kern_synch.c revision 104295
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 104295 2002-10-01 14:10:08Z 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 * If we have no queued work to do, then 478 * upcall to the UTS to see if it has more to do. 479 * We don't need to upcall now, just make it and 480 * queue it. 481 */ 482 mtx_lock_spin(&sched_lock); 483 if (TAILQ_FIRST(&td->td_ksegrp->kg_runq) == NULL) { 484 /* Don't recurse here! */ 485 td->td_flags |= TDF_INMSLEEP; 486 thread_schedule_upcall(td, td->td_kse); 487 td->td_flags &= ~TDF_INMSLEEP; 488 } 489 mtx_unlock_spin(&sched_lock); 490 } 491 } 492 mtx_lock_spin(&sched_lock); 493 if (cold ) { 494 /* 495 * During autoconfiguration, just give interrupts 496 * a chance, then just return. 497 * Don't run any other procs or panic below, 498 * in case this is the idle process and already asleep. 499 */ 500 if (mtx != NULL && priority & PDROP) 501 mtx_unlock(mtx); 502 mtx_unlock_spin(&sched_lock); 503 return (0); 504 } 505 506 DROP_GIANT(); 507 508 if (mtx != NULL) { 509 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 510 WITNESS_SAVE(&mtx->mtx_object, mtx); 511 mtx_unlock(mtx); 512 if (priority & PDROP) 513 mtx = NULL; 514 } 515 516 KASSERT(p != NULL, ("msleep1")); 517 KASSERT(ident != NULL && TD_IS_RUNNING(td), ("msleep")); 518 519 CTR5(KTR_PROC, "msleep: thread %p (pid %d, %s) on %s (%p)", 520 td, p->p_pid, p->p_comm, wmesg, ident); 521 522 td->td_wchan = ident; 523 td->td_wmesg = wmesg; 524 td->td_ksegrp->kg_slptime = 0; 525 td->td_priority = priority & PRIMASK; 526 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], td, td_slpq); 527 TD_SET_ON_SLEEPQ(td); 528 if (timo) 529 callout_reset(&td->td_slpcallout, timo, endtsleep, td); 530 /* 531 * We put ourselves on the sleep queue and start our timeout 532 * before calling thread_suspend_check, as we could stop there, and 533 * a wakeup or a SIGCONT (or both) could occur while we were stopped. 534 * without resuming us, thus we must be ready for sleep 535 * when cursig is called. If the wakeup happens while we're 536 * stopped, td->td_wchan will be 0 upon return from cursig. 537 */ 538 if (catch) { 539 CTR3(KTR_PROC, "msleep caught: thread %p (pid %d, %s)", td, 540 p->p_pid, p->p_comm); 541 td->td_flags |= TDF_SINTR; 542 mtx_unlock_spin(&sched_lock); 543 PROC_LOCK(p); 544 sig = cursig(td); 545 if (sig == 0 && thread_suspend_check(1)) 546 sig = SIGSTOP; 547 mtx_lock_spin(&sched_lock); 548 PROC_UNLOCK(p); 549 if (sig != 0) { 550 if (TD_ON_SLEEPQ(td)) 551 unsleep(td); 552 } else if (!TD_ON_SLEEPQ(td)) 553 catch = 0; 554 } else 555 sig = 0; 556 if (TD_ON_SLEEPQ(td)) { 557 p->p_stats->p_ru.ru_nvcsw++; 558 TD_SET_SLEEPING(td); 559 mi_switch(); 560 } 561 CTR3(KTR_PROC, "msleep resume: thread %p (pid %d, %s)", td, p->p_pid, 562 p->p_comm); 563 KASSERT(TD_IS_RUNNING(td), ("running but not TDS_RUNNING")); 564 td->td_flags &= ~TDF_SINTR; 565 if (td->td_flags & TDF_TIMEOUT) { 566 td->td_flags &= ~TDF_TIMEOUT; 567 if (sig == 0) 568 rval = EWOULDBLOCK; 569 } else if (td->td_flags & TDF_TIMOFAIL) { 570 td->td_flags &= ~TDF_TIMOFAIL; 571 } else if (timo && callout_stop(&td->td_slpcallout) == 0) { 572 /* 573 * This isn't supposed to be pretty. If we are here, then 574 * the endtsleep() callout is currently