kern_switch.c revision 136170
1/* 2 * Copyright (c) 2001 Jake Burkholder <jake@FreeBSD.org> 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24 * SUCH DAMAGE. 25 */ 26 27/*** 28Here is the logic.. 29 30If there are N processors, then there are at most N KSEs (kernel 31schedulable entities) working to process threads that belong to a 32KSEGROUP (kg). If there are X of these KSEs actually running at the 33moment in question, then there are at most M (N-X) of these KSEs on 34the run queue, as running KSEs are not on the queue. 35 36Runnable threads are queued off the KSEGROUP in priority order. 37If there are M or more threads runnable, the top M threads 38(by priority) are 'preassigned' to the M KSEs not running. The KSEs take 39their priority from those threads and are put on the run queue. 40 41The last thread that had a priority high enough to have a KSE associated 42with it, AND IS ON THE RUN QUEUE is pointed to by 43kg->kg_last_assigned. If no threads queued off the KSEGROUP have KSEs 44assigned as all the available KSEs are activly running, or because there 45are no threads queued, that pointer is NULL. 46 47When a KSE is removed from the run queue to become runnable, we know 48it was associated with the highest priority thread in the queue (at the head 49of the queue). If it is also the last assigned we know M was 1 and must 50now be 0. Since the thread is no longer queued that pointer must be 51removed from it. Since we know there were no more KSEs available, 52(M was 1 and is now 0) and since we are not FREEING our KSE 53but using it, we know there are STILL no more KSEs available, we can prove 54that the next thread in the ksegrp list will not have a KSE to assign to 55it, so we can show that the pointer must be made 'invalid' (NULL). 56 57The pointer exists so that when a new thread is made runnable, it can 58have its priority compared with the last assigned thread to see if 59it should 'steal' its KSE or not.. i.e. is it 'earlier' 60on the list than that thread or later.. If it's earlier, then the KSE is 61removed from the last assigned (which is now not assigned a KSE) 62and reassigned to the new thread, which is placed earlier in the list. 63The pointer is then backed up to the previous thread (which may or may not 64be the new thread). 65 66When a thread sleeps or is removed, the KSE becomes available and if there 67are queued threads that are not assigned KSEs, the highest priority one of 68them is assigned the KSE, which is then placed back on the run queue at 69the approipriate place, and the kg->kg_last_assigned pointer is adjusted down 70to point to it. 71 72The following diagram shows 2 KSEs and 3 threads from a single process. 