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 * 4. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35#include <sys/cdefs.h> 36__FBSDID("$FreeBSD: releng/10.3/sys/kern/sched_4bsd.c 288463 2015-10-01 21:54:43Z jhb $"); 37 38#include "opt_hwpmc_hooks.h" 39#include "opt_sched.h" 40#include "opt_kdtrace.h" 41 42#include <sys/param.h> 43#include <sys/systm.h> 44#include <sys/cpuset.h> 45#include <sys/kernel.h> 46#include <sys/ktr.h> 47#include <sys/lock.h> 48#include <sys/kthread.h> 49#include <sys/mutex.h> 50#include <sys/proc.h> 51#include <sys/resourcevar.h> 52#include <sys/sched.h> 53#include <sys/sdt.h> 54#include <sys/smp.h> 55#include <sys/sysctl.h> 56#include <sys/sx.h> 57#include <sys/turnstile.h> 58#include <sys/umtx.h> 59#include <machine/pcb.h> 60#include <machine/smp.h> 61 62#ifdef HWPMC_HOOKS 63#include <sys/pmckern.h> 64#endif 65 66#ifdef KDTRACE_HOOKS 67#include <sys/dtrace_bsd.h> 68int dtrace_vtime_active; 69dtrace_vtime_switch_func_t dtrace_vtime_switch_func; 70#endif 71 72/* 73 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in 74 * the range 100-256 Hz (approximately). 75 */ 76#define ESTCPULIM(e) \ 77 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \ 78 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1) 79#ifdef SMP 80#define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus) 81#else 82#define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */ 83#endif 84#define NICE_WEIGHT 1 /* Priorities per nice level. */ 85 86#define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX))) 87 88/* 89 * The schedulable entity that runs a context. 90 * This is an extension to the thread structure and is tailored to 91 * the requirements of this scheduler 92 */ 93struct td_sched { 94 fixpt_t ts_pctcpu; /* (j) %cpu during p_swtime. */ 95 int ts_cpticks; /* (j) Ticks of cpu time. */ 96 int ts_slptime; /* (j) Seconds !RUNNING. */ 97 int ts_slice; /* Remaining part of time slice. */ 98 int ts_flags; 99 struct runq *ts_runq; /* runq the thread is currently on */ 100#ifdef KTR 101 char ts_name[TS_NAME_LEN]; 102#endif 103}; 104 105/* flags kept in td_flags */ 106#define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */ 107#define TDF_BOUND TDF_SCHED1 /* Bound to one CPU. */ 108#define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */ 109 110/* flags kept in ts_flags */ 111#define TSF_AFFINITY 0x0001 /* Has a non-"full" CPU set. */ 112 113#define SKE_RUNQ_PCPU(ts) \ 114 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq) 115 116#define THREAD_CAN_SCHED(td, cpu) \ 117 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 118 119static struct td_sched td_sched0; 120struct mtx sched_lock; 121 122static int realstathz = 127; /* stathz is sometimes 0 and run off of hz. */ 123static int sched_tdcnt; /* Total runnable threads in the system. */ 124static int sched_slice = 12; /* Thread run time before rescheduling. */ 125 126static void setup_runqs(void); 127static void schedcpu(void); 128static void schedcpu_thread(void); 129static void sched_priority(struct thread *td, u_char prio); 130static void sched_setup(void *dummy); 131static void maybe_resched(struct thread *td); 132static void updatepri(struct thread *td); 133static void resetpriority(struct thread *td); 134static void resetpriority_thread(struct thread *td); 135#ifdef SMP 136static int sched_pickcpu(struct thread *td); 137static int forward_wakeup(int cpunum); 138static void kick_other_cpu(int pri, int cpuid); 139#endif 140 141static struct kproc_desc sched_kp = { 142 "schedcpu", 143 schedcpu_thread, 144 NULL 145}; 146SYSINIT(schedcpu, SI_SUB_LAST, SI_ORDER_FIRST, kproc_start, 147 &sched_kp); 148SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 149 150static void sched_initticks(void *dummy); 151SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 152 NULL); 153 154/* 155 * Global run queue. 156 */ 157static struct runq runq; 158 159#ifdef SMP 160/* 161 * Per-CPU run queues 162 */ 163static struct runq runq_pcpu[MAXCPU]; 164long runq_length[MAXCPU]; 165 166static cpuset_t idle_cpus_mask; 167#endif 168 169struct pcpuidlestat { 170 u_int idlecalls; 171 u_int oldidlecalls; 172}; 173static DPCPU_DEFINE(struct pcpuidlestat, idlestat); 174 175static void 176setup_runqs(void) 177{ 178#ifdef SMP 179 int i; 180 181 for (i = 0; i < MAXCPU; ++i) 182 runq_init(&runq_pcpu[i]); 183#endif 184 185 runq_init(&runq); 186} 187 188static int 189sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 190{ 191 int error, new_val, period; 192 193 period = 1000000 / realstathz; 194 new_val = period * sched_slice; 195 error = sysctl_handle_int(oidp, &new_val, 0, req); 196 if (error != 0 || req->newptr == NULL) 197 return (error); 198 if (new_val <= 0) 199 return (EINVAL); 200 sched_slice = imax(1, (new_val + period / 2) / period); 201 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 202 realstathz); 203 return (0); 204} 205 206SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler"); 207 208SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0, 209 "Scheduler name"); 210SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW, 211 NULL, 0, sysctl_kern_quantum, "I", 212 "Quantum for timeshare threads in microseconds"); 213SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 214 "Quantum for timeshare threads in stathz ticks"); 215#ifdef SMP 216/* Enable forwarding of wakeups to all other cpus */ 217static SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, 218 "Kernel SMP"); 219 220static int runq_fuzz = 1; 221SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, ""); 222 223static int forward_wakeup_enabled = 1; 224SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW, 225 &forward_wakeup_enabled, 0, 226 "Forwarding of wakeup to idle CPUs"); 227 228static int forward_wakeups_requested = 0; 229SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD, 230 &forward_wakeups_requested, 0, 231 "Requests for Forwarding of wakeup to idle CPUs"); 232 233static int forward_wakeups_delivered = 0; 234SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD, 235 &forward_wakeups_delivered, 0, 236 "Completed Forwarding of wakeup to