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 1.55 2005/12/01 18:30:08 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 #include <sys/lock.h> 55 #ifdef KTRACE 56 #include <sys/uio.h> 57 #include <sys/ktrace.h> 58 #endif 59 #include <sys/xwait.h> 60 61 #include <machine/cpu.h> 62 #include <machine/ipl.h> 63 #include <machine/smp.h> 64 65 TAILQ_HEAD(tslpque, thread); 66 67 static void sched_setup (void *dummy); 68 SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 69 70 int hogticks; 71 int lbolt; 72 int lbolt_syncer; 73 int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 74 int ncpus; 75 int ncpus2, ncpus2_shift, ncpus2_mask; 76 int safepri; 77 78 static struct callout loadav_callout; 79 static struct callout schedcpu_callout; 80 MALLOC_DEFINE(M_TSLEEP, "tslpque", "tsleep queues"); 81 82 struct loadavg averunnable = 83 { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ 84 /* 85 * Constants for averages over 1, 5, and 15 minutes 86 * when sampling at 5 second intervals. 87 */ 88 static fixpt_t cexp[3] = { 89 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 90 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 91 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 92 }; 93 94 static void endtsleep (void *); 95 static void unsleep_and_wakeup_thread(struct thread *td); 96 static void loadav (void *arg); 97 static void schedcpu (void *arg); 98 99 /* 100 * Adjust the scheduler quantum. The quantum is specified in microseconds. 101 * Note that 'tick' is in microseconds per tick. 102 */ 103 static int 104 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 105 { 106 int error, new_val; 107 108 new_val = sched_quantum * tick; 109 error = sysctl_handle_int(oidp, &new_val, 0, req); 110 if (error != 0 || req->newptr == NULL) 111 return (error); 112 if (new_val < tick) 113 return (EINVAL); 114 sched_quantum = new_val / tick; 115 hogticks = 2 * sched_quantum; 116 return (0); 117 } 118 119 SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 120 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 121 122 /* 123 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 124 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 125 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 126 * 127 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 128 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 129 * 130 * If you don't want to bother with the faster/more-accurate formula, you 131 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 132 * (more general) method of calculating the %age of CPU used by a process. 133 * 134 * decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing 135 */ 136 #define CCPU_SHIFT 11 137 138 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 139 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 140 141 /* 142 * kernel uses `FSCALE', userland (SHOULD) use kern.fscale 143 */ 144 static int fscale __unused = FSCALE; 145 SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 146 147 /* 148 * Recompute process priorities, once a second. 149 * 150 * Since the userland schedulers are typically event oriented, if the 151 * estcpu calculation at wakeup() time is not sufficient to make a 152 * process runnable relative to other processes in the system we have 153 * a 1-second recalc to help out. 154 * 155 * This code also allows us to store sysclock_t data in the process structure 156 * without fear of an overrun, since sysclock_t are guarenteed to hold 157 * several seconds worth of count. 