375 376/* 377 * We're only looking at 7 bits of the address; everything is 378 * aligned to 4, lots of things are aligned to greater powers 379 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 380 */ 381#define TABLESIZE 128 382static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE]; 383#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 384 385void 386sleepinit(void) 387{ 388 int i; 389 390 sched_quantum = hz/10; 391 hogticks = 2 * sched_quantum; 392 for (i = 0; i < TABLESIZE; i++) 393 TAILQ_INIT(&slpque[i]); 394} 395 396/* 397 * General sleep call. Suspends the current process until a wakeup is 398 * performed on the specified identifier. The process will then be made 399 * runnable with the specified priority. Sleeps at most timo/hz seconds 400 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 401 * before and after sleeping, else signals are not checked. Returns 0 if 402 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 403 * signal needs to be delivered, ERESTART is returned if the current system 404 * call should be restarted if possible, and EINTR is returned if the system 405 * call should be interrupted by the signal (return EINTR). 406 * 407 * The mutex argument is exited before the caller is suspended, and 408 * entered before msleep returns. If priority includes the PDROP 409 * flag the mutex is not entered before returning. 410 */ 411int 412msleep(ident, mtx, priority, wmesg, timo) 413 void *ident; 414 struct mtx *mtx; 415 int priority, timo; 416 const char *wmesg; 417{ 418 struct proc *p = curproc; 419 struct thread *td = curthread; 420 int sig, catch = priority & PCATCH; 421 int rval = 0; 422 WITNESS_SAVE_DECL(mtx); 423 424#ifdef KTRACE 425 if (p && KTRPOINT(p, KTR_CSW)) 426 ktrcsw(p->p_tracep, 1, 0); 427#endif 428 WITNESS_SLEEP(0, &mtx->mtx_object); 429 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 430 ("sleeping without a mutex")); 431 mtx_lock_spin(&sched_lock); 432 if (cold || panicstr) { 433 /* 434 * After a panic, or during autoconfiguration, 435 * just give interrupts a chance, then just return; 436 * don't run any other procs or panic below, 437 * in case this is the idle process and already asleep. 438 */ 439 if (mtx != NULL && priority & PDROP) 440 mtx_unlock_flags(mtx, MTX_NOSWITCH); 441 mtx_unlock_spin(&sched_lock); 442 return (0); 443 } 444 445 DROP_GIANT_NOSWITCH(); 446 447 if (mtx != NULL) { 448 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 449 WITNESS_SAVE(&mtx->mtx_object, mtx); 450 mtx_unlock_flags(mtx, MTX_NOSWITCH); 451 if (priority & PDROP) 452 mtx = NULL; 453 } 454 455 KASSERT(p != NULL, ("msleep1")); 456 KASSERT(ident != NULL && td->td_proc->p_stat == SRUN, ("msleep")); 457 458 td->td_wchan = ident; 459 td->td_wmesg = wmesg; 460 td->td_kse->ke_slptime = 0; /* XXXKSE */ 461 td->td_ksegrp->kg_slptime = 0; 462 td->td_ksegrp->kg_pri.pri_level = priority & PRIMASK; 463 CTR5(KTR_PROC, "msleep: thread %p (pid %d, %s) on %s (%p)", 464 td, p->p_pid, p->p_comm, wmesg, ident); 465 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], td, td_slpq); 466 if (timo) 467 callout_reset(&td->td_slpcallout, timo, endtsleep, td); 468 /* 469 * We put ourselves on the sleep queue and start our timeout 470 * before calling CURSIG, as we could stop there, and a wakeup 471 * or a SIGCONT (or both) could occur while we were stopped. 