kern_synch.c revision 106180
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 106180 2002-10-30 02:28:41Z davidxu $ 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; 78 79struct loadavg averunnable = 80 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 81/* 82 * Constants for averages over 1, 5, and 15 minutes 83 * when sampling at 5 second intervals. 84 */ 85static fixpt_t cexp[3] = { 86 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 87 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 88 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 89}; 90 91static void endtsleep(void *); 92static void loadav(void *arg); 93 94/* 95 * We're only looking at 7 bits of the address; everything is 96 * aligned to 4, lots of things are aligned to greater powers 97 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 98 */ 99#define TABLESIZE 128 100static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE]; 101#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 102 103void 104sleepinit(void) 105{ 106 int i; 107 108 hogticks = (hz / 10) * 2; /* Default only. */ 109 for (i = 0; i < TABLESIZE; i++) 110 TAILQ_INIT(&slpque[i]); 111} 112 113/* 114 * General sleep call. Suspends the current process until a wakeup is 115 * performed on the specified identifier. The process will then be made 116 * runnable with the specified priority. Sleeps at most timo/hz seconds 117 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 118 * before and after sleeping, else signals are not checked. Returns 0 if 119 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 120 * signal needs to be delivered, ERESTART is returned if the current system 121 * call should be restarted if possible, and EINTR is returned if the system 122 * call should be interrupted by the signal (return EINTR). 123 * 124 * The mutex argument is exited before the caller is suspended, and 125 * entered before msleep returns. If priority includes the PDROP 126 * flag the mutex is not entered before returning. 127 */ 128 129int 130msleep(ident, mtx, priority, wmesg, timo) 131 void *ident; 132 struct mtx *mtx; 133 int priority, timo; 134 const char *wmesg; 135{ 136 struct thread *td = curthread; 137 struct proc *p = td->td_proc; 138 int sig, catch = priority & PCATCH; 139 int rval = 0; 140 WITNESS_SAVE_DECL(mtx); 141 142#ifdef KTRACE 143 if (KTRPOINT(td, KTR_CSW)) 144 ktrcsw(1, 0); 145#endif 146 WITNESS_SLEEP(0, &mtx->mtx_object); 147 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 148 ("sleeping without a mutex")); 149 /* 150 * If we are capable of async syscalls and there isn't already 151 * another one ready to return, start a new thread 152 * and queue it as ready to run. Note that there is danger here 153 * because we need to make sure that we don't sleep allocating 154 * the thread (recursion here might be bad). 155 * Hence the TDF_INMSLEEP flag. 156 */ 157 if (p->p_flag & P_KSES) { 158 /* 159 * Just don't bother if we are exiting 160 * and not the exiting thread or thread was marked as 161 * interrupted. 162 */ 163 if (catch && 164 (((p->p_flag & P_WEXIT) && (p->p_singlethread != td)) || 165 (td->td_flags & TDF_INTERRUPT))) { 166 td->td_flags &= ~TDF_INTERRUPT; 167 return (EINTR); 168 } 169 mtx_lock_spin(&sched_lock); 170 if ((td->td_flags & (TDF_UNBOUND|TDF_INMSLEEP)) == 171 TDF_UNBOUND) { 172 /* 173 * Arrange for an upcall to be readied. 