158 */ 159 /* ARGSUSED */ 160 static void 161 schedcpu(void *arg) 162 { 163 struct rlimit *rlim; 164 struct proc *p; 165 u_int64_t ttime; 166 167 /* 168 * General process statistics once a second 169 */ 170 FOREACH_PROC_IN_SYSTEM(p) { 171 crit_enter(); 172 p->p_swtime++; 173 if (p->p_stat == SSLEEP) 174 p->p_slptime++; 175 176 /* 177 * Only recalculate processes that are active or have slept 178 * less then 2 seconds. The schedulers understand this. 179 */ 180 if (p->p_slptime <= 1) { 181 p->p_usched->recalculate(&p->p_lwp); 182 } else { 183 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 184 } 185 crit_exit(); 186 } 187 188 /* 189 * Resource checks. XXX break out since psignal/killproc can block, 190 * limiting us to one process killed per second. There is probably 191 * a better way. 192 */ 193 FOREACH_PROC_IN_SYSTEM(p) { 194 crit_enter(); 195 if (p->p_stat == SIDL || 196 (p->p_flag & P_ZOMBIE) || 197 p->p_limit == NULL || 198 p->p_thread == NULL 199 ) { 200 crit_exit(); 201 continue; 202 } 203 ttime = p->p_thread->td_sticks + p->p_thread->td_uticks; 204 if (p->p_limit->p_cpulimit != RLIM_INFINITY && 205 ttime > p->p_limit->p_cpulimit 206 ) { 207 rlim = &p->p_rlimit[RLIMIT_CPU]; 208 if (ttime / (rlim_t)1000000 >= rlim->rlim_max) { 209 killproc(p, "exceeded maximum CPU limit"); 210 } else { 211 psignal(p, SIGXCPU); 212 if (rlim->rlim_cur < rlim->rlim_max) { 213 /* XXX: we should make a private copy */ 214 rlim->rlim_cur += 5; 215 } 216 } 217 crit_exit(); 218 break; 219 } 220 crit_exit(); 221 } 222 223 wakeup((caddr_t)&lbolt); 224 wakeup((caddr_t)&lbolt_syncer); 225 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 226 } 227 228 /* 229 * This is only used by ps. Generate a cpu percentage use over 230 * a period of one second. 231 */ 232 void 233 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 234 { 235 fixpt_t acc; 236 int remticks; 237 238 acc = (cpticks << FSHIFT) / ttlticks; 239 if (ttlticks >= ESTCPUFREQ) { 240 lp->lwp_pctcpu = acc; 241 } else { 242 remticks = ESTCPUFREQ - ttlticks; 243 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 244 ESTCPUFREQ; 245 } 246 } 247 248 /* 249 * We're only looking at 7 bits of the address; everything is 250 * aligned to 4, lots of things are aligned to greater powers 251 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 252 */ 253 #define TABLESIZE 128 254 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 255 256 static cpumask_t slpque_cpumasks[TABLESIZE]; 257 258 /* 259 * General scheduler initialization. We force a reschedule 25 times 260 * a second by default. Note that cpu0 is initialized in early boot and 261 * cannot make any high level calls. 262 * 263 * Each cpu has its own sleep queue. 264 */ 265 void 266 sleep_gdinit(globaldata_t gd) 267 { 268 static struct tslpque slpque_cpu0[TABLESIZE]; 269 int i; 270 271 if (gd->gd_cpuid == 0) { 272 sched_quantum = (hz + 24) / 25; 273 hogticks = 2 * sched_quantum; 274 275 gd->gd_tsleep_hash = slpque_cpu0; 276 } else { 277 gd->gd_tsleep_hash = malloc(sizeof(slpque_cpu0), 278 M_TSLEEP, M_WAITOK | M_ZERO); 279 } 280 for (i = 0; i < TABLESIZE; ++i) 281 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 282 } 283 284 /* 285 * General sleep call. Suspends the current process until a wakeup is 286 * performed on the specified identifier. The process will then be made 287 * runnable with the specified priority. Sleeps at most timo/hz seconds 288 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 289 * before and after sleeping, else signals are not checked. Returns 0 if 290 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 291 * signal needs to be delivered, ERESTART is returned if the current system 292 * call should be restarted if possible, and EINTR is returned if the system 293 * call should be interrupted by the signal (return EINTR). 294 * 295 * Note that if we are a process, we release_curproc() before messing with 296 * the LWKT scheduler. 297 * 298 * During autoconfiguration or after a panic, a sleep will simply 299 * lower the priority briefly to allow interrupts, then return. 