472 * A SIGCONT would cause us to be marked as SSLEEP 473 * without resuming us, thus we must be ready for sleep 474 * when CURSIG is called. If the wakeup happens while we're 475 * stopped, td->td_wchan will be 0 upon return from CURSIG. 476 */ 477 if (catch) { 478 CTR3(KTR_PROC, "msleep caught: proc %p (pid %d, %s)", p, 479 p->p_pid, p->p_comm); 480 td->td_flags |= TDF_SINTR; 481 mtx_unlock_spin(&sched_lock); 482 PROC_LOCK(p); 483 sig = CURSIG(p); 484 mtx_lock_spin(&sched_lock); 485 PROC_UNLOCK_NOSWITCH(p); 486 if (sig != 0) { 487 if (td->td_wchan != NULL) 488 unsleep(td); 489 } else if (td->td_wchan == NULL) 490 catch = 0; 491 } else 492 sig = 0; 493 if (td->td_wchan != NULL) { 494 td->td_proc->p_stat = SSLEEP; 495 p->p_stats->p_ru.ru_nvcsw++; 496 mi_switch(); 497 } 498 CTR3(KTR_PROC, "msleep resume: proc %p (pid %d, %s)", td, p->p_pid, 499 p->p_comm); 500 KASSERT(td->td_proc->p_stat == SRUN, ("running but not SRUN")); 501 td->td_flags &= ~TDF_SINTR; 502 if (td->td_flags & TDF_TIMEOUT) { 503 td->td_flags &= ~TDF_TIMEOUT; 504 if (sig == 0) 505 rval = EWOULDBLOCK; 506 } else if (td->td_flags & TDF_TIMOFAIL) 507 td->td_flags &= ~TDF_TIMOFAIL; 508 else if (timo && callout_stop(&td->td_slpcallout) == 0) { 509 /* 510 * This isn't supposed to be pretty. If we are here, then 511 * the endtsleep() callout is currently executing on another 512 * CPU and is either spinning on the sched_lock or will be 513 * soon. If we don't synchronize here, there is a chance 514 * that this process may msleep() again before the callout 515 * has a chance to run and the callout may end up waking up 516 * the wrong msleep(). Yuck. 517 */ 518 td->td_flags |= TDF_TIMEOUT; 519 p->p_stats->p_ru.ru_nivcsw++; 520 mi_switch(); 521 } 522 mtx_unlock_spin(&sched_lock); 523 524 if (rval == 0 && catch) { 525 PROC_LOCK(p); 526 /* XXX: shouldn't we always be calling CURSIG() */ 527 if (sig != 0 || (sig = CURSIG(p))) { 528 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 529 rval = EINTR; 530 else 531 rval = ERESTART; 532 } 533 PROC_UNLOCK(p); 534 } 535 PICKUP_GIANT(); 536#ifdef KTRACE 537 mtx_lock(&Giant); 538 if (KTRPOINT(p, KTR_CSW)) 539 ktrcsw(p->p_tracep, 0, 0); 540 mtx_unlock(&Giant); 541#endif 542 if (mtx != NULL) { 543 mtx_lock(mtx); 544 WITNESS_RESTORE(&mtx->mtx_object, mtx); 545 } 546 return (rval); 547} 548 549/* 550 * Implement timeout for msleep() 551 * 552 * If process hasn't been awakened (wchan non-zero), 553 * set timeout flag and undo the sleep. If proc 554 * is stopped, just unsleep so it will remain stopped. 555 * MP-safe, called without the Giant mutex. 556 */ 557static void 558endtsleep(arg) 559 void *arg; 560{ 561 register struct thread *td = arg; 562 563 CTR3(KTR_PROC, "endtsleep: thread %p (pid %d, %s)", td, td->td_proc->p_pid, 564 td->td_proc->p_comm); 565 mtx_lock_spin(&sched_lock); 566 /* 567 * This is the other half of the synchronization with msleep() 568 * described above. If the PS_TIMEOUT flag is set, we lost the 569 * race and just need to put the process back on the runqueue. 