1 /*- 2 * Copyright (c) 1982, 1986, 1990, 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 4. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 #include <sys/cdefs.h> 36 __FBSDID("$FreeBSD$"); 37 38 #define kse td_sched 39 40 #include <sys/param.h> 41 #include <sys/systm.h> 42 #include <sys/kernel.h> 43 #include <sys/ktr.h> 44 #include <sys/lock.h> 45 #include <sys/kthread.h> 46 #include <sys/mutex.h> 47 #include <sys/proc.h> 48 #include <sys/resourcevar.h> 49 #include <sys/sched.h> 50 #include <sys/smp.h> 51 #include <sys/sysctl.h> 52 #include <sys/sx.h> 53 #include <machine/smp.h> 54 55 /* 56 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in 57 * the range 100-256 Hz (approximately). 58 */ 59 #define ESTCPULIM(e) \ 60 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \ 61 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1) 62 #ifdef SMP 63 #define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus) 64 #else 65 #define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */ 66 #endif 67 #define NICE_WEIGHT 1 /* Priorities per nice level. */ 68 69 /* 70 * The schedulable entity that can be given a context to run. 71 * A process may have several of these. Probably one per processor 72 * but posibly a few more. In this universe they are grouped 73 * with a KSEG that contains the priority and niceness 74 * for the group. 75 */ 76 struct kse { 77 TAILQ_ENTRY(kse) ke_kglist; /* (*) Queue of KSEs in ke_ksegrp. */ 78 TAILQ_ENTRY(kse) ke_kgrlist; /* (*) Queue of KSEs in this state. */ 79 TAILQ_ENTRY(kse) ke_procq; /* (j/z) Run queue. */ 80 struct thread *ke_thread; /* (*) Active associated thread. */ 81 fixpt_t ke_pctcpu; /* (j) %cpu during p_swtime. */ 82 u_char ke_oncpu; /* (j) Which cpu we are on. */ 83 char ke_rqindex; /* (j) Run queue index. */ 84 enum { 85 KES_THREAD = 0x0, /* slaved to thread state */ 86 KES_ONRUNQ 87 } ke_state; /* (j) KSE status. */ 88 int ke_cpticks; /* (j) Ticks of cpu time. */ 89 struct runq *ke_runq; /* runq the kse is currently on */ 90 }; 91 92 #define ke_proc ke_thread->td_proc 93 #define ke_ksegrp ke_thread->td_ksegrp 94 95 #define td_kse td_sched 96 97 /* flags kept in td_flags */ 98 #define TDF_DIDRUN TDF_SCHED0 /* KSE actually ran. */ 99 #define TDF_EXIT TDF_SCHED1 /* KSE is being killed. */ 100 #define TDF_BOUND TDF_SCHED2 101 102 #define ke_flags ke_thread->td_flags 103 #define KEF_DIDRUN TDF_DIDRUN /* KSE actually ran. */ 104 #define KEF_EXIT TDF_EXIT /* KSE is being killed. */ 105 #define KEF_BOUND TDF_BOUND /* stuck to one CPU */ 106 107 #define SKE_RUNQ_PCPU(ke) \ 108 ((ke)->ke_runq != 0 && (ke)->ke_runq != &runq) 109 110 struct kg_sched { 111 struct thread *skg_last_assigned; /* (j) Last thread assigned to */ 112 /* the system scheduler. */ 113 int skg_avail_opennings; /* (j) Num KSEs requested in group. */ 114 int skg_concurrency; /* (j) Num KSEs requested in group. */ 115 int skg_runq_kses; /* (j) Num KSEs on runq. */ 116 }; 117 #define kg_last_assigned kg_sched->skg_last_assigned 118 #define kg_avail_opennings kg_sched->skg_avail_opennings 119 #define kg_concurrency kg_sched->skg_concurrency 120 #define kg_runq_kses kg_sched->skg_runq_kses 121 122 /* 123 * KSE_CAN_MIGRATE macro returns true if the kse can migrate between 124 * cpus. 125 */ 126 #define KSE_CAN_MIGRATE(ke) \ 127 ((ke)->ke_thread->td_pinned == 0 && ((ke)->ke_flags & KEF_BOUND) == 0) 128 129 static struct kse kse0; 130 static struct kg_sched kg_sched0; 131 132 static int sched_tdcnt; /* Total runnable threads in the system. */ 133 static int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 134 #define SCHED_QUANTUM (hz / 10) /* Default sched quantum */ 135 136 static struct callout roundrobin_callout; 137 138 static void slot_fill(struct ksegrp *kg); 139 static struct kse *sched_choose(void); /* XXX Should be thread * */ 140 141 static void setup_runqs(void); 142 static void roundrobin(void *arg); 143 static void schedcpu(void); 144 static void schedcpu_thread(void); 145 static void sched_setup(void *dummy); 146 static void maybe_resched(struct thread *td); 147 static void updatepri(struct ksegrp *kg); 148 static void resetpriority(struct ksegrp *kg); 149 #ifdef SMP 150 static int forward_wakeup(int cpunum); 151 #endif 152 153 static struct kproc_desc sched_kp = { 154 "schedcpu", 155 schedcpu_thread, 156 NULL 157 }; 158 SYSINIT(schedcpu, SI_SUB_RUN_SCHEDULER, SI_ORDER_FIRST, kproc_start, &sched_kp) 159 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL) 160 161 /* 162 * Global run queue. 