300 */ 301 int 302 tsleep(void *ident, int flags, const char *wmesg, int timo) 303 { 304 struct thread *td = curthread; 305 struct proc *p = td->td_proc; /* may be NULL */ 306 globaldata_t gd; 307 int sig; 308 int catch; 309 int id; 310 int error; 311 int oldpri; 312 struct callout thandle; 313 314 /* 315 * NOTE: removed KTRPOINT, it could cause races due to blocking 316 * even in stable. Just scrap it for now. 317 */ 318 if (cold || panicstr) { 319 /* 320 * After a panic, or during autoconfiguration, 321 * just give interrupts a chance, then just return; 322 * don't run any other procs or panic below, 323 * in case this is the idle process and already asleep. 324 */ 325 splz(); 326 oldpri = td->td_pri & TDPRI_MASK; 327 lwkt_setpri_self(safepri); 328 lwkt_switch(); 329 lwkt_setpri_self(oldpri); 330 return (0); 331 } 332 gd = td->td_gd; 333 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 334 335 /* 336 * NOTE: all of this occurs on the current cpu, including any 337 * callout-based wakeups, so a critical section is a sufficient 338 * interlock. 339 * 340 * The entire sequence through to where we actually sleep must 341 * run without breaking the critical section. 342 */ 343 id = LOOKUP(ident); 344 catch = flags & PCATCH; 345 error = 0; 346 sig = 0; 347 348 crit_enter_quick(td); 349 350 KASSERT(ident != NULL, ("tsleep: no ident")); 351 KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d", 352 ident, wmesg, p->p_stat)); 353 354 /* 355 * Setup for the current process (if this is a process). 356 */ 357 if (p) { 358 if (catch) { 359 /* 360 * Early termination if PCATCH was set and a 361 * signal is pending, interlocked with the 362 * critical section. 363 * 364 * Early termination only occurs when tsleep() is 365 * entered while in a normal SRUN state. 366 */ 367 if ((sig = CURSIG(p)) != 0) 368 goto resume; 369 370 /* 371 * Causes psignal to wake us up when. 372 */ 373 p->p_flag |= P_SINTR; 374 } 375 376 /* 377 * Make sure the current process has been untangled from 378 * the userland scheduler and initialize slptime to start 379 * counting. 380 */ 381 if (flags & PNORESCHED) 382 td->td_flags |= TDF_NORESCHED; 383 p->p_usched->release_curproc(&p->p_lwp); 384 p->p_slptime = 0; 385 } 386 387 /* 388 * Move our thread to the correct queue and setup our wchan, etc. 389 */ 390 lwkt_deschedule_self(td); 391 td->td_flags |= TDF_TSLEEPQ; 392 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_threadq); 393 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask); 394 395 td->td_wchan = ident; 396 td->td_wmesg = wmesg; 397 td->td_wdomain = flags & PDOMAIN_MASK; 398 399 /* 400 * Setup the timeout, if any 401 */ 402 if (timo) { 403 callout_init(&thandle); 404 callout_reset(&thandle, timo, endtsleep, td); 405 } 406 407 /* 408 * Beddy bye bye. 409 */ 410 if (p) { 411 /* 412 * Ok, we are sleeping. Remove us from the userland runq 413 * and place us in the SSLEEP state. 414 */ 415 if (p->p_flag & P_ONRUNQ) 416 p->p_usched->remrunqueue(&p->p_lwp); 417 p->p_stat = SSLEEP; 418 p->p_stats->p_ru.ru_nvcsw++; 419 lwkt_switch(); 420 p->p_stat = SRUN; 421 } else { 422 lwkt_switch(); 423 } 424 425 /* 426 * Make sure we haven't switched cpus while we were asleep. It's 427 * not supposed to happen. Cleanup our temporary flags. 428 */ 429 KKASSERT(gd == td->td_gd); 430 td->td_flags &= ~TDF_NORESCHED; 431 432 /* 433 * Cleanup the timeout. 434 */ 435 if (timo) { 436 if (td->td_flags & TDF_TIMEOUT) { 437 td->td_flags &= ~TDF_TIMEOUT; 438 if (sig == 0) 439 error = EWOULDBLOCK; 440 } else { 441 callout_stop(&thandle); 442 } 443 } 444 445 /* 446 * Since td_threadq is used both for our run queue AND for the 447 * tsleep hash queue, we can't still be on it at this point because 448 * we've gotten cpu back. 449 */ 450 KKASSERT((td->td_flags & TDF_TSLEEPQ) == 0); 451 td->td_wchan = NULL; 452 td->td_wmesg = NULL; 453 td->td_wdomain = 0; 454 455 /* 456 * Figure out the correct error return 457 */ 458 resume: 459 if (p) { 460 p->p_flag &= ~(P_BREAKTSLEEP | P_SINTR); 461 if (catch && error == 0 && (sig != 0 || (sig = CURSIG(p)))) { 462 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 463 error = EINTR; 464 else 465 error = ERESTART; 466 } 467 } 468 crit_exit_quick(td); 469 return (error); 470 } 471 472 /* 473 * This is a dandy function that allows us to interlock tsleep/wakeup 474 * operations with unspecified upper level locks, such as lockmgr locks, 475 * simply by holding a critical section. The sequence is: 476 * 477 * (enter critical section) 478 * (acquire upper level lock) 479 * tsleep_interlock(blah) 480 * (release upper level lock) 481 * tsleep(blah, ...) 482 * (exit critical section) 483 * 484 * Basically this function sets our cpumask for the ident which informs 485 * other cpus that our cpu 'might' be waiting (or about to wait on) the 486 * hash index related to the ident. The critical section prevents another 487 * cpu's wakeup() from being processed on our cpu until we are actually 488 * able to enter the tsleep(). Thus, no race occurs between our attempt 489 * to release a resource and sleep, and another cpu's attempt to acquire 490 * a resource and call wakeup. 491 * 492 * There isn't much of a point to this function unless you call it while 493 * holding a critical section. 494 */ 495 void 496 tsleep_interlock(void *ident) 497 { 498 int id = LOOKUP(ident); 499 500 atomic_set_int(&slpque_cpumasks[id], mycpu->gd_cpumask); 501 } 502 503 /* 504 * Implement the timeout for tsleep. 505 * 506 * We set P_BREAKTSLEEP to indicate that an event has occured, but 507 * we only call setrunnable if the process is not stopped. 508 * 509 * This type of callout timeout is scheduled on the same cpu the process 510 * is sleeping on. Also, at the moment, the MP lock is held. 511 */ 512 static void 513 endtsleep(void *arg) 514 { 515 thread_t td = arg; 516 struct proc *p; 517 518 ASSERT_MP_LOCK_HELD(curthread); 519 crit_enter(); 520 521 /* 522 * cpu interlock. Thread flags are only manipulated on 523 * the cpu owning the thread. proc flags are only manipulated 524 * by the older of the MP lock. We have both. 525 */ 526 if (td->td_flags & TDF_TSLEEPQ) { 527 td->td_flags |= TDF_TIMEOUT; 528 529 if ((p = td->td_proc) != NULL) { 530 p->p_flag |= P_BREAKTSLEEP; 531 if ((p->p_flag & P_STOPPED) == 0) 532 setrunnable(p); 533 } else { 534 unsleep_and_wakeup_thread(td); 535 } 536 } 537 crit_exit(); 538 } 539 540 /* 541 * Unsleep and wakeup a thread. This function runs without the MP lock 542 * which means that it can only manipulate thread state on the owning cpu, 543 * and cannot touch the process state at all. 544 */ 545 static 546 void 547 unsleep_and_wakeup_thread(struct thread *td) 548 { 549 globaldata_t gd = mycpu; 550 int id; 551 552 #ifdef SMP 553 if (td->td_gd != gd) { 554 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)unsleep_and_wakeup_thread, td); 555 return; 556 } 557 #endif 558 crit_enter(); 559 if (td->td_flags & TDF_TSLEEPQ) { 560 td->td_flags &= ~TDF_TSLEEPQ; 561 id = LOOKUP(td->td_wchan); 562 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_threadq); 563 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 564 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask); 565 lwkt_schedule(td); 566 } 567 crit_exit(); 568 } 569 570 /* 571 * Make all processes sleeping on the specified identifier runnable. 