1 /*- 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 1982, 1986, 1990, 1991, 1993 5 * The Regents of the University of California. All rights reserved. 6 * (c) UNIX System Laboratories, Inc. 7 * All or some portions of this file are derived from material licensed 8 * to the University of California by American Telephone and Telegraph 9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 10 * the permission of UNIX System Laboratories, Inc. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95 37 */ 38 39 #include <sys/cdefs.h> 40 __FBSDID("$FreeBSD$"); 41 42 #include "opt_ktrace.h" 43 #include "opt_sched.h" 44 45 #include <sys/param.h> 46 #include <sys/systm.h> 47 #include <sys/condvar.h> 48 #include <sys/kdb.h> 49 #include <sys/kernel.h> 50 #include <sys/ktr.h> 51 #include <sys/lock.h> 52 #include <sys/mutex.h> 53 #include <sys/proc.h> 54 #include <sys/resourcevar.h> 55 #include <sys/sched.h> 56 #include <sys/sdt.h> 57 #include <sys/signalvar.h> 58 #include <sys/sleepqueue.h> 59 #include <sys/smp.h> 60 #include <sys/sx.h> 61 #include <sys/sysctl.h> 62 #include <sys/sysproto.h> 63 #include <sys/vmmeter.h> 64 #ifdef KTRACE 65 #include <sys/uio.h> 66 #include <sys/ktrace.h> 67 #endif 68 69 #include <machine/cpu.h> 70 71 static void synch_setup(void *dummy); 72 SYSINIT(synch_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, synch_setup, 73 NULL); 74 75 int hogticks; 76 static uint8_t pause_wchan[MAXCPU]; 77 78 static struct callout loadav_callout; 79 80 struct 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 */ 86 static 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 */ 93 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, SYSCTL_NULL_INT_PTR, FSCALE, ""); 94 95 static void loadav(void *arg); 96 97 SDT_PROVIDER_DECLARE(sched); 98 SDT_PROBE_DEFINE(sched, , , preempt); 99 100 static void 101 sleepinit(void *unused) 102 { 103 104 hogticks = (hz / 10) * 2; /* Default only. */ 105 init_sleepqueues(); 106 } 107 108 /* 109 * vmem tries to lock the sleepq mutexes when free'ing kva, so make sure 110 * it is available. 111 */ 112 SYSINIT(sleepinit, SI_SUB_KMEM, SI_ORDER_ANY, sleepinit, 0); 113 114 /* 115 * General sleep call. Suspends the current thread until a wakeup is 116 * performed on the specified identifier. The thread will then be made 117 * runnable with the specified priority. Sleeps at most sbt units of time 118 * (0 means no timeout). If pri includes the PCATCH flag, let signals 119 * interrupt the sleep, otherwise ignore them while sleeping. Returns 0 if 120 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 121 * signal becomes pending, ERESTART is returned if the current system 122 * call should be restarted if possible, and EINTR is returned if the system 123 * call should be interrupted by the signal (return EINTR). 124 * 125 * The lock argument is unlocked before the caller is suspended, and 126 * re-locked before _sleep() returns. If priority includes the PDROP 127 * flag the lock is not re-locked before returning. 128 */ 129 int 130 _sleep(void *ident, struct lock_object *lock, int priority, 131 const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags) 132 { 133 struct thread *td; 134 struct proc *p; 135 struct lock_class *class; 136 uintptr_t lock_state; 137 int catch, pri, rval, sleepq_flags; 138 WITNESS_SAVE_DECL(lock_witness); 139 140 td = curthread; 141 p = td->td_proc; 142 #ifdef KTRACE 143 if (KTRPOINT(td, KTR_CSW)) 144 ktrcsw(1, 0, wmesg); 145 #endif 146 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, lock, 147 "Sleeping on \"%s\"", wmesg); 148 KASSERT(sbt != 0 || mtx_owned(&Giant) || lock != NULL, 149 ("sleeping without a lock")); 150 KASSERT(ident != NULL, ("_sleep: NULL ident")); 151 KASSERT(TD_IS_RUNNING(td), ("_sleep: curthread not running")); 152 if (priority & PDROP) 153 KASSERT(lock != NULL && lock != &Giant.lock_object, 154 ("PDROP requires a non-Giant