sched_ule.c revision 233814
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 233814 2012-04-02 20:34:15Z jhb $"); 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/smp.h> 57#include <sys/sx.h> 58#include <sys/sysctl.h> 59#include <sys/sysproto.h> 60#include <sys/turnstile.h> 61#include <sys/umtx.h> 62#include <sys/vmmeter.h> 63#include <sys/cpuset.h> 64#include <sys/sbuf.h> 65 66#ifdef HWPMC_HOOKS 67#include <sys/pmckern.h> 68#endif 69 70#ifdef KDTRACE_HOOKS 71#include <sys/dtrace_bsd.h> 72int dtrace_vtime_active; 73dtrace_vtime_switch_func_t dtrace_vtime_switch_func; 74#endif 75 76#include <machine/cpu.h> 77#include <machine/smp.h> 78 79#if defined(__powerpc__) && defined(E500) 80#error "This architecture is not currently compatible with ULE" 81#endif 82 83#define KTR_ULE 0 84 85#define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX))) 86#define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU))) 87#define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load")) 88 89/* 90 * Thread scheduler specific section. All fields are protected 91 * by the thread lock. 92 */ 93struct td_sched { 94 struct runq *ts_runq; /* Run-queue we're queued on. */ 95 short ts_flags; /* TSF_* flags. */ 96 u_char ts_cpu; /* CPU that we have affinity for. */ 97 int ts_rltick; /* Real last tick, for affinity. */ 98 int ts_slice; /* Ticks of slice remaining. */ 99 u_int ts_slptime; /* Number of ticks we vol. slept */ 100 u_int ts_runtime; /* Number of ticks we were running */ 101 int ts_ltick; /* Last tick that we were running on */ 102 int ts_incrtick; /* Last tick that we incremented 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 = 16; 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 *); 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); 330 331/* 332 * Print the threads waiting on a run-queue. 333 */ 334static void 335runq_print(struct runq *rq) 336{ 337 struct rqhead *rqh; 338 struct thread *td; 339 int pri; 340 int j; 341 int i; 342 343 for (i = 0; i < RQB_LEN; i++) { 344 printf("\t\trunq bits %d 0x%zx\n", 345 i, rq->rq_status.rqb_bits[i]); 346 for (j = 0; j < RQB_BPW; j++) 347 if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 348 pri = j + (i << RQB_L2BPW); 349 rqh = &rq->rq_queues[pri]; 350 TAILQ_FOREACH(td, rqh, td_runq) { 351 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 352 td, td->td_name, td->td_priority, 353 td->td_rqindex, pri); 354 } 355 } 356 } 357} 358 359/* 360 * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 361 */ 362void 363tdq_print(int cpu) 364{ 365 struct tdq *tdq; 366 367 tdq = TDQ_CPU(cpu); 368 369 printf("tdq %d:\n", TDQ_ID(tdq)); 370 printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 371 printf("\tLock name: %s\n", tdq->tdq_name); 372 printf("\tload: %d\n", tdq->tdq_load); 373 printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt); 374 printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt); 375 printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 376 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 377 printf("\tload transferable: %d\n", tdq->tdq_transferable); 378 printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 379 printf("\trealtime runq:\n"); 380 runq_print(&tdq->tdq_realtime); 381 printf("\ttimeshare runq:\n"); 382 runq_print(&tdq->tdq_timeshare); 383 printf("\tidle runq:\n"); 384 runq_print(&tdq->tdq_idle); 385} 386 387static inline int 388sched_shouldpreempt(int pri, int cpri, int remote) 389{ 390 /* 391 * If the new priority is not better than the current priority there is 392 * nothing to do. 393 */ 394 if (pri >= cpri) 395 return (0); 396 /* 397 * Always preempt idle. 398 */ 399 if (cpri >= PRI_MIN_IDLE) 400 return (1); 401 /* 402 * If preemption is disabled don't preempt others. 403 */ 404 if (preempt_thresh == 0) 405 return (0); 406 /* 407 * Preempt if we exceed the threshold. 408 */ 409 if (pri <= preempt_thresh) 410 return (1); 411 /* 412 * If we're interactive or better and there is non-interactive 413 * or worse running preempt only remote processors. 414 */ 415 if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT) 416 return (1); 417 return (0); 418} 419 420/* 421 * Add a thread to the actual run-queue. Keeps transferable counts up to 422 * date with what is actually on the run-queue. Selects the correct 423 * queue position for timeshare threads. 424 */ 425static __inline void 426tdq_runq_add(struct tdq *tdq, struct thread *td, int flags) 427{ 428 struct td_sched *ts; 429 u_char pri; 430 431 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 432 THREAD_LOCK_ASSERT(td, MA_OWNED); 433 434 pri = td->td_priority; 435 ts = td->td_sched; 436 TD_SET_RUNQ(td); 437 if (THREAD_CAN_MIGRATE(td)) { 438 tdq->tdq_transferable++; 439 ts->ts_flags |= TSF_XFERABLE; 440 } 441 if (pri < PRI_MIN_BATCH) { 442 ts->ts_runq = &tdq->tdq_realtime; 443 } else if (pri <= PRI_MAX_BATCH) { 444 ts->ts_runq = &tdq->tdq_timeshare; 445 KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH, 446 ("Invalid priority %d on timeshare runq", pri)); 447 /* 448 * This queue contains only priorities between MIN and MAX 449 * realtime. Use the whole queue to represent these values. 450 */ 451 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 452 pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE; 453 pri = (pri + tdq->tdq_idx) % RQ_NQS; 454 /* 455 * This effectively shortens the queue by one so we 456 * can have a one slot difference between idx and 457 * ridx while we wait for threads to drain. 458 */ 459 if (tdq->tdq_ridx != tdq->tdq_idx && 460 pri == tdq->tdq_ridx) 461 pri = (unsigned char)(pri - 1) % RQ_NQS; 462 } else 463 pri = tdq->tdq_ridx; 464 runq_add_pri(ts->ts_runq, td, pri, flags); 465 return; 466 } else 467 ts->ts_runq = &tdq->tdq_idle; 468 runq_add(ts->ts_runq, td, flags); 469} 470 471/* 472 * Remove a thread from a run-queue. This typically happens when a thread 473 * is selected to run. Running threads are not on the queue and the 474 * transferable count does not reflect them. 475 */ 476static __inline void 477tdq_runq_rem(struct tdq *tdq, struct thread *td) 478{ 479 struct td_sched *ts; 480 481 ts = td->td_sched; 482 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 483 KASSERT(ts->ts_runq != NULL, 484 ("tdq_runq_remove: thread %p null ts_runq", td)); 485 if (ts->ts_flags & TSF_XFERABLE) { 486 tdq->tdq_transferable--; 487 ts->ts_flags &= ~TSF_XFERABLE; 488 } 489 if (ts->ts_runq == &tdq->tdq_timeshare) { 490 if (tdq->tdq_idx != tdq->tdq_ridx) 491 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx); 492 else 493 runq_remove_idx(ts->ts_runq, td, NULL); 494 } else 495 runq_remove(ts->ts_runq, td); 496} 497 498/* 499 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 500 * for this thread to the referenced thread queue. 501 */ 502static void 503tdq_load_add(struct tdq *tdq, struct thread *td) 504{ 505 506 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 507 THREAD_LOCK_ASSERT(td, MA_OWNED); 508 509 tdq->tdq_load++; 510 if ((td->td_flags & TDF_NOLOAD) == 0) 511 tdq->tdq_sysload++; 512 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 513} 514 515/* 516 * Remove the load from a thread that is transitioning to a sleep state or 517 * exiting. 518 */ 519static void 520tdq_load_rem(struct tdq *tdq, struct thread *td) 521{ 522 523 THREAD_LOCK_ASSERT(td, MA_OWNED); 524 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 525 KASSERT(tdq->tdq_load != 0, 526 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 527 528 tdq->tdq_load--; 529 if ((td->td_flags & TDF_NOLOAD) == 0) 530 tdq->tdq_sysload--; 531 KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 532} 533 534/* 535 * Set lowpri to its exact value by searching the run-queue and 536 * evaluating curthread. curthread may be passed as an optimization. 537 */ 538static void 539tdq_setlowpri(struct tdq *tdq, struct thread *ctd) 540{ 541 struct thread *td; 542 543 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 544 if (ctd == NULL) 545 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 546 td = tdq_choose(tdq); 547 if (td == NULL || td->td_priority > ctd->td_priority) 548 tdq->tdq_lowpri = ctd->td_priority; 549 else 550 tdq->tdq_lowpri = td->td_priority; 551} 552 553#ifdef SMP 554struct cpu_search { 555 cpuset_t cs_mask; 556 u_int cs_prefer; 557 int cs_pri; /* Min priority for low. */ 558 int cs_limit; /* Max load for low, min load for high. */ 559 int cs_cpu; 560 int cs_load; 561}; 562 563#define CPU_SEARCH_LOWEST 0x1 564#define CPU_SEARCH_HIGHEST 0x2 565#define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 566 567#define CPUSET_FOREACH(cpu, mask) \ 568 for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \ 569 if (CPU_ISSET(cpu, &mask)) 570 571static __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low, 572 struct cpu_search *high, const int match); 573int cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low); 574int cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high); 575int cpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 576 struct cpu_search *high); 577 578/* 579 * Search the tree of cpu_groups for the lowest or highest loaded cpu 580 * according to the match argument. This routine actually compares the 581 * load on all paths through the tree and finds the least loaded cpu on 582 * the least loaded path, which may differ from the least loaded cpu in 583 * the system. This balances work among caches and busses. 