kern_synch.c revision 100209
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 100209 2002-07-17 02:23:44Z gallatin $ 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 /* XXXKSE **WRONG***/ 281 /* 282 * the kse slptimes are not touched in wakeup 283 * because the thread may not HAVE a KSE 284 */ 285 if ((ke->ke_state == KES_ONRUNQ) || 286 ((ke->ke_state == KES_THREAD) && 287 (ke->ke_thread->td_state == TDS_RUNNING))) { 288 ke->ke_slptime++; 289 } else { 290 ke->ke_slptime = 0; 291 awake = 1; 292 } 293 294 /* 295 * pctcpu is only for ps? 296 * Do it per kse.. and add them up at the end? 297 * XXXKSE 298 */ 299 ke->ke_pctcpu = (ke->ke_pctcpu * ccpu) >> FSHIFT; 300 /* 301 * If the kse has been idle the entire second, 302 * stop recalculating its priority until 303 * it wakes up. 304 */ 305 if (ke->ke_slptime > 1) { 306 continue; 307 } 308 309#if (FSHIFT >= CCPU_SHIFT) 310 ke->ke_pctcpu += (realstathz == 100) ? 311 ((fixpt_t) ke->ke_cpticks) << 312 (FSHIFT - CCPU_SHIFT) : 313 100 * (((fixpt_t) ke->ke_cpticks) << 314 (FSHIFT - CCPU_SHIFT)) / realstathz; 315#else 316 ke->ke_pctcpu += ((FSCALE - ccpu) * 317 (ke->ke_cpticks * FSCALE / realstathz)) >> 318 FSHIFT; 319#endif 320 ke->ke_cpticks = 0; 321 } /* end of kse loop */ 322 if (awake == 0) { 323 kg->kg_slptime++; 324 } else { 325 kg->kg_slptime = 0; 326 } 327 kg->kg_estcpu = decay_cpu(loadfac, kg->kg_estcpu); 328 resetpriority(kg); 329 FOREACH_THREAD_IN_GROUP(kg, td) { 330 int changedqueue; 331 if (td->td_priority >= PUSER) { 332 /* 333 * Only change the priority 334 * of threads that are still at their 335 * user priority. 336 * XXXKSE This is problematic 337 * as we may need to re-order 338 * the threads on the KSEG list. 339 */ 340 changedqueue = 341 ((td->td_priority / RQ_PPQ) != 342 (kg->kg_user_pri / RQ_PPQ)); 343 344 td->td_priority = kg->kg_user_pri; 345 if (changedqueue && 346 td->td_state == TDS_RUNQ) { 347 /* this could be optimised */ 348 remrunqueue(td); 349 td->td_priority = 350 kg->kg_user_pri; 351 setrunqueue(td); 352 } else { 353 td->td_priority = kg->kg_user_pri; 354 } 355 } 356 } 357 } /* end of ksegrp loop */ 358 mtx_unlock_spin(&sched_lock); 359 } /* end of process loop */ 360 sx_sunlock(&allproc_lock); 361 wakeup(&lbolt); 362 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 363} 364 365/* 366 * Recalculate the priority of a process after it has slept for a while. 367 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 368 * least six times the loadfactor will decay p_estcpu to zero. 369 */ 370void 371updatepri(td) 372 register struct thread *td; 373{ 374 register struct ksegrp *kg; 375 register unsigned int newcpu; 376 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 377 378 if (td == NULL) 379 return; 380 kg = td->td_ksegrp; 381 newcpu = kg->kg_estcpu; 382 if (kg->kg_slptime > 5 * loadfac) 383 kg->kg_estcpu = 0; 384 else { 385 kg->kg_slptime--; /* the first time was done in schedcpu */ 386 while (newcpu && --kg->kg_slptime) 387 newcpu = decay_cpu(loadfac, newcpu); 388 kg->kg_estcpu = newcpu; 389 } 390 resetpriority(td->td_ksegrp); 391} 392 393/* 394 * We're only looking at 7 bits of the address; everything is 395 * aligned to 4, lots of things are