73 74 RUNQ: --->KSE---KSE--... (KSEs queued at priorities from threads) 75 \ \____ 76 \ \ 77 KSEGROUP---thread--thread--thread (queued in priority order) 78 \ / 79 \_______________/ 80 (last_assigned) 81 82The result of this scheme is that the M available KSEs are always 83queued at the priorities they have inherrited from the M highest priority 84threads for that KSEGROUP. If this situation changes, the KSEs are 85reassigned to keep this true. 86***/ 87 88#include <sys/cdefs.h> 89__FBSDID("$FreeBSD: head/sys/kern/kern_switch.c 136170 2004-10-05 22:03:10Z julian $"); 90 91#include "opt_sched.h" 92 93#ifndef KERN_SWITCH_INCLUDE 94#include <sys/param.h> 95#include <sys/systm.h> 96#include <sys/kdb.h> 97#include <sys/kernel.h> 98#include <sys/ktr.h> 99#include <sys/lock.h> 100#include <sys/mutex.h> 101#include <sys/proc.h> 102#include <sys/queue.h> 103#include <sys/sched.h> 104#else /* KERN_SWITCH_INCLUDE */ 105#if defined(SMP) && (defined(__i386__) || defined(__amd64__)) 106#include <sys/smp.h> 107#endif 108#include <machine/critical.h> 109#if defined(SMP) && defined(SCHED_4BSD) 110#include <sys/sysctl.h> 111#endif 112 113#ifdef FULL_PREEMPTION 114#ifndef PREEMPTION 115#error "The FULL_PREEMPTION option requires the PREEMPTION option" 116#endif 117#endif 118 119CTASSERT((RQB_BPW * RQB_LEN) == RQ_NQS); 120 121#define td_kse td_sched 122 123/************************************************************************ 124 * Functions that manipulate runnability from a thread perspective. * 125 ************************************************************************/ 126/* 127 * Select the KSE that will be run next. From that find the thread, and 128 * remove it from the KSEGRP's run queue. If there is thread clustering, 129 * this will be what does it. 130 */ 131struct thread * 132choosethread(void) 133{ 134 struct kse *ke; 135 struct thread *td; 136 struct ksegrp *kg; 137 138#if defined(SMP) && (defined(__i386__) || defined(__amd64__)) 139 if (smp_active == 0 && PCPU_GET(cpuid) != 0) { 140 /* Shutting down, run idlethread on AP's */ 141 td = PCPU_GET(idlethread); 142 ke = td->td_kse; 143 CTR1(KTR_RUNQ, "choosethread: td=%p (idle)", td); 144 ke->ke_flags |= KEF_DIDRUN; 145 TD_SET_RUNNING(td); 146 return (td); 147 } 148#endif 149 150retry: 151 ke = sched_choose(); 152 if (ke) { 153 td = ke->ke_thread; 154 KASSERT((td->td_kse == ke), ("kse/thread mismatch")); 155 kg = ke->ke_ksegrp; 156 if (td->td_proc->p_flag & P_HADTHREADS) { 157 if (kg->kg_last_assigned == td) { 158 kg->kg_last_assigned = TAILQ_PREV(td, 159 threadqueue, td_runq); 160 } 161 TAILQ_REMOVE(&kg->kg_runq, td, td_runq); 162 kg->kg_runnable--; 163 } 164 CTR2(KTR_RUNQ, "choosethread: td=%p pri=%d", 165 td, td->td_priority); 166 } else { 167 /* Simulate runq_choose() having returned the idle thread */ 168 td = PCPU_GET(idlethread); 169 ke = td->td_kse; 170 CTR1(KTR_RUNQ, "choosethread: td=%p (idle)", td); 171 } 172 ke->ke_flags |= KEF_DIDRUN; 173 174 /* 175 * If we are in panic, only allow system threads, 176 * plus the one we are running in, to be run. 177 */ 178 if (panicstr && ((td->td_proc->p_flag & P_SYSTEM) == 0 && 179 (td->td_flags & TDF_INPANIC) == 0)) { 180 /* note that it is no longer on the run queue */ 181 TD_SET_CAN_RUN(td); 182 goto retry; 183 } 184 185 TD_SET_RUNNING(td); 186 return (td); 187} 188 189/* 190 * Given a surplus system slot, try assign a new runnable thread to it. 191 * Called from: 192 * sched_thread_exit() (local) 193 * sched_switch() (local) 194 * sched_thread_exit() (local) 195 * remrunqueue() (local) (not at the moment) 196 */ 197static void 198slot_fill(struct ksegrp *kg) 199{ 200 struct thread *td; 201 202 mtx_assert(&sched_lock, MA_OWNED); 203 while (kg->kg_avail_opennings > 0) { 204 /* 205 * Find the first unassigned thread 206 */ 207 if ((td = kg->kg_last_assigned) != NULL) 208 td = TAILQ_NEXT(td, td_runq); 209 else 210 td = TAILQ_FIRST(&kg->kg_runq); 211 212 /* 213 * If we found one, send it to the system scheduler. 