sched_ule.c revision 236546
1/*- 2 * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org> 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice unmodified, this list of conditions, and the following 10 * disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 16 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 17 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 18 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 19 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 20 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 21 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 22 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 23 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 24 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 25 */ 26 27/* 28 * This file implements the ULE scheduler. ULE supports independent CPU 29 * run queues and fine grain locking. It has superior interactive 30 * performance under load even on uni-processor systems. 31 * 32 * etymology: 33 * ULE is the last three letters in schedule. It owes its name to a 34 * generic user created for a scheduling system by Paul Mikesell at 35 * Isilon Systems and a general lack of creativity on the part of the author. 36 */ 37 38#include <sys/cdefs.h> 39__FBSDID("$FreeBSD: stable/9/sys/kern/sched_ule.c 236546 2012-06-04 07:12:36Z mav $"); 40 41#include "opt_hwpmc_hooks.h" 42#include "opt_kdtrace.h" 43#include "opt_sched.h" 44 45#include <sys/param.h> 46#include <sys/systm.h> 47#include <sys/kdb.h> 48#include <sys/kernel.h> 49#include <sys/ktr.h> 50#include <sys/lock.h> 51#include <sys/mutex.h> 52#include <sys/proc.h> 53#include <sys/resource.h> 54#include <sys/resourcevar.h> 55#include <sys/sched.h> 56#include <sys/sdt.h> 57#include <sys/smp.h> 58#include <sys/sx.h> 59#include <sys/sysctl.h> 60#include <sys/sysproto.h> 61#include <sys/turnstile.h> 62#include <sys/umtx.h> 63#include <sys/vmmeter.h> 64#include <sys/cpuset.h> 65#include <sys/sbuf.h> 66 67#ifdef HWPMC_HOOKS 68#include <sys/pmckern.h> 69#endif 70 71#ifdef KDTRACE_HOOKS 72#include <sys/dtrace_bsd.h> 73int dtrace_vtime_active; 74dtrace_vtime_switch_func_t dtrace_vtime_switch_func; 75#endif 76 77#include <machine/cpu.h> 78#include <machine/smp.h> 79 80#if defined(__powerpc__) && defined(E500) 81#error "This architecture is not currently compatible with ULE" 82#endif 83 84#define KTR_ULE 0 85 86#define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX))) 87#define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU))) 88#define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load")) 89 90/* 91 * Thread scheduler specific section. All fields are protected 92 * by the thread lock. 93 */ 94struct td_sched { 95 struct runq *ts_runq; /* Run-queue we're queued on. */ 96 short ts_flags; /* TSF_* flags. */ 97 u_char ts_cpu; /* CPU that we have affinity for. */ 98 int ts_rltick; /* Real last tick, for affinity. */ 99 int ts_slice; /* Ticks of slice remaining. */ 100 u_int ts_slptime; /* Number of ticks we vol. slept */ 101 u_int ts_runtime; /* Number of ticks we were running */ 102 int ts_ltick; /* Last tick that we were running on */ 103 int ts_ftick; /* First tick that we were running on */ 104 int ts_ticks; /* Tick count */ 105#ifdef KTR 106 char ts_name[TS_NAME_LEN]; 107#endif 108}; 109/* flags kept in ts_flags */ 110#define TSF_BOUND 0x0001 /* Thread can not migrate. */ 111#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 112 113static struct td_sched td_sched0; 114 115#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 116#define THREAD_CAN_SCHED(td, cpu) \ 117 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 118 119/* 120 * Priority ranges used for interactive and non-interactive timeshare 121 * threads. The timeshare priorities are split up into four ranges. 122 * The first range handles interactive threads. The last three ranges 123 * (NHALF, x, and NHALF) handle non-interactive threads with the outer 124 * ranges supporting nice values. 125 */ 126#define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1) 127#define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2) 128#define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE) 129 130#define PRI_MIN_INTERACT PRI_MIN_TIMESHARE 131#define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1) 132#define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE) 133#define PRI_MAX_BATCH PRI_MAX_TIMESHARE 134 135/* 136 * Cpu percentage computation macros and defines. 137 * 138 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 139 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 140 * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 141 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 142 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 143 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 144 */ 145#define SCHED_TICK_SECS 10 146#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 147#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 148#define SCHED_TICK_SHIFT 10 149#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 150#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 151 152/* 153 * These macros determine priorities for non-interactive threads. They are 154 * assigned a priority based on their recent cpu utilization as expressed 155 * by the ratio of ticks to the tick total. NHALF priorities at the start 156 * and end of the MIN to MAX timeshare range are only reachable with negative 157 * or positive nice respectively. 158 * 159 * PRI_RANGE: Priority range for utilization dependent priorities. 160 * PRI_NRESV: Number of nice values. 161 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 162 * PRI_NICE: Determines the part of the priority inherited from nice. 163 */ 164#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 165#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 166#define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF) 167#define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF) 168#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1) 169#define SCHED_PRI_TICKS(ts) \ 170 (SCHED_TICK_HZ((ts)) / \ 171 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 172#define SCHED_PRI_NICE(nice) (nice) 173 174/* 175 * These determine the interactivity of a process. Interactivity differs from 176 * cpu utilization in that it expresses the voluntary time slept vs time ran 177 * while cpu utilization includes all time not running. This more accurately 178 * models the intent of the thread. 179 * 180 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 181 * before throttling back. 182 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 183 * INTERACT_MAX: Maximum interactivity value. Smaller is better. 184 * INTERACT_THRESH: Threshold for placement on the current runq. 185 */ 186#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 187#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 188#define SCHED_INTERACT_MAX (100) 189#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 190#define SCHED_INTERACT_THRESH (30) 191 192/* 193 * tickincr: Converts a stathz tick into a hz domain scaled by 194 * the shift factor. Without the shift the error rate 195 * due to rounding would be unacceptably high. 196 * realstathz: stathz is sometimes 0 and run off of hz. 197 * sched_slice: Runtime of each thread before rescheduling. 198 * preempt_thresh: Priority threshold for preemption and remote IPIs. 199 */ 200static int sched_interact = SCHED_INTERACT_THRESH; 201static int realstathz; 202static int tickincr; 203static int sched_slice = 1; 204#ifdef PREEMPTION 205#ifdef FULL_PREEMPTION 206static int preempt_thresh = PRI_MAX_IDLE; 207#else 208static int preempt_thresh = PRI_MIN_KERN; 209#endif 210#else 211static int preempt_thresh = 0; 212#endif 213static int static_boost = PRI_MIN_BATCH; 214static int sched_idlespins = 10000; 215static int sched_idlespinthresh = -1; 216 217/* 218 * tdq - per processor runqs and statistics. All fields are protected by the 219 * tdq_lock. The load and lowpri may be accessed without to avoid excess 220 * locking in sched_pickcpu(); 221 */ 222struct tdq { 223 /* Ordered to improve efficiency of cpu_search() and switch(). */ 224 struct mtx tdq_lock; /* run queue lock. */ 225 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */ 226 volatile int tdq_load; /* Aggregate load. */ 227 volatile int tdq_cpu_idle; /* cpu_idle() is active. */ 228 int tdq_sysload; /* For loadavg, !ITHD load. */ 229 int tdq_transferable; /* Transferable thread count. */ 230 short tdq_switchcnt; /* Switches this tick. */ 231 short tdq_oldswitchcnt; /* Switches last tick. */ 232 u_char tdq_lowpri; /* Lowest priority thread. */ 233 u_char tdq_ipipending; /* IPI pending. */ 234 u_char tdq_idx; /* Current insert index. */ 235 u_char tdq_ridx; /* Current removal index. */ 236 struct runq tdq_realtime; /* real-time run queue. */ 237 struct runq tdq_timeshare; /* timeshare run queue. */ 238 struct runq tdq_idle; /* Queue of IDLE threads. */ 239 char tdq_name[TDQ_NAME_LEN]; 240#ifdef KTR 241 char tdq_loadname[TDQ_LOADNAME_LEN]; 242#endif 243} __aligned(64); 244 245/* Idle thread states and config. */ 246#define TDQ_RUNNING 1 247#define TDQ_IDLE 2 248 249#ifdef SMP 250struct cpu_group *cpu_top; /* CPU topology */ 251 252#define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000)) 253#define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity)) 254 255/* 256 * Run-time tunables. 257 */ 258static int rebalance = 1; 259static int balance_interval = 128; /* Default set in sched_initticks(). */ 260static int affinity; 261static int steal_idle = 1; 262static int steal_thresh = 2; 263 264/* 265 * One thread queue per processor. 