584 * 585 * This inline is instantiated in three forms below using constants for the 586 * match argument. It is reduced to the minimum set for each case. It is 587 * also recursive to the depth of the tree. 588 */ 589static __inline int 590cpu_search(const struct cpu_group *cg, struct cpu_search *low, 591 struct cpu_search *high, const int match) 592{ 593 struct cpu_search lgroup; 594 struct cpu_search hgroup; 595 cpuset_t cpumask; 596 struct cpu_group *child; 597 struct tdq *tdq; 598 int cpu, i, hload, lload, load, total, rnd; 599 600 total = 0; 601 cpumask = cg->cg_mask; 602 if (match & CPU_SEARCH_LOWEST) { 603 lload = INT_MAX; 604 low->cs_load = INT_MAX; 605 lgroup = *low; 606 } 607 if (match & CPU_SEARCH_HIGHEST) { 608 hload = -1; 609 high->cs_load = -1; 610 hgroup = *high; 611 } 612 613 /* Iterate through the child CPU groups and then remaining CPUs. */ 614 for (i = 0, cpu = 0; i <= cg->cg_children; ) { 615 if (i >= cg->cg_children) { 616 while (cpu <= mp_maxid && !CPU_ISSET(cpu, &cpumask)) 617 cpu++; 618 if (cpu > mp_maxid) 619 break; 620 child = NULL; 621 } else 622 child = &cg->cg_child[i]; 623 624 if (child) { /* Handle child CPU group. */ 625 CPU_NAND(&cpumask, &child->cg_mask); 626 switch (match) { 627 case CPU_SEARCH_LOWEST: 628 load = cpu_search_lowest(child, &lgroup); 629 break; 630 case CPU_SEARCH_HIGHEST: 631 load = cpu_search_highest(child, &hgroup); 632 break; 633 case CPU_SEARCH_BOTH: 634 load = cpu_search_both(child, &lgroup, &hgroup); 635 break; 636 } 637 } else { /* Handle child CPU. */ 638 tdq = TDQ_CPU(cpu); 639 load = tdq->tdq_load * 256; 640 rnd = DPCPU_SET(randomval, 641 DPCPU_GET(randomval) * 69069 + 5) >> 26; 642 if (match & CPU_SEARCH_LOWEST) { 643 if (cpu == low->cs_prefer) 644 load -= 64; 645 /* If that CPU is allowed and get data. */ 646 if (CPU_ISSET(cpu, &lgroup.cs_mask) && 647 tdq->tdq_lowpri > lgroup.cs_pri && 648 tdq->tdq_load <= lgroup.cs_limit) { 649 lgroup.cs_cpu = cpu; 650 lgroup.cs_load = load - rnd; 651 } 652 } 653 if (match & CPU_SEARCH_HIGHEST) 654 if (CPU_ISSET(cpu, &hgroup.cs_mask) && 655 tdq->tdq_load >= hgroup.cs_limit && 656 tdq->tdq_transferable) { 657 hgroup.cs_cpu = cpu; 658 hgroup.cs_load = load - rnd; 659 } 660 } 661 total += load; 662 663 /* We have info about child item. Compare it. */ 664 if (match & CPU_SEARCH_LOWEST) { 665 if (lgroup.cs_load != INT_MAX && 666 (load < lload || 667 (load == lload && lgroup.cs_load < low->cs_load))) { 668 lload = load; 669 low->cs_cpu = lgroup.cs_cpu; 670 low->cs_load = lgroup.cs_load; 671 } 672 } 673 if (match & CPU_SEARCH_HIGHEST) 674 if (hgroup.cs_load != -1 && 675 (load > hload || 676 (load == hload && hgroup.cs_load > high->cs_load))) { 677 hload = load; 678 high->cs_cpu = hgroup.cs_cpu; 679 high->cs_load = hgroup.cs_load; 680 } 681 if (child) 682 i++; 683 else 684 cpu++; 685 } 686 return (total); 687} 688 689/* 690 * cpu_search instantiations must pass constants to maintain the inline 691 * optimization. 692 */ 693int 694cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low) 695{ 696 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 697} 698 699int 700cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high) 701{ 702 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 703} 704 705int 706cpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 707 struct cpu_search *high) 708{ 709 return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 710} 711 712/* 713 * Find the cpu with the least load via the least loaded path that has a 714 * lowpri greater than pri pri. A pri of -1 indicates any priority is 715 * acceptable. 716 */ 717static inline int 718sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload, 719 int prefer) 720{ 721 struct cpu_search low; 722 723 low.cs_cpu = -1; 724 low.cs_prefer = prefer; 725 low.cs_mask = mask; 726 low.cs_pri = pri; 727 low.cs_limit = maxload; 728 cpu_search_lowest(cg, &low); 729 return low.cs_cpu; 730} 731 732/* 733 * Find the cpu with the highest load via the highest loaded path. 734 */ 735static inline int 736sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload) 737{ 738 struct cpu_search high; 739 740 high.cs_cpu = -1; 741 high.cs_mask = mask; 742 high.cs_limit = minload; 743 cpu_search_highest(cg, &high); 744 return high.cs_cpu; 745} 746 747/* 748 * Simultaneously find the highest and lowest loaded cpu reachable via 749 * cg. 750 */ 751static inline void 752sched_both(const struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu) 753{ 754 struct cpu_search high; 755 struct cpu_search low; 756 757 low.cs_cpu = -1; 758 low.cs_prefer = -1; 759 low.cs_pri = -1; 760 low.cs_limit = INT_MAX; 761 low.cs_mask = mask; 762 high.cs_cpu = -1; 763 high.cs_limit = -1; 764 high.cs_mask = mask; 765 cpu_search_both(cg, &low, &high); 766 *lowcpu = low.cs_cpu; 767 *highcpu = high.cs_cpu; 768 return; 769} 770 771static void 772sched_balance_group(struct cpu_group *cg) 773{ 774 cpuset_t hmask, lmask; 775 int high, low, anylow; 776 777 CPU_FILL(&hmask); 778 for (;;) { 779 high = sched_highest(cg, hmask, 1); 780 /* Stop if there is no more CPU with transferrable threads. */ 781 if (high == -1) 782 break; 783 CPU_CLR(high, &hmask); 784 CPU_COPY(&hmask, &lmask); 785 /* Stop if there is no more CPU left for low. */ 786 if (CPU_EMPTY(&lmask)) 787 break; 788 anylow = 1; 789nextlow: 790 low = sched_lowest(cg, lmask, -1, 791 TDQ_CPU(high)->tdq_load - 1, high); 792 /* Stop if we looked well and found no less loaded CPU. */ 793 if (anylow && low == -1) 794 break; 795 /* Go to next high if we found no less loaded CPU. */ 796 if (low == -1) 797 continue; 798 /* Transfer thread from high to low. */ 799 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) { 800 /* CPU that got thread can no longer be a donor. */ 801 CPU_CLR(low, &hmask); 802 } else { 803 /* 804 * If failed, then there is no threads on high 805 * that can run on this low. Drop low from low 806 * mask and look for different one. 807 */ 808 CPU_CLR(low, &lmask); 809 anylow = 0; 810 goto nextlow; 811 } 812 } 813} 814 815static void 816sched_balance(void) 817{ 818 struct tdq *tdq; 819 820 /* 821 * Select a random time between .5 * balance_interval and 822 * 1.5 * balance_interval. 823 */ 824 balance_ticks = max(balance_interval / 2, 1); 825 balance_ticks += random() % balance_interval; 826 if (smp_started == 0 || rebalance == 0) 827 return; 828 tdq = TDQ_SELF(); 829 TDQ_UNLOCK(tdq); 830 sched_balance_group(cpu_top); 831 TDQ_LOCK(tdq); 832} 833 834/* 835 * Lock two thread queues using their address to maintain lock order. 836 */ 837static void 838tdq_lock_pair(struct tdq *one, struct tdq *two) 839{ 840 if (one < two) { 841 TDQ_LOCK(one); 842 TDQ_LOCK_FLAGS(two, MTX_DUPOK); 843 } else { 844 TDQ_LOCK(two); 845 TDQ_LOCK_FLAGS(one, MTX_DUPOK); 846 } 847} 848 849/* 850 * Unlock two thread queues. Order is not important here. 851 */ 852static void 853tdq_unlock_pair(struct tdq *one, struct tdq *two) 854{ 855 TDQ_UNLOCK(one); 856 TDQ_UNLOCK(two); 857} 858 859/* 860 * Transfer load between two imbalanced thread queues. 861 */ 862static int 863sched_balance_pair(struct tdq *high, struct tdq *low) 864{ 865 int moved; 866 int cpu; 867 868 tdq_lock_pair(high, low); 869 moved = 0; 870 /* 871 * Determine what the imbalance is and then adjust that to how many 872 * threads we actually have to give up (transferable). 873 */ 874 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load && 875 (moved = tdq_move(high, low)) > 0) { 876 /* 877 * In case the target isn't the current cpu IPI it to force a 878 * reschedule with the new workload. 879 */ 880 cpu = TDQ_ID(low); 881 sched_pin(); 882 if (cpu != PCPU_GET(cpuid)) 883 ipi_cpu(cpu, IPI_PREEMPT); 884 sched_unpin(); 885 } 886 tdq_unlock_pair(high, low); 887 return (moved); 888} 889 890/* 891 * Move a thread from one thread queue to another. 