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/umtx.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, ""); 441 442 /* 443 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 444 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 445 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 446 * 447 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 448 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 449 * 450 * If you don't want to bother with the faster/more-accurate formula, you 451 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 452 * (more general) method of calculating the %age of CPU used by a process. 453 */ 454 #define CCPU_SHIFT 11 455 456 /* 457 * Recompute process priorities, every hz ticks. 458 * MP-safe, called without the Giant mutex. 459 */ 460 /* ARGSUSED */ 461 static void 462 schedcpu(void) 463 { 464 fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 465 struct thread *td; 466 struct proc *p; 467 struct td_sched *ts; 468 int awake; 469 470 sx_slock(&allproc_lock); 471 FOREACH_PROC_IN_SYSTEM(p) { 472 PROC_LOCK(p); 473 if (p->p_state == PRS_NEW) { 474 PROC_UNLOCK(p); 475 continue; 476 } 477 FOREACH_THREAD_IN_PROC(p, td) { 478 awake = 0; 479 ts = td_get_sched(td); 480 thread_lock(td); 481 /* 482 * Increment sleep time (if sleeping). We 483 * ignore overflow, as above. 484 */ 485 /* 486 * The td_sched slptimes are not touched in wakeup 487 * because the thread may not HAVE everything in 488 * memory? XXX I think this is out of date. 489 */ 490 if (TD_ON_RUNQ(td)) { 491 awake = 1; 492 td->td_flags &= ~TDF_DIDRUN; 493 } else if (TD_IS_RUNNING(td)) { 494 awake = 1; 495 /* Do not clear TDF_DIDRUN */ 496 } else if (td->td_flags & TDF_DIDRUN) { 497 awake = 1; 498 td->td_flags &= ~TDF_DIDRUN; 499 } 500 501 /* 502 * ts_pctcpu is only for ps and ttyinfo(). 503 */ 504 ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT; 505 /* 506 * If the td_sched has been idle the entire second, 507 * stop recalculating its priority until 508 * it wakes up. 509 */ 510 if (ts->ts_cpticks != 0) { 511 #if (FSHIFT >= CCPU_SHIFT) 512 ts->ts_pctcpu += (realstathz == 100) 513 ? ((fixpt_t) ts->ts_cpticks) << 514 (FSHIFT - CCPU_SHIFT) : 515 100 * (((fixpt_t) ts->ts_cpticks) 516 << (FSHIFT - CCPU_SHIFT)) / realstathz; 517 #else 518 ts->ts_pctcpu += ((FSCALE - ccpu) * 519 (ts->ts_cpticks * 520 FSCALE / realstathz)) >> FSHIFT; 521 #endif 522 ts->ts_cpticks = 0; 523 } 524 /* 525 * If there are ANY running threads in this process, 526 * then don't count it as sleeping. 527 * XXX: this is broken. 528 */ 529 if (awake) { 530 if (ts->ts_slptime > 1) { 531 /* 532 * In an ideal world, this should not 533 * happen, because whoever woke us 534 * up from the long sleep should have 535 * unwound the slptime and reset our 536 * priority before we run at the stale 537 * priority. Should KASSERT at some 538 * point when all the cases are fixed. 539 */ 540 updatepri(td); 541 } 542 ts->ts_slptime = 0; 543 } else 544 ts->ts_slptime++; 545 if (ts->ts_slptime > 1) { 546 thread_unlock(td); 547 continue; 548 } 549 ts->ts_estcpu = decay_cpu(loadfac, ts->ts_estcpu); 550 resetpriority(td); 551 resetpriority_thread(td); 552 thread_unlock(td); 553 } 554 PROC_UNLOCK(p); 555 } 556 sx_sunlock(&allproc_lock); 557 } 558 559 /* 560 * Main loop for a kthread that executes schedcpu once a second. 561 */ 562 static void 563 schedcpu_thread(void) 564 { 565 566 for (;;) { 567 schedcpu(); 568 pause("-", hz); 569 } 570 } 571 572 /* 573 * Recalculate the priority of a process after it has slept for a while. 574 * For all load averages >= 1 and max ts_estcpu of 255, sleeping for at 575 * least six times the loadfactor will decay ts_estcpu to zero. 576 */ 577 static void 578 updatepri(struct thread *td) 579 { 580 struct td_sched *ts; 581 fixpt_t loadfac; 582 unsigned int newcpu; 583 584 ts = td_get_sched(td); 585 loadfac = loadfactor(averunnable.ldavg[0]); 586 if (ts->ts_slptime > 5 * loadfac) 587 ts->ts_estcpu = 0; 588 else { 589 newcpu = ts->ts_estcpu; 590 ts->ts_slptime--; /* was incremented in schedcpu() */ 591 while (newcpu && --ts->ts_slptime) 592 newcpu = decay_cpu(loadfac, newcpu); 593 ts->ts_estcpu = newcpu; 594 } 595 } 596 597 /* 598 * Compute the priority of a process when running in user mode. 599 * Arrange to reschedule if the resulting priority is better 600 * than that of the current process. 