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", ""); 116 117 /* 118 * Arrange to reschedule if necessary, taking the priorities and 119 * schedulers into account. 120 */ 121 void 122 maybe_resched(kg) 123 struct ksegrp *kg; 124 { 125 126 mtx_assert(&sched_lock, MA_OWNED); 127 if (kg->kg_pri.pri_level < curthread->td_ksegrp->kg_pri.pri_level) 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 register struct proc *p; 260 register struct kse *ke; 261 register struct ksegrp *kg; 262 register int realstathz; 263 int awake; 264 265 realstathz = stathz ? stathz : hz; 266 sx_slock(&allproc_lock); 267 FOREACH_PROC_IN_SYSTEM(p) { 268 mtx_lock_spin(&sched_lock); 269 p->p_swtime++; 270 FOREACH_KSEGRP_IN_PROC(p, kg) { 271 awake = 0; 272 FOREACH_KSE_IN_GROUP(kg, ke) { 273 /* 274 * Increment time in/out of memory and sleep 275 * time (if sleeping). We ignore overflow; 276 * with 16-bit int's (remember them?) 277 * overflow takes 45 days. 278 */ 279 /* XXXKSE */ 280 /* if ((ke->ke_flags & KEF_ONRUNQ) == 0) */ 281 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) { 282 ke->ke_slptime++; 283 } else { 284 ke->ke_slptime = 0; 285 awake = 1; 286 } 287 288 /* 289 * pctcpu is only for ps? 290 * Do it per kse.. and add them up at the end? 291 * XXXKSE 292 */ 293 ke->ke_pctcpu = (ke->ke_pctcpu * ccpu) >> FSHIFT; 294 /* 295 * If the kse has been idle the entire second, 296 * stop recalculating its priority until 297 * it wakes up. 298 */ 299 if (ke->ke_slptime > 1) { 300 continue; 301 } 302 303 #if (FSHIFT >= CCPU_SHIFT) 304 ke->ke_pctcpu += (realstathz == 100) ? 305 ((fixpt_t) ke->ke_cpticks) << 306 (FSHIFT - CCPU_SHIFT) : 307 100 * (((fixpt_t) ke->ke_cpticks) << 308 (FSHIFT - CCPU_SHIFT)) / realstathz; 309 #else 310 ke->ke_pctcpu += ((FSCALE - ccpu) * 311 (ke->ke_cpticks * FSCALE / realstathz)) >> 312 FSHIFT; 313 #endif 314 ke->ke_cpticks = 0; 315 } /* end of kse loop */ 316 if (awake == 0) { 317 kg->kg_slptime++; 318 } else { 319 kg->kg_slptime = 0; 320 } 321 kg->kg_estcpu = decay_cpu(loadfac, kg->kg_estcpu); 322 resetpriority(kg); 323 if (kg->kg_pri.pri_level >= PUSER && 324 (p->p_sflag & PS_INMEM)) { 325 int changedqueue = 326 ((kg->kg_pri.pri_level / RQ_PPQ) != 327 (kg->kg_pri.pri_user / RQ_PPQ)); 328 329 kg->kg_pri.pri_level = kg->kg_pri.pri_user; 330 FOREACH_KSE_IN_GROUP(kg, ke) { 331 if ((ke->ke_oncpu == NOCPU) && /* idle */ 332 (p->p_stat == SRUN) && /* XXXKSE */ 333 changedqueue) { 334 remrunqueue(ke->ke_thread); 335 setrunqueue(ke->ke_thread); 336 } 337 } 338 } 339 } /* end of ksegrp loop */ 340 mtx_unlock_spin(&sched_lock); 341 } /* end of process loop */ 342 sx_sunlock(&allproc_lock); 343 wakeup((caddr_t)&lbolt); 344 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 345 } 346 347 /* 348 * Recalculate the priority of a process after it has slept for a while. 