idle CPUs"); 237 238static int forward_wakeup_use_mask = 1; 239SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW, 240 &forward_wakeup_use_mask, 0, 241 "Use the mask of idle cpus"); 242 243static int forward_wakeup_use_loop = 0; 244SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW, 245 &forward_wakeup_use_loop, 0, 246 "Use a loop to find idle cpus"); 247 248#endif 249#if 0 250static int sched_followon = 0; 251SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW, 252 &sched_followon, 0, 253 "allow threads to share a quantum"); 254#endif 255 256SDT_PROVIDER_DEFINE(sched); 257 258SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *", 259 "struct proc *", "uint8_t"); 260SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *", 261 "struct proc *", "void *"); 262SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *", 263 "struct proc *", "void *", "int"); 264SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *", 265 "struct proc *", "uint8_t", "struct thread *"); 266SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int"); 267SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *", 268 "struct proc *"); 269SDT_PROBE_DEFINE(sched, , , on__cpu); 270SDT_PROBE_DEFINE(sched, , , remain__cpu); 271SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *", 272 "struct proc *"); 273 274static __inline void 275sched_load_add(void) 276{ 277 278 sched_tdcnt++; 279 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt); 280 SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt); 281} 282 283static __inline void 284sched_load_rem(void) 285{ 286 287 sched_tdcnt--; 288 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt); 289 SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt); 290} 291/* 292 * Arrange to reschedule if necessary, taking the priorities and 293 * schedulers into account. 294 */ 295static void 296maybe_resched(struct thread *td) 297{ 298 299 THREAD_LOCK_ASSERT(td, MA_OWNED); 300 if (td->td_priority < curthread->td_priority) 301 curthread->td_flags |= TDF_NEEDRESCHED; 302} 303 304/* 305 * This function is called when a thread is about to be put on run queue 306 * because it has been made runnable or its priority has been adjusted. It 307 * determines if the new thread should be immediately preempted to. If so, 308 * it switches to it and eventually returns true. If not, it returns false 309 * so that the caller may place the thread on an appropriate run queue. 310 */ 311int 312maybe_preempt(struct thread *td) 313{ 314#ifdef PREEMPTION 315 struct thread *ctd; 316 int cpri, pri; 317 318 /* 319 * The new thread should not preempt the current thread if any of the 320 * following conditions are true: 321 * 322 * - The kernel is in the throes of crashing (panicstr). 323 * - The current thread has a higher (numerically lower) or 324 * equivalent priority. Note that this prevents curthread from 325 * trying to preempt to itself. 326 * - It is too early in the boot for context switches (cold is set). 327 * - The current thread has an inhibitor set or is in the process of 328 * exiting. In this case, the current thread is about to switch 329 * out anyways, so there's no point in preempting. If we did, 330 * the current thread would not be properly resumed as well, so 331 * just avoid that whole landmine. 332 * - If the new thread's priority is not a realtime priority and 333 * the current thread's priority is not an idle priority and 334 * FULL_PREEMPTION is disabled. 335 * 336 * If all of these conditions are false, but the current thread is in 337 * a nested critical section, then we have to defer the preemption 338 * until we exit the critical section. Otherwise, switch immediately 339 * to the new thread. 340 */ 341 ctd = curthread; 342 THREAD_LOCK_ASSERT(td, MA_OWNED); 343 KASSERT((td->td_inhibitors == 0), 344 ("maybe_preempt: trying to run inhibited thread")); 345 pri = td->td_priority; 346 cpri = ctd->td_priority; 347 if (panicstr != NULL || pri >= cpri || cold /* || dumping */ || 348 TD_IS_INHIBITED(ctd)) 349 return (0); 350#ifndef FULL_PREEMPTION 351 if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE) 352 return (0); 353#endif 354 355 if (ctd->td_critnest > 1) { 356 CTR1(KTR_PROC, "maybe_preempt: in critical section %d", 357 ctd->td_critnest); 358 ctd->td_owepreempt = 1; 359 return (0); 360 } 361 /* 362 * Thread is runnable but not yet put on system run queue. 363 */ 364 MPASS(ctd->td_lock == td->td_lock); 365 MPASS(TD_ON_RUNQ(td)); 366 TD_SET_RUNNING(td); 367 CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td, 368 td->td_proc->p_pid, td->td_name); 369 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, td); 370 /* 371 * td's lock pointer may have changed. We have to return with it 372 * locked. 373 */ 374 spinlock_enter(); 375 thread_unlock(ctd); 376 thread_lock(td); 377 spinlock_exit(); 378 return (1); 379#else 380 return (0); 381#endif 382} 383 384/* 385 * Constants for digital decay and forget: 386 * 90% of (td_estcpu) usage in 5 * loadav time 387 * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive) 388 * Note that, as ps(1) mentions, this can let percentages 389 * total over 100% (I've seen 137.9% for 3 processes). 390 * 391 * Note that schedclock() updates td_estcpu and p_cpticks asynchronously. 