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 40 #include <sys/param.h> 41 #include <sys/systm.h> 42 #include <sys/kernel.h> 43 #include <sys/ktr.h> 44 #include <sys/lock.h> 45 #include <sys/kthread.h> 46 #include <sys/mutex.h> 47 #include <sys/proc.h> 48 #include <sys/resourcevar.h> 49 #include <sys/sched.h> 50 #include <sys/smp.h> 51 #include <sys/sysctl.h> 52 #include <sys/sx.h> 53 #include <sys/turnstile.h> 54 #include <sys/umtx.h> 55 #include <machine/pcb.h> 56 #include <machine/smp.h> 57 58 #ifdef HWPMC_HOOKS 59 #include <sys/pmckern.h> 60 #endif 61 62 /* 63 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in 64 * the range 100-256 Hz (approximately). 65 */ 66 #define ESTCPULIM(e) \ 67 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \ 68 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1) 69 #ifdef SMP 70 #define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus) 71 #else 72 #define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */ 73 #endif 74 #define NICE_WEIGHT 1 /* Priorities per nice level. */ 75 76 /* 77 * The schedulable entity that runs a context. 78 * This is an extension to the thread structure and is tailored to 79 * the requirements of this scheduler 80 */ 81 struct td_sched { 82 TAILQ_ENTRY(td_sched) ts_procq; /* (j/z) Run queue. */ 83 struct thread *ts_thread; /* (*) Active associated thread. */ 84 fixpt_t ts_pctcpu; /* (j) %cpu during p_swtime. */ 85 u_char ts_rqindex; /* (j) Run queue index. */ 86 int ts_cpticks; /* (j) Ticks of cpu time. */ 87 int ts_slptime; /* (j) Seconds !RUNNING. */ 88 struct runq *ts_runq; /* runq the thread is currently on */ 89 }; 90 91 /* flags kept in td_flags */ 92 #define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */ 93 #define TDF_EXIT TDF_SCHED1 /* thread is being killed. */ 94 #define TDF_BOUND TDF_SCHED2 95 96 #define ts_flags ts_thread->td_flags 97 #define TSF_DIDRUN TDF_DIDRUN /* thread actually ran. */ 98 #define TSF_EXIT TDF_EXIT /* thread is being killed. */ 99 #define TSF_BOUND TDF_BOUND /* stuck to one CPU */ 100 101 #define SKE_RUNQ_PCPU(ts) \ 102 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq) 103 104 static struct td_sched td_sched0; 105 struct mtx sched_lock; 106 107 static int sched_tdcnt; /* Total runnable threads in the system. */ 108 static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 109 #define SCHED_QUANTUM (hz / 10) /* Default sched quantum */ 110 111 static void setup_runqs(void); 112 static void schedcpu(void); 113 static void schedcpu_thread(void); 114 static void sched_priority(struct thread *td, u_char prio); 115 static void sched_setup(void *dummy); 116 static void maybe_resched(struct thread *td); 117 static void updatepri(struct thread *td); 118 static void resetpriority(struct thread *td); 119 static void resetpriority_thread(struct thread *td); 120 #ifdef SMP 121 static int forward_wakeup(int cpunum); 122 #endif 123 124 static struct kproc_desc sched_kp = { 125 "schedcpu", 126 schedcpu_thread, 127 NULL 128 }; 129 SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start, &sched_kp) 130 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL) 131 132 /* 133 * Global run queue. 134 */ 135 static struct runq runq; 136 137 #ifdef SMP 138 /* 139 * Per-CPU run queues 140 */ 141 static struct runq runq_pcpu[MAXCPU]; 142 #endif 143 144 static void 145 setup_runqs(void) 146 { 147 #ifdef SMP 148 int i; 149 150 for (i = 0; i < MAXCPU; ++i) 151 runq_init(&runq_pcpu[i]); 152 #endif 153 154 runq_init(&runq); 155 } 156 157 static int 158 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 159 { 160 int error, new_val; 161 162 new_val = sched_quantum * tick; 163 error = sysctl_handle_int(oidp, &new_val, 0, req); 164 if (error != 0 || req->newptr == NULL) 165 return (error); 166 if (new_val < tick) 167 return (EINVAL); 168 sched_quantum = new_val / tick; 169 hogticks = 2 * sched_quantum; 170 return (0); 171 } 172 173 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler"); 174 175 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0, 176 "Scheduler name"); 177 178 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW, 179 0, sizeof sched_quantum, sysctl_kern_quantum, "I", 180 "Roundrobin scheduling quantum in microseconds"); 181 182 #ifdef SMP 183 /* Enable forwarding of wakeups to all other cpus */ 184 SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP"); 185 186 static int forward_wakeup_enabled = 1; 187 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW, 188 &forward_wakeup_enabled, 0, 189 "Forwarding of wakeup to idle CPUs"); 190 191 static int forward_wakeups_requested = 0; 192 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD, 193 &forward_wakeups_requested, 0, 