kern_synch.c revision 78638
112795Swpaul/*- 212795Swpaul * Copyright (c) 1982, 1986, 1990, 1991, 1993 312795Swpaul * The Regents of the University of California. All rights reserved. 412795Swpaul * (c) UNIX System Laboratories, Inc. 512795Swpaul * All or some portions of this file are derived from material licensed 612795Swpaul * to the University of California by American Telephone and Telegraph 712795Swpaul * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8100441Scharnier * the permission of UNIX System Laboratories, Inc. 912795Swpaul * 1012795Swpaul * Redistribution and use in source and binary forms, with or without 1112795Swpaul * modification, are permitted provided that the following conditions 12100441Scharnier * are met: 1312795Swpaul * 1. Redistributions of source code must retain the above copyright 1412795Swpaul * notice, this list of conditions and the following disclaimer. 1512795Swpaul * 2. Redistributions in binary form must reproduce the above copyright 16100441Scharnier * notice, this list of conditions and the following disclaimer in the 1712795Swpaul * documentation and/or other materials provided with the distribution. 1812795Swpaul * 3. All advertising materials mentioning features or use of this software 1912795Swpaul * must display the following acknowledgement: 20100441Scharnier * This product includes software developed by the University of 2112795Swpaul * California, Berkeley and its contributors. 2212795Swpaul * 4. Neither the name of the University nor the names of its contributors 2312795Swpaul * may be used to endorse or promote products derived from this software 24100441Scharnier * without specific prior written permission. 2512795Swpaul * 2612795Swpaul * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 2712795Swpaul * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 2812795Swpaul * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 2912795Swpaul * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30100441Scharnier * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 3112795Swpaul * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32146833Sstefanf * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 3312795Swpaul * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 3412795Swpaul * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 3527935Scharnier * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 3612795Swpaul * SUCH DAMAGE. 37100441Scharnier * 38100441Scharnier * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95 39100441Scharnier * $FreeBSD: head/sys/kern/kern_synch.c 78638 2001-06-22 23:11:26Z jhb $ 4012795Swpaul */ 4112795Swpaul 4212795Swpaul#include "opt_ktrace.h" 4312795Swpaul 4427935Scharnier#include <sys/param.h> 4512795Swpaul#include <sys/systm.h> 46200462Sdelphij#include <sys/condvar.h> 4712795Swpaul#include <sys/kernel.h> 48149682Sstefanf#include <sys/ktr.h> 4912795Swpaul#include <sys/lock.h> 5012795Swpaul#include <sys/mutex.h> 5112795Swpaul#include <sys/proc.h> 5212795Swpaul#include <sys/resourcevar.h> 5312795Swpaul#include <sys/signalvar.h> 5412795Swpaul#include <sys/smp.h> 5512795Swpaul#include <sys/sx.h> 5612795Swpaul#include <sys/sysctl.h> 5712795Swpaul#include <sys/sysproto.h> 58141614Salfred#include <sys/vmmeter.h> 5912795Swpaul#include <vm/vm.h> 60141614Salfred#include <vm/vm_extern.h> 6112795Swpaul#ifdef KTRACE 62141614Salfred#include <sys/uio.h> 6312795Swpaul#include <sys/ktrace.h> 6412795Swpaul#endif 6512795Swpaul 6612795Swpaul#include <machine/cpu.h> 6792921Simp 68152398Sdwmalonestatic void sched_setup __P((void *dummy)); 6912795SwpaulSYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 7012795Swpaul 71152398Sdwmaloneint hogticks; 7212795Swpaulint lbolt; 7312795Swpaulint sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 7412795Swpaul 7512795Swpaulstatic struct callout schedcpu_callout; 7612795Swpaulstatic struct callout roundrobin_callout; 