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 1.54 2005/11/19 17:19:47 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 == SZOMB || 196 p->p_limit == NULL || 197 p->p_thread == NULL 198 ) { 199 crit_exit(); 200 continue; 201 } 202 ttime = p->p_thread->td_sticks + p->p_thread->td_uticks; 203 if (p->p_limit->p_cpulimit != RLIM_INFINITY && 204 ttime > p->p_limit->p_cpulimit 205 ) { 206 rlim = &p->p_rlimit[RLIMIT_CPU]; 207 if (ttime / (rlim_t)1000000 >= rlim->rlim_max) { 208 killproc(p, "exceeded maximum CPU limit"); 209 } else { 210 psignal(p, SIGXCPU); 211 if (rlim->rlim_cur < rlim->rlim_max) { 212 /* XXX: we should make a private copy */ 213 rlim->rlim_cur += 5; 214 } 215 } 216 crit_exit(); 217 break; 218 } 219 crit_exit(); 220 } 221 222 wakeup((caddr_t)&lbolt); 223 wakeup((caddr_t)&lbolt_syncer); 224 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 225 } 226 227 /* 228 * This is only used by ps. Generate a cpu percentage use over 229 * a period of one second. 230 */ 231 void 232 updatepcpu(struct lwp *lp, int cpticks, int ttlticks) 233 { 234 fixpt_t acc; 235 int remticks; 236 237 acc = (cpticks << FSHIFT) / ttlticks; 238 if (ttlticks >= ESTCPUFREQ) { 239 lp->lwp_pctcpu = acc; 240 } else { 241 remticks = ESTCPUFREQ - ttlticks; 242 lp->lwp_pctcpu = (acc * ttlticks + lp->lwp_pctcpu * remticks) / 243 ESTCPUFREQ; 244 } 245 } 246 247 /* 248 * We're only looking at 7 bits of the address; everything is 249 * aligned to 4, lots of things are aligned to greater powers 250 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 251 */ 252 #define TABLESIZE 128 253 #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 254 255 static cpumask_t slpque_cpumasks[TABLESIZE]; 256 257 /* 258 * General scheduler initialization. We force a reschedule 25 times 259 * a second by default. Note that cpu0 is initialized in early boot and 260 * cannot make any high level calls. 261 * 262 * Each cpu has its own sleep queue. 263 */ 264 void 265 sleep_gdinit(globaldata_t gd) 266 { 267 static struct tslpque slpque_cpu0[TABLESIZE]; 268 int i; 269 270 if (gd->gd_cpuid == 0) { 271 sched_quantum = (hz + 24) / 25; 272 hogticks = 2 * sched_quantum; 273 274 gd->gd_tsleep_hash = slpque_cpu0; 275 } else { 276 gd->gd_tsleep_hash = malloc(sizeof(slpque_cpu0), 277 M_TSLEEP, M_WAITOK | M_ZERO); 278 } 279 for (i = 0; i < TABLESIZE; ++i) 280 TAILQ_INIT(&gd->gd_tsleep_hash[i]); 281 } 282 283 /* 284 * General sleep call. Suspends the current process until a wakeup is 285 * performed on the specified identifier. The process will then be made 286 * runnable with the specified priority. Sleeps at most timo/hz seconds 287 * (0 means no timeout). If flags includes PCATCH flag, signals are checked 288 * before and after sleeping, else signals are not checked. Returns 0 if 289 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 290 * signal needs to be delivered, ERESTART is returned if the current system 291 * call should be restarted if possible, and EINTR is returned if the system 292 * call should be interrupted by the signal (return EINTR). 293 * 294 * Note that if we are a process, we release_curproc() before messing with 295 * the LWKT scheduler. 296 * 297 * During autoconfiguration or after a panic, a sleep will simply 298 * lower the priority briefly to allow interrupts, then return. 299 */ 300 int 301 tsleep(void *ident, int flags, const char *wmesg, int timo) 302 { 303 struct thread *td = curthread; 304 struct proc *p = td->td_proc; /* may be NULL */ 305 globaldata_t gd; 306 int sig; 307 int catch; 308 int id; 309 int error; 310 int oldpri; 311 struct callout thandle; 312 313 /* 314 * NOTE: removed KTRPOINT, it could cause races due to blocking 315 * even in stable. Just scrap it for now. 316 */ 317 if (cold || panicstr) { 318 /* 319 * After a panic, or during autoconfiguration, 320 * just give interrupts a chance, then just return; 321 * don't run any other procs or panic below, 322 * in case this is the idle process and already asleep. 