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 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 *
| 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 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 *
|
69 70struct ke_sched *kse0_sched = &ke_sched; 71struct kg_sched *ksegrp0_sched = NULL; 72struct p_sched *proc0_sched = NULL; 73struct td_sched *thread0_sched = NULL; 74 75static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 76#define SCHED_QUANTUM (hz / 10) /* Default sched quantum */ 77 78static struct callout schedcpu_callout; 79static struct callout roundrobin_callout; 80 81static void roundrobin(void *arg); 82static void schedcpu(void *arg); 83static void sched_setup(void *dummy); 84static void maybe_resched(struct thread *td); 85static void updatepri(struct ksegrp *kg); 86static void resetpriority(struct ksegrp *kg); 87 88SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 89 90/* 91 * Global run queue. 92 */ 93static struct runq runq; 94SYSINIT(runq, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, runq_init, &runq) 95 96static int 97sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 98{ 99 int error, new_val; 100 101 new_val = sched_quantum * tick; 102 error = sysctl_handle_int(oidp, &new_val, 0, req); 103 if (error != 0 || req->newptr == NULL) 104 return (error); 105 if (new_val < tick) 106 return (EINVAL); 107 sched_quantum = new_val / tick; 108 hogticks = 2 * sched_quantum; 109 return (0); 110} 111 112SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 113 0, sizeof sched_quantum, sysctl_kern_quantum, "I", 114 "Roundrobin scheduling quantum in microseconds"); 115 116/* 117 * Arrange to reschedule if necessary, taking the priorities and 118 * schedulers into account. 119 */ 120static void 121maybe_resched(struct thread *td) 122{ 123 124 mtx_assert(&sched_lock, MA_OWNED); 125 if (td->td_priority < curthread->td_priority && curthread->td_kse) 126 curthread->td_flags |= TDF_NEEDRESCHED; 127} 128 129/* 130 * Force switch among equal priority processes every 100ms. 131 * We don't actually need to force a context switch of the current process. 132 * The act of firing the event triggers a context switch to softclock() and 133 * then switching back out again which is equivalent to a preemption, thus 134 * no further work is needed on the local CPU. 135 */ 136/* ARGSUSED */ 137static void 138roundrobin(void *arg) 139{ 140 141#ifdef SMP 142 mtx_lock_spin(&sched_lock); 143 forward_roundrobin(); 144 mtx_unlock_spin(&sched_lock); 145#endif 146 147 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 148} 149 150/* 151 * Constants for digital decay and forget: 152 * 90% of (p_estcpu) usage in 5 * loadav time 153 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 154 * Note that, as ps(1) mentions, this can let percentages 155 * total over 100% (I've seen 137.9% for 3 processes). 156 * 157 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 158 * 159 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 160 * That is, the system wants to compute a value of decay such 161 * that the following for loop: 162 * for (i = 0; i < (5 * loadavg); i++) 163 * p_estcpu *= decay; 164 * will compute 165 * p_estcpu *= 0.1; 166 * for all values of loadavg: 167 * 168 * Mathematically this loop can be expressed by saying: 169 * decay ** (5 * loadavg) ~= .1 170 * 171 * The system computes decay as: 172 * decay = (2 * loadavg) / (2 * loadavg + 1) 173 * 174 * We wish to prove that the system's computation of decay 175 * will always fulfill the equation: 176 * decay ** (5 * loadavg) ~= .1 177 * 178 * If we compute b as: 179 * b = 2 * loadavg 180 * then 181 * decay = b / (b + 1) 182 * 183 * We now need to prove two things: 184 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 185 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 186 * 187 * Facts: 188 * For x close to zero, exp(x) =~ 1 + x, since 189 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 190 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 191 * For x close to zero, ln(1+x) =~ x, since 192 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 193 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 194 * ln(.1) =~ -2.30 195 * 196 * Proof of (1): 197 * Solve (factor)**(power) =~ .1 given power (5*loadav): 198 * solving for factor, 199 * ln(factor) =~ (-2.30/5*loadav), or 200 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 201 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 202 * 203 * Proof of (2): 204 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 205 * solving