kern_synch.c revision 104719
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: head/sys/kern/kern_synch.c 104719 2002-10-09 17:17:24Z jhb $ 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 70static void sched_setup(void *dummy); 71SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 72 73int hogticks; 74int lbolt; 75int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 76 77static struct callout loadav_callout; 78static struct callout schedcpu_callout; 79static struct callout roundrobin_callout; 80 81struct 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 */ 87static 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 93static void endtsleep(void *); 94static void loadav(void *arg); 95static void roundrobin(void *arg); 96static void schedcpu(void *arg); 97 98static int 99sysctl_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 114SYSCTL_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 */ 122void 123maybe_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 131int 132roundrobin_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 */ 145static void 146roundrobin(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 */ 228static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 229SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 230 231/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 232static int fscale __unused = FSCALE; 233SYSCTL_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 */ 254static void 255schedcpu(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 /* 281 * The kse slptimes are not touched in wakeup 282 * because the thread may not HAVE a KSE. 283 */ 284 if (ke->ke_state == KES_ONRUNQ) { 285 awake = 1; 286 ke->ke_flags &= ~KEF_DIDRUN; 287 } else if ((ke->ke_state == KES_THREAD) && 288 (TD_IS_RUNNING(ke->ke_thread))) { 289 awake = 1; 290 /* Do not clear KEF_DIDRUN */ 291 } else if (ke->ke_flags & KEF_DIDRUN) { 292 awake = 1; 293 ke->ke_flags &= ~KEF_DIDRUN; 294 } 295 296 /* 297 * pctcpu is only for ps? 298 * Do it per kse.. and add them up at the end? 299 * XXXKSE 300 */ 301 ke->ke_pctcpu 302 = (ke->ke_pctcpu * ccpu) >> FSHIFT; 303 /* 304 * If the kse has been idle the entire second, 305 * stop recalculating its priority until 306 * it wakes up. 307 */ 308 if (ke->ke_cpticks == 0) 309 continue; 310#if (FSHIFT >= CCPU_SHIFT) 311 ke->ke_pctcpu += (realstathz == 100) ? 312 ((fixpt_t) ke->ke_cpticks) << 313 (FSHIFT - CCPU_SHIFT) : 314 100 * (((fixpt_t) ke->ke_cpticks) << 315 (FSHIFT - CCPU_SHIFT)) / realstathz; 316#else 317 ke->ke_pctcpu += ((FSCALE - ccpu) * 318 (ke->ke_cpticks * FSCALE / realstathz)) >> 319 FSHIFT; 320#endif 321 ke->ke_cpticks = 0; 322 } /* end of kse loop */ 323 /* 324 * If there are ANY running threads in this KSEGRP, 325 * then don't count it as sleeping. 326 */ 327 if (awake) { 328 if (kg->kg_slptime > 1) { 329 /* 330 * In an ideal world, this should not 331 * happen, because whoever woke us 332 * up from the long sleep should have 333 * unwound the slptime and reset our 334 * priority before we run at the stale 335 * priority. Should KASSERT at some 336 * point when all the cases are fixed. 337 */ 338 updatepri(kg); 339 } 340 kg->kg_slptime = 0; 341 } else { 342 kg->kg_slptime++; 343 } 344 if (kg->kg_slptime > 1) 345 continue; 346 kg->kg_estcpu = decay_cpu(loadfac, kg->kg_estcpu); 347 resetpriority(kg); 348 FOREACH_THREAD_IN_GROUP(kg, td) { 349 int changedqueue; 350 if (td->td_priority >= PUSER) { 351 /* 352 * Only change the priority 353 * of threads that are still at their 354 * user priority. 