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 __P((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 __P((void *)); 94 static void loadav __P((void *arg)); 95 static void roundrobin __P((void *arg)); 96 static void schedcpu __P((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(kg) 124 struct ksegrp *kg; 125 { 126 127 mtx_assert(&sched_lock, MA_OWNED); 128 if (kg->kg_pri.pri_level < curthread->td_ksegrp->kg_pri.pri_level) 129 curthread->td_kse->ke_flags |= KEF_NEEDRESCHED; 130 } 131 132 int 133 roundrobin_interval(void) 134 { 135 return (sched_quantum); 136 } 137 138 /* 139 * Force switch among equal priority processes every 100ms. 140 * We don't actually need to force a context switch of the current process. 141 * The act of firing the event triggers a context switch to softclock() and 142 * then switching back out again which is equivalent to a preemption, thus 143 * no further work is needed on the local CPU. 144 */ 145 /* ARGSUSED */ 146 static void 147 roundrobin(arg) 148 void *arg; 149 { 150 151 #ifdef SMP 152 mtx_lock_spin(&sched_lock); 153 forward_roundrobin(); 154 mtx_unlock_spin(&sched_lock); 155 #endif 156 157 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 158 } 159 160 /* 161 * Constants for digital decay and forget: 162 * 90% of (p_estcpu) usage in 5 * loadav time 163 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 164 * Note that, as ps(1) mentions, this can let percentages 165 * total over 100% (I've seen 137.9% for 3 processes). 166 * 167 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 168 * 169 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 170 * That is, the system wants to compute a value of decay such 171 * that the following for loop: 172 * for (i = 0; i < (5 * loadavg); i++) 173 * p_estcpu *= decay; 174 * will compute 175 * p_estcpu *= 0.1; 176 * for all values of loadavg: 177 * 178 * Mathematically this loop can be expressed by saying: 179 * decay ** (5 * loadavg) ~= .1 180 * 181 * The system computes decay as: 182 * decay = (2 * loadavg) / (2 * loadavg + 1) 183 * 184 * We wish to prove that the system's computation of decay 185 * will always fulfill the equation: 186 * decay ** (5 * loadavg) ~= .1 187 * 188 * If we compute b as: 189 * b = 2 * loadavg 190 * then 191 * decay = b / (b + 1) 192 * 193 * We now need to prove two things: 194 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 195 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 196 * 197 * Facts: 198 * For x close to zero, exp(x) =~ 1 + x, since 199 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 200 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 201 * For x close to zero, ln(1+x) =~ x, since 202 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 203 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 204 * ln(.1) =~ -2.30 205 * 206 * Proof of (1): 207 * Solve (factor)**(power) =~ .1 given power (5*loadav): 208 * solving for factor, 209 * ln(factor) =~ (-2.30/5*loadav), or 210 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 211 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 212 * 213 * Proof of (2): 214 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 215 * solving for power, 216 * power*ln(b/(b+1)) =~ -2.30, or 217 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 218 * 219 * Actual power values for the implemented algorithm are as