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 * $FreeBSD$ 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 70 static void sched_setup(void *dummy); 71 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 72 73 int hogticks; 74 int lbolt; 75 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 76 77 static struct callout loadav_callout; 78 static struct callout schedcpu_callout; 79 static struct callout roundrobin_callout; 80 81 struct loadavg averunnable = 82 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 83 /* 84 * Constants for averages over 1, 5, and 15 minutes 85 * when sampling at 5 second intervals. 86 */ 87 static fixpt_t cexp[3] = { 88 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 89 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 90 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 91 }; 92 93 static void endtsleep(void *); 94 static void loadav(void *arg); 95 static void roundrobin(void *arg); 96 static void schedcpu(void *arg); 97 98 static int 99 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 100 { 101 int error, new_val; 102 103 new_val = sched_quantum * tick; 104 error = sysctl_handle_int(oidp, &new_val, 0, req); 105 if (error != 0 || req->newptr == NULL) 106 return (error); 107 if (new_val < tick) 108 return (EINVAL); 109 sched_quantum = new_val / tick; 110 hogticks = 2 * sched_quantum; 111 return (0); 112 } 113 114 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 115 0, sizeof sched_quantum, sysctl_kern_quantum, "I", 116 "Roundrobin scheduling quantum in microseconds"); 117 118 /* 119 * Arrange to reschedule if necessary, taking the priorities and 120 * schedulers into account. 121 */ 122 void 123 maybe_resched(struct thread *td) 124 { 125 126 mtx_assert(&sched_lock, MA_OWNED); 127 if (td->td_priority < curthread->td_priority) 128 curthread->td_kse->ke_flags |= KEF_NEEDRESCHED; 129 } 130 131 int 132 roundrobin_interval(void) 133 { 134 return (sched_quantum); 135 } 136 137 /* 138 * Force switch among equal priority processes every 100ms. 139 * We don't actually need to force a context switch of the current process. 140 * The act of firing the event triggers a context switch to softclock() and 141 * then switching back out again which is equivalent to a preemption, thus 142 * no further work is needed on the local CPU. 143 */ 144 /* ARGSUSED */ 145 static void 146 roundrobin(arg) 147 void *arg; 148 { 149 150 #ifdef SMP 151 mtx_lock_spin(&sched_lock); 152 forward_roundrobin(); 153 mtx_unlock_spin(&sched_lock); 154 #endif 155 156 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 157 } 158 159 /* 160 * Constants for digital decay and forget: 161 * 90% of (p_estcpu) usage in 5 * loadav time 162 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 163 * Note that, as ps(1) mentions, this can let percentages 164 * total over 100% (I've seen 137.9% for 3 processes). 165 * 166 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 167 * 168 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 169 * That is, the system wants to compute a value of decay such 170 * that the following for loop: 171 * for (i = 0; i < (5 * loadavg); i++) 172 * p_estcpu *= decay; 173 * will compute 174 * p_estcpu *= 0.1; 175 * for all values of loadavg: 176 * 177 * Mathematically this loop can be expressed by saying: 178 * decay ** (5 * loadavg) ~= .1 179 * 180 * The system computes decay as: 181 * decay = (2 * loadavg) / (2 * loadavg + 1) 182 * 183 * We wish to prove that the system's computation of decay 184 * will always fulfill the equation: 185 * decay ** (5 * loadavg) ~= .1 186 * 187 * If we compute b as: 188 * b = 2 * loadavg 189 * then 190 * decay = b / (b + 1) 191 * 192 * We now need to prove two things: 193 