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 /*** 28 Here is the logic.. 29 30 If there are N processors, then there are at most N KSEs (kernel 31 schedulable entities) working to process threads that belong to a 32 KSEGROUP (kg). If there are X of these KSEs actually running at the 33 moment in question, then there are at most M (N-X) of these KSEs on 34 the run queue, as running KSEs are not on the queue. 35 36 Runnable threads are queued off the KSEGROUP in priority order. 37 If 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 39 their priority from those threads and are put on the run queue. 40 41 The last thread that had a priority high enough to have a KSE associated 42 with it, AND IS ON THE RUN QUEUE is pointed to by 43 kg->kg_last_assigned. If no threads queued off the KSEGROUP have KSEs 44 assigned as all the available KSEs are activly running, or because there 45 are no threads queued, that pointer is NULL. 46 47 When a KSE is removed from the run queue to become runnable, we know 48 it was associated with the highest priority thread in the queue (at the head 49 of the queue). If it is also the last assigned we know M was 1 and must 50 now be 0. Since the thread is no longer queued that pointer must be 51 removed 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 53 but using it, we know there are STILL no more KSEs available, we can prove 54 that the next thread in the ksegrp list will not have a KSE to assign to 55 it, so we can show that the pointer must be made 'invalid' (NULL). 56 57 The pointer exists so that when a new thread is made runnable, it can 58 have its priority compared with the last assigned thread to see if 59 it should 'steal' its KSE or not.. i.e. is it 'earlier' 60 on the list than that thread or later.. If it's earlier, then the KSE is 61 removed from the last assigned (which is now not assigned a KSE) 62 and reassigned to the new thread, which is placed earlier in the list. 63 The pointer is then backed up to the previous thread (which may or may not 64 be the new thread). 65 66 When a thread sleeps or is removed, the KSE becomes available and if there 67 are queued threads that are not assigned KSEs, the highest priority one of 68 them is assigned the KSE, which is then placed back on the run queue at 69 the approipriate place, and the kg->kg_last_assigned pointer is adjusted down 70 to point to it. 71 72 The 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 82 The result of this scheme is that the M available KSEs are always 83 queued at the priorities they have inherrited from the M highest priority 84 threads for that KSEGROUP. If this situation changes, the KSEs are 85 reassigned to keep this true. 86 ***/ 87 88 #include <sys/cdefs.h> 89 __FBSDID("$FreeBSD$"); 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 #if defined(SMP) && defined(SCHED_4BSD) 109 #include <sys/sysctl.h> 110 #endif 111 112 #ifdef FULL_PREEMPTION 113 #ifndef PREEMPTION 114 #error "The FULL_PREEMPTION option requires the PREEMPTION option" 115 #endif 116 #endif 117 118 CTASSERT((RQB_BPW * RQB_LEN) == RQ_NQS); 119 120 #define td_kse td_sched 121 122 /* 123 * kern.sched.preemption allows user space to determine if preemption support 124 * is compiled in or not. It is not currently a boot or runtime flag that 125 * can be changed. 