kern_synch.c revision 111883
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 111883 2003-03-04 21:03:05Z 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/sched.h> 55#include <sys/signalvar.h> 56#include <sys/smp.h> 57#include <sys/sx.h> 58#include <sys/sysctl.h> 59#include <sys/sysproto.h> 60#include <sys/vmmeter.h> 61#ifdef DDB 62#include <ddb/ddb.h> 63#endif 64#ifdef KTRACE 65#include <sys/uio.h> 66#include <sys/ktrace.h> 67#endif 68 69#include <machine/cpu.h> 70 71static void sched_setup(void *dummy); 72SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 73 74int hogticks; 75int lbolt; 76 77static struct callout loadav_callout; 78static struct callout lbolt_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 92/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 93static int fscale __unused = FSCALE; 94SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 95 96static void endtsleep(void *); 97static void loadav(void *arg); 98static void lboltcb(void *arg); 99 100/* 101 * We're only looking at 7 bits of the address; everything is 102 * aligned to 4, lots of things are aligned to greater powers 103 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 104 */ 105#define TABLESIZE 128 106static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE]; 107#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 108 109void 110sleepinit(void) 111{ 112 int i; 113 114 hogticks = (hz / 10) * 2; /* Default only. */ 115 for (i = 0; i < TABLESIZE; i++) 116 TAILQ_INIT(&slpque[i]); 117} 118 119/* 120 * General sleep call. Suspends the current process until a wakeup is 121 * performed on the specified identifier. The process will then be made 122 * runnable with the specified priority. Sleeps at most timo/hz seconds 123 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 124 * before and after sleeping, else signals are not checked. Returns 0 if 125 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 126 * signal needs to be delivered, ERESTART is returned if the current system 127 * call should be restarted if possible, and EINTR is returned if the system 128 * call should be interrupted by the signal (return EINTR). 129 * 130 * The mutex argument is exited before the caller is suspended, and 131 * entered before msleep returns. If priority includes the PDROP 132 * flag the mutex is not entered before returning. 133 */ 134 135int 136msleep(ident, mtx, priority, wmesg, timo) 137 void *ident; 138 struct mtx *mtx; 139 int priority, timo; 140 const char *wmesg; 141{ 142 struct thread *td = curthread; 143 struct proc *p = td->td_proc; 144 int sig, catch = priority & PCATCH; 145 int rval = 0; 146 WITNESS_SAVE_DECL(mtx); 147 148#ifdef KTRACE 149 if (KTRPOINT(td, KTR_CSW)) 150 ktrcsw(1, 0); 151#endif 152 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, &mtx->mtx_object, 153 "Sleeping on \"%s\"", wmesg); 154 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 155 ("sleeping without a mutex")); 156 /* 157 * If we are capable of async syscalls and there isn't already 158 * another one ready to return, start a new thread 159 * and queue it as ready to run. Note that there is danger here 160 * because we need to make sure that we don't sleep allocating 161 * the thread (recursion here might be bad). 