892 */ 893static int 894tdq_move(struct tdq *from, struct tdq *to) 895{ 896 struct td_sched *ts; 897 struct thread *td; 898 struct tdq *tdq; 899 int cpu; 900 901 TDQ_LOCK_ASSERT(from, MA_OWNED); 902 TDQ_LOCK_ASSERT(to, MA_OWNED); 903 904 tdq = from; 905 cpu = TDQ_ID(to); 906 td = tdq_steal(tdq, cpu); 907 if (td == NULL) 908 return (0); 909 ts = td->td_sched; 910 /* 911 * Although the run queue is locked the thread may be blocked. Lock 912 * it to clear this and acquire the run-queue lock. 913 */ 914 thread_lock(td); 915 /* Drop recursive lock on from acquired via thread_lock(). */ 916 TDQ_UNLOCK(from); 917 sched_rem(td); 918 ts->ts_cpu = cpu; 919 td->td_lock = TDQ_LOCKPTR(to); 920 tdq_add(to, td, SRQ_YIELDING); 921 return (1); 922} 923 924/* 925 * This tdq has idled. Try to steal a thread from another cpu and switch 926 * to it. 927 */ 928static int 929tdq_idled(struct tdq *tdq) 930{ 931 struct cpu_group *cg; 932 struct tdq *steal; 933 cpuset_t mask; 934 int thresh; 935 int cpu; 936 937 if (smp_started == 0 || steal_idle == 0) 938 return (1); 939 CPU_FILL(&mask); 940 CPU_CLR(PCPU_GET(cpuid), &mask); 941 /* We don't want to be preempted while we're iterating. */ 942 spinlock_enter(); 943 for (cg = tdq->tdq_cg; cg != NULL; ) { 944 if ((cg->cg_flags & CG_FLAG_THREAD) == 0) 945 thresh = steal_thresh; 946 else 947 thresh = 1; 948 cpu = sched_highest(cg, mask, thresh); 949 if (cpu == -1) { 950 cg = cg->cg_parent; 951 continue; 952 } 953 steal = TDQ_CPU(cpu); 954 CPU_CLR(cpu, &mask); 955 tdq_lock_pair(tdq, steal); 956 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 957 tdq_unlock_pair(tdq, steal); 958 continue; 959 } 960 /* 961 * If a thread was added while interrupts were disabled don't 962 * steal one here. If we fail to acquire one due to affinity 963 * restrictions loop again with this cpu removed from the 964 * set. 965 */ 966 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 967 tdq_unlock_pair(tdq, steal); 968 continue; 969 } 970 spinlock_exit(); 971 TDQ_UNLOCK(steal); 972 mi_switch(SW_VOL | SWT_IDLE, NULL); 973 thread_unlock(curthread); 974 975 return (0); 976 } 977 spinlock_exit(); 978 return (1); 979} 980 981/* 982 * Notify a remote cpu of new work. Sends an IPI if criteria are met. 983 */ 984static void 985tdq_notify(struct tdq *tdq, struct thread *td) 986{ 987 struct thread *ctd; 988 int pri; 989 int cpu; 990 991 if (tdq->tdq_ipipending) 992 return; 993 cpu = td->td_sched->ts_cpu; 994 pri = td->td_priority; 995 ctd = pcpu_find(cpu)->pc_curthread; 996 if (!sched_shouldpreempt(pri, ctd->td_priority, 1)) 997 return; 998 if (TD_IS_IDLETHREAD(ctd)) { 999 /* 1000 * If the MD code has an idle wakeup routine try that before 1001 * falling back to IPI. 1002 */ 1003 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu)) 1004 return; 1005 } 1006 tdq->tdq_ipipending = 1; 1007 ipi_cpu(cpu, IPI_PREEMPT); 1008} 1009 1010/* 1011 * Steals load from a timeshare queue. Honors the rotating queue head 1012 * index. 1013 */ 1014static struct thread * 1015runq_steal_from(struct runq *rq, int cpu, u_char start) 1016{ 1017 struct rqbits *rqb; 1018 struct rqhead *rqh; 1019 struct thread *td, *first; 1020 int bit; 1021 int pri; 1022 int i; 1023 1024 rqb = &rq->rq_status; 1025 bit = start & (RQB_BPW -1); 1026 pri = 0; 1027 first = NULL; 1028again: 1029 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 1030 if (rqb->rqb_bits[i] == 0) 1031 continue; 1032 if (bit != 0) { 1033 for (pri = bit; pri < RQB_BPW; pri++) 1034 if (rqb->rqb_bits[i] & (1ul << pri)) 1035 break; 1036 if (pri >= RQB_BPW) 1037 continue; 1038 } else 1039 pri = RQB_FFS(rqb->rqb_bits[i]); 1040 pri += (i << RQB_L2BPW); 1041 rqh = &rq->rq_queues[pri]; 1042 TAILQ_FOREACH(td, rqh, td_runq) { 1043 if (first && THREAD_CAN_MIGRATE(td) && 1044 THREAD_CAN_SCHED(td, cpu)) 1045 return (td); 1046 first = td; 1047 } 1048 } 1049 if (start != 0) { 1050 start = 0; 1051 goto again; 1052 } 1053 1054 if (first && THREAD_CAN_MIGRATE(first) && 1055 THREAD_CAN_SCHED(first, cpu)) 1056 return (first); 1057 return (NULL); 1058} 1059 1060/* 1061 * Steals load from a standard linear queue. 1062 */ 1063static struct thread * 1064runq_steal(struct runq *rq, int cpu) 1065{ 1066 struct rqhead *rqh; 1067 struct rqbits *rqb; 1068 struct thread *td; 1069 int word; 1070 int bit; 1071 1072 rqb = &rq->rq_status; 1073 for (word = 0; word < RQB_LEN; word++) { 1074 if (rqb->rqb_bits[word] == 0) 1075 continue; 1076 for (bit = 0; bit < RQB_BPW; bit++) { 1077 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1078 continue; 1079 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1080 TAILQ_FOREACH(td, rqh, td_runq) 1081 if (THREAD_CAN_MIGRATE(td) && 1082 THREAD_CAN_SCHED(td, cpu)) 1083 return (td); 1084 } 1085 } 1086 return (NULL); 1087} 1088 1089/* 1090 * Attempt to steal a thread in priority order from a thread queue. 1091 */ 1092static struct thread * 1093tdq_steal(struct tdq *tdq, int cpu) 1094{ 1095 struct thread *td; 1096 1097 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1098 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1099 return (td); 1100 if ((td = runq_steal_from(&tdq->tdq_timeshare, 1101 cpu, tdq->tdq_ridx)) != NULL) 1102 return (td); 1103 return (runq_steal(&tdq->tdq_idle, cpu)); 1104} 1105 1106/* 1107 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1108 * current lock and returns with the assigned queue locked. 1109 */ 1110static inline struct tdq * 1111sched_setcpu(struct thread *td, int cpu, int flags) 1112{ 1113 1114 struct tdq *tdq; 1115 1116 THREAD_LOCK_ASSERT(td, MA_OWNED); 1117 tdq = TDQ_CPU(cpu); 1118 td->td_sched->ts_cpu = cpu; 1119 /* 1120 * If the lock matches just return the queue. 1121 */ 1122 if (td->td_lock == TDQ_LOCKPTR(tdq)) 1123 return (tdq); 1124#ifdef notyet 1125 /* 1126 * If the thread isn't running its lockptr is a 1127 * turnstile or a sleepqueue. We can just lock_set without 1128 * blocking. 1129 */ 1130 if (TD_CAN_RUN(td)) { 1131 TDQ_LOCK(tdq); 1132 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1133 return (tdq); 1134 } 1135#endif 1136 /* 1137 * The hard case, migration, we need to block the thread first to 1138 * prevent order reversals with other cpus locks. 1139 */ 1140 spinlock_enter(); 1141 thread_lock_block(td); 1142 TDQ_LOCK(tdq); 1143 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1144 spinlock_exit(); 1145 return (tdq); 1146} 1147 1148SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1149SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1150SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1151SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1152SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1153SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1154 1155static int 1156sched_pickcpu(struct thread *td, int flags) 1157{ 1158 struct cpu_group *cg, *ccg; 1159 struct td_sched *ts; 1160 struct tdq *tdq; 1161 cpuset_t mask; 1162 int cpu, pri, self; 1163 1164 self = PCPU_GET(cpuid); 1165 ts = td->td_sched; 1166 if (smp_started == 0) 1167 return (self); 1168 /* 1169 * Don't migrate a running thread from sched_switch(). 1170 */ 1171 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1172 return (ts->ts_cpu); 1173 /* 1174 * Prefer to run interrupt threads on the processors that generate 1175 * the interrupt. 1176 */ 1177 pri = td->td_priority; 1178 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1179 curthread->td_intr_nesting_level && ts->ts_cpu != self) { 1180 SCHED_STAT_INC(pickcpu_intrbind); 1181 ts->ts_cpu = self; 1182 if (TDQ_CPU(self)->tdq_lowpri > pri) { 1183 SCHED_STAT_INC(pickcpu_affinity); 1184 return (ts->ts_cpu); 1185 } 1186 } 1187 /* 1188 * If the thread can run on the last cpu and the affinity has not 1189 * expired or it is idle run it there. 1190 */ 1191 tdq = TDQ_CPU(ts->ts_cpu); 1192 cg = tdq->tdq_cg; 1193 if (THREAD_CAN_SCHED(td, ts->ts_cpu) && 1194 tdq->tdq_lowpri >= PRI_MIN_IDLE && 1195 SCHED_AFFINITY(ts, CG_SHARE_L2)) { 1196 if (cg->cg_flags & CG_FLAG_THREAD) { 1197 CPUSET_FOREACH(cpu, cg->cg_mask) { 1198 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) 1199 break; 1200 } 1201 } else 1202 cpu = INT_MAX; 1203 if (cpu > mp_maxid) { 1204 SCHED_STAT_INC(pickcpu_idle_affinity); 1205 return (ts->ts_cpu); 1206 } 1207 } 1208 /* 1209 * Search for the last level cache CPU group in the tree. 1210 * Skip caches with expired affinity time and SMT groups. 1211 * Affinity to higher level caches will be handled less aggressively. 1212 */ 1213 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) { 1214 if (cg->cg_flags & CG_FLAG_THREAD) 1215 continue; 1216 if (!SCHED_AFFINITY(ts, cg->cg_level)) 1217 continue; 1218 ccg = cg; 1219 } 1220 if (ccg != NULL) 1221 cg = ccg; 1222 cpu = -1; 1223 /* Search the group for the less loaded idle CPU we can run now. */ 1224 mask = td->td_cpuset->cs_mask; 1225 if (cg != NULL && cg != cpu_top && 1226 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0) 1227 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE), 1228 INT_MAX, ts->ts_cpu); 1229 /* Search globally for the less loaded CPU we can run now. */ 1230 if (cpu == -1) 1231 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu); 1232 /* Search globally for the less loaded CPU. */ 1233 if (cpu == -1) 1234 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu); 1235 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1236 /* 1237 * Compare the lowest loaded cpu to current cpu. 1238 */ 1239 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1240 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE && 1241 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) { 1242 SCHED_STAT_INC(pickcpu_local); 1243 cpu = self; 1244 } else 1245 SCHED_STAT_INC(pickcpu_lowest); 1246 if (cpu != ts->ts_cpu) 1247 SCHED_STAT_INC(pickcpu_migration); 1248 return (cpu); 1249} 1250#endif 1251 1252/* 1253 * Pick the highest priority task we have and return it. 1254 */ 1255static struct thread * 1256tdq_choose(struct tdq *tdq) 1257{ 1258 struct thread *td; 1259 1260 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1261 td = runq_choose(&tdq->tdq_realtime); 1262 if (td != NULL) 1263 return (td); 1264 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1265 if (td != NULL) { 1266 KASSERT(td->td_priority >= PRI_MIN_BATCH, 1267 ("tdq_choose: Invalid priority on timeshare queue %d", 1268 td->td_priority)); 1269 return (td); 1270 } 1271 td = runq_choose(&tdq->tdq_idle); 1272 if (td != NULL) { 1273 KASSERT(td->td_priority >= PRI_MIN_IDLE, 1274 ("tdq_choose: Invalid priority on idle queue %d", 1275 td->td_priority)); 1276 return (td); 1277 } 1278 1279 return (NULL); 1280} 1281 1282/* 1283 * Initialize a thread queue. 