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/6.2/sys/kern/sched_4bsd.c 164286 2006-11-14 20:42:41Z cvs2svn $"); 37 38 #include "opt_hwpmc_hooks.h" 39 40 #define kse td_sched 41 42 #include <sys/param.h> 43 #include <sys/systm.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 <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 can be given a context to run. 78 * A process may have several of these. Probably one per processor 79 * but posibly a few more. In this universe they are grouped 80 * with a KSEG that contains the priority and niceness 81 * for the group. 82 */ 83 struct kse { 84 TAILQ_ENTRY(kse) ke_procq; /* (j/z) Run queue. */ 85 struct thread *ke_thread; /* (*) Active associated thread. */ 86 fixpt_t ke_pctcpu; /* (j) %cpu during p_swtime. */ 87 char ke_rqindex; /* (j) Run queue index. */ 88 enum { 89 KES_THREAD = 0x0, /* slaved to thread state */ 90 KES_ONRUNQ 91 } ke_state; /* (j) KSE status. */ 92 int ke_cpticks; /* (j) Ticks of cpu time. */ 93 struct runq *ke_runq; /* runq the kse is currently on */ 94 }; 95 96 #define ke_proc ke_thread->td_proc 97 #define ke_ksegrp ke_thread->td_ksegrp 98 99 #define td_kse td_sched 100 101 /* flags kept in td_flags */ 102 #define TDF_DIDRUN TDF_SCHED0 /* KSE actually ran. */ 103 #define TDF_EXIT TDF_SCHED1 /* KSE is being killed. */ 104 #define TDF_BOUND TDF_SCHED2 105 106 #define ke_flags ke_thread->td_flags 107 #define KEF_DIDRUN TDF_DIDRUN /* KSE actually ran. */ 108 #define KEF_EXIT TDF_EXIT /* KSE is being killed. */ 109 #define KEF_BOUND TDF_BOUND /* stuck to one CPU */ 110 111 #define SKE_RUNQ_PCPU(ke) \ 112 ((ke)->ke_runq != 0 && (ke)->ke_runq != &runq) 113 114 struct kg_sched { 115 struct thread *skg_last_assigned; /* (j) Last thread assigned to */ 116 /* the system scheduler. */ 117 int skg_avail_opennings; /* (j) Num KSEs requested in group. */ 118 int skg_concurrency; /* (j) Num KSEs requested in group. */ 119 }; 120 #define kg_last_assigned kg_sched->skg_last_assigned 121 #define kg_avail_opennings kg_sched->skg_avail_opennings 122 #define kg_concurrency kg_sched->skg_concurrency 123 124 #define SLOT_RELEASE(kg) \ 125 do { \ 126 kg->kg_avail_opennings++; \ 127 CTR3(KTR_RUNQ, "kg %p(%d) Slot released (->%d)", \ 128 kg, \ 129 kg->kg_concurrency, \ 130 kg->kg_avail_opennings); \ 131 /* KASSERT((kg->kg_avail_opennings <= kg->kg_concurrency), \ 132 ("slots out of whack"));*/ \ 133 } while (0) 134 135 #define SLOT_USE(kg) \ 136 do { \ 137 kg->kg_avail_opennings--; \ 138 CTR3(KTR_RUNQ, "kg %p(%d) Slot used (->%d)", \ 139 kg, \ 140 kg->kg_concurrency, \ 141 kg->kg_avail_opennings); \ 142 /* KASSERT((kg->kg_avail_opennings >= 0), \ 143 ("slots out of whack"));*/ \ 144 } while (0) 145 146 /* 147 * KSE_CAN_MIGRATE macro returns true if the kse can migrate between 148 * cpus. 149 */ 150 #define KSE_CAN_MIGRATE(ke) \ 151 ((ke)->ke_thread->td_pinned == 0 && ((ke)->ke_flags & KEF_BOUND) == 0) 152 153 static struct kse kse0; 154 static struct kg_sched kg_sched0; 155 156 static int sched_tdcnt; /* Total runnable threads in the system. */ 157 static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 158 #define SCHED_QUANTUM (hz / 10) /* Default sched quantum */ 159 160 static struct callout roundrobin_callout; 161 162 static void slot_fill(struct ksegrp *kg); 163 static struct kse *sched_choose(void); /* XXX Should be thread * */ 164 165 static void setup_runqs(void); 166 static void roundrobin(void *arg); 167 static void schedcpu(void); 168 static void schedcpu_thread(void); 169 static void sched_priority(struct thread *td, u_char prio); 170 static void sched_setup(void *dummy); 171 static void maybe_resched(struct thread *td); 172 static void updatepri(struct ksegrp *kg); 173 static void resetpriority(struct ksegrp *kg); 174 static void resetpriority_thread(struct thread *td, struct ksegrp *kg); 175 #ifdef SMP 176 static int forward_wakeup(int cpunum); 177 #endif 178 179 static struct kproc_desc sched_kp = { 180 "schedcpu", 181 schedcpu_thread, 182 NULL 183 }; 184 SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start, &sched_kp) 185 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL) 186 187 /* 188 * Global run queue. 