266 */ 267static struct tdq tdq_cpu[MAXCPU]; 268static struct tdq *balance_tdq; 269static int balance_ticks; 270static DPCPU_DEFINE(uint32_t, randomval); 271 272#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 273#define TDQ_CPU(x) (&tdq_cpu[(x)]) 274#define TDQ_ID(x) ((int)((x) - tdq_cpu)) 275#else /* !SMP */ 276static struct tdq tdq_cpu; 277 278#define TDQ_ID(x) (0) 279#define TDQ_SELF() (&tdq_cpu) 280#define TDQ_CPU(x) (&tdq_cpu) 281#endif 282 283#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 284#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 285#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 286#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 287#define TDQ_LOCKPTR(t) (&(t)->tdq_lock) 288 289static void sched_priority(struct thread *); 290static void sched_thread_priority(struct thread *, u_char); 291static int sched_interact_score(struct thread *); 292static void sched_interact_update(struct thread *); 293static void sched_interact_fork(struct thread *); 294static void sched_pctcpu_update(struct td_sched *, int); 295 296/* Operations on per processor queues */ 297static struct thread *tdq_choose(struct tdq *); 298static void tdq_setup(struct tdq *); 299static void tdq_load_add(struct tdq *, struct thread *); 300static void tdq_load_rem(struct tdq *, struct thread *); 301static __inline void tdq_runq_add(struct tdq *, struct thread *, int); 302static __inline void tdq_runq_rem(struct tdq *, struct thread *); 303static inline int sched_shouldpreempt(int, int, int); 304void tdq_print(int cpu); 305static void runq_print(struct runq *rq); 306static void tdq_add(struct tdq *, struct thread *, int); 307#ifdef SMP 308static int tdq_move(struct tdq *, struct tdq *); 309static int tdq_idled(struct tdq *); 310static void tdq_notify(struct tdq *, struct thread *); 311static struct thread *tdq_steal(struct tdq *, int); 312static struct thread *runq_steal(struct runq *, int); 313static int sched_pickcpu(struct thread *, int); 314static void sched_balance(void); 315static int sched_balance_pair(struct tdq *, struct tdq *); 316static inline struct tdq *sched_setcpu(struct thread *, int, int); 317static inline void thread_unblock_switch(struct thread *, struct mtx *); 318static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 319static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS); 320static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, 321 struct cpu_group *cg, int indent); 322#endif 323 324static void sched_setup(void *dummy); 325SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 326 327static void sched_initticks(void *dummy); 328SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 329 NULL); 330 331SDT_PROVIDER_DEFINE(sched); 332 333SDT_PROBE_DEFINE3(sched, , , change_pri, change-pri, "struct thread *", 334 "struct proc *", "uint8_t"); 335SDT_PROBE_DEFINE3(sched, , , dequeue, dequeue, "struct thread *", 336 "struct proc *", "void *"); 337SDT_PROBE_DEFINE4(sched, , , enqueue, enqueue, "struct thread *", 338 "struct proc *", "void *", "int"); 339SDT_PROBE_DEFINE4(sched, , , lend_pri, lend-pri, "struct thread *", 340 "struct proc *", "uint8_t", "struct thread *"); 341SDT_PROBE_DEFINE2(sched, , , load_change, load-change, "int", "int"); 342SDT_PROBE_DEFINE2(sched, , , off_cpu, off-cpu, "struct thread *", 343 "struct proc *"); 344SDT_PROBE_DEFINE(sched, , , on_cpu, on-cpu); 345SDT_PROBE_DEFINE(sched, , , remain_cpu, remain-cpu); 346SDT_PROBE_DEFINE2(sched, , , surrender, surrender, "struct thread *", 347 "struct proc *"); 348 349/* 350 * Print the threads waiting on a run-queue. 351 */ 352static void 353runq_print(struct runq *rq) 354{ 355 struct rqhead *rqh; 356 struct thread *td; 357 int pri; 358 int j; 359 int i; 360 361 for (i = 0; i < RQB_LEN; i++) { 362 printf("\t\trunq bits %d 0x%zx\n", 363 i, rq->rq_status.rqb_bits[i]); 364 for (j = 0; j < RQB_BPW; j++) 365 if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 366 pri = j + (i << RQB_L2BPW); 367 rqh = &rq->rq_queues[pri]; 368 TAILQ_FOREACH(td, rqh, td_runq) { 369 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 370 td, td->td_name, td->td_priority, 371 td->td_rqindex, pri); 372 } 373 } 374 } 375} 376 377/* 378 * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 379 */ 380void 381tdq_print(int cpu) 382{ 383 struct tdq *tdq; 384 385 tdq = TDQ_CPU(cpu); 386 387 printf("tdq %d:\n", TDQ_ID(tdq)); 388 printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 389 printf("\tLock name: %s\n", tdq->tdq_name); 390 printf("\tload: %d\n", tdq->tdq_load); 391 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt); 392 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt); 393 printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 394 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 395 printf("\tload transferable: %d\n", tdq->tdq_transferable); 396 printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 397 printf("\trealtime runq:\n"); 398 runq_print(&tdq->tdq_realtime); 399 printf("\ttimeshare runq:\n"); 400 runq_print(&tdq->tdq_timeshare); 401 printf("\tidle runq:\n"); 402 runq_print(&tdq->tdq_idle); 403} 404 405static inline int 406sched_shouldpreempt(int pri, int cpri, int remote) 407{ 408 /* 409 * If the new priority is not better than the current priority there is 410 * nothing to do. 411 */ 412 if (pri >= cpri) 413 return (0); 414 /* 415 * Always preempt idle. 416 */ 417 if (cpri >= PRI_MIN_IDLE) 418 return (1); 419 /* 420 * If preemption is disabled don't preempt others. 421 */ 422 if (preempt_thresh == 0) 423 return (0); 424 /* 425 * Preempt if we exceed the threshold. 426 */ 427 if (pri <= preempt_thresh) 428 return (1); 429 /* 430 * If we're interactive or better and there is non-interactive 431 * or worse running preempt only remote processors. 432 */ 433 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT) 434 return (1); 435 return (0); 436} 437 438/* 439 * Add a thread to the actual run-queue. Keeps transferable counts up to 440 * date with what is actually on the run-queue. Selects the correct 441 * queue position for timeshare threads. 442 */ 443static __inline void 444tdq_runq_add(struct tdq *tdq, struct thread *td, int flags) 445{ 446 struct td_sched *ts; 447 u_char pri; 448 449 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 450 THREAD_LOCK_ASSERT(td, MA_OWNED); 451 452 pri = td->td_priority; 453 ts = td->td_sched; 454 TD_SET_RUNQ(td); 455 if (THREAD_CAN_MIGRATE(td)) { 456 tdq->tdq_transferable++; 457 ts->ts_flags |= TSF_XFERABLE; 458 } 459 if (pri < PRI_MIN_BATCH) { 460 ts->ts_runq = &tdq->tdq_realtime; 461 } else if (pri <= PRI_MAX_BATCH) { 462 ts->ts_runq = &tdq->tdq_timeshare; 463 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH, 464 ("Invalid priority %d on timeshare runq", pri)); 465 /* 466 * This queue contains only priorities between MIN and MAX 467 * realtime. Use the whole queue to represent these values. 468 */ 469 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 470 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE; 471 pri = (pri + tdq->tdq_idx) % RQ_NQS; 472 /* 473 * This effectively shortens the queue by one so we 474 * can have a one slot difference between idx and 475 * ridx while we wait for threads to drain. 476 */ 477 if (tdq->tdq_ridx != tdq->tdq_idx && 478 pri == tdq->tdq_ridx) 479 pri = (unsigned char)(pri - 1) % RQ_NQS; 480 } else 481 pri = tdq->tdq_ridx; 482 runq_add_pri(ts->ts_runq, td, pri, flags); 483 return; 484 } else 485 ts->ts_runq = &tdq->tdq_idle; 486 runq_add(ts->ts_runq, td, flags); 487} 488 489/* 490 * Remove a thread from a run-queue. This typically happens when a thread 491 * is selected to run. Running threads are not on the queue and the 492 * transferable count does not reflect them. 493 */ 494static __inline void 495tdq_runq_rem(struct tdq *tdq, struct thread *td) 496{ 497 struct td_sched *ts; 498 499 ts = td->td_sched; 500 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 501 KASSERT(ts->ts_runq != NULL, 502 ("tdq_runq_remove: thread %p null ts_runq", td)); 503 if (ts->ts_flags & TSF_XFERABLE) { 504 tdq->tdq_transferable--; 505 ts->ts_flags &= ~TSF_XFERABLE; 506 } 507 if (ts->ts_runq == &tdq->tdq_timeshare) { 508 if (tdq->tdq_idx != tdq->tdq_ridx) 509 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx); 510 else 511 runq_remove_idx(ts->ts_runq, td, NULL); 512 } else 513 runq_remove(ts->ts_runq, td); 514} 515 516/* 517 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 518 * for this thread to the referenced thread queue. 519 */ 520static void 521tdq_load_add(struct tdq *tdq, struct thread *td) 522{ 523 524 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 525 THREAD_LOCK_ASSERT(td, MA_OWNED); 526 527 tdq->tdq_load++; 528 if ((td->td_flags & TDF_NOLOAD) == 0) 529 tdq->tdq_sysload++; 530 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 531 SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load); 532} 533 534/* 535 * Remove the load from a thread that is transitioning to a sleep state or 536 * exiting. 537 */ 538static void 539tdq_load_rem(struct tdq *tdq, struct thread *td) 540{ 541 542 THREAD_LOCK_ASSERT(td, MA_OWNED); 543 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 544 KASSERT(tdq->tdq_load != 0, 545 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 546 547 tdq->tdq_load--; 548 if ((td->td_flags & TDF_NOLOAD) == 0) 549 tdq->tdq_sysload--; 550 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 551 SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load); 552} 553 554/* 555 * Set lowpri to its exact value by searching the run-queue and 556 * evaluating curthread. curthread may be passed as an optimization. 557 */ 558static void 559tdq_setlowpri(struct tdq *tdq, struct thread *ctd) 560{ 561 struct thread *td; 562 563 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 564 if (ctd == NULL) 565 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 566 td = tdq_choose(tdq); 567 if (td == NULL || td->td_priority > ctd->td_priority) 568 tdq->tdq_lowpri = ctd->td_priority; 569 else 570 tdq->tdq_lowpri = td->td_priority; 571} 572 573#ifdef SMP 574struct cpu_search { 575 cpuset_t cs_mask; 576 u_int cs_prefer; 577 int cs_pri; /* Min priority for low. */ 578 int cs_limit; /* Max load for low, min load for high. */ 579 int cs_cpu; 580 int cs_load; 581}; 582 583#define CPU_SEARCH_LOWEST 0x1 584#define CPU_SEARCH_HIGHEST 0x2 585#define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 586 587#define CPUSET_FOREACH(cpu, mask) \ 588 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \ 589 if (CPU_ISSET(cpu, &mask)) 590 591static __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low, 592 struct cpu_search *high, const int match); 593int cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low); 594int cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high); 595int cpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 596 struct cpu_search *high); 597 598/* 599 * Search the tree of cpu_groups for the lowest or highest loaded cpu 600 * according to the match argument. This routine actually compares the 601 * load on all paths through the tree and finds the least loaded cpu on 602 * the least loaded path, which may differ from the least loaded cpu in 603 * the system. This balances work among caches and busses. 