126 */ 127 #ifdef PREEMPTION 128 static int kern_sched_preemption = 1; 129 #else 130 static int kern_sched_preemption = 0; 131 #endif 132 SYSCTL_INT(_kern_sched, OID_AUTO, preemption, CTLFLAG_RD, 133 &kern_sched_preemption, 0, "Kernel preemption enabled"); 134 135 /************************************************************************ 136 * Functions that manipulate runnability from a thread perspective. * 137 ************************************************************************/ 138 /* 139 * Select the KSE that will be run next. From that find the thread, and 140 * remove it from the KSEGRP's run queue. If there is thread clustering, 141 * this will be what does it. 142 */ 143 struct thread * 144 choosethread(void) 145 { 146 struct kse *ke; 147 struct thread *td; 148 struct ksegrp *kg; 149 150 #if defined(SMP) && (defined(__i386__) || defined(__amd64__)) 151 if (smp_active == 0 && PCPU_GET(cpuid) != 0) { 152 /* Shutting down, run idlethread on AP's */ 153 td = PCPU_GET(idlethread); 154 ke = td->td_kse; 155 CTR1(KTR_RUNQ, "choosethread: td=%p (idle)", td); 156 ke->ke_flags |= KEF_DIDRUN; 157 TD_SET_RUNNING(td); 158 return (td); 159 } 160 #endif 161 162 retry: 163 ke = sched_choose(); 164 if (ke) { 165 td = ke->ke_thread; 166 KASSERT((td->td_kse == ke), ("kse/thread mismatch")); 167 kg = ke->ke_ksegrp; 168 if (td->td_proc->p_flag & P_HADTHREADS) { 169 if (kg->kg_last_assigned == td) { 170 kg->kg_last_assigned = TAILQ_PREV(td, 171 threadqueue, td_runq); 172 } 173 TAILQ_REMOVE(&kg->kg_runq, td, td_runq); 174 } 175 CTR2(KTR_RUNQ, "choosethread: td=%p pri=%d", 176 td, td->td_priority); 177 } else { 178 /* Simulate runq_choose() having returned the idle thread */ 179 td = PCPU_GET(idlethread); 180 ke = td->td_kse; 181 CTR1(KTR_RUNQ, "choosethread: td=%p (idle)", td); 182 } 183 ke->ke_flags |= KEF_DIDRUN; 184 185 /* 186 * If we are in panic, only allow system threads, 187 * plus the one we are running in, to be run. 188 */ 189 if (panicstr && ((td->td_proc->p_flag & P_SYSTEM) == 0 && 190 (td->td_flags & TDF_INPANIC) == 0)) { 191 /* note that it is no longer on the run queue */ 192 TD_SET_CAN_RUN(td); 193 goto retry; 194 } 195 196 TD_SET_RUNNING(td); 197 return (td); 198 } 199 200 /* 201 * Given a surplus system slot, try assign a new runnable thread to it. 202 * Called from: 203 * sched_thread_exit() (local) 204 * sched_switch() (local) 205 * sched_thread_exit() (local) 206 * remrunqueue() (local) (not at the moment) 207 */ 208 static void 209 slot_fill(struct ksegrp *kg) 210 { 211 struct thread *td; 212 213 mtx_assert(&sched_lock, MA_OWNED); 214 while (kg->kg_avail_opennings > 0) { 215 /* 216 * Find the first unassigned thread 217 */ 218 if ((td = kg->kg_last_assigned) != NULL) 219 td = TAILQ_NEXT(td, td_runq); 220 else 221 td = TAILQ_FIRST(&kg->kg_runq); 222 223 /* 224 * If we found one, send it to the system scheduler. 225 */ 226 if (td) { 227 kg->kg_last_assigned = td; 228 sched_add(td, SRQ_YIELDING); 229 CTR2(KTR_RUNQ, "slot_fill: td%p -> kg%p", td, kg); 230 } else { 231 /* no threads to use up the slots. quit now */ 232 break; 233 } 234 } 235 } 236 237 #ifdef SCHED_4BSD 238 /* 239 * Remove a thread from its KSEGRP's run queue. 240 * This in turn may remove it from a KSE if it was already assigned 241 * to one, possibly causing a new thread to be assigned to the KSE 242 * and the KSE getting a new priority. 243 */ 244 static void 245 remrunqueue(struct thread *td) 246 { 247 struct thread *td2, *td3; 248 struct ksegrp *kg; 249 struct kse *ke; 250 251 mtx_assert(&sched_lock, MA_OWNED); 252 KASSERT((TD_ON_RUNQ(td)), ("remrunqueue: Bad state on run queue")); 253 kg = td->td_ksegrp; 254 ke = td->td_kse; 255 CTR1(KTR_RUNQ, "remrunqueue: td%p", td); 256 TD_SET_CAN_RUN(td); 257 /* 258 * If it is not a threaded process, take the shortcut. 