194 "Requests for Forwarding of wakeup to idle CPUs"); 195 196 static int forward_wakeups_delivered = 0; 197 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD, 198 &forward_wakeups_delivered, 0, 199 "Completed Forwarding of wakeup to idle CPUs"); 200 201 static int forward_wakeup_use_mask = 1; 202 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW, 203 &forward_wakeup_use_mask, 0, 204 "Use the mask of idle cpus"); 205 206 static int forward_wakeup_use_loop = 0; 207 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW, 208 &forward_wakeup_use_loop, 0, 209 "Use a loop to find idle cpus"); 210 211 static int forward_wakeup_use_single = 0; 212 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, onecpu, CTLFLAG_RW, 213 &forward_wakeup_use_single, 0, 214 "Only signal one idle cpu"); 215 216 static int forward_wakeup_use_htt = 0; 217 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, htt2, CTLFLAG_RW, 218 &forward_wakeup_use_htt, 0, 219 "account for htt"); 220 221 #endif 222 #if 0 223 static int sched_followon = 0; 224 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW, 225 &sched_followon, 0, 226 "allow threads to share a quantum"); 227 #endif 228 229 static __inline void 230 sched_load_add(void) 231 { 232 sched_tdcnt++; 233 CTR1(KTR_SCHED, "global load: %d", sched_tdcnt); 234 } 235 236 static __inline void 237 sched_load_rem(void) 238 { 239 sched_tdcnt--; 240 CTR1(KTR_SCHED, "global load: %d", sched_tdcnt); 241 } 242 /* 243 * Arrange to reschedule if necessary, taking the priorities and 244 * schedulers into account. 245 */ 246 static void 247 maybe_resched(struct thread *td) 248 { 249 250 THREAD_LOCK_ASSERT(td, MA_OWNED); 251 if (td->td_priority < curthread->td_priority) 252 curthread->td_flags |= TDF_NEEDRESCHED; 253 } 254 255 /* 256 * Constants for digital decay and forget: 257 * 90% of (td_estcpu) usage in 5 * loadav time 258 * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive) 259 * Note that, as ps(1) mentions, this can let percentages 260 * total over 100% (I've seen 137.9% for 3 processes). 261 * 262 * Note that schedclock() updates td_estcpu and p_cpticks asynchronously. 263 * 264 * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds. 265 * That is, the system wants to compute a value of decay such 266 * that the following for loop: 267 * for (i = 0; i < (5 * loadavg); i++) 268 * td_estcpu *= decay; 269 * will compute 270 * td_estcpu *= 0.1; 271 * for all values of loadavg: 272 * 273 * Mathematically this loop can be expressed by saying: 274 * decay ** (5 * loadavg) ~= .1 275 * 276 * The system computes decay as: 277 * decay = (2 * loadavg) / (2 * loadavg + 1) 278 * 279 * We wish to prove that the system's computation of decay 280 * will always fulfill the equation: 281 * decay ** (5 * loadavg) ~= .1 282 * 283 * If we compute b as: 284 * b = 2 * loadavg 285 * then 286 * decay = b / (b + 1) 287 * 288 * We now need to prove two things: 289 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 290 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 291 * 292 * Facts: 293 * For x close to zero, exp(x) =~ 1 + x, since 294 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 295 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 296 * For x close to zero, ln(1+x) =~ x, since 297 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 298 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 299 * ln(.1) =~ -2.30 300 * 301 * Proof of (1): 302 * Solve (factor)**(power) =~ .1 given power (5*loadav): 303 * solving for factor, 304 * ln(factor) =~ (-2.30/5*loadav), or 305 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 306 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 307 * 308 * Proof of (2): 309 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 310 * solving for power, 311 * power*ln(b/(b+1)) =~ -2.30, or 312 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 313 * 314 * Actual power values for the implemented algorithm are as follows: 315 * loadav: 1 2 3 4 316 * power: 5.68 10.32 14.94 19.55 317 */ 318 319 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 320 #define loadfactor(loadav) (2 * (loadav)) 321 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 322 323 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 324 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 325 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 326 327 /* 328 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 329 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 330 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 331 * 332 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 333 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 334 * 335 * If you don't want to bother with the faster/more-accurate formula, you 336 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 337 * (more general) method of calculating the %age of CPU used by a process. 