392 * 393 * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds. 394 * That is, the system wants to compute a value of decay such 395 * that the following for loop: 396 * for (i = 0; i < (5 * loadavg); i++) 397 * td_estcpu *= decay; 398 * will compute 399 * td_estcpu *= 0.1; 400 * for all values of loadavg: 401 * 402 * Mathematically this loop can be expressed by saying: 403 * decay ** (5 * loadavg) ~= .1 404 * 405 * The system computes decay as: 406 * decay = (2 * loadavg) / (2 * loadavg + 1) 407 * 408 * We wish to prove that the system's computation of decay 409 * will always fulfill the equation: 410 * decay ** (5 * loadavg) ~= .1 411 * 412 * If we compute b as: 413 * b = 2 * loadavg 414 * then 415 * decay = b / (b + 1) 416 * 417 * We now need to prove two things: 418 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 419 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 420 * 421 * Facts: 422 * For x close to zero, exp(x) =~ 1 + x, since 423 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 424 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 425 * For x close to zero, ln(1+x) =~ x, since 426 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 427 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 428 * ln(.1) =~ -2.30 429 * 430 * Proof of (1): 431 * Solve (factor)**(power) =~ .1 given power (5*loadav): 432 * solving for factor, 433 * ln(factor) =~ (-2.30/5*loadav), or 434 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 435 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 436 * 437 * Proof of (2): 438 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 439 * solving for power, 440 * power*ln(b/(b+1)) =~ -2.30, or 441 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 442 * 443 * Actual power values for the implemented algorithm are as follows: 444 * loadav: 1 2 3 4 445 * power: 5.68 10.32 14.94 19.55 446 */ 447 448/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 449#define loadfactor(loadav) (2 * (loadav)) 450#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 451 452/* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 453static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 454SYSCTL_UINT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 455 456/* 457 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 458 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 459 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 460 * 461 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 462 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 463 * 464 * If you don't want to bother with the faster/more-accurate formula, you 465 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 466 * (more general) method of calculating the %age of CPU used by a process. 467 */ 468#define CCPU_SHIFT 11 469 470/* 471 * Recompute process priorities, every hz ticks. 472 * MP-safe, called without the Giant mutex. 473 */ 474/* ARGSUSED */ 475static void 476schedcpu(void) 477{ 478 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 479 struct thread *td; 480 struct proc *p; 481 struct td_sched *ts; 482 int awake; 483 484 sx_slock(&allproc_lock); 485 FOREACH_PROC_IN_SYSTEM(p) { 486 PROC_LOCK(p); 487 if (p->p_state == PRS_NEW) { 488 PROC_UNLOCK(p); 489 continue; 490 } 491 FOREACH_THREAD_IN_PROC(p, td) { 492 awake = 0; 493 thread_lock(td); 494 ts = td->td_sched; 495 /* 496 * Increment sleep time (if sleeping). We 497 * ignore overflow, as above. 498 */ 499 /* 500 * The td_sched slptimes are not touched in wakeup 501 * because the thread may not HAVE everything in 502 * memory? XXX I think this is out of date. 503 */ 504 if (TD_ON_RUNQ(td)) { 505 awake = 1; 506 td->td_flags &= ~TDF_DIDRUN; 507 } else if (TD_IS_RUNNING(td)) { 508 awake = 1; 509 /* Do not clear TDF_DIDRUN */ 510 } else if (td->td_flags & TDF_DIDRUN) { 511 awake = 1; 512 td->td_flags &= ~TDF_DIDRUN; 513 } 514 515 /* 516 * ts_pctcpu is only for ps and ttyinfo(). 517 */ 518 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT; 519 /* 520 * If the td_sched has been idle the entire second, 521 * stop recalculating its priority until 522 * it wakes up. 523 */ 524 if (ts->ts_cpticks != 0) { 525#if (FSHIFT >= CCPU_SHIFT) 526 ts->ts_pctcpu += (realstathz == 100) 527 ? ((fixpt_t) ts->ts_cpticks) << 528 (FSHIFT - CCPU_SHIFT) : 529 100 * (((fixpt_t) ts->ts_cpticks) 530 << (FSHIFT - CCPU_SHIFT)) / realstathz; 531#else 532 ts->ts_pctcpu += ((FSCALE - ccpu) * 533 (ts->ts_cpticks * 534 FSCALE / realstathz)) >> FSHIFT; 535#endif 536 ts->ts_cpticks = 0; 537 } 538 /* 539 * If there are ANY running threads in this process, 540 * then don't count it as sleeping. 541 * XXX: this is broken. 542 */ 543 if (awake) { 544 if (ts->ts_slptime > 1) { 545 /* 546 * In an ideal world, this should not 547 * happen, because whoever woke us 548 * up from the long sleep should have 549 * unwound the slptime and reset our 550 * priority before we run at the stale 551 * priority. Should KASSERT at some 552 * point when all the cases are fixed. 553 */ 554 updatepri(td); 555 } 556 ts->ts_slptime = 0; 557 } else 558 ts->ts_slptime++; 559 if (ts->ts_slptime > 1) { 560 thread_unlock(td); 561 continue; 562 } 563 td->td_estcpu = decay_cpu(loadfac, td->td_estcpu); 564 resetpriority(td); 565 resetpriority_thread(td); 566 thread_unlock(td); 567 } 568 PROC_UNLOCK(p); 569 } 570 sx_sunlock(&allproc_lock); 571} 572 573/* 574 * Main loop for a kthread that executes schedcpu once a second. 575 */ 576static void 577schedcpu_thread(void) 578{ 579 580 for (;;) { 581 schedcpu(); 582 pause("-", hz); 583 } 584} 585 586/* 587 * Recalculate the priority of a process after it has slept for a while. 588 * For all load averages >= 1 and max td_estcpu of 255, sleeping for at 589 * least six times the loadfactor will decay td_estcpu to zero. 590 */ 591static void 592updatepri(struct thread *td) 593{ 594 struct td_sched *ts; 595 fixpt_t loadfac; 596 unsigned int newcpu; 597 598 ts = td->td_sched; 599 loadfac = loadfactor(averunnable.ldavg[0]); 600 if (ts->ts_slptime > 5 * loadfac) 601 td->td_estcpu = 0; 602 else { 603 newcpu = td->td_estcpu; 604 ts->ts_slptime--; /* was incremented in schedcpu() */ 605 while (newcpu && --ts->ts_slptime) 606 newcpu = decay_cpu(loadfac, newcpu); 607 td->td_estcpu = newcpu; 608 } 609} 610 611/* 612 * Compute the priority of a process when running in user mode. 613 * Arrange to reschedule if the resulting priority is better 614 * than that of the current process. 615 */ 616static void 617resetpriority(struct thread *td) 618{ 619 register unsigned int newpriority; 620 621 if (td->td_pri_class == PRI_TIMESHARE) { 622 newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT + 623 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN); 624 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 625 PRI_MAX_TIMESHARE); 626 sched_user_prio(td, newpriority); 627 } 628} 629 630/* 631 * Update the thread's priority when the associated process's user 632 * priority changes. 