7712795Swpaul 7812795Swpaulstatic void endtsleep __P((void *)); 7912795Swpaulstatic void roundrobin __P((void *arg)); 8012795Swpaulstatic void schedcpu __P((void *arg)); 8112795Swpaul 8212795Swpaulstatic int 8312795Swpaulsysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 8412795Swpaul{ 8517142Sjkh int error, new_val; 86152398Sdwmalone 8712795Swpaul new_val = sched_quantum * tick; 8812795Swpaul error = sysctl_handle_int(oidp, &new_val, 0, req); 8912795Swpaul if (error != 0 || req->newptr == NULL) 9012795Swpaul return (error); 9112795Swpaul if (new_val < tick) 9212795Swpaul return (EINVAL); 9312795Swpaul sched_quantum = new_val / tick; 9412795Swpaul hogticks = 2 * sched_quantum; 9512795Swpaul return (0); 9612795Swpaul} 9712795Swpaul 9812795SwpaulSYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 9912795Swpaul 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 10012795Swpaul 10112795Swpaul/* 10212795Swpaul * Arrange to reschedule if necessary, taking the priorities and 10312795Swpaul * schedulers into account. 10412795Swpaul */ 10512795Swpaulvoid 106222080Sbenlmaybe_resched(p) 10712795Swpaul struct proc *p; 10812795Swpaul{ 10912795Swpaul 11012795Swpaul mtx_assert(&sched_lock, MA_OWNED); 11112795Swpaul if (p->p_pri.pri_level < curproc->p_pri.pri_level) 11212795Swpaul need_resched(curproc); 11312795Swpaul} 11427935Scharnier 11512795Swpaulint 11612795Swpaulroundrobin_interval(void) 11712795Swpaul{ 11812795Swpaul return (sched_quantum); 11912795Swpaul} 120141614Salfred 12112795Swpaul/* 12212795Swpaul * Force switch among equal priority processes every 100ms. 123149709Sstefanf */ 12412795Swpaul/* ARGSUSED */ 12512795Swpaulstatic void 12612795Swpaulroundrobin(arg) 12712795Swpaul void *arg; 12812795Swpaul{ 12912795Swpaul 13012795Swpaul mtx_lock_spin(&sched_lock); 13112795Swpaul need_resched(curproc); 13212795Swpaul#ifdef SMP 13312795Swpaul forward_roundrobin(); 13412795Swpaul#endif 135100441Scharnier mtx_unlock_spin(&sched_lock); 13612795Swpaul 13712795Swpaul callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 13812795Swpaul} 13912795Swpaul 14012795Swpaul/* 14112795Swpaul * Constants for digital decay and forget: 14212795Swpaul * 90% of (p_estcpu) usage in 5 * loadav time 14312795Swpaul * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 144222080Sbenl * Note that, as ps(1) mentions, this can let percentages 14512795Swpaul * total over 100% (I've seen 137.9% for 3 processes). 14612795Swpaul * 14712795Swpaul * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 14812795Swpaul * 14917142Sjkh * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 150152398Sdwmalone * That is, the system wants to compute a value of decay such 15112795Swpaul * that the following for loop: 15212795Swpaul * for (i = 0; i < (5 * loadavg); i++) 15312795Swpaul * p_estcpu *= decay; 15412795Swpaul * will compute 15512795Swpaul * p_estcpu *= 0.1; 156141614Salfred * for all values of loadavg: 15712795Swpaul * 15812795Swpaul * Mathematically this loop can be expressed by saying: 15912795Swpaul * decay ** (5 * loadavg) ~= .1 16012795Swpaul * 16112795Swpaul * The system computes decay as: 16212795Swpaul * decay = (2 * loadavg) / (2 * loadavg + 1) 16312795Swpaul * 16412795Swpaul * We wish to prove that the system's computation of decay 16512795Swpaul * will always fulfill the equation: 166141614Salfred * decay ** (5 * loadavg) ~= .1 16712795Swpaul * 16812795Swpaul * If we compute b as: 16912795Swpaul * b = 2 * loadavg 17012795Swpaul * then 17112795Swpaul * decay = b / (b + 1) 17212795Swpaul * 173 * We now need to prove two things: 174 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 175 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 176 * 177 * Facts: 178 * For x close to zero, exp(x) =~ 1 + x, since 179 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 180 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 181 * For x close to zero, ln(1+x) =~ x, since 