for power, 206 * power*ln(b/(b+1)) =~ -2.30, or 207 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 208 * 209 * Actual power values for the implemented algorithm are as follows: 210 * loadav: 1 2 3 4 211 * power: 5.68 10.32 14.94 19.55 212 */ 213 214/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 215#define loadfactor(loadav) (2 * (loadav)) 216#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 217 218/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 219static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 220SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 221 222/* 223 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 224 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 225 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 226 * 227 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 228 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 229 * 230 * If you don't want to bother with the faster/more-accurate formula, you 231 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 232 * (more general) method of calculating the %age of CPU used by a process. 233 */ 234#define CCPU_SHIFT 11 235 236/* 237 * Recompute process priorities, every hz ticks. 238 * MP-safe, called without the Giant mutex. 239 */ 240/* ARGSUSED */ 241static void 242schedcpu(void *arg) 243{ 244 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 245 struct thread *td; 246 struct proc *p; 247 struct kse *ke; 248 struct ksegrp *kg; 249 int realstathz; 250 int awake; 251 252 realstathz = stathz ? stathz : hz; 253 sx_slock(&allproc_lock); 254 FOREACH_PROC_IN_SYSTEM(p) { 255 mtx_lock_spin(&sched_lock); 256 p->p_swtime++; 257 FOREACH_KSEGRP_IN_PROC(p, kg) { 258 awake = 0; 259 FOREACH_KSE_IN_GROUP(kg, ke) { 260 /* 261 * Increment time in/out of memory and sleep 262 * time (if sleeping). We ignore overflow; 263 * with 16-bit int's (remember them?) 264 * overflow takes 45 days. 265 */ 266 /* 267 * The kse slptimes are not touched in wakeup 268 * because the thread may not HAVE a KSE. 269 */ 270 if (ke->ke_state == KES_ONRUNQ) { 271 awake = 1; 272 ke->ke_flags &= ~KEF_DIDRUN; 273 } else if ((ke->ke_state == KES_THREAD) && 274 (TD_IS_RUNNING(ke->ke_thread))) { 275 awake = 1; 276 /* Do not clear KEF_DIDRUN */ 277 } else if (ke->ke_flags & KEF_DIDRUN) { 278 awake = 1; 279 ke->ke_flags &= ~KEF_DIDRUN; 280 } 281 282 /* 283 * pctcpu is only for ps? 284 * Do it per kse.. and add them up at the end? 285 * XXXKSE 286 */ 287 ke->ke_pctcpu 288 = (ke->ke_pctcpu * ccpu) >> 289 FSHIFT; 290 /* 291 * If the kse has been idle the entire second, 292 * stop recalculating its priority until 293 * it wakes up. 294 */ 295 if (ke->ke_sched->ske_cpticks == 0) 296 continue; 297#if (FSHIFT >= CCPU_SHIFT) 298 ke->ke_pctcpu += (realstathz == 100) 299 ? ((fixpt_t) ke->ke_sched->ske_cpticks) << 300 (FSHIFT - CCPU_SHIFT) : 301 100 * (((fixpt_t) ke->ke_sched->ske_cpticks) 302 << (FSHIFT - CCPU_SHIFT)) / realstathz; 303#else 304 ke->ke_pctcpu += ((FSCALE - ccpu) * 305 (ke->ke_sched->ske_cpticks * 306 FSCALE / realstathz)) >> FSHIFT; 307#endif 308 ke->ke_sched->ske_cpticks = 0; 309 } /* end of kse loop */ 310 /* 311 * If there are ANY running threads in this KSEGRP, 312 * then don't count it as sleeping. 313 */ 314 if (awake) { 315 if (kg->kg_slptime > 1) { 316 /* 317 * In an ideal world, this should not 318 * happen, because whoever woke us 319 * up from the long sleep should have 320 * unwound the slptime and reset our 321 * priority before we run at the stale 322 * priority. Should KASSERT at some 323 * point when all the cases are fixed. 324 */ 325 updatepri(kg); 326 } 327 kg->kg_slptime = 0; 328 } else { 329 kg->kg_slptime++; 330 } 331 if (kg->kg_slptime > 1) 332 continue; 333 kg->kg_estcpu = decay_cpu(loadfac, kg->kg_estcpu); 334 resetpriority(kg); 335 FOREACH_THREAD_IN_GROUP(kg, td) { 336 if (td->td_priority >= PUSER) { 337 sched_prio(td, kg->kg_user_pri); 338 } 339 } 340 } /* end of ksegrp loop */ 341 mtx_unlock_spin(&sched_lock); 342 } /* end of process loop */ 343 sx_sunlock(&allproc_lock); 344 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 345} 346 347/* 348 * Recalculate the priority of a process after it has slept for a while. 