174 * it will not actually happen until all 175 * pending in-kernel work for this KSEGRP 176 * has been done. 177 */ 178 /* Don't recurse here! */ 179 td->td_flags |= TDF_INMSLEEP; 180 thread_schedule_upcall(td, td->td_kse); 181 td->td_flags &= ~TDF_INMSLEEP; 182 } 183 } else { 184 mtx_lock_spin(&sched_lock); 185 } 186 if (cold ) { 187 /* 188 * During autoconfiguration, just give interrupts 189 * a chance, then just return. 190 * Don't run any other procs or panic below, 191 * in case this is the idle process and already asleep. 192 */ 193 if (mtx != NULL && priority & PDROP) 194 mtx_unlock(mtx); 195 mtx_unlock_spin(&sched_lock); 196 return (0); 197 } 198 199 DROP_GIANT(); 200 201 if (mtx != NULL) { 202 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 203 WITNESS_SAVE(&mtx->mtx_object, mtx); 204 mtx_unlock(mtx); 205 if (priority & PDROP) 206 mtx = NULL; 207 } 208 209 KASSERT(p != NULL, ("msleep1")); 210 KASSERT(ident != NULL && TD_IS_RUNNING(td), ("msleep")); 211 212 CTR5(KTR_PROC, "msleep: thread %p (pid %d, %s) on %s (%p)", 213 td, p->p_pid, p->p_comm, wmesg, ident); 214 215 td->td_wchan = ident; 216 td->td_wmesg = wmesg; 217 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], td, td_slpq); 218 TD_SET_ON_SLEEPQ(td); 219 if (timo) 220 callout_reset(&td->td_slpcallout, timo, endtsleep, td); 221 /* 222 * We put ourselves on the sleep queue and start our timeout 223 * before calling thread_suspend_check, as we could stop there, and 224 * a wakeup or a SIGCONT (or both) could occur while we were stopped. 225 * without resuming us, thus we must be ready for sleep 226 * when cursig is called. If the wakeup happens while we're 227 * stopped, td->td_wchan will be 0 upon return from cursig. 228 */ 229 if (catch) { 230 CTR3(KTR_PROC, "msleep caught: thread %p (pid %d, %s)", td, 231 p->p_pid, p->p_comm); 232 td->td_flags |= TDF_SINTR; 233 mtx_unlock_spin(&sched_lock); 234 PROC_LOCK(p); 235 sig = cursig(td); 236 if (sig == 0 && thread_suspend_check(1)) 237 sig = SIGSTOP; 238 mtx_lock_spin(&sched_lock); 239 PROC_UNLOCK(p); 240 if (sig != 0) { 241 if (TD_ON_SLEEPQ(td)) 242 unsleep(td); 243 } else if (!TD_ON_SLEEPQ(td)) 244 catch = 0; 245 } else 246 sig = 0; 247 248 /* 249 * Let the scheduler know we're about to voluntarily go to sleep. 250 */ 251 sched_sleep(td, priority & PRIMASK); 252 253 if (TD_ON_SLEEPQ(td)) { 254 p->p_stats->p_ru.ru_nvcsw++; 255 TD_SET_SLEEPING(td); 256 mi_switch(); 257 } 258 /* 259 * We're awake from voluntary sleep. 260 */ 261 CTR3(KTR_PROC, "msleep resume: thread %p (pid %d, %s)", td, p->p_pid, 262 p->p_comm); 263 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 264 td->td_flags &= ~TDF_SINTR; 265 if (td->td_flags & TDF_TIMEOUT) { 266 td->td_flags &= ~TDF_TIMEOUT; 267 if (sig == 0) 268 rval = EWOULDBLOCK; 269 } else if (td->td_flags & TDF_TIMOFAIL) { 270 td->td_flags &= ~TDF_TIMOFAIL; 271 } else if (timo && callout_stop(&td->td_slpcallout) == 0) { 272 /* 273 * This isn't supposed to be pretty. If we are here, then 274 * the endtsleep() callout is currently executing on another 275 * CPU and is either spinning on the sched_lock