182 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 183 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 184 * ln(.1) =~ -2.30 185 * 186 * Proof of (1): 187 * Solve (factor)**(power) =~ .1 given power (5*loadav): 188 * solving for factor, 189 * ln(factor) =~ (-2.30/5*loadav), or 190 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 191 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 192 * 193 * Proof of (2): 194 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 195 * solving for power, 196 * power*ln(b/(b+1)) =~ -2.30, or 197 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 198 * 199 * Actual power values for the implemented algorithm are as follows: 200 * loadav: 1 2 3 4 201 * power: 5.68 10.32 14.94 19.55 202 */ 203 204/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 205#define loadfactor(loadav) (2 * (loadav)) 206#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 207 208/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 209static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 210SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 211 212/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 213static int fscale __unused = FSCALE; 214SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 215 216/* 217 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 218 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 219 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 220 * 221 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 222 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 223 * 224 * If you don't want to bother with the faster/more-accurate formula, you 225 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 226 * (more general) method of calculating the %age of CPU used by a process. 227 */ 228#define CCPU_SHIFT 11 229 230/* 231 * Recompute process priorities, every hz ticks. 232 * MP-safe, called without the Giant mutex. 233 */ 234/* ARGSUSED */ 235static void 236schedcpu(arg) 237 void *arg; 238{ 239 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 240 register struct proc *p; 241 register int realstathz, s; 242 243 realstathz = stathz ? stathz : hz; 244 sx_slock(&allproc_lock); 245 LIST_FOREACH(p, &allproc, p_list) { 246 /* 247 * Increment time in/out of memory and sleep time 248 * (if sleeping). We ignore overflow; with 16-bit int's 249 * (remember them?) overflow takes 45 days. 250 if (p->p_stat == SWAIT) 251 continue; 252 */ 253 mtx_lock_spin(&sched_lock); 254 p->p_swtime++; 255 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 256 p->p_slptime++; 257 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 258 /* 259 * If the process has slept the entire second, 260 * stop recalculating its priority until it wakes up. 261 */ 262 if (p->p_slptime > 1) { 263 mtx_unlock_spin(&sched_lock); 264 continue; 265 } 266 267 /* 268 * prevent state changes and protect run queue 269 */ 270 s = splhigh(); 271 272 /* 273 * p_pctcpu is only for ps. 274 */ 275#if (FSHIFT >= CCPU_SHIFT) 276 p->p_pctcpu += (realstathz == 100)? 277 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 278 100 * (((fixpt_t) p->p_cpticks) 279 << (FSHIFT - CCPU_SHIFT)) / realstathz; 280#else 281 p->p_pctcpu += ((FSCALE - ccpu) * 282 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT; 283#endif 284 p->p_cpticks = 0; 285 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu); 286 resetpriority(p); 287 if (p->p_pri.pri_level >= PUSER) { 288 if ((p != curproc) && 289#ifdef SMP 290 p->p_oncpu == NOCPU && /* idle */ 291#endif 292 p->p_stat == SRUN && 293 (p->p_sflag & PS_INMEM) && 294 (p->p_pri.pri_level / RQ_PPQ) != 295 (p->p_pri.pri_user / RQ_PPQ)) { 296 remrunqueue(p); 297 p->p_pri.pri_level = p->p_pri.pri_user; 298 setrunqueue(p); 299 } else 300 p->p_pri.pri_level = p->p_pri.pri_user; 301 } 302 mtx_unlock_spin(&sched_lock); 303 splx(s); 304 } 305 sx_sunlock(&allproc_lock); 306 vmmeter(); 307 wakeup((caddr_t)&lbolt); 308 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 309} 310 311/* 312 * Recalculate the priority of a process after it has slept for a while. 