572 * count may be zero or one only. 573 * 574 * The domain encodes the sleep/wakeup domain AND the first cpu to check 575 * (which is always the current cpu). As we iterate across cpus 576 * 577 * This call may run without the MP lock held. We can only manipulate thread 578 * state on the cpu owning the thread. We CANNOT manipulate process state 579 * at all. 580 */ 581 static void 582 _wakeup(void *ident, int domain) 583 { 584 struct tslpque *qp; 585 struct thread *td; 586 struct thread *ntd; 587 globaldata_t gd; 588 #ifdef SMP 589 cpumask_t mask; 590 cpumask_t tmask; 591 int startcpu; 592 int nextcpu; 593 #endif 594 int id; 595 596 crit_enter(); 597 gd = mycpu; 598 id = LOOKUP(ident); 599 qp = &gd->gd_tsleep_hash[id]; 600 restart: 601 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 602 ntd = TAILQ_NEXT(td, td_threadq); 603 if (td->td_wchan == ident && 604 td->td_wdomain == (domain & PDOMAIN_MASK) 605 ) { 606 KKASSERT(td->td_flags & TDF_TSLEEPQ); 607 td->td_flags &= ~TDF_TSLEEPQ; 608 TAILQ_REMOVE(qp, td, td_threadq); 609 if (TAILQ_FIRST(qp) == NULL) { 610 atomic_clear_int(&slpque_cpumasks[id], 611 gd->gd_cpumask); 612 } 613 lwkt_schedule(td); 614 if (domain & PWAKEUP_ONE) 615 goto done; 616 goto restart; 617 } 618 } 619 620 #ifdef SMP 621 /* 622 * We finished checking the current cpu but there still may be 623 * more work to do. Either wakeup_one was requested and no matching 624 * thread was found, or a normal wakeup was requested and we have 625 * to continue checking cpus. 626 * 627 * The cpu that started the wakeup sequence is encoded in the domain. 628 * We use this information to determine which cpus still need to be 629 * checked, locate a candidate cpu, and chain the wakeup 630 * asynchronously with an IPI message. 631 * 632 * It should be noted that this scheme is actually less expensive then 633 * the old scheme when waking up multiple threads, since we send 634 * only one IPI message per target candidate which may then schedule 635 * multiple threads. Before we could have wound up sending an IPI 636 * message for each thread on the target cpu (!= current cpu) that 637 * needed to be woken up. 638 * 639 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 640 * should be ok since we are passing idents in the IPI rather then 641 * thread pointers. 642 */ 643 if ((mask = slpque_cpumasks[id]) != 0) { 644 /* 645 * Look for a cpu that might have work to do. Mask out cpus 646 * which have already been processed. 647 * 648 * 31xxxxxxxxxxxxxxxxxxxxxxxxxxxxx0 649 * ^ ^ ^ 650 * start currentcpu start 651 * case2 case1 652 * * * * 653 * 11111111111111110000000000000111 case1 654 * 00000000111111110000000000000000 case2 655 * 656 * case1: We started at start_case1 and processed through 657 * to the current cpu. We have to check any bits 658 * after the current cpu, then check bits before 659 * the starting cpu. 660 * 661 * case2: We have already checked all the bits from 662 * start_case2 to the end, and from 0 to the current 663 * cpu. We just have the bits from the current cpu 664 * to start_case2 left to check. 665 */ 666 startcpu = PWAKEUP_DECODE(domain); 667 if (gd->gd_cpuid >= startcpu) { 668 /* 669 * CASE1 670 */ 671 tmask = mask & ~((gd->gd_cpumask << 1) - 1); 672 if (mask & tmask) { 673 nextcpu = bsfl(mask & tmask); 674 lwkt_send_ipiq2(globaldata_find(nextcpu), 675 _wakeup, ident, domain); 676 } else { 677 tmask = (1 << startcpu) - 1; 678 if (mask & tmask) { 679 nextcpu = bsfl(mask & tmask); 680 lwkt_send_ipiq2( 681 globaldata_find(nextcpu), 682 _wakeup, ident, domain); 683 } 684 } 685 } else { 686 /* 687 * CASE2 688 */ 689 tmask = ~((gd->gd_cpumask << 1) - 1) & 690 ((1 << startcpu) - 1); 691 if (mask & tmask) { 692 nextcpu = bsfl(mask & tmask); 693 lwkt_send_ipiq2(globaldata_find(nextcpu), 694 _wakeup, ident, domain); 695 } 696 } 697 } 698 #endif 699 done: 700 crit_exit(); 701 } 702 703 void 704 wakeup(void *ident) 705 { 706 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid)); 707 } 708 709 void 710 wakeup_one(void *ident) 711 { 712 /* XXX potentially round-robin the first responding cpu */ 713 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE); 714 } 715 716 void 717 wakeup_domain(void *ident, int domain) 718 { 719 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 720 } 721 722 void 723 wakeup_domain_one(void *ident, int domain) 724 { 725 /* XXX potentially round-robin the first responding cpu */ 726 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 727 } 728 729 /* 730 * setrunnable() 731 * 732 * Make a process runnable. The MP lock must be held on call. This only 733 * has an effect if we are in SSLEEP. We only break out of the 734 * tsleep if P_BREAKTSLEEP is set, otherwise we just fix-up the state. 