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 1.41 2005/01/14 02:20:22 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 73 static struct callout loadav_callout; 74 static struct callout roundrobin_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 roundrobin (void *arg); 92 static void schedcpu (void *arg); 93 static void updatepri (struct proc *p); 94 static void crit_panicints(void); 95 96 static int 97 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 98 { 99 int error, new_val; 100 101 new_val = sched_quantum * tick; 102 error = sysctl_handle_int(oidp, &new_val, 0, req); 103 if (error != 0 || req->newptr == NULL) 104 return (error); 105 if (new_val < tick) 106 return (EINVAL); 107 sched_quantum = new_val / tick; 108 hogticks = 2 * sched_quantum; 109 return (0); 110 } 111 112 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 113 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 114 115 int 116 roundrobin_interval(void) 117 { 118 return (sched_quantum); 119 } 120 121 /* 122 * Force switch among equal priority processes every 100ms. 123 * 124 * WARNING! The MP lock is not held on ipi message remotes. 125 */ 126 #ifdef SMP 127 128 static void 129 roundrobin_remote(void *arg) 130 { 131 struct proc *p = lwkt_preempted_proc(); 132 if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type)) 133 need_user_resched(); 134 } 135 136 #endif 137 138 static void 139 roundrobin(void *arg) 140 { 141 struct proc *p = lwkt_preempted_proc(); 142 if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type)) 143 need_user_resched(); 144 #ifdef SMP 145 lwkt_send_ipiq_mask(mycpu->gd_other_cpus, roundrobin_remote, NULL); 146 #endif 147 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 148 } 149 150 #ifdef SMP 151 152 void 153 resched_cpus(u_int32_t mask) 154 { 155 lwkt_send_ipiq_mask(mask, roundrobin_remote, NULL); 156 } 157 158 #endif 159 160 /* 161 * The load average is scaled by FSCALE (2048 typ). The estimated cpu is 162 * incremented at a rate of ESTCPUVFREQ per second (40hz typ), but this is 163 * divided up across all cpu bound processes running in the system so an 164 * individual process will get less under load. ESTCPULIM typicaly caps 165 * out at ESTCPUMAX (around 376, or 11 nice levels). 166 * 167 * Generally speaking the decay equation needs to break-even on growth 168 * at the limit at all load levels >= 1.0, so if the estimated cpu for 169 * a process increases by (ESTVCPUFREQ / load) per second, then the decay 170 * should reach this value when estcpu reaches ESTCPUMAX. That calculation 171 * is: 172 * 173 * ESTCPUMAX * decay = ESTCPUVFREQ / load 174 * decay = ESTCPUVFREQ / (load * ESTCPUMAX) 175 * decay = estcpu * 0.053 / load 176 * 177 * If the load is less then 1.0 we assume a load of 1.0. 178 */ 179 180 #define cload(loadav) ((loadav) < FSCALE ? FSCALE : (loadav)) 181 #define decay_cpu(loadav,estcpu) \ 182 ((estcpu) * (FSCALE * ESTCPUVFREQ / ESTCPUMAX) / cload(loadav)) 183 184 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 185 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 186 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 187 188 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 189 static int fscale __unused = FSCALE; 190 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 191 192 /* 193 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 194 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 195 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 196 * 197 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 198 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 199 * 200 * If you don't want to bother with the faster/more-accurate formula, you 201 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 202 * (more general) method of calculating the %age of CPU used by a process. 