570 */ 571 if ((td->td_flags & TDF_TIMEOUT) != 0) { 572 td->td_flags &= ~TDF_TIMEOUT; 573 setrunqueue(td); 574 } else if (td->td_wchan != NULL) { 575 if (td->td_proc->p_stat == SSLEEP) /* XXXKSE */ 576 setrunnable(td); 577 else 578 unsleep(td); 579 td->td_flags |= TDF_TIMEOUT; 580 } else { 581 td->td_flags |= TDF_TIMOFAIL; 582 } 583 mtx_unlock_spin(&sched_lock); 584} 585 586/* 587 * Remove a process from its wait queue 588 */ 589void 590unsleep(struct thread *td) 591{ 592 593 mtx_lock_spin(&sched_lock); 594 if (td->td_wchan != NULL) { 595 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_slpq); 596 td->td_wchan = NULL; 597 } 598 mtx_unlock_spin(&sched_lock); 599} 600 601/* 602 * Make all processes sleeping on the specified identifier runnable. 603 */ 604void 605wakeup(ident) 606 register void *ident; 607{ 608 register struct slpquehead *qp; 609 register struct thread *td; 610 struct proc *p; 611 612 mtx_lock_spin(&sched_lock); 613 qp = &slpque[LOOKUP(ident)]; 614restart: 615 TAILQ_FOREACH(td, qp, td_slpq) { 616 p = td->td_proc; 617 if (td->td_wchan == ident) { 618 TAILQ_REMOVE(qp, td, td_slpq); 619 td->td_wchan = NULL; 620 if (td->td_proc->p_stat == SSLEEP) { 621 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 622 CTR3(KTR_PROC, "wakeup: thread %p (pid %d, %s)", 623 td, p->p_pid, p->p_comm); 624 if (td->td_ksegrp->kg_slptime > 1) 625 updatepri(td); 626 td->td_ksegrp->kg_slptime = 0; 627 td->td_kse->ke_slptime = 0; 628 td->td_proc->p_stat = SRUN; 629 if (p->p_sflag & PS_INMEM) { 630 setrunqueue(td); 631 maybe_resched(td->td_ksegrp); 632 } else { 633 p->p_sflag |= PS_SWAPINREQ; 634 wakeup((caddr_t)&proc0); 635 } 636 /* END INLINE EXPANSION */ 637 goto restart; 638 } 639 } 640 } 641 mtx_unlock_spin(&sched_lock); 642} 643 644/* 645 * Make a process sleeping on the specified identifier runnable. 646 * May wake more than one process if a target process is currently 647 * swapped out. 648 */ 649void 650wakeup_one(ident) 651 register void *ident; 652{ 653 register struct slpquehead *qp; 654 register struct thread *td; 655 register struct proc *p; 656 657 mtx_lock_spin(&sched_lock); 658 qp = &slpque[LOOKUP(ident)]; 659 660 TAILQ_FOREACH(td, qp, td_slpq) { 661 p = td->td_proc; 662 if (td->td_wchan == ident) { 663 TAILQ_REMOVE(qp, td, td_slpq); 664 td->td_wchan = NULL; 665 if (td->td_proc->p_stat == SSLEEP) { 666 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 667 CTR3(KTR_PROC, "wakeup1: proc %p (pid %d, %s)", 668 p, p->p_pid, p->p_comm); 669 if (td->td_ksegrp->kg_slptime > 1) 670 updatepri(td); 671 td->td_ksegrp->kg_slptime = 0; 672 td->td_kse->ke_slptime = 0; 673 td->td_proc->p_stat = SRUN; 674 if (p->p_sflag & PS_INMEM) { 675 setrunqueue(td); 676 maybe_resched(td->td_ksegrp); 677 break; 678 } else { 679 p->p_sflag |= PS_SWAPINREQ; 680 wakeup((caddr_t)&proc0); 681 } 682 /* END INLINE EXPANSION */ 683 } 684 } 685 } 686 mtx_unlock_spin(&sched_lock); 687} 688 689/* 690 * The machine independent parts of mi_switch(). 691 */ 692void 693mi_switch() 694{ 695 struct timeval new_switchtime; 696 struct thread *td = curthread; /* XXX */ 697 register struct proc *p = td->td_proc; /* XXX */ 698#if 0 699 register struct rlimit *rlim; 700#endif 701 critical_t sched_crit; 702 u_int sched_nest; 703 704 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 705 706 /* 707 * Compute the amount of time during which the current 708 * process was running, and add that to its total so far. 709 */ 710 microuptime(&new_switchtime); 711 if (timevalcmp(&new_switchtime, PCPU_PTR(switchtime), <)) { 712#if 0 713 /* XXX: This doesn't play well with sched_lock right now. */ 714 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 715 PCPU_GET(switchtime.tv_sec), PCPU_GET(switchtime.tv_usec), 716 new_switchtime.tv_sec, new_switchtime.tv_usec); 717#endif 718 new_switchtime = PCPU_GET(switchtime); 719 } else { 720 p->p_runtime += (new_switchtime.tv_usec - PCPU_GET(switchtime.tv_usec)) + 721 (new_switchtime.tv_sec - PCPU_GET(switchtime.tv_sec)) * 722 (int64_t)1000000; 723 } 724 725#ifdef DDB 726 /* 727 * Don't perform context switches from the debugger. 