lock")); 155 if (lock != NULL) 156 class = LOCK_CLASS(lock); 157 else 158 class = NULL; 159 160 if (SCHEDULER_STOPPED_TD(td)) { 161 if (lock != NULL && priority & PDROP) 162 class->lc_unlock(lock); 163 return (0); 164 } 165 catch = priority & PCATCH; 166 pri = priority & PRIMASK; 167 168 KASSERT(!TD_ON_SLEEPQ(td), ("recursive sleep")); 169 170 if ((uint8_t *)ident >= &pause_wchan[0] && 171 (uint8_t *)ident <= &pause_wchan[MAXCPU - 1]) 172 sleepq_flags = SLEEPQ_PAUSE; 173 else 174 sleepq_flags = SLEEPQ_SLEEP; 175 if (catch) 176 sleepq_flags |= SLEEPQ_INTERRUPTIBLE; 177 178 sleepq_lock(ident); 179 CTR5(KTR_PROC, "sleep: thread %ld (pid %ld, %s) on %s (%p)", 180 td->td_tid, p->p_pid, td->td_name, wmesg, ident); 181 182 if (lock == &Giant.lock_object) 183 mtx_assert(&Giant, MA_OWNED); 184 DROP_GIANT(); 185 if (lock != NULL && lock != &Giant.lock_object && 186 !(class->lc_flags & LC_SLEEPABLE)) { 187 WITNESS_SAVE(lock, lock_witness); 188 lock_state = class->lc_unlock(lock); 189 } else 190 /* GCC needs to follow the Yellow Brick Road */ 191 lock_state = -1; 192 193 /* 194 * We put ourselves on the sleep queue and start our timeout 195 * before calling thread_suspend_check, as we could stop there, 196 * and a wakeup or a SIGCONT (or both) could occur while we were 197 * stopped without resuming us. Thus, we must be ready for sleep 198 * when cursig() is called. If the wakeup happens while we're 199 * stopped, then td will no longer be on a sleep queue upon 200 * return from cursig(). 201 */ 202 sleepq_add(ident, lock, wmesg, sleepq_flags, 0); 203 if (sbt != 0) 204 sleepq_set_timeout_sbt(ident, sbt, pr, flags); 205 if (lock != NULL && class->lc_flags & LC_SLEEPABLE) { 206 sleepq_release(ident); 207 WITNESS_SAVE(lock, lock_witness); 208 lock_state = class->lc_unlock(lock); 209 sleepq_lock(ident); 210 } 211 if (sbt != 0 && catch) 212 rval = sleepq_timedwait_sig(ident, pri); 213 else if (sbt != 0) 214 rval = sleepq_timedwait(ident, pri); 215 else if (catch) 216 rval = sleepq_wait_sig(ident, pri); 217 else { 218 sleepq_wait(ident, pri); 219 rval = 0; 220 } 221 #ifdef KTRACE 222 if (KTRPOINT(td, KTR_CSW)) 223 ktrcsw(0, 0, wmesg); 224 #endif 225 PICKUP_GIANT(); 226 if (lock != NULL && lock != &Giant.lock_object && !(priority & PDROP)) { 227 class->lc_lock(lock, lock_state); 228 WITNESS_RESTORE(lock, lock_witness); 229 } 230 return (rval); 231 } 232 233 int 234 msleep_spin_sbt(void *ident, struct mtx *mtx, const char *wmesg, 235 sbintime_t sbt, sbintime_t pr, int flags) 236 { 237 struct thread *td; 238 struct proc *p; 239 int rval; 240 WITNESS_SAVE_DECL(mtx); 241 242 td = curthread; 243 p = td->td_proc; 244 KASSERT(mtx != NULL, ("sleeping without a mutex")); 245 KASSERT(ident != NULL, ("msleep_spin_sbt: NULL ident")); 246 KASSERT(TD_IS_RUNNING(td), ("msleep_spin_sbt: curthread not running")); 247 248 if (SCHEDULER_STOPPED_TD(td)) 249 return (0); 250 251 sleepq_lock(ident); 252 CTR5(KTR_PROC, "msleep_spin: thread %ld (pid %ld, %s) on %s (%p)", 253 td->td_tid, p->p_pid, td->td_name, wmesg, ident); 254 255 DROP_GIANT(); 256 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 257 WITNESS_SAVE(&mtx->lock_object, mtx); 258 mtx_unlock_spin(mtx); 259 260 /* 261 * We put ourselves on the sleep queue and start our timeout. 262 */ 263 sleepq_add(ident, &mtx->lock_object, wmesg, SLEEPQ_SLEEP, 0); 264 if (sbt != 0) 265 sleepq_set_timeout_sbt(ident, sbt, pr, flags); 266 267 /* 268 * Can't call ktrace with any spin locks held so it can lock the 269 * ktrace_mtx lock, and WITNESS_WARN considers it an error to hold 270 * any spin lock. Thus, we have to drop the sleepq spin lock while 271 * we handle those requests. This is safe since we have placed our 272 * thread on the sleep queue already. 