executing on another 575 * CPU and is either spinning on the sched_lock or will be 576 * soon. If we don't synchronize here, there is a chance 577 * that this process may msleep() again before the callout 578 * has a chance to run and the callout may end up waking up 579 * the wrong msleep(). Yuck. 580 */ 581 TD_SET_SLEEPING(td); 582 p->p_stats->p_ru.ru_nivcsw++; 583 mi_switch(); 584 td->td_flags &= ~TDF_TIMOFAIL; 585 } 586 mtx_unlock_spin(&sched_lock); 587 588 if (rval == 0 && catch) { 589 PROC_LOCK(p); 590 /* XXX: shouldn't we always be calling cursig() */ 591 if (sig != 0 || (sig = cursig(td))) { 592 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 593 rval = EINTR; 594 else 595 rval = ERESTART; 596 } 597 PROC_UNLOCK(p); 598 } 599#ifdef KTRACE 600 if (KTRPOINT(td, KTR_CSW)) 601 ktrcsw(0, 0); 602#endif 603 PICKUP_GIANT(); 604 if (mtx != NULL) { 605 mtx_lock(mtx); 606 WITNESS_RESTORE(&mtx->mtx_object, mtx); 607 } 608 return (rval); 609} 610 611/* 612 * Implement timeout for msleep() 613 * 614 * If process hasn't been awakened (wchan non-zero), 615 * set timeout flag and undo the sleep. If proc 616 * is stopped, just unsleep so it will remain stopped. 617 * MP-safe, called without the Giant mutex. 618 */ 619static void 620endtsleep(arg) 621 void *arg; 622{ 623 register struct thread *td = arg; 624 625 CTR3(KTR_PROC, "endtsleep: thread %p (pid %d, %s)", 626 td, td->td_proc->p_pid, td->td_proc->p_comm); 627 mtx_lock_spin(&sched_lock); 628 /* 629 * This is the other half of the synchronization with msleep() 630 * described above. If the TDS_TIMEOUT flag is set, we lost the 631 * race and just need to put the process back on the runqueue. 632 */ 633 if (TD_ON_SLEEPQ(td)) { 634 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_slpq); 635 TD_CLR_ON_SLEEPQ(td); 636 td->td_flags |= TDF_TIMEOUT; 637 } else { 638 td->td_flags |= TDF_TIMOFAIL; 639 } 640 TD_CLR_SLEEPING(td); 641 setrunnable(td); 642 mtx_unlock_spin(&sched_lock); 643} 644 645/* 646 * Abort a thread, as if an interrupt had occured. Only abort 647 * interruptable waits (unfortunatly it isn't only safe to abort others). 648 * This is about identical to cv_abort(). 649 * Think about merging them? 650 * Also, whatever the signal code does... 651 */ 652void 653abortsleep(struct thread *td) 654{ 655 656 mtx_assert(&sched_lock, MA_OWNED); 657 /* 658 * If the TDF_TIMEOUT flag is set, just leave. A 659 * timeout is scheduled anyhow. 660 */ 661 if ((td->td_flags & (TDF_TIMEOUT | TDF_SINTR)) == TDF_SINTR) { 662 if (TD_ON_SLEEPQ(td)) { 663 unsleep(td); 664 TD_CLR_SLEEPING(td); 665 setrunnable(td); 666 } 667 } 668} 669 670/* 671 * Remove a process from its wait queue 672 */ 673void 674unsleep(struct thread *td) 675{ 676 677 mtx_lock_spin(&sched_lock); 678 if (TD_ON_SLEEPQ(td)) { 679 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_slpq); 680 TD_CLR_ON_SLEEPQ(td); 681 } 682 mtx_unlock_spin(&sched_lock); 683} 684 685/* 686 * Make all processes sleeping on the specified identifier runnable. 