601 */ 602 static void 603 resetpriority(struct thread *td) 604 { 605 u_int newpriority; 606 607 if (td->td_pri_class != PRI_TIMESHARE) 608 return; 609 newpriority = PUSER + 610 td_get_sched(td)->ts_estcpu / INVERSE_ESTCPU_WEIGHT + 611 NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN); 612 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 613 PRI_MAX_TIMESHARE); 614 sched_user_prio(td, newpriority); 615 } 616 617 /* 618 * Update the thread's priority when the associated process's user 619 * priority changes. 620 */ 621 static void 622 resetpriority_thread(struct thread *td) 623 { 624 625 /* Only change threads with a time sharing user priority. */ 626 if (td->td_priority < PRI_MIN_TIMESHARE || 627 td->td_priority > PRI_MAX_TIMESHARE) 628 return; 629 630 /* XXX the whole needresched thing is broken, but not silly. */ 631 maybe_resched(td); 632 633 sched_prio(td, td->td_user_pri); 634 } 635 636 /* ARGSUSED */ 637 static void 638 sched_setup(void *dummy) 639 { 640 641 setup_runqs(); 642 643 /* Account for thread0. */ 644 sched_load_add(); 645 } 646 647 /* 648 * This routine determines time constants after stathz and hz are setup. 649 */ 650 static void 651 sched_initticks(void *dummy) 652 { 653 654 realstathz = stathz ? stathz : hz; 655 sched_slice = realstathz / 10; /* ~100ms */ 656 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 657 realstathz); 658 } 659 660 /* External interfaces start here */ 661 662 /* 663 * Very early in the boot some setup of scheduler-specific 664 * parts of proc0 and of some scheduler resources needs to be done. 665 * Called from: 666 * proc0_init() 667 */ 668 void 669 schedinit(void) 670 { 671 672 /* 673 * Set up the scheduler specific parts of thread0. 674 */ 675 thread0.td_lock = &sched_lock; 676 td_get_sched(&thread0)->ts_slice = sched_slice; 677 mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN); 678 } 679 680 int 681 sched_runnable(void) 682 { 683 #ifdef SMP 684 return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]); 685 #else 686 return runq_check(&runq); 687 #endif 688 } 689 690 int 691 sched_rr_interval(void) 692 { 693 694 /* Convert sched_slice from stathz to hz. */ 695 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz)); 696 } 697 698 /* 699 * We adjust the priority of the current process. The priority of a 700 * process gets worse as it accumulates CPU time. The cpu usage 701 * estimator (ts_estcpu) is increased here. resetpriority() will 702 * compute a different priority each time ts_estcpu increases by 703 * INVERSE_ESTCPU_WEIGHT (until PRI_MAX_TIMESHARE is reached). The 704 * cpu usage estimator ramps up quite quickly when the process is 705 * running (linearly), and decays away exponentially, at a rate which 706 * is proportionally slower when the system is busy. The basic 707 * principle is that the system will 90% forget that the process used 708 * a lot of CPU time in 5 * loadav seconds. This causes the system to 709 * favor processes which haven't run much recently, and to round-robin 710 * among other processes. 711 */ 712 static void 713 sched_clock_tick(struct thread *td) 714 { 715 struct pcpuidlestat *stat; 716 struct td_sched *ts; 717 718 THREAD_LOCK_ASSERT(td, MA_OWNED); 719 ts = td_get_sched(td); 720 721 ts->ts_cpticks++; 722 ts->ts_estcpu = ESTCPULIM(ts->ts_estcpu + 1); 723 if ((ts->ts_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 724 resetpriority(td); 725 resetpriority_thread(td); 726 } 727 728 /* 729 * Force a context switch if the current thread has used up a full 730 * time slice (default is 100ms). 731 */ 732 if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) { 733 ts->ts_slice = sched_slice; 734 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND; 735 } 736 737 stat = DPCPU_PTR(idlestat); 738 stat->oldidlecalls = stat->idlecalls; 739 stat->idlecalls = 0; 740 } 741 742 void 743 sched_clock(struct thread *td, int cnt) 744 { 745 746 for ( ; cnt > 0; cnt--) 747 sched_clock_tick(td); 748 } 749 750 /* 751 * Charge child's scheduling CPU usage to parent. 