kern_synch.c revision 102544
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 102544 2002-08-28 23:45:15Z peter $ 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 ((ke->ke_state == KES_THREAD) && 286 (ke->ke_thread->td_state == TDS_RUNNING))) { 287 ke->ke_slptime = 0; 288 awake = 1; 289 } else { 290 /* XXXKSE 291 * This is probably a pointless 292 * statistic in a KSE world. 293 */ 294 ke->ke_slptime++; 295 } 296 297 /* 298 * pctcpu is only for ps? 299 * Do it per kse.. and add them up at the end? 300 * XXXKSE 301 */ 302 ke->ke_pctcpu 303 = (ke->ke_pctcpu * ccpu) >> FSHIFT; 304 /* 305 * If the kse has been idle the entire second, 306 * stop recalculating its priority until 307 * it wakes up. 308 */ 309 if (ke->ke_slptime > 1) { 310 continue; 311 } 312 313#if (FSHIFT >= CCPU_SHIFT) 314 ke->ke_pctcpu += (realstathz == 100) ? 315 ((fixpt_t) ke->ke_cpticks) << 316 (FSHIFT - CCPU_SHIFT) : 317 100 * (((fixpt_t) ke->ke_cpticks) << 318 (FSHIFT - CCPU_SHIFT)) / realstathz; 319#else 320 ke->ke_pctcpu += ((FSCALE - ccpu) * 321 (ke->ke_cpticks * FSCALE / realstathz)) >> 322 FSHIFT; 323#endif 324 ke->ke_cpticks = 0; 325 } /* end of kse loop */ 326 /* 327 * If there are ANY running threads in this KSEGRP, 328 * then don't count it as sleeping. 329 */ 330 if (awake == 0) { 331 kg->kg_slptime++; 332 } else { 333 kg->kg_slptime = 0; 334 } 335 kg->kg_estcpu = decay_cpu(loadfac, kg->kg_estcpu); 336 resetpriority(kg); 337 FOREACH_THREAD_IN_GROUP(kg, td) { 338 int changedqueue; 339 if (td->td_priority >= PUSER) { 340 /* 341 * Only change the priority 342 * of threads that are still at their 343 * user priority. 344 * XXXKSE This is problematic 345 * as we may need to re-order 346 * the threads on the KSEG list. 347 */ 348 changedqueue = 349 ((td->td_priority / RQ_PPQ) != 350 (kg->kg_user_pri / RQ_PPQ)); 351 352 td->td_priority = kg->kg_user_pri; 353 if (changedqueue && 354 td->td_state == TDS_RUNQ) { 355 /* this could be optimised */ 356 remrunqueue(td); 357 td->td_priority = 358 kg->kg_user_pri; 359 setrunqueue(td); 360 } else { 361 td->td_priority = kg->kg_user_pri; 362 } 363 } 364 } 365 } /* end of ksegrp loop */ 366 mtx_unlock_spin(&sched_lock); 367 } /* end of process loop */ 368 sx_sunlock(&allproc_lock); 369 wakeup(&lbolt); 370 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 371} 372 373/* 374 * Recalculate the priority of a process after it has slept for a while. 375 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 376 * least six times the loadfactor will decay p_estcpu to zero. 377 */ 378void 379updatepri(struct ksegrp *kg) 380{ 381 register unsigned int newcpu; 382 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 383 384 newcpu = kg->kg_estcpu; 385 if (kg->kg_slptime > 5 * loadfac) 386 kg->kg_estcpu = 0; 387 else { 388 kg->kg_slptime--; /* the first time was done in schedcpu */ 389 while (newcpu && --kg->kg_slptime) 390 newcpu = decay_cpu(loadfac, newcpu); 391 kg->kg_estcpu = newcpu; 392 } 393 resetpriority(kg); 394} 395 396/* 397 * We're only looking at 7 bits of the address; everything is 398 * aligned to 4, lots of things are aligned to greater powers 399 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 400 */ 401#define TABLESIZE 128 402static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE]; 403#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 404 405void 406sleepinit(void) 407{ 408 int i; 409 410 sched_quantum = hz/10; 411 hogticks = 2 * sched_quantum; 412 for (i = 0; i < TABLESIZE; i++) 413 TAILQ_INIT(&slpque[i]); 414} 415 416/* 417 * General sleep call. Suspends the current process until a wakeup is 418 * performed on the specified identifier. The process will then be made 419 * runnable with the specified priority. Sleeps at most timo/hz seconds 420 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 421 * before and after sleeping, else signals are not checked. Returns 0 if 422 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 423 * signal needs to be delivered, ERESTART is returned if the current system 424 * call should be restarted if possible, and EINTR is returned if the system 425 * call should be interrupted by the signal (return EINTR). 426 * 427 * The mutex argument is exited before the caller is suspended, and 428 * entered before msleep returns. If priority includes the PDROP 429 * flag the mutex is not entered before returning. 430 */ 431 432int 433msleep(ident, mtx, priority, wmesg, timo) 434 void *ident; 435 struct mtx *mtx; 436 int priority, timo; 437 const char *wmesg; 438{ 439 struct thread *td = curthread; 440 struct proc *p = td->td_proc; 441 int sig, catch = priority & PCATCH; 442 int rval = 0; 443 WITNESS_SAVE_DECL(mtx); 444 445#ifdef KTRACE 446 if (KTRPOINT(td, KTR_CSW)) 447 ktrcsw(1, 0); 448#endif 449 WITNESS_SLEEP(0, &mtx->mtx_object); 450 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 451 ("sleeping without a mutex")); 452 /* 453 * If we are capable of async syscalls and there isn't already 454 * another one ready to return, start a new thread 455 * and queue it as ready to run. Note that there is danger here 456 * because we need to make sure that we don't sleep allocating 457 * the thread (recursion here might be bad). 458 * Hence the TDF_INMSLEEP flag. 459 */ 460 if (p->p_flag & P_KSES) { 461 /* Just don't bother if we are exiting 462 and not the exiting thread. */ 463 if ((p->p_flag & P_WEXIT) && catch && p->p_singlethread != td) 464 return (EINTR); 465 if (td->td_mailbox && (!(td->td_flags & TDF_INMSLEEP))) { 466 /* 467 * If we have no queued work to do, then 468 * upcall to the UTS to see if it has more to do. 469 * We don't need to upcall now, just make it and 470 * queue it. 471 */ 472 mtx_lock_spin(&sched_lock); 473 if (TAILQ_FIRST(&td->td_ksegrp->kg_runq) == NULL) { 474 /* Don't recurse here! */ 475 td->td_flags |= TDF_INMSLEEP; 476 thread_schedule_upcall(td, td->td_kse); 477 td->td_flags &= ~TDF_INMSLEEP; 478 } 479 mtx_unlock_spin(&sched_lock); 480 } 481 } 482 mtx_lock_spin(&sched_lock); 483 if (cold ) { 484 /* 485 * During autoconfiguration, just give interrupts 486 * a chance, then just return. 487 * Don't run any other procs or panic below, 488 * in case this is the idle process and already asleep. 489 */ 490 if (mtx != NULL && priority & PDROP) 491 mtx_unlock(mtx); 492 mtx_unlock_spin(&sched_lock); 493 return (0); 494 } 495 496 DROP_GIANT(); 497 498 if (mtx != NULL) { 499 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 500 WITNESS_SAVE(&mtx->mtx_object, mtx); 501 mtx_unlock(mtx); 502 if (priority & PDROP) 503 mtx = NULL; 504 } 505 506 KASSERT(p != NULL, ("msleep1")); 507 KASSERT(ident != NULL && td->td_state == TDS_RUNNING, ("msleep")); 508 509 td->td_wchan = ident; 510 td->td_wmesg = wmesg; 511 td->td_kse->ke_slptime = 0; /* XXXKSE */ 512 td->td_ksegrp->kg_slptime = 0; 513 td->td_priority = priority & PRIMASK; 514 CTR5(KTR_PROC, "msleep: thread %p (pid %d, %s) on %s (%p)", 515 td, p->p_pid, p->p_comm, wmesg, ident); 516 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], td, td_slpq); 517 if (timo) 518 callout_reset(&td->td_slpcallout, timo, endtsleep, td); 519 /* 520 * We put ourselves on the sleep queue and start our timeout 521 * before calling thread_suspend_check, as we could stop there, and 522 * a wakeup or a SIGCONT (or both) could occur while we were stopped. 