or will be 276 * soon. If we don't synchronize here, there is a chance 277 * that this process may msleep() again before the callout 278 * has a chance to run and the callout may end up waking up 279 * the wrong msleep(). Yuck. 280 */ 281 TD_SET_SLEEPING(td); 282 p->p_stats->p_ru.ru_nivcsw++; 283 mi_switch(); 284 td->td_flags &= ~TDF_TIMOFAIL; 285 } 286 if ((td->td_flags & TDF_INTERRUPT) && (priority & PCATCH) && 287 (rval == 0)) { 288 td->td_flags &= ~TDF_INTERRUPT; 289 rval = EINTR; 290 } 291 mtx_unlock_spin(&sched_lock); 292 293 if (rval == 0 && catch) { 294 PROC_LOCK(p); 295 /* XXX: shouldn't we always be calling cursig() */ 296 if (sig != 0 || (sig = cursig(td))) { 297 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 298 rval = EINTR; 299 else 300 rval = ERESTART; 301 } 302 PROC_UNLOCK(p); 303 } 304#ifdef KTRACE 305 if (KTRPOINT(td, KTR_CSW)) 306 ktrcsw(0, 0); 307#endif 308 PICKUP_GIANT(); 309 if (mtx != NULL) { 310 mtx_lock(mtx); 311 WITNESS_RESTORE(&mtx->mtx_object, mtx); 312 } 313 return (rval); 314} 315 316/* 317 * Implement timeout for msleep() 318 * 319 * If process hasn't been awakened (wchan non-zero), 320 * set timeout flag and undo the sleep. If proc 321 * is stopped, just unsleep so it will remain stopped. 322 * MP-safe, called without the Giant mutex. 323 */ 324static void 325endtsleep(arg) 326 void *arg; 327{ 328 register struct thread *td = arg; 329 330 CTR3(KTR_PROC, "endtsleep: thread %p (pid %d, %s)", 331 td, td->td_proc->p_pid, td->td_proc->p_comm); 332 mtx_lock_spin(&sched_lock); 333 /* 334 * This is the other half of the synchronization with msleep() 335 * described above. If the TDS_TIMEOUT flag is set, we lost the 336 * race and just need to put the process back on the runqueue. 337 */ 338 if (TD_ON_SLEEPQ(td)) { 339 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_slpq); 340 TD_CLR_ON_SLEEPQ(td); 341 td->td_flags |= TDF_TIMEOUT; 342 } else { 343 td->td_flags |= TDF_TIMOFAIL; 344 } 345 TD_CLR_SLEEPING(td); 346 setrunnable(td); 347 mtx_unlock_spin(&sched_lock); 348} 349 350/* 351 * Abort a thread, as if an interrupt had occured. Only abort 352 * interruptable waits (unfortunatly it isn't only safe to abort others). 353 * This is about identical to cv_abort(). 354 * Think about merging them? 355 * Also, whatever the signal code does... 356 */ 357void 358abortsleep(struct thread *td) 359{ 360 361 mtx_assert(&sched_lock, MA_OWNED); 362 /* 363 * If the TDF_TIMEOUT flag is set, just leave. A 364 * timeout is scheduled anyhow. 365 */ 366 if ((td->td_flags & (TDF_TIMEOUT | TDF_SINTR)) == TDF_SINTR) { 367 if (TD_ON_SLEEPQ(td)) { 368 unsleep(td); 369 TD_CLR_SLEEPING(td); 370 setrunnable(td); 371 } 372 } 373} 374 375/* 376 * Remove a process from its wait queue 377 */ 378void 379unsleep(struct thread *td) 380{ 381 382 mtx_lock_spin(&sched_lock); 383 if (TD_ON_SLEEPQ(td)) { 384 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_slpq); 385 TD_CLR_ON_SLEEPQ(td); 386 } 387 mtx_unlock_spin(&sched_lock); 388} 389 390/* 391 * Make all processes sleeping on the specified identifier runnable. 