323 */ 324 splz(); 325 oldpri = td->td_pri & TDPRI_MASK; 326 lwkt_setpri_self(safepri); 327 lwkt_switch(); 328 lwkt_setpri_self(oldpri); 329 return (0); 330 } 331 gd = td->td_gd; 332 KKASSERT(td != &gd->gd_idlethread); /* you must be kidding! */ 333 334 /* 335 * NOTE: all of this occurs on the current cpu, including any 336 * callout-based wakeups, so a critical section is a sufficient 337 * interlock. 338 * 339 * The entire sequence through to where we actually sleep must 340 * run without breaking the critical section. 341 */ 342 id = LOOKUP(ident); 343 catch = flags & PCATCH; 344 error = 0; 345 sig = 0; 346 347 crit_enter_quick(td); 348 349 KASSERT(ident != NULL, ("tsleep: no ident")); 350 KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d", 351 ident, wmesg, p->p_stat)); 352 353 /* 354 * Setup for the current process (if this is a process). 355 */ 356 if (p) { 357 if (catch) { 358 /* 359 * Early termination if PCATCH was set and a 360 * signal is pending, interlocked with the 361 * critical section. 362 * 363 * Early termination only occurs when tsleep() is 364 * entered while in a normal SRUN state. 365 */ 366 if ((sig = CURSIG(p)) != 0) 367 goto resume; 368 369 /* 370 * Causes psignal to wake us up when. 371 */ 372 p->p_flag |= P_SINTR; 373 } 374 375 /* 376 * Make sure the current process has been untangled from 377 * the userland scheduler and initialize slptime to start 378 * counting. 379 */ 380 if (flags & PNORESCHED) 381 td->td_flags |= TDF_NORESCHED; 382 p->p_usched->release_curproc(&p->p_lwp); 383 p->p_slptime = 0; 384 } 385 386 /* 387 * Move our thread to the correct queue and setup our wchan, etc. 388 */ 389 lwkt_deschedule_self(td); 390 td->td_flags |= TDF_TSLEEPQ; 391 TAILQ_INSERT_TAIL(&gd->gd_tsleep_hash[id], td, td_threadq); 392 atomic_set_int(&slpque_cpumasks[id], gd->gd_cpumask); 393 394 td->td_wchan = ident; 395 td->td_wmesg = wmesg; 396 td->td_wdomain = flags & PDOMAIN_MASK; 397 398 /* 399 * Setup the timeout, if any 400 */ 401 if (timo) { 402 callout_init(&thandle); 403 callout_reset(&thandle, timo, endtsleep, td); 404 } 405 406 /* 407 * Beddy bye bye. 408 */ 409 if (p) { 410 /* 411 * Ok, we are sleeping. Remove us from the userland runq 412 * and place us in the SSLEEP state. 413 */ 414 if (p->p_flag & P_ONRUNQ) 415 p->p_usched->remrunqueue(&p->p_lwp); 416 p->p_stat = SSLEEP; 417 p->p_stats->p_ru.ru_nvcsw++; 418 lwkt_switch(); 419 p->p_stat = SRUN; 420 } else { 421 lwkt_switch(); 422 } 423 424 /* 425 * Make sure we haven't switched cpus while we were asleep. It's 426 * not supposed to happen. Cleanup our temporary flags. 427 */ 428 KKASSERT(gd == td->td_gd); 429 td->td_flags &= ~TDF_NORESCHED; 430 431 /* 432 * Cleanup the timeout. 433 */ 434 if (timo) { 435 if (td->td_flags & TDF_TIMEOUT) { 436 td->td_flags &= ~TDF_TIMEOUT; 437 if (sig == 0) 438 error = EWOULDBLOCK; 439 } else { 440 callout_stop(&thandle); 441 } 442 } 443 444 /* 445 * Since td_threadq is used both for our run queue AND for the 446 * tsleep hash queue, we can't still be on it at this point because 447 * we've gotten cpu back. 448 */ 449 KKASSERT((td->td_flags & TDF_TSLEEPQ) == 0); 450 td->td_wchan = NULL; 451 td->td_wmesg = NULL; 452 td->td_wdomain = 0; 453 454 /* 455 * Figure out the correct error return 456 */ 457 resume: 458 if (p) { 459 p->p_flag &= ~(P_BREAKTSLEEP | P_SINTR); 460 if (catch && error == 0 && (sig != 0 || (sig = CURSIG(p)))) { 461 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 462 error = EINTR; 463 else 464 error = ERESTART; 465 } 466 } 467 crit_exit_quick(td); 468 return (error); 469 } 470 471 /* 472 * This is a dandy function that allows us to interlock tsleep/wakeup 473 * operations with unspecified upper level locks, such as lockmgr locks, 474 * simply by holding a critical section. The sequence is: 475 * 476 * (enter critical section) 477 * (acquire upper level lock) 478 * tsleep_interlock(blah) 479 * (release upper level lock) 480 * tsleep(blah, ...) 