349 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 350 * least six times the loadfactor will decay p_estcpu to zero. 351 */ 352static void 353updatepri(struct ksegrp *kg) 354{ 355 register unsigned int newcpu; 356 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 357 358 newcpu = kg->kg_estcpu; 359 if (kg->kg_slptime > 5 * loadfac) 360 kg->kg_estcpu = 0; 361 else { 362 kg->kg_slptime--; /* the first time was done in schedcpu */ 363 while (newcpu && --kg->kg_slptime) 364 newcpu = decay_cpu(loadfac, newcpu); 365 kg->kg_estcpu = newcpu; 366 } 367 resetpriority(kg); 368} 369 370/* 371 * Compute the priority of a process when running in user mode. 372 * Arrange to reschedule if the resulting priority is better 373 * than that of the current process. 374 */ 375static void 376resetpriority(struct ksegrp *kg) 377{ 378 register unsigned int newpriority; 379 struct thread *td; 380 381 if (kg->kg_pri_class == PRI_TIMESHARE) { 382 newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT + 383 NICE_WEIGHT * (kg->kg_nice - PRIO_MIN); 384 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 385 PRI_MAX_TIMESHARE); 386 kg->kg_user_pri = newpriority; 387 } 388 FOREACH_THREAD_IN_GROUP(kg, td) { 389 maybe_resched(td); /* XXXKSE silly */ 390 } 391} 392 393/* ARGSUSED */ 394static void 395sched_setup(void *dummy) 396{ 397 if (sched_quantum == 0) 398 sched_quantum = SCHED_QUANTUM; 399 hogticks = 2 * sched_quantum; 400 401 callout_init(&schedcpu_callout, 1); 402 callout_init(&roundrobin_callout, 0); 403 404 /* Kick off timeout driven events by calling first time. */ 405 roundrobin(NULL); 406 schedcpu(NULL); 407} 408 409/* External interfaces start here */ 410int 411sched_runnable(void) 412{ 413 return runq_check(&runq); 414} 415 416int 417sched_rr_interval(void) 418{ 419 if (sched_quantum == 0) 420 sched_quantum = SCHED_QUANTUM; 421 return (sched_quantum); 422} 423 424/* 425 * We adjust the priority of the current process. The priority of 426 * a process gets worse as it accumulates CPU time. The cpu usage 427 * estimator (p_estcpu) is increased here. resetpriority() will 428 * compute a different priority each time p_estcpu increases by 429 * INVERSE_ESTCPU_WEIGHT 430 * (until MAXPRI is reached). The cpu usage estimator ramps up 431 * quite quickly when the process is running (linearly), and decays 432 * away exponentially, at a rate which is proportionally slower when 433 * the system is busy. The basic principle is that the system will 434 * 90% forget that the process used a lot of CPU time in 5 * loadav 435 * seconds. This causes the system to favor processes which haven't 436 * run much recently, and to round-robin among other processes. 437 */ 438void 439sched_clock(struct kse *ke) 440{ 441 struct ksegrp *kg; 442 struct thread *td; 443 444 mtx_assert(&sched_lock, MA_OWNED); 445 kg = ke->ke_ksegrp; 446 td = ke->ke_thread; 447 448 ke->ke_sched->ske_cpticks++; 449 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1); 450 if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 451 resetpriority(kg); 452 if (td->td_priority >= PUSER) 453 td->td_priority = kg->kg_user_pri; 454 } 455} 456/* 457 * charge childs scheduling cpu usage to parent. 458 * 459 * XXXKSE assume only one thread & kse & ksegrp keep estcpu in each ksegrp. 460 * Charge it to the ksegrp that did the wait since process estcpu is sum of 461 * all ksegrps, this is strictly as expected. Assume that the child process 462 * aggregated all the estcpu into the 'built-in' ksegrp. 463 */ 464void 465sched_exit(struct proc *p, struct proc *p1) 466{ 467 sched_exit_kse(FIRST_KSE_IN_PROC(p), FIRST_KSE_IN_PROC(p1)); 468 sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), FIRST_KSEGRP_IN_PROC(p1)); 469 sched_exit_thread(FIRST_THREAD_IN_PROC(p), FIRST_THREAD_IN_PROC(p1)); 470} 471 472void 473sched_exit_kse(struct kse *ke, struct kse *child) 474{ 475} 476 477void 478sched_exit_ksegrp(struct ksegrp *kg, struct ksegrp *child) 479{ 480 481 mtx_assert(&sched_lock, MA_OWNED); 482 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + child->kg_estcpu); 483} 484 485void 486sched_exit_thread(struct thread *td, struct thread *child) 487{ 488} 489 490void 491sched_fork(struct proc *p, struct proc *p1) 492{ 493 sched_fork_kse(FIRST_KSE_IN_PROC(p), FIRST_KSE_IN_PROC(p1)); 494 sched_fork_ksegrp(FIRST_KSEGRP_IN_PROC(p), FIRST_KSEGRP_IN_PROC(p1)); 495 sched_fork_thread(FIRST_THREAD_IN_PROC(p), FIRST_THREAD_IN_PROC(p1)); 496} 497 498void 499sched_fork_kse(struct kse *ke, struct kse *child) 500{ 501 child->ke_sched->ske_cpticks = 0; 502} 503 504void 505sched_fork_ksegrp(struct ksegrp *kg, struct ksegrp *child) 506{ 507 mtx_assert(&sched_lock, MA_OWNED); 508 child->kg_estcpu = kg->kg_estcpu; 509} 510 511void 512sched_fork_thread(struct thread *td, struct thread *child) 513{ 514} 515 516void 517sched_nice(struct ksegrp *kg, int nice) 518{ 519 520 PROC_LOCK_ASSERT(kg->kg_proc, MA_OWNED); 521 mtx_assert(&sched_lock, MA_OWNED); 522 kg->kg_nice = nice; 523 resetpriority(kg); 524} 525 526void 527sched_class(struct ksegrp *kg, int class) 528{ 529 mtx_assert(&sched_lock, MA_OWNED); 530 kg->kg_pri_class = class; 531} 532 533/* 534 * Adjust the priority of a thread. 