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 194 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 195 * 196 * Facts: 197 * For x close to zero, exp(x) =~ 1 + x, since 198 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 199 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 200 * For x close to zero, ln(1+x) =~ x, since 201 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 202 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 203 * ln(.1) =~ -2.30 204 * 205 * Proof of (1): 206 * Solve (factor)**(power) =~ .1 given power (5*loadav): 207 * solving for factor, 208 * ln(factor) =~ (-2.30/5*loadav), or 209 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 210 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 211 * 212 * Proof of (2): 213 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 214 * solving for power, 215 * power*ln(b/(b+1)) =~ -2.30, or 216 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 217 * 218 * Actual power values for the implemented algorithm are as follows: 219 * loadav: 1 2 3 4 220 * power: 5.68 10.32 14.94 19.55 221 */ 222 223 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 224 #define loadfactor(loadav) (2 * (loadav)) 225 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 226 227 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 228 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 229 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 230 231 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 232 static int fscale __unused = FSCALE; 233 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 234 235 /* 236 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 237 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 238 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 239 * 240 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 241 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 242 * 243 * If you don't want to bother with the faster/more-accurate formula, you 244 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 245 * (more general) method of calculating the %age of CPU used by a process. 246 */ 247 #define CCPU_SHIFT 11 248 249 /* 250 * Recompute process priorities, every hz ticks. 251 * MP-safe, called without the Giant mutex. 252 */ 253 /* ARGSUSED */ 254 static void 255 schedcpu(arg) 256 void *arg; 257 { 258 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 259 struct thread *td; 260 struct proc *p; 261 struct kse *ke; 262 struct ksegrp *kg; 263 int realstathz; 264 int awake; 265 266 realstathz = stathz ? stathz : hz; 267 sx_slock(&allproc_lock); 268 FOREACH_PROC_IN_SYSTEM(p) { 269 mtx_lock_spin(&sched_lock); 270 p->p_swtime++; 271 FOREACH_KSEGRP_IN_PROC(p, kg) { 272 awake = 0; 273 FOREACH_KSE_IN_GROUP(kg, ke) { 274 /* 275 * Increment time in/out of memory and sleep 276 * time (if sleeping). We ignore overflow; 277 * with 16-bit int's (remember them?) 278 * overflow takes 45 days. 279 */ 280 /* XXXKSE **WRONG***/ 281 /* 282 * the kse slptimes are not touched in wakeup 283 * because the thread may not HAVE a KSE 284 */ 285 if (ke->ke_state == KES_ONRUNQ && 286 ke->ke_state == KES_RUNNING) { 287 ke->ke_slptime++; 288 } else { 289 ke->ke_slptime = 0; 290 awake = 1; 291 } 292 293 /* 294 * pctcpu is only for ps? 295 * Do it per kse.. and add them up at the end? 296 * XXXKSE 297 */ 298 ke->ke_pctcpu = (ke->ke_pctcpu * ccpu) >> FSHIFT; 299 /* 300 * If the kse has been idle the entire second, 301 * stop recalculating its priority until 302 * it wakes up. 303 */ 304 if (ke->ke_slptime > 1) { 305 continue; 306 } 307 308 #if (FSHIFT >= CCPU_SHIFT) 309 ke->ke_pctcpu += (realstathz == 100) ? 