sched_ule.c revision 241248
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 241248 2012-10-06 12:51:16Z mav $"); 40 41#include "opt_hwpmc_hooks.h" 42#include "opt_kdtrace.h" 43#include "opt_sched.h" 44 45#include <sys/param.h> 46#include <sys/systm.h> 47#include <sys/kdb.h> 48#include <sys/kernel.h> 49#include <sys/ktr.h> 50#include <sys/lock.h> 51#include <sys/mutex.h> 52#include <sys/proc.h> 53#include <sys/resource.h> 54#include <sys/resourcevar.h> 55#include <sys/sched.h> 56#include <sys/sdt.h> 57#include <sys/smp.h> 58#include <sys/sx.h> 59#include <sys/sysctl.h> 60#include <sys/sysproto.h> 61#include <sys/turnstile.h> 62#include <sys/umtx.h> 63#include <sys/vmmeter.h> 64#include <sys/cpuset.h> 65#include <sys/sbuf.h> 66 67#ifdef HWPMC_HOOKS 68#include <sys/pmckern.h> 69#endif 70 71#ifdef KDTRACE_HOOKS 72#include <sys/dtrace_bsd.h> 73int dtrace_vtime_active; 74dtrace_vtime_switch_func_t dtrace_vtime_switch_func; 75#endif 76 77#include <machine/cpu.h> 78#include <machine/smp.h> 79 80#if defined(__powerpc__) && defined(E500) 81#error "This architecture is not currently compatible with ULE" 82#endif 83 84#define KTR_ULE 0 85 86#define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX))) 87#define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU))) 88#define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load")) 89 90/* 91 * Thread scheduler specific section. All fields are protected 92 * by the thread lock. 93 */ 94struct td_sched { 95 struct runq *ts_runq; /* Run-queue we're queued on. */ 96 short ts_flags; /* TSF_* flags. */ 97 u_char ts_cpu; /* CPU that we have affinity for. */ 98 int ts_rltick; /* Real last tick, for affinity. */ 99 int ts_slice; /* Ticks of slice remaining. */ 100 u_int ts_slptime; /* Number of ticks we vol. slept */ 101 u_int ts_runtime; /* Number of ticks we were running */ 102 int ts_ltick; /* Last tick that we were running on */ 103 int ts_ftick; /* First tick that we were running on */ 104 int ts_ticks; /* Tick count */ 105#ifdef KTR 106 char ts_name[TS_NAME_LEN]; 107#endif 108}; 109/* flags kept in ts_flags */ 110#define TSF_BOUND 0x0001 /* Thread can not migrate. */ 111#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 112 113static struct td_sched td_sched0; 114 115#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 116#define THREAD_CAN_SCHED(td, cpu) \ 117 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 118 119/* 120 * Priority ranges used for interactive and non-interactive timeshare 121 * threads. The timeshare priorities are split up into four ranges. 122 * The first range handles interactive threads. The last three ranges 123 * (NHALF, x, and NHALF) handle non-interactive threads with the outer 124 * ranges supporting nice values. 125 */ 126#define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1) 127#define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2) 128#define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE) 129 130#define PRI_MIN_INTERACT PRI_MIN_TIMESHARE 131#define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1) 132#define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE) 133#define PRI_MAX_BATCH PRI_MAX_TIMESHARE 134 135/* 136 * Cpu percentage computation macros and defines. 137 * 138 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 139 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 140 * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 141 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 142 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 143 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 144 */ 145#define SCHED_TICK_SECS 10 146#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 147#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 148#define SCHED_TICK_SHIFT 10 149#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 150#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 151 152/* 153 * These macros determine priorities for non-interactive threads. They are 154 * assigned a priority based on their recent cpu utilization as expressed 155 * by the ratio of ticks to the tick total. NHALF priorities at the start 156 * and end of the MIN to MAX timeshare range are only reachable with negative 157 * or positive nice respectively. 158 * 159 * PRI_RANGE: Priority range for utilization dependent priorities. 160 * PRI_NRESV: Number of nice values. 161 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 162 * PRI_NICE: Determines the part of the priority inherited from nice. 163 */ 164#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 165#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 166#define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF) 167#define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF) 168#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1) 169#define SCHED_PRI_TICKS(ts) \ 170 (SCHED_TICK_HZ((ts)) / \ 171 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 172#define SCHED_PRI_NICE(nice) (nice) 173 174/* 175 * These determine the interactivity of a process. Interactivity differs from 176 * cpu utilization in that it expresses the voluntary time slept vs time ran 177 * while cpu utilization includes all time not running. This more accurately 178 * models the intent of the thread. 179 * 180 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 181 * before throttling back. 182 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 183 * INTERACT_MAX: Maximum interactivity value. Smaller is better. 184 * INTERACT_THRESH: Threshold for placement on the current runq. 185 */ 186#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 187#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 188#define SCHED_INTERACT_MAX (100) 189#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 190#define SCHED_INTERACT_THRESH (30) 191 192/* 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; 205static int tickincr; 206static int sched_slice = 1; 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, change-pri, "struct thread *", 337 "struct proc *", "uint8_t"); 338SDT_PROBE_DEFINE3(sched, , , dequeue, dequeue, "struct thread *", 339 "struct proc *", "void *"); 340SDT_PROBE_DEFINE4(sched, , , enqueue, enqueue, "struct thread *", 341 "struct proc *", "void *", "int"); 342SDT_PROBE_DEFINE4(sched, , , lend_pri, lend-pri, "struct thread *", 343 "struct proc *", "uint8_t", "struct thread *"); 344SDT_PROBE_DEFINE2(sched, , , load_change, load-change, "int", "int"); 345SDT_PROBE_DEFINE2(sched, , , off_cpu, off-cpu, "struct thread *", 346 "struct proc *"); 347SDT_PROBE_DEFINE(sched, , , on_cpu, on-cpu); 348SDT_PROBE_DEFINE(sched, , , remain_cpu, remain-cpu); 349SDT_PROBE_DEFINE2(sched, , , surrender, 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; i >= 0; ) { 636 if (i == 0) { 637 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask)) 638 cpu--; 639 if (cpu < 0) 640 break; 641 child = NULL; 642 } else 643 child = &cg->cg_child[i - 1]; 644 645 if (match & CPU_SEARCH_LOWEST) 646 lgroup.cs_cpu = -1; 647 if (match & CPU_SEARCH_HIGHEST) 648 hgroup.cs_cpu = -1; 649 if (child) { /* Handle child CPU group. */ 650 CPU_NAND(&cpumask, &child->cg_mask); 651 switch (match) { 652 case CPU_SEARCH_LOWEST: 653 load = cpu_search_lowest(child, &lgroup); 654 break; 655 case CPU_SEARCH_HIGHEST: 656 load = cpu_search_highest(child, &hgroup); 657 break; 658 case CPU_SEARCH_BOTH: 659 load = cpu_search_both(child, &lgroup, &hgroup); 660 break; 661 } 662 } else { /* Handle child CPU. */ 663 tdq = TDQ_CPU(cpu); 664 load = tdq->tdq_load * 256; 665 rndptr = DPCPU_PTR(randomval); 666 rnd = (*rndptr = *rndptr * 69069 + 5) >> 26; 667 if (match & CPU_SEARCH_LOWEST) { 668 if (cpu == low->cs_prefer) 669 load -= 64; 670 /* If that CPU is allowed and get data. */ 671 if (tdq->tdq_lowpri > lgroup.cs_pri && 672 tdq->tdq_load <= lgroup.cs_limit && 673 CPU_ISSET(cpu, &lgroup.cs_mask)) { 674 lgroup.cs_cpu = cpu; 675 lgroup.cs_load = load - rnd; 676 } 677 } 678 if (match & CPU_SEARCH_HIGHEST) 679 if (tdq->tdq_load >= hgroup.cs_limit && 680 tdq->tdq_transferable && 681 CPU_ISSET(cpu, &hgroup.cs_mask)) { 682 hgroup.cs_cpu = cpu; 683 hgroup.cs_load = load - rnd; 684 } 685 } 686 total += load; 687 688 /* We have info about child item. Compare it. */ 689 if (match & CPU_SEARCH_LOWEST) { 690 if (lgroup.cs_cpu >= 0 && 691 (load < lload || 692 (load == lload && lgroup.cs_load < low->cs_load))) { 693 lload = load; 694 low->cs_cpu = lgroup.cs_cpu; 695 low->cs_load = lgroup.cs_load; 696 } 697 } 698 if (match & CPU_SEARCH_HIGHEST) 699 if (hgroup.cs_cpu >= 0 && 700 (load > hload || 701 (load == hload && hgroup.cs_load > high->cs_load))) { 702 hload = load; 703 high->cs_cpu = hgroup.cs_cpu; 704 high->cs_load = hgroup.cs_load; 705 } 706 if (child) { 707 i--; 708 if (i == 0 && CPU_EMPTY(&cpumask)) 709 break; 710 } else 711 cpu--; 712 } 713 return (total); 714} 715 716/* 717 * cpu_search instantiations must pass constants to maintain the inline 718 * optimization. 