338 */ 339 #define CCPU_SHIFT 11 340 341 /* 342 * Recompute process priorities, every hz ticks. 343 * MP-safe, called without the Giant mutex. 344 */ 345 /* ARGSUSED */ 346 static void 347 schedcpu(void) 348 { 349 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 350 struct thread *td; 351 struct proc *p; 352 struct td_sched *ts; 353 int awake, realstathz; 354 355 realstathz = stathz ? stathz : hz; 356 sx_slock(&allproc_lock); 357 FOREACH_PROC_IN_SYSTEM(p) { 358 PROC_SLOCK(p); 359 FOREACH_THREAD_IN_PROC(p, td) { 360 awake = 0; 361 thread_lock(td); 362 ts = td->td_sched; 363 /* 364 * Increment sleep time (if sleeping). We 365 * ignore overflow, as above. 366 */ 367 /* 368 * The td_sched slptimes are not touched in wakeup 369 * because the thread may not HAVE everything in 370 * memory? XXX I think this is out of date. 371 */ 372 if (TD_ON_RUNQ(td)) { 373 awake = 1; 374 ts->ts_flags &= ~TSF_DIDRUN; 375 } else if (TD_IS_RUNNING(td)) { 376 awake = 1; 377 /* Do not clear TSF_DIDRUN */ 378 } else if (ts->ts_flags & TSF_DIDRUN) { 379 awake = 1; 380 ts->ts_flags &= ~TSF_DIDRUN; 381 } 382 383 /* 384 * ts_pctcpu is only for ps and ttyinfo(). 385 * Do it per td_sched, and add them up at the end? 386 * XXXKSE 387 */ 388 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT; 389 /* 390 * If the td_sched has been idle the entire second, 391 * stop recalculating its priority until 392 * it wakes up. 393 */ 394 if (ts->ts_cpticks != 0) { 395 #if (FSHIFT >= CCPU_SHIFT) 396 ts->ts_pctcpu += (realstathz == 100) 397 ? ((fixpt_t) ts->ts_cpticks) << 398 (FSHIFT - CCPU_SHIFT) : 399 100 * (((fixpt_t) ts->ts_cpticks) 400 << (FSHIFT - CCPU_SHIFT)) / realstathz; 401 #else 402 ts->ts_pctcpu += ((FSCALE - ccpu) * 403 (ts->ts_cpticks * 404 FSCALE / realstathz)) >> FSHIFT; 405 #endif 406 ts->ts_cpticks = 0; 407 } 408 /* 409 * If there are ANY running threads in this process, 410 * then don't count it as sleeping. 411 XXX this is broken 412 413 */ 414 if (awake) { 415 if (ts->ts_slptime > 1) { 416 /* 417 * In an ideal world, this should not 418 * happen, because whoever woke us 419 * up from the long sleep should have 420 * unwound the slptime and reset our 421 * priority before we run at the stale 422 * priority. Should KASSERT at some 423 * point when all the cases are fixed. 424 */ 425 updatepri(td); 426 } 427 ts->ts_slptime = 0; 428 } else 429 ts->ts_slptime++; 430 if (ts->ts_slptime > 1) { 431 thread_unlock(td); 432 continue; 433 } 434 td->td_estcpu = decay_cpu(loadfac, td->td_estcpu); 435 resetpriority(td); 436 resetpriority_thread(td); 437 thread_unlock(td); 438 } /* end of thread loop */ 439 PROC_SUNLOCK(p); 440 } /* end of process loop */ 441 sx_sunlock(&allproc_lock); 442 } 443 444 /* 445 * Main loop for a kthread that executes schedcpu once a second. 446 */ 447 static void 448 schedcpu_thread(void) 449 { 450 451 for (;;) { 452 schedcpu(); 453 pause("-", hz); 454 } 455 } 456 457 /* 458 * Recalculate the priority of a process after it has slept for a while. 459 * For all load averages >= 1 and max td_estcpu of 255, sleeping for at 460 * least six times the loadfactor will decay td_estcpu to zero. 461 */ 462 static void 463 updatepri(struct thread *td) 464 { 465 struct td_sched *ts; 466 fixpt_t loadfac; 467 unsigned int newcpu; 468 469 ts = td->td_sched; 470 loadfac = loadfactor(averunnable.ldavg[0]); 471 if (ts->ts_slptime > 5 * loadfac) 472 td->td_estcpu = 0; 473 else { 474 newcpu = td->td_estcpu; 475 ts->ts_slptime--; /* was incremented in schedcpu() */ 476 while (newcpu && --ts->ts_slptime) 477 newcpu = decay_cpu(loadfac, newcpu); 478 td->td_estcpu = newcpu; 479 } 480 } 481 482 /* 483 * Compute the priority of a process when running in user mode. 484 * Arrange to reschedule if the resulting priority is better 485 * than that of the current process. 486 */ 487 static void 488 resetpriority(struct thread *td) 489 { 490 register unsigned int newpriority; 491 492 if (td->td_pri_class == PRI_TIMESHARE) { 493 newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT + 494 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN); 495 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 496 PRI_MAX_TIMESHARE); 497 sched_user_prio(td, newpriority); 498 } 499 } 500 501 /* 502 * Update the thread's priority when the associated process's user 503 * priority changes. 