203 */ 204 #define CCPU_SHIFT 11 205 206 /* 207 * Recompute process priorities, once a second. 208 */ 209 /* ARGSUSED */ 210 static void 211 schedcpu(void *arg) 212 { 213 fixpt_t loadfac = averunnable.ldavg[0]; 214 struct proc *p; 215 int s; 216 unsigned int ndecay; 217 218 FOREACH_PROC_IN_SYSTEM(p) { 219 /* 220 * Increment time in/out of memory and sleep time 221 * (if sleeping). We ignore overflow; with 16-bit int's 222 * (remember them?) overflow takes 45 days. 223 */ 224 p->p_swtime++; 225 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 226 p->p_slptime++; 227 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 228 229 /* 230 * If the process has slept the entire second, 231 * stop recalculating its priority until it wakes up. 232 * 233 * Note that interactive calculations do not occur for 234 * long sleeps (because that isn't necessarily indicative 235 * of an interactive process). 236 */ 237 if (p->p_slptime > 1) 238 continue; 239 /* prevent state changes and protect run queue */ 240 s = splhigh(); 241 /* 242 * p_cpticks runs at ESTCPUFREQ but must be divided by the 243 * load average for par-100% use. Higher p_interactive 244 * values mean less interactive, lower values mean more 245 * interactive. 246 */ 247 if ((((fixpt_t)p->p_cpticks * cload(loadfac)) >> FSHIFT) > 248 ESTCPUFREQ / 4) { 249 if (p->p_interactive < 127) 250 ++p->p_interactive; 251 } else { 252 if (p->p_interactive > -127) 253 --p->p_interactive; 254 } 255 /* 256 * p_pctcpu is only for ps. 257 */ 258 #if (FSHIFT >= CCPU_SHIFT) 259 p->p_pctcpu += (ESTCPUFREQ == 100)? 260 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 261 100 * (((fixpt_t) p->p_cpticks) 262 << (FSHIFT - CCPU_SHIFT)) / ESTCPUFREQ; 263 #else 264 p->p_pctcpu += ((FSCALE - ccpu) * 265 (p->p_cpticks * FSCALE / ESTCPUFREQ)) >> FSHIFT; 266 #endif 267 p->p_cpticks = 0; 268 ndecay = decay_cpu(loadfac, p->p_estcpu); 269 if (p->p_estcpu > ndecay) 270 p->p_estcpu -= ndecay; 271 else 272 p->p_estcpu = 0; 273 resetpriority(p); 274 splx(s); 275 } 276 wakeup((caddr_t)&lbolt); 277 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 278 } 279 280 /* 281 * Recalculate the priority of a process after it has slept for a while. 282 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 283 * least six times the loadfactor will decay p_estcpu to zero. 284 */ 285 static void 286 updatepri(struct proc *p) 287 { 288 unsigned int ndecay; 289 290 ndecay = decay_cpu(averunnable.ldavg[0], p->p_estcpu) * p->p_slptime; 291 if (p->p_estcpu > ndecay) 292 p->p_estcpu -= ndecay; 293 else 294 p->p_estcpu = 0; 295 resetpriority(p); 296 } 297 298 /* 299 * We're only looking at 7 bits of the address; everything is 300 * aligned to 4, lots of things are aligned to greater powers 301 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 302 */ 303 #define TABLESIZE 128 304 static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE]; 305 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 306 307 /* 308 * During autoconfiguration or after a panic, a sleep will simply 309 * lower the priority briefly to allow interrupts, then return. 310 * The priority to be used (safepri) is machine-dependent, thus this 311 * value is initialized and maintained in the machine-dependent layers. 312 * This priority will typically be 0, or the lowest priority 313 * that is safe for use on the interrupt stack; it can be made 314 * higher to block network software interrupts after panics. 