214 */ 215 if (td) { 216 kg->kg_last_assigned = td; 217 sched_add(td, SRQ_BORING); 218 CTR2(KTR_RUNQ, "slot_fill: td%p -> kg%p", td, kg); 219 } else { 220 /* no threads to use up the slots. quit now */ 221 break; 222 } 223 } 224} 225 226#ifdef SCHED_4BSD 227/* 228 * Remove a thread from its KSEGRP's run queue. 229 * This in turn may remove it from a KSE if it was already assigned 230 * to one, possibly causing a new thread to be assigned to the KSE 231 * and the KSE getting a new priority. 232 */ 233static void 234remrunqueue(struct thread *td) 235{ 236 struct thread *td2, *td3; 237 struct ksegrp *kg; 238 struct kse *ke; 239 240 mtx_assert(&sched_lock, MA_OWNED); 241 KASSERT((TD_ON_RUNQ(td)), ("remrunqueue: Bad state on run queue")); 242 kg = td->td_ksegrp; 243 ke = td->td_kse; 244 CTR1(KTR_RUNQ, "remrunqueue: td%p", td); 245 TD_SET_CAN_RUN(td); 246 /* 247 * If it is not a threaded process, take the shortcut. 248 */ 249 if ((td->td_proc->p_flag & P_HADTHREADS) == 0) { 250 /* remve from sys run queue and free up a slot */ 251 sched_rem(td); 252 ke->ke_state = KES_THREAD; 253 return; 254 } 255 td3 = TAILQ_PREV(td, threadqueue, td_runq); 256 TAILQ_REMOVE(&kg->kg_runq, td, td_runq); 257 kg->kg_runnable--; 258 if (ke->ke_state == KES_ONRUNQ) { 259 /* 260 * This thread has been assigned to the system run queue. 261 * We need to dissociate it and try assign the 262 * KSE to the next available thread. Then, we should 263 * see if we need to move the KSE in the run queues. 264 */ 265 sched_rem(td); 266 ke->ke_state = KES_THREAD; 267 td2 = kg->kg_last_assigned; 268 KASSERT((td2 != NULL), ("last assigned has wrong value")); 269 if (td2 == td) 270 kg->kg_last_assigned = td3; 271 /* slot_fill(kg); */ /* will replace it with another */ 272 } 273} 274#endif 275 276/* 277 * Change the priority of a thread that is on the run queue. 278 */ 279void 280adjustrunqueue( struct thread *td, int newpri) 281{ 282 struct ksegrp *kg; 283 struct kse *ke; 284 285 mtx_assert(&sched_lock, MA_OWNED); 286 KASSERT((TD_ON_RUNQ(td)), ("adjustrunqueue: Bad state on run queue")); 287 288 ke = td->td_kse; 289 CTR1(KTR_RUNQ, "adjustrunqueue: td%p", td); 290 /* 291 * If it is not a threaded process, take the shortcut. 292 */ 293 if ((td->td_proc->p_flag & P_HADTHREADS) == 0) { 294 /* We only care about the kse in the run queue. */ 295 td->td_priority = newpri; 296 if (ke->ke_rqindex != (newpri / RQ_PPQ)) { 297 sched_rem(td); 298 sched_add(td, SRQ_BORING); 299 } 300 return; 301 } 302 303 /* It is a threaded process */ 304 kg = td->td_ksegrp; 305 if (ke->ke_state == KES_ONRUNQ) { 306 if (kg->kg_last_assigned == td) { 307 kg->kg_last_assigned = 308 TAILQ_PREV(td, threadqueue, td_runq); 309 } 310 sched_rem(td); 311 } 312 TAILQ_REMOVE(&kg->kg_runq, td, td_runq); 313 kg->kg_runnable--; 314 TD_SET_CAN_RUN(td); 315 td->td_priority = newpri; 316 setrunqueue(td, SRQ_BORING); 317} 318int limitcount; 319void 320setrunqueue(struct thread *td, int flags) 321{ 322 struct ksegrp *kg; 323 struct thread *td2; 324 struct thread *tda; 325 326 CTR3(KTR_RUNQ, "setrunqueue: td:%p kg:%p pid:%d", 327 td, td->td_ksegrp, td->td_proc->p_pid); 328 mtx_assert(&sched_lock, MA_OWNED); 329 KASSERT((td->td_inhibitors == 0), 330 ("setrunqueue: trying to run inhibitted thread")); 331 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 332 ("setrunqueue: bad thread state")); 333 TD_SET_RUNQ(td); 334 kg = td->td_ksegrp; 335 if ((td->td_proc->p_flag & P_HADTHREADS) == 0) { 336 /* 337 * Common path optimisation: Only one of everything 338 * and the KSE is always already attached. 339 * Totally ignore the ksegrp run queue. 340 */ 341 if (kg->kg_avail_opennings != 1) { 342 if (limitcount < 1) { 343 limitcount++; 344 printf("pid %d: corrected slot count (%d->1)\n", 345 td->td_proc->p_pid, kg->kg_avail_opennings); 346 347 } 348 kg->kg_avail_opennings = 1; 349 } 350 sched_add(td, flags); 351 return; 352 } 353 354 /* 355 * If the concurrency has reduced, and we would go in the 356 * assigned section, then keep removing entries from the 357 * system run queue, until we are not in that section 358 * or there is room for us to be put in that section. 359 * What we MUST avoid is the case where there are threads of less 360 * priority than the new one scheduled, but it can not 361 * be scheduled itself. That would lead to a non contiguous set 362 * of scheduled threads, and everything would break. 363 */ 364 tda = kg->kg_last_assigned; 365 while ((kg->kg_avail_opennings <= 0) && 366 (tda && (tda->td_priority > td->td_priority))) { 367 /* 368 * None free, but there is one we can commandeer. 369 */ 370 CTR2(KTR_RUNQ, 371 "setrunqueue: kg:%p: take slot from td: %p", kg, tda); 372 sched_rem(tda); 373 tda = kg->kg_last_assigned = 374 TAILQ_PREV(tda, threadqueue, td_runq); 375 SLOT_RELEASE(kg); 376 } 377 378 /* 379 * Add the thread to the ksegrp's run queue at 380 * the appropriate place. 381 */ 382 TAILQ_FOREACH(td2, &kg->kg_runq, td_runq) { 383 if (td2->td_priority > td->td_priority) { 384 kg->kg_runnable++; 385 TAILQ_INSERT_BEFORE(td2, td, td_runq); 386 break; 387 } 388 } 389 if (td2 == NULL) { 390 /* We ran off the end of the TAILQ or it was empty. */ 391 kg->kg_runnable++; 392 TAILQ_INSERT_TAIL(&kg->kg_runq, td, td_runq); 393 } 394 395 /* 396 * If we have a slot to use, then put the thread on the system 397 * run queue and if needed, readjust the last_assigned pointer. 398 * it may be that we need to schedule something anyhow 399 * even if the availabel slots are -ve so that 400 * all the items < last_assigned are scheduled. 401 */ 402 if (kg->kg_avail_opennings > 0) { 403 if (tda == NULL) { 404 /* 405 * No pre-existing last assigned so whoever is first 406 * gets the slot.. (maybe us) 407 */ 408 td2 = TAILQ_FIRST(&kg->kg_runq); 409 kg->kg_last_assigned = td2; 410 } else if (tda->td_priority > td->td_priority) { 411 td2 = td; 412 } else { 413 /* 414 * We are past last_assigned, so 415 * give the next slot to whatever is next, 416 * which may or may not be us. 417 */ 418 td2 = TAILQ_NEXT(tda, td_runq); 419 kg->kg_last_assigned = td2; 420 } 421 sched_add(td2, flags); 422 } else { 423 CTR3(KTR_RUNQ, "setrunqueue: held: td%p kg%p pid%d", 424 td, td->td_ksegrp, td->td_proc->p_pid); 425 } 426} 427 428/* 429 * Kernel thread preemption implementation. Critical sections mark 430 * regions of code in which preemptions are not allowed. 