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 */ 40 41 #include <sys/cdefs.h> 42 __FBSDID("$FreeBSD$"); 43 44 #include "opt_ddb.h" 45 #include "opt_ktrace.h" 46 47 #include <sys/param.h> 48 #include <sys/systm.h> 49 #include <sys/condvar.h> 50 #include <sys/kernel.h> 51 #include <sys/ktr.h> 52 #include <sys/lock.h> 53 #include <sys/mutex.h> 54 #include <sys/proc.h> 55 #include <sys/resourcevar.h> 56 #include <sys/sched.h> 57 #include <sys/signalvar.h> 58 #include <sys/smp.h> 59 #include <sys/sx.h> 60 #include <sys/sysctl.h> 61 #include <sys/sysproto.h> 62 #include <sys/vmmeter.h> 63 #ifdef DDB 64 #include <ddb/ddb.h> 65 #endif 66 #ifdef KTRACE 67 #include <sys/uio.h> 68 #include <sys/ktrace.h> 69 #endif 70 71 #include <machine/cpu.h> 72 73 static void sched_setup(void *dummy); 74 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 75 76 int hogticks; 77 int lbolt; 78 79 static struct callout loadav_callout; 80 static struct callout lbolt_callout; 81 82 struct loadavg averunnable = 83 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 84 /* 85 * Constants for averages over 1, 5, and 15 minutes 86 * when sampling at 5 second intervals. 87 */ 88 static fixpt_t cexp[3] = { 89 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 90 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 91 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 92 }; 93 94 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 95 static int fscale __unused = FSCALE; 96 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 97 98 static void endtsleep(void *); 99 static void loadav(void *arg); 100 static void lboltcb(void *arg); 101 102 /* 103 * We're only looking at 7 bits of the address; everything is 104 * aligned to 4, lots of things are aligned to greater powers 105 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 106 */ 107 #define TABLESIZE 128 108 static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE]; 109 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 110 111 void 112 sleepinit(void) 113 { 114 int i; 115 116 hogticks = (hz / 10) * 2; /* Default only. */ 117 for (i = 0; i < TABLESIZE; i++) 118 TAILQ_INIT(&slpque[i]); 119 } 120 121 /* 122 * General sleep call. Suspends the current process until a wakeup is 123 * performed on the specified identifier. The process will then be made 124 * runnable with the specified priority. Sleeps at most timo/hz seconds 125 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 126 * before and after sleeping, else signals are not checked. Returns 0 if 127 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 128 * signal needs to be delivered, ERESTART is returned if the current system 129 * call should be restarted if possible, and EINTR is returned if the system 130 * call should be interrupted by the signal (return EINTR). 131 * 132 * The mutex argument is exited before the caller is suspended, and 133 * entered before msleep returns. If priority includes the PDROP 134 * flag the mutex is not entered before returning. 135 */ 136 137 int 138 msleep(ident, mtx, priority, wmesg, timo) 139 void *ident; 140 struct mtx *mtx; 141 int priority, timo; 142 const char *wmesg; 143 { 144 struct thread *td = curthread; 145 struct proc *p = td->td_proc; 146 int sig, catch = priority & PCATCH; 147 int rval = 0; 148 WITNESS_SAVE_DECL(mtx); 149 150 #ifdef KTRACE 151 if (KTRPOINT(td, KTR_CSW)) 152 ktrcsw(1, 0); 153 #endif 154 /* XXX: mtx == NULL ?? */ 155 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, &mtx->mtx_object, 156 "Sleeping on \"%s\"", wmesg); 157 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 158 ("sleeping without a mutex")); 159 /* 160 * If we are capable of async syscalls and there isn't already 161 * another one ready to return, start a new thread 162 * and queue it as ready to run. Note that there is danger here 163 * because we need to make sure that we don't sleep allocating 164 * the thread (recursion here might be bad). 