1 /* $NetBSD: kern_synch.c,v 1.299 2012/03/03 00:22:24 matt Exp $ */ 2 3 /*- 4 * Copyright (c) 1999, 2000, 2004, 2006, 2007, 2008, 2009 5 * The NetBSD Foundation, Inc. 6 * All rights reserved. 7 * 8 * This code is derived from software contributed to The NetBSD Foundation 9 * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility, 10 * NASA Ames Research Center, by Charles M. Hannum, Andrew Doran and 11 * Daniel Sieger. 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 24 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 25 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 26 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 27 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 28 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 29 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 30 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 31 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 32 * POSSIBILITY OF SUCH DAMAGE. 33 */ 34 35 /*- 36 * Copyright (c) 1982, 1986, 1990, 1991, 1993 37 * The Regents of the University of California. All rights reserved. 38 * (c) UNIX System Laboratories, Inc. 39 * All or some portions of this file are derived from material licensed 40 * to the University of California by American Telephone and Telegraph 41 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 42 * the permission of UNIX System Laboratories, Inc. 43 * 44 * Redistribution and use in source and binary forms, with or without 45 * modification, are permitted provided that the following conditions 46 * are met: 47 * 1. Redistributions of source code must retain the above copyright 48 * notice, this list of conditions and the following disclaimer. 49 * 2. Redistributions in binary form must reproduce the above copyright 50 * notice, this list of conditions and the following disclaimer in the 51 * documentation and/or other materials provided with the distribution. 52 * 3. Neither the name of the University nor the names of its contributors 53 * may be used to endorse or promote products derived from this software 54 * without specific prior written permission. 55 * 56 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 57 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 58 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 59 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 60 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 61 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 62 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 63 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 64 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 65 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 66 * SUCH DAMAGE. 67 * 68 * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95 69 */ 70 71 #include <sys/cdefs.h> 72 __KERNEL_RCSID(0, "$NetBSD: kern_synch.c,v 1.299 2012/03/03 00:22:24 matt Exp $"); 73 74 #include "opt_kstack.h" 75 #include "opt_perfctrs.h" 76 #include "opt_dtrace.h" 77 78 #define __MUTEX_PRIVATE 79 80 #include <sys/param.h> 81 #include <sys/systm.h> 82 #include <sys/proc.h> 83 #include <sys/kernel.h> 84 #if defined(PERFCTRS) 85 #include <sys/pmc.h> 86 #endif 87 #include <sys/cpu.h> 88 #include <sys/pserialize.h> 89 #include <sys/resourcevar.h> 90 #include <sys/sched.h> 91 #include <sys/syscall_stats.h> 92 #include <sys/sleepq.h> 93 #include <sys/lockdebug.h> 94 #include <sys/evcnt.h> 95 #include <sys/intr.h> 96 #include <sys/lwpctl.h> 97 #include <sys/atomic.h> 98 #include <sys/simplelock.h> 99 #include <sys/syslog.h> 100 101 #include <uvm/uvm_extern.h> 102 103 #include <dev/lockstat.h> 104 105 #include <sys/dtrace_bsd.h> 106 int dtrace_vtime_active=0; 107 dtrace_vtime_switch_func_t dtrace_vtime_switch_func; 108 109 static void sched_unsleep(struct lwp *, bool); 110 static void sched_changepri(struct lwp *, pri_t); 111 static void sched_lendpri(struct lwp *, pri_t); 112 static void resched_cpu(struct lwp *); 113 114 syncobj_t sleep_syncobj = { 115 SOBJ_SLEEPQ_SORTED, 116 sleepq_unsleep, 117 sleepq_changepri, 118 sleepq_lendpri, 119 syncobj_noowner, 120 }; 121 122 syncobj_t