Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]
FreeBSD/Linux Kernel Cross Reference
sys/kern/sched_4bsd.c
Version:
- FREEBSD - FREEBSD-13-STABLE - FREEBSD-13-0 - FREEBSD-12-STABLE - FREEBSD-12-0 - FREEBSD-11-STABLE - FREEBSD-11-0 - FREEBSD-10-STABLE - FREEBSD-10-0 - FREEBSD-9-STABLE - FREEBSD-9-0 - FREEBSD-8-STABLE - FREEBSD-8-0 - FREEBSD-7-STABLE - FREEBSD-7-0 - FREEBSD-6-STABLE - FREEBSD-6-0 - FREEBSD-5-STABLE - FREEBSD-5-0 - FREEBSD-4-STABLE - FREEBSD-3-STABLE - FREEBSD22 - l41 - OPENBSD - linux-2.6 - MK84 - PLAN9 - xnu-8792
SearchContext: - none - 3 - 10
SearchContext: - none - 3 - 10
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/7.3/sys/kern/sched_4bsd.c 203704 2010-02-09 14:56:10Z attilio $"); 37 38 #include "opt_hwpmc_hooks.h" 39 #include "opt_kdtrace.h" 40 41 #include <sys/param.h> 42 #include <sys/systm.h> 43 #include <sys/cpuset.h> 44 #include <sys/kernel.h> 45 #include <sys/ktr.h> 46 #include <sys/lock.h> 47 #include <sys/kthread.h> 48 #include <sys/mutex.h> 49 #include <sys/proc.h> 50 #include <sys/resourcevar.h> 51 #include <sys/sched.h> 52 #include <sys/smp.h> 53 #include <sys/sysctl.h> 54 #include <sys/sx.h> 55 #include <sys/turnstile.h> 56 #include <sys/umtx.h> 57 #include <machine/pcb.h> 58 #include <machine/smp.h> 59 60 #ifdef HWPMC_HOOKS 61 #include <sys/pmckern.h> 62 #endif 63 64 #ifdef KDTRACE_HOOKS 65 #include <sys/dtrace_bsd.h> 66 int dtrace_vtime_active; 67 dtrace_vtime_switch_func_t dtrace_vtime_switch_func; 68 #endif 69 70 /* 71 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in 72 * the range 100-256 Hz (approximately). 73 */ 74 #define ESTCPULIM(e) \ 75 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \ 76 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1) 77 #ifdef SMP 78 #define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus) 79 #else 80 #define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */ 81 #endif 82 #define NICE_WEIGHT 1 /* Priorities per nice level. */ 83 84 /* 85 * The schedulable entity that runs a context. 86 * This is an extension to the thread structure and is tailored to 87 * the requirements of this scheduler 88 */ 89 struct td_sched { 90 TAILQ_ENTRY(td_sched) ts_procq; /* (j/z) Run queue. */ 91 struct thread *ts_thread; /* (*) Active associated thread. */ 92 fixpt_t ts_pctcpu; /* (j) %cpu during p_swtime. */ 93 u_char ts_rqindex; /* (j) Run queue index. */ 94 int ts_cpticks; /* (j) Ticks of cpu time. */ 95 int ts_slptime; /* (j) Seconds !RUNNING. */ 96 int ts_flags; 97 struct runq *ts_runq; /* runq the thread is currently on */ 98 }; 99 100 /* flags kept in td_flags */ 101 #define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */ 102 #define TDF_BOUND TDF_SCHED1 /* Bound to one CPU. */ 103 104 /* flags kept in ts_flags */ 105 #define TSF_AFFINITY 0x0001 /* Has a non-"full" CPU set. */ 106 107 #define SKE_RUNQ_PCPU(ts) \ 108 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq) 109 110 #define THREAD_CAN_SCHED(td, cpu) \ 111 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 112 113 static struct td_sched td_sched0; 114 struct mtx sched_lock; 115 116 static int sched_tdcnt; /* Total runnable threads in the system. */ 117 static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 118 #define SCHED_QUANTUM (hz / 10) /* Default sched quantum */ 119 120 static void setup_runqs(void); 121 static void schedcpu(void); 122 static void schedcpu_thread(void); 123 static void sched_priority(struct thread *td, u_char prio); 124 static void sched_setup(void *dummy); 125 static void maybe_resched(struct thread *td); 126 static void updatepri(struct thread *td); 127 static void resetpriority(struct thread *td); 128 static void resetpriority_thread(struct thread *td); 129 #ifdef SMP 130 static int sched_pickcpu(struct thread *td); 131 static int forward_wakeup(int cpunum); 132 static void kick_other_cpu(int pri, int cpuid); 133 #endif 134 135 static struct kproc_desc sched_kp = { 136 "schedcpu", 137 schedcpu_thread, 138 NULL 139 }; 140 SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start, 141 &sched_kp); 142 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 143 144 /* 145 * Global run queue. 146 */ 147 static struct runq runq; 148 149 #ifdef SMP 150 /* 151 * Per-CPU run queues 152 */ 153 static struct runq runq_pcpu[MAXCPU]; 154 long runq_length[MAXCPU]; 155 #endif 156 157 static void 158 setup_runqs(void) 159 { 160 #ifdef SMP 161 int i; 162 163 for (i = 0; i < MAXCPU; ++i) 164 runq_init(&runq_pcpu[i]); 165 #endif 166 167 runq_init(&runq); 168 } 169 170 static int 171 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 172 { 173 int error, new_val; 174 175 new_val = sched_quantum * tick; 176 error = sysctl_handle_int(oidp, &new_val, 0, req); 177 if (error != 0 || req->newptr == NULL) 178 return (error); 179 if (new_val < tick) 180 return (EINVAL); 181 sched_quantum = new_val / tick; 182 hogticks = 2 * sched_quantum; 183 return (0); 184 } 185 186 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler"); 187 188 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0, 189 "Scheduler name"); 190 191 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW, 192 0, sizeof sched_quantum, sysctl_kern_quantum, "I", 193 "Roundrobin scheduling quantum in microseconds"); 