523 * without resuming us, thus we must be ready for sleep 524 * when cursig is called. If the wakeup happens while we're 525 * stopped, td->td_wchan will be 0 upon return from cursig. 526 */ 527 if (catch) { 528 CTR3(KTR_PROC, "msleep caught: thread %p (pid %d, %s)", td, 529 p->p_pid, p->p_comm); 530 td->td_flags |= TDF_SINTR; 531 mtx_unlock_spin(&sched_lock); 532 PROC_LOCK(p); 533 sig = cursig(td); 534 if (sig == 0 && thread_suspend_check(1)) 535 sig = SIGSTOP; 536 mtx_lock_spin(&sched_lock); 537 PROC_UNLOCK(p); 538 if (sig != 0) { 539 if (td->td_wchan != NULL) 540 unsleep(td); 541 } else if (td->td_wchan == NULL) 542 catch = 0; 543 } else 544 sig = 0; 545 if (td->td_wchan != NULL) { 546 p->p_stats->p_ru.ru_nvcsw++; 547 td->td_state = TDS_SLP; 548 mi_switch(); 549 } 550 CTR3(KTR_PROC, "msleep resume: thread %p (pid %d, %s)", td, p->p_pid, 551 p->p_comm); 552 KASSERT(td->td_state == TDS_RUNNING, ("running but not TDS_RUNNING")); 553 td->td_flags &= ~TDF_SINTR; 554 if (td->td_flags & TDF_TIMEOUT) { 555 td->td_flags &= ~TDF_TIMEOUT; 556 if (sig == 0) 557 rval = EWOULDBLOCK; 558 } else if (td->td_flags & TDF_TIMOFAIL) { 559 td->td_flags &= ~TDF_TIMOFAIL; 560 } else if (timo && callout_stop(&td->td_slpcallout) == 0) { 561 /* 562 * This isn't supposed to be pretty. If we are here, then 563 * the endtsleep() callout is currently executing on another 564 * CPU and is either spinning on the sched_lock or will be 565 * soon. If we don't synchronize here, there is a chance 566 * that this process may msleep() again before the callout 567 * has a chance to run and the callout may end up waking up 568 * the wrong msleep(). Yuck. 569 */ 570 td->td_flags |= TDF_TIMEOUT; 571 td->td_state = TDS_SLP; 572 p->p_stats->p_ru.ru_nivcsw++; 573 mi_switch(); 574 } 575 mtx_unlock_spin(&sched_lock); 576 577 if (rval == 0 && catch) { 578 PROC_LOCK(p); 579 /* XXX: shouldn't we always be calling cursig() */ 580 if (sig != 0 || (sig = cursig(td))) { 581 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 582 rval = EINTR; 583 else 584 rval = ERESTART; 585 } 586 PROC_UNLOCK(p); 587 } 588#ifdef KTRACE 589 if (KTRPOINT(td, KTR_CSW)) 590 ktrcsw(0, 0); 591#endif 592 PICKUP_GIANT(); 593 if (mtx != NULL) { 594 mtx_lock(mtx); 595 WITNESS_RESTORE(&mtx->mtx_object, mtx); 596 } 597 return (rval); 598} 599 600/* 601 * Implement timeout for msleep() 602 * 603 * If process hasn't been awakened (wchan non-zero), 604 * set timeout flag and undo the sleep. If proc 605 * is stopped, just unsleep so it will remain stopped. 606 * MP-safe, called without the Giant mutex. 607 */ 608static void 609endtsleep(arg) 610 void *arg; 611{ 612 register struct thread *td = arg; 613 614 CTR3(KTR_PROC, "endtsleep: thread %p (pid %d, %s)", td, td->td_proc->p_pid, 615 td->td_proc->p_comm); 616 mtx_lock_spin(&sched_lock); 617 /* 618 * This is the other half of the synchronization with msleep() 619 * described above. If the TDS_TIMEOUT flag is set, we lost the 620 * race and just need to put the process back on the runqueue. 