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 */ 281 /* if ((ke->ke_flags & KEF_ONRUNQ) == 0) */ 282 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) { 283 ke->ke_slptime++; 284 } else { 285 ke->ke_slptime = 0; 286 awake = 1; 287 } 288 289 /* 290 * pctcpu is only for ps? 291 * Do it per kse.. and add them up at the end? 292 * XXXKSE 293 */ 294 ke->ke_pctcpu = (ke->ke_pctcpu * ccpu) >> FSHIFT; 295 /* 296 * If the kse has been idle the entire second, 297 * stop recalculating its priority until 298 * it wakes up. 299 */ 300 if (ke->ke_slptime > 1) { 301 continue; 302 } 303 304 #if (FSHIFT >= CCPU_SHIFT) 305 ke->ke_pctcpu += (realstathz == 100) ? 306 ((fixpt_t) ke->ke_cpticks) << 307 (FSHIFT - CCPU_SHIFT) : 308 100 * (((fixpt_t) ke->ke_cpticks) << 309 (FSHIFT - CCPU_SHIFT)) / realstathz; 310 #else 311 ke->ke_pctcpu += ((FSCALE - ccpu) * 312 (ke->ke_cpticks * FSCALE / realstathz)) >> 313 FSHIFT; 314 #endif 315 ke->ke_cpticks = 0; 316 } /* end of kse loop */ 317 if (awake == 0) { 318 kg->kg_slptime++; 319 } else { 320 kg->kg_slptime = 0; 321 } 322 kg->kg_estcpu = decay_cpu(loadfac, kg->kg_estcpu); 323 resetpriority(kg); 324 td = FIRST_THREAD_IN_PROC(p); 325 if (td->td_priority >= PUSER && 326 (p->p_sflag & PS_INMEM)) { 327 int changedqueue = 328 ((td->td_priority / RQ_PPQ) != 329 (kg->kg_user_pri / RQ_PPQ)); 330 331 td->td_priority = kg->kg_user_pri; 332 FOREACH_KSE_IN_GROUP(kg, ke) { 333 if ((ke->ke_oncpu == NOCPU) && 334 (p->p_stat == SRUN) && /* XXXKSE */ 335 changedqueue) { 336 remrunqueue(ke->ke_thread); 337 setrunqueue(ke->ke_thread); 338 } 339 } 340 } 341 } /* end of ksegrp loop */ 342 mtx_unlock_spin(&sched_lock); 343 } /* end of process loop */ 344 sx_sunlock(&allproc_lock); 345 wakeup((caddr_t)&lbolt); 346 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 347 } 348 349 /* 350 * Recalculate the priority of a process after it has slept for a while. 351 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 352 * least six times the loadfactor will decay p_estcpu to zero. 353 */ 354 void 355 updatepri(td) 356 register struct thread *td; 357 { 358 register struct ksegrp *kg; 359 register unsigned int newcpu; 360 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 361 362 if (td == NULL) 363 return; 364 kg = td->td_ksegrp; 365 newcpu = kg->kg_estcpu; 366 if (kg->kg_slptime > 5 * loadfac) 367 kg->kg_estcpu = 0; 368 else { 369 kg->kg_slptime--; /* the first time was done in schedcpu */ 370 while (newcpu && --kg->kg_slptime) 371 newcpu = decay_cpu(loadfac, newcpu); 372 kg->kg_estcpu = newcpu; 373 } 374 resetpriority(td->td_ksegrp); 375 } 376 377 /* 378 * We're only looking at 7 bits of the address; everything is 379 * aligned to 4, lots of things are aligned to greater powers 380 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 381 */ 382 #define TABLESIZE 128 383 static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE]; 384 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 385 386 void 387 sleepinit(void) 388 { 389 int i; 390 391 sched_quantum = hz/10; 392 hogticks = 2 * sched_quantum; 393 for (i = 0; i < TABLESIZE; i++) 394 TAILQ_INIT(&slpque[i]); 395 } 396 397 /* 398 * General sleep call. Suspends the current process until a wakeup is 399 * performed on the specified identifier. The process will then be made 400 * runnable with the specified priority. Sleeps at most timo/hz seconds 401 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 402 * before and after sleeping, else signals are not checked. Returns 0 if 403 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 404 * signal needs to be delivered, ERESTART is returned if the current system 405 * call should be restarted if possible, and EINTR is returned if the system 406 * call should be interrupted by the signal (return EINTR). 407 * 408 * The mutex argument is exited before the caller is suspended, and 409 * entered before msleep returns. If priority includes the PDROP 410 * flag the mutex is not entered before returning. 411 */ 412 int 413 msleep(ident, mtx, priority, wmesg, timo) 414 void *ident; 415 struct mtx *mtx; 416 int priority, timo; 417 const char *wmesg; 418 { 419 struct proc *p = curproc; 420 struct thread *td = curthread; 421 int sig, catch = priority & PCATCH; 422 int rval = 0; 423 WITNESS_SAVE_DECL(mtx); 424 425 #ifdef KTRACE 426 if (p && KTRPOINT(p, KTR_CSW)) 427 ktrcsw(p->p_tracep, 1, 0); 428 #endif 429 WITNESS_SLEEP(0, &mtx->mtx_object); 430 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 431 ("sleeping without a mutex")); 432 mtx_lock_spin(&sched_lock); 433 if (cold || panicstr) { 434 /* 435 * After a panic, or during autoconfiguration, 436 * just give interrupts a chance, then just return; 437 * don't run any other procs or panic below, 438 * in case this is the idle process and already asleep. 439 */ 440 if (mtx != NULL && priority & PDROP) 441 mtx_unlock(mtx); 442 mtx_unlock_spin(&sched_lock); 443 return (0); 444 } 445 446 DROP_GIANT(); 447 448 if (mtx != NULL) { 449 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 450 WITNESS_SAVE(&mtx->mtx_object, mtx); 451 mtx_unlock(mtx); 452 if (priority & PDROP) 453 mtx = NULL; 454 } 455 456 KASSERT(p != NULL, ("msleep1")); 457 KASSERT(ident != NULL && td->td_proc->p_stat == SRUN, ("msleep")); 458 459 td->td_wchan = ident; 460 td->td_wmesg = wmesg; 461 td->td_kse->ke_slptime = 0; /* XXXKSE */ 462 td->td_ksegrp->kg_slptime = 0; 463 td->td_priority = priority & PRIMASK; 464 CTR5(KTR_PROC, "msleep: thread %p (pid %d, %s) on %s (%p)", 465 td, p->p_pid, p->p_comm, wmesg, ident); 466 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], td, td_slpq); 467 if (timo) 468 callout_reset(&td->td_slpcallout, timo, endtsleep, td); 469 /* 470 * We put ourselves on the sleep queue and start our timeout 471 * before calling CURSIG, as we could stop there, and a wakeup 472 * or a SIGCONT (or both) could occur while we were stopped. 473 * A SIGCONT would cause us to be marked as SSLEEP 474 * without resuming us, thus we must be ready for sleep 475 * when CURSIG is called. If the wakeup happens while we're 476 * stopped, td->td_wchan will be 0 upon return from CURSIG. 