719 */ 720int 721cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low) 722{ 723 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 724} 725 726int 727cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high) 728{ 729 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 730} 731 732int 733cpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 734 struct cpu_search *high) 735{ 736 return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 737} 738 739/* 740 * Find the cpu with the least load via the least loaded path that has a 741 * lowpri greater than pri pri. A pri of -1 indicates any priority is 742 * acceptable. 743 */ 744static inline int 745sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload, 746 int prefer) 747{ 748 struct cpu_search low; 749 750 low.cs_cpu = -1; 751 low.cs_prefer = prefer; 752 low.cs_mask = mask; 753 low.cs_pri = pri; 754 low.cs_limit = maxload; 755 cpu_search_lowest(cg, &low); 756 return low.cs_cpu; 757} 758 759/* 760 * Find the cpu with the highest load via the highest loaded path. 761 */ 762static inline int 763sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload) 764{ 765 struct cpu_search high; 766 767 high.cs_cpu = -1; 768 high.cs_mask = mask; 769 high.cs_limit = minload; 770 cpu_search_highest(cg, &high); 771 return high.cs_cpu; 772} 773 774/* 775 * Simultaneously find the highest and lowest loaded cpu reachable via 776 * cg. 777 */ 778static inline void 779sched_both(const struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu) 780{ 781 struct cpu_search high; 782 struct cpu_search low; 783 784 low.cs_cpu = -1; 785 low.cs_prefer = -1; 786 low.cs_pri = -1; 787 low.cs_limit = INT_MAX; 788 low.cs_mask = mask; 789 high.cs_cpu = -1; 790 high.cs_limit = -1; 791 high.cs_mask = mask; 792 cpu_search_both(cg, &low, &high); 793 *lowcpu = low.cs_cpu; 794 *highcpu = high.cs_cpu; 795 return; 796} 797 798static void 799sched_balance_group(struct cpu_group *cg) 800{ 801 cpuset_t hmask, lmask; 802 int high, low, anylow; 803 804 CPU_FILL(&hmask); 805 for (;;) { 806 high = sched_highest(cg, hmask, 1); 807 /* Stop if there is no more CPU with transferrable threads. */ 808 if (high == -1) 809 break; 810 CPU_CLR(high, &hmask); 811 CPU_COPY(&hmask, &lmask); 812 /* Stop if there is no more CPU left for low. */ 813 if (CPU_EMPTY(&lmask)) 814 break; 815 anylow = 1; 816nextlow: 817 low = sched_lowest(cg, lmask, -1, 818 TDQ_CPU(high)->tdq_load - 1, high); 819 /* Stop if we looked well and found no less loaded CPU. */ 820 if (anylow && low == -1) 821 break; 822 /* Go to next high if we found no less loaded CPU. */ 823 if (low == -1) 824 continue; 825 /* Transfer thread from high to low. */ 826 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) { 827 /* CPU that got thread can no longer be a donor. */ 828 CPU_CLR(low, &hmask); 829 } else { 830 /* 831 * If failed, then there is no threads on high 832 * that can run on this low. Drop low from low 833 * mask and look for different one. 834 */ 835 CPU_CLR(low, &lmask); 836 anylow = 0; 837 goto nextlow; 838 } 839 } 840} 841 842static void 843sched_balance(void) 844{ 845 struct tdq *tdq; 846 847 /* 848 * Select a random time between .5 * balance_interval and 849 * 1.5 * balance_interval. 850 */ 851 balance_ticks = max(balance_interval / 2, 1); 852 balance_ticks += random() % balance_interval; 853 if (smp_started == 0 || rebalance == 0) 854 return; 855 tdq = TDQ_SELF(); 856 TDQ_UNLOCK(tdq); 857 sched_balance_group(cpu_top); 858 TDQ_LOCK(tdq); 859} 860 861/* 862 * Lock two thread queues using their address to maintain lock order. 863 */ 864static void 865tdq_lock_pair(struct tdq *one, struct tdq *two) 866{ 867 if (one < two) { 868 TDQ_LOCK(one); 869 TDQ_LOCK_FLAGS(two, MTX_DUPOK); 870 } else { 871 TDQ_LOCK(two); 872 TDQ_LOCK_FLAGS(one, MTX_DUPOK); 873 } 874} 875 876/* 877 * Unlock two thread queues. Order is not important here. 878 */ 879static void 880tdq_unlock_pair(struct tdq *one, struct tdq *two) 881{ 882 TDQ_UNLOCK(one); 883 TDQ_UNLOCK(two); 884} 885 886/* 887 * Transfer load between two imbalanced thread queues. 888 */ 889static int 890sched_balance_pair(struct tdq *high, struct tdq *low) 891{ 892 int moved; 893 int cpu; 894 895 tdq_lock_pair(high, low); 896 moved = 0; 897 /* 898 * Determine what the imbalance is and then adjust that to how many 899 * threads we actually have to give up (transferable). 900 */ 901 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load && 902 (moved = tdq_move(high, low)) > 0) { 903 /* 904 * In case the target isn't the current cpu IPI it to force a 905 * reschedule with the new workload. 906 */ 907 cpu = TDQ_ID(low); 908 sched_pin(); 909 if (cpu != PCPU_GET(cpuid)) 910 ipi_cpu(cpu, IPI_PREEMPT); 911 sched_unpin(); 912 } 913 tdq_unlock_pair(high, low); 914 return (moved); 915} 916 917/* 918 * Move a thread from one thread queue to another. 919 */ 920static int 921tdq_move(struct tdq *from, struct tdq *to) 922{ 923 struct td_sched *ts; 924 struct thread *td; 925 struct tdq *tdq; 926 int cpu; 927 928 TDQ_LOCK_ASSERT(from, MA_OWNED); 929 TDQ_LOCK_ASSERT(to, MA_OWNED); 930 931 tdq = from; 932 cpu = TDQ_ID(to); 933 td = tdq_steal(tdq, cpu); 934 if (td == NULL) 935 return (0); 936 ts = td->td_sched; 937 /* 938 * Although the run queue is locked the thread may be blocked. Lock 939 * it to clear this and acquire the run-queue lock. 940 */ 941 thread_lock(td); 942 /* Drop recursive lock on from acquired via thread_lock(). */ 943 TDQ_UNLOCK(from); 944 sched_rem(td); 945 ts->ts_cpu = cpu; 946 td->td_lock = TDQ_LOCKPTR(to); 947 tdq_add(to, td, SRQ_YIELDING); 948 return (1); 949} 950 951/* 952 * This tdq has idled. Try to steal a thread from another cpu and switch 953 * to it. 954 */ 955static int 956tdq_idled(struct tdq *tdq) 957{ 958 struct cpu_group *cg; 959 struct tdq *steal; 960 cpuset_t mask; 961 int thresh; 962 int cpu; 963 964 if (smp_started == 0 || steal_idle == 0) 965 return (1); 966 CPU_FILL(&mask); 967 CPU_CLR(PCPU_GET(cpuid), &mask); 968 /* We don't want to be preempted while we're iterating. */ 969 spinlock_enter(); 970 for (cg = tdq->tdq_cg; cg != NULL; ) { 971 if ((cg->cg_flags & CG_FLAG_THREAD) == 0) 972 thresh = steal_thresh; 973 else 974 thresh = 1; 975 cpu = sched_highest(cg, mask, thresh); 976 if (cpu == -1) { 977 cg = cg->cg_parent; 978 continue; 979 } 980 steal = TDQ_CPU(cpu); 981 CPU_CLR(cpu, &mask); 982 tdq_lock_pair(tdq, steal); 983 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 984 tdq_unlock_pair(tdq, steal); 985 continue; 986 } 987 /* 988 * If a thread was added while interrupts were disabled don't 989 * steal one here. If we fail to acquire one due to affinity 990 * restrictions loop again with this cpu removed from the 991 * set. 992 */ 993 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 994 tdq_unlock_pair(tdq, steal); 995 continue; 996 } 997 spinlock_exit(); 998 TDQ_UNLOCK(steal); 999 mi_switch(SW_VOL | SWT_IDLE, NULL); 1000 thread_unlock(curthread); 1001 1002 return (0); 1003 } 1004 spinlock_exit(); 1005 return (1); 1006} 1007 1008/* 1009 * Notify a remote cpu of new work. Sends an IPI if criteria are met. 1010 */ 1011static void 1012tdq_notify(struct tdq *tdq, struct thread *td) 1013{ 1014 struct thread *ctd; 1015 int pri; 1016 int cpu; 1017 1018 if (tdq->tdq_ipipending) 1019 return; 1020 cpu = td->td_sched->ts_cpu; 1021 pri = td->td_priority; 1022 ctd = pcpu_find(cpu)->pc_curthread; 1023 if (!sched_shouldpreempt(pri, ctd->td_priority, 1)) 1024 return; 1025 if (TD_IS_IDLETHREAD(ctd)) { 1026 /* 1027 * If the MD code has an idle wakeup routine try that before 1028 * falling back to IPI. 1029 */ 1030 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu)) 1031 return; 1032 } 1033 tdq->tdq_ipipending = 1; 1034 ipi_cpu(cpu, IPI_PREEMPT); 1035} 1036 1037/* 1038 * Steals load from a timeshare queue. Honors the rotating queue head 1039 * index. 