621 */ 622 if ((td->td_flags & TDF_TIMEOUT) != 0) { 623 td->td_flags &= ~TDF_TIMEOUT; 624 if (td->td_proc->p_sflag & PS_INMEM) { 625 setrunqueue(td); 626 maybe_resched(td); 627 } else { 628 td->td_state = TDS_SWAPPED; 629 if ((td->td_proc->p_sflag & PS_SWAPPINGIN) == 0) { 630 td->td_proc->p_sflag |= PS_SWAPINREQ; 631 wakeup(&proc0); 632 } 633 } 634 } else if (td->td_wchan != NULL) { 635 if (td->td_state == TDS_SLP) /* XXXKSE */ 636 setrunnable(td); 637 else 638 unsleep(td); 639 td->td_flags |= TDF_TIMEOUT; 640 } else { 641 td->td_flags |= TDF_TIMOFAIL; 642 } 643 mtx_unlock_spin(&sched_lock); 644} 645 646/* 647 * Abort a thread, as if an interrupt had occured. Only abort 648 * interruptable waits (unfortunatly it isn't only safe to abort others). 649 * This is about identical to cv_abort(). 650 * Think about merging them? 651 * Also, whatever the signal code does... 652 */ 653void 654abortsleep(struct thread *td) 655{ 656 657 mtx_lock_spin(&sched_lock); 658 /* 659 * If the TDF_TIMEOUT flag is set, just leave. A 660 * timeout is scheduled anyhow. 661 */ 662 if ((td->td_flags & (TDF_TIMEOUT | TDF_SINTR)) == TDF_SINTR) { 663 if (td->td_wchan != NULL) { 664 if (td->td_state == TDS_SLP) { /* XXXKSE */ 665 setrunnable(td); 666 } else { 667 /* 668 * Probably in a suspended state.. 669 * um.. dunno XXXKSE 670 */ 671 unsleep(td); 672 } 673 } 674 } 675 mtx_unlock_spin(&sched_lock); 676} 677 678/* 679 * Remove a process from its wait queue 680 */ 681void 682unsleep(struct thread *td) 683{ 684 685 mtx_lock_spin(&sched_lock); 686 if (td->td_wchan != NULL) { 687 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_slpq); 688 td->td_wchan = NULL; 689 } 690 mtx_unlock_spin(&sched_lock); 691} 692 693/* 694 * Make all processes sleeping on the specified identifier runnable. 695 */ 696void 697wakeup(ident) 698 register void *ident; 699{ 700 register struct slpquehead *qp; 701 register struct thread *td; 702 struct thread *ntd; 703 struct ksegrp *kg; 704 struct proc *p; 705 706 mtx_lock_spin(&sched_lock); 707 qp = &slpque[LOOKUP(ident)]; 708restart: 709 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 710 ntd = TAILQ_NEXT(td, td_slpq); 711 p = td->td_proc; 712 if (td->td_wchan == ident) { 713 TAILQ_REMOVE(qp, td, td_slpq); 714 td->td_wchan = NULL; 715 if (td->td_state == TDS_SLP) { 716 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 717 CTR3(KTR_PROC, "wakeup: thread %p (pid %d, %s)", 718 td, p->p_pid, p->p_comm); 719 kg = td->td_ksegrp; 720 if (kg->kg_slptime > 1) 721 updatepri(kg); 722 kg->kg_slptime = 0; 723 if (p->p_sflag & PS_INMEM) { 724 setrunqueue(td); 725 maybe_resched(td); 726 } else { 727 td->td_state = TDS_SWAPPED; 728 if ((p->p_sflag & PS_SWAPPINGIN) == 0) { 729 p->p_sflag |= PS_SWAPINREQ; 730 wakeup(&proc0); 731 } 732 } 733 /* END INLINE EXPANSION */ 734 } 735 goto restart; 736 } 737 } 738 mtx_unlock_spin(&sched_lock); 739} 740 741/* 742 * Make a process sleeping on the specified identifier runnable. 