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 ast_sched_locked(curthread, TDA_SCHED); 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_SLICEEND; 760 ast_sched_locked(td, TDA_SCHED); 761 } 762 } 763 764 stat = DPCPU_PTR(idlestat); 765 stat->oldidlecalls = stat->idlecalls; 766 stat->idlecalls = 0; 767 } 768 769 void 770 sched_clock(struct thread *td, int cnt) 771 { 772 773 for ( ; cnt > 0; cnt--) 774 sched_clock_tick(td); 775 } 776 777 /* 778 * Charge child's scheduling CPU usage to parent. 779 */ 780 void 781 sched_exit(struct proc *p, struct thread *td) 782 { 783 784 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit", 785 "prio:%d", td->td_priority); 786 787 PROC_LOCK_ASSERT(p, MA_OWNED); 788 sched_exit_thread(FIRST_THREAD_IN_PROC(p), td); 789 } 790 791 void 792 sched_exit_thread(struct thread *td, struct thread *child) 793 { 794 795 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit", 796 "prio:%d", child->td_priority); 797 thread_lock(td); 798 td_get_sched(td)->ts_estcpu = ESTCPULIM(td_get_sched(td)->ts_estcpu + 799 td_get_sched(child)->ts_estcpu); 800 thread_unlock(td); 801 thread_lock(child); 802 if ((child->td_flags & TDF_NOLOAD) == 0) 803 sched_load_rem(); 804 thread_unlock(child); 805 } 806 807 void 808 sched_fork(struct thread *td, struct thread *childtd) 809 { 810 sched_fork_thread(td, childtd); 811 } 812 813 void 814 sched_fork_thread(struct thread *td, struct thread *childtd) 815 { 816 struct td_sched *ts, *tsc; 817 818 childtd->td_oncpu = NOCPU; 819 childtd->td_lastcpu = NOCPU; 820 childtd->td_lock = &sched_lock; 821 childtd->td_cpuset = cpuset_ref(td->td_cpuset); 822 childtd->td_domain.dr_policy = td->td_cpuset->cs_domain; 823 childtd->td_priority = childtd->td_base_pri; 824 ts = td_get_sched(childtd); 825 bzero(ts, sizeof(*ts)); 826 tsc = td_get_sched(td); 827 ts->ts_estcpu = tsc->ts_estcpu; 828 ts->ts_flags |= (tsc->ts_flags & TSF_AFFINITY); 829 ts->ts_slice = 1; 830 } 831 832 void 833 sched_nice(struct proc *p, int nice) 834 { 835 struct thread *td; 836 837 PROC_LOCK_ASSERT(p, MA_OWNED); 838 p->p_nice = nice; 839 FOREACH_THREAD_IN_PROC(p, td) { 840 thread_lock(td); 841 resetpriority(td); 842 resetpriority_thread(td); 843 thread_unlock(td); 844 } 845 } 846 847 void 848 sched_class(struct thread *td, int class) 849 { 850 THREAD_LOCK_ASSERT(td, MA_OWNED); 851 td->td_pri_class = class; 852 } 853 854 /* 855 * Adjust the priority of a thread. 856 */ 857 static void 858 sched_priority(struct thread *td, u_char prio) 859 { 860 861 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change", 862 "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED, 863 sched_tdname(curthread)); 864 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio); 865 if (td != curthread && prio > td->td_priority) { 866 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 867 "lend prio", "prio:%d", td->td_priority, "new prio:%d", 868 prio, KTR_ATTR_LINKED, sched_tdname(td)); 869 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio, 870 curthread); 871 } 872 THREAD_LOCK_ASSERT(td, MA_OWNED); 873 if (td->td_priority == prio) 874 return; 875 td->td_priority = prio; 876 if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) { 877 sched_rem(td); 878 sched_add(td, SRQ_BORING | SRQ_HOLDTD); 879 } 880 } 881 882 /* 883 * Update a thread's priority when it is lent another thread's 884 * priority. 