633 */ 634static void 635resetpriority_thread(struct thread *td) 636{ 637 638 /* Only change threads with a time sharing user priority. */ 639 if (td->td_priority < PRI_MIN_TIMESHARE || 640 td->td_priority > PRI_MAX_TIMESHARE) 641 return; 642 643 /* XXX the whole needresched thing is broken, but not silly. */ 644 maybe_resched(td); 645 646 sched_prio(td, td->td_user_pri); 647} 648 649/* ARGSUSED */ 650static void 651sched_setup(void *dummy) 652{ 653 654 setup_runqs(); 655 656 /* Account for thread0. */ 657 sched_load_add(); 658} 659 660/* 661 * This routine determines time constants after stathz and hz are setup. 662 */ 663static void 664sched_initticks(void *dummy) 665{ 666 667 realstathz = stathz ? stathz : hz; 668 sched_slice = realstathz / 10; /* ~100ms */ 669 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 670 realstathz); 671} 672 673/* External interfaces start here */ 674 675/* 676 * Very early in the boot some setup of scheduler-specific 677 * parts of proc0 and of some scheduler resources needs to be done. 678 * Called from: 679 * proc0_init() 680 */ 681void 682schedinit(void) 683{ 684 /* 685 * Set up the scheduler specific parts of proc0. 686 */ 687 proc0.p_sched = NULL; /* XXX */ 688 thread0.td_sched = &td_sched0; 689 thread0.td_lock = &sched_lock; 690 td_sched0.ts_slice = sched_slice; 691 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 692} 693 694int 695sched_runnable(void) 696{ 697#ifdef SMP 698 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]); 699#else 700 return runq_check(&runq); 701#endif 702} 703 704int 705sched_rr_interval(void) 706{ 707 708 /* Convert sched_slice from stathz to hz. */ 709 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz)); 710} 711 712/* 713 * We adjust the priority of the current process. The priority of 714 * a process gets worse as it accumulates CPU time. The cpu usage 715 * estimator (td_estcpu) is increased here. resetpriority() will 716 * compute a different priority each time td_estcpu increases by 717 * INVERSE_ESTCPU_WEIGHT 718 * (until MAXPRI is reached). The cpu usage estimator ramps up 719 * quite quickly when the process is running (linearly), and decays 720 * away exponentially, at a rate which is proportionally slower when 721 * the system is busy. The basic principle is that the system will 722 * 90% forget that the process used a lot of CPU time in 5 * loadav 723 * seconds. This causes the system to favor processes which haven't 724 * run much recently, and to round-robin among other processes. 725 */ 726void 727sched_clock(struct thread *td) 728{ 729 struct pcpuidlestat *stat; 730 struct td_sched *ts; 731 732 THREAD_LOCK_ASSERT(td, MA_OWNED); 733 ts = td->td_sched; 734 735 ts->ts_cpticks++; 736 td->td_estcpu = ESTCPULIM(td->td_estcpu + 1); 737 if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 738 resetpriority(td); 739 resetpriority_thread(td); 740 } 741 742 /* 743 * Force a context switch if the current thread has used up a full 744 * time slice (default is 100ms). 745 */ 746 if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) { 747 ts->ts_slice = sched_slice; 748 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND; 749 } 750 751 stat = DPCPU_PTR(idlestat); 752 stat->oldidlecalls = stat->idlecalls; 753 stat->idlecalls = 0; 754} 755 756/* 757 * Charge child's scheduling CPU usage to parent. 758 */ 759void 760sched_exit(struct proc *p, struct thread *td) 761{ 762 763 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit", 764 "prio:%d", td->td_priority); 765 766 PROC_LOCK_ASSERT(p, MA_OWNED); 767 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td); 768} 769 770void 771sched_exit_thread(struct thread *td, struct thread *child) 772{ 773 774 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit", 775 "prio:%d", child->td_priority); 776 thread_lock(td); 777 td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu); 778 thread_unlock(td); 779 thread_lock(child); 780 if ((child->td_flags & TDF_NOLOAD) == 0) 781 sched_load_rem(); 782 thread_unlock(child); 783} 784 785void 786sched_fork(struct thread *td, struct thread *childtd) 787{ 788 sched_fork_thread(td, childtd); 789} 790 791void 792sched_fork_thread(struct thread *td, struct thread *childtd) 793{ 794 struct td_sched *ts; 795 796 childtd->td_oncpu = NOCPU; 797 childtd->td_lastcpu = NOCPU; 798 childtd->td_estcpu = td->td_estcpu; 799 childtd->td_lock = &sched_lock; 800 childtd->td_cpuset = cpuset_ref(td->td_cpuset); 801 childtd->td_priority = childtd->td_base_pri; 802 ts = childtd->td_sched; 803 bzero(ts, sizeof(*ts)); 804 ts->ts_flags |= (td->td_sched->ts_flags & TSF_AFFINITY); 805 ts->ts_slice = 1; 806} 807 808void 809sched_nice(struct proc *p, int nice) 810{ 811 struct thread *td; 812 813 PROC_LOCK_ASSERT(p, MA_OWNED); 814 p->p_nice = nice; 815 FOREACH_THREAD_IN_PROC(p, td) { 816 thread_lock(td); 817 resetpriority(td); 818 resetpriority_thread(td); 819 thread_unlock(td); 820 } 821} 822 823void 824sched_class(struct thread *td, int class) 825{ 826 THREAD_LOCK_ASSERT(td, MA_OWNED); 827 td->td_pri_class = class; 828} 829 830/* 831 * Adjust the priority of a thread. 832 */ 833static void 834sched_priority(struct thread *td, u_char prio) 835{ 836 837 838 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change", 839 "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED, 840 sched_tdname(curthread)); 841 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio); 842 if (td != curthread && prio > td->td_priority) { 843 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 844 "lend prio", "prio:%d", td->td_priority, "new prio:%d", 845 prio, KTR_ATTR_LINKED, sched_tdname(td)); 846 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio, 847 curthread); 848 } 849 THREAD_LOCK_ASSERT(td, MA_OWNED); 850 if (td->td_priority == prio) 851 return; 852 td->td_priority = prio; 853 if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) { 854 sched_rem(td); 855 sched_add(td, SRQ_BORING); 856 } 857} 858 859/* 860 * Update a thread's priority when it is lent another thread's 861 * priority. 