752 */ 753 void 754 sched_exit(struct proc *p, struct thread *td) 755 { 756 757 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit", 758 "prio:%d", td->td_priority); 759 760 PROC_LOCK_ASSERT(p, MA_OWNED); 761 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td); 762 } 763 764 void 765 sched_exit_thread(struct thread *td, struct thread *child) 766 { 767 768 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit", 769 "prio:%d", child->td_priority); 770 thread_lock(td); 771 td_get_sched(td)->ts_estcpu = ESTCPULIM(td_get_sched(td)->ts_estcpu + 772 td_get_sched(child)->ts_estcpu); 773 thread_unlock(td); 774 thread_lock(child); 775 if ((child->td_flags & TDF_NOLOAD) == 0) 776 sched_load_rem(); 777 thread_unlock(child); 778 } 779 780 void 781 sched_fork(struct thread *td, struct thread *childtd) 782 { 783 sched_fork_thread(td, childtd); 784 } 785 786 void 787 sched_fork_thread(struct thread *td, struct thread *childtd) 788 { 789 struct td_sched *ts, *tsc; 790 791 childtd->td_oncpu = NOCPU; 792 childtd->td_lastcpu = NOCPU; 793 childtd->td_lock = &sched_lock; 794 childtd->td_cpuset = cpuset_ref(td->td_cpuset); 795 childtd->td_domain.dr_policy = td->td_cpuset->cs_domain; 796 childtd->td_priority = childtd->td_base_pri; 797 ts = td_get_sched(childtd); 798 bzero(ts, sizeof(*ts)); 799 tsc = td_get_sched(td); 800 ts->ts_estcpu = tsc->ts_estcpu; 801 ts->ts_flags |= (tsc->ts_flags & TSF_AFFINITY); 802 ts->ts_slice = 1; 803 } 804 805 void 806 sched_nice(struct proc *p, int nice) 807 { 808 struct thread *td; 809 810 PROC_LOCK_ASSERT(p, MA_OWNED); 811 p->p_nice = nice; 812 FOREACH_THREAD_IN_PROC(p, td) { 813 thread_lock(td); 814 resetpriority(td); 815 resetpriority_thread(td); 816 thread_unlock(td); 817 } 818 } 819 820 void 821 sched_class(struct thread *td, int class) 822 { 823 THREAD_LOCK_ASSERT(td, MA_OWNED); 824 td->td_pri_class = class; 825 } 826 827 /* 828 * Adjust the priority of a thread. 829 */ 830 static void 831 sched_priority(struct thread *td, u_char prio) 832 { 833 834 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change", 835 "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED, 836 sched_tdname(curthread)); 837 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio); 838 if (td != curthread && prio > td->td_priority) { 839 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 840 "lend prio", "prio:%d", td->td_priority, "new prio:%d", 841 prio, KTR_ATTR_LINKED, sched_tdname(td)); 842 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio, 843 curthread); 844 } 845 THREAD_LOCK_ASSERT(td, MA_OWNED); 846 if (td->td_priority == prio) 847 return; 848 td->td_priority = prio; 849 if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) { 850 sched_rem(td); 851 sched_add(td, SRQ_BORING | SRQ_HOLDTD); 852 } 853 } 854 855 /* 856 * Update a thread's priority when it is lent another thread's 857 * priority. 858 */ 859 void 860 sched_lend_prio(struct thread *td, u_char prio) 861 { 862 863 td->td_flags |= TDF_BORROWING; 864 sched_priority(td, prio); 865 } 866 867 /* 868 * Restore a thread's priority when priority propagation is 869 * over. The prio argument is the minimum priority the thread 870 * needs to have to satisfy other possible priority lending 871 * requests. If the thread's regulary priority is less 872 * important than prio the thread will keep a priority boost 873 * of prio. 874 */ 875 void 876 sched_unlend_prio(struct thread *td, u_char prio) 877 { 878 u_char base_pri; 879 880 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 881 td->td_base_pri <= PRI_MAX_TIMESHARE) 882 base_pri = td->td_user_pri; 883 else 884 base_pri = td->td_base_pri; 885 if (prio >= base_pri) { 886 td->td_flags &= ~TDF_BORROWING; 887 sched_prio(td, base_pri); 888 } else 889 sched_lend_prio(td, prio); 890 } 891 892 void 893 sched_prio(struct thread *td, u_char prio) 894 { 895 u_char oldprio; 896 897 /* First, update the base priority. */ 898 td->td_base_pri = prio; 899 900 /* 901 * If the thread is borrowing another thread's priority, don't ever 902 * lower the priority. 903 */ 904 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 905 return; 906 907 /* Change the real priority. */ 908 oldprio = td->td_priority; 909 sched_priority(td, prio); 910 911 /* 912 * If the thread is on a turnstile, then let the turnstile update 913 * its state. 914 */ 915 if (TD_ON_LOCK(td) && oldprio != prio) 916 turnstile_adjust(td, oldprio); 917 } 918 919 void 920 sched_user_prio(struct thread *td, u_char prio) 921 { 922 923 THREAD_LOCK_ASSERT(td, MA_OWNED); 924 td->td_base_user_pri = prio; 925 if (td->td_lend_user_pri <= prio) 926 return; 927 td->td_user_pri = prio; 928 } 929 930 void 931 sched_lend_user_prio(struct thread *td, u_char prio) 932 { 933 934 THREAD_LOCK_ASSERT(td, MA_OWNED); 935 td->td_lend_user_pri = prio; 936 td->td_user_pri = min(prio, td->td_base_user_pri); 937 if (td->td_priority > td->td_user_pri) 938 sched_prio(td, td->td_user_pri); 939 else if (td->td_priority != td->td_user_pri) 940 td->td_flags |= TDF_NEEDRESCHED; 941 } 942 943 /* 944 * Like the above but first check if there is anything to do. 