273 */ 274 #ifdef KTRACE 275 if (KTRPOINT(td, KTR_CSW)) { 276 sleepq_release(ident); 277 ktrcsw(1, 0, wmesg); 278 sleepq_lock(ident); 279 } 280 #endif 281 #ifdef WITNESS 282 sleepq_release(ident); 283 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "Sleeping on \"%s\"", 284 wmesg); 285 sleepq_lock(ident); 286 #endif 287 if (sbt != 0) 288 rval = sleepq_timedwait(ident, 0); 289 else { 290 sleepq_wait(ident, 0); 291 rval = 0; 292 } 293 #ifdef KTRACE 294 if (KTRPOINT(td, KTR_CSW)) 295 ktrcsw(0, 0, wmesg); 296 #endif 297 PICKUP_GIANT(); 298 mtx_lock_spin(mtx); 299 WITNESS_RESTORE(&mtx->lock_object, mtx); 300 return (rval); 301 } 302 303 /* 304 * pause() delays the calling thread by the given number of system ticks. 305 * During cold bootup, pause() uses the DELAY() function instead of 306 * the tsleep() function to do the waiting. The "timo" argument must be 307 * greater than or equal to zero. A "timo" value of zero is equivalent 308 * to a "timo" value of one. 309 */ 310 int 311 pause_sbt(const char *wmesg, sbintime_t sbt, sbintime_t pr, int flags) 312 { 313 KASSERT(sbt >= 0, ("pause: timeout must be >= 0")); 314 315 /* silently convert invalid timeouts */ 316 if (sbt == 0) 317 sbt = tick_sbt; 318 319 if ((cold && curthread == &thread0) || kdb_active || 320 SCHEDULER_STOPPED()) { 321 /* 322 * We delay one second at a time to avoid overflowing the 323 * system specific DELAY() function(s): 324 */ 325 while (sbt >= SBT_1S) { 326 DELAY(1000000); 327 sbt -= SBT_1S; 328 } 329 /* Do the delay remainder, if any */ 330 sbt = howmany(sbt, SBT_1US); 331 if (sbt > 0) 332 DELAY(sbt); 333 return (0); 334 } 335 return (_sleep(&pause_wchan[curcpu], NULL, 0, wmesg, sbt, pr, flags)); 336 } 337 338 /* 339 * Make all threads sleeping on the specified identifier runnable. 340 */ 341 void 342 wakeup(void *ident) 343 { 344 int wakeup_swapper; 345 346 sleepq_lock(ident); 347 wakeup_swapper = sleepq_broadcast(ident, SLEEPQ_SLEEP, 0, 0); 348 sleepq_release(ident); 349 if (wakeup_swapper) { 350 KASSERT(ident != &proc0, 351 ("wakeup and wakeup_swapper and proc0")); 352 kick_proc0(); 353 } 354 } 355 356 /* 357 * Make a thread sleeping on the specified identifier runnable. 358 * May wake more than one thread if a target thread is currently 359 * swapped out. 360 */ 361 void 362 wakeup_one(void *ident) 363 { 364 int wakeup_swapper; 365 366 sleepq_lock(ident); 367 wakeup_swapper = sleepq_signal(ident, SLEEPQ_SLEEP, 0, 0); 368 sleepq_release(ident); 369 if (wakeup_swapper) 370 kick_proc0(); 371 } 372 373 static void 374 kdb_switch(void) 375 { 376 thread_unlock(curthread); 377 kdb_backtrace(); 378 kdb_reenter(); 379 panic("%s: did not reenter debugger", __func__); 380 } 381 382 /* 383 * The machine independent parts of context switching. 384 */ 385 void 386 mi_switch(int flags, struct thread *newtd) 387 { 388 uint64_t runtime, new_switchtime; 389 struct thread *td; 390 391 td = curthread; /* XXX */ 392 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED); 393 KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code")); 394 #ifdef INVARIANTS 395 if (!TD_ON_LOCK(td) && !TD_IS_RUNNING(td)) 396 mtx_assert(&Giant, MA_NOTOWNED); 397 #endif 398 KASSERT(td->td_critnest == 1 || panicstr, 399 ("mi_switch: switch in a critical section")); 400 KASSERT((flags & (SW_INVOL | SW_VOL)) != 0, 401 ("mi_switch: switch must be voluntary or involuntary")); 402 KASSERT(newtd != curthread, ("mi_switch: preempting back to ourself")); 403 404 /* 405 * Don't perform context switches from the debugger. 