604 * 605 * This inline is instantiated in three forms below using constants for the 606 * match argument. It is reduced to the minimum set for each case. It is 607 * also recursive to the depth of the tree. 608 */ 609static __inline int 610cpu_search(const struct cpu_group *cg, struct cpu_search *low, 611 struct cpu_search *high, const int match) 612{ 613 struct cpu_search lgroup; 614 struct cpu_search hgroup; 615 cpuset_t cpumask; 616 struct cpu_group *child; 617 struct tdq *tdq; 618 int cpu, i, hload, lload, load, total, rnd; 619 620 total = 0; 621 cpumask = cg->cg_mask; 622 if (match & CPU_SEARCH_LOWEST) { 623 lload = INT_MAX; 624 low->cs_load = INT_MAX; 625 lgroup = *low; 626 } 627 if (match & CPU_SEARCH_HIGHEST) { 628 hload = -1; 629 high->cs_load = -1; 630 hgroup = *high; 631 } 632 633 /* Iterate through the child CPU groups and then remaining CPUs. */ 634 for (i = 0, cpu = 0; i <= cg->cg_children; ) { 635 if (i >= cg->cg_children) { 636 while (cpu <= mp_maxid && !CPU_ISSET(cpu, &cpumask)) 637 cpu++; 638 if (cpu > mp_maxid) 639 break; 640 child = NULL; 641 } else 642 child = &cg->cg_child[i]; 643 644 if (child) { /* Handle child CPU group. */ 645 CPU_NAND(&cpumask, &child->cg_mask); 646 switch (match) { 647 case CPU_SEARCH_LOWEST: 648 load = cpu_search_lowest(child, &lgroup); 649 break; 650 case CPU_SEARCH_HIGHEST: 651 load = cpu_search_highest(child, &hgroup); 652 break; 653 case CPU_SEARCH_BOTH: 654 load = cpu_search_both(child, &lgroup, &hgroup); 655 break; 656 } 657 } else { /* Handle child CPU. */ 658 tdq = TDQ_CPU(cpu); 659 load = tdq->tdq_load * 256; 660 rnd = DPCPU_SET(randomval, 661 DPCPU_GET(randomval) * 69069 + 5) >> 26; 662 if (match & CPU_SEARCH_LOWEST) { 663 if (cpu == low->cs_prefer) 664 load -= 64; 665 /* If that CPU is allowed and get data. */ 666 if (CPU_ISSET(cpu, &lgroup.cs_mask) && 667 tdq->tdq_lowpri > lgroup.cs_pri && 668 tdq->tdq_load <= lgroup.cs_limit) { 669 lgroup.cs_cpu = cpu; 670 lgroup.cs_load = load - rnd; 671 } 672 } 673 if (match & CPU_SEARCH_HIGHEST) 674 if (CPU_ISSET(cpu, &hgroup.cs_mask) && 675 tdq->tdq_load >= hgroup.cs_limit && 676 tdq->tdq_transferable) { 677 hgroup.cs_cpu = cpu; 678 hgroup.cs_load = load - rnd; 679 } 680 } 681 total += load; 682 683 /* We have info about child item. Compare it. */ 684 if (match & CPU_SEARCH_LOWEST) { 685 if (lgroup.cs_load != INT_MAX && 686 (load < lload || 687 (load == lload && lgroup.cs_load < low->cs_load))) { 688 lload = load; 689 low->cs_cpu = lgroup.cs_cpu; 690 low->cs_load = lgroup.cs_load; 691 } 692 } 693 if (match & CPU_SEARCH_HIGHEST) 694 if (hgroup.cs_load >= 0 && 695 (load > hload || 696 (load == hload && hgroup.cs_load > high->cs_load))) { 697 hload = load; 698 high->cs_cpu = hgroup.cs_cpu; 699 high->cs_load = hgroup.cs_load; 700 } 701 if (child) 702 i++; 703 else 704 cpu++; 705 } 706 return (total); 707} 708 709/* 710 * cpu_search instantiations must pass constants to maintain the inline 711 * optimization. 712 */ 713int 714cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low) 715{ 716 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 717} 718 719int 720cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high) 721{ 722 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 723} 724 725int 726cpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 727 struct cpu_search *high) 728{ 729 return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 730} 731 732/* 733 * Find the cpu with the least load via the least loaded path that has a 734 * lowpri greater than pri pri. A pri of -1 indicates any priority is 735 * acceptable. 736 */ 737static inline int 738sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload, 739 int prefer) 740{ 741 struct cpu_search low; 742 743 low.cs_cpu = -1; 744 low.cs_prefer = prefer; 745 low.cs_mask = mask; 746 low.cs_pri = pri; 747 low.cs_limit = maxload; 748 cpu_search_lowest(cg, &low); 749 return low.cs_cpu; 750} 751 752/* 753 * Find the cpu with the highest load via the highest loaded path. 754 */ 755static inline int 756sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload) 757{ 758 struct cpu_search high; 759 760 high.cs_cpu = -1; 761 high.cs_mask = mask; 762 high.cs_limit = minload; 763 cpu_search_highest(cg, &high); 764 return high.cs_cpu; 765} 766 767/* 768 * Simultaneously find the highest and lowest loaded cpu reachable via 769 * cg. 770 */ 771static inline void 772sched_both(const struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu) 773{ 774 struct cpu_search high; 775 struct cpu_search low; 776 777 low.cs_cpu = -1; 778 low.cs_prefer = -1; 779 low.cs_pri = -1; 780 low.cs_limit = INT_MAX; 781 low.cs_mask = mask; 782 high.cs_cpu = -1; 783 high.cs_limit = -1; 784 high.cs_mask = mask; 785 cpu_search_both(cg, &low, &high); 786 *lowcpu = low.cs_cpu; 787 *highcpu = high.cs_cpu; 788 return; 789} 790 791static void 792sched_balance_group(struct cpu_group *cg) 793{ 794 cpuset_t hmask, lmask; 795 int high, low, anylow; 796 797 CPU_FILL(&hmask); 798 for (;;) { 799 high = sched_highest(cg, hmask, 1); 800 /* Stop if there is no more CPU with transferrable threads. */ 801 if (high == -1) 802 break; 803 CPU_CLR(high, &hmask); 804 CPU_COPY(&hmask, &lmask); 805 /* Stop if there is no more CPU left for low. */ 806 if (CPU_EMPTY(&lmask)) 807 break; 808 anylow = 1; 809nextlow: 810 low = sched_lowest(cg, lmask, -1, 811 TDQ_CPU(high)->tdq_load - 1, high); 812 /* Stop if we looked well and found no less loaded CPU. */ 813 if (anylow && low == -1) 814 break; 815 /* Go to next high if we found no less loaded CPU. */ 816 if (low == -1) 817 continue; 818 /* Transfer thread from high to low. */ 819 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) { 820 /* CPU that got thread can no longer be a donor. */ 821 CPU_CLR(low, &hmask); 822 } else { 823 /* 824 * If failed, then there is no threads on high 825 * that can run on this low. Drop low from low 826 * mask and look for different one. 827 */ 828 CPU_CLR(low, &lmask); 829 anylow = 0; 830 goto nextlow; 831 } 832 } 833} 834 835static void 836sched_balance(void) 837{ 838 struct tdq *tdq; 839 840 /* 841 * Select a random time between .5 * balance_interval and 842 * 1.5 * balance_interval. 843 */ 844 balance_ticks = max(balance_interval / 2, 1); 845 balance_ticks += random() % balance_interval; 846 if (smp_started == 0 || rebalance == 0) 847 return; 848 tdq = TDQ_SELF(); 849 TDQ_UNLOCK(tdq); 850 sched_balance_group(cpu_top); 851 TDQ_LOCK(tdq); 852} 853 854/* 855 * Lock two thread queues using their address to maintain lock order. 856 */ 857static void 858tdq_lock_pair(struct tdq *one, struct tdq *two) 859{ 860 if (one < two) { 861 TDQ_LOCK(one); 862 TDQ_LOCK_FLAGS(two, MTX_DUPOK); 863 } else { 864 TDQ_LOCK(two); 865 TDQ_LOCK_FLAGS(one, MTX_DUPOK); 866 } 867} 868 869/* 870 * Unlock two thread queues. Order is not important here. 871 */ 872static void 873tdq_unlock_pair(struct tdq *one, struct tdq *two) 874{ 875 TDQ_UNLOCK(one); 876 TDQ_UNLOCK(two); 877} 878 879/* 880 * Transfer load between two imbalanced thread queues. 881 */ 882static int 883sched_balance_pair(struct tdq *high, struct tdq *low) 884{ 885 int moved; 886 int cpu; 887 888 tdq_lock_pair(high, low); 889 moved = 0; 890 /* 891 * Determine what the imbalance is and then adjust that to how many 892 * threads we actually have to give up (transferable). 893 */ 894 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load && 895 (moved = tdq_move(high, low)) > 0) { 896 /* 897 * In case the target isn't the current cpu IPI it to force a 898 * reschedule with the new workload. 899 */ 900 cpu = TDQ_ID(low); 901 sched_pin(); 902 if (cpu != PCPU_GET(cpuid)) 903 ipi_cpu(cpu, IPI_PREEMPT); 904 sched_unpin(); 905 } 906 tdq_unlock_pair(high, low); 907 return (moved); 908} 909 910/* 911 * Move a thread from one thread queue to another. 912 */ 913static int 914tdq_move(struct tdq *from, struct tdq *to) 915{ 916 struct td_sched *ts; 917 struct thread *td; 918 struct tdq *tdq; 919 int cpu; 920 921 TDQ_LOCK_ASSERT(from, MA_OWNED); 922 TDQ_LOCK_ASSERT(to, MA_OWNED); 923 924 tdq = from; 925 cpu = TDQ_ID(to); 926 td = tdq_steal(tdq, cpu); 927 if (td == NULL) 928 return (0); 929 ts = td->td_sched; 930 /* 931 * Although the run queue is locked the thread may be blocked. Lock 932 * it to clear this and acquire the run-queue lock. 933 */ 934 thread_lock(td); 935 /* Drop recursive lock on from acquired via thread_lock(). */ 936 TDQ_UNLOCK(from); 937 sched_rem(td); 938 ts->ts_cpu = cpu; 939 td->td_lock = TDQ_LOCKPTR(to); 940 tdq_add(to, td, SRQ_YIELDING); 941 return (1); 942} 943 944/* 945 * This tdq has idled. Try to steal a thread from another cpu and switch 946 * to it. 947 */ 948static int 949tdq_idled(struct tdq *tdq) 950{ 951 struct cpu_group *cg; 952 struct tdq *steal; 953 cpuset_t mask; 954 int thresh; 955 int cpu; 956 957 if (smp_started == 0 || steal_idle == 0) 958 return (1); 959 CPU_FILL(&mask); 960 CPU_CLR(PCPU_GET(cpuid), &mask); 961 /* We don't want to be preempted while we're iterating. */ 962 spinlock_enter(); 963 for (cg = tdq->tdq_cg; cg != NULL; ) { 964 if ((cg->cg_flags & CG_FLAG_THREAD) == 0) 965 thresh = steal_thresh; 966 else 967 thresh = 1; 968 cpu = sched_highest(cg, mask, thresh); 969 if (cpu == -1) { 970 cg = cg->cg_parent; 971 continue; 972 } 973 steal = TDQ_CPU(cpu); 974 CPU_CLR(cpu, &mask); 975 tdq_lock_pair(tdq, steal); 976 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 977 tdq_unlock_pair(tdq, steal); 978 continue; 979 } 980 /* 981 * If a thread was added while interrupts were disabled don't 982 * steal one here. If we fail to acquire one due to affinity 983 * restrictions loop again with this cpu removed from the 984 * set. 985 */ 986 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 987 tdq_unlock_pair(tdq, steal); 988 continue; 989 } 990 spinlock_exit(); 991 TDQ_UNLOCK(steal); 992 mi_switch(SW_VOL | SWT_IDLE, NULL); 993 thread_unlock(curthread); 994 995 return (0); 996 } 997 spinlock_exit(); 998 return (1); 999} 1000 1001/* 1002 * Notify a remote cpu of new work. Sends an IPI if criteria are met. 1003 */ 1004static void 1005tdq_notify(struct tdq *tdq, struct thread *td) 1006{ 1007 struct thread *ctd; 1008 int pri; 1009 int cpu; 1010 1011 if (tdq->tdq_ipipending) 1012 return; 1013 cpu = td->td_sched->ts_cpu; 1014 pri = td->td_priority; 1015 ctd = pcpu_find(cpu)->pc_curthread; 1016 if (!sched_shouldpreempt(pri, ctd->td_priority, 1)) 1017 return; 1018 if (TD_IS_IDLETHREAD(ctd)) { 1019 /* 1020 * If the MD code has an idle wakeup routine try that before 1021 * falling back to IPI. 1022 */ 1023 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu)) 1024 return; 1025 } 1026 tdq->tdq_ipipending = 1; 1027 ipi_cpu(cpu, IPI_PREEMPT); 1028} 1029 1030/* 1031 * Steals load from a timeshare queue. Honors the rotating queue head 1032 * index. 