355 * XXXKSE This is problematic 356 * as we may need to re-order 357 * the threads on the KSEG list. 358 */ 359 changedqueue = 360 ((td->td_priority / RQ_PPQ) != 361 (kg->kg_user_pri / RQ_PPQ)); 362 363 td->td_priority = kg->kg_user_pri; 364 if (changedqueue && TD_ON_RUNQ(td)) { 365 /* this could be optimised */ 366 remrunqueue(td); 367 td->td_priority = 368 kg->kg_user_pri; 369 setrunqueue(td); 370 } else { 371 td->td_priority = kg->kg_user_pri; 372 } 373 } 374 } 375 } /* end of ksegrp loop */ 376 mtx_unlock_spin(&sched_lock); 377 } /* end of process loop */ 378 sx_sunlock(&allproc_lock); 379 wakeup(&lbolt); 380 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 381} 382 383/* 384 * Recalculate the priority of a process after it has slept for a while. 385 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 386 * least six times the loadfactor will decay p_estcpu to zero. 387 */ 388void 389updatepri(struct ksegrp *kg) 390{ 391 register unsigned int newcpu; 392 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 393 394 newcpu = kg->kg_estcpu; 395 if (kg->kg_slptime > 5 * loadfac) 396 kg->kg_estcpu = 0; 397 else { 398 kg->kg_slptime--; /* the first time was done in schedcpu */ 399 while (newcpu && --kg->kg_slptime) 400 newcpu = decay_cpu(loadfac, newcpu); 401 kg->kg_estcpu = newcpu; 402 } 403 resetpriority(kg); 404} 405 406/* 407 * We're only looking at 7 bits of the address; everything is 408 * aligned to 4, lots of things are aligned to greater powers 409 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 410 */ 411#define TABLESIZE 128 412static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE]; 413#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 414 415void 416sleepinit(void) 417{ 418 int i; 419 420 sched_quantum = hz/10; 421 hogticks = 2 * sched_quantum; 422 for (i = 0; i < TABLESIZE; i++) 423 TAILQ_INIT(&slpque[i]); 424} 425 426/* 427 * General sleep call. Suspends the current process until a wakeup is 428 * performed on the specified identifier. The process will then be made 429 * runnable with the specified priority. Sleeps at most timo/hz seconds 430 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 431 * before and after sleeping, else signals are not checked. Returns 0 if 432 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 433 * signal needs to be delivered, ERESTART is returned if the current system 434 * call should be restarted if possible, and EINTR is returned if the system 435 * call should be interrupted by the signal (return EINTR). 436 * 437 * The mutex argument is exited before the caller is suspended, and 438 * entered before msleep returns. If priority includes the PDROP 439 * flag the mutex is not entered before returning. 440 */ 441 442int 443msleep(ident, mtx, priority, wmesg, timo) 444 void *ident; 445 struct mtx *mtx; 446 int priority, timo; 447 const char *wmesg; 448{ 449 struct thread *td = curthread; 450 struct proc *p = td->td_proc; 451 int sig, catch = priority & PCATCH; 452 int rval = 0; 453 WITNESS_SAVE_DECL(mtx); 454 455#ifdef KTRACE 456 if (KTRPOINT(td, KTR_CSW)) 457 ktrcsw(1, 0); 458#endif 459 WITNESS_SLEEP(0, &mtx->mtx_object); 460 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 461 ("sleeping without a mutex")); 462 /* 463 * If we are capable of async syscalls and there isn't already 464 * another one ready to return, start a new thread 465 * and queue it as ready to run. Note that there is danger here 466 * because we need to make sure that we don't sleep allocating 467 * the thread (recursion here might be bad). 