431 */ 432void 433critical_enter(void) 434{ 435 struct thread *td; 436 437 td = curthread; 438 if (td->td_critnest == 0) 439 cpu_critical_enter(td); 440 td->td_critnest++; 441} 442 443void 444critical_exit(void) 445{ 446 struct thread *td; 447 448 td = curthread; 449 KASSERT(td->td_critnest != 0, 450 ("critical_exit: td_critnest == 0")); 451 if (td->td_critnest == 1) { 452#ifdef PREEMPTION 453 mtx_assert(&sched_lock, MA_NOTOWNED); 454 if (td->td_pflags & TDP_OWEPREEMPT) { 455 mtx_lock_spin(&sched_lock); 456 mi_switch(SW_INVOL, NULL); 457 mtx_unlock_spin(&sched_lock); 458 } 459#endif 460 td->td_critnest = 0; 461 cpu_critical_exit(td); 462 } else { 463 td->td_critnest--; 464 } 465} 466 467/* 468 * This function is called when a thread is about to be put on run queue 469 * because it has been made runnable or its priority has been adjusted. It 470 * determines if the new thread should be immediately preempted to. If so, 471 * it switches to it and eventually returns true. If not, it returns false 472 * so that the caller may place the thread on an appropriate run queue. 473 */ 474int 475maybe_preempt(struct thread *td) 476{ 477#ifdef PREEMPTION 478 struct thread *ctd; 479 int cpri, pri; 480#endif 481 482 mtx_assert(&sched_lock, MA_OWNED); 483#ifdef PREEMPTION 484 /* 485 * The new thread should not preempt the current thread if any of the 486 * following conditions are true: 487 * 488 * - The current thread has a higher (numerically lower) or 489 * equivalent priority. Note that this prevents curthread from 490 * trying to preempt to itself. 491 * - It is too early in the boot for context switches (cold is set). 492 * - The current thread has an inhibitor set or is in the process of 493 * exiting. In this case, the current thread is about to switch 494 * out anyways, so there's no point in preempting. If we did, 495 * the current thread would not be properly resumed as well, so 496 * just avoid that whole landmine. 497 * - If the new thread's priority is not a realtime priority and 498 * the current thread's priority is not an idle priority and 499 * FULL_PREEMPTION is disabled. 500 * 501 * If all of these conditions are false, but the current thread is in 502 * a nested critical section, then we have to defer the preemption 503 * until we exit the critical section. Otherwise, switch immediately 504 * to the new thread. 505 */ 506 ctd = curthread; 507 KASSERT ((ctd->td_kse != NULL && ctd->td_kse->ke_thread == ctd), 508 ("thread has no (or wrong) sched-private part.")); 509 KASSERT((td->td_inhibitors == 0), 510 ("maybe_preempt: trying to run inhibitted thread")); 511 pri = td->td_priority; 512 cpri = ctd->td_priority; 513 if (pri >= cpri || cold /* || dumping */ || TD_IS_INHIBITED(ctd) || 514 td->td_kse->ke_state != KES_THREAD) 515 return (0); 516#ifndef FULL_PREEMPTION 517 if (!