sched_syncobj = { 123 SOBJ_SLEEPQ_SORTED, 124 sched_unsleep, 125 sched_changepri, 126 sched_lendpri, 127 syncobj_noowner, 128 }; 129 130 /* "Lightning bolt": once a second sleep address. */ 131 kcondvar_t lbolt __cacheline_aligned; 132 133 u_int sched_pstats_ticks __cacheline_aligned; 134 135 /* Preemption event counters. */ 136 static struct evcnt kpreempt_ev_crit __cacheline_aligned; 137 static struct evcnt kpreempt_ev_klock __cacheline_aligned; 138 static struct evcnt kpreempt_ev_immed __cacheline_aligned; 139 140 /* 141 * During autoconfiguration or after a panic, a sleep will simply lower the 142 * priority briefly to allow interrupts, then return. The priority to be 143 * used (safepri) is machine-dependent, thus this value is initialized and 144 * maintained in the machine-dependent layers. This priority will typically 145 * be 0, or the lowest priority that is safe for use on the interrupt stack; 146 * it can be made higher to block network software interrupts after panics. 147 */ 148 #ifdef IPL_SAFEPRI 149 int safepri = IPL_SAFEPRI; 150 #else 151 int safepri; 152 #endif 153 154 void 155 synch_init(void) 156 { 157 158 cv_init(&lbolt, "lbolt"); 159 160 evcnt_attach_dynamic(&kpreempt_ev_crit, EVCNT_TYPE_MISC, NULL, 161 "kpreempt", "defer: critical section"); 162 evcnt_attach_dynamic(&kpreempt_ev_klock, EVCNT_TYPE_MISC, NULL, 163 "kpreempt", "defer: kernel_lock"); 164 evcnt_attach_dynamic(&kpreempt_ev_immed, EVCNT_TYPE_MISC, NULL, 165 "kpreempt", "immediate"); 166 } 167 168 /* 169 * OBSOLETE INTERFACE 170 * 171 * General sleep call. Suspends the current LWP until a wakeup is 172 * performed on the specified identifier. The LWP will then be made 173 * runnable with the specified priority. Sleeps at most timo/hz seconds (0 174 * means no timeout). If pri includes PCATCH flag, signals are checked 175 * before and after sleeping, else signals are not checked. Returns 0 if 176 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 177 * signal needs to be delivered, ERESTART is returned if the current system 178 * call should be restarted if possible, and EINTR is returned if the system 179 * call should be interrupted by the signal (return EINTR). 180 */ 181 int 182 tsleep(wchan_t ident, pri_t priority, const char *wmesg, int timo) 183 { 184 struct lwp *l = curlwp; 185 sleepq_t *sq; 186 kmutex_t *mp; 187 188 KASSERT((l->l_pflag & LP_INTR) == 0); 189 KASSERT(ident != &lbolt); 190 191 if (sleepq_dontsleep(l)) { 192 (void)sleepq_abort(NULL, 0); 193 return 0; 194 } 195 196 l->l_kpriority = true; 197 sq = sleeptab_lookup(&sleeptab, ident, &mp); 198 sleepq_enter(sq, l, mp); 199 sleepq_enqueue(sq, ident, wmesg, &sleep_syncobj); 200 return sleepq_block(timo, priority & PCATCH); 201 } 202 203 int 204 mtsleep(wchan_t ident, pri_t priority, const char *wmesg, int timo, 205 kmutex_t *mtx) 206 { 207 struct lwp *l = curlwp; 208 sleepq_t *sq; 209 kmutex_t *mp; 210 int error; 211 212 KASSERT((l->l_pflag & LP_INTR) == 0); 213 KASSERT(ident != &lbolt); 214 215 if (sleepq_dontsleep(l)) { 216 (void)sleepq_abort(mtx, (priority & PNORELOCK) != 0); 217 return 0; 218 } 219 220 l->l_kpriority = true; 221 sq = sleeptab_lookup(&sleeptab, ident, &mp); 222 sleepq_enter(sq, l, mp); 223 sleepq_enqueue(sq, ident, wmesg, &sleep_syncobj); 224 mutex_exit(mtx); 225 error = sleepq_block(timo, priority & PCATCH); 226 227 if ((priority & PNORELOCK) == 0) 228 mutex_enter(mtx); 229 230 return error; 231 } 232 233 /* 234 * General sleep call for situations where a wake-up is not expected. 235 */ 236 int 237 kpause(const char *wmesg, bool intr, int timo, kmutex_t *mtx) 238 { 239 struct lwp *l = curlwp; 240 kmutex_t *mp; 241 sleepq_t *sq; 242 int error; 243 244 KASSERT(!(timo == 0 && intr == false)); 245 246 if (sleepq_dontsleep(l)) 247 return sleepq_abort(NULL, 0); 248 249 if (mtx != NULL) 250 mutex_exit(mtx); 251 l->l_kpriority = true; 252 sq = sleeptab_lookup(&sleeptab, l, &mp); 253 sleepq_enter(sq, l, mp); 254 sleepq_enqueue(sq, l, wmesg, &sleep_syncobj); 255 error = sleepq_block(timo, intr); 256 if (mtx != NULL) 257 mutex_enter(mtx); 258 259 return error; 260 } 261 262 /* 263 * OBSOLETE INTERFACE 264 * 265 * Make all LWPs sleeping on the specified identifier runnable. 