follows: 220 * loadav: 1 2 3 4 221 * power: 5.68 10.32 14.94 19.55 222 */ 223 224 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 225 #define loadfactor(loadav) (2 * (loadav)) 226 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 227 228 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 229 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 230 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 231 232 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 233 static int fscale __unused = FSCALE; 234 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 235 236 /* 237 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 238 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 239 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 240 * 241 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 242 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 243 * 244 * If you don't want to bother with the faster/more-accurate formula, you 245 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 246 * (more general) method of calculating the %age of CPU used by a process. 247 */ 248 #define CCPU_SHIFT 11 249 250 /* 251 * Recompute process priorities, every hz ticks. 252 * MP-safe, called without the Giant mutex. 253 */ 254 /* ARGSUSED */ 255 static void 256 schedcpu(arg) 257 void *arg; 258 { 259 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 260 register struct proc *p; 261 register struct kse *ke; 262 register struct ksegrp *kg; 263 register 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 if (kg->kg_pri.pri_level >= PUSER && 325 (p->p_sflag & PS_INMEM)) { 326 int changedqueue = 327 ((kg->kg_pri.pri_level / RQ_PPQ) != 328 (kg->kg_pri.pri_user / RQ_PPQ)); 329 330 kg->kg_pri.pri_level = kg->kg_pri.pri_user; 331 FOREACH_KSE_IN_GROUP(kg, ke) { 332 if ((ke->ke_oncpu == NOCPU) && /* idle */ 333 (p->p_stat == SRUN) && /* XXXKSE */ 334 changedqueue) { 335 remrunqueue(ke->ke_thread); 336 setrunqueue(ke->ke_thread); 337 } 338 } 339 } 340 } /* end of ksegrp loop */ 341 mtx_unlock_spin(&sched_lock); 342 } /* end of process loop */ 343 sx_sunlock(&allproc_lock); 344 wakeup((caddr_t)&lbolt); 345 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 346 } 347 348 /* 349 * Recalculate the priority of a process after it has slept for a while. 350 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 351 * least six times the loadfactor will decay p_estcpu to zero. 352 */ 353 void 354 updatepri(td) 355 register struct thread *td; 356 { 357 register struct ksegrp *kg; 358 register unsigned int newcpu; 359 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 360 361 if (td == NULL) 362 return; 363 kg = td->td_ksegrp; 364 newcpu = kg->kg_estcpu; 365 if (kg->kg_slptime > 5 * loadfac) 366 kg->kg_estcpu = 0; 367 else { 368 kg->kg_slptime--; /* the first time was done in schedcpu */ 369 while (newcpu && --kg->kg_slptime) 370 newcpu = decay_cpu(loadfac, newcpu); 371 kg->kg_estcpu = newcpu; 372 } 373 resetpriority(td->td_ksegrp); 374 } 375 376 /* 377 * We're only looking at 7 bits of the address; everything is 378 * aligned to 4, lots of things are aligned to greater powers 379 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 380 */ 381 #define TABLESIZE 128 382 static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE]; 383 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 384 385 void 386 