1 /*- 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 1982, 1986, 1990, 1991, 1993 5 * The Regents of the University of California. All rights reserved. 6 * (c) UNIX System Laboratories, Inc. 7 * All or some portions of this file are derived from material licensed 8 * to the University of California by American Telephone and Telegraph 9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 10 * the permission of UNIX System Laboratories, Inc. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 */ 36 37 #include <sys/cdefs.h> 38 __FBSDID("$FreeBSD$"); 39 40 #include "opt_hwpmc_hooks.h" 41 #include "opt_sched.h" 42 43 #include <sys/param.h> 44 #include <sys/systm.h> 45 #include <sys/cpuset.h> 46 #include <sys/kernel.h> 47 #include <sys/ktr.h> 48 #include <sys/lock.h> 49 #include <sys/kthread.h> 50 #include <sys/mutex.h> 51 #include <sys/proc.h> 52 #include <sys/resourcevar.h> 53 #include <sys/sched.h> 54 #include <sys/sdt.h> 55 #include <sys/smp.h> 56 #include <sys/sysctl.h> 57 #include <sys/sx.h> 58 #include <sys/turnstile.h> 59 #include <sys/umtxvar.h> 60 #include <machine/pcb.h> 61 #include <machine/smp.h> 62 63 #ifdef HWPMC_HOOKS 64 #include <sys/pmckern.h> 65 #endif 66 67 #ifdef KDTRACE_HOOKS 68 #include <sys/dtrace_bsd.h> 69 int __read_mostly dtrace_vtime_active; 70 dtrace_vtime_switch_func_t dtrace_vtime_switch_func; 71 #endif 72 73 /* 74 * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in 75 * the range 100-256 Hz (approximately). 76 */ 77 #define ESTCPULIM(e) \ 78 min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \ 79 RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1) 80 #ifdef SMP 81 #define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus) 82 #else 83 #define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */ 84 #endif 85 #define NICE_WEIGHT 1 /* Priorities per nice level. */ 86 87 #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX))) 88 89 /* 90 * The schedulable entity that runs a context. 91 * This is an extension to the thread structure and is tailored to 92 * the requirements of this scheduler. 93 * All fields are protected by the scheduler lock. 94 */ 95 struct td_sched { 96 fixpt_t ts_pctcpu; /* %cpu during p_swtime. */ 97 u_int ts_estcpu; /* Estimated cpu utilization. */ 98 int ts_cpticks; /* Ticks of cpu time. */ 99 int ts_slptime; /* Seconds !RUNNING. */ 100 int ts_slice; /* Remaining part of time slice. */ 101 int ts_flags; 102 struct runq *ts_runq; /* runq the thread is currently on */ 103 #ifdef KTR 104 char ts_name[TS_NAME_LEN]; 105 #endif 106 }; 107 108 /* flags kept in td_flags */ 109 #define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */ 110 #define TDF_BOUND TDF_SCHED1 /* Bound to one CPU. */ 111 #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */ 112 113 /* flags kept in ts_flags */ 114 #define TSF_AFFINITY 0x0001 /* Has a non-"full" CPU set. */ 115 116 #define SKE_RUNQ_PCPU(ts) \ 117 ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq) 118 119 #define THREAD_CAN_SCHED(td, cpu) \ 120 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 121 122 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <= 123 sizeof(struct thread0_storage), 124 "increase struct thread0_storage.t0st_sched size"); 125 126 static struct mtx sched_lock; 127 128 static int realstathz = 127; /* stathz is sometimes 0 and run off of hz. */ 129 static int sched_tdcnt; /* Total runnable threads in the system. */ 130 static int sched_slice = 12; /* Thread run time before rescheduling. */ 131 132 static void setup_runqs(void); 133 static void schedcpu(void); 134 static void schedcpu_thread(void); 135 static void sched_priority(struct thread *td, u_char prio); 136 static void sched_setup(void *dummy); 137 static void maybe_resched(struct thread *td); 138 static void updatepri(struct thread *td); 139 static void resetpriority(struct thread *td); 140 static void resetpriority_thread(struct thread *td); 141 #ifdef SMP 142 static int sched_pickcpu(struct thread *td); 143 static int forward_wakeup(int cpunum); 144 static void kick_other_cpu(int pri, int cpuid); 145 #endif 146 147 static struct kproc_desc sched_kp = { 148 "schedcpu", 149 schedcpu_thread, 150 NULL 151 }; 152 SYSINIT(schedcpu, SI_SUB_LAST, SI_ORDER_FIRST, kproc_start, 153 &sched_kp); 154 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 155 156 static void sched_initticks(void *dummy); 157 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 158 NULL); 159 160 /* 161 * Global run queue. 162 */ 163 static struct runq runq; 164 165 #ifdef SMP 166 /* 167 * Per-CPU run queues 168 */ 169 static struct runq runq_pcpu[MAXCPU]; 170 long runq_length[MAXCPU]; 171 172 static cpuset_t idle_cpus_mask; 173 #endif 174 175 struct pcpuidlestat { 176 u_int idlecalls; 177 u_int oldidlecalls; 178 }; 179 DPCPU_DEFINE_STATIC(struct pcpuidlestat, idlestat); 180 181 static void 182 setup_runqs(void) 183 { 184 #ifdef SMP 185 int i; 186 187 for (i = 0; i < MAXCPU; ++i) 188 runq_init(&runq_pcpu[i]); 189 #endif 190 191 runq_init(&runq); 192 } 193 194 static int 195 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 196 { 197 int error, new_val, period; 198 199 period = 1000000 / realstathz; 200 new_val = period * sched_slice; 201 error = sysctl_handle_int(oidp, &new_val, 0, req); 202 if (error != 0 || req->newptr == NULL) 203 return (error); 204 if (new_val <= 0) 205 return (EINVAL); 206 sched_slice = imax(1, (new_val + period / 2) / period); 207 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 208 realstathz); 209 return (0); 210 } 211 212 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 213 "Scheduler"); 214 215 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0, 216 "Scheduler name"); 217 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, 218 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0, 219 sysctl_kern_quantum, "I", 220 "Quantum for timeshare threads in microseconds"); 221 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 222 "Quantum for timeshare threads in stathz ticks"); 223 #ifdef SMP 224 /* Enable forwarding of wakeups to all other cpus */ 225 static SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, 226 CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 227 "Kernel SMP"); 228 229 static int runq_fuzz = 1; 230 SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, ""); 231 232 static int forward_wakeup_enabled = 1; 233 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW, 234 &forward_wakeup_enabled, 0, 235 "Forwarding of wakeup to idle CPUs"); 236 237 static int forward_wakeups_requested = 0; 238 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD, 239 &forward_wakeups_requested, 0, 240 "Requests for Forwarding of wakeup to idle CPUs"); 241 242 static int forward_wakeups_delivered = 0; 243 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD, 244 &forward_wakeups_delivered, 0, 245 "Completed Forwarding of wakeup to idle CPUs"); 246 247 static int forward_wakeup_use_mask = 1; 248 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW, 249 &forward_wakeup_use_mask, 0, 250 "Use the mask of idle cpus"); 251 252 static int forward_wakeup_use_loop = 0; 253 SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW, 254 &forward_wakeup_use_loop, 0, 255 "Use a loop to find idle cpus"); 256 257 #endif 258 #if 0 259 static int sched_followon = 0; 260 SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW, 261 &sched_followon, 0, 262 "allow threads to share a quantum"); 263 #endif 264 265 SDT_PROVIDER_DEFINE(sched); 266 267 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *", 268 "struct proc *", "uint8_t"); 269 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *", 270 "struct proc *", "void *"); 271 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *", 272 "struct proc *", "void *", "int"); 273 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *", 274 "struct proc *", "uint8_t", "struct thread *"); 275 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int"); 276 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *", 277 "struct proc *"); 278 SDT_PROBE_DEFINE(sched, , , on__cpu); 279 SDT_PROBE_DEFINE(sched, , , remain__cpu); 280 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *", 281 "struct proc *"); 282 283 static __inline void 284 sched_load_add(void) 285 { 286 287 sched_tdcnt++; 288 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt); 289 SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt); 290 } 291 292 static __inline void 293 sched_load_rem(void) 294 { 295 296 sched_tdcnt--; 297 KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt); 298 SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt); 299 } 300 /* 301 * Arrange to reschedule if necessary, taking the priorities and 302 * schedulers into account. 303 */ 304 static void 305 maybe_resched(struct thread *td) 306 { 307 308 THREAD_LOCK_ASSERT(td, MA_OWNED); 309 if (td->td_priority < curthread->td_priority) 310 curthread->td_flags |= TDF_NEEDRESCHED; 311 } 312 313 /* 314 * This function is called when a thread is about to be put on run queue 315 * because it has been made runnable or its priority has been adjusted. It 316 * determines if the new thread should preempt the current thread. If so, 317 * it sets td_owepreempt to request a preemption. 318 */ 319 int 320 maybe_preempt(struct thread *td) 321 { 322 #ifdef PREEMPTION 323 struct thread *ctd; 324 int cpri, pri; 325 326 /* 327 * The new thread should not preempt the current thread if any of the 328 * following conditions are true: 329 * 330 * - The kernel is in the throes of crashing (panicstr). 331 * - The current thread has a higher (numerically lower) or 332 * equivalent priority. Note that this prevents curthread from 333 * trying to preempt to itself. 334 * - The current thread has an inhibitor set or is in the process of 335 * exiting. In this case, the current thread is about to switch 336 * out anyways, so there's no point in preempting. If we did, 337 * the current thread would not be properly resumed as well, so 338 * just avoid that whole landmine. 339 * - If the new thread's priority is not a realtime priority and 340 * the current thread's priority is not an idle priority and 341 * FULL_PREEMPTION is disabled. 342 * 343 * If all of these conditions are false, but the current thread is in 344 * a nested critical section, then we have to defer the preemption 345 * until we exit the critical section. Otherwise, switch immediately 346 * to the new thread. 347 */ 348 ctd = curthread; 349 THREAD_LOCK_ASSERT(td, MA_OWNED); 350 KASSERT((td->td_inhibitors == 0), 351 ("maybe_preempt: trying to run inhibited thread")); 352 pri = td->td_priority; 353 cpri = ctd->td_priority; 354 if (KERNEL_PANICKED() || pri >= cpri /* || dumping */ || 355 TD_IS_INHIBITED(ctd)) 356 return (0); 357 #ifndef FULL_PREEMPTION 358 if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE) 359 return (0); 360 #endif 361 362 CTR0(KTR_PROC, "maybe_preempt: scheduling preemption"); 363 ctd->td_owepreempt = 1; 364 return (1); 365 #else 366 return (0); 367 #endif 368 } 369 370 /* 371 * Constants for digital decay and forget: 372 * 90% of (ts_estcpu) usage in 5 * loadav time 373 * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive) 374 * Note that, as ps(1) mentions, this can let percentages 375 * total over 100% (I've seen 137.9% for 3 processes). 376 * 377 * Note that schedclock() updates ts_estcpu and p_cpticks asynchronously. 378 * 379 * We wish to decay away 90% of ts_estcpu in (5 * loadavg) seconds. 380 * That is, the system wants to compute a value of decay such 381 * that the following for loop: 382 * for (i = 0; i < (5 * loadavg); i++) 383 * ts_estcpu *= decay; 384 * will compute 385 * ts_estcpu *= 0.1; 386 * for all values of loadavg: 387 * 388 * Mathematically this loop can be expressed by saying: 389 * decay ** (5 * loadavg) ~= .1 390 * 391 * The system computes decay as: 392 * decay = (2 * loadavg) / (2 * loadavg + 1) 393 * 394 * We wish to prove that the system's computation of decay 395 * will always fulfill the equation: 396 * decay ** (5 * loadavg) ~= .1 397 * 398 * If we compute b as: 399 * b = 2 * loadavg 400 * then 401 * decay = b / (b + 1) 402 * 403 * We now need to prove two things: 404 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 405 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 406 * 407 * Facts: 408 * For x close to zero, exp(x) =~ 1 + x, since 409 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 410 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 411 * For x close to zero, ln(1+x) =~ x, since 412 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 413 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 414 * ln(.1) =~ -2.30 415 * 416 * Proof of (1): 417 * Solve (factor)**(power) =~ .1 given power (5*loadav): 418 * solving for factor, 419 * ln(factor) =~ (-2.30/5*loadav), or 420 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 421 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 422 * 423 * Proof of (2): 424 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 425 * solving for power, 426 * power*ln(b/(b+1)) =~ -2.30, or 427 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 428 * 429 * Actual power values for the implemented algorithm are as follows: 430 * loadav: 1 2 3 4 431 * power: 5.68 10.32 14.94 19.55 432 */ 433 434 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 435 #define loadfactor(loadav) (2 * (loadav)) 436 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 437 438 /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 439 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 440 SYSCTL_UINT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, 441 "Decay factor used for updating %CPU"); 442 443 /* 444 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 445 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 446 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 447 * 448 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 449 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 450 * 451 * If you don't want to bother with the faster/more-accurate formula, you 452 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 453 * (more general) method of calculating the %age of CPU used by a process. 454 */ 455 #define CCPU_SHIFT 11 456 457 /* 458 * Recompute process priorities, every hz ticks. 459 * MP-safe, called without the Giant mutex. 460 */ 461 /* ARGSUSED */ 462 static void 463 schedcpu(void) 464 { 465 fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 466 struct thread *td; 467 struct proc *p; 468 struct td_sched *ts; 469 int awake; 470 471 sx_slock(&allproc_lock); 472 FOREACH_PROC_IN_SYSTEM(p) { 473 PROC_LOCK(p); 474 if (p->p_state == PRS_NEW) { 475 PROC_UNLOCK(p); 476 continue; 477 } 478 FOREACH_THREAD_IN_PROC(p, td) { 479 awake = 0; 480 ts = td_get_sched(td); 481 thread_lock(td); 482 /* 483 * Increment sleep time (if sleeping). We 484 * ignore overflow, as above. 485 */ 486 /* 487 * The td_sched slptimes are not touched in wakeup 488 * because the thread may not HAVE everything in 489 * memory? XXX I think this is out of date. 490 */ 491 if (TD_ON_RUNQ(td)) { 492 awake = 1; 493 td->td_flags &= ~TDF_DIDRUN; 494 } else if (TD_IS_RUNNING(td)) { 495 awake = 1; 496 /* Do not clear TDF_DIDRUN */ 497 } else if (td->td_flags & TDF_DIDRUN) { 498 awake = 1; 499 td->td_flags &= ~TDF_DIDRUN; 500 } 501 502 /* 503 * ts_pctcpu is only for ps and ttyinfo(). 504 */ 505 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT; 506 /* 507 * If the td_sched has been idle the entire second, 508 * stop recalculating its priority until 509 * it wakes up. 510 */ 511 if (ts->ts_cpticks != 0) { 512 #if (FSHIFT >= CCPU_SHIFT) 513 ts->ts_pctcpu += (realstathz == 100) 514 ? ((fixpt_t) ts->ts_cpticks) << 515 (FSHIFT - CCPU_SHIFT) : 516 100 * (((fixpt_t) ts->ts_cpticks) 517 << (FSHIFT - CCPU_SHIFT)) / realstathz; 518 #else 519 ts->ts_pctcpu += ((FSCALE - ccpu) * 520 (ts->ts_cpticks * 521 FSCALE / realstathz)) >> FSHIFT; 522 #endif 523 ts->ts_cpticks = 0; 524 } 525 /* 526 * If there are ANY running threads in this process, 527 * then don't count it as sleeping. 528 * XXX: this is broken. 529 */ 530 if (awake) { 531 if (ts->ts_slptime > 1) { 532 /* 533 * In an ideal world, this should not 534 * happen, because whoever woke us 535 * up from the long sleep should have 536 * unwound the slptime and reset our 537 * priority before we run at the stale 538 * priority. Should KASSERT at some 539 * point when all the cases are fixed. 540 */ 541 updatepri(td); 542 } 543 ts->ts_slptime = 0; 544 } else 545 ts->ts_slptime++; 546 if (ts->ts_slptime > 1) { 547 thread_unlock(td); 548 continue; 549 } 550 ts->ts_estcpu = decay_cpu(loadfac, ts->ts_estcpu); 551 resetpriority(td); 552 resetpriority_thread(td); 553 thread_unlock(td); 554 } 555 PROC_UNLOCK(p); 556 } 557 sx_sunlock(&allproc_lock); 558 } 559 560 /* 561 * Main loop for a kthread that executes schedcpu once a second. 562 */ 563 static void 564 schedcpu_thread(void) 565 { 566 567 for (;;) { 568 schedcpu(); 569 pause("-", hz); 570 } 571 } 572 573 /* 574 * Recalculate the priority of a process after it has slept for a while. 575 * For all load averages >= 1 and max ts_estcpu of 255, sleeping for at 576 * least six times the loadfactor will decay ts_estcpu to zero. 577 */ 578 static void 579 updatepri(struct thread *td) 580 { 581 struct td_sched *ts; 582 fixpt_t loadfac; 583 unsigned int newcpu; 584 585 ts = td_get_sched(td); 586 loadfac = loadfactor(averunnable.ldavg[0]); 587 if (ts->ts_slptime > 5 * loadfac) 588 ts->ts_estcpu = 0; 589 else { 590 newcpu = ts->ts_estcpu; 591 ts->ts_slptime--; /* was incremented in schedcpu() */ 592 while (newcpu && --ts->ts_slptime) 593 newcpu = decay_cpu(loadfac, newcpu); 594 ts->ts_estcpu = newcpu; 595 } 596 } 597 598 /* 599 * Compute the priority of a process when running in user mode. 600 * Arrange to reschedule if the resulting priority is better 601 * than that of the current process. 602 */ 603 static void 604 resetpriority(struct thread *td) 605 { 606 u_int newpriority; 607 608 if (td->td_pri_class != PRI_TIMESHARE) 609 return; 610 newpriority = PUSER + 611 td_get_sched(td)->ts_estcpu / INVERSE_ESTCPU_WEIGHT + 612 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN); 613 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 614 PRI_MAX_TIMESHARE); 615 sched_user_prio(td, newpriority); 616 } 617 618 /* 619 * Update the thread's priority when the associated process's user 620 * priority changes. 621 */ 622 static void 623 resetpriority_thread(struct thread *td) 624 { 625 626 /* Only change threads with a time sharing user priority. */ 627 if (td->td_priority < PRI_MIN_TIMESHARE || 628 td->td_priority > PRI_MAX_TIMESHARE) 629 return; 630 631 /* XXX the whole needresched thing is broken, but not silly. */ 632 maybe_resched(td); 633 634 sched_prio(td, td->td_user_pri); 635 } 636 637 /* ARGSUSED */ 638 static void 639 sched_setup(void *dummy) 640 { 641 642 setup_runqs(); 643 644 /* Account for thread0. */ 645 sched_load_add(); 646 } 647 648 /* 649 * This routine determines time constants after stathz and hz are setup. 