315 */ 316 int safepri; 317 318 void 319 sleepinit(void) 320 { 321 int i; 322 323 sched_quantum = hz/10; 324 hogticks = 2 * sched_quantum; 325 for (i = 0; i < TABLESIZE; i++) 326 TAILQ_INIT(&slpque[i]); 327 } 328 329 /* 330 * General sleep call. Suspends the current process until a wakeup is 331 * performed on the specified identifier. The process will then be made 332 * runnable with the specified priority. Sleeps at most timo/hz seconds 333 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 334 * before and after sleeping, else signals are not checked. Returns 0 if 335 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 336 * signal needs to be delivered, ERESTART is returned if the current system 337 * call should be restarted if possible, and EINTR is returned if the system 338 * call should be interrupted by the signal (return EINTR). 339 * 340 * Note that if we are a process, we release_curproc() before messing with 341 * the LWKT scheduler. 342 */ 343 int 344 tsleep(void *ident, int flags, const char *wmesg, int timo) 345 { 346 struct thread *td = curthread; 347 struct proc *p = td->td_proc; /* may be NULL */ 348 int sig = 0, catch = flags & PCATCH; 349 int id = LOOKUP(ident); 350 struct callout thandle; 351 352 /* 353 * NOTE: removed KTRPOINT, it could cause races due to blocking 354 * even in stable. Just scrap it for now. 355 */ 356 if (cold || panicstr) { 357 /* 358 * After a panic, or during autoconfiguration, 359 * just give interrupts a chance, then just return; 360 * don't run any other procs or panic below, 361 * in case this is the idle process and already asleep. 362 */ 363 crit_panicints(); 364 return (0); 365 } 366 KKASSERT(td != &mycpu->gd_idlethread); /* you must be kidding! */ 367 crit_enter_quick(td); 368 KASSERT(ident != NULL, ("tsleep: no ident")); 369 KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d", 370 ident, wmesg, p->p_stat)); 371 372 td->td_wchan = ident; 373 td->td_wmesg = wmesg; 374 td->td_wdomain = flags & PDOMAIN_MASK; 375 if (p) { 376 if (flags & PNORESCHED) 377 td->td_flags |= TDF_NORESCHED; 378 release_curproc(p); 379 p->p_slptime = 0; 380 } 381 lwkt_deschedule_self(td); 382 TAILQ_INSERT_TAIL(&slpque[id], td, td_threadq); 383 if (timo) { 384 callout_init(&thandle); 385 callout_reset(&thandle, timo, endtsleep, td); 386 } 387 /* 388 * We put ourselves on the sleep queue and start our timeout 389 * before calling CURSIG, as we could stop there, and a wakeup 390 * or a SIGCONT (or both) could occur while we were stopped. 391 * A SIGCONT would cause us to be marked as SSLEEP 392 * without resuming us, thus we must be ready for sleep 393 * when CURSIG is called. If the wakeup happens while we're 394 * stopped, td->td_wchan will be 0 upon return from CURSIG. 395 */ 396 if (p) { 397 if (catch) { 398 p->p_flag |= P_SINTR; 399 if ((sig = CURSIG(p))) { 400 if (td->td_wchan) { 401 unsleep(td); 402 lwkt_schedule_self(td); 403 } 404 p->p_stat = SRUN; 405 goto resume; 406 } 407 if (td->td_wchan == NULL) { 408 catch = 0; 409 goto resume; 410 } 411 } else { 412 sig = 0; 413 } 414 415 /* 416 * If we are not the current process we have to remove ourself 417 * from the run queue. 418 */ 419 KASSERT(p->p_stat == SRUN, ("PSTAT NOT SRUN %d %d", p->p_pid, p->p_stat)); 420 /* 421 * If this is the current 'user' process schedule another one. 422 */ 423 clrrunnable(p, SSLEEP); 424 p->p_stats->p_ru.ru_nvcsw++; 425 mi_switch(p); 426 KASSERT(p->p_stat == SRUN, ("tsleep: stat not srun")); 427 } else { 428 lwkt_switch(); 429 } 430 resume: 431 if (p) 432 p->p_flag &= ~P_SINTR; 433 crit_exit_quick(td); 434 td->td_flags &= ~TDF_NORESCHED; 435 if (td->td_flags & TDF_TIMEOUT) { 436 td->td_flags &= ~TDF_TIMEOUT; 437 if (sig == 0) 438 return (EWOULDBLOCK); 439 } else if (timo) { 440 callout_stop(&thandle); 441 } else if (td->td_wmesg) { 442 /* 443 * This can happen if a thread is woken up directly. Clear 444 * wmesg to avoid debugging confusion. 