743 * May wake more than one process if a target process is currently 744 * swapped out. 745 */ 746void 747wakeup_one(ident) 748 register void *ident; 749{ 750 register struct slpquehead *qp; 751 register struct thread *td; 752 register struct proc *p; 753 struct thread *ntd; 754 struct ksegrp *kg; 755 756 mtx_lock_spin(&sched_lock); 757 qp = &slpque[LOOKUP(ident)]; 758restart: 759 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 760 ntd = TAILQ_NEXT(td, td_slpq); 761 p = td->td_proc; 762 if (td->td_wchan == ident) { 763 TAILQ_REMOVE(qp, td, td_slpq); 764 td->td_wchan = NULL; 765 if (td->td_state == TDS_SLP) { 766 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 767 CTR3(KTR_PROC,"wakeup1: thread %p (pid %d, %s)", 768 td, p->p_pid, p->p_comm); 769 kg = td->td_ksegrp; 770 if (kg->kg_slptime > 1) 771 updatepri(kg); 772 kg->kg_slptime = 0; 773 if (p->p_sflag & PS_INMEM) { 774 setrunqueue(td); 775 maybe_resched(td); 776 break; 777 } else { 778 td->td_state = TDS_SWAPPED; 779 if ((p->p_sflag & PS_SWAPPINGIN) == 0) { 780 p->p_sflag |= PS_SWAPINREQ; 781 wakeup(&proc0); 782 } 783 } 784 /* END INLINE EXPANSION */ 785 goto restart; 786 } 787 } 788 } 789 mtx_unlock_spin(&sched_lock); 790} 791 792/* 793 * The machine independent parts of mi_switch(). 794 */ 795void 796mi_switch(void) 797{ 798 struct bintime new_switchtime; 799 struct thread *td = curthread; /* XXX */ 800 struct proc *p = td->td_proc; /* XXX */ 801 struct kse *ke = td->td_kse; 802#if 0 803 register struct rlimit *rlim; 804#endif 805 u_int sched_nest; 806 807 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 808 KASSERT((ke->ke_state == KES_THREAD), ("mi_switch: kse state?")); 809 KASSERT((td->td_state != TDS_RUNQ), ("mi_switch: called by old code")); 810#ifdef INVARIANTS 811 if (td->td_state != TDS_MTX && 812 td->td_state != TDS_RUNQ && 813 td->td_state != TDS_RUNNING) 814 mtx_assert(&Giant, MA_NOTOWNED); 815#endif 816 KASSERT(td->td_critnest == 1, 817 ("mi_switch: switch in a critical section")); 818 819 /* 820 * Compute the amount of time during which the current 821 * process was running, and add that to its total so far. 822 */ 823 binuptime(&new_switchtime); 824 bintime_add(&p->p_runtime, &new_switchtime); 825 bintime_sub(&p->p_runtime, PCPU_PTR(switchtime)); 826 827#ifdef DDB 828 /* 829 * Don't perform context switches from the debugger. 830 */ 831 if (db_active) { 832 mtx_unlock_spin(&sched_lock); 833 db_error("Context switches not allowed in the debugger."); 834 } 835#endif 836 837#if 0 838 /* 839 * Check if the process exceeds its cpu resource allocation. 840 * If over max, kill it. 841 * 842 * XXX drop sched_lock, pickup Giant 843 */ 844 if (p->p_state != PRS_ZOMBIE && 845 p->p_limit->p_cpulimit != RLIM_INFINITY && 846 p->p_runtime > p->p_limit->p_cpulimit) { 847 rlim = &p->p_rlimit[RLIMIT_CPU]; 848 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 849 mtx_unlock_spin(&sched_lock); 850 PROC_LOCK(p); 851 killproc(p, "exceeded maximum CPU limit"); 852 mtx_lock_spin(&sched_lock); 853 PROC_UNLOCK(p); 854 } else { 855 mtx_unlock_spin(&sched_lock); 856 PROC_LOCK(p); 857 psignal(p, SIGXCPU); 858 mtx_lock_spin(&sched_lock); 859 PROC_UNLOCK(p); 860 if (rlim->rlim_cur < rlim->rlim_max) { 861 /* XXX: we should make a private copy */ 862 rlim->rlim_cur += 5; 863 } 864 } 865 } 866#endif 867 868 /* 869 * Finish up stats for outgoing thread. 870 */ 871 cnt.v_swtch++; 872 PCPU_SET(switchtime, new_switchtime); 873 CTR3(KTR_PROC, "mi_switch: old thread %p (pid %d, %s)", td, p->p_pid, 874 p->p_comm); 875 sched_nest = sched_lock.mtx_recurse; 876 td->td_lastcpu = ke->ke_oncpu; 877 ke->ke_oncpu = NOCPU; 878 ke->ke_flags &= ~KEF_NEEDRESCHED; 879 /* 880 * At the last moment, if this thread is still marked RUNNING, 881 * then put it back on the run queue as it has not been suspended 882 * or stopped or any thing else similar. 