189 */ 190 static struct runq runq; 191 192 #ifdef SMP 193 /* 194 * Per-CPU run queues 195 */ 196 static struct runq runq_pcpu[MAXCPU]; 197 #endif 198 199 static void 200 setup_runqs(void) 201 { 202 #ifdef SMP 203 int i; 204 205 for (i = 0; i < MAXCPU; ++i) 206 runq_init(&runq_pcpu[i]); 207 #endif 208 209 runq_init(&runq); 210 } 211 212 static int 213 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 214 { 215 int error, new_val; 216 217 new_val = sched_quantum * tick; 218 error = sysctl_handle_int(oidp, &new_val, 0, req); 219 if (error != 0 || req->newptr == NULL) 220 return (error); 221 if (new_val < tick) 222 return (EINVAL); 223 sched_quantum = new_val / tick; 224 hogticks = 2 * sched_quantum; 225 return (0); 226 } 227 228 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler"); 229 230 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0, 231 "Scheduler name"); 232 233 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW, 234 0, sizeof sched_quantum, sysctl_kern_quantum, "I", 235 "Roundrobin scheduling quantum in microseconds"); 236 237 #ifdef SMP 238 /* Enable forwarding of wakeups to all other cpus */ 239 SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP"); 240 241 static int forward_wakeup_enabled = 1; 242 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW, 243 &forward_wakeup_enabled, 0, 244 "Forwarding of wakeup to idle CPUs"); 245 246 static int forward_wakeups_requested = 0; 247 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD, 248 &forward_wakeups_requested, 0, 249 "Requests for Forwarding of wakeup to idle CPUs"); 250 251 static int forward_wakeups_delivered = 0; 252 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD, 253 &forward_wakeups_delivered, 0, 254 "Completed Forwarding of wakeup to idle CPUs"); 255 256 static int forward_wakeup_use_mask = 1; 257 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW, 258 &forward_wakeup_use_mask, 0, 259 "Use the mask of idle cpus"); 260 261 static int forward_wakeup_use_loop = 0; 262 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW, 263 &forward_wakeup_use_loop, 0, 264 "Use a loop to find idle cpus"); 265 266 static int forward_wakeup_use_single = 0; 267 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, onecpu, CTLFLAG_RW, 268 &forward_wakeup_use_single, 0, 269 "Only signal one idle cpu"); 270 271 static int forward_wakeup_use_htt = 0; 272 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, htt2, CTLFLAG_RW, 273 &forward_wakeup_use_htt, 0, 274 "account for htt"); 275 276 #endif 277 static int sched_followon = 0; 278 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW, 279 &sched_followon, 0, 280 "allow threads to share a quantum"); 281 282 static int sched_pfollowons = 0; 283 SYSCTL_INT(_kern_sched, OID_AUTO, pfollowons, CTLFLAG_RD, 284 &sched_pfollowons, 0, 285 "number of followons done to a different ksegrp"); 286 287 static int sched_kgfollowons = 0; 288 SYSCTL_INT(_kern_sched, OID_AUTO, kgfollowons, CTLFLAG_RD, 289 &sched_kgfollowons, 0, 290 "number of followons done in a ksegrp"); 291 292 static __inline void 293 sched_load_add(void) 294 { 295 sched_tdcnt++; 296 CTR1(KTR_SCHED, "global load: %d", sched_tdcnt); 297 } 298 299 static __inline void 300 sched_load_rem(void) 301 { 302 sched_tdcnt--; 303 CTR1(KTR_SCHED, "global load: %d", sched_tdcnt); 304 } 305 /* 306 * Arrange to reschedule if necessary, taking the priorities and 307 * schedulers into account. 308 */ 309 static void 310 maybe_resched(struct thread *td) 311 { 312 313 mtx_assert(&sched_lock, MA_OWNED); 314 if (td->td_priority < curthread->td_priority) 315 curthread->td_flags |= TDF_NEEDRESCHED; 316 } 317 318 /* 319 * Force switch among equal priority processes every 100ms. 320 * We don't actually need to force a context switch of the current process. 321 * The act of firing the event triggers a context switch to softclock() and 322 * then switching back out again which is equivalent to a preemption, thus 323 * no further work is needed on the local CPU. 324 */ 325 /* ARGSUSED */ 326 static void 327 roundrobin(void *arg) 328 { 329 330 #ifdef SMP 331 mtx_lock_spin(&sched_lock); 332 forward_roundrobin(); 333 mtx_unlock_spin(&sched_lock); 334 #endif 335 336 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 337 } 338 339 /* 340 * Constants for digital decay and forget: 341 * 90% of (kg_estcpu) usage in 5 * loadav time 342 * 95% of (ke_pctcpu) usage in 60 seconds (load insensitive) 343 * Note that, as ps(1) mentions, this can let percentages 344 * total over 100% (I've seen 137.9% for 3 processes). 345 * 346 * Note that schedclock() updates kg_estcpu and p_cpticks asynchronously. 347 * 348 * We wish to decay away 90% of kg_estcpu in (5 * loadavg) seconds. 