266 */ 267 void 268 wakeup(wchan_t ident) 269 { 270 sleepq_t *sq; 271 kmutex_t *mp; 272 273 if (__predict_false(cold)) 274 return; 275 276 sq = sleeptab_lookup(&sleeptab, ident, &mp); 277 sleepq_wake(sq, ident, (u_int)-1, mp); 278 } 279 280 /* 281 * General yield call. Puts the current LWP back on its run queue and 282 * performs a voluntary context switch. Should only be called when the 283 * current LWP explicitly requests it (eg sched_yield(2)). 284 */ 285 void 286 yield(void) 287 { 288 struct lwp *l = curlwp; 289 290 KERNEL_UNLOCK_ALL(l, &l->l_biglocks); 291 lwp_lock(l); 292 KASSERT(lwp_locked(l, l->l_cpu->ci_schedstate.spc_lwplock)); 293 KASSERT(l->l_stat == LSONPROC); 294 l->l_kpriority = false; 295 (void)mi_switch(l); 296 KERNEL_LOCK(l->l_biglocks, l); 297 } 298 299 /* 300 * General preemption call. Puts the current LWP back on its run queue 301 * and performs an involuntary context switch. 302 */ 303 void 304 preempt(void) 305 { 306 struct lwp *l = curlwp; 307 308 KERNEL_UNLOCK_ALL(l, &l->l_biglocks); 309 lwp_lock(l); 310 KASSERT(lwp_locked(l, l->l_cpu->ci_schedstate.spc_lwplock)); 311 KASSERT(l->l_stat == LSONPROC); 312 l->l_kpriority = false; 313 l->l_nivcsw++; 314 (void)mi_switch(l); 315 KERNEL_LOCK(l->l_biglocks, l); 316 } 317 318 /* 319 * Handle a request made by another agent to preempt the current LWP 320 * in-kernel. Usually called when l_dopreempt may be non-zero. 321 * 322 * Character addresses for lockstat only. 323 */ 324 static char in_critical_section; 325 static char kernel_lock_held; 326 static char is_softint; 327 static char cpu_kpreempt_enter_fail; 328 329 bool 330 kpreempt(uintptr_t where) 331 { 332 uintptr_t failed; 333 lwp_t *l; 334 int s, dop, lsflag; 335 336 l = curlwp; 337 failed = 0; 338 while ((dop = l->l_dopreempt) != 0) { 339 if (l->l_stat != LSONPROC) { 340 /* 341 * About to block (or die), let it happen. 342 * Doesn't really count as "preemption has 343 * been blocked", since we're going to 344 * context switch. 345 */ 346 l->l_dopreempt = 0; 347 return true; 348 } 349 if (__predict_false((l->l_flag & LW_IDLE) != 0)) { 350 /* Can't preempt idle loop, don't count as failure. */ 351 l->l_dopreempt = 0; 352 return true; 353 } 354 if (__predict_false(l->l_nopreempt != 0)) { 355 /* LWP holds preemption disabled, explicitly. */ 356 if ((dop & DOPREEMPT_COUNTED) == 0) { 357 kpreempt_ev_crit.ev_count++; 358 } 359 failed = (uintptr_t)&in_critical_section; 360 break; 361 } 362 if (__predict_false((l->l_pflag & LP_INTR) != 0)) { 363 /* Can't preempt soft interrupts yet. */ 364 l->l_dopreempt = 0; 365 failed = (uintptr_t)&is_softint; 366 break; 367 } 368 s = splsched(); 369 if (__predict_false(l->l_blcnt != 0 || 370 curcpu()->ci_biglock_wanted != NULL)) { 371 /* Hold or want kernel_lock, code is not MT safe. */ 372 splx(s); 373 if ((dop & DOPREEMPT_COUNTED) == 0) { 374 kpreempt_ev_klock.ev_count++; 375 } 376 failed = (uintptr_t)&kernel_lock_held; 377 break; 378 } 379 if (__predict_false(!cpu_kpreempt_enter(where, s))) { 380 /* 381 * It may be that the IPL is too high. 382 * kpreempt_enter() can schedule an 383 * interrupt to retry later. 384 */ 385 splx(s); 386 failed = (uintptr_t)&cpu_kpreempt_enter_fail; 387 break; 388 } 389 /* Do it! */ 390 if (__predict_true((dop & DOPREEMPT_COUNTED) == 0)) { 391 kpreempt_ev_immed.ev_count++; 392 } 393 lwp_lock(l); 394 mi_switch(l); 395 l->l_nopreempt++; 396 splx(s); 397 398 /* Take care of any MD cleanup. */ 399 cpu_kpreempt_exit(where); 400 l->l_nopreempt--; 401 } 402 403 if (__predict_true(!failed)) { 404 return false; 405 } 406 407 /* Record preemption failure for reporting via lockstat. */ 408 atomic_or_uint(&l->l_dopreempt, DOPREEMPT_COUNTED); 409 lsflag = 0; 410 LOCKSTAT_ENTER(lsflag); 411 if (__predict_false(lsflag)) { 412 if (where == 0) { 413 where = (uintptr_t)__builtin_return_address(0); 414 } 415 /* Preemption is on, might recurse, so make it atomic. */ 416 if (atomic_cas_ptr_ni((void *)&l->l_pfailaddr, NULL, 417 (void *)where) == NULL) { 418 LOCKSTAT_START_TIMER(lsflag, l->l_pfailtime); 419 l->l_pfaillock = failed; 420 } 421 } 422 LOCKSTAT_EXIT(lsflag); 423 return true; 424 } 425 426 /* 427 * Return true if preemption is explicitly disabled. 