349 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 350 * least six times the loadfactor will decay p_estcpu to zero. 351 */ 352 void 353 updatepri(td) 354 register struct thread *td; 355 { 356 register struct ksegrp *kg; 357 register unsigned int newcpu; 358 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 359 360 if (td == NULL) 361 return; 362 kg = td->td_ksegrp; 363 newcpu = kg->kg_estcpu; 364 if (kg->kg_slptime > 5 * loadfac) 365 kg->kg_estcpu = 0; 366 else { 367 kg->kg_slptime--; /* the first time was done in schedcpu */ 368 while (newcpu && --kg->kg_slptime) 369 newcpu = decay_cpu(loadfac, newcpu); 370 kg->kg_estcpu = newcpu; 371 } 372 resetpriority(td->td_ksegrp); 373 } 374 375 /* 376 * We're only looking at 7 bits of the address; everything is 377 * aligned to 4, lots of things are aligned to greater powers 378 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 379 */ 380 #define TABLESIZE 128 381 static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE]; 382 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 383 384 void 385 sleepinit(void) 386 { 387 int i; 388 389 sched_quantum = hz/10; 390 hogticks = 2 * sched_quantum; 391 for (i = 0; i < TABLESIZE; i++) 392 TAILQ_INIT(&slpque[i]); 393 } 394 395 /* 396 * General sleep call. Suspends the current process until a wakeup is 397 * performed on the specified identifier. The process will then be made 398 * runnable with the specified priority. Sleeps at most timo/hz seconds 399 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 400 * before and after sleeping, else signals are not checked. Returns 0 if 401 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 402 * signal needs to be delivered, ERESTART is returned if the current system 403 * call should be restarted if possible, and EINTR is returned if the system 404 * call should be interrupted by the signal (return EINTR). 405 * 406 * The mutex argument is exited before the caller is suspended, and 407 * entered before msleep returns. If priority includes the PDROP 408 * flag the mutex is not entered before returning. 409 */ 410 int 411 msleep(ident, mtx, priority, wmesg, timo) 412 void *ident; 413 struct mtx *mtx; 414 int priority, timo; 415 const char *wmesg; 416 { 417 struct proc *p = curproc; 418 struct thread *td = curthread; 419 int sig, catch = priority & PCATCH; 420 int rval = 0; 421 WITNESS_SAVE_DECL(mtx); 422 423 #ifdef KTRACE 424 if (p && KTRPOINT(p, KTR_CSW)) 425 ktrcsw(p->p_tracep, 1, 0); 426 #endif 427 WITNESS_SLEEP(0, &mtx->mtx_object); 428 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 429 ("sleeping without a mutex")); 430 mtx_lock_spin(&sched_lock); 431 if (cold || panicstr) { 432 /* 433 * After a panic, or during autoconfiguration, 434 * just give interrupts a chance, then just return; 435 * don't run any other procs or panic below, 436 * in case this is the idle process and already asleep. 437 */ 438 if (mtx != NULL && priority & PDROP) 439 mtx_unlock_flags(mtx, MTX_NOSWITCH); 440 mtx_unlock_spin(&sched_lock); 441 return (0); 442 } 443 444 DROP_GIANT_NOSWITCH(); 445 446 if (mtx != NULL) { 447 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 448 WITNESS_SAVE(&mtx->mtx_object, mtx); 449 mtx_unlock_flags(mtx, MTX_NOSWITCH); 450 if (priority & PDROP) 451 mtx = NULL; 452 } 453 454 KASSERT(p != NULL, ("msleep1")); 455 KASSERT(ident != NULL && td->td_proc->p_stat == SRUN, ("msleep")); 456 457 td->td_wchan = ident; 458 td->td_wmesg = wmesg; 459 td->td_kse->ke_slptime = 0; /* XXXKSE */ 460 td->td_ksegrp->kg_slptime = 0; 461 td->td_ksegrp->kg_pri.pri_level = priority & PRIMASK; 462 CTR5(KTR_PROC, "msleep: thread %p (pid %d, %s) on %s (%p)", 463 td, p->p_pid, p->p_comm, wmesg, ident); 464 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], td, td_slpq); 465 if (timo) 466 callout_reset(&td->td_slpcallout, timo, endtsleep, td); 467 /* 468 * We put ourselves on the sleep queue and start our timeout 469 * before calling CURSIG, as we could stop there, and a wakeup 470 * or a SIGCONT (or both) could occur while we were stopped. 471 * A SIGCONT would cause us to be marked as SSLEEP 472 * without resuming us, thus we must be ready for sleep 473 * when CURSIG is called. If the wakeup happens while we're 474 * stopped, td->td_wchan will be 0 upon return from CURSIG. 475 */ 476 if (catch) { 477 CTR3(KTR_PROC, "msleep caught: proc %p (pid %d, %s)", p, 478 p->p_pid, p->p_comm); 479 td->td_flags |= TDF_SINTR; 480 mtx_unlock_spin(&sched_lock); 481 PROC_LOCK(p); 482 sig = CURSIG(p); 483 mtx_lock_spin(&sched_lock); 484 PROC_UNLOCK_NOSWITCH(p); 485 if (sig != 0) { 486 if (td->td_wchan != NULL) 487 unsleep(td); 488 } else if (td->td_wchan == NULL) 489 catch = 0; 490 } else 491 sig = 0; 492 if (td->td_wchan != NULL) { 493 td->td_proc->p_stat = SSLEEP; 494 p->p_stats->p_ru.ru_nvcsw++; 495 mi_switch(); 496 } 497 CTR3(KTR_PROC, "msleep resume: proc %p (pid %d, %s)", td, p->p_pid, 498 p->p_comm); 499 KASSERT(td->td_proc->p_stat == SRUN, ("running but not SRUN")); 500 td->td_flags &= ~TDF_SINTR; 501 if (td->td_flags & TDF_TIMEOUT) { 502 td->td_flags &= ~TDF_TIMEOUT; 503 if (sig == 0) 504 rval = EWOULDBLOCK; 505 } else if (td->td_flags & TDF_TIMOFAIL) 506 td->td_flags &= ~TDF_TIMOFAIL; 507 else if (timo && callout_stop(&td->td_slpcallout) == 0) { 508 /* 509 * This isn't supposed to be pretty. If we are here, then 510 * the endtsleep() callout is currently executing on another 511 * CPU and is either spinning on the sched_lock or will be 512 * soon. If we don't synchronize here, there is a chance 513 * that this process may msleep() again before the callout 514 * has a chance to run and the callout may end up waking up 515 * the wrong msleep(). Yuck. 516 */ 517 td->td_flags |= TDF_TIMEOUT; 518 p->p_stats->p_ru.ru_nivcsw++; 519 mi_switch(); 520 } 521 mtx_unlock_spin(&sched_lock); 522 523 if (rval == 0 && catch) { 524 PROC_LOCK(p); 525 /* XXX: shouldn't we always be calling CURSIG() */ 526 if (sig != 0 || (sig = CURSIG(p))) { 527 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 528 rval = EINTR; 529 else 530 rval = ERESTART; 531 } 532 PROC_UNLOCK(p); 533 } 534 PICKUP_GIANT(); 535 #ifdef KTRACE 536 mtx_lock(&Giant); 537 if (KTRPOINT(p, KTR_CSW)) 538 ktrcsw(p->p_tracep, 0, 0); 539 mtx_unlock(&Giant); 540 #endif 541 if (mtx != NULL) { 542 mtx_lock(mtx); 543 WITNESS_RESTORE(&mtx->mtx_object, mtx); 544 } 545 return (rval); 546 } 547 548 /* 549 * Implement timeout for msleep() 550 * 551 * If process hasn't been awakened (wchan non-zero), 552 * set timeout flag and undo the sleep. If proc 553 * is stopped, just unsleep so it will remain stopped. 554 * MP-safe, called without the Giant mutex. 