aligned to greater powers 396 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 397 */ 398#define TABLESIZE 128 399static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE]; 400#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 401 402void 403sleepinit(void) 404{ 405 int i; 406 407 sched_quantum = hz/10; 408 hogticks = 2 * sched_quantum; 409 for (i = 0; i < TABLESIZE; i++) 410 TAILQ_INIT(&slpque[i]); 411} 412 413/* 414 * General sleep call. Suspends the current process until a wakeup is 415 * performed on the specified identifier. The process will then be made 416 * runnable with the specified priority. Sleeps at most timo/hz seconds 417 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 418 * before and after sleeping, else signals are not checked. Returns 0 if 419 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 420 * signal needs to be delivered, ERESTART is returned if the current system 421 * call should be restarted if possible, and EINTR is returned if the system 422 * call should be interrupted by the signal (return EINTR). 423 * 424 * The mutex argument is exited before the caller is suspended, and 425 * entered before msleep returns. If priority includes the PDROP 426 * flag the mutex is not entered before returning. 427 */ 428 429int 430msleep(ident, mtx, priority, wmesg, timo) 431 void *ident; 432 struct mtx *mtx; 433 int priority, timo; 434 const char *wmesg; 435{ 436 struct thread *td = curthread; 437 struct proc *p = td->td_proc; 438 int sig, catch = priority & PCATCH; 439 int rval = 0; 440 WITNESS_SAVE_DECL(mtx); 441 442#ifdef KTRACE 443 if (KTRPOINT(td, KTR_CSW)) 444 ktrcsw(1, 0); 445#endif 446 WITNESS_SLEEP(0, &mtx->mtx_object); 447 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 448 ("sleeping without a mutex")); 449 /* 450 * If we are capable of async syscalls and there isn't already 451 * another one ready to return, start a new thread 452 * and queue it as ready to run. Note that there is danger here 453 * because we need to make sure that we don't sleep allocating 454 * the thread (recursion here might be bad). 455 * Hence the TDF_INMSLEEP flag. 456 */ 457 if (p->p_flag & P_KSES) { 458 /* Just don't bother if we are exiting 459 and not the exiting thread. */ 460 if ((p->p_flag & P_WEXIT) && catch && p->p_singlethread != td) 461 return (EINTR); 462 if (td->td_mailbox && (!(td->td_flags & TDF_INMSLEEP))) { 463 /* 464 * If we have no queued work to do, then 465 * upcall to the UTS to see if it has more to do. 466 * We don't need to upcall now, just make it and 467 * queue it. 468 */ 469 mtx_lock_spin(&sched_lock); 470 if (TAILQ_FIRST(&td->td_ksegrp->kg_runq) == NULL) { 471 /* Don't recurse here! */ 472 td->td_flags |= TDF_INMSLEEP; 473 thread_schedule_upcall(td, td->td_kse); 474 td->td_flags &= ~TDF_INMSLEEP; 475 } 476 mtx_unlock_spin(&sched_lock); 477 } 478 } 479 mtx_lock_spin(&sched_lock); 480 if (cold ) { 481 /* 482 * During autoconfiguration, just give interrupts 483 * a chance, then just return. 484 * Don't run any other procs or panic below, 485 * in case this is the idle process and already asleep. 