259 */ 260 if ((td->td_proc->p_flag & P_HADTHREADS) == 0) { 261 /* remve from sys run queue and free up a slot */ 262 sched_rem(td); 263 ke->ke_state = KES_THREAD; 264 return; 265 } 266 td3 = TAILQ_PREV(td, threadqueue, td_runq); 267 TAILQ_REMOVE(&kg->kg_runq, td, td_runq); 268 if (ke->ke_state == KES_ONRUNQ) { 269 /* 270 * This thread has been assigned to the system run queue. 271 * We need to dissociate it and try assign the 272 * KSE to the next available thread. Then, we should 273 * see if we need to move the KSE in the run queues. 274 */ 275 sched_rem(td); 276 ke->ke_state = KES_THREAD; 277 td2 = kg->kg_last_assigned; 278 KASSERT((td2 != NULL), ("last assigned has wrong value")); 279 if (td2 == td) 280 kg->kg_last_assigned = td3; 281 /* slot_fill(kg); */ /* will replace it with another */ 282 } 283 } 284 #endif 285 286 /* 287 * Change the priority of a thread that is on the run queue. 288 */ 289 void 290 adjustrunqueue( struct thread *td, int newpri) 291 { 292 struct ksegrp *kg; 293 struct kse *ke; 294 295 mtx_assert(&sched_lock, MA_OWNED); 296 KASSERT((TD_ON_RUNQ(td)), ("adjustrunqueue: Bad state on run queue")); 297 298 ke = td->td_kse; 299 CTR1(KTR_RUNQ, "adjustrunqueue: td%p", td); 300 /* 301 * If it is not a threaded process, take the shortcut. 302 */ 303 if ((td->td_proc->p_flag & P_HADTHREADS) == 0) { 304 /* We only care about the kse in the run queue. */ 305 td->td_priority = newpri; 306 if (ke->ke_rqindex != (newpri / RQ_PPQ)) { 307 sched_rem(td); 308 sched_add(td, SRQ_BORING); 309 } 310 return; 311 } 312 313 /* It is a threaded process */ 314 kg = td->td_ksegrp; 315 if (ke->ke_state == KES_ONRUNQ) { 316 if (kg->kg_last_assigned == td) { 317 kg->kg_last_assigned = 318 TAILQ_PREV(td, threadqueue, td_runq); 319 } 320 sched_rem(td); 321 } 322 TAILQ_REMOVE(&kg->kg_runq, td, td_runq); 323 TD_SET_CAN_RUN(td); 324 td->td_priority = newpri; 325 setrunqueue(td, SRQ_BORING); 326 } 327 328 /* 329 * This function is called when a thread is about to be put on a 330 * ksegrp run queue because it has been made runnable or its 331 * priority has been adjusted and the ksegrp does not have a 332 * free kse slot. It determines if a thread from the same ksegrp 333 * should be preempted. If so, it tries to switch threads 334 * if the thread is on the same cpu or notifies another cpu that 335 * it should switch threads. 336 */ 337 338 static void 339 maybe_preempt_in_ksegrp(struct thread *td) 340 #if !defined(SMP) 341 { 342 struct thread *running_thread; 343 344 mtx_assert(&sched_lock, MA_OWNED); 345 running_thread = curthread; 346 347 if (running_thread->td_ksegrp != td->td_ksegrp) 348 return; 349 350 if (td->td_priority >= running_thread->td_priority) 351 return; 352 #ifdef PREEMPTION 353 #ifndef FULL_PREEMPTION 354 if (td->td_priority > PRI_MAX_ITHD) { 355 running_thread->td_flags |= TDF_NEEDRESCHED; 356 return; 357 } 358 #endif /* FULL_PREEMPTION */ 359 360 if (running_thread->td_critnest > 1) 361 running_thread->td_owepreempt = 1; 362 else 363 mi_switch(SW_INVOL, NULL); 364 365 #else /* PREEMPTION */ 366 running_thread->td_flags |= TDF_NEEDRESCHED; 367 #endif /* PREEMPTION */ 368 return; 369 } 370 371 #else /* SMP */ 372 { 373 struct thread *running_thread; 374 int worst_pri; 375 struct ksegrp *kg; 376 cpumask_t cpumask,dontuse; 377 struct pcpu *pc; 378 struct pcpu *best_pcpu; 379 struct thread *cputhread; 380 381 mtx_assert(&sched_lock, MA_OWNED); 382 383 running_thread = curthread; 384 385 #if !defined(KSEG_PEEMPT_BEST_CPU) 386 if (running_thread->td_ksegrp != td->td_ksegrp) { 387 #endif 388 kg = td->td_ksegrp; 389 390 /* if someone is ahead of this thread, wait our turn */ 391 if (td != TAILQ_FIRST(&kg->kg_runq)) 392 return; 393 394 worst_pri = td->td_priority; 395 best_pcpu = NULL; 396 dontuse = stopped_cpus | idle_cpus_mask; 397 398 /* 399 * Find a cpu with the worst priority that runs at thread from 400 * the same ksegrp - if multiple exist give first the last run 401 * cpu and then the current cpu priority 402 */ 403 404 SLIST_FOREACH(pc, &cpuhead, pc_allcpu) { 405 cpumask = pc->pc_cpumask; 406 cputhread = pc->pc_curthread; 407 408 if ((cpumask & dontuse) || 409 cputhread->td_ksegrp != kg) 410 continue; 411 412 if (cputhread->td_priority > worst_pri) { 413 worst_pri = cputhread->td_priority; 414 best_pcpu = pc; 415 continue; 416 } 417 418 if (cputhread->td_priority == worst_pri && 419 best_pcpu != NULL && 420 (td->td_lastcpu == pc->pc_cpuid || 421 (PCPU_GET(cpumask) == cpumask && 422 td->td_lastcpu != best_pcpu->pc_cpuid))) 423 best_pcpu = pc; 424 } 425 426 /* Check if we need to preempt someone */ 427 if (best_pcpu == NULL) 428 return; 429 430 #if defined(IPI_PREEMPTION) && defined(PREEMPTION) 431 #if !defined(FULL_PREEMPTION) 432 if (td->td_priority <= PRI_MAX_ITHD) 433 #endif /* ! FULL_PREEMPTION */ 434 { 435 ipi_selected(best_pcpu->pc_cpumask, IPI_PREEMPT); 436 return; 437 } 438 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */ 439 440 if (PCPU_GET(cpuid) != best_pcpu->pc_cpuid) { 441 best_pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED; 442 ipi_selected(best_pcpu->pc_cpumask, IPI_AST); 443 return; 444 } 445 #if !defined(KSEG_PEEMPT_BEST_CPU) 446 } 447 #endif 448 449 if (td->td_priority >= running_thread->td_priority) 450 return; 451 #ifdef PREEMPTION 452 453 #if !defined(FULL_PREEMPTION) 454 if (td->td_priority > PRI_MAX_ITHD) { 455 running_thread->td_flags |= TDF_NEEDRESCHED; 456 } 457 #endif /* ! FULL_PREEMPTION */ 458 459 if (running_thread->td_critnest > 1) 460 running_thread->td_owepreempt = 1; 461 else 462 mi_switch(SW_INVOL, NULL); 463 464 #else /* PREEMPTION */ 465 running_thread->td_flags |= TDF_NEEDRESCHED; 466 #endif /* PREEMPTION */ 467 return; 468 } 469 #endif /* !SMP */ 470 471 472 int limitcount; 473 void 474 setrunqueue(struct thread *td, int flags) 475 { 476 struct ksegrp *kg; 477 struct thread *td2; 478 struct thread *tda; 479 480 CTR3(KTR_RUNQ, "setrunqueue: td:%p kg:%p pid:%d", 481 td, td->td_ksegrp, td->td_proc->p_pid); 482 CTR5(KTR_SCHED, "setrunqueue: %p(%s) prio %d by %p(%s)", 483 td, td->td_proc->p_comm, td->td_priority, curthread, 484 curthread->td_proc->p_comm); 485 mtx_assert(&sched_lock, MA_OWNED); 486 KASSERT((td->td_inhibitors == 0), 487 ("setrunqueue: trying to run inhibitted thread")); 488 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 489 ("setrunqueue: bad thread state")); 490 TD_SET_RUNQ(td); 491 kg = td->td_ksegrp; 492 if ((td->td_proc->p_flag & P_HADTHREADS) == 0) { 493 /* 494 * Common path optimisation: Only one of everything 495 * and the KSE is always already attached. 496 * Totally ignore the ksegrp run queue. 497 */ 498 if (kg->kg_avail_opennings != 1) { 499 if (limitcount < 1) { 500 limitcount++; 501 printf("pid %d: corrected slot count (%d->1)\n", 502 td->td_proc->p_pid, kg->kg_avail_opennings); 503 504 } 505 kg->kg_avail_opennings = 1; 506 } 507 sched_add(td, flags); 508 return; 509 } 510 511 /* 512 * If the concurrency has reduced, and we would go in the 513 * assigned section, then keep removing entries from the 514 * system run queue, until we are not in that section 515 * or there is room for us to be put in that section. 516 * What we MUST avoid is the case where there are threads of less 517 * priority than the new one scheduled, but it can not 518 * be scheduled itself. That would lead to a non contiguous set 519 * of scheduled threads, and everything would break. 