862 */ 863void 864sched_lend_prio(struct thread *td, u_char prio) 865{ 866 867 td->td_flags |= TDF_BORROWING; 868 sched_priority(td, prio); 869} 870 871/* 872 * Restore a thread's priority when priority propagation is 873 * over. The prio argument is the minimum priority the thread 874 * needs to have to satisfy other possible priority lending 875 * requests. If the thread's regulary priority is less 876 * important than prio the thread will keep a priority boost 877 * of prio. 878 */ 879void 880sched_unlend_prio(struct thread *td, u_char prio) 881{ 882 u_char base_pri; 883 884 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 885 td->td_base_pri <= PRI_MAX_TIMESHARE) 886 base_pri = td->td_user_pri; 887 else 888 base_pri = td->td_base_pri; 889 if (prio >= base_pri) { 890 td->td_flags &= ~TDF_BORROWING; 891 sched_prio(td, base_pri); 892 } else 893 sched_lend_prio(td, prio); 894} 895 896void 897sched_prio(struct thread *td, u_char prio) 898{ 899 u_char oldprio; 900 901 /* First, update the base priority. */ 902 td->td_base_pri = prio; 903 904 /* 905 * If the thread is borrowing another thread's priority, don't ever 906 * lower the priority. 907 */ 908 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 909 return; 910 911 /* Change the real priority. */ 912 oldprio = td->td_priority; 913 sched_priority(td, prio); 914 915 /* 916 * If the thread is on a turnstile, then let the turnstile update 917 * its state. 918 */ 919 if (TD_ON_LOCK(td) && oldprio != prio) 920 turnstile_adjust(td, oldprio); 921} 922 923void 924sched_user_prio(struct thread *td, u_char prio) 925{ 926 927 THREAD_LOCK_ASSERT(td, MA_OWNED); 928 td->td_base_user_pri = prio; 929 if (td->td_lend_user_pri <= prio) 930 return; 931 td->td_user_pri = prio; 932} 933 934void 935sched_lend_user_prio(struct thread *td, u_char prio) 936{ 937 938 THREAD_LOCK_ASSERT(td, MA_OWNED); 939 td->td_lend_user_pri = prio; 940 td->td_user_pri = min(prio, td->td_base_user_pri); 941 if (td->td_priority > td->td_user_pri) 942 sched_prio(td, td->td_user_pri); 943 else if (td->td_priority != td->td_user_pri) 944 td->td_flags |= TDF_NEEDRESCHED; 945} 946 947void 948sched_sleep(struct thread *td, int pri) 949{ 950 951 THREAD_LOCK_ASSERT(td, MA_OWNED); 952 td->td_slptick = ticks; 953 td->td_sched->ts_slptime = 0; 954 if (pri != 0 && PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 955 sched_prio(td, pri); 956 if (TD_IS_SUSPENDED(td) || pri >= PSOCK) 957 td->td_flags |= TDF_CANSWAP; 958} 959 960void 961sched_switch(struct thread *td, struct thread *newtd, int flags) 962{ 963 struct mtx *tmtx; 964 struct td_sched *ts; 965 struct proc *p; 966 int preempted; 967 968 tmtx = NULL; 969 ts = td->td_sched; 970 p = td->td_proc; 971 972 THREAD_LOCK_ASSERT(td, MA_OWNED); 973 974 /* 975 * Switch to the sched lock to fix things up and pick 976 * a new thread. 977 * Block the td_lock in order to avoid breaking the critical path. 978 */ 979 if (td->td_lock != &sched_lock) { 980 mtx_lock_spin(&sched_lock); 981 tmtx = thread_lock_block(td); 982 } 983 984 if ((td->td_flags & TDF_NOLOAD) == 0) 985 sched_load_rem(); 986 987 td->td_lastcpu = td->td_oncpu; 988 preempted = !((td->td_flags & TDF_SLICEEND) || 989 (flags & SWT_RELINQUISH)); 990 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND); 991 td->td_owepreempt = 0; 992 td->td_oncpu = NOCPU; 993 994 /* 995 * At the last moment, if this thread is still marked RUNNING, 996 * then put it back on the run queue as it has not been suspended 997 * or stopped or any thing else similar. We never put the idle 998 * threads on the run queue, however. 999 */ 1000 if (td->td_flags & TDF_IDLETD) { 1001 TD_SET_CAN_RUN(td); 1002#ifdef SMP 1003 CPU_CLR(PCPU_GET(cpuid), &idle_cpus_mask); 1004#endif 1005 } else { 1006 if (TD_IS_RUNNING(td)) { 1007 /* Put us back on the run queue. */ 1008 sched_add(td, preempted ? 1009 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1010 SRQ_OURSELF|SRQ_YIELDING); 1011 } 1012 } 1013 if (newtd) { 1014 /* 1015 * The thread we are about to run needs to be counted 1016 * as if it had been added to the run queue and selected. 1017 * It came from: 1018 * * A preemption 1019 * * An upcall 1020 * * A followon 1021 */ 1022 KASSERT((newtd->td_inhibitors == 0), 1023 ("trying to run inhibited thread")); 1024 newtd->td_flags |= TDF_DIDRUN; 1025 TD_SET_RUNNING(newtd); 1026 if ((newtd->td_flags & TDF_NOLOAD) == 0) 1027 sched_load_add(); 1028 } else { 1029 newtd = choosethread(); 1030 MPASS(newtd->td_lock == &sched_lock); 1031 } 1032 1033 if (td != newtd) { 1034#ifdef HWPMC_HOOKS 1035 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1036 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1037#endif 1038 1039 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc); 1040 1041 /* I feel sleepy */ 1042 lock_profile_release_lock(&sched_lock.lock_object); 1043#ifdef KDTRACE_HOOKS 1044 /* 1045 * If DTrace has set the active vtime enum to anything 1046 * other than INACTIVE (0), then it should have set the 1047 * function to call. 1048 */ 1049 if (dtrace_vtime_active) 1050 (*dtrace_vtime_switch_func)(newtd); 1051#endif 1052 1053 cpu_switch(td, newtd, tmtx != NULL ? tmtx : td->td_lock); 1054 lock_profile_obtain_lock_success(&sched_lock.lock_object, 1055 0, 0, __FILE__, __LINE__); 1056 /* 1057 * Where am I? What year is it? 1058 * We are in the same thread that went to sleep above, 1059 * but any amount of time may have passed. All our context 1060 * will still be available as will local variables. 