945 */ 946 void 947 sched_lend_user_prio_cond(struct thread *td, u_char prio) 948 { 949 950 if (td->td_lend_user_pri != prio) 951 goto lend; 952 if (td->td_user_pri != min(prio, td->td_base_user_pri)) 953 goto lend; 954 if (td->td_priority >= td->td_user_pri) 955 goto lend; 956 return; 957 958 lend: 959 thread_lock(td); 960 sched_lend_user_prio(td, prio); 961 thread_unlock(td); 962 } 963 964 void 965 sched_sleep(struct thread *td, int pri) 966 { 967 968 THREAD_LOCK_ASSERT(td, MA_OWNED); 969 td->td_slptick = ticks; 970 td_get_sched(td)->ts_slptime = 0; 971 if (pri != 0 && PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 972 sched_prio(td, pri); 973 if (TD_IS_SUSPENDED(td) || pri >= PSOCK) 974 td->td_flags |= TDF_CANSWAP; 975 } 976 977 void 978 sched_switch(struct thread *td, int flags) 979 { 980 struct thread *newtd; 981 struct mtx *tmtx; 982 struct td_sched *ts; 983 struct proc *p; 984 int preempted; 985 986 tmtx = &sched_lock; 987 ts = td_get_sched(td); 988 p = td->td_proc; 989 990 THREAD_LOCK_ASSERT(td, MA_OWNED); 991 992 td->td_lastcpu = td->td_oncpu; 993 preempted = (td->td_flags & TDF_SLICEEND) == 0 && 994 (flags & SW_PREEMPT) != 0; 995 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND); 996 td->td_owepreempt = 0; 997 td->td_oncpu = NOCPU; 998 999 /* 1000 * At the last moment, if this thread is still marked RUNNING, 1001 * then put it back on the run queue as it has not been suspended 1002 * or stopped or any thing else similar. We never put the idle 1003 * threads on the run queue, however. 1004 */ 1005 if (td->td_flags & TDF_IDLETD) { 1006 TD_SET_CAN_RUN(td); 1007 #ifdef SMP 1008 CPU_CLR(PCPU_GET(cpuid), &idle_cpus_mask); 1009 #endif 1010 } else { 1011 if (TD_IS_RUNNING(td)) { 1012 /* Put us back on the run queue. */ 1013 sched_add(td, preempted ? 1014 SRQ_HOLDTD|SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1015 SRQ_HOLDTD|SRQ_OURSELF|SRQ_YIELDING); 1016 } 1017 } 1018 1019 /* 1020 * Switch to the sched lock to fix things up and pick 1021 * a new thread. Block the td_lock in order to avoid 1022 * breaking the critical path. 1023 */ 1024 if (td->td_lock != &sched_lock) { 1025 mtx_lock_spin(&sched_lock); 1026 tmtx = thread_lock_block(td); 1027 mtx_unlock_spin(tmtx); 1028 } 1029 1030 if ((td->td_flags & TDF_NOLOAD) == 0) 1031 sched_load_rem(); 1032 1033 newtd = choosethread(); 1034 MPASS(newtd->td_lock == &sched_lock); 1035 1036 #if (KTR_COMPILE & KTR_SCHED) != 0 1037 if (TD_IS_IDLETHREAD(td)) 1038 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle", 1039 "prio:%d", td->td_priority); 1040 else 1041 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td), 1042 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg, 1043 "lockname:\"%s\"", td->td_lockname); 1044 #endif 1045 1046 if (td != newtd) { 1047 #ifdef HWPMC_HOOKS 1048 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1049 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1050 #endif 1051 1052 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc); 1053 1054 /* I feel sleepy */ 1055 lock_profile_release_lock(&sched_lock.lock_object); 1056 #ifdef KDTRACE_HOOKS 1057 /* 1058 * If DTrace has set the active vtime enum to anything 1059 * other than INACTIVE (0), then it should have set the 1060 * function to call. 1061 */ 1062 if (dtrace_vtime_active) 1063 (*dtrace_vtime_switch_func)(newtd); 1064 #endif 1065 1066 cpu_switch(td, newtd, tmtx); 1067 lock_profile_obtain_lock_success(&sched_lock.lock_object, 1068 0, 0, __FILE__, __LINE__); 1069 /* 1070 * Where am I? What year is it? 1071 * We are in the same thread that went to sleep above, 1072 * but any amount of time may have passed. All our context 1073 * will still be available as will local variables. 1074 * PCPU values however may have changed as we may have 1075 * changed CPU so don't trust cached values of them. 1076 * New threads will go to fork_exit() instead of here 1077 * so if you change things here you may need to change 1078 * things there too. 1079 * 1080 * If the thread above was exiting it will never wake 1081 * up again here, so either it has saved everything it 1082 * needed to, or the thread_wait() or wait() will 1083 * need to reap it. 1084 */ 1085 1086 SDT_PROBE0(sched, , , on__cpu); 1087 #ifdef HWPMC_HOOKS 1088 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1089 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1090 #endif 1091 } else { 1092 td->td_lock = &sched_lock; 1093 SDT_PROBE0(sched, , , remain__cpu); 1094 } 1095 1096 