445 */ 446 td->td_wmesg = NULL; 447 } 448 /* inline of iscaught() */ 449 if (p) { 450 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 451 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 452 return (EINTR); 453 return (ERESTART); 454 } 455 } 456 return (0); 457 } 458 459 /* 460 * Implement the timeout for tsleep. We interlock against 461 * wchan when setting TDF_TIMEOUT. For processes we remove 462 * the sleep if the process is stopped rather then sleeping, 463 * so it remains stopped. 464 */ 465 static void 466 endtsleep(void *arg) 467 { 468 thread_t td = arg; 469 struct proc *p; 470 471 crit_enter(); 472 if (td->td_wchan) { 473 td->td_flags |= TDF_TIMEOUT; 474 if ((p = td->td_proc) != NULL) { 475 if (p->p_stat == SSLEEP) 476 setrunnable(p); 477 else 478 unsleep(td); 479 } else { 480 unsleep(td); 481 lwkt_schedule(td); 482 } 483 } 484 crit_exit(); 485 } 486 487 /* 488 * Remove a process from its wait queue 489 */ 490 void 491 unsleep(struct thread *td) 492 { 493 crit_enter(); 494 if (td->td_wchan) { 495 #if 0 496 if (p->p_flag & P_XSLEEP) { 497 struct xwait *w = p->p_wchan; 498 TAILQ_REMOVE(&w->waitq, p, p_procq); 499 p->p_flag &= ~P_XSLEEP; 500 } else 501 #endif 502 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_threadq); 503 td->td_wchan = NULL; 504 } 505 crit_exit(); 506 } 507 508 #if 0 509 /* 510 * Make all processes sleeping on the explicit lock structure runnable. 511 */ 512 void 513 xwakeup(struct xwait *w) 514 { 515 struct proc *p; 516 517 crit_enter(); 518 ++w->gen; 519 while ((p = TAILQ_FIRST(&w->waitq)) != NULL) { 520 TAILQ_REMOVE(&w->waitq, p, p_procq); 521 KASSERT(p->p_wchan == w && (p->p_flag & P_XSLEEP), 522 ("xwakeup: wchan mismatch for %p (%p/%p) %08x", p, p->p_wchan, w, p->p_flag & P_XSLEEP)); 523 p->p_wchan = NULL; 524 p->p_flag &= ~P_XSLEEP; 525 if (p->p_stat == SSLEEP) { 526 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 527 if (p->p_slptime > 1) 528 updatepri(p); 529 p->p_slptime = 0; 530 p->p_stat = SRUN; 531 if (p->p_flag & P_INMEM) { 532 lwkt_schedule(td); 533 } else { 534 p->p_flag |= P_SWAPINREQ; 535 wakeup((caddr_t)&proc0); 536 } 537 } 538 } 539 crit_exit(); 540 } 541 #endif 542 543 /* 544 * Make all processes sleeping on the specified identifier runnable. 545 */ 546 static void 547 _wakeup(void *ident, int domain, int count) 548 { 549 struct slpquehead *qp; 550 struct thread *td; 551 struct thread *ntd; 552 struct proc *p; 553 int id = LOOKUP(ident); 554 555 crit_enter(); 556 qp = &slpque[id]; 557 restart: 558 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 559 ntd = TAILQ_NEXT(td, td_threadq); 560 if (td->td_wchan == ident && td->td_wdomain == domain) { 561 TAILQ_REMOVE(qp, td, td_threadq); 562 td->td_wchan = NULL; 563 if ((p = td->td_proc) != NULL && p->p_stat == SSLEEP) { 564 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 565 if (p->p_slptime > 1) 566 updatepri(p); 567 p->p_slptime = 0; 568 p->p_stat = SRUN; 569 if (p->p_flag & P_INMEM) { 570 /* 571 * LWKT scheduled now, there is no 572 * userland runq interaction until 573 * the thread tries to return to user 574 * mode. 575 * 576 * setrunqueue(p); 577 */ 578 lwkt_schedule(td); 579 } else { 580 p->p_flag |= P_SWAPINREQ; 581 wakeup((caddr_t)&proc0); 582 } 583 /* END INLINE EXPANSION */ 584 } else if (p == NULL) { 585 lwkt_schedule(td); 586 } 587 if (--count == 0) 588 break; 589 goto restart; 590 } 591 } 592 crit_exit(); 593 } 594 595 void 596 wakeup(void *ident) 597 { 598 _wakeup(ident, 0, 0); 599 } 600 601 void 602 wakeup_one(void *ident) 603 { 604 _wakeup(ident, 0, 1); 605 } 606 607 void 608 wakeup_domain(void *ident, int domain) 609 { 610 _wakeup(ident, domain, 0); 611 } 612 613 void 614 wakeup_domain_one(void *ident, int domain) 615 { 616 _wakeup(ident, domain, 1); 617 } 618 619 /* 620 * The machine independent parts of mi_switch(). 621 * 622 * 'p' must be the current process. 