165 */ 166 mtx_lock_spin(&sched_lock); 167 if (p->p_flag & P_SA || p->p_numthreads > 1) { 168 /* 169 * Just don't bother if we are exiting 170 * and not the exiting thread or thread was marked as 171 * interrupted. 172 */ 173 if (catch) { 174 if ((p->p_flag & P_WEXIT) && p->p_singlethread != td) { 175 mtx_unlock_spin(&sched_lock); 176 return (EINTR); 177 } 178 if (td->td_flags & TDF_INTERRUPT) { 179 mtx_unlock_spin(&sched_lock); 180 return (td->td_intrval); 181 } 182 } 183 } 184 if (cold ) { 185 /* 186 * During autoconfiguration, just return; 187 * don't run any other procs or panic below, 188 * in case this is the idle process and already asleep. 189 * XXX: this used to do "s = splhigh(); splx(safepri); 190 * splx(s);" to give interrupts a chance, but there is 191 * no way to give interrupts a chance now. 192 */ 193 if (mtx != NULL && priority & PDROP) 194 mtx_unlock(mtx); 195 mtx_unlock_spin(&sched_lock); 196 return (0); 197 } 198 DROP_GIANT(); 199 if (mtx != NULL) { 200 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 201 WITNESS_SAVE(&mtx->mtx_object, mtx); 202 mtx_unlock(mtx); 203 if (priority & PDROP) 204 mtx = NULL; 205 } 206 KASSERT(p != NULL, ("msleep1")); 207 KASSERT(ident != NULL && TD_IS_RUNNING(td), ("msleep")); 208 209 CTR5(KTR_PROC, "msleep: thread %p (pid %d, %s) on %s (%p)", 210 td, p->p_pid, p->p_comm, wmesg, ident); 211 212 td->td_wchan = ident; 213 td->td_wmesg = wmesg; 214 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], td, td_slpq); 215 TD_SET_ON_SLEEPQ(td); 216 if (timo) 217 callout_reset(&td->td_slpcallout, timo, endtsleep, td); 218 /* 219 * We put ourselves on the sleep queue and start our timeout 220 * before calling thread_suspend_check, as we could stop there, and 221 * a wakeup or a SIGCONT (or both) could occur while we were stopped. 222 * without resuming us, thus we must be ready for sleep 223 * when cursig is called. If the wakeup happens while we're 224 * stopped, td->td_wchan will be 0 upon return from cursig. 225 */ 226 if (catch) { 227 CTR3(KTR_PROC, "msleep caught: thread %p (pid %d, %s)", td, 228 p->p_pid, p->p_comm); 229 td->td_flags |= TDF_SINTR; 230 mtx_unlock_spin(&sched_lock); 231 PROC_LOCK(p); 232 mtx_lock(&p->p_sigacts->ps_mtx); 233 sig = cursig(td); 234 mtx_unlock(&p->p_sigacts->ps_mtx); 235 if (sig == 0 && thread_suspend_check(1)) 236 sig = SIGSTOP; 237 mtx_lock_spin(&sched_lock); 238 PROC_UNLOCK(p); 239 if (sig != 0) { 240 if (TD_ON_SLEEPQ(td)) 241 unsleep(td); 242 } else if (!TD_ON_SLEEPQ(td)) 243 catch = 0; 244 } else 245 sig = 0; 246 247 /* 248 * Let the scheduler know we're about to voluntarily go to sleep. 249 */ 250 sched_sleep(td, priority & PRIMASK); 251 252 if (TD_ON_SLEEPQ(td)) { 253 p->p_stats->p_ru.ru_nvcsw++; 254 TD_SET_SLEEPING(td); 255 mi_switch(); 256 } 257 /* 258 * We're awake from voluntary sleep. 259 */ 260 CTR3(KTR_PROC, "msleep resume: thread %p (pid %d, %s)", td, p->p_pid, 261 p->p_comm); 262 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 263 td->td_flags &= ~TDF_SINTR; 264 if (td->td_flags & TDF_TIMEOUT) { 265 td->td_flags &= ~TDF_TIMEOUT; 266 if (sig == 0) 267 rval = EWOULDBLOCK; 268 } else if (td->td_flags & TDF_TIMOFAIL) { 269 td->td_flags &= ~TDF_TIMOFAIL; 270 } else if (timo && callout_stop(&td->td_slpcallout) == 0) { 271 /* 272 * This isn't supposed to be pretty. If we are here, then 273 * the endtsleep() callout is currently executing on another 274 * CPU and is either spinning on the sched_lock or will be 275 * soon. If we don't synchronize here, there is a chance 276 * that this process may msleep() again before the callout 277 * has a chance to run and the callout may end up waking up 278 * the wrong msleep(). Yuck. 279 */ 280 TD_SET_SLEEPING(td); 281 p->p_stats->p_ru.ru_nivcsw++; 282 mi_switch(); 283 td->td_flags &= ~TDF_TIMOFAIL; 284 } 285 if ((td->td_flags & TDF_INTERRUPT) && (priority & PCATCH) && 286 (rval == 0)) { 287 rval = td->td_intrval; 288 } 289 mtx_unlock_spin(&sched_lock); 