194 195 #ifdef SMP 196 /* Enable forwarding of wakeups to all other cpus */ 197 SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP"); 198 199 static int forward_wakeup_enabled = 1; 200 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW, 201 &forward_wakeup_enabled, 0, 202 "Forwarding of wakeup to idle CPUs"); 203 204 static int forward_wakeups_requested = 0; 205 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD, 206 &forward_wakeups_requested, 0, 207 "Requests for Forwarding of wakeup to idle CPUs"); 208 209 static int forward_wakeups_delivered = 0; 210 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD, 211 &forward_wakeups_delivered, 0, 212 "Completed Forwarding of wakeup to idle CPUs"); 213 214 static int forward_wakeup_use_mask = 1; 215 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW, 216 &forward_wakeup_use_mask, 0, 217 "Use the mask of idle cpus"); 218 219 static int forward_wakeup_use_loop = 0; 220 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW, 221 &forward_wakeup_use_loop, 0, 222 "Use a loop to find idle cpus"); 223 224 static int forward_wakeup_use_single = 0; 225 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, onecpu, CTLFLAG_RW, 226 &forward_wakeup_use_single, 0, 227 "Only signal one idle cpu"); 228 229 static int forward_wakeup_use_htt = 0; 230 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, htt2, CTLFLAG_RW, 231 &forward_wakeup_use_htt, 0, 232 "account for htt"); 233 234 #endif 235 #if 0 236 static int sched_followon = 0; 237 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW, 238 &sched_followon, 0, 239 "allow threads to share a quantum"); 240 #endif 241 242 static __inline void 243 sched_load_add(void) 244 { 245 sched_tdcnt++; 246 CTR1(KTR_SCHED, "global load: %d", sched_tdcnt); 247 } 248 249 static __inline void 250 sched_load_rem(void) 251 { 252 sched_tdcnt--; 253 CTR1(KTR_SCHED, "global load: %d", sched_tdcnt); 254 } 255 /* 256 * Arrange to reschedule if necessary, taking the priorities and 257 * schedulers into account. 258 */ 259 static void 260 maybe_resched(struct thread *td) 261 { 262 263 THREAD_LOCK_ASSERT(td, MA_OWNED); 264 if (td->td_priority < curthread->td_priority) 265 curthread->td_flags |= TDF_NEEDRESCHED; 266 } 267 268 /* 269 * Constants for digital decay and forget: 270 * 90% of (td_estcpu) usage in 5 * loadav time 271 * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive) 272 * Note that, as ps(1) mentions, this can let percentages 273 * total over 100% (I've seen 137.9% for 3 processes). 274 * 275 * Note that schedclock() updates td_estcpu and p_cpticks asynchronously. 276 * 277 * We wish to decay away 90% of td_estcpu in (5 * loadavg) seconds. 278 * That is, the system wants to compute a value of decay such 279 * that the following for loop: 280 * for (i = 0; i < (5 * loadavg); i++) 281 * td_estcpu *= decay; 282 * will compute 283 * td_estcpu *= 0.1; 284 * for all values of loadavg: 285 * 286 * Mathematically this loop can be expressed by saying: 287 * decay ** (5 * loadavg) ~= .1 288 * 289 * The system computes decay as: 290 * decay = (2 * loadavg) / (2 * loadavg + 1) 291 * 292 * We wish to prove that the system's computation of decay 293 * will always fulfill the equation: 294 * decay ** (5 * loadavg) ~= .1 295 * 296 * If we compute b as: 297 * b = 2 * loadavg 298 * then 299 * decay = b / (b + 1) 300 * 301 * We now need to prove two things: 302 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 303 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 304 * 305 * Facts: 306 * For x close to zero, exp(x) =~ 1 + x, since 307 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 308 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 309 * For x close to zero, ln(1+x) =~ x, since 310 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 311 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 312 * ln(.1) =~ -2.30 313 * 314 * Proof of (1): 315 * Solve (factor)**(power) =~ .1 given power (5*loadav): 316 * solving for factor, 317 * ln(factor) =~ (-2.30/5*loadav), or 318 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 319 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 320 * 321 * Proof of (2): 322 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 323 * solving for power, 324 * power*ln(b/(b+1)) =~ -2.30, or 325 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 326 * 327 * Actual power values for the implemented algorithm are as follows: 328 * loadav: 1 2 3 4 329 * power: 5.68 10.32 14.94 19.55 330 */ 331 332 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 333 #define loadfactor(loadav) (2 * (loadav)) 334 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 335 336 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 337 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 338 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 339 340 /* 341 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 342 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 343 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 344 * 345 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 346 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 347 * 348 * If you don't want to bother with the faster/more-accurate formula, you 349 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 350 * (more general) method of calculating the %age of CPU used by a process. 