486 */ 487 if (mtx != NULL && priority & PDROP) 488 mtx_unlock(mtx); 489 mtx_unlock_spin(&sched_lock); 490 return (0); 491 } 492 493 DROP_GIANT(); 494 495 if (mtx != NULL) { 496 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 497 WITNESS_SAVE(&mtx->mtx_object, mtx); 498 mtx_unlock(mtx); 499 if (priority & PDROP) 500 mtx = NULL; 501 } 502 503 KASSERT(p != NULL, ("msleep1")); 504 KASSERT(ident != NULL && td->td_state == TDS_RUNNING, ("msleep")); 505 506 td->td_wchan = ident; 507 td->td_wmesg = wmesg; 508 td->td_kse->ke_slptime = 0; /* XXXKSE */ 509 td->td_ksegrp->kg_slptime = 0; 510 td->td_priority = priority & PRIMASK; 511 CTR5(KTR_PROC, "msleep: thread %p (pid %d, %s) on %s (%p)", 512 td, p->p_pid, p->p_comm, wmesg, ident); 513 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], td, td_slpq); 514 if (timo) 515 callout_reset(&td->td_slpcallout, timo, endtsleep, td); 516 /* 517 * We put ourselves on the sleep queue and start our timeout 518 * before calling thread_suspend_check, as we could stop there, and 519 * a wakeup or a SIGCONT (or both) could occur while we were stopped. 520 * without resuming us, thus we must be ready for sleep 521 * when cursig is called. If the wakeup happens while we're 522 * stopped, td->td_wchan will be 0 upon return from cursig. 523 */ 524 if (catch) { 525 CTR3(KTR_PROC, "msleep caught: thread %p (pid %d, %s)", td, 526 p->p_pid, p->p_comm); 527 td->td_flags |= TDF_SINTR; 528 mtx_unlock_spin(&sched_lock); 529 PROC_LOCK(p); 530 sig = cursig(td); 531 if (sig == 0) { 532 if (thread_suspend_check(1)) { 533 sig = SIGSTOP; 534 } 535 } 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 } 546 if (td->td_wchan != NULL) { 547 p->p_stats->p_ru.ru_nvcsw++; 548 td->td_state = TDS_SLP; 549 mi_switch(); 550 } 551 CTR3(KTR_PROC, "msleep resume: thread %p (pid %d, %s)", td, p->p_pid, 552 p->p_comm); 553 KASSERT(td->td_state == TDS_RUNNING, ("running but not TDS_RUNNING")); 554 td->td_flags &= ~TDF_SINTR; 555 if (td->td_flags & TDF_TIMEOUT) { 556 td->td_flags &= ~TDF_TIMEOUT; 557 if (sig == 0) 558 rval = EWOULDBLOCK; 559 } else if (td->td_flags & TDF_TIMOFAIL) { 560 td->td_flags &= ~TDF_TIMOFAIL; 561 } else if (timo && callout_stop(&td->td_slpcallout) == 0) { 562 /* 563 * This isn't supposed to be pretty. If we are here, then 564 * the endtsleep() callout is currently executing on another 565 * CPU and is either spinning on the sched_lock or will be 566 * soon. If we don't synchronize here, there is a chance 567 * that this process may msleep() again before the callout 568 * has a chance to run and the callout may end up waking up 569 * the wrong msleep(). Yuck. 570 */ 571 td->td_flags |= TDF_TIMEOUT; 572 td->td_state = TDS_SLP; 573 p->p_stats->p_ru.ru_nivcsw++; 574 mi_switch(); 575 } 576 mtx_unlock_spin(&sched_lock); 577 578 if (rval == 0 && catch) { 579 PROC_LOCK(p); 580 /* XXX: shouldn't we always be calling cursig() */ 581 if (sig != 0 || (sig = cursig(td))) { 582 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 583 rval = EINTR; 584 else 585 rval = ERESTART; 586 } 587 PROC_UNLOCK(p); 588 } 589#ifdef KTRACE 590 if (KTRPOINT(td, KTR_CSW)) 591 ktrcsw(0, 0); 592#endif 593 PICKUP_GIANT(); 594 if (mtx != NULL) { 595 mtx_lock(mtx); 596 WITNESS_RESTORE(&mtx->mtx_object, mtx); 597 } 598 return (rval); 599} 600 601/* 602 * Implement timeout for msleep() 603 * 604 * If process hasn't been awakened (wchan non-zero), 605 * set timeout flag and undo the sleep. If proc 606 * is stopped, just unsleep so it will remain stopped. 