885 */ 886 void 887 sched_lend_prio(struct thread *td, u_char prio) 888 { 889 890 td->td_flags |= TDF_BORROWING; 891 sched_priority(td, prio); 892 } 893 894 /* 895 * Restore a thread's priority when priority propagation is 896 * over. The prio argument is the minimum priority the thread 897 * needs to have to satisfy other possible priority lending 898 * requests. If the thread's regulary priority is less 899 * important than prio the thread will keep a priority boost 900 * of prio. 901 */ 902 void 903 sched_unlend_prio(struct thread *td, u_char prio) 904 { 905 u_char base_pri; 906 907 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 908 td->td_base_pri <= PRI_MAX_TIMESHARE) 909 base_pri = td->td_user_pri; 910 else 911 base_pri = td->td_base_pri; 912 if (prio >= base_pri) { 913 td->td_flags &= ~TDF_BORROWING; 914 sched_prio(td, base_pri); 915 } else 916 sched_lend_prio(td, prio); 917 } 918 919 void 920 sched_prio(struct thread *td, u_char prio) 921 { 922 u_char oldprio; 923 924 /* First, update the base priority. */ 925 td->td_base_pri = prio; 926 927 /* 928 * If the thread is borrowing another thread's priority, don't ever 929 * lower the priority. 930 */ 931 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 932 return; 933 934 /* Change the real priority. */ 935 oldprio = td->td_priority; 936 sched_priority(td, prio); 937 938 /* 939 * If the thread is on a turnstile, then let the turnstile update 940 * its state. 941 */ 942 if (TD_ON_LOCK(td) && oldprio != prio) 943 turnstile_adjust(td, oldprio); 944 } 945 946 void 947 sched_ithread_prio(struct thread *td, u_char prio) 948 { 949 THREAD_LOCK_ASSERT(td, MA_OWNED); 950 MPASS(td->td_pri_class == PRI_ITHD); 951 td->td_base_ithread_pri = prio; 952 sched_prio(td, prio); 953 } 954 955 void 956 sched_user_prio(struct thread *td, u_char prio) 957 { 958 959 THREAD_LOCK_ASSERT(td, MA_OWNED); 960 td->td_base_user_pri = prio; 961 if (td->td_lend_user_pri <= prio) 962 return; 963 td->td_user_pri = prio; 964 } 965 966 void 967 sched_lend_user_prio(struct thread *td, u_char prio) 968 { 969 970 THREAD_LOCK_ASSERT(td, MA_OWNED); 971 td->td_lend_user_pri = prio; 972 td->td_user_pri = min(prio, td->td_base_user_pri); 973 if (td->td_priority > td->td_user_pri) 974 sched_prio(td, td->td_user_pri); 975 else if (td->td_priority != td->td_user_pri) 976 ast_sched_locked(td, TDA_SCHED); 977 } 978 979 /* 980 * Like the above but first check if there is anything to do. 981 */ 982 void 983 sched_lend_user_prio_cond(struct thread *td, u_char prio) 984 { 985 986 if (td->td_lend_user_pri != prio) 987 goto lend; 988 if (td->td_user_pri != min(prio, td->td_base_user_pri)) 989 goto lend; 990 if (td->td_priority != td->td_user_pri) 991 goto lend; 992 return; 993 994 lend: 995 thread_lock(td); 996 sched_lend_user_prio(td, prio); 997 thread_unlock(td); 998 } 999 1000 void 1001 sched_sleep(struct thread *td, int pri) 1002 { 1003 1004 THREAD_LOCK_ASSERT(td, MA_OWNED); 1005 td->td_slptick = ticks; 1006 td_get_sched(td)->ts_slptime = 0; 1007 if (pri != 0 && PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 1008 sched_prio(td, pri); 1009 if (TD_IS_SUSPENDED(td) || pri >= PSOCK) 1010 td->td_flags |= TDF_CANSWAP; 1011 } 1012 1013 void 1014 sched_switch(struct thread *td, int flags) 1015 { 1016 struct thread *newtd; 1017 struct mtx *tmtx; 1018 int preempted; 1019 1020 tmtx = &sched_lock; 1021 1022 THREAD_LOCK_ASSERT(td, MA_OWNED); 1023 1024 td->td_lastcpu = td->td_oncpu; 1025 preempted = (td->td_flags & TDF_SLICEEND) == 0 && 1026 (flags & SW_PREEMPT) != 0; 1027 td->td_flags &= ~TDF_SLICEEND; 1028 ast_unsched_locked(td, TDA_SCHED); 1029 td->td_owepreempt = 0; 1030 td->td_oncpu = NOCPU; 1031 1032 /* 1033 * At the last moment, if this thread is still marked RUNNING, 1034 * then put it back on the run queue as it has not been suspended 1035 * or stopped or any thing else similar. We never put the idle 1036 * threads on the run queue, however. 