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 1.25 2003/10/17 07:30:42 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 72 static struct callout loadav_callout; 73 74 struct loadavg averunnable = 75 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 76 /* 77 * Constants for averages over 1, 5, and 15 minutes 78 * when sampling at 5 second intervals. 79 */ 80 static fixpt_t cexp[3] = { 81 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 82 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 83 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 84 }; 85 86 static void endtsleep (void *); 87 static void loadav (void *arg); 88 static void roundrobin (void *arg); 89 static void schedcpu (void *arg); 90 static void updatepri (struct proc *p); 91 static void crit_panicints(void); 92 93 static int 94 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 95 { 96 int error, new_val; 97 98 new_val = sched_quantum * tick; 99 error = sysctl_handle_int(oidp, &new_val, 0, req); 100 if (error != 0 || req->newptr == NULL) 101 return (error); 102 if (new_val < tick) 103 return (EINVAL); 104 sched_quantum = new_val / tick; 105 hogticks = 2 * sched_quantum; 106 return (0); 107 } 108 109 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 110 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 111 112 int 113 roundrobin_interval(void) 114 { 115 return (sched_quantum); 116 } 117 118 /* 119 * Force switch among equal priority processes every 100ms. 120 * 121 * WARNING! The MP lock is not held on ipi message remotes. 122 */ 123 #ifdef SMP 124 125 static void 126 roundrobin_remote(void *arg) 127 { 128 struct proc *p = lwkt_preempted_proc(); 129 if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type)) 130 need_resched(); 131 } 132 133 #endif 134 135 static void 136 roundrobin(void *arg) 137 { 138 struct proc *p = lwkt_preempted_proc(); 139 if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type)) 140 need_resched(); 141 #ifdef SMP 142 lwkt_send_ipiq_mask(mycpu->gd_other_cpus, roundrobin_remote, NULL); 143 #endif 144 timeout(roundrobin, NULL, sched_quantum); 145 } 146 147 #ifdef SMP 148 149 void 150 resched_cpus(u_int32_t mask) 151 { 152 lwkt_send_ipiq_mask(mask, roundrobin_remote, NULL); 153 } 154 155 #endif 156 157 /* 158 * Constants for digital decay and forget: 159 * 90% of (p_estcpu) usage in 5 * loadav time 160 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 161 * Note that, as ps(1) mentions, this can let percentages 162 * total over 100% (I've seen 137.9% for 3 processes). 163 * 164 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 165 * 166 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 167 * That is, the system wants to compute a value of decay such 168 * that the following for loop: 169 * for (i = 0; i < (5 * loadavg); i++) 170 * p_estcpu *= decay; 171 * will compute 172 * p_estcpu *= 0.1; 173 * for all values of loadavg: 174 * 175 * Mathematically this loop can be expressed by saying: 176 * decay ** (5 * loadavg) ~= .1 177 * 178 * The system computes decay as: 179 * decay = (2 * loadavg) / (2 * loadavg + 1) 180 * 181 * We wish to prove that the system's computation of decay 182 * will always fulfill the equation: 183 * decay ** (5 * loadavg) ~= .1 184 * 185 * If we compute b as: 186 * b = 2 * loadavg 187 * then 188 * decay = b / (b + 1) 189 * 190 * We now need to prove two things: 191 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 192 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 193 * 194 * Facts: 195 * For x close to zero, exp(x) =~ 1 + x, since 196 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 197 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 198 * For x close to zero, ln(1+x) =~ x, since 199 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 200 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 201 * ln(.1) =~ -2.30 202 * 203 * Proof of (1): 204 * Solve (factor)**(power) =~ .1 given power (5*loadav): 205 * solving for factor, 206 * ln(factor) =~ (-2.30/5*loadav), or 207 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 208 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 209 * 210 * Proof of (2): 211 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 212 * solving for power, 213 * power*ln(b/(b+1)) =~ -2.30, or 214 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 215 * 216 * Actual power values for the implemented algorithm are as follows: 217 * loadav: 1 2 3 4 218 * power: 5.68 10.32 14.94 19.55 219 */ 220 221 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 222 #define loadfactor(loadav) (2 * (loadav)) 223 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 224 225 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 226 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 227 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 228 229 /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 230 static int fscale __unused = FSCALE; 231 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 232 233 /* 234 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 235 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 236 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 237 * 238 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 239 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 240 * 241 * If you don't want to bother with the faster/more-accurate formula, you 242 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 243 * (more general) method of calculating the %age of CPU used by a process. 244 */ 245 #define CCPU_SHIFT 11 246 247 /* 248 * Recompute process priorities, every hz ticks. 249 */ 250 /* ARGSUSED */ 251 static void 252 schedcpu(void *arg) 253 { 254 fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 255 struct proc *p; 256 int realstathz, s; 257 258 realstathz = stathz ? stathz : hz; 259 FOREACH_PROC_IN_SYSTEM(p) { 260 /* 261 * Increment time in/out of memory and sleep time 262 * (if sleeping). We ignore overflow; with 16-bit int's 263 * (remember them?) overflow takes 45 days. 264 */ 265 p->p_swtime++; 266 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 267 p->p_slptime++; 268 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 269 /* 270 * If the process has slept the entire second, 271 * stop recalculating its priority until it wakes up. 272 */ 273 if (p->p_slptime > 1) 274 continue; 275 s = splhigh(); /* prevent state changes and protect run queue */ 276 /* 277 * p_pctcpu is only for ps. 278 */ 279 #if (FSHIFT >= CCPU_SHIFT) 280 p->p_pctcpu += (realstathz == 100)? 281 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 282 100 * (((fixpt_t) p->p_cpticks) 283 << (FSHIFT - CCPU_SHIFT)) / realstathz; 284 #else 285 p->p_pctcpu += ((FSCALE - ccpu) * 286 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT; 287 #endif 288 p->p_cpticks = 0; 289 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu); 290 resetpriority(p); 291 splx(s); 292 } 293 wakeup((caddr_t)&lbolt); 294 timeout(schedcpu, (void *)0, hz); 295 } 296 297 /* 298 * Recalculate the priority of a process after it has slept for a while. 299 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 300 * least six times the loadfactor will decay p_estcpu to zero. 301 */ 302 static void 303 updatepri(struct proc *p) 304 { 305 unsigned int newcpu = p->p_estcpu; 306 fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 307 308 if (p->p_slptime > 5 * loadfac) { 309 p->p_estcpu = 0; 310 } else { 311 p->p_slptime--; /* the first time was done in schedcpu */ 312 while (newcpu && --p->p_slptime) 313 newcpu = decay_cpu(loadfac, newcpu); 314 p->p_estcpu = newcpu; 315 } 316 resetpriority(p); 317 } 318 319 /* 320 * We're only looking at 7 bits of the address; everything is 321 * aligned to 4, lots of things are aligned to greater powers 322 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 323 */ 324 #define TABLESIZE 128 325 static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE]; 326 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 327 328 /* 329 * During autoconfiguration or after a panic, a sleep will simply 330 * lower the priority briefly to allow interrupts, then return. 