163 */ 164 static struct runq runq; 165 166 #ifdef SMP 167 /* 168 * Per-CPU run queues 169 */ 170 static struct runq runq_pcpu[MAXCPU]; 171 #endif 172 173 static void 174 setup_runqs(void) 175 { 176 #ifdef SMP 177 int i; 178 179 for (i = 0; i < MAXCPU; ++i) 180 runq_init(&runq_pcpu[i]); 181 #endif 182 183 runq_init(&runq); 184 } 185 186 static int 187 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 188 { 189 int error, new_val; 190 191 new_val = sched_quantum * tick; 192 error = sysctl_handle_int(oidp, &new_val, 0, req); 193 if (error != 0 || req->newptr == NULL) 194 return (error); 195 if (new_val < tick) 196 return (EINVAL); 197 sched_quantum = new_val / tick; 198 hogticks = 2 * sched_quantum; 199 return (0); 200 } 201 202 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD, 0, "Scheduler"); 203 204 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0, 205 "Scheduler name"); 206 207 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW, 208 0, sizeof sched_quantum, sysctl_kern_quantum, "I", 209 "Roundrobin scheduling quantum in microseconds"); 210 211 #ifdef SMP 212 /* Enable forwarding of wakeups to all other cpus */ 213 SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD, NULL, "Kernel SMP"); 214 215 static int forward_wakeup_enabled = 1; 216 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW, 217 &forward_wakeup_enabled, 0, 218 "Forwarding of wakeup to idle CPUs"); 219 220 static int forward_wakeups_requested = 0; 221 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD, 222 &forward_wakeups_requested, 0, 223 "Requests for Forwarding of wakeup to idle CPUs"); 224 225 static int forward_wakeups_delivered = 0; 226 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD, 227 &forward_wakeups_delivered, 0, 228 "Completed Forwarding of wakeup to idle CPUs"); 229 230 static int forward_wakeup_use_mask = 1; 231 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW, 232 &forward_wakeup_use_mask, 0, 233 "Use the mask of idle cpus"); 234 235 static int forward_wakeup_use_loop = 0; 236 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW, 237 &forward_wakeup_use_loop, 0, 238 "Use a loop to find idle cpus"); 239 240 static int forward_wakeup_use_single = 0; 241 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, onecpu, CTLFLAG_RW, 242 &forward_wakeup_use_single, 0, 243 "Only signal one idle cpu"); 244 245 static int forward_wakeup_use_htt = 0; 246 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, htt2, CTLFLAG_RW, 247 &forward_wakeup_use_htt, 0, 248 "account for htt"); 249 250 #endif 251 static int sched_followon = 0; 252 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW, 253 &sched_followon, 0, 254 "allow threads to share a quantum"); 255 256 static int sched_pfollowons = 0; 257 SYSCTL_INT(_kern_sched, OID_AUTO, pfollowons, CTLFLAG_RD, 258 &sched_pfollowons, 0, 259 "number of followons done to a different ksegrp"); 260 261 static int sched_kgfollowons = 0; 262 SYSCTL_INT(_kern_sched, OID_AUTO, kgfollowons, CTLFLAG_RD, 263 &sched_kgfollowons, 0, 264 "number of followons done in a ksegrp"); 265 266 /* 267 * Arrange to reschedule if necessary, taking the priorities and 268 * schedulers into account. 269 */ 270 static void 271 maybe_resched(struct thread *td) 272 { 273 274 mtx_assert(&sched_lock, MA_OWNED); 275 if (td->td_priority < curthread->td_priority) 276 curthread->td_flags |= TDF_NEEDRESCHED; 277 } 278 279 /* 280 * Force switch among equal priority processes every 100ms. 281 * We don't actually need to force a context switch of the current process. 