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("$FreeBSD$"); 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/* Flags kept in td_flags. */ 193#define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */ 194 195/* 196 * tickincr: Converts a stathz tick into a hz domain scaled by 197 * the shift factor. Without the shift the error rate 198 * due to rounding would be unacceptably high. 199 * realstathz: stathz is sometimes 0 and run off of hz. 200 * sched_slice: Runtime of each thread before rescheduling. 201 * preempt_thresh: Priority threshold for preemption and remote IPIs. 202 */ 203static int sched_interact = SCHED_INTERACT_THRESH; 204static int realstathz = 127; 205static int tickincr = 8 << SCHED_TICK_SHIFT; 206static int sched_slice = 12; 207#ifdef PREEMPTION 208#ifdef FULL_PREEMPTION 209static int preempt_thresh = PRI_MAX_IDLE; 210#else 211static int preempt_thresh = PRI_MIN_KERN; 212#endif 213#else 214static int preempt_thresh = 0; 215#endif 216static int static_boost = PRI_MIN_BATCH; 217static int sched_idlespins = 10000; 218static int sched_idlespinthresh = -1; 219 220/* 221 * tdq - per processor runqs and statistics. All fields are protected by the 222 * tdq_lock. The load and lowpri may be accessed without to avoid excess 223 * locking in sched_pickcpu(); 224 */ 225struct tdq { 226 /* Ordered to improve efficiency of cpu_search() and switch(). */ 227 struct mtx tdq_lock; /* run queue lock. */ 228 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */ 229 volatile int tdq_load; /* Aggregate load. */ 230 volatile int tdq_cpu_idle; /* cpu_idle() is active. */ 231 int tdq_sysload; /* For loadavg, !ITHD load. */ 232 int tdq_transferable; /* Transferable thread count. */ 233 short tdq_switchcnt; /* Switches this tick. */ 234 short tdq_oldswitchcnt; /* Switches last tick. */ 235 u_char tdq_lowpri; /* Lowest priority thread. */ 236 u_char tdq_ipipending; /* IPI pending. */ 237 u_char tdq_idx; /* Current insert index. */ 238 u_char tdq_ridx; /* Current removal index. */ 239 struct runq tdq_realtime; /* real-time run queue. */ 240 struct runq tdq_timeshare; /* timeshare run queue. */ 241 struct runq tdq_idle; /* Queue of IDLE threads. */ 242 char tdq_name[TDQ_NAME_LEN]; 243#ifdef KTR 244 char tdq_loadname[TDQ_LOADNAME_LEN]; 245#endif 246} __aligned(64); 247 248/* Idle thread states and config. */ 249#define TDQ_RUNNING 1 250#define TDQ_IDLE 2 251 252#ifdef SMP 253struct cpu_group *cpu_top; /* CPU topology */ 254 255#define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000)) 256#define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity)) 257 258/* 259 * Run-time tunables. 260 */ 261static int rebalance = 1; 262static int balance_interval = 128; /* Default set in sched_initticks(). */ 263static int affinity; 264static int steal_idle = 1; 265static int steal_thresh = 2; 266 267/* 268 * One thread queue per processor. 269 */ 270static struct tdq tdq_cpu[MAXCPU]; 271static struct tdq *balance_tdq; 272static int balance_ticks; 273static DPCPU_DEFINE(uint32_t, randomval); 274 275#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 276#define TDQ_CPU(x) (&tdq_cpu[(x)]) 277#define TDQ_ID(x) ((int)((x) - tdq_cpu)) 278#else /* !SMP */ 279static struct tdq tdq_cpu; 280 281#define TDQ_ID(x) (0) 282#define TDQ_SELF() (&tdq_cpu) 283#define TDQ_CPU(x) (&tdq_cpu) 284#endif 285 286#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 287#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 288#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 289#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 290#define TDQ_LOCKPTR(t) (&(t)->tdq_lock) 291 292static void sched_priority(struct thread *); 293static void sched_thread_priority(struct thread *, u_char); 294static int sched_interact_score(struct thread *); 295static void sched_interact_update(struct thread *); 296static void sched_interact_fork(struct thread *); 297static void sched_pctcpu_update(struct td_sched *, int); 298 299/* Operations on per processor queues */ 300static struct thread *tdq_choose(struct tdq *); 301static void tdq_setup(struct tdq *); 302static void tdq_load_add(struct tdq *, struct thread *); 303static void tdq_load_rem(struct tdq *, struct thread *); 304static __inline void tdq_runq_add(struct tdq *, struct thread *, int); 305static __inline void tdq_runq_rem(struct tdq *, struct thread *); 306static inline int sched_shouldpreempt(int, int, int); 307void tdq_print(int cpu); 308static void runq_print(struct runq *rq); 309static void tdq_add(struct tdq *, struct thread *, int); 310#ifdef SMP 311static int tdq_move(struct tdq *, struct tdq *); 312static int tdq_idled(struct tdq *); 313static void tdq_notify(struct tdq *, struct thread *); 314static struct thread *tdq_steal(struct tdq *, int); 315static struct thread *runq_steal(struct runq *, int); 316static int sched_pickcpu(struct thread *, int); 317static void sched_balance(void); 318static int sched_balance_pair(struct tdq *, struct tdq *); 319static inline struct tdq *sched_setcpu(struct thread *, int, int); 320static inline void thread_unblock_switch(struct thread *, struct mtx *); 321static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 322static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS); 323static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, 324 struct cpu_group *cg, int indent); 325#endif 326 327static void sched_setup(void *dummy); 328SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 329 330static void sched_initticks(void *dummy); 331SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 332 NULL); 333 334SDT_PROVIDER_DEFINE(sched); 335 336SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *", 337 "struct proc *", "uint8_t"); 338SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *", 339 "struct proc *", "void *"); 340SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *", 341 "struct proc *", "void *", "int"); 342SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *", 343 "struct proc *", "uint8_t", "struct thread *"); 344SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int"); 345SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *", 346 "struct proc *"); 347SDT_PROBE_DEFINE(sched, , , on__cpu); 348SDT_PROBE_DEFINE(sched, , , remain__cpu); 349SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *", 350 "struct proc *"); 351 352/* 353 * Print the threads waiting on a run-queue. 354 */ 355static void 356runq_print(struct runq *rq) 357{ 358 struct rqhead *rqh; 359 struct thread *td; 360 int pri; 361 int j; 362 int i; 363 364 for (i = 0; i < RQB_LEN; i++) { 365 printf("\t\trunq bits %d 0x%zx\n", 366 i, rq->rq_status.rqb_bits[i]); 367 for (j = 0; j < RQB_BPW; j++) 368 if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 369 pri = j + (i << RQB_L2BPW); 370 rqh = &rq->rq_queues[pri]; 371 TAILQ_FOREACH(td, rqh, td_runq) { 372 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 373 td, td->td_name, td->td_priority, 374 td->td_rqindex, pri); 375 } 376 } 377 } 378} 379 380/* 381 * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 382 */ 383void 384tdq_print(int cpu) 385{ 386 struct tdq *tdq; 387 388 tdq = TDQ_CPU(cpu); 389 390 printf("tdq %d:\n", TDQ_ID(tdq)); 391 printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 392 printf("\tLock name: %s\n", tdq->tdq_name); 393 printf("\tload: %d\n", tdq->tdq_load); 394 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt); 395 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt); 396 printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 397 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 398 printf("\tload transferable: %d\n", tdq->tdq_transferable); 399 printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 400 printf("\trealtime runq:\n"); 401 runq_print(&tdq->tdq_realtime); 402 printf("\ttimeshare runq:\n"); 403 runq_print(&tdq->tdq_timeshare); 404 printf("\tidle runq:\n"); 405 runq_print(&tdq->tdq_idle); 406} 407 408static inline int 409sched_shouldpreempt(int pri, int cpri, int remote) 410{ 411 /* 412 * If the new priority is not better than the current priority there is 413 * nothing to do. 414 */ 415 if (pri >= cpri) 416 return (0); 417 /* 418 * Always preempt idle. 419 */ 420 if (cpri >= PRI_MIN_IDLE) 421 return (1); 422 /* 423 * If preemption is disabled don't preempt others. 424 */ 425 if (preempt_thresh == 0) 426 return (0); 427 /* 428 * Preempt if we exceed the threshold. 429 */ 430 if (pri <= preempt_thresh) 431 return (1); 432 /* 433 * If we're interactive or better and there is non-interactive 434 * or worse running preempt only remote processors. 435 */ 436 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT) 437 return (1); 438 return (0); 439} 440 441/* 442 * Add a thread to the actual run-queue. Keeps transferable counts up to 443 * date with what is actually on the run-queue. Selects the correct 444 * queue position for timeshare threads. 445 */ 446static __inline void 447tdq_runq_add(struct tdq *tdq, struct thread *td, int flags) 448{ 449 struct td_sched *ts; 450 u_char pri; 451 452 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 453 THREAD_LOCK_ASSERT(td, MA_OWNED); 454 455 pri = td->td_priority; 456 ts = td->td_sched; 457 TD_SET_RUNQ(td); 458 if (THREAD_CAN_MIGRATE(td)) { 459 tdq->tdq_transferable++; 460 ts->ts_flags |= TSF_XFERABLE; 461 } 462 if (pri < PRI_MIN_BATCH) { 463 ts->ts_runq = &tdq->tdq_realtime; 464 } else if (pri <= PRI_MAX_BATCH) { 465 ts->ts_runq = &tdq->tdq_timeshare; 466 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH, 467 ("Invalid priority %d on timeshare runq", pri)); 468 /* 469 * This queue contains only priorities between MIN and MAX 470 * realtime. Use the whole queue to represent these values. 