735 * 736 * NOTE: With the MP lock held we can only safely manipulate the process 737 * structure. We cannot safely manipulate the thread structure. 738 */ 739 void 740 setrunnable(struct proc *p) 741 { 742 crit_enter(); 743 ASSERT_MP_LOCK_HELD(curthread); 744 p->p_flag &= ~P_STOPPED; 745 if (p->p_stat == SSLEEP && (p->p_flag & P_BREAKTSLEEP)) { 746 unsleep_and_wakeup_thread(p->p_thread); 747 } 748 crit_exit(); 749 } 750 751 /* 752 * The process is stopped due to some condition, usually because P_STOPPED 753 * is set but also possibly due to being traced. 754 * 755 * NOTE! If the caller sets P_STOPPED, the caller must also clear P_WAITED 756 * because the parent may check the child's status before the child actually 757 * gets to this routine. 758 * 759 * This routine is called with the current process only, typically just 760 * before returning to userland. 761 * 762 * Setting P_BREAKTSLEEP before entering the tsleep will cause a passive 763 * SIGCONT to break out of the tsleep. 764 */ 765 void 766 tstop(struct proc *p) 767 { 768 wakeup((caddr_t)p->p_pptr); 769 p->p_flag |= P_BREAKTSLEEP; 770 tsleep(p, 0, "stop", 0); 771 } 772 773 /* 774 * Yield / synchronous reschedule. This is a bit tricky because the trap 775 * code might have set a lazy release on the switch function. Setting 776 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call 777 * switch, and that we are given a greater chance of affinity with our 778 * current cpu. 779 * 780 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt 781 * run queue. lwkt_switch() will also execute any assigned passive release 782 * (which usually calls release_curproc()), allowing a same/higher priority 783 * process to be designated as the current process. 784 * 785 * While it is possible for a lower priority process to be designated, 786 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely 787 * round-robin back to us and we will be able to re-acquire the current 788 * process designation. 789 */ 790 void 791 uio_yield(void) 792 { 793 struct thread *td = curthread; 794 struct proc *p = td->td_proc; 795 796 lwkt_setpri_self(td->td_pri & TDPRI_MASK); 797 if (p) { 798 p->p_flag |= P_PASSIVE_ACQ; 799 lwkt_switch(); 800 p->p_flag &= ~P_PASSIVE_ACQ; 801 } else { 802 lwkt_switch(); 803 } 804 } 805 806 /* 807 * Compute a tenex style load average of a quantity on 808 * 1, 5 and 15 minute intervals. 809 */ 810 static void 811 loadav(void *arg) 812 { 813 int i, nrun; 814 struct loadavg *avg; 815 struct proc *p; 816 thread_t td; 817 818 avg = &averunnable; 819 nrun = 0; 820 FOREACH_PROC_IN_SYSTEM(p) { 821 switch (p->p_stat) { 822 case SRUN: 823 if ((td = p->p_thread) == NULL) 824 break; 825 if (td->td_flags & TDF_BLOCKED) 826 break; 827 /* fall through */ 828 case SIDL: 829 nrun++; 830 break; 831 default: 832 break; 833 } 834 } 835 for (i = 0; i < 3; i++) 836 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 837 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 838 839 /* 840 * Schedule the next update to occur after 5 seconds, but add a 841 * random variation to avoid synchronisation with processes that 842 * run at regular intervals. 843 */ 844 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)), 845 loadav, NULL); 846 } 847 848 /* ARGSUSED */ 849 static void 850 sched_setup(void *dummy) 851 { 852 callout_init(&loadav_callout); 853 callout_init(&schedcpu_callout); 854 855 /* Kick off timeout driven events by calling first time. */ 856 schedcpu(NULL); 857 loadav(NULL); 858 } 859 860