282 * The act of firing the event triggers a context switch to softclock() and 283 * then switching back out again which is equivalent to a preemption, thus 284 * no further work is needed on the local CPU. 285 */ 286 /* ARGSUSED */ 287 static void 288 roundrobin(void *arg) 289 { 290 291 #ifdef SMP 292 mtx_lock_spin(&sched_lock); 293 forward_roundrobin(); 294 mtx_unlock_spin(&sched_lock); 295 #endif 296 297 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 298 } 299 300 /* 301 * Constants for digital decay and forget: 302 * 90% of (kg_estcpu) usage in 5 * loadav time 303 * 95% of (ke_pctcpu) usage in 60 seconds (load insensitive) 304 * Note that, as ps(1) mentions, this can let percentages 305 * total over 100% (I've seen 137.9% for 3 processes). 306 * 307 * Note that schedclock() updates kg_estcpu and p_cpticks asynchronously. 308 * 309 * We wish to decay away 90% of kg_estcpu in (5 * loadavg) seconds. 310 * That is, the system wants to compute a value of decay such 311 * that the following for loop: 312 * for (i = 0; i < (5 * loadavg); i++) 313 * kg_estcpu *= decay; 314 * will compute 315 * kg_estcpu *= 0.1; 316 * for all values of loadavg: 317 * 318 * Mathematically this loop can be expressed by saying: 319 * decay ** (5 * loadavg) ~= .1 320 * 321 * The system computes decay as: 322 * decay = (2 * loadavg) / (2 * loadavg + 1) 323 * 324 * We wish to prove that the system's computation of decay 325 * will always fulfill the equation: 326 * decay ** (5 * loadavg) ~= .1 327 * 328 * If we compute b as: 329 * b = 2 * loadavg 330 * then 331 * decay = b / (b + 1) 332 * 333 * We now need to prove two things: 334 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 335 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 336 * 337 * Facts: 338 * For x close to zero, exp(x) =~ 1 + x, since 339 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 340 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 341 * For x close to zero, ln(1+x) =~ x, since 342 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 343 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 344 * ln(.1) =~ -2.30 345 * 346 * Proof of (1): 347 * Solve (factor)**(power) =~ .1 given power (5*loadav): 348 * solving for factor, 349 * ln(factor) =~ (-2.30/5*loadav), or 350 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 351 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 352 * 353 * Proof of (2): 354 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 355 * solving for power, 356 * power*ln(b/(b+1)) =~ -2.30, or 357 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 358 * 359 * Actual power values for the implemented algorithm are as follows: 360 * loadav: 1 2 3 4 361 * power: 5.68 10.32 14.94 19.55 362 */ 363 364 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 365 #define loadfactor(loadav) (2 * (loadav)) 366 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 367 368 /* decay 95% of `ke_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 369 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 370 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 371 372 /* 373 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 374 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 375 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 376 * 377 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 378 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 379 * 380 * If you don't want to bother with the faster/more-accurate formula, you 381 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 382 * (more general) method of calculating the %age of CPU used by a process. 383 */ 384 #define CCPU_SHIFT 11 385 386 /* 387 * Recompute process priorities, every hz ticks. 388 * MP-safe, called without the Giant mutex. 389 */ 390 /* ARGSUSED */ 391 static void 392 schedcpu(void) 393 { 394 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 395 struct thread *td; 396 struct proc *p; 397 struct kse *ke; 398 struct ksegrp *kg; 399 int awake, realstathz; 400 401 realstathz = stathz ? stathz : hz; 402 sx_slock(&allproc_lock); 403 FOREACH_PROC_IN_SYSTEM(p) { 404 /* 405 * Prevent state changes and protect run queue. 