1284 */ 1285static void 1286tdq_setup(struct tdq *tdq) 1287{ 1288 1289 if (bootverbose) 1290 printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1291 runq_init(&tdq->tdq_realtime); 1292 runq_init(&tdq->tdq_timeshare); 1293 runq_init(&tdq->tdq_idle); 1294 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1295 "sched lock %d", (int)TDQ_ID(tdq)); 1296 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1297 MTX_SPIN | MTX_RECURSE); 1298#ifdef KTR 1299 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname), 1300 "CPU %d load", (int)TDQ_ID(tdq)); 1301#endif 1302} 1303 1304#ifdef SMP 1305static void 1306sched_setup_smp(void) 1307{ 1308 struct tdq *tdq; 1309 int i; 1310 1311 cpu_top = smp_topo(); 1312 CPU_FOREACH(i) { 1313 tdq = TDQ_CPU(i); 1314 tdq_setup(tdq); 1315 tdq->tdq_cg = smp_topo_find(cpu_top, i); 1316 if (tdq->tdq_cg == NULL) 1317 panic("Can't find cpu group for %d\n", i); 1318 } 1319 balance_tdq = TDQ_SELF(); 1320 sched_balance(); 1321} 1322#endif 1323 1324/* 1325 * Setup the thread queues and initialize the topology based on MD 1326 * information. 1327 */ 1328static void 1329sched_setup(void *dummy) 1330{ 1331 struct tdq *tdq; 1332 1333 tdq = TDQ_SELF(); 1334#ifdef SMP 1335 sched_setup_smp(); 1336#else 1337 tdq_setup(tdq); 1338#endif 1339 /* 1340 * To avoid divide-by-zero, we set realstathz a dummy value 1341 * in case which sched_clock() called before sched_initticks(). 1342 */ 1343 realstathz = hz; 1344 sched_slice = (realstathz/10); /* ~100ms */ 1345 tickincr = 1 << SCHED_TICK_SHIFT; 1346 1347 /* Add thread0's load since it's running. */ 1348 TDQ_LOCK(tdq); 1349 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1350 tdq_load_add(tdq, &thread0); 1351 tdq->tdq_lowpri = thread0.td_priority; 1352 TDQ_UNLOCK(tdq); 1353} 1354 1355/* 1356 * This routine determines the tickincr after stathz and hz are setup. 1357 */ 1358/* ARGSUSED */ 1359static void 1360sched_initticks(void *dummy) 1361{ 1362 int incr; 1363 1364 realstathz = stathz ? stathz : hz; 1365 sched_slice = (realstathz/10); /* ~100ms */ 1366 1367 /* 1368 * tickincr is shifted out by 10 to avoid rounding errors due to 1369 * hz not being evenly divisible by stathz on all platforms. 1370 */ 1371 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1372 /* 1373 * This does not work for values of stathz that are more than 1374 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1375 */ 1376 if (incr == 0) 1377 incr = 1; 1378 tickincr = incr; 1379#ifdef SMP 1380 /* 1381 * Set the default balance interval now that we know 1382 * what realstathz is. 1383 */ 1384 balance_interval = realstathz; 1385 /* 1386 * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4. 1387 * This prevents excess thrashing on large machines and excess idle 1388 * on smaller machines. 1389 */ 1390 steal_thresh = min(fls(mp_ncpus) - 1, 3); 1391 affinity = SCHED_AFFINITY_DEFAULT; 1392#endif 1393} 1394 1395 1396/* 1397 * This is the core of the interactivity algorithm. Determines a score based 1398 * on past behavior. It is the ratio of sleep time to run time scaled to 1399 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1400 * differs from the cpu usage because it does not account for time spent 1401 * waiting on a run-queue. Would be prettier if we had floating point. 1402 */ 1403static int 1404sched_interact_score(struct thread *td) 1405{ 1406 struct td_sched *ts; 1407 int div; 1408 1409 ts = td->td_sched; 1410 /* 1411 * The score is only needed if this is likely to be an interactive 1412 * task. Don't go through the expense of computing it if there's 1413 * no chance. 1414 */ 1415 if (sched_interact <= SCHED_INTERACT_HALF && 1416 ts->ts_runtime >= ts->ts_slptime) 1417 return (SCHED_INTERACT_HALF); 1418 1419 if (ts->ts_runtime > ts->ts_slptime) { 1420 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1421 return (SCHED_INTERACT_HALF + 1422 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1423 } 1424 if (ts->ts_slptime > ts->ts_runtime) { 1425 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1426 return (ts->ts_runtime / div); 1427 } 1428 /* runtime == slptime */ 1429 if (ts->ts_runtime) 1430 return (SCHED_INTERACT_HALF); 1431 1432 /* 1433 * This can happen if slptime and runtime are 0. 1434 */ 1435 return (0); 1436 1437} 1438 1439/* 1440 * Scale the scheduling priority according to the "interactivity" of this 1441 * process. 1442 */ 1443static void 1444sched_priority(struct thread *td) 1445{ 1446 int score; 1447 int pri; 1448 1449 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1450 return; 1451 /* 1452 * If the score is interactive we place the thread in the realtime 1453 * queue with a priority that is less than kernel and interrupt 1454 * priorities. These threads are not subject to nice restrictions. 1455 * 1456 * Scores greater than this are placed on the normal timeshare queue 1457 * where the priority is partially decided by the most recent cpu 1458 * utilization and the rest is decided by nice value. 1459 * 1460 * The nice value of the process has a linear effect on the calculated 1461 * score. Negative nice values make it easier for a thread to be 1462 * considered interactive. 1463 */ 1464 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice); 1465 if (score < sched_interact) { 1466 pri = PRI_MIN_INTERACT; 1467 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) / 1468 sched_interact) * score; 1469 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT, 1470 ("sched_priority: invalid interactive priority %d score %d", 1471 pri, score)); 1472 } else { 1473 pri = SCHED_PRI_MIN; 1474 if (td->td_sched->ts_ticks) 1475 pri += min(SCHED_PRI_TICKS(td->td_sched), 1476 SCHED_PRI_RANGE); 1477 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1478 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH, 1479 ("sched_priority: invalid priority %d: nice %d, " 1480 "ticks %d ftick %d ltick %d tick pri %d", 1481 pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1482 td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1483 SCHED_PRI_TICKS(td->td_sched))); 1484 } 1485 sched_user_prio(td, pri); 1486 1487 return; 1488} 1489 1490/* 1491 * This routine enforces a maximum limit on the amount of scheduling history 1492 * kept. It is called after either the slptime or runtime is adjusted. This 1493 * function is ugly due to integer math. 1494 */ 1495static void 1496sched_interact_update(struct thread *td) 1497{ 1498 struct td_sched *ts; 1499 u_int sum; 1500 1501 ts = td->td_sched; 1502 sum = ts->ts_runtime + ts->ts_slptime; 1503 if (sum < SCHED_SLP_RUN_MAX) 1504 return; 1505 /* 1506 * This only happens from two places: 1507 * 1) We have added an unusual amount of run time from fork_exit. 1508 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1509 */ 1510 if (sum > SCHED_SLP_RUN_MAX * 2) { 1511 if (ts->ts_runtime > ts->ts_slptime) { 1512 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1513 ts->ts_slptime = 1; 1514 } else { 1515 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1516 ts->ts_runtime = 1; 1517 } 1518 return; 1519 } 1520 /* 1521 * If we have exceeded by more than 1/5th then the algorithm below 1522 * will not bring us back into range. Dividing by two here forces 1523 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1524 */ 1525 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1526 ts->ts_runtime /= 2; 1527 ts->ts_slptime /= 2; 1528 return; 1529 } 1530 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1531 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1532} 1533 1534/* 1535 * Scale back the interactivity history when a child thread is created. The 1536 * history is inherited from the parent but the thread may behave totally 1537 * differently. For example, a shell spawning a compiler process. We want 1538 * to learn that the compiler is behaving badly very quickly. 