349 * That is, the system wants to compute a value of decay such 350 * that the following for loop: 351 * for (i = 0; i < (5 * loadavg); i++) 352 * kg_estcpu *= decay; 353 * will compute 354 * kg_estcpu *= 0.1; 355 * for all values of loadavg: 356 * 357 * Mathematically this loop can be expressed by saying: 358 * decay ** (5 * loadavg) ~= .1 359 * 360 * The system computes decay as: 361 * decay = (2 * loadavg) / (2 * loadavg + 1) 362 * 363 * We wish to prove that the system's computation of decay 364 * will always fulfill the equation: 365 * decay ** (5 * loadavg) ~= .1 366 * 367 * If we compute b as: 368 * b = 2 * loadavg 369 * then 370 * decay = b / (b + 1) 371 * 372 * We now need to prove two things: 373 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 374 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 375 * 376 * Facts: 377 * For x close to zero, exp(x) =~ 1 + x, since 378 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 379 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 380 * For x close to zero, ln(1+x) =~ x, since 381 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 382 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 383 * ln(.1) =~ -2.30 384 * 385 * Proof of (1): 386 * Solve (factor)**(power) =~ .1 given power (5*loadav): 387 * solving for factor, 388 * ln(factor) =~ (-2.30/5*loadav), or 389 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 390 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 391 * 392 * Proof of (2): 393 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 394 * solving for power, 395 * power*ln(b/(b+1)) =~ -2.30, or 396 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 397 * 398 * Actual power values for the implemented algorithm are as follows: 399 * loadav: 1 2 3 4 400 * power: 5.68 10.32 14.94 19.55 401 */ 402 403 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 404 #define loadfactor(loadav) (2 * (loadav)) 405 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 406 407 /* decay 95% of `ke_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 408 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 409 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 410 411 /* 412 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 413 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 414 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 415 * 416 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 417 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 418 * 419 * If you don't want to bother with the faster/more-accurate formula, you 420 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 421 * (more general) method of calculating the %age of CPU used by a process. 422 */ 423 #define CCPU_SHIFT 11 424 425 /* 426 * Recompute process priorities, every hz ticks. 427 * MP-safe, called without the Giant mutex. 428 */ 429 /* ARGSUSED */ 430 static void 431 schedcpu(void) 432 { 433 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 434 struct thread *td; 435 struct proc *p; 436 struct kse *ke; 437 struct ksegrp *kg; 438 int awake, realstathz; 439 440 realstathz = stathz ? stathz : hz; 441 sx_slock(&allproc_lock); 442 FOREACH_PROC_IN_SYSTEM(p) { 443 /* 444 * Prevent state changes and protect run queue. 445 */ 446 mtx_lock_spin(&sched_lock); 447 /* 448 * Increment time in/out of memory. We ignore overflow; with 449 * 16-bit int's (remember them?) overflow takes 45 days. 450 */ 451 p->p_swtime++; 452 FOREACH_KSEGRP_IN_PROC(p, kg) { 453 awake = 0; 454 FOREACH_THREAD_IN_GROUP(kg, td) { 455 ke = td->td_kse; 456 /* 457 * Increment sleep time (if sleeping). We 458 * ignore overflow, as above. 459 */ 460 /* 461 * The kse slptimes are not touched in wakeup 462 * because the thread may not HAVE a KSE. 463 */ 464 if (ke->ke_state == KES_ONRUNQ) { 465 awake = 1; 466 ke->ke_flags &= ~KEF_DIDRUN; 467 } else if ((ke->ke_state == KES_THREAD) && 468 (TD_IS_RUNNING(td))) { 469 awake = 1; 470 /* Do not clear KEF_DIDRUN */ 471 } else if (ke->ke_flags & KEF_DIDRUN) { 472 awake = 1; 473 ke->ke_flags &= ~KEF_DIDRUN; 474 } 475 476 /* 477 * ke_pctcpu is only for ps and ttyinfo(). 478 * Do it per kse, and add them up at the end? 479 * XXXKSE 480 */ 481 ke->ke_pctcpu = (ke->ke_pctcpu * ccpu) >> 482 FSHIFT; 483 /* 484 * If the kse has been idle the entire second, 485 * stop recalculating its priority until 486 * it wakes up. 487 */ 488 if (ke->ke_cpticks == 0) 489 continue; 490 #if (FSHIFT >= CCPU_SHIFT) 491 ke->ke_pctcpu += (realstathz == 100) 492 ? ((fixpt_t) ke->ke_cpticks) << 493 (FSHIFT - CCPU_SHIFT) : 494 100 * (((fixpt_t) ke->ke_cpticks) 495 << (FSHIFT - CCPU_SHIFT)) / realstathz; 496 #else 497 ke->ke_pctcpu += ((FSCALE - ccpu) * 498 (ke->ke_cpticks * 499 FSCALE / realstathz)) >> FSHIFT; 500 #endif 501 ke->ke_cpticks = 0; 502 } /* end of kse loop */ 503 /* 504 * If there are ANY running threads in this KSEGRP, 505 * then don't count it as sleeping. 506 */ 507 if (awake) { 508 if (kg->kg_slptime > 1) { 509 /* 510 * In an ideal world, this should not 511 * happen, because whoever woke us 512 * up from the long sleep should have 513 * unwound the slptime and reset our 514 * priority before we run at the stale 515 * priority. Should KASSERT at some 516 * point when all the cases are fixed. 