406 */ 407 mtx_lock_spin(&sched_lock); 408 /* 409 * Increment time in/out of memory. We ignore overflow; with 410 * 16-bit int's (remember them?) overflow takes 45 days. 411 */ 412 p->p_swtime++; 413 FOREACH_KSEGRP_IN_PROC(p, kg) { 414 awake = 0; 415 FOREACH_THREAD_IN_GROUP(kg, td) { 416 ke = td->td_kse; 417 /* 418 * Increment sleep time (if sleeping). We 419 * ignore overflow, as above. 420 */ 421 /* 422 * The kse slptimes are not touched in wakeup 423 * because the thread may not HAVE a KSE. 424 */ 425 if (ke->ke_state == KES_ONRUNQ) { 426 awake = 1; 427 ke->ke_flags &= ~KEF_DIDRUN; 428 } else if ((ke->ke_state == KES_THREAD) && 429 (TD_IS_RUNNING(td))) { 430 awake = 1; 431 /* Do not clear KEF_DIDRUN */ 432 } else if (ke->ke_flags & KEF_DIDRUN) { 433 awake = 1; 434 ke->ke_flags &= ~KEF_DIDRUN; 435 } 436 437 /* 438 * ke_pctcpu is only for ps and ttyinfo(). 439 * Do it per kse, and add them up at the end? 440 * XXXKSE 441 */ 442 ke->ke_pctcpu = (ke->ke_pctcpu * ccpu) >> 443 FSHIFT; 444 /* 445 * If the kse has been idle the entire second, 446 * stop recalculating its priority until 447 * it wakes up. 448 */ 449 if (ke->ke_cpticks == 0) 450 continue; 451 #if (FSHIFT >= CCPU_SHIFT) 452 ke->ke_pctcpu += (realstathz == 100) 453 ? ((fixpt_t) ke->ke_cpticks) << 454 (FSHIFT - CCPU_SHIFT) : 455 100 * (((fixpt_t) ke->ke_cpticks) 456 << (FSHIFT - CCPU_SHIFT)) / realstathz; 457 #else 458 ke->ke_pctcpu += ((FSCALE - ccpu) * 459 (ke->ke_cpticks * 460 FSCALE / realstathz)) >> FSHIFT; 461 #endif 462 ke->ke_cpticks = 0; 463 } /* end of kse loop */ 464 /* 465 * If there are ANY running threads in this KSEGRP, 466 * then don't count it as sleeping. 467 */ 468 if (awake) { 469 if (kg->kg_slptime > 1) { 470 /* 471 * In an ideal world, this should not 472 * happen, because whoever woke us 473 * up from the long sleep should have 474 * unwound the slptime and reset our 475 * priority before we run at the stale 476 * priority. Should KASSERT at some 477 * point when all the cases are fixed. 478 */ 479 updatepri(kg); 480 } 481 kg->kg_slptime = 0; 482 } else 483 kg->kg_slptime++; 484 if (kg->kg_slptime > 1) 485 continue; 486 kg->kg_estcpu = decay_cpu(loadfac, kg->kg_estcpu); 487 resetpriority(kg); 488 FOREACH_THREAD_IN_GROUP(kg, td) { 489 if (td->td_priority >= PUSER) { 490 sched_prio(td, kg->kg_user_pri); 491 } 492 } 493 } /* end of ksegrp loop */ 494 mtx_unlock_spin(&sched_lock); 495 } /* end of process loop */ 496 sx_sunlock(&allproc_lock); 497 } 498 499 /* 500 * Main loop for a kthread that executes schedcpu once a second. 501 */ 502 static void 503 schedcpu_thread(void) 504 { 505 int nowake; 506 507 for (;;) { 508 schedcpu(); 509 tsleep(&nowake, curthread->td_priority, "-", hz); 510 } 511 } 512 513 /* 514 * Recalculate the priority of a process after it has slept for a while. 515 * For all load averages >= 1 and max kg_estcpu of 255, sleeping for at 516 * least six times the loadfactor will decay kg_estcpu to zero. 517 */ 518 static void 519 updatepri(struct ksegrp *kg) 520 { 521 register fixpt_t loadfac; 522 register unsigned int newcpu; 523 524 loadfac = loadfactor(averunnable.ldavg[0]); 525 if (kg->kg_slptime > 5 * loadfac) 526 kg->kg_estcpu = 0; 527 else { 528 newcpu = kg->kg_estcpu; 529 kg->kg_slptime--; /* was incremented in schedcpu() */ 530 while (newcpu && --kg->kg_slptime) 531 newcpu = decay_cpu(loadfac, newcpu); 532 kg->kg_estcpu = newcpu; 533 } 534 resetpriority(kg); 535 } 536 537 /* 538 * Compute the priority of a process when running in user mode. 539 * Arrange to reschedule if the resulting priority is better 540 * than that of the current process. 