1539 */ 1540static void 1541sched_interact_fork(struct thread *td) 1542{ 1543 int ratio; 1544 int sum; 1545 1546 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1547 if (sum > SCHED_SLP_RUN_FORK) { 1548 ratio = sum / SCHED_SLP_RUN_FORK; 1549 td->td_sched->ts_runtime /= ratio; 1550 td->td_sched->ts_slptime /= ratio; 1551 } 1552} 1553 1554/* 1555 * Called from proc0_init() to setup the scheduler fields. 1556 */ 1557void 1558schedinit(void) 1559{ 1560 1561 /* 1562 * Set up the scheduler specific parts of proc0. 1563 */ 1564 proc0.p_sched = NULL; /* XXX */ 1565 thread0.td_sched = &td_sched0; 1566 td_sched0.ts_ltick = ticks; 1567 td_sched0.ts_ftick = ticks; 1568 td_sched0.ts_slice = sched_slice; 1569} 1570 1571/* 1572 * This is only somewhat accurate since given many processes of the same 1573 * priority they will switch when their slices run out, which will be 1574 * at most sched_slice stathz ticks. 1575 */ 1576int 1577sched_rr_interval(void) 1578{ 1579 1580 /* Convert sched_slice to hz */ 1581 return (hz/(realstathz/sched_slice)); 1582} 1583 1584/* 1585 * Update the percent cpu tracking information when it is requested or 1586 * the total history exceeds the maximum. We keep a sliding history of 1587 * tick counts that slowly decays. This is less precise than the 4BSD 1588 * mechanism since it happens with less regular and frequent events. 1589 */ 1590static void 1591sched_pctcpu_update(struct td_sched *ts) 1592{ 1593 1594 if (ts->ts_ticks == 0) 1595 return; 1596 if (ticks - (hz / 10) < ts->ts_ltick && 1597 SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1598 return; 1599 /* 1600 * Adjust counters and watermark for pctcpu calc. 1601 */ 1602 if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1603 ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1604 SCHED_TICK_TARG; 1605 else 1606 ts->ts_ticks = 0; 1607 ts->ts_ltick = ticks; 1608 ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1609} 1610 1611/* 1612 * Adjust the priority of a thread. Move it to the appropriate run-queue 1613 * if necessary. This is the back-end for several priority related 1614 * functions. 1615 */ 1616static void 1617sched_thread_priority(struct thread *td, u_char prio) 1618{ 1619 struct td_sched *ts; 1620 struct tdq *tdq; 1621 int oldpri; 1622 1623 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio", 1624 "prio:%d", td->td_priority, "new prio:%d", prio, 1625 KTR_ATTR_LINKED, sched_tdname(curthread)); 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 } 1631 ts = td->td_sched; 1632 THREAD_LOCK_ASSERT(td, MA_OWNED); 1633 if (td->td_priority == prio) 1634 return; 1635 /* 1636 * If the priority has been elevated due to priority 1637 * propagation, we may have to move ourselves to a new 1638 * queue. This could be optimized to not re-add in some 1639 * cases. 1640 */ 1641 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1642 sched_rem(td); 1643 td->td_priority = prio; 1644 sched_add(td, SRQ_BORROWING); 1645 return; 1646 } 1647 /* 1648 * If the thread is currently running we may have to adjust the lowpri 1649 * information so other cpus are aware of our current priority. 1650 */ 1651 if (TD_IS_RUNNING(td)) { 1652 tdq = TDQ_CPU(ts->ts_cpu); 1653 oldpri = td->td_priority; 1654 td->td_priority = prio; 1655 if (prio < tdq->tdq_lowpri) 1656 tdq->tdq_lowpri = prio; 1657 else if (tdq->tdq_lowpri == oldpri) 1658 tdq_setlowpri(tdq, td); 1659 return; 1660 } 1661 td->td_priority = prio; 1662} 1663 1664/* 1665 * Update a thread's priority when it is lent another thread's 1666 * priority. 1667 */ 1668void 1669sched_lend_prio(struct thread *td, u_char prio) 1670{ 1671 1672 td->td_flags |= TDF_BORROWING; 1673 sched_thread_priority(td, prio); 1674} 1675 1676/* 1677 * Restore a thread's priority when priority propagation is 1678 * over. The prio argument is the minimum priority the thread 1679 * needs to have to satisfy other possible priority lending 1680 * requests. If the thread's regular priority is less 1681 * important than prio, the thread will keep a priority boost 1682 * of prio. 1683 */ 1684void 1685sched_unlend_prio(struct thread *td, u_char prio) 1686{ 1687 u_char base_pri; 1688 1689 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1690 td->td_base_pri <= PRI_MAX_TIMESHARE) 1691 base_pri = td->td_user_pri; 1692 else 1693 base_pri = td->td_base_pri; 1694 if (prio >= base_pri) { 1695 td->td_flags &= ~TDF_BORROWING; 1696 sched_thread_priority(td, base_pri); 1697 } else 1698 sched_lend_prio(td, prio); 1699} 1700 1701/* 1702 * Standard entry for setting the priority to an absolute value. 1703 */ 1704void 1705sched_prio(struct thread *td, u_char prio) 1706{ 1707 u_char oldprio; 1708 1709 /* First, update the base priority. */ 1710 td->td_base_pri = prio; 1711 1712 /* 1713 * If the thread is borrowing another thread's priority, don't 1714 * ever lower the priority. 1715 */ 1716 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1717 return; 1718 1719 /* Change the real priority. */ 1720 oldprio = td->td_priority; 1721 sched_thread_priority(td, prio); 1722 1723 /* 1724 * If the thread is on a turnstile, then let the turnstile update 1725 * its state. 1726 */ 1727 if (TD_ON_LOCK(td) && oldprio != prio) 1728 turnstile_adjust(td, oldprio); 1729} 1730 1731/* 1732 * Set the base user priority, does not effect current running priority. 1733 */ 1734void 1735sched_user_prio(struct thread *td, u_char prio) 1736{ 1737 1738 td->td_base_user_pri = prio; 1739 if (td->td_lend_user_pri <= prio) 1740 return; 1741 td->td_user_pri = prio; 1742} 1743 1744void 1745sched_lend_user_prio(struct thread *td, u_char prio) 1746{ 1747 1748 THREAD_LOCK_ASSERT(td, MA_OWNED); 1749 td->td_lend_user_pri = prio; 1750 td->td_user_pri = min(prio, td->td_base_user_pri); 1751 if (td->td_priority > td->td_user_pri) 1752 sched_prio(td, td->td_user_pri); 1753 else if (td->td_priority != td->td_user_pri) 1754 td->td_flags |= TDF_NEEDRESCHED; 1755} 1756 1757/* 1758 * Handle migration from sched_switch(). This happens only for 1759 * cpu binding. 1760 */ 1761static struct mtx * 1762sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1763{ 1764 struct tdq *tdn; 1765 1766 tdn = TDQ_CPU(td->td_sched->ts_cpu); 1767#ifdef SMP 1768 tdq_load_rem(tdq, td); 1769 /* 1770 * Do the lock dance required to avoid LOR. We grab an extra 1771 * spinlock nesting to prevent preemption while we're 1772 * not holding either run-queue lock. 1773 */ 1774 spinlock_enter(); 1775 thread_lock_block(td); /* This releases the lock on tdq. */ 1776 1777 /* 1778 * Acquire both run-queue locks before placing the thread on the new 1779 * run-queue to avoid deadlocks created by placing a thread with a 1780 * blocked lock on the run-queue of a remote processor. The deadlock 1781 * occurs when a third processor attempts to lock the two queues in 1782 * question while the target processor is spinning with its own 1783 * run-queue lock held while waiting for the blocked lock to clear. 1784 */ 1785 tdq_lock_pair(tdn, tdq); 1786 tdq_add(tdn, td, flags); 1787 tdq_notify(tdn, td); 1788 TDQ_UNLOCK(tdn); 1789 spinlock_exit(); 1790#endif 1791 return (TDQ_LOCKPTR(tdn)); 1792} 1793 1794/* 1795 * Variadic version of thread_lock_unblock() that does not assume td_lock 1796 * is blocked. 1797 */ 1798static inline void 1799thread_unblock_switch(struct thread *td, struct mtx *mtx) 1800{ 1801 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1802 (uintptr_t)mtx); 1803} 1804 1805/* 1806 * Switch threads. This function has to handle threads coming in while 1807 * blocked for some reason, running, or idle. It also must deal with 1808 * migrating a thread from one queue to another as running threads may 1809 * be assigned elsewhere via binding. 1810 */ 1811void 1812sched_switch(struct thread *td, struct thread *newtd, int flags) 1813{ 1814 struct tdq *tdq; 1815 struct td_sched *ts; 1816 struct mtx *mtx; 1817 int srqflag; 1818 int cpuid; 1819 1820 THREAD_LOCK_ASSERT(td, MA_OWNED); 1821 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1822 1823 cpuid = PCPU_GET(cpuid); 1824 tdq = TDQ_CPU(cpuid); 1825 ts = td->td_sched; 1826 mtx = td->td_lock; 1827 ts->ts_rltick = ticks; 1828 td->td_lastcpu = td->td_oncpu; 1829 td->td_oncpu = NOCPU; 1830 if (!(flags & SW_PREEMPT)) 1831 td->td_flags &= ~TDF_NEEDRESCHED; 1832 td->td_owepreempt = 0; 1833 tdq->tdq_switchcnt++; 1834 /* 1835 * The lock pointer in an idle thread should never change. Reset it 1836 * to CAN_RUN as well. 1837 */ 1838 if (TD_IS_IDLETHREAD(td)) { 1839 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1840 TD_SET_CAN_RUN(td); 1841 } else if (TD_IS_RUNNING(td)) { 1842 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1843 srqflag = (flags & SW_PREEMPT) ? 