687 */ 688void 689wakeup(ident) 690 register void *ident; 691{ 692 register struct slpquehead *qp; 693 register struct thread *td; 694 struct thread *ntd; 695 struct proc *p; 696 697 mtx_lock_spin(&sched_lock); 698 qp = &slpque[LOOKUP(ident)]; 699restart: 700 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 701 ntd = TAILQ_NEXT(td, td_slpq); 702 if (td->td_wchan == ident) { 703 unsleep(td); 704 TD_CLR_SLEEPING(td); 705 setrunnable(td); 706 p = td->td_proc; 707 CTR3(KTR_PROC,"wakeup: thread %p (pid %d, %s)", 708 td, p->p_pid, p->p_comm); 709 goto restart; 710 } 711 } 712 mtx_unlock_spin(&sched_lock); 713} 714 715/* 716 * Make a process sleeping on the specified identifier runnable. 717 * May wake more than one process if a target process is currently 718 * swapped out. 719 */ 720void 721wakeup_one(ident) 722 register void *ident; 723{ 724 register struct slpquehead *qp; 725 register struct thread *td; 726 register struct proc *p; 727 struct thread *ntd; 728 729 mtx_lock_spin(&sched_lock); 730 qp = &slpque[LOOKUP(ident)]; 731 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 732 ntd = TAILQ_NEXT(td, td_slpq); 733 if (td->td_wchan == ident) { 734 unsleep(td); 735 TD_CLR_SLEEPING(td); 736 setrunnable(td); 737 p = td->td_proc; 738 CTR3(KTR_PROC,"wakeup1: thread %p (pid %d, %s)", 739 td, p->p_pid, p->p_comm); 740 break; 741 } 742 } 743 mtx_unlock_spin(&sched_lock); 744} 745 746/* 747 * The machine independent parts of mi_switch(). 748 */ 749void 750mi_switch(void) 751{ 752 struct bintime new_switchtime; 753 struct thread *td = curthread; /* XXX */ 754 struct proc *p = td->td_proc; /* XXX */ 755 struct kse *ke = td->td_kse; 756 u_int sched_nest; 757 758 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 759 KASSERT((ke->ke_state == KES_THREAD), ("mi_switch: kse state?")); 760 KASSERT(!TD_ON_RUNQ(td), ("mi_switch: called by old code")); 761#ifdef INVARIANTS 762 if (!TD_ON_MUTEX(td) && 763 !TD_ON_RUNQ(td) && 764 !TD_IS_RUNNING(td)) 765 mtx_assert(&Giant, MA_NOTOWNED); 766#endif 767 KASSERT(td->td_critnest == 1, 768 ("mi_switch: switch in a critical section")); 769 770 /* 771 * Compute the amount of time during which the current 772 * process was running, and add that to its total so far. 773 */ 774 binuptime(&new_switchtime); 775 bintime_add(&p->p_runtime, &new_switchtime); 776 bintime_sub(&p->p_runtime, PCPU_PTR(switchtime)); 777 778#ifdef DDB 779 /* 780 * Don't perform context switches from the debugger. 781 */ 782 if (db_active) { 783 mtx_unlock_spin(&sched_lock); 784 db_error("Context switches not allowed in the debugger."); 785 } 786#endif 787 788 /* 789 * Check if the process exceeds its cpu resource allocation. If 790 * over max, arrange to kill the process in ast(). 791 */ 792 if (p->p_state != PRS_ZOMBIE && 793 p->p_limit->p_cpulimit != RLIM_INFINITY && 794 p->p_runtime.sec > p->p_limit->p_cpulimit) { 795 p->p_sflag |= PS_XCPU; 796 ke->ke_flags |= KEF_ASTPENDING; 797 } 798 799 /* 800 * Finish up stats for outgoing thread. 801 */ 802 cnt.v_swtch++; 803 PCPU_SET(switchtime, new_switchtime); 804 CTR3(KTR_PROC, "mi_switch: old thread %p (pid %d, %s)", td, p->p_pid, 805 p->p_comm); 806 sched_nest = sched_lock.mtx_recurse; 807 td->td_lastcpu = ke->ke_oncpu; 808 ke->ke_oncpu = NOCPU; 809 ke->ke_flags &= ~KEF_NEEDRESCHED; 810 /* 811 * At the last moment, if this thread is still marked RUNNING, 812 * then put it back on the run queue as it has not been suspended 813 * or stopped or any thing else similar. 