331 * The priority to be used (safepri) is machine-dependent, thus this 332 * value is initialized and maintained in the machine-dependent layers. 333 * This priority will typically be 0, or the lowest priority 334 * that is safe for use on the interrupt stack; it can be made 335 * higher to block network software interrupts after panics. 336 */ 337 int safepri; 338 339 void 340 sleepinit(void) 341 { 342 int i; 343 344 sched_quantum = hz/10; 345 hogticks = 2 * sched_quantum; 346 for (i = 0; i < TABLESIZE; i++) 347 TAILQ_INIT(&slpque[i]); 348 } 349 350 /* 351 * General sleep call. Suspends the current process until a wakeup is 352 * performed on the specified identifier. The process will then be made 353 * runnable with the specified priority. Sleeps at most timo/hz seconds 354 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 355 * before and after sleeping, else signals are not checked. Returns 0 if 356 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 357 * signal needs to be delivered, ERESTART is returned if the current system 358 * call should be restarted if possible, and EINTR is returned if the system 359 * call should be interrupted by the signal (return EINTR). 360 * 361 * If the process has P_CURPROC set mi_switch() will not re-queue it to 362 * the userland scheduler queues because we are in a SSLEEP state. If 363 * we are not the current process then we have to remove ourselves from 364 * the scheduler queues. 365 * 366 * YYY priority now unused 367 */ 368 int 369 tsleep(ident, flags, wmesg, timo) 370 void *ident; 371 int flags, timo; 372 const char *wmesg; 373 { 374 struct thread *td = curthread; 375 struct proc *p = td->td_proc; /* may be NULL */ 376 int s, sig = 0, catch = flags & PCATCH; 377 int id = LOOKUP(ident); 378 struct callout_handle thandle; 379 380 /* 381 * NOTE: removed KTRPOINT, it could cause races due to blocking 382 * even in stable. Just scrap it for now. 383 */ 384 if (cold || panicstr) { 385 /* 386 * After a panic, or during autoconfiguration, 387 * just give interrupts a chance, then just return; 388 * don't run any other procs or panic below, 389 * in case this is the idle process and already asleep. 390 */ 391 crit_panicints(); 392 return (0); 393 } 394 KKASSERT(td != &mycpu->gd_idlethread); /* you must be kidding! */ 395 s = splhigh(); 396 KASSERT(ident != NULL, ("tsleep: no ident")); 397 KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d", 398 ident, wmesg, p->p_stat)); 399 400 crit_enter(); 401 td->td_wchan = ident; 402 td->td_wmesg = wmesg; 403 if (p) 404 p->p_slptime = 0; 405 lwkt_deschedule_self(); 406 TAILQ_INSERT_TAIL(&slpque[id], td, td_threadq); 407 if (timo) 408 thandle = timeout(endtsleep, (void *)td, timo); 409 /* 410 * We put ourselves on the sleep queue and start our timeout 411 * before calling CURSIG, as we could stop there, and a wakeup 412 * or a SIGCONT (or both) could occur while we were stopped. 413 * A SIGCONT would cause us to be marked as SSLEEP 414 * without resuming us, thus we must be ready for sleep 415 * when CURSIG is called. If the wakeup happens while we're 416 * stopped, td->td_wchan will be 0 upon return from CURSIG. 417 */ 418 if (p) { 419 if (catch) { 420 p->p_flag |= P_SINTR; 421 if ((sig = CURSIG(p))) { 422 if (td->td_wchan) { 423 unsleep(td); 424 lwkt_schedule_self(); 425 } 426 p->p_stat = SRUN; 427 goto resume; 428 } 429 if (td->td_wchan == NULL) { 430 catch = 0; 431 goto resume; 432 } 433 } else { 434 sig = 0; 435 } 436 437 /* 438 * If we are not the current process we have to remove ourself 439 * from the run queue. 