(pri >= PRI_MIN_ITHD && pri <= PRI_MAX_ITHD) && 518 !(cpri >= PRI_MIN_IDLE)) 519 return (0); 520#endif 521 if (ctd->td_critnest > 1) { 522 CTR1(KTR_PROC, "maybe_preempt: in critical section %d", 523 ctd->td_critnest); 524 ctd->td_pflags |= TDP_OWEPREEMPT; 525 return (0); 526 } 527 528 /* 529 * Thread is runnable but not yet put on system run queue. 530 */ 531 MPASS(TD_ON_RUNQ(td)); 532 MPASS(td->td_sched->ke_state != KES_ONRUNQ); 533 if (td->td_proc->p_flag & P_HADTHREADS) { 534 /* 535 * If this is a threaded process we actually ARE on the 536 * ksegrp run queue so take it off that first. 537 * Also undo any damage done to the last_assigned pointer. 538 * XXX Fix setrunqueue so this isn't needed 539 */ 540 struct ksegrp *kg; 541 542 kg = td->td_ksegrp; 543 if (kg->kg_last_assigned == td) 544 kg->kg_last_assigned = 545 TAILQ_PREV(td, threadqueue, td_runq); 546 TAILQ_REMOVE(&kg->kg_runq, td, td_runq); 547 } 548 549 TD_SET_RUNNING(td); 550 CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td, 551 td->td_proc->p_pid, td->td_proc->p_comm); 552 mi_switch(SW_INVOL|SW_PREEMPT, td); 553 return (1); 554#else 555 return (0); 556#endif 557} 558 559#if 0 560#ifndef PREEMPTION 561/* XXX: There should be a non-static version of this. */ 562static void 563printf_caddr_t(void *data) 564{ 565 printf("%s", (char *)data); 566} 567static char preempt_warning[] = 568 "WARNING: Kernel preemption is disabled, expect reduced performance.\n"; 569SYSINIT(preempt_warning, SI_SUB_COPYRIGHT, SI_ORDER_ANY, printf_caddr_t, 570 preempt_warning) 571#endif 572#endif 573 574/************************************************************************ 575 * SYSTEM RUN QUEUE manipulations and tests * 576 ************************************************************************/ 577/* 578 * Initialize a run structure. 579 */ 580void 581runq_init(struct runq *rq) 582{ 583 int i; 584 585 bzero(rq, sizeof *rq); 586 for (i = 0; i < RQ_NQS; i++) 587 TAILQ_INIT(&rq->rq_queues[i]); 588} 589 590/* 591 * Clear the status bit of the queue corresponding to priority level pri, 592 * indicating that it is empty. 593 */ 594static __inline void 595runq_clrbit(struct runq *rq, int pri) 596{ 597 struct rqbits *rqb; 598 599 rqb = &rq->rq_status; 600 CTR4(KTR_RUNQ, "runq_clrbit: bits=%#x %#x bit=%#x word=%d", 601 rqb->rqb_bits[RQB_WORD(pri)], 602 rqb->rqb_bits[RQB_WORD(pri)] & ~RQB_BIT(pri), 603 RQB_BIT(pri), RQB_WORD(pri)); 604 rqb->rqb_bits[RQB_WORD(pri)] &= ~RQB_BIT(pri); 605} 606 607/* 608 * Find the index of the first non-empty run queue. This is done by 609 * scanning the status bits, a set bit indicates a non-empty queue. 610 */ 611static __inline int 612runq_findbit(struct runq *rq) 613{ 614 struct rqbits *rqb; 615 int pri; 616 int i; 617 618 rqb = &rq->rq_status; 619 for (i = 0; i < RQB_LEN; i++) 620 if (rqb->rqb_bits[i]) { 621 pri = RQB_FFS(rqb->rqb_bits[i]) + (i << RQB_L2BPW); 622 CTR3(KTR_RUNQ, "runq_findbit: bits=%#x i=%d pri=%d", 623 rqb->rqb_bits[i], i, pri); 624 return (pri); 625 } 626 627 return (-1); 628} 629 630/* 631 * Set the status bit of the queue corresponding to priority level pri, 632 * indicating that it is non-empty. 