814 */ 815 if (TD_IS_RUNNING(td)) { 816 KASSERT(((ke->ke_flags & KEF_IDLEKSE) == 0), 817 ("Idle thread in mi_switch with wrong state")); 818 /* Put us back on the run queue (kse and all). */ 819 setrunqueue(td); 820 } else if (td->td_flags & TDF_UNBOUND) { 821 /* 822 * We will not be on the run queue. So we must be 823 * sleeping or similar. If it's available, 824 * someone else can use the KSE if they need it. 825 * XXXKSE KSE loaning will change this. 826 */ 827 td->td_kse = NULL; 828 kse_reassign(ke); 829 } else if (p->p_flag & P_KSES) { 830 KASSERT(((ke->ke_bound == NULL) || (ke->ke_bound == td)), 831 ("mi_switch: bad bound state")); 832 ke->ke_bound = td; 833 kse_reassign(ke); 834 } 835 836 cpu_switch(); /* SHAZAM!!*/ 837 838 /* 839 * Start setting up stats etc. for the incoming thread. 840 * Similar code in fork_exit() is returned to by cpu_switch() 841 * in the case of a new thread/process. 842 */ 843 td->td_kse->ke_oncpu = PCPU_GET(cpuid); 844 sched_lock.mtx_recurse = sched_nest; 845 sched_lock.mtx_lock = (uintptr_t)td; 846 CTR3(KTR_PROC, "mi_switch: new thread %p (pid %d, %s)", td, p->p_pid, 847 p->p_comm); 848 if (PCPU_GET(switchtime.sec) == 0) 849 binuptime(PCPU_PTR(switchtime)); 850 PCPU_SET(switchticks, ticks); 851 852 /* 853 * Call the switchin function while still holding the scheduler lock 854 * (used by the idlezero code and the general page-zeroing code) 855 */ 856 if (td->td_switchin) 857 td->td_switchin(); 858} 859 860/* 861 * Change process state to be runnable, 862 * placing it on the run queue if it is in memory, 863 * and awakening the swapper if it isn't in memory. 864 */ 865void 866setrunnable(struct thread *td) 867{ 868 struct proc *p = td->td_proc; 869 struct ksegrp *kg; 870 871 mtx_assert(&sched_lock, MA_OWNED); 872 switch (p->p_state) { 873 case PRS_ZOMBIE: 874 panic("setrunnable(1)"); 875 default: 876 break; 877 } 878 switch (td->td_state) { 879 case TDS_RUNNING: 880 case TDS_RUNQ: 881 return; 882 case TDS_INHIBITED: 883 /* 884 * If we are only inhibited because we are swapped out 885 * then arange to swap in this process. Otherwise just return. 886 */ 887 if (td->td_inhibitors != TDI_SWAPPED) 888 return; 889 case TDS_CAN_RUN: 890 break; 891 default: 892 printf("state is 0x%x", td->td_state); 893 panic("setrunnable(2)"); 894 } 895 if ((p->p_sflag & PS_INMEM) == 0) { 896 if ((p->p_sflag & PS_SWAPPINGIN) == 0) { 897 p->p_sflag |= PS_SWAPINREQ; 898 wakeup(&proc0); 899 } 900 } else { 901 kg = td->td_ksegrp; 902 if (kg->kg_slptime > 1) 903 updatepri(kg); 904 kg->kg_slptime = 0; 905 setrunqueue(td); 906 maybe_resched(td); 907 } 908} 909 910/* 911 * Compute the priority of a process when running in user mode. 912 * Arrange to reschedule if the resulting priority is better 913 * than that of the current process. 914 */ 915void 916resetpriority(kg) 917 register struct ksegrp *kg; 918{ 919 register unsigned int newpriority; 920 struct thread *td; 921 922 mtx_lock_spin(&sched_lock); 923 if (kg->kg_pri_class == PRI_TIMESHARE) { 924 newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT + 925 NICE_WEIGHT * (kg->kg_nice - PRIO_MIN); 926 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 927 PRI_MAX_TIMESHARE); 928 kg->kg_user_pri = newpriority; 929 } 930 FOREACH_THREAD_IN_GROUP(kg, td) { 931 maybe_resched(td); /* XXXKSE silly */ 932 } 933 mtx_unlock_spin(&sched_lock); 934} 935 936/* 937 * Compute a tenex style load average of a quantity on 938 * 1, 5 and 15 minute intervals. 