650 */ 651 static void 652 sched_initticks(void *dummy) 653 { 654 655 realstathz = stathz ? stathz : hz; 656 sched_slice = realstathz / 10; /* ~100ms */ 657 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 658 realstathz); 659 } 660 661 /* External interfaces start here */ 662 663 /* 664 * Very early in the boot some setup of scheduler-specific 665 * parts of proc0 and of some scheduler resources needs to be done. 666 * Called from: 667 * proc0_init() 668 */ 669 void 670 schedinit(void) 671 { 672 673 /* 674 * Set up the scheduler specific parts of thread0. 675 */ 676 thread0.td_lock = &sched_lock; 677 td_get_sched(&thread0)->ts_slice = sched_slice; 678 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN); 679 } 680 681 void 682 schedinit_ap(void) 683 { 684 685 /* Nothing needed. */ 686 } 687 688 int 689 sched_runnable(void) 690 { 691 #ifdef SMP 692 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]); 693 #else 694 return runq_check(&runq); 695 #endif 696 } 697 698 int 699 sched_rr_interval(void) 700 { 701 702 /* Convert sched_slice from stathz to hz. */ 703 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz)); 704 } 705 706 SCHED_STAT_DEFINE(ithread_demotions, "Interrupt thread priority demotions"); 707 SCHED_STAT_DEFINE(ithread_preemptions, 708 "Interrupt thread preemptions due to time-sharing"); 709 710 /* 711 * We adjust the priority of the current process. The priority of a 712 * process gets worse as it accumulates CPU time. The cpu usage 713 * estimator (ts_estcpu) is increased here. resetpriority() will 714 * compute a different priority each time ts_estcpu increases by 715 * INVERSE_ESTCPU_WEIGHT (until PRI_MAX_TIMESHARE is reached). The 716 * cpu usage estimator ramps up quite quickly when the process is 717 * running (linearly), and decays away exponentially, at a rate which 718 * is proportionally slower when the system is busy. The basic 719 * principle is that the system will 90% forget that the process used 720 * a lot of CPU time in 5 * loadav seconds. This causes the system to 721 * favor processes which haven't run much recently, and to round-robin 722 * among other processes. 723 */ 724 static void 725 sched_clock_tick(struct thread *td) 726 { 727 struct pcpuidlestat *stat; 728 struct td_sched *ts; 729 730 THREAD_LOCK_ASSERT(td, MA_OWNED); 731 ts = td_get_sched(td); 732 733 ts->ts_cpticks++; 734 ts->ts_estcpu = ESTCPULIM(ts->ts_estcpu + 1); 735 if ((ts->ts_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 736 resetpriority(td); 737 resetpriority_thread(td); 738 } 739 740 /* 741 * Force a context switch if the current thread has used up a full 742 * time slice (default is 100ms). 743 */ 744 if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) { 745 ts->ts_slice = sched_slice; 746 747 /* 748 * If an ithread uses a full quantum, demote its 749 * priority and preempt it. 750 */ 751 if (PRI_BASE(td->td_pri_class) == PRI_ITHD) { 752 SCHED_STAT_INC(ithread_preemptions); 753 td->td_owepreempt = 1; 754 if (td->td_base_pri + RQ_PPQ < PRI_MAX_ITHD) { 755 SCHED_STAT_INC(ithread_demotions); 756 sched_prio(td, td->td_base_pri + RQ_PPQ); 757 } 758 } else 759 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND; 760 } 761 762 stat = DPCPU_PTR(idlestat); 763 stat->oldidlecalls = stat->idlecalls; 764 stat->idlecalls = 0; 765 } 766 767 void 768 sched_clock(struct thread *td, int cnt) 769 { 770 771 for ( ; cnt > 0; cnt--) 772 sched_clock_tick(td); 773 } 774 775 /* 776 * Charge child's scheduling CPU usage to parent. 777 */ 778 void 779 sched_exit(struct proc *p, struct thread *td) 780 { 781 782 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit", 783 "prio:%d", td->td_priority); 784 785 PROC_LOCK_ASSERT(p, MA_OWNED); 786 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td); 787 } 788 789 void 790 sched_exit_thread(struct thread *td, struct thread *child) 791 { 792 793 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit", 794 "prio:%d", child->td_priority); 795 thread_lock(td); 796 td_get_sched(td)->ts_estcpu = ESTCPULIM(td_get_sched(td)->ts_estcpu + 797 td_get_sched(child)->ts_estcpu); 798 thread_unlock(td); 799 thread_lock(child); 800 if ((child->td_flags & TDF_NOLOAD) == 0) 801 sched_load_rem(); 802 thread_unlock(child); 803 } 804 805 void 806 sched_fork(struct thread *td, struct thread *childtd) 807 { 808 sched_fork_thread(td, childtd); 809 } 810 811 void 812 sched_fork_thread(struct thread *td, struct thread *childtd) 813 { 814 struct td_sched *ts, *tsc; 815 816 childtd->td_oncpu = NOCPU; 817 childtd->td_lastcpu = NOCPU; 818 childtd->td_lock = &sched_lock; 819 childtd->td_cpuset = cpuset_ref(td->td_cpuset); 820 childtd->td_domain.dr_policy = td->td_cpuset->cs_domain; 821 childtd->td_priority = childtd->td_base_pri; 822 ts = td_get_sched(childtd); 823 bzero(ts, sizeof(*ts)); 824 tsc = td_get_sched(td); 825 ts->ts_estcpu = tsc->ts_estcpu; 826 ts->ts_flags |= (tsc->ts_flags & TSF_AFFINITY); 827 ts->ts_slice = 1; 828 } 829 830 void 831 sched_nice(struct proc *p, int nice) 832 { 833 struct thread *td; 834 835 PROC_LOCK_ASSERT(p, MA_OWNED); 836 p->p_nice = nice; 837 FOREACH_THREAD_IN_PROC(p, td) { 838 thread_lock(td); 839 resetpriority(td); 840 resetpriority_thread(td); 841 thread_unlock(td); 842 } 843 } 844 845 void 846 sched_class(struct thread *td, int class) 847 { 848 THREAD_LOCK_ASSERT(td, MA_OWNED); 849 td->td_pri_class = class; 850 } 851 852 /* 853 * Adjust the priority of a thread. 854 */ 855 static void 856 sched_priority(struct thread *td, u_char prio) 857 { 858 859 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change", 860 "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED, 861 sched_tdname(curthread)); 862 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio); 863 if (td != curthread && prio > td->td_priority) { 864 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 865 "lend prio", "prio:%d", td->td_priority, "new prio:%d", 866 prio, KTR_ATTR_LINKED, sched_tdname(td)); 867 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio, 868 curthread); 869 } 870 THREAD_LOCK_ASSERT(td, MA_OWNED); 871 if (td->td_priority == prio) 872 return; 873 td->td_priority = prio; 874 if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) { 875 sched_rem(td); 876 sched_add(td, SRQ_BORING | SRQ_HOLDTD); 877 } 878 } 879 880 /* 881 * Update a thread's priority when it is lent another thread's 882 * priority. 