428 */ 429 bool 430 kpreempt_disabled(void) 431 { 432 const lwp_t *l = curlwp; 433 434 return l->l_nopreempt != 0 || l->l_stat == LSZOMB || 435 (l->l_flag & LW_IDLE) != 0 || cpu_kpreempt_disabled(); 436 } 437 438 /* 439 * Disable kernel preemption. 440 */ 441 void 442 kpreempt_disable(void) 443 { 444 445 KPREEMPT_DISABLE(curlwp); 446 } 447 448 /* 449 * Reenable kernel preemption. 450 */ 451 void 452 kpreempt_enable(void) 453 { 454 455 KPREEMPT_ENABLE(curlwp); 456 } 457 458 /* 459 * Compute the amount of time during which the current lwp was running. 460 * 461 * - update l_rtime unless it's an idle lwp. 462 */ 463 464 void 465 updatertime(lwp_t *l, const struct bintime *now) 466 { 467 468 if (__predict_false(l->l_flag & LW_IDLE)) 469 return; 470 471 /* rtime += now - stime */ 472 bintime_add(&l->l_rtime, now); 473 bintime_sub(&l->l_rtime, &l->l_stime); 474 } 475 476 /* 477 * Select next LWP from the current CPU to run.. 478 */ 479 static inline lwp_t * 480 nextlwp(struct cpu_info *ci, struct schedstate_percpu *spc) 481 { 482 lwp_t *newl; 483 484 /* 485 * Let sched_nextlwp() select the LWP to run the CPU next. 486 * If no LWP is runnable, select the idle LWP. 487 * 488 * Note that spc_lwplock might not necessary be held, and 489 * new thread would be unlocked after setting the LWP-lock. 490 */ 491 newl = sched_nextlwp(); 492 if (newl != NULL) { 493 sched_dequeue(newl); 494 KASSERT(lwp_locked(newl, spc->spc_mutex)); 495 KASSERT(newl->l_cpu == ci); 496 newl->l_stat = LSONPROC; 497 newl->l_pflag |= LP_RUNNING; 498 lwp_setlock(newl, spc->spc_lwplock); 499 } else { 500 newl = ci->ci_data.cpu_idlelwp; 501 newl->l_stat = LSONPROC; 502 newl->l_pflag |= LP_RUNNING; 503 } 504 505 /* 506 * Only clear want_resched if there are no pending (slow) 507 * software interrupts. 508 */ 509 ci->ci_want_resched = ci->ci_data.cpu_softints; 510 spc->spc_flags &= ~SPCF_SWITCHCLEAR; 511 spc->spc_curpriority = lwp_eprio(newl); 512 513 return newl; 514 } 515 516 /* 517 * The machine independent parts of context switch. 518 * 519 * Returns 1 if another LWP was actually run. 520 */ 521 int 522 mi_switch(lwp_t *l) 523 { 524 struct cpu_info *ci; 525 struct schedstate_percpu *spc; 526 struct lwp *newl; 527 int retval, oldspl; 528 struct bintime bt; 529 bool returning; 530 531 KASSERT(lwp_locked(l, NULL)); 532 KASSERT(kpreempt_disabled()); 533 LOCKDEBUG_BARRIER(l->l_mutex, 1); 534 535 kstack_check_magic(l); 536 537 binuptime(&bt); 538 539 KASSERT((l->l_pflag & LP_RUNNING) != 0); 540 KASSERT(l->l_cpu == curcpu()); 541 ci = l->l_cpu; 542 spc = &ci->ci_schedstate; 543 returning = false; 544 newl = NULL; 545 546 /* 547 * If we have been asked to switch to a specific LWP, then there 548 * is no need to inspect the run queues. If a soft interrupt is 549 * blocking, then return to the interrupted thread without adjusting 550 * VM context or its start time: neither have been changed in order 551 * to take the interrupt. 552 */ 553 if (l->l_switchto != NULL) { 554 if ((l->l_pflag & LP_INTR) != 0) { 555 returning = true; 556 softint_block(l); 557 if ((l->l_pflag & LP_TIMEINTR) != 0) 558 updatertime(l, &bt); 559 } 560 newl = l->l_switchto; 561 l->l_switchto = NULL; 562 } 563 #ifndef __HAVE_FAST_SOFTINTS 564 else if (ci->ci_data.cpu_softints != 0) { 565 /* There are pending soft interrupts, so pick one. */ 566 newl = softint_picklwp(); 567 newl->l_stat = LSONPROC; 568 newl->l_pflag |= LP_RUNNING; 569 } 570 #endif /* !__HAVE_FAST_SOFTINTS */ 571 572 /* Count time spent in current system call */ 573 if (!returning) { 574 SYSCALL_TIME_SLEEP(l); 575 576 /* 577 * XXXSMP If we are using h/w performance counters, 578 * save context. 579 */ 580 #if PERFCTRS 581 if (PMC_ENABLED(l->l_proc)) { 582 pmc_save_context(l->l_proc); 583 } 584 #endif 585 updatertime(l, &bt); 586 } 587 588 /* Lock the runqueue */ 589 KASSERT(l->l_stat != LSRUN); 590 mutex_spin_enter(spc->spc_mutex); 591 592 /* 593 * If on the CPU and we have gotten this far, then we must yield. 