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running", 1097 "prio:%d", td->td_priority); 1098 1099 #ifdef SMP 1100 if (td->td_flags & TDF_IDLETD) 1101 CPU_SET(PCPU_GET(cpuid), &idle_cpus_mask); 1102 #endif 1103 sched_lock.mtx_lock = (uintptr_t)td; 1104 td->td_oncpu = PCPU_GET(cpuid); 1105 spinlock_enter(); 1106 mtx_unlock_spin(&sched_lock); 1107 } 1108 1109 void 1110 sched_wakeup(struct thread *td, int srqflags) 1111 { 1112 struct td_sched *ts; 1113 1114 THREAD_LOCK_ASSERT(td, MA_OWNED); 1115 ts = td_get_sched(td); 1116 td->td_flags &= ~TDF_CANSWAP; 1117 if (ts->ts_slptime > 1) { 1118 updatepri(td); 1119 resetpriority(td); 1120 } 1121 td->td_slptick = 0; 1122 ts->ts_slptime = 0; 1123 ts->ts_slice = sched_slice; 1124 sched_add(td, srqflags); 1125 } 1126 1127 #ifdef SMP 1128 static int 1129 forward_wakeup(int cpunum) 1130 { 1131 struct pcpu *pc; 1132 cpuset_t dontuse, map, map2; 1133 u_int id, me; 1134 int iscpuset; 1135 1136 mtx_assert(&sched_lock, MA_OWNED); 1137 1138 CTR0(KTR_RUNQ, "forward_wakeup()"); 1139 1140 if ((!forward_wakeup_enabled) || 1141 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0)) 1142 return (0); 1143 if (!smp_started || KERNEL_PANICKED()) 1144 return (0); 1145 1146 forward_wakeups_requested++; 1147 1148 /* 1149 * Check the idle mask we received against what we calculated 1150 * before in the old version. 1151 */ 1152 me = PCPU_GET(cpuid); 1153 1154 /* Don't bother if we should be doing it ourself. */ 1155 if (CPU_ISSET(me, &idle_cpus_mask) && 1156 (cpunum == NOCPU || me == cpunum)) 1157 return (0); 1158 1159 CPU_SETOF(me, &dontuse); 1160 CPU_OR(&dontuse, &stopped_cpus); 1161 CPU_OR(&dontuse, &hlt_cpus_mask); 1162 CPU_ZERO(&map2); 1163 if (forward_wakeup_use_loop) { 1164 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { 1165 id = pc->pc_cpuid; 1166 if (!CPU_ISSET(id, &dontuse) && 1167 pc->pc_curthread == pc->pc_idlethread) { 1168 CPU_SET(id, &map2); 1169 } 1170 } 1171 } 1172 1173 if (forward_wakeup_use_mask) { 1174 map = idle_cpus_mask; 1175 CPU_ANDNOT(&map, &dontuse); 1176 1177 /* If they are both on, compare and use loop if different. */ 1178 if (forward_wakeup_use_loop) { 1179 if (CPU_CMP(&map, &map2)) { 1180 printf("map != map2, loop method preferred\n"); 1181 map = map2; 1182 } 1183 } 1184 } else { 1185 map = map2; 1186 } 1187 1188 /* If we only allow a specific CPU, then mask off all the others. */ 1189 if (cpunum != NOCPU) { 1190 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum.")); 1191 iscpuset = CPU_ISSET(cpunum, &map); 1192 if (iscpuset == 0) 1193 CPU_ZERO(&map); 1194 else 1195 CPU_SETOF(cpunum, &map); 1196 } 1197 if (!CPU_EMPTY(&map)) { 1198 forward_wakeups_delivered++; 1199 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { 1200 id = pc->pc_cpuid; 1201 if (!CPU_ISSET(id, &map)) 1202 continue; 1203 if (cpu_idle_wakeup(pc->pc_cpuid)) 1204 CPU_CLR(id, &map); 1205 } 1206 if (!CPU_EMPTY(&map)) 1207 ipi_selected(map, IPI_AST); 1208 return (1); 1209 } 1210 if (cpunum == NOCPU) 1211 printf("forward_wakeup: Idle processor not found\n"); 1212 return (0); 1213 } 1214 1215 static void 1216 kick_other_cpu(int pri, int cpuid) 1217 { 1218 struct pcpu *pcpu; 1219 int cpri; 1220 1221 pcpu = pcpu_find(cpuid); 1222 if (CPU_ISSET(cpuid, &idle_cpus_mask)) { 1223 forward_wakeups_delivered++; 1224 if (!cpu_idle_wakeup(cpuid)) 1225 ipi_cpu(cpuid, IPI_AST); 1226 return; 1227 } 1228 1229 cpri = pcpu->pc_curthread->td_priority; 1230 if (pri >= cpri) 1231 return; 1232 1233 #if defined(IPI_PREEMPTION) && defined(PREEMPTION) 1234 #if !defined(FULL_PREEMPTION) 1235 if (pri <= PRI_MAX_ITHD) 1236 #endif /* ! FULL_PREEMPTION */ 1237 { 1238 ipi_cpu(cpuid, IPI_PREEMPT); 1239 return; 1240 } 1241 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */ 1242 1243 pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED; 1244 ipi_cpu(cpuid, IPI_AST); 1245 return; 1246 } 1247 #endif /* SMP */ 1248 1249 #ifdef SMP 1250 static int 1251 sched_pickcpu(struct thread *td) 1252 { 1253 int best, cpu; 1254 1255 mtx_assert(&sched_lock, MA_OWNED); 1256 1257 if (td->td_lastcpu != NOCPU && THREAD_CAN_SCHED(td, td->td_lastcpu)) 1258 best = td->td_lastcpu; 1259 else 1260 best = NOCPU; 1261 CPU_FOREACH(cpu) { 1262 if (!THREAD_CAN_SCHED(td, cpu)) 1263 continue; 1264 1265 if (best == NOCPU) 1266 best = cpu; 1267 else if (runq_length[cpu] < runq_length[best]) 1268 best = cpu; 1269 } 1270 KASSERT(best != NOCPU, ("no valid CPUs")); 1271 1272 return (best); 1273 } 1274 #endif 1275 1276 void 1277 sched_add(struct thread *td, int flags) 