1033 */ 1034static struct thread * 1035runq_steal_from(struct runq *rq, int cpu, u_char start) 1036{ 1037 struct rqbits *rqb; 1038 struct rqhead *rqh; 1039 struct thread *td, *first; 1040 int bit; 1041 int pri; 1042 int i; 1043 1044 rqb = &rq->rq_status; 1045 bit = start & (RQB_BPW -1); 1046 pri = 0; 1047 first = NULL; 1048again: 1049 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 1050 if (rqb->rqb_bits[i] == 0) 1051 continue; 1052 if (bit != 0) { 1053 for (pri = bit; pri < RQB_BPW; pri++) 1054 if (rqb->rqb_bits[i] & (1ul << pri)) 1055 break; 1056 if (pri >= RQB_BPW) 1057 continue; 1058 } else 1059 pri = RQB_FFS(rqb->rqb_bits[i]); 1060 pri += (i << RQB_L2BPW); 1061 rqh = &rq->rq_queues[pri]; 1062 TAILQ_FOREACH(td, rqh, td_runq) { 1063 if (first && THREAD_CAN_MIGRATE(td) && 1064 THREAD_CAN_SCHED(td, cpu)) 1065 return (td); 1066 first = td; 1067 } 1068 } 1069 if (start != 0) { 1070 start = 0; 1071 goto again; 1072 } 1073 1074 if (first && THREAD_CAN_MIGRATE(first) && 1075 THREAD_CAN_SCHED(first, cpu)) 1076 return (first); 1077 return (NULL); 1078} 1079 1080/* 1081 * Steals load from a standard linear queue. 1082 */ 1083static struct thread * 1084runq_steal(struct runq *rq, int cpu) 1085{ 1086 struct rqhead *rqh; 1087 struct rqbits *rqb; 1088 struct thread *td; 1089 int word; 1090 int bit; 1091 1092 rqb = &rq->rq_status; 1093 for (word = 0; word < RQB_LEN; word++) { 1094 if (rqb->rqb_bits[word] == 0) 1095 continue; 1096 for (bit = 0; bit < RQB_BPW; bit++) { 1097 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1098 continue; 1099 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1100 TAILQ_FOREACH(td, rqh, td_runq) 1101 if (THREAD_CAN_MIGRATE(td) && 1102 THREAD_CAN_SCHED(td, cpu)) 1103 return (td); 1104 } 1105 } 1106 return (NULL); 1107} 1108 1109/* 1110 * Attempt to steal a thread in priority order from a thread queue. 1111 */ 1112static struct thread * 1113tdq_steal(struct tdq *tdq, int cpu) 1114{ 1115 struct thread *td; 1116 1117 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1118 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1119 return (td); 1120 if ((td = runq_steal_from(&tdq->tdq_timeshare, 1121 cpu, tdq->tdq_ridx)) != NULL) 1122 return (td); 1123 return (runq_steal(&tdq->tdq_idle, cpu)); 1124} 1125 1126/* 1127 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1128 * current lock and returns with the assigned queue locked. 1129 */ 1130static inline struct tdq * 1131sched_setcpu(struct thread *td, int cpu, int flags) 1132{ 1133 1134 struct tdq *tdq; 1135 1136 THREAD_LOCK_ASSERT(td, MA_OWNED); 1137 tdq = TDQ_CPU(cpu); 1138 td->td_sched->ts_cpu = cpu; 1139 /* 1140 * If the lock matches just return the queue. 1141 */ 1142 if (td->td_lock == TDQ_LOCKPTR(tdq)) 1143 return (tdq); 1144#ifdef notyet 1145 /* 1146 * If the thread isn't running its lockptr is a 1147 * turnstile or a sleepqueue. We can just lock_set without 1148 * blocking. 1149 */ 1150 if (TD_CAN_RUN(td)) { 1151 TDQ_LOCK(tdq); 1152 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1153 return (tdq); 1154 } 1155#endif 1156 /* 1157 * The hard case, migration, we need to block the thread first to 1158 * prevent order reversals with other cpus locks. 1159 */ 1160 spinlock_enter(); 1161 thread_lock_block(td); 1162 TDQ_LOCK(tdq); 1163 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1164 spinlock_exit(); 1165 return (tdq); 1166} 1167 1168SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1169SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1170SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1171SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1172SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1173SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1174 1175static int 1176sched_pickcpu(struct thread *td, int flags) 1177{ 1178 struct cpu_group *cg, *ccg; 1179 struct td_sched *ts; 1180 struct tdq *tdq; 1181 cpuset_t mask; 1182 int cpu, pri, self; 1183 1184 self = PCPU_GET(cpuid); 1185 ts = td->td_sched; 1186 if (smp_started == 0) 1187 return (self); 1188 /* 1189 * Don't migrate a running thread from sched_switch(). 1190 */ 1191 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1192 return (ts->ts_cpu); 1193 /* 1194 * Prefer to run interrupt threads on the processors that generate 1195 * the interrupt. 1196 */ 1197 pri = td->td_priority; 1198 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1199 curthread->td_intr_nesting_level && ts->ts_cpu != self) { 1200 SCHED_STAT_INC(pickcpu_intrbind); 1201 ts->ts_cpu = self; 1202 if (TDQ_CPU(self)->tdq_lowpri > pri) { 1203 SCHED_STAT_INC(pickcpu_affinity); 1204 return (ts->ts_cpu); 1205 } 1206 } 1207 /* 1208 * If the thread can run on the last cpu and the affinity has not 1209 * expired or it is idle run it there. 1210 */ 1211 tdq = TDQ_CPU(ts->ts_cpu); 1212 cg = tdq->tdq_cg; 1213 if (THREAD_CAN_SCHED(td, ts->ts_cpu) && 1214 tdq->tdq_lowpri >= PRI_MIN_IDLE && 1215 SCHED_AFFINITY(ts, CG_SHARE_L2)) { 1216 if (cg->cg_flags & CG_FLAG_THREAD) { 1217 CPUSET_FOREACH(cpu, cg->cg_mask) { 1218 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) 1219 break; 1220 } 1221 } else 1222 cpu = INT_MAX; 1223 if (cpu > mp_maxid) { 1224 SCHED_STAT_INC(pickcpu_idle_affinity); 1225 return (ts->ts_cpu); 1226 } 1227 } 1228 /* 1229 * Search for the last level cache CPU group in the tree. 1230 * Skip caches with expired affinity time and SMT groups. 1231 * Affinity to higher level caches will be handled less aggressively. 1232 */ 1233 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) { 1234 if (cg->cg_flags & CG_FLAG_THREAD) 1235 continue; 1236 if (!SCHED_AFFINITY(ts, cg->cg_level)) 1237 continue; 1238 ccg = cg; 1239 } 1240 if (ccg != NULL) 1241 cg = ccg; 1242 cpu = -1; 1243 /* Search the group for the less loaded idle CPU we can run now. */ 1244 mask = td->td_cpuset->cs_mask; 1245 if (cg != NULL && cg != cpu_top && 1246 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0) 1247 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE), 1248 INT_MAX, ts->ts_cpu); 1249 /* Search globally for the less loaded CPU we can run now. */ 1250 if (cpu == -1) 1251 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu); 1252 /* Search globally for the less loaded CPU. */ 1253 if (cpu == -1) 1254 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu); 1255 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1256 /* 1257 * Compare the lowest loaded cpu to current cpu. 1258 */ 1259 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1260 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE && 1261 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) { 1262 SCHED_STAT_INC(pickcpu_local); 1263 cpu = self; 1264 } else 1265 SCHED_STAT_INC(pickcpu_lowest); 1266 if (cpu != ts->ts_cpu) 1267 SCHED_STAT_INC(pickcpu_migration); 1268 return (cpu); 1269} 1270#endif 1271 1272/* 1273 * Pick the highest priority task we have and return it. 1274 */ 1275static struct thread * 1276tdq_choose(struct tdq *tdq) 1277{ 1278 struct thread *td; 1279 1280 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1281 td = runq_choose(&tdq->tdq_realtime); 1282 if (td != NULL) 1283 return (td); 1284 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1285 if (td != NULL) { 1286 KASSERT(td->td_priority >= PRI_MIN_BATCH, 1287 ("tdq_choose: Invalid priority on timeshare queue %d", 1288 td->td_priority)); 1289 return (td); 1290 } 1291 td = runq_choose(&tdq->tdq_idle); 1292 if (td != NULL) { 1293 KASSERT(td->td_priority >= PRI_MIN_IDLE, 1294 ("tdq_choose: Invalid priority on idle queue %d", 1295 td->td_priority)); 1296 return (td); 1297 } 1298 1299 return (NULL); 1300} 1301 1302/* 1303 * Initialize a thread queue. 1304 */ 1305static void 1306tdq_setup(struct tdq *tdq) 1307{ 1308 1309 if (bootverbose) 1310 printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1311 runq_init(&tdq->tdq_realtime); 1312 runq_init(&tdq->tdq_timeshare); 1313 runq_init(&tdq->tdq_idle); 1314 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1315 "sched lock %d", (int)TDQ_ID(tdq)); 1316 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1317 MTX_SPIN | MTX_RECURSE); 1318#ifdef KTR 1319 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname), 1320 "CPU %d load", (int)TDQ_ID(tdq)); 1321#endif 1322} 1323 1324#ifdef SMP 1325static void 1326sched_setup_smp(void) 1327{ 1328 struct tdq *tdq; 1329 int i; 1330 1331 cpu_top = smp_topo(); 1332 CPU_FOREACH(i) { 1333 tdq = TDQ_CPU(i); 1334 tdq_setup(tdq); 1335 tdq->tdq_cg = smp_topo_find(cpu_top, i); 1336 if (tdq->tdq_cg == NULL) 1337 panic("Can't find cpu group for %d\n", i); 1338 } 1339 balance_tdq = TDQ_SELF(); 1340 sched_balance(); 1341} 1342#endif 1343 1344/* 1345 * Setup the thread queues and initialize the topology based on MD 1346 * information. 1347 */ 1348static void 1349sched_setup(void *dummy) 1350{ 1351 struct tdq *tdq; 1352 1353 tdq = TDQ_SELF(); 1354#ifdef SMP 1355 sched_setup_smp(); 1356#else 1357 tdq_setup(tdq); 1358#endif 1359 /* 1360 * To avoid divide-by-zero, we set realstathz a dummy value 1361 * in case which sched_clock() called before sched_initticks(). 1362 */ 1363 realstathz = hz; 1364 sched_slice = (realstathz/10); /* ~100ms */ 1365 tickincr = 1 << SCHED_TICK_SHIFT; 1366 1367 /* Add thread0's load since it's running. */ 1368 TDQ_LOCK(tdq); 1369 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1370 tdq_load_add(tdq, &thread0); 1371 tdq->tdq_lowpri = thread0.td_priority; 1372 TDQ_UNLOCK(tdq); 1373} 1374 1375/* 1376 * This routine determines the tickincr after stathz and hz are setup. 1377 */ 1378/* ARGSUSED */ 1379static void 1380sched_initticks(void *dummy) 1381{ 1382 int incr; 1383 1384 realstathz = stathz ? stathz : hz; 1385 sched_slice = (realstathz/10); /* ~100ms */ 1386 1387 /* 1388 * tickincr is shifted out by 10 to avoid rounding errors due to 1389 * hz not being evenly divisible by stathz on all platforms. 