392 */ 393void 394wakeup(ident) 395 register void *ident; 396{ 397 register struct slpquehead *qp; 398 register struct thread *td; 399 struct thread *ntd; 400 struct proc *p; 401 402 mtx_lock_spin(&sched_lock); 403 qp = &slpque[LOOKUP(ident)]; 404restart: 405 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 406 ntd = TAILQ_NEXT(td, td_slpq); 407 if (td->td_wchan == ident) { 408 unsleep(td); 409 TD_CLR_SLEEPING(td); 410 setrunnable(td); 411 p = td->td_proc; 412 CTR3(KTR_PROC,"wakeup: thread %p (pid %d, %s)", 413 td, p->p_pid, p->p_comm); 414 goto restart; 415 } 416 } 417 mtx_unlock_spin(&sched_lock); 418} 419 420/* 421 * Make a process sleeping on the specified identifier runnable. 422 * May wake more than one process if a target process is currently 423 * swapped out. 424 */ 425void 426wakeup_one(ident) 427 register void *ident; 428{ 429 register struct slpquehead *qp; 430 register struct thread *td; 431 register struct proc *p; 432 struct thread *ntd; 433 434 mtx_lock_spin(&sched_lock); 435 qp = &slpque[LOOKUP(ident)]; 436 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 437 ntd = TAILQ_NEXT(td, td_slpq); 438 if (td->td_wchan == ident) { 439 unsleep(td); 440 TD_CLR_SLEEPING(td); 441 setrunnable(td); 442 p = td->td_proc; 443 CTR3(KTR_PROC,"wakeup1: thread %p (pid %d, %s)", 444 td, p->p_pid, p->p_comm); 445 break; 446 } 447 } 448 mtx_unlock_spin(&sched_lock); 449} 450 451/* 452 * The machine independent parts of mi_switch(). 453 */ 454void 455mi_switch(void) 456{ 457 struct bintime new_switchtime; 458 struct thread *td = curthread; /* XXX */ 459 struct proc *p = td->td_proc; /* XXX */ 460 struct kse *ke = td->td_kse; 461 u_int sched_nest; 462 463 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 464 465 KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code")); 466#ifdef INVARIANTS 467 if (!TD_ON_LOCK(td) && 468 !TD_ON_RUNQ(td) && 469 !TD_IS_RUNNING(td)) 470 mtx_assert(&Giant, MA_NOTOWNED); 471#endif 472 KASSERT(td->td_critnest == 1, 473 ("mi_switch: switch in a critical section")); 474 475 /* 476 * Compute the amount of time during which the current 477 * process was running, and add that to its total so far. 478 */ 479 binuptime(&new_switchtime); 480 bintime_add(&p->p_runtime, &new_switchtime); 481 bintime_sub(&p->p_runtime, PCPU_PTR(switchtime)); 482 483#ifdef DDB 484 /* 485 * Don't perform context switches from the debugger. 486 */ 487 if (db_active) { 488 mtx_unlock_spin(&sched_lock); 489 db_error("Context switches not allowed in the debugger."); 490 } 491#endif 492 493 /* 494 * Check if the process exceeds its cpu resource allocation. If 495 * over max, arrange to kill the process in ast(). 496 */ 497 if (p->p_cpulimit != RLIM_INFINITY && 498 p->p_runtime.sec > p->p_cpulimit) { 499 p->p_sflag |= PS_XCPU; 500 ke->ke_flags |= KEF_ASTPENDING; 501 } 502 503 /* 504 * Finish up stats for outgoing thread. 505 */ 506 cnt.v_swtch++; 507 PCPU_SET(switchtime, new_switchtime); 508 CTR3(KTR_PROC, "mi_switch: old thread %p (pid %d, %s)", td, p->p_pid, 509 p->p_comm); 510 511 sched_nest = sched_lock.mtx_recurse; 512 sched_switchout(td); 513 514 cpu_switch(); /* SHAZAM!!*/ 515 516 sched_lock.mtx_recurse = sched_nest; 517 sched_lock.mtx_lock = (uintptr_t)td; 518 sched_switchin(td); 519 520 /* 521 * Start setting up stats etc. for the incoming thread. 522 * Similar code in fork_exit() is returned to by cpu_switch() 523 * in the case of a new thread/process. 