1278 #ifdef SMP 1279 { 1280 cpuset_t tidlemsk; 1281 struct td_sched *ts; 1282 u_int cpu, cpuid; 1283 int forwarded = 0; 1284 int single_cpu = 0; 1285 1286 ts = td_get_sched(td); 1287 THREAD_LOCK_ASSERT(td, MA_OWNED); 1288 KASSERT((td->td_inhibitors == 0), 1289 ("sched_add: trying to run inhibited thread")); 1290 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 1291 ("sched_add: bad thread state")); 1292 KASSERT(td->td_flags & TDF_INMEM, 1293 ("sched_add: thread swapped out")); 1294 1295 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 1296 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 1297 sched_tdname(curthread)); 1298 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 1299 KTR_ATTR_LINKED, sched_tdname(td)); 1300 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 1301 flags & SRQ_PREEMPTED); 1302 1303 /* 1304 * Now that the thread is moving to the run-queue, set the lock 1305 * to the scheduler's lock. 1306 */ 1307 if (td->td_lock != &sched_lock) { 1308 mtx_lock_spin(&sched_lock); 1309 if ((flags & SRQ_HOLD) != 0) 1310 td->td_lock = &sched_lock; 1311 else 1312 thread_lock_set(td, &sched_lock); 1313 1314 } 1315 TD_SET_RUNQ(td); 1316 1317 /* 1318 * If SMP is started and the thread is pinned or otherwise limited to 1319 * a specific set of CPUs, queue the thread to a per-CPU run queue. 1320 * Otherwise, queue the thread to the global run queue. 1321 * 1322 * If SMP has not yet been started we must use the global run queue 1323 * as per-CPU state may not be initialized yet and we may crash if we 1324 * try to access the per-CPU run queues. 1325 */ 1326 if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND || 1327 ts->ts_flags & TSF_AFFINITY)) { 1328 if (td->td_pinned != 0) 1329 cpu = td->td_lastcpu; 1330 else if (td->td_flags & TDF_BOUND) { 1331 /* Find CPU from bound runq. */ 1332 KASSERT(SKE_RUNQ_PCPU(ts), 1333 ("sched_add: bound td_sched not on cpu runq")); 1334 cpu = ts->ts_runq - &runq_pcpu[0]; 1335 } else 1336 /* Find a valid CPU for our cpuset */ 1337 cpu = sched_pickcpu(td); 1338 ts->ts_runq = &runq_pcpu[cpu]; 1339 single_cpu = 1; 1340 CTR3(KTR_RUNQ, 1341 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, 1342 cpu); 1343 } else { 1344 CTR2(KTR_RUNQ, 1345 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts, 1346 td); 1347 cpu = NOCPU; 1348 ts->ts_runq = &runq; 1349 } 1350 1351 if ((td->td_flags & TDF_NOLOAD) == 0) 1352 sched_load_add(); 1353 runq_add(ts->ts_runq, td, flags); 1354 if (cpu != NOCPU) 1355 runq_length[cpu]++; 1356 1357 cpuid = PCPU_GET(cpuid); 1358 if (single_cpu && cpu != cpuid) { 1359 kick_other_cpu(td->td_priority, cpu); 1360 } else { 1361 if (!single_cpu) { 1362 tidlemsk = idle_cpus_mask; 1363 CPU_ANDNOT(&tidlemsk, &hlt_cpus_mask); 1364 CPU_CLR(cpuid, &tidlemsk); 1365 1366 if (!CPU_ISSET(cpuid, &idle_cpus_mask) && 1367 ((flags & SRQ_INTR) == 0) && 1368 !CPU_EMPTY(&tidlemsk)) 1369 forwarded = forward_wakeup(cpu); 1370 } 1371 1372 if (!forwarded) { 1373 if (!maybe_preempt(td)) 1374 maybe_resched(td); 1375 } 1376 } 1377 if ((flags & SRQ_HOLDTD) == 0) 1378 thread_unlock(td); 1379 } 1380 #else /* SMP */ 1381 { 1382 struct td_sched *ts; 1383 1384 ts = td_get_sched(td); 1385 THREAD_LOCK_ASSERT(td, MA_OWNED); 1386 KASSERT((td->td_inhibitors == 0), 1387 ("sched_add: trying to run inhibited thread")); 1388 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 1389 ("sched_add: bad thread state")); 1390 KASSERT(td->td_flags & TDF_INMEM, 1391 ("sched_add: thread swapped out")); 1392 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 1393 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 1394 sched_tdname(curthread)); 1395 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 1396 KTR_ATTR_LINKED, sched_tdname(td)); 1397 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 1398 flags & SRQ_PREEMPTED); 1399 1400 /* 1401 * Now that the thread is moving to the run-queue, set the lock 1402 * to the scheduler's lock. 1403 */ 1404 if (td->td_lock != &sched_lock) { 1405 mtx_lock_spin(&sched_lock); 1406 if ((flags & SRQ_HOLD) != 0) 1407 td->td_lock = &sched_lock; 1408 else 1409 thread_lock_set(td, &sched_lock); 1410 } 1411 TD_SET_RUNQ(td); 1412 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td); 1413 ts->ts_runq = &runq; 1414 1415 if ((td->td_flags & TDF_NOLOAD) == 0) 1416 sched_load_add(); 1417 runq_add(ts->ts_runq, td, flags); 1418 if (!maybe_preempt(td)) 1419 maybe_resched(td); 1420 if ((flags & SRQ_HOLDTD) == 0) 1421 thread_unlock(td); 1422 } 1423 #endif /* SMP */ 1424 1425 void 1426 sched_rem(struct thread *td) 1427 { 1428 struct td_sched *ts; 1429 1430 ts = td_get_sched(td); 1431 KASSERT(td->td_flags & TDF_INMEM, 1432 ("sched_rem: thread swapped out")); 1433 KASSERT(TD_ON_RUNQ(td), 1434 ("sched_rem: thread not on run queue")); 1435 mtx_assert(&sched_lock, MA_OWNED); 1436 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 1437 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 1438 sched_tdname(curthread)); 1439 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); 1440 1441 if ((td->td_flags & TDF_NOLOAD) == 0) 1442 sched_load_rem(); 1443 #ifdef SMP 1444 if (ts->ts_runq != &runq) 1445 runq_length[ts->ts_runq - runq_pcpu]--; 1446 #endif 1447 runq_remove(ts->ts_runq, td); 1448 TD_SET_CAN_RUN(td); 1449 } 1450 1451 /* 1452 * Select threads to run. Note that running threads still consume a 1453 * slot. 1454 */ 1455 struct thread * 1456 sched_choose(void) 1457 { 1458 struct thread *td; 1459 struct runq *rq; 1460 1461 mtx_assert(&sched_lock, MA_OWNED); 1462 #ifdef SMP 1463 struct thread *tdcpu; 1464 1465 rq = &runq; 1466 td = runq_choose_fuzz(&runq, runq_fuzz); 1467 tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]); 1468 1469 if (td == NULL || 1470 (tdcpu != NULL && 1471 tdcpu->td_priority < td->td_priority)) { 1472 CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu, 1473 PCPU_GET(cpuid)); 1474 td = tdcpu; 1475 rq = &runq_pcpu[PCPU_GET(cpuid)]; 1476 } else { 1477 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td); 1478 } 1479 1480 #else 1481 rq = &runq; 1482 td = runq_choose(&runq); 1483 #endif 1484 1485 if (td) { 1486 #ifdef SMP 1487 if (td == tdcpu) 1488 runq_length[PCPU_GET(cpuid)]--; 1489 #endif 1490 runq_remove(rq, td); 1491 td->td_flags |= TDF_DIDRUN; 1492 1493 KASSERT(td->td_flags & TDF_INMEM, 1494 ("sched_choose: thread swapped out")); 1495 return (td); 1496 } 1497 return (PCPU_GET(idlethread)); 1498 } 1499 1500 void 1501 sched_preempt(struct thread *td) 1502 { 1503 1504 SDT_PROBE2(sched, , , surrender, td, td->td_proc); 1505 if (td->td_critnest > 1) { 1506 td->td_owepreempt = 1; 1507 } else { 1508 thread_lock(td); 1509 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT); 1510 } 1511 } 1512 1513 void 1514 sched_userret_slowpath(struct thread *td) 1515 { 1516 1517 thread_lock(td); 1518 td->td_priority = td->td_user_pri; 1519 td->td_base_pri = td->td_user_pri; 1520 thread_unlock(td); 1521 } 1522 1523 void 1524 sched_bind(struct thread *td, int cpu) 1525 { 1526 struct td_sched *ts; 1527 1528 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 1529 KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 1530 1531 ts = td_get_sched(td); 1532 1533 td->td_flags |= TDF_BOUND; 1534 #ifdef SMP 1535 ts->ts_runq = &runq_pcpu[cpu]; 1536 if (PCPU_GET(cpuid) == cpu) 1537 return; 1538 1539 mi_switch(SW_VOL); 1540 thread_lock(td); 1541 #endif 1542 } 1543 1544 void 1545 sched_unbind(struct thread* td) 1546 { 1547 THREAD_LOCK_ASSERT(td, MA_OWNED); 1548 KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 1549 td->td_flags &= ~TDF_BOUND; 1550 } 1551 1552 int 1553 sched_is_bound(struct thread *td) 1554 { 1555 THREAD_LOCK_ASSERT(td, MA_OWNED); 1556 return (td->td_flags & TDF_BOUND); 1557 } 1558 1559 void 1560 sched_relinquish(struct thread *td) 1561 { 1562 thread_lock(td); 1563 mi_switch(SW_VOL | SWT_RELINQUISH); 1564 } 1565 1566 int 1567 sched_load(void) 1568 { 1569 return (sched_tdcnt); 1570 } 1571 1572 int 1573 sched_sizeof_proc(void) 1574 { 1575 return (sizeof(struct proc)); 1576 } 1577 1578 int 1579 sched_sizeof_thread(void) 1580 { 1581 return (sizeof(struct thread) + sizeof(struct td_sched)); 1582 } 1583 1584 fixpt_t 1585 sched_pctcpu(struct thread *td) 1586 { 1587 struct td_sched *ts; 1588 1589 THREAD_LOCK_ASSERT(td, MA_OWNED); 1590 ts = td_get_sched(td); 1591 return (ts->ts_pctcpu); 1592 } 1593 1594 #ifdef RACCT 1595 /* 1596 * Calculates the contribution to the thread cpu usage for the latest 1597 * (unfinished) second. 1598 */ 1599 fixpt_t 1600 sched_pctcpu_delta(struct thread *td) 1601 { 1602 struct td_sched *ts; 1603 fixpt_t delta; 1604 int realstathz; 1605 1606 THREAD_LOCK_ASSERT(td, MA_OWNED); 1607 ts = td_get_sched(td); 1608 delta = 0; 1609 realstathz = stathz ? stathz : hz; 1610 if (ts->ts_cpticks != 0) { 1611 #if (FSHIFT >= CCPU_SHIFT) 1612 delta = (realstathz == 100) 1613 ? ((fixpt_t) ts->ts_cpticks) << 1614 (FSHIFT - CCPU_SHIFT) : 1615 100 * (((fixpt_t) ts->ts_cpticks) 1616 << (FSHIFT - CCPU_SHIFT)) / realstathz; 1617 #else 1618 delta = ((FSCALE - ccpu) * 1619 (ts->ts_cpticks * 1620 FSCALE / realstathz)) >> FSHIFT; 1621 #endif 1622 } 1623 1624 return (delta); 1625 } 1626 #endif 1627 1628 u_int 1629 sched_estcpu(struct thread *td) 1630 { 1631 1632 return (td_get_sched(td)->ts_estcpu); 1633 } 1634 1635 /* 1636 * The actual idle process. 