541 */ 542 static void 543 resetpriority(struct ksegrp *kg) 544 { 545 register unsigned int newpriority; 546 struct thread *td; 547 548 if (kg->kg_pri_class == PRI_TIMESHARE) { 549 newpriority = PUSER + kg->kg_estcpu / INVERSE_ESTCPU_WEIGHT + 550 NICE_WEIGHT * (kg->kg_proc->p_nice - PRIO_MIN); 551 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 552 PRI_MAX_TIMESHARE); 553 kg->kg_user_pri = newpriority; 554 } 555 FOREACH_THREAD_IN_GROUP(kg, td) { 556 maybe_resched(td); /* XXXKSE silly */ 557 } 558 } 559 560 /* ARGSUSED */ 561 static void 562 sched_setup(void *dummy) 563 { 564 setup_runqs(); 565 566 if (sched_quantum == 0) 567 sched_quantum = SCHED_QUANTUM; 568 hogticks = 2 * sched_quantum; 569 570 callout_init(&roundrobin_callout, CALLOUT_MPSAFE); 571 572 /* Kick off timeout driven events by calling first time. */ 573 roundrobin(NULL); 574 575 /* Account for thread0. */ 576 sched_tdcnt++; 577 } 578 579 /* External interfaces start here */ 580 /* 581 * Very early in the boot some setup of scheduler-specific 582 * parts of proc0 and of soem scheduler resources needs to be done. 583 * Called from: 584 * proc0_init() 585 */ 586 void 587 schedinit(void) 588 { 589 /* 590 * Set up the scheduler specific parts of proc0. 591 */ 592 proc0.p_sched = NULL; /* XXX */ 593 ksegrp0.kg_sched = &kg_sched0; 594 thread0.td_sched = &kse0; 595 kse0.ke_thread = &thread0; 596 kse0.ke_oncpu = NOCPU; /* wrong.. can we use PCPU(cpuid) yet? */ 597 kse0.ke_state = KES_THREAD; 598 kg_sched0.skg_concurrency = 1; 599 kg_sched0.skg_avail_opennings = 0; /* we are already running */ 600 } 601 602 int 603 sched_runnable(void) 604 { 605 #ifdef SMP 606 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]); 607 #else 608 return runq_check(&runq); 609 #endif 610 } 611 612 int 613 sched_rr_interval(void) 614 { 615 if (sched_quantum == 0) 616 sched_quantum = SCHED_QUANTUM; 617 return (sched_quantum); 618 } 619 620 /* 621 * We adjust the priority of the current process. The priority of 622 * a process gets worse as it accumulates CPU time. The cpu usage 623 * estimator (kg_estcpu) is increased here. resetpriority() will 624 * compute a different priority each time kg_estcpu increases by 625 * INVERSE_ESTCPU_WEIGHT 626 * (until MAXPRI is reached). The cpu usage estimator ramps up 627 * quite quickly when the process is running (linearly), and decays 628 * away exponentially, at a rate which is proportionally slower when 629 * the system is busy. The basic principle is that the system will 630 * 90% forget that the process used a lot of CPU time in 5 * loadav 631 * seconds. This causes the system to favor processes which haven't 632 * run much recently, and to round-robin among other processes. 633 */ 634 void 635 sched_clock(struct thread *td) 636 { 637 struct ksegrp *kg; 638 struct kse *ke; 639 640 mtx_assert(&sched_lock, MA_OWNED); 641 kg = td->td_ksegrp; 642 ke = td->td_kse; 643 644 ke->ke_cpticks++; 645 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + 1); 646 if ((kg->kg_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 647 resetpriority(kg); 648 if (td->td_priority >= PUSER) 649 td->td_priority = kg->kg_user_pri; 650 } 651 } 652 653 /* 654 * charge childs scheduling cpu usage to parent. 655 * 656 * XXXKSE assume only one thread & kse & ksegrp keep estcpu in each ksegrp. 657 * Charge it to the ksegrp that did the wait since process estcpu is sum of 658 * all ksegrps, this is strictly as expected. Assume that the child process 659 * aggregated all the estcpu into the 'built-in' ksegrp. 660 */ 661 void 662 sched_exit(struct proc *p, struct thread *td) 663 { 664 sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), td); 665 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td); 666 } 667 668 void 669 sched_exit_ksegrp(struct ksegrp *kg, struct thread *childtd) 670 { 671 672 mtx_assert(&sched_lock, MA_OWNED); 673 kg->kg_estcpu = ESTCPULIM(kg->kg_estcpu + childtd->td_ksegrp->kg_estcpu); 674 } 675 676 void 677 sched_exit_thread(struct thread *td, struct thread *child) 678 { 679 if ((child->td_proc->p_flag & P_NOLOAD) == 0) 680 sched_tdcnt--; 681 } 682 683 void 684 sched_fork(struct thread *td, struct thread *childtd) 685 { 686 sched_fork_ksegrp(td, childtd->td_ksegrp); 687 sched_fork_thread(td, childtd); 688 } 689 690 void 691 sched_fork_ksegrp(struct thread *td, struct ksegrp *child) 692 { 693 mtx_assert(&sched_lock, MA_OWNED); 694 child->kg_estcpu = td->td_ksegrp->kg_estcpu; 695 } 696 697 void 698 sched_fork_thread(struct thread *td, struct thread *childtd) 699 { 700 sched_newthread(childtd); 701 } 702 703 void 704 sched_nice(struct proc *p, int nice) 705 { 706 struct ksegrp *kg; 707 708 PROC_LOCK_ASSERT(p, MA_OWNED); 709 mtx_assert(&sched_lock, MA_OWNED); 710 p->p_nice = nice; 711 FOREACH_KSEGRP_IN_PROC(p, kg) { 712 resetpriority(kg); 713 } 714 } 715 716 void 717 sched_class(struct ksegrp *kg, int class) 718 { 719 mtx_assert(&sched_lock, MA_OWNED); 720 kg->kg_pri_class = class; 721 } 722 723 /* 724 * Adjust the priority of a thread. 