162 * Hence the TDF_INMSLEEP flag. 163 */ 164 if (p->p_flag & P_THREADED) { 165 /* 166 * Just don't bother if we are exiting 167 * and not the exiting thread or thread was marked as 168 * interrupted. 169 */ 170 if (catch && 171 (((p->p_flag & P_WEXIT) && (p->p_singlethread != td)) || 172 (td->td_flags & TDF_INTERRUPT))) { 173 td->td_flags &= ~TDF_INTERRUPT; 174 return (EINTR); 175 } 176 } 177 mtx_lock_spin(&sched_lock); 178 if (cold ) { 179 /* 180 * During autoconfiguration, just give interrupts 181 * a chance, then just return. 182 * Don't run any other procs or panic below, 183 * in case this is the idle process and already asleep. 184 */ 185 if (mtx != NULL && priority & PDROP) 186 mtx_unlock(mtx); 187 mtx_unlock_spin(&sched_lock); 188 return (0); 189 } 190 191 DROP_GIANT(); 192 193 if (mtx != NULL) { 194 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 195 WITNESS_SAVE(&mtx->mtx_object, mtx); 196 mtx_unlock(mtx); 197 if (priority & PDROP) 198 mtx = NULL; 199 } 200 201 KASSERT(p != NULL, ("msleep1")); 202 KASSERT(ident != NULL && TD_IS_RUNNING(td), ("msleep")); 203 204 CTR5(KTR_PROC, "msleep: thread %p (pid %d, %s) on %s (%p)", 205 td, p->p_pid, p->p_comm, wmesg, ident); 206 207 td->td_wchan = ident; 208 td->td_wmesg = wmesg; 209 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], td, td_slpq); 210 TD_SET_ON_SLEEPQ(td); 211 if (timo) 212 callout_reset(&td->td_slpcallout, timo, endtsleep, td); 213 /* 214 * We put ourselves on the sleep queue and start our timeout 215 * before calling thread_suspend_check, as we could stop there, and 216 * a wakeup or a SIGCONT (or both) could occur while we were stopped. 217 * without resuming us, thus we must be ready for sleep 218 * when cursig is called. If the wakeup happens while we're 219 * stopped, td->td_wchan will be 0 upon return from cursig. 220 */ 221 if (catch) { 222 CTR3(KTR_PROC, "msleep caught: thread %p (pid %d, %s)", td, 223 p->p_pid, p->p_comm); 224 td->td_flags |= TDF_SINTR; 225 mtx_unlock_spin(&sched_lock); 226 PROC_LOCK(p); 227 sig = cursig(td); 228 if (sig == 0 && thread_suspend_check(1)) 229 sig = SIGSTOP; 230 mtx_lock_spin(&sched_lock); 231 PROC_UNLOCK(p); 232 if (sig != 0) { 233 if (TD_ON_SLEEPQ(td)) 234 unsleep(td); 235 } else if (!TD_ON_SLEEPQ(td)) 236 catch = 0; 237 } else 238 sig = 0; 239 240 /* 241 * Let the scheduler know we're about to voluntarily go to sleep. 242 */ 243 sched_sleep(td, priority & PRIMASK); 244 245 if (TD_ON_SLEEPQ(td)) { 246 p->p_stats->p_ru.ru_nvcsw++; 247 TD_SET_SLEEPING(td); 248 mi_switch(); 249 } 250 /* 251 * We're awake from voluntary sleep. 252 */ 253 CTR3(KTR_PROC, "msleep resume: thread %p (pid %d, %s)", td, p->p_pid, 254 p->p_comm); 255 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 256 td->td_flags &= ~TDF_SINTR; 257 if (td->td_flags & TDF_TIMEOUT) { 258 td->td_flags &= ~TDF_TIMEOUT; 259 if (sig == 0) 260 rval = EWOULDBLOCK; 261 } else if (td->td_flags & TDF_TIMOFAIL) { 262 td->td_flags &= ~TDF_TIMOFAIL; 263 } else if (timo && callout_stop(&td->td_slpcallout) == 0) { 264 /* 265 * This isn't supposed to be pretty. If we are here, then 266 * the endtsleep() callout is currently executing on another 267 * CPU and is either spinning on the sched_lock or will be 268 * soon. If we don't synchronize here, there is a chance 269 * that this process may msleep() again before the callout 270 * has a chance to run and