1040 */ 1041static struct thread * 1042runq_steal_from(struct runq *rq, int cpu, u_char start) 1043{ 1044 struct rqbits *rqb; 1045 struct rqhead *rqh; 1046 struct thread *td, *first; 1047 int bit; 1048 int pri; 1049 int i; 1050 1051 rqb = &rq->rq_status; 1052 bit = start & (RQB_BPW -1); 1053 pri = 0; 1054 first = NULL; 1055again: 1056 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 1057 if (rqb->rqb_bits[i] == 0) 1058 continue; 1059 if (bit != 0) { 1060 for (pri = bit; pri < RQB_BPW; pri++) 1061 if (rqb->rqb_bits[i] & (1ul << pri)) 1062 break; 1063 if (pri >= RQB_BPW) 1064 continue; 1065 } else 1066 pri = RQB_FFS(rqb->rqb_bits[i]); 1067 pri += (i << RQB_L2BPW); 1068 rqh = &rq->rq_queues[pri]; 1069 TAILQ_FOREACH(td, rqh, td_runq) { 1070 if (first && THREAD_CAN_MIGRATE(td) && 1071 THREAD_CAN_SCHED(td, cpu)) 1072 return (td); 1073 first = td; 1074 } 1075 } 1076 if (start != 0) { 1077 start = 0; 1078 goto again; 1079 } 1080 1081 if (first && THREAD_CAN_MIGRATE(first) && 1082 THREAD_CAN_SCHED(first, cpu)) 1083 return (first); 1084 return (NULL); 1085} 1086 1087/* 1088 * Steals load from a standard linear queue. 1089 */ 1090static struct thread * 1091runq_steal(struct runq *rq, int cpu) 1092{ 1093 struct rqhead *rqh; 1094 struct rqbits *rqb; 1095 struct thread *td; 1096 int word; 1097 int bit; 1098 1099 rqb = &rq->rq_status; 1100 for (word = 0; word < RQB_LEN; word++) { 1101 if (rqb->rqb_bits[word] == 0) 1102 continue; 1103 for (bit = 0; bit < RQB_BPW; bit++) { 1104 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1105 continue; 1106 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1107 TAILQ_FOREACH(td, rqh, td_runq) 1108 if (THREAD_CAN_MIGRATE(td) && 1109 THREAD_CAN_SCHED(td, cpu)) 1110 return (td); 1111 } 1112 } 1113 return (NULL); 1114} 1115 1116/* 1117 * Attempt to steal a thread in priority order from a thread queue. 1118 */ 1119static struct thread * 1120tdq_steal(struct tdq *tdq, int cpu) 1121{ 1122 struct thread *td; 1123 1124 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1125 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1126 return (td); 1127 if ((td = runq_steal_from(&tdq->tdq_timeshare, 1128 cpu, tdq->tdq_ridx)) != NULL) 1129 return (td); 1130 return (runq_steal(&tdq->tdq_idle, cpu)); 1131} 1132 1133/* 1134 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1135 * current lock and returns with the assigned queue locked. 1136 */ 1137static inline struct tdq * 1138sched_setcpu(struct thread *td, int cpu, int flags) 1139{ 1140 1141 struct tdq *tdq; 1142 1143 THREAD_LOCK_ASSERT(td, MA_OWNED); 1144 tdq = TDQ_CPU(cpu); 1145 td->td_sched->ts_cpu = cpu; 1146 /* 1147 * If the lock matches just return the queue. 1148 */ 1149 if (td->td_lock == TDQ_LOCKPTR(tdq)) 1150 return (tdq); 1151#ifdef notyet 1152 /* 1153 * If the thread isn't running its lockptr is a 1154 * turnstile or a sleepqueue. We can just lock_set without 1155 * blocking. 1156 */ 1157 if (TD_CAN_RUN(td)) { 1158 TDQ_LOCK(tdq); 1159 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1160 return (tdq); 1161 } 1162#endif 1163 /* 1164 * The hard case, migration, we need to block the thread first to 1165 * prevent order reversals with other cpus locks. 1166 */ 1167 spinlock_enter(); 1168 thread_lock_block(td); 1169 TDQ_LOCK(tdq); 1170 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1171 spinlock_exit(); 1172 return (tdq); 1173} 1174 1175SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1176SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1177SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1178SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1179SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1180SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1181 1182static int 1183sched_pickcpu(struct thread *td, int flags) 1184{ 1185 struct cpu_group *cg, *ccg; 1186 struct td_sched *ts; 1187 struct tdq *tdq; 1188 cpuset_t mask; 1189 int cpu, pri, self; 1190 1191 self = PCPU_GET(cpuid); 1192 ts = td->td_sched; 1193 if (smp_started == 0) 1194 return (self); 1195 /* 1196 * Don't migrate a running thread from sched_switch(). 1197 */ 1198 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1199 return (ts->ts_cpu); 1200 /* 1201 * Prefer to run interrupt threads on the processors that generate 1202 * the interrupt. 1203 */ 1204 pri = td->td_priority; 1205 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1206 curthread->td_intr_nesting_level && ts->ts_cpu != self) { 1207 SCHED_STAT_INC(pickcpu_intrbind); 1208 ts->ts_cpu = self; 1209 if (TDQ_CPU(self)->tdq_lowpri > pri) { 1210 SCHED_STAT_INC(pickcpu_affinity); 1211 return (ts->ts_cpu); 1212 } 1213 } 1214 /* 1215 * If the thread can run on the last cpu and the affinity has not 1216 * expired or it is idle run it there. 1217 */ 1218 tdq = TDQ_CPU(ts->ts_cpu); 1219 cg = tdq->tdq_cg; 1220 if (THREAD_CAN_SCHED(td, ts->ts_cpu) && 1221 tdq->tdq_lowpri >= PRI_MIN_IDLE && 1222 SCHED_AFFINITY(ts, CG_SHARE_L2)) { 1223 if (cg->cg_flags & CG_FLAG_THREAD) { 1224 CPUSET_FOREACH(cpu, cg->cg_mask) { 1225 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) 1226 break; 1227 } 1228 } else 1229 cpu = INT_MAX; 1230 if (cpu > mp_maxid) { 1231 SCHED_STAT_INC(pickcpu_idle_affinity); 1232 return (ts->ts_cpu); 1233 } 1234 } 1235 /* 1236 * Search for the last level cache CPU group in the tree. 1237 * Skip caches with expired affinity time and SMT groups. 1238 * Affinity to higher level caches will be handled less aggressively. 1239 */ 1240 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) { 1241 if (cg->cg_flags & CG_FLAG_THREAD) 1242 continue; 1243 if (!SCHED_AFFINITY(ts, cg->cg_level)) 1244 continue; 1245 ccg = cg; 1246 } 1247 if (ccg != NULL) 1248 cg = ccg; 1249 cpu = -1; 1250 /* Search the group for the less loaded idle CPU we can run now. */ 1251 mask = td->td_cpuset->cs_mask; 1252 if (cg != NULL && cg != cpu_top && 1253 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0) 1254 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE), 1255 INT_MAX, ts->ts_cpu); 1256 /* Search globally for the less loaded CPU we can run now. */ 1257 if (cpu == -1) 1258 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu); 1259 /* Search globally for the less loaded CPU. */ 1260 if (cpu == -1) 1261 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu); 1262 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1263 /* 1264 * Compare the lowest loaded cpu to current cpu. 1265 */ 1266 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1267 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE && 1268 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) { 1269 SCHED_STAT_INC(pickcpu_local); 1270 cpu = self; 1271 } else 1272 SCHED_STAT_INC(pickcpu_lowest); 1273 if (cpu != ts->ts_cpu) 1274 SCHED_STAT_INC(pickcpu_migration); 1275 return (cpu); 1276} 1277#endif 1278 1279/* 1280 * Pick the highest priority task we have and return it. 1281 */ 1282static struct thread * 1283tdq_choose(struct tdq *tdq) 1284{ 1285 struct thread *td; 1286 1287 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1288 td = runq_choose(&tdq->tdq_realtime); 1289 if (td != NULL) 1290 return (td); 1291 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1292 if (td != NULL) { 1293 KASSERT(td->td_priority >= PRI_MIN_BATCH, 1294 ("tdq_choose: Invalid priority on timeshare queue %d", 1295 td->td_priority)); 1296 return (td); 1297 } 1298 td = runq_choose(&tdq->tdq_idle); 1299 if (td != NULL) { 1300 KASSERT(td->td_priority >= PRI_MIN_IDLE, 1301 ("tdq_choose: Invalid priority on idle queue %d", 1302 td->td_priority)); 1303 return (td); 1304 } 1305 1306 return (NULL); 1307} 1308 1309/* 1310 * Initialize a thread queue. 