883 */ 884 if (td->td_state == TDS_RUNNING) { 885 KASSERT(((ke->ke_flags & KEF_IDLEKSE) == 0), 886 ("Idle thread in mi_switch with wrong state")); 887 /* Put us back on the run queue (kse and all). */ 888 setrunqueue(td); 889 } else if (td->td_flags & TDF_UNBOUND) { 890 /* 891 * We will not be on the run queue. So we must be 892 * sleeping or similar. If it's available, 893 * someone else can use the KSE if they need it. 894 * XXXKSE KSE loaning will change this. 895 */ 896 td->td_kse = NULL; 897 kse_reassign(ke); 898 } 899 900 cpu_switch(); /* SHAZAM!!*/ 901 902 /* 903 * Start setting up stats etc. for the incoming thread. 904 * Similar code in fork_exit() is returned to by cpu_switch() 905 * in the case of a new thread/process. 906 */ 907 td->td_kse->ke_oncpu = PCPU_GET(cpuid); 908 sched_lock.mtx_recurse = sched_nest; 909 sched_lock.mtx_lock = (uintptr_t)td; 910 CTR3(KTR_PROC, "mi_switch: new thread %p (pid %d, %s)", td, p->p_pid, 911 p->p_comm); 912 if (PCPU_GET(switchtime.sec) == 0) 913 binuptime(PCPU_PTR(switchtime)); 914 PCPU_SET(switchticks, ticks); 915 916 /* 917 * Call the switchin function while still holding the scheduler lock 918 * (used by the idlezero code and the general page-zeroing code) 919 */ 920 if (td->td_switchin) 921 td->td_switchin(); 922} 923 924/* 925 * Change process state to be runnable, 926 * placing it on the run queue if it is in memory, 927 * and awakening the swapper if it isn't in memory. 928 */ 929void 930setrunnable(struct thread *td) 931{ 932 struct proc *p = td->td_proc; 933 struct ksegrp *kg; 934 935 mtx_assert(&sched_lock, MA_OWNED); 936 switch (p->p_state) { 937 case PRS_ZOMBIE: 938 panic("setrunnable(1)"); 939 default: 940 break; 941 } 942 switch (td->td_state) { 943 case 0: 944 case TDS_RUNNING: 945 case TDS_IWAIT: 946 case TDS_SWAPPED: 947 default: 948 printf("state is %d", td->td_state); 949 panic("setrunnable(2)"); 950 case TDS_SUSPENDED: 951 thread_unsuspend(p); 952 break; 953 case TDS_SLP: /* e.g. when sending signals */ 954 if (td->td_flags & TDF_CVWAITQ) 955 cv_waitq_remove(td); 956 else 957 unsleep(td); 958 case TDS_UNQUEUED: /* being put back onto the queue */ 959 case TDS_NEW: /* not yet had time to suspend */ 960 case TDS_RUNQ: /* not yet had time to suspend */ 961 break; 962 } 963 kg = td->td_ksegrp; 964 if (kg->kg_slptime > 1) 965 updatepri(kg); 966 kg->kg_slptime = 0; 967 if ((p->p_sflag & PS_INMEM) == 0) { 968 td->td_state = TDS_SWAPPED; 969 if ((p->p_sflag & PS_SWAPPINGIN) == 0) { 970 p->p_sflag |= PS_SWAPINREQ; 971 wakeup(&proc0); 972 } 973 } else { 974 if (td->td_state != TDS_RUNQ) 975 setrunqueue(td); /* XXXKSE */ 976 maybe_resched(td); 977 } 978} 979 980/* 981 * Compute the priority of a process when running in user mode. 982 * Arrange to reschedule if the resulting priority is better 983 * than that of the current process. 984 */ 985void 986resetpriority(kg) 987 register struct ksegrp *kg; 988{ 989 register unsigned int newpriority; 990 struct thread *td; 991 992 mtx_lock_spin(&sched_lock); 993 if (kg->kg_pri_class == PRI_TIMESHARE) { 994 newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT + 995 NICE_WEIGHT * (kg->kg_nice - PRIO_MIN); 996 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 997 PRI_MAX_TIMESHARE); 998 kg->kg_user_pri = newpriority; 999 } 1000 FOREACH_THREAD_IN_GROUP(kg, td) { 1001 maybe_resched(td); /* XXXKSE silly */ 1002 } 1003 mtx_unlock_spin(&sched_lock); 1004} 1005 1006/* 1007 * Compute a tenex style load average of a quantity on 1008 * 1, 5 and 15 minute intervals. 