524 */ 525 CTR3(KTR_PROC, "mi_switch: new thread %p (pid %d, %s)", td, p->p_pid, 526 p->p_comm); 527 if (PCPU_GET(switchtime.sec) == 0) 528 binuptime(PCPU_PTR(switchtime)); 529 PCPU_SET(switchticks, ticks); 530 531 /* 532 * Call the switchin function while still holding the scheduler lock 533 * (used by the idlezero code and the general page-zeroing code) 534 */ 535 if (td->td_switchin) 536 td->td_switchin(); 537} 538 539/* 540 * Change process state to be runnable, 541 * placing it on the run queue if it is in memory, 542 * and awakening the swapper if it isn't in memory. 543 */ 544void 545setrunnable(struct thread *td) 546{ 547 struct proc *p = td->td_proc; 548 549 mtx_assert(&sched_lock, MA_OWNED); 550 switch (p->p_state) { 551 case PRS_ZOMBIE: 552 panic("setrunnable(1)"); 553 default: 554 break; 555 } 556 switch (td->td_state) { 557 case TDS_RUNNING: 558 case TDS_RUNQ: 559 return; 560 case TDS_INHIBITED: 561 /* 562 * If we are only inhibited because we are swapped out 563 * then arange to swap in this process. Otherwise just return. 564 */ 565 if (td->td_inhibitors != TDI_SWAPPED) 566 return; 567 case TDS_CAN_RUN: 568 break; 569 default: 570 printf("state is 0x%x", td->td_state); 571 panic("setrunnable(2)"); 572 } 573 if ((p->p_sflag & PS_INMEM) == 0) { 574 if ((p->p_sflag & PS_SWAPPINGIN) == 0) { 575 p->p_sflag |= PS_SWAPINREQ; 576 wakeup(&proc0); 577 } 578 } else 579 sched_wakeup(td); 580} 581 582/* 583 * Compute a tenex style load average of a quantity on 584 * 1, 5 and 15 minute intervals. 585 * XXXKSE Needs complete rewrite when correct info is available. 586 * Completely Bogus.. only works with 1:1 (but compiles ok now :-) 587 */ 588static void 589loadav(void *arg) 590{ 591 int i, nrun; 592 struct loadavg *avg; 593 struct proc *p; 594 struct thread *td; 595 596 avg = &averunnable; 597 sx_slock(&allproc_lock); 598 nrun = 0; 599 FOREACH_PROC_IN_SYSTEM(p) { 600 FOREACH_THREAD_IN_PROC(p, td) { 601 switch (td->td_state) { 602 case TDS_RUNQ: 603 case TDS_RUNNING: 604 if ((p->p_flag & P_NOLOAD) != 0) 605 goto nextproc; 606 nrun++; /* XXXKSE */ 607 default: 608 break; 609 } 610nextproc: 611 continue; 612 } 613 } 614 sx_sunlock(&allproc_lock); 615 for (i = 0; i < 3; i++) 616 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 617 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 618 619 /* 620 * Schedule the next update to occur after 5 seconds, but add a 621 * random variation to avoid synchronisation with processes that 622 * run at regular intervals. 623 */ 624 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)), 625 loadav, NULL); 626} 627 628/* ARGSUSED */ 629static void 630sched_setup(dummy) 631 void *dummy; 632{ 633 callout_init(&loadav_callout, 0); 634 635 /* Kick off timeout driven events by calling first time. */ 636 loadav(NULL); 637} 638 639/* 640 * General purpose yield system call 641 */ 642int 643yield(struct thread *td, struct yield_args *uap) 644{ 645 struct ksegrp *kg = td->td_ksegrp; 646 647 mtx_assert(&Giant, MA_NOTOWNED); 648 mtx_lock_spin(&sched_lock); 649 kg->kg_proc->p_stats->p_ru.ru_nvcsw++; 650 sched_prio(td, PRI_MAX_TIMESHARE); 651 mi_switch(); 652 mtx_unlock_spin(&sched_lock); 653 td->td_retval[0] = 0; 654 655 return (0); 656} 657 658