504 */ 505 static void 506 resetpriority_thread(struct thread *td) 507 { 508 509 /* Only change threads with a time sharing user priority. */ 510 if (td->td_priority < PRI_MIN_TIMESHARE || 511 td->td_priority > PRI_MAX_TIMESHARE) 512 return; 513 514 /* XXX the whole needresched thing is broken, but not silly. */ 515 maybe_resched(td); 516 517 sched_prio(td, td->td_user_pri); 518 } 519 520 /* ARGSUSED */ 521 static void 522 sched_setup(void *dummy) 523 { 524 setup_runqs(); 525 526 if (sched_quantum == 0) 527 sched_quantum = SCHED_QUANTUM; 528 hogticks = 2 * sched_quantum; 529 530 /* Account for thread0. */ 531 sched_load_add(); 532 } 533 534 /* External interfaces start here */ 535 /* 536 * Very early in the boot some setup of scheduler-specific 537 * parts of proc0 and of some scheduler resources needs to be done. 538 * Called from: 539 * proc0_init() 540 */ 541 void 542 schedinit(void) 543 { 544 /* 545 * Set up the scheduler specific parts of proc0. 546 */ 547 proc0.p_sched = NULL; /* XXX */ 548 thread0.td_sched = &td_sched0; 549 thread0.td_lock = &sched_lock; 550 td_sched0.ts_thread = &thread0; 551 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 552 } 553 554 int 555 sched_runnable(void) 556 { 557 #ifdef SMP 558 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]); 559 #else 560 return runq_check(&runq); 561 #endif 562 } 563 564 int 565 sched_rr_interval(void) 566 { 567 if (sched_quantum == 0) 568 sched_quantum = SCHED_QUANTUM; 569 return (sched_quantum); 570 } 571 572 /* 573 * We adjust the priority of the current process. The priority of 574 * a process gets worse as it accumulates CPU time. The cpu usage 575 * estimator (td_estcpu) is increased here. resetpriority() will 576 * compute a different priority each time td_estcpu increases by 577 * INVERSE_ESTCPU_WEIGHT 578 * (until MAXPRI is reached). The cpu usage estimator ramps up 579 * quite quickly when the process is running (linearly), and decays 580 * away exponentially, at a rate which is proportionally slower when 581 * the system is busy. The basic principle is that the system will 582 * 90% forget that the process used a lot of CPU time in 5 * loadav 583 * seconds. This causes the system to favor processes which haven't 584 * run much recently, and to round-robin among other processes. 585 */ 586 void 587 sched_clock(struct thread *td) 588 { 589 struct td_sched *ts; 590 591 THREAD_LOCK_ASSERT(td, MA_OWNED); 592 ts = td->td_sched; 593 594 ts->ts_cpticks++; 595 td->td_estcpu = ESTCPULIM(td->td_estcpu + 1); 596 if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 597 resetpriority(td); 598 resetpriority_thread(td); 599 } 600 601 /* 602 * Force a context switch if the current thread has used up a full 603 * quantum (default quantum is 100ms). 604 */ 605 if (!TD_IS_IDLETHREAD(td) && 606 ticks - PCPU_GET(switchticks) >= sched_quantum) 607 td->td_flags |= TDF_NEEDRESCHED; 608 } 609 610 /* 611 * charge childs scheduling cpu usage to parent. 612 */ 613 void 614 sched_exit(struct proc *p, struct thread *td) 615 { 616 617 CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d", 618 td, td->td_name, td->td_priority); 619 PROC_SLOCK_ASSERT(p, MA_OWNED); 620 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td); 621 } 622 623 void 624 sched_exit_thread(struct thread *td, struct thread *child) 625 { 626 627 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d", 628 child, child->td_name, child->td_priority); 629 thread_lock(td); 630 td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu); 631 thread_unlock(td); 632 mtx_lock_spin(&sched_lock); 633 if ((child->td_proc->p_flag & P_NOLOAD) == 0) 634 sched_load_rem(); 635 mtx_unlock_spin(&sched_lock); 636 } 637 638 void 639 sched_fork(struct thread *td, struct thread *childtd) 640 { 641 sched_fork_thread(td, childtd); 642 } 643 644 void 645 sched_fork_thread(struct thread *td, struct thread *childtd) 646 { 647 childtd->td_estcpu = td->td_estcpu; 648 childtd->td_lock = &sched_lock; 649 sched_newthread(childtd); 650 } 651 652 void 653 sched_nice(struct proc *p, int nice) 654 { 655 struct thread *td; 656 657 PROC_LOCK_ASSERT(p, MA_OWNED); 658 PROC_SLOCK_ASSERT(p, MA_OWNED); 659 p->p_nice = nice; 660 FOREACH_THREAD_IN_PROC(p, td) { 661 thread_lock(td); 662 resetpriority(td); 663 resetpriority_thread(td); 664 thread_unlock(td); 665 } 666 } 667 668 void 669 sched_class(struct thread *td, int class) 670 { 671 THREAD_LOCK_ASSERT(td, MA_OWNED); 672 td->td_pri_class = class; 673 } 674 675 /* 676 * Adjust the priority of a thread. 677 */ 678 static void 679 sched_priority(struct thread *td, u_char prio) 680 { 681 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 682 td, td->td_name, td->td_priority, prio, curthread, 683 curthread->td_name); 684 685 THREAD_LOCK_ASSERT(td, MA_OWNED); 686 if (td->td_priority == prio) 687 return; 688 td->td_priority = prio; 689 if (TD_ON_RUNQ(td) && 690 td->td_sched->ts_rqindex != (prio / RQ_PPQ)) { 691 sched_rem(td); 692 sched_add(td, SRQ_BORING); 693 } 694 } 695 696 /* 697 * Update a thread's priority when it is lent another thread's 698 * priority. 