1311 */ 1312static void 1313tdq_setup(struct tdq *tdq) 1314{ 1315 1316 if (bootverbose) 1317 printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1318 runq_init(&tdq->tdq_realtime); 1319 runq_init(&tdq->tdq_timeshare); 1320 runq_init(&tdq->tdq_idle); 1321 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1322 "sched lock %d", (int)TDQ_ID(tdq)); 1323 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1324 MTX_SPIN | MTX_RECURSE); 1325#ifdef KTR 1326 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname), 1327 "CPU %d load", (int)TDQ_ID(tdq)); 1328#endif 1329} 1330 1331#ifdef SMP 1332static void 1333sched_setup_smp(void) 1334{ 1335 struct tdq *tdq; 1336 int i; 1337 1338 cpu_top = smp_topo(); 1339 CPU_FOREACH(i) { 1340 tdq = TDQ_CPU(i); 1341 tdq_setup(tdq); 1342 tdq->tdq_cg = smp_topo_find(cpu_top, i); 1343 if (tdq->tdq_cg == NULL) 1344 panic("Can't find cpu group for %d\n", i); 1345 } 1346 balance_tdq = TDQ_SELF(); 1347 sched_balance(); 1348} 1349#endif 1350 1351/* 1352 * Setup the thread queues and initialize the topology based on MD 1353 * information. 1354 */ 1355static void 1356sched_setup(void *dummy) 1357{ 1358 struct tdq *tdq; 1359 1360 tdq = TDQ_SELF(); 1361#ifdef SMP 1362 sched_setup_smp(); 1363#else 1364 tdq_setup(tdq); 1365#endif 1366 /* 1367 * To avoid divide-by-zero, we set realstathz a dummy value 1368 * in case which sched_clock() called before sched_initticks(). 1369 */ 1370 realstathz = hz; 1371 sched_slice = (realstathz/10); /* ~100ms */ 1372 tickincr = 1 << SCHED_TICK_SHIFT; 1373 1374 /* Add thread0's load since it's running. */ 1375 TDQ_LOCK(tdq); 1376 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1377 tdq_load_add(tdq, &thread0); 1378 tdq->tdq_lowpri = thread0.td_priority; 1379 TDQ_UNLOCK(tdq); 1380} 1381 1382/* 1383 * This routine determines the tickincr after stathz and hz are setup. 1384 */ 1385/* ARGSUSED */ 1386static void 1387sched_initticks(void *dummy) 1388{ 1389 int incr; 1390 1391 realstathz = stathz ? stathz : hz; 1392 sched_slice = (realstathz/10); /* ~100ms */ 1393 1394 /* 1395 * tickincr is shifted out by 10 to avoid rounding errors due to 1396 * hz not being evenly divisible by stathz on all platforms. 1397 */ 1398 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1399 /* 1400 * This does not work for values of stathz that are more than 1401 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1402 */ 1403 if (incr == 0) 1404 incr = 1; 1405 tickincr = incr; 1406#ifdef SMP 1407 /* 1408 * Set the default balance interval now that we know 1409 * what realstathz is. 1410 */ 1411 balance_interval = realstathz; 1412 /* 1413 * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4. 1414 * This prevents excess thrashing on large machines and excess idle 1415 * on smaller machines. 1416 */ 1417 steal_thresh = min(fls(mp_ncpus) - 1, 3); 1418 affinity = SCHED_AFFINITY_DEFAULT; 1419#endif 1420 if (sched_idlespinthresh < 0) 1421 sched_idlespinthresh = max(16, 2 * hz / realstathz); 1422} 1423 1424 1425/* 1426 * This is the core of the interactivity algorithm. Determines a score based 1427 * on past behavior. It is the ratio of sleep time to run time scaled to 1428 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1429 * differs from the cpu usage because it does not account for time spent 1430 * waiting on a run-queue. Would be prettier if we had floating point. 1431 */ 1432static int 1433sched_interact_score(struct thread *td) 1434{ 1435 struct td_sched *ts; 1436 int div; 1437 1438 ts = td->td_sched; 1439 /* 1440 * The score is only needed if this is likely to be an interactive 1441 * task. Don't go through the expense of computing it if there's 1442 * no chance. 1443 */ 1444 if (sched_interact <= SCHED_INTERACT_HALF && 1445 ts->ts_runtime >= ts->ts_slptime) 1446 return (SCHED_INTERACT_HALF); 1447 1448 if (ts->ts_runtime > ts->ts_slptime) { 1449 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1450 return (SCHED_INTERACT_HALF + 1451 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1452 } 1453 if (ts->ts_slptime > ts->ts_runtime) { 1454 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1455 return (ts->ts_runtime / div); 1456 } 1457 /* runtime == slptime */ 1458 if (ts->ts_runtime) 1459 return (SCHED_INTERACT_HALF); 1460 1461 /* 1462 * This can happen if slptime and runtime are 0. 1463 */ 1464 return (0); 1465 1466} 1467 1468/* 1469 * Scale the scheduling priority according to the "interactivity" of this 1470 * process. 1471 */ 1472static void 1473sched_priority(struct thread *td) 1474{ 1475 int score; 1476 int pri; 1477 1478 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1479 return; 1480 /* 1481 * If the score is interactive we place the thread in the realtime 1482 * queue with a priority that is less than kernel and interrupt 1483 * priorities. These threads are not subject to nice restrictions. 1484 * 1485 * Scores greater than this are placed on the normal timeshare queue 1486 * where the priority is partially decided by the most recent cpu 1487 * utilization and the rest is decided by nice value. 1488 * 1489 * The nice value of the process has a linear effect on the calculated 1490 * score. Negative nice values make it easier for a thread to be 1491 * considered interactive. 1492 */ 1493 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice); 1494 if (score < sched_interact) { 1495 pri = PRI_MIN_INTERACT; 1496 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) / 1497 sched_interact) * score; 1498 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT, 1499 ("sched_priority: invalid interactive priority %d score %d", 1500 pri, score)); 1501 } else { 1502 pri = SCHED_PRI_MIN; 1503 if (td->td_sched->ts_ticks) 1504 pri += min(SCHED_PRI_TICKS(td->td_sched), 1505 SCHED_PRI_RANGE); 1506 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1507 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH, 1508 ("sched_priority: invalid priority %d: nice %d, " 1509 "ticks %d ftick %d ltick %d tick pri %d", 1510 pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1511 td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1512 SCHED_PRI_TICKS(td->td_sched))); 1513 } 1514 sched_user_prio(td, pri); 1515 1516 return; 1517} 1518 1519/* 1520 * This routine enforces a maximum limit on the amount of scheduling history 1521 * kept. It is called after either the slptime or runtime is adjusted. This 1522 * function is ugly due to integer math. 1523 */ 1524static void 1525sched_interact_update(struct thread *td) 1526{ 1527 struct td_sched *ts; 1528 u_int sum; 1529 1530 ts = td->td_sched; 1531 sum = ts->ts_runtime + ts->ts_slptime; 1532 if (sum < SCHED_SLP_RUN_MAX) 1533 return; 1534 /* 1535 * This only happens from two places: 1536 * 1) We have added an unusual amount of run time from fork_exit. 1537 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1538 */ 1539 if (sum > SCHED_SLP_RUN_MAX * 2) { 1540 if (ts->ts_runtime > ts->ts_slptime) { 1541 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1542 ts->ts_slptime = 1; 1543 } else { 1544 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1545 ts->ts_runtime = 1; 1546 } 1547 return; 1548 } 1549 /* 1550 * If we have exceeded by more than 1/5th then the algorithm below 1551 * will not bring us back into range. Dividing by two here forces 1552 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1553 */ 1554 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1555 ts->ts_runtime /= 2; 1556 ts->ts_slptime /= 2; 1557 return; 1558 } 1559 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1560 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1561} 1562 1563/* 1564 * Scale back the interactivity history when a child thread is created. The 1565 * history is inherited from the parent but the thread may behave totally 1566 * differently. For example, a shell spawning a compiler process. We want 1567 * to learn that the compiler is behaving badly very quickly. 