1037 */ 1038 if (td->td_flags & TDF_IDLETD) { 1039 TD_SET_CAN_RUN(td); 1040 #ifdef SMP 1041 CPU_CLR(PCPU_GET(cpuid), &idle_cpus_mask); 1042 #endif 1043 } else { 1044 if (TD_IS_RUNNING(td)) { 1045 /* Put us back on the run queue. */ 1046 sched_add(td, preempted ? 1047 SRQ_HOLDTD|SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1048 SRQ_HOLDTD|SRQ_OURSELF|SRQ_YIELDING); 1049 } 1050 } 1051 1052 /* 1053 * Switch to the sched lock to fix things up and pick 1054 * a new thread. Block the td_lock in order to avoid 1055 * breaking the critical path. 1056 */ 1057 if (td->td_lock != &sched_lock) { 1058 mtx_lock_spin(&sched_lock); 1059 tmtx = thread_lock_block(td); 1060 mtx_unlock_spin(tmtx); 1061 } 1062 1063 if ((td->td_flags & TDF_NOLOAD) == 0) 1064 sched_load_rem(); 1065 1066 newtd = choosethread(); 1067 MPASS(newtd->td_lock == &sched_lock); 1068 1069 #if (KTR_COMPILE & KTR_SCHED) != 0 1070 if (TD_IS_IDLETHREAD(td)) 1071 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle", 1072 "prio:%d", td->td_priority); 1073 else 1074 KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td), 1075 "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg, 1076 "lockname:\"%s\"", td->td_lockname); 1077 #endif 1078 1079 if (td != newtd) { 1080 #ifdef HWPMC_HOOKS 1081 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1082 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1083 #endif 1084 1085 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc); 1086 1087 /* I feel sleepy */ 1088 lock_profile_release_lock(&sched_lock.lock_object, true); 1089 #ifdef KDTRACE_HOOKS 1090 /* 1091 * If DTrace has set the active vtime enum to anything 1092 * other than INACTIVE (0), then it should have set the 1093 * function to call. 1094 */ 1095 if (dtrace_vtime_active) 1096 (*dtrace_vtime_switch_func)(newtd); 1097 #endif 1098 1099 cpu_switch(td, newtd, tmtx); 1100 lock_profile_obtain_lock_success(&sched_lock.lock_object, true, 1101 0, 0, __FILE__, __LINE__); 1102 /* 1103 * Where am I? What year is it? 1104 * We are in the same thread that went to sleep above, 1105 * but any amount of time may have passed. All our context 1106 * will still be available as will local variables. 1107 * PCPU values however may have changed as we may have 1108 * changed CPU so don't trust cached values of them. 1109 * New threads will go to fork_exit() instead of here 1110 * so if you change things here you may need to change 1111 * things there too. 1112 * 1113 * If the thread above was exiting it will never wake 1114 * up again here, so either it has saved everything it 1115 * needed to, or the thread_wait() or wait() will 1116 * need to reap it. 1117 */ 1118 1119 SDT_PROBE0(sched, , , on__cpu); 1120 #ifdef HWPMC_HOOKS 1121 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1122 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1123 #endif 1124 } else { 1125 td->td_lock = &sched_lock; 1126 SDT_PROBE0(sched, , , remain__cpu); 1127 } 1128 1129 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running", 1130 "prio:%d", td->td_priority); 1131 1132 #ifdef SMP 1133 if (td->td_flags & TDF_IDLETD) 1134 CPU_SET(PCPU_GET(cpuid), &idle_cpus_mask); 1135 #endif 1136 sched_lock.mtx_lock = (uintptr_t)td; 1137 td->td_oncpu = PCPU_GET(cpuid); 1138 spinlock_enter(); 1139 mtx_unlock_spin(&sched_lock); 1140 } 1141 1142 void 1143 sched_wakeup(struct thread *td, int srqflags) 1144 { 1145 struct td_sched *ts; 1146 1147 THREAD_LOCK_ASSERT(td, MA_OWNED); 1148 ts = td_get_sched(td); 1149 td->td_flags &= ~TDF_CANSWAP; 1150 if (ts->ts_slptime > 1) { 1151 updatepri(td); 1152 resetpriority(td); 1153 } 1154 td->td_slptick = 0; 1155 ts->ts_slptime = 0; 1156 ts->ts_slice = sched_slice; 1157 1158 /* 1159 * When resuming an idle ithread, restore its base ithread 1160 * priority. 