883 */ 884 void 885 sched_lend_prio(struct thread *td, u_char prio) 886 { 887 888 td->td_flags |= TDF_BORROWING; 889 sched_priority(td, prio); 890 } 891 892 /* 893 * Restore a thread's priority when priority propagation is 894 * over. The prio argument is the minimum priority the thread 895 * needs to have to satisfy other possible priority lending 896 * requests. If the thread's regulary priority is less 897 * important than prio the thread will keep a priority boost 898 * of prio. 899 */ 900 void 901 sched_unlend_prio(struct thread *td, u_char prio) 902 { 903 u_char base_pri; 904 905 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 906 td->td_base_pri <= PRI_MAX_TIMESHARE) 907 base_pri = td->td_user_pri; 908 else 909 base_pri = td->td_base_pri; 910 if (prio >= base_pri) { 911 td->td_flags &= ~TDF_BORROWING; 912 sched_prio(td, base_pri); 913 } else 914 sched_lend_prio(td, prio); 915 } 916 917 void 918 sched_prio(struct thread *td, u_char prio) 919 { 920 u_char oldprio; 921 922 /* First, update the base priority. */ 923 td->td_base_pri = prio; 924 925 /* 926 * If the thread is borrowing another thread's priority, don't ever 927 * lower the priority. 928 */ 929 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 930 return; 931 932 /* Change the real priority. */ 933 oldprio = td->td_priority; 934 sched_priority(td, prio); 935 936 /* 937 * If the thread is on a turnstile, then let the turnstile update 938 * its state. 939 */ 940 if (TD_ON_LOCK(td) && oldprio != prio) 941 turnstile_adjust(td, oldprio); 942 } 943 944 void 945 sched_ithread_prio(struct thread *td, u_char prio) 946 { 947 THREAD_LOCK_ASSERT(td, MA_OWNED); 948 MPASS(td->td_pri_class == PRI_ITHD); 949 td->td_base_ithread_pri = prio; 950 sched_prio(td, prio); 951 } 952 953 void 954 sched_user_prio(struct thread *td, u_char prio) 955 { 956 957 THREAD_LOCK_ASSERT(td, MA_OWNED); 958 td->td_base_user_pri = prio; 959 if (td->td_lend_user_pri <= prio) 960 return; 961 td->td_user_pri = prio; 962 } 963 964 void 965 sched_lend_user_prio(struct thread *td, u_char prio) 966 { 967 968 THREAD_LOCK_ASSERT(td, MA_OWNED); 969 td->td_lend_user_pri = prio; 970 td->td_user_pri = min(prio, td->td_base_user_pri); 971 if (td->td_priority > td->td_user_pri) 972 sched_prio(td, td->td_user_pri); 973 else if (td->td_priority != td->td_user_pri) 974 td->td_flags |= TDF_NEEDRESCHED; 975 } 976 977 /* 978 * Like the above but first check if there is anything to do. 979 */ 980 void 981 sched_lend_user_prio_cond(struct thread *td, u_char prio) 982 { 983 984 if (td->td_lend_user_pri != prio) 985 goto lend; 986 if (td->td_user_pri != min(prio, td->td_base_user_pri)) 987 goto lend; 988 if (td->td_priority != td->td_user_pri) 989 goto lend; 990 return; 991 992 lend: 993 thread_lock(td); 994 sched_lend_user_prio(td, prio); 995 thread_unlock(td); 996 } 997 998 void 999 sched_sleep(struct thread *td, int pri) 1000 { 1001 1002 THREAD_LOCK_ASSERT(td, MA_OWNED); 1003 td->td_slptick = ticks; 1004 td_get_sched(td)->ts_slptime = 0; 1005 if (pri != 0 && PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 1006 sched_prio(td, pri); 1007 if (TD_IS_SUSPENDED(td) || pri >= PSOCK) 1008 td->td_flags |= TDF_CANSWAP; 1009 } 1010 1011 void 1012 sched_switch(struct thread *td, int flags) 1013 { 1014 struct thread *newtd; 1015 struct mtx *tmtx; 1016 int preempted; 1017 1018 tmtx = &sched_lock; 1019 1020 THREAD_LOCK_ASSERT(td, MA_OWNED); 1021 1022 td->td_lastcpu = td->td_oncpu; 1023 preempted = (td->td_flags & TDF_SLICEEND) == 0 && 1024 (flags & SW_PREEMPT) != 0; 1025 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND); 1026 td->td_owepreempt = 0; 1027 td->td_oncpu = NOCPU; 1028 1029 /* 1030 * At the last moment, if this thread is still marked RUNNING, 1031 * then put it back on the run queue as it has not been suspended 1032 * or stopped or any thing else similar. We never put the idle 1033 * threads on the run queue, however. 1034 */ 1035 if (td->td_flags & TDF_IDLETD) { 1036 TD_SET_CAN_RUN(td); 1037 #ifdef SMP 1038 CPU_CLR(PCPU_GET(cpuid), &idle_cpus_mask); 1039 #endif 1040 } else { 1041 if (TD_IS_RUNNING(td)) { 1042 /* Put us back on the run queue. */ 1043 sched_add(td, preempted ? 1044 SRQ_HOLDTD|SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1045 SRQ_HOLDTD|SRQ_OURSELF|SRQ_YIELDING); 1046 } 1047 } 1048 1049 /* 1050 * Switch to the sched lock to fix things up and pick 1051 * a new thread. Block the td_lock in order to avoid 1052 * breaking the critical path. 1053 */ 1054 if (td->td_lock != &sched_lock) { 1055 mtx_lock_spin(&sched_lock); 1056 tmtx = thread_lock_block(td); 1057 mtx_unlock_spin(tmtx); 1058 } 1059 1060 if ((td->td_flags & TDF_NOLOAD) == 0) 1061 sched_load_rem(); 1062 1063 newtd = choosethread(); 1064 MPASS(newtd->td_lock == &sched_lock); 1065 1066 #if (KTR_COMPILE & KTR_SCHED) != 0 1067 if (TD_IS_IDLETHREAD(td)) 1068 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle", 1069 "prio:%d", td->td_priority); 1070 else 1071 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td), 1072 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg, 1073 "lockname:\"%s\"", td->td_lockname); 1074 #endif 1075 1076 if (td != newtd) { 1077 #ifdef HWPMC_HOOKS 1078 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1079 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1080 #endif 1081 1082 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc); 1083 1084 /* I feel sleepy */ 1085 lock_profile_release_lock(&sched_lock.lock_object, true); 1086 #ifdef KDTRACE_HOOKS 1087 /* 1088 * If DTrace has set the active vtime enum to anything 1089 * other than INACTIVE (0), then it should have set the 1090 * function to call. 1091 */ 1092 if (dtrace_vtime_active) 1093 (*dtrace_vtime_switch_func)(newtd); 1094 #endif 1095 1096 cpu_switch(td, newtd, tmtx); 1097 lock_profile_obtain_lock_success(&sched_lock.lock_object, true, 1098 0, 0, __FILE__, __LINE__); 1099 /* 1100 * Where am I? What year is it? 1101 * We are in the same thread that went to sleep above, 1102 * but any amount of time may have passed. All our context 1103 * will still be available as will local variables. 