535 * This may include moving the thread within the KSEGRP, 536 * changing the assignment of a kse to the thread, 537 * and moving a KSE in the system run queue. 538 */ 539void 540sched_prio(struct thread *td, u_char prio) 541{ 542 543 mtx_assert(&sched_lock, MA_OWNED); 544 if (TD_ON_RUNQ(td)) { 545 adjustrunqueue(td, prio); 546 } else { 547 td->td_priority = prio; 548 } 549} 550 551void 552sched_sleep(struct thread *td, u_char prio) 553{ 554 555 mtx_assert(&sched_lock, MA_OWNED); 556 td->td_ksegrp->kg_slptime = 0; 557 td->td_priority = prio; 558} 559 560void 561sched_switchin(struct thread *td) 562{ 563 564 mtx_assert(&sched_lock, MA_OWNED); 565 td->td_oncpu = PCPU_GET(cpuid); 566} 567 568void 569sched_switchout(struct thread *td) 570{ 571 struct kse *ke; 572 struct proc *p; 573 574 ke = td->td_kse; 575 p = td->td_proc; 576 577 mtx_assert(&sched_lock, MA_OWNED); 578 KASSERT((ke->ke_state == KES_THREAD), ("mi_switch: kse state?")); 579 580 td->td_lastcpu = td->td_oncpu; 581 td->td_last_kse = ke; 582 td->td_oncpu = NOCPU; 583 td->td_flags &= ~TDF_NEEDRESCHED; 584 /* 585 * At the last moment, if this thread is still marked RUNNING, 586 * then put it back on the run queue as it has not been suspended 587 * or stopped or any thing else similar. 588 */ 589 if (TD_IS_RUNNING(td)) { 590 /* Put us back on the run queue (kse and all). */ 591 setrunqueue(td); 592 } else if (p->p_flag & P_THREADED) { 593 /* 594 * We will not be on the run queue. So we must be 595 * sleeping or similar. As it's available, 596 * someone else can use the KSE if they need it. 597 */ 598 kse_reassign(ke); 599 } 600} 601 602void 603sched_wakeup(struct thread *td) 604{ 605 struct ksegrp *kg; 606 607 mtx_assert(&sched_lock, MA_OWNED); 608 kg = td->td_ksegrp; 609 if (kg->kg_slptime > 1) 610 updatepri(kg); 611 kg->kg_slptime = 0; 612 setrunqueue(td); 613 maybe_resched(td); 614} 615 616void 617sched_add(struct kse *ke) 618{ 619 mtx_assert(&sched_lock, MA_OWNED); 620 KASSERT((ke->ke_thread != NULL), ("runq_add: No thread on KSE")); 621 KASSERT((ke->ke_thread->td_kse != NULL), 622 ("runq_add: No KSE on thread")); 623 KASSERT(ke->ke_state != KES_ONRUNQ, 624 ("runq_add: kse %p (%s) already in run queue", ke, 625 ke->ke_proc->p_comm)); 626 KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 627 ("runq_add: process swapped out")); 628 ke->ke_ksegrp->kg_runq_kses++; 629 ke->ke_state = KES_ONRUNQ; 630 631 runq_add(&runq, ke); 632} 633 634void 635sched_rem(struct kse *ke) 636{ 637 KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 638 ("runq_remove: process swapped out")); 639 KASSERT((ke->ke_state == KES_ONRUNQ), ("KSE not on run queue")); 640 mtx_assert(&sched_lock, MA_OWNED); 641 642 runq_remove(&runq, ke); 643 ke->ke_state = KES_THREAD; 644 ke->ke_ksegrp->kg_runq_kses--; 645} 646 647struct kse * 648sched_choose(void) 649{ 650 struct kse *ke; 651 652 ke = runq_choose(&runq); 653 654 if (ke != NULL) { 655 runq_remove(&runq, ke); 656 ke->ke_state = KES_THREAD; 657 658 KASSERT((ke->ke_thread != NULL), 659 ("runq_choose: No thread on KSE")); 660 KASSERT((ke->ke_thread->td_kse != NULL), 661 ("runq_choose: No KSE on thread")); 662 KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 663 ("runq_choose: process swapped out")); 664 } 665 return (ke); 666} 667 668void 669sched_userret(struct thread *td) 670{ 671 struct ksegrp *kg; 672 /* 673 * XXX we cheat slightly on the locking here to avoid locking in 674 * the usual case. Setting td_priority here is essentially an 675 * incomplete workaround for not setting it properly elsewhere. 676 * Now that some interrupt handlers are threads, not setting it 677 * properly elsewhere can clobber it in the window between setting 678 * it here and returning to user mode, so don't waste time setting 679 * it perfectly here. 680 */ 681 kg = td->td_ksegrp; 682 if (td->td_priority != kg->kg_user_pri) { 683 mtx_lock_spin(&sched_lock); 684 td->td_priority = kg->kg_user_pri; 685 mtx_unlock_spin(&sched_lock); 686 } 687} 688 689int 690sched_sizeof_kse(void) 691{ 692 return (sizeof(struct kse) + sizeof(struct ke_sched)); 693} 694int 695sched_sizeof_ksegrp(void) 696{ 697 return (sizeof(struct ksegrp)); 698} 699int 700sched_sizeof_proc(void) 701{ 702 return (sizeof(struct proc)); 703} 704int 705sched_sizeof_thread(void) 706{ 707 return (sizeof(struct thread)); 708} 709 710fixpt_t 711sched_pctcpu(struct kse *ke) 712{ 713 return (ke->ke_pctcpu); 714}