313 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 314 * least six times the loadfactor will decay p_estcpu to zero. 315 */ 316void 317updatepri(p) 318 register struct proc *p; 319{ 320 register unsigned int newcpu = p->p_estcpu; 321 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 322 323 if (p->p_slptime > 5 * loadfac) 324 p->p_estcpu = 0; 325 else { 326 p->p_slptime--; /* the first time was done in schedcpu */ 327 while (newcpu && --p->p_slptime) 328 newcpu = decay_cpu(loadfac, newcpu); 329 p->p_estcpu = newcpu; 330 } 331 resetpriority(p); 332} 333 334/* 335 * We're only looking at 7 bits of the address; everything is 336 * aligned to 4, lots of things are aligned to greater powers 337 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 338 */ 339#define TABLESIZE 128 340static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE]; 341#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 342 343void 344sleepinit(void) 345{ 346 int i; 347 348 sched_quantum = hz/10; 349 hogticks = 2 * sched_quantum; 350 for (i = 0; i < TABLESIZE; i++) 351 TAILQ_INIT(&slpque[i]); 352} 353 354/* 355 * General sleep call. Suspends the current process until a wakeup is 356 * performed on the specified identifier. The process will then be made 357 * runnable with the specified priority. Sleeps at most timo/hz seconds 358 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 359 * before and after sleeping, else signals are not checked. Returns 0 if 360 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 361 * signal needs to be delivered, ERESTART is returned if the current system 362 * call should be restarted if possible, and EINTR is returned if the system 363 * call should be interrupted by the signal (return EINTR). 364 * 365 * The mutex argument is exited before the caller is suspended, and 366 * entered before msleep returns. If priority includes the PDROP 367 * flag the mutex is not entered before returning. 368 */ 369int 370msleep(ident, mtx, priority, wmesg, timo) 371 void *ident; 372 struct mtx *mtx; 373 int priority, timo; 374 const char *wmesg; 375{ 376 struct proc *p = curproc; 377 int sig, catch = priority & PCATCH; 378 int rval = 0; 379 WITNESS_SAVE_DECL(mtx); 380 381#ifdef KTRACE 382 if (p && KTRPOINT(p, KTR_CSW)) 383 ktrcsw(p->p_tracep, 1, 0); 384#endif 385 WITNESS_SLEEP(0, &mtx->mtx_object); 386 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 387 ("sleeping without a mutex")); 388 mtx_lock_spin(&sched_lock); 389 if (cold || panicstr) { 390 /* 391 * After a panic, or during autoconfiguration, 392 * just give interrupts a chance, then just return; 393 * don't run any other procs or panic below, 394 * in case this is the idle process and already asleep. 395 */ 396 if (mtx != NULL && priority & PDROP) 397 mtx_unlock_flags(mtx, MTX_NOSWITCH); 398 mtx_unlock_spin(&sched_lock); 399 return (0); 400 } 401 402 DROP_GIANT_NOSWITCH(); 403 404 if (mtx != NULL) { 405 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 406 WITNESS_SAVE(&mtx->mtx_object, mtx); 407 mtx_unlock_flags(mtx, MTX_NOSWITCH); 408 if (priority & PDROP) 409 mtx = NULL; 410 } 411 412 KASSERT(p != NULL, ("msleep1")); 413 KASSERT(ident != NULL && p->p_stat == SRUN, ("msleep")); 414 /* 415 * Process may be sitting on a slpque if asleep() was called, remove 416 * it before re-adding. 417 */ 418 if (p->p_wchan != NULL) 419 unsleep(p); 420 421 p->p_wchan = ident; 422 p->p_wmesg = wmesg; 423 p->p_slptime = 0; 424 p->p_pri.pri_level = priority & PRIMASK; 425 CTR3(KTR_PROC, "msleep: proc %p (pid %d, %s)", p, p->p_pid, p->p_comm); 426 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 427 if (timo) 428 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 429 /* 430 * We put ourselves on the sleep queue and start our timeout 431 * before calling CURSIG, as we could stop there, and a wakeup 432 * or a SIGCONT (or both) could occur while we were stopped. 