1061 * PCPU values however may have changed as we may have 1062 * changed CPU so don't trust cached values of them. 1063 * New threads will go to fork_exit() instead of here 1064 * so if you change things here you may need to change 1065 * things there too. 1066 * 1067 * If the thread above was exiting it will never wake 1068 * up again here, so either it has saved everything it 1069 * needed to, or the thread_wait() or wait() will 1070 * need to reap it. 1071 */ 1072 1073 SDT_PROBE0(sched, , , on__cpu); 1074#ifdef HWPMC_HOOKS 1075 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1076 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1077#endif 1078 } else 1079 SDT_PROBE0(sched, , , remain__cpu); 1080 1081#ifdef SMP 1082 if (td->td_flags & TDF_IDLETD) 1083 CPU_SET(PCPU_GET(cpuid), &idle_cpus_mask); 1084#endif 1085 sched_lock.mtx_lock = (uintptr_t)td; 1086 td->td_oncpu = PCPU_GET(cpuid); 1087 MPASS(td->td_lock == &sched_lock); 1088} 1089 1090void 1091sched_wakeup(struct thread *td) 1092{ 1093 struct td_sched *ts; 1094 1095 THREAD_LOCK_ASSERT(td, MA_OWNED); 1096 ts = td->td_sched; 1097 td->td_flags &= ~TDF_CANSWAP; 1098 if (ts->ts_slptime > 1) { 1099 updatepri(td); 1100 resetpriority(td); 1101 } 1102 td->td_slptick = 0; 1103 ts->ts_slptime = 0; 1104 ts->ts_slice = sched_slice; 1105 sched_add(td, SRQ_BORING); 1106} 1107 1108#ifdef SMP 1109static int 1110forward_wakeup(int cpunum) 1111{ 1112 struct pcpu *pc; 1113 cpuset_t dontuse, map, map2; 1114 u_int id, me; 1115 int iscpuset; 1116 1117 mtx_assert(&sched_lock, MA_OWNED); 1118 1119 CTR0(KTR_RUNQ, "forward_wakeup()"); 1120 1121 if ((!forward_wakeup_enabled) || 1122 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0)) 1123 return (0); 1124 if (!smp_started || cold || panicstr) 1125 return (0); 1126 1127 forward_wakeups_requested++; 1128 1129 /* 1130 * Check the idle mask we received against what we calculated 1131 * before in the old version. 1132 */ 1133 me = PCPU_GET(cpuid); 1134 1135 /* Don't bother if we should be doing it ourself. */ 1136 if (CPU_ISSET(me, &idle_cpus_mask) && 1137 (cpunum == NOCPU || me == cpunum)) 1138 return (0); 1139 1140 CPU_SETOF(me, &dontuse); 1141 CPU_OR(&dontuse, &stopped_cpus); 1142 CPU_OR(&dontuse, &hlt_cpus_mask); 1143 CPU_ZERO(&map2); 1144 if (forward_wakeup_use_loop) { 1145 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { 1146 id = pc->pc_cpuid; 1147 if (!CPU_ISSET(id, &dontuse) && 1148 pc->pc_curthread == pc->pc_idlethread) { 1149 CPU_SET(id, &map2); 1150 } 1151 } 1152 } 1153 1154 if (forward_wakeup_use_mask) { 1155 map = idle_cpus_mask; 1156 CPU_NAND(&map, &dontuse); 1157 1158 /* If they are both on, compare and use loop if different. */ 1159 if (forward_wakeup_use_loop) { 1160 if (CPU_CMP(&map, &map2)) { 1161 printf("map != map2, loop method preferred\n"); 1162 map = map2; 1163 } 1164 } 1165 } else { 1166 map = map2; 1167 } 1168 1169 /* If we only allow a specific CPU, then mask off all the others. */ 1170 if (cpunum != NOCPU) { 1171 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum.")); 1172 iscpuset = CPU_ISSET(cpunum, &map); 1173 if (iscpuset == 0) 1174 CPU_ZERO(&map); 1175 else 1176 CPU_SETOF(cpunum, &map); 1177 } 1178 if (!CPU_EMPTY(&map)) { 1179 forward_wakeups_delivered++; 1180 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { 1181 id = pc->pc_cpuid; 1182 if (!CPU_ISSET(id, &map)) 1183 continue; 1184 if (cpu_idle_wakeup(pc->pc_cpuid)) 1185 CPU_CLR(id, &map); 1186 } 1187 if (!CPU_EMPTY(&map)) 1188 ipi_selected(map, IPI_AST); 1189 return (1); 1190 } 1191 if (cpunum == NOCPU) 1192 printf("forward_wakeup: Idle processor not found\n"); 1193 return (0); 1194} 1195 1196static void 1197kick_other_cpu(int pri, int cpuid) 1198{ 1199 struct pcpu *pcpu; 1200 int cpri; 1201 1202 pcpu = pcpu_find(cpuid); 1203 if (CPU_ISSET(cpuid, &idle_cpus_mask)) { 1204 forward_wakeups_delivered++; 1205 if (!cpu_idle_wakeup(cpuid)) 1206 ipi_cpu(cpuid, IPI_AST); 1207 return; 1208 } 1209 1210 cpri = pcpu->pc_curthread->td_priority; 1211 if (pri >= cpri) 1212 return; 1213 1214#if defined(IPI_PREEMPTION) && defined(PREEMPTION) 1215#if !defined(FULL_PREEMPTION) 1216 if (pri <= PRI_MAX_ITHD) 1217#endif /* ! FULL_PREEMPTION */ 1218 { 1219 ipi_cpu(cpuid, IPI_PREEMPT); 1220 return; 1221 } 1222#endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */ 1223 1224 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED; 1225 ipi_cpu(cpuid, IPI_AST); 1226 return; 1227} 1228#endif /* SMP */ 1229 1230#ifdef SMP 1231static int 1232sched_pickcpu(struct thread *td) 1233{ 1234 int best, cpu; 1235 1236 mtx_assert(&sched_lock, MA_OWNED); 1237 1238 if (THREAD_CAN_SCHED(td, td->td_lastcpu)) 1239 best = td->td_lastcpu; 1240 else 1241 best = NOCPU; 1242 CPU_FOREACH(cpu) { 1243 if (!THREAD_CAN_SCHED(td, cpu)) 1244 continue; 1245 1246 if (best == NOCPU) 1247 best = cpu; 1248 else if (runq_length[cpu] < runq_length[best]) 1249 best = cpu; 1250 } 1251 KASSERT(best != NOCPU, ("no valid CPUs")); 1252 1253 return (best); 1254} 1255#endif 1256 1257void 1258sched_add(struct thread *td, int flags) 1259#ifdef SMP 1260{ 1261 cpuset_t tidlemsk; 1262 struct td_sched *ts; 1263 u_int cpu, cpuid; 1264 int forwarded = 0; 1265 int single_cpu = 0; 1266 1267 ts = td->td_sched; 1268 THREAD_LOCK_ASSERT(td, MA_OWNED); 1269 KASSERT((td->td_inhibitors == 0), 1270 ("sched_add: trying to run inhibited thread")); 1271 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 1272 ("sched_add: bad thread state")); 1273 KASSERT(td->td_flags & TDF_INMEM, 1274 ("sched_add: thread swapped out")); 1275 1276 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 1277 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 1278 sched_tdname(curthread)); 1279 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 1280 KTR_ATTR_LINKED, sched_tdname(td)); 1281 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 1282 flags & SRQ_PREEMPTED); 1283 1284 1285 /* 1286 * Now that the thread is moving to the run-queue, set the lock 1287 * to the scheduler's lock. 1288 */ 1289 if (td->td_lock != &sched_lock) { 1290 mtx_lock_spin(&sched_lock); 1291 thread_lock_set(td, &sched_lock); 1292 } 1293 TD_SET_RUNQ(td); 1294 1295 /* 1296 * If SMP is started and the thread is pinned or otherwise limited to 1297 * a specific set of CPUs, queue the thread to a per-CPU run queue. 1298 * Otherwise, queue the thread to the global run queue. 1299 * 1300 * If SMP has not yet been started we must use the global run queue 1301 * as per-CPU state may not be initialized yet and we may crash if we 1302 * try to access the per-CPU run queues. 