555 */ 556 static void 557 endtsleep(arg) 558 void *arg; 559 { 560 register struct thread *td = arg; 561 562 CTR3(KTR_PROC, "endtsleep: thread %p (pid %d, %s)", td, td->td_proc->p_pid, 563 td->td_proc->p_comm); 564 mtx_lock_spin(&sched_lock); 565 /* 566 * This is the other half of the synchronization with msleep() 567 * described above. If the PS_TIMEOUT flag is set, we lost the 568 * race and just need to put the process back on the runqueue. 569 */ 570 if ((td->td_flags & TDF_TIMEOUT) != 0) { 571 td->td_flags &= ~TDF_TIMEOUT; 572 setrunqueue(td); 573 } else if (td->td_wchan != NULL) { 574 if (td->td_proc->p_stat == SSLEEP) /* XXXKSE */ 575 setrunnable(td); 576 else 577 unsleep(td); 578 td->td_flags |= TDF_TIMEOUT; 579 } else { 580 td->td_flags |= TDF_TIMOFAIL; 581 } 582 mtx_unlock_spin(&sched_lock); 583 } 584 585 /* 586 * Remove a process from its wait queue 587 */ 588 void 589 unsleep(struct thread *td) 590 { 591 592 mtx_lock_spin(&sched_lock); 593 if (td->td_wchan != NULL) { 594 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_slpq); 595 td->td_wchan = NULL; 596 } 597 mtx_unlock_spin(&sched_lock); 598 } 599 600 /* 601 * Make all processes sleeping on the specified identifier runnable. 602 */ 603 void 604 wakeup(ident) 605 register void *ident; 606 { 607 register struct slpquehead *qp; 608 register struct thread *td; 609 struct proc *p; 610 611 mtx_lock_spin(&sched_lock); 612 qp = &slpque[LOOKUP(ident)]; 613 restart: 614 TAILQ_FOREACH(td, qp, td_slpq) { 615 p = td->td_proc; 616 if (td->td_wchan == ident) { 617 TAILQ_REMOVE(qp, td, td_slpq); 618 td->td_wchan = NULL; 619 if (td->td_proc->p_stat == SSLEEP) { 620 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 621 CTR3(KTR_PROC, "wakeup: thread %p (pid %d, %s)", 622 td, p->p_pid, p->p_comm); 623 if (td->td_ksegrp->kg_slptime > 1) 624 updatepri(td); 625 td->td_ksegrp->kg_slptime = 0; 626 td->td_kse->ke_slptime = 0; 627 td->td_proc->p_stat = SRUN; 628 if (p->p_sflag & PS_INMEM) { 629 setrunqueue(td); 630 maybe_resched(td->td_ksegrp); 631 } else { 632 p->p_sflag |= PS_SWAPINREQ; 633 wakeup((caddr_t)&proc0); 634 } 635 /* END INLINE EXPANSION */ 636 goto restart; 637 } 638 } 639 } 640 mtx_unlock_spin(&sched_lock); 641 } 642 643 /* 644 * Make a process sleeping on the specified identifier runnable. 645 * May wake more than one process if a target process is currently 646 * swapped out. 647 */ 648 void 649 wakeup_one(ident) 650 register void *ident; 651 { 652 register struct slpquehead *qp; 653 register struct thread *td; 654 register struct proc *p; 655 656 mtx_lock_spin(&sched_lock); 657 qp = &slpque[LOOKUP(ident)]; 658 659 TAILQ_FOREACH(td, qp, td_slpq) { 660 p = td->td_proc; 661 if (td->td_wchan == ident) { 662 TAILQ_REMOVE(qp, td, td_slpq); 663 td->td_wchan = NULL; 664 if (td->td_proc->p_stat == SSLEEP) { 665 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 666 CTR3(KTR_PROC, "wakeup1: proc %p (pid %d, %s)", 667 p, p->p_pid, p->p_comm); 668 if (td->td_ksegrp->kg_slptime > 1) 669 updatepri(td); 670 td->td_ksegrp->kg_slptime = 0; 671 td->td_kse->ke_slptime = 0; 672 td->td_proc->p_stat = SRUN; 673 if (p->p_sflag & PS_INMEM) { 674 setrunqueue(td); 675 maybe_resched(td->td_ksegrp); 676 break; 677 } else { 678 p->p_sflag |= PS_SWAPINREQ; 679 wakeup((caddr_t)&proc0); 680 } 681 /* END INLINE EXPANSION */ 682 } 683 } 684 } 685 mtx_unlock_spin(&sched_lock); 686 } 687 688 /* 689 * The machine independent parts of mi_switch(). 