1568 */ 1569static void 1570sched_interact_fork(struct thread *td) 1571{ 1572 int ratio; 1573 int sum; 1574 1575 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1576 if (sum > SCHED_SLP_RUN_FORK) { 1577 ratio = sum / SCHED_SLP_RUN_FORK; 1578 td->td_sched->ts_runtime /= ratio; 1579 td->td_sched->ts_slptime /= ratio; 1580 } 1581} 1582 1583/* 1584 * Called from proc0_init() to setup the scheduler fields. 1585 */ 1586void 1587schedinit(void) 1588{ 1589 1590 /* 1591 * Set up the scheduler specific parts of proc0. 1592 */ 1593 proc0.p_sched = NULL; /* XXX */ 1594 thread0.td_sched = &td_sched0; 1595 td_sched0.ts_ltick = ticks; 1596 td_sched0.ts_ftick = ticks; 1597 td_sched0.ts_slice = sched_slice; 1598} 1599 1600/* 1601 * This is only somewhat accurate since given many processes of the same 1602 * priority they will switch when their slices run out, which will be 1603 * at most sched_slice stathz ticks. 1604 */ 1605int 1606sched_rr_interval(void) 1607{ 1608 1609 /* Convert sched_slice to hz */ 1610 return (hz/(realstathz/sched_slice)); 1611} 1612 1613/* 1614 * Update the percent cpu tracking information when it is requested or 1615 * the total history exceeds the maximum. We keep a sliding history of 1616 * tick counts that slowly decays. This is less precise than the 4BSD 1617 * mechanism since it happens with less regular and frequent events. 1618 */ 1619static void 1620sched_pctcpu_update(struct td_sched *ts, int run) 1621{ 1622 int t = ticks; 1623 1624 if (t - ts->ts_ltick >= SCHED_TICK_TARG) { 1625 ts->ts_ticks = 0; 1626 ts->ts_ftick = t - SCHED_TICK_TARG; 1627 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) { 1628 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) * 1629 (ts->ts_ltick - (t - SCHED_TICK_TARG)); 1630 ts->ts_ftick = t - SCHED_TICK_TARG; 1631 } 1632 if (run) 1633 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT; 1634 ts->ts_ltick = t; 1635} 1636 1637/* 1638 * Adjust the priority of a thread. Move it to the appropriate run-queue 1639 * if necessary. This is the back-end for several priority related 1640 * functions. 1641 */ 1642static void 1643sched_thread_priority(struct thread *td, u_char prio) 1644{ 1645 struct td_sched *ts; 1646 struct tdq *tdq; 1647 int oldpri; 1648 1649 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio", 1650 "prio:%d", td->td_priority, "new prio:%d", prio, 1651 KTR_ATTR_LINKED, sched_tdname(curthread)); 1652 SDT_PROBE3(sched, , , change_pri, td, td->td_proc, prio); 1653 if (td != curthread && prio < td->td_priority) { 1654 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 1655 "lend prio", "prio:%d", td->td_priority, "new prio:%d", 1656 prio, KTR_ATTR_LINKED, sched_tdname(td)); 1657 SDT_PROBE4(sched, , , lend_pri, td, td->td_proc, prio, 1658 curthread); 1659 } 1660 ts = td->td_sched; 1661 THREAD_LOCK_ASSERT(td, MA_OWNED); 1662 if (td->td_priority == prio) 1663 return; 1664 /* 1665 * If the priority has been elevated due to priority 1666 * propagation, we may have to move ourselves to a new 1667 * queue. This could be optimized to not re-add in some 1668 * cases. 1669 */ 1670 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1671 sched_rem(td); 1672 td->td_priority = prio; 1673 sched_add(td, SRQ_BORROWING); 1674 return; 1675 } 1676 /* 1677 * If the thread is currently running we may have to adjust the lowpri 1678 * information so other cpus are aware of our current priority. 1679 */ 1680 if (TD_IS_RUNNING(td)) { 1681 tdq = TDQ_CPU(ts->ts_cpu); 1682 oldpri = td->td_priority; 1683 td->td_priority = prio; 1684 if (prio < tdq->tdq_lowpri) 1685 tdq->tdq_lowpri = prio; 1686 else if (tdq->tdq_lowpri == oldpri) 1687 tdq_setlowpri(tdq, td); 1688 return; 1689 } 1690 td->td_priority = prio; 1691} 1692 1693/* 1694 * Update a thread's priority when it is lent another thread's 1695 * priority. 1696 */ 1697void 1698sched_lend_prio(struct thread *td, u_char prio) 1699{ 1700 1701 td->td_flags |= TDF_BORROWING; 1702 sched_thread_priority(td, prio); 1703} 1704 1705/* 1706 * Restore a thread's priority when priority propagation is 1707 * over. The prio argument is the minimum priority the thread 1708 * needs to have to satisfy other possible priority lending 1709 * requests. If the thread's regular priority is less 1710 * important than prio, the thread will keep a priority boost 1711 * of prio. 1712 */ 1713void 1714sched_unlend_prio(struct thread *td, u_char prio) 1715{ 1716 u_char base_pri; 1717 1718 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1719 td->td_base_pri <= PRI_MAX_TIMESHARE) 1720 base_pri = td->td_user_pri; 1721 else 1722 base_pri = td->td_base_pri; 1723 if (prio >= base_pri) { 1724 td->td_flags &= ~TDF_BORROWING; 1725 sched_thread_priority(td, base_pri); 1726 } else 1727 sched_lend_prio(td, prio); 1728} 1729 1730/* 1731 * Standard entry for setting the priority to an absolute value. 1732 */ 1733void 1734sched_prio(struct thread *td, u_char prio) 1735{ 1736 u_char oldprio; 1737 1738 /* First, update the base priority. */ 1739 td->td_base_pri = prio; 1740 1741 /* 1742 * If the thread is borrowing another thread's priority, don't 1743 * ever lower the priority. 1744 */ 1745 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1746 return; 1747 1748 /* Change the real priority. */ 1749 oldprio = td->td_priority; 1750 sched_thread_priority(td, prio); 1751 1752 /* 1753 * If the thread is on a turnstile, then let the turnstile update 1754 * its state. 1755 */ 1756 if (TD_ON_LOCK(td) && oldprio != prio) 1757 turnstile_adjust(td, oldprio); 1758} 1759 1760/* 1761 * Set the base user priority, does not effect current running priority. 1762 */ 1763void 1764sched_user_prio(struct thread *td, u_char prio) 1765{ 1766 1767 td->td_base_user_pri = prio; 1768 if (td->td_lend_user_pri <= prio) 1769 return; 1770 td->td_user_pri = prio; 1771} 1772 1773void 1774sched_lend_user_prio(struct thread *td, u_char prio) 1775{ 1776 1777 THREAD_LOCK_ASSERT(td, MA_OWNED); 1778 td->td_lend_user_pri = prio; 1779 td->td_user_pri = min(prio, td->td_base_user_pri); 1780 if (td->td_priority > td->td_user_pri) 1781 sched_prio(td, td->td_user_pri); 1782 else if (td->td_priority != td->td_user_pri) 1783 td->td_flags |= TDF_NEEDRESCHED; 1784} 1785 1786/* 1787 * Handle migration from sched_switch(). This happens only for 1788 * cpu binding. 1789 */ 1790static struct mtx * 1791sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1792{ 1793 struct tdq *tdn; 1794 1795 tdn = TDQ_CPU(td->td_sched->ts_cpu); 1796#ifdef SMP 1797 tdq_load_rem(tdq, td); 1798 /* 1799 * Do the lock dance required to avoid LOR. We grab an extra 1800 * spinlock nesting to prevent preemption while we're 1801 * not holding either run-queue lock. 1802 */ 1803 spinlock_enter(); 1804 thread_lock_block(td); /* This releases the lock on tdq. */ 1805 1806 /* 1807 * Acquire both run-queue locks before placing the thread on the new 1808 * run-queue to avoid deadlocks created by placing a thread with a 1809 * blocked lock on the run-queue of a remote processor. The deadlock 1810 * occurs when a third processor attempts to lock the two queues in 1811 * question while the target processor is spinning with its own 1812 * run-queue lock held while waiting for the blocked lock to clear. 1813 */ 1814 tdq_lock_pair(tdn, tdq); 1815 tdq_add(tdn, td, flags); 1816 tdq_notify(tdn, td); 1817 TDQ_UNLOCK(tdn); 1818 spinlock_exit(); 1819#endif 1820 return (TDQ_LOCKPTR(tdn)); 1821} 1822 1823/* 1824 * Variadic version of thread_lock_unblock() that does not assume td_lock 1825 * is blocked. 