440 */ 441 KASSERT(p->p_stat == SRUN, ("PSTAT NOT SRUN %d %d", p->p_pid, p->p_stat)); 442 /* 443 * If this is the current 'user' process schedule another one. 444 */ 445 clrrunnable(p, SSLEEP); 446 p->p_stats->p_ru.ru_nvcsw++; 447 KKASSERT(td->td_release || (p->p_flag & P_CURPROC) == 0); 448 mi_switch(); 449 KASSERT(p->p_stat == SRUN, ("tsleep: stat not srun")); 450 } else { 451 lwkt_switch(); 452 } 453 resume: 454 crit_exit(); 455 if (p) 456 p->p_flag &= ~P_SINTR; 457 splx(s); 458 if (td->td_flags & TDF_TIMEOUT) { 459 td->td_flags &= ~TDF_TIMEOUT; 460 if (sig == 0) 461 return (EWOULDBLOCK); 462 } else if (timo) { 463 untimeout(endtsleep, (void *)td, thandle); 464 } else if (td->td_wmesg) { 465 /* 466 * This can happen if a thread is woken up directly. Clear 467 * wmesg to avoid debugging confusion. 468 */ 469 td->td_wmesg = NULL; 470 } 471 if (p) { 472 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 473 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 474 return (EINTR); 475 return (ERESTART); 476 } 477 } 478 return (0); 479 } 480 481 /* 482 * Implement the timeout for tsleep. We interlock against 483 * wchan when setting TDF_TIMEOUT. For processes we remove 484 * the sleep if the process is stopped rather then sleeping, 485 * so it remains stopped. 486 */ 487 static void 488 endtsleep(void *arg) 489 { 490 thread_t td = arg; 491 struct proc *p; 492 int s; 493 494 s = splhigh(); 495 if (td->td_wchan) { 496 td->td_flags |= TDF_TIMEOUT; 497 if ((p = td->td_proc) != NULL) { 498 if (p->p_stat == SSLEEP) 499 setrunnable(p); 500 else 501 unsleep(td); 502 } else { 503 unsleep(td); 504 lwkt_schedule(td); 505 } 506 } 507 splx(s); 508 } 509 510 /* 511 * Remove a process from its wait queue 512 */ 513 void 514 unsleep(struct thread *td) 515 { 516 int s; 517 518 s = splhigh(); 519 if (td->td_wchan) { 520 #if 0 521 if (p->p_flag & P_XSLEEP) { 522 struct xwait *w = p->p_wchan; 523 TAILQ_REMOVE(&w->waitq, p, p_procq); 524 p->p_flag &= ~P_XSLEEP; 525 } else 526 #endif 527 TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_threadq); 528 td->td_wchan = NULL; 529 } 530 splx(s); 531 } 532 533 #if 0 534 /* 535 * Make all processes sleeping on the explicit lock structure runnable. 536 */ 537 void 538 xwakeup(struct xwait *w) 539 { 540 struct proc *p; 541 int s; 542 543 s = splhigh(); 544 ++w->gen; 545 while ((p = TAILQ_FIRST(&w->waitq)) != NULL) { 546 TAILQ_REMOVE(&w->waitq, p, p_procq); 547 KASSERT(p->p_wchan == w && (p->p_flag & P_XSLEEP), 548 ("xwakeup: wchan mismatch for %p (%p/%p) %08x", p, p->p_wchan, w, p->p_flag & P_XSLEEP)); 549 p->p_wchan = NULL; 550 p->p_flag &= ~P_XSLEEP; 551 if (p->p_stat == SSLEEP) { 552 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 553 if (p->p_slptime > 1) 554 updatepri(p); 555 p->p_slptime = 0; 556 p->p_stat = SRUN; 557 if (p->p_flag & P_INMEM) { 558 setrunqueue(p); 559 } else { 560 p->p_flag |= P_SWAPINREQ; 561 wakeup((caddr_t)&proc0); 562 } 563 } 564 } 565 splx(s); 566 } 567 #endif 568 569 /* 570 * Make all processes sleeping on the specified identifier runnable. 571 */ 572 static void 573 _wakeup(void *ident, int count) 574 { 575 struct slpquehead *qp; 576 struct thread *td; 577 struct thread *ntd; 578 struct proc *p; 579 int s; 580 int id = LOOKUP(ident); 581 582 s = splhigh(); 583 qp = &slpque[id]; 584 restart: 585 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 586 ntd = TAILQ_NEXT(td, td_threadq); 587 if (td->td_wchan == ident) { 588 TAILQ_REMOVE(qp, td, td_threadq); 589 td->td_wchan = NULL; 590 if ((p = td->td_proc) != NULL && p->p_stat == SSLEEP) { 591 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 592 if (p->p_slptime > 1) 593 updatepri(p); 594 p->p_slptime = 0; 595 p->p_stat = SRUN; 596 if (p->p_flag & P_INMEM) { 597 setrunqueue(p); 598 } else { 599 p->p_flag |= P_SWAPINREQ; 600 wakeup((caddr_t)&proc0); 601 } 602 /* END INLINE EXPANSION */ 603 } else if (p == NULL) { 604 lwkt_schedule(td); 605 } 606 if (--count == 0) 607 break; 608 goto restart; 609 } 610 } 611 splx(s); 612 } 613 614 void 615 wakeup(void *ident) 616 { 617 _wakeup(ident, 0); 618 } 619 620 void 621 wakeup_one(void *ident) 622 { 623 _wakeup(ident, 1); 624 } 625 626 /* 627 * The machine independent parts of mi_switch(). 