471 */ 472 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 473 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE; 474 pri = (pri + tdq->tdq_idx) % RQ_NQS; 475 /* 476 * This effectively shortens the queue by one so we 477 * can have a one slot difference between idx and 478 * ridx while we wait for threads to drain. 479 */ 480 if (tdq->tdq_ridx != tdq->tdq_idx && 481 pri == tdq->tdq_ridx) 482 pri = (unsigned char)(pri - 1) % RQ_NQS; 483 } else 484 pri = tdq->tdq_ridx; 485 runq_add_pri(ts->ts_runq, td, pri, flags); 486 return; 487 } else 488 ts->ts_runq = &tdq->tdq_idle; 489 runq_add(ts->ts_runq, td, flags); 490} 491 492/* 493 * Remove a thread from a run-queue. This typically happens when a thread 494 * is selected to run. Running threads are not on the queue and the 495 * transferable count does not reflect them. 496 */ 497static __inline void 498tdq_runq_rem(struct tdq *tdq, struct thread *td) 499{ 500 struct td_sched *ts; 501 502 ts = td->td_sched; 503 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 504 KASSERT(ts->ts_runq != NULL, 505 ("tdq_runq_remove: thread %p null ts_runq", td)); 506 if (ts->ts_flags & TSF_XFERABLE) { 507 tdq->tdq_transferable--; 508 ts->ts_flags &= ~TSF_XFERABLE; 509 } 510 if (ts->ts_runq == &tdq->tdq_timeshare) { 511 if (tdq->tdq_idx != tdq->tdq_ridx) 512 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx); 513 else 514 runq_remove_idx(ts->ts_runq, td, NULL); 515 } else 516 runq_remove(ts->ts_runq, td); 517} 518 519/* 520 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 521 * for this thread to the referenced thread queue. 522 */ 523static void 524tdq_load_add(struct tdq *tdq, struct thread *td) 525{ 526 527 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 528 THREAD_LOCK_ASSERT(td, MA_OWNED); 529 530 tdq->tdq_load++; 531 if ((td->td_flags & TDF_NOLOAD) == 0) 532 tdq->tdq_sysload++; 533 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 534 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load); 535} 536 537/* 538 * Remove the load from a thread that is transitioning to a sleep state or 539 * exiting. 540 */ 541static void 542tdq_load_rem(struct tdq *tdq, struct thread *td) 543{ 544 545 THREAD_LOCK_ASSERT(td, MA_OWNED); 546 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 547 KASSERT(tdq->tdq_load != 0, 548 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 549 550 tdq->tdq_load--; 551 if ((td->td_flags & TDF_NOLOAD) == 0) 552 tdq->tdq_sysload--; 553 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 554 SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load); 555} 556 557/* 558 * Set lowpri to its exact value by searching the run-queue and 559 * evaluating curthread. curthread may be passed as an optimization. 560 */ 561static void 562tdq_setlowpri(struct tdq *tdq, struct thread *ctd) 563{ 564 struct thread *td; 565 566 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 567 if (ctd == NULL) 568 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 569 td = tdq_choose(tdq); 570 if (td == NULL || td->td_priority > ctd->td_priority) 571 tdq->tdq_lowpri = ctd->td_priority; 572 else 573 tdq->tdq_lowpri = td->td_priority; 574} 575 576#ifdef SMP 577struct cpu_search { 578 cpuset_t cs_mask; 579 u_int cs_prefer; 580 int cs_pri; /* Min priority for low. */ 581 int cs_limit; /* Max load for low, min load for high. */ 582 int cs_cpu; 583 int cs_load; 584}; 585 586#define CPU_SEARCH_LOWEST 0x1 587#define CPU_SEARCH_HIGHEST 0x2 588#define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 589 590#define CPUSET_FOREACH(cpu, mask) \ 591 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \ 592 if (CPU_ISSET(cpu, &mask)) 593 594static __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low, 595 struct cpu_search *high, const int match); 596int cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low); 597int cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high); 598int cpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 599 struct cpu_search *high); 600 601/* 602 * Search the tree of cpu_groups for the lowest or highest loaded cpu 603 * according to the match argument. This routine actually compares the 604 * load on all paths through the tree and finds the least loaded cpu on 605 * the least loaded path, which may differ from the least loaded cpu in 606 * the system. This balances work among caches and busses. 607 * 608 * This inline is instantiated in three forms below using constants for the 609 * match argument. It is reduced to the minimum set for each case. It is 610 * also recursive to the depth of the tree. 611 */ 612static __inline int 613cpu_search(const struct cpu_group *cg, struct cpu_search *low, 614 struct cpu_search *high, const int match) 615{ 616 struct cpu_search lgroup; 617 struct cpu_search hgroup; 618 cpuset_t cpumask; 619 struct cpu_group *child; 620 struct tdq *tdq; 621 int cpu, i, hload, lload, load, total, rnd, *rndptr; 622 623 total = 0; 624 cpumask = cg->cg_mask; 625 if (match & CPU_SEARCH_LOWEST) { 626 lload = INT_MAX; 627 lgroup = *low; 628 } 629 if (match & CPU_SEARCH_HIGHEST) { 630 hload = INT_MIN; 631 hgroup = *high; 632 } 633 634 /* Iterate through the child CPU groups and then remaining CPUs. */ 635 for (i = cg->cg_children, cpu = mp_maxid; ; ) { 636 if (i == 0) { 637#ifdef HAVE_INLINE_FFSL 638 cpu = CPU_FFS(&cpumask) - 1; 639#else 640 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask)) 641 cpu--; 642#endif 643 if (cpu < 0) 644 break; 645 child = NULL; 646 } else 647 child = &cg->cg_child[i - 1]; 648 649 if (match & CPU_SEARCH_LOWEST) 650 lgroup.cs_cpu = -1; 651 if (match & CPU_SEARCH_HIGHEST) 652 hgroup.cs_cpu = -1; 653 if (child) { /* Handle child CPU group. */ 654 CPU_NAND(&cpumask, &child->cg_mask); 655 switch (match) { 656 case CPU_SEARCH_LOWEST: 657 load = cpu_search_lowest(child, &lgroup); 658 break; 659 case CPU_SEARCH_HIGHEST: 660 load = cpu_search_highest(child, &hgroup); 661 break; 662 case CPU_SEARCH_BOTH: 663 load = cpu_search_both(child, &lgroup, &hgroup); 664 break; 665 } 666 } else { /* Handle child CPU. */ 667 CPU_CLR(cpu, &cpumask); 668 tdq = TDQ_CPU(cpu); 669 load = tdq->tdq_load * 256; 670 rndptr = DPCPU_PTR(randomval); 671 rnd = (*rndptr = *rndptr * 69069 + 5) >> 26; 672 if (match & CPU_SEARCH_LOWEST) { 673 if (cpu == low->cs_prefer) 674 load -= 64; 675 /* If that CPU is allowed and get data. */ 676 if (tdq->tdq_lowpri > lgroup.cs_pri && 677 tdq->tdq_load <= lgroup.cs_limit && 678 CPU_ISSET(cpu, &lgroup.cs_mask)) { 679 lgroup.cs_cpu = cpu; 680 lgroup.cs_load = load - rnd; 681 } 682 } 683 if (match & CPU_SEARCH_HIGHEST) 684 if (tdq->tdq_load >= hgroup.cs_limit && 685 tdq->tdq_transferable && 686 CPU_ISSET(cpu, &hgroup.cs_mask)) { 687 hgroup.cs_cpu = cpu; 688 hgroup.cs_load = load - rnd; 689 } 690 } 691 total += load; 692 693 /* We have info about child item. Compare it. */ 694 if (match & CPU_SEARCH_LOWEST) { 695 if (lgroup.cs_cpu >= 0 && 696 (load < lload || 697 (load == lload && lgroup.cs_load < low->cs_load))) { 698 lload = load; 699 low->cs_cpu = lgroup.cs_cpu; 700 low->cs_load = lgroup.cs_load; 701 } 702 } 703 if (match & CPU_SEARCH_HIGHEST) 704 if (hgroup.cs_cpu >= 0 && 705 (load > hload || 706 (load == hload && hgroup.cs_load > high->cs_load))) { 707 hload = load; 708 high->cs_cpu = hgroup.cs_cpu; 709 high->cs_load = hgroup.cs_load; 710 } 711 if (child) { 712 i--; 713 if (i == 0 && CPU_EMPTY(&cpumask)) 714 break; 715 } 716#ifndef HAVE_INLINE_FFSL 717 else 718 cpu--; 719#endif 720 } 721 return (total); 722} 723 724/* 725 * cpu_search instantiations must pass constants to maintain the inline 726 * optimization. 727 */ 728int 729cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low) 730{ 731 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 732} 733 734int 735cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high) 736{ 737 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 738} 739 740int 741cpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 742 struct cpu_search *high) 743{ 744 return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 745} 746 747/* 748 * Find the cpu with the least load via the least loaded path that has a 749 * lowpri greater than pri pri. A pri of -1 indicates any priority is 750 * acceptable. 751 */ 752static inline int 753sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload, 754 int prefer) 755{ 756 struct cpu_search low; 757 758 low.cs_cpu = -1; 759 low.cs_prefer = prefer; 760 low.cs_mask = mask; 761 low.cs_pri = pri; 762 low.cs_limit = maxload; 763 cpu_search_lowest(cg, &low); 764 return low.cs_cpu; 765} 766 767/* 768 * Find the cpu with the highest load via the highest loaded path. 769 */ 770static inline int 771sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload) 772{ 773 struct cpu_search high; 774 775 high.cs_cpu = -1; 776 high.cs_mask = mask; 777 high.cs_limit = minload; 778 cpu_search_highest(cg, &high); 779 return high.cs_cpu; 780} 781 782static void 783sched_balance_group(struct cpu_group *cg) 784{ 785 cpuset_t hmask, lmask; 786 int high, low, anylow; 787 788 CPU_FILL(&hmask); 789 for (;;) { 790 high = sched_highest(cg, hmask, 1); 791 /* Stop if there is no more CPU with transferrable threads. */ 792 if (high == -1) 793 break; 794 CPU_CLR(high, &hmask); 795 CPU_COPY(&hmask, &lmask); 796 /* Stop if there is no more CPU left for low. */ 797 if (CPU_EMPTY(&lmask)) 798 break; 799 anylow = 1; 800nextlow: 801 low = sched_lowest(cg, lmask, -1, 802 TDQ_CPU(high)->tdq_load - 1, high); 803 /* Stop if we looked well and found no less loaded CPU. */ 804 if (anylow && low == -1) 805 break; 806 /* Go to next high if we found no less loaded CPU. */ 807 if (low == -1) 808 continue; 809 /* Transfer thread from high to low. */ 810 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) { 811 /* CPU that got thread can no longer be a donor. */ 812 CPU_CLR(low, &hmask); 813 } else { 814 /* 815 * If failed, then there is no threads on high 816 * that can run on this low. Drop low from low 817 * mask and look for different one. 818 */ 819 CPU_CLR(low, &lmask); 820 anylow = 0; 821 goto nextlow; 822 } 823 } 824} 825 826static void 827sched_balance(void) 828{ 829 struct tdq *tdq; 830 831 /* 832 * Select a random time between .5 * balance_interval and 833 * 1.5 * balance_interval. 