| 69 70struct ke_sched *kse0_sched = &ke_sched; 71struct kg_sched *ksegrp0_sched = NULL; 72struct p_sched *proc0_sched = NULL; 73struct td_sched *thread0_sched = NULL; 74 75static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 76#define SCHED_QUANTUM (hz / 10) /* Default sched quantum */ 77 78static struct callout schedcpu_callout; 79static struct callout roundrobin_callout; 80 81static void roundrobin(void *arg); 82static void schedcpu(void *arg); 83static void sched_setup(void *dummy); 84static void maybe_resched(struct thread *td); 85static void updatepri(struct ksegrp *kg); 86static void resetpriority(struct ksegrp *kg); 87 88SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 89 90/* 91 * Global run queue. 92 */ 93static struct runq runq; 94SYSINIT(runq, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, runq_init, &runq) 95 96static int 97sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 98{ 99 int error, new_val; 100 101 new_val = sched_quantum * tick; 102 error = sysctl_handle_int(oidp, &new_val, 0, req); 103 if (error != 0 || req->newptr == NULL) 104 return (error); 105 if (new_val < tick) 106 return (EINVAL); 107 sched_quantum = new_val / tick; 108 hogticks = 2 * sched_quantum; 109 return (0); 110} 111 112SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 113 0, sizeof sched_quantum, sysctl_kern_quantum, "I", 114 "Roundrobin scheduling quantum in microseconds"); 115 116/* 117 * Arrange to reschedule if necessary, taking the priorities and 118 * schedulers into account. 119 */ 120static void 121maybe_resched(struct thread *td) 122{ 123 124 mtx_assert(&sched_lock, MA_OWNED); 125 if (td->td_priority < curthread->td_priority && curthread->td_kse) 126 curthread->td_flags |= TDF_NEEDRESCHED; 127} 128 129/* 130 * Force switch among equal priority processes every 100ms. 131 * We don't actually need to force a context switch of the current process. 132 * The act of firing the event triggers a context switch to softclock() and 133 * then switching back out again which is equivalent to a preemption, thus 134 * no further work is needed on the local CPU. 135 */ 136/* ARGSUSED */ 137static void 138roundrobin(void *arg) 139{ 140 141#ifdef SMP 142 mtx_lock_spin(&sched_lock); 143 forward_roundrobin(); 144 mtx_unlock_spin(&sched_lock); 145#endif 146 147 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 148} 149 150/* 151 * Constants for digital decay and forget: 152 * 90% of (p_estcpu) usage in 5 * loadav time 153 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 154 * Note that, as ps(1) mentions, this can let percentages 155 * total over 100% (I've seen 137.9% for 3 processes). 156 * 157 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 158 * 159 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 160 * That is, the system wants to compute a value of decay such 161 * that the following for loop: 162 * for (i = 0; i < (5 * loadavg); i++) 163 * p_estcpu *= decay; 164 * will compute 165 * p_estcpu *= 0.1; 166 * for all values of loadavg: 167 * 168 * Mathematically this loop can be expressed by saying: 169 * decay ** (5 * loadavg) ~= .1 170 * 171 * The system computes decay as: 172 * decay = (2 * loadavg) / (2 * loadavg + 1) 173 * 174 * We wish to prove that the system's computation of decay 175 * will always fulfill the equation: 176 * decay ** (5 * loadavg) ~= .1 177 * 178 * If we compute b as: 179 * b = 2 * loadavg 180 * then 181 * decay = b / (b + 1) 182 * 183 * We now need to prove two things: 184 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 185 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 186 * 187 * Facts: 188 * For x close to zero, exp(x) =~ 1 + x, since 189 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 190 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 191 * For x close to zero, ln(1+x) =~ x, since 192 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 193 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 194 * ln(.1) =~ -2.30 195 * 196 * Proof of (1): 197 * Solve (factor)**(power) =~ .1 given power (5*loadav): 198 * solving for factor, 199 * ln(factor) =~ (-2.30/5*loadav), or 200 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 201 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 202 * 203 * Proof of (2): 204 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 205 * solving for power, 206 * power*ln(b/(b+1)) =~ -2.30, or 207 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 208 * 