1303 */ 1304 if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND || 1305 ts->ts_flags & TSF_AFFINITY)) { 1306 if (td->td_pinned != 0) 1307 cpu = td->td_lastcpu; 1308 else if (td->td_flags & TDF_BOUND) { 1309 /* Find CPU from bound runq. */ 1310 KASSERT(SKE_RUNQ_PCPU(ts), 1311 ("sched_add: bound td_sched not on cpu runq")); 1312 cpu = ts->ts_runq - &runq_pcpu[0]; 1313 } else 1314 /* Find a valid CPU for our cpuset */ 1315 cpu = sched_pickcpu(td); 1316 ts->ts_runq = &runq_pcpu[cpu]; 1317 single_cpu = 1; 1318 CTR3(KTR_RUNQ, 1319 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, 1320 cpu); 1321 } else { 1322 CTR2(KTR_RUNQ, 1323 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts, 1324 td); 1325 cpu = NOCPU; 1326 ts->ts_runq = &runq; 1327 } 1328 1329 cpuid = PCPU_GET(cpuid); 1330 if (single_cpu && cpu != cpuid) { 1331 kick_other_cpu(td->td_priority, cpu); 1332 } else { 1333 if (!single_cpu) { 1334 tidlemsk = idle_cpus_mask; 1335 CPU_NAND(&tidlemsk, &hlt_cpus_mask); 1336 CPU_CLR(cpuid, &tidlemsk); 1337 1338 if (!CPU_ISSET(cpuid, &idle_cpus_mask) && 1339 ((flags & SRQ_INTR) == 0) && 1340 !CPU_EMPTY(&tidlemsk)) 1341 forwarded = forward_wakeup(cpu); 1342 } 1343 1344 if (!forwarded) { 1345 if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td)) 1346 return; 1347 else 1348 maybe_resched(td); 1349 } 1350 } 1351 1352 if ((td->td_flags & TDF_NOLOAD) == 0) 1353 sched_load_add(); 1354 runq_add(ts->ts_runq, td, flags); 1355 if (cpu != NOCPU) 1356 runq_length[cpu]++; 1357} 1358#else /* SMP */ 1359{ 1360 struct td_sched *ts; 1361 1362 ts = td->td_sched; 1363 THREAD_LOCK_ASSERT(td, MA_OWNED); 1364 KASSERT((td->td_inhibitors == 0), 1365 ("sched_add: trying to run inhibited thread")); 1366 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 1367 ("sched_add: bad thread state")); 1368 KASSERT(td->td_flags & TDF_INMEM, 1369 ("sched_add: thread swapped out")); 1370 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 1371 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 1372 sched_tdname(curthread)); 1373 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 1374 KTR_ATTR_LINKED, sched_tdname(td)); 1375 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 1376 flags & SRQ_PREEMPTED); 1377 1378 /* 1379 * Now that the thread is moving to the run-queue, set the lock 1380 * to the scheduler's lock. 1381 */ 1382 if (td->td_lock != &sched_lock) { 1383 mtx_lock_spin(&sched_lock); 1384 thread_lock_set(td, &sched_lock); 1385 } 1386 TD_SET_RUNQ(td); 1387 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td); 1388 ts->ts_runq = &runq; 1389 1390 /* 1391 * If we are yielding (on the way out anyhow) or the thread 1392 * being saved is US, then don't try be smart about preemption 1393 * or kicking off another CPU as it won't help and may hinder. 1394 * In the YIEDLING case, we are about to run whoever is being 1395 * put in the queue anyhow, and in the OURSELF case, we are 1396 * puting ourself on the run queue which also only happens 1397 * when we are about to yield. 1398 */ 1399 if ((flags & SRQ_YIELDING) == 0) { 1400 if (maybe_preempt(td)) 1401 return; 1402 } 1403 if ((td->td_flags & TDF_NOLOAD) == 0) 1404 sched_load_add(); 1405 runq_add(ts->ts_runq, td, flags); 1406 maybe_resched(td); 1407} 1408#endif /* SMP */ 1409 1410void 1411sched_rem(struct thread *td) 1412{ 1413 struct td_sched *ts; 1414 1415 ts = td->td_sched; 1416 KASSERT(td->td_flags & TDF_INMEM, 1417 ("sched_rem: thread swapped out")); 1418 KASSERT(TD_ON_RUNQ(td), 1419 ("sched_rem: thread not on run queue")); 1420 mtx_assert(&sched_lock, MA_OWNED); 1421 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 1422 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 1423 sched_tdname(curthread)); 1424 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); 1425 1426 if ((td->td_flags & TDF_NOLOAD) == 0) 1427 sched_load_rem(); 1428#ifdef SMP 1429 if (ts->ts_runq != &runq) 1430 runq_length[ts->ts_runq - runq_pcpu]--; 1431#endif 1432 runq_remove(ts->ts_runq, td); 1433 TD_SET_CAN_RUN(td); 1434} 1435 1436/* 1437 * Select threads to run. Note that running threads still consume a 1438 * slot. 1439 */ 1440struct thread * 1441sched_choose(void) 1442{ 1443 struct thread *td; 1444 struct runq *rq; 1445 1446 mtx_assert(&sched_lock, MA_OWNED); 1447#ifdef SMP 1448 struct thread *tdcpu; 1449 1450 rq = &runq; 1451 td = runq_choose_fuzz(&runq, runq_fuzz); 1452 tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]); 1453 1454 if (td == NULL || 1455 (tdcpu != NULL && 1456 tdcpu->td_priority < td->td_priority)) { 1457 CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu, 1458 PCPU_GET(cpuid)); 1459 td = tdcpu; 1460 rq = &runq_pcpu[PCPU_GET(cpuid)]; 1461 } else { 1462 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td); 1463 } 1464 1465#else 1466 rq = &runq; 1467 td = runq_choose(&runq); 1468#endif 1469 1470 if (td) { 1471#ifdef SMP 1472 if (td == tdcpu) 1473 runq_length[PCPU_GET(cpuid)]--; 1474#endif 1475 runq_remove(rq, td); 1476 td->td_flags |= TDF_DIDRUN; 1477 1478 KASSERT(td->td_flags & TDF_INMEM, 1479 ("sched_choose: thread swapped out")); 1480 return (td); 1481 } 1482 return (PCPU_GET(idlethread)); 1483} 1484 1485void 1486sched_preempt(struct thread *td) 1487{ 1488 1489 SDT_PROBE2(sched, , , surrender, td, td->td_proc); 1490 thread_lock(td); 1491 if (td->td_critnest > 1) 1492 td->td_owepreempt = 1; 1493 else 1494 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT, NULL); 1495 thread_unlock(td); 1496} 1497 1498void 1499sched_userret(struct thread *td) 1500{ 1501 /* 1502 * XXX we cheat slightly on the locking here to avoid locking in 1503 * the usual case. Setting td_priority here is essentially an 1504 * incomplete workaround for not setting it properly elsewhere. 