290 if (rval == 0 && catch) { 291 PROC_LOCK(p); 292 /* XXX: shouldn't we always be calling cursig()? */ 293 mtx_lock(&p->p_sigacts->ps_mtx); 294 if (sig != 0 || (sig = cursig(td))) { 295 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 296 rval = EINTR; 297 else 298 rval = ERESTART; 299 } 300 mtx_unlock(&p->p_sigacts->ps_mtx); 301 PROC_UNLOCK(p); 302 } 303 #ifdef KTRACE 304 if (KTRPOINT(td, KTR_CSW)) 305 ktrcsw(0, 0); 306 #endif 307 PICKUP_GIANT(); 308 if (mtx != NULL) { 309 mtx_lock(mtx); 310 WITNESS_RESTORE(&mtx->mtx_object, mtx); 311 } 312 return (rval); 313 } 314 315 /* 316 * Implement timeout for msleep(). 317 * 318 * If process hasn't been awakened (wchan non-zero), 319 * set timeout flag and undo the sleep. If proc 320 * is stopped, just unsleep so it will remain stopped. 321 * MP-safe, called without the Giant mutex. 322 */ 323 static void 324 endtsleep(arg) 325 void *arg; 326 { 327 register struct thread *td; 328 329 td = (struct thread *)arg; 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 td->td_wmesg = NULL; 343 } else 344 td->td_flags |= TDF_TIMOFAIL; 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 */ 357 void 358 abortsleep(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 */ 378 void 379 unsleep(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 td->td_wmesg = NULL; 387 } 388 mtx_unlock_spin(&sched_lock); 389 } 390 391 /* 392 * Make all processes sleeping on the specified identifier runnable. 393 */ 394 void 395 wakeup(ident) 396 register void *ident; 397 { 398 register struct slpquehead *qp; 399 register struct thread *td; 400 struct thread *ntd; 401 struct proc *p; 402 403 mtx_lock_spin(&sched_lock); 404 qp = &slpque[LOOKUP(ident)]; 405 restart: 406 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 407 ntd = TAILQ_NEXT(td, td_slpq); 408 if (td->td_wchan == ident) { 409 unsleep(td); 410 TD_CLR_SLEEPING(td); 411 setrunnable(td); 412 p = td->td_proc; 413 CTR3(KTR_PROC,"wakeup: thread %p (pid %d, %s)", 414 td, p->p_pid, p->p_comm); 415 goto restart; 416 } 417 } 418 mtx_unlock_spin(&sched_lock); 419 } 420 421 /* 422 * Make a process sleeping on the specified identifier runnable. 423 * May wake more than one process if a target process is currently 424 * swapped out. 425 */ 426 void 427 wakeup_one(ident) 428 register void *ident; 429 { 430 register struct proc *p; 431 register struct slpquehead *qp; 432 register struct thread *td; 433 struct thread *ntd; 434 435 mtx_lock_spin(&sched_lock); 436 qp = &slpque[LOOKUP(ident)]; 437 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 438 ntd = TAILQ_NEXT(td, td_slpq); 439 if (td->td_wchan == ident) { 440 unsleep(td); 441 TD_CLR_SLEEPING(td); 442 setrunnable(td); 443 p = td->td_proc; 444 CTR3(KTR_PROC,"wakeup1: thread %p (pid %d, %s)", 445 td, p->p_pid, p->p_comm); 446 break; 447 } 448 } 449 mtx_unlock_spin(&sched_lock); 450 } 451 452 /* 453 * The machine independent parts of mi_switch(). 454 */ 455 void 456 mi_switch(void) 457 { 458 struct bintime new_switchtime; 459 struct thread *td; 460 struct proc *p; 461 462 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 463 td = curthread; /* XXX */ 464 p = td->td_proc; /* XXX */ 465 KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code")); 466 #ifdef INVARIANTS 467 if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td)) 468 mtx_assert(&Giant, MA_NOTOWNED); 469 #endif 470 KASSERT(td->td_critnest == 1, 471 ("mi_switch: switch in a critical section")); 472 473 /* 474 * Compute the amount of time during which the current 475 * process was running, and add that to its total so far. 476 */ 477 binuptime(&new_switchtime); 478 bintime_add(&p->p_runtime, &new_switchtime); 479 bintime_sub(&p->p_runtime, PCPU_PTR(switchtime)); 480 481 td->td_generation++; /* bump preempt-detect counter */ 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_print_backtrace(); 490 db_error("Context switches not allowed in the debugger"); 491 } 492 #endif 493 494 /* 495 * Check if the process exceeds its cpu resource allocation. If 496 * over max, arrange to kill the process in ast(). 