209 * Actual power values for the implemented algorithm are as follows: 210 * loadav: 1 2 3 4 211 * power: 5.68 10.32 14.94 19.55 212 */ 213 214/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 215#define loadfactor(loadav) (2 * (loadav)) 216#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 217 218/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 219static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 220SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 221 222/* 223 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 224 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 225 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 226 * 227 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 228 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 229 * 230 * If you don't want to bother with the faster/more-accurate formula, you 231 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 232 * (more general) method of calculating the %age of CPU used by a process. 233 */ 234#define CCPU_SHIFT 11 235 236/* 237 * Recompute process priorities, every hz ticks. 238 * MP-safe, called without the Giant mutex. 239 */ 240/* ARGSUSED */ 241static void 242schedcpu(void *arg) 243{ 244 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 245 struct thread *td; 246 struct proc *p; 247 struct kse *ke; 248 struct ksegrp *kg; 249 int realstathz; 250 int awake; 251 252 realstathz = stathz ? stathz : hz; 253 sx_slock(&allproc_lock); 254 FOREACH_PROC_IN_SYSTEM(p) { 255 mtx_lock_spin(&sched_lock); 256 p->p_swtime++; 257 FOREACH_KSEGRP_IN_PROC(p, kg) { 258 awake = 0; 259 FOREACH_KSE_IN_GROUP(kg, ke) { 260 /* 261 * Increment time in/out of memory and sleep 262 * time (if sleeping). We ignore overflow; 263 * with 16-bit int's (remember them?) 264 * overflow takes 45 days. 265 */ 266 /* 267 * The kse slptimes are not touched in wakeup 268 * because the thread may not HAVE a KSE. 269 */ 270 if (ke->ke_state == KES_ONRUNQ) { 271 awake = 1; 272 ke->ke_flags &= ~KEF_DIDRUN; 273 } else if ((ke->ke_state == KES_THREAD) && 274 (TD_IS_RUNNING(ke->ke_thread))) { 275 awake = 1; 276 /* Do not clear KEF_DIDRUN */ 277 } else if (ke->ke_flags & KEF_DIDRUN) { 278 awake = 1; 279 ke->ke_flags &= ~KEF_DIDRUN; 280 } 281 282 /* 283 * pctcpu is only for ps? 284 * Do it per kse.. and add them up at the end? 285 * XXXKSE 286 */ 287 ke->ke_pctcpu 288 = (ke->ke_pctcpu * ccpu) >> 289 FSHIFT; 290 /* 291 * If the kse has been idle the entire second, 292 * stop recalculating its priority until 293 * it wakes up. 294 */ 295 if (ke->ke_sched->ske_cpticks == 0) 296 continue; 297#if (FSHIFT >= CCPU_SHIFT) 298 ke->ke_pctcpu += (realstathz == 100) 299 ? ((fixpt_t) ke->ke_sched->ske_cpticks) << 300 (FSHIFT - CCPU_SHIFT) : 301 100 * (((fixpt_t) ke->ke_sched->ske_cpticks) 302 << (FSHIFT - CCPU_SHIFT)) / realstathz; 303#else 304 ke->ke_pctcpu += ((FSCALE - ccpu) * 305 (ke->ke_sched->ske_cpticks * 306 FSCALE / realstathz)) >> FSHIFT; 307#endif 308 ke->ke_sched->ske_cpticks = 0; 309 } /* end of kse loop */ 310 /* 311 * If there are ANY running threads in this KSEGRP, 312 * then don't count it as sleeping. 313 */ 314 if (awake) { 315 if (kg->kg_slptime > 1) { 316 /* 317 * In an ideal world, this should not 318 * happen, because whoever woke us 319 * up from the long sleep should have 320 * unwound the slptime and reset our 321 * priority before we run at the stale 322 * priority. Should KASSERT at some 323 * point when all the cases are fixed. 324 */ 325 updatepri(kg); 326 } 327 kg->kg_slptime = 0; 328 } else { 329 kg->kg_slptime++; 330 } 331 if (kg->kg_slptime > 1) 332 continue; 333 kg->kg_estcpu = decay_cpu(loadfac, kg->kg_estcpu); 334 resetpriority(kg); 335 FOREACH_THREAD_IN_GROUP(kg, td) { 336 if (td->td_priority >= PUSER) { 337 sched_prio(td, kg->kg_user_pri); 338 } 339 } 340 } /* end of ksegrp loop */ 341 mtx_unlock_spin(&sched_lock); 342 } /* end of process loop */ 343 sx_sunlock(&allproc_lock); 344 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 345} 346 347/* 348 * Recalculate the priority of a process after it has slept for a while. 349 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 350 * least six times the loadfactor will decay p_estcpu to zero. 351 */ 352static void 353updatepri(struct ksegrp *kg) 354{ 355 register unsigned int newcpu; 356 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 357 358 newcpu = kg->kg_estcpu; 359 if (kg->kg_slptime > 5 * loadfac) 360 kg->kg_estcpu = 0; 361 else { 362 kg->kg_slptime--; /* the first time was done in schedcpu */ 363 while (newcpu && --kg->kg_slptime) 364 newcpu = decay_cpu(loadfac, newcpu); 365 kg->kg_estcpu = newcpu; 366 } 367 resetpriority(kg); 368} 369 370/* 371 * Compute the priority of a process when running in user mode. 372 * Arrange to reschedule if the resulting priority is better 373 * than that of the current process. 