623 */ 624 void 625 mi_switch(struct proc *p) 626 { 627 thread_t td = p->p_thread; 628 struct rlimit *rlim; 629 u_int64_t ttime; 630 631 KKASSERT(td == mycpu->gd_curthread); 632 633 crit_enter_quick(td); 634 635 /* 636 * Check if the process exceeds its cpu resource allocation. 637 * If over max, kill it. Time spent in interrupts is not 638 * included. YYY 64 bit match is expensive. Ick. 639 */ 640 ttime = td->td_sticks + td->td_uticks; 641 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 642 ttime > p->p_limit->p_cpulimit) { 643 rlim = &p->p_rlimit[RLIMIT_CPU]; 644 if (ttime / (rlim_t)1000000 >= rlim->rlim_max) { 645 killproc(p, "exceeded maximum CPU limit"); 646 } else { 647 psignal(p, SIGXCPU); 648 if (rlim->rlim_cur < rlim->rlim_max) { 649 /* XXX: we should make a private copy */ 650 rlim->rlim_cur += 5; 651 } 652 } 653 } 654 655 /* 656 * If we are in a SSTOPped state we deschedule ourselves. 657 * YYY this needs to be cleaned up, remember that LWKTs stay on 658 * their run queue which works differently then the user scheduler 659 * which removes the process from the runq when it runs it. 660 */ 661 mycpu->gd_cnt.v_swtch++; 662 if (p->p_stat == SSTOP) 663 lwkt_deschedule_self(td); 664 lwkt_switch(); 665 crit_exit_quick(td); 666 } 667 668 /* 669 * Change process state to be runnable, 670 * placing it on the run queue if it is in memory, 671 * and awakening the swapper if it isn't in memory. 672 */ 673 void 674 setrunnable(struct proc *p) 675 { 676 int s; 677 678 s = splhigh(); 679 switch (p->p_stat) { 680 case 0: 681 case SRUN: 682 case SZOMB: 683 default: 684 panic("setrunnable"); 685 case SSTOP: 686 case SSLEEP: 687 unsleep(p->p_thread); /* e.g. when sending signals */ 688 break; 689 690 case SIDL: 691 break; 692 } 693 p->p_stat = SRUN; 694 695 /* 696 * The process is controlled by LWKT at this point, we do not mess 697 * around with the userland scheduler until the thread tries to 698 * return to user mode. 699 */ 700 #if 0 701 if (p->p_flag & P_INMEM) 702 setrunqueue(p); 703 #endif 704 if (p->p_flag & P_INMEM) 705 lwkt_schedule(p->p_thread); 706 splx(s); 707 if (p->p_slptime > 1) 708 updatepri(p); 709 p->p_slptime = 0; 710 if ((p->p_flag & P_INMEM) == 0) { 711 p->p_flag |= P_SWAPINREQ; 712 wakeup((caddr_t)&proc0); 713 } 714 } 715 716 /* 717 * Change the process state to NOT be runnable, removing it from the run 718 * queue. 719 */ 720 void 721 clrrunnable(struct proc *p, int stat) 722 { 723 crit_enter_quick(p->p_thread); 724 if (p->p_stat == SRUN && (p->p_flag & P_ONRUNQ)) 725 remrunqueue(p); 726 p->p_stat = stat; 727 crit_exit_quick(p->p_thread); 728 } 729 730 /* 731 * Compute the priority of a process when running in user mode. 732 * Arrange to reschedule if the resulting priority is better 733 * than that of the current process. 734 */ 735 void 736 resetpriority(struct proc *p) 737 { 738 int newpriority; 739 int interactive; 740 int opq; 741 int npq; 742 743 /* 744 * Set p_priority for general process comparisons 745 */ 746 switch(p->p_rtprio.type) { 747 case RTP_PRIO_REALTIME: 748 p->p_priority = PRIBASE_REALTIME + p->p_rtprio.prio; 749 return; 750 case RTP_PRIO_NORMAL: 751 break; 752 case RTP_PRIO_IDLE: 753 p->p_priority = PRIBASE_IDLE + p->p_rtprio.prio; 754 return; 755 case RTP_PRIO_THREAD: 756 p->p_priority = PRIBASE_THREAD + p->p_rtprio.prio; 757 return; 758 } 759 760 /* 761 * NORMAL priorities fall through. These are based on niceness 762 * and cpu use. Lower numbers == higher priorities. 763 */ 764 newpriority = (int)(NICE_ADJUST(p->p_nice - PRIO_MIN) + 765 p->p_estcpu / ESTCPURAMP); 766 767 /* 768 * p_interactive is -128 to +127 and represents very long term 769 * interactivity or batch (whereas estcpu is a much faster variable). 770 * Interactivity can modify the priority by up to 8 units either way. 771 * (8 units == approximately 4 nice levels). 