633 */ 634static __inline void 635runq_setbit(struct runq *rq, int pri) 636{ 637 struct rqbits *rqb; 638 639 rqb = &rq->rq_status; 640 CTR4(KTR_RUNQ, "runq_setbit: bits=%#x %#x bit=%#x word=%d", 641 rqb->rqb_bits[RQB_WORD(pri)], 642 rqb->rqb_bits[RQB_WORD(pri)] | RQB_BIT(pri), 643 RQB_BIT(pri), RQB_WORD(pri)); 644 rqb->rqb_bits[RQB_WORD(pri)] |= RQB_BIT(pri); 645} 646 647/* 648 * Add the KSE to the queue specified by its priority, and set the 649 * corresponding status bit. 650 */ 651void 652runq_add(struct runq *rq, struct kse *ke, int flags) 653{ 654 struct rqhead *rqh; 655 int pri; 656 657 pri = ke->ke_thread->td_priority / RQ_PPQ; 658 ke->ke_rqindex = pri; 659 runq_setbit(rq, pri); 660 rqh = &rq->rq_queues[pri]; 661 CTR5(KTR_RUNQ, "runq_add: td=%p ke=%p pri=%d %d rqh=%p", 662 ke->ke_thread, ke, ke->ke_thread->td_priority, pri, rqh); 663 if (flags & SRQ_PREEMPTED) { 664 TAILQ_INSERT_HEAD(rqh, ke, ke_procq); 665 } else { 666 TAILQ_INSERT_TAIL(rqh, ke, ke_procq); 667 } 668} 669 670/* 671 * Return true if there are runnable processes of any priority on the run 672 * queue, false otherwise. Has no side effects, does not modify the run 673 * queue structure. 674 */ 675int 676runq_check(struct runq *rq) 677{ 678 struct rqbits *rqb; 679 int i; 680 681 rqb = &rq->rq_status; 682 for (i = 0; i < RQB_LEN; i++) 683 if (rqb->rqb_bits[i]) { 684 CTR2(KTR_RUNQ, "runq_check: bits=%#x i=%d", 685 rqb->rqb_bits[i], i); 686 return (1); 687 } 688 CTR0(KTR_RUNQ, "runq_check: empty"); 689 690 return (0); 691} 692 693#if defined(SMP) && defined(SCHED_4BSD) 694int runq_fuzz = 1; 695SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, ""); 696#endif 697 698/* 699 * Find the highest priority process on the run queue. 700 */ 701struct kse * 702runq_choose(struct runq *rq) 703{ 704 struct rqhead *rqh; 705 struct kse *ke; 706 int pri; 707 708 mtx_assert(&sched_lock, MA_OWNED); 709 while ((pri = runq_findbit(rq)) != -1) { 710 rqh = &rq->rq_queues[pri]; 711#if defined(SMP) && defined(SCHED_4BSD) 712 /* fuzz == 1 is normal.. 0 or less are ignored */ 713 if (runq_fuzz > 1) { 714 /* 715 * In the first couple of entries, check if 716 * there is one for our CPU as a preference. 717 */ 718 int count = runq_fuzz; 719 int cpu = PCPU_GET(cpuid); 720 struct kse *ke2; 721 ke2 = ke = TAILQ_FIRST(rqh); 722 723 while (count-- && ke2) { 724 if (ke->ke_thread->td_lastcpu == cpu) { 725 ke = ke2; 726 break; 727 } 728 ke2 = TAILQ_NEXT(ke2, ke_procq); 729 } 730 } else 731#endif 732 ke = TAILQ_FIRST(rqh); 733 KASSERT(ke != NULL, ("runq_choose: no proc on busy queue")); 734 CTR3(KTR_RUNQ, 735 "runq_choose: pri=%d kse=%p rqh=%p", pri, ke, rqh); 736 return (ke); 737 } 738 CTR1(KTR_RUNQ, "runq_choose: idleproc pri=%d", pri); 739 740 return (NULL); 741} 742 743/* 744 * Remove the KSE from the queue specified by its priority, and clear the 745 * corresponding status bit if the queue becomes empty. 746 * Caller must set ke->ke_state afterwards. 