374 */ 375static void 376resetpriority(struct ksegrp *kg) 377{ 378 register unsigned int newpriority; 379 struct thread *td; 380 381 if (kg->kg_pri_class == PRI_TIMESHARE) { 382 newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT + 383 NICE_WEIGHT * (kg->kg_nice - PRIO_MIN); 384 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 385 PRI_MAX_TIMESHARE); 386 kg->kg_user_pri = newpriority; 387 } 388 FOREACH_THREAD_IN_GROUP(kg, td) { 389 maybe_resched(td); /* XXXKSE silly */ 390 } 391} 392 393/* ARGSUSED */ 394static void 395sched_setup(void *dummy) 396{ 397 if (sched_quantum == 0) 398 sched_quantum = SCHED_QUANTUM; 399 hogticks = 2 * sched_quantum; 400 401 callout_init(&schedcpu_callout, 1); 402 callout_init(&roundrobin_callout, 0); 403 404 /* Kick off timeout driven events by calling first time. */ 405 roundrobin(NULL); 406 schedcpu(NULL); 407} 408 409/* External interfaces start here */ 410int 411sched_runnable(void) 412{ 413 return runq_check(&runq); 414} 415 416int 417sched_rr_interval(void) 418{ 419 if (sched_quantum == 0) 420 sched_quantum = SCHED_QUANTUM; 421 return (sched_quantum); 422} 423 424/* 425 * We adjust the priority of the current process. The priority of 426 * a process gets worse as it accumulates CPU time. The cpu usage 427 * estimator (p_estcpu) is increased here. resetpriority() will 428 * compute a different priority each time p_estcpu increases by 429 * INVERSE_ESTCPU_WEIGHT 430 * (until MAXPRI is reached). The cpu usage estimator ramps up 431 * quite quickly when the process is running (linearly), and decays 432 * away exponentially, at a rate which is proportionally slower when 433 * the system is busy. The basic principle is that the system will 434 * 90% forget that the process used a lot of CPU time in 5 * loadav 435 * seconds. This causes the system to favor processes which haven't 436 * run much recently, and to round-robin among other processes. 437 */ 438void 439sched_clock(struct kse *ke) 440{ 441 struct ksegrp *kg; 442 struct thread *td; 443 444 mtx_assert(&sched_lock, MA_OWNED); 445 kg = ke->ke_ksegrp; 446 td = ke->ke_thread; 447 448 ke->ke_sched->ske_cpticks++; 449 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1); 450 if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 451 resetpriority(kg); 452 if (td->td_priority >= PUSER) 453 td->td_priority = kg->kg_user_pri; 454 } 455} 456/* 457 * charge childs scheduling cpu usage to parent. 458 * 459 * XXXKSE assume only one thread & kse & ksegrp keep estcpu in each ksegrp. 460 * Charge it to the ksegrp that did the wait since process estcpu is sum of 461 * all ksegrps, this is strictly as expected. Assume that the child process 462 * aggregated all the estcpu into the 'built-in' ksegrp. 463 */ 464void 465sched_exit(struct proc *p, struct proc *p1) 466{ 467 sched_exit_kse(FIRST_KSE_IN_PROC(p), FIRST_KSE_IN_PROC(p1)); 468 sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), FIRST_KSEGRP_IN_PROC(p1)); 469 sched_exit_thread(FIRST_THREAD_IN_PROC(p), FIRST_THREAD_IN_PROC(p1)); 470} 471 472void 473sched_exit_kse(struct kse *ke, struct kse *child) 474{ 475} 476 477void 478sched_exit_ksegrp(struct ksegrp *kg, struct ksegrp *child) 479{ 480 481 mtx_assert(&sched_lock, MA_OWNED); 482 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + child->kg_estcpu); 483} 484 485void 486sched_exit_thread(struct thread *td, struct thread *child) 487{ 488} 489 490void 491sched_fork(struct proc *p, struct proc *p1) 492{ 493 sched_fork_kse(FIRST_KSE_IN_PROC(p), FIRST_KSE_IN_PROC(p1)); 494 sched_fork_ksegrp(FIRST_KSEGRP_IN_PROC(p), FIRST_KSEGRP_IN_PROC(p1)); 495 sched_fork_thread(FIRST_THREAD_IN_PROC(p), FIRST_THREAD_IN_PROC(p1)); 496} 497 498void 499sched_fork_kse(struct kse *ke, struct kse *child) 500{ 501 child->ke_sched->ske_cpticks = 0; 502} 503 504void 505sched_fork_ksegrp(struct ksegrp *kg, struct ksegrp *child) 506{ 507 mtx_assert(&sched_lock, MA_OWNED); 508 child->kg_estcpu = kg->kg_estcpu; 509} 510 511void 512sched_fork_thread(struct thread *td, struct thread *child) 513{ 514} 515 516void 517sched_nice(struct ksegrp *kg, int nice) 518{ 519 520 PROC_LOCK_ASSERT(kg->kg_proc, MA_OWNED); 521 mtx_assert(&sched_lock, MA_OWNED); 522 kg->kg_nice = nice; 523 resetpriority(kg); 524} 525 526void 527sched_class(struct ksegrp *kg, int class) 528{ 529 mtx_assert(&sched_lock, MA_OWNED); 530 kg->kg_pri_class = class; 531} 532 533/* 534 * Adjust the priority of a thread. 