699 */ 700 void 701 sched_lend_prio(struct thread *td, u_char prio) 702 { 703 704 td->td_flags |= TDF_BORROWING; 705 sched_priority(td, prio); 706 } 707 708 /* 709 * Restore a thread's priority when priority propagation is 710 * over. The prio argument is the minimum priority the thread 711 * needs to have to satisfy other possible priority lending 712 * requests. If the thread's regulary priority is less 713 * important than prio the thread will keep a priority boost 714 * of prio. 715 */ 716 void 717 sched_unlend_prio(struct thread *td, u_char prio) 718 { 719 u_char base_pri; 720 721 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 722 td->td_base_pri <= PRI_MAX_TIMESHARE) 723 base_pri = td->td_user_pri; 724 else 725 base_pri = td->td_base_pri; 726 if (prio >= base_pri) { 727 td->td_flags &= ~TDF_BORROWING; 728 sched_prio(td, base_pri); 729 } else 730 sched_lend_prio(td, prio); 731 } 732 733 void 734 sched_prio(struct thread *td, u_char prio) 735 { 736 u_char oldprio; 737 738 /* First, update the base priority. */ 739 td->td_base_pri = prio; 740 741 /* 742 * If the thread is borrowing another thread's priority, don't ever 743 * lower the priority. 744 */ 745 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 746 return; 747 748 /* Change the real priority. */ 749 oldprio = td->td_priority; 750 sched_priority(td, prio); 751 752 /* 753 * If the thread is on a turnstile, then let the turnstile update 754 * its state. 755 */ 756 if (TD_ON_LOCK(td) && oldprio != prio) 757 turnstile_adjust(td, oldprio); 758 } 759 760 void 761 sched_user_prio(struct thread *td, u_char prio) 762 { 763 u_char oldprio; 764 765 td->td_base_user_pri = prio; 766 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio) 767 return; 768 oldprio = td->td_user_pri; 769 td->td_user_pri = prio; 770 771 if (TD_ON_UPILOCK(td) && oldprio != prio) 772 umtx_pi_adjust(td, oldprio); 773 } 774 775 void 776 sched_lend_user_prio(struct thread *td, u_char prio) 777 { 778 u_char oldprio; 779 780 td->td_flags |= TDF_UBORROWING; 781 782 oldprio = td->td_user_pri; 783 td->td_user_pri = prio; 784 785 if (TD_ON_UPILOCK(td) && oldprio != prio) 786 umtx_pi_adjust(td, oldprio); 787 } 788 789 void 790 sched_unlend_user_prio(struct thread *td, u_char prio) 791 { 792 u_char base_pri; 793 794 base_pri = td->td_base_user_pri; 795 if (prio >= base_pri) { 796 td->td_flags &= ~TDF_UBORROWING; 797 sched_user_prio(td, base_pri); 798 } else 799 sched_lend_user_prio(td, prio); 800 } 801 802 void 803 sched_sleep(struct thread *td) 804 { 805 806 THREAD_LOCK_ASSERT(td, MA_OWNED); 807 td->td_slptick = ticks; 808 td->td_sched->ts_slptime = 0; 809 } 810 811 void 812 sched_switch(struct thread *td, struct thread *newtd, int flags) 813 { 814 struct td_sched *ts; 815 struct proc *p; 816 817 ts = td->td_sched; 818 p = td->td_proc; 819 820 THREAD_LOCK_ASSERT(td, MA_OWNED); 821 /* 822 * Switch to the sched lock to fix things up and pick 823 * a new thread. 824 */ 825 if (td->td_lock != &sched_lock) { 826 mtx_lock_spin(&sched_lock); 827 thread_unlock(td); 828 } 829 830 if ((p->p_flag & P_NOLOAD) == 0) 831 sched_load_rem(); 832 833 if (newtd) 834 newtd->td_flags |= (td->td_flags & TDF_NEEDRESCHED); 835 836 td->td_lastcpu = td->td_oncpu; 837 td->td_flags &= ~TDF_NEEDRESCHED; 838 td->td_owepreempt = 0; 839 td->td_oncpu = NOCPU; 840 /* 841 * At the last moment, if this thread is still marked RUNNING, 842 * then put it back on the run queue as it has not been suspended 843 * or stopped or any thing else similar. We never put the idle 844 * threads on the run queue, however. 845 */ 846 if (td->td_flags & TDF_IDLETD) { 847 TD_SET_CAN_RUN(td); 848 #ifdef SMP 849 idle_cpus_mask &= ~PCPU_GET(cpumask); 850 #endif 851 } else { 852 if (TD_IS_RUNNING(td)) { 853 /* Put us back on the run queue. */ 854 sched_add(td, (flags & SW_PREEMPT) ? 855 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 856 SRQ_OURSELF|SRQ_YIELDING); 857 } 858 } 859 if (newtd) { 860 /* 861 * The thread we are about to run needs to be counted 862 * as if it had been added to the run queue and selected. 