1826 */ 1827static inline void 1828thread_unblock_switch(struct thread *td, struct mtx *mtx) 1829{ 1830 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1831 (uintptr_t)mtx); 1832} 1833 1834/* 1835 * Switch threads. This function has to handle threads coming in while 1836 * blocked for some reason, running, or idle. It also must deal with 1837 * migrating a thread from one queue to another as running threads may 1838 * be assigned elsewhere via binding. 1839 */ 1840void 1841sched_switch(struct thread *td, struct thread *newtd, int flags) 1842{ 1843 struct tdq *tdq; 1844 struct td_sched *ts; 1845 struct mtx *mtx; 1846 int srqflag; 1847 int cpuid, preempted; 1848 1849 THREAD_LOCK_ASSERT(td, MA_OWNED); 1850 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1851 1852 cpuid = PCPU_GET(cpuid); 1853 tdq = TDQ_CPU(cpuid); 1854 ts = td->td_sched; 1855 mtx = td->td_lock; 1856 sched_pctcpu_update(ts, 1); 1857 ts->ts_rltick = ticks; 1858 td->td_lastcpu = td->td_oncpu; 1859 td->td_oncpu = NOCPU; 1860 preempted = !(td->td_flags & TDF_SLICEEND); 1861 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND); 1862 td->td_owepreempt = 0; 1863 tdq->tdq_switchcnt++; 1864 /* 1865 * The lock pointer in an idle thread should never change. Reset it 1866 * to CAN_RUN as well. 1867 */ 1868 if (TD_IS_IDLETHREAD(td)) { 1869 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1870 TD_SET_CAN_RUN(td); 1871 } else if (TD_IS_RUNNING(td)) { 1872 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1873 srqflag = preempted ? 1874 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1875 SRQ_OURSELF|SRQ_YIELDING; 1876#ifdef SMP 1877 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu)) 1878 ts->ts_cpu = sched_pickcpu(td, 0); 1879#endif 1880 if (ts->ts_cpu == cpuid) 1881 tdq_runq_add(tdq, td, srqflag); 1882 else { 1883 KASSERT(THREAD_CAN_MIGRATE(td) || 1884 (ts->ts_flags & TSF_BOUND) != 0, 1885 ("Thread %p shouldn't migrate", td)); 1886 mtx = sched_switch_migrate(tdq, td, srqflag); 1887 } 1888 } else { 1889 /* This thread must be going to sleep. */ 1890 TDQ_LOCK(tdq); 1891 mtx = thread_lock_block(td); 1892 tdq_load_rem(tdq, td); 1893 } 1894 /* 1895 * We enter here with the thread blocked and assigned to the 1896 * appropriate cpu run-queue or sleep-queue and with the current 1897 * thread-queue locked. 1898 */ 1899 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1900 newtd = choosethread(); 1901 /* 1902 * Call the MD code to switch contexts if necessary. 1903 */ 1904 if (td != newtd) { 1905#ifdef HWPMC_HOOKS 1906 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1907 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1908#endif 1909 SDT_PROBE2(sched, , , off_cpu, td, td->td_proc); 1910 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1911 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1912 sched_pctcpu_update(newtd->td_sched, 0); 1913 1914#ifdef KDTRACE_HOOKS 1915 /* 1916 * If DTrace has set the active vtime enum to anything 1917 * other than INACTIVE (0), then it should have set the 1918 * function to call. 1919 */ 1920 if (dtrace_vtime_active) 1921 (*dtrace_vtime_switch_func)(newtd); 1922#endif 1923 1924 cpu_switch(td, newtd, mtx); 1925 /* 1926 * We may return from cpu_switch on a different cpu. However, 1927 * we always return with td_lock pointing to the current cpu's 1928 * run queue lock. 1929 */ 1930 cpuid = PCPU_GET(cpuid); 1931 tdq = TDQ_CPU(cpuid); 1932 lock_profile_obtain_lock_success( 1933 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1934 1935 SDT_PROBE0(sched, , , on_cpu); 1936#ifdef HWPMC_HOOKS 1937 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1938 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1939#endif 1940 } else { 1941 thread_unblock_switch(td, mtx); 1942 SDT_PROBE0(sched, , , remain_cpu); 1943 } 1944 /* 1945 * Assert that all went well and return. 1946 */ 1947 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1948 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1949 td->td_oncpu = cpuid; 1950} 1951 1952/* 1953 * Adjust thread priorities as a result of a nice request. 1954 */ 1955void 1956sched_nice(struct proc *p, int nice) 1957{ 1958 struct thread *td; 1959 1960 PROC_LOCK_ASSERT(p, MA_OWNED); 1961 1962 p->p_nice = nice; 1963 FOREACH_THREAD_IN_PROC(p, td) { 1964 thread_lock(td); 1965 sched_priority(td); 1966 sched_prio(td, td->td_base_user_pri); 1967 thread_unlock(td); 1968 } 1969} 1970 1971/* 1972 * Record the sleep time for the interactivity scorer. 1973 */ 1974void 1975sched_sleep(struct thread *td, int prio) 1976{ 1977 1978 THREAD_LOCK_ASSERT(td, MA_OWNED); 1979 1980 td->td_slptick = ticks; 1981 if (TD_IS_SUSPENDED(td) || prio >= PSOCK) 1982 td->td_flags |= TDF_CANSWAP; 1983 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1984 return; 1985 if (static_boost == 1 && prio) 1986 sched_prio(td, prio); 1987 else if (static_boost && td->td_priority > static_boost) 1988 sched_prio(td, static_boost); 1989} 1990 1991/* 1992 * Schedule a thread to resume execution and record how long it voluntarily 1993 * slept. We also update the pctcpu, interactivity, and priority. 1994 */ 1995void 1996sched_wakeup(struct thread *td) 1997{ 1998 struct td_sched *ts; 1999 int slptick; 2000 2001 THREAD_LOCK_ASSERT(td, MA_OWNED); 2002 ts = td->td_sched; 2003 td->td_flags &= ~TDF_CANSWAP; 2004 /* 2005 * If we slept for more than a tick update our interactivity and 2006 * priority. 2007 */ 2008 slptick = td->td_slptick; 2009 td->td_slptick = 0; 2010 if (slptick && slptick != ticks) { 2011 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT; 2012 sched_interact_update(td); 2013 sched_pctcpu_update(ts, 0); 2014 } 2015 /* Reset the slice value after we sleep. */ 2016 ts->ts_slice = sched_slice; 2017 sched_add(td, SRQ_BORING); 2018} 2019 2020/* 2021 * Penalize the parent for creating a new child and initialize the child's 2022 * priority. 2023 */ 2024void 2025sched_fork(struct thread *td, struct thread *child) 2026{ 2027 THREAD_LOCK_ASSERT(td, MA_OWNED); 2028 sched_pctcpu_update(td->td_sched, 1); 2029 sched_fork_thread(td, child); 2030 /* 2031 * Penalize the parent and child for forking. 2032 */ 2033 sched_interact_fork(child); 2034 sched_priority(child); 2035 td->td_sched->ts_runtime += tickincr; 2036 sched_interact_update(td); 2037 sched_priority(td); 2038} 2039 2040/* 2041 * Fork a new thread, may be within the same process. 2042 */ 2043void 2044sched_fork_thread(struct thread *td, struct thread *child) 2045{ 2046 struct td_sched *ts; 2047 struct td_sched *ts2; 2048 2049 THREAD_LOCK_ASSERT(td, MA_OWNED); 2050 /* 2051 * Initialize child. 2052 */ 2053 ts = td->td_sched; 2054 ts2 = child->td_sched; 2055 child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 2056 child->td_cpuset = cpuset_ref(td->td_cpuset); 2057 ts2->ts_cpu = ts->ts_cpu; 2058 ts2->ts_flags = 0; 2059 /* 2060 * Grab our parents cpu estimation information. 2061 */ 2062 ts2->ts_ticks = ts->ts_ticks; 2063 ts2->ts_ltick = ts->ts_ltick; 2064 ts2->ts_ftick = ts->ts_ftick; 2065 /* 2066 * Do not inherit any borrowed priority from the parent. 2067 */ 2068 child->td_priority = child->td_base_pri; 2069 /* 2070 * And update interactivity score. 2071 */ 2072 ts2->ts_slptime = ts->ts_slptime; 2073 ts2->ts_runtime = ts->ts_runtime; 2074 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 2075#ifdef KTR 2076 bzero(ts2->ts_name, sizeof(ts2->ts_name)); 2077#endif 2078} 2079 2080/* 2081 * Adjust the priority class of a thread. 2082 */ 2083void 2084sched_class(struct thread *td, int class) 2085{ 2086 2087 THREAD_LOCK_ASSERT(td, MA_OWNED); 2088 if (td->td_pri_class == class) 2089 return; 2090 td->td_pri_class = class; 2091} 2092 2093/* 2094 * Return some of the child's priority and interactivity to the parent. 2095 */ 2096void 2097sched_exit(struct proc *p, struct thread *child) 2098{ 2099 struct thread *td; 2100 2101 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit", 2102 "prio:%d", child->td_priority); 2103 PROC_LOCK_ASSERT(p, MA_OWNED); 2104 td = FIRST_THREAD_IN_PROC(p); 2105 sched_exit_thread(td, child); 2106} 2107 2108/* 2109 * Penalize another thread for the time spent on this one. This helps to 2110 * worsen the priority and interactivity of processes which schedule batch 2111 * jobs such as make. This has little effect on the make process itself but 2112 * causes new processes spawned by it to receive worse scores immediately. 