468 * Hence the TDF_INMSLEEP flag. 469 */ 470 if (p->p_flag & P_KSES) { 471 /* Just don't bother if we are exiting 472 and not the exiting thread. */ 473 if ((p->p_flag & P_WEXIT) && catch && p->p_singlethread != td) 474 return (EINTR); 475 if (td->td_mailbox && (!(td->td_flags & TDF_INMSLEEP))) { 476 /* 477 * Arrange for an upcall to be readied. 478 * it will not actually happen until all 479 * pending in-kernel work for this KSEGRP 480 * has been done. 481 */ 482 mtx_lock_spin(&sched_lock); 483 /* Don't recurse here! */ 484 td->td_flags |= TDF_INMSLEEP; 485 thread_schedule_upcall(td, td->td_kse); 486 td->td_flags &= ~TDF_INMSLEEP; 487 mtx_unlock_spin(&sched_lock); 488 } 489 } 490 mtx_lock_spin(&sched_lock); 491 if (cold ) { 492 /* 493 * During autoconfiguration, just give interrupts 494 * a chance, then just return. 495 * Don't run any other procs or panic below, 496 * in case this is the idle process and already asleep. 497 */ 498 if (mtx != NULL && priority & PDROP) 499 mtx_unlock(mtx); 500 mtx_unlock_spin(&sched_lock); 501 return (0); 502 } 503 504 DROP_GIANT(); 505 506 if (mtx != NULL) { 507 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 508 WITNESS_SAVE(&mtx->mtx_object, mtx); 509 mtx_unlock(mtx); 510 if (priority & PDROP) 511 mtx = NULL; 512 } 513 514 KASSERT(p != NULL, ("msleep1")); 515 KASSERT(ident != NULL && TD_IS_RUNNING(td), ("msleep")); 516 517 CTR5(KTR_PROC, "msleep: thread %p (pid %d, %s) on %s (%p)", 518 td, p->p_pid, p->p_comm, wmesg, ident); 519 520 td->td_wchan = ident; 521 td->td_wmesg = wmesg; 522 td->td_ksegrp->kg_slptime = 0; 523 td->td_priority = priority & PRIMASK; 524 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], td, td_slpq); 525 TD_SET_ON_SLEEPQ(td); 526 if (timo) 527 callout_reset(&td->td_slpcallout, timo, endtsleep, td); 528 /* 529 * We put ourselves on the sleep queue and start our timeout 530 * before calling thread_suspend_check, as we could stop there, and 531 * a wakeup or a SIGCONT (or both) could occur while we were stopped. 532 * without resuming us, thus we must be ready for sleep 533 * when cursig is called. If the wakeup happens while we're 534 * stopped, td->td_wchan will be 0 upon return from cursig. 535 */ 536 if (catch) { 537 CTR3(KTR_PROC, "msleep caught: thread %p (pid %d, %s)", td, 538 p->p_pid, p->p_comm); 539 td->td_flags |= TDF_SINTR; 540 mtx_unlock_spin(&sched_lock); 541 PROC_LOCK(p); 542 sig = cursig(td); 543 if (sig == 0 && thread_suspend_check(1)) 544 sig = SIGSTOP; 545 mtx_lock_spin(&sched_lock); 546 PROC_UNLOCK(p); 547 if (sig != 0) { 548 if (TD_ON_SLEEPQ(td)) 549 unsleep(td); 550 } else if (!TD_ON_SLEEPQ(td)) 551 catch = 0; 552 } else 553 sig = 0; 554 if (TD_ON_SLEEPQ(td)) { 555 p->p_stats->p_ru.ru_nvcsw++; 556 TD_SET_SLEEPING(td); 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_IS_RUNNING(td), ("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_SET_SLEEPING(td); 580 p->p_stats->p_ru.ru_nivcsw++; 581 mi_switch(); 582 td->td_flags &= ~TDF_TIMOFAIL; 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 */ 617static void 618endtsleep(arg) 619 void *arg; 620{ 621 register struct thread *td = arg; 622 623 CTR3(KTR_PROC, "endtsleep: thread %p (pid %d, %s)", 624 td, td->td_proc->p_pid, 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 TDS_TIMEOUT flag is set, we lost the 629 * race and just need to put the process back on the runqueue. 630 */ 631 if (TD_ON_SLEEPQ(td)) { 632 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_slpq); 633 TD_CLR_ON_SLEEPQ(td); 634 td->td_flags |= TDF_TIMEOUT; 635 } else { 636 td->td_flags |= TDF_TIMOFAIL; 637 } 638 TD_CLR_SLEEPING(td); 639 setrunnable(td); 640 mtx_unlock_spin(&sched_lock); 641} 642 643/* 644 * Abort a thread, as if an interrupt had occured. Only abort 645 * interruptable waits (unfortunatly it isn't only safe to abort others). 