1844 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1845 SRQ_OURSELF|SRQ_YIELDING; 1846#ifdef SMP 1847 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu)) 1848 ts->ts_cpu = sched_pickcpu(td, 0); 1849#endif 1850 if (ts->ts_cpu == cpuid) 1851 tdq_runq_add(tdq, td, srqflag); 1852 else { 1853 KASSERT(THREAD_CAN_MIGRATE(td) || 1854 (ts->ts_flags & TSF_BOUND) != 0, 1855 ("Thread %p shouldn't migrate", td)); 1856 mtx = sched_switch_migrate(tdq, td, srqflag); 1857 } 1858 } else { 1859 /* This thread must be going to sleep. */ 1860 TDQ_LOCK(tdq); 1861 mtx = thread_lock_block(td); 1862 tdq_load_rem(tdq, td); 1863 } 1864 /* 1865 * We enter here with the thread blocked and assigned to the 1866 * appropriate cpu run-queue or sleep-queue and with the current 1867 * thread-queue locked. 1868 */ 1869 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1870 newtd = choosethread(); 1871 /* 1872 * Call the MD code to switch contexts if necessary. 1873 */ 1874 if (td != newtd) { 1875#ifdef HWPMC_HOOKS 1876 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1877 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1878#endif 1879 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1880 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1881 1882#ifdef KDTRACE_HOOKS 1883 /* 1884 * If DTrace has set the active vtime enum to anything 1885 * other than INACTIVE (0), then it should have set the 1886 * function to call. 1887 */ 1888 if (dtrace_vtime_active) 1889 (*dtrace_vtime_switch_func)(newtd); 1890#endif 1891 1892 cpu_switch(td, newtd, mtx); 1893 /* 1894 * We may return from cpu_switch on a different cpu. However, 1895 * we always return with td_lock pointing to the current cpu's 1896 * run queue lock. 1897 */ 1898 cpuid = PCPU_GET(cpuid); 1899 tdq = TDQ_CPU(cpuid); 1900 lock_profile_obtain_lock_success( 1901 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1902#ifdef HWPMC_HOOKS 1903 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1904 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1905#endif 1906 } else 1907 thread_unblock_switch(td, mtx); 1908 /* 1909 * Assert that all went well and return. 1910 */ 1911 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1912 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1913 td->td_oncpu = cpuid; 1914} 1915 1916/* 1917 * Adjust thread priorities as a result of a nice request. 1918 */ 1919void 1920sched_nice(struct proc *p, int nice) 1921{ 1922 struct thread *td; 1923 1924 PROC_LOCK_ASSERT(p, MA_OWNED); 1925 1926 p->p_nice = nice; 1927 FOREACH_THREAD_IN_PROC(p, td) { 1928 thread_lock(td); 1929 sched_priority(td); 1930 sched_prio(td, td->td_base_user_pri); 1931 thread_unlock(td); 1932 } 1933} 1934 1935/* 1936 * Record the sleep time for the interactivity scorer. 1937 */ 1938void 1939sched_sleep(struct thread *td, int prio) 1940{ 1941 1942 THREAD_LOCK_ASSERT(td, MA_OWNED); 1943 1944 td->td_slptick = ticks; 1945 if (TD_IS_SUSPENDED(td) || prio >= PSOCK) 1946 td->td_flags |= TDF_CANSWAP; 1947 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1948 return; 1949 if (static_boost == 1 && prio) 1950 sched_prio(td, prio); 1951 else if (static_boost && td->td_priority > static_boost) 1952 sched_prio(td, static_boost); 1953} 1954 1955/* 1956 * Schedule a thread to resume execution and record how long it voluntarily 1957 * slept. We also update the pctcpu, interactivity, and priority. 1958 */ 1959void 1960sched_wakeup(struct thread *td) 1961{ 1962 struct td_sched *ts; 1963 int slptick; 1964 1965 THREAD_LOCK_ASSERT(td, MA_OWNED); 1966 ts = td->td_sched; 1967 td->td_flags &= ~TDF_CANSWAP; 1968 /* 1969 * If we slept for more than a tick update our interactivity and 1970 * priority. 1971 */ 1972 slptick = td->td_slptick; 1973 td->td_slptick = 0; 1974 if (slptick && slptick != ticks) { 1975 u_int hzticks; 1976 1977 hzticks = (ticks - slptick) << SCHED_TICK_SHIFT; 1978 ts->ts_slptime += hzticks; 1979 sched_interact_update(td); 1980 sched_pctcpu_update(ts); 1981 } 1982 /* Reset the slice value after we sleep. */ 1983 ts->ts_slice = sched_slice; 1984 sched_add(td, SRQ_BORING); 1985} 1986 1987/* 1988 * Penalize the parent for creating a new child and initialize the child's 1989 * priority. 1990 */ 1991void 1992sched_fork(struct thread *td, struct thread *child) 1993{ 1994 THREAD_LOCK_ASSERT(td, MA_OWNED); 1995 sched_fork_thread(td, child); 1996 /* 1997 * Penalize the parent and child for forking. 1998 */ 1999 sched_interact_fork(child); 2000 sched_priority(child); 2001 td->td_sched->ts_runtime += tickincr; 2002 sched_interact_update(td); 2003 sched_priority(td); 2004} 2005 2006/* 2007 * Fork a new thread, may be within the same process. 2008 */ 2009void 2010sched_fork_thread(struct thread *td, struct thread *child) 2011{ 2012 struct td_sched *ts; 2013 struct td_sched *ts2; 2014 2015 THREAD_LOCK_ASSERT(td, MA_OWNED); 2016 /* 2017 * Initialize child. 2018 */ 2019 ts = td->td_sched; 2020 ts2 = child->td_sched; 2021 child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 2022 child->td_cpuset = cpuset_ref(td->td_cpuset); 2023 ts2->ts_cpu = ts->ts_cpu; 2024 ts2->ts_flags = 0; 2025 /* 2026 * Grab our parents cpu estimation information. 2027 */ 2028 ts2->ts_ticks = ts->ts_ticks; 2029 ts2->ts_ltick = ts->ts_ltick; 2030 ts2->ts_incrtick = ts->ts_incrtick; 2031 ts2->ts_ftick = ts->ts_ftick; 2032 /* 2033 * Do not inherit any borrowed priority from the parent. 2034 */ 2035 child->td_priority = child->td_base_pri; 2036 /* 2037 * And update interactivity score. 2038 */ 2039 ts2->ts_slptime = ts->ts_slptime; 2040 ts2->ts_runtime = ts->ts_runtime; 2041 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 2042#ifdef KTR 2043 bzero(ts2->ts_name, sizeof(ts2->ts_name)); 2044#endif 2045} 2046 2047/* 2048 * Adjust the priority class of a thread. 2049 */ 2050void 2051sched_class(struct thread *td, int class) 2052{ 2053 2054 THREAD_LOCK_ASSERT(td, MA_OWNED); 2055 if (td->td_pri_class == class) 2056 return; 2057 td->td_pri_class = class; 2058} 2059 2060/* 2061 * Return some of the child's priority and interactivity to the parent. 2062 */ 2063void 2064sched_exit(struct proc *p, struct thread *child) 2065{ 2066 struct thread *td; 2067 2068 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit", 2069 "prio:%d", child->td_priority); 2070 PROC_LOCK_ASSERT(p, MA_OWNED); 2071 td = FIRST_THREAD_IN_PROC(p); 2072 sched_exit_thread(td, child); 2073} 2074 2075/* 2076 * Penalize another thread for the time spent on this one. This helps to 2077 * worsen the priority and interactivity of processes which schedule batch 2078 * jobs such as make. This has little effect on the make process itself but 2079 * causes new processes spawned by it to receive worse scores immediately. 2080 */ 2081void 2082sched_exit_thread(struct thread *td, struct thread *child) 2083{ 2084 2085 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit", 2086 "prio:%d", child->td_priority); 2087 /* 2088 * Give the child's runtime to the parent without returning the 2089 * sleep time as a penalty to the parent. This causes shells that 2090 * launch expensive things to mark their children as expensive. 2091 */ 2092 thread_lock(td); 2093 td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2094 sched_interact_update(td); 2095 sched_priority(td); 2096 thread_unlock(td); 2097} 2098 2099void 2100sched_preempt(struct thread *td) 2101{ 2102 struct tdq *tdq; 2103 2104 thread_lock(td); 2105 tdq = TDQ_SELF(); 2106 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2107 tdq->tdq_ipipending = 0; 2108 if (td->td_priority > tdq->tdq_lowpri) { 2109 int flags; 2110 2111 flags = SW_INVOL | SW_PREEMPT; 2112 if (td->td_critnest > 1) 2113 td->td_owepreempt = 1; 2114 else if (TD_IS_IDLETHREAD(td)) 2115 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL); 2116 else 2117 mi_switch(flags | SWT_REMOTEPREEMPT, NULL); 2118 } 2119 thread_unlock(td); 2120} 2121 2122/* 2123 * Fix priorities on return to user-space. Priorities may be elevated due 2124 * to static priorities in msleep() or similar. 2125 */ 2126void 2127sched_userret(struct thread *td) 2128{ 2129 /* 2130 * XXX we cheat slightly on the locking here to avoid locking in 2131 * the usual case. Setting td_priority here is essentially an 2132 * incomplete workaround for not setting it properly elsewhere. 2133 * Now that some interrupt handlers are threads, not setting it 2134 * properly elsewhere can clobber it in the window between setting 2135 * it here and returning to user mode, so don't waste time setting 2136 * it perfectly here. 2137 */ 2138 KASSERT((td->td_flags & TDF_BORROWING) == 0, 2139 ("thread with borrowed priority returning to userland")); 2140 if (td->td_priority != td->td_user_pri) { 2141 thread_lock(td); 2142 td->td_priority = td->td_user_pri; 2143 td->td_base_pri = td->td_user_pri; 2144 tdq_setlowpri(TDQ_SELF(), td); 2145 thread_unlock(td); 2146 } 2147} 2148 2149/* 2150 * Handle a stathz tick. This is really only relevant for timeshare 2151 * threads. 