1390 */ 1391 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1392 /* 1393 * This does not work for values of stathz that are more than 1394 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1395 */ 1396 if (incr == 0) 1397 incr = 1; 1398 tickincr = incr; 1399#ifdef SMP 1400 /* 1401 * Set the default balance interval now that we know 1402 * what realstathz is. 1403 */ 1404 balance_interval = realstathz; 1405 /* 1406 * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4. 1407 * This prevents excess thrashing on large machines and excess idle 1408 * on smaller machines. 1409 */ 1410 steal_thresh = min(fls(mp_ncpus) - 1, 3); 1411 affinity = SCHED_AFFINITY_DEFAULT; 1412#endif 1413 if (sched_idlespinthresh < 0) 1414 sched_idlespinthresh = max(16, 2 * hz / realstathz); 1415} 1416 1417 1418/* 1419 * This is the core of the interactivity algorithm. Determines a score based 1420 * on past behavior. It is the ratio of sleep time to run time scaled to 1421 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1422 * differs from the cpu usage because it does not account for time spent 1423 * waiting on a run-queue. Would be prettier if we had floating point. 1424 */ 1425static int 1426sched_interact_score(struct thread *td) 1427{ 1428 struct td_sched *ts; 1429 int div; 1430 1431 ts = td->td_sched; 1432 /* 1433 * The score is only needed if this is likely to be an interactive 1434 * task. Don't go through the expense of computing it if there's 1435 * no chance. 1436 */ 1437 if (sched_interact <= SCHED_INTERACT_HALF && 1438 ts->ts_runtime >= ts->ts_slptime) 1439 return (SCHED_INTERACT_HALF); 1440 1441 if (ts->ts_runtime > ts->ts_slptime) { 1442 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1443 return (SCHED_INTERACT_HALF + 1444 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1445 } 1446 if (ts->ts_slptime > ts->ts_runtime) { 1447 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1448 return (ts->ts_runtime / div); 1449 } 1450 /* runtime == slptime */ 1451 if (ts->ts_runtime) 1452 return (SCHED_INTERACT_HALF); 1453 1454 /* 1455 * This can happen if slptime and runtime are 0. 1456 */ 1457 return (0); 1458 1459} 1460 1461/* 1462 * Scale the scheduling priority according to the "interactivity" of this 1463 * process. 1464 */ 1465static void 1466sched_priority(struct thread *td) 1467{ 1468 int score; 1469 int pri; 1470 1471 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1472 return; 1473 /* 1474 * If the score is interactive we place the thread in the realtime 1475 * queue with a priority that is less than kernel and interrupt 1476 * priorities. These threads are not subject to nice restrictions. 1477 * 1478 * Scores greater than this are placed on the normal timeshare queue 1479 * where the priority is partially decided by the most recent cpu 1480 * utilization and the rest is decided by nice value. 1481 * 1482 * The nice value of the process has a linear effect on the calculated 1483 * score. Negative nice values make it easier for a thread to be 1484 * considered interactive. 1485 */ 1486 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice); 1487 if (score < sched_interact) { 1488 pri = PRI_MIN_INTERACT; 1489 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) / 1490 sched_interact) * score; 1491 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT, 1492 ("sched_priority: invalid interactive priority %d score %d", 1493 pri, score)); 1494 } else { 1495 pri = SCHED_PRI_MIN; 1496 if (td->td_sched->ts_ticks) 1497 pri += min(SCHED_PRI_TICKS(td->td_sched), 1498 SCHED_PRI_RANGE); 1499 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1500 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH, 1501 ("sched_priority: invalid priority %d: nice %d, " 1502 "ticks %d ftick %d ltick %d tick pri %d", 1503 pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1504 td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1505 SCHED_PRI_TICKS(td->td_sched))); 1506 } 1507 sched_user_prio(td, pri); 1508 1509 return; 1510} 1511 1512/* 1513 * This routine enforces a maximum limit on the amount of scheduling history 1514 * kept. It is called after either the slptime or runtime is adjusted. This 1515 * function is ugly due to integer math. 1516 */ 1517static void 1518sched_interact_update(struct thread *td) 1519{ 1520 struct td_sched *ts; 1521 u_int sum; 1522 1523 ts = td->td_sched; 1524 sum = ts->ts_runtime + ts->ts_slptime; 1525 if (sum < SCHED_SLP_RUN_MAX) 1526 return; 1527 /* 1528 * This only happens from two places: 1529 * 1) We have added an unusual amount of run time from fork_exit. 1530 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1531 */ 1532 if (sum > SCHED_SLP_RUN_MAX * 2) { 1533 if (ts->ts_runtime > ts->ts_slptime) { 1534 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1535 ts->ts_slptime = 1; 1536 } else { 1537 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1538 ts->ts_runtime = 1; 1539 } 1540 return; 1541 } 1542 /* 1543 * If we have exceeded by more than 1/5th then the algorithm below 1544 * will not bring us back into range. Dividing by two here forces 1545 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1546 */ 1547 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1548 ts->ts_runtime /= 2; 1549 ts->ts_slptime /= 2; 1550 return; 1551 } 1552 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1553 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1554} 1555 1556/* 1557 * Scale back the interactivity history when a child thread is created. The 1558 * history is inherited from the parent but the thread may behave totally 1559 * differently. For example, a shell spawning a compiler process. We want 1560 * to learn that the compiler is behaving badly very quickly. 1561 */ 1562static void 1563sched_interact_fork(struct thread *td) 1564{ 1565 int ratio; 1566 int sum; 1567 1568 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1569 if (sum > SCHED_SLP_RUN_FORK) { 1570 ratio = sum / SCHED_SLP_RUN_FORK; 1571 td->td_sched->ts_runtime /= ratio; 1572 td->td_sched->ts_slptime /= ratio; 1573 } 1574} 1575 1576/* 1577 * Called from proc0_init() to setup the scheduler fields. 1578 */ 1579void 1580schedinit(void) 1581{ 1582 1583 /* 1584 * Set up the scheduler specific parts of proc0. 1585 */ 1586 proc0.p_sched = NULL; /* XXX */ 1587 thread0.td_sched = &td_sched0; 1588 td_sched0.ts_ltick = ticks; 1589 td_sched0.ts_ftick = ticks; 1590 td_sched0.ts_slice = sched_slice; 1591} 1592 1593/* 1594 * This is only somewhat accurate since given many processes of the same 1595 * priority they will switch when their slices run out, which will be 1596 * at most sched_slice stathz ticks. 1597 */ 1598int 1599sched_rr_interval(void) 1600{ 1601 1602 /* Convert sched_slice to hz */ 1603 return (hz/(realstathz/sched_slice)); 1604} 1605 1606/* 1607 * Update the percent cpu tracking information when it is requested or 1608 * the total history exceeds the maximum. We keep a sliding history of 1609 * tick counts that slowly decays. This is less precise than the 4BSD 1610 * mechanism since it happens with less regular and frequent events. 1611 */ 1612static void 1613sched_pctcpu_update(struct td_sched *ts, int run) 1614{ 1615 int t = ticks; 1616 1617 if (t - ts->ts_ltick >= SCHED_TICK_TARG) { 1618 ts->ts_ticks = 0; 1619 ts->ts_ftick = t - SCHED_TICK_TARG; 1620 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) { 1621 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) * 1622 (ts->ts_ltick - (t - SCHED_TICK_TARG)); 1623 ts->ts_ftick = t - SCHED_TICK_TARG; 1624 } 1625 if (run) 1626 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT; 1627 ts->ts_ltick = t; 1628} 1629 1630/* 1631 * Adjust the priority of a thread. Move it to the appropriate run-queue 1632 * if necessary. This is the back-end for several priority related 1633 * functions. 1634 */ 1635static void 1636sched_thread_priority(struct thread *td, u_char prio) 1637{ 1638 struct td_sched *ts; 1639 struct tdq *tdq; 1640 int oldpri; 1641 1642 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio", 1643 "prio:%d", td->td_priority, "new prio:%d", prio, 1644 KTR_ATTR_LINKED, sched_tdname(curthread)); 1645 SDT_PROBE3(sched, , , change_pri, td, td->td_proc, prio); 1646 if (td != curthread && prio > td->td_priority) { 1647 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 1648 "lend prio", "prio:%d", td->td_priority, "new prio:%d", 1649 prio, KTR_ATTR_LINKED, sched_tdname(td)); 1650 SDT_PROBE4(sched, , , lend_pri, td, td->td_proc, prio, 1651 curthread); 1652 } 1653 ts = td->td_sched; 1654 THREAD_LOCK_ASSERT(td, MA_OWNED); 1655 if (td->td_priority == prio) 1656 return; 1657 /* 1658 * If the priority has been elevated due to priority 1659 * propagation, we may have to move ourselves to a new 1660 * queue. This could be optimized to not re-add in some 1661 * cases. 1662 */ 1663 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1664 sched_rem(td); 1665 td->td_priority = prio; 1666 sched_add(td, SRQ_BORROWING); 1667 return; 1668 } 1669 /* 1670 * If the thread is currently running we may have to adjust the lowpri 1671 * information so other cpus are aware of our current priority. 1672 */ 1673 if (TD_IS_RUNNING(td)) { 1674 tdq = TDQ_CPU(ts->ts_cpu); 1675 oldpri = td->td_priority; 1676 td->td_priority = prio; 1677 if (prio < tdq->tdq_lowpri) 1678 tdq->tdq_lowpri = prio; 1679 else if (tdq->tdq_lowpri == oldpri) 1680 tdq_setlowpri(tdq, td); 1681 return; 1682 } 1683 td->td_priority = prio; 1684} 1685 1686/* 1687 * Update a thread's priority when it is lent another thread's 1688 * priority. 1689 */ 1690void 1691sched_lend_prio(struct thread *td, u_char prio) 1692{ 1693 1694 td->td_flags |= TDF_BORROWING; 1695 sched_thread_priority(td, prio); 1696} 1697 1698/* 1699 * Restore a thread's priority when priority propagation is 1700 * over. The prio argument is the minimum priority the thread 1701 * needs to have to satisfy other possible priority lending 1702 * requests. If the thread's regular priority is less 1703 * important than prio, the thread will keep a priority boost 1704 * of prio. 1705 */ 1706void 1707sched_unlend_prio(struct thread *td, u_char prio) 1708{ 1709 u_char base_pri; 1710 1711 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1712 td->td_base_pri <= PRI_MAX_TIMESHARE) 1713 base_pri = td->td_user_pri; 1714 else 1715 base_pri = td->td_base_pri; 1716 if (prio >= base_pri) { 1717 td->td_flags &= ~TDF_BORROWING; 1718 sched_thread_priority(td, base_pri); 1719 } else 1720 sched_lend_prio(td, prio); 1721} 1722 1723/* 1724 * Standard entry for setting the priority to an absolute value. 