477 */ 478 if (catch) { 479 CTR3(KTR_PROC, "msleep caught: proc %p (pid %d, %s)", p, 480 p->p_pid, p->p_comm); 481 td->td_flags |= TDF_SINTR; 482 mtx_unlock_spin(&sched_lock); 483 PROC_LOCK(p); 484 sig = CURSIG(p); 485 mtx_lock_spin(&sched_lock); 486 PROC_UNLOCK(p); 487 if (sig != 0) { 488 if (td->td_wchan != NULL) 489 unsleep(td); 490 } else if (td->td_wchan == NULL) 491 catch = 0; 492 } else 493 sig = 0; 494 if (td->td_wchan != NULL) { 495 td->td_proc->p_stat = SSLEEP; 496 p->p_stats->p_ru.ru_nvcsw++; 497 mi_switch(); 498 } 499 CTR3(KTR_PROC, "msleep resume: proc %p (pid %d, %s)", td, p->p_pid, 500 p->p_comm); 501 KASSERT(td->td_proc->p_stat == SRUN, ("running but not SRUN")); 502 td->td_flags &= ~TDF_SINTR; 503 if (td->td_flags & TDF_TIMEOUT) { 504 td->td_flags &= ~TDF_TIMEOUT; 505 if (sig == 0) 506 rval = EWOULDBLOCK; 507 } else if (td->td_flags & TDF_TIMOFAIL) 508 td->td_flags &= ~TDF_TIMOFAIL; 509 else if (timo && callout_stop(&td->td_slpcallout) == 0) { 510 /* 511 * This isn't supposed to be pretty. If we are here, then 512 * the endtsleep() callout is currently executing on another 513 * CPU and is either spinning on the sched_lock or will be 514 * soon. If we don't synchronize here, there is a chance 515 * that this process may msleep() again before the callout 516 * has a chance to run and the callout may end up waking up 517 * the wrong msleep(). Yuck. 518 */ 519 td->td_flags |= TDF_TIMEOUT; 520 p->p_stats->p_ru.ru_nivcsw++; 521 mi_switch(); 522 } 523 mtx_unlock_spin(&sched_lock); 524 525 if (rval == 0 && catch) { 526 PROC_LOCK(p); 527 /* XXX: shouldn't we always be calling CURSIG() */ 528 if (sig != 0 || (sig = CURSIG(p))) { 529 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 530 rval = EINTR; 531 else 532 rval = ERESTART; 533 } 534 PROC_UNLOCK(p); 535 } 536 PICKUP_GIANT(); 537 #ifdef KTRACE 538 mtx_lock(&Giant); 539 if (KTRPOINT(p, KTR_CSW)) 540 ktrcsw(p->p_tracep, 0, 0); 541 mtx_unlock(&Giant); 542 #endif 543 if (mtx != NULL) { 544 mtx_lock(mtx); 545 WITNESS_RESTORE(&mtx->mtx_object, mtx); 546 } 547 return (rval); 548 } 549 550 /* 551 * Implement timeout for msleep() 552 * 553 * If process hasn't been awakened (wchan non-zero), 554 * set timeout flag and undo the sleep. If proc 555 * is stopped, just unsleep so it will remain stopped. 556 * MP-safe, called without the Giant mutex. 557 */ 558 static void 559 endtsleep(arg) 560 void *arg; 561 { 562 register struct thread *td = arg; 563 564 CTR3(KTR_PROC, "endtsleep: thread %p (pid %d, %s)", td, td->td_proc->p_pid, 565 td->td_proc->p_comm); 566 mtx_lock_spin(&sched_lock); 567 /* 568 * This is the other half of the synchronization with msleep() 569 * described above. If the PS_TIMEOUT flag is set, we lost the 570 * race and just need to put the process back on the runqueue. 571 */ 572 if ((td->td_flags & TDF_TIMEOUT) != 0) { 573 td->td_flags &= ~TDF_TIMEOUT; 574 setrunqueue(td); 575 } else if (td->td_wchan != NULL) { 576 if (td->td_proc->p_stat == SSLEEP) /* XXXKSE */ 577 setrunnable(td); 578 else 579 unsleep(td); 580 td->td_flags |= TDF_TIMEOUT; 581 } else { 582 td->td_flags |= TDF_TIMOFAIL; 583 } 584 mtx_unlock_spin(&sched_lock); 585 } 586 587 /* 588 * Remove a process from its wait queue 589 */ 590 void 591 unsleep(struct thread *td) 592 { 593 594 mtx_lock_spin(&sched_lock); 595 if (td->td_wchan != NULL) { 596 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_slpq); 597 td->td_wchan = NULL; 598 } 599 mtx_unlock_spin(&sched_lock); 600 } 601 602 /* 603 * Make all processes sleeping on the specified identifier runnable. 