1104 * PCPU values however may have changed as we may have 1105 * changed CPU so don't trust cached values of them. 1106 * New threads will go to fork_exit() instead of here 1107 * so if you change things here you may need to change 1108 * things there too. 1109 * 1110 * If the thread above was exiting it will never wake 1111 * up again here, so either it has saved everything it 1112 * needed to, or the thread_wait() or wait() will 1113 * need to reap it. 1114 */ 1115 1116 SDT_PROBE0(sched, , , on__cpu); 1117 #ifdef HWPMC_HOOKS 1118 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1119 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1120 #endif 1121 } else { 1122 td->td_lock = &sched_lock; 1123 SDT_PROBE0(sched, , , remain__cpu); 1124 } 1125 1126 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running", 1127 "prio:%d", td->td_priority); 1128 1129 #ifdef SMP 1130 if (td->td_flags & TDF_IDLETD) 1131 CPU_SET(PCPU_GET(cpuid), &idle_cpus_mask); 1132 #endif 1133 sched_lock.mtx_lock = (uintptr_t)td; 1134 td->td_oncpu = PCPU_GET(cpuid); 1135 spinlock_enter(); 1136 mtx_unlock_spin(&sched_lock); 1137 } 1138 1139 void 1140 sched_wakeup(struct thread *td, int srqflags) 1141 { 1142 struct td_sched *ts; 1143 1144 THREAD_LOCK_ASSERT(td, MA_OWNED); 1145 ts = td_get_sched(td); 1146 td->td_flags &= ~TDF_CANSWAP; 1147 if (ts->ts_slptime > 1) { 1148 updatepri(td); 1149 resetpriority(td); 1150 } 1151 td->td_slptick = 0; 1152 ts->ts_slptime = 0; 1153 ts->ts_slice = sched_slice; 1154 1155 /* 1156 * When resuming an idle ithread, restore its base ithread 1157 * priority. 1158 */ 1159 if (PRI_BASE(td->td_pri_class) == PRI_ITHD && 1160 td->td_base_pri != td->td_base_ithread_pri) 1161 sched_prio(td, td->td_base_ithread_pri); 1162 1163 sched_add(td, srqflags); 1164 } 1165 1166 #ifdef SMP 1167 static int 1168 forward_wakeup(int cpunum) 1169 { 1170 struct pcpu *pc; 1171 cpuset_t dontuse, map, map2; 1172 u_int id, me; 1173 int iscpuset; 1174 1175 mtx_assert(&sched_lock, MA_OWNED); 1176 1177 CTR0(KTR_RUNQ, "forward_wakeup()"); 1178 1179 if ((!forward_wakeup_enabled) || 1180 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0)) 1181 return (0); 1182 if (!smp_started || KERNEL_PANICKED()) 1183 return (0); 1184 1185 forward_wakeups_requested++; 1186 1187 /* 1188 * Check the idle mask we received against what we calculated 1189 * before in the old version. 1190 */ 1191 me = PCPU_GET(cpuid); 1192 1193 /* Don't bother if we should be doing it ourself. */ 1194 if (CPU_ISSET(me, &idle_cpus_mask) && 1195 (cpunum == NOCPU || me == cpunum)) 1196 return (0); 1197 1198 CPU_SETOF(me, &dontuse); 1199 CPU_OR(&dontuse, &dontuse, &stopped_cpus); 1200 CPU_OR(&dontuse, &dontuse, &hlt_cpus_mask); 1201 CPU_ZERO(&map2); 1202 if (forward_wakeup_use_loop) { 1203 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { 1204 id = pc->pc_cpuid; 1205 if (!CPU_ISSET(id, &dontuse) && 1206 pc->pc_curthread == pc->pc_idlethread) { 1207 CPU_SET(id, &map2); 1208 } 1209 } 1210 } 1211 1212 if (forward_wakeup_use_mask) { 1213 map = idle_cpus_mask; 1214 CPU_ANDNOT(&map, &map, &dontuse); 1215 1216 /* If they are both on, compare and use loop if different. */ 1217 if (forward_wakeup_use_loop) { 1218 if (CPU_CMP(&map, &map2)) { 1219 printf("map != map2, loop method preferred\n"); 1220 map = map2; 1221 } 1222 } 1223 } else { 1224 map = map2; 1225 } 1226 1227 /* If we only allow a specific CPU, then mask off all the others. */ 1228 if (cpunum != NOCPU) { 1229 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum.")); 1230 iscpuset = CPU_ISSET(cpunum, &map); 1231 if (iscpuset == 0) 1232 CPU_ZERO(&map); 1233 else 1234 CPU_SETOF(cpunum, &map); 1235 } 1236 if (!CPU_EMPTY(&map)) { 1237 forward_wakeups_delivered++; 1238 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { 1239 id = pc->pc_cpuid; 1240 if (!CPU_ISSET(id, &map)) 1241 continue; 1242 if (cpu_idle_wakeup(pc->pc_cpuid)) 1243 CPU_CLR(id, &map); 1244 } 1245 if (!CPU_EMPTY(&map)) 1246 ipi_selected(map, IPI_AST); 1247 return (1); 1248 } 1249 if (cpunum == NOCPU) 1250 printf("forward_wakeup: Idle processor not found\n"); 1251 return (0); 1252 } 1253 1254 static void 1255 kick_other_cpu(int pri, int cpuid) 1256 { 1257 struct pcpu *pcpu; 1258 int cpri; 1259 1260 pcpu = pcpu_find(cpuid); 1261 if (CPU_ISSET(cpuid, &idle_cpus_mask)) { 1262 forward_wakeups_delivered++; 1263 if (!cpu_idle_wakeup(cpuid)) 1264 ipi_cpu(cpuid, IPI_AST); 1265 return; 1266 } 1267 1268 cpri = pcpu->pc_curthread->td_priority; 1269 if (pri >= cpri) 1270 return; 1271 1272 #if defined(IPI_PREEMPTION) && defined(PREEMPTION) 1273 #if !defined(FULL_PREEMPTION) 1274 if (pri <= PRI_MAX_ITHD) 1275 #endif /* ! FULL_PREEMPTION */ 1276 { 1277 ipi_cpu(cpuid, IPI_PREEMPT); 1278 return; 1279 } 1280 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */ 1281 1282 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED; 1283 ipi_cpu(cpuid, IPI_AST); 1284 return; 1285 } 1286 #endif /* SMP */ 1287 1288 #ifdef SMP 1289 static int 1290 sched_pickcpu(struct thread *td) 1291 { 1292 int best, cpu; 1293 1294 mtx_assert(&sched_lock, MA_OWNED); 1295 1296 if (td->td_lastcpu != NOCPU && THREAD_CAN_SCHED(td, td->td_lastcpu)) 1297 best = td->td_lastcpu; 1298 else 1299 best = NOCPU; 1300 CPU_FOREACH(cpu) { 1301 if (!THREAD_CAN_SCHED(td, cpu)) 1302 continue; 1303 1304 if (best == NOCPU) 1305 best = cpu; 1306 else if (runq_length[cpu] < runq_length[best]) 1307 best = cpu; 1308 } 1309 KASSERT(best != NOCPU, ("no valid CPUs")); 1310 1311 return (best); 1312 } 1313 #endif 1314 1315 void 1316 sched_add(struct thread *td, int flags) 1317 #ifdef SMP 1318 { 1319 cpuset_t tidlemsk; 1320 struct td_sched *ts; 1321 u_int cpu, cpuid; 1322 int forwarded = 0; 1323 int single_cpu = 0; 1324 1325 ts = td_get_sched(td); 1326 THREAD_LOCK_ASSERT(td, MA_OWNED); 1327 KASSERT((td->td_inhibitors == 0), 1328 ("sched_add: trying to run inhibited thread")); 1329 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 1330 ("sched_add: bad thread state")); 1331 KASSERT(td->td_flags & TDF_INMEM, 1332 ("sched_add: thread swapped out")); 1333 1334 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 1335 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 1336 sched_tdname(curthread)); 1337 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 1338 KTR_ATTR_LINKED, sched_tdname(td)); 1339 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 1340 flags & SRQ_PREEMPTED); 1341 1342 /* 1343 * Now that the thread is moving to the run-queue, set the lock 1344 * to the scheduler's lock. 1345 */ 1346 if (td->td_lock != &sched_lock) { 1347 mtx_lock_spin(&sched_lock); 1348 if ((flags & SRQ_HOLD) != 0) 1349 td->td_lock = &sched_lock; 1350 else 1351 thread_lock_set(td, &sched_lock); 1352 } 1353 TD_SET_RUNQ(td); 1354 1355 /* 1356 * If SMP is started and the thread is pinned or otherwise limited to 1357 * a specific set of CPUs, queue the thread to a per-CPU run queue. 