594 */ 595 if (l->l_stat == LSONPROC && l != newl) { 596 KASSERT(lwp_locked(l, spc->spc_lwplock)); 597 if ((l->l_flag & LW_IDLE) == 0) { 598 l->l_stat = LSRUN; 599 lwp_setlock(l, spc->spc_mutex); 600 sched_enqueue(l, true); 601 /* 602 * Handle migration. Note that "migrating LWP" may 603 * be reset here, if interrupt/preemption happens 604 * early in idle LWP. 605 */ 606 if (l->l_target_cpu != NULL) { 607 KASSERT((l->l_pflag & LP_INTR) == 0); 608 spc->spc_migrating = l; 609 } 610 } else 611 l->l_stat = LSIDL; 612 } 613 614 /* Pick new LWP to run. */ 615 if (newl == NULL) { 616 newl = nextlwp(ci, spc); 617 } 618 619 /* Items that must be updated with the CPU locked. */ 620 if (!returning) { 621 /* Update the new LWP's start time. */ 622 newl->l_stime = bt; 623 624 /* 625 * ci_curlwp changes when a fast soft interrupt occurs. 626 * We use cpu_onproc to keep track of which kernel or 627 * user thread is running 'underneath' the software 628 * interrupt. This is important for time accounting, 629 * itimers and forcing user threads to preempt (aston). 630 */ 631 ci->ci_data.cpu_onproc = newl; 632 } 633 634 /* 635 * Preemption related tasks. Must be done with the current 636 * CPU locked. 637 */ 638 cpu_did_resched(l); 639 l->l_dopreempt = 0; 640 if (__predict_false(l->l_pfailaddr != 0)) { 641 LOCKSTAT_FLAG(lsflag); 642 LOCKSTAT_ENTER(lsflag); 643 LOCKSTAT_STOP_TIMER(lsflag, l->l_pfailtime); 644 LOCKSTAT_EVENT_RA(lsflag, l->l_pfaillock, LB_NOPREEMPT|LB_SPIN, 645 1, l->l_pfailtime, l->l_pfailaddr); 646 LOCKSTAT_EXIT(lsflag); 647 l->l_pfailtime = 0; 648 l->l_pfaillock = 0; 649 l->l_pfailaddr = 0; 650 } 651 652 if (l != newl) { 653 struct lwp *prevlwp; 654 655 /* Release all locks, but leave the current LWP locked */ 656 if (l->l_mutex == spc->spc_mutex) { 657 /* 658 * Drop spc_lwplock, if the current LWP has been moved 659 * to the run queue (it is now locked by spc_mutex). 660 */ 661 mutex_spin_exit(spc->spc_lwplock); 662 } else { 663 /* 664 * Otherwise, drop the spc_mutex, we are done with the 665 * run queues. 666 */ 667 mutex_spin_exit(spc->spc_mutex); 668 } 669 670 /* 671 * Mark that context switch is going to be performed 672 * for this LWP, to protect it from being switched 673 * to on another CPU. 674 */ 675 KASSERT(l->l_ctxswtch == 0); 676 l->l_ctxswtch = 1; 677 l->l_ncsw++; 678 KASSERT((l->l_pflag & LP_RUNNING) != 0); 679 l->l_pflag &= ~LP_RUNNING; 680 681 /* 682 * Increase the count of spin-mutexes before the release 683 * of the last lock - we must remain at IPL_SCHED during 684 * the context switch. 685 */ 686 KASSERTMSG(ci->ci_mtx_count == -1, 687 "%s: cpu%u: ci_mtx_count (%d) != -1", 688 __func__, cpu_index(ci), ci->ci_mtx_count); 689 oldspl = MUTEX_SPIN_OLDSPL(ci); 690 ci->ci_mtx_count--; 691 lwp_unlock(l); 692 693 /* Count the context switch on this CPU. */ 694 ci->ci_data.cpu_nswtch++; 695 696 /* Update status for lwpctl, if present. */ 697 if (l->l_lwpctl != NULL) 698 l->l_lwpctl->lc_curcpu = LWPCTL_CPU_NONE; 699 700 /* 701 * Save old VM context, unless a soft interrupt 702 * handler is blocking. 703 */ 704 if (!returning) 705 pmap_deactivate(l); 706 707 /* 708 * We may need to spin-wait if 'newl' is still 709 * context switching on another CPU. 710 */ 711 if (__predict_false(newl->l_ctxswtch != 0)) { 712 u_int count; 713 count = SPINLOCK_BACKOFF_MIN; 714 while (newl->l_ctxswtch) 715 SPINLOCK_BACKOFF(count); 716 } 717 718 /* 719 * If DTrace has set the active vtime enum to anything 720 * other than INACTIVE (0), then it should have set the 721 * function to call. 722 */ 723 if (__predict_false(dtrace_vtime_active)) { 724 (*dtrace_vtime_switch_func)(newl); 725 } 726 727 /* Switch to the new LWP.. */ 728 prevlwp = cpu_switchto(l, newl, returning); 729 ci = curcpu(); 730 731 /* 732 * Switched away - we have new curlwp. 