sleepinit(void) 387 { 388 int i; 389 390 sched_quantum = hz/10; 391 hogticks = 2 * sched_quantum; 392 for (i = 0; i < TABLESIZE; i++) 393 TAILQ_INIT(&slpque[i]); 394 } 395 396 /* 397 * General sleep call. Suspends the current process until a wakeup is 398 * performed on the specified identifier. The process will then be made 399 * runnable with the specified priority. Sleeps at most timo/hz seconds 400 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 401 * before and after sleeping, else signals are not checked. Returns 0 if 402 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 403 * signal needs to be delivered, ERESTART is returned if the current system 404 * call should be restarted if possible, and EINTR is returned if the system 405 * call should be interrupted by the signal (return EINTR). 406 * 407 * The mutex argument is exited before the caller is suspended, and 408 * entered before msleep returns. If priority includes the PDROP 409 * flag the mutex is not entered before returning. 410 */ 411 int 412 msleep(ident, mtx, priority, wmesg, timo) 413 void *ident; 414 struct mtx *mtx; 415 int priority, timo; 416 const char *wmesg; 417 { 418 struct proc *p = curproc; 419 struct thread *td = curthread; 420 int sig, catch = priority & PCATCH; 421 int rval = 0; 422 WITNESS_SAVE_DECL(mtx); 423 424 #ifdef KTRACE 425 if (p && KTRPOINT(p, KTR_CSW)) 426 ktrcsw(p->p_tracep, 1, 0); 427 #endif 428 WITNESS_SLEEP(0, &mtx->mtx_object); 429 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 430 ("sleeping without a mutex")); 431 mtx_lock_spin(&sched_lock); 432 if (cold || panicstr) { 433 /* 434 * After a panic, or during autoconfiguration, 435 * just give interrupts a chance, then just return; 436 * don't run any other procs or panic below, 437 * in case this is the idle process and already asleep. 438 */ 439 if (mtx != NULL && priority & PDROP) 440 mtx_unlock(mtx); 441 mtx_unlock_spin(&sched_lock); 442 return (0); 443 } 444 445 DROP_GIANT(); 446 447 if (mtx != NULL) { 448 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 449 WITNESS_SAVE(&mtx->mtx_object, mtx); 450 mtx_unlock(mtx); 451 if (priority & PDROP) 452 mtx = NULL; 453 } 454 455 KASSERT(p != NULL, ("msleep1")); 456 KASSERT(ident != NULL && td->td_proc->p_stat == SRUN, ("msleep")); 457 458 td->td_wchan = ident; 459 td->td_wmesg = wmesg; 460 td->td_kse->ke_slptime = 0; /* XXXKSE */ 461 td->td_ksegrp->kg_slptime = 0; 462 td->td_ksegrp->kg_pri.pri_level = priority & PRIMASK; 463 CTR5(KTR_PROC, "msleep: thread %p (pid %d, %s) on %s (%p)", 464 td, p->p_pid, p->p_comm, wmesg, ident); 465 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], td, td_slpq); 466 if (timo) 467 callout_reset(&td->td_slpcallout, timo, endtsleep, td); 468 /* 469 * We put ourselves on the sleep queue and start our timeout 470 * before calling CURSIG, as we could stop there, and a wakeup 471 * or a SIGCONT (or both) could occur while we were stopped. 472 * A SIGCONT would cause us to be marked as SSLEEP 473 * without resuming us, thus we must be ready for sleep 474 * when CURSIG is called. If the wakeup happens while we're 475 * stopped, td->td_wchan will be 0 upon return from CURSIG. 