646 * This is about identical to cv_abort(). 647 * Think about merging them? 648 * Also, whatever the signal code does... 649 */ 650void 651abortsleep(struct thread *td) 652{ 653 654 mtx_assert(&sched_lock, MA_OWNED); 655 /* 656 * If the TDF_TIMEOUT flag is set, just leave. A 657 * timeout is scheduled anyhow. 658 */ 659 if ((td->td_flags & (TDF_TIMEOUT | TDF_SINTR)) == TDF_SINTR) { 660 if (TD_ON_SLEEPQ(td)) { 661 unsleep(td); 662 TD_CLR_SLEEPING(td); 663 setrunnable(td); 664 } 665 } 666} 667 668/* 669 * Remove a process from its wait queue 670 */ 671void 672unsleep(struct thread *td) 673{ 674 675 mtx_lock_spin(&sched_lock); 676 if (TD_ON_SLEEPQ(td)) { 677 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_slpq); 678 TD_CLR_ON_SLEEPQ(td); 679 } 680 mtx_unlock_spin(&sched_lock); 681} 682 683/* 684 * Make all processes sleeping on the specified identifier runnable. 685 */ 686void 687wakeup(ident) 688 register void *ident; 689{ 690 register struct slpquehead *qp; 691 register struct thread *td; 692 struct thread *ntd; 693 struct proc *p; 694 695 mtx_lock_spin(&sched_lock); 696 qp = &slpque[LOOKUP(ident)]; 697restart: 698 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 699 ntd = TAILQ_NEXT(td, td_slpq); 700 if (td->td_wchan == ident) { 701 unsleep(td); 702 TD_CLR_SLEEPING(td); 703 setrunnable(td); 704 p = td->td_proc; 705 CTR3(KTR_PROC,"wakeup: thread %p (pid %d, %s)", 706 td, p->p_pid, p->p_comm); 707 goto restart; 708 } 709 } 710 mtx_unlock_spin(&sched_lock); 711} 712 713/* 714 * Make a process sleeping on the specified identifier runnable. 715 * May wake more than one process if a target process is currently 716 * swapped out. 717 */ 718void 719wakeup_one(ident) 720 register void *ident; 721{ 722 register struct slpquehead *qp; 723 register struct thread *td; 724 register struct proc *p; 725 struct thread *ntd; 726 727 mtx_lock_spin(&sched_lock); 728 qp = &slpque[LOOKUP(ident)]; 729 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 730 ntd = TAILQ_NEXT(td, td_slpq); 731 if (td->td_wchan == ident) { 732 unsleep(td); 733 TD_CLR_SLEEPING(td); 734 setrunnable(td); 735 p = td->td_proc; 736 CTR3(KTR_PROC,"wakeup1: thread %p (pid %d, %s)", 737 td, p->p_pid, p->p_comm); 738 break; 739 } 740 } 741 mtx_unlock_spin(&sched_lock); 742} 743 744/* 745 * The machine independent parts of mi_switch(). 746 */ 747void 748mi_switch(void) 749{ 750 struct bintime new_switchtime; 751 struct thread *td = curthread; /* XXX */ 752 struct proc *p = td->td_proc; /* XXX */ 753 struct kse *ke = td->td_kse; 754 u_int sched_nest; 755 756 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 757 KASSERT((ke->ke_state == KES_THREAD), ("mi_switch: kse state?")); 758 KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code")); 759#ifdef INVARIANTS 760 if (!TD_ON_LOCK(td) && 761 !TD_ON_RUNQ(td) && 762 !TD_IS_RUNNING(td)) 763 mtx_assert(&Giant, MA_NOTOWNED); 764#endif 765 KASSERT(td->td_critnest == 1, 766 ("mi_switch: switch in a critical section")); 767 768 /* 769 * Compute the amount of time during which the current 770 * process was running, and add that to its total so far. 771 */ 772 binuptime(&new_switchtime); 773 bintime_add(&p->p_runtime, &new_switchtime); 774 bintime_sub(&p->p_runtime, PCPU_PTR(switchtime)); 775 776#ifdef DDB 777 /* 778 * Don't perform context switches from the debugger. 