517 */ 518 updatepri(kg); 519 } 520 kg->kg_slptime = 0; 521 } else 522 kg->kg_slptime++; 523 if (kg->kg_slptime > 1) 524 continue; 525 kg->kg_estcpu = decay_cpu(loadfac, kg->kg_estcpu); 526 resetpriority(kg); 527 FOREACH_THREAD_IN_GROUP(kg, td) { 528 resetpriority_thread(td, kg); 529 } 530 } /* end of ksegrp loop */ 531 mtx_unlock_spin(&sched_lock); 532 } /* end of process loop */ 533 sx_sunlock(&allproc_lock); 534 } 535 536 /* 537 * Main loop for a kthread that executes schedcpu once a second. 538 */ 539 static void 540 schedcpu_thread(void) 541 { 542 int nowake; 543 544 for (;;) { 545 schedcpu(); 546 tsleep(&nowake, 0, "-", hz); 547 } 548 } 549 550 /* 551 * Recalculate the priority of a process after it has slept for a while. 552 * For all load averages >= 1 and max kg_estcpu of 255, sleeping for at 553 * least six times the loadfactor will decay kg_estcpu to zero. 554 */ 555 static void 556 updatepri(struct ksegrp *kg) 557 { 558 register fixpt_t loadfac; 559 register unsigned int newcpu; 560 561 loadfac = loadfactor(averunnable.ldavg[0]); 562 if (kg->kg_slptime > 5 * loadfac) 563 kg->kg_estcpu = 0; 564 else { 565 newcpu = kg->kg_estcpu; 566 kg->kg_slptime--; /* was incremented in schedcpu() */ 567 while (newcpu && --kg->kg_slptime) 568 newcpu = decay_cpu(loadfac, newcpu); 569 kg->kg_estcpu = newcpu; 570 } 571 } 572 573 /* 574 * Compute the priority of a process when running in user mode. 575 * Arrange to reschedule if the resulting priority is better 576 * than that of the current process. 577 */ 578 static void 579 resetpriority(struct ksegrp *kg) 580 { 581 register unsigned int newpriority; 582 583 if (kg->kg_pri_class == PRI_TIMESHARE) { 584 newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT + 585 NICE_WEIGHT * (kg->kg_proc->p_nice - PRIO_MIN); 586 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 587 PRI_MAX_TIMESHARE); 588 kg->kg_user_pri = newpriority; 589 } 590 } 591 592 /* 593 * Update the thread's priority when the associated ksegroup's user 594 * priority changes. 595 */ 596 static void 597 resetpriority_thread(struct thread *td, struct ksegrp *kg) 598 { 599 600 /* Only change threads with a time sharing user priority. */ 601 if (td->td_priority < PRI_MIN_TIMESHARE || 602 td->td_priority > PRI_MAX_TIMESHARE) 603 return; 604 605 /* XXX the whole needresched thing is broken, but not silly. */ 606 maybe_resched(td); 607 608 sched_prio(td, kg->kg_user_pri); 609 } 610 611 /* ARGSUSED */ 612 static void 613 sched_setup(void *dummy) 614 { 615 setup_runqs(); 616 617 if (sched_quantum == 0) 618 sched_quantum = SCHED_QUANTUM; 619 hogticks = 2 * sched_quantum; 620 621 callout_init(&roundrobin_callout, CALLOUT_MPSAFE); 622 623 /* Kick off timeout driven events by calling first time. */ 624 roundrobin(NULL); 625 626 /* Account for thread0. */ 627 sched_load_add(); 628 } 629 630 /* External interfaces start here */ 631 /* 632 * Very early in the boot some setup of scheduler-specific 633 * parts of proc0 and of some scheduler resources needs to be done. 634 * Called from: 635 * proc0_init() 636 */ 637 void 638 schedinit(void) 639 { 640 /* 641 * Set up the scheduler specific parts of proc0. 642 */ 643 proc0.p_sched = NULL; /* XXX */ 644 ksegrp0.kg_sched = &kg_sched0; 645 thread0.td_sched = &kse0; 646 kse0.ke_thread = &thread0; 647 kse0.ke_state = KES_THREAD; 648 kg_sched0.skg_concurrency = 1; 649 kg_sched0.skg_avail_opennings = 0; /* we are already running */ 650 } 651 652 int 653 sched_runnable(void) 654 { 655 #ifdef SMP 656 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]); 657 #else 658 return runq_check(&runq); 659 #endif 660 } 661 662 int 663 sched_rr_interval(void) 664 { 665 if (sched_quantum == 0) 666 sched_quantum = SCHED_QUANTUM; 667 return (sched_quantum); 668 } 669 670 /* 671 * We adjust the priority of the current process. The priority of 672 * a process gets worse as it accumulates CPU time. The cpu usage 673 * estimator (kg_estcpu) is increased here. resetpriority() will 674 * compute a different priority each time kg_estcpu increases by 675 * INVERSE_ESTCPU_WEIGHT 676 * (until MAXPRI is reached). The cpu usage estimator ramps up 677 * quite quickly when the process is running (linearly), and decays 678 * away exponentially, at a rate which is proportionally slower when 679 * the system is busy. The basic principle is that the system will 680 * 90% forget that the process used a lot of CPU time in 5 * loadav 681 * seconds. This causes the system to favor processes which haven't 682 * run much recently, and to round-robin among other processes. 