834 */ 835 balance_ticks = max(balance_interval / 2, 1); 836 balance_ticks += random() % balance_interval; 837 if (smp_started == 0 || rebalance == 0) 838 return; 839 tdq = TDQ_SELF(); 840 TDQ_UNLOCK(tdq); 841 sched_balance_group(cpu_top); 842 TDQ_LOCK(tdq); 843} 844 845/* 846 * Lock two thread queues using their address to maintain lock order. 847 */ 848static void 849tdq_lock_pair(struct tdq *one, struct tdq *two) 850{ 851 if (one < two) { 852 TDQ_LOCK(one); 853 TDQ_LOCK_FLAGS(two, MTX_DUPOK); 854 } else { 855 TDQ_LOCK(two); 856 TDQ_LOCK_FLAGS(one, MTX_DUPOK); 857 } 858} 859 860/* 861 * Unlock two thread queues. Order is not important here. 862 */ 863static void 864tdq_unlock_pair(struct tdq *one, struct tdq *two) 865{ 866 TDQ_UNLOCK(one); 867 TDQ_UNLOCK(two); 868} 869 870/* 871 * Transfer load between two imbalanced thread queues. 872 */ 873static int 874sched_balance_pair(struct tdq *high, struct tdq *low) 875{ 876 int moved; 877 int cpu; 878 879 tdq_lock_pair(high, low); 880 moved = 0; 881 /* 882 * Determine what the imbalance is and then adjust that to how many 883 * threads we actually have to give up (transferable). 884 */ 885 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load && 886 (moved = tdq_move(high, low)) > 0) { 887 /* 888 * In case the target isn't the current cpu IPI it to force a 889 * reschedule with the new workload. 890 */ 891 cpu = TDQ_ID(low); 892 sched_pin(); 893 if (cpu != PCPU_GET(cpuid)) 894 ipi_cpu(cpu, IPI_PREEMPT); 895 sched_unpin(); 896 } 897 tdq_unlock_pair(high, low); 898 return (moved); 899} 900 901/* 902 * Move a thread from one thread queue to another. 903 */ 904static int 905tdq_move(struct tdq *from, struct tdq *to) 906{ 907 struct td_sched *ts; 908 struct thread *td; 909 struct tdq *tdq; 910 int cpu; 911 912 TDQ_LOCK_ASSERT(from, MA_OWNED); 913 TDQ_LOCK_ASSERT(to, MA_OWNED); 914 915 tdq = from; 916 cpu = TDQ_ID(to); 917 td = tdq_steal(tdq, cpu); 918 if (td == NULL) 919 return (0); 920 ts = td->td_sched; 921 /* 922 * Although the run queue is locked the thread may be blocked. Lock 923 * it to clear this and acquire the run-queue lock. 924 */ 925 thread_lock(td); 926 /* Drop recursive lock on from acquired via thread_lock(). */ 927 TDQ_UNLOCK(from); 928 sched_rem(td); 929 ts->ts_cpu = cpu; 930 td->td_lock = TDQ_LOCKPTR(to); 931 tdq_add(to, td, SRQ_YIELDING); 932 return (1); 933} 934 935/* 936 * This tdq has idled. Try to steal a thread from another cpu and switch 937 * to it. 938 */ 939static int 940tdq_idled(struct tdq *tdq) 941{ 942 struct cpu_group *cg; 943 struct tdq *steal; 944 cpuset_t mask; 945 int thresh; 946 int cpu; 947 948 if (smp_started == 0 || steal_idle == 0) 949 return (1); 950 CPU_FILL(&mask); 951 CPU_CLR(PCPU_GET(cpuid), &mask); 952 /* We don't want to be preempted while we're iterating. */ 953 spinlock_enter(); 954 for (cg = tdq->tdq_cg; cg != NULL; ) { 955 if ((cg->cg_flags & CG_FLAG_THREAD) == 0) 956 thresh = steal_thresh; 957 else 958 thresh = 1; 959 cpu = sched_highest(cg, mask, thresh); 960 if (cpu == -1) { 961 cg = cg->cg_parent; 962 continue; 963 } 964 steal = TDQ_CPU(cpu); 965 CPU_CLR(cpu, &mask); 966 tdq_lock_pair(tdq, steal); 967 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 968 tdq_unlock_pair(tdq, steal); 969 continue; 970 } 971 /* 972 * If a thread was added while interrupts were disabled don't 973 * steal one here. If we fail to acquire one due to affinity 974 * restrictions loop again with this cpu removed from the 975 * set. 976 */ 977 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 978 tdq_unlock_pair(tdq, steal); 979 continue; 980 } 981 spinlock_exit(); 982 TDQ_UNLOCK(steal); 983 mi_switch(SW_VOL | SWT_IDLE, NULL); 984 thread_unlock(curthread); 985 986 return (0); 987 } 988 spinlock_exit(); 989 return (1); 990} 991 992/* 993 * Notify a remote cpu of new work. Sends an IPI if criteria are met. 994 */ 995static void 996tdq_notify(struct tdq *tdq, struct thread *td) 997{ 998 struct thread *ctd; 999 int pri; 1000 int cpu; 1001 1002 if (tdq->tdq_ipipending) 1003 return; 1004 cpu = td->td_sched->ts_cpu; 1005 pri = td->td_priority; 1006 ctd = pcpu_find(cpu)->pc_curthread; 1007 if (!sched_shouldpreempt(pri, ctd->td_priority, 1)) 1008 return; 1009 if (TD_IS_IDLETHREAD(ctd)) { 1010 /* 1011 * If the MD code has an idle wakeup routine try that before 1012 * falling back to IPI. 1013 */ 1014 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu)) 1015 return; 1016 } 1017 tdq->tdq_ipipending = 1; 1018 ipi_cpu(cpu, IPI_PREEMPT); 1019} 1020 1021/* 1022 * Steals load from a timeshare queue. Honors the rotating queue head 1023 * index. 1024 */ 1025static struct thread * 1026runq_steal_from(struct runq *rq, int cpu, u_char start) 1027{ 1028 struct rqbits *rqb; 1029 struct rqhead *rqh; 1030 struct thread *td, *first; 1031 int bit; 1032 int pri; 1033 int i; 1034 1035 rqb = &rq->rq_status; 1036 bit = start & (RQB_BPW -1); 1037 pri = 0; 1038 first = NULL; 1039again: 1040 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 1041 if (rqb->rqb_bits[i] == 0) 1042 continue; 1043 if (bit != 0) { 1044 for (pri = bit; pri < RQB_BPW; pri++) 1045 if (rqb->rqb_bits[i] & (1ul << pri)) 1046 break; 1047 if (pri >= RQB_BPW) 1048 continue; 1049 } else 1050 pri = RQB_FFS(rqb->rqb_bits[i]); 1051 pri += (i << RQB_L2BPW); 1052 rqh = &rq->rq_queues[pri]; 1053 TAILQ_FOREACH(td, rqh, td_runq) { 1054 if (first && THREAD_CAN_MIGRATE(td) && 1055 THREAD_CAN_SCHED(td, cpu)) 1056 return (td); 1057 first = td; 1058 } 1059 } 1060 if (start != 0) { 1061 start = 0; 1062 goto again; 1063 } 1064 1065 if (first && THREAD_CAN_MIGRATE(first) && 1066 THREAD_CAN_SCHED(first, cpu)) 1067 return (first); 1068 return (NULL); 1069} 1070 1071/* 1072 * Steals load from a standard linear queue. 1073 */ 1074static struct thread * 1075runq_steal(struct runq *rq, int cpu) 1076{ 1077 struct rqhead *rqh; 1078 struct rqbits *rqb; 1079 struct thread *td; 1080 int word; 1081 int bit; 1082 1083 rqb = &rq->rq_status; 1084 for (word = 0; word < RQB_LEN; word++) { 1085 if (rqb->rqb_bits[word] == 0) 1086 continue; 1087 for (bit = 0; bit < RQB_BPW; bit++) { 1088 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1089 continue; 1090 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1091 TAILQ_FOREACH(td, rqh, td_runq) 1092 if (THREAD_CAN_MIGRATE(td) && 1093 THREAD_CAN_SCHED(td, cpu)) 1094 return (td); 1095 } 1096 } 1097 return (NULL); 1098} 1099 1100/* 1101 * Attempt to steal a thread in priority order from a thread queue. 1102 */ 1103static struct thread * 1104tdq_steal(struct tdq *tdq, int cpu) 1105{ 1106 struct thread *td; 1107 1108 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1109 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1110 return (td); 1111 if ((td = runq_steal_from(&tdq->tdq_timeshare, 1112 cpu, tdq->tdq_ridx)) != NULL) 1113 return (td); 1114 return (runq_steal(&tdq->tdq_idle, cpu)); 1115} 1116 1117/* 1118 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1119 * current lock and returns with the assigned queue locked. 1120 */ 1121static inline struct tdq * 1122sched_setcpu(struct thread *td, int cpu, int flags) 1123{ 1124 1125 struct tdq *tdq; 1126 1127 THREAD_LOCK_ASSERT(td, MA_OWNED); 1128 tdq = TDQ_CPU(cpu); 1129 td->td_sched->ts_cpu = cpu; 1130 /* 1131 * If the lock matches just return the queue. 1132 */ 1133 if (td->td_lock == TDQ_LOCKPTR(tdq)) 1134 return (tdq); 1135#ifdef notyet 1136 /* 1137 * If the thread isn't running its lockptr is a 1138 * turnstile or a sleepqueue. We can just lock_set without 1139 * blocking. 1140 */ 1141 if (TD_CAN_RUN(td)) { 1142 TDQ_LOCK(tdq); 1143 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1144 return (tdq); 1145 } 1146#endif 1147 /* 1148 * The hard case, migration, we need to block the thread first to 1149 * prevent order reversals with other cpus locks. 1150 */ 1151 spinlock_enter(); 1152 thread_lock_block(td); 1153 TDQ_LOCK(tdq); 1154 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1155 spinlock_exit(); 1156 return (tdq); 1157} 1158 1159SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1160SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1161SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1162SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1163SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1164SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1165 1166static int 1167sched_pickcpu(struct thread *td, int flags) 1168{ 1169 struct cpu_group *cg, *ccg; 1170 struct td_sched *ts; 1171 struct tdq *tdq; 1172 cpuset_t mask; 1173 int cpu, pri, self; 1174 1175 self = PCPU_GET(cpuid); 1176 ts = td->td_sched; 1177 if (smp_started == 0) 1178 return (self); 1179 /* 1180 * Don't migrate a running thread from sched_switch(). 1181 */ 1182 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1183 return (ts->ts_cpu); 1184 /* 1185 * Prefer to run interrupt threads on the processors that generate 1186 * the interrupt. 