939 * XXXKSE Needs complete rewrite when correct info is available. 940 * Completely Bogus.. only works with 1:1 (but compiles ok now :-) 941 */ 942static void 943loadav(void *arg) 944{ 945 int i, nrun; 946 struct loadavg *avg; 947 struct proc *p; 948 struct thread *td; 949 950 avg = &averunnable; 951 sx_slock(&allproc_lock); 952 nrun = 0; 953 FOREACH_PROC_IN_SYSTEM(p) { 954 FOREACH_THREAD_IN_PROC(p, td) { 955 switch (td->td_state) { 956 case TDS_RUNQ: 957 case TDS_RUNNING: 958 if ((p->p_flag & P_NOLOAD) != 0) 959 goto nextproc; 960 nrun++; /* XXXKSE */ 961 default: 962 break; 963 } 964nextproc: 965 continue; 966 } 967 } 968 sx_sunlock(&allproc_lock); 969 for (i = 0; i < 3; i++) 970 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 971 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 972 973 /* 974 * Schedule the next update to occur after 5 seconds, but add a 975 * random variation to avoid synchronisation with processes that 976 * run at regular intervals. 977 */ 978 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)), 979 loadav, NULL); 980} 981 982/* ARGSUSED */ 983static void 984sched_setup(dummy) 985 void *dummy; 986{ 987 988 callout_init(&schedcpu_callout, 1); 989 callout_init(&roundrobin_callout, 0); 990 callout_init(&loadav_callout, 0); 991 992 /* Kick off timeout driven events by calling first time. */ 993 roundrobin(NULL); 994 schedcpu(NULL); 995 loadav(NULL); 996} 997 998/* 999 * We adjust the priority of the current process. The priority of 1000 * a process gets worse as it accumulates CPU time. The cpu usage 1001 * estimator (p_estcpu) is increased here. resetpriority() will 1002 * compute a different priority each time p_estcpu increases by 1003 * INVERSE_ESTCPU_WEIGHT 1004 * (until MAXPRI is reached). The cpu usage estimator ramps up 1005 * quite quickly when the process is running (linearly), and decays 1006 * away exponentially, at a rate which is proportionally slower when 1007 * the system is busy. The basic principle is that the system will 1008 * 90% forget that the process used a lot of CPU time in 5 * loadav 1009 * seconds. This causes the system to favor processes which haven't 1010 * run much recently, and to round-robin among other processes. 1011 */ 1012void 1013schedclock(td) 1014 struct thread *td; 1015{ 1016 struct kse *ke; 1017 struct ksegrp *kg; 1018 1019 KASSERT((td != NULL), ("schedclock: null thread pointer")); 1020 ke = td->td_kse; 1021 kg = td->td_ksegrp; 1022 ke->ke_cpticks++; 1023 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1); 1024 if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1025 resetpriority(kg); 1026 if (td->td_priority >= PUSER) 1027 td->td_priority = kg->kg_user_pri; 1028 } 1029} 1030 1031/* 1032 * General purpose yield system call 1033 */ 1034int 1035yield(struct thread *td, struct yield_args *uap) 1036{ 1037 struct ksegrp *kg = td->td_ksegrp; 1038 1039 mtx_assert(&Giant, MA_NOTOWNED); 1040 mtx_lock_spin(&sched_lock); 1041 td->td_priority = PRI_MAX_TIMESHARE; 1042 kg->kg_proc->p_stats->p_ru.ru_nvcsw++; 1043 mi_switch(); 1044 mtx_unlock_spin(&sched_lock); 1045 td->td_retval[0] = 0; 1046 1047 return (0); 1048} 1049 1050