476 */ 477 if (catch) { 478 CTR3(KTR_PROC, "msleep caught: proc %p (pid %d, %s)", p, 479 p->p_pid, p->p_comm); 480 td->td_flags |= TDF_SINTR; 481 mtx_unlock_spin(&sched_lock); 482 PROC_LOCK(p); 483 sig = CURSIG(p); 484 mtx_lock_spin(&sched_lock); 485 PROC_UNLOCK(p); 486 if (sig != 0) { 487 if (td->td_wchan != NULL) 488 unsleep(td); 489 } else if (td->td_wchan == NULL) 490 catch = 0; 491 } else 492 sig = 0; 493 if (td->td_wchan != NULL) { 494 td->td_proc->p_stat = SSLEEP; 495 p->p_stats->p_ru.ru_nvcsw++; 496 mi_switch(); 497 } 498 CTR3(KTR_PROC, "msleep resume: proc %p (pid %d, %s)", td, p->p_pid, 499 p->p_comm); 500 KASSERT(td->td_proc->p_stat == SRUN, ("running but not SRUN")); 501 td->td_flags &= ~TDF_SINTR; 502 if (td->td_flags & TDF_TIMEOUT) { 503 td->td_flags &= ~TDF_TIMEOUT; 504 if (sig == 0) 505 rval = EWOULDBLOCK; 506 } else if (td->td_flags & TDF_TIMOFAIL) 507 td->td_flags &= ~TDF_TIMOFAIL; 508 else if (timo && callout_stop(&td->td_slpcallout) == 0) { 509 /* 510 * This isn't supposed to be pretty. If we are here, then 511 * the endtsleep() callout is currently executing on another 512 * CPU and is either spinning on the sched_lock or will be 513 * soon. If we don't synchronize here, there is a chance 514 * that this process may msleep() again before the callout 515 * has a chance to run and the callout may end up waking up 516 * the wrong msleep(). Yuck. 517 */ 518 td->td_flags |= TDF_TIMEOUT; 519 p->p_stats->p_ru.ru_nivcsw++; 520 mi_switch(); 521 } 522 mtx_unlock_spin(&sched_lock); 523 524 if (rval == 0 && catch) { 525 PROC_LOCK(p); 526 /* XXX: shouldn't we always be calling CURSIG() */ 527 if (sig != 0 || (sig = CURSIG(p))) { 528 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 529 rval = EINTR; 530 else 531 rval = ERESTART; 532 } 533 PROC_UNLOCK(p); 534 } 535 PICKUP_GIANT(); 536 #ifdef KTRACE 537 mtx_lock(&Giant); 538 if (KTRPOINT(p, KTR_CSW)) 539 ktrcsw(p->p_tracep, 0, 0); 540 mtx_unlock(&Giant); 541 #endif 542 if (mtx != NULL) { 543 mtx_lock(mtx); 544 WITNESS_RESTORE(&mtx->mtx_object, mtx); 545 } 546 return (rval); 547 } 548 549 /* 550 * Implement timeout for msleep() 551 * 552 * If process hasn't been awakened (wchan non-zero), 553 * set timeout flag and undo the sleep. If proc 554 * is stopped, just unsleep so it will remain stopped. 555 * MP-safe, called without the Giant mutex. 556 */ 557 static void 558 endtsleep(arg) 559 void *arg; 560 { 561 register struct thread *td = arg; 562 563 CTR3(KTR_PROC, "endtsleep: thread %p (pid %d, %s)", td, td->td_proc->p_pid, 564 td->td_proc->p_comm); 565 mtx_lock_spin(&sched_lock); 566 /* 567 * This is the other half of the synchronization with msleep() 568 * described above. If the PS_TIMEOUT flag is set, we lost the 569 * race and just need to put the process back on the runqueue. 570 */ 571 if ((td->td_flags & TDF_TIMEOUT) != 0) { 572 td->td_flags &= ~TDF_TIMEOUT; 573 setrunqueue(td); 574 } else if (td->td_wchan != NULL) { 575 if (td->td_proc->p_stat == SSLEEP) /* XXXKSE */ 576 setrunnable(td); 577 else 578 unsleep(td); 579 td->td_flags |= TDF_TIMEOUT; 580 } else { 581 td->td_flags |= TDF_TIMOFAIL; 582 } 583 mtx_unlock_spin(&sched_lock); 584 } 585 586 /* 587 * Remove a process from its wait queue 588 */ 589 void 590 unsleep(struct thread *td) 591 { 592 593 mtx_lock_spin(&sched_lock); 594 if (td->td_wchan != NULL) { 595 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_slpq); 596 td->td_wchan = NULL; 597 } 598 mtx_unlock_spin(&sched_lock); 599 } 600 601 /* 602 * Make all processes sleeping on the specified identifier runnable. 