433 * A SIGCONT would cause us to be marked as SSLEEP 434 * without resuming us, thus we must be ready for sleep 435 * when CURSIG is called. If the wakeup happens while we're 436 * stopped, p->p_wchan will be 0 upon return from CURSIG. 437 */ 438 if (catch) { 439 CTR3(KTR_PROC, "msleep caught: proc %p (pid %d, %s)", p, 440 p->p_pid, p->p_comm); 441 p->p_sflag |= PS_SINTR; 442 mtx_unlock_spin(&sched_lock); 443 PROC_LOCK(p); 444 sig = CURSIG(p); 445 mtx_lock_spin(&sched_lock); 446 PROC_UNLOCK_NOSWITCH(p); 447 if (sig != 0) { 448 if (p->p_wchan) 449 unsleep(p); 450 } else if (p->p_wchan == NULL) 451 catch = 0; 452 } else 453 sig = 0; 454 if (p->p_wchan != NULL) { 455 p->p_stat = SSLEEP; 456 p->p_stats->p_ru.ru_nvcsw++; 457 mi_switch(); 458 } 459 CTR3(KTR_PROC, "msleep resume: proc %p (pid %d, %s)", p, p->p_pid, 460 p->p_comm); 461 KASSERT(p->p_stat == SRUN, ("running but not SRUN")); 462 p->p_sflag &= ~PS_SINTR; 463 if (p->p_sflag & PS_TIMEOUT) { 464 p->p_sflag &= ~PS_TIMEOUT; 465 if (sig == 0) 466 rval = EWOULDBLOCK; 467 } else if (timo) 468 callout_stop(&p->p_slpcallout); 469 mtx_unlock_spin(&sched_lock); 470 471 if (rval == 0 && catch) { 472 PROC_LOCK(p); 473 /* XXX: shouldn't we always be calling CURSIG() */ 474 if (sig != 0 || (sig = CURSIG(p))) { 475 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 476 rval = EINTR; 477 else 478 rval = ERESTART; 479 } 480 PROC_UNLOCK(p); 481 } 482 PICKUP_GIANT(); 483#ifdef KTRACE 484 mtx_lock(&Giant); 485 if (KTRPOINT(p, KTR_CSW)) 486 ktrcsw(p->p_tracep, 0, 0); 487 mtx_unlock(&Giant); 488#endif 489 if (mtx != NULL) { 490 mtx_lock(mtx); 491 WITNESS_RESTORE(&mtx->mtx_object, mtx); 492 } 493 return (rval); 494} 495 496/* 497 * asleep() - async sleep call. Place process on wait queue and return 498 * immediately without blocking. The process stays runnable until mawait() 499 * is called. If ident is NULL, remove process from wait queue if it is still 500 * on one. 501 * 502 * Only the most recent sleep condition is effective when making successive 503 * calls to asleep() or when calling msleep(). 504 * 505 * The timeout, if any, is not initiated until mawait() is called. The sleep 506 * priority, signal, and timeout is specified in the asleep() call but may be 507 * overriden in the mawait() call. 508 * 509 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 510 */ 511 512int 513asleep(void *ident, int priority, const char *wmesg, int timo) 514{ 515 struct proc *p = curproc; 516 517 /* 518 * Remove preexisting wait condition (if any) and place process 519 * on appropriate slpque, but do not put process to sleep. 520 */ 521 522 mtx_lock_spin(&sched_lock); 523 524 if (p->p_wchan != NULL) 525 unsleep(p); 526 527 if (ident) { 528 p->p_wchan = ident; 529 p->p_wmesg = wmesg; 530 p->p_slptime = 0; 531 p->p_asleep.as_priority = priority; 532 p->p_asleep.as_timo = timo; 533 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 534 } 535 536 mtx_unlock_spin(&sched_lock); 537 538 return(0); 539} 540 541/* 542 * mawait() - wait for async condition to occur. The process blocks until 543 * wakeup() is called on the most recent asleep() address. If wakeup is called 544 * prior to mawait(), mawait() winds up being a NOP. 545 * 546 * If mawait() is called more then once (without an intervening asleep() call), 547 * mawait() is still effectively a NOP but it calls mi_switch() to give other 548 * processes some cpu before returning. The process is left runnable. 