733 * Restore VM context and IPL. 734 */ 735 pmap_activate(l); 736 uvm_emap_switch(l); 737 pcu_switchpoint(l); 738 739 if (prevlwp != NULL) { 740 /* Normalize the count of the spin-mutexes */ 741 ci->ci_mtx_count++; 742 /* Unmark the state of context switch */ 743 membar_exit(); 744 prevlwp->l_ctxswtch = 0; 745 } 746 747 /* Update status for lwpctl, if present. */ 748 if (l->l_lwpctl != NULL) { 749 l->l_lwpctl->lc_curcpu = (int)cpu_index(ci); 750 l->l_lwpctl->lc_pctr++; 751 } 752 753 /* Note trip through cpu_switchto(). */ 754 pserialize_switchpoint(); 755 756 KASSERT(l->l_cpu == ci); 757 splx(oldspl); 758 retval = 1; 759 } else { 760 /* Nothing to do - just unlock and return. */ 761 mutex_spin_exit(spc->spc_mutex); 762 lwp_unlock(l); 763 retval = 0; 764 } 765 766 KASSERT(l == curlwp); 767 KASSERT(l->l_stat == LSONPROC); 768 769 /* 770 * XXXSMP If we are using h/w performance counters, restore context. 771 * XXXSMP preemption problem. 772 */ 773 #if PERFCTRS 774 if (PMC_ENABLED(l->l_proc)) { 775 pmc_restore_context(l->l_proc); 776 } 777 #endif 778 SYSCALL_TIME_WAKEUP(l); 779 LOCKDEBUG_BARRIER(NULL, 1); 780 781 return retval; 782 } 783 784 /* 785 * The machine independent parts of context switch to oblivion. 786 * Does not return. Call with the LWP unlocked. 787 */ 788 void 789 lwp_exit_switchaway(lwp_t *l) 790 { 791 struct cpu_info *ci; 792 struct lwp *newl; 793 struct bintime bt; 794 795 ci = l->l_cpu; 796 797 KASSERT(kpreempt_disabled()); 798 KASSERT(l->l_stat == LSZOMB || l->l_stat == LSIDL); 799 KASSERT(ci == curcpu()); 800 LOCKDEBUG_BARRIER(NULL, 0); 801 802 kstack_check_magic(l); 803 804 /* Count time spent in current system call */ 805 SYSCALL_TIME_SLEEP(l); 806 binuptime(&bt); 807 updatertime(l, &bt); 808 809 /* Must stay at IPL_SCHED even after releasing run queue lock. */ 810 (void)splsched(); 811 812 /* 813 * Let sched_nextlwp() select the LWP to run the CPU next. 814 * If no LWP is runnable, select the idle LWP. 815 * 816 * Note that spc_lwplock might not necessary be held, and 817 * new thread would be unlocked after setting the LWP-lock. 818 */ 819 spc_lock(ci); 820 #ifndef __HAVE_FAST_SOFTINTS 821 if (ci->ci_data.cpu_softints != 0) { 822 /* There are pending soft interrupts, so pick one. */ 823 newl = softint_picklwp(); 824 newl->l_stat = LSONPROC; 825 newl->l_pflag |= LP_RUNNING; 826 } else 827 #endif /* !__HAVE_FAST_SOFTINTS */ 828 { 829 newl = nextlwp(ci, &ci->ci_schedstate); 830 } 831 832 /* Update the new LWP's start time. */ 833 newl->l_stime = bt; 834 l->l_pflag &= ~LP_RUNNING; 835 836 /* 837 * ci_curlwp changes when a fast soft interrupt occurs. 838 * We use cpu_onproc to keep track of which kernel or 839 * user thread is running 'underneath' the software 840 * interrupt. This is important for time accounting, 841 * itimers and forcing user threads to preempt (aston). 842 */ 843 ci->ci_data.cpu_onproc = newl; 844 845 /* 846 * Preemption related tasks. Must be done with the current 847 * CPU locked. 848 */ 849 cpu_did_resched(l); 850 851 /* Unlock the run queue. */ 852 spc_unlock(ci); 853 854 /* Count the context switch on this CPU. */ 855 ci->ci_data.cpu_nswtch++; 856 857 /* Update status for lwpctl, if present. */ 858 if (l->l_lwpctl != NULL) 859 l->l_lwpctl->lc_curcpu = LWPCTL_CPU_EXITED; 860 861 /* 862 * We may need to spin-wait if 'newl' is still 863 * context switching on another CPU. 864 */ 865 if (__predict_false(newl->l_ctxswtch != 0)) { 866 u_int count; 867 count = SPINLOCK_BACKOFF_MIN; 868 while (newl->l_ctxswtch) 869 SPINLOCK_BACKOFF(count); 870 } 871 872 /* 873 * If DTrace has set the active vtime enum to anything 874 * other than INACTIVE (0), then it should have set the 875 * function to call. 876 */ 877 if (__predict_false(dtrace_vtime_active)) { 878 (*dtrace_vtime_switch_func)(newl); 879 } 880 881 /* Switch to the new LWP.. */ 882 (void)cpu_switchto(NULL, newl, false); 883 884 for (;;) continue; /* XXX: convince gcc about "noreturn" */ 885 /* NOTREACHED */ 886 } 887 888 /* 889 * setrunnable: change LWP state to be runnable, placing it on the run queue. 890 * 891 * Call with the process and LWP locked. Will return with the LWP unlocked. 892 */ 893 void 894 setrunnable(struct lwp *l) 895 { 896 struct proc *p = l->l_proc; 897 struct cpu_info *ci; 898 899 KASSERT((l->l_flag & LW_IDLE) == 0); 900 KASSERT(mutex_owned(p->p_lock)); 901 KASSERT(lwp_locked(l, NULL)); 902 KASSERT(l->l_mutex != l->l_cpu->ci_schedstate.spc_mutex); 903 904 switch (l->l_stat) { 905 case LSSTOP: 906 /* 907 * If we're being traced (possibly because someone attached us 908 * while we were stopped), check for a signal from the debugger. 