1358 * Otherwise, queue the thread to the global run queue. 1359 * 1360 * If SMP has not yet been started we must use the global run queue 1361 * as per-CPU state may not be initialized yet and we may crash if we 1362 * try to access the per-CPU run queues. 1363 */ 1364 if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND || 1365 ts->ts_flags & TSF_AFFINITY)) { 1366 if (td->td_pinned != 0) 1367 cpu = td->td_lastcpu; 1368 else if (td->td_flags & TDF_BOUND) { 1369 /* Find CPU from bound runq. */ 1370 KASSERT(SKE_RUNQ_PCPU(ts), 1371 ("sched_add: bound td_sched not on cpu runq")); 1372 cpu = ts->ts_runq - &runq_pcpu[0]; 1373 } else 1374 /* Find a valid CPU for our cpuset */ 1375 cpu = sched_pickcpu(td); 1376 ts->ts_runq = &runq_pcpu[cpu]; 1377 single_cpu = 1; 1378 CTR3(KTR_RUNQ, 1379 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, 1380 cpu); 1381 } else { 1382 CTR2(KTR_RUNQ, 1383 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts, 1384 td); 1385 cpu = NOCPU; 1386 ts->ts_runq = &runq; 1387 } 1388 1389 if ((td->td_flags & TDF_NOLOAD) == 0) 1390 sched_load_add(); 1391 runq_add(ts->ts_runq, td, flags); 1392 if (cpu != NOCPU) 1393 runq_length[cpu]++; 1394 1395 cpuid = PCPU_GET(cpuid); 1396 if (single_cpu && cpu != cpuid) { 1397 kick_other_cpu(td->td_priority, cpu); 1398 } else { 1399 if (!single_cpu) { 1400 tidlemsk = idle_cpus_mask; 1401 CPU_ANDNOT(&tidlemsk, &tidlemsk, &hlt_cpus_mask); 1402 CPU_CLR(cpuid, &tidlemsk); 1403 1404 if (!CPU_ISSET(cpuid, &idle_cpus_mask) && 1405 ((flags & SRQ_INTR) == 0) && 1406 !CPU_EMPTY(&tidlemsk)) 1407 forwarded = forward_wakeup(cpu); 1408 } 1409 1410 if (!forwarded) { 1411 if (!maybe_preempt(td)) 1412 maybe_resched(td); 1413 } 1414 } 1415 if ((flags & SRQ_HOLDTD) == 0) 1416 thread_unlock(td); 1417 } 1418 #else /* SMP */ 1419 { 1420 struct td_sched *ts; 1421 1422 ts = td_get_sched(td); 1423 THREAD_LOCK_ASSERT(td, MA_OWNED); 1424 KASSERT((td->td_inhibitors == 0), 1425 ("sched_add: trying to run inhibited thread")); 1426 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 1427 ("sched_add: bad thread state")); 1428 KASSERT(td->td_flags & TDF_INMEM, 1429 ("sched_add: thread swapped out")); 1430 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 1431 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 1432 sched_tdname(curthread)); 1433 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 1434 KTR_ATTR_LINKED, sched_tdname(td)); 1435 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 1436 flags & SRQ_PREEMPTED); 1437 1438 /* 1439 * Now that the thread is moving to the run-queue, set the lock 1440 * to the scheduler's lock. 1441 */ 1442 if (td->td_lock != &sched_lock) { 1443 mtx_lock_spin(&sched_lock); 1444 if ((flags & SRQ_HOLD) != 0) 1445 td->td_lock = &sched_lock; 1446 else 1447 thread_lock_set(td, &sched_lock); 1448 } 1449 TD_SET_RUNQ(td); 1450 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td); 1451 ts->ts_runq = &runq; 1452 1453 if ((td->td_flags & TDF_NOLOAD) == 0) 1454 sched_load_add(); 1455 runq_add(ts->ts_runq, td, flags); 1456 if (!maybe_preempt(td)) 1457 maybe_resched(td); 1458 if ((flags & SRQ_HOLDTD) == 0) 1459 thread_unlock(td); 1460 } 1461 #endif /* SMP */ 1462 1463 void 1464 sched_rem(struct thread *td) 1465 { 1466 struct td_sched *ts; 1467 1468 ts = td_get_sched(td); 1469 KASSERT(td->td_flags & TDF_INMEM, 1470 ("sched_rem: thread swapped out")); 1471 KASSERT(TD_ON_RUNQ(td), 1472 ("sched_rem: thread not on run queue")); 1473 mtx_assert(&sched_lock, MA_OWNED); 1474 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 1475 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 1476 sched_tdname(curthread)); 1477 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); 1478 1479 if ((td->td_flags & TDF_NOLOAD) == 0) 1480 sched_load_rem(); 1481 #ifdef SMP 1482 if (ts->ts_runq != &runq) 1483 runq_length[ts->ts_runq - runq_pcpu]--; 1484 #endif 1485 runq_remove(ts->ts_runq, td); 1486 TD_SET_CAN_RUN(td); 1487 } 1488 1489 /* 1490 * Select threads to run. Note that running threads still consume a 1491 * slot. 1492 */ 1493 struct thread * 1494 sched_choose(void) 1495 { 1496 struct thread *td; 1497 struct runq *rq; 1498 1499 mtx_assert(&sched_lock, MA_OWNED); 1500 #ifdef SMP 1501 struct thread *tdcpu; 1502 1503 rq = &runq; 1504 td = runq_choose_fuzz(&runq, runq_fuzz); 1505 tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]); 1506 1507 if (td == NULL || 1508 (tdcpu != NULL && 1509 tdcpu->td_priority < td->td_priority)) { 1510 CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu, 1511 PCPU_GET(cpuid)); 1512 td = tdcpu; 1513 rq = &runq_pcpu[PCPU_GET(cpuid)]; 1514 } else { 1515 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td); 1516 } 1517 1518 #else 1519 rq = &runq; 1520 td = runq_choose(&runq); 1521 #endif 1522 1523 if (td) { 1524 #ifdef SMP 1525 if (td == tdcpu) 1526 runq_length[PCPU_GET(cpuid)]--; 1527 #endif 1528 runq_remove(rq, td); 1529 td->td_flags |= TDF_DIDRUN; 1530 1531 KASSERT(td->td_flags & TDF_INMEM, 1532 ("sched_choose: thread swapped out")); 1533 return (td); 1534 } 1535 return (PCPU_GET(idlethread)); 1536 } 1537 1538 void 1539 sched_preempt(struct thread *td) 1540 { 1541 1542 SDT_PROBE2(sched, , , surrender, td, td->td_proc); 1543 if (td->td_critnest > 1) { 1544 td->td_owepreempt = 1; 1545 } else { 1546 thread_lock(td); 1547 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT); 1548 } 1549 } 1550 1551 void 1552 sched_userret_slowpath(struct thread *td) 1553 { 1554 1555 thread_lock(td); 1556 td->td_priority = td->td_user_pri; 1557 td->td_base_pri = td->td_user_pri; 1558 thread_unlock(td); 1559 } 1560 1561 void 1562 sched_bind(struct thread *td, int cpu) 1563 { 1564 #ifdef SMP 1565 struct td_sched *ts = td_get_sched(td); 1566 #endif 1567 1568 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 1569 KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 1570 1571 td->td_flags |= TDF_BOUND; 1572 #ifdef SMP 1573 ts->ts_runq = &runq_pcpu[cpu]; 1574 if (PCPU_GET(cpuid) == cpu) 1575 return; 1576 1577 mi_switch(SW_VOL); 1578 thread_lock(td); 1579 #endif 1580 } 