481 * (exit critical section) 482 * 483 * Basically this function sets our cpumask for the ident which informs 484 * other cpus that our cpu 'might' be waiting (or about to wait on) the 485 * hash index related to the ident. The critical section prevents another 486 * cpu's wakeup() from being processed on our cpu until we are actually 487 * able to enter the tsleep(). Thus, no race occurs between our attempt 488 * to release a resource and sleep, and another cpu's attempt to acquire 489 * a resource and call wakeup. 490 * 491 * There isn't much of a point to this function unless you call it while 492 * holding a critical section. 493 */ 494 void 495 tsleep_interlock(void *ident) 496 { 497 int id = LOOKUP(ident); 498 499 atomic_set_int(&slpque_cpumasks[id], mycpu->gd_cpumask); 500 } 501 502 /* 503 * Implement the timeout for tsleep. 504 * 505 * We set P_BREAKTSLEEP to indicate that an event has occured, but 506 * we only call setrunnable if the process is not stopped. 507 * 508 * This type of callout timeout is scheduled on the same cpu the process 509 * is sleeping on. Also, at the moment, the MP lock is held. 510 */ 511 static void 512 endtsleep(void *arg) 513 { 514 thread_t td = arg; 515 struct proc *p; 516 517 ASSERT_MP_LOCK_HELD(curthread); 518 crit_enter(); 519 520 /* 521 * cpu interlock. Thread flags are only manipulated on 522 * the cpu owning the thread. proc flags are only manipulated 523 * by the older of the MP lock. We have both. 524 */ 525 if (td->td_flags & TDF_TSLEEPQ) { 526 td->td_flags |= TDF_TIMEOUT; 527 528 if ((p = td->td_proc) != NULL) { 529 p->p_flag |= P_BREAKTSLEEP; 530 if ((p->p_flag & P_STOPPED) == 0) 531 setrunnable(p); 532 } else { 533 unsleep_and_wakeup_thread(td); 534 } 535 } 536 crit_exit(); 537 } 538 539 /* 540 * Unsleep and wakeup a thread. This function runs without the MP lock 541 * which means that it can only manipulate thread state on the owning cpu, 542 * and cannot touch the process state at all. 543 */ 544 static 545 void 546 unsleep_and_wakeup_thread(struct thread *td) 547 { 548 globaldata_t gd = mycpu; 549 int id; 550 551 #ifdef SMP 552 if (td->td_gd != gd) { 553 lwkt_send_ipiq(td->td_gd, (ipifunc1_t)unsleep_and_wakeup_thread, td); 554 return; 555 } 556 #endif 557 crit_enter(); 558 if (td->td_flags & TDF_TSLEEPQ) { 559 td->td_flags &= ~TDF_TSLEEPQ; 560 id = LOOKUP(td->td_wchan); 561 TAILQ_REMOVE(&gd->gd_tsleep_hash[id], td, td_threadq); 562 if (TAILQ_FIRST(&gd->gd_tsleep_hash[id]) == NULL) 563 atomic_clear_int(&slpque_cpumasks[id], gd->gd_cpumask); 564 lwkt_schedule(td); 565 } 566 crit_exit(); 567 } 568 569 /* 570 * Make all processes sleeping on the specified identifier runnable. 571 * count may be zero or one only. 572 * 573 * The domain encodes the sleep/wakeup domain AND the first cpu to check 574 * (which is always the current cpu). As we iterate across cpus 575 * 576 * This call may run without the MP lock held. We can only manipulate thread 577 * state on the cpu owning the thread. We CANNOT manipulate process state 578 * at all. 579 */ 580 static void 581 _wakeup(void *ident, int domain) 582 { 583 struct tslpque *qp; 584 struct thread *td; 585 struct thread *ntd; 586 globaldata_t gd; 587 #ifdef SMP 588 cpumask_t mask; 589 cpumask_t tmask; 590 int startcpu; 591 int nextcpu; 592 #endif 593 int id; 594 595 crit_enter(); 596 gd = mycpu; 597 id = LOOKUP(ident); 598 qp = &gd->gd_tsleep_hash[id]; 599 restart: 600 