the callout may end up waking up 271 * the wrong msleep(). Yuck. 272 */ 273 TD_SET_SLEEPING(td); 274 p->p_stats->p_ru.ru_nivcsw++; 275 mi_switch(); 276 td->td_flags &= ~TDF_TIMOFAIL; 277 } 278 if ((td->td_flags & TDF_INTERRUPT) && (priority & PCATCH) && 279 (rval == 0)) { 280 td->td_flags &= ~TDF_INTERRUPT; 281 rval = EINTR; 282 } 283 mtx_unlock_spin(&sched_lock); 284 285 if (rval == 0 && catch) { 286 PROC_LOCK(p); 287 /* XXX: shouldn't we always be calling cursig() */ 288 if (sig != 0 || (sig = cursig(td))) { 289 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 290 rval = EINTR; 291 else 292 rval = ERESTART; 293 } 294 PROC_UNLOCK(p); 295 } 296#ifdef KTRACE 297 if (KTRPOINT(td, KTR_CSW)) 298 ktrcsw(0, 0); 299#endif 300 PICKUP_GIANT(); 301 if (mtx != NULL) { 302 mtx_lock(mtx); 303 WITNESS_RESTORE(&mtx->mtx_object, mtx); 304 } 305 return (rval); 306} 307 308/* 309 * Implement timeout for msleep() 310 * 311 * If process hasn't been awakened (wchan non-zero), 312 * set timeout flag and undo the sleep. If proc 313 * is stopped, just unsleep so it will remain stopped. 314 * MP-safe, called without the Giant mutex. 315 */ 316static void 317endtsleep(arg) 318 void *arg; 319{ 320 register struct thread *td = arg; 321 322 CTR3(KTR_PROC, "endtsleep: thread %p (pid %d, %s)", 323 td, td->td_proc->p_pid, td->td_proc->p_comm); 324 mtx_lock_spin(&sched_lock); 325 /* 326 * This is the other half of the synchronization with msleep() 327 * described above. If the TDS_TIMEOUT flag is set, we lost the 328 * race and just need to put the process back on the runqueue. 329 */ 330 if (TD_ON_SLEEPQ(td)) { 331 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_slpq); 332 TD_CLR_ON_SLEEPQ(td); 333 td->td_flags |= TDF_TIMEOUT; 334 td->td_wmesg = NULL; 335 } else { 336 td->td_flags |= TDF_TIMOFAIL; 337 } 338 TD_CLR_SLEEPING(td); 339 setrunnable(td); 340 mtx_unlock_spin(&sched_lock); 341} 342 343/* 344 * Abort a thread, as if an interrupt had occured. Only abort 345 * interruptable waits (unfortunatly it isn't only safe to abort others). 346 * This is about identical to cv_abort(). 347 * Think about merging them? 348 * Also, whatever the signal code does... 349 */ 350void 351abortsleep(struct thread *td) 352{ 353 354 mtx_assert(&sched_lock, MA_OWNED); 355 /* 356 * If the TDF_TIMEOUT flag is set, just leave. A 357 * timeout is scheduled anyhow. 358 */ 359 if ((td->td_flags & (TDF_TIMEOUT | TDF_SINTR)) == TDF_SINTR) { 360 if (TD_ON_SLEEPQ(td)) { 361 unsleep(td); 362 TD_CLR_SLEEPING(td); 363 setrunnable(td); 364 } 365 } 366} 367 368/* 369 * Remove a process from its wait queue 370 */ 371void 372unsleep(struct thread *td) 373{ 374 375 mtx_lock_spin(&sched_lock); 376 if (TD_ON_SLEEPQ(td)) { 377 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_slpq); 378 TD_CLR_ON_SLEEPQ(td); 379 td->td_wmesg = NULL; 380 } 381 mtx_unlock_spin(&sched_lock); 382} 383 384/* 385 * Make all processes sleeping on the specified identifier runnable. 386 */ 387void 388wakeup(ident) 389 register void *ident; 390{ 391 register struct slpquehead *qp; 392 register struct thread *td; 393 struct thread *ntd; 394 struct proc *p; 395 396 mtx_lock_spin(&sched_lock); 397 qp = &slpque[LOOKUP(ident)]; 398restart: 399 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 400 ntd = TAILQ_NEXT(td, td_slpq); 401 if (td->td_wchan == ident) { 402 unsleep(td); 403 TD_CLR_SLEEPING(td); 404 setrunnable(td); 405 p = td->td_proc; 406 CTR3(KTR_PROC,"wakeup: thread %p (pid %d, %s)", 407 td, p->p_pid, p->p_comm); 408 goto restart; 409 } 410 } 411 mtx_unlock_spin(&sched_lock); 412} 413 414/* 415 * Make a process sleeping on the specified identifier runnable. 