603 */ 604 void 605 wakeup(ident) 606 register void *ident; 607 { 608 register struct slpquehead *qp; 609 register struct thread *td; 610 struct proc *p; 611 612 mtx_lock_spin(&sched_lock); 613 qp = &slpque[LOOKUP(ident)]; 614 restart: 615 TAILQ_FOREACH(td, qp, td_slpq) { 616 p = td->td_proc; 617 if (td->td_wchan == ident) { 618 TAILQ_REMOVE(qp, td, td_slpq); 619 td->td_wchan = NULL; 620 if (td->td_proc->p_stat == SSLEEP) { 621 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 622 CTR3(KTR_PROC, "wakeup: thread %p (pid %d, %s)", 623 td, p->p_pid, p->p_comm); 624 if (td->td_ksegrp->kg_slptime > 1) 625 updatepri(td); 626 td->td_ksegrp->kg_slptime = 0; 627 td->td_kse->ke_slptime = 0; 628 td->td_proc->p_stat = SRUN; 629 if (p->p_sflag & PS_INMEM) { 630 setrunqueue(td); 631 maybe_resched(td->td_ksegrp); 632 } else { 633 p->p_sflag |= PS_SWAPINREQ; 634 wakeup((caddr_t)&proc0); 635 } 636 /* END INLINE EXPANSION */ 637 goto restart; 638 } 639 } 640 } 641 mtx_unlock_spin(&sched_lock); 642 } 643 644 /* 645 * Make a process sleeping on the specified identifier runnable. 646 * May wake more than one process if a target process is currently 647 * swapped out. 648 */ 649 void 650 wakeup_one(ident) 651 register void *ident; 652 { 653 register struct slpquehead *qp; 654 register struct thread *td; 655 register struct proc *p; 656 657 mtx_lock_spin(&sched_lock); 658 qp = &slpque[LOOKUP(ident)]; 659 660 TAILQ_FOREACH(td, qp, td_slpq) { 661 p = td->td_proc; 662 if (td->td_wchan == ident) { 663 TAILQ_REMOVE(qp, td, td_slpq); 664 td->td_wchan = NULL; 665 if (td->td_proc->p_stat == SSLEEP) { 666 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 667 CTR3(KTR_PROC, "wakeup1: proc %p (pid %d, %s)", 668 p, p->p_pid, p->p_comm); 669 if (td->td_ksegrp->kg_slptime > 1) 670 updatepri(td); 671 td->td_ksegrp->kg_slptime = 0; 672 td->td_kse->ke_slptime = 0; 673 td->td_proc->p_stat = SRUN; 674 if (p->p_sflag & PS_INMEM) { 675 setrunqueue(td); 676 maybe_resched(td->td_ksegrp); 677 break; 678 } else { 679 p->p_sflag |= PS_SWAPINREQ; 680 wakeup((caddr_t)&proc0); 681 } 682 /* END INLINE EXPANSION */ 683 } 684 } 685 } 686 mtx_unlock_spin(&sched_lock); 687 } 688 689 /* 690 * The machine independent parts of mi_switch(). 691 */ 692 void 693 mi_switch() 694 { 695 struct timeval new_switchtime; 696 struct thread *td = curthread; /* XXX */ 697 register struct proc *p = td->td_proc; /* XXX */ 698 #if 0 699 register struct rlimit *rlim; 700 #endif 701 u_int sched_nest; 702 703 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 704 #ifdef INVARIANTS 705 if (p->p_stat != SMTX && p->p_stat != SRUN) 706 mtx_assert(&Giant, MA_NOTOWNED); 707 #endif 708 709 /* 710 * Compute the amount of time during which the current 711 * process was running, and add that to its total so far. 712 */ 713 microuptime(&new_switchtime); 714 if (timevalcmp(&new_switchtime, PCPU_PTR(switchtime), <)) { 715 #if 0 716 /* XXX: This doesn't play well with sched_lock right now. */ 717 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 718 PCPU_GET(switchtime.tv_sec), PCPU_GET(switchtime.tv_usec), 719 new_switchtime.tv_sec, new_switchtime.tv_usec); 720 #endif 721 new_switchtime = PCPU_GET(switchtime); 722 } else { 723 p->p_runtime += (new_switchtime.tv_usec - PCPU_GET(switchtime.tv_usec)) + 724 (new_switchtime.tv_sec - PCPU_GET(switchtime.tv_sec)) * 725 (int64_t)1000000; 726 } 727 728 #ifdef DDB 729 /* 730 * Don't perform context switches from the debugger. 