863 * It came from: 864 * * A preemption 865 * * An upcall 866 * * A followon 867 */ 868 KASSERT((newtd->td_inhibitors == 0), 869 ("trying to run inhibited thread")); 870 newtd->td_sched->ts_flags |= TSF_DIDRUN; 871 TD_SET_RUNNING(newtd); 872 if ((newtd->td_proc->p_flag & P_NOLOAD) == 0) 873 sched_load_add(); 874 } else { 875 newtd = choosethread(); 876 } 877 MPASS(newtd->td_lock == &sched_lock); 878 879 if (td != newtd) { 880 #ifdef HWPMC_HOOKS 881 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 882 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 883 #endif 884 885 /* I feel sleepy */ 886 cpu_switch(td, newtd, td->td_lock); 887 /* 888 * Where am I? What year is it? 889 * We are in the same thread that went to sleep above, 890 * but any amount of time may have passed. All out context 891 * will still be available as will local variables. 892 * PCPU values however may have changed as we may have 893 * changed CPU so don't trust cached values of them. 894 * New threads will go to fork_exit() instead of here 895 * so if you change things here you may need to change 896 * things there too. 897 * If the thread above was exiting it will never wake 898 * up again here, so either it has saved everything it 899 * needed to, or the thread_wait() or wait() will 900 * need to reap it. 901 */ 902 #ifdef HWPMC_HOOKS 903 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 904 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 905 #endif 906 } 907 908 #ifdef SMP 909 if (td->td_flags & TDF_IDLETD) 910 idle_cpus_mask |= PCPU_GET(cpumask); 911 #endif 912 sched_lock.mtx_lock = (uintptr_t)td; 913 td->td_oncpu = PCPU_GET(cpuid); 914 MPASS(td->td_lock == &sched_lock); 915 } 916 917 void 918 sched_wakeup(struct thread *td) 919 { 920 struct td_sched *ts; 921 922 THREAD_LOCK_ASSERT(td, MA_OWNED); 923 ts = td->td_sched; 924 if (ts->ts_slptime > 1) { 925 updatepri(td); 926 resetpriority(td); 927 } 928 td->td_slptick = ticks; 929 ts->ts_slptime = 0; 930 sched_add(td, SRQ_BORING); 931 } 932 933 #ifdef SMP 934 /* enable HTT_2 if you have a 2-way HTT cpu.*/ 935 static int 936 forward_wakeup(int cpunum) 937 { 938 cpumask_t map, me, dontuse; 939 cpumask_t map2; 940 struct pcpu *pc; 941 cpumask_t id, map3; 942 943 mtx_assert(&sched_lock, MA_OWNED); 944 945 CTR0(KTR_RUNQ, "forward_wakeup()"); 946 947 if ((!forward_wakeup_enabled) || 948 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0)) 949 return (0); 950 if (!smp_started || cold || panicstr) 951 return (0); 952 953 forward_wakeups_requested++; 954 955 /* 956 * check the idle mask we received against what we calculated before 957 * in the old version. 958 */ 959 me = PCPU_GET(cpumask); 960 /* 961 * don't bother if we should be doing it ourself.. 962 */ 963 if ((me & idle_cpus_mask) && (cpunum == NOCPU || me == (1 << cpunum))) 964 return (0); 965 966 dontuse = me | stopped_cpus | hlt_cpus_mask; 967 map3 = 0; 968 if (forward_wakeup_use_loop) { 969 SLIST_FOREACH(pc, &cpuhead, pc_allcpu) { 970 id = pc->pc_cpumask; 971 if ( (id & dontuse) == 0 && 972 pc->pc_curthread == pc->pc_idlethread) { 973 map3 |= id; 974 } 975 } 976 } 977 978 if (forward_wakeup_use_mask) { 979 map = 0; 980 map = idle_cpus_mask & ~dontuse; 981 982 /* If they are both on, compare and use loop if different */ 983 if (forward_wakeup_use_loop) { 984 if (map != map3) { 985 printf("map (%02X) != map3 (%02X)\n", 986 map, map3); 987 map = map3; 988 } 989 } 990 } else { 991 map = map3; 992 } 993 /* If we only allow a specific CPU, then mask off all the others */ 994 if (cpunum != NOCPU) { 995 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum.")); 996 map &= (1 << cpunum); 997 } else { 998 /* Try choose an idle die. */ 999 if (forward_wakeup_use_htt) { 1000 map2 = (map & (map >> 1)) & 0x5555; 1001 if (map2) { 1002 map = map2; 1003 } 1004 } 1005 1006 /* set only one bit */ 1007 if (forward_wakeup_use_single) { 1008 map = map & ((~map) + 1); 1009 } 1010 } 1011 if (map) { 1012 forward_wakeups_delivered++; 1013 ipi_selected(map, IPI_AST); 1014 return (1); 1015 } 1016 if (cpunum == NOCPU) 1017 printf("forward_wakeup: Idle processor not found\n"); 1018 return (0); 1019 } 1020 #endif 1021 1022 #ifdef SMP 1023 static void kick_other_cpu(int pri,int cpuid); 1024 1025 static void 1026 kick_other_cpu(int pri,int cpuid) 1027 { 1028 struct pcpu * pcpu = pcpu_find(cpuid); 1029 int cpri = pcpu->pc_curthread->td_priority; 1030 1031 if (idle_cpus_mask & pcpu->pc_cpumask) { 1032 forward_wakeups_delivered++; 1033 ipi_selected(pcpu->pc_cpumask, IPI_AST); 1034 return; 1035 } 1036 1037 if (pri >= cpri) 1038 return; 1039 1040 #if defined(IPI_PREEMPTION) && defined(PREEMPTION) 1041 #if !defined(FULL_PREEMPTION) 1042 if (pri <= PRI_MAX_ITHD) 1043 #endif /* ! FULL_PREEMPTION */ 1044 { 1045 ipi_selected(pcpu->pc_cpumask, IPI_PREEMPT); 1046 return; 1047 } 1048 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */ 1049 1050 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED; 1051 ipi_selected( pcpu->pc_cpumask , IPI_AST); 1052 return; 1053 } 1054 #endif /* SMP */ 1055 1056 void 1057 sched_add(struct thread *td, int flags) 1058 #ifdef SMP 1059 { 1060 struct