310 ((fixpt_t) ke->ke_cpticks) << 311 (FSHIFT - CCPU_SHIFT) : 312 100 * (((fixpt_t) ke->ke_cpticks) << 313 (FSHIFT - CCPU_SHIFT)) / realstathz; 314 #else 315 ke->ke_pctcpu += ((FSCALE - ccpu) * 316 (ke->ke_cpticks * FSCALE / realstathz)) >> 317 FSHIFT; 318 #endif 319 ke->ke_cpticks = 0; 320 } /* end of kse loop */ 321 if (awake == 0) { 322 kg->kg_slptime++; 323 } else { 324 kg->kg_slptime = 0; 325 } 326 kg->kg_estcpu = decay_cpu(loadfac, kg->kg_estcpu); 327 resetpriority(kg); 328 FOREACH_THREAD_IN_GROUP(kg, td) { 329 int changedqueue; 330 if (td->td_priority >= PUSER) { 331 /* 332 * Only change the priority 333 * of threads that are still at their 334 * user priority. 335 * XXXKSE This is problematic 336 * as we may need to re-order 337 * the threads on the KSEG list. 338 */ 339 changedqueue = 340 ((td->td_priority / RQ_PPQ) != 341 (kg->kg_user_pri / RQ_PPQ)); 342 343 td->td_priority = kg->kg_user_pri; 344 if (changedqueue && 345 td->td_state == TDS_RUNQ) { 346 /* this could be optimised */ 347 remrunqueue(td); 348 td->td_priority = 349 kg->kg_user_pri; 350 setrunqueue(td); 351 } else { 352 td->td_priority = kg->kg_user_pri; 353 } 354 } 355 } 356 } /* end of ksegrp loop */ 357 mtx_unlock_spin(&sched_lock); 358 } /* end of process loop */ 359 sx_sunlock(&allproc_lock); 360 wakeup(&lbolt); 361 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 362 } 363 364 /* 365 * Recalculate the priority of a process after it has slept for a while. 366 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 367 * least six times the loadfactor will decay p_estcpu to zero. 368 */ 369 void 370 updatepri(td) 371 register struct thread *td; 372 { 373 register struct ksegrp *kg; 374 register unsigned int newcpu; 375 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 376 377 if (td == NULL) 378 return; 379 kg = td->td_ksegrp; 380 newcpu = kg->kg_estcpu; 381 if (kg->kg_slptime > 5 * loadfac) 382 kg->kg_estcpu = 0; 383 else { 384 kg->kg_slptime--; /* the first time was done in schedcpu */ 385 while (newcpu && --kg->kg_slptime) 386 newcpu = decay_cpu(loadfac, newcpu); 387 kg->kg_estcpu = newcpu; 388 } 389 resetpriority(td->td_ksegrp); 390 } 391 392 /* 393 * We're only looking at 7 bits of the address; everything is 394 * aligned to 4, lots of things are aligned to greater powers 395 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 396 */ 397 #define TABLESIZE 128 398 static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE]; 399 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 400 401 void 402 sleepinit(void) 403 { 404 int i; 405 406 sched_quantum = hz/10; 407 hogticks = 2 * sched_quantum; 408 for (i = 0; i < TABLESIZE; i++) 409 TAILQ_INIT(&slpque[i]); 410 } 411 412 /* 413 * General sleep call. Suspends the current process until a wakeup is 414 * performed on the specified identifier. The process will then be made 415 * runnable with the specified priority. Sleeps at most timo/hz seconds 416 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 417 * before and after sleeping, else signals are not checked. Returns 0 if 418 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 419 * signal needs to be delivered, ERESTART is returned if the current system 420 * call should be restarted if possible, and EINTR is returned if the system 421 * call should be interrupted by the signal (return EINTR). 422 * 423 * The mutex argument is exited before the caller is suspended, and 424 * entered before msleep returns. If priority includes the PDROP 425 * flag the mutex is not entered before returning. 