1725 */ 1726void 1727sched_prio(struct thread *td, u_char prio) 1728{ 1729 u_char oldprio; 1730 1731 /* First, update the base priority. */ 1732 td->td_base_pri = prio; 1733 1734 /* 1735 * If the thread is borrowing another thread's priority, don't 1736 * ever lower the priority. 1737 */ 1738 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1739 return; 1740 1741 /* Change the real priority. */ 1742 oldprio = td->td_priority; 1743 sched_thread_priority(td, prio); 1744 1745 /* 1746 * If the thread is on a turnstile, then let the turnstile update 1747 * its state. 1748 */ 1749 if (TD_ON_LOCK(td) && oldprio != prio) 1750 turnstile_adjust(td, oldprio); 1751} 1752 1753/* 1754 * Set the base user priority, does not effect current running priority. 1755 */ 1756void 1757sched_user_prio(struct thread *td, u_char prio) 1758{ 1759 1760 td->td_base_user_pri = prio; 1761 if (td->td_lend_user_pri <= prio) 1762 return; 1763 td->td_user_pri = prio; 1764} 1765 1766void 1767sched_lend_user_prio(struct thread *td, u_char prio) 1768{ 1769 1770 THREAD_LOCK_ASSERT(td, MA_OWNED); 1771 td->td_lend_user_pri = prio; 1772 td->td_user_pri = min(prio, td->td_base_user_pri); 1773 if (td->td_priority > td->td_user_pri) 1774 sched_prio(td, td->td_user_pri); 1775 else if (td->td_priority != td->td_user_pri) 1776 td->td_flags |= TDF_NEEDRESCHED; 1777} 1778 1779/* 1780 * Handle migration from sched_switch(). This happens only for 1781 * cpu binding. 1782 */ 1783static struct mtx * 1784sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1785{ 1786 struct tdq *tdn; 1787 1788 tdn = TDQ_CPU(td->td_sched->ts_cpu); 1789#ifdef SMP 1790 tdq_load_rem(tdq, td); 1791 /* 1792 * Do the lock dance required to avoid LOR. We grab an extra 1793 * spinlock nesting to prevent preemption while we're 1794 * not holding either run-queue lock. 1795 */ 1796 spinlock_enter(); 1797 thread_lock_block(td); /* This releases the lock on tdq. */ 1798 1799 /* 1800 * Acquire both run-queue locks before placing the thread on the new 1801 * run-queue to avoid deadlocks created by placing a thread with a 1802 * blocked lock on the run-queue of a remote processor. The deadlock 1803 * occurs when a third processor attempts to lock the two queues in 1804 * question while the target processor is spinning with its own 1805 * run-queue lock held while waiting for the blocked lock to clear. 1806 */ 1807 tdq_lock_pair(tdn, tdq); 1808 tdq_add(tdn, td, flags); 1809 tdq_notify(tdn, td); 1810 TDQ_UNLOCK(tdn); 1811 spinlock_exit(); 1812#endif 1813 return (TDQ_LOCKPTR(tdn)); 1814} 1815 1816/* 1817 * Variadic version of thread_lock_unblock() that does not assume td_lock 1818 * is blocked. 1819 */ 1820static inline void 1821thread_unblock_switch(struct thread *td, struct mtx *mtx) 1822{ 1823 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1824 (uintptr_t)mtx); 1825} 1826 1827/* 1828 * Switch threads. This function has to handle threads coming in while 1829 * blocked for some reason, running, or idle. It also must deal with 1830 * migrating a thread from one queue to another as running threads may 1831 * be assigned elsewhere via binding. 1832 */ 1833void 1834sched_switch(struct thread *td, struct thread *newtd, int flags) 1835{ 1836 struct tdq *tdq; 1837 struct td_sched *ts; 1838 struct mtx *mtx; 1839 int srqflag; 1840 int cpuid; 1841 1842 THREAD_LOCK_ASSERT(td, MA_OWNED); 1843 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1844 1845 cpuid = PCPU_GET(cpuid); 1846 tdq = TDQ_CPU(cpuid); 1847 ts = td->td_sched; 1848 mtx = td->td_lock; 1849 sched_pctcpu_update(ts, 1); 1850 ts->ts_rltick = ticks; 1851 td->td_lastcpu = td->td_oncpu; 1852 td->td_oncpu = NOCPU; 1853 if (!(flags & SW_PREEMPT)) 1854 td->td_flags &= ~TDF_NEEDRESCHED; 1855 td->td_owepreempt = 0; 1856 tdq->tdq_switchcnt++; 1857 /* 1858 * The lock pointer in an idle thread should never change. Reset it 1859 * to CAN_RUN as well. 1860 */ 1861 if (TD_IS_IDLETHREAD(td)) { 1862 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1863 TD_SET_CAN_RUN(td); 1864 } else if (TD_IS_RUNNING(td)) { 1865 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1866 srqflag = (flags & SW_PREEMPT) ? 1867 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1868 SRQ_OURSELF|SRQ_YIELDING; 1869#ifdef SMP 1870 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu)) 1871 ts->ts_cpu = sched_pickcpu(td, 0); 1872#endif 1873 if (ts->ts_cpu == cpuid) 1874 tdq_runq_add(tdq, td, srqflag); 1875 else { 1876 KASSERT(THREAD_CAN_MIGRATE(td) || 1877 (ts->ts_flags & TSF_BOUND) != 0, 1878 ("Thread %p shouldn't migrate", td)); 1879 mtx = sched_switch_migrate(tdq, td, srqflag); 1880 } 1881 } else { 1882 /* This thread must be going to sleep. */ 1883 TDQ_LOCK(tdq); 1884 mtx = thread_lock_block(td); 1885 tdq_load_rem(tdq, td); 1886 } 1887 /* 1888 * We enter here with the thread blocked and assigned to the 1889 * appropriate cpu run-queue or sleep-queue and with the current 1890 * thread-queue locked. 1891 */ 1892 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1893 newtd = choosethread(); 1894 /* 1895 * Call the MD code to switch contexts if necessary. 1896 */ 1897 if (td != newtd) { 1898#ifdef HWPMC_HOOKS 1899 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1900 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1901#endif 1902 SDT_PROBE2(sched, , , off_cpu, td, td->td_proc); 1903 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1904 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1905 sched_pctcpu_update(newtd->td_sched, 0); 1906 1907#ifdef KDTRACE_HOOKS 1908 /* 1909 * If DTrace has set the active vtime enum to anything 1910 * other than INACTIVE (0), then it should have set the 1911 * function to call. 1912 */ 1913 if (dtrace_vtime_active) 1914 (*dtrace_vtime_switch_func)(newtd); 1915#endif 1916 1917 cpu_switch(td, newtd, mtx); 1918 /* 1919 * We may return from cpu_switch on a different cpu. However, 1920 * we always return with td_lock pointing to the current cpu's 1921 * run queue lock. 1922 */ 1923 cpuid = PCPU_GET(cpuid); 1924 tdq = TDQ_CPU(cpuid); 1925 lock_profile_obtain_lock_success( 1926 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1927 1928 SDT_PROBE0(sched, , , on_cpu); 1929#ifdef HWPMC_HOOKS 1930 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1931 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1932#endif 1933 } else { 1934 thread_unblock_switch(td, mtx); 1935 SDT_PROBE0(sched, , , remain_cpu); 1936 } 1937 /* 1938 * Assert that all went well and return. 1939 */ 1940 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1941 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1942 td->td_oncpu = cpuid; 1943} 1944 1945/* 1946 * Adjust thread priorities as a result of a nice request. 1947 */ 1948void 1949sched_nice(struct proc *p, int nice) 1950{ 1951 struct thread *td; 1952 1953 PROC_LOCK_ASSERT(p, MA_OWNED); 1954 1955 p->p_nice = nice; 1956 FOREACH_THREAD_IN_PROC(p, td) { 1957 thread_lock(td); 1958 sched_priority(td); 1959 sched_prio(td, td->td_base_user_pri); 1960 thread_unlock(td); 1961 } 1962} 1963 1964/* 1965 * Record the sleep time for the interactivity scorer. 1966 */ 1967void 1968sched_sleep(struct thread *td, int prio) 1969{ 1970 1971 THREAD_LOCK_ASSERT(td, MA_OWNED); 1972 1973 td->td_slptick = ticks; 1974 if (TD_IS_SUSPENDED(td) || prio >= PSOCK) 1975 td->td_flags |= TDF_CANSWAP; 1976 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1977 return; 1978 if (static_boost == 1 && prio) 1979 sched_prio(td, prio); 1980 else if (static_boost && td->td_priority > static_boost) 1981 sched_prio(td, static_boost); 1982} 1983 1984/* 1985 * Schedule a thread to resume execution and record how long it voluntarily 1986 * slept. We also update the pctcpu, interactivity, and priority. 1987 */ 1988void 1989sched_wakeup(struct thread *td) 1990{ 1991 struct td_sched *ts; 1992 int slptick; 1993 1994 THREAD_LOCK_ASSERT(td, MA_OWNED); 1995 ts = td->td_sched; 1996 td->td_flags &= ~TDF_CANSWAP; 1997 /* 1998 * If we slept for more than a tick update our interactivity and 1999 * priority. 2000 */ 2001 slptick = td->td_slptick; 2002 td->td_slptick = 0; 2003 if (slptick && slptick != ticks) { 2004 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT; 2005 sched_interact_update(td); 2006 sched_pctcpu_update(ts, 0); 2007 } 2008 /* Reset the slice value after we sleep. */ 2009 ts->ts_slice = sched_slice; 2010 sched_add(td, SRQ_BORING); 2011} 2012 2013/* 2014 * Penalize the parent for creating a new child and initialize the child's 2015 * priority. 2016 */ 2017void 2018sched_fork(struct thread *td, struct thread *child) 2019{ 2020 THREAD_LOCK_ASSERT(td, MA_OWNED); 2021 sched_pctcpu_update(td->td_sched, 1); 2022 sched_fork_thread(td, child); 2023 /* 2024 * Penalize the parent and child for forking. 2025 */ 2026 sched_interact_fork(child); 2027 sched_priority(child); 2028 td->td_sched->ts_runtime += tickincr; 2029 sched_interact_update(td); 2030 sched_priority(td); 2031} 2032 2033/* 2034 * Fork a new thread, may be within the same process. 2035 */ 2036void 2037sched_fork_thread(struct thread *td, struct thread *child) 2038{ 2039 struct td_sched *ts; 2040 struct td_sched *ts2; 2041 2042 THREAD_LOCK_ASSERT(td, MA_OWNED); 2043 /* 2044 * Initialize child. 2045 */ 2046 ts = td->td_sched; 2047 ts2 = child->td_sched; 2048 child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 2049 child->td_cpuset = cpuset_ref(td->td_cpuset); 2050 ts2->ts_cpu = ts->ts_cpu; 2051 ts2->ts_flags = 0; 2052 /* 2053 * Grab our parents cpu estimation information. 2054 */ 2055 ts2->ts_ticks = ts->ts_ticks; 2056 ts2->ts_ltick = ts->ts_ltick; 2057 ts2->ts_ftick = ts->ts_ftick; 2058 /* 2059 * Do not inherit any borrowed priority from the parent. 2060 */ 2061 child->td_priority = child->td_base_pri; 2062 /* 2063 * And update interactivity score. 2064 */ 2065 ts2->ts_slptime = ts->ts_slptime; 2066 ts2->ts_runtime = ts->ts_runtime; 2067 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 2068#ifdef KTR 2069 bzero(ts2->ts_name, sizeof(ts2->ts_name)); 2070#endif 2071} 2072 2073/* 2074 * Adjust the priority class of a thread. 2075 */ 2076void 2077sched_class(struct thread *td, int class) 2078{ 2079 2080 THREAD_LOCK_ASSERT(td, MA_OWNED); 2081 if (td->td_pri_class == class) 2082 return; 2083 td->td_pri_class = class; 2084} 2085 2086/* 2087 * Return some of the child's priority and interactivity to the parent. 