779 */ 780 if (db_active) { 781 mtx_unlock_spin(&sched_lock); 782 db_error("Context switches not allowed in the debugger."); 783 } 784#endif 785 786 /* 787 * Check if the process exceeds its cpu resource allocation. If 788 * over max, arrange to kill the process in ast(). 789 */ 790 if (p->p_cpulimit != RLIM_INFINITY && 791 p->p_runtime.sec > p->p_cpulimit) { 792 p->p_sflag |= PS_XCPU; 793 ke->ke_flags |= KEF_ASTPENDING; 794 } 795 796 /* 797 * Finish up stats for outgoing thread. 798 */ 799 cnt.v_swtch++; 800 PCPU_SET(switchtime, new_switchtime); 801 CTR3(KTR_PROC, "mi_switch: old thread %p (pid %d, %s)", td, p->p_pid, 802 p->p_comm); 803 sched_nest = sched_lock.mtx_recurse; 804 td->td_lastcpu = ke->ke_oncpu; 805 ke->ke_oncpu = NOCPU; 806 ke->ke_flags &= ~KEF_NEEDRESCHED; 807 /* 808 * At the last moment, if this thread is still marked RUNNING, 809 * then put it back on the run queue as it has not been suspended 810 * or stopped or any thing else similar. 811 */ 812 if (TD_IS_RUNNING(td)) { 813 /* Put us back on the run queue (kse and all). */ 814 setrunqueue(td); 815 } else if (p->p_flag & P_KSES) { 816 /* 817 * We will not be on the run queue. So we must be 818 * sleeping or similar. As it's available, 819 * someone else can use the KSE if they need it. 820 * (If bound LOANING can still occur). 821 */ 822 kse_reassign(ke); 823 } 824 825 cpu_switch(); /* SHAZAM!!*/ 826 827 /* 828 * Start setting up stats etc. for the incoming thread. 829 * Similar code in fork_exit() is returned to by cpu_switch() 830 * in the case of a new thread/process. 831 */ 832 td->td_kse->ke_oncpu = PCPU_GET(cpuid); 833 sched_lock.mtx_recurse = sched_nest; 834 sched_lock.mtx_lock = (uintptr_t)td; 835 CTR3(KTR_PROC, "mi_switch: new thread %p (pid %d, %s)", td, p->p_pid, 836 p->p_comm); 837 if (PCPU_GET(switchtime.sec) == 0) 838 binuptime(PCPU_PTR(switchtime)); 839 PCPU_SET(switchticks, ticks); 840 841 /* 842 * Call the switchin function while still holding the scheduler lock 843 * (used by the idlezero code and the general page-zeroing code) 844 */ 845 if (td->td_switchin) 846 td->td_switchin(); 847} 848 849/* 850 * Change process state to be runnable, 851 * placing it on the run queue if it is in memory, 852 * and awakening the swapper if it isn't in memory. 853 */ 854void 855setrunnable(struct thread *td) 856{ 857 struct proc *p = td->td_proc; 858 struct ksegrp *kg; 859 860 mtx_assert(&sched_lock, MA_OWNED); 861 switch (p->p_state) { 862 case PRS_ZOMBIE: 863 panic("setrunnable(1)"); 864 default: 865 break; 866 } 867 switch (td->td_state) { 868 case TDS_RUNNING: 869 case TDS_RUNQ: 870 return; 871 case TDS_INHIBITED: 872 /* 873 * If we are only inhibited because we are swapped out 874 * then arange to swap in this process. Otherwise just return. 875 */ 876 if (td->td_inhibitors != TDI_SWAPPED) 877 return; 878 case TDS_CAN_RUN: 879 break; 880 default: 881 printf("state is 0x%x", td->td_state); 882 panic("setrunnable(2)"); 883 } 884 if ((p->p_sflag & PS_INMEM) == 0) { 885 if ((p->p_sflag & PS_SWAPPINGIN) == 0) { 886 p->p_sflag |= PS_SWAPINREQ; 887 wakeup(&proc0); 888 } 889 } else { 890 kg = td->td_ksegrp; 891 if (kg->kg_slptime > 1) 892 updatepri(kg); 893 kg->kg_slptime = 0; 894 setrunqueue(td); 895 maybe_resched(td); 896 } 897} 898 899/* 900 * Compute the priority of a process when running in user mode. 901 * Arrange to reschedule if the resulting priority is better 902 * than that of the current process. 