td_sched *ts; 1061 int forwarded = 0; 1062 int cpu; 1063 int single_cpu = 0; 1064 1065 ts = td->td_sched; 1066 THREAD_LOCK_ASSERT(td, MA_OWNED); 1067 KASSERT((td->td_inhibitors == 0), 1068 ("sched_add: trying to run inhibited thread")); 1069 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 1070 ("sched_add: bad thread state")); 1071 KASSERT(td->td_flags & TDF_INMEM, 1072 ("sched_add: thread swapped out")); 1073 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 1074 td, td->td_name, td->td_priority, curthread, 1075 curthread->td_name); 1076 /* 1077 * Now that the thread is moving to the run-queue, set the lock 1078 * to the scheduler's lock. 1079 */ 1080 if (td->td_lock != &sched_lock) { 1081 mtx_lock_spin(&sched_lock); 1082 thread_lock_set(td, &sched_lock); 1083 } 1084 TD_SET_RUNQ(td); 1085 1086 if (td->td_pinned != 0) { 1087 cpu = td->td_lastcpu; 1088 ts->ts_runq = &runq_pcpu[cpu]; 1089 single_cpu = 1; 1090 CTR3(KTR_RUNQ, 1091 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, cpu); 1092 } else if ((ts)->ts_flags & TSF_BOUND) { 1093 /* Find CPU from bound runq */ 1094 KASSERT(SKE_RUNQ_PCPU(ts),("sched_add: bound td_sched not on cpu runq")); 1095 cpu = ts->ts_runq - &runq_pcpu[0]; 1096 single_cpu = 1; 1097 CTR3(KTR_RUNQ, 1098 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, cpu); 1099 } else { 1100 CTR2(KTR_RUNQ, 1101 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts, td); 1102 cpu = NOCPU; 1103 ts->ts_runq = &runq; 1104 } 1105 1106 if (single_cpu && (cpu != PCPU_GET(cpuid))) { 1107 kick_other_cpu(td->td_priority,cpu); 1108 } else { 1109 1110 if (!single_cpu) { 1111 cpumask_t me = PCPU_GET(cpumask); 1112 int idle = idle_cpus_mask & me; 1113 1114 if (!idle && ((flags & SRQ_INTR) == 0) && 1115 (idle_cpus_mask & ~(hlt_cpus_mask | me))) 1116 forwarded = forward_wakeup(cpu); 1117 } 1118 1119 if (!forwarded) { 1120 if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td)) 1121 return; 1122 else 1123 maybe_resched(td); 1124 } 1125 } 1126 1127 if ((td->td_proc->p_flag & P_NOLOAD) == 0) 1128 sched_load_add(); 1129 runq_add(ts->ts_runq, ts, flags); 1130 } 1131 #else /* SMP */ 1132 { 1133 struct td_sched *ts; 1134 ts = td->td_sched; 1135 THREAD_LOCK_ASSERT(td, MA_OWNED); 1136 KASSERT((td->td_inhibitors == 0), 1137 ("sched_add: trying to run inhibited thread")); 1138 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 1139 ("sched_add: bad thread state")); 1140 KASSERT(td->td_flags & TDF_INMEM, 1141 ("sched_add: thread swapped out")); 1142 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 1143 td, td->td_name, td->td_priority, curthread, 1144 curthread->td_name); 1145 /* 1146 * Now that the thread is moving to the run-queue, set the lock 1147 * to the scheduler's lock. 1148 */ 1149 if (td->td_lock != &sched_lock) { 1150 mtx_lock_spin(&sched_lock); 1151 thread_lock_set(td, &sched_lock); 1152 } 1153 TD_SET_RUNQ(td); 1154 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td); 1155 ts->ts_runq = &runq; 1156 1157 /* 1158 * If we are yielding (on the way out anyhow) 1159 * or the thread being saved is US, 1160 * then don't try be smart about preemption 1161 * or kicking off another CPU 1162 * as it won't help and may hinder. 1163 * In the YIEDLING case, we are about to run whoever is 1164 * being put in the queue anyhow, and in the 1165 * OURSELF case, we are puting ourself on the run queue 1166 * which also only happens when we are about to yield. 1167 */ 1168 if((flags & SRQ_YIELDING) == 0) { 1169 if (maybe_preempt(td)) 1170 return; 1171 } 1172 if ((td->td_proc->p_flag & P_NOLOAD) == 0) 1173 sched_load_add(); 1174 runq_add(ts->ts_runq, ts, flags); 1175 maybe_resched(td); 1176 } 1177 #endif /* SMP */ 1178 1179 void 1180 sched_rem(struct thread *td) 1181 { 1182 struct td_sched *ts; 1183 1184 ts = td->td_sched; 1185 KASSERT(td->td_flags & TDF_INMEM, 1186 ("sched_rem: thread swapped out")); 1187 KASSERT(TD_ON_RUNQ(td), 1188 ("sched_rem: thread not on run queue")); 1189 mtx_assert(&sched_lock, MA_OWNED); 1190 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 1191 td, td->td_name, td->td_priority, curthread, 1192 curthread->td_name); 1193 1194 if ((td->td_proc->p_flag & P_NOLOAD) == 0) 1195 sched_load_rem(); 1196 runq_remove(ts->ts_runq, ts); 1197 TD_SET_CAN_RUN(td); 1198 } 1199 1200 /* 1201 * Select threads to run. 1202 * Notice that the running threads still consume a slot. 1203 */ 1204 struct thread * 1205 sched_choose(void) 1206 { 1207 struct td_sched *ts; 1208 struct runq *rq; 1209 1210 mtx_assert(&sched_lock, MA_OWNED); 1211 #ifdef SMP 1212 struct td_sched *kecpu; 1213 1214 rq = &runq; 1215 ts = runq_choose(&runq); 1216 kecpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]); 1217 