520 */ 521 tda = kg->kg_last_assigned; 522 while ((kg->kg_avail_opennings <= 0) && 523 (tda && (tda->td_priority > td->td_priority))) { 524 /* 525 * None free, but there is one we can commandeer. 526 */ 527 CTR2(KTR_RUNQ, 528 "setrunqueue: kg:%p: take slot from td: %p", kg, tda); 529 sched_rem(tda); 530 tda = kg->kg_last_assigned = 531 TAILQ_PREV(tda, threadqueue, td_runq); 532 } 533 534 /* 535 * Add the thread to the ksegrp's run queue at 536 * the appropriate place. 537 */ 538 TAILQ_FOREACH(td2, &kg->kg_runq, td_runq) { 539 if (td2->td_priority > td->td_priority) { 540 TAILQ_INSERT_BEFORE(td2, td, td_runq); 541 break; 542 } 543 } 544 if (td2 == NULL) { 545 /* We ran off the end of the TAILQ or it was empty. */ 546 TAILQ_INSERT_TAIL(&kg->kg_runq, td, td_runq); 547 } 548 549 /* 550 * If we have a slot to use, then put the thread on the system 551 * run queue and if needed, readjust the last_assigned pointer. 552 * it may be that we need to schedule something anyhow 553 * even if the availabel slots are -ve so that 554 * all the items < last_assigned are scheduled. 555 */ 556 if (kg->kg_avail_opennings > 0) { 557 if (tda == NULL) { 558 /* 559 * No pre-existing last assigned so whoever is first 560 * gets the slot.. (maybe us) 561 */ 562 td2 = TAILQ_FIRST(&kg->kg_runq); 563 kg->kg_last_assigned = td2; 564 } else if (tda->td_priority > td->td_priority) { 565 td2 = td; 566 } else { 567 /* 568 * We are past last_assigned, so 569 * give the next slot to whatever is next, 570 * which may or may not be us. 571 */ 572 td2 = TAILQ_NEXT(tda, td_runq); 573 kg->kg_last_assigned = td2; 574 } 575 sched_add(td2, flags); 576 } else { 577 CTR3(KTR_RUNQ, "setrunqueue: held: td%p kg%p pid%d", 578 td, td->td_ksegrp, td->td_proc->p_pid); 579 if ((flags & SRQ_YIELDING) == 0) 580 maybe_preempt_in_ksegrp(td); 581 } 582 } 583 584 /* 585 * Kernel thread preemption implementation. Critical sections mark 586 * regions of code in which preemptions are not allowed. 587 */ 588 void 589 critical_enter(void) 590 { 591 struct thread *td; 592 593 td = curthread; 594 td->td_critnest++; 595 CTR4(KTR_CRITICAL, "critical_enter by thread %p (%ld, %s) to %d", td, 596 (long)td->td_proc->p_pid, td->td_proc->p_comm, td->td_critnest); 597 } 598 599 void 600 critical_exit(void) 601 { 602 struct thread *td; 603 604 td = curthread; 605 KASSERT(td->td_critnest != 0, 606 ("critical_exit: td_critnest == 0")); 607 #ifdef PREEMPTION 608 if (td->td_critnest == 1) { 609 td->td_critnest = 0; 610 mtx_assert(&sched_lock, MA_NOTOWNED); 611 if (td->td_owepreempt) { 612 td->td_critnest = 1; 613 mtx_lock_spin(&sched_lock); 614 td->td_critnest--; 615 mi_switch(SW_INVOL, NULL); 616 mtx_unlock_spin(&sched_lock); 617 } 618 } else 619 #endif 620 td->td_critnest--; 621 622 623 CTR4(KTR_CRITICAL, "critical_exit by thread %p (%ld, %s) to %d", td, 624 (long)td->td_proc->p_pid, td->td_proc->p_comm, td->td_critnest); 625 } 626 627 /* 628 * This function is called when a thread is about to be put on run queue 629 * because it has been made runnable or its priority has been adjusted. It 630 * determines if the new thread should be immediately preempted to. If so, 631 * it switches to it and eventually returns true. If not, it returns false 632 * so that the caller may place the thread on an appropriate run queue. 633 */ 634 int 635 maybe_preempt(struct thread *td) 636 { 637 #ifdef PREEMPTION 638 struct thread *ctd; 639 int cpri, pri; 640 #endif 641 642 mtx_assert(&sched_lock, MA_OWNED); 643 #ifdef PREEMPTION 644 /* 645 * The new thread should not preempt the current thread if any of the 646 * following conditions are true: 647 * 648 * - The kernel is in the throes of crashing (panicstr). 