903 */ 904void 905resetpriority(kg) 906 register struct ksegrp *kg; 907{ 908 register unsigned int newpriority; 909 struct thread *td; 910 911 mtx_lock_spin(&sched_lock); 912 if (kg->kg_pri_class == PRI_TIMESHARE) { 913 newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT + 914 NICE_WEIGHT * (kg->kg_nice - PRIO_MIN); 915 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 916 PRI_MAX_TIMESHARE); 917 kg->kg_user_pri = newpriority; 918 } 919 FOREACH_THREAD_IN_GROUP(kg, td) { 920 maybe_resched(td); /* XXXKSE silly */ 921 } 922 mtx_unlock_spin(&sched_lock); 923} 924 925/* 926 * Compute a tenex style load average of a quantity on 927 * 1, 5 and 15 minute intervals. 928 * XXXKSE Needs complete rewrite when correct info is available. 929 * Completely Bogus.. only works with 1:1 (but compiles ok now :-) 930 */ 931static void 932loadav(void *arg) 933{ 934 int i, nrun; 935 struct loadavg *avg; 936 struct proc *p; 937 struct thread *td; 938 939 avg = &averunnable; 940 sx_slock(&allproc_lock); 941 nrun = 0; 942 FOREACH_PROC_IN_SYSTEM(p) { 943 FOREACH_THREAD_IN_PROC(p, td) { 944 switch (td->td_state) { 945 case TDS_RUNQ: 946 case TDS_RUNNING: 947 if ((p->p_flag & P_NOLOAD) != 0) 948 goto nextproc; 949 nrun++; /* XXXKSE */ 950 default: 951 break; 952 } 953nextproc: 954 continue; 955 } 956 } 957 sx_sunlock(&allproc_lock); 958 for (i = 0; i < 3; i++) 959 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 960 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 961 962 /* 963 * Schedule the next update to occur after 5 seconds, but add a 964 * random variation to avoid synchronisation with processes that 965 * run at regular intervals. 966 */ 967 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)), 968 loadav, NULL); 969} 970 971/* ARGSUSED */ 972static void 973sched_setup(dummy) 974 void *dummy; 975{ 976 977 callout_init(&schedcpu_callout, 1); 978 callout_init(&roundrobin_callout, 0); 979 callout_init(&loadav_callout, 0); 980 981 /* Kick off timeout driven events by calling first time. */ 982 roundrobin(NULL); 983 schedcpu(NULL); 984 loadav(NULL); 985} 986 987/* 988 * We adjust the priority of the current process. The priority of 989 * a process gets worse as it accumulates CPU time. The cpu usage 990 * estimator (p_estcpu) is increased here. resetpriority() will 991 * compute a different priority each time p_estcpu increases by 992 * INVERSE_ESTCPU_WEIGHT 993 * (until MAXPRI is reached). The cpu usage estimator ramps up 994 * quite quickly when the process is running (linearly), and decays 995 * away exponentially, at a rate which is proportionally slower when 996 * the system is busy. The basic principle is that the system will 997 * 90% forget that the process used a lot of CPU time in 5 * loadav 998 * seconds. This causes the system to favor processes which haven't 999 * run much recently, and to round-robin among other processes. 1000 */ 1001void 1002schedclock(td) 1003 struct thread *td; 1004{ 1005 struct kse *ke; 1006 struct ksegrp *kg; 1007 1008 KASSERT((td != NULL), ("schedclock: null thread pointer")); 1009 ke = td->td_kse; 1010 kg = td->td_ksegrp; 1011 ke->ke_cpticks++; 1012 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1); 1013 if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1014 resetpriority(kg); 1015 if (td->td_priority >= PUSER) 1016 td->td_priority = kg->kg_user_pri; 1017 } 1018} 1019 1020/* 1021 * General purpose yield system call 1022 */ 1023int 1024yield(struct thread *td, struct yield_args *uap) 1025{ 1026 struct ksegrp *kg = td->td_ksegrp; 1027 1028 mtx_assert(&Giant, MA_NOTOWNED); 1029 mtx_lock_spin(&sched_lock); 1030 td->td_priority = PRI_MAX_TIMESHARE; 1031 kg->kg_proc->p_stats->p_ru.ru_nvcsw++; 1032 mi_switch(); 1033 mtx_unlock_spin(&sched_lock); 1034 td->td_retval[0] = 0; 1035 1036 return (0); 1037} 1038 1039