2113 */ 2114void 2115sched_exit_thread(struct thread *td, struct thread *child) 2116{ 2117 2118 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit", 2119 "prio:%d", child->td_priority); 2120 /* 2121 * Give the child's runtime to the parent without returning the 2122 * sleep time as a penalty to the parent. This causes shells that 2123 * launch expensive things to mark their children as expensive. 2124 */ 2125 thread_lock(td); 2126 td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2127 sched_interact_update(td); 2128 sched_priority(td); 2129 thread_unlock(td); 2130} 2131 2132void 2133sched_preempt(struct thread *td) 2134{ 2135 struct tdq *tdq; 2136 2137 SDT_PROBE2(sched, , , surrender, td, td->td_proc); 2138 2139 thread_lock(td); 2140 tdq = TDQ_SELF(); 2141 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2142 tdq->tdq_ipipending = 0; 2143 if (td->td_priority > tdq->tdq_lowpri) { 2144 int flags; 2145 2146 flags = SW_INVOL | SW_PREEMPT; 2147 if (td->td_critnest > 1) 2148 td->td_owepreempt = 1; 2149 else if (TD_IS_IDLETHREAD(td)) 2150 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL); 2151 else 2152 mi_switch(flags | SWT_REMOTEPREEMPT, NULL); 2153 } 2154 thread_unlock(td); 2155} 2156 2157/* 2158 * Fix priorities on return to user-space. Priorities may be elevated due 2159 * to static priorities in msleep() or similar. 2160 */ 2161void 2162sched_userret(struct thread *td) 2163{ 2164 /* 2165 * XXX we cheat slightly on the locking here to avoid locking in 2166 * the usual case. Setting td_priority here is essentially an 2167 * incomplete workaround for not setting it properly elsewhere. 2168 * Now that some interrupt handlers are threads, not setting it 2169 * properly elsewhere can clobber it in the window between setting 2170 * it here and returning to user mode, so don't waste time setting 2171 * it perfectly here. 2172 */ 2173 KASSERT((td->td_flags & TDF_BORROWING) == 0, 2174 ("thread with borrowed priority returning to userland")); 2175 if (td->td_priority != td->td_user_pri) { 2176 thread_lock(td); 2177 td->td_priority = td->td_user_pri; 2178 td->td_base_pri = td->td_user_pri; 2179 tdq_setlowpri(TDQ_SELF(), td); 2180 thread_unlock(td); 2181 } 2182} 2183 2184/* 2185 * Handle a stathz tick. This is really only relevant for timeshare 2186 * threads. 2187 */ 2188void 2189sched_clock(struct thread *td) 2190{ 2191 struct tdq *tdq; 2192 struct td_sched *ts; 2193 2194 THREAD_LOCK_ASSERT(td, MA_OWNED); 2195 tdq = TDQ_SELF(); 2196#ifdef SMP 2197 /* 2198 * We run the long term load balancer infrequently on the first cpu. 2199 */ 2200 if (balance_tdq == tdq) { 2201 if (balance_ticks && --balance_ticks == 0) 2202 sched_balance(); 2203 } 2204#endif 2205 /* 2206 * Save the old switch count so we have a record of the last ticks 2207 * activity. Initialize the new switch count based on our load. 2208 * If there is some activity seed it to reflect that. 2209 */ 2210 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2211 tdq->tdq_switchcnt = tdq->tdq_load; 2212 /* 2213 * Advance the insert index once for each tick to ensure that all 2214 * threads get a chance to run. 2215 */ 2216 if (tdq->tdq_idx == tdq->tdq_ridx) { 2217 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2218 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2219 tdq->tdq_ridx = tdq->tdq_idx; 2220 } 2221 ts = td->td_sched; 2222 sched_pctcpu_update(ts, 1); 2223 if (td->td_pri_class & PRI_FIFO_BIT) 2224 return; 2225 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) { 2226 /* 2227 * We used a tick; charge it to the thread so 2228 * that we can compute our interactivity. 2229 */ 2230 td->td_sched->ts_runtime += tickincr; 2231 sched_interact_update(td); 2232 sched_priority(td); 2233 } 2234 /* 2235 * We used up one time slice. 2236 */ 2237 if (--ts->ts_slice > 0) 2238 return; 2239 /* 2240 * We're out of time, force a requeue at userret(). 2241 */ 2242 ts->ts_slice = sched_slice; 2243 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND; 2244} 2245 2246/* 2247 * Called once per hz tick. 2248 */ 2249void 2250sched_tick(int cnt) 2251{ 2252 2253} 2254 2255/* 2256 * Return whether the current CPU has runnable tasks. Used for in-kernel 2257 * cooperative idle threads. 2258 */ 2259int 2260sched_runnable(void) 2261{ 2262 struct tdq *tdq; 2263 int load; 2264 2265 load = 1; 2266 2267 tdq = TDQ_SELF(); 2268 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2269 if (tdq->tdq_load > 0) 2270 goto out; 2271 } else 2272 if (tdq->tdq_load - 1 > 0) 2273 goto out; 2274 load = 0; 2275out: 2276 return (load); 2277} 2278 2279/* 2280 * Choose the highest priority thread to run. The thread is removed from 2281 * the run-queue while running however the load remains. For SMP we set 2282 * the tdq in the global idle bitmask if it idles here. 2283 */ 2284struct thread * 2285sched_choose(void) 2286{ 2287 struct thread *td; 2288 struct tdq *tdq; 2289 2290 tdq = TDQ_SELF(); 2291 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2292 td = tdq_choose(tdq); 2293 if (td) { 2294 tdq_runq_rem(tdq, td); 2295 tdq->tdq_lowpri = td->td_priority; 2296 return (td); 2297 } 2298 tdq->tdq_lowpri = PRI_MAX_IDLE; 2299 return (PCPU_GET(idlethread)); 2300} 2301 2302/* 2303 * Set owepreempt if necessary. Preemption never happens directly in ULE, 2304 * we always request it once we exit a critical section. 2305 */ 2306static inline void 2307sched_setpreempt(struct thread *td) 2308{ 2309 struct thread *ctd; 2310 int cpri; 2311 int pri; 2312 2313 THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2314 2315 ctd = curthread; 2316 pri = td->td_priority; 2317 cpri = ctd->td_priority; 2318 if (pri < cpri) 2319 ctd->td_flags |= TDF_NEEDRESCHED; 2320 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2321 return; 2322 if (!sched_shouldpreempt(pri, cpri, 0)) 2323 return; 2324 ctd->td_owepreempt = 1; 2325} 2326 2327/* 2328 * Add a thread to a thread queue. Select the appropriate runq and add the 2329 * thread to it. This is the internal function called when the tdq is 2330 * predetermined. 2331 */ 2332void 2333tdq_add(struct tdq *tdq, struct thread *td, int flags) 2334{ 2335 2336 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2337 KASSERT((td->td_inhibitors == 0), 2338 ("sched_add: trying to run inhibited thread")); 2339 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2340 ("sched_add: bad thread state")); 2341 KASSERT(td->td_flags & TDF_INMEM, 2342 ("sched_add: thread swapped out")); 2343 2344 if (td->td_priority < tdq->tdq_lowpri) 2345 tdq->tdq_lowpri = td->td_priority; 2346 tdq_runq_add(tdq, td, flags); 2347 tdq_load_add(tdq, td); 2348} 2349 2350/* 2351 * Select the target thread queue and add a thread to it. Request 2352 * preemption or IPI a remote processor if required. 2353 */ 2354void 2355sched_add(struct thread *td, int flags) 2356{ 2357 struct tdq *tdq; 2358#ifdef SMP 2359 int cpu; 2360#endif 2361 2362 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 2363 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 2364 sched_tdname(curthread)); 2365 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 2366 KTR_ATTR_LINKED, sched_tdname(td)); 2367 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 2368 flags & SRQ_PREEMPTED); 2369 THREAD_LOCK_ASSERT(td, MA_OWNED); 2370 /* 2371 * Recalculate the priority before we select the target cpu or 2372 * run-queue. 2373 */ 2374 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2375 sched_priority(td); 2376#ifdef SMP 2377 /* 2378 * Pick the destination cpu and if it isn't ours transfer to the 2379 * target cpu. 2380 */ 2381 cpu = sched_pickcpu(td, flags); 2382 tdq = sched_setcpu(td, cpu, flags); 2383 tdq_add(tdq, td, flags); 2384 if (cpu != PCPU_GET(cpuid)) { 2385 tdq_notify(tdq, td); 2386 return; 2387 } 2388#else 2389 tdq = TDQ_SELF(); 2390 TDQ_LOCK(tdq); 2391 /* 2392 * Now that the thread is moving to the run-queue, set the lock 2393 * to the scheduler's lock. 2394 */ 2395 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2396 tdq_add(tdq, td, flags); 2397#endif 2398 if (!