649 * - The current thread has a higher (numerically lower) or 650 * equivalent priority. Note that this prevents curthread from 651 * trying to preempt to itself. 652 * - It is too early in the boot for context switches (cold is set). 653 * - The current thread has an inhibitor set or is in the process of 654 * exiting. In this case, the current thread is about to switch 655 * out anyways, so there's no point in preempting. If we did, 656 * the current thread would not be properly resumed as well, so 657 * just avoid that whole landmine. 658 * - If the new thread's priority is not a realtime priority and 659 * the current thread's priority is not an idle priority and 660 * FULL_PREEMPTION is disabled. 661 * 662 * If all of these conditions are false, but the current thread is in 663 * a nested critical section, then we have to defer the preemption 664 * until we exit the critical section. Otherwise, switch immediately 665 * to the new thread. 666 */ 667 ctd = curthread; 668 KASSERT ((ctd->td_kse != NULL && ctd->td_kse->ke_thread == ctd), 669 ("thread has no (or wrong) sched-private part.")); 670 KASSERT((td->td_inhibitors == 0), 671 ("maybe_preempt: trying to run inhibitted thread")); 672 pri = td->td_priority; 673 cpri = ctd->td_priority; 674 if (panicstr != NULL || pri >= cpri || cold /* || dumping */ || 675 TD_IS_INHIBITED(ctd) || td->td_kse->ke_state != KES_THREAD) 676 return (0); 677 #ifndef FULL_PREEMPTION 678 if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE) 679 return (0); 680 #endif 681 682 if (ctd->td_critnest > 1) { 683 CTR1(KTR_PROC, "maybe_preempt: in critical section %d", 684 ctd->td_critnest); 685 ctd->td_owepreempt = 1; 686 return (0); 687 } 688 689 /* 690 * Thread is runnable but not yet put on system run queue. 691 */ 692 MPASS(TD_ON_RUNQ(td)); 693 MPASS(td->td_sched->ke_state != KES_ONRUNQ); 694 if (td->td_proc->p_flag & P_HADTHREADS) { 695 /* 696 * If this is a threaded process we actually ARE on the 697 * ksegrp run queue so take it off that first. 698 * Also undo any damage done to the last_assigned pointer. 699 * XXX Fix setrunqueue so this isn't needed 700 */ 701 struct ksegrp *kg; 702 703 kg = td->td_ksegrp; 704 if (kg->kg_last_assigned == td) 705 kg->kg_last_assigned = 706 TAILQ_PREV(td, threadqueue, td_runq); 707 TAILQ_REMOVE(&kg->kg_runq, td, td_runq); 708 } 709 710 TD_SET_RUNNING(td); 711 CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td, 712 td->td_proc->p_pid, td->td_proc->p_comm); 713 mi_switch(SW_INVOL|SW_PREEMPT, td); 714 return (1); 715 #else 716 return (0); 717 #endif 718 } 719 720 #if 0 721 #ifndef PREEMPTION 722 /* XXX: There should be a non-static version of this. */ 723 static void 724 printf_caddr_t(void *data) 725 { 726 printf("%s", (char *)data); 727 } 728 static char preempt_warning[] = 729 "WARNING: Kernel preemption is disabled, expect reduced performance.\n"; 730 SYSINIT(preempt_warning, SI_SUB_COPYRIGHT, SI_ORDER_ANY, printf_caddr_t, 731 preempt_warning) 732 #endif 733 #endif 734 735 /************************************************************************ 736 * SYSTEM RUN QUEUE manipulations and tests * 737 ************************************************************************/ 738 /* 739 * Initialize a run structure. 740 */ 741 void 742 runq_init(struct runq *rq) 743 { 744 int i; 745 746 bzero(rq, sizeof *rq); 747 for (i = 0; i < RQ_NQS; i++) 748 TAILQ_INIT(&rq->rq_queues[i]); 749 } 750 751 /* 752 * Clear the status bit of the queue corresponding to priority level pri, 753 * indicating that it is empty. 