1505 * Now that some interrupt handlers are threads, not setting it 1506 * properly elsewhere can clobber it in the window between setting 1507 * it here and returning to user mode, so don't waste time setting 1508 * it perfectly here. 1509 */ 1510 KASSERT((td->td_flags & TDF_BORROWING) == 0, 1511 ("thread with borrowed priority returning to userland")); 1512 if (td->td_priority != td->td_user_pri) { 1513 thread_lock(td); 1514 td->td_priority = td->td_user_pri; 1515 td->td_base_pri = td->td_user_pri; 1516 thread_unlock(td); 1517 } 1518} 1519 1520void 1521sched_bind(struct thread *td, int cpu) 1522{ 1523 struct td_sched *ts; 1524 1525 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 1526 KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 1527 1528 ts = td->td_sched; 1529 1530 td->td_flags |= TDF_BOUND; 1531#ifdef SMP 1532 ts->ts_runq = &runq_pcpu[cpu]; 1533 if (PCPU_GET(cpuid) == cpu) 1534 return; 1535 1536 mi_switch(SW_VOL, NULL); 1537#endif 1538} 1539 1540void 1541sched_unbind(struct thread* td) 1542{ 1543 THREAD_LOCK_ASSERT(td, MA_OWNED); 1544 KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 1545 td->td_flags &= ~TDF_BOUND; 1546} 1547 1548int 1549sched_is_bound(struct thread *td) 1550{ 1551 THREAD_LOCK_ASSERT(td, MA_OWNED); 1552 return (td->td_flags & TDF_BOUND); 1553} 1554 1555void 1556sched_relinquish(struct thread *td) 1557{ 1558 thread_lock(td); 1559 mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 1560 thread_unlock(td); 1561} 1562 1563int 1564sched_load(void) 1565{ 1566 return (sched_tdcnt); 1567} 1568 1569int 1570sched_sizeof_proc(void) 1571{ 1572 return (sizeof(struct proc)); 1573} 1574 1575int 1576sched_sizeof_thread(void) 1577{ 1578 return (sizeof(struct thread) + sizeof(struct td_sched)); 1579} 1580 1581fixpt_t 1582sched_pctcpu(struct thread *td) 1583{ 1584 struct td_sched *ts; 1585 1586 THREAD_LOCK_ASSERT(td, MA_OWNED); 1587 ts = td->td_sched; 1588 return (ts->ts_pctcpu); 1589} 1590 1591#ifdef RACCT 1592/* 1593 * Calculates the contribution to the thread cpu usage for the latest 1594 * (unfinished) second. 1595 */ 1596fixpt_t 1597sched_pctcpu_delta(struct thread *td) 1598{ 1599 struct td_sched *ts; 1600 fixpt_t delta; 1601 int realstathz; 1602 1603 THREAD_LOCK_ASSERT(td, MA_OWNED); 1604 ts = td->td_sched; 1605 delta = 0; 1606 realstathz = stathz ? stathz : hz; 1607 if (ts->ts_cpticks != 0) { 1608#if (FSHIFT >= CCPU_SHIFT) 1609 delta = (realstathz == 100) 1610 ? ((fixpt_t) ts->ts_cpticks) << 1611 (FSHIFT - CCPU_SHIFT) : 1612 100 * (((fixpt_t) ts->ts_cpticks) 1613 << (FSHIFT - CCPU_SHIFT)) / realstathz; 1614#else 1615 delta = ((FSCALE - ccpu) * 1616 (ts->ts_cpticks * 1617 FSCALE / realstathz)) >> FSHIFT; 1618#endif 1619 } 1620 1621 return (delta); 1622} 1623#endif 1624 1625void 1626sched_tick(int cnt) 1627{ 1628} 1629 1630/* 1631 * The actual idle process. 1632 */ 1633void 1634sched_idletd(void *dummy) 1635{ 1636 struct pcpuidlestat *stat; 1637 1638 THREAD_NO_SLEEPING(); 1639 stat = DPCPU_PTR(idlestat); 1640 for (;;) { 1641 mtx_assert(&Giant, MA_NOTOWNED); 1642 1643 while (sched_runnable() == 0) { 1644 cpu_idle(stat->idlecalls + stat->oldidlecalls > 64); 1645 stat->idlecalls++; 1646 } 1647 1648 mtx_lock_spin(&sched_lock); 1649 mi_switch(SW_VOL | SWT_IDLE, NULL); 1650 mtx_unlock_spin(&sched_lock); 1651 } 1652} 1653 1654/* 1655 * A CPU is entering for the first time or a thread is exiting. 1656 */ 1657void 1658sched_throw(struct thread *td) 1659{ 1660 /* 1661 * Correct spinlock nesting. The idle thread context that we are 1662 * borrowing was created so that it would start out with a single 1663 * spin lock (sched_lock) held in fork_trampoline(). Since we've 1664 * explicitly acquired locks in this function, the nesting count 1665 * is now 2 rather than 1. Since we are nested, calling 1666 * spinlock_exit() will simply adjust the counts without allowing 1667 * spin lock using code to interrupt us. 1668 */ 1669 if (td == NULL) { 1670 mtx_lock_spin(&sched_lock); 1671 spinlock_exit(); 1672 PCPU_SET(switchtime, cpu_ticks()); 1673 PCPU_SET(switchticks, ticks); 1674 } else { 1675 lock_profile_release_lock(&sched_lock.lock_object); 1676 MPASS(td->td_lock == &sched_lock); 1677 td->td_lastcpu = td->td_oncpu; 1678 td->td_oncpu = NOCPU; 1679 } 1680 mtx_assert(&sched_lock, MA_OWNED); 1681 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 1682 cpu_throw(td, choosethread()); /* doesn't return */ 1683} 1684 1685void 1686sched_fork_exit(struct thread *td) 1687{ 1688 1689 /* 1690 * Finish setting up thread glue so that it begins execution in a 1691 * non-nested critical section with sched_lock held but not recursed. 1692 */ 1693 td->td_oncpu = PCPU_GET(cpuid); 1694 sched_lock.mtx_lock = (uintptr_t)td; 1695 lock_profile_obtain_lock_success(&sched_lock.lock_object, 1696 0, 0, __FILE__, __LINE__); 1697 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED); 1698} 1699 1700char * 1701sched_tdname(struct thread *td) 1702{ 1703#ifdef KTR 1704 struct td_sched *ts; 1705 1706 ts = td->td_sched; 1707 if (ts->ts_name[0] == '\0') 1708 snprintf(ts->ts_name, sizeof(ts->ts_name), 1709 "%s tid %d", td->td_name, td->td_tid); 1710 return (ts->ts_name); 1711#else 1712 return (td->td_name); 1713#endif 1714} 1715 1716#ifdef KTR 1717void 1718sched_clear_tdname(struct thread *td) 1719{ 1720 struct td_sched *ts; 1721 1722 ts = td->td_sched; 1723 ts->ts_name[0] = '\0'; 1724} 1725#endif 1726 1727void 1728sched_affinity(struct thread *td) 1729{ 1730#ifdef SMP 1731 struct td_sched *ts; 1732 int cpu; 1733 1734 THREAD_LOCK_ASSERT(td, MA_OWNED); 1735 1736 /* 1737 * Set the TSF_AFFINITY flag if there is at least one CPU this 1738 * thread can't run on. 1739 */ 1740 ts = td->td_sched; 1741 ts->ts_flags &= ~TSF_AFFINITY; 1742 CPU_FOREACH(cpu) { 1743 if (!THREAD_CAN_SCHED(td, cpu)) { 1744 ts->ts_flags |= TSF_AFFINITY; 1745 break; 1746 } 1747 } 1748 1749 /* 1750 * If this thread can run on all CPUs, nothing else to do. 1751 */ 1752 if (!(ts->ts_flags & TSF_AFFINITY)) 1753 return; 1754 1755 /* Pinned threads and bound threads should be left alone. */ 1756 if (td->td_pinned != 0 || td->td_flags & TDF_BOUND) 1757 return; 1758 1759 switch (td->td_state) { 1760 case TDS_RUNQ: 1761 /* 1762 * If we are on a per-CPU runqueue that is in the set, 1763 * then nothing needs to be done. 1764 */ 1765 if (ts->ts_runq != &runq && 1766 THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu)) 1767 return; 1768 1769 /* Put this thread on a valid per-CPU runqueue. */ 1770 sched_rem(td); 1771 sched_add(td, SRQ_BORING); 1772 break; 1773 case TDS_RUNNING: 1774 /* 1775 * See if our current CPU is in the set. If not, force a 1776 * context switch. 1777 */ 1778 if (THREAD_CAN_SCHED(td, td->td_oncpu)) 1779 return; 1780 1781 td->td_flags |= TDF_NEEDRESCHED; 1782 if (td != curthread) 1783 ipi_cpu(cpu, IPI_AST); 1784 break; 1785 default: 1786 break; 1787 } 1788#endif 1789} 1790