1581 1582 void 1583 sched_unbind(struct thread* td) 1584 { 1585 THREAD_LOCK_ASSERT(td, MA_OWNED); 1586 KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 1587 td->td_flags &= ~TDF_BOUND; 1588 } 1589 1590 int 1591 sched_is_bound(struct thread *td) 1592 { 1593 THREAD_LOCK_ASSERT(td, MA_OWNED); 1594 return (td->td_flags & TDF_BOUND); 1595 } 1596 1597 void 1598 sched_relinquish(struct thread *td) 1599 { 1600 thread_lock(td); 1601 mi_switch(SW_VOL | SWT_RELINQUISH); 1602 } 1603 1604 int 1605 sched_load(void) 1606 { 1607 return (sched_tdcnt); 1608 } 1609 1610 int 1611 sched_sizeof_proc(void) 1612 { 1613 return (sizeof(struct proc)); 1614 } 1615 1616 int 1617 sched_sizeof_thread(void) 1618 { 1619 return (sizeof(struct thread) + sizeof(struct td_sched)); 1620 } 1621 1622 fixpt_t 1623 sched_pctcpu(struct thread *td) 1624 { 1625 struct td_sched *ts; 1626 1627 THREAD_LOCK_ASSERT(td, MA_OWNED); 1628 ts = td_get_sched(td); 1629 return (ts->ts_pctcpu); 1630 } 1631 1632 #ifdef RACCT 1633 /* 1634 * Calculates the contribution to the thread cpu usage for the latest 1635 * (unfinished) second. 1636 */ 1637 fixpt_t 1638 sched_pctcpu_delta(struct thread *td) 1639 { 1640 struct td_sched *ts; 1641 fixpt_t delta; 1642 int realstathz; 1643 1644 THREAD_LOCK_ASSERT(td, MA_OWNED); 1645 ts = td_get_sched(td); 1646 delta = 0; 1647 realstathz = stathz ? stathz : hz; 1648 if (ts->ts_cpticks != 0) { 1649 #if (FSHIFT >= CCPU_SHIFT) 1650 delta = (realstathz == 100) 1651 ? ((fixpt_t) ts->ts_cpticks) << 1652 (FSHIFT - CCPU_SHIFT) : 1653 100 * (((fixpt_t) ts->ts_cpticks) 1654 << (FSHIFT - CCPU_SHIFT)) / realstathz; 1655 #else 1656 delta = ((FSCALE - ccpu) * 1657 (ts->ts_cpticks * 1658 FSCALE / realstathz)) >> FSHIFT; 1659 #endif 1660 } 1661 1662 return (delta); 1663 } 1664 #endif 1665 1666 u_int 1667 sched_estcpu(struct thread *td) 1668 { 1669 1670 return (td_get_sched(td)->ts_estcpu); 1671 } 1672 1673 /* 1674 * The actual idle process. 1675 */ 1676 void 1677 sched_idletd(void *dummy) 1678 { 1679 struct pcpuidlestat *stat; 1680 1681 THREAD_NO_SLEEPING(); 1682 stat = DPCPU_PTR(idlestat); 1683 for (;;) { 1684 mtx_assert(&Giant, MA_NOTOWNED); 1685 1686 while (sched_runnable() == 0) { 1687 cpu_idle(stat->idlecalls + stat->oldidlecalls > 64); 1688 stat->idlecalls++; 1689 } 1690 1691 mtx_lock_spin(&sched_lock); 1692 mi_switch(SW_VOL | SWT_IDLE); 1693 } 1694 } 1695 1696 static void 1697 sched_throw_tail(struct thread *td) 1698 { 1699 1700 mtx_assert(&sched_lock, MA_OWNED); 1701 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 1702 cpu_throw(td, choosethread()); /* doesn't return */ 1703 } 1704 1705 /* 1706 * A CPU is entering for the first time. 1707 */ 1708 void 1709 sched_ap_entry(void) 1710 { 1711 1712 /* 1713 * Correct spinlock nesting. The idle thread context that we are 1714 * borrowing was created so that it would start out with a single 1715 * spin lock (sched_lock) held in fork_trampoline(). Since we've 1716 * explicitly acquired locks in this function, the nesting count 1717 * is now 2 rather than 1. Since we are nested, calling 1718 * spinlock_exit() will simply adjust the counts without allowing 1719 * spin lock using code to interrupt us. 1720 */ 1721 mtx_lock_spin(&sched_lock); 1722 spinlock_exit(); 1723 PCPU_SET(switchtime, cpu_ticks()); 1724 PCPU_SET(switchticks, ticks); 1725 1726 sched_throw_tail(NULL); 1727 } 1728 1729 /* 1730 * A thread is exiting. 1731 */ 1732 void 1733 sched_throw(struct thread *td) 1734 { 1735 1736 MPASS(td != NULL); 1737 MPASS(td->td_lock == &sched_lock); 1738 1739 lock_profile_release_lock(&sched_lock.lock_object, true); 1740 td->td_lastcpu = td->td_oncpu; 1741 td->td_oncpu = NOCPU; 1742 1743 sched_throw_tail(td); 1744 } 1745 1746 void 1747 sched_fork_exit(struct thread *td) 1748 { 1749 1750 /* 1751 * Finish setting up thread glue so that it begins execution in a 1752 * non-nested critical section with sched_lock held but not recursed. 1753 */ 1754 td->td_oncpu = PCPU_GET(cpuid); 1755 sched_lock.mtx_lock = (uintptr_t)td; 1756 lock_profile_obtain_lock_success(&sched_lock.lock_object, true, 1757 0, 0, __FILE__, __LINE__); 1758 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED); 1759 1760 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running", 1761 "prio:%d", td->td_priority); 1762 SDT_PROBE0(sched, , , on__cpu); 1763 } 1764 1765 char * 1766 sched_tdname(struct thread *td) 1767 { 1768 #ifdef KTR 1769 struct td_sched *ts; 1770 1771 ts = td_get_sched(td); 1772 if (ts->ts_name[0] == '\0') 1773 snprintf(ts->ts_name, sizeof(ts->ts_name), 1774 "%s tid %d", td->td_name, td->td_tid); 1775 return (ts->ts_name); 1776 #else 1777 return (td->td_name); 1778 #endif 1779 } 1780 1781 #ifdef KTR 1782 void 1783 sched_clear_tdname(struct thread *td) 1784 { 1785 struct td_sched *ts; 1786 1787 ts = td_get_sched(td); 1788 ts->ts_name[0] = '\0'; 1789 } 1790 #endif 1791 1792 void 1793 sched_affinity(struct thread *td) 1794 { 1795 #ifdef SMP 1796 struct td_sched *ts; 1797 int cpu; 1798 1799 THREAD_LOCK_ASSERT(td, MA_OWNED); 1800 1801 /* 1802 * Set the TSF_AFFINITY flag if there is at least one CPU this 1803 * thread can't run on. 1804 */ 1805 ts = td_get_sched(td); 1806 ts->ts_flags &= ~TSF_AFFINITY; 1807 CPU_FOREACH(cpu) { 1808 if (!THREAD_CAN_SCHED(td, cpu)) { 1809 ts->ts_flags |= TSF_AFFINITY; 1810 break; 1811 } 1812 } 1813 1814 /* 1815 * If this thread can run on all CPUs, nothing else to do. 1816 */ 1817 if (!(ts->ts_flags & TSF_AFFINITY)) 1818 return; 1819 1820 /* Pinned threads and bound threads should be left alone. */ 1821 if (td->td_pinned != 0 || td->td_flags & TDF_BOUND) 1822 return; 1823 1824 switch (TD_GET_STATE(td)) { 1825 case TDS_RUNQ: 1826 /* 1827 * If we are on a per-CPU runqueue that is in the set, 1828 * then nothing needs to be done. 1829 */ 1830 if (ts->ts_runq != &runq && 1831 THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu)) 1832 return; 1833 1834 /* Put this thread on a valid per-CPU runqueue. */ 1835 sched_rem(td); 1836 sched_add(td, SRQ_HOLDTD | SRQ_BORING); 1837 break; 1838 case TDS_RUNNING: 1839 /* 1840 * See if our current CPU is in the set. If not, force a 1841 * context switch. 1842 */ 1843 if (THREAD_CAN_SCHED(td, td->td_oncpu)) 1844 return; 1845 1846 td->td_flags |= TDF_NEEDRESCHED; 1847 if (td != curthread) 1848 ipi_cpu(cpu, IPI_AST); 1849 break; 1850 default: 1851 break; 1852 } 1853 #endif 1854 } 1855