2152 */ 2153void 2154sched_clock(struct thread *td) 2155{ 2156 struct tdq *tdq; 2157 struct td_sched *ts; 2158 2159 THREAD_LOCK_ASSERT(td, MA_OWNED); 2160 tdq = TDQ_SELF(); 2161#ifdef SMP 2162 /* 2163 * We run the long term load balancer infrequently on the first cpu. 2164 */ 2165 if (balance_tdq == tdq) { 2166 if (balance_ticks && --balance_ticks == 0) 2167 sched_balance(); 2168 } 2169#endif 2170 /* 2171 * Save the old switch count so we have a record of the last ticks 2172 * activity. Initialize the new switch count based on our load. 2173 * If there is some activity seed it to reflect that. 2174 */ 2175 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2176 tdq->tdq_switchcnt = tdq->tdq_load; 2177 /* 2178 * Advance the insert index once for each tick to ensure that all 2179 * threads get a chance to run. 2180 */ 2181 if (tdq->tdq_idx == tdq->tdq_ridx) { 2182 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2183 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2184 tdq->tdq_ridx = tdq->tdq_idx; 2185 } 2186 ts = td->td_sched; 2187 if (td->td_pri_class & PRI_FIFO_BIT) 2188 return; 2189 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) { 2190 /* 2191 * We used a tick; charge it to the thread so 2192 * that we can compute our interactivity. 2193 */ 2194 td->td_sched->ts_runtime += tickincr; 2195 sched_interact_update(td); 2196 sched_priority(td); 2197 } 2198 /* 2199 * We used up one time slice. 2200 */ 2201 if (--ts->ts_slice > 0) 2202 return; 2203 /* 2204 * We're out of time, force a requeue at userret(). 2205 */ 2206 ts->ts_slice = sched_slice; 2207 td->td_flags |= TDF_NEEDRESCHED; 2208} 2209 2210/* 2211 * Called once per hz tick. Used for cpu utilization information. This 2212 * is easier than trying to scale based on stathz. 2213 */ 2214void 2215sched_tick(int cnt) 2216{ 2217 struct td_sched *ts; 2218 2219 ts = curthread->td_sched; 2220 /* 2221 * Ticks is updated asynchronously on a single cpu. Check here to 2222 * avoid incrementing ts_ticks multiple times in a single tick. 2223 */ 2224 if (ts->ts_incrtick == ticks) 2225 return; 2226 /* Adjust ticks for pctcpu */ 2227 ts->ts_ticks += cnt << SCHED_TICK_SHIFT; 2228 ts->ts_ltick = ticks; 2229 ts->ts_incrtick = ticks; 2230 /* 2231 * Update if we've exceeded our desired tick threshold by over one 2232 * second. 2233 */ 2234 if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2235 sched_pctcpu_update(ts); 2236} 2237 2238/* 2239 * Return whether the current CPU has runnable tasks. Used for in-kernel 2240 * cooperative idle threads. 2241 */ 2242int 2243sched_runnable(void) 2244{ 2245 struct tdq *tdq; 2246 int load; 2247 2248 load = 1; 2249 2250 tdq = TDQ_SELF(); 2251 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2252 if (tdq->tdq_load > 0) 2253 goto out; 2254 } else 2255 if (tdq->tdq_load - 1 > 0) 2256 goto out; 2257 load = 0; 2258out: 2259 return (load); 2260} 2261 2262/* 2263 * Choose the highest priority thread to run. The thread is removed from 2264 * the run-queue while running however the load remains. For SMP we set 2265 * the tdq in the global idle bitmask if it idles here. 2266 */ 2267struct thread * 2268sched_choose(void) 2269{ 2270 struct thread *td; 2271 struct tdq *tdq; 2272 2273 tdq = TDQ_SELF(); 2274 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2275 td = tdq_choose(tdq); 2276 if (td) { 2277 td->td_sched->ts_ltick = ticks; 2278 tdq_runq_rem(tdq, td); 2279 tdq->tdq_lowpri = td->td_priority; 2280 return (td); 2281 } 2282 tdq->tdq_lowpri = PRI_MAX_IDLE; 2283 return (PCPU_GET(idlethread)); 2284} 2285 2286/* 2287 * Set owepreempt if necessary. Preemption never happens directly in ULE, 2288 * we always request it once we exit a critical section. 2289 */ 2290static inline void 2291sched_setpreempt(struct thread *td) 2292{ 2293 struct thread *ctd; 2294 int cpri; 2295 int pri; 2296 2297 THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2298 2299 ctd = curthread; 2300 pri = td->td_priority; 2301 cpri = ctd->td_priority; 2302 if (pri < cpri) 2303 ctd->td_flags |= TDF_NEEDRESCHED; 2304 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2305 return; 2306 if (!sched_shouldpreempt(pri, cpri, 0)) 2307 return; 2308 ctd->td_owepreempt = 1; 2309} 2310 2311/* 2312 * Add a thread to a thread queue. Select the appropriate runq and add the 2313 * thread to it. This is the internal function called when the tdq is 2314 * predetermined. 2315 */ 2316void 2317tdq_add(struct tdq *tdq, struct thread *td, int flags) 2318{ 2319 2320 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2321 KASSERT((td->td_inhibitors == 0), 2322 ("sched_add: trying to run inhibited thread")); 2323 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2324 ("sched_add: bad thread state")); 2325 KASSERT(td->td_flags & TDF_INMEM, 2326 ("sched_add: thread swapped out")); 2327 2328 if (td->td_priority < tdq->tdq_lowpri) 2329 tdq->tdq_lowpri = td->td_priority; 2330 tdq_runq_add(tdq, td, flags); 2331 tdq_load_add(tdq, td); 2332} 2333 2334/* 2335 * Select the target thread queue and add a thread to it. Request 2336 * preemption or IPI a remote processor if required. 2337 */ 2338void 2339sched_add(struct thread *td, int flags) 2340{ 2341 struct tdq *tdq; 2342#ifdef SMP 2343 int cpu; 2344#endif 2345 2346 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 2347 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 2348 sched_tdname(curthread)); 2349 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 2350 KTR_ATTR_LINKED, sched_tdname(td)); 2351 THREAD_LOCK_ASSERT(td, MA_OWNED); 2352 /* 2353 * Recalculate the priority before we select the target cpu or 2354 * run-queue. 2355 */ 2356 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2357 sched_priority(td); 2358#ifdef SMP 2359 /* 2360 * Pick the destination cpu and if it isn't ours transfer to the 2361 * target cpu. 2362 */ 2363 cpu = sched_pickcpu(td, flags); 2364 tdq = sched_setcpu(td, cpu, flags); 2365 tdq_add(tdq, td, flags); 2366 if (cpu != PCPU_GET(cpuid)) { 2367 tdq_notify(tdq, td); 2368 return; 2369 } 2370#else 2371 tdq = TDQ_SELF(); 2372 TDQ_LOCK(tdq); 2373 /* 2374 * Now that the thread is moving to the run-queue, set the lock 2375 * to the scheduler's lock. 2376 */ 2377 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2378 tdq_add(tdq, td, flags); 2379#endif 2380 if (!(flags & SRQ_YIELDING)) 2381 sched_setpreempt(td); 2382} 2383 2384/* 2385 * Remove a thread from a run-queue without running it. This is used 2386 * when we're stealing a thread from a remote queue. Otherwise all threads 2387 * exit by calling sched_exit_thread() and sched_throw() themselves. 2388 */ 2389void 2390sched_rem(struct thread *td) 2391{ 2392 struct tdq *tdq; 2393 2394 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 2395 "prio:%d", td->td_priority); 2396 tdq = TDQ_CPU(td->td_sched->ts_cpu); 2397 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2398 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2399 KASSERT(TD_ON_RUNQ(td), 2400 ("sched_rem: thread not on run queue")); 2401 tdq_runq_rem(tdq, td); 2402 tdq_load_rem(tdq, td); 2403 TD_SET_CAN_RUN(td); 2404 if (td->td_priority == tdq->tdq_lowpri) 2405 tdq_setlowpri(tdq, NULL); 2406} 2407 2408/* 2409 * Fetch cpu utilization information. Updates on demand. 2410 */ 2411fixpt_t 2412sched_pctcpu(struct thread *td) 2413{ 2414 fixpt_t pctcpu; 2415 struct td_sched *ts; 2416 2417 pctcpu = 0; 2418 ts = td->td_sched; 2419 if (ts == NULL) 2420 return (0); 2421 2422 THREAD_LOCK_ASSERT(td, MA_OWNED); 2423 if (ts->ts_ticks) { 2424 int rtick; 2425 2426 sched_pctcpu_update(ts); 2427 /* How many rtick per second ? */ 2428 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2429 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2430 } 2431 2432 return (pctcpu); 2433} 2434 2435/* 2436 * Enforce affinity settings for a thread. Called after adjustments to 2437 * cpumask. 2438 */ 2439void 2440sched_affinity(struct thread *td) 2441{ 2442#ifdef SMP 2443 struct td_sched *ts; 2444 2445 THREAD_LOCK_ASSERT(td, MA_OWNED); 2446 ts = td->td_sched; 2447 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2448 return; 2449 if (TD_ON_RUNQ(td)) { 2450 sched_rem(td); 2451 sched_add(td, SRQ_BORING); 2452 return; 2453 } 2454 if (!TD_IS_RUNNING(td)) 2455 return; 2456 /* 2457 * Force a switch before returning to userspace. If the 2458 * target thread is not running locally send an ipi to force 2459 * the issue. 2460 */ 2461 td->td_flags |= TDF_NEEDRESCHED; 2462 if (td != curthread) 2463 ipi_cpu(ts->ts_cpu, IPI_PREEMPT); 2464#endif 2465} 2466 2467/* 2468 * Bind a thread to a target cpu. 2469 */ 2470void 2471sched_bind(struct thread *td, int cpu) 2472{ 2473 struct td_sched *ts; 2474 2475 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2476 KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 2477 ts = td->td_sched; 2478 if (ts->ts_flags & TSF_BOUND) 2479 sched_unbind(td); 2480 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td)); 2481 ts->ts_flags |= TSF_BOUND; 2482 sched_pin(); 2483 if (PCPU_GET(cpuid) == cpu) 2484 return; 2485 ts->ts_cpu = cpu; 2486 /* When we return from mi_switch we'll be on the correct cpu. */ 2487 mi_switch(SW_VOL, NULL); 2488} 2489 2490/* 2491 * Release a bound thread. 