690 */ 691 void 692 mi_switch() 693 { 694 struct timeval new_switchtime; 695 struct thread *td = curthread; /* XXX */ 696 register struct proc *p = td->td_proc; /* XXX */ 697 #if 0 698 register struct rlimit *rlim; 699 #endif 700 critical_t sched_crit; 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_NOSWITCH(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_NOSWITCH(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_crit = sched_lock.mtx_savecrit; 776 sched_nest = sched_lock.mtx_recurse; 777 td->td_lastcpu = td->td_kse->ke_oncpu; 778 td->td_kse->ke_oncpu = NOCPU; 779 td->td_kse->ke_flags &= ~KEF_NEEDRESCHED; 780 cpu_switch(); 781 td->td_kse->ke_oncpu = PCPU_GET(cpuid); 782 sched_lock.mtx_savecrit = sched_crit; 783 sched_lock.mtx_recurse = sched_nest; 784 sched_lock.mtx_lock = (uintptr_t)td; 785 CTR3(KTR_PROC, "mi_switch: new proc %p (pid %d, %s)", p, p->p_pid, 786 p->p_comm); 787 if (PCPU_GET(switchtime.tv_sec) == 0) 788 microuptime(PCPU_PTR(switchtime)); 789 PCPU_SET(switchticks, ticks); 790 } 791 792 /* 793 * Change process state to be runnable, 794 * placing it on the run queue if it is in memory, 795 * and awakening the swapper if it isn't in memory. 796 */ 797 void 798 setrunnable(struct thread *td) 799 { 800 struct proc *p = td->td_proc; 801 802 mtx_lock_spin(&sched_lock); 803 switch (p->p_stat) { 804 case SZOMB: /* not a thread flag XXXKSE */ 805 panic("setrunnable(1)"); 806 } 807 switch (td->td_proc->p_stat) { 808 case 0: 809 case SRUN: 810 case SWAIT: 811 default: 812 panic("setrunnable(2)"); 813 case SSTOP: 814 case SSLEEP: /* e.g. when sending signals */ 815 if (td->td_flags & TDF_CVWAITQ) 816 cv_waitq_remove(td); 817 else 818 unsleep(td); 819 break; 820 821 case SIDL: 822 break; 823 } 824 td->td_proc->p_stat = SRUN; 825 if (td->td_ksegrp->kg_slptime > 1) 826 updatepri(td); 827 td->td_ksegrp->kg_slptime = 0; 828 td->td_kse->ke_slptime = 0; 829 if ((p->p_sflag & PS_INMEM) == 0) { 830 p->p_sflag |= PS_SWAPINREQ; 831 wakeup((caddr_t)&proc0); 832 } else { 833 setrunqueue(td); 834 maybe_resched(td->td_ksegrp); 835 } 836 mtx_unlock_spin(&sched_lock); 837 } 838 839 /* 840 * Compute the priority of a process when running in user mode. 841 * Arrange to reschedule if the resulting priority is better 842 * than that of the current process. 843 */ 844 void 845 resetpriority(kg) 846 register struct ksegrp *kg; 847 { 848 register unsigned int newpriority; 849 850 mtx_lock_spin(&sched_lock); 851 if (kg->kg_pri.pri_class == PRI_TIMESHARE) { 852 newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT + 853 NICE_WEIGHT * (kg->kg_nice - PRIO_MIN); 854 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 855 PRI_MAX_TIMESHARE); 856 kg->kg_pri.pri_user = newpriority; 857 } 858 maybe_resched(kg); 859 mtx_unlock_spin(&sched_lock); 860 } 861 862 /* 863 * Compute a tenex style load average of a quantity on 864 * 1, 5 and 15 minute intervals. 