725 * This may include moving the thread within the KSEGRP, 726 * changing the assignment of a kse to the thread, 727 * and moving a KSE in the system run queue. 728 */ 729 void 730 sched_prio(struct thread *td, u_char prio) 731 { 732 733 mtx_assert(&sched_lock, MA_OWNED); 734 if (TD_ON_RUNQ(td)) { 735 adjustrunqueue(td, prio); 736 } else { 737 td->td_priority = prio; 738 } 739 } 740 741 void 742 sched_sleep(struct thread *td) 743 { 744 745 mtx_assert(&sched_lock, MA_OWNED); 746 td->td_ksegrp->kg_slptime = 0; 747 td->td_base_pri = td->td_priority; 748 } 749 750 static void remrunqueue(struct thread *td); 751 752 void 753 sched_switch(struct thread *td, struct thread *newtd, int flags) 754 { 755 struct kse *ke; 756 struct ksegrp *kg; 757 struct proc *p; 758 759 ke = td->td_kse; 760 p = td->td_proc; 761 762 mtx_assert(&sched_lock, MA_OWNED); 763 764 if ((p->p_flag & P_NOLOAD) == 0) 765 sched_tdcnt--; 766 /* 767 * We are volunteering to switch out so we get to nominate 768 * a successor for the rest of our quantum 769 * First try another thread in our ksegrp, and then look for 770 * other ksegrps in our process. 771 */ 772 if (sched_followon && 773 (p->p_flag & P_HADTHREADS) && 774 (flags & SW_VOL) && 775 newtd == NULL) { 776 /* lets schedule another thread from this process */ 777 kg = td->td_ksegrp; 778 if ((newtd = TAILQ_FIRST(&kg->kg_runq))) { 779 remrunqueue(newtd); 780 sched_kgfollowons++; 781 } else { 782 FOREACH_KSEGRP_IN_PROC(p, kg) { 783 if ((newtd = TAILQ_FIRST(&kg->kg_runq))) { 784 sched_pfollowons++; 785 remrunqueue(newtd); 786 break; 787 } 788 } 789 } 790 } 791 792 /* 793 * The thread we are about to run needs to be counted as if it had been 794 * added to the run queue and selected. 795 * it came from: 796 * A preemption 797 * An upcall 798 * A followon 799 * Do this before saving curthread so that the slot count 800 * doesn't give an overly optimistic view when that happens. 801 */ 802 if (newtd) { 803 KASSERT((newtd->td_inhibitors == 0), 804 ("trying to run inhibitted thread")); 805 newtd->td_ksegrp->kg_avail_opennings--; 806 newtd->td_kse->ke_flags |= KEF_DIDRUN; 807 TD_SET_RUNNING(newtd); 808 if ((newtd->td_proc->p_flag & P_NOLOAD) == 0) 809 sched_tdcnt++; 810 } 811 812 td->td_lastcpu = td->td_oncpu; 813 td->td_flags &= ~TDF_NEEDRESCHED; 814 td->td_pflags &= ~TDP_OWEPREEMPT; 815 td->td_oncpu = NOCPU; 816 /* 817 * At the last moment, if this thread is still marked RUNNING, 818 * then put it back on the run queue as it has not been suspended 819 * or stopped or any thing else similar. We never put the idle 820 * threads on the run queue, however. 821 */ 822 if (td == PCPU_GET(idlethread)) 823 TD_SET_CAN_RUN(td); 824 else { 825 td->td_ksegrp->kg_avail_opennings++; 826 if (TD_IS_RUNNING(td)) { 827 /* Put us back on the run queue (kse and all). */ 828 setrunqueue(td, SRQ_OURSELF|SRQ_YIELDING); 829 } else if (p->p_flag & P_HADTHREADS) { 830 /* 831 * We will not be on the run queue. So we must be 832 * sleeping or similar. As it's available, 833 * someone else can use the KSE if they need it. 834 */ 835 slot_fill(td->td_ksegrp); 836 } 837 } 838 if (newtd == NULL) 839 newtd = choosethread(); 840 if (td != newtd) 841 cpu_switch(td, newtd); 842 sched_lock.mtx_lock = (uintptr_t)td; 843 td->td_oncpu = PCPU_GET(cpuid); 844 } 845 846 void 847 sched_wakeup(struct thread *td) 848 { 849 struct ksegrp *kg; 850 851 mtx_assert(&sched_lock, MA_OWNED); 852 kg = td->td_ksegrp; 853 if (kg->kg_slptime > 1) 854 updatepri(kg); 855 kg->kg_slptime = 0; 856 setrunqueue(td, SRQ_BORING); 857 } 858 859 #ifdef SMP 860 /* enable HTT_2 if you have a 2-way HTT cpu.