549 * 550 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 551 */ 552 553int 554mawait(struct mtx *mtx, int priority, int timo) 555{ 556 struct proc *p = curproc; 557 int rval = 0; 558 WITNESS_SAVE_DECL(mtx); 559 560 WITNESS_SLEEP(0, &mtx->mtx_object); 561 KASSERT(timo > 0 || mtx_owned(&Giant) || mtx != NULL, 562 ("sleeping without a mutex")); 563 mtx_lock_spin(&sched_lock); 564 DROP_GIANT_NOSWITCH(); 565 if (mtx != NULL) { 566 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 567 WITNESS_SAVE(&mtx->mtx_object, mtx); 568 mtx_unlock_flags(mtx, MTX_NOSWITCH); 569 if (priority & PDROP) 570 mtx = NULL; 571 } 572 573 if (p->p_wchan != NULL) { 574 int sig; 575 int catch; 576 577#ifdef KTRACE 578 if (p && KTRPOINT(p, KTR_CSW)) 579 ktrcsw(p->p_tracep, 1, 0); 580#endif 581 /* 582 * The call to mawait() can override defaults specified in 583 * the original asleep(). 584 */ 585 if (priority < 0) 586 priority = p->p_asleep.as_priority; 587 if (timo < 0) 588 timo = p->p_asleep.as_timo; 589 590 /* 591 * Install timeout 592 */ 593 594 if (timo) 595 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 596 597 sig = 0; 598 catch = priority & PCATCH; 599 600 if (catch) { 601 p->p_sflag |= PS_SINTR; 602 mtx_unlock_spin(&sched_lock); 603 PROC_LOCK(p); 604 sig = CURSIG(p); 605 mtx_lock_spin(&sched_lock); 606 PROC_UNLOCK_NOSWITCH(p); 607 if (sig != 0) { 608 if (p->p_wchan) 609 unsleep(p); 610 } else if (p->p_wchan == NULL) 611 catch = 0; 612 } 613 if (p->p_wchan != NULL) { 614 p->p_stat = SSLEEP; 615 p->p_stats->p_ru.ru_nvcsw++; 616 mi_switch(); 617 } 618 KASSERT(p->p_stat == SRUN, ("running but not SRUN")); 619 p->p_sflag &= ~PS_SINTR; 620 if (p->p_sflag & PS_TIMEOUT) { 621 p->p_sflag &= ~PS_TIMEOUT; 622 if (sig == 0) 623 rval = EWOULDBLOCK; 624 } else if (timo) 625 callout_stop(&p->p_slpcallout); 626 mtx_unlock_spin(&sched_lock); 627 if (rval == 0 && catch) { 628 PROC_LOCK(p); 629 if (sig != 0 || (sig = CURSIG(p))) { 630 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 631 rval = EINTR; 632 else 633 rval = ERESTART; 634 } 635 PROC_UNLOCK(p); 636 } 637#ifdef KTRACE 638 mtx_lock(&Giant); 639 if (KTRPOINT(p, KTR_CSW)) 640 ktrcsw(p->p_tracep, 0, 0); 641 mtx_unlock(&Giant); 642#endif 643 } else { 644 /* 645 * If as_priority is 0, mawait() has been called without an 646 * intervening asleep(). We are still effectively a NOP, 647 * but we call mi_switch() for safety. 648 */ 649 650 if (p->p_asleep.as_priority == 0) { 651 p->p_stats->p_ru.ru_nvcsw++; 652 mi_switch(); 653 } 654 mtx_unlock_spin(&sched_lock); 655 } 656 657 /* 658 * clear p_asleep.as_priority as an indication that mawait() has been 659 * called. If mawait() is called again without an intervening asleep(), 660 * mawait() is still effectively a NOP but the above mi_switch() code 661 * is triggered as a safety. 662 */ 663 if (rval == 0) 664 p->p_asleep.as_priority = 0; 665 666 PICKUP_GIANT(); 667 if (mtx != NULL) { 668 mtx_lock(mtx); 669 WITNESS_RESTORE(&mtx->mtx_object, mtx); 670 } 671 return (rval); 672} 673 674/* 675 * Implement timeout for msleep or asleep()/mawait() 676 * 677 * If process hasn't been awakened (wchan non-zero), 678 * set timeout flag and undo the sleep. If proc 679 * is stopped, just unsleep so it will remain stopped. 680 * MP-safe, called without the Giant mutex. 681 */ 682static void 683endtsleep(arg) 684 void *arg; 685{ 686 register struct proc *p; 687 int s; 688 689 p = (struct proc *)arg; 690 CTR3(KTR_PROC, "endtsleep: proc %p (pid %d, %s)", p, p->p_pid, 691 p->p_comm); 692 s = splhigh(); 693 mtx_lock_spin(&sched_lock); 694 if (p->p_wchan) { 695 if (p->p_stat == SSLEEP) 696 setrunnable(p); 697 else 698 unsleep(p); 699 p->p_sflag |= PS_TIMEOUT; 700 } 701 mtx_unlock_spin(&sched_lock); 702 splx(s); 703} 704 705/* 706 * Remove a process from its wait queue 707 */ 708void 709unsleep(p) 710 register struct proc *p; 711{ 712 int s; 713 714 s = splhigh(); 715 mtx_lock_spin(&sched_lock); 716 if (p->p_wchan) { 717 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_slpq); 718 p->p_wchan = NULL; 719 } 720 mtx_unlock_spin(&sched_lock); 721 splx(s); 722} 723 724/* 725 * Make all processes sleeping on the specified identifier runnable. 