628 * Must be called at splstatclock() or higher. 629 */ 630 void 631 mi_switch() 632 { 633 struct thread *td = curthread; 634 struct proc *p = td->td_proc; /* XXX */ 635 struct rlimit *rlim; 636 int x; 637 u_int64_t ttime; 638 639 /* 640 * XXX this spl is almost unnecessary. It is partly to allow for 641 * sloppy callers that don't do it (issignal() via CURSIG() is the 642 * main offender). It is partly to work around a bug in the i386 643 * cpu_switch() (the ipl is not preserved). We ran for years 644 * without it. I think there was only a interrupt latency problem. 645 * The main caller, tsleep(), does an splx() a couple of instructions 646 * after calling here. The buggy caller, issignal(), usually calls 647 * here at spl0() and sometimes returns at splhigh(). The process 648 * then runs for a little too long at splhigh(). The ipl gets fixed 649 * when the process returns to user mode (or earlier). 650 * 651 * It would probably be better to always call here at spl0(). Callers 652 * are prepared to give up control to another process, so they must 653 * be prepared to be interrupted. The clock stuff here may not 654 * actually need splstatclock(). 655 */ 656 x = splstatclock(); 657 clear_resched(); 658 659 /* 660 * Check if the process exceeds its cpu resource allocation. 661 * If over max, kill it. Time spent in interrupts is not 662 * included. YYY 64 bit match is expensive. Ick. 663 */ 664 ttime = td->td_sticks + td->td_uticks; 665 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 666 ttime > p->p_limit->p_cpulimit) { 667 rlim = &p->p_rlimit[RLIMIT_CPU]; 668 if (ttime / (rlim_t)1000000 >= rlim->rlim_max) { 669 killproc(p, "exceeded maximum CPU limit"); 670 } else { 671 psignal(p, SIGXCPU); 672 if (rlim->rlim_cur < rlim->rlim_max) { 673 /* XXX: we should make a private copy */ 674 rlim->rlim_cur += 5; 675 } 676 } 677 } 678 679 /* 680 * Pick a new current process and record its start time. If we 681 * are in a SSTOPped state we deschedule ourselves. YYY this needs 682 * to be cleaned up, remember that LWKTs stay on their run queue 683 * which works differently then the user scheduler which removes 684 * the process from the runq when it runs it. 685 */ 686 mycpu->gd_cnt.v_swtch++; 687 if (p->p_stat == SSTOP) 688 lwkt_deschedule_self(); 689 lwkt_switch(); 690 691 splx(x); 692 } 693 694 /* 695 * Change process state to be runnable, 696 * placing it on the run queue if it is in memory, 697 * and awakening the swapper if it isn't in memory. 698 */ 699 void 700 setrunnable(struct proc *p) 701 { 702 int s; 703 704 s = splhigh(); 705 switch (p->p_stat) { 706 case 0: 707 case SRUN: 708 case SZOMB: 709 default: 710 panic("setrunnable"); 711 case SSTOP: 712 case SSLEEP: 713 unsleep(p->p_thread); /* e.g. when sending signals */ 714 break; 715 716 case SIDL: 717 break; 718 } 719 p->p_stat = SRUN; 720 if (p->p_flag & P_INMEM) 721 setrunqueue(p); 722 splx(s); 723 if (p->p_slptime > 1) 724 updatepri(p); 725 p->p_slptime = 0; 726 if ((p->p_flag & P_INMEM) == 0) { 727 p->p_flag |= P_SWAPINREQ; 728 wakeup((caddr_t)&proc0); 729 } 730 } 731 732 /* 733 * Change the process state to NOT be runnable, removing it from the run 734 * queue. If P_CURPROC is not set and we are in SRUN the process is on the 735 * run queue (If P_INMEM is not set then it isn't because it is swapped). 736 */ 737 void 738 clrrunnable(struct proc *p, int stat) 739 { 740 int s; 741 742 s = splhigh(); 743 switch(p->p_stat) { 744 case SRUN: 745 if (p->p_flag & P_ONRUNQ) 746 remrunqueue(p); 747 break; 748 default: 749 break; 750 } 751 p->p_stat = stat; 752 splx(s); 753 } 754 755 /* 756 * Compute the priority of a process when running in user mode. 757 * Arrange to reschedule if the resulting priority is better 758 * than that of the current process. 