1187 */ 1188 pri = td->td_priority; 1189 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1190 curthread->td_intr_nesting_level && ts->ts_cpu != self) { 1191 SCHED_STAT_INC(pickcpu_intrbind); 1192 ts->ts_cpu = self; 1193 if (TDQ_CPU(self)->tdq_lowpri > pri) { 1194 SCHED_STAT_INC(pickcpu_affinity); 1195 return (ts->ts_cpu); 1196 } 1197 } 1198 /* 1199 * If the thread can run on the last cpu and the affinity has not 1200 * expired or it is idle run it there. 1201 */ 1202 tdq = TDQ_CPU(ts->ts_cpu); 1203 cg = tdq->tdq_cg; 1204 if (THREAD_CAN_SCHED(td, ts->ts_cpu) && 1205 tdq->tdq_lowpri >= PRI_MIN_IDLE && 1206 SCHED_AFFINITY(ts, CG_SHARE_L2)) { 1207 if (cg->cg_flags & CG_FLAG_THREAD) { 1208 CPUSET_FOREACH(cpu, cg->cg_mask) { 1209 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) 1210 break; 1211 } 1212 } else 1213 cpu = INT_MAX; 1214 if (cpu > mp_maxid) { 1215 SCHED_STAT_INC(pickcpu_idle_affinity); 1216 return (ts->ts_cpu); 1217 } 1218 } 1219 /* 1220 * Search for the last level cache CPU group in the tree. 1221 * Skip caches with expired affinity time and SMT groups. 1222 * Affinity to higher level caches will be handled less aggressively. 1223 */ 1224 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) { 1225 if (cg->cg_flags & CG_FLAG_THREAD) 1226 continue; 1227 if (!SCHED_AFFINITY(ts, cg->cg_level)) 1228 continue; 1229 ccg = cg; 1230 } 1231 if (ccg != NULL) 1232 cg = ccg; 1233 cpu = -1; 1234 /* Search the group for the less loaded idle CPU we can run now. */ 1235 mask = td->td_cpuset->cs_mask; 1236 if (cg != NULL && cg != cpu_top && 1237 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0) 1238 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE), 1239 INT_MAX, ts->ts_cpu); 1240 /* Search globally for the less loaded CPU we can run now. */ 1241 if (cpu == -1) 1242 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu); 1243 /* Search globally for the less loaded CPU. */ 1244 if (cpu == -1) 1245 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu); 1246 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1247 /* 1248 * Compare the lowest loaded cpu to current cpu. 1249 */ 1250 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1251 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE && 1252 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) { 1253 SCHED_STAT_INC(pickcpu_local); 1254 cpu = self; 1255 } else 1256 SCHED_STAT_INC(pickcpu_lowest); 1257 if (cpu != ts->ts_cpu) 1258 SCHED_STAT_INC(pickcpu_migration); 1259 return (cpu); 1260} 1261#endif 1262 1263/* 1264 * Pick the highest priority task we have and return it. 1265 */ 1266static struct thread * 1267tdq_choose(struct tdq *tdq) 1268{ 1269 struct thread *td; 1270 1271 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1272 td = runq_choose(&tdq->tdq_realtime); 1273 if (td != NULL) 1274 return (td); 1275 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1276 if (td != NULL) { 1277 KASSERT(td->td_priority >= PRI_MIN_BATCH, 1278 ("tdq_choose: Invalid priority on timeshare queue %d", 1279 td->td_priority)); 1280 return (td); 1281 } 1282 td = runq_choose(&tdq->tdq_idle); 1283 if (td != NULL) { 1284 KASSERT(td->td_priority >= PRI_MIN_IDLE, 1285 ("tdq_choose: Invalid priority on idle queue %d", 1286 td->td_priority)); 1287 return (td); 1288 } 1289 1290 return (NULL); 1291} 1292 1293/* 1294 * Initialize a thread queue. 1295 */ 1296static void 1297tdq_setup(struct tdq *tdq) 1298{ 1299 1300 if (bootverbose) 1301 printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1302 runq_init(&tdq->tdq_realtime); 1303 runq_init(&tdq->tdq_timeshare); 1304 runq_init(&tdq->tdq_idle); 1305 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1306 "sched lock %d", (int)TDQ_ID(tdq)); 1307 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1308 MTX_SPIN | MTX_RECURSE); 1309#ifdef KTR 1310 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname), 1311 "CPU %d load", (int)TDQ_ID(tdq)); 1312#endif 1313} 1314 1315#ifdef SMP 1316static void 1317sched_setup_smp(void) 1318{ 1319 struct tdq *tdq; 1320 int i; 1321 1322 cpu_top = smp_topo(); 1323 CPU_FOREACH(i) { 1324 tdq = TDQ_CPU(i); 1325 tdq_setup(tdq); 1326 tdq->tdq_cg = smp_topo_find(cpu_top, i); 1327 if (tdq->tdq_cg == NULL) 1328 panic("Can't find cpu group for %d\n", i); 1329 } 1330 balance_tdq = TDQ_SELF(); 1331 sched_balance(); 1332} 1333#endif 1334 1335/* 1336 * Setup the thread queues and initialize the topology based on MD 1337 * information. 1338 */ 1339static void 1340sched_setup(void *dummy) 1341{ 1342 struct tdq *tdq; 1343 1344 tdq = TDQ_SELF(); 1345#ifdef SMP 1346 sched_setup_smp(); 1347#else 1348 tdq_setup(tdq); 1349#endif 1350 1351 /* Add thread0's load since it's running. */ 1352 TDQ_LOCK(tdq); 1353 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1354 tdq_load_add(tdq, &thread0); 1355 tdq->tdq_lowpri = thread0.td_priority; 1356 TDQ_UNLOCK(tdq); 1357} 1358 1359/* 1360 * This routine determines time constants after stathz and hz are setup. 1361 */ 1362/* ARGSUSED */ 1363static void 1364sched_initticks(void *dummy) 1365{ 1366 int incr; 1367 1368 realstathz = stathz ? stathz : hz; 1369 sched_slice = realstathz / 10; /* ~100ms */ 1370 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 1371 realstathz); 1372 1373 /* 1374 * tickincr is shifted out by 10 to avoid rounding errors due to 1375 * hz not being evenly divisible by stathz on all platforms. 1376 */ 1377 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1378 /* 1379 * This does not work for values of stathz that are more than 1380 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1381 */ 1382 if (incr == 0) 1383 incr = 1; 1384 tickincr = incr; 1385#ifdef SMP 1386 /* 1387 * Set the default balance interval now that we know 1388 * what realstathz is. 1389 */ 1390 balance_interval = realstathz; 1391 affinity = SCHED_AFFINITY_DEFAULT; 1392#endif 1393 if (sched_idlespinthresh < 0) 1394 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz; 1395} 1396 1397 1398/* 1399 * This is the core of the interactivity algorithm. Determines a score based 1400 * on past behavior. It is the ratio of sleep time to run time scaled to 1401 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1402 * differs from the cpu usage because it does not account for time spent 1403 * waiting on a run-queue. Would be prettier if we had floating point. 1404 */ 1405static int 1406sched_interact_score(struct thread *td) 1407{ 1408 struct td_sched *ts; 1409 int div; 1410 1411 ts = td->td_sched; 1412 /* 1413 * The score is only needed if this is likely to be an interactive 1414 * task. Don't go through the expense of computing it if there's 1415 * no chance. 1416 */ 1417 if (sched_interact <= SCHED_INTERACT_HALF && 1418 ts->ts_runtime >= ts->ts_slptime) 1419 return (SCHED_INTERACT_HALF); 1420 1421 if (ts->ts_runtime > ts->ts_slptime) { 1422 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1423 return (SCHED_INTERACT_HALF + 1424 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1425 } 1426 if (ts->ts_slptime > ts->ts_runtime) { 1427 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1428 return (ts->ts_runtime / div); 1429 } 1430 /* runtime == slptime */ 1431 if (ts->ts_runtime) 1432 return (SCHED_INTERACT_HALF); 1433 1434 /* 1435 * This can happen if slptime and runtime are 0. 1436 */ 1437 return (0); 1438 1439} 1440 1441/* 1442 * Scale the scheduling priority according to the "interactivity" of this 1443 * process. 1444 */ 1445static void 1446sched_priority(struct thread *td) 1447{ 1448 int score; 1449 int pri; 1450 1451 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1452 return; 1453 /* 1454 * If the score is interactive we place the thread in the realtime 1455 * queue with a priority that is less than kernel and interrupt 1456 * priorities. These threads are not subject to nice restrictions. 1457 * 1458 * Scores greater than this are placed on the normal timeshare queue 1459 * where the priority is partially decided by the most recent cpu 1460 * utilization and the rest is decided by nice value. 1461 * 1462 * The nice value of the process has a linear effect on the calculated 1463 * score. Negative nice values make it easier for a thread to be 1464 * considered interactive. 1465 */ 1466 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice); 1467 if (score < sched_interact) { 1468 pri = PRI_MIN_INTERACT; 1469 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) / 1470 sched_interact) * score; 1471 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT, 1472 ("sched_priority: invalid interactive priority %d score %d", 1473 pri, score)); 1474 } else { 1475 pri = SCHED_PRI_MIN; 1476 if (td->td_sched->ts_ticks) 1477 pri += min(SCHED_PRI_TICKS(td->td_sched), 1478 SCHED_PRI_RANGE - 1); 1479 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1480 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH, 1481 ("sched_priority: invalid priority %d: nice %d, " 1482 "ticks %d ftick %d ltick %d tick pri %d", 1483 pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1484 td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1485 SCHED_PRI_TICKS(td->td_sched))); 1486 } 1487 sched_user_prio(td, pri); 1488 1489 return; 1490} 1491 1492/* 1493 * This routine enforces a maximum limit on the amount of scheduling history 1494 * kept. It is called after either the slptime or runtime is adjusted. This 1495 * function is ugly due to integer math. 1496 */ 1497static void 1498sched_interact_update(struct thread *td) 1499{ 1500 struct td_sched *ts; 1501 u_int sum; 1502 1503 ts = td->td_sched; 1504 sum = ts->ts_runtime + ts->ts_slptime; 1505 if (sum < SCHED_SLP_RUN_MAX) 1506 return; 1507 /* 1508 * This only happens from two places: 1509 * 1) We have added an unusual amount of run time from fork_exit. 1510 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1511 */ 1512 if (sum > SCHED_SLP_RUN_MAX * 2) { 1513 if (ts->ts_runtime > ts->ts_slptime) { 1514 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1515 ts->ts_slptime = 1; 1516 } else { 1517 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1518 ts->ts_runtime = 1; 1519 } 1520 return; 1521 } 1522 /* 1523 * If we have exceeded by more than 1/5th then the algorithm below 1524 * will not bring us back into range. Dividing by two here forces 1525 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1526 */ 1527 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1528 ts->ts_runtime /= 2; 1529 ts->ts_slptime /= 2; 1530 return; 1531 } 1532 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1533 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1534} 1535 1536/* 1537 * Scale back the interactivity history when a child thread is created. The 1538 * history is inherited from the parent but the thread may behave totally 1539 * differently. For example, a shell spawning a compiler process. We want 1540 * to learn that the compiler is behaving badly very quickly. 