604 */ 605 void 606 wakeup(ident) 607 register void *ident; 608 { 609 register struct slpquehead *qp; 610 register struct thread *td; 611 struct proc *p; 612 613 mtx_lock_spin(&sched_lock); 614 qp = &slpque[LOOKUP(ident)]; 615 restart: 616 TAILQ_FOREACH(td, qp, td_slpq) { 617 p = td->td_proc; 618 if (td->td_wchan == ident) { 619 TAILQ_REMOVE(qp, td, td_slpq); 620 td->td_wchan = NULL; 621 if (td->td_proc->p_stat == SSLEEP) { 622 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 623 CTR3(KTR_PROC, "wakeup: thread %p (pid %d, %s)", 624 td, p->p_pid, p->p_comm); 625 if (td->td_ksegrp->kg_slptime > 1) 626 updatepri(td); 627 td->td_ksegrp->kg_slptime = 0; 628 td->td_kse->ke_slptime = 0; 629 td->td_proc->p_stat = SRUN; 630 if (p->p_sflag & PS_INMEM) { 631 setrunqueue(td); 632 maybe_resched(td); 633 } else { 634 p->p_sflag |= PS_SWAPINREQ; 635 wakeup((caddr_t)&proc0); 636 } 637 /* END INLINE EXPANSION */ 638 goto restart; 639 } 640 } 641 } 642 mtx_unlock_spin(&sched_lock); 643 } 644 645 /* 646 * Make a process sleeping on the specified identifier runnable. 647 * May wake more than one process if a target process is currently 648 * swapped out. 649 */ 650 void 651 wakeup_one(ident) 652 register void *ident; 653 { 654 register struct slpquehead *qp; 655 register struct thread *td; 656 register struct proc *p; 657 658 mtx_lock_spin(&sched_lock); 659 qp = &slpque[LOOKUP(ident)]; 660 661 TAILQ_FOREACH(td, qp, td_slpq) { 662 p = td->td_proc; 663 if (td->td_wchan == ident) { 664 TAILQ_REMOVE(qp, td, td_slpq); 665 td->td_wchan = NULL; 666 if (td->td_proc->p_stat == SSLEEP) { 667 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 668 CTR3(KTR_PROC, "wakeup1: proc %p (pid %d, %s)", 669 p, p->p_pid, p->p_comm); 670 if (td->td_ksegrp->kg_slptime > 1) 671 updatepri(td); 672 td->td_ksegrp->kg_slptime = 0; 673 td->td_kse->ke_slptime = 0; 674 td->td_proc->p_stat = SRUN; 675 if (p->p_sflag & PS_INMEM) { 676 setrunqueue(td); 677 maybe_resched(td); 678 break; 679 } else { 680 p->p_sflag |= PS_SWAPINREQ; 681 wakeup((caddr_t)&proc0); 682 } 683 /* END INLINE EXPANSION */ 684 } 685 } 686 } 687 mtx_unlock_spin(&sched_lock); 688 } 689 690 /* 691 * The machine independent parts of mi_switch(). 692 */ 693 void 694 mi_switch() 695 { 696 struct bintime new_switchtime; 697 struct thread *td = curthread; /* XXX */ 698 register struct proc *p = td->td_proc; /* XXX */ 699 #if 0 700 register struct rlimit *rlim; 701 #endif 702 u_int sched_nest; 703 704 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 705 #ifdef INVARIANTS 706 if (p->p_stat != SMTX && p->p_stat != SRUN) 707 mtx_assert(&Giant, MA_NOTOWNED); 708 #endif 709 710 /* 711 * Compute the amount of time during which the current 712 * process was running, and add that to its total so far. 713 */ 714 binuptime(&new_switchtime); 715 bintime_add(&p->p_runtime, &new_switchtime); 716 bintime_sub(&p->p_runtime, PCPU_PTR(switchtime)); 717 718 #ifdef DDB 719 /* 720 * Don't perform context switches from the debugger. 721 */ 722 if (db_active) { 723 mtx_unlock_spin(&sched_lock); 724 db_error("Context switches not allowed in the debugger."); 725 } 726 #endif 727 728 #if 0 729 /* 730 * Check if the process exceeds its cpu resource allocation. 