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: src/sys/kern/kern_synch.c,v 1.87.2.6 2002/10/13 07:29:53 kbyanc Exp $ 40 * $DragonFly: src/sys/kern/kern_synch.c,v 1.47 2005/06/27 18:37:57 dillon Exp $ 41 */ 42 43 #include "opt_ktrace.h" 44 45 #include <sys/param.h> 46 #include <sys/systm.h> 47 #include <sys/proc.h> 48 #include <sys/kernel.h> 49 #include <sys/signalvar.h> 50 #include <sys/resourcevar.h> 51 #include <sys/vmmeter.h> 52 #include <sys/sysctl.h> 53 #include <sys/thread2.h> 54 #ifdef KTRACE 55 #include <sys/uio.h> 56 #include <sys/ktrace.h> 57 #endif 58 #include <sys/xwait.h> 59 60 #include <machine/cpu.h> 61 #include <machine/ipl.h> 62 #include <machine/smp.h> 63 64 static void sched_setup (void *dummy); 65 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 66 67 int hogticks; 68 int lbolt; 69 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 70 int ncpus; 71 int ncpus2, ncpus2_shift, ncpus2_mask; 72 int safepri; 73 74 static struct callout loadav_callout; 75 static struct callout schedcpu_callout; 76 77 struct loadavg averunnable = 78 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 79 /* 80 * Constants for averages over 1, 5, and 15 minutes 81 * when sampling at 5 second intervals. 82 */ 83 static fixpt_t cexp[3] = { 84 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 85 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 86 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 87 }; 88 89 static void endtsleep (void *); 90 static void loadav (void *arg); 91 static void schedcpu (void *arg); 92 93 /* 94 * Adjust the scheduler quantum. The quantum is specified in microseconds. 95 * Note that 'tick' is in microseconds per tick. 96 */ 97 static int 98 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 99 { 100 int error, new_val; 101 102 new_val = sched_quantum * tick; 103 error = sysctl_handle_int(oidp, &new_val, 0, req); 104 if (error != 0 || req->newptr == NULL) 105 return (error); 106 if (new_val < tick) 107 return (EINVAL); 108 sched_quantum = new_val / tick; 109 hogticks = 2 * sched_quantum; 110 return (0); 111 } 112 113 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 114 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 115 116 /* 117 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 118 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 119 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 120 * 121 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 122 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 123 * 124 * If you don't want to bother with the faster/more-accurate formula, you 125 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 126 * (more general) method of calculating the %age of CPU used by a process. 127 * 128 * decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing 129 */ 130 #define CCPU_SHIFT 11 131 132 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 133 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 134 135 /* 136 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 137 */ 138 static int fscale __unused = FSCALE; 139 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 140 141 /* 142 * Recompute process priorities, once a second. 143 * 144 * Since the userland schedulers are typically event oriented, if the 145 * estcpu calculation at wakeup() time is not sufficient to make a 146 * process runnable relative to other processes in the system we have 147 * a 1-second recalc to help out. 148 * 149 * This code also allows us to store sysclock_t data in the process structure 150 * without fear of an overrun, since sysclock_t are guarenteed to hold 151 * several seconds worth of count. 152 */ 153 /* ARGSUSED */ 154 static void 155 schedcpu(void *arg) 156 { 157 struct proc *p; 158 159 FOREACH_PROC_IN_SYSTEM(p) { 160 /* 161 * Increment time in/out of memory and sleep time 162 * (if sleeping). We ignore overflow; with 16-bit int's 163 * (remember them?) overflow takes 45 days. 164 */ 165 crit_enter(); 166 p->p_swtime++; 167 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 168 p->p_slptime++; 169 170 /* 171 * Only recalculate processes that are active or have slept 172 * less then 2 seconds. The schedulers understand this. 173 */ 174 if (p->p_slptime <= 1) { 175 p->p_usched->recalculate(p); 176 } else { 177 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 178 } 179 crit_exit(); 180 } 181 wakeup((caddr_t)&lbolt); 182 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 183 } 184 185 /* 186 * This is only used by ps. Generate a cpu percentage use over 187 * a period of one second. 