1541 */ 1542static void 1543sched_interact_fork(struct thread *td) 1544{ 1545 int ratio; 1546 int sum; 1547 1548 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1549 if (sum > SCHED_SLP_RUN_FORK) { 1550 ratio = sum / SCHED_SLP_RUN_FORK; 1551 td->td_sched->ts_runtime /= ratio; 1552 td->td_sched->ts_slptime /= ratio; 1553 } 1554} 1555 1556/* 1557 * Called from proc0_init() to setup the scheduler fields. 1558 */ 1559void 1560schedinit(void) 1561{ 1562 1563 /* 1564 * Set up the scheduler specific parts of proc0. 1565 */ 1566 proc0.p_sched = NULL; /* XXX */ 1567 thread0.td_sched = &td_sched0; 1568 td_sched0.ts_ltick = ticks; 1569 td_sched0.ts_ftick = ticks; 1570 td_sched0.ts_slice = sched_slice; 1571} 1572 1573/* 1574 * This is only somewhat accurate since given many processes of the same 1575 * priority they will switch when their slices run out, which will be 1576 * at most sched_slice stathz ticks. 1577 */ 1578int 1579sched_rr_interval(void) 1580{ 1581 1582 /* Convert sched_slice from stathz to hz. */ 1583 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz)); 1584} 1585 1586/* 1587 * Update the percent cpu tracking information when it is requested or 1588 * the total history exceeds the maximum. We keep a sliding history of 1589 * tick counts that slowly decays. This is less precise than the 4BSD 1590 * mechanism since it happens with less regular and frequent events. 1591 */ 1592static void 1593sched_pctcpu_update(struct td_sched *ts, int run) 1594{ 1595 int t = ticks; 1596 1597 if (t - ts->ts_ltick >= SCHED_TICK_TARG) { 1598 ts->ts_ticks = 0; 1599 ts->ts_ftick = t - SCHED_TICK_TARG; 1600 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) { 1601 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) * 1602 (ts->ts_ltick - (t - SCHED_TICK_TARG)); 1603 ts->ts_ftick = t - SCHED_TICK_TARG; 1604 } 1605 if (run) 1606 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT; 1607 ts->ts_ltick = t; 1608} 1609 1610/* 1611 * Adjust the priority of a thread. Move it to the appropriate run-queue 1612 * if necessary. This is the back-end for several priority related 1613 * functions. 1614 */ 1615static void 1616sched_thread_priority(struct thread *td, u_char prio) 1617{ 1618 struct td_sched *ts; 1619 struct tdq *tdq; 1620 int oldpri; 1621 1622 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio", 1623 "prio:%d", td->td_priority, "new prio:%d", prio, 1624 KTR_ATTR_LINKED, sched_tdname(curthread)); 1625 SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio); 1626 if (td != curthread && prio < td->td_priority) { 1627 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 1628 "lend prio", "prio:%d", td->td_priority, "new prio:%d", 1629 prio, KTR_ATTR_LINKED, sched_tdname(td)); 1630 SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio, 1631 curthread); 1632 } 1633 ts = td->td_sched; 1634 THREAD_LOCK_ASSERT(td, MA_OWNED); 1635 if (td->td_priority == prio) 1636 return; 1637 /* 1638 * If the priority has been elevated due to priority 1639 * propagation, we may have to move ourselves to a new 1640 * queue. This could be optimized to not re-add in some 1641 * cases. 1642 */ 1643 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1644 sched_rem(td); 1645 td->td_priority = prio; 1646 sched_add(td, SRQ_BORROWING); 1647 return; 1648 } 1649 /* 1650 * If the thread is currently running we may have to adjust the lowpri 1651 * information so other cpus are aware of our current priority. 1652 */ 1653 if (TD_IS_RUNNING(td)) { 1654 tdq = TDQ_CPU(ts->ts_cpu); 1655 oldpri = td->td_priority; 1656 td->td_priority = prio; 1657 if (prio < tdq->tdq_lowpri) 1658 tdq->tdq_lowpri = prio; 1659 else if (tdq->tdq_lowpri == oldpri) 1660 tdq_setlowpri(tdq, td); 1661 return; 1662 } 1663 td->td_priority = prio; 1664} 1665 1666/* 1667 * Update a thread's priority when it is lent another thread's 1668 * priority. 1669 */ 1670void 1671sched_lend_prio(struct thread *td, u_char prio) 1672{ 1673 1674 td->td_flags |= TDF_BORROWING; 1675 sched_thread_priority(td, prio); 1676} 1677 1678/* 1679 * Restore a thread's priority when priority propagation is 1680 * over. The prio argument is the minimum priority the thread 1681 * needs to have to satisfy other possible priority lending 1682 * requests. If the thread's regular priority is less 1683 * important than prio, the thread will keep a priority boost 1684 * of prio. 1685 */ 1686void 1687sched_unlend_prio(struct thread *td, u_char prio) 1688{ 1689 u_char base_pri; 1690 1691 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1692 td->td_base_pri <= PRI_MAX_TIMESHARE) 1693 base_pri = td->td_user_pri; 1694 else 1695 base_pri = td->td_base_pri; 1696 if (prio >= base_pri) { 1697 td->td_flags &= ~TDF_BORROWING; 1698 sched_thread_priority(td, base_pri); 1699 } else 1700 sched_lend_prio(td, prio); 1701} 1702 1703/* 1704 * Standard entry for setting the priority to an absolute value. 1705 */ 1706void 1707sched_prio(struct thread *td, u_char prio) 1708{ 1709 u_char oldprio; 1710 1711 /* First, update the base priority. */ 1712 td->td_base_pri = prio; 1713 1714 /* 1715 * If the thread is borrowing another thread's priority, don't 1716 * ever lower the priority. 1717 */ 1718 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1719 return; 1720 1721 /* Change the real priority. */ 1722 oldprio = td->td_priority; 1723 sched_thread_priority(td, prio); 1724 1725 /* 1726 * If the thread is on a turnstile, then let the turnstile update 1727 * its state. 1728 */ 1729 if (TD_ON_LOCK(td) && oldprio != prio) 1730 turnstile_adjust(td, oldprio); 1731} 1732 1733/* 1734 * Set the base user priority, does not effect current running priority. 1735 */ 1736void 1737sched_user_prio(struct thread *td, u_char prio) 1738{ 1739 1740 td->td_base_user_pri = prio; 1741 if (td->td_lend_user_pri <= prio) 1742 return; 1743 td->td_user_pri = prio; 1744} 1745 1746void 1747sched_lend_user_prio(struct thread *td, u_char prio) 1748{ 1749 1750 THREAD_LOCK_ASSERT(td, MA_OWNED); 1751 td->td_lend_user_pri = prio; 1752 td->td_user_pri = min(prio, td->td_base_user_pri); 1753 if (td->td_priority > td->td_user_pri) 1754 sched_prio(td, td->td_user_pri); 1755 else if (td->td_priority != td->td_user_pri) 1756 td->td_flags |= TDF_NEEDRESCHED; 1757} 1758 1759/* 1760 * Handle migration from sched_switch(). This happens only for 1761 * cpu binding. 1762 */ 1763static struct mtx * 1764sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1765{ 1766 struct tdq *tdn; 1767 1768 tdn = TDQ_CPU(td->td_sched->ts_cpu); 1769#ifdef SMP 1770 tdq_load_rem(tdq, td); 1771 /* 1772 * Do the lock dance required to avoid LOR. We grab an extra 1773 * spinlock nesting to prevent preemption while we're 1774 * not holding either run-queue lock. 1775 */ 1776 spinlock_enter(); 1777 thread_lock_block(td); /* This releases the lock on tdq. */ 1778 1779 /* 1780 * Acquire both run-queue locks before placing the thread on the new 1781 * run-queue to avoid deadlocks created by placing a thread with a 1782 * blocked lock on the run-queue of a remote processor. The deadlock 1783 * occurs when a third processor attempts to lock the two queues in 1784 * question while the target processor is spinning with its own 1785 * run-queue lock held while waiting for the blocked lock to clear. 1786 */ 1787 tdq_lock_pair(tdn, tdq); 1788 tdq_add(tdn, td, flags); 1789 tdq_notify(tdn, td); 1790 TDQ_UNLOCK(tdn); 1791 spinlock_exit(); 1792#endif 1793 return (TDQ_LOCKPTR(tdn)); 1794} 1795 1796/* 1797 * Variadic version of thread_lock_unblock() that does not assume td_lock 1798 * is blocked. 1799 */ 1800static inline void 1801thread_unblock_switch(struct thread *td, struct mtx *mtx) 1802{ 1803 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1804 (uintptr_t)mtx); 1805} 1806 1807/* 1808 * Switch threads. This function has to handle threads coming in while 1809 * blocked for some reason, running, or idle. It also must deal with 1810 * migrating a thread from one queue to another as running threads may 1811 * be assigned elsewhere via binding. 1812 */ 1813void 1814sched_switch(struct thread *td, struct thread *newtd, int flags) 1815{ 1816 struct tdq *tdq; 1817 struct td_sched *ts; 1818 struct mtx *mtx; 1819 int srqflag; 1820 int cpuid, preempted; 1821 1822 THREAD_LOCK_ASSERT(td, MA_OWNED); 1823 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1824 1825 cpuid = PCPU_GET(cpuid); 1826 tdq = TDQ_CPU(cpuid); 1827 ts = td->td_sched; 1828 mtx = td->td_lock; 1829 sched_pctcpu_update(ts, 1); 1830 ts->ts_rltick = ticks; 1831 td->td_lastcpu = td->td_oncpu; 1832 td->td_oncpu = NOCPU; 1833 preempted = !