607 * MP-safe, called without the Giant mutex. 608 */ 609static void 610endtsleep(arg) 611 void *arg; 612{ 613 register struct thread *td = arg; 614 615 CTR3(KTR_PROC, "endtsleep: thread %p (pid %d, %s)", td, td->td_proc->p_pid, 616 td->td_proc->p_comm); 617 mtx_lock_spin(&sched_lock); 618 /* 619 * This is the other half of the synchronization with msleep() 620 * described above. If the PS_TIMEOUT flag is set, we lost the 621 * race and just need to put the process back on the runqueue. 622 */ 623 if ((td->td_flags & TDF_TIMEOUT) != 0) { 624 td->td_flags &= ~TDF_TIMEOUT; 625 setrunqueue(td); 626 } else if (td->td_wchan != NULL) { 627 if (td->td_state == TDS_SLP) /* XXXKSE */ 628 setrunnable(td); 629 else 630 unsleep(td); 631 td->td_flags |= TDF_TIMEOUT; 632 } else { 633 td->td_flags |= TDF_TIMOFAIL; 634 } 635 mtx_unlock_spin(&sched_lock); 636} 637 638/* 639 * Abort a thread, as if an interrupt had occured. Only abort 640 * interruptable waits (unfortunatly it isn't only safe to abort others). 641 * This is about identical to cv_abort(). 642 * Think about merging them? 643 * Also, whatever the signal code does... 644 */ 645void 646abortsleep(struct thread *td) 647{ 648 649 mtx_lock_spin(&sched_lock); 650 /* 651 * If the TDF_TIMEOUT flag is set, just leave. A 652 * timeout is scheduled anyhow. 653 */ 654 if ((td->td_flags & (TDF_TIMEOUT | TDF_SINTR)) == TDF_SINTR) { 655 if (td->td_wchan != NULL) { 656 if (td->td_state == TDS_SLP) { /* XXXKSE */ 657 setrunnable(td); 658 } else { 659 /* 660 * Probably in a suspended state.. 661 * um.. dunno XXXKSE 662 */ 663 unsleep(td); 664 } 665 } 666 } 667 mtx_unlock_spin(&sched_lock); 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->td_wchan != NULL) { 679 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_slpq); 680 td->td_wchan = NULL; 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 p = td->td_proc; 703 if (td->td_wchan == ident) { 704 TAILQ_REMOVE(qp, td, td_slpq); 705 td->td_wchan = NULL; 706 if (td->td_state == TDS_SLP) { 707 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 708 CTR3(KTR_PROC, "wakeup: thread %p (pid %d, %s)", 709 td, p->p_pid, p->p_comm); 710 if (td->td_ksegrp->kg_slptime > 1) 711 updatepri(td); 712 td->td_ksegrp->kg_slptime = 0; 713 if (p->p_sflag & PS_INMEM) { 714 setrunqueue(td); 715 maybe_resched(td); 716 } else { 717/* XXXKSE Wrong! */ td->td_state = TDS_RUNQ; 718 p->p_sflag |= PS_SWAPINREQ; 719 wakeup(&proc0); 720 } 721 /* END INLINE EXPANSION */ 722 } 723 goto restart; 724 } 725 } 726 mtx_unlock_spin(&sched_lock); 727} 728 729/* 730 * Make a process sleeping on the specified identifier runnable. 