728 */ 729 if (db_active) { 730 mtx_unlock_spin(&sched_lock); 731 db_error("Context switches not allowed in the debugger."); 732 } 733#endif 734 735#if 0 736 /* 737 * Check if the process exceeds its cpu resource allocation. 738 * If over max, kill it. 739 * 740 * XXX drop sched_lock, pickup Giant 741 */ 742 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 743 p->p_runtime > p->p_limit->p_cpulimit) { 744 rlim = &p->p_rlimit[RLIMIT_CPU]; 745 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 746 mtx_unlock_spin(&sched_lock); 747 PROC_LOCK(p); 748 killproc(p, "exceeded maximum CPU limit"); 749 mtx_lock_spin(&sched_lock); 750 PROC_UNLOCK_NOSWITCH(p); 751 } else { 752 mtx_unlock_spin(&sched_lock); 753 PROC_LOCK(p); 754 psignal(p, SIGXCPU); 755 mtx_lock_spin(&sched_lock); 756 PROC_UNLOCK_NOSWITCH(p); 757 if (rlim->rlim_cur < rlim->rlim_max) { 758 /* XXX: we should make a private copy */ 759 rlim->rlim_cur += 5; 760 } 761 } 762 } 763#endif 764 765 /* 766 * Pick a new current process and record its start time. 767 */ 768 cnt.v_swtch++; 769 PCPU_SET(switchtime, new_switchtime); 770 CTR3(KTR_PROC, "mi_switch: old proc %p (pid %d, %s)", p, p->p_pid, 771 p->p_comm); 772 sched_crit = sched_lock.mtx_savecrit; 773 sched_nest = sched_lock.mtx_recurse; 774 td->td_lastcpu = td->td_kse->ke_oncpu; 775 td->td_kse->ke_oncpu = NOCPU; 776 td->td_kse->ke_flags &= ~KEF_NEEDRESCHED; 777 cpu_switch(); 778 td->td_kse->ke_oncpu = PCPU_GET(cpuid); 779 sched_lock.mtx_savecrit = sched_crit; 780 sched_lock.mtx_recurse = sched_nest; 781 sched_lock.mtx_lock = (uintptr_t)td; 782 CTR3(KTR_PROC, "mi_switch: new proc %p (pid %d, %s)", p, p->p_pid, 783 p->p_comm); 784 if (PCPU_GET(switchtime.tv_sec) == 0) 785 microuptime(PCPU_PTR(switchtime)); 786 PCPU_SET(switchticks, ticks); 787} 788 789/* 790 * Change process state to be runnable, 791 * placing it on the run queue if it is in memory, 792 * and awakening the swapper if it isn't in memory. 793 */ 794void 795setrunnable(struct thread *td) 796{ 797 struct proc *p = td->td_proc; 798 799 mtx_lock_spin(&sched_lock); 800 switch (p->p_stat) { 801 case SZOMB: /* not a thread flag XXXKSE */ 802 panic("setrunnable(1)"); 803 } 804 switch (td->td_proc->p_stat) { 805 case 0: 806 case SRUN: 807 case SWAIT: 808 default: 809 panic("setrunnable(2)"); 810 case SSTOP: 811 case SSLEEP: /* e.g. when sending signals */ 812 if (td->td_flags & TDF_CVWAITQ) 813 cv_waitq_remove(td); 814 else 815 unsleep(td); 816 break; 817 818 case SIDL: 819 break; 820 } 821 td->td_proc->p_stat = SRUN; 822 if (td->td_ksegrp->kg_slptime > 1) 823 updatepri(td); 824 td->td_ksegrp->kg_slptime = 0; 825 td->td_kse->ke_slptime = 0; 826 if ((p->p_sflag & PS_INMEM) == 0) { 827 p->p_sflag |= PS_SWAPINREQ; 828 wakeup((caddr_t)&proc0); 829 } else { 830 setrunqueue(td); 831 maybe_resched(td->td_ksegrp); 832 } 833 mtx_unlock_spin(&sched_lock); 834} 835 836/* 837 * Compute the priority of a process when running in user mode. 838 * Arrange to reschedule if the resulting priority is better 839 * than that of the current process. 