(td->td_flags & TDF_SLICEEND); 1834 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND); 1835 td->td_owepreempt = 0; 1836 if (!TD_IS_IDLETHREAD(td)) 1837 tdq->tdq_switchcnt++; 1838 /* 1839 * The lock pointer in an idle thread should never change. Reset it 1840 * to CAN_RUN as well. 1841 */ 1842 if (TD_IS_IDLETHREAD(td)) { 1843 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1844 TD_SET_CAN_RUN(td); 1845 } else if (TD_IS_RUNNING(td)) { 1846 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1847 srqflag = preempted ? 1848 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1849 SRQ_OURSELF|SRQ_YIELDING; 1850#ifdef SMP 1851 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu)) 1852 ts->ts_cpu = sched_pickcpu(td, 0); 1853#endif 1854 if (ts->ts_cpu == cpuid) 1855 tdq_runq_add(tdq, td, srqflag); 1856 else { 1857 KASSERT(THREAD_CAN_MIGRATE(td) || 1858 (ts->ts_flags & TSF_BOUND) != 0, 1859 ("Thread %p shouldn't migrate", td)); 1860 mtx = sched_switch_migrate(tdq, td, srqflag); 1861 } 1862 } else { 1863 /* This thread must be going to sleep. */ 1864 TDQ_LOCK(tdq); 1865 mtx = thread_lock_block(td); 1866 tdq_load_rem(tdq, td); 1867 } 1868 /* 1869 * We enter here with the thread blocked and assigned to the 1870 * appropriate cpu run-queue or sleep-queue and with the current 1871 * thread-queue locked. 1872 */ 1873 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1874 newtd = choosethread(); 1875 /* 1876 * Call the MD code to switch contexts if necessary. 1877 */ 1878 if (td != newtd) { 1879#ifdef HWPMC_HOOKS 1880 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1881 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1882#endif 1883 SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc); 1884 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1885 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1886 sched_pctcpu_update(newtd->td_sched, 0); 1887 1888#ifdef KDTRACE_HOOKS 1889 /* 1890 * If DTrace has set the active vtime enum to anything 1891 * other than INACTIVE (0), then it should have set the 1892 * function to call. 1893 */ 1894 if (dtrace_vtime_active) 1895 (*dtrace_vtime_switch_func)(newtd); 1896#endif 1897 1898 cpu_switch(td, newtd, mtx); 1899 /* 1900 * We may return from cpu_switch on a different cpu. However, 1901 * we always return with td_lock pointing to the current cpu's 1902 * run queue lock. 1903 */ 1904 cpuid = PCPU_GET(cpuid); 1905 tdq = TDQ_CPU(cpuid); 1906 lock_profile_obtain_lock_success( 1907 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1908 1909 SDT_PROBE0(sched, , , on__cpu); 1910#ifdef HWPMC_HOOKS 1911 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1912 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1913#endif 1914 } else { 1915 thread_unblock_switch(td, mtx); 1916 SDT_PROBE0(sched, , , remain__cpu); 1917 } 1918 /* 1919 * Assert that all went well and return. 1920 */ 1921 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1922 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1923 td->td_oncpu = cpuid; 1924} 1925 1926/* 1927 * Adjust thread priorities as a result of a nice request. 1928 */ 1929void 1930sched_nice(struct proc *p, int nice) 1931{ 1932 struct thread *td; 1933 1934 PROC_LOCK_ASSERT(p, MA_OWNED); 1935 1936 p->p_nice = nice; 1937 FOREACH_THREAD_IN_PROC(p, td) { 1938 thread_lock(td); 1939 sched_priority(td); 1940 sched_prio(td, td->td_base_user_pri); 1941 thread_unlock(td); 1942 } 1943} 1944 1945/* 1946 * Record the sleep time for the interactivity scorer. 1947 */ 1948void 1949sched_sleep(struct thread *td, int prio) 1950{ 1951 1952 THREAD_LOCK_ASSERT(td, MA_OWNED); 1953 1954 td->td_slptick = ticks; 1955 if (TD_IS_SUSPENDED(td) || prio >= PSOCK) 1956 td->td_flags |= TDF_CANSWAP; 1957 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1958 return; 1959 if (static_boost == 1 && prio) 1960 sched_prio(td, prio); 1961 else if (static_boost && td->td_priority > static_boost) 1962 sched_prio(td, static_boost); 1963} 1964 1965/* 1966 * Schedule a thread to resume execution and record how long it voluntarily 1967 * slept. We also update the pctcpu, interactivity, and priority. 1968 */ 1969void 1970sched_wakeup(struct thread *td) 1971{ 1972 struct td_sched *ts; 1973 int slptick; 1974 1975 THREAD_LOCK_ASSERT(td, MA_OWNED); 1976 ts = td->td_sched; 1977 td->td_flags &= ~TDF_CANSWAP; 1978 /* 1979 * If we slept for more than a tick update our interactivity and 1980 * priority. 1981 */ 1982 slptick = td->td_slptick; 1983 td->td_slptick = 0; 1984 if (slptick && slptick != ticks) { 1985 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT; 1986 sched_interact_update(td); 1987 sched_pctcpu_update(ts, 0); 1988 } 1989 /* Reset the slice value after we sleep. */ 1990 ts->ts_slice = sched_slice; 1991 sched_add(td, SRQ_BORING); 1992} 1993 1994/* 1995 * Penalize the parent for creating a new child and initialize the child's 1996 * priority. 1997 */ 1998void 1999sched_fork(struct thread *td, struct thread *child) 2000{ 2001 THREAD_LOCK_ASSERT(td, MA_OWNED); 2002 sched_pctcpu_update(td->td_sched, 1); 2003 sched_fork_thread(td, child); 2004 /* 2005 * Penalize the parent and child for forking. 2006 */ 2007 sched_interact_fork(child); 2008 sched_priority(child); 2009 td->td_sched->ts_runtime += tickincr; 2010 sched_interact_update(td); 2011 sched_priority(td); 2012} 2013 2014/* 2015 * Fork a new thread, may be within the same process. 2016 */ 2017void 2018sched_fork_thread(struct thread *td, struct thread *child) 2019{ 2020 struct td_sched *ts; 2021 struct td_sched *ts2; 2022 2023 THREAD_LOCK_ASSERT(td, MA_OWNED); 2024 /* 2025 * Initialize child. 2026 */ 2027 ts = td->td_sched; 2028 ts2 = child->td_sched; 2029 child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 2030 child->td_cpuset = cpuset_ref(td->td_cpuset); 2031 ts2->ts_cpu = ts->ts_cpu; 2032 ts2->ts_flags = 0; 2033 /* 2034 * Grab our parents cpu estimation information. 2035 */ 2036 ts2->ts_ticks = ts->ts_ticks; 2037 ts2->ts_ltick = ts->ts_ltick; 2038 ts2->ts_ftick = ts->ts_ftick; 2039 /* 2040 * Do not inherit any borrowed priority from the parent. 2041 */ 2042 child->td_priority = child->td_base_pri; 2043 /* 2044 * And update interactivity score. 2045 */ 2046 ts2->ts_slptime = ts->ts_slptime; 2047 ts2->ts_runtime = ts->ts_runtime; 2048 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 2049#ifdef KTR 2050 bzero(ts2->ts_name, sizeof(ts2->ts_name)); 2051#endif 2052} 2053 2054/* 2055 * Adjust the priority class of a thread. 2056 */ 2057void 2058sched_class(struct thread *td, int class) 2059{ 2060 2061 THREAD_LOCK_ASSERT(td, MA_OWNED); 2062 if (td->td_pri_class == class) 2063 return; 2064 td->td_pri_class = class; 2065} 2066 2067/* 2068 * Return some of the child's priority and interactivity to the parent. 2069 */ 2070void 2071sched_exit(struct proc *p, struct thread *child) 2072{ 2073 struct thread *td; 2074 2075 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit", 2076 "prio:%d", child->td_priority); 2077 PROC_LOCK_ASSERT(p, MA_OWNED); 2078 td = FIRST_THREAD_IN_PROC(p); 2079 sched_exit_thread(td, child); 2080} 2081 2082/* 2083 * Penalize another thread for the time spent on this one. This helps to 2084 * worsen the priority and interactivity of processes which schedule batch 2085 * jobs such as make. This has little effect on the make process itself but 2086 * causes new processes spawned by it to receive worse scores immediately. 2087 */ 2088void 2089sched_exit_thread(struct thread *td, struct thread *child) 2090{ 2091 2092 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit", 2093 "prio:%d", child->td_priority); 2094 /* 2095 * Give the child's runtime to the parent without returning the 2096 * sleep time as a penalty to the parent. This causes shells that 2097 * launch expensive things to mark their children as expensive. 2098 */ 2099 thread_lock(td); 2100 td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2101 sched_interact_update(td); 2102 sched_priority(td); 2103 thread_unlock(td); 2104} 2105 2106void 2107sched_preempt(struct thread *td) 2108{ 2109 struct tdq *tdq; 2110 2111 SDT_PROBE2(sched, , , surrender, td, td->td_proc); 2112 2113 thread_lock(td); 2114 tdq = TDQ_SELF(); 2115 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2116 tdq->tdq_ipipending = 0; 2117 if (td->td_priority > tdq->tdq_lowpri) { 2118 int flags; 2119 2120 flags = SW_INVOL | SW_PREEMPT; 2121 if (td->td_critnest > 1) 2122 td->td_owepreempt = 1; 2123 else if (TD_IS_IDLETHREAD(td)) 2124 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL); 2125 else 2126 mi_switch(flags | SWT_REMOTEPREEMPT, NULL); 2127 } 2128 thread_unlock(td); 2129} 2130 2131/* 2132 * Fix priorities on return to user-space. Priorities may be elevated due 2133 * to static priorities in msleep() or similar. 2134 */ 2135void 2136sched_userret(struct thread *td) 2137{ 2138 /* 2139 * XXX we cheat slightly on the locking here to avoid locking in 2140 * the usual case. Setting td_priority here is essentially an 2141 * incomplete workaround for not setting it properly elsewhere. 2142 * Now that some interrupt handlers are threads, not setting it 2143 * properly elsewhere can clobber it in the window between setting 2144 * it here and returning to user mode, so don't waste time setting 2145 * it perfectly here. 2146 */ 2147 KASSERT((td->td_flags & TDF_BORROWING) == 0, 2148 ("thread with borrowed priority returning to userland")); 2149 if (td->td_priority != td->td_user_pri) { 2150 thread_lock(td); 2151 td->td_priority = td->td_user_pri; 2152 td->td_base_pri = td->td_user_pri; 2153 tdq_setlowpri(TDQ_SELF(), td); 2154 thread_unlock(td); 2155 } 2156} 2157 2158/* 2159 * Handle a stathz tick. This is really only relevant for timeshare 2160 * threads. 