1009 * XXXKSE Needs complete rewrite when correct info is available. 1010 * Completely Bogus.. only works with 1:1 (but compiles ok now :-) 1011 */ 1012static void 1013loadav(void *arg) 1014{ 1015 int i, nrun; 1016 struct loadavg *avg; 1017 struct proc *p; 1018 struct thread *td; 1019 1020 avg = &averunnable; 1021 sx_slock(&allproc_lock); 1022 nrun = 0; 1023 FOREACH_PROC_IN_SYSTEM(p) { 1024 FOREACH_THREAD_IN_PROC(p, td) { 1025 switch (td->td_state) { 1026 case TDS_RUNQ: 1027 case TDS_RUNNING: 1028 if ((p->p_flag & P_NOLOAD) != 0) 1029 goto nextproc; 1030 nrun++; /* XXXKSE */ 1031 default: 1032 break; 1033 } 1034nextproc: 1035 continue; 1036 } 1037 } 1038 sx_sunlock(&allproc_lock); 1039 for (i = 0; i < 3; i++) 1040 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1041 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1042 1043 /* 1044 * Schedule the next update to occur after 5 seconds, but add a 1045 * random variation to avoid synchronisation with processes that 1046 * run at regular intervals. 1047 */ 1048 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)), 1049 loadav, NULL); 1050} 1051 1052/* ARGSUSED */ 1053static void 1054sched_setup(dummy) 1055 void *dummy; 1056{ 1057 1058 callout_init(&schedcpu_callout, 1); 1059 callout_init(&roundrobin_callout, 0); 1060 callout_init(&loadav_callout, 0); 1061 1062 /* Kick off timeout driven events by calling first time. */ 1063 roundrobin(NULL); 1064 schedcpu(NULL); 1065 loadav(NULL); 1066} 1067 1068/* 1069 * We adjust the priority of the current process. The priority of 1070 * a process gets worse as it accumulates CPU time. The cpu usage 1071 * estimator (p_estcpu) is increased here. resetpriority() will 1072 * compute a different priority each time p_estcpu increases by 1073 * INVERSE_ESTCPU_WEIGHT 1074 * (until MAXPRI is reached). The cpu usage estimator ramps up 1075 * quite quickly when the process is running (linearly), and decays 1076 * away exponentially, at a rate which is proportionally slower when 1077 * the system is busy. The basic principle is that the system will 1078 * 90% forget that the process used a lot of CPU time in 5 * loadav 1079 * seconds. This causes the system to favor processes which haven't 1080 * run much recently, and to round-robin among other processes. 1081 */ 1082void 1083schedclock(td) 1084 struct thread *td; 1085{ 1086 struct kse *ke; 1087 struct ksegrp *kg; 1088 1089 KASSERT((td != NULL), ("schedclock: null thread pointer")); 1090 ke = td->td_kse; 1091 kg = td->td_ksegrp; 1092 ke->ke_cpticks++; 1093 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1); 1094 if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1095 resetpriority(kg); 1096 if (td->td_priority >= PUSER) 1097 td->td_priority = kg->kg_user_pri; 1098 } 1099} 1100 1101/* 1102 * General purpose yield system call 1103 */ 1104int 1105yield(struct thread *td, struct yield_args *uap) 1106{ 1107 struct ksegrp *kg = td->td_ksegrp; 1108 1109 mtx_assert(&Giant, MA_NOTOWNED); 1110 mtx_lock_spin(&sched_lock); 1111 td->td_priority = PRI_MAX_TIMESHARE; 1112 kg->kg_proc->p_stats->p_ru.ru_nvcsw++; 1113 mi_switch(); 1114 mtx_unlock_spin(&sched_lock); 1115 td->td_retval[0] = 0; 1116 1117 return (0); 1118} 1119 1120