351 */ 352 #define CCPU_SHIFT 11 353 354 /* 355 * Recompute process priorities, every hz ticks. 356 * MP-safe, called without the Giant mutex. 357 */ 358 /* ARGSUSED */ 359 static void 360 schedcpu(void) 361 { 362 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 363 struct thread *td; 364 struct proc *p; 365 struct td_sched *ts; 366 int awake, realstathz; 367 368 realstathz = stathz ? stathz : hz; 369 sx_slock(&allproc_lock); 370 FOREACH_PROC_IN_SYSTEM(p) { 371 PROC_SLOCK(p); 372 FOREACH_THREAD_IN_PROC(p, td) { 373 awake = 0; 374 thread_lock(td); 375 ts = td->td_sched; 376 /* 377 * Increment sleep time (if sleeping). We 378 * ignore overflow, as above. 379 */ 380 /* 381 * The td_sched slptimes are not touched in wakeup 382 * because the thread may not HAVE everything in 383 * memory? XXX I think this is out of date. 384 */ 385 if (TD_ON_RUNQ(td)) { 386 awake = 1; 387 td->td_flags &= ~TDF_DIDRUN; 388 } else if (TD_IS_RUNNING(td)) { 389 awake = 1; 390 /* Do not clear TDF_DIDRUN */ 391 } else if (td->td_flags & TDF_DIDRUN) { 392 awake = 1; 393 td->td_flags &= ~TDF_DIDRUN; 394 } 395 396 /* 397 * ts_pctcpu is only for ps and ttyinfo(). 398 * Do it per td_sched, and add them up at the end? 399 * XXXKSE 400 */ 401 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT; 402 /* 403 * If the td_sched has been idle the entire second, 404 * stop recalculating its priority until 405 * it wakes up. 406 */ 407 if (ts->ts_cpticks != 0) { 408 #if (FSHIFT >= CCPU_SHIFT) 409 ts->ts_pctcpu += (realstathz == 100) 410 ? ((fixpt_t) ts->ts_cpticks) << 411 (FSHIFT - CCPU_SHIFT) : 412 100 * (((fixpt_t) ts->ts_cpticks) 413 << (FSHIFT - CCPU_SHIFT)) / realstathz; 414 #else 415 ts->ts_pctcpu += ((FSCALE - ccpu) * 416 (ts->ts_cpticks * 417 FSCALE / realstathz)) >> FSHIFT; 418 #endif 419 ts->ts_cpticks = 0; 420 } 421 /* 422 * If there are ANY running threads in this process, 423 * then don't count it as sleeping. 424 * XXX: this is broken. 425 */ 426 if (awake) { 427 if (ts->ts_slptime > 1) { 428 /* 429 * In an ideal world, this should not 430 * happen, because whoever woke us 431 * up from the long sleep should have 432 * unwound the slptime and reset our 433 * priority before we run at the stale 434 * priority. Should KASSERT at some 435 * point when all the cases are fixed. 436 */ 437 updatepri(td); 438 } 439 ts->ts_slptime = 0; 440 } else 441 ts->ts_slptime++; 442 if (ts->ts_slptime > 1) { 443 thread_unlock(td); 444 continue; 445 } 446 td->td_estcpu = decay_cpu(loadfac, td->td_estcpu); 447 resetpriority(td); 448 resetpriority_thread(td); 449 thread_unlock(td); 450 } 451 PROC_SUNLOCK(p); 452 } 453 sx_sunlock(&allproc_lock); 454 } 455 456 /* 457 * Main loop for a kthread that executes schedcpu once a second. 458 */ 459 static void 460 schedcpu_thread(void) 461 { 462 463 for (;;) { 464 schedcpu(); 465 pause("-", hz); 466 } 467 } 468 469 /* 470 * Recalculate the priority of a process after it has slept for a while. 471 * For all load averages >= 1 and max td_estcpu of 255, sleeping for at 472 * least six times the loadfactor will decay td_estcpu to zero. 473 */ 474 static void 475 updatepri(struct thread *td) 476 { 477 struct td_sched *ts; 478 fixpt_t loadfac; 479 unsigned int newcpu; 480 481 ts = td->td_sched; 482 loadfac = loadfactor(averunnable.ldavg[0]); 483 if (ts->ts_slptime > 5 * loadfac) 484 td->td_estcpu = 0; 485 else { 486 newcpu = td->td_estcpu; 487 ts->ts_slptime--; /* was incremented in schedcpu() */ 488 while (newcpu && --ts->ts_slptime) 489 newcpu = decay_cpu(loadfac, newcpu); 490 td->td_estcpu = newcpu; 491 } 492 } 493 494 /* 495 * Compute the priority of a process when running in user mode. 496 * Arrange to reschedule if the resulting priority is better 497 * than that of the current process. 498 */ 499 static void 500 resetpriority(struct thread *td) 501 { 502 register unsigned int newpriority; 503 504 if (td->td_pri_class == PRI_TIMESHARE) { 505 newpriority = PUSER + td->td_estcpu / INVERSE_ESTCPU_WEIGHT + 506 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN); 507 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 508 PRI_MAX_TIMESHARE); 509 sched_user_prio(td, newpriority); 510 } 511 } 512 513 /* 514 * Update the thread's priority when the associated process's user 515 * priority changes. 516 */ 517 static void 518 resetpriority_thread(struct thread *td) 519 { 520 521 /* Only change threads with a time sharing user priority. */ 522 if (td->td_priority < PRI_MIN_TIMESHARE || 523 td->td_priority > PRI_MAX_TIMESHARE) 524 return; 525 526 /* XXX the whole needresched thing is broken, but not silly. */ 527 maybe_resched(td); 528 529 sched_prio(td, td->td_user_pri); 530 } 531 532 /* ARGSUSED */ 533 static void 534 sched_setup(void *dummy) 535 { 536 setup_runqs(); 537 538 if (sched_quantum == 0) 539 sched_quantum = SCHED_QUANTUM; 540 hogticks = 2 * sched_quantum; 541 542 /* Account for thread0. */ 543 sched_load_add(); 544 } 545 546 /* External interfaces start here */ 547 548 /* 549 * Very early in the boot some setup of scheduler-specific 550 * parts of proc0 and of some scheduler resources needs to be done. 551 * Called from: 552 * proc0_init() 553 */ 554 void 555 schedinit(void) 556 { 557 /* 558 * Set up the scheduler specific parts of proc0. 