731 * May wake more than one process if a target process is currently 732 * swapped out. 733 */ 734void 735wakeup_one(ident) 736 register void *ident; 737{ 738 register struct slpquehead *qp; 739 register struct thread *td; 740 register struct proc *p; 741 struct thread *ntd; 742 743 mtx_lock_spin(&sched_lock); 744 qp = &slpque[LOOKUP(ident)]; 745restart: 746 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 747 ntd = TAILQ_NEXT(td, td_slpq); 748 p = td->td_proc; 749 if (td->td_wchan == ident) { 750 TAILQ_REMOVE(qp, td, td_slpq); 751 td->td_wchan = NULL; 752 if (td->td_state == TDS_SLP) { 753 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 754 CTR3(KTR_PROC,"wakeup1: thread %p (pid %d, %s)", 755 td, p->p_pid, p->p_comm); 756 if (td->td_ksegrp->kg_slptime > 1) 757 updatepri(td); 758 td->td_ksegrp->kg_slptime = 0; 759 if (p->p_sflag & PS_INMEM) { 760 setrunqueue(td); 761 maybe_resched(td); 762 break; 763 } else { 764/* XXXKSE Wrong */ td->td_state = TDS_RUNQ; 765 p->p_sflag |= PS_SWAPINREQ; 766 wakeup(&proc0); 767 } 768 /* END INLINE EXPANSION */ 769 goto restart; 770 } 771 } 772 } 773 mtx_unlock_spin(&sched_lock); 774} 775 776/* 777 * The machine independent parts of mi_switch(). 778 */ 779void 780mi_switch() 781{ 782 struct bintime new_switchtime; 783 struct thread *td = curthread; /* XXX */ 784 struct proc *p = td->td_proc; /* XXX */ 785 struct kse *ke = td->td_kse; 786#if 0 787 register struct rlimit *rlim; 788#endif 789 u_int sched_nest; 790 791 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 792 KASSERT((ke->ke_state == KES_THREAD), ("mi_switch: kse state?")); 793#ifdef INVARIANTS 794 if (td->td_state != TDS_MTX && 795 td->td_state != TDS_RUNQ && 796 td->td_state != TDS_RUNNING) 797 mtx_assert(&Giant, MA_NOTOWNED); 798#endif 799 800 /* 801 * Compute the amount of time during which the current 802 * process was running, and add that to its total so far. 803 */ 804 binuptime(&new_switchtime); 805 bintime_add(&p->p_runtime, &new_switchtime); 806 bintime_sub(&p->p_runtime, PCPU_PTR(switchtime)); 807 808#ifdef DDB 809 /* 810 * Don't perform context switches from the debugger. 811 */ 812 if (db_active) { 813 mtx_unlock_spin(&sched_lock); 814 db_error("Context switches not allowed in the debugger."); 815 } 816#endif 817 818#if 0 819 /* 820 * Check if the process exceeds its cpu resource allocation. 821 * If over max, kill it. 822 * 823 * XXX drop sched_lock, pickup Giant 824 */ 825 if (p->p_state != PRS_ZOMBIE && 826 p->p_limit->p_cpulimit != RLIM_INFINITY && 827 p->p_runtime > p->p_limit->p_cpulimit) { 828 rlim = &p->p_rlimit[RLIMIT_CPU]; 829 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 830 mtx_unlock_spin(&sched_lock); 831 PROC_LOCK(p); 832 killproc(p, "exceeded maximum CPU limit"); 833 mtx_lock_spin(&sched_lock); 834 PROC_UNLOCK(p); 835 } else { 836 mtx_unlock_spin(&sched_lock); 837 PROC_LOCK(p); 838 psignal(p, SIGXCPU); 839 mtx_lock_spin(&sched_lock); 840 PROC_UNLOCK(p); 841 if (rlim->rlim_cur < rlim->rlim_max) { 842 /* XXX: we should make a private copy */ 843 rlim->rlim_cur += 5; 844 } 845 } 846 } 847#endif 848 849 /* 850 * Pick a new current process and record its start time. 851 */ 852 cnt.v_swtch++; 853 PCPU_SET(switchtime, new_switchtime); 854 CTR3(KTR_PROC, "mi_switch: old thread %p (pid %d, %s)", td, p->p_pid, 855 p->p_comm); 856 sched_nest = sched_lock.mtx_recurse; 857 td->td_lastcpu = ke->ke_oncpu; 858 ke->ke_oncpu = NOCPU; 859 ke->ke_flags &= ~KEF_NEEDRESCHED; 860 /* 861 * At the last moment: if this KSE is not on the run queue, 862 * it needs to be freed correctly and the thread treated accordingly. 