2161 */ 2162void 2163sched_clock(struct thread *td) 2164{ 2165 struct tdq *tdq; 2166 struct td_sched *ts; 2167 2168 THREAD_LOCK_ASSERT(td, MA_OWNED); 2169 tdq = TDQ_SELF(); 2170#ifdef SMP 2171 /* 2172 * We run the long term load balancer infrequently on the first cpu. 2173 */ 2174 if (balance_tdq == tdq) { 2175 if (balance_ticks && --balance_ticks == 0) 2176 sched_balance(); 2177 } 2178#endif 2179 /* 2180 * Save the old switch count so we have a record of the last ticks 2181 * activity. Initialize the new switch count based on our load. 2182 * If there is some activity seed it to reflect that. 2183 */ 2184 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2185 tdq->tdq_switchcnt = tdq->tdq_load; 2186 /* 2187 * Advance the insert index once for each tick to ensure that all 2188 * threads get a chance to run. 2189 */ 2190 if (tdq->tdq_idx == tdq->tdq_ridx) { 2191 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2192 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2193 tdq->tdq_ridx = tdq->tdq_idx; 2194 } 2195 ts = td->td_sched; 2196 sched_pctcpu_update(ts, 1); 2197 if (td->td_pri_class & PRI_FIFO_BIT) 2198 return; 2199 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) { 2200 /* 2201 * We used a tick; charge it to the thread so 2202 * that we can compute our interactivity. 2203 */ 2204 td->td_sched->ts_runtime += tickincr; 2205 sched_interact_update(td); 2206 sched_priority(td); 2207 } 2208 2209 /* 2210 * Force a context switch if the current thread has used up a full 2211 * time slice (default is 100ms). 2212 */ 2213 if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) { 2214 ts->ts_slice = sched_slice; 2215 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND; 2216 } 2217} 2218 2219/* 2220 * Called once per hz tick. 2221 */ 2222void 2223sched_tick(int cnt) 2224{ 2225 2226} 2227 2228/* 2229 * Return whether the current CPU has runnable tasks. Used for in-kernel 2230 * cooperative idle threads. 2231 */ 2232int 2233sched_runnable(void) 2234{ 2235 struct tdq *tdq; 2236 int load; 2237 2238 load = 1; 2239 2240 tdq = TDQ_SELF(); 2241 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2242 if (tdq->tdq_load > 0) 2243 goto out; 2244 } else 2245 if (tdq->tdq_load - 1 > 0) 2246 goto out; 2247 load = 0; 2248out: 2249 return (load); 2250} 2251 2252/* 2253 * Choose the highest priority thread to run. The thread is removed from 2254 * the run-queue while running however the load remains. For SMP we set 2255 * the tdq in the global idle bitmask if it idles here. 2256 */ 2257struct thread * 2258sched_choose(void) 2259{ 2260 struct thread *td; 2261 struct tdq *tdq; 2262 2263 tdq = TDQ_SELF(); 2264 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2265 td = tdq_choose(tdq); 2266 if (td) { 2267 tdq_runq_rem(tdq, td); 2268 tdq->tdq_lowpri = td->td_priority; 2269 return (td); 2270 } 2271 tdq->tdq_lowpri = PRI_MAX_IDLE; 2272 return (PCPU_GET(idlethread)); 2273} 2274 2275/* 2276 * Set owepreempt if necessary. Preemption never happens directly in ULE, 2277 * we always request it once we exit a critical section. 2278 */ 2279static inline void 2280sched_setpreempt(struct thread *td) 2281{ 2282 struct thread *ctd; 2283 int cpri; 2284 int pri; 2285 2286 THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2287 2288 ctd = curthread; 2289 pri = td->td_priority; 2290 cpri = ctd->td_priority; 2291 if (pri < cpri) 2292 ctd->td_flags |= TDF_NEEDRESCHED; 2293 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2294 return; 2295 if (!sched_shouldpreempt(pri, cpri, 0)) 2296 return; 2297 ctd->td_owepreempt = 1; 2298} 2299 2300/* 2301 * Add a thread to a thread queue. Select the appropriate runq and add the 2302 * thread to it. This is the internal function called when the tdq is 2303 * predetermined. 2304 */ 2305void 2306tdq_add(struct tdq *tdq, struct thread *td, int flags) 2307{ 2308 2309 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2310 KASSERT((td->td_inhibitors == 0), 2311 ("sched_add: trying to run inhibited thread")); 2312 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2313 ("sched_add: bad thread state")); 2314 KASSERT(td->td_flags & TDF_INMEM, 2315 ("sched_add: thread swapped out")); 2316 2317 if (td->td_priority < tdq->tdq_lowpri) 2318 tdq->tdq_lowpri = td->td_priority; 2319 tdq_runq_add(tdq, td, flags); 2320 tdq_load_add(tdq, td); 2321} 2322 2323/* 2324 * Select the target thread queue and add a thread to it. Request 2325 * preemption or IPI a remote processor if required. 2326 */ 2327void 2328sched_add(struct thread *td, int flags) 2329{ 2330 struct tdq *tdq; 2331#ifdef SMP 2332 int cpu; 2333#endif 2334 2335 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 2336 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 2337 sched_tdname(curthread)); 2338 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 2339 KTR_ATTR_LINKED, sched_tdname(td)); 2340 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 2341 flags & SRQ_PREEMPTED); 2342 THREAD_LOCK_ASSERT(td, MA_OWNED); 2343 /* 2344 * Recalculate the priority before we select the target cpu or 2345 * run-queue. 2346 */ 2347 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2348 sched_priority(td); 2349#ifdef SMP 2350 /* 2351 * Pick the destination cpu and if it isn't ours transfer to the 2352 * target cpu. 2353 */ 2354 cpu = sched_pickcpu(td, flags); 2355 tdq = sched_setcpu(td, cpu, flags); 2356 tdq_add(tdq, td, flags); 2357 if (cpu != PCPU_GET(cpuid)) { 2358 tdq_notify(tdq, td); 2359 return; 2360 } 2361#else 2362 tdq = TDQ_SELF(); 2363 TDQ_LOCK(tdq); 2364 /* 2365 * Now that the thread is moving to the run-queue, set the lock 2366 * to the scheduler's lock. 2367 */ 2368 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2369 tdq_add(tdq, td, flags); 2370#endif 2371 if (!(flags & SRQ_YIELDING)) 2372 sched_setpreempt(td); 2373} 2374 2375/* 2376 * Remove a thread from a run-queue without running it. This is used 2377 * when we're stealing a thread from a remote queue. Otherwise all threads 2378 * exit by calling sched_exit_thread() and sched_throw() themselves. 2379 */ 2380void 2381sched_rem(struct thread *td) 2382{ 2383 struct tdq *tdq; 2384 2385 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 2386 "prio:%d", td->td_priority); 2387 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); 2388 tdq = TDQ_CPU(td->td_sched->ts_cpu); 2389 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2390 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2391 KASSERT(TD_ON_RUNQ(td), 2392 ("sched_rem: thread not on run queue")); 2393 tdq_runq_rem(tdq, td); 2394 tdq_load_rem(tdq, td); 2395 TD_SET_CAN_RUN(td); 2396 if (td->td_priority == tdq->tdq_lowpri) 2397 tdq_setlowpri(tdq, NULL); 2398} 2399 2400/* 2401 * Fetch cpu utilization information. Updates on demand. 2402 */ 2403fixpt_t 2404sched_pctcpu(struct thread *td) 2405{ 2406 fixpt_t pctcpu; 2407 struct td_sched *ts; 2408 2409 pctcpu = 0; 2410 ts = td->td_sched; 2411 if (ts == NULL) 2412 return (0); 2413 2414 THREAD_LOCK_ASSERT(td, MA_OWNED); 2415 sched_pctcpu_update(ts, TD_IS_RUNNING(td)); 2416 if (ts->ts_ticks) { 2417 int rtick; 2418 2419 /* How many rtick per second ? */ 2420 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2421 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2422 } 2423 2424 return (pctcpu); 2425} 2426 2427/* 2428 * Enforce affinity settings for a thread. Called after adjustments to 2429 * cpumask. 2430 */ 2431void 2432sched_affinity(struct thread *td) 2433{ 2434#ifdef SMP 2435 struct td_sched *ts; 2436 2437 THREAD_LOCK_ASSERT(td, MA_OWNED); 2438 ts = td->td_sched; 2439 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2440 return; 2441 if (TD_ON_RUNQ(td)) { 2442 sched_rem(td); 2443 sched_add(td, SRQ_BORING); 2444 return; 2445 } 2446 if (!TD_IS_RUNNING(td)) 2447 return; 2448 /* 2449 * Force a switch before returning to userspace. If the 2450 * target thread is not running locally send an ipi to force 2451 * the issue. 2452 */ 2453 td->td_flags |= TDF_NEEDRESCHED; 2454 if (td != curthread) 2455 ipi_cpu(ts->ts_cpu, IPI_PREEMPT); 2456#endif 2457} 2458 2459/* 2460 * Bind a thread to a target cpu. 2461 */ 2462void 2463sched_bind(struct thread *td, int cpu) 2464{ 2465 struct td_sched *ts; 2466 2467 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2468 KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 2469 ts = td->td_sched; 2470 if (ts->ts_flags & TSF_BOUND) 2471 sched_unbind(td); 2472 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td)); 2473 ts->ts_flags |= TSF_BOUND; 2474 sched_pin(); 2475 if (PCPU_GET(cpuid) == cpu) 2476 return; 2477 ts->ts_cpu = cpu; 2478 /* When we return from mi_switch we'll be on the correct cpu. */ 2479 mi_switch(SW_VOL, NULL); 2480} 2481 2482/* 2483 * Release a bound thread. 