754 */ 755 static __inline void 756 runq_clrbit(struct runq *rq, int pri) 757 { 758 struct rqbits *rqb; 759 760 rqb = &rq->rq_status; 761 CTR4(KTR_RUNQ, "runq_clrbit: bits=%#x %#x bit=%#x word=%d", 762 rqb->rqb_bits[RQB_WORD(pri)], 763 rqb->rqb_bits[RQB_WORD(pri)] & ~RQB_BIT(pri), 764 RQB_BIT(pri), RQB_WORD(pri)); 765 rqb->rqb_bits[RQB_WORD(pri)] &= ~RQB_BIT(pri); 766 } 767 768 /* 769 * Find the index of the first non-empty run queue. This is done by 770 * scanning the status bits, a set bit indicates a non-empty queue. 771 */ 772 static __inline int 773 runq_findbit(struct runq *rq) 774 { 775 struct rqbits *rqb; 776 int pri; 777 int i; 778 779 rqb = &rq->rq_status; 780 for (i = 0; i < RQB_LEN; i++) 781 if (rqb->rqb_bits[i]) { 782 pri = RQB_FFS(rqb->rqb_bits[i]) + (i << RQB_L2BPW); 783 CTR3(KTR_RUNQ, "runq_findbit: bits=%#x i=%d pri=%d", 784 rqb->rqb_bits[i], i, pri); 785 return (pri); 786 } 787 788 return (-1); 789 } 790 791 /* 792 * Set the status bit of the queue corresponding to priority level pri, 793 * indicating that it is non-empty. 794 */ 795 static __inline void 796 runq_setbit(struct runq *rq, int pri) 797 { 798 struct rqbits *rqb; 799 800 rqb = &rq->rq_status; 801 CTR4(KTR_RUNQ, "runq_setbit: bits=%#x %#x bit=%#x word=%d", 802 rqb->rqb_bits[RQB_WORD(pri)], 803 rqb->rqb_bits[RQB_WORD(pri)] | RQB_BIT(pri), 804 RQB_BIT(pri), RQB_WORD(pri)); 805 rqb->rqb_bits[RQB_WORD(pri)] |= RQB_BIT(pri); 806 } 807 808 /* 809 * Add the KSE to the queue specified by its priority, and set the 810 * corresponding status bit. 811 */ 812 void 813 runq_add(struct runq *rq, struct kse *ke, int flags) 814 { 815 struct rqhead *rqh; 816 int pri; 817 818 pri = ke->ke_thread->td_priority / RQ_PPQ; 819 ke->ke_rqindex = pri; 820 runq_setbit(rq, pri); 821 rqh = &rq->rq_queues[pri]; 822 CTR5(KTR_RUNQ, "runq_add: td=%p ke=%p pri=%d %d rqh=%p", 823 ke->ke_thread, ke, ke->ke_thread->td_priority, pri, rqh); 824 if (flags & SRQ_PREEMPTED) { 825 TAILQ_INSERT_HEAD(rqh, ke, ke_procq); 826 } else { 827 TAILQ_INSERT_TAIL(rqh, ke, ke_procq); 828 } 829 } 830 831 /* 832 * Return true if there are runnable processes of any priority on the run 833 * queue, false otherwise. Has no side effects, does not modify the run 834 * queue structure. 835 */ 836 int 837 runq_check(struct runq *rq) 838 { 839 struct rqbits *rqb; 840 int i; 841 842 rqb = &rq->rq_status; 843 for (i = 0; i < RQB_LEN; i++) 844 if (rqb->rqb_bits[i]) { 845 CTR2(KTR_RUNQ, "runq_check: bits=%#x i=%d", 846 rqb->rqb_bits[i], i); 847 return (1); 848 } 849 CTR0(KTR_RUNQ, "runq_check: empty"); 850 851 return (0); 852 } 853 854 #if defined(SMP) && defined(SCHED_4BSD) 855 int runq_fuzz = 1; 856 SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, ""); 857 #endif 858 859 /* 860 * Find the highest priority process on the run queue. 861 */ 862 struct kse * 863 runq_choose(struct runq *rq) 864 { 865 struct rqhead *rqh; 866 struct kse *ke; 867 int pri; 868 869 mtx_assert(&sched_lock, MA_OWNED); 870 while ((pri = runq_findbit(rq)) != -1) { 871 rqh = &rq->rq_queues[pri]; 872 #if defined(SMP) && defined(SCHED_4BSD) 873 /* fuzz == 1 is normal.. 0 or less are ignored */ 874 if (runq_fuzz > 1) { 875 /* 876 * In the first couple of entries, check if 877 * there is one for our CPU as a preference. 878 */ 879 int count = runq_fuzz; 880 int cpu = PCPU_GET(cpuid); 881 struct kse *ke2; 882 ke2 = ke = TAILQ_FIRST(rqh); 883 884 while (count-- && ke2) { 885 if (ke->ke_thread->td_lastcpu == cpu) { 886 ke = ke2; 887 break; 888 } 889 ke2 = TAILQ_NEXT(ke2, ke_procq); 890 } 891 } else 892 #endif 893 ke = TAILQ_FIRST(rqh); 894 KASSERT(ke != NULL, ("runq_choose: no proc on busy queue")); 895 CTR3(KTR_RUNQ, 896 "runq_choose: pri=%d kse=%p rqh=%p", pri, ke, rqh); 897 return (ke); 898 } 899 CTR1(KTR_RUNQ, "runq_choose: idleproc pri=%d", pri); 900 901 return (NULL); 902 } 903 904 /* 905 * Remove the KSE from the queue specified by its priority, and clear the 906 * corresponding status bit if the queue becomes empty. 