731 */ 732 if (db_active) { 733 mtx_unlock_spin(&sched_lock); 734 db_error("Context switches not allowed in the debugger."); 735 } 736 #endif 737 738 #if 0 739 /* 740 * Check if the process exceeds its cpu resource allocation. 741 * If over max, kill it. 742 * 743 * XXX drop sched_lock, pickup Giant 744 */ 745 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 746 p->p_runtime > p->p_limit->p_cpulimit) { 747 rlim = &p->p_rlimit[RLIMIT_CPU]; 748 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 749 mtx_unlock_spin(&sched_lock); 750 PROC_LOCK(p); 751 killproc(p, "exceeded maximum CPU limit"); 752 mtx_lock_spin(&sched_lock); 753 PROC_UNLOCK(p); 754 } else { 755 mtx_unlock_spin(&sched_lock); 756 PROC_LOCK(p); 757 psignal(p, SIGXCPU); 758 mtx_lock_spin(&sched_lock); 759 PROC_UNLOCK(p); 760 if (rlim->rlim_cur < rlim->rlim_max) { 761 /* XXX: we should make a private copy */ 762 rlim->rlim_cur += 5; 763 } 764 } 765 } 766 #endif 767 768 /* 769 * Pick a new current process and record its start time. 770 */ 771 cnt.v_swtch++; 772 PCPU_SET(switchtime, new_switchtime); 773 CTR3(KTR_PROC, "mi_switch: old proc %p (pid %d, %s)", p, p->p_pid, 774 p->p_comm); 775 sched_nest = sched_lock.mtx_recurse; 776 td->td_lastcpu = td->td_kse->ke_oncpu; 777 td->td_kse->ke_oncpu = NOCPU; 778 td->td_kse->ke_flags &= ~KEF_NEEDRESCHED; 779 cpu_switch(); 780 td->td_kse->ke_oncpu = PCPU_GET(cpuid); 781 sched_lock.mtx_recurse = sched_nest; 782 sched_lock.mtx_lock = (uintptr_t)td; 783 CTR3(KTR_PROC, "mi_switch: new proc %p (pid %d, %s)", p, p->p_pid, 784 p->p_comm); 785 if (PCPU_GET(switchtime.tv_sec) == 0) 786 microuptime(PCPU_PTR(switchtime)); 787 PCPU_SET(switchticks, ticks); 788 } 789 790 /* 791 * Change process state to be runnable, 792 * placing it on the run queue if it is in memory, 793 * and awakening the swapper if it isn't in memory. 794 */ 795 void 796 setrunnable(struct thread *td) 797 { 798 struct proc *p = td->td_proc; 799 800 mtx_lock_spin(&sched_lock); 801 switch (p->p_stat) { 802 case SZOMB: /* not a thread flag XXXKSE */ 803 panic("setrunnable(1)"); 804 } 805 switch (td->td_proc->p_stat) { 806 case 0: 807 case SRUN: 808 case SWAIT: 809 default: 810 panic("setrunnable(2)"); 811 case SSTOP: 812 case SSLEEP: /* e.g. when sending signals */ 813 if (td->td_flags & TDF_CVWAITQ) 814 cv_waitq_remove(td); 815 else 816 unsleep(td); 817 break; 818 819 case SIDL: 820 break; 821 } 822 td->td_proc->p_stat = SRUN; 823 if (td->td_ksegrp->kg_slptime > 1) 824 updatepri(td); 825 td->td_ksegrp->kg_slptime = 0; 826 td->td_kse->ke_slptime = 0; 827 if ((p->p_sflag & PS_INMEM) == 0) { 828 p->p_sflag |= PS_SWAPINREQ; 829 wakeup((caddr_t)&proc0); 830 } else { 831 setrunqueue(td); 832 maybe_resched(td->td_ksegrp); 833 } 834 mtx_unlock_spin(&sched_lock); 835 } 836 837 /* 838 * Compute the priority of a process when running in user mode. 839 * Arrange to reschedule if the resulting priority is better 840 * than that of the current process. 841 */ 842 void 843 resetpriority(kg) 844 register struct ksegrp *kg; 845 { 846 register unsigned int newpriority; 847 848 mtx_lock_spin(&sched_lock); 849 if (kg->kg_pri.pri_class == PRI_TIMESHARE) { 850 newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT + 851 NICE_WEIGHT * (kg->kg_nice - PRIO_MIN); 852 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 853 PRI_MAX_TIMESHARE); 854 kg->kg_pri.pri_user = newpriority; 855 } 856 maybe_resched(kg); 857 mtx_unlock_spin(&sched_lock); 858 } 859 860 /* 861 * Compute a tenex style load average of a quantity on 862 * 1, 5 and 15 minute intervals. 