726 */ 727void 728wakeup(ident) 729 register void *ident; 730{ 731 register struct slpquehead *qp; 732 register struct proc *p; 733 int s; 734 735 s = splhigh(); 736 mtx_lock_spin(&sched_lock); 737 qp = &slpque[LOOKUP(ident)]; 738restart: 739 TAILQ_FOREACH(p, qp, p_slpq) { 740 if (p->p_wchan == ident) { 741 TAILQ_REMOVE(qp, p, p_slpq); 742 p->p_wchan = NULL; 743 if (p->p_stat == SSLEEP) { 744 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 745 CTR3(KTR_PROC, "wakeup: proc %p (pid %d, %s)", 746 p, p->p_pid, p->p_comm); 747 if (p->p_slptime > 1) 748 updatepri(p); 749 p->p_slptime = 0; 750 p->p_stat = SRUN; 751 if (p->p_sflag & PS_INMEM) { 752 setrunqueue(p); 753 maybe_resched(p); 754 } else { 755 p->p_sflag |= PS_SWAPINREQ; 756 wakeup((caddr_t)&proc0); 757 } 758 /* END INLINE EXPANSION */ 759 goto restart; 760 } 761 } 762 } 763 mtx_unlock_spin(&sched_lock); 764 splx(s); 765} 766 767/* 768 * Make a process sleeping on the specified identifier runnable. 769 * May wake more than one process if a target process is currently 770 * swapped out. 771 */ 772void 773wakeup_one(ident) 774 register void *ident; 775{ 776 register struct slpquehead *qp; 777 register struct proc *p; 778 int s; 779 780 s = splhigh(); 781 mtx_lock_spin(&sched_lock); 782 qp = &slpque[LOOKUP(ident)]; 783 784 TAILQ_FOREACH(p, qp, p_slpq) { 785 if (p->p_wchan == ident) { 786 TAILQ_REMOVE(qp, p, p_slpq); 787 p->p_wchan = NULL; 788 if (p->p_stat == SSLEEP) { 789 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 790 CTR3(KTR_PROC, "wakeup1: proc %p (pid %d, %s)", 791 p, p->p_pid, p->p_comm); 792 if (p->p_slptime > 1) 793 updatepri(p); 794 p->p_slptime = 0; 795 p->p_stat = SRUN; 796 if (p->p_sflag & PS_INMEM) { 797 setrunqueue(p); 798 maybe_resched(p); 799 break; 800 } else { 801 p->p_sflag |= PS_SWAPINREQ; 802 wakeup((caddr_t)&proc0); 803 } 804 /* END INLINE EXPANSION */ 805 } 806 } 807 } 808 mtx_unlock_spin(&sched_lock); 809 splx(s); 810} 811 812/* 813 * The machine independent parts of mi_switch(). 814 * Must be called at splstatclock() or higher. 815 */ 816void 817mi_switch() 818{ 819 struct timeval new_switchtime; 820 register struct proc *p = curproc; /* XXX */ 821#if 0 822 register struct rlimit *rlim; 823#endif 824 u_int sched_nest; 825 826 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 827 828 /* 829 * Compute the amount of time during which the current 830 * process was running, and add that to its total so far. 831 */ 832 microuptime(&new_switchtime); 833 if (timevalcmp(&new_switchtime, PCPU_PTR(switchtime), <)) { 834#if 0 835 /* XXX: This doesn't play well with sched_lock right now. */ 836 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 837 PCPU_GET(switchtime.tv_sec), PCPU_GET(switchtime.tv_usec), 838 new_switchtime.tv_sec, new_switchtime.tv_usec); 839#endif 840 new_switchtime = PCPU_GET(switchtime); 841 } else { 842 p->p_runtime += (new_switchtime.tv_usec - PCPU_GET(switchtime.tv_usec)) + 843 (new_switchtime.tv_sec - PCPU_GET(switchtime.tv_sec)) * 844 (int64_t)1000000; 845 } 846 847#if 0 848 /* 849 * Check if the process exceeds its cpu resource allocation. 850 * If over max, kill it. 851 * 852 * XXX drop sched_lock, pickup Giant 853 */ 854 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 855 p->p_runtime > p->p_limit->p_cpulimit) { 856 rlim = &p->p_rlimit[RLIMIT_CPU]; 857 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 858 mtx_unlock_spin(&sched_lock); 859 PROC_LOCK(p); 860 killproc(p, "exceeded maximum CPU limit"); 861 mtx_lock_spin(&sched_lock); 862 PROC_UNLOCK_NOSWITCH(p); 863 } else { 864 mtx_unlock_spin(&sched_lock); 865 PROC_LOCK(p); 866 psignal(p, SIGXCPU); 867 mtx_lock_spin(&sched_lock); 868 PROC_UNLOCK_NOSWITCH(p); 869 if (rlim->rlim_cur < rlim->rlim_max) { 870 /* XXX: we should make a private copy */ 871 rlim->rlim_cur += 5; 872 } 873 } 874 } 875#endif 876 877 /* 878 * Pick a new current process and record its start time. 879 */ 880 cnt.v_swtch++; 881 PCPU_SET(switchtime, new_switchtime); 