865 * XXXKSE Needs complete rewrite when correct info is available. 866 * Completely Bogus.. only works with 1:1 (but compiles ok now :-) 867 */ 868 static void 869 loadav(void *arg) 870 { 871 int i, nrun; 872 struct loadavg *avg; 873 struct proc *p; 874 struct ksegrp *kg; 875 876 avg = &averunnable; 877 sx_slock(&allproc_lock); 878 nrun = 0; 879 FOREACH_PROC_IN_SYSTEM(p) { 880 FOREACH_KSEGRP_IN_PROC(p, kg) { 881 switch (p->p_stat) { 882 case SRUN: 883 if ((p->p_flag & P_NOLOAD) != 0) 884 goto nextproc; 885 /* FALLTHROUGH */ 886 case SIDL: 887 nrun++; 888 } 889 nextproc: 890 } 891 } 892 sx_sunlock(&allproc_lock); 893 for (i = 0; i < 3; i++) 894 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 895 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 896 897 /* 898 * Schedule the next update to occur after 5 seconds, but add a 899 * random variation to avoid synchronisation with processes that 900 * run at regular intervals. 901 */ 902 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)), 903 loadav, NULL); 904 } 905 906 /* ARGSUSED */ 907 static void 908 sched_setup(dummy) 909 void *dummy; 910 { 911 912 callout_init(&schedcpu_callout, 1); 913 callout_init(&roundrobin_callout, 0); 914 callout_init(&loadav_callout, 0); 915 916 /* Kick off timeout driven events by calling first time. */ 917 roundrobin(NULL); 918 schedcpu(NULL); 919 loadav(NULL); 920 } 921 922 /* 923 * We adjust the priority of the current process. The priority of 924 * a process gets worse as it accumulates CPU time. The cpu usage 925 * estimator (p_estcpu) is increased here. resetpriority() will 926 * compute a different priority each time p_estcpu increases by 927 * INVERSE_ESTCPU_WEIGHT 928 * (until MAXPRI is reached). The cpu usage estimator ramps up 929 * quite quickly when the process is running (linearly), and decays 930 * away exponentially, at a rate which is proportionally slower when 931 * the system is busy. The basic principle is that the system will 932 * 90% forget that the process used a lot of CPU time in 5 * loadav 933 * seconds. This causes the system to favor processes which haven't 934 * run much recently, and to round-robin among other processes. 935 */ 936 void 937 schedclock(td) 938 struct thread *td; 939 { 940 struct kse *ke = td->td_kse; 941 struct ksegrp *kg = td->td_ksegrp; 942 943 if (td) { 944 ke->ke_cpticks++; 945 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1); 946 if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 947 resetpriority(td->td_ksegrp); 948 if (kg->kg_pri.pri_level >= PUSER) 949 kg->kg_pri.pri_level = kg->kg_pri.pri_user; 950 } 951 } else { 952 panic("schedclock"); 953 } 954 } 955 956 /* 957 * General purpose yield system call 958 */ 959 int 960 yield(struct thread *td, struct yield_args *uap) 961 { 962 struct ksegrp *kg = td->td_ksegrp; 963 964 mtx_assert(&Giant, MA_NOTOWNED); 965 mtx_lock_spin(&sched_lock); 966 kg->kg_pri.pri_level = PRI_MAX_TIMESHARE; 967 setrunqueue(td); 968 kg->kg_proc->p_stats->p_ru.ru_nvcsw++; 969 mi_switch(); 970 mtx_unlock_spin(&sched_lock); 971 td->td_retval[0] = 0; 972 973 return (0); 974 } 975 976