863 * XXXKSE Needs complete rewrite when correct info is available. 864 * Completely Bogus.. only works with 1:1 (but compiles ok now :-) 865 */ 866 static void 867 loadav(void *arg) 868 { 869 int i, nrun; 870 struct loadavg *avg; 871 struct proc *p; 872 struct ksegrp *kg; 873 874 avg = &averunnable; 875 sx_slock(&allproc_lock); 876 nrun = 0; 877 FOREACH_PROC_IN_SYSTEM(p) { 878 FOREACH_KSEGRP_IN_PROC(p, kg) { 879 switch (p->p_stat) { 880 case SRUN: 881 if ((p->p_flag & P_NOLOAD) != 0) 882 goto nextproc; 883 /* FALLTHROUGH */ 884 case SIDL: 885 nrun++; 886 } 887 nextproc: 888 } 889 } 890 sx_sunlock(&allproc_lock); 891 for (i = 0; i < 3; i++) 892 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 893 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 894 895 /* 896 * Schedule the next update to occur after 5 seconds, but add a 897 * random variation to avoid synchronisation with processes that 898 * run at regular intervals. 899 */ 900 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)), 901 loadav, NULL); 902 } 903 904 /* ARGSUSED */ 905 static void 906 sched_setup(dummy) 907 void *dummy; 908 { 909 910 callout_init(&schedcpu_callout, 1); 911 callout_init(&roundrobin_callout, 0); 912 callout_init(&loadav_callout, 0); 913 914 /* Kick off timeout driven events by calling first time. */ 915 roundrobin(NULL); 916 schedcpu(NULL); 917 loadav(NULL); 918 } 919 920 /* 921 * We adjust the priority of the current process. The priority of 922 * a process gets worse as it accumulates CPU time. The cpu usage 923 * estimator (p_estcpu) is increased here. resetpriority() will 924 * compute a different priority each time p_estcpu increases by 925 * INVERSE_ESTCPU_WEIGHT 926 * (until MAXPRI is reached). The cpu usage estimator ramps up 927 * quite quickly when the process is running (linearly), and decays 928 * away exponentially, at a rate which is proportionally slower when 929 * the system is busy. The basic principle is that the system will 930 * 90% forget that the process used a lot of CPU time in 5 * loadav 931 * seconds. This causes the system to favor processes which haven't 932 * run much recently, and to round-robin among other processes. 933 */ 934 void 935 schedclock(td) 936 struct thread *td; 937 { 938 struct kse *ke = td->td_kse; 939 struct ksegrp *kg = td->td_ksegrp; 940 941 if (td) { 942 ke->ke_cpticks++; 943 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1); 944 if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 945 resetpriority(td->td_ksegrp); 946 if (kg->kg_pri.pri_level >= PUSER) 947 kg->kg_pri.pri_level = kg->kg_pri.pri_user; 948 } 949 } else { 950 panic("schedclock"); 951 } 952 } 953 954 /* 955 * General purpose yield system call 956 */ 957 int 958 yield(struct thread *td, struct yield_args *uap) 959 { 960 struct ksegrp *kg = td->td_ksegrp; 961 962 mtx_assert(&Giant, MA_NOTOWNED); 963 mtx_lock_spin(&sched_lock); 964 kg->kg_pri.pri_level = PRI_MAX_TIMESHARE; 965 setrunqueue(td); 966 kg->kg_proc->p_stats->p_ru.ru_nvcsw++; 967 mi_switch(); 968 mtx_unlock_spin(&sched_lock); 969 td->td_retval[0] = 0; 970 971 return (0); 972 } 973 974