535 * This may include moving the thread within the KSEGRP, 536 * changing the assignment of a kse to the thread, 537 * and moving a KSE in the system run queue. 538 */ 539void 540sched_prio(struct thread *td, u_char prio) 541{ 542 543 mtx_assert(&sched_lock, MA_OWNED); 544 if (TD_ON_RUNQ(td)) { 545 adjustrunqueue(td, prio); 546 } else { 547 td->td_priority = prio; 548 } 549} 550 551void 552sched_sleep(struct thread *td, u_char prio) 553{ 554 555 mtx_assert(&sched_lock, MA_OWNED); 556 td->td_ksegrp->kg_slptime = 0; 557 td->td_priority = prio; 558} 559 560void 561sched_switchin(struct thread *td) 562{ 563 564 mtx_assert(&sched_lock, MA_OWNED); 565 td->td_oncpu = PCPU_GET(cpuid); 566} 567 568void 569sched_switchout(struct thread *td) 570{ 571 struct kse *ke; 572 struct proc *p; 573 574 ke = td->td_kse; 575 p = td->td_proc; 576 577 mtx_assert(&sched_lock, MA_OWNED); 578 KASSERT((ke->ke_state == KES_THREAD), ("mi_switch: kse state?")); 579 580 td->td_lastcpu = td->td_oncpu; 581 td->td_last_kse = ke; 582 td->td_oncpu = NOCPU; 583 td->td_flags &= ~TDF_NEEDRESCHED; 584 /* 585 * At the last moment, if this thread is still marked RUNNING, 586 * then put it back on the run queue as it has not been suspended 587 * or stopped or any thing else similar. 588 */ 589 if (TD_IS_RUNNING(td)) { 590 /* Put us back on the run queue (kse and all). */ 591 setrunqueue(td); 592 } else if (p->p_flag & P_THREADED) { 593 /* 594 * We will not be on the run queue. So we must be 595 * sleeping or similar. As it's available, 596 * someone else can use the KSE if they need it. 597 */ 598 kse_reassign(ke); 599 } 600} 601 602void 603sched_wakeup(struct thread *td) 604{ 605 struct ksegrp *kg; 606 607 mtx_assert(&sched_lock, MA_OWNED); 608 kg = td->td_ksegrp; 609 if (kg->kg_slptime > 1) 610 updatepri(kg); 611 kg->kg_slptime = 0; 612 setrunqueue(td); 613 maybe_resched(td); 614} 615 616void 617sched_add(struct kse *ke) 618{ 619 mtx_assert(&sched_lock, MA_OWNED); 620 KASSERT((ke->ke_thread != NULL), ("runq_add: No thread on KSE")); 621 KASSERT((ke->ke_thread->td_kse != NULL), 622 ("runq_add: No KSE on thread")); 623 KASSERT(ke->ke_state != KES_ONRUNQ, 624 ("runq_add: kse %p (%s) already in run queue", ke, 625 ke->ke_proc->p_comm)); 626 KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 627 ("runq_add: process swapped out")); 628 ke->ke_ksegrp->kg_runq_kses++; 629 ke->ke_state = KES_ONRUNQ; 630 631 runq_add(&runq, ke); 632} 633 634void 635sched_rem(struct kse *ke) 636{ 637 KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 638 ("runq_remove: process swapped out")); 639 KASSERT((ke->ke_state == KES_ONRUNQ), ("KSE not on run queue")); 640 mtx_assert(&sched_lock, MA_OWNED); 641 642 runq_remove(&runq, ke); 643 ke->ke_state = KES_THREAD; 644 ke->ke_ksegrp->kg_runq_kses--; 645} 646 647struct kse * 648sched_choose(void) 649{ 650 struct kse *ke; 651 652 ke = runq_choose(&runq); 653 654 if (ke != NULL) { 655 runq_remove(&runq, ke); 656 ke->ke_state = KES_THREAD; 657 658 KASSERT((ke->ke_thread != NULL), 659 ("runq_choose: No thread on KSE")); 660 KASSERT((ke->ke_thread->td_kse != NULL), 661 ("runq_choose: No KSE on thread")); 662 KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 663 ("runq_choose: process swapped out")); 664 } 665 return (ke); 666} 667 668void 669sched_userret(struct thread *td) 670{ 671 struct ksegrp *kg; 672 /* 673 * XXX we cheat slightly on the locking here to avoid locking in 674 * the usual case. Setting td_priority here is essentially an 675 * incomplete workaround for not setting it properly elsewhere. 676 * Now that some interrupt handlers are threads, not setting it 677 * properly elsewhere can clobber it in the window between setting 678 * it here and returning to user mode, so don't waste time setting 679 * it perfectly here. 680 */ 681 kg = td->td_ksegrp; 682 if (td->td_priority != kg->kg_user_pri) { 683 mtx_lock_spin(&sched_lock); 684 td->td_priority = kg->kg_user_pri; 685 mtx_unlock_spin(&sched_lock); 686 } 687} 688 689int 690sched_sizeof_kse(void) 691{ 692 return (sizeof(struct kse) + sizeof(struct ke_sched)); 693} 694int 695sched_sizeof_ksegrp(void) 696{ 697 return (sizeof(struct ksegrp)); 698} 699int 700sched_sizeof_proc(void) 701{ 702 return (sizeof(struct proc)); 703} 704int 705sched_sizeof_thread(void) 706{ 707 return (sizeof(struct thread)); 708} 709 710fixpt_t 711sched_pctcpu(struct kse *ke) 712{ 713 return (ke->ke_pctcpu); 714}
|