426 */ 427 428 int 429 msleep(ident, mtx, priority, wmesg, timo) 430 void *ident; 431 struct mtx *mtx; 432 int priority, timo; 433 const char *wmesg; 434 { 435 struct thread *td = curthread; 436 struct proc *p = td->td_proc; 437 int sig, catch = priority & PCATCH; 438 int rval = 0; 439 WITNESS_SAVE_DECL(mtx); 440 441 #ifdef KTRACE 442 if (KTRPOINT(td, KTR_CSW)) 443 ktrcsw(1, 0); 444 #endif 445 KASSERT((td->td_kse != NULL), ("msleep: NULL KSE?")); 446 KASSERT((td->td_kse->ke_state == KES_RUNNING), ("msleep: kse state?")); 447 WITNESS_SLEEP(0, &mtx->mtx_object); 448 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 449 ("sleeping without a mutex")); 450 /* 451 * If we are capable of async syscalls and there isn't already 452 * another one ready to return, start a new thread 453 * and queue it as ready to run. Note that there is danger here 454 * because we need to make sure that we don't sleep allocating 455 * the thread (recursion here might be bad). 456 * Hence the TDF_INMSLEEP flag. 457 */ 458 if (p->p_flag & P_KSES) { 459 /* Just don't bother if we are exiting 460 and not the exiting thread. */ 461 if ((p->p_flag & P_WEXIT) && catch && p->p_singlethread != td) 462 return (EINTR); 463 if (td->td_mailbox && (!(td->td_flags & TDF_INMSLEEP))) { 464 /* 465 * If we have no queued work to do, then 466 * upcall to the UTS to see if it has more to do. 467 * We don't need to upcall now, just make it and 468 * queue it. 469 */ 470 mtx_lock_spin(&sched_lock); 471 if (TAILQ_FIRST(&td->td_ksegrp->kg_runq) == NULL) { 472 /* Don't recurse here! */ 473 KASSERT((td->td_kse->ke_state == KES_RUNNING), ("msleep: kse stateX?")); 474 td->td_flags |= TDF_INMSLEEP; 475 thread_schedule_upcall(td, td->td_kse); 476 td->td_flags &= ~TDF_INMSLEEP; 477 KASSERT((td->td_kse->ke_state == KES_RUNNING), ("msleep: kse stateY?")); 478 } 479 mtx_unlock_spin(&sched_lock); 480 } 481 KASSERT((td->td_kse != NULL), ("msleep: NULL KSE2?")); 482 KASSERT((td->td_kse->ke_state == KES_RUNNING), 483 ("msleep: kse state2?")); 484 KASSERT((td->td_kse->ke_thread == td), 485 ("msleep: kse/thread mismatch?")); 486 } 487 mtx_lock_spin(&sched_lock); 488 if (cold || panicstr) { 489 /* 490 * After a panic, or during autoconfiguration, 491 * just give interrupts a chance, then just return; 492 * don't run any other procs or panic below, 493 * in case this is the idle process and already asleep. 494 */ 495 if (mtx != NULL && priority & PDROP) 496 mtx_unlock(mtx); 497 mtx_unlock_spin(&sched_lock); 498 return (0); 499 } 500 501 DROP_GIANT(); 502 503 if (mtx != NULL) { 504 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 505 WITNESS_SAVE(&mtx->mtx_object, mtx); 506 mtx_unlock(mtx); 507 if (priority & PDROP) 508 mtx = NULL; 509 } 510 511 KASSERT(p != NULL, ("msleep1")); 512 KASSERT(ident != NULL && td->td_state == TDS_RUNNING, ("msleep")); 513 514 td->td_wchan = ident; 515 td->td_wmesg = wmesg; 516 td->td_kse->ke_slptime = 0; /* XXXKSE */ 517 td->td_ksegrp->kg_slptime = 0; 518 td->td_priority = priority & PRIMASK; 519 CTR5(KTR_PROC, "msleep: thread %p (pid %d, %s) on %s (%p)", 520 td, p->p_pid, p->p_comm, wmesg, ident); 521 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], td, td_slpq); 522 if (timo) 523 callout_reset(&td->td_slpcallout, timo, endtsleep, td); 524 /* 525 * We put ourselves on the sleep queue and start our timeout 526 * before calling thread_suspend_check, as we could stop there, and 527 * a wakeup or a SIGCONT (or both) could occur while we were stopped. 528 * without resuming us, thus we must be ready for sleep 529 * when cursig is called. If the wakeup happens while we're 530 * stopped, td->td_wchan will be 0 upon return from cursig. 