1218 if (ts == NULL || 1219 (kecpu != NULL && 1220 kecpu->ts_thread->td_priority < ts->ts_thread->td_priority)) { 1221 CTR2(KTR_RUNQ, "choosing td_sched %p from pcpu runq %d", kecpu, 1222 PCPU_GET(cpuid)); 1223 ts = kecpu; 1224 rq = &runq_pcpu[PCPU_GET(cpuid)]; 1225 } else { 1226 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", ts); 1227 } 1228 1229 #else 1230 rq = &runq; 1231 ts = runq_choose(&runq); 1232 #endif 1233 1234 if (ts) { 1235 runq_remove(rq, ts); 1236 ts->ts_flags |= TSF_DIDRUN; 1237 1238 KASSERT(ts->ts_thread->td_flags & TDF_INMEM, 1239 ("sched_choose: thread swapped out")); 1240 return (ts->ts_thread); 1241 } 1242 return (PCPU_GET(idlethread)); 1243 } 1244 1245 void 1246 sched_userret(struct thread *td) 1247 { 1248 /* 1249 * XXX we cheat slightly on the locking here to avoid locking in 1250 * the usual case. Setting td_priority here is essentially an 1251 * incomplete workaround for not setting it properly elsewhere. 1252 * Now that some interrupt handlers are threads, not setting it 1253 * properly elsewhere can clobber it in the window between setting 1254 * it here and returning to user mode, so don't waste time setting 1255 * it perfectly here. 1256 */ 1257 KASSERT((td->td_flags & TDF_BORROWING) == 0, 1258 ("thread with borrowed priority returning to userland")); 1259 if (td->td_priority != td->td_user_pri) { 1260 thread_lock(td); 1261 td->td_priority = td->td_user_pri; 1262 td->td_base_pri = td->td_user_pri; 1263 thread_unlock(td); 1264 } 1265 } 1266 1267 void 1268 sched_bind(struct thread *td, int cpu) 1269 { 1270 struct td_sched *ts; 1271 1272 THREAD_LOCK_ASSERT(td, MA_OWNED); 1273 KASSERT(TD_IS_RUNNING(td), 1274 ("sched_bind: cannot bind non-running thread")); 1275 1276 ts = td->td_sched; 1277 1278 ts->ts_flags |= TSF_BOUND; 1279 #ifdef SMP 1280 ts->ts_runq = &runq_pcpu[cpu]; 1281 if (PCPU_GET(cpuid) == cpu) 1282 return; 1283 1284 mi_switch(SW_VOL, NULL); 1285 #endif 1286 } 1287 1288 void 1289 sched_unbind(struct thread* td) 1290 { 1291 THREAD_LOCK_ASSERT(td, MA_OWNED); 1292 td->td_sched->ts_flags &= ~TSF_BOUND; 1293 } 1294 1295 int 1296 sched_is_bound(struct thread *td) 1297 { 1298 THREAD_LOCK_ASSERT(td, MA_OWNED); 1299 return (td->td_sched->ts_flags & TSF_BOUND); 1300 } 1301 1302 void 1303 sched_relinquish(struct thread *td) 1304 { 1305 thread_lock(td); 1306 SCHED_STAT_INC(switch_relinquish); 1307 mi_switch(SW_VOL, NULL); 1308 thread_unlock(td); 1309 } 1310 1311 int 1312 sched_load(void) 1313 { 1314 return (sched_tdcnt); 1315 } 1316 1317 int 1318 sched_sizeof_proc(void) 1319 { 1320 return (sizeof(struct proc)); 1321 } 1322 1323 int 1324 sched_sizeof_thread(void) 1325 { 1326 return (sizeof(struct thread) + sizeof(struct td_sched)); 1327 } 1328 1329 fixpt_t 1330 sched_pctcpu(struct thread *td) 1331 { 1332 struct td_sched *ts; 1333 1334 ts = td->td_sched; 1335 return (ts->ts_pctcpu); 1336 } 1337 1338 void 1339 sched_tick(void) 1340 { 1341 } 1342 1343 /* 1344 * The actual idle process. 1345 */ 1346 void 1347 sched_idletd(void *dummy) 1348 { 1349 1350 for (;;) { 1351 mtx_assert(&Giant, MA_NOTOWNED); 1352 1353 while (sched_runnable() == 0) 1354 cpu_idle(); 1355 1356 mtx_lock_spin(&sched_lock); 1357 mi_switch(SW_VOL, NULL); 1358 mtx_unlock_spin(&sched_lock); 1359 } 1360 } 1361 1362 /* 1363 * A CPU is entering for the first time or a thread is exiting. 1364 */ 1365 void 1366 sched_throw(struct thread *td) 1367 { 1368 /* 1369 * Correct spinlock nesting. The idle thread context that we are 1370 * borrowing was created so that it would start out with a single 1371 * spin lock (sched_lock) held in fork_trampoline(). Since we've 1372 * explicitly acquired locks in this function, the nesting count 1373 * is now 2 rather than 1. Since we are nested, calling 1374 * spinlock_exit() will simply adjust the counts without allowing 1375 * spin lock using code to interrupt us. 1376 */ 1377 if (td == NULL) { 1378 mtx_lock_spin(&sched_lock); 1379 spinlock_exit(); 1380 } else { 1381 MPASS(td->td_lock == &sched_lock); 1382 } 1383 mtx_assert(&sched_lock, MA_OWNED); 1384 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 1385 PCPU_SET(switchtime, cpu_ticks()); 1386 PCPU_SET(switchticks, ticks); 1387 cpu_throw(td, choosethread()); /* doesn't return */ 1388 } 1389 1390 void 1391 sched_fork_exit(struct thread *td) 1392 { 1393 1394 /* 1395 * Finish setting up thread glue so that it begins execution in a 1396 * non-nested critical section with sched_lock held but not recursed. 1397 */ 1398 td->td_oncpu = PCPU_GET(cpuid); 1399 sched_lock.mtx_lock = (uintptr_t)td; 1400 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED); 1401 } 1402 1403 #define KERN_SWITCH_INCLUDE 1 1404 #include "kern/kern_switch.c" 1405