683 */ 684 void 685 sched_clock(struct thread *td) 686 { 687 struct ksegrp *kg; 688 struct kse *ke; 689 690 mtx_assert(&sched_lock, MA_OWNED); 691 kg = td->td_ksegrp; 692 ke = td->td_kse; 693 694 ke->ke_cpticks++; 695 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1); 696 if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 697 resetpriority(kg); 698 resetpriority_thread(td, kg); 699 } 700 } 701 702 /* 703 * charge childs scheduling cpu usage to parent. 704 * 705 * XXXKSE assume only one thread & kse & ksegrp keep estcpu in each ksegrp. 706 * Charge it to the ksegrp that did the wait since process estcpu is sum of 707 * all ksegrps, this is strictly as expected. Assume that the child process 708 * aggregated all the estcpu into the 'built-in' ksegrp. 709 */ 710 void 711 sched_exit(struct proc *p, struct thread *td) 712 { 713 sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), td); 714 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td); 715 } 716 717 void 718 sched_exit_ksegrp(struct ksegrp *kg, struct thread *childtd) 719 { 720 721 mtx_assert(&sched_lock, MA_OWNED); 722 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + childtd->td_ksegrp->kg_estcpu); 723 } 724 725 void 726 sched_exit_thread(struct thread *td, struct thread *child) 727 { 728 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d", 729 child, child->td_proc->p_comm, child->td_priority); 730 if ((child->td_proc->p_flag & P_NOLOAD) == 0) 731 sched_load_rem(); 732 } 733 734 void 735 sched_fork(struct thread *td, struct thread *childtd) 736 { 737 sched_fork_ksegrp(td, childtd->td_ksegrp); 738 sched_fork_thread(td, childtd); 739 } 740 741 void 742 sched_fork_ksegrp(struct thread *td, struct ksegrp *child) 743 { 744 mtx_assert(&sched_lock, MA_OWNED); 745 child->kg_estcpu = td->td_ksegrp->kg_estcpu; 746 } 747 748 void 749 sched_fork_thread(struct thread *td, struct thread *childtd) 750 { 751 sched_newthread(childtd); 752 } 753 754 void 755 sched_nice(struct proc *p, int nice) 756 { 757 struct ksegrp *kg; 758 struct thread *td; 759 760 PROC_LOCK_ASSERT(p, MA_OWNED); 761 mtx_assert(&sched_lock, MA_OWNED); 762 p->p_nice = nice; 763 FOREACH_KSEGRP_IN_PROC(p, kg) { 764 resetpriority(kg); 765 FOREACH_THREAD_IN_GROUP(kg, td) { 766 resetpriority_thread(td, kg); 767 } 768 } 769 } 770 771 void 772 sched_class(struct ksegrp *kg, int class) 773 { 774 mtx_assert(&sched_lock, MA_OWNED); 775 kg->kg_pri_class = class; 776 } 777 778 /* 779 * Adjust the priority of a thread. 780 * This may include moving the thread within the KSEGRP, 781 * changing the assignment of a kse to the thread, 782 * and moving a KSE in the system run queue. 783 */ 784 static void 785 sched_priority(struct thread *td, u_char prio) 786 { 787 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 788 td, td->td_proc->p_comm, td->td_priority, prio, curthread, 789 curthread->td_proc->p_comm); 790 791 mtx_assert(&sched_lock, MA_OWNED); 792 if (td->td_priority == prio) 793 return; 794 if (TD_ON_RUNQ(td)) { 795 adjustrunqueue(td, prio); 796 } else { 797 td->td_priority = prio; 798 } 799 } 800 801 /* 802 * Update a thread's priority when it is lent another thread's 803 * priority. 804 */ 805 void 806 sched_lend_prio(struct thread *td, u_char prio) 807 { 808 809 td->td_flags |= TDF_BORROWING; 810 sched_priority(td, prio); 811 } 812 813 /* 814 * Restore a thread's priority when priority propagation is 815 * over. The prio argument is the minimum priority the thread 816 * needs to have to satisfy other possible priority lending 817 * requests. If the thread's regulary priority is less 818 * important than prio the thread will keep a priority boost 819 * of prio. 820 */ 821 void 822 sched_unlend_prio(struct thread *td, u_char prio) 823 { 824 u_char base_pri; 825 826 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 827 td->td_base_pri <= PRI_MAX_TIMESHARE) 828 base_pri = td->td_ksegrp->kg_user_pri; 829 else 830 base_pri = td->td_base_pri; 831 if (prio >= base_pri) { 832 td->td_flags &= ~TDF_BORROWING; 833 sched_prio(td, base_pri); 834 } else 835 sched_lend_prio(td, prio); 836 } 837 838 void 839 sched_prio(struct thread *td, u_char prio) 840 { 841 u_char oldprio; 842 843 /* First, update the base priority. */ 844 td->td_base_pri = prio; 845 846 /* 847 * If the thread is borrowing another thread's priority, don't ever 848 * lower the priority. 849 */ 850 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 851 return; 852 853 /* Change the real priority. */ 854 oldprio = td->td_priority; 855 sched_priority(td, prio); 856 857 /* 858 * If the thread is on a turnstile, then let the turnstile update 859 * its state. 