531 */ 532 if (catch) { 533 CTR3(KTR_PROC, "msleep caught: thread %p (pid %d, %s)", td, 534 p->p_pid, p->p_comm); 535 td->td_flags |= TDF_SINTR; 536 mtx_unlock_spin(&sched_lock); 537 PROC_LOCK(p); 538 sig = cursig(td); 539 if (sig == 0) { 540 if (thread_suspend_check(1)) { 541 sig = SIGSTOP; 542 } 543 } 544 mtx_lock_spin(&sched_lock); 545 PROC_UNLOCK(p); 546 if (sig != 0) { 547 if (td->td_wchan != NULL) 548 unsleep(td); 549 } else if (td->td_wchan == NULL) 550 catch = 0; 551 } else { 552 sig = 0; 553 } 554 if (td->td_wchan != NULL) { 555 p->p_stats->p_ru.ru_nvcsw++; 556 td->td_state = TDS_SLP; 557 mi_switch(); 558 } 559 CTR3(KTR_PROC, "msleep resume: thread %p (pid %d, %s)", td, p->p_pid, 560 p->p_comm); 561 KASSERT(td->td_state == TDS_RUNNING, ("running but not TDS_RUNNING")); 562 td->td_flags &= ~TDF_SINTR; 563 if (td->td_flags & TDF_TIMEOUT) { 564 td->td_flags &= ~TDF_TIMEOUT; 565 if (sig == 0) 566 rval = EWOULDBLOCK; 567 } else if (td->td_flags & TDF_TIMOFAIL) { 568 td->td_flags &= ~TDF_TIMOFAIL; 569 } else if (timo && callout_stop(&td->td_slpcallout) == 0) { 570 /* 571 * This isn't supposed to be pretty. If we are here, then 572 * the endtsleep() callout is currently executing on another 573 * CPU and is either spinning on the sched_lock or will be 574 * soon. If we don't synchronize here, there is a chance 575 * that this process may msleep() again before the callout 576 * has a chance to run and the callout may end up waking up 577 * the wrong msleep(). Yuck. 578 */ 579 td->td_flags |= TDF_TIMEOUT; 580 td->td_state = TDS_SLP; 581 p->p_stats->p_ru.ru_nivcsw++; 582 mi_switch(); 583 } 584 mtx_unlock_spin(&sched_lock); 585 586 if (rval == 0 && catch) { 587 PROC_LOCK(p); 588 /* XXX: shouldn't we always be calling cursig() */ 589 if (sig != 0 || (sig = cursig(td))) { 590 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 591 rval = EINTR; 592 else 593 rval = ERESTART; 594 } 595 PROC_UNLOCK(p); 596 } 597 #ifdef KTRACE 598 if (KTRPOINT(td, KTR_CSW)) 599 ktrcsw(0, 0); 600 #endif 601 PICKUP_GIANT(); 602 if (mtx != NULL) { 603 mtx_lock(mtx); 604 WITNESS_RESTORE(&mtx->mtx_object, mtx); 605 } 606 return (rval); 607 } 608 609 /* 610 * Implement timeout for msleep() 611 * 612 * If process hasn't been awakened (wchan non-zero), 613 * set timeout flag and undo the sleep. If proc 614 * is stopped, just unsleep so it will remain stopped. 615 * MP-safe, called without the Giant mutex. 616 */ 617 static void 618 endtsleep(arg) 619 void *arg; 620 { 621 register struct thread *td = arg; 622 623 CTR3(KTR_PROC, "endtsleep: thread %p (pid %d, %s)", td, td->td_proc->p_pid, 624 td->td_proc->p_comm); 625 mtx_lock_spin(&sched_lock); 626 /* 627 * This is the other half of the synchronization with msleep() 628 * described above. If the PS_TIMEOUT flag is set, we lost the 629 * race and just need to put the process back on the runqueue. 630 */ 631 if ((td->td_flags & TDF_TIMEOUT) != 0) { 632 td->td_flags &= ~TDF_TIMEOUT; 633 setrunqueue(td); 634 } else if (td->td_wchan != NULL) { 635 if (td->td_state == TDS_SLP) /* XXXKSE */ 636 setrunnable(td); 637 else 638 unsleep(td); 639 td->td_flags |= TDF_TIMEOUT; 640 } else { 641 td->td_flags |= TDF_TIMOFAIL; 642 } 643 mtx_unlock_spin(&sched_lock); 644 } 645 646 /* 647 * Abort a thread, as if an interrupt had occured. Only abort 648 * interruptable waits (unfortunatly it isn't only safe to abort others). 649 * This is about identical to cv_abort(). 650 * Think about merging them? 651 * Also, whatever the signal code does... 652 */ 653 void 654 abortsleep(struct thread *td) 655 { 656 657 mtx_lock_spin(&sched_lock); 658 /* 659 * If the TDF_TIMEOUT flag is set, just leave. A 660 * timeout is scheduled anyhow. 661 */ 662 if ((td->td_flags & (TDF_TIMEOUT | TDF_SINTR)) == TDF_SINTR) { 663 if (td->td_wchan != NULL) { 664 if (td->td_state == TDS_SLP) { /* XXXKSE */ 665 setrunnable(td); 666 } else { 667 /* 668 * Probably in a suspended state.. 