497 */ 498 if (p->p_cpulimit != RLIM_INFINITY && 499 p->p_runtime.sec > p->p_cpulimit) { 500 p->p_sflag |= PS_XCPU; 501 td->td_flags |= TDF_ASTPENDING; 502 } 503 504 /* 505 * Finish up stats for outgoing thread. 506 */ 507 cnt.v_swtch++; 508 PCPU_SET(switchtime, new_switchtime); 509 PCPU_SET(switchticks, ticks); 510 CTR3(KTR_PROC, "mi_switch: old thread %p (pid %d, %s)", td, p->p_pid, 511 p->p_comm); 512 if (td->td_proc->p_flag & P_SA) 513 thread_switchout(td); 514 sched_switch(td); 515 516 CTR3(KTR_PROC, "mi_switch: new thread %p (pid %d, %s)", td, p->p_pid, 517 p->p_comm); 518 519 /* 520 * If the last thread was exiting, finish cleaning it up. 521 */ 522 if ((td = PCPU_GET(deadthread))) { 523 PCPU_SET(deadthread, NULL); 524 thread_stash(td); 525 } 526 } 527 528 /* 529 * Change process state to be runnable, 530 * placing it on the run queue if it is in memory, 531 * and awakening the swapper if it isn't in memory. 532 */ 533 void 534 setrunnable(struct thread *td) 535 { 536 struct proc *p; 537 538 p = td->td_proc; 539 mtx_assert(&sched_lock, MA_OWNED); 540 switch (p->p_state) { 541 case PRS_ZOMBIE: 542 panic("setrunnable(1)"); 543 default: 544 break; 545 } 546 switch (td->td_state) { 547 case TDS_RUNNING: 548 case TDS_RUNQ: 549 return; 550 case TDS_INHIBITED: 551 /* 552 * If we are only inhibited because we are swapped out 553 * then arange to swap in this process. Otherwise just return. 554 */ 555 if (td->td_inhibitors != TDI_SWAPPED) 556 return; 557 /* XXX: intentional fall-through ? */ 558 case TDS_CAN_RUN: 559 break; 560 default: 561 printf("state is 0x%x", td->td_state); 562 panic("setrunnable(2)"); 563 } 564 if ((p->p_sflag & PS_INMEM) == 0) { 565 if ((p->p_sflag & PS_SWAPPINGIN) == 0) { 566 p->p_sflag |= PS_SWAPINREQ; 567 wakeup(&proc0); 568 } 569 } else 570 sched_wakeup(td); 571 } 572 573 /* 574 * Compute a tenex style load average of a quantity on 575 * 1, 5 and 15 minute intervals. 576 * XXXKSE Needs complete rewrite when correct info is available. 577 * Completely Bogus.. only works with 1:1 (but compiles ok now :-) 578 */ 579 static void 580 loadav(void *arg) 581 { 582 int i, nrun; 583 struct loadavg *avg; 584 struct proc *p; 585 struct thread *td; 586 587 avg = &averunnable; 588 sx_slock(&allproc_lock); 589 nrun = 0; 590 FOREACH_PROC_IN_SYSTEM(p) { 591 FOREACH_THREAD_IN_PROC(p, td) { 592 switch (td->td_state) { 593 case TDS_RUNQ: 594 case TDS_RUNNING: 595 if ((p->p_flag & P_NOLOAD) != 0) 596 goto nextproc; 597 nrun++; /* XXXKSE */ 598 default: 599 break; 600 } 601 nextproc: 602 continue; 603 } 604 } 605 sx_sunlock(&allproc_lock); 606 for (i = 0; i < 3; i++) 607 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 608 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 609 610 /* 611 * Schedule the next update to occur after 5 seconds, but add a 612 * random variation to avoid synchronisation with processes that 613 * run at regular intervals. 614 */ 615 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)), 616 loadav, NULL); 617 } 618 619 static void 620 lboltcb(void *arg) 621 { 622 wakeup(&lbolt); 623 callout_reset(&lbolt_callout, hz, lboltcb, NULL); 624 } 625 626 /* ARGSUSED */ 627 static void 628 sched_setup(dummy) 629 void *dummy; 630 { 631 callout_init(&loadav_callout, 0); 632 callout_init(&lbolt_callout, CALLOUT_MPSAFE); 633 634 /* Kick off timeout driven events by calling first time. */ 635 loadav(NULL); 636 lboltcb(NULL); 637 } 638 639 /* 640 * General purpose yield system call 641 */ 642 int 643 yield(struct thread *td, struct yield_args *uap) 644 { 645 struct ksegrp *kg; 646 647 kg = td->td_ksegrp; 648 mtx_assert(&Giant, MA_NOTOWNED); 649 mtx_lock_spin(&sched_lock); 650 kg->kg_proc->p_stats->p_ru.ru_nvcsw++; 651 sched_prio(td, PRI_MAX_TIMESHARE); 652 mi_switch(); 653 mtx_unlock_spin(&sched_lock); 654 td->td_retval[0] = 0; 655 return (0); 656 } 657