for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { 601 ntd = TAILQ_NEXT(td, td_threadq); 602 if (td->td_wchan == ident && 603 td->td_wdomain == (domain & PDOMAIN_MASK) 604 ) { 605 KKASSERT(td->td_flags & TDF_TSLEEPQ); 606 td->td_flags &= ~TDF_TSLEEPQ; 607 TAILQ_REMOVE(qp, td, td_threadq); 608 if (TAILQ_FIRST(qp) == NULL) { 609 atomic_clear_int(&slpque_cpumasks[id], 610 gd->gd_cpumask); 611 } 612 lwkt_schedule(td); 613 if (domain & PWAKEUP_ONE) 614 goto done; 615 goto restart; 616 } 617 } 618 619 #ifdef SMP 620 /* 621 * We finished checking the current cpu but there still may be 622 * more work to do. Either wakeup_one was requested and no matching 623 * thread was found, or a normal wakeup was requested and we have 624 * to continue checking cpus. 625 * 626 * The cpu that started the wakeup sequence is encoded in the domain. 627 * We use this information to determine which cpus still need to be 628 * checked, locate a candidate cpu, and chain the wakeup 629 * asynchronously with an IPI message. 630 * 631 * It should be noted that this scheme is actually less expensive then 632 * the old scheme when waking up multiple threads, since we send 633 * only one IPI message per target candidate which may then schedule 634 * multiple threads. Before we could have wound up sending an IPI 635 * message for each thread on the target cpu (!= current cpu) that 636 * needed to be woken up. 637 * 638 * NOTE: Wakeups occuring on remote cpus are asynchronous. This 639 * should be ok since we are passing idents in the IPI rather then 640 * thread pointers. 641 */ 642 if ((mask = slpque_cpumasks[id]) != 0) { 643 /* 644 * Look for a cpu that might have work to do. Mask out cpus 645 * which have already been processed. 646 * 647 * 31xxxxxxxxxxxxxxxxxxxxxxxxxxxxx0 648 * ^ ^ ^ 649 * start currentcpu start 650 * case2 case1 651 * * * * 652 * 11111111111111110000000000000111 case1 653 * 00000000111111110000000000000000 case2 654 * 655 * case1: We started at start_case1 and processed through 656 * to the current cpu. We have to check any bits 657 * after the current cpu, then check bits before 658 * the starting cpu. 659 * 660 * case2: We have already checked all the bits from 661 * start_case2 to the end, and from 0 to the current 662 * cpu. We just have the bits from the current cpu 663 * to start_case2 left to check. 664 */ 665 startcpu = PWAKEUP_DECODE(domain); 666 if (gd->gd_cpuid >= startcpu) { 667 /* 668 * CASE1 669 */ 670 tmask = mask & ~((gd->gd_cpumask << 1) - 1); 671 if (mask & tmask) { 672 nextcpu = bsfl(mask & tmask); 673 lwkt_send_ipiq2(globaldata_find(nextcpu), 674 _wakeup, ident, domain); 675 } else { 676 tmask = (1 << startcpu) - 1; 677 if (mask & tmask) { 678 nextcpu = bsfl(mask & tmask); 679 lwkt_send_ipiq2( 680 globaldata_find(nextcpu), 681 _wakeup, ident, domain); 682 } 683 } 684 } else { 685 /* 686 * CASE2 687 */ 688 tmask = ~((gd->gd_cpumask << 1) - 1) & 689 ((1 << startcpu) - 1); 690 if (mask & tmask) { 691 nextcpu = bsfl(mask & tmask); 692 lwkt_send_ipiq2(globaldata_find(nextcpu), 693 _wakeup, ident, domain); 694 } 695 } 696 } 697 #endif 698 done: 699 crit_exit(); 700 } 701 702 void 703 wakeup(void *ident) 704 { 705 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid)); 706 } 707 708 void 709 wakeup_one(void *ident) 710 { 711 /* XXX potentially round-robin the first responding cpu */ 712 _wakeup(ident, PWAKEUP_ENCODE(0, mycpu->gd_cpuid) | PWAKEUP_ONE); 713 } 714 715 void 716 wakeup_domain(void *ident, int domain) 717 { 718 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid)); 719 } 720 721 void 722 wakeup_domain_one(void *ident, int domain) 723 { 724 /* XXX potentially round-robin the first responding cpu */ 725 _wakeup(ident, PWAKEUP_ENCODE(domain, mycpu->gd_cpuid) | PWAKEUP_ONE); 726 } 727 728 /* 729 * setrunnable() 730 * 731 * Make a process runnable. The MP lock must be held on call. This only 732 * has an effect if we are in SSLEEP. We only break out of the 733 * tsleep if P_BREAKTSLEEP is set, otherwise we just fix-up the state. 