669 * um.. dunno XXXKSE 670 */ 671 unsleep(td); 672 } 673 } 674 } 675 mtx_unlock_spin(&sched_lock); 676 } 677 678 /* 679 * Remove a process from its wait queue 680 */ 681 void 682 unsleep(struct thread *td) 683 { 684 685 mtx_lock_spin(&sched_lock); 686 if (td->td_wchan != NULL) { 687 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_slpq); 688 td->td_wchan = NULL; 689 } 690 mtx_unlock_spin(&sched_lock); 691 } 692 693 /* 694 * Make all processes sleeping on the specified identifier runnable. 695 */ 696 void 697 wakeup(ident) 698 register void *ident; 699 { 700 register struct slpquehead *qp; 701 register struct thread *td; 702 struct thread *ntd; 703 struct proc *p; 704 705 mtx_lock_spin(&sched_lock); 706 qp = &slpque[LOOKUP(ident)]; 707 restart: 708 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 709 ntd = TAILQ_NEXT(td, td_slpq); 710 p = td->td_proc; 711 if (td->td_wchan == ident) { 712 TAILQ_REMOVE(qp, td, td_slpq); 713 td->td_wchan = NULL; 714 if (td->td_state == TDS_SLP) { 715 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 716 CTR3(KTR_PROC, "wakeup: thread %p (pid %d, %s)", 717 td, p->p_pid, p->p_comm); 718 if (td->td_ksegrp->kg_slptime > 1) 719 updatepri(td); 720 td->td_ksegrp->kg_slptime = 0; 721 if (p->p_sflag & PS_INMEM) { 722 setrunqueue(td); 723 maybe_resched(td); 724 } else { 725 /* XXXKSE Wrong! */ td->td_state = TDS_RUNQ; 726 p->p_sflag |= PS_SWAPINREQ; 727 wakeup(&proc0); 728 } 729 /* END INLINE EXPANSION */ 730 } 731 goto restart; 732 } 733 } 734 mtx_unlock_spin(&sched_lock); 735 } 736 737 /* 738 * Make a process sleeping on the specified identifier runnable. 739 * May wake more than one process if a target process is currently 740 * swapped out. 741 */ 742 void 743 wakeup_one(ident) 744 register void *ident; 745 { 746 register struct slpquehead *qp; 747 register struct thread *td; 748 register struct proc *p; 749 struct thread *ntd; 750 751 mtx_lock_spin(&sched_lock); 752 qp = &slpque[LOOKUP(ident)]; 753 restart: 754 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 755 ntd = TAILQ_NEXT(td, td_slpq); 756 p = td->td_proc; 757 if (td->td_wchan == ident) { 758 TAILQ_REMOVE(qp, td, td_slpq); 759 td->td_wchan = NULL; 760 if (td->td_state == TDS_SLP) { 761 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 762 CTR3(KTR_PROC,"wakeup1: thread %p (pid %d, %s)", 763 td, p->p_pid, p->p_comm); 764 if (td->td_ksegrp->kg_slptime > 1) 765 updatepri(td); 766 td->td_ksegrp->kg_slptime = 0; 767 if (p->p_sflag & PS_INMEM) { 768 setrunqueue(td); 769 maybe_resched(td); 770 break; 771 } else { 772 /* XXXKSE Wrong */ td->td_state = TDS_RUNQ; 773 p->p_sflag |= PS_SWAPINREQ; 774 wakeup(&proc0); 775 } 776 /* END INLINE EXPANSION */ 777 goto restart; 778 } 779 } 780 } 781 mtx_unlock_spin(&sched_lock); 782 } 783 784 /* 785 * The machine independent parts of mi_switch(). 786 */ 787 void 788 mi_switch() 789 { 790 struct bintime new_switchtime; 791 struct thread *td = curthread; /* XXX */ 792 struct proc *p = td->td_proc; /* XXX */ 793 struct kse *ke = td->td_kse; 794 #if 0 795 register struct rlimit *rlim; 796 #endif 797 u_int sched_nest; 798 799 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 800 KASSERT((ke->ke_state == KES_RUNNING), ("mi_switch: kse state?")); 801 #ifdef INVARIANTS 802 if (td->td_state != TDS_MTX && 803 td->td_state != TDS_RUNQ && 804 td->td_state != TDS_RUNNING) 805 mtx_assert(&Giant, MA_NOTOWNED); 806 #endif 807 808 /* 809 * Compute the amount of time during which the current 810 * process was running, and add that to its total so far. 811 */ 812 binuptime(&new_switchtime); 813 bintime_add(&p->p_runtime, &new_switchtime); 814 bintime_sub(&p->p_runtime, PCPU_PTR(switchtime)); 815 816 #ifdef DDB 817 /* 818 * Don't perform context switches from the debugger. 819 */ 820 if (db_active) { 821 mtx_unlock_spin(&sched_lock); 822 db_error("Context switches not allowed in the debugger."); 823 } 824 #endif 825 826 #if 0 827 /* 828 * Check if the process exceeds its cpu resource allocation. 