559 */ 560 proc0.p_sched = NULL; /* XXX */ 561 thread0.td_sched = &td_sched0; 562 thread0.td_lock = &sched_lock; 563 td_sched0.ts_thread = &thread0; 564 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN | MTX_RECURSE); 565 } 566 567 int 568 sched_runnable(void) 569 { 570 #ifdef SMP 571 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]); 572 #else 573 return runq_check(&runq); 574 #endif 575 } 576 577 int 578 sched_rr_interval(void) 579 { 580 if (sched_quantum == 0) 581 sched_quantum = SCHED_QUANTUM; 582 return (sched_quantum); 583 } 584 585 /* 586 * We adjust the priority of the current process. The priority of 587 * a process gets worse as it accumulates CPU time. The cpu usage 588 * estimator (td_estcpu) is increased here. resetpriority() will 589 * compute a different priority each time td_estcpu increases by 590 * INVERSE_ESTCPU_WEIGHT 591 * (until MAXPRI is reached). The cpu usage estimator ramps up 592 * quite quickly when the process is running (linearly), and decays 593 * away exponentially, at a rate which is proportionally slower when 594 * the system is busy. The basic principle is that the system will 595 * 90% forget that the process used a lot of CPU time in 5 * loadav 596 * seconds. This causes the system to favor processes which haven't 597 * run much recently, and to round-robin among other processes. 598 */ 599 void 600 sched_clock(struct thread *td) 601 { 602 struct td_sched *ts; 603 604 THREAD_LOCK_ASSERT(td, MA_OWNED); 605 ts = td->td_sched; 606 607 ts->ts_cpticks++; 608 td->td_estcpu = ESTCPULIM(td->td_estcpu + 1); 609 if ((td->td_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 610 resetpriority(td); 611 resetpriority_thread(td); 612 } 613 614 /* 615 * Force a context switch if the current thread has used up a full 616 * quantum (default quantum is 100ms). 617 */ 618 if (!TD_IS_IDLETHREAD(td) && 619 ticks - PCPU_GET(switchticks) >= sched_quantum) 620 td->td_flags |= TDF_NEEDRESCHED; 621 } 622 623 /* 624 * Charge child's scheduling CPU usage to parent. 625 */ 626 void 627 sched_exit(struct proc *p, struct thread *td) 628 { 629 630 CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d", 631 td, td->td_proc->p_comm, td->td_priority); 632 PROC_SLOCK_ASSERT(p, MA_OWNED); 633 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td); 634 } 635 636 void 637 sched_exit_thread(struct thread *td, struct thread *child) 638 { 639 640 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d", 641 child, child->td_proc->p_comm, child->td_priority); 642 thread_lock(td); 643 td->td_estcpu = ESTCPULIM(td->td_estcpu + child->td_estcpu); 644 thread_unlock(td); 645 mtx_lock_spin(&sched_lock); 646 if ((child->td_proc->p_flag & P_NOLOAD) == 0) 647 sched_load_rem(); 648 mtx_unlock_spin(&sched_lock); 649 } 650 651 void 652 sched_fork(struct thread *td, struct thread *childtd) 653 { 654 sched_fork_thread(td, childtd); 655 } 656 657 void 658 sched_fork_thread(struct thread *td, struct thread *childtd) 659 { 660 childtd->td_estcpu = td->td_estcpu; 661 childtd->td_lock = &sched_lock; 662 childtd->td_cpuset = cpuset_ref(td->td_cpuset); 663 sched_newthread(childtd); 664 childtd->td_sched->ts_flags |= (td->td_sched->ts_flags & TSF_AFFINITY); 665 } 666 667 void 668 sched_nice(struct proc *p, int nice) 669 { 670 struct thread *td; 671 672 PROC_LOCK_ASSERT(p, MA_OWNED); 673 PROC_SLOCK_ASSERT(p, MA_OWNED); 674 p->p_nice = nice; 675 FOREACH_THREAD_IN_PROC(p, td) { 676 thread_lock(td); 677 resetpriority(td); 678 resetpriority_thread(td); 679 thread_unlock(td); 680 } 681 } 682 683 void 684 sched_class(struct thread *td, int class) 685 { 686 THREAD_LOCK_ASSERT(td, MA_OWNED); 687 td->td_pri_class = class; 688 } 689 690 /* 691 * Adjust the priority of a thread. 692 */ 693 static void 694 sched_priority(struct thread *td, u_char prio) 695 { 696 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 697 td, td->td_proc->p_comm, td->td_priority, prio, curthread, 698 curthread->td_proc->p_comm); 699 700 THREAD_LOCK_ASSERT(td, MA_OWNED); 701 if (td->td_priority == prio) 702 return; 703 td->td_priority = prio; 704 if (TD_ON_RUNQ(td) && 705 td->td_sched->ts_rqindex != (prio / RQ_PPQ)) { 706 sched_rem(td); 707 sched_add(td, SRQ_BORING); 708 } 709 } 710 711 /* 712 * Update a thread's priority when it is lent another thread's 713 * priority. 714 */ 715 void 716 sched_lend_prio(struct thread *td, u_char prio) 717 { 718 719 td->td_flags |= TDF_BORROWING; 720 sched_priority(td, prio); 721 } 722 723 /* 724 * Restore a thread's priority when priority propagation is 725 * over. The prio argument is the minimum priority the thread 726 * needs to have to satisfy other possible priority lending 727 * requests. If the thread's regulary priority is less 728 * important than prio the thread will keep a priority boost 729 * of prio. 730 */ 731 void 732 sched_unlend_prio(struct thread *td, u_char prio) 733 { 734 u_char base_pri; 735 736 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 737 td->td_base_pri <= PRI_MAX_TIMESHARE) 738 base_pri = td->td_user_pri; 739 else 740 base_pri = td->td_base_pri; 741 if (prio >= base_pri) { 742 td->td_flags &= ~TDF_BORROWING; 743 sched_prio(td, base_pri); 744 } else 745 sched_lend_prio(td, prio); 746 } 747 748 void 749 sched_prio(struct thread *td, u_char prio) 750 { 751 u_char oldprio; 752 753 /* First, update the base priority. */ 754 td->td_base_pri = prio; 755 756 /* 757 * If the thread is borrowing another thread's priority, don't ever 758 * lower the priority. 