772 */ 773 interactive = p->p_interactive / 10; 774 newpriority += interactive; 775 776 newpriority = MIN(newpriority, MAXPRI); 777 newpriority = MAX(newpriority, 0); 778 npq = newpriority / PPQ; 779 crit_enter(); 780 opq = (p->p_priority & PRIMASK) / PPQ; 781 if (p->p_stat == SRUN && (p->p_flag & P_ONRUNQ) && opq != npq) { 782 /* 783 * We have to move the process to another queue 784 */ 785 remrunqueue(p); 786 p->p_priority = PRIBASE_NORMAL + newpriority; 787 setrunqueue(p); 788 } else { 789 /* 790 * We can just adjust the priority and it will be picked 791 * up later. 792 */ 793 KKASSERT(opq == npq || (p->p_flag & P_ONRUNQ) == 0); 794 p->p_priority = PRIBASE_NORMAL + newpriority; 795 } 796 crit_exit(); 797 } 798 799 /* 800 * Compute a tenex style load average of a quantity on 801 * 1, 5 and 15 minute intervals. 802 */ 803 static void 804 loadav(void *arg) 805 { 806 int i, nrun; 807 struct loadavg *avg; 808 struct proc *p; 809 thread_t td; 810 811 avg = &averunnable; 812 nrun = 0; 813 FOREACH_PROC_IN_SYSTEM(p) { 814 switch (p->p_stat) { 815 case SRUN: 816 if ((td = p->p_thread) == NULL) 817 break; 818 if (td->td_flags & TDF_BLOCKED) 819 break; 820 /* fall through */ 821 case SIDL: 822 nrun++; 823 break; 824 default: 825 break; 826 } 827 } 828 for (i = 0; i < 3; i++) 829 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 830 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 831 832 /* 833 * Schedule the next update to occur after 5 seconds, but add a 834 * random variation to avoid synchronisation with processes that 835 * run at regular intervals. 836 */ 837 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)), 838 loadav, NULL); 839 } 840 841 /* ARGSUSED */ 842 static void 843 sched_setup(void *dummy) 844 { 845 callout_init(&loadav_callout); 846 callout_init(&roundrobin_callout); 847 callout_init(&schedcpu_callout); 848 849 /* Kick off timeout driven events by calling first time. */ 850 roundrobin(NULL); 851 schedcpu(NULL); 852 loadav(NULL); 853 } 854 855 /* 856 * We adjust the priority of the current process. The priority of 857 * a process gets worse as it accumulates CPU time. The cpu usage 858 * estimator (p_estcpu) is increased here. resetpriority() will 859 * compute a different priority each time p_estcpu increases by 860 * INVERSE_ESTCPU_WEIGHT * (until MAXPRI is reached). 861 * 862 * The cpu usage estimator ramps up quite quickly when the process is 863 * running (linearly), and decays away exponentially, at a rate which 864 * is proportionally slower when the system is busy. The basic principle 865 * is that the system will 90% forget that the process used a lot of CPU 866 * time in 5 * loadav seconds. This causes the system to favor processes 867 * which haven't run much recently, and to round-robin among other processes. 868 * 869 * The actual schedulerclock interrupt rate is ESTCPUFREQ, but we generally 870 * want to ramp-up at a faster rate, ESTCPUVFREQ, so p_estcpu is scaled 871 * by (ESTCPUVFREQ / ESTCPUFREQ). You can control the ramp-up/ramp-down 872 * rate by adjusting ESTCPUVFREQ in sys/proc.h in integer multiples 873 * of ESTCPUFREQ. 874 * 875 * WARNING! called from a fast-int or an IPI, the MP lock MIGHT NOT BE HELD 876 * and we cannot block. 877 */ 878 void 879 schedulerclock(void *dummy) 880 { 881 struct thread *td; 882 struct proc *p; 883 884 td = curthread; 885 if ((p = td->td_proc) != NULL) { 886 p->p_cpticks++; /* cpticks runs at ESTCPUFREQ */ 887 p->p_estcpu = ESTCPULIM(p->p_estcpu + ESTCPUVFREQ / ESTCPUFREQ); 888 if (try_mplock()) { 889 resetpriority(p); 890 rel_mplock(); 891 } 892 } 893 } 894 895 static 896 void 897 crit_panicints(void) 898 { 899 int s; 900 int cpri; 901 902 s = splhigh(); 903 cpri = crit_panic_save(); 904 splx(safepri); 905 crit_panic_restore(cpri); 906 splx(s); 907 } 908 909