829 * If over max, kill it. 830 * 831 * XXX drop sched_lock, pickup Giant 832 */ 833 if (p->p_state != PRS_ZOMBIE && 834 p->p_limit->p_cpulimit != RLIM_INFINITY && 835 p->p_runtime > p->p_limit->p_cpulimit) { 836 rlim = &p->p_rlimit[RLIMIT_CPU]; 837 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 838 mtx_unlock_spin(&sched_lock); 839 PROC_LOCK(p); 840 killproc(p, "exceeded maximum CPU limit"); 841 mtx_lock_spin(&sched_lock); 842 PROC_UNLOCK(p); 843 } else { 844 mtx_unlock_spin(&sched_lock); 845 PROC_LOCK(p); 846 psignal(p, SIGXCPU); 847 mtx_lock_spin(&sched_lock); 848 PROC_UNLOCK(p); 849 if (rlim->rlim_cur < rlim->rlim_max) { 850 /* XXX: we should make a private copy */ 851 rlim->rlim_cur += 5; 852 } 853 } 854 } 855 #endif 856 857 /* 858 * Pick a new current process and record its start time. 859 */ 860 cnt.v_swtch++; 861 PCPU_SET(switchtime, new_switchtime); 862 CTR3(KTR_PROC, "mi_switch: old thread %p (pid %d, %s)", td, p->p_pid, 863 p->p_comm); 864 sched_nest = sched_lock.mtx_recurse; 865 td->td_lastcpu = ke->ke_oncpu; 866 ke->ke_oncpu = NOCPU; 867 ke->ke_flags &= ~KEF_NEEDRESCHED; 868 /* 869 * At the last moment: if this KSE is not on the run queue, 870 * it needs to be freed correctly and the thread treated accordingly. 871 */ 872 if ((td->td_state == TDS_RUNNING) && 873 ((ke->ke_flags & KEF_IDLEKSE) == 0)) { 874 /* Put us back on the run queue (kse and all). */ 875 setrunqueue(td); 876 } else if ((td->td_flags & TDF_UNBOUND) && 877 (td->td_state != TDS_RUNQ)) { /* in case of old code */ 878 /* 879 * We will not be on the run queue. 880 * Someone else can use the KSE if they need it. 881 */ 882 td->td_kse = NULL; 883 kse_reassign(ke); 884 } 885 cpu_switch(); 886 td->td_kse->ke_oncpu = PCPU_GET(cpuid); 887 td->td_kse->ke_state = KES_RUNNING; 888 sched_lock.mtx_recurse = sched_nest; 889 sched_lock.mtx_lock = (uintptr_t)td; 890 CTR3(KTR_PROC, "mi_switch: new thread %p (pid %d, %s)", td, p->p_pid, 891 p->p_comm); 892 if (PCPU_GET(switchtime.sec) == 0) 893 binuptime(PCPU_PTR(switchtime)); 894 PCPU_SET(switchticks, ticks); 895 } 896 897 /* 898 * Change process state to be runnable, 899 * placing it on the run queue if it is in memory, 900 * and awakening the swapper if it isn't in memory. 901 */ 902 void 903 setrunnable(struct thread *td) 904 { 905 struct proc *p = td->td_proc; 906 907 mtx_assert(&sched_lock, MA_OWNED); 908 switch (p->p_state) { 909 case PRS_ZOMBIE: 910 panic("setrunnable(1)"); 911 default: 912 break; 913 } 914 switch (td->td_state) { 915 case 0: 916 case TDS_RUNNING: 917 case TDS_IWAIT: 918 default: 919 printf("state is %d", td->td_state); 920 panic("setrunnable(2)"); 921 case TDS_SUSPENDED: 922 thread_unsuspend(p); 923 break; 924 case TDS_SLP: /* e.g. when sending signals */ 925 if (td->td_flags & TDF_CVWAITQ) 926 cv_waitq_remove(td); 927 else 928 unsleep(td); 929 case TDS_UNQUEUED: /* being put back onto the queue */ 930 case TDS_NEW: /* not yet had time to suspend */ 931 case TDS_RUNQ: /* not yet had time to suspend */ 932 break; 933 } 934 if (td->td_ksegrp->kg_slptime > 1) 935 updatepri(td); 936 td->td_ksegrp->kg_slptime = 0; 937 if ((p->p_sflag & PS_INMEM) == 0) { 938 td->td_state = TDS_RUNQ; /* XXXKSE not a good idea */ 939 p->p_sflag |= PS_SWAPINREQ; 940 wakeup(&proc0); 941 } else { 942 if (td->td_state != TDS_RUNQ) 943 setrunqueue(td); /* XXXKSE */ 944 maybe_resched(td); 945 } 946 } 947 948 /* 949 * Compute the priority of a process when running in user mode. 950 * Arrange to reschedule if the resulting priority is better 951 * than that of the current process. 