860 */ 861 if (TD_ON_LOCK(td) && oldprio != prio) 862 turnstile_adjust(td, oldprio); 863 } 864 865 void 866 sched_sleep(struct thread *td) 867 { 868 869 mtx_assert(&sched_lock, MA_OWNED); 870 td->td_ksegrp->kg_slptime = 0; 871 } 872 873 static void remrunqueue(struct thread *td); 874 875 void 876 sched_switch(struct thread *td, struct thread *newtd, int flags) 877 { 878 struct kse *ke; 879 struct ksegrp *kg; 880 struct proc *p; 881 882 ke = td->td_kse; 883 p = td->td_proc; 884 885 mtx_assert(&sched_lock, MA_OWNED); 886 887 if ((p->p_flag & P_NOLOAD) == 0) 888 sched_load_rem(); 889 /* 890 * We are volunteering to switch out so we get to nominate 891 * a successor for the rest of our quantum 892 * First try another thread in our ksegrp, and then look for 893 * other ksegrps in our process. 894 */ 895 if (sched_followon && 896 (p->p_flag & P_HADTHREADS) && 897 (flags & SW_VOL) && 898 newtd == NULL) { 899 /* lets schedule another thread from this process */ 900 kg = td->td_ksegrp; 901 if ((newtd = TAILQ_FIRST(&kg->kg_runq))) { 902 remrunqueue(newtd); 903 sched_kgfollowons++; 904 } else { 905 FOREACH_KSEGRP_IN_PROC(p, kg) { 906 if ((newtd = TAILQ_FIRST(&kg->kg_runq))) { 907 sched_pfollowons++; 908 remrunqueue(newtd); 909 break; 910 } 911 } 912 } 913 } 914 915 if (newtd) 916 newtd->td_flags |= (td->td_flags & TDF_NEEDRESCHED); 917 918 td->td_lastcpu = td->td_oncpu; 919 td->td_flags &= ~TDF_NEEDRESCHED; 920 td->td_owepreempt = 0; 921 td->td_oncpu = NOCPU; 922 /* 923 * At the last moment, if this thread is still marked RUNNING, 924 * then put it back on the run queue as it has not been suspended 925 * or stopped or any thing else similar. We never put the idle 926 * threads on the run queue, however. 927 */ 928 if (td == PCPU_GET(idlethread)) 929 TD_SET_CAN_RUN(td); 930 else { 931 SLOT_RELEASE(td->td_ksegrp); 932 if (TD_IS_RUNNING(td)) { 933 /* Put us back on the run queue (kse and all). */ 934 setrunqueue(td, (flags & SW_PREEMPT) ? 935 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 936 SRQ_OURSELF|SRQ_YIELDING); 937 } else if (p->p_flag & P_HADTHREADS) { 938 /* 939 * We will not be on the run queue. So we must be 940 * sleeping or similar. As it's available, 941 * someone else can use the KSE if they need it. 942 * It's NOT available if we are about to need it 943 */ 944 if (newtd == NULL || newtd->td_ksegrp != td->td_ksegrp) 945 slot_fill(td->td_ksegrp); 946 } 947 } 948 if (newtd) { 949 /* 950 * The thread we are about to run needs to be counted 951 * as if it had been added to the run queue and selected. 952 * It came from: 953 * * A preemption 954 * * An upcall 955 * * A followon 956 */ 957 KASSERT((newtd->td_inhibitors == 0), 958 ("trying to run inhibitted thread")); 959 SLOT_USE(newtd->td_ksegrp); 960 newtd->td_kse->ke_flags |= KEF_DIDRUN; 961 TD_SET_RUNNING(newtd); 962 if ((newtd->td_proc->p_flag & P_NOLOAD) == 0) 963 sched_load_add(); 964 } else { 965 newtd = choosethread(); 966 } 967 968 if (td != newtd) { 969 #ifdef HWPMC_HOOKS 970 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 971 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 972 #endif 973 cpu_switch(td, newtd); 974 #ifdef HWPMC_HOOKS 975 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 976 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 977 #endif 978 } 979 980 sched_lock.mtx_lock = (uintptr_t)td; 981 td->td_oncpu = PCPU_GET(cpuid); 982 } 983 984 void 985 sched_wakeup(struct thread *td) 986 { 987 struct ksegrp *kg; 988 989 mtx_assert(&sched_lock, MA_OWNED); 990 kg = td->td_ksegrp; 991 if (kg->kg_slptime > 1) { 992 updatepri(kg); 993 resetpriority(kg); 994 } 995 kg->kg_slptime = 0; 996 setrunqueue(td, SRQ_BORING); 997 } 998 999 #ifdef SMP 1000 /* enable HTT_2 if you have a 2-way HTT cpu.*/ 1001 static int 1002 forward_wakeup(int cpunum) 1003 { 1004 cpumask_t map, me, dontuse; 1005 cpumask_t map2; 1006 struct pcpu *pc; 1007 cpumask_t id, map3; 1008 1009 mtx_assert(&sched_lock, MA_OWNED); 1010 1011 CTR0(KTR_RUNQ, "forward_wakeup()"); 1012 1013 if ((!forward_wakeup_enabled) || 1014 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0)) 1015 return (0); 1016 if (!smp_started || cold || panicstr) 1017 return (0); 1018 1019 forward_wakeups_requested++; 1020 1021 /* 1022 * check the idle mask we received against what we calculated before 1023 * in the old version. 1024 */ 1025 me = PCPU_GET(cpumask); 1026 /* 1027 * don't bother if we should be doing it ourself.. 