863 */ 864 if ((td->td_state == TDS_RUNNING) && 865 ((ke->ke_flags & KEF_IDLEKSE) == 0)) { 866 /* Put us back on the run queue (kse and all). */ 867 setrunqueue(td); 868 } else if ((td->td_flags & TDF_UNBOUND) && 869 (td->td_state != TDS_RUNQ)) { /* in case of old code */ 870 /* 871 * We will not be on the run queue. 872 * Someone else can use the KSE if they need it. 873 */ 874 td->td_kse = NULL; 875 kse_reassign(ke); 876 } 877 cpu_switch(); 878 td->td_kse->ke_oncpu = PCPU_GET(cpuid); 879 sched_lock.mtx_recurse = sched_nest; 880 sched_lock.mtx_lock = (uintptr_t)td; 881 CTR3(KTR_PROC, "mi_switch: new thread %p (pid %d, %s)", td, p->p_pid, 882 p->p_comm); 883 if (PCPU_GET(switchtime.sec) == 0) 884 binuptime(PCPU_PTR(switchtime)); 885 PCPU_SET(switchticks, ticks); 886 887 /* 888 * Call the switchin function while still holding the scheduler lock 889 * (used by the idlezero code and the general page-zeroing code) 890 */ 891 if (td->td_switchin) 892 td->td_switchin(); 893} 894 895/* 896 * Change process state to be runnable, 897 * placing it on the run queue if it is in memory, 898 * and awakening the swapper if it isn't in memory. 899 */ 900void 901setrunnable(struct thread *td) 902{ 903 struct proc *p = td->td_proc; 904 905 mtx_assert(&sched_lock, MA_OWNED); 906 switch (p->p_state) { 907 case PRS_ZOMBIE: 908 panic("setrunnable(1)"); 909 default: 910 break; 911 } 912 switch (td->td_state) { 913 case 0: 914 case TDS_RUNNING: 915 case TDS_IWAIT: 916 default: 917 printf("state is %d", td->td_state); 918 panic("setrunnable(2)"); 919 case TDS_SUSPENDED: 920 thread_unsuspend(p); 921 break; 922 case TDS_SLP: /* e.g. when sending signals */ 923 if (td->td_flags & TDF_CVWAITQ) 924 cv_waitq_remove(td); 925 else 926 unsleep(td); 927 case TDS_UNQUEUED: /* being put back onto the queue */ 928 case TDS_NEW: /* not yet had time to suspend */ 929 case TDS_RUNQ: /* not yet had time to suspend */ 930 break; 931 } 932 if (td->td_ksegrp->kg_slptime > 1) 933 updatepri(td); 934 td->td_ksegrp->kg_slptime = 0; 935 if ((p->p_sflag & PS_INMEM) == 0) { 936 td->td_state = TDS_RUNQ; /* XXXKSE not a good idea */ 937 p->p_sflag |= PS_SWAPINREQ; 938 wakeup(&proc0); 939 } else { 940 if (td->td_state != TDS_RUNQ) 941 setrunqueue(td); /* XXXKSE */ 942 maybe_resched(td); 943 } 944} 945 946/* 947 * Compute the priority of a process when running in user mode. 948 * Arrange to reschedule if the resulting priority is better 949 * than that of the current process. 950 */ 951void 952resetpriority(kg) 953 register struct ksegrp *kg; 954{ 955 register unsigned int newpriority; 956 struct thread *td; 957 958 mtx_lock_spin(&sched_lock); 959 if (kg->kg_pri_class == PRI_TIMESHARE) { 960 newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT + 961 NICE_WEIGHT * (kg->kg_nice - PRIO_MIN); 962 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 963 PRI_MAX_TIMESHARE); 964 kg->kg_user_pri = newpriority; 965 } 966 FOREACH_THREAD_IN_GROUP(kg, td) { 967 maybe_resched(td); /* XXXKSE silly */ 968 } 969 mtx_unlock_spin(&sched_lock); 970} 971 972/* 973 * Compute a tenex style load average of a quantity on 974 * 1, 5 and 15 minute intervals. 