759 */ 760 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 761 return; 762 763 /* Change the real priority. */ 764 oldprio = td->td_priority; 765 sched_priority(td, prio); 766 767 /* 768 * If the thread is on a turnstile, then let the turnstile update 769 * its state. 770 */ 771 if (TD_ON_LOCK(td) && oldprio != prio) 772 turnstile_adjust(td, oldprio); 773 } 774 775 void 776 sched_user_prio(struct thread *td, u_char prio) 777 { 778 u_char oldprio; 779 780 THREAD_LOCK_ASSERT(td, MA_OWNED); 781 td->td_base_user_pri = prio; 782 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio) 783 return; 784 oldprio = td->td_user_pri; 785 td->td_user_pri = prio; 786 } 787 788 void 789 sched_lend_user_prio(struct thread *td, u_char prio) 790 { 791 u_char oldprio; 792 793 THREAD_LOCK_ASSERT(td, MA_OWNED); 794 td->td_flags |= TDF_UBORROWING; 795 796 oldprio = td->td_user_pri; 797 td->td_user_pri = prio; 798 } 799 800 void 801 sched_unlend_user_prio(struct thread *td, u_char prio) 802 { 803 u_char base_pri; 804 805 THREAD_LOCK_ASSERT(td, MA_OWNED); 806 base_pri = td->td_base_user_pri; 807 if (prio >= base_pri) { 808 td->td_flags &= ~TDF_UBORROWING; 809 sched_user_prio(td, base_pri); 810 } else { 811 sched_lend_user_prio(td, prio); 812 } 813 } 814 815 void 816 sched_sleep(struct thread *td) 817 { 818 819 THREAD_LOCK_ASSERT(td, MA_OWNED); 820 td->td_slptick = ticks; 821 td->td_sched->ts_slptime = 0; 822 } 823 824 void 825 sched_switch(struct thread *td, struct thread *newtd, int flags) 826 { 827 struct mtx *tmtx; 828 struct td_sched *ts; 829 struct proc *p; 830 831 tmtx = NULL; 832 ts = td->td_sched; 833 p = td->td_proc; 834 835 THREAD_LOCK_ASSERT(td, MA_OWNED); 836 837 /* 838 * Switch to the sched lock to fix things up and pick 839 * a new thread. 840 * Block the td_lock in order to avoid breaking the critical path. 841 */ 842 if (td->td_lock != &sched_lock) { 843 mtx_lock_spin(&sched_lock); 844 tmtx = thread_lock_block(td); 845 } 846 847 if ((p->p_flag & P_NOLOAD) == 0) 848 sched_load_rem(); 849 850 if (newtd) { 851 MPASS(newtd->td_lock == &sched_lock); 852 newtd->td_flags |= (td->td_flags & TDF_NEEDRESCHED); 853 } 854 855 td->td_lastcpu = td->td_oncpu; 856 td->td_flags &= ~TDF_NEEDRESCHED; 857 td->td_owepreempt = 0; 858 td->td_oncpu = NOCPU; 859 860 /* 861 * At the last moment, if this thread is still marked RUNNING, 862 * then put it back on the run queue as it has not been suspended 863 * or stopped or any thing else similar. We never put the idle 864 * threads on the run queue, however. 865 */ 866 if (td->td_flags & TDF_IDLETD) { 867 TD_SET_CAN_RUN(td); 868 #ifdef SMP 869 idle_cpus_mask &= ~PCPU_GET(cpumask); 870 #endif 871 } else { 872 if (TD_IS_RUNNING(td)) { 873 /* Put us back on the run queue. */ 874 sched_add(td, (flags & SW_PREEMPT) ? 875 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 876 SRQ_OURSELF|SRQ_YIELDING); 877 } 878 } 879 if (newtd) { 880 /* 881 * The thread we are about to run needs to be counted 882 * as if it had been added to the run queue and selected. 883 * It came from: 884 * * A preemption 885 * * An upcall 886 * * A followon 887 */ 888 KASSERT((newtd->td_inhibitors == 0), 889 ("trying to run inhibited thread")); 890 newtd->td_flags |= TDF_DIDRUN; 891 TD_SET_RUNNING(newtd); 892 if ((newtd->td_proc->p_flag & P_NOLOAD) == 0) 893 sched_load_add(); 894 } else { 895 newtd = choosethread(); 896 MPASS(newtd->td_lock == &sched_lock); 897 } 898 899 if (td != newtd) { 900 #ifdef HWPMC_HOOKS 901 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 902 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 903 #endif 904 905 #ifdef KDTRACE_HOOKS 906 /* 907 * If DTrace has set the active vtime enum to anything 908 * other than INACTIVE (0), then it should have set the 909 * function to call. 910 */ 911 if (dtrace_vtime_active) 912 (*dtrace_vtime_switch_func)(newtd); 913 #endif 914 /* I feel sleepy */ 915 cpu_switch(td, newtd, tmtx != NULL ? tmtx : td->td_lock); 916 /* 917 * Where am I? What year is it? 918 * We are in the same thread that went to sleep above, 919 * but any amount of time may have passed. All our context 920 * will still be available as will local variables. 