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