731 * If over max, kill it. 732 * 733 * XXX drop sched_lock, pickup Giant 734 */ 735 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 736 p->p_runtime > p->p_limit->p_cpulimit) { 737 rlim = &p->p_rlimit[RLIMIT_CPU]; 738 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 739 mtx_unlock_spin(&sched_lock); 740 PROC_LOCK(p); 741 killproc(p, "exceeded maximum CPU limit"); 742 mtx_lock_spin(&sched_lock); 743 PROC_UNLOCK(p); 744 } else { 745 mtx_unlock_spin(&sched_lock); 746 PROC_LOCK(p); 747 psignal(p, SIGXCPU); 748 mtx_lock_spin(&sched_lock); 749 PROC_UNLOCK(p); 750 if (rlim->rlim_cur < rlim->rlim_max) { 751 /* XXX: we should make a private copy */ 752 rlim->rlim_cur += 5; 753 } 754 } 755 } 756 #endif 757 758 /* 759 * Pick a new current process and record its start time. 760 */ 761 cnt.v_swtch++; 762 PCPU_SET(switchtime, new_switchtime); 763 CTR3(KTR_PROC, "mi_switch: old proc %p (pid %d, %s)", p, p->p_pid, 764 p->p_comm); 765 sched_nest = sched_lock.mtx_recurse; 766 td->td_lastcpu = td->td_kse->ke_oncpu; 767 td->td_kse->ke_oncpu = NOCPU; 768 td->td_kse->ke_flags &= ~KEF_NEEDRESCHED; 769 cpu_switch(); 770 td->td_kse->ke_oncpu = PCPU_GET(cpuid); 771 sched_lock.mtx_recurse = sched_nest; 772 sched_lock.mtx_lock = (uintptr_t)td; 773 CTR3(KTR_PROC, "mi_switch: new proc %p (pid %d, %s)", p, p->p_pid, 774 p->p_comm); 775 if (PCPU_GET(switchtime.sec) == 0) 776 binuptime(PCPU_PTR(switchtime)); 777 PCPU_SET(switchticks, ticks); 778 } 779 780 /* 781 * Change process state to be runnable, 782 * placing it on the run queue if it is in memory, 783 * and awakening the swapper if it isn't in memory. 784 */ 785 void 786 setrunnable(struct thread *td) 787 { 788 struct proc *p = td->td_proc; 789 790 mtx_lock_spin(&sched_lock); 791 switch (p->p_stat) { 792 case SZOMB: /* not a thread flag XXXKSE */ 793 panic("setrunnable(1)"); 794 } 795 switch (td->td_proc->p_stat) { 796 case 0: 797 case SRUN: 798 case SWAIT: 799 default: 800 panic("setrunnable(2)"); 801 case SSTOP: 802 case SSLEEP: /* e.g. when sending signals */ 803 if (td->td_flags & TDF_CVWAITQ) 804 cv_waitq_remove(td); 805 else 806 unsleep(td); 807 break; 808 809 case SIDL: 810 break; 811 } 812 td->td_proc->p_stat = SRUN; 813 if (td->td_ksegrp->kg_slptime > 1) 814 updatepri(td); 815 td->td_ksegrp->kg_slptime = 0; 816 td->td_kse->ke_slptime = 0; 817 if ((p->p_sflag & PS_INMEM) == 0) { 818 p->p_sflag |= PS_SWAPINREQ; 819 wakeup((caddr_t)&proc0); 820 } else { 821 setrunqueue(td); 822 maybe_resched(td); 823 } 824 mtx_unlock_spin(&sched_lock); 825 } 826 827 /* 828 * Compute the priority of a process when running in user mode. 829 * Arrange to reschedule if the resulting priority is better 830 * than that of the current process. 831 */ 832 void 833 resetpriority(kg) 834 register struct ksegrp *kg; 835 { 836 register unsigned int newpriority; 837 struct thread *td; 838 839 mtx_lock_spin(&sched_lock); 840 if (kg->kg_pri_class == PRI_TIMESHARE) { 841 newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT + 842 NICE_WEIGHT * (kg->kg_nice - PRIO_MIN); 843 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 844 PRI_MAX_TIMESHARE); 845 kg->kg_user_pri = newpriority; 846 } 847 FOREACH_THREAD_IN_GROUP(kg, td) { 848 maybe_resched(td); 849 } 850 mtx_unlock_spin(&sched_lock); 851 } 852 853 /* 854 * Compute a tenex style load average of a quantity on 855 * 1, 5 and 15 minute intervals. 