975 * XXXKSE Needs complete rewrite when correct info is available. 976 * Completely Bogus.. only works with 1:1 (but compiles ok now :-) 977 */ 978static void 979loadav(void *arg) 980{ 981 int i, nrun; 982 struct loadavg *avg; 983 struct proc *p; 984 struct thread *td; 985 986 avg = &averunnable; 987 sx_slock(&allproc_lock); 988 nrun = 0; 989 FOREACH_PROC_IN_SYSTEM(p) { 990 FOREACH_THREAD_IN_PROC(p, td) { 991 switch (td->td_state) { 992 case TDS_RUNQ: 993 case TDS_RUNNING: 994 if ((p->p_flag & P_NOLOAD) != 0) 995 goto nextproc; 996 nrun++; /* XXXKSE */ 997 default: 998 break; 999 } 1000nextproc: 1001 continue; 1002 } 1003 } 1004 sx_sunlock(&allproc_lock); 1005 for (i = 0; i < 3; i++) 1006 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1007 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1008 1009 /* 1010 * Schedule the next update to occur after 5 seconds, but add a 1011 * random variation to avoid synchronisation with processes that 1012 * run at regular intervals. 1013 */ 1014 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)), 1015 loadav, NULL); 1016} 1017 1018/* ARGSUSED */ 1019static void 1020sched_setup(dummy) 1021 void *dummy; 1022{ 1023 1024 callout_init(&schedcpu_callout, 1); 1025 callout_init(&roundrobin_callout, 0); 1026 callout_init(&loadav_callout, 0); 1027 1028 /* Kick off timeout driven events by calling first time. */ 1029 roundrobin(NULL); 1030 schedcpu(NULL); 1031 loadav(NULL); 1032} 1033 1034/* 1035 * We adjust the priority of the current process. The priority of 1036 * a process gets worse as it accumulates CPU time. The cpu usage 1037 * estimator (p_estcpu) is increased here. resetpriority() will 1038 * compute a different priority each time p_estcpu increases by 1039 * INVERSE_ESTCPU_WEIGHT 1040 * (until MAXPRI is reached). The cpu usage estimator ramps up 1041 * quite quickly when the process is running (linearly), and decays 1042 * away exponentially, at a rate which is proportionally slower when 1043 * the system is busy. The basic principle is that the system will 1044 * 90% forget that the process used a lot of CPU time in 5 * loadav 1045 * seconds. This causes the system to favor processes which haven't 1046 * run much recently, and to round-robin among other processes. 1047 */ 1048void 1049schedclock(td) 1050 struct thread *td; 1051{ 1052 struct kse *ke; 1053 struct ksegrp *kg; 1054 1055 KASSERT((td != NULL), ("schedlock: null thread pointer")); 1056 ke = td->td_kse; 1057 kg = td->td_ksegrp; 1058 ke->ke_cpticks++; 1059 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1); 1060 if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1061 resetpriority(kg); 1062 if (td->td_priority >= PUSER) 1063 td->td_priority = kg->kg_user_pri; 1064 } 1065} 1066 1067/* 1068 * General purpose yield system call 1069 */ 1070int 1071yield(struct thread *td, struct yield_args *uap) 1072{ 1073 struct ksegrp *kg = td->td_ksegrp; 1074 1075 mtx_assert(&Giant, MA_NOTOWNED); 1076 mtx_lock_spin(&sched_lock); 1077 td->td_priority = PRI_MAX_TIMESHARE; 1078 kg->kg_proc->p_stats->p_ru.ru_nvcsw++; 1079 mi_switch(); 1080 mtx_unlock_spin(&sched_lock); 1081 td->td_retval[0] = 0; 1082 1083 return (0); 1084} 1085 1086