747 */ 748void 749runq_remove(struct runq *rq, struct kse *ke) 750{ 751 struct rqhead *rqh; 752 int pri; 753 754 KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 755 ("runq_remove: process swapped out")); 756 pri = ke->ke_rqindex; 757 rqh = &rq->rq_queues[pri]; 758 CTR5(KTR_RUNQ, "runq_remove: td=%p, ke=%p pri=%d %d rqh=%p", 759 ke->ke_thread, ke, ke->ke_thread->td_priority, pri, rqh); 760 KASSERT(ke != NULL, ("runq_remove: no proc on busy queue")); 761 TAILQ_REMOVE(rqh, ke, ke_procq); 762 if (TAILQ_EMPTY(rqh)) { 763 CTR0(KTR_RUNQ, "runq_remove: empty"); 764 runq_clrbit(rq, pri); 765 } 766} 767 768/****** functions that are temporarily here ***********/ 769#include <vm/uma.h> 770#define RANGEOF(type, start, end) (offsetof(type, end) - offsetof(type, start)) 771extern struct mtx kse_zombie_lock; 772 773/* 774 * Allocate scheduler specific per-process resources. 775 * The thread and ksegrp have already been linked in. 776 * In this case just set the default concurrency value. 777 * 778 * Called from: 779 * proc_init() (UMA init method) 780 */ 781void 782sched_newproc(struct proc *p, struct ksegrp *kg, struct thread *td) 783{ 784 785 /* This can go in sched_fork */ 786 sched_init_concurrency(kg); 787} 788 789#define RANGEOF(type, start, end) (offsetof(type, end) - offsetof(type, start)) 790/* 791 * thread is being either created or recycled. 792 * Fix up the per-scheduler resources associated with it. 793 * Called from: 794 * sched_fork_thread() 795 * thread_dtor() (*may go away) 796 * thread_init() (*may go away) 797 */ 798void 799sched_newthread(struct thread *td) 800{ 801 struct td_sched *ke; 802 803 ke = (struct td_sched *) (td + 1); 804 bzero(ke, sizeof(*ke)); 805 td->td_sched = ke; 806 ke->ke_thread = td; 807 ke->ke_oncpu = NOCPU; 808 ke->ke_state = KES_THREAD; 809} 810 811/* 812 * Set up an initial concurrency of 1 813 * and set the given thread (if given) to be using that 814 * concurrency slot. 815 * May be used "offline"..before the ksegrp is attached to the world 816 * and thus wouldn't need schedlock in that case. 817 * Called from: 818 * thr_create() 819 * proc_init() (UMA) via sched_newproc() 820 */ 821void 822sched_init_concurrency(struct ksegrp *kg) 823{ 824 825 CTR1(KTR_RUNQ,"kg %p init slots and concurrency to 1", kg); 826 kg->kg_concurrency = 1; 827 kg->kg_avail_opennings = 1; 828} 829 830/* 831 * Change the concurrency of an existing ksegrp to N 832 * Called from: 833 * kse_create() 834 * kse_exit() 835 * thread_exit() 836 * thread_single() 837 */ 838void 839sched_set_concurrency(struct ksegrp *kg, int concurrency) 840{ 841 842 CTR4(KTR_RUNQ,"kg %p set concurrency to %d, slots %d -> %d", 843 kg, 844 concurrency, 845 kg->kg_avail_opennings, 846 kg->kg_avail_opennings + (concurrency - kg->kg_concurrency)); 847 kg->kg_avail_opennings += (concurrency - kg->kg_concurrency); 848 kg->kg_concurrency = concurrency; 849} 850 851/* 852 * Called from thread_exit() for all exiting thread 853 * 854 * Not to be confused with sched_exit_thread() 855 * that is only called from thread_exit() for threads exiting 856 * without the rest of the process exiting because it is also called from 857 * sched_exit() and we wouldn't want to call it twice. 858 * XXX This can probably be fixed. 859 */ 860void 861sched_thread_exit(struct thread *td) 862{ 863 864 SLOT_RELEASE(td->td_ksegrp); 865 slot_fill(td->td_ksegrp); 866} 867 868#endif /* KERN_SWITCH_INCLUDE */ 869