856 * XXXKSE Needs complete rewrite when correct info is available. 857 * Completely Bogus.. only works with 1:1 (but compiles ok now :-) 858 */ 859 static void 860 loadav(void *arg) 861 { 862 int i, nrun; 863 struct loadavg *avg; 864 struct proc *p; 865 struct ksegrp *kg; 866 867 avg = &averunnable; 868 sx_slock(&allproc_lock); 869 nrun = 0; 870 FOREACH_PROC_IN_SYSTEM(p) { 871 FOREACH_KSEGRP_IN_PROC(p, kg) { 872 switch (p->p_stat) { 873 case SRUN: 874 if ((p->p_flag & P_NOLOAD) != 0) 875 goto nextproc; 876 /* FALLTHROUGH */ 877 case SIDL: 878 nrun++; 879 } 880 nextproc: 881 continue; 882 } 883 } 884 sx_sunlock(&allproc_lock); 885 for (i = 0; i < 3; i++) 886 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 887 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 888 889 /* 890 * Schedule the next update to occur after 5 seconds, but add a 891 * random variation to avoid synchronisation with processes that 892 * run at regular intervals. 893 */ 894 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)), 895 loadav, NULL); 896 } 897 898 /* ARGSUSED */ 899 static void 900 sched_setup(dummy) 901 void *dummy; 902 { 903 904 callout_init(&schedcpu_callout, 1); 905 callout_init(&roundrobin_callout, 0); 906 callout_init(&loadav_callout, 0); 907 908 /* Kick off timeout driven events by calling first time. */ 909 roundrobin(NULL); 910 schedcpu(NULL); 911 loadav(NULL); 912 } 913 914 /* 915 * We adjust the priority of the current process. The priority of 916 * a process gets worse as it accumulates CPU time. The cpu usage 917 * estimator (p_estcpu) is increased here. resetpriority() will 918 * compute a different priority each time p_estcpu increases by 919 * INVERSE_ESTCPU_WEIGHT 920 * (until MAXPRI is reached). The cpu usage estimator ramps up 921 * quite quickly when the process is running (linearly), and decays 922 * away exponentially, at a rate which is proportionally slower when 923 * the system is busy. The basic principle is that the system will 924 * 90% forget that the process used a lot of CPU time in 5 * loadav 925 * seconds. This causes the system to favor processes which haven't 926 * run much recently, and to round-robin among other processes. 927 */ 928 void 929 schedclock(td) 930 struct thread *td; 931 { 932 struct kse *ke = td->td_kse; 933 struct ksegrp *kg = td->td_ksegrp; 934 935 if (td) { 936 ke->ke_cpticks++; 937 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1); 938 if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 939 resetpriority(td->td_ksegrp); 940 if (td->td_priority >= PUSER) 941 td->td_priority = kg->kg_user_pri; 942 } 943 } else { 944 panic("schedclock"); 945 } 946 } 947 948 /* 949 * General purpose yield system call 950 */ 951 int 952 yield(struct thread *td, struct yield_args *uap) 953 { 954 struct ksegrp *kg = td->td_ksegrp; 955 956 mtx_assert(&Giant, MA_NOTOWNED); 957 mtx_lock_spin(&sched_lock); 958 td->td_priority = PRI_MAX_TIMESHARE; 959 setrunqueue(td); 960 kg->kg_proc->p_stats->p_ru.ru_nvcsw++; 961 mi_switch(); 962 mtx_unlock_spin(&sched_lock); 963 td->td_retval[0] = 0; 964 965 return (0); 966 } 967 968