952 */ 953 void 954 resetpriority(kg) 955 register struct ksegrp *kg; 956 { 957 register unsigned int newpriority; 958 struct thread *td; 959 960 mtx_lock_spin(&sched_lock); 961 if (kg->kg_pri_class == PRI_TIMESHARE) { 962 newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT + 963 NICE_WEIGHT * (kg->kg_nice - PRIO_MIN); 964 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 965 PRI_MAX_TIMESHARE); 966 kg->kg_user_pri = newpriority; 967 } 968 FOREACH_THREAD_IN_GROUP(kg, td) { 969 maybe_resched(td); /* XXXKSE silly */ 970 } 971 mtx_unlock_spin(&sched_lock); 972 } 973 974 /* 975 * Compute a tenex style load average of a quantity on 976 * 1, 5 and 15 minute intervals. 977 * XXXKSE Needs complete rewrite when correct info is available. 978 * Completely Bogus.. only works with 1:1 (but compiles ok now :-) 979 */ 980 static void 981 loadav(void *arg) 982 { 983 int i, nrun; 984 struct loadavg *avg; 985 struct proc *p; 986 struct thread *td; 987 988 avg = &averunnable; 989 sx_slock(&allproc_lock); 990 nrun = 0; 991 FOREACH_PROC_IN_SYSTEM(p) { 992 FOREACH_THREAD_IN_PROC(p, td) { 993 switch (td->td_state) { 994 case TDS_RUNQ: 995 case TDS_RUNNING: 996 if ((p->p_flag & P_NOLOAD) != 0) 997 goto nextproc; 998 nrun++; /* XXXKSE */ 999 default: 1000 break; 1001 } 1002 nextproc: 1003 continue; 1004 } 1005 } 1006 sx_sunlock(&allproc_lock); 1007 for (i = 0; i < 3; i++) 1008 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1009 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1010 1011 /* 1012 * Schedule the next update to occur after 5 seconds, but add a 1013 * random variation to avoid synchronisation with processes that 1014 * run at regular intervals. 1015 */ 1016 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)), 1017 loadav, NULL); 1018 } 1019 1020 /* ARGSUSED */ 1021 static void 1022 sched_setup(dummy) 1023 void *dummy; 1024 { 1025 1026 callout_init(&schedcpu_callout, 1); 1027 callout_init(&roundrobin_callout, 0); 1028 callout_init(&loadav_callout, 0); 1029 1030 /* Kick off timeout driven events by calling first time. */ 1031 roundrobin(NULL); 1032 schedcpu(NULL); 1033 loadav(NULL); 1034 } 1035 1036 /* 1037 * We adjust the priority of the current process. The priority of 1038 * a process gets worse as it accumulates CPU time. The cpu usage 1039 * estimator (p_estcpu) is increased here. resetpriority() will 1040 * compute a different priority each time p_estcpu increases by 1041 * INVERSE_ESTCPU_WEIGHT 1042 * (until MAXPRI is reached). The cpu usage estimator ramps up 1043 * quite quickly when the process is running (linearly), and decays 1044 * away exponentially, at a rate which is proportionally slower when 1045 * the system is busy. The basic principle is that the system will 1046 * 90% forget that the process used a lot of CPU time in 5 * loadav 1047 * seconds. This causes the system to favor processes which haven't 1048 * run much recently, and to round-robin among other processes. 1049 */ 1050 void 1051 schedclock(td) 1052 struct thread *td; 1053 { 1054 struct kse *ke; 1055 struct ksegrp *kg; 1056 1057 KASSERT((td != NULL), ("schedlock: null thread pointer")); 1058 ke = td->td_kse; 1059 kg = td->td_ksegrp; 1060 ke->ke_cpticks++; 1061 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1); 1062 if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1063 resetpriority(kg); 1064 if (td->td_priority >= PUSER) 1065 td->td_priority = kg->kg_user_pri; 1066 } 1067 } 1068 1069 /* 1070 * General purpose yield system call 1071 */ 1072 int 1073 yield(struct thread *td, struct yield_args *uap) 1074 { 1075 struct ksegrp *kg = td->td_ksegrp; 1076 1077 mtx_assert(&Giant, MA_NOTOWNED); 1078 mtx_lock_spin(&sched_lock); 1079 td->td_priority = PRI_MAX_TIMESHARE; 1080 kg->kg_proc->p_stats->p_ru.ru_nvcsw++; 1081 mi_switch(); 1082 mtx_unlock_spin(&sched_lock); 1083 td->td_retval[0] = 0; 1084 1085 return (0); 1086 } 1087 1088