909 */ 910 if ((p->p_slflag & PSL_TRACED) != 0 && p->p_xstat != 0) 911 signotify(l); 912 p->p_nrlwps++; 913 break; 914 case LSSUSPENDED: 915 l->l_flag &= ~LW_WSUSPEND; 916 p->p_nrlwps++; 917 cv_broadcast(&p->p_lwpcv); 918 break; 919 case LSSLEEP: 920 KASSERT(l->l_wchan != NULL); 921 break; 922 default: 923 panic("setrunnable: lwp %p state was %d", l, l->l_stat); 924 } 925 926 /* 927 * If the LWP was sleeping, start it again. 928 */ 929 if (l->l_wchan != NULL) { 930 l->l_stat = LSSLEEP; 931 /* lwp_unsleep() will release the lock. */ 932 lwp_unsleep(l, true); 933 return; 934 } 935 936 /* 937 * If the LWP is still on the CPU, mark it as LSONPROC. It may be 938 * about to call mi_switch(), in which case it will yield. 939 */ 940 if ((l->l_pflag & LP_RUNNING) != 0) { 941 l->l_stat = LSONPROC; 942 l->l_slptime = 0; 943 lwp_unlock(l); 944 return; 945 } 946 947 /* 948 * Look for a CPU to run. 949 * Set the LWP runnable. 950 */ 951 ci = sched_takecpu(l); 952 l->l_cpu = ci; 953 spc_lock(ci); 954 lwp_unlock_to(l, ci->ci_schedstate.spc_mutex); 955 sched_setrunnable(l); 956 l->l_stat = LSRUN; 957 l->l_slptime = 0; 958 959 sched_enqueue(l, false); 960 resched_cpu(l); 961 lwp_unlock(l); 962 } 963 964 /* 965 * suspendsched: 966 * 967 * Convert all non-LW_SYSTEM LSSLEEP or LSRUN LWPs to LSSUSPENDED. 968 */ 969 void 970 suspendsched(void) 971 { 972 CPU_INFO_ITERATOR cii; 973 struct cpu_info *ci; 974 struct lwp *l; 975 struct proc *p; 976 977 /* 978 * We do this by process in order not to violate the locking rules. 979 */ 980 mutex_enter(proc_lock); 981 PROCLIST_FOREACH(p, &allproc) { 982 mutex_enter(p->p_lock); 983 if ((p->p_flag & PK_SYSTEM) != 0) { 984 mutex_exit(p->p_lock); 985 continue; 986 } 987 988 p->p_stat = SSTOP; 989 990 LIST_FOREACH(l, &p->p_lwps, l_sibling) { 991 if (l == curlwp) 992 continue; 993 994 lwp_lock(l); 995 996 /* 997 * Set L_WREBOOT so that the LWP will suspend itself 998 * when it tries to return to user mode. We want to 999 * try and get to get as many LWPs as possible to 1000 * the user / kernel boundary, so that they will 1001 * release any locks that they hold. 1002 */ 1003 l->l_flag |= (LW_WREBOOT | LW_WSUSPEND); 1004 1005 if (l->l_stat == LSSLEEP && 1006 (l->l_flag & LW_SINTR) != 0) { 1007 /* setrunnable() will release the lock. */ 1008 setrunnable(l); 1009 continue; 1010 } 1011 1012 lwp_unlock(l); 1013 } 1014 1015 mutex_exit(p->p_lock); 1016 } 1017 mutex_exit(proc_lock); 1018 1019 /* 1020 * Kick all CPUs to make them preempt any LWPs running in user mode. 1021 * They'll trap into the kernel and suspend themselves in userret(). 1022 */ 1023 for (CPU_INFO_FOREACH(cii, ci)) { 1024 spc_lock(ci); 1025 cpu_need_resched(ci, RESCHED_IMMED); 1026 spc_unlock(ci); 1027 } 1028 } 1029 1030 /* 1031 * sched_unsleep: 1032 * 1033 * The is called when the LWP has not been awoken normally but instead 1034 * interrupted: for example, if the sleep timed out. Because of this, 1035 * it's not a valid action for running or idle LWPs. 1036 */ 1037 static void 1038 sched_unsleep(struct lwp *l, bool cleanup) 1039 { 1040 1041 lwp_unlock(l); 1042 panic("sched_unsleep"); 1043 } 1044 1045 static void 1046 resched_cpu(struct lwp *l) 1047 { 1048 struct cpu_info *ci = l->l_cpu; 1049 1050 KASSERT(lwp_locked(l, NULL)); 1051 if (lwp_eprio(l) > ci->ci_schedstate.spc_curpriority) 1052 cpu_need_resched(ci, 0); 1053 } 1054 1055 static void 1056 sched_changepri(struct lwp *l, pri_t pri) 1057 { 1058 1059 KASSERT(lwp_locked(l, NULL)); 1060 1061 if (l->l_stat == LSRUN) { 1062 KASSERT(lwp_locked(l, l->l_cpu->ci_schedstate.spc_mutex)); 1063 sched_dequeue(l); 1064 l->l_priority = pri; 1065 sched_enqueue(l, false); 1066 } else { 1067 l->l_priority = pri; 1068 } 1069 resched_cpu(l); 1070 } 1071 1072 static void 1073 sched_lendpri(struct lwp *l, pri_t pri) 1074 { 1075 1076 KASSERT(lwp_locked(l, NULL)); 1077 1078 if (l->l_stat == LSRUN) { 1079 KASSERT(lwp_locked(l, l->l_cpu->ci_schedstate.spc_mutex)); 1080 sched_dequeue(l); 1081 l->l_inheritedprio = pri; 1082 sched_enqueue(l, false); 1083 } else { 1084 l->l_inheritedprio = pri; 1085 } 1086 resched_cpu(l); 1087 } 1088 1089 struct lwp * 1090 syncobj_noowner(wchan_t wchan) 1091 { 1092 1093 return NULL; 1094 } 1095 1096 /* Decay 95% of proc::p_pctcpu in 60 seconds, ccpu = exp(-1/20) */ 1097 const fixpt_t ccpu = 0.95122942450071400909 * FSCALE; 1098 1099 /* 1100 * Constants for averages over 1, 5 and 15 minutes when sampling at 1101 * 5 second intervals. 