1028 */ 1029 if ((me & idle_cpus_mask) && (cpunum == NOCPU || me == (1 << cpunum))) 1030 return (0); 1031 1032 dontuse = me | stopped_cpus | hlt_cpus_mask; 1033 map3 = 0; 1034 if (forward_wakeup_use_loop) { 1035 SLIST_FOREACH(pc, &cpuhead, pc_allcpu) { 1036 id = pc->pc_cpumask; 1037 if ( (id & dontuse) == 0 && 1038 pc->pc_curthread == pc->pc_idlethread) { 1039 map3 |= id; 1040 } 1041 } 1042 } 1043 1044 if (forward_wakeup_use_mask) { 1045 map = 0; 1046 map = idle_cpus_mask & ~dontuse; 1047 1048 /* If they are both on, compare and use loop if different */ 1049 if (forward_wakeup_use_loop) { 1050 if (map != map3) { 1051 printf("map (%02X) != map3 (%02X)\n", 1052 map, map3); 1053 map = map3; 1054 } 1055 } 1056 } else { 1057 map = map3; 1058 } 1059 /* If we only allow a specific CPU, then mask off all the others */ 1060 if (cpunum != NOCPU) { 1061 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum.")); 1062 map &= (1 << cpunum); 1063 } else { 1064 /* Try choose an idle die. */ 1065 if (forward_wakeup_use_htt) { 1066 map2 = (map & (map >> 1)) & 0x5555; 1067 if (map2) { 1068 map = map2; 1069 } 1070 } 1071 1072 /* set only one bit */ 1073 if (forward_wakeup_use_single) { 1074 map = map & ((~map) + 1); 1075 } 1076 } 1077 if (map) { 1078 forward_wakeups_delivered++; 1079 ipi_selected(map, IPI_AST); 1080 return (1); 1081 } 1082 if (cpunum == NOCPU) 1083 printf("forward_wakeup: Idle processor not found\n"); 1084 return (0); 1085 } 1086 #endif 1087 1088 #ifdef SMP 1089 static void kick_other_cpu(int pri,int cpuid); 1090 1091 static void 1092 kick_other_cpu(int pri,int cpuid) 1093 { 1094 struct pcpu * pcpu = pcpu_find(cpuid); 1095 int cpri = pcpu->pc_curthread->td_priority; 1096 1097 if (idle_cpus_mask & pcpu->pc_cpumask) { 1098 forward_wakeups_delivered++; 1099 ipi_selected(pcpu->pc_cpumask, IPI_AST); 1100 return; 1101 } 1102 1103 if (pri >= cpri) 1104 return; 1105 1106 #if defined(IPI_PREEMPTION) && defined(PREEMPTION) 1107 #if !defined(FULL_PREEMPTION) 1108 if (pri <= PRI_MAX_ITHD) 1109 #endif /* ! FULL_PREEMPTION */ 1110 { 1111 ipi_selected(pcpu->pc_cpumask, IPI_PREEMPT); 1112 return; 1113 } 1114 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */ 1115 1116 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED; 1117 ipi_selected( pcpu->pc_cpumask , IPI_AST); 1118 return; 1119 } 1120 #endif /* SMP */ 1121 1122 void 1123 sched_add(struct thread *td, int flags) 1124 #ifdef SMP 1125 { 1126 struct kse *ke; 1127 int forwarded = 0; 1128 int cpu; 1129 int single_cpu = 0; 1130 1131 ke = td->td_kse; 1132 mtx_assert(&sched_lock, MA_OWNED); 1133 KASSERT(ke->ke_state != KES_ONRUNQ, 1134 ("sched_add: kse %p (%s) already in run queue", ke, 1135 ke->ke_proc->p_comm)); 1136 KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 1137 ("sched_add: process swapped out")); 1138 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 1139 td, td->td_proc->p_comm, td->td_priority, curthread, 1140 curthread->td_proc->p_comm); 1141 1142 1143 if (td->td_pinned != 0) { 1144 cpu = td->td_lastcpu; 1145 ke->ke_runq = &runq_pcpu[cpu]; 1146 single_cpu = 1; 1147 CTR3(KTR_RUNQ, 1148 "sched_add: Put kse:%p(td:%p) on cpu%d runq", ke, td, cpu); 1149 } else if ((ke)->ke_flags & KEF_BOUND) { 1150 /* Find CPU from bound runq */ 1151 KASSERT(SKE_RUNQ_PCPU(ke),("sched_add: bound kse not on cpu runq")); 1152 cpu = ke->ke_runq - &runq_pcpu[0]; 1153 single_cpu = 1; 1154 CTR3(KTR_RUNQ, 1155 "sched_add: Put kse:%p(td:%p) on cpu%d runq", ke, td, cpu); 1156 } else { 1157 CTR2(KTR_RUNQ, 1158 "sched_add: adding kse:%p (td:%p) to gbl runq", ke, td); 1159 cpu = NOCPU; 1160 ke->ke_runq = &runq; 1161 } 1162 1163 if (single_cpu && (cpu != PCPU_GET(cpuid))) { 1164 kick_other_cpu(td->td_priority,cpu); 1165 } else { 1166 1167 if (!single_cpu) { 1168 cpumask_t me = PCPU_GET(cpumask); 1169 int idle = idle_cpus_mask & me; 1170 1171 if (!idle && ((flags & SRQ_INTR) == 0) && 1172 (idle_cpus_mask & ~(hlt_cpus_mask | me))) 1173 forwarded = forward_wakeup(cpu); 1174 } 1175 1176 if (!forwarded) { 1177 if ((flags & SRQ_YIELDING) == 0 && maybe_preempt(td)) 1178 return; 1179 else 1180 maybe_resched(td); 1181 } 1182 } 1183 1184 if ((td->td_proc->p_flag & P_NOLOAD) == 0) 1185 sched_load_add(); 1186 SLOT_USE(td->td_ksegrp); 1187 runq_add(ke->ke_runq, ke, flags); 1188 ke->ke_state = KES_ONRUNQ; 1189 } 1190 #else /* SMP */ 1191 { 1192 struct kse *ke; 1193 ke = td->td_kse; 1194 mtx_assert(&sched_lock, MA_OWNED); 1195 KASSERT(ke->ke_state != KES_ONRUNQ, 1196 ("sched_add: kse %p (%s) already in run queue", ke, 1197 ke->ke_proc->p_comm)); 1198 KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 1199 ("sched_add: process swapped out")); 1200 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 1201 td, td->td_proc->p_comm, td->td_priority, curthread, 1202 curthread->td_proc->p_comm); 1203 CTR2(KTR_RUNQ, "sched_add: adding kse:%p (td:%p) to runq", ke, td); 1204 ke->ke_runq = &runq; 1205 1206 /* 1207 * If we are yielding (on the way out anyhow) 1208 * or the thread being saved is US, 1209 * then don't try be smart about preemption 1210 * or kicking off another CPU 1211 * as it won't help and may hinder. 