2492 */ 2493void 2494sched_unbind(struct thread *td) 2495{ 2496 struct td_sched *ts; 2497 2498 THREAD_LOCK_ASSERT(td, MA_OWNED); 2499 KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 2500 ts = td->td_sched; 2501 if ((ts->ts_flags & TSF_BOUND) == 0) 2502 return; 2503 ts->ts_flags &= ~TSF_BOUND; 2504 sched_unpin(); 2505} 2506 2507int 2508sched_is_bound(struct thread *td) 2509{ 2510 THREAD_LOCK_ASSERT(td, MA_OWNED); 2511 return (td->td_sched->ts_flags & TSF_BOUND); 2512} 2513 2514/* 2515 * Basic yield call. 2516 */ 2517void 2518sched_relinquish(struct thread *td) 2519{ 2520 thread_lock(td); 2521 mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 2522 thread_unlock(td); 2523} 2524 2525/* 2526 * Return the total system load. 2527 */ 2528int 2529sched_load(void) 2530{ 2531#ifdef SMP 2532 int total; 2533 int i; 2534 2535 total = 0; 2536 CPU_FOREACH(i) 2537 total += TDQ_CPU(i)->tdq_sysload; 2538 return (total); 2539#else 2540 return (TDQ_SELF()->tdq_sysload); 2541#endif 2542} 2543 2544int 2545sched_sizeof_proc(void) 2546{ 2547 return (sizeof(struct proc)); 2548} 2549 2550int 2551sched_sizeof_thread(void) 2552{ 2553 return (sizeof(struct thread) + sizeof(struct td_sched)); 2554} 2555 2556#ifdef SMP 2557#define TDQ_IDLESPIN(tdq) \ 2558 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0) 2559#else 2560#define TDQ_IDLESPIN(tdq) 1 2561#endif 2562 2563/* 2564 * The actual idle process. 2565 */ 2566void 2567sched_idletd(void *dummy) 2568{ 2569 struct thread *td; 2570 struct tdq *tdq; 2571 int switchcnt; 2572 int i; 2573 2574 mtx_assert(&Giant, MA_NOTOWNED); 2575 td = curthread; 2576 tdq = TDQ_SELF(); 2577 for (;;) { 2578#ifdef SMP 2579 if (tdq_idled(tdq) == 0) 2580 continue; 2581#endif 2582 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2583 /* 2584 * If we're switching very frequently, spin while checking 2585 * for load rather than entering a low power state that 2586 * may require an IPI. However, don't do any busy 2587 * loops while on SMT machines as this simply steals 2588 * cycles from cores doing useful work. 2589 */ 2590 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) { 2591 for (i = 0; i < sched_idlespins; i++) { 2592 if (tdq->tdq_load) 2593 break; 2594 cpu_spinwait(); 2595 } 2596 } 2597 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2598 if (tdq->tdq_load == 0) { 2599 tdq->tdq_cpu_idle = 1; 2600 if (tdq->tdq_load == 0) { 2601 cpu_idle(switchcnt > sched_idlespinthresh * 4); 2602 tdq->tdq_switchcnt++; 2603 } 2604 tdq->tdq_cpu_idle = 0; 2605 } 2606 if (tdq->tdq_load) { 2607 thread_lock(td); 2608 mi_switch(SW_VOL | SWT_IDLE, NULL); 2609 thread_unlock(td); 2610 } 2611 } 2612} 2613 2614/* 2615 * A CPU is entering for the first time or a thread is exiting. 2616 */ 2617void 2618sched_throw(struct thread *td) 2619{ 2620 struct thread *newtd; 2621 struct tdq *tdq; 2622 2623 tdq = TDQ_SELF(); 2624 if (td == NULL) { 2625 /* Correct spinlock nesting and acquire the correct lock. */ 2626 TDQ_LOCK(tdq); 2627 spinlock_exit(); 2628 PCPU_SET(switchtime, cpu_ticks()); 2629 PCPU_SET(switchticks, ticks); 2630 } else { 2631 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2632 tdq_load_rem(tdq, td); 2633 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2634 } 2635 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2636 newtd = choosethread(); 2637 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2638 cpu_throw(td, newtd); /* doesn't return */ 2639} 2640 2641/* 2642 * This is called from fork_exit(). Just acquire the correct locks and 2643 * let fork do the rest of the work. 2644 */ 2645void 2646sched_fork_exit(struct thread *td) 2647{ 2648 struct td_sched *ts; 2649 struct tdq *tdq; 2650 int cpuid; 2651 2652 /* 2653 * Finish setting up thread glue so that it begins execution in a 2654 * non-nested critical section with the scheduler lock held. 2655 */ 2656 cpuid = PCPU_GET(cpuid); 2657 tdq = TDQ_CPU(cpuid); 2658 ts = td->td_sched; 2659 if (TD_IS_IDLETHREAD(td)) 2660 td->td_lock = TDQ_LOCKPTR(tdq); 2661 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2662 td->td_oncpu = cpuid; 2663 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2664 lock_profile_obtain_lock_success( 2665 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2666} 2667 2668/* 2669 * Create on first use to catch odd startup conditons. 2670 */ 2671char * 2672sched_tdname(struct thread *td) 2673{ 2674#ifdef KTR 2675 struct td_sched *ts; 2676 2677 ts = td->td_sched; 2678 if (ts->ts_name[0] == '\0') 2679 snprintf(ts->ts_name, sizeof(ts->ts_name), 2680 "%s tid %d", td->td_name, td->td_tid); 2681 return (ts->ts_name); 2682#else 2683 return (td->td_name); 2684#endif 2685} 2686 2687#ifdef KTR 2688void 2689sched_clear_tdname(struct thread *td) 2690{ 2691 struct td_sched *ts; 2692 2693 ts = td->td_sched; 2694 ts->ts_name[0] = '\0'; 2695} 2696#endif 2697 2698#ifdef SMP 2699 2700/* 2701 * Build the CPU topology dump string. Is recursively called to collect 2702 * the topology tree. 2703 */ 2704static int 2705sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg, 2706 int indent) 2707{ 2708 char cpusetbuf[CPUSETBUFSIZ]; 2709 int i, first; 2710 2711 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent, 2712 "", 1 + indent / 2, cg->cg_level); 2713 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "", 2714 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask)); 2715 first = TRUE; 2716 for (i = 0; i < MAXCPU; i++) { 2717 if (CPU_ISSET(i, &cg->cg_mask)) { 2718 if (!first) 2719 sbuf_printf(sb, ", "); 2720 else 2721 first = FALSE; 2722 sbuf_printf(sb, "%d", i); 2723 } 2724 } 2725 sbuf_printf(sb, "</cpu>\n"); 2726 2727 if (cg->cg_flags != 0) { 2728 sbuf_printf(sb, "%*s <flags>", indent, ""); 2729 if ((cg->cg_flags & CG_FLAG_HTT) != 0) 2730 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>"); 2731 if ((cg->cg_flags & CG_FLAG_THREAD) != 0) 2732 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>"); 2733 if ((cg->cg_flags & CG_FLAG_SMT) != 0) 2734 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>"); 2735 sbuf_printf(sb, "</flags>\n"); 2736 } 2737 2738 if (cg->cg_children > 0) { 2739 sbuf_printf(sb, "%*s <children>\n", indent, ""); 2740 for (i = 0; i < cg->cg_children; i++) 2741 sysctl_kern_sched_topology_spec_internal(sb, 2742 &cg->cg_child[i], indent+2); 2743 sbuf_printf(sb, "%*s </children>\n", indent, ""); 2744 } 2745 sbuf_printf(sb, "%*s</group>\n", indent, ""); 2746 return (0); 2747} 2748 2749/* 2750 * Sysctl handler for retrieving topology dump. It's a wrapper for 2751 * the recursive sysctl_kern_smp_topology_spec_internal(). 2752 */ 2753static int 2754sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS) 2755{ 2756 struct sbuf *topo; 2757 int err; 2758 2759 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized")); 2760 2761 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND); 2762 if (topo == NULL) 2763 return (ENOMEM); 2764 2765 sbuf_printf(topo, "<groups>\n"); 2766 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1); 2767 sbuf_printf(topo, "</groups>\n"); 2768 2769 if (err == 0) { 2770 sbuf_finish(topo); 2771 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo)); 2772 } 2773 sbuf_delete(topo); 2774 return (err); 2775} 2776 2777#endif 2778 2779SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2780SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2781 "Scheduler name"); 2782SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2783 "Slice size for timeshare threads"); 2784SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2785 "Interactivity score threshold"); 2786SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2787 0,"Min priority for preemption, lower priorities have greater precedence"); 2788SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 2789 0,"Controls whether static kernel priorities are assigned to sleeping threads."); 2790SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 2791 0,"Number of times idle will spin waiting for new work."); 2792SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, &sched_idlespinthresh, 2793 0,"Threshold before we will permit idle spinning."); 2794#ifdef SMP 2795SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2796 "Number of hz ticks to keep thread affinity for"); 2797SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2798 "Enables the long-term load balancer"); 2799SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2800 &balance_interval, 0, 2801 "Average frequency in stathz ticks to run the long-term balancer"); 2802SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2803 "Attempts to steal work from other cores before idling"); 2804SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2805 "Minimum load on remote cpu before we'll steal"); 2806 2807/* Retrieve SMP topology */ 2808SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING | 2809 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A", 2810 "XML dump of detected CPU topology"); 2811 2812#endif 2813 2814/* ps compat. All cpu percentages from ULE are weighted. */ 2815static int ccpu = 0; 2816SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2817