406 */ 407 if (kdb_active) 408 kdb_switch(); 409 if (SCHEDULER_STOPPED_TD(td)) 410 return; 411 if (flags & SW_VOL) { 412 td->td_ru.ru_nvcsw++; 413 td->td_swvoltick = ticks; 414 } else { 415 td->td_ru.ru_nivcsw++; 416 td->td_swinvoltick = ticks; 417 } 418 #ifdef SCHED_STATS 419 SCHED_STAT_INC(sched_switch_stats[flags & SW_TYPE_MASK]); 420 #endif 421 /* 422 * Compute the amount of time during which the current 423 * thread was running, and add that to its total so far. 424 */ 425 new_switchtime = cpu_ticks(); 426 runtime = new_switchtime - PCPU_GET(switchtime); 427 td->td_runtime += runtime; 428 td->td_incruntime += runtime; 429 PCPU_SET(switchtime, new_switchtime); 430 td->td_generation++; /* bump preempt-detect counter */ 431 VM_CNT_INC(v_swtch); 432 PCPU_SET(switchticks, ticks); 433 CTR4(KTR_PROC, "mi_switch: old thread %ld (td_sched %p, pid %ld, %s)", 434 td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name); 435 #ifdef KDTRACE_HOOKS 436 if ((flags & SW_PREEMPT) != 0 || ((flags & SW_INVOL) != 0 && 437 (flags & SW_TYPE_MASK) == SWT_NEEDRESCHED)) 438 SDT_PROBE0(sched, , , preempt); 439 #endif 440 sched_switch(td, newtd, flags); 441 CTR4(KTR_PROC, "mi_switch: new thread %ld (td_sched %p, pid %ld, %s)", 442 td->td_tid, td_get_sched(td), td->td_proc->p_pid, td->td_name); 443 444 /* 445 * If the last thread was exiting, finish cleaning it up. 446 */ 447 if ((td = PCPU_GET(deadthread))) { 448 PCPU_SET(deadthread, NULL); 449 thread_stash(td); 450 } 451 } 452 453 /* 454 * Change thread state to be runnable, placing it on the run queue if 455 * it is in memory. If it is swapped out, return true so our caller 456 * will know to awaken the swapper. 457 */ 458 int 459 setrunnable(struct thread *td) 460 { 461 462 THREAD_LOCK_ASSERT(td, MA_OWNED); 463 KASSERT(td->td_proc->p_state != PRS_ZOMBIE, 464 ("setrunnable: pid %d is a zombie", td->td_proc->p_pid)); 465 switch (td->td_state) { 466 case TDS_RUNNING: 467 case TDS_RUNQ: 468 return (0); 469 case TDS_INHIBITED: 470 /* 471 * If we are only inhibited because we are swapped out 472 * then arange to swap in this process. Otherwise just return. 473 */ 474 if (td->td_inhibitors != TDI_SWAPPED) 475 return (0); 476 /* FALLTHROUGH */ 477 case TDS_CAN_RUN: 478 break; 479 default: 480 printf("state is 0x%x", td->td_state); 481 panic("setrunnable(2)"); 482 } 483 if ((td->td_flags & TDF_INMEM) == 0) { 484 if ((td->td_flags & TDF_SWAPINREQ) == 0) { 485 td->td_flags |= TDF_SWAPINREQ; 486 return (1); 487 } 488 } else 489 sched_wakeup(td); 490 return (0); 491 } 492 493 /* 494 * Compute a tenex style load average of a quantity on 495 * 1, 5 and 15 minute intervals. 496 */ 497 static void 498 loadav(void *arg) 499 { 500 int i, nrun; 501 struct loadavg *avg; 502 503 nrun = sched_load(); 504 avg = &averunnable; 505 506 for (i = 0; i < 3; i++) 507 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 508 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 509 510 /* 511 * Schedule the next update to occur after 5 seconds, but add a 512 * random variation to avoid synchronisation with processes that 513 * run at regular intervals. 514 */ 515 callout_reset_sbt(&loadav_callout, 516 SBT_1US * (4000000 + (int)(random() % 2000001)), SBT_1US, 517 loadav, NULL, C_DIRECT_EXEC | C_PREL(32)); 518 } 519 520 /* ARGSUSED */ 521 static void 522 synch_setup(void *dummy) 523 { 524 callout_init(&loadav_callout, 1); 525 526 /* Kick off timeout driven events by calling first time. */ 527 loadav(NULL); 528 } 529 530 int 531 should_yield(void) 532 { 533 534 return ((u_int)ticks - (u_int)curthread->td_swvoltick >= hogticks); 535 } 536 537 void 538 maybe_yield(void) 539 { 540 541 if (should_yield()) 542 kern_yield(PRI_USER); 543 } 544 545 void 546 kern_yield(int prio) 547 { 548 struct thread *td; 549 550 td = curthread; 551 DROP_GIANT(); 552 thread_lock(td); 553 if (prio == PRI_USER) 554 prio = td->td_user_pri; 555 if (prio >= 0) 556 sched_prio(td, prio); 557 mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 558 thread_unlock(td); 559 PICKUP_GIANT(); 560 } 561 562 /* 563 * General purpose yield system call. 564 */ 565 int 566 sys_yield(struct thread *td, struct yield_args *uap) 567 { 568 569 thread_lock(td); 570 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 571 sched_prio(td, PRI_MAX_TIMESHARE); 572 mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 573 thread_unlock(td); 574 td->td_retval[0] = 0; 575 return (0); 576 } 577