734 * 735 * NOTE: With the MP lock held we can only safely manipulate the process 736 * structure. We cannot safely manipulate the thread structure. 737 */ 738 void 739 setrunnable(struct proc *p) 740 { 741 crit_enter(); 742 ASSERT_MP_LOCK_HELD(curthread); 743 p->p_flag &= ~P_STOPPED; 744 if (p->p_stat == SSLEEP && (p->p_flag & P_BREAKTSLEEP)) { 745 unsleep_and_wakeup_thread(p->p_thread); 746 } 747 crit_exit(); 748 } 749 750 /* 751 * The process is stopped due to some condition, usually because P_STOPPED 752 * is set but also possibly due to being traced. 753 * 754 * NOTE! If the caller sets P_STOPPED, the caller must also clear P_WAITED 755 * because the parent may check the child's status before the child actually 756 * gets to this routine. 757 * 758 * This routine is called with the current process only, typically just 759 * before returning to userland. 760 * 761 * Setting P_BREAKTSLEEP before entering the tsleep will cause a passive 762 * SIGCONT to break out of the tsleep. 763 */ 764 void 765 tstop(struct proc *p) 766 { 767 wakeup((caddr_t)p->p_pptr); 768 p->p_flag |= P_BREAKTSLEEP; 769 tsleep(p, 0, "stop", 0); 770 } 771 772 /* 773 * Yield / synchronous reschedule. This is a bit tricky because the trap 774 * code might have set a lazy release on the switch function. Setting 775 * P_PASSIVE_ACQ will ensure that the lazy release executes when we call 776 * switch, and that we are given a greater chance of affinity with our 777 * current cpu. 778 * 779 * We call lwkt_setpri_self() to rotate our thread to the end of the lwkt 780 * run queue. lwkt_switch() will also execute any assigned passive release 781 * (which usually calls release_curproc()), allowing a same/higher priority 782 * process to be designated as the current process. 783 * 784 * While it is possible for a lower priority process to be designated, 785 * it's call to lwkt_maybe_switch() in acquire_curproc() will likely 786 * round-robin back to us and we will be able to re-acquire the current 787 * process designation. 788 */ 789 void 790 uio_yield(void) 791 { 792 struct thread *td = curthread; 793 struct proc *p = td->td_proc; 794 795 lwkt_setpri_self(td->td_pri & TDPRI_MASK); 796 if (p) { 797 p->p_flag |= P_PASSIVE_ACQ; 798 lwkt_switch(); 799 p->p_flag &= ~P_PASSIVE_ACQ; 800 } else { 801 lwkt_switch(); 802 } 803 } 804 805 /* 806 * Compute a tenex style load average of a quantity on 807 * 1, 5 and 15 minute intervals. 808 */ 809 static void 810 loadav(void *arg) 811 { 812 int i, nrun; 813 struct loadavg *avg; 814 struct proc *p; 815 thread_t td; 816 817 avg = &averunnable; 818 nrun = 0; 819 FOREACH_PROC_IN_SYSTEM(p) { 820 switch (p->p_stat) { 821 case SRUN: 822 if ((td = p->p_thread) == NULL) 823 break; 824 if (td->td_flags & TDF_BLOCKED) 825 break; 826 /* fall through */ 827 case SIDL: 828 nrun++; 829 break; 830 default: 831 break; 832 } 833 } 834 for (i = 0; i < 3; i++) 835 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 836 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 837 838 /* 839 * Schedule the next update to occur after 5 seconds, but add a 840 * random variation to avoid synchronisation with processes that 841 * run at regular intervals. 842 */ 843 callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)), 844 loadav, NULL); 845 } 846 847 /* ARGSUSED */ 848 static void 849 sched_setup(void *dummy) 850 { 851 callout_init(&loadav_callout); 852 callout_init(&schedcpu_callout); 853 854 /* Kick off timeout driven events by calling first time. */ 855 schedcpu(NULL); 856 loadav(NULL); 857 } 858 859