2484 */ 2485void 2486sched_unbind(struct thread *td) 2487{ 2488 struct td_sched *ts; 2489 2490 THREAD_LOCK_ASSERT(td, MA_OWNED); 2491 KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 2492 ts = td->td_sched; 2493 if ((ts->ts_flags & TSF_BOUND) == 0) 2494 return; 2495 ts->ts_flags &= ~TSF_BOUND; 2496 sched_unpin(); 2497} 2498 2499int 2500sched_is_bound(struct thread *td) 2501{ 2502 THREAD_LOCK_ASSERT(td, MA_OWNED); 2503 return (td->td_sched->ts_flags & TSF_BOUND); 2504} 2505 2506/* 2507 * Basic yield call. 2508 */ 2509void 2510sched_relinquish(struct thread *td) 2511{ 2512 thread_lock(td); 2513 mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 2514 thread_unlock(td); 2515} 2516 2517/* 2518 * Return the total system load. 2519 */ 2520int 2521sched_load(void) 2522{ 2523#ifdef SMP 2524 int total; 2525 int i; 2526 2527 total = 0; 2528 CPU_FOREACH(i) 2529 total += TDQ_CPU(i)->tdq_sysload; 2530 return (total); 2531#else 2532 return (TDQ_SELF()->tdq_sysload); 2533#endif 2534} 2535 2536int 2537sched_sizeof_proc(void) 2538{ 2539 return (sizeof(struct proc)); 2540} 2541 2542int 2543sched_sizeof_thread(void) 2544{ 2545 return (sizeof(struct thread) + sizeof(struct td_sched)); 2546} 2547 2548#ifdef SMP 2549#define TDQ_IDLESPIN(tdq) \ 2550 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0) 2551#else 2552#define TDQ_IDLESPIN(tdq) 1 2553#endif 2554 2555/* 2556 * The actual idle process. 2557 */ 2558void 2559sched_idletd(void *dummy) 2560{ 2561 struct thread *td; 2562 struct tdq *tdq; 2563 int oldswitchcnt, switchcnt; 2564 int i; 2565 2566 mtx_assert(&Giant, MA_NOTOWNED); 2567 td = curthread; 2568 tdq = TDQ_SELF(); 2569 oldswitchcnt = -1; 2570 for (;;) { 2571 if (tdq->tdq_load) { 2572 thread_lock(td); 2573 mi_switch(SW_VOL | SWT_IDLE, NULL); 2574 thread_unlock(td); 2575 } 2576 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2577#ifdef SMP 2578 if (switchcnt != oldswitchcnt) { 2579 oldswitchcnt = switchcnt; 2580 if (tdq_idled(tdq) == 0) 2581 continue; 2582 } 2583 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2584#else 2585 oldswitchcnt = switchcnt; 2586#endif 2587 /* 2588 * If we're switching very frequently, spin while checking 2589 * for load rather than entering a low power state that 2590 * may require an IPI. However, don't do any busy 2591 * loops while on SMT machines as this simply steals 2592 * cycles from cores doing useful work. 2593 */ 2594 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) { 2595 for (i = 0; i < sched_idlespins; i++) { 2596 if (tdq->tdq_load) 2597 break; 2598 cpu_spinwait(); 2599 } 2600 } 2601 2602 /* If there was context switch during spin, restart it. */ 2603 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2604 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt) 2605 continue; 2606 2607 /* Run main MD idle handler. */ 2608 tdq->tdq_cpu_idle = 1; 2609 cpu_idle(switchcnt * 4 > sched_idlespinthresh); 2610 tdq->tdq_cpu_idle = 0; 2611 2612 /* 2613 * Account thread-less hardware interrupts and 2614 * other wakeup reasons equal to context switches. 2615 */ 2616 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2617 if (switchcnt != oldswitchcnt) 2618 continue; 2619 tdq->tdq_switchcnt++; 2620 oldswitchcnt++; 2621 } 2622} 2623 2624/* 2625 * A CPU is entering for the first time or a thread is exiting. 2626 */ 2627void 2628sched_throw(struct thread *td) 2629{ 2630 struct thread *newtd; 2631 struct tdq *tdq; 2632 2633 tdq = TDQ_SELF(); 2634 if (td == NULL) { 2635 /* Correct spinlock nesting and acquire the correct lock. */ 2636 TDQ_LOCK(tdq); 2637 spinlock_exit(); 2638 PCPU_SET(switchtime, cpu_ticks()); 2639 PCPU_SET(switchticks, ticks); 2640 } else { 2641 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2642 tdq_load_rem(tdq, td); 2643 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2644 } 2645 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2646 newtd = choosethread(); 2647 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2648 cpu_throw(td, newtd); /* doesn't return */ 2649} 2650 2651/* 2652 * This is called from fork_exit(). Just acquire the correct locks and 2653 * let fork do the rest of the work. 2654 */ 2655void 2656sched_fork_exit(struct thread *td) 2657{ 2658 struct td_sched *ts; 2659 struct tdq *tdq; 2660 int cpuid; 2661 2662 /* 2663 * Finish setting up thread glue so that it begins execution in a 2664 * non-nested critical section with the scheduler lock held. 2665 */ 2666 cpuid = PCPU_GET(cpuid); 2667 tdq = TDQ_CPU(cpuid); 2668 ts = td->td_sched; 2669 if (TD_IS_IDLETHREAD(td)) 2670 td->td_lock = TDQ_LOCKPTR(tdq); 2671 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2672 td->td_oncpu = cpuid; 2673 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2674 lock_profile_obtain_lock_success( 2675 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2676} 2677 2678/* 2679 * Create on first use to catch odd startup conditons. 2680 */ 2681char * 2682sched_tdname(struct thread *td) 2683{ 2684#ifdef KTR 2685 struct td_sched *ts; 2686 2687 ts = td->td_sched; 2688 if (ts->ts_name[0] == '\0') 2689 snprintf(ts->ts_name, sizeof(ts->ts_name), 2690 "%s tid %d", td->td_name, td->td_tid); 2691 return (ts->ts_name); 2692#else 2693 return (td->td_name); 2694#endif 2695} 2696 2697#ifdef KTR 2698void 2699sched_clear_tdname(struct thread *td) 2700{ 2701 struct td_sched *ts; 2702 2703 ts = td->td_sched; 2704 ts->ts_name[0] = '\0'; 2705} 2706#endif 2707 2708#ifdef SMP 2709 2710/* 2711 * Build the CPU topology dump string. Is recursively called to collect 2712 * the topology tree. 2713 */ 2714static int 2715sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg, 2716 int indent) 2717{ 2718 char cpusetbuf[CPUSETBUFSIZ]; 2719 int i, first; 2720 2721 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent, 2722 "", 1 + indent / 2, cg->cg_level); 2723 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "", 2724 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask)); 2725 first = TRUE; 2726 for (i = 0; i < MAXCPU; i++) { 2727 if (CPU_ISSET(i, &cg->cg_mask)) { 2728 if (!first) 2729 sbuf_printf(sb, ", "); 2730 else 2731 first = FALSE; 2732 sbuf_printf(sb, "%d", i); 2733 } 2734 } 2735 sbuf_printf(sb, "</cpu>\n"); 2736 2737 if (cg->cg_flags != 0) { 2738 sbuf_printf(sb, "%*s <flags>", indent, ""); 2739 if ((cg->cg_flags & CG_FLAG_HTT) != 0) 2740 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>"); 2741 if ((cg->cg_flags & CG_FLAG_THREAD) != 0) 2742 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>"); 2743 if ((cg->cg_flags & CG_FLAG_SMT) != 0) 2744 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>"); 2745 sbuf_printf(sb, "</flags>\n"); 2746 } 2747 2748 if (cg->cg_children > 0) { 2749 sbuf_printf(sb, "%*s <children>\n", indent, ""); 2750 for (i = 0; i < cg->cg_children; i++) 2751 sysctl_kern_sched_topology_spec_internal(sb, 2752 &cg->cg_child[i], indent+2); 2753 sbuf_printf(sb, "%*s </children>\n", indent, ""); 2754 } 2755 sbuf_printf(sb, "%*s</group>\n", indent, ""); 2756 return (0); 2757} 2758 2759/* 2760 * Sysctl handler for retrieving topology dump. It's a wrapper for 2761 * the recursive sysctl_kern_smp_topology_spec_internal(). 2762 */ 2763static int 2764sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS) 2765{ 2766 struct sbuf *topo; 2767 int err; 2768 2769 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized")); 2770 2771 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND); 2772 if (topo == NULL) 2773 return (ENOMEM); 2774 2775 sbuf_printf(topo, "<groups>\n"); 2776 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1); 2777 sbuf_printf(topo, "</groups>\n"); 2778 2779 if (err == 0) { 2780 sbuf_finish(topo); 2781 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo)); 2782 } 2783 sbuf_delete(topo); 2784 return (err); 2785} 2786 2787#endif 2788 2789static int 2790sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 2791{ 2792 int error, new_val, period; 2793 2794 period = 1000000 / realstathz; 2795 new_val = period * sched_slice; 2796 error = sysctl_handle_int(oidp, &new_val, 0, req); 2797 if (error != 0 || req->newptr == NULL) 2798 return (error); 2799 if (new_val <= 0) 2800 return (EINVAL); 2801 sched_slice = imax(1, (new_val + period / 2) / period); 2802 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 2803 realstathz); 2804 return (0); 2805} 2806 2807SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2808SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2809 "Scheduler name"); 2810SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW, 2811 NULL, 0, sysctl_kern_quantum, "I", 2812 "Quantum for timeshare threads in microseconds"); 2813SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2814 "Quantum for timeshare threads in stathz ticks"); 2815SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2816 "Interactivity score threshold"); 2817SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, 2818 &preempt_thresh, 0, 2819 "Maximal (lowest) priority for preemption"); 2820SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0, 2821 "Assign static kernel priorities to sleeping threads"); 2822SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0, 2823 "Number of times idle thread will spin waiting for new work"); 2824SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, 2825 &sched_idlespinthresh, 0, 2826 "Threshold before we will permit idle thread spinning"); 2827#ifdef SMP 2828SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2829 "Number of hz ticks to keep thread affinity for"); 2830SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2831 "Enables the long-term load balancer"); 2832SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2833 &balance_interval, 0, 2834 "Average period in stathz ticks to run the long-term balancer"); 2835SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2836 "Attempts to steal work from other cores before idling"); 2837SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2838 "Minimum load on remote CPU before we'll steal"); 2839SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING | 2840 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A", 2841 "XML dump of detected CPU topology"); 2842#endif 2843 2844/* ps compat. All cpu percentages from ULE are weighted. */ 2845static int ccpu = 0; 2846SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2847