907 * Caller must set ke->ke_state afterwards. 908 */ 909 void 910 runq_remove(struct runq *rq, struct kse *ke) 911 { 912 struct rqhead *rqh; 913 int pri; 914 915 KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 916 ("runq_remove: process swapped out")); 917 pri = ke->ke_rqindex; 918 rqh = &rq->rq_queues[pri]; 919 CTR5(KTR_RUNQ, "runq_remove: td=%p, ke=%p pri=%d %d rqh=%p", 920 ke->ke_thread, ke, ke->ke_thread->td_priority, pri, rqh); 921 KASSERT(ke != NULL, ("runq_remove: no proc on busy queue")); 922 TAILQ_REMOVE(rqh, ke, ke_procq); 923 if (TAILQ_EMPTY(rqh)) { 924 CTR0(KTR_RUNQ, "runq_remove: empty"); 925 runq_clrbit(rq, pri); 926 } 927 } 928 929 /****** functions that are temporarily here ***********/ 930 #include <vm/uma.h> 931 extern struct mtx kse_zombie_lock; 932 933 /* 934 * Allocate scheduler specific per-process resources. 935 * The thread and ksegrp have already been linked in. 936 * In this case just set the default concurrency value. 937 * 938 * Called from: 939 * proc_init() (UMA init method) 940 */ 941 void 942 sched_newproc(struct proc *p, struct ksegrp *kg, struct thread *td) 943 { 944 945 /* This can go in sched_fork */ 946 sched_init_concurrency(kg); 947 } 948 949 /* 950 * thread is being either created or recycled. 951 * Fix up the per-scheduler resources associated with it. 952 * Called from: 953 * sched_fork_thread() 954 * thread_dtor() (*may go away) 955 * thread_init() (*may go away) 956 */ 957 void 958 sched_newthread(struct thread *td) 959 { 960 struct td_sched *ke; 961 962 ke = (struct td_sched *) (td + 1); 963 bzero(ke, sizeof(*ke)); 964 td->td_sched = ke; 965 ke->ke_thread = td; 966 ke->ke_state = KES_THREAD; 967 } 968 969 /* 970 * Set up an initial concurrency of 1 971 * and set the given thread (if given) to be using that 972 * concurrency slot. 973 * May be used "offline"..before the ksegrp is attached to the world 974 * and thus wouldn't need schedlock in that case. 975 * Called from: 976 * thr_create() 977 * proc_init() (UMA) via sched_newproc() 978 */ 979 void 980 sched_init_concurrency(struct ksegrp *kg) 981 { 982 983 CTR1(KTR_RUNQ,"kg %p init slots and concurrency to 1", kg); 984 kg->kg_concurrency = 1; 985 kg->kg_avail_opennings = 1; 986 } 987 988 /* 989 * Change the concurrency of an existing ksegrp to N 990 * Called from: 991 * kse_create() 992 * kse_exit() 993 * thread_exit() 994 * thread_single() 995 */ 996 void 997 sched_set_concurrency(struct ksegrp *kg, int concurrency) 998 { 999 1000 CTR4(KTR_RUNQ,"kg %p set concurrency to %d, slots %d -> %d", 1001 kg, 1002 concurrency, 1003 kg->kg_avail_opennings, 1004 kg->kg_avail_opennings + (concurrency - kg->kg_concurrency)); 1005 kg->kg_avail_opennings += (concurrency - kg->kg_concurrency); 1006 kg->kg_concurrency = concurrency; 1007 } 1008 1009 /* 1010 * Called from thread_exit() for all exiting thread 1011 * 1012 * Not to be confused with sched_exit_thread() 1013 * that is only called from thread_exit() for threads exiting 1014 * without the rest of the process exiting because it is also called from 1015 * sched_exit() and we wouldn't want to call it twice. 1016 * XXX This can probably be fixed. 1017 */ 1018 void 1019 sched_thread_exit(struct thread *td) 1020 { 1021 1022 SLOT_RELEASE(td->td_ksegrp); 1023 slot_fill(td->td_ksegrp); 1024 } 1025 1026 #endif /* KERN_SWITCH_INCLUDE */ 1027