1102 */ 1103 static const fixpt_t cexp[ ] = { 1104 0.9200444146293232 * FSCALE, /* exp(-1/12) */ 1105 0.9834714538216174 * FSCALE, /* exp(-1/60) */ 1106 0.9944598480048967 * FSCALE, /* exp(-1/180) */ 1107 }; 1108 1109 /* 1110 * sched_pstats: 1111 * 1112 * => Update process statistics and check CPU resource allocation. 1113 * => Call scheduler-specific hook to eventually adjust LWP priorities. 1114 * => Compute load average of a quantity on 1, 5 and 15 minute intervals. 1115 */ 1116 void 1117 sched_pstats(void) 1118 { 1119 extern struct loadavg averunnable; 1120 struct loadavg *avg = &averunnable; 1121 const int clkhz = (stathz != 0 ? stathz : hz); 1122 static bool backwards = false; 1123 static u_int lavg_count = 0; 1124 struct proc *p; 1125 int nrun; 1126 1127 sched_pstats_ticks++; 1128 if (++lavg_count >= 5) { 1129 lavg_count = 0; 1130 nrun = 0; 1131 } 1132 mutex_enter(proc_lock); 1133 PROCLIST_FOREACH(p, &allproc) { 1134 struct lwp *l; 1135 struct rlimit *rlim; 1136 time_t runtm; 1137 int sig; 1138 1139 /* Increment sleep time (if sleeping), ignore overflow. */ 1140 mutex_enter(p->p_lock); 1141 runtm = p->p_rtime.sec; 1142 LIST_FOREACH(l, &p->p_lwps, l_sibling) { 1143 fixpt_t lpctcpu; 1144 u_int lcpticks; 1145 1146 if (__predict_false((l->l_flag & LW_IDLE) != 0)) 1147 continue; 1148 lwp_lock(l); 1149 runtm += l->l_rtime.sec; 1150 l->l_swtime++; 1151 sched_lwp_stats(l); 1152 1153 /* For load average calculation. */ 1154 if (__predict_false(lavg_count == 0) && 1155 (l->l_flag & (LW_SINTR | LW_SYSTEM)) == 0) { 1156 switch (l->l_stat) { 1157 case LSSLEEP: 1158 if (l->l_slptime > 1) { 1159 break; 1160 } 1161 case LSRUN: 1162 case LSONPROC: 1163 case LSIDL: 1164 nrun++; 1165 } 1166 } 1167 lwp_unlock(l); 1168 1169 l->l_pctcpu = (l->l_pctcpu * ccpu) >> FSHIFT; 1170 if (l->l_slptime != 0) 1171 continue; 1172 1173 lpctcpu = l->l_pctcpu; 1174 lcpticks = atomic_swap_uint(&l->l_cpticks, 0); 1175 lpctcpu += ((FSCALE - ccpu) * 1176 (lcpticks * FSCALE / clkhz)) >> FSHIFT; 1177 l->l_pctcpu = lpctcpu; 1178 } 1179 /* Calculating p_pctcpu only for ps(1) */ 1180 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 1181 1182 /* 1183 * Check if the process exceeds its CPU resource allocation. 1184 * If over the hard limit, kill it with SIGKILL. 1185 * If over the soft limit, send SIGXCPU and raise 1186 * the soft limit a little. 1187 */ 1188 rlim = &p->p_rlimit[RLIMIT_CPU]; 1189 sig = 0; 1190 if (__predict_false(runtm >= rlim->rlim_cur)) { 1191 if (runtm >= rlim->rlim_max) { 1192 sig = SIGKILL; 1193 log(LOG_NOTICE, "pid %d is killed: %s\n", 1194 p->p_pid, "exceeded RLIMIT_CPU"); 1195 uprintf("pid %d, command %s, is killed: %s\n", 1196 p->p_pid, p->p_comm, 1197 "exceeded RLIMIT_CPU"); 1198 } else { 1199 sig = SIGXCPU; 1200 if (rlim->rlim_cur < rlim->rlim_max) 1201 rlim->rlim_cur += 5; 1202 } 1203 } 1204 mutex_exit(p->p_lock); 1205 if (__predict_false(runtm < 0)) { 1206 if (!backwards) { 1207 backwards = true; 1208 printf("WARNING: negative runtime; " 1209 "monotonic clock has gone backwards\n"); 1210 } 1211 } else if (__predict_false(sig)) { 1212 KASSERT((p->p_flag & PK_SYSTEM) == 0); 1213 psignal(p, sig); 1214 } 1215 } 1216 mutex_exit(proc_lock); 1217 1218 /* Load average calculation. */ 1219 if (__predict_false(lavg_count == 0)) { 1220 int i; 1221 CTASSERT(__arraycount(cexp) == __arraycount(avg->ldavg)); 1222 for (i = 0; i < __arraycount(cexp); i++) { 1223 avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + 1224 nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; 1225 } 1226 } 1227 1228 /* Lightning bolt. */ 1229 cv_broadcast(&lbolt); 1230 } 1231