882 CTR3(KTR_PROC, "mi_switch: old proc %p (pid %d, %s)", p, p->p_pid, 883 p->p_comm); 884 sched_nest = sched_lock.mtx_recurse; 885 curproc->p_lastcpu = curproc->p_oncpu; 886 curproc->p_oncpu = NOCPU; 887 clear_resched(curproc); 888 cpu_switch(); 889 curproc->p_oncpu = PCPU_GET(cpuid); 890 sched_lock.mtx_recurse = sched_nest; 891 sched_lock.mtx_lock = (uintptr_t)curproc; 892 CTR3(KTR_PROC, "mi_switch: new proc %p (pid %d, %s)", p, p->p_pid, 893 p->p_comm); 894 if (PCPU_GET(switchtime.tv_sec) == 0) 895 microuptime(PCPU_PTR(switchtime)); 896 PCPU_SET(switchticks, ticks); 897} 898 899/* 900 * Change process state to be runnable, 901 * placing it on the run queue if it is in memory, 902 * and awakening the swapper if it isn't in memory. 903 */ 904void 905setrunnable(p) 906 register struct proc *p; 907{ 908 register int s; 909 910 s = splhigh(); 911 mtx_lock_spin(&sched_lock); 912 switch (p->p_stat) { 913 case 0: 914 case SRUN: 915 case SZOMB: 916 case SWAIT: 917 default: 918 panic("setrunnable"); 919 case SSTOP: 920 case SSLEEP: /* e.g. when sending signals */ 921 if (p->p_sflag & PS_CVWAITQ) 922 cv_waitq_remove(p); 923 else 924 unsleep(p); 925 break; 926 927 case SIDL: 928 break; 929 } 930 p->p_stat = SRUN; 931 if (p->p_sflag & PS_INMEM) 932 setrunqueue(p); 933 splx(s); 934 if (p->p_slptime > 1) 935 updatepri(p); 936 p->p_slptime = 0; 937 if ((p->p_sflag & PS_INMEM) == 0) { 938 p->p_sflag |= PS_SWAPINREQ; 939 wakeup((caddr_t)&proc0); 940 } 941 else 942 maybe_resched(p); 943 mtx_unlock_spin(&sched_lock); 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(p) 953 register struct proc *p; 954{ 955 register unsigned int newpriority; 956 957 mtx_lock_spin(&sched_lock); 958 if (p->p_pri.pri_class == PRI_TIMESHARE) { 959 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 960 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 961 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 962 PRI_MAX_TIMESHARE); 963 p->p_pri.pri_user = newpriority; 964 } 965 maybe_resched(p); 966 mtx_unlock_spin(&sched_lock); 967} 968 969/* ARGSUSED */ 970static void 971sched_setup(dummy) 972 void *dummy; 973{ 974 975 callout_init(&schedcpu_callout, 1); 976 callout_init(&roundrobin_callout, 0); 977 978 /* Kick off timeout driven events by calling first time. */ 979 roundrobin(NULL); 980 schedcpu(NULL); 981} 982 983/* 984 * We adjust the priority of the current process. The priority of 985 * a process gets worse as it accumulates CPU time. The cpu usage 986 * estimator (p_estcpu) is increased here. resetpriority() will 987 * compute a different priority each time p_estcpu increases by 988 * INVERSE_ESTCPU_WEIGHT 989 * (until MAXPRI is reached). The cpu usage estimator ramps up 990 * quite quickly when the process is running (linearly), and decays 991 * away exponentially, at a rate which is proportionally slower when 992 * the system is busy. The basic principle is that the system will 993 * 90% forget that the process used a lot of CPU time in 5 * loadav 994 * seconds. This causes the system to favor processes which haven't 995 * run much recently, and to round-robin among other processes. 996 */ 997void 998schedclock(p) 999 struct proc *p; 1000{ 1001 1002 p->p_cpticks++; 1003 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 1004 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1005 resetpriority(p); 1006 if (p->p_pri.pri_level >= PUSER) 1007 p->p_pri.pri_level = p->p_pri.pri_user; 1008 } 1009} 1010 1011/* 1012 * General purpose yield system call 1013 */ 1014int 1015yield(struct proc *p, struct yield_args *uap) 1016{ 1017 int s; 1018 1019 p->p_retval[0] = 0; 1020 1021 s = splhigh(); 1022 mtx_lock_spin(&sched_lock); 1023 DROP_GIANT_NOSWITCH(); 1024 p->p_pri.pri_level = PRI_MAX_TIMESHARE; 1025 setrunqueue(p); 1026 p->p_stats->p_ru.ru_nvcsw++; 1027 mi_switch(); 1028 mtx_unlock_spin(&sched_lock); 1029 PICKUP_GIANT(); 1030 splx(s); 1031 1032 return (0); 1033} 1034