1637 */ 1638 void 1639 sched_idletd(void *dummy) 1640 { 1641 struct pcpuidlestat *stat; 1642 1643 THREAD_NO_SLEEPING(); 1644 stat = DPCPU_PTR(idlestat); 1645 for (;;) { 1646 mtx_assert(&Giant, MA_NOTOWNED); 1647 1648 while (sched_runnable() == 0) { 1649 cpu_idle(stat->idlecalls + stat->oldidlecalls > 64); 1650 stat->idlecalls++; 1651 } 1652 1653 mtx_lock_spin(&sched_lock); 1654 mi_switch(SW_VOL | SWT_IDLE); 1655 } 1656 } 1657 1658 /* 1659 * A CPU is entering for the first time or a thread is exiting. 1660 */ 1661 void 1662 sched_throw(struct thread *td) 1663 { 1664 /* 1665 * Correct spinlock nesting. The idle thread context that we are 1666 * borrowing was created so that it would start out with a single 1667 * spin lock (sched_lock) held in fork_trampoline(). Since we've 1668 * explicitly acquired locks in this function, the nesting count 1669 * is now 2 rather than 1. Since we are nested, calling 1670 * spinlock_exit() will simply adjust the counts without allowing 1671 * spin lock using code to interrupt us. 1672 */ 1673 if (td == NULL) { 1674 mtx_lock_spin(&sched_lock); 1675 spinlock_exit(); 1676 PCPU_SET(switchtime, cpu_ticks()); 1677 PCPU_SET(switchticks, ticks); 1678 } else { 1679 lock_profile_release_lock(&sched_lock.lock_object); 1680 MPASS(td->td_lock == &sched_lock); 1681 td->td_lastcpu = td->td_oncpu; 1682 td->td_oncpu = NOCPU; 1683 } 1684 mtx_assert(&sched_lock, MA_OWNED); 1685 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 1686 cpu_throw(td, choosethread()); /* doesn't return */ 1687 } 1688 1689 void 1690 sched_fork_exit(struct thread *td) 1691 { 1692 1693 /* 1694 * Finish setting up thread glue so that it begins execution in a 1695 * non-nested critical section with sched_lock held but not recursed. 1696 */ 1697 td->td_oncpu = PCPU_GET(cpuid); 1698 sched_lock.mtx_lock = (uintptr_t)td; 1699 lock_profile_obtain_lock_success(&sched_lock.lock_object, 1700 0, 0, __FILE__, __LINE__); 1701 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED); 1702 1703 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running", 1704 "prio:%d", td->td_priority); 1705 SDT_PROBE0(sched, , , on__cpu); 1706 } 1707 1708 char * 1709 sched_tdname(struct thread *td) 1710 { 1711 #ifdef KTR 1712 struct td_sched *ts; 1713 1714 ts = td_get_sched(td); 1715 if (ts->ts_name[0] == '\0') 1716 snprintf(ts->ts_name, sizeof(ts->ts_name), 1717 "%s tid %d", td->td_name, td->td_tid); 1718 return (ts->ts_name); 1719 #else 1720 return (td->td_name); 1721 #endif 1722 } 1723 1724 #ifdef KTR 1725 void 1726 sched_clear_tdname(struct thread *td) 1727 { 1728 struct td_sched *ts; 1729 1730 ts = td_get_sched(td); 1731 ts->ts_name[0] = '\0'; 1732 } 1733 #endif 1734 1735 void 1736 sched_affinity(struct thread *td) 1737 { 1738 #ifdef SMP 1739 struct td_sched *ts; 1740 int cpu; 1741 1742 THREAD_LOCK_ASSERT(td, MA_OWNED); 1743 1744 /* 1745 * Set the TSF_AFFINITY flag if there is at least one CPU this 1746 * thread can't run on. 1747 */ 1748 ts = td_get_sched(td); 1749 ts->ts_flags &= ~TSF_AFFINITY; 1750 CPU_FOREACH(cpu) { 1751 if (!THREAD_CAN_SCHED(td, cpu)) { 1752 ts->ts_flags |= TSF_AFFINITY; 1753 break; 1754 } 1755 } 1756 1757 /* 1758 * If this thread can run on all CPUs, nothing else to do. 1759 */ 1760 if (!(ts->ts_flags & TSF_AFFINITY)) 1761 return; 1762 1763 /* Pinned threads and bound threads should be left alone. */ 1764 if (td->td_pinned != 0 || td->td_flags & TDF_BOUND) 1765 return; 1766 1767 switch (td->td_state) { 1768 case TDS_RUNQ: 1769 /* 1770 * If we are on a per-CPU runqueue that is in the set, 1771 * then nothing needs to be done. 1772 */ 1773 if (ts->ts_runq != &runq && 1774 THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu)) 1775 return; 1776 1777 /* Put this thread on a valid per-CPU runqueue. */ 1778 sched_rem(td); 1779 sched_add(td, SRQ_HOLDTD | SRQ_BORING); 1780 break; 1781 case TDS_RUNNING: 1782 /* 1783 * See if our current CPU is in the set. If not, force a 1784 * context switch. 1785 */ 1786 if (THREAD_CAN_SCHED(td, td->td_oncpu)) 1787 return; 1788 1789 td->td_flags |= TDF_NEEDRESCHED; 1790 if (td != curthread) 1791 ipi_cpu(cpu, IPI_AST); 1792 break; 1793 default: 1794 break; 1795 } 1796 #endif 1797 } 1798