188 */ 189 void 190 updatepcpu(struct proc *p, int cpticks, int ttlticks) 191 { 192 fixpt_t acc; 193 int remticks; 194 195 acc = (cpticks << FSHIFT) / ttlticks; 196 if (ttlticks >= ESTCPUFREQ) { 197 p->p_pctcpu = acc; 198 } else { 199 remticks = ESTCPUFREQ - ttlticks; 200 p->p_pctcpu = (acc * ttlticks + p->p_pctcpu * remticks) / 201 ESTCPUFREQ; 202 } 203 } 204 205 206 /* 207 * We're only looking at 7 bits of the address; everything is 208 * aligned to 4, lots of things are aligned to greater powers 209 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 210 */ 211 #define TABLESIZE 128 212 static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE]; 213 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 214 215 /* 216 * General scheduler initialization. We force a reschedule 25 times 217 * a second by default. 218 */ 219 void 220 sleepinit(void) 221 { 222 int i; 223 224 sched_quantum = (hz + 24) / 25; 225 hogticks = 2 * sched_quantum; 226 for (i = 0; i < TABLESIZE; i++) 227 TAILQ_INIT(&slpque[i]); 228 } 229 230 /* 231 * General sleep call. Suspends the current process until a wakeup is 232 * performed on the specified identifier. The process will then be made 233 * runnable with the specified priority. Sleeps at most timo/hz seconds 234 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 235 * before and after sleeping, else signals are not checked. Returns 0 if 236 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 237 * signal needs to be delivered, ERESTART is returned if the current system 238 * call should be restarted if possible, and EINTR is returned if the system 239 * call should be interrupted by the signal (return EINTR). 240 * 241 * Note that if we are a process, we release_curproc() before messing with 242 * the LWKT scheduler. 243 * 244 * During autoconfiguration or after a panic, a sleep will simply 245 * lower the priority briefly to allow interrupts, then return. 246 */ 247 int 248 tsleep(void *ident, int flags, const char *wmesg, int timo) 249 { 250 struct thread *td = curthread; 251 struct proc *p = td->td_proc; /* may be NULL */ 252 int sig = 0, catch = flags & PCATCH; 253 int id = LOOKUP(ident); 254 int oldpri; 255 struct callout thandle; 256 257 /* 258 * NOTE: removed KTRPOINT, it could cause races due to blocking 259 * even in stable. Just scrap it for now. 260 */ 261 if (cold || panicstr) { 262 /* 263 * After a panic, or during autoconfiguration, 264 * just give interrupts a chance, then just return; 265 * don't run any other procs or panic below, 266 * in case this is the idle process and already asleep. 267 */ 268 splz(); 269 oldpri = td->td_pri & TDPRI_MASK; 270 lwkt_setpri_self(safepri); 271 lwkt_switch(); 272 lwkt_setpri_self(oldpri); 273 return (0); 274 } 275 KKASSERT(td != &mycpu->gd_idlethread); /* you must be kidding! */ 276 crit_enter_quick(td); 277 KASSERT(ident != NULL, ("tsleep: no ident")); 278 KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d", 279 ident, wmesg, p->p_stat)); 280 281 td->td_wchan = ident; 282 td->td_wmesg = wmesg; 283 td->td_wdomain = flags & PDOMAIN_MASK; 284 if (p) { 285 if (flags & PNORESCHED) 286 td->td_flags |= TDF_NORESCHED; 287 p->p_usched->release_curproc(p); 288 p->p_slptime = 0; 289 } 290 lwkt_deschedule_self(td); 291 TAILQ_INSERT_TAIL(&slpque[id], td, td_threadq); 292 if (timo) { 293 callout_init(&thandle); 294 callout_reset(&thandle, timo, endtsleep, td); 295 } 296 /* 297 * We put ourselves on the sleep queue and start our timeout 298 * before calling CURSIG, as we could stop there, and a wakeup 299 * or a SIGCONT (or both) could occur while we were stopped. 300 * A SIGCONT would cause us to be marked as SSLEEP 301 * without resuming us, thus we must be ready for sleep 302 * when CURSIG is called. If the wakeup happens while we're 303 * stopped, td->td_wchan will be 0 upon return from CURSIG. 