921 * PCPU values however may have changed as we may have 922 * changed CPU so don't trust cached values of them. 923 * New threads will go to fork_exit() instead of here 924 * so if you change things here you may need to change 925 * things there too. 926 * 927 * If the thread above was exiting it will never wake 928 * up again here, so either it has saved everything it 929 * needed to, or the thread_wait() or wait() will 930 * need to reap it. 931 */ 932 #ifdef HWPMC_HOOKS 933 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 934 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 935 #endif 936 } 937 938 #ifdef SMP 939 if (td->td_flags & TDF_IDLETD) 940 idle_cpus_mask |= PCPU_GET(cpumask); 941 #endif 942 sched_lock.mtx_lock = (uintptr_t)td; 943 td->td_oncpu = PCPU_GET(cpuid); 944 MPASS(td->td_lock == &sched_lock); 945 } 946 947 void 948 sched_wakeup(struct thread *td) 949 { 950 struct td_sched *ts; 951 952 THREAD_LOCK_ASSERT(td, MA_OWNED); 953 ts = td->td_sched; 954 if (ts->ts_slptime > 1) { 955 updatepri(td); 956 resetpriority(td); 957 } 958 td->td_slptick = ticks; 959 ts->ts_slptime = 0; 960 sched_add(td, SRQ_BORING); 961 } 962 963 #ifdef SMP 964 static int 965 forward_wakeup(int cpunum) 966 { 967 struct pcpu *pc; 968 cpumask_t dontuse, id, map, map2, map3, me; 969 970 mtx_assert(&sched_lock, MA_OWNED); 971 972 CTR0(KTR_RUNQ, "forward_wakeup()"); 973 974 if ((!forward_wakeup_enabled) || 975 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0)) 976 return (0); 977 if (!smp_started || cold || panicstr) 978 return (0); 979 980 forward_wakeups_requested++; 981 982 /* 983 * Check the idle mask we received against what we calculated 984 * before in the old version. 985 */ 986 me = PCPU_GET(cpumask); 987 988 /* Don't bother if we should be doing it ourself. */ 989 if ((me & idle_cpus_mask) && (cpunum == NOCPU || me == (1 << cpunum))) 990 return (0); 991 992 dontuse = me | stopped_cpus | hlt_cpus_mask; 993 map3 = 0; 994 if (forward_wakeup_use_loop) { 995 SLIST_FOREACH(pc, &cpuhead, pc_allcpu) { 996 id = pc->pc_cpumask; 997 if ((id & dontuse) == 0 && 998 pc->pc_curthread == pc->pc_idlethread) { 999 map3 |= id; 1000 } 1001 } 1002 } 1003 1004 if (forward_wakeup_use_mask) { 1005 map = 0; 1006 map = idle_cpus_mask & ~dontuse; 1007 1008 /* If they are both on, compare and use loop if different. */ 1009 if (forward_wakeup_use_loop) { 1010 if (map != map3) { 1011 printf("map (%02X) != map3 (%02X)\n", map, 1012 map3); 1013 map = map3; 1014 } 1015 } 1016 } else { 1017 map = map3; 1018 } 1019 1020 /* If we only allow a specific CPU, then mask off all the others. */ 1021 if (cpunum != NOCPU) { 1022 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum.")); 1023 map &= (1 << cpunum); 1024 } else { 1025 /* Try choose an idle die. */ 1026 if (forward_wakeup_use_htt) { 1027 map2 = (map & (map >> 1)) & 0x5555; 1028 if (map2) { 1029 map = map2; 1030 } 1031 } 1032 1033 /* Set only one bit. */ 1034 if (forward_wakeup_use_single) { 1035 map = map & ((~map) + 1); 1036 } 1037 } 1038 if (map) { 1039 forward_wakeups_delivered++; 1040 ipi_selected(map, IPI_AST); 1041 return (1); 1042 } 1043 if (cpunum == NOCPU) 1044 printf("forward_wakeup: Idle processor not found\n"); 1045 return (0); 1046 } 1047 1048 static void 1049 kick_other_cpu(int pri, int cpuid) 1050 { 1051 struct pcpu *pcpu; 1052 int cpri; 1053 1054 pcpu = pcpu_find(cpuid); 1055 if (idle_cpus_mask & pcpu->pc_cpumask) { 1056 forward_wakeups_delivered++; 1057 ipi_selected(pcpu->pc_cpumask, IPI_AST); 1058 return; 1059 } 1060 1061 cpri = pcpu->pc_curthread->td_priority; 1062 if (pri >= cpri) 1063 return; 1064 1065 #if defined(IPI_PREEMPTION) && defined(PREEMPTION) 1066 #if !defined(FULL_PREEMPTION) 1067 if (pri <= PRI_MAX_ITHD) 1068 #endif /* ! FULL_PREEMPTION */ 1069 { 1070 ipi_selected(pcpu->pc_cpumask, IPI_PREEMPT); 1071 return; 1072 } 1073 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */ 1074 1075 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED; 1076 ipi_selected(pcpu->pc_cpumask, IPI_AST); 1077 return; 1078 } 1079 #endif /* SMP */ 1080 1081 #ifdef SMP 1082 static int 1083 sched_pickcpu(struct thread *td) 1084 { 1085 int best, cpu; 1086 1087 mtx_assert(&sched_lock, MA_OWNED); 1088 1089 if (THREAD_CAN_SCHED(td, td->td_lastcpu)) 1090 best = td->td_lastcpu; 1091 else 1092 best = NOCPU; 1093 for (cpu = 0; cpu <= mp_maxid; cpu++) { 1094 if (CPU_ABSENT(cpu)) 1095 continue; 1096 if (!THREAD_CAN_SCHED(td, cpu)) 1097 continue; 1098 1099 if (best == NOCPU) 1100 best = cpu; 1101 else if (runq_length[cpu] < runq_length[best]) 1102 best = cpu; 1103 } 1104 KASSERT(best != NOCPU, ("no valid CPUs")); 1105 1106 return (best); 1107 } 1108 #endif 1109 1110 void 1111 sched_add(struct thread *td, int flags) 1112 #ifdef SMP 1113 { 1114 struct td_sched *ts; 1115 int forwarded = 0; 1116 int cpu; 1117 int single_cpu = 0; 1118 1119 ts = td->td_sched; 1120 THREAD_LOCK_ASSERT(td, MA_OWNED); 1121 KASSERT((td->td_inhibitors == 0), 1122 ("sched_add: trying to run inhibited thread")); 1123 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 1124 ("sched_add: bad thread state")); 1125 KASSERT(td->td_flags & TDF_INMEM, 1126 ("sched_add: thread swapped out")); 1127 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 1128 td, td->td_proc->p_comm, td->td_priority, curthread, 1129 curthread->td_proc->p_comm); 1130 1131 /* 1132 * Now that the thread is moving to the run-queue, set the lock 1133 * to the scheduler's lock. 1134 */ 1135 if (td->td_lock != &sched_lock) { 1136 mtx_lock_spin(&sched_lock); 1137 thread_lock_set(td, &sched_lock); 1138 } 1139 TD_SET_RUNQ(td); 1140 1141 if (td->td_pinned != 0) { 1142 cpu = td->td_lastcpu; 1143 ts->ts_runq = &runq_pcpu[cpu]; 1144 single_cpu = 1; 1145 CTR3(KTR_RUNQ, 1146 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, 1147 cpu); 1148 } else if (td->td_flags & TDF_BOUND) { 1149 /* Find CPU from bound runq. */ 1150 KASSERT(SKE_RUNQ_PCPU(ts), 1151 ("sched_add: bound td_sched not on cpu runq")); 1152 