759 */ 760 void 761 resetpriority(struct proc *p) 762 { 763 unsigned int newpriority; 764 int opq; 765 int npq; 766 767 /* 768 * Set p_priority for general process comparisons 769 */ 770 switch(p->p_rtprio.type) { 771 case RTP_PRIO_REALTIME: 772 p->p_priority = PRIBASE_REALTIME + p->p_rtprio.prio; 773 return; 774 case RTP_PRIO_NORMAL: 775 break; 776 case RTP_PRIO_IDLE: 777 p->p_priority = PRIBASE_IDLE + p->p_rtprio.prio; 778 return; 779 case RTP_PRIO_THREAD: 780 p->p_priority = PRIBASE_THREAD + p->p_rtprio.prio; 781 return; 782 } 783 784 /* 785 * NORMAL priorities fall through. These are based on niceness 786 * and cpu use. 787 */ 788 newpriority = NICE_ADJUST(p->p_nice - PRIO_MIN) + 789 p->p_estcpu / ESTCPURAMP; 790 newpriority = min(newpriority, MAXPRI); 791 npq = newpriority / PPQ; 792 crit_enter(); 793 opq = (p->p_priority & PRIMASK) / PPQ; 794 if (p->p_stat == SRUN && (p->p_flag & P_ONRUNQ) && opq != npq) { 795 /* 796 * We have to move the process to another queue 797 */ 798 remrunqueue(p); 799 p->p_priority = PRIBASE_NORMAL + newpriority; 800 setrunqueue(p); 801 } else { 802 /* 803 * We can just adjust the priority and it will be picked 804 * up later. 805 */ 806 KKASSERT(opq == npq || (p->p_flag & P_ONRUNQ) == 0); 807 p->p_priority = PRIBASE_NORMAL + newpriority; 808 } 809 crit_exit(); 810 } 811 812 /* 813 * Compute a tenex style load average of a quantity on 814 * 1, 5 and 15 minute intervals. 815 */ 816 static void 817 loadav(void *arg) 818 { 819 int i, nrun; 820 struct loadavg *avg; 821 struct proc *p; 822 823 avg = &averunnable; 824 nrun = 0; 825 FOREACH_PROC_IN_SYSTEM(p) { 826 switch (p->p_stat) { 827 case SRUN: 828 case SIDL: 829 nrun++; 830 } 831 } 832 for (i = 0; i < 3; i++) 833 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 834 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 835 836 /* 837 * Schedule the next update to occur after 5 seconds, but add a 838 * random variation to avoid synchronisation with processes that 839 * run at regular intervals. 840 */ 841 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)), 842 loadav, NULL); 843 } 844 845 /* ARGSUSED */ 846 static void 847 sched_setup(dummy) 848 void *dummy; 849 { 850 851 callout_init(&loadav_callout); 852 853 /* Kick off timeout driven events by calling first time. */ 854 roundrobin(NULL); 855 schedcpu(NULL); 856 loadav(NULL); 857 } 858 859 /* 860 * We adjust the priority of the current process. The priority of 861 * a process gets worse as it accumulates CPU time. The cpu usage 862 * estimator (p_estcpu) is increased here. resetpriority() will 863 * compute a different priority each time p_estcpu increases by 864 * INVERSE_ESTCPU_WEIGHT * (until MAXPRI is reached). 865 * 866 * The cpu usage estimator ramps up quite quickly when the process is 867 * running (linearly), and decays away exponentially, at a rate which 868 * is proportionally slower when the system is busy. The basic principle 869 * is that the system will 90% forget that the process used a lot of CPU 870 * time in 5 * loadav seconds. This causes the system to favor processes 871 * which haven't run much recently, and to round-robin among other processes. 872 * 873 * WARNING! 874 */ 875 void 876 schedclock(void *dummy) 877 { 878 struct thread *td; 879 struct proc *p; 880 881 td = curthread; 882 if ((p = td->td_proc) != NULL) { 883 p->p_cpticks++; 884 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 885 if ((p->p_estcpu % PPQ) == 0 && try_mplock()) { 886 resetpriority(p); 887 rel_mplock(); 888 } 889 } 890 } 891 892 static 893 void 894 crit_panicints(void) 895 { 896 int s; 897 int cpri; 898 899 s = splhigh(); 900 cpri = crit_panic_save(); 901 splx(safepri); 902 crit_panic_restore(cpri); 903 splx(s); 904 } 905 906