2088 */ 2089void 2090sched_exit(struct proc *p, struct thread *child) 2091{ 2092 struct thread *td; 2093 2094 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit", 2095 "prio:%d", child->td_priority); 2096 PROC_LOCK_ASSERT(p, MA_OWNED); 2097 td = FIRST_THREAD_IN_PROC(p); 2098 sched_exit_thread(td, child); 2099} 2100 2101/* 2102 * Penalize another thread for the time spent on this one. This helps to 2103 * worsen the priority and interactivity of processes which schedule batch 2104 * jobs such as make. This has little effect on the make process itself but 2105 * causes new processes spawned by it to receive worse scores immediately. 2106 */ 2107void 2108sched_exit_thread(struct thread *td, struct thread *child) 2109{ 2110 2111 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit", 2112 "prio:%d", child->td_priority); 2113 /* 2114 * Give the child's runtime to the parent without returning the 2115 * sleep time as a penalty to the parent. This causes shells that 2116 * launch expensive things to mark their children as expensive. 2117 */ 2118 thread_lock(td); 2119 td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2120 sched_interact_update(td); 2121 sched_priority(td); 2122 thread_unlock(td); 2123} 2124 2125void 2126sched_preempt(struct thread *td) 2127{ 2128 struct tdq *tdq; 2129 2130 SDT_PROBE2(sched, , , surrender, td, td->td_proc); 2131 2132 thread_lock(td); 2133 tdq = TDQ_SELF(); 2134 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2135 tdq->tdq_ipipending = 0; 2136 if (td->td_priority > tdq->tdq_lowpri) { 2137 int flags; 2138 2139 flags = SW_INVOL | SW_PREEMPT; 2140 if (td->td_critnest > 1) 2141 td->td_owepreempt = 1; 2142 else if (TD_IS_IDLETHREAD(td)) 2143 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL); 2144 else 2145 mi_switch(flags | SWT_REMOTEPREEMPT, NULL); 2146 } 2147 thread_unlock(td); 2148} 2149 2150/* 2151 * Fix priorities on return to user-space. Priorities may be elevated due 2152 * to static priorities in msleep() or similar. 2153 */ 2154void 2155sched_userret(struct thread *td) 2156{ 2157 /* 2158 * XXX we cheat slightly on the locking here to avoid locking in 2159 * the usual case. Setting td_priority here is essentially an 2160 * incomplete workaround for not setting it properly elsewhere. 2161 * Now that some interrupt handlers are threads, not setting it 2162 * properly elsewhere can clobber it in the window between setting 2163 * it here and returning to user mode, so don't waste time setting 2164 * it perfectly here. 2165 */ 2166 KASSERT((td->td_flags & TDF_BORROWING) == 0, 2167 ("thread with borrowed priority returning to userland")); 2168 if (td->td_priority != td->td_user_pri) { 2169 thread_lock(td); 2170 td->td_priority = td->td_user_pri; 2171 td->td_base_pri = td->td_user_pri; 2172 tdq_setlowpri(TDQ_SELF(), td); 2173 thread_unlock(td); 2174 } 2175} 2176 2177/* 2178 * Handle a stathz tick. This is really only relevant for timeshare 2179 * threads. 2180 */ 2181void 2182sched_clock(struct thread *td) 2183{ 2184 struct tdq *tdq; 2185 struct td_sched *ts; 2186 2187 THREAD_LOCK_ASSERT(td, MA_OWNED); 2188 tdq = TDQ_SELF(); 2189#ifdef SMP 2190 /* 2191 * We run the long term load balancer infrequently on the first cpu. 2192 */ 2193 if (balance_tdq == tdq) { 2194 if (balance_ticks && --balance_ticks == 0) 2195 sched_balance(); 2196 } 2197#endif 2198 /* 2199 * Save the old switch count so we have a record of the last ticks 2200 * activity. Initialize the new switch count based on our load. 2201 * If there is some activity seed it to reflect that. 2202 */ 2203 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2204 tdq->tdq_switchcnt = tdq->tdq_load; 2205 /* 2206 * Advance the insert index once for each tick to ensure that all 2207 * threads get a chance to run. 2208 */ 2209 if (tdq->tdq_idx == tdq->tdq_ridx) { 2210 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2211 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2212 tdq->tdq_ridx = tdq->tdq_idx; 2213 } 2214 ts = td->td_sched; 2215 sched_pctcpu_update(ts, 1); 2216 if (td->td_pri_class & PRI_FIFO_BIT) 2217 return; 2218 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) { 2219 /* 2220 * We used a tick; charge it to the thread so 2221 * that we can compute our interactivity. 2222 */ 2223 td->td_sched->ts_runtime += tickincr; 2224 sched_interact_update(td); 2225 sched_priority(td); 2226 } 2227 /* 2228 * We used up one time slice. 2229 */ 2230 if (--ts->ts_slice > 0) 2231 return; 2232 /* 2233 * We're out of time, force a requeue at userret(). 2234 */ 2235 ts->ts_slice = sched_slice; 2236 td->td_flags |= TDF_NEEDRESCHED; 2237} 2238 2239/* 2240 * Called once per hz tick. 2241 */ 2242void 2243sched_tick(int cnt) 2244{ 2245 2246} 2247 2248/* 2249 * Return whether the current CPU has runnable tasks. Used for in-kernel 2250 * cooperative idle threads. 2251 */ 2252int 2253sched_runnable(void) 2254{ 2255 struct tdq *tdq; 2256 int load; 2257 2258 load = 1; 2259 2260 tdq = TDQ_SELF(); 2261 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2262 if (tdq->tdq_load > 0) 2263 goto out; 2264 } else 2265 if (tdq->tdq_load - 1 > 0) 2266 goto out; 2267 load = 0; 2268out: 2269 return (load); 2270} 2271 2272/* 2273 * Choose the highest priority thread to run. The thread is removed from 2274 * the run-queue while running however the load remains. For SMP we set 2275 * the tdq in the global idle bitmask if it idles here. 2276 */ 2277struct thread * 2278sched_choose(void) 2279{ 2280 struct thread *td; 2281 struct tdq *tdq; 2282 2283 tdq = TDQ_SELF(); 2284 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2285 td = tdq_choose(tdq); 2286 if (td) { 2287 tdq_runq_rem(tdq, td); 2288 tdq->tdq_lowpri = td->td_priority; 2289 return (td); 2290 } 2291 tdq->tdq_lowpri = PRI_MAX_IDLE; 2292 return (PCPU_GET(idlethread)); 2293} 2294 2295/* 2296 * Set owepreempt if necessary. Preemption never happens directly in ULE, 2297 * we always request it once we exit a critical section. 2298 */ 2299static inline void 2300sched_setpreempt(struct thread *td) 2301{ 2302 struct thread *ctd; 2303 int cpri; 2304 int pri; 2305 2306 THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2307 2308 ctd = curthread; 2309 pri = td->td_priority; 2310 cpri = ctd->td_priority; 2311 if (pri < cpri) 2312 ctd->td_flags |= TDF_NEEDRESCHED; 2313 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2314 return; 2315 if (!sched_shouldpreempt(pri, cpri, 0)) 2316 return; 2317 ctd->td_owepreempt = 1; 2318} 2319 2320/* 2321 * Add a thread to a thread queue. Select the appropriate runq and add the 2322 * thread to it. This is the internal function called when the tdq is 2323 * predetermined. 2324 */ 2325void 2326tdq_add(struct tdq *tdq, struct thread *td, int flags) 2327{ 2328 2329 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2330 KASSERT((td->td_inhibitors == 0), 2331 ("sched_add: trying to run inhibited thread")); 2332 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2333 ("sched_add: bad thread state")); 2334 KASSERT(td->td_flags & TDF_INMEM, 2335 ("sched_add: thread swapped out")); 2336 2337 if (td->td_priority < tdq->tdq_lowpri) 2338 tdq->tdq_lowpri = td->td_priority; 2339 tdq_runq_add(tdq, td, flags); 2340 tdq_load_add(tdq, td); 2341} 2342 2343/* 2344 * Select the target thread queue and add a thread to it. Request 2345 * preemption or IPI a remote processor if required. 2346 */ 2347void 2348sched_add(struct thread *td, int flags) 2349{ 2350 struct tdq *tdq; 2351#ifdef SMP 2352 int cpu; 2353#endif 2354 2355 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 2356 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 2357 sched_tdname(curthread)); 2358 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 2359 KTR_ATTR_LINKED, sched_tdname(td)); 2360 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 2361 flags & SRQ_PREEMPTED); 2362 THREAD_LOCK_ASSERT(td, MA_OWNED); 2363 /* 2364 * Recalculate the priority before we select the target cpu or 2365 * run-queue. 2366 */ 2367 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2368 sched_priority(td); 2369#ifdef SMP 2370 /* 2371 * Pick the destination cpu and if it isn't ours transfer to the 2372 * target cpu. 2373 */ 2374 cpu = sched_pickcpu(td, flags); 2375 tdq = sched_setcpu(td, cpu, flags); 2376 tdq_add(tdq, td, flags); 2377 if (cpu != PCPU_GET(cpuid)) { 2378 tdq_notify(tdq, td); 2379 return; 2380 } 2381#else 2382 tdq = TDQ_SELF(); 2383 TDQ_LOCK(tdq); 2384 /* 2385 * Now that the thread is moving to the run-queue, set the lock 2386 * to the scheduler's lock. 2387 */ 2388 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2389 tdq_add(tdq, td, flags); 2390#endif 2391 if (!(flags & SRQ_YIELDING)) 2392 sched_setpreempt(td); 2393} 2394 2395/* 2396 * Remove a thread from a run-queue without running it. This is used 2397 * when we're stealing a thread from a remote queue. Otherwise all threads 2398 * exit by calling sched_exit_thread() and sched_throw() themselves. 2399 */ 2400void 2401sched_rem(struct thread *td) 2402{ 2403 struct tdq *tdq; 2404 2405 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 2406 "prio:%d", td->td_priority); 2407 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); 2408 tdq = TDQ_CPU(td->td_sched->ts_cpu); 2409 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2410 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2411 KASSERT(TD_ON_RUNQ(td), 2412 ("sched_rem: thread not on run queue")); 2413 tdq_runq_rem(tdq, td); 2414 tdq_load_rem(tdq, td); 2415 TD_SET_CAN_RUN(td); 2416 if (td->td_priority == tdq->tdq_lowpri) 2417 tdq_setlowpri(tdq, NULL); 2418} 2419 2420/* 2421 * Fetch cpu utilization information. Updates on demand. 2422 */ 2423fixpt_t 2424sched_pctcpu(struct thread *td) 2425{ 2426 fixpt_t pctcpu; 2427 struct td_sched *ts; 2428 2429 pctcpu = 0; 2430 ts = td->td_sched; 2431 if (ts == NULL) 2432 return (0); 2433 2434 THREAD_LOCK_ASSERT(td, MA_OWNED); 2435 sched_pctcpu_update(ts, TD_IS_RUNNING(td)); 2436 if (ts->ts_ticks) { 2437 int rtick; 2438 2439 /* How many rtick per second ? */ 2440 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2441 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2442 } 2443 2444 return (pctcpu); 2445} 2446 2447/* 2448 * Enforce affinity settings for a thread. Called after adjustments to 2449 * cpumask. 