cpu = ts->ts_runq - &runq_pcpu[0]; 1153 single_cpu = 1; 1154 CTR3(KTR_RUNQ, 1155 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, 1156 cpu); 1157 } else if (ts->ts_flags & TSF_AFFINITY) { 1158 /* Find a valid CPU for our cpuset */ 1159 cpu = sched_pickcpu(td); 1160 ts->ts_runq = &runq_pcpu[cpu]; 1161 single_cpu = 1; 1162 CTR3(KTR_RUNQ, 1163 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, 1164 cpu); 1165 } else { 1166 CTR2(KTR_RUNQ, 1167 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts, 1168 td); 1169 cpu = NOCPU; 1170 ts->ts_runq = &runq; 1171 } 1172 1173 if (single_cpu && (cpu != PCPU_GET(cpuid))) { 1174 kick_other_cpu(td->td_priority, cpu); 1175 } else { 1176 if (!single_cpu) { 1177 cpumask_t me = PCPU_GET(cpumask); 1178 cpumask_t idle = idle_cpus_mask & me; 1179 1180 if (!idle && ((flags & SRQ_INTR) == 0) && 1181 (idle_cpus_mask & ~(hlt_cpus_mask | me))) 1182 forwarded = forward_wakeup(cpu); 1183 } 1184 1185 if (!forwarded) { 1186 if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td)) 1187 return; 1188 else 1189 maybe_resched(td); 1190 } 1191 } 1192 1193 if ((td->td_proc->p_flag & P_NOLOAD) == 0) 1194 sched_load_add(); 1195 runq_add(ts->ts_runq, ts, flags); 1196 if (cpu != NOCPU) 1197 runq_length[cpu]++; 1198 } 1199 #else /* SMP */ 1200 { 1201 struct td_sched *ts; 1202 1203 ts = td->td_sched; 1204 THREAD_LOCK_ASSERT(td, MA_OWNED); 1205 KASSERT((td->td_inhibitors == 0), 1206 ("sched_add: trying to run inhibited thread")); 1207 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 1208 ("sched_add: bad thread state")); 1209 KASSERT(td->td_flags & TDF_INMEM, 1210 ("sched_add: thread swapped out")); 1211 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 1212 td, td->td_proc->p_comm, td->td_priority, curthread, 1213 curthread->td_proc->p_comm); 1214 1215 /* 1216 * Now that the thread is moving to the run-queue, set the lock 1217 * to the scheduler's lock. 1218 */ 1219 if (td->td_lock != &sched_lock) { 1220 mtx_lock_spin(&sched_lock); 1221 thread_lock_set(td, &sched_lock); 1222 } 1223 TD_SET_RUNQ(td); 1224 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td); 1225 ts->ts_runq = &runq; 1226 1227 /* 1228 * If we are yielding (on the way out anyhow) or the thread 1229 * being saved is US, then don't try be smart about preemption 1230 * or kicking off another CPU as it won't help and may hinder. 1231 * In the YIEDLING case, we are about to run whoever is being 1232 * put in the queue anyhow, and in the OURSELF case, we are 1233 * puting ourself on the run queue which also only happens 1234 * when we are about to yield. 1235 */ 1236 if ((flags & SRQ_YIELDING) == 0) { 1237 if (maybe_preempt(td)) 1238 return; 1239 } 1240 if ((td->td_proc->p_flag & P_NOLOAD) == 0) 1241 sched_load_add(); 1242 runq_add(ts->ts_runq, ts, flags); 1243 maybe_resched(td); 1244 } 1245 #endif /* SMP */ 1246 1247 void 1248 sched_rem(struct thread *td) 1249 { 1250 struct td_sched *ts; 1251 1252 ts = td->td_sched; 1253 KASSERT(td->td_flags & TDF_INMEM, 1254 ("sched_rem: thread swapped out")); 1255 KASSERT(TD_ON_RUNQ(td), 1256 ("sched_rem: thread not on run queue")); 1257 mtx_assert(&sched_lock, MA_OWNED); 1258 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 1259 td, td->td_proc->p_comm, td->td_priority, curthread, 1260 curthread->td_proc->p_comm); 1261 1262 if ((td->td_proc->p_flag & P_NOLOAD) == 0) 1263 sched_load_rem(); 1264 #ifdef SMP 1265 if (ts->ts_runq != &runq) 1266 runq_length[ts->ts_runq - runq_pcpu]--; 1267 #endif 1268 runq_remove(ts->ts_runq, ts); 1269 TD_SET_CAN_RUN(td); 1270 } 1271 1272 /* 1273 * Select threads to run. Note that running threads still consume a 1274 * slot. 1275 */ 1276 struct thread * 1277 sched_choose(void) 1278 { 1279 struct td_sched *ts; 1280 struct runq *rq; 1281 1282 mtx_assert(&sched_lock, MA_OWNED); 1283 #ifdef SMP 1284 struct td_sched *kecpu; 1285 1286 rq = &runq; 1287 ts = runq_choose(&runq); 1288 kecpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]); 1289 1290 if (ts == NULL || 1291 (kecpu != NULL && 1292 kecpu->ts_thread->td_priority < ts->ts_thread->td_priority)) { 1293 CTR2(KTR_RUNQ, "choosing td_sched %p from pcpu runq %d", kecpu, 1294 PCPU_GET(cpuid)); 1295 ts = kecpu; 1296 rq = &runq_pcpu[PCPU_GET(cpuid)]; 1297 } else { 1298 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", ts); 1299 } 1300 1301 #else 1302 rq = &runq; 1303 ts = runq_choose(&runq); 1304 #endif 1305 1306 if (ts) { 1307 #ifdef SMP 1308 if (ts == kecpu) 1309 runq_length[PCPU_GET(cpuid)]--; 1310 #endif 1311 runq_remove(rq, ts); 1312 ts->ts_thread->td_flags |= TDF_DIDRUN; 1313 1314 KASSERT(ts->ts_thread->td_flags & TDF_INMEM, 1315 ("sched_choose: thread swapped out")); 1316 return (ts->ts_thread); 1317 } 1318 return (PCPU_GET(idlethread)); 1319 } 1320 1321 void 1322 sched_userret(struct thread *td) 1323 { 1324 /* 1325 * XXX we cheat slightly on the locking here to avoid locking in 1326 * the usual case. Setting td_priority here is essentially an 1327 * incomplete workaround for not setting it properly elsewhere. 