304 */ 305 if (p) { 306 if (catch) { 307 p->p_flag |= P_SINTR; 308 if ((sig = CURSIG(p))) { 309 if (td->td_wchan) { 310 unsleep(td); 311 lwkt_schedule_self(td); 312 } 313 p->p_stat = SRUN; 314 goto resume; 315 } 316 if (td->td_wchan == NULL) { 317 catch = 0; 318 goto resume; 319 } 320 } else { 321 sig = 0; 322 } 323 324 /* 325 * If we are not the current process we have to remove ourself 326 * from the run queue. 327 */ 328 KASSERT(p->p_stat == SRUN, ("PSTAT NOT SRUN %d %d", p->p_pid, p->p_stat)); 329 /* 330 * If this is the current 'user' process schedule another one. 331 */ 332 clrrunnable(p, SSLEEP); 333 p->p_stats->p_ru.ru_nvcsw++; 334 mi_switch(p); 335 KASSERT(p->p_stat == SRUN, ("tsleep: stat not srun")); 336 } else { 337 lwkt_switch(); 338 } 339 resume: 340 if (p) 341 p->p_flag &= ~P_SINTR; 342 crit_exit_quick(td); 343 td->td_flags &= ~TDF_NORESCHED; 344 if (td->td_flags & TDF_TIMEOUT) { 345 td->td_flags &= ~TDF_TIMEOUT; 346 if (sig == 0) 347 return (EWOULDBLOCK); 348 } else if (timo) { 349 callout_stop(&thandle); 350 } else if (td->td_wmesg) { 351 /* 352 * This can happen if a thread is woken up directly. Clear 353 * wmesg to avoid debugging confusion. 354 */ 355 td->td_wmesg = NULL; 356 } 357 /* inline of iscaught() */ 358 if (p) { 359 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 360 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 361 return (EINTR); 362 return (ERESTART); 363 } 364 } 365 return (0); 366 } 367 368 /* 369 * Implement the timeout for tsleep. We interlock against 370 * wchan when setting TDF_TIMEOUT. For processes we remove 371 * the sleep if the process is stopped rather then sleeping, 372 * so it remains stopped. 373 */ 374 static void 375 endtsleep(void *arg) 376 { 377 thread_t td = arg; 378 struct proc *p; 379 380 crit_enter(); 381 if (td->td_wchan) { 382 td->td_flags |= TDF_TIMEOUT; 383 if ((p = td->td_proc) != NULL) { 384 if (p->p_stat == SSLEEP) 385 setrunnable(p); 386 else 387 unsleep(td); 388 } else { 389 unsleep(td); 390 lwkt_schedule(td); 391 } 392 } 393 crit_exit(); 394 } 395 396 /* 397 * Remove a process from its wait queue 398 */ 399 void 400 unsleep(struct thread *td) 401 { 402 crit_enter(); 403 if (td->td_wchan) { 404 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_threadq); 405 td->td_wchan = NULL; 406 } 407 crit_exit(); 408 } 409 410 /* 411 * Make all processes sleeping on the specified identifier runnable. 412 */ 413 static void 414 _wakeup(void *ident, int domain, int count) 415 { 416 struct slpquehead *qp; 417 struct thread *td; 418 struct thread *ntd; 419 struct proc *p; 420 int id = LOOKUP(ident); 421 422 crit_enter(); 423 qp = &slpque[id]; 424 restart: 425 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 426 ntd = TAILQ_NEXT(td, td_threadq); 427 if (td->td_wchan == ident && td->td_wdomain == domain) { 428 TAILQ_REMOVE(qp, td, td_threadq); 429 td->td_wchan = NULL; 430 if ((p = td->td_proc) != NULL && p->p_stat == SSLEEP) { 431 p->p_stat = SRUN; 432 if (p->p_flag & P_INMEM) { 433 /* 434 * LWKT scheduled now, there is no 435 * userland runq interaction until 436 * the thread tries to return to user 437 * mode. We do NOT call setrunqueue(). 438 */ 439 lwkt_schedule(td); 440 } else { 441 p->p_flag |= P_SWAPINREQ; 442 wakeup((caddr_t)&proc0); 443 } 444 /* END INLINE EXPANSION */ 445 } else if (p == NULL) { 446 lwkt_schedule(td); 447 } 448 if (--count == 0) 449 break; 450 goto restart; 451 } 452 } 453 crit_exit(); 454 } 455 456 void 457 wakeup(void *ident) 458 { 459 _wakeup(ident, 0, 0); 460 } 461 462 void 463 wakeup_one(void *ident) 464 { 465 _wakeup(ident, 0, 1); 466 } 467 468 void 469 wakeup_domain(void *ident, int domain) 470 { 471 _wakeup(ident, domain, 0); 472 } 473 474 void 475 wakeup_domain_one(void *ident, int domain) 476 { 477 _wakeup(ident, domain, 1); 478 } 479 480 /* 481 * The machine independent parts of mi_switch(). 482 * 483 * 'p' must be the current process. 484 */ 485 void 486 mi_switch(struct proc *p) 487 { 488 thread_t td = p->p_thread; 489 struct rlimit *rlim; 490 u_int64_t ttime; 491 492 KKASSERT(td == mycpu->gd_curthread); 493 494 crit_enter_quick(td); 495 496 /* 497 * Check if the process exceeds its cpu resource allocation. 