1161 */ 1162 if (PRI_BASE(td->td_pri_class) == PRI_ITHD && 1163 td->td_base_pri != td->td_base_ithread_pri) 1164 sched_prio(td, td->td_base_ithread_pri); 1165 1166 sched_add(td, srqflags); 1167 } 1168 1169 #ifdef SMP 1170 static int 1171 forward_wakeup(int cpunum) 1172 { 1173 struct pcpu *pc; 1174 cpuset_t dontuse, map, map2; 1175 u_int id, me; 1176 int iscpuset; 1177 1178 mtx_assert(&sched_lock, MA_OWNED); 1179 1180 CTR0(KTR_RUNQ, "forward_wakeup()"); 1181 1182 if ((!forward_wakeup_enabled) || 1183 (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0)) 1184 return (0); 1185 if (!smp_started || KERNEL_PANICKED()) 1186 return (0); 1187 1188 forward_wakeups_requested++; 1189 1190 /* 1191 * Check the idle mask we received against what we calculated 1192 * before in the old version. 1193 */ 1194 me = PCPU_GET(cpuid); 1195 1196 /* Don't bother if we should be doing it ourself. */ 1197 if (CPU_ISSET(me, &idle_cpus_mask) && 1198 (cpunum == NOCPU || me == cpunum)) 1199 return (0); 1200 1201 CPU_SETOF(me, &dontuse); 1202 CPU_OR(&dontuse, &dontuse, &stopped_cpus); 1203 CPU_OR(&dontuse, &dontuse, &hlt_cpus_mask); 1204 CPU_ZERO(&map2); 1205 if (forward_wakeup_use_loop) { 1206 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { 1207 id = pc->pc_cpuid; 1208 if (!CPU_ISSET(id, &dontuse) && 1209 pc->pc_curthread == pc->pc_idlethread) { 1210 CPU_SET(id, &map2); 1211 } 1212 } 1213 } 1214 1215 if (forward_wakeup_use_mask) { 1216 map = idle_cpus_mask; 1217 CPU_ANDNOT(&map, &map, &dontuse); 1218 1219 /* If they are both on, compare and use loop if different. */ 1220 if (forward_wakeup_use_loop) { 1221 if (CPU_CMP(&map, &map2)) { 1222 printf("map != map2, loop method preferred\n"); 1223 map = map2; 1224 } 1225 } 1226 } else { 1227 map = map2; 1228 } 1229 1230 /* If we only allow a specific CPU, then mask off all the others. */ 1231 if (cpunum != NOCPU) { 1232 KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum.")); 1233 iscpuset = CPU_ISSET(cpunum, &map); 1234 if (iscpuset == 0) 1235 CPU_ZERO(&map); 1236 else 1237 CPU_SETOF(cpunum, &map); 1238 } 1239 if (!CPU_EMPTY(&map)) { 1240 forward_wakeups_delivered++; 1241 STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { 1242 id = pc->pc_cpuid; 1243 if (!CPU_ISSET(id, &map)) 1244 continue; 1245 if (cpu_idle_wakeup(pc->pc_cpuid)) 1246 CPU_CLR(id, &map); 1247 } 1248 if (!CPU_EMPTY(&map)) 1249 ipi_selected(map, IPI_AST); 1250 return (1); 1251 } 1252 if (cpunum == NOCPU) 1253 printf("forward_wakeup: Idle processor not found\n"); 1254 return (0); 1255 } 1256 1257 static void 1258 kick_other_cpu(int pri, int cpuid) 1259 { 1260 struct pcpu *pcpu; 1261 int cpri; 1262 1263 pcpu = pcpu_find(cpuid); 1264 if (CPU_ISSET(cpuid, &idle_cpus_mask)) { 1265 forward_wakeups_delivered++; 1266 if (!cpu_idle_wakeup(cpuid)) 1267 ipi_cpu(cpuid, IPI_AST); 1268 return; 1269 } 1270 1271 cpri = pcpu->pc_curthread->td_priority; 1272 if (pri >= cpri) 1273 return; 1274 1275 #if defined(IPI_PREEMPTION) && defined(PREEMPTION) 1276 #if !defined(FULL_PREEMPTION) 1277 if (pri <= PRI_MAX_ITHD) 1278 #endif /* ! FULL_PREEMPTION */ 1279 { 1280 ipi_cpu(cpuid, IPI_PREEMPT); 1281 return; 1282 } 1283 #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */ 1284 1285 ast_sched_locked(pcpu->pc_curthread, TDA_SCHED); 1286 ipi_cpu(cpuid, IPI_AST); 1287 return; 1288 } 1289 #endif /* SMP */ 1290 1291 #ifdef SMP 1292 static int 1293 sched_pickcpu(struct thread *td) 1294 { 1295 int best, cpu; 1296 1297 mtx_assert(&sched_lock, MA_OWNED); 1298 1299 if (td->td_lastcpu != NOCPU && THREAD_CAN_SCHED(td, td->td_lastcpu)) 1300 best = td->td_lastcpu; 1301 else 1302 best = NOCPU; 1303 CPU_FOREACH(cpu) { 1304 if (!THREAD_CAN_SCHED(td, cpu)) 1305 continue; 1306 1307 if (best == NOCPU) 1308 best = cpu; 1309 else if (runq_length[cpu] < runq_length[best]) 1310 best = cpu; 1311 } 1312 KASSERT(best != NOCPU, ("no valid CPUs")); 1313 1314 return (best); 1315 } 1316 #endif 1317 1318 void 1319 sched_add(struct thread *td, int flags) 1320 #ifdef SMP 1321 { 1322 cpuset_t tidlemsk; 1323 struct td_sched *ts; 1324 u_int cpu, cpuid; 1325 int forwarded = 0; 1326 int single_cpu = 0; 1327 1328 ts = td_get_sched(td); 1329 THREAD_LOCK_ASSERT(td, MA_OWNED); 1330 KASSERT((td->td_inhibitors == 0), 1331 ("sched_add: trying to run inhibited thread")); 1332 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 1333 ("sched_add: bad thread state")); 1334 KASSERT(td->td_flags & TDF_INMEM, 1335 ("sched_add: thread swapped out")); 1336 1337 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 1338 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 1339 sched_tdname(curthread)); 1340 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 1341 KTR_ATTR_LINKED, sched_tdname(td)); 1342 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 1343 flags & SRQ_PREEMPTED); 1344 1345 /* 1346 * Now that the thread is moving to the run-queue, set the lock 1347 * to the scheduler's lock. 1348 */ 1349 if (td->td_lock != &sched_lock) { 1350 mtx_lock_spin(&sched_lock); 1351 if ((flags & SRQ_HOLD) != 0) 1352 td->td_lock = &sched_lock; 1353 else 1354 thread_lock_set(td, &sched_lock); 1355 } 1356 TD_SET_RUNQ(td); 1357 1358 /* 1359 * If SMP is started and the thread is pinned or otherwise limited to 1360 * a specific set of CPUs, queue the thread to a per-CPU run queue. 1361 * Otherwise, queue the thread to the global run queue. 1362 * 1363 * If SMP has not yet been started we must use the global run queue 1364 * as per-CPU state may not be initialized yet and we may crash if we 1365 * try to access the per-CPU run queues. 1366 */ 1367 if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND || 1368 ts->ts_flags & TSF_AFFINITY)) { 1369 if (td->td_pinned != 0) 1370 cpu = td->td_lastcpu; 1371 else if (td->td_flags & TDF_BOUND) { 1372 /* Find CPU from bound runq. */ 1373 KASSERT(SKE_RUNQ_PCPU(ts), 1374 ("sched_add: bound td_sched not on cpu runq")); 1375 cpu = ts->ts_runq - &runq_pcpu[0]; 1376 } else 1377 /* Find a valid CPU for our cpuset */ 1378 cpu = sched_pickcpu(td); 1379 ts->ts_runq = &runq_pcpu[cpu]; 1380 single_cpu = 1; 1381 CTR3(KTR_RUNQ, 1382 "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, 1383 cpu); 1384 } else { 1385 CTR2(KTR_RUNQ, 1386 "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts, 1387 td); 1388 cpu = NOCPU; 1389 ts->ts_runq = &runq; 1390 } 1391 1392 if ((td->td_flags & TDF_NOLOAD) == 0) 1393 sched_load_add(); 1394 runq_add(ts->ts_runq, td, flags); 1395 if (cpu != NOCPU) 1396 runq_length[cpu]++; 1397 1398 cpuid = PCPU_GET(cpuid); 1399 if (single_cpu && cpu != cpuid) { 1400 kick_other_cpu(td->td_priority, cpu); 1401 } else { 1402 if (!single_cpu) { 1403 tidlemsk = idle_cpus_mask; 1404 CPU_ANDNOT(&tidlemsk, &tidlemsk, &hlt_cpus_mask); 1405 CPU_CLR(cpuid, &tidlemsk); 1406 1407 if (!CPU_ISSET(cpuid, &idle_cpus_mask) && 1408 ((flags & SRQ_INTR) == 0) && 1409 !CPU_EMPTY(&tidlemsk)) 1410 forwarded = forward_wakeup(cpu); 1411 } 1412 1413 if (!forwarded) { 1414 if (!maybe_preempt(td)) 1415 maybe_resched(td); 1416 } 1417 } 1418 if ((flags & SRQ_HOLDTD) == 0) 1419 thread_unlock(td); 1420 } 1421 #else /* SMP */ 1422 { 1423 struct td_sched *ts; 1424 1425 ts = td_get_sched(td); 1426 THREAD_LOCK_ASSERT(td, MA_OWNED); 1427 KASSERT((td->td_inhibitors == 0), 1428 ("sched_add: trying to run inhibited thread")); 1429 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 1430 ("sched_add: bad thread state")); 1431 KASSERT(td->td_flags & TDF_INMEM, 1432 ("sched_add: thread swapped out")); 1433 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 1434 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 1435 sched_tdname(curthread)); 1436 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 1437 KTR_ATTR_LINKED, sched_tdname(td)); 1438 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 1439 flags & SRQ_PREEMPTED); 1440 1441 /* 1442 * Now that the thread is moving to the run-queue, set the lock 1443 * to the scheduler's lock. 1444 */ 1445 if (td->td_lock != &sched_lock) { 1446 mtx_lock_spin(&sched_lock); 1447 if ((flags & SRQ_HOLD) != 0) 1448 td->td_lock = &sched_lock; 1449 else 1450 thread_lock_set(td, &sched_lock); 1451 } 1452 TD_SET_RUNQ(td); 1453 CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td); 1454 ts->ts_runq = &runq; 1455 1456 if ((td->td_flags & TDF_NOLOAD) == 0) 1457 sched_load_add(); 1458 runq_add(ts->ts_runq, td, flags); 1459 if (!maybe_preempt(td)) 1460 maybe_resched(td); 1461 if ((flags & SRQ_HOLDTD) == 0) 1462 thread_unlock(td); 1463 } 1464 #endif /* SMP */ 1465 1466 void 1467 sched_rem(struct thread *td) 1468 { 1469 struct td_sched *ts; 1470 1471 ts = td_get_sched(td); 1472 KASSERT(td->td_flags & TDF_INMEM, 1473 ("sched_rem: thread swapped out")); 1474 KASSERT(TD_ON_RUNQ(td), 1475 ("sched_rem: thread not on run queue")); 1476 mtx_assert(&sched_lock, MA_OWNED); 1477 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 1478 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 1479 sched_tdname(curthread)); 1480 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); 1481 1482 if ((td->td_flags & TDF_NOLOAD) == 0) 1483 sched_load_rem(); 1484 #ifdef SMP 1485 if (ts->ts_runq != &runq) 1486 runq_length[ts->ts_runq - runq_pcpu]--; 1487 #endif 1488 runq_remove(ts->ts_runq, td); 1489 TD_SET_CAN_RUN(td); 1490 } 1491 1492 /* 1493 * Select threads to run. Note that running threads still consume a 1494 * slot. 1495 */ 1496 struct thread * 1497 sched_choose(void) 1498 { 1499 struct thread *td; 1500 struct runq *rq; 1501 1502 mtx_assert(&sched_lock, MA_OWNED); 1503 #ifdef SMP 1504 struct thread *tdcpu; 1505 1506 rq = &runq; 1507 td = runq_choose_fuzz(&runq, runq_fuzz); 1508 tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]); 1509 1510 if (td == NULL || 1511 (tdcpu != NULL && 1512 tdcpu->td_priority < td->td_priority)) { 1513 CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu, 1514 PCPU_GET(cpuid)); 1515 td = tdcpu; 1516 rq = &runq_pcpu[PCPU_GET(cpuid)]; 1517 } else { 1518 CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td); 1519 } 1520 1521 #else 1522 rq = &runq; 1523 td = runq_choose(&runq); 1524 #endif 1525 1526 if (td) { 1527 #ifdef SMP 1528 if (td == tdcpu) 1529 runq_length[PCPU_GET(cpuid)]--; 1530 #endif 1531 runq_remove(rq, td); 1532 td->td_flags |= TDF_DIDRUN; 1533 1534 KASSERT(td->td_flags & TDF_INMEM, 1535 ("sched_choose: thread swapped out")); 1536 return (td); 1537 } 1538 return (PCPU_GET(idlethread)); 1539 } 1540 1541 void 1542 sched_preempt(struct thread *td) 1543 { 1544 1545 SDT_PROBE2(sched, , , surrender, td, td->td_proc); 1546 if (td->td_critnest > 1) { 1547 td->td_owepreempt = 1; 1548 } else { 1549 thread_lock(td); 1550 mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT); 1551 } 1552 } 1553 1554 void 1555 sched_userret_slowpath(struct thread *td) 1556 { 1557 1558 thread_lock(td); 1559 td->td_priority = td->td_user_pri; 1560 td->td_base_pri = td->td_user_pri; 1561 thread_unlock(td); 1562 } 1563 1564 void 1565 sched_bind(struct thread *td, int cpu) 1566 { 1567 #ifdef SMP 1568 struct td_sched *ts = td_get_sched(td); 1569 #endif 1570 1571 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 1572 KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 1573 1574 td->td_flags |= TDF_BOUND; 1575 #ifdef SMP 1576 ts->ts_runq = &runq_pcpu[cpu]; 1577 if (PCPU_GET(cpuid) == cpu) 1578 return; 1579 1580 mi_switch(SW_VOL); 1581 thread_lock(td); 1582 #endif 1583 } 1584 1585 void 1586 sched_unbind(struct thread* td) 1587 { 1588 THREAD_LOCK_ASSERT(td, MA_OWNED); 1589 KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 1590 td->td_flags &= ~TDF_BOUND; 1591 } 1592 1593 int 1594 sched_is_bound(struct thread *td) 1595 { 1596 THREAD_LOCK_ASSERT(td, MA_OWNED); 1597 return (td->td_flags & TDF_BOUND); 1598 } 1599 1600 void 1601 sched_relinquish(struct thread *td) 1602 { 1603 thread_lock(td); 1604 mi_switch(SW_VOL | SWT_RELINQUISH); 1605 } 1606 1607 int 1608 sched_load(void) 1609 { 1610 return (sched_tdcnt); 1611 } 1612 1613 int 1614 sched_sizeof_proc(void) 1615 { 1616 return (sizeof(struct proc)); 1617 } 1618 1619 int 1620 sched_sizeof_thread(void) 1621 { 1622 return (sizeof(struct thread) + sizeof(struct td_sched)); 1623 } 1624 1625 fixpt_t 1626 sched_pctcpu(struct thread *td) 1627 { 1628 struct td_sched *ts; 1629 1630 THREAD_LOCK_ASSERT(td, MA_OWNED); 1631 ts = td_get_sched(td); 1632 return (ts->ts_pctcpu); 1633 } 1634 1635 #ifdef RACCT 1636 /* 1637 * Calculates the contribution to the thread cpu usage for the latest 1638 * (unfinished) second. 1639 */ 1640 fixpt_t 1641 sched_pctcpu_delta(struct thread *td) 1642 { 1643 struct td_sched *ts; 1644 fixpt_t delta; 1645 int realstathz; 1646 1647 THREAD_LOCK_ASSERT(td, MA_OWNED); 1648 ts = td_get_sched(td); 1649 delta = 0; 1650 realstathz = stathz ? stathz : hz; 1651 if (ts->ts_cpticks != 0) { 1652 #if (FSHIFT >= CCPU_SHIFT) 1653 delta = (realstathz == 100) 1654 ? ((fixpt_t) ts->ts_cpticks) << 1655 (FSHIFT - CCPU_SHIFT) : 1656 100 * (((fixpt_t) ts->ts_cpticks) 1657 << (FSHIFT - CCPU_SHIFT)) / realstathz; 1658 #else 1659 delta = ((FSCALE - ccpu) * 1660 (ts->ts_cpticks * 1661 FSCALE / realstathz)) >> FSHIFT; 1662 #endif 1663 } 1664 1665 return (delta); 1666 } 1667 #endif 1668 1669 u_int 1670 sched_estcpu(struct thread *td) 1671 { 1672 1673 return (td_get_sched(td)->ts_estcpu); 1674 } 1675 1676 /* 1677 * The actual idle process. 1678 */ 1679 void 1680 sched_idletd(void *dummy) 1681 { 1682 struct pcpuidlestat *stat; 1683 1684 THREAD_NO_SLEEPING(); 1685 stat = DPCPU_PTR(idlestat); 1686 for (;;) { 1687 mtx_assert(&Giant, MA_NOTOWNED); 1688 1689 while (sched_runnable() == 0) { 1690 cpu_idle(stat->idlecalls + stat->oldidlecalls > 64); 1691 stat->idlecalls++; 1692 } 1693 1694 mtx_lock_spin(&sched_lock); 1695 mi_switch(SW_VOL | SWT_IDLE); 1696 } 1697 } 1698 1699 static void 1700 sched_throw_tail(struct thread *td) 1701 { 1702 1703 mtx_assert(&sched_lock, MA_OWNED); 1704 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 1705 cpu_throw(td, choosethread()); /* doesn't return */ 1706 } 1707 1708 /* 1709 * A CPU is entering for the first time. 1710 */ 1711 void 1712 sched_ap_entry(void) 1713 { 1714 1715 /* 1716 * Correct spinlock nesting. The