1212 * In the YIEDLING case, we are about to run whoever is 1213 * being put in the queue anyhow, and in the 1214 * OURSELF case, we are puting ourself on the run queue 1215 * which also only happens when we are about to yield. 1216 */ 1217 if((flags & SRQ_YIELDING) == 0) { 1218 if (maybe_preempt(td)) 1219 return; 1220 } 1221 if ((td->td_proc->p_flag & P_NOLOAD) == 0) 1222 sched_load_add(); 1223 SLOT_USE(td->td_ksegrp); 1224 runq_add(ke->ke_runq, ke, flags); 1225 ke->ke_state = KES_ONRUNQ; 1226 maybe_resched(td); 1227 } 1228 #endif /* SMP */ 1229 1230 void 1231 sched_rem(struct thread *td) 1232 { 1233 struct kse *ke; 1234 1235 ke = td->td_kse; 1236 KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 1237 ("sched_rem: process swapped out")); 1238 KASSERT((ke->ke_state == KES_ONRUNQ), 1239 ("sched_rem: KSE not on run queue")); 1240 mtx_assert(&sched_lock, MA_OWNED); 1241 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 1242 td, td->td_proc->p_comm, td->td_priority, curthread, 1243 curthread->td_proc->p_comm); 1244 1245 if ((td->td_proc->p_flag & P_NOLOAD) == 0) 1246 sched_load_rem(); 1247 SLOT_RELEASE(td->td_ksegrp); 1248 runq_remove(ke->ke_runq, ke); 1249 1250 ke->ke_state = KES_THREAD; 1251 } 1252 1253 /* 1254 * Select threads to run. 1255 * Notice that the running threads still consume a slot. 1256 */ 1257 struct kse * 1258 sched_choose(void) 1259 { 1260 struct kse *ke; 1261 struct runq *rq; 1262 1263 #ifdef SMP 1264 struct kse *kecpu; 1265 1266 rq = &runq; 1267 ke = runq_choose(&runq); 1268 kecpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]); 1269 1270 if (ke == NULL || 1271 (kecpu != NULL && 1272 kecpu->ke_thread->td_priority < ke->ke_thread->td_priority)) { 1273 CTR2(KTR_RUNQ, "choosing kse %p from pcpu runq %d", kecpu, 1274 PCPU_GET(cpuid)); 1275 ke = kecpu; 1276 rq = &runq_pcpu[PCPU_GET(cpuid)]; 1277 } else { 1278 CTR1(KTR_RUNQ, "choosing kse %p from main runq", ke); 1279 } 1280 1281 #else 1282 rq = &runq; 1283 ke = runq_choose(&runq); 1284 #endif 1285 1286 if (ke != NULL) { 1287 runq_remove(rq, ke); 1288 ke->ke_state = KES_THREAD; 1289 1290 KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 1291 ("sched_choose: process swapped out")); 1292 } 1293 return (ke); 1294 } 1295 1296 void 1297 sched_userret(struct thread *td) 1298 { 1299 struct ksegrp *kg; 1300 /* 1301 * XXX we cheat slightly on the locking here to avoid locking in 1302 * the usual case. Setting td_priority here is essentially an 1303 * incomplete workaround for not setting it properly elsewhere. 1304 * Now that some interrupt handlers are threads, not setting it 1305 * properly elsewhere can clobber it in the window between setting 1306 * it here and returning to user mode, so don't waste time setting 1307 * it perfectly here. 1308 */ 1309 KASSERT((td->td_flags & TDF_BORROWING) == 0, 1310 ("thread with borrowed priority returning to userland")); 1311 kg = td->td_ksegrp; 1312 if (td->td_priority != kg->kg_user_pri) { 1313 mtx_lock_spin(&sched_lock); 1314 td->td_priority = kg->kg_user_pri; 1315 td->td_base_pri = kg->kg_user_pri; 1316 mtx_unlock_spin(&sched_lock); 1317 } 1318 } 1319 1320 void 1321 sched_bind(struct thread *td, int cpu) 1322 { 1323 struct kse *ke; 1324 1325 mtx_assert(&sched_lock, MA_OWNED); 1326 KASSERT(TD_IS_RUNNING(td), 1327 ("sched_bind: cannot bind non-running thread")); 1328 1329 ke = td->td_kse; 1330 1331 ke->ke_flags |= KEF_BOUND; 1332 #ifdef SMP 1333 ke->ke_runq = &runq_pcpu[cpu]; 1334 if (PCPU_GET(cpuid) == cpu) 1335 return; 1336 1337 ke->ke_state = KES_THREAD; 1338 1339 mi_switch(SW_VOL, NULL); 1340 #endif 1341 } 1342 1343 void 1344 sched_unbind(struct thread* td) 1345 { 1346 mtx_assert(&sched_lock, MA_OWNED); 1347 td->td_kse->ke_flags &= ~KEF_BOUND; 1348 } 1349 1350 int 1351 sched_is_bound(struct thread *td) 1352 { 1353 mtx_assert(&sched_lock, MA_OWNED); 1354 return (td->td_kse->ke_flags & KEF_BOUND); 1355 } 1356 1357 int 1358 sched_load(void) 1359 { 1360 return (sched_tdcnt); 1361 } 1362 1363 int 1364 sched_sizeof_ksegrp(void) 1365 { 1366 return (sizeof(struct ksegrp) + sizeof(struct kg_sched)); 1367 } 1368 int 1369 sched_sizeof_proc(void) 1370 { 1371 return (sizeof(struct proc)); 1372 } 1373 int 1374 sched_sizeof_thread(void) 1375 { 1376 return (sizeof(struct thread) + sizeof(struct kse)); 1377 } 1378 1379 fixpt_t 1380 sched_pctcpu(struct thread *td) 1381 { 1382 struct kse *ke; 1383 1384 ke = td->td_kse; 1385 return (ke->ke_pctcpu); 1386 1387 return (0); 1388 } 1389 #define KERN_SWITCH_INCLUDE 1 1390 #include "kern/kern_switch.c"
Cache object: fcf4fa5af682bdcb7493548d76945a66
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]