416 * May wake more than one process if a target process is currently 417 * swapped out. 418 */ 419void 420wakeup_one(ident) 421 register void *ident; 422{ 423 register struct slpquehead *qp; 424 register struct thread *td; 425 register struct proc *p; 426 struct thread *ntd; 427 428 mtx_lock_spin(&sched_lock); 429 qp = &slpque[LOOKUP(ident)]; 430 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 431 ntd = TAILQ_NEXT(td, td_slpq); 432 if (td->td_wchan == ident) { 433 unsleep(td); 434 TD_CLR_SLEEPING(td); 435 setrunnable(td); 436 p = td->td_proc; 437 CTR3(KTR_PROC,"wakeup1: thread %p (pid %d, %s)", 438 td, p->p_pid, p->p_comm); 439 break; 440 } 441 } 442 mtx_unlock_spin(&sched_lock); 443} 444 445/* 446 * The machine independent parts of mi_switch(). 447 */ 448void 449mi_switch(void) 450{ 451 struct bintime new_switchtime; 452 struct thread *td = curthread; /* XXX */ 453 struct proc *p = td->td_proc; /* XXX */ 454 u_int sched_nest; 455 456 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 457 458 KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code")); 459#ifdef INVARIANTS 460 if (!TD_ON_LOCK(td) && 461 !TD_ON_RUNQ(td) && 462 !TD_IS_RUNNING(td)) 463 mtx_assert(&Giant, MA_NOTOWNED); 464#endif 465 KASSERT(td->td_critnest == 1, 466 ("mi_switch: switch in a critical section")); 467 468 /* 469 * Compute the amount of time during which the current 470 * process was running, and add that to its total so far. 471 */ 472 binuptime(&new_switchtime); 473 bintime_add(&p->p_runtime, &new_switchtime); 474 bintime_sub(&p->p_runtime, PCPU_PTR(switchtime)); 475 476#ifdef DDB 477 /* 478 * Don't perform context switches from the debugger. 479 */ 480 if (db_active) { 481 mtx_unlock_spin(&sched_lock); 482 db_print_backtrace(); 483 db_error("Context switches not allowed in the debugger."); 484 } 485#endif 486 487 /* 488 * Check if the process exceeds its cpu resource allocation. If 489 * over max, arrange to kill the process in ast(). 490 */ 491 if (p->p_cpulimit != RLIM_INFINITY && 492 p->p_runtime.sec > p->p_cpulimit) { 493 p->p_sflag |= PS_XCPU; 494 td->td_flags |= TDF_ASTPENDING; 495 } 496 497 /* 498 * Finish up stats for outgoing thread. 499 */ 500 cnt.v_swtch++; 501 PCPU_SET(switchtime, new_switchtime); 502 CTR3(KTR_PROC, "mi_switch: old thread %p (pid %d, %s)", td, p->p_pid, 503 p->p_comm); 504 505 sched_nest = sched_lock.mtx_recurse; 506 sched_switchout(td); 507 508 cpu_switch(); /* SHAZAM!!*/ 509 510 sched_lock.mtx_recurse = sched_nest; 511 sched_lock.mtx_lock = (uintptr_t)td; 512 sched_switchin(td); 513 514 /* 515 * Start setting up stats etc. for the incoming thread. 516 * Similar code in fork_exit() is returned to by cpu_switch() 517 * in the case of a new thread/process. 