idle thread context that we are 1717 * borrowing was created so that it would start out with a single 1718 * spin lock (sched_lock) held in fork_trampoline(). Since we've 1719 * explicitly acquired locks in this function, the nesting count 1720 * is now 2 rather than 1. Since we are nested, calling 1721 * spinlock_exit() will simply adjust the counts without allowing 1722 * spin lock using code to interrupt us. 1723 */ 1724 mtx_lock_spin(&sched_lock); 1725 spinlock_exit(); 1726 PCPU_SET(switchtime, cpu_ticks()); 1727 PCPU_SET(switchticks, ticks); 1728 1729 sched_throw_tail(NULL); 1730 } 1731 1732 /* 1733 * A thread is exiting. 1734 */ 1735 void 1736 sched_throw(struct thread *td) 1737 { 1738 1739 MPASS(td != NULL); 1740 MPASS(td->td_lock == &sched_lock); 1741 1742 lock_profile_release_lock(&sched_lock.lock_object, true); 1743 td->td_lastcpu = td->td_oncpu; 1744 td->td_oncpu = NOCPU; 1745 1746 sched_throw_tail(td); 1747 } 1748 1749 void 1750 sched_fork_exit(struct thread *td) 1751 { 1752 1753 /* 1754 * Finish setting up thread glue so that it begins execution in a 1755 * non-nested critical section with sched_lock held but not recursed. 1756 */ 1757 td->td_oncpu = PCPU_GET(cpuid); 1758 sched_lock.mtx_lock = (uintptr_t)td; 1759 lock_profile_obtain_lock_success(&sched_lock.lock_object, true, 1760 0, 0, __FILE__, __LINE__); 1761 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED); 1762 1763 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running", 1764 "prio:%d", td->td_priority); 1765 SDT_PROBE0(sched, , , on__cpu); 1766 } 1767 1768 char * 1769 sched_tdname(struct thread *td) 1770 { 1771 #ifdef KTR 1772 struct td_sched *ts; 1773 1774 ts = td_get_sched(td); 1775 if (ts->ts_name[0] == '\0') 1776 snprintf(ts->ts_name, sizeof(ts->ts_name), 1777 "%s tid %d", td->td_name, td->td_tid); 1778 return (ts->ts_name); 1779 #else 1780 return (td->td_name); 1781 #endif 1782 } 1783 1784 #ifdef KTR 1785 void 1786 sched_clear_tdname(struct thread *td) 1787 { 1788 struct td_sched *ts; 1789 1790 ts = td_get_sched(td); 1791 ts->ts_name[0] = '\0'; 1792 } 1793 #endif 1794 1795 void 1796 sched_affinity(struct thread *td) 1797 { 1798 #ifdef SMP 1799 struct td_sched *ts; 1800 int cpu; 1801 1802 THREAD_LOCK_ASSERT(td, MA_OWNED); 1803 1804 /* 1805 * Set the TSF_AFFINITY flag if there is at least one CPU this 1806 * thread can't run on. 1807 */ 1808 ts = td_get_sched(td); 1809 ts->ts_flags &= ~TSF_AFFINITY; 1810 CPU_FOREACH(cpu) { 1811 if (!THREAD_CAN_SCHED(td, cpu)) { 1812 ts->ts_flags |= TSF_AFFINITY; 1813 break; 1814 } 1815 } 1816 1817 /* 1818 * If this thread can run on all CPUs, nothing else to do. 1819 */ 1820 if (!(ts->ts_flags & TSF_AFFINITY)) 1821 return; 1822 1823 /* Pinned threads and bound threads should be left alone. */ 1824 if (td->td_pinned != 0 || td->td_flags & TDF_BOUND) 1825 return; 1826 1827 switch (TD_GET_STATE(td)) { 1828 case TDS_RUNQ: 1829 /* 1830 * If we are on a per-CPU runqueue that is in the set, 1831 * then nothing needs to be done. 1832 */ 1833 if (ts->ts_runq != &runq && 1834 THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu)) 1835 return; 1836 1837 /* Put this thread on a valid per-CPU runqueue. */ 1838 sched_rem(td); 1839 sched_add(td, SRQ_HOLDTD | SRQ_BORING); 1840 break; 1841 case TDS_RUNNING: 1842 /* 1843 * See if our current CPU is in the set. If not, force a 1844 * context switch. 1845 */ 1846 if (THREAD_CAN_SCHED(td, td->td_oncpu)) 1847 return; 1848 1849 ast_sched_locked(td, TDA_SCHED); 1850 if (td != curthread) 1851 ipi_cpu(cpu, IPI_AST); 1852 break; 1853 default: 1854 break; 1855 } 1856 #endif 1857 } 1858