840 */ 841void 842resetpriority(kg) 843 register struct ksegrp *kg; 844{ 845 register unsigned int newpriority; 846 847 mtx_lock_spin(&sched_lock); 848 if (kg->kg_pri.pri_class == PRI_TIMESHARE) { 849 newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT + 850 NICE_WEIGHT * (kg->kg_nice - PRIO_MIN); 851 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 852 PRI_MAX_TIMESHARE); 853 kg->kg_pri.pri_user = newpriority; 854 } 855 maybe_resched(kg); 856 mtx_unlock_spin(&sched_lock); 857} 858 859/* 860 * Compute a tenex style load average of a quantity on 861 * 1, 5 and 15 minute intervals. 862 * XXXKSE Needs complete rewrite when correct info is available. 863 * Completely Bogus.. only works with 1:1 (but compiles ok now :-) 864 */ 865static void 866loadav(struct loadavg *avg) 867{ 868 int i, nrun; 869 struct proc *p; 870 struct ksegrp *kg; 871 872 sx_slock(&allproc_lock); 873 nrun = 0; 874 FOREACH_PROC_IN_SYSTEM(p) { 875 FOREACH_KSEGRP_IN_PROC(p, kg) { 876 switch (p->p_stat) { 877 case SRUN: 878 if ((p->p_flag & P_NOLOAD) != 0) 879 goto nextproc; 880 /* FALLTHROUGH */ 881 case SIDL: 882 nrun++; 883 } 884nextproc: 885 } 886 } 887 sx_sunlock(&allproc_lock); 888 for (i = 0; i < 3; i++) 889 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 890 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 891} 892 893/* ARGSUSED */ 894static void 895sched_setup(dummy) 896 void *dummy; 897{ 898 899 callout_init(&schedcpu_callout, 1); 900 callout_init(&roundrobin_callout, 0); 901 902 /* Kick off timeout driven events by calling first time. */ 903 roundrobin(NULL); 904 schedcpu(NULL); 905} 906 907/* 908 * We adjust the priority of the current process. The priority of 909 * a process gets worse as it accumulates CPU time. The cpu usage 910 * estimator (p_estcpu) is increased here. resetpriority() will 911 * compute a different priority each time p_estcpu increases by 912 * INVERSE_ESTCPU_WEIGHT 913 * (until MAXPRI is reached). The cpu usage estimator ramps up 914 * quite quickly when the process is running (linearly), and decays 915 * away exponentially, at a rate which is proportionally slower when 916 * the system is busy. The basic principle is that the system will 917 * 90% forget that the process used a lot of CPU time in 5 * loadav 918 * seconds. This causes the system to favor processes which haven't 919 * run much recently, and to round-robin among other processes. 920 */ 921void 922schedclock(td) 923 struct thread *td; 924{ 925 struct kse *ke = td->td_kse; 926 struct ksegrp *kg = td->td_ksegrp; 927 928 if (td) { 929 ke->ke_cpticks++; 930 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1); 931 if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 932 resetpriority(td->td_ksegrp); 933 if (kg->kg_pri.pri_level >= PUSER) 934 kg->kg_pri.pri_level = kg->kg_pri.pri_user; 935 } 936 } else { 937 panic("schedclock"); 938 } 939} 940 941/* 942 * General purpose yield system call 943 */ 944int 945yield(struct thread *td, struct yield_args *uap) 946{ 947 struct ksegrp *kg = td->td_ksegrp; 948 949 mtx_assert(&Giant, MA_NOTOWNED); 950 mtx_lock_spin(&sched_lock); 951 kg->kg_pri.pri_level = PRI_MAX_TIMESHARE; 952 setrunqueue(td); 953 kg->kg_proc->p_stats->p_ru.ru_nvcsw++; 954 mi_switch(); 955 mtx_unlock_spin(&sched_lock); 956 td->td_retval[0] = 0; 957 958 return (0); 959} 960 961