518 */ 519 CTR3(KTR_PROC, "mi_switch: new thread %p (pid %d, %s)", td, p->p_pid, 520 p->p_comm); 521 if (PCPU_GET(switchtime.sec) == 0) 522 binuptime(PCPU_PTR(switchtime)); 523 PCPU_SET(switchticks, ticks); 524 525 /* 526 * Call the switchin function while still holding the scheduler lock 527 * (used by the idlezero code and the general page-zeroing code) 528 */ 529 if (td->td_switchin) 530 td->td_switchin(); 531 532 /* 533 * If the last thread was exiting, finish cleaning it up. 534 */ 535 if ((td = PCPU_GET(deadthread))) { 536 PCPU_SET(deadthread, NULL); 537 thread_stash(td); 538 } 539} 540 541/* 542 * Change process state to be runnable, 543 * placing it on the run queue if it is in memory, 544 * and awakening the swapper if it isn't in memory. 545 */ 546void 547setrunnable(struct thread *td) 548{ 549 struct proc *p = td->td_proc; 550 551 mtx_assert(&sched_lock, MA_OWNED); 552 switch (p->p_state) { 553 case PRS_ZOMBIE: 554 panic("setrunnable(1)"); 555 default: 556 break; 557 } 558 switch (td->td_state) { 559 case TDS_RUNNING: 560 case TDS_RUNQ: 561 return; 562 case TDS_INHIBITED: 563 /* 564 * If we are only inhibited because we are swapped out 565 * then arange to swap in this process. Otherwise just return. 566 */ 567 if (td->td_inhibitors != TDI_SWAPPED) 568 return; 569 case TDS_CAN_RUN: 570 break; 571 default: 572 printf("state is 0x%x", td->td_state); 573 panic("setrunnable(2)"); 574 } 575 if ((p->p_sflag & PS_INMEM) == 0) { 576 if ((p->p_sflag & PS_SWAPPINGIN) == 0) { 577 p->p_sflag |= PS_SWAPINREQ; 578 wakeup(&proc0); 579 } 580 } else 581 sched_wakeup(td); 582} 583 584/* 585 * Compute a tenex style load average of a quantity on 586 * 1, 5 and 15 minute intervals. 587 * XXXKSE Needs complete rewrite when correct info is available. 588 * Completely Bogus.. only works with 1:1 (but compiles ok now :-) 589 */ 590static void 591loadav(void *arg) 592{ 593 int i, nrun; 594 struct loadavg *avg; 595 struct proc *p; 596 struct thread *td; 597 598 avg = &averunnable; 599 sx_slock(&allproc_lock); 600 nrun = 0; 601 FOREACH_PROC_IN_SYSTEM(p) { 602 FOREACH_THREAD_IN_PROC(p, td) { 603 switch (td->td_state) { 604 case TDS_RUNQ: 605 case TDS_RUNNING: 606 if ((p->p_flag & P_NOLOAD) != 0) 607 goto nextproc; 608 nrun++; /* XXXKSE */ 609 default: 610 break; 611 } 612nextproc: 613 continue; 614 } 615 } 616 sx_sunlock(&allproc_lock); 617 for (i = 0; i < 3; i++) 618 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 619 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 620 621 /* 622 * Schedule the next update to occur after 5 seconds, but add a 623 * random variation to avoid synchronisation with processes that 624 * run at regular intervals. 625 */ 626 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)), 627 loadav, NULL); 628} 629 630static void 631lboltcb(void *arg) 632{ 633 wakeup(&lbolt); 634 callout_reset(&lbolt_callout, hz, lboltcb, NULL); 635} 636 637/* ARGSUSED */ 638static void 639sched_setup(dummy) 640 void *dummy; 641{ 642 callout_init(&loadav_callout, 0); 643 callout_init(&lbolt_callout, 1); 644 645 /* Kick off timeout driven events by calling first time. */ 646 loadav(NULL); 647 lboltcb(NULL); 648} 649 650/* 651 * General purpose yield system call 652 */ 653int 654yield(struct thread *td, struct yield_args *uap) 655{ 656 struct ksegrp *kg = td->td_ksegrp; 657 658 mtx_assert(&Giant, MA_NOTOWNED); 659 mtx_lock_spin(&sched_lock); 660 kg->kg_proc->p_stats->p_ru.ru_nvcsw++; 661 sched_prio(td, PRI_MAX_TIMESHARE); 662 mi_switch(); 663 mtx_unlock_spin(&sched_lock); 664 td->td_retval[0] = 0; 665 666 return (0); 667} 668 669