*/ 861 static int 862 forward_wakeup(int cpunum) 863 { 864 cpumask_t map, me, dontuse; 865 cpumask_t map2; 866 struct pcpu *pc; 867 cpumask_t id, map3; 868 869 mtx_assert(&sched_lock, MA_OWNED); 870 871 CTR0(KTR_RUNQ, "forward_wakeup()"); 872 873 if ((!forward_wakeup_enabled) || 874 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0)) 875 return (0); 876 if (!smp_started || cold || panicstr) 877 return (0); 878 879 forward_wakeups_requested++; 880 881 /* 882 * check the idle mask we received against what we calculated before 883 * in the old version. 884 */ 885 me = PCPU_GET(cpumask); 886 /* 887 * don't bother if we should be doing it ourself.. 888 */ 889 if ((me & idle_cpus_mask) && (cpunum == NOCPU || me == (1 << cpunum))) 890 return (0); 891 892 dontuse = me | stopped_cpus | hlt_cpus_mask; 893 map3 = 0; 894 if (forward_wakeup_use_loop) { 895 SLIST_FOREACH(pc, &cpuhead, pc_allcpu) { 896 id = pc->pc_cpumask; 897 if ( (id & dontuse) == 0 && 898 pc->pc_curthread == pc->pc_idlethread) { 899 map3 |= id; 900 } 901 } 902 } 903 904 if (forward_wakeup_use_mask) { 905 map = 0; 906 map = idle_cpus_mask & ~dontuse; 907 908 /* If they are both on, compare and use loop if different */ 909 if (forward_wakeup_use_loop) { 910 if (map != map3) { 911 printf("map (%02X) != map3 (%02X)\n", 912 map, map3); 913 map = map3; 914 } 915 } 916 } else { 917 map = map3; 918 } 919 /* If we only allow a specific CPU, then mask off all the others */ 920 if (cpunum != NOCPU) { 921 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum.")); 922 map &= (1 << cpunum); 923 } else { 924 /* Try choose an idle die. */ 925 if (forward_wakeup_use_htt) { 926 map2 = (map & (map >> 1)) & 0x5555; 927 if (map2) { 928 map = map2; 929 } 930 } 931 932 /* set only one bit */ 933 if (forward_wakeup_use_single) { 934 map = map & ((~map) + 1); 935 } 936 } 937 if (map) { 938 forward_wakeups_delivered++; 939 ipi_selected(map, IPI_AST); 940 return (1); 941 } 942 if (cpunum == NOCPU) 943 printf("forward_wakeup: Idle processor not found\n"); 944 return (0); 945 } 946 #endif 947 948 void 949 sched_add(struct thread *td, int flags) 950 { 951 struct kse *ke; 952 #ifdef SMP 953 int forwarded = 0; 954 int cpu; 955 #endif 956 957 ke = td->td_kse; 958 mtx_assert(&sched_lock, MA_OWNED); 959 KASSERT(ke->ke_state != KES_ONRUNQ, 960 ("sched_add: kse %p (%s) already in run queue", ke, 961 ke->ke_proc->p_comm)); 962 KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 963 ("sched_add: process swapped out")); 964 965 #ifdef SMP 966 if (KSE_CAN_MIGRATE(ke)) { 967 CTR2(KTR_RUNQ, 968 "sched_add: adding kse:%p (td:%p) to gbl runq", ke, td); 969 cpu = NOCPU; 970 ke->ke_runq = &runq; 971 } else { 972 if (!SKE_RUNQ_PCPU(ke)) 973 ke->ke_runq = &runq_pcpu[(cpu = PCPU_GET(cpuid))]; 974 else 975 cpu = td->td_lastcpu; 976 CTR3(KTR_RUNQ, 977 "sched_add: Put kse:%p(td:%p) on cpu%d runq", ke, td, cpu); 978 } 979 #else 980 CTR2(KTR_RUNQ, "sched_add: adding kse:%p (td:%p) to runq", ke, td); 981 ke->ke_runq = &runq; 982 983 #endif 984 /* 985 * If we are yielding (on the way out anyhow) 986 * or the thread being saved is US, 987 * then don't try be smart about preemption 988 * or kicking off another CPU 989 * as it won't help and may hinder. 990 * In the YIEDLING case, we are about to run whoever is 991 * being put in the queue anyhow, and in the 992 * OURSELF case, we are puting ourself on the run queue 993 * which also only happens when we are about to yield. 994 */ 995 if((flags & SRQ_YIELDING) == 0) { 996 #ifdef SMP 997 cpumask_t me = PCPU_GET(cpumask); 998 int idle = idle_cpus_mask & me; 999 /* 1000 * Only try to kick off another CPU if 1001 * the thread is unpinned 1002 * or pinned to another cpu, 1003 * and there are other available and idle CPUs. 1004 * if we are idle, or it's an interrupt, 1005 * then skip straight to preemption. 1006 */ 1007 if ( (! idle) && ((flags & SRQ_INTR) == 0) && 1008 (idle_cpus_mask & ~(hlt_cpus_mask | me)) && 1009 ( KSE_CAN_MIGRATE(ke) || 1010 ke->ke_runq != &runq_pcpu[PCPU_GET(cpuid)])) { 1011 forwarded = forward_wakeup(cpu); 1012 } 1013 /* 1014 * If we failed to kick off another cpu, then look to 1015 * see if we should preempt this CPU. Only allow this 1016 * if it is not pinned or IS pinned to this CPU. 1017 * If we are the idle thread, we also try do preempt. 