2450 */ 2451void 2452sched_affinity(struct thread *td) 2453{ 2454#ifdef SMP 2455 struct td_sched *ts; 2456 2457 THREAD_LOCK_ASSERT(td, MA_OWNED); 2458 ts = td->td_sched; 2459 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2460 return; 2461 if (TD_ON_RUNQ(td)) { 2462 sched_rem(td); 2463 sched_add(td, SRQ_BORING); 2464 return; 2465 } 2466 if (!TD_IS_RUNNING(td)) 2467 return; 2468 /* 2469 * Force a switch before returning to userspace. If the 2470 * target thread is not running locally send an ipi to force 2471 * the issue. 2472 */ 2473 td->td_flags |= TDF_NEEDRESCHED; 2474 if (td != curthread) 2475 ipi_cpu(ts->ts_cpu, IPI_PREEMPT); 2476#endif 2477} 2478 2479/* 2480 * Bind a thread to a target cpu. 2481 */ 2482void 2483sched_bind(struct thread *td, int cpu) 2484{ 2485 struct td_sched *ts; 2486 2487 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2488 KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 2489 ts = td->td_sched; 2490 if (ts->ts_flags & TSF_BOUND) 2491 sched_unbind(td); 2492 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td)); 2493 ts->ts_flags |= TSF_BOUND; 2494 sched_pin(); 2495 if (PCPU_GET(cpuid) == cpu) 2496 return; 2497 ts->ts_cpu = cpu; 2498 /* When we return from mi_switch we'll be on the correct cpu. */ 2499 mi_switch(SW_VOL, NULL); 2500} 2501 2502/* 2503 * Release a bound thread. 2504 */ 2505void 2506sched_unbind(struct thread *td) 2507{ 2508 struct td_sched *ts; 2509 2510 THREAD_LOCK_ASSERT(td, MA_OWNED); 2511 KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 2512 ts = td->td_sched; 2513 if ((ts->ts_flags & TSF_BOUND) == 0) 2514 return; 2515 ts->ts_flags &= ~TSF_BOUND; 2516 sched_unpin(); 2517} 2518 2519int 2520sched_is_bound(struct thread *td) 2521{ 2522 THREAD_LOCK_ASSERT(td, MA_OWNED); 2523 return (td->td_sched->ts_flags & TSF_BOUND); 2524} 2525 2526/* 2527 * Basic yield call. 2528 */ 2529void 2530sched_relinquish(struct thread *td) 2531{ 2532 thread_lock(td); 2533 mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 2534 thread_unlock(td); 2535} 2536 2537/* 2538 * Return the total system load. 2539 */ 2540int 2541sched_load(void) 2542{ 2543#ifdef SMP 2544 int total; 2545 int i; 2546 2547 total = 0; 2548 CPU_FOREACH(i) 2549 total += TDQ_CPU(i)->tdq_sysload; 2550 return (total); 2551#else 2552 return (TDQ_SELF()->tdq_sysload); 2553#endif 2554} 2555 2556int 2557sched_sizeof_proc(void) 2558{ 2559 return (sizeof(struct proc)); 2560} 2561 2562int 2563sched_sizeof_thread(void) 2564{ 2565 return (sizeof(struct thread) + sizeof(struct td_sched)); 2566} 2567 2568#ifdef SMP 2569#define TDQ_IDLESPIN(tdq) \ 2570 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0) 2571#else 2572#define TDQ_IDLESPIN(tdq) 1 2573#endif 2574 2575/* 2576 * The actual idle process. 2577 */ 2578void 2579sched_idletd(void *dummy) 2580{ 2581 struct thread *td; 2582 struct tdq *tdq; 2583 int switchcnt; 2584 int i; 2585 2586 mtx_assert(&Giant, MA_NOTOWNED); 2587 td = curthread; 2588 tdq = TDQ_SELF(); 2589 for (;;) { 2590#ifdef SMP 2591 if (tdq_idled(tdq) == 0) 2592 continue; 2593#endif 2594 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2595 /* 2596 * If we're switching very frequently, spin while checking 2597 * for load rather than entering a low power state that 2598 * may require an IPI. However, don't do any busy 2599 * loops while on SMT machines as this simply steals 2600 * cycles from cores doing useful work. 2601 */ 2602 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) { 2603 for (i = 0; i < sched_idlespins; i++) { 2604 if (tdq->tdq_load) 2605 break; 2606 cpu_spinwait(); 2607 } 2608 } 2609 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2610 if (tdq->tdq_load == 0) { 2611 tdq->tdq_cpu_idle = 1; 2612 if (tdq->tdq_load == 0) { 2613 cpu_idle(switchcnt > sched_idlespinthresh * 4); 2614 tdq->tdq_switchcnt++; 2615 } 2616 tdq->tdq_cpu_idle = 0; 2617 } 2618 if (tdq->tdq_load) { 2619 thread_lock(td); 2620 mi_switch(SW_VOL | SWT_IDLE, NULL); 2621 thread_unlock(td); 2622 } 2623 } 2624} 2625 2626/* 2627 * A CPU is entering for the first time or a thread is exiting. 2628 */ 2629void 2630sched_throw(struct thread *td) 2631{ 2632 struct thread *newtd; 2633 struct tdq *tdq; 2634 2635 tdq = TDQ_SELF(); 2636 if (td == NULL) { 2637 /* Correct spinlock nesting and acquire the correct lock. */ 2638 TDQ_LOCK(tdq); 2639 spinlock_exit(); 2640 PCPU_SET(switchtime, cpu_ticks()); 2641 PCPU_SET(switchticks, ticks); 2642 } else { 2643 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2644 tdq_load_rem(tdq, td); 2645 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2646 } 2647 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2648 newtd = choosethread(); 2649 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2650 cpu_throw(td, newtd); /* doesn't return */ 2651} 2652 2653/* 2654 * This is called from fork_exit(). Just acquire the correct locks and 2655 * let fork do the rest of the work. 2656 */ 2657void 2658sched_fork_exit(struct thread *td) 2659{ 2660 struct td_sched *ts; 2661 struct tdq *tdq; 2662 int cpuid; 2663 2664 /* 2665 * Finish setting up thread glue so that it begins execution in a 2666 * non-nested critical section with the scheduler lock held. 2667 */ 2668 cpuid = PCPU_GET(cpuid); 2669 tdq = TDQ_CPU(cpuid); 2670 ts = td->td_sched; 2671 if (TD_IS_IDLETHREAD(td)) 2672 td->td_lock = TDQ_LOCKPTR(tdq); 2673 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2674 td->td_oncpu = cpuid; 2675 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2676 lock_profile_obtain_lock_success( 2677 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2678} 2679 2680/* 2681 * Create on first use to catch odd startup conditons. 2682 */ 2683char * 2684sched_tdname(struct thread *td) 2685{ 2686#ifdef KTR 2687 struct td_sched *ts; 2688 2689 ts = td->td_sched; 2690 if (ts->ts_name[0] == '\0') 2691 snprintf(ts->ts_name, sizeof(ts->ts_name), 2692 "%s tid %d", td->td_name, td->td_tid); 2693 return (ts->ts_name); 2694#else 2695 return (td->td_name); 2696#endif 2697} 2698 2699#ifdef KTR 2700void 2701sched_clear_tdname(struct thread *td) 2702{ 2703 struct td_sched *ts; 2704 2705 ts = td->td_sched; 2706 ts->ts_name[0] = '\0'; 2707} 2708#endif 2709 2710#ifdef SMP 2711 2712/* 2713 * Build the CPU topology dump string. Is recursively called to collect 2714 * the topology tree. 2715 */ 2716static int 2717sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg, 2718 int indent) 2719{ 2720 char cpusetbuf[CPUSETBUFSIZ]; 2721 int i, first; 2722 2723 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent, 2724 "", 1 + indent / 2, cg->cg_level); 2725 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "", 2726 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask)); 2727 first = TRUE; 2728 for (i = 0; i < MAXCPU; i++) { 2729 if (CPU_ISSET(i, &cg->cg_mask)) { 2730 if (!first) 2731 sbuf_printf(sb, ", "); 2732 else 2733 first = FALSE; 2734 sbuf_printf(sb, "%d", i); 2735 } 2736 } 2737 sbuf_printf(sb, "</cpu>\n"); 2738 2739 if (cg->cg_flags != 0) { 2740 sbuf_printf(sb, "%*s <flags>", indent, ""); 2741 if ((cg->cg_flags & CG_FLAG_HTT) != 0) 2742 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>"); 2743 if ((cg->cg_flags & CG_FLAG_THREAD) != 0) 2744 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>"); 2745 if ((cg->cg_flags & CG_FLAG_SMT) != 0) 2746 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>"); 2747 sbuf_printf(sb, "</flags>\n"); 2748 } 2749 2750 if (cg->cg_children > 0) { 2751 sbuf_printf(sb, "%*s <children>\n", indent, ""); 2752 for (i = 0; i < cg->cg_children; i++) 2753 sysctl_kern_sched_topology_spec_internal(sb, 2754 &cg->cg_child[i], indent+2); 2755 sbuf_printf(sb, "%*s </children>\n", indent, ""); 2756 } 2757 sbuf_printf(sb, "%*s</group>\n", indent, ""); 2758 return (0); 2759} 2760 2761/* 2762 * Sysctl handler for retrieving topology dump. It's a wrapper for 2763 * the recursive sysctl_kern_smp_topology_spec_internal(). 2764 */ 2765static int 2766sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS) 2767{ 2768 struct sbuf *topo; 2769 int err; 2770 2771 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized")); 2772 2773 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND); 2774 if (topo == NULL) 2775 return (ENOMEM); 2776 2777 sbuf_printf(topo, "<groups>\n"); 2778 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1); 2779 sbuf_printf(topo, "</groups>\n"); 2780 2781 if (err == 0) { 2782 sbuf_finish(topo); 2783 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo)); 2784 } 2785 sbuf_delete(topo); 2786 return (err); 2787} 2788 2789#endif 2790 2791SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2792SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2793 "Scheduler name"); 2794SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2795 "Slice size for timeshare threads"); 2796SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2797 "Interactivity score threshold"); 2798SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2799 0,"Min priority for preemption, lower priorities have greater precedence"); 2800SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 2801 0,"Controls whether static kernel priorities are assigned to sleeping threads."); 2802SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 2803 0,"Number of times idle will spin waiting for new work."); 2804SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, &sched_idlespinthresh, 2805 0,"Threshold before we will permit idle spinning."); 2806#ifdef SMP 2807SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2808 "Number of hz ticks to keep thread affinity for"); 2809SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2810 "Enables the long-term load balancer"); 2811SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2812 &balance_interval, 0, 2813 "Average frequency in stathz ticks to run the long-term balancer"); 2814SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2815 "Attempts to steal work from other cores before idling"); 2816SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2817 "Minimum load on remote cpu before we'll steal"); 2818 2819/* Retrieve SMP topology */ 2820SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING | 2821 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A", 2822 "XML dump of detected CPU topology"); 2823 2824#endif 2825 2826/* ps compat. All cpu percentages from ULE are weighted. */ 2827static int ccpu = 0; 2828SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2829