498 * If over max, kill it. Time spent in interrupts is not 499 * included. YYY 64 bit match is expensive. Ick. 500 */ 501 ttime = td->td_sticks + td->td_uticks; 502 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 503 ttime > p->p_limit->p_cpulimit) { 504 rlim = &p->p_rlimit[RLIMIT_CPU]; 505 if (ttime / (rlim_t)1000000 >= rlim->rlim_max) { 506 killproc(p, "exceeded maximum CPU limit"); 507 } else { 508 psignal(p, SIGXCPU); 509 if (rlim->rlim_cur < rlim->rlim_max) { 510 /* XXX: we should make a private copy */ 511 rlim->rlim_cur += 5; 512 } 513 } 514 } 515 516 /* 517 * If we are in a SSTOPped state we deschedule ourselves. 518 * YYY this needs to be cleaned up, remember that LWKTs stay on 519 * their run queue which works differently then the user scheduler 520 * which removes the process from the runq when it runs it. 521 */ 522 mycpu->gd_cnt.v_swtch++; 523 if (p->p_stat == SSTOP) 524 lwkt_deschedule_self(td); 525 lwkt_switch(); 526 crit_exit_quick(td); 527 } 528 529 /* 530 * Change process state to be runnable, 531 * placing it on the run queue if it is in memory, 532 * and awakening the swapper if it isn't in memory. 533 */ 534 void 535 setrunnable(struct proc *p) 536 { 537 crit_enter(); 538 539 switch (p->p_stat) { 540 case 0: 541 case SRUN: 542 case SZOMB: 543 default: 544 panic("setrunnable"); 545 case SSTOP: 546 case SSLEEP: 547 unsleep(p->p_thread); /* e.g. when sending signals */ 548 break; 549 550 case SIDL: 551 break; 552 } 553 p->p_stat = SRUN; 554 555 /* 556 * The process is controlled by LWKT at this point, we do not mess 557 * around with the userland scheduler until the thread tries to 558 * return to user mode. We do not clear p_slptime or call 559 * setrunqueue(). 560 */ 561 if (p->p_flag & P_INMEM) { 562 lwkt_schedule(p->p_thread); 563 } else { 564 p->p_flag |= P_SWAPINREQ; 565 wakeup((caddr_t)&proc0); 566 } 567 crit_exit(); 568 } 569 570 /* 571 * Yield / synchronous reschedule. This is a bit tricky because the trap 572 * code might have set a lazy release on the switch function. Setting 573 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call 574 * switch, and that we are given a greater chance of affinity with our 575 * current cpu. 576 * 577 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt 578 * run queue. lwkt_switch() will also execute any assigned passive release 579 * (which usually calls release_curproc()), allowing a same/higher priority 580 * process to be designated as the current process. 581 * 582 * While it is possible for a lower priority process to be designated, 583 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely 584 * round-robin back to us and we will be able to re-acquire the current 585 * process designation. 586 */ 587 void 588 uio_yield(void) 589 { 590 struct thread *td = curthread; 591 struct proc *p = td->td_proc; 592 593 lwkt_setpri_self(td->td_pri & TDPRI_MASK); 594 if (p) { 595 p->p_flag |= P_PASSIVE_ACQ; 596 lwkt_switch(); 597 p->p_flag &= ~P_PASSIVE_ACQ; 598 } else { 599 lwkt_switch(); 600 } 601 } 602 603 /* 604 * Change the process state to NOT be runnable, removing it from the run 605 * queue. 606 */ 607 void 608 clrrunnable(struct proc *p, int stat) 609 { 610 crit_enter_quick(p->p_thread); 611 if (p->p_stat == SRUN && (p->p_flag & P_ONRUNQ)) 612 p->p_usched->remrunqueue(p); 613 p->p_stat = stat; 614 crit_exit_quick(p->p_thread); 615 } 616 617 /* 618 * Compute a tenex style load average of a quantity on 619 * 1, 5 and 15 minute intervals. 620 */ 621 static void 622 loadav(void *arg) 623 { 624 int i, nrun; 625 struct loadavg *avg; 626 struct proc *p; 627 thread_t td; 628 629 avg = &averunnable; 630 nrun = 0; 631 FOREACH_PROC_IN_SYSTEM(p) { 632 switch (p->p_stat) { 633 case SRUN: 634 if ((td = p->p_thread) == NULL) 635 break; 636 if (td->td_flags & TDF_BLOCKED) 637 break; 638 /* fall through */ 639 case SIDL: 640 nrun++; 641 break; 642 default: 643 break; 644 } 645 } 646 for (i = 0; i < 3; i++) 647 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 648 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 649 650 /* 651 * Schedule the next update to occur after 5 seconds, but add a 652 * random variation to avoid synchronisation with processes that 653 * run at regular intervals. 654 */ 655 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)), 656 loadav, NULL); 657 } 658 659 /* ARGSUSED */ 660 static void 661 sched_setup(void *dummy) 662 { 663 callout_init(&loadav_callout); 664 callout_init(&schedcpu_callout); 665 666 /* Kick off timeout driven events by calling first time. */ 667 schedcpu(NULL); 668 loadav(NULL); 669 } 670 671