1328 * Now that some interrupt handlers are threads, not setting it 1329 * properly elsewhere can clobber it in the window between setting 1330 * it here and returning to user mode, so don't waste time setting 1331 * it perfectly here. 1332 */ 1333 KASSERT((td->td_flags & TDF_BORROWING) == 0, 1334 ("thread with borrowed priority returning to userland")); 1335 if (td->td_priority != td->td_user_pri) { 1336 thread_lock(td); 1337 td->td_priority = td->td_user_pri; 1338 td->td_base_pri = td->td_user_pri; 1339 thread_unlock(td); 1340 } 1341 } 1342 1343 void 1344 sched_bind(struct thread *td, int cpu) 1345 { 1346 struct td_sched *ts; 1347 1348 THREAD_LOCK_ASSERT(td, MA_OWNED); 1349 KASSERT(TD_IS_RUNNING(td), 1350 ("sched_bind: cannot bind non-running thread")); 1351 1352 ts = td->td_sched; 1353 1354 td->td_flags |= TDF_BOUND; 1355 #ifdef SMP 1356 ts->ts_runq = &runq_pcpu[cpu]; 1357 if (PCPU_GET(cpuid) == cpu) 1358 return; 1359 1360 mi_switch(SW_VOL, NULL); 1361 #endif 1362 } 1363 1364 void 1365 sched_unbind(struct thread* td) 1366 { 1367 THREAD_LOCK_ASSERT(td, MA_OWNED); 1368 td->td_flags &= ~TDF_BOUND; 1369 } 1370 1371 int 1372 sched_is_bound(struct thread *td) 1373 { 1374 THREAD_LOCK_ASSERT(td, MA_OWNED); 1375 return (td->td_flags & TDF_BOUND); 1376 } 1377 1378 void 1379 sched_relinquish(struct thread *td) 1380 { 1381 thread_lock(td); 1382 SCHED_STAT_INC(switch_relinquish); 1383 mi_switch(SW_VOL, NULL); 1384 thread_unlock(td); 1385 } 1386 1387 int 1388 sched_load(void) 1389 { 1390 return (sched_tdcnt); 1391 } 1392 1393 int 1394 sched_sizeof_proc(void) 1395 { 1396 return (sizeof(struct proc)); 1397 } 1398 1399 int 1400 sched_sizeof_thread(void) 1401 { 1402 return (sizeof(struct thread) + sizeof(struct td_sched)); 1403 } 1404 1405 fixpt_t 1406 sched_pctcpu(struct thread *td) 1407 { 1408 struct td_sched *ts; 1409 1410 ts = td->td_sched; 1411 return (ts->ts_pctcpu); 1412 } 1413 1414 void 1415 sched_tick(void) 1416 { 1417 } 1418 1419 /* 1420 * The actual idle process. 1421 */ 1422 void 1423 sched_idletd(void *dummy) 1424 { 1425 struct proc *p; 1426 struct thread *td; 1427 1428 td = curthread; 1429 p = td->td_proc; 1430 for (;;) { 1431 mtx_assert(&Giant, MA_NOTOWNED); 1432 1433 while (sched_runnable() == 0) 1434 cpu_idle(); 1435 1436 mtx_lock_spin(&sched_lock); 1437 mi_switch(SW_VOL, NULL); 1438 mtx_unlock_spin(&sched_lock); 1439 } 1440 } 1441 1442 /* 1443 * A CPU is entering for the first time or a thread is exiting. 1444 */ 1445 void 1446 sched_throw(struct thread *td) 1447 { 1448 /* 1449 * Correct spinlock nesting. The idle thread context that we are 1450 * borrowing was created so that it would start out with a single 1451 * spin lock (sched_lock) held in fork_trampoline(). Since we've 1452 * explicitly acquired locks in this function, the nesting count 1453 * is now 2 rather than 1. Since we are nested, calling 1454 * spinlock_exit() will simply adjust the counts without allowing 1455 * spin lock using code to interrupt us. 1456 */ 1457 if (td == NULL) { 1458 mtx_lock_spin(&sched_lock); 1459 spinlock_exit(); 1460 } else { 1461 MPASS(td->td_lock == &sched_lock); 1462 } 1463 mtx_assert(&sched_lock, MA_OWNED); 1464 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 1465 PCPU_SET(switchtime, cpu_ticks()); 1466 PCPU_SET(switchticks, ticks); 1467 cpu_throw(td, choosethread()); /* doesn't return */ 1468 } 1469 1470 void 1471 sched_fork_exit(struct thread *td) 1472 { 1473 1474 /* 1475 * Finish setting up thread glue so that it begins execution in a 1476 * non-nested critical section with sched_lock held but not recursed. 1477 */ 1478 td->td_oncpu = PCPU_GET(cpuid); 1479 sched_lock.mtx_lock = (uintptr_t)td; 1480 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED); 1481 } 1482 1483 void 1484 sched_affinity(struct thread *td) 1485 { 1486 #ifdef SMP 1487 struct td_sched *ts; 1488 int cpu; 1489 1490 THREAD_LOCK_ASSERT(td, MA_OWNED); 1491 1492 /* 1493 * Set the TSF_AFFINITY flag if there is at least one CPU this 1494 * thread can't run on. 1495 */ 1496 ts = td->td_sched; 1497 ts->ts_flags &= ~TSF_AFFINITY; 1498 for (cpu = 0; cpu <= mp_maxid; cpu++) { 1499 if (CPU_ABSENT(cpu)) 1500 continue; 1501 if (!THREAD_CAN_SCHED(td, cpu)) { 1502 ts->ts_flags |= TSF_AFFINITY; 1503 break; 1504 } 1505 } 1506 1507 /* 1508 * If this thread can run on all CPUs, nothing else to do. 1509 */ 1510 if (!(ts->ts_flags & TSF_AFFINITY)) 1511 return; 1512 1513 /* Pinned threads and bound threads should be left alone. */ 1514 if (td->td_pinned != 0 || td->td_flags & TDF_BOUND) 1515 return; 1516 1517 switch (td->td_state) { 1518 case TDS_RUNQ: 1519 /* 1520 * If we are on a per-CPU runqueue that is in the set, 1521 * then nothing needs to be done. 1522 */ 1523 if (ts->ts_runq != &runq && 1524 THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu)) 1525 return; 1526 1527 /* Put this thread on a valid per-CPU runqueue. */ 1528 sched_rem(td); 1529 sched_add(td, SRQ_BORING); 1530 break; 1531 case TDS_RUNNING: 1532 /* 1533 * See if our current CPU is in the set. If not, force a 1534 * context switch. 1535 */ 1536 if (THREAD_CAN_SCHED(td, td->td_oncpu)) 1537 return; 1538 1539 td->td_flags |= TDF_NEEDRESCHED; 1540 if (td != curthread) 1541 ipi_selected(1 << cpu, IPI_AST); 1542 break; 1543 default: 1544 break; 1545 } 1546 #endif 1547 } 1548 1549 #define KERN_SWITCH_INCLUDE 1 1550 #include "kern/kern_switch.c"
Cache object: 075587f6447d28e0a579323a5421f35a
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]