1018 * as it will be quicker and being idle, we won't 1019 * lose in doing so.. 1020 */ 1021 if ((!forwarded) && 1022 (ke->ke_runq == &runq || 1023 ke->ke_runq == &runq_pcpu[PCPU_GET(cpuid)])) 1024 #endif 1025 1026 { 1027 if (maybe_preempt(td)) 1028 return; 1029 } 1030 } 1031 if ((td->td_proc->p_flag & P_NOLOAD) == 0) 1032 sched_tdcnt++; 1033 td->td_ksegrp->kg_avail_opennings--; 1034 runq_add(ke->ke_runq, ke); 1035 ke->ke_ksegrp->kg_runq_kses++; 1036 ke->ke_state = KES_ONRUNQ; 1037 maybe_resched(td); 1038 } 1039 1040 void 1041 sched_rem(struct thread *td) 1042 { 1043 struct kse *ke; 1044 1045 ke = td->td_kse; 1046 KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 1047 ("sched_rem: process swapped out")); 1048 KASSERT((ke->ke_state == KES_ONRUNQ), 1049 ("sched_rem: KSE not on run queue")); 1050 mtx_assert(&sched_lock, MA_OWNED); 1051 1052 if ((td->td_proc->p_flag & P_NOLOAD) == 0) 1053 sched_tdcnt--; 1054 td->td_ksegrp->kg_avail_opennings++; 1055 runq_remove(ke->ke_runq, ke); 1056 1057 ke->ke_state = KES_THREAD; 1058 td->td_ksegrp->kg_runq_kses--; 1059 } 1060 1061 /* 1062 * Select threads to run. 1063 * Notice that the running threads still consume a slot. 1064 */ 1065 struct kse * 1066 sched_choose(void) 1067 { 1068 struct kse *ke; 1069 struct runq *rq; 1070 1071 #ifdef SMP 1072 struct kse *kecpu; 1073 1074 rq = &runq; 1075 ke = runq_choose(&runq); 1076 kecpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]); 1077 1078 if (ke == NULL || 1079 (kecpu != NULL && 1080 kecpu->ke_thread->td_priority < ke->ke_thread->td_priority)) { 1081 CTR2(KTR_RUNQ, "choosing kse %p from pcpu runq %d", kecpu, 1082 PCPU_GET(cpuid)); 1083 ke = kecpu; 1084 rq = &runq_pcpu[PCPU_GET(cpuid)]; 1085 } else { 1086 CTR1(KTR_RUNQ, "choosing kse %p from main runq", ke); 1087 } 1088 1089 #else 1090 rq = &runq; 1091 ke = runq_choose(&runq); 1092 #endif 1093 1094 if (ke != NULL) { 1095 runq_remove(rq, ke); 1096 ke->ke_state = KES_THREAD; 1097 ke->ke_ksegrp->kg_runq_kses--; 1098 1099 KASSERT(ke->ke_proc->p_sflag & PS_INMEM, 1100 ("sched_choose: process swapped out")); 1101 } 1102 return (ke); 1103 } 1104 1105 void 1106 sched_userret(struct thread *td) 1107 { 1108 struct ksegrp *kg; 1109 /* 1110 * XXX we cheat slightly on the locking here to avoid locking in 1111 * the usual case. Setting td_priority here is essentially an 1112 * incomplete workaround for not setting it properly elsewhere. 1113 * Now that some interrupt handlers are threads, not setting it 1114 * properly elsewhere can clobber it in the window between setting 1115 * it here and returning to user mode, so don't waste time setting 1116 * it perfectly here. 1117 */ 1118 kg = td->td_ksegrp; 1119 if (td->td_priority != kg->kg_user_pri) { 1120 mtx_lock_spin(&sched_lock); 1121 td->td_priority = kg->kg_user_pri; 1122 mtx_unlock_spin(&sched_lock); 1123 } 1124 } 1125 1126 void 1127 sched_bind(struct thread *td, int cpu) 1128 { 1129 struct kse *ke; 1130 1131 mtx_assert(&sched_lock, MA_OWNED); 1132 KASSERT(TD_IS_RUNNING(td), 1133 ("sched_bind: cannot bind non-running thread")); 1134 1135 ke = td->td_kse; 1136 1137 ke->ke_flags |= KEF_BOUND; 1138 #ifdef SMP 1139 ke->ke_runq = &runq_pcpu[cpu]; 1140 if (PCPU_GET(cpuid) == cpu) 1141 return; 1142 1143 ke->ke_state = KES_THREAD; 1144 1145 mi_switch(SW_VOL, NULL); 1146 #endif 1147 } 1148 1149 void 1150 sched_unbind(struct thread* td) 1151 { 1152 mtx_assert(&sched_lock, MA_OWNED); 1153 td->td_kse->ke_flags &= ~KEF_BOUND; 1154 } 1155 1156 int 1157 sched_load(void) 1158 { 1159 return (sched_tdcnt); 1160 } 1161 1162 int 1163 sched_sizeof_ksegrp(void) 1164 { 1165 return (sizeof(struct ksegrp) + sizeof(struct kg_sched)); 1166 } 1167 int 1168 sched_sizeof_proc(void) 1169 { 1170 return (sizeof(struct proc)); 1171 } 1172 int 1173 sched_sizeof_thread(void) 1174 { 1175 return (sizeof(struct thread) + sizeof(struct kse)); 1176 } 1177 1178 fixpt_t 1179 sched_pctcpu(struct thread *td) 1180 { 1181 struct kse *ke; 1182 1183 ke = td->td_kse; 1184 return (ke->ke_pctcpu); 1185 1186 return (0); 1187 } 1188 #define KERN_SWITCH_INCLUDE 1 1189 #include "kern/kern_switch.c" 1190