(flags & SRQ_YIELDING)) 2399 sched_setpreempt(td); 2400} 2401 2402/* 2403 * Remove a thread from a run-queue without running it. This is used 2404 * when we're stealing a thread from a remote queue. Otherwise all threads 2405 * exit by calling sched_exit_thread() and sched_throw() themselves. 2406 */ 2407void 2408sched_rem(struct thread *td) 2409{ 2410 struct tdq *tdq; 2411 2412 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 2413 "prio:%d", td->td_priority); 2414 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); 2415 tdq = TDQ_CPU(td->td_sched->ts_cpu); 2416 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2417 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2418 KASSERT(TD_ON_RUNQ(td), 2419 ("sched_rem: thread not on run queue")); 2420 tdq_runq_rem(tdq, td); 2421 tdq_load_rem(tdq, td); 2422 TD_SET_CAN_RUN(td); 2423 if (td->td_priority == tdq->tdq_lowpri) 2424 tdq_setlowpri(tdq, NULL); 2425} 2426 2427/* 2428 * Fetch cpu utilization information. Updates on demand. 2429 */ 2430fixpt_t 2431sched_pctcpu(struct thread *td) 2432{ 2433 fixpt_t pctcpu; 2434 struct td_sched *ts; 2435 2436 pctcpu = 0; 2437 ts = td->td_sched; 2438 if (ts == NULL) 2439 return (0); 2440 2441 THREAD_LOCK_ASSERT(td, MA_OWNED); 2442 sched_pctcpu_update(ts, TD_IS_RUNNING(td)); 2443 if (ts->ts_ticks) { 2444 int rtick; 2445 2446 /* How many rtick per second ? */ 2447 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2448 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2449 } 2450 2451 return (pctcpu); 2452} 2453 2454/* 2455 * Enforce affinity settings for a thread. Called after adjustments to 2456 * cpumask. 2457 */ 2458void 2459sched_affinity(struct thread *td) 2460{ 2461#ifdef SMP 2462 struct td_sched *ts; 2463 2464 THREAD_LOCK_ASSERT(td, MA_OWNED); 2465 ts = td->td_sched; 2466 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2467 return; 2468 if (TD_ON_RUNQ(td)) { 2469 sched_rem(td); 2470 sched_add(td, SRQ_BORING); 2471 return; 2472 } 2473 if (!TD_IS_RUNNING(td)) 2474 return; 2475 /* 2476 * Force a switch before returning to userspace. If the 2477 * target thread is not running locally send an ipi to force 2478 * the issue. 2479 */ 2480 td->td_flags |= TDF_NEEDRESCHED; 2481 if (td != curthread) 2482 ipi_cpu(ts->ts_cpu, IPI_PREEMPT); 2483#endif 2484} 2485 2486/* 2487 * Bind a thread to a target cpu. 2488 */ 2489void 2490sched_bind(struct thread *td, int cpu) 2491{ 2492 struct td_sched *ts; 2493 2494 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2495 KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 2496 ts = td->td_sched; 2497 if (ts->ts_flags & TSF_BOUND) 2498 sched_unbind(td); 2499 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td)); 2500 ts->ts_flags |= TSF_BOUND; 2501 sched_pin(); 2502 if (PCPU_GET(cpuid) == cpu) 2503 return; 2504 ts->ts_cpu = cpu; 2505 /* When we return from mi_switch we'll be on the correct cpu. */ 2506 mi_switch(SW_VOL, NULL); 2507} 2508 2509/* 2510 * Release a bound thread. 2511 */ 2512void 2513sched_unbind(struct thread *td) 2514{ 2515 struct td_sched *ts; 2516 2517 THREAD_LOCK_ASSERT(td, MA_OWNED); 2518 KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 2519 ts = td->td_sched; 2520 if ((ts->ts_flags & TSF_BOUND) == 0) 2521 return; 2522 ts->ts_flags &= ~TSF_BOUND; 2523 sched_unpin(); 2524} 2525 2526int 2527sched_is_bound(struct thread *td) 2528{ 2529 THREAD_LOCK_ASSERT(td, MA_OWNED); 2530 return (td->td_sched->ts_flags & TSF_BOUND); 2531} 2532 2533/* 2534 * Basic yield call. 2535 */ 2536void 2537sched_relinquish(struct thread *td) 2538{ 2539 thread_lock(td); 2540 mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 2541 thread_unlock(td); 2542} 2543 2544/* 2545 * Return the total system load. 2546 */ 2547int 2548sched_load(void) 2549{ 2550#ifdef SMP 2551 int total; 2552 int i; 2553 2554 total = 0; 2555 CPU_FOREACH(i) 2556 total += TDQ_CPU(i)->tdq_sysload; 2557 return (total); 2558#else 2559 return (TDQ_SELF()->tdq_sysload); 2560#endif 2561} 2562 2563int 2564sched_sizeof_proc(void) 2565{ 2566 return (sizeof(struct proc)); 2567} 2568 2569int 2570sched_sizeof_thread(void) 2571{ 2572 return (sizeof(struct thread) + sizeof(struct td_sched)); 2573} 2574 2575#ifdef SMP 2576#define TDQ_IDLESPIN(tdq) \ 2577 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0) 2578#else 2579#define TDQ_IDLESPIN(tdq) 1 2580#endif 2581 2582/* 2583 * The actual idle process. 2584 */ 2585void 2586sched_idletd(void *dummy) 2587{ 2588 struct thread *td; 2589 struct tdq *tdq; 2590 int switchcnt; 2591 int i; 2592 2593 mtx_assert(&Giant, MA_NOTOWNED); 2594 td = curthread; 2595 tdq = TDQ_SELF(); 2596 for (;;) { 2597#ifdef SMP 2598 if (tdq_idled(tdq) == 0) 2599 continue; 2600#endif 2601 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2602 /* 2603 * If we're switching very frequently, spin while checking 2604 * for load rather than entering a low power state that 2605 * may require an IPI. However, don't do any busy 2606 * loops while on SMT machines as this simply steals 2607 * cycles from cores doing useful work. 2608 */ 2609 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) { 2610 for (i = 0; i < sched_idlespins; i++) { 2611 if (tdq->tdq_load) 2612 break; 2613 cpu_spinwait(); 2614 } 2615 } 2616 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2617 if (tdq->tdq_load == 0) { 2618 tdq->tdq_cpu_idle = 1; 2619 if (tdq->tdq_load == 0) { 2620 cpu_idle(switchcnt > sched_idlespinthresh * 4); 2621 tdq->tdq_switchcnt++; 2622 } 2623 tdq->tdq_cpu_idle = 0; 2624 } 2625 if (tdq->tdq_load) { 2626 thread_lock(td); 2627 mi_switch(SW_VOL | SWT_IDLE, NULL); 2628 thread_unlock(td); 2629 } 2630 } 2631} 2632 2633/* 2634 * A CPU is entering for the first time or a thread is exiting. 2635 */ 2636void 2637sched_throw(struct thread *td) 2638{ 2639 struct thread *newtd; 2640 struct tdq *tdq; 2641 2642 tdq = TDQ_SELF(); 2643 if (td == NULL) { 2644 /* Correct spinlock nesting and acquire the correct lock. */ 2645 TDQ_LOCK(tdq); 2646 spinlock_exit(); 2647 PCPU_SET(switchtime, cpu_ticks()); 2648 PCPU_SET(switchticks, ticks); 2649 } else { 2650 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2651 tdq_load_rem(tdq, td); 2652 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2653 } 2654 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2655 newtd = choosethread(); 2656 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2657 cpu_throw(td, newtd); /* doesn't return */ 2658} 2659 2660/* 2661 * This is called from fork_exit(). Just acquire the correct locks and 2662 * let fork do the rest of the work. 2663 */ 2664void 2665sched_fork_exit(struct thread *td) 2666{ 2667 struct td_sched *ts; 2668 struct tdq *tdq; 2669 int cpuid; 2670 2671 /* 2672 * Finish setting up thread glue so that it begins execution in a 2673 * non-nested critical section with the scheduler lock held. 2674 */ 2675 cpuid = PCPU_GET(cpuid); 2676 tdq = TDQ_CPU(cpuid); 2677 ts = td->td_sched; 2678 if (TD_IS_IDLETHREAD(td)) 2679 td->td_lock = TDQ_LOCKPTR(tdq); 2680 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2681 td->td_oncpu = cpuid; 2682 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2683 lock_profile_obtain_lock_success( 2684 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2685} 2686 2687/* 2688 * Create on first use to catch odd startup conditons. 2689 */ 2690char * 2691sched_tdname(struct thread *td) 2692{ 2693#ifdef KTR 2694 struct td_sched *ts; 2695 2696 ts = td->td_sched; 2697 if (ts->ts_name[0] == '\0') 2698 snprintf(ts->ts_name, sizeof(ts->ts_name), 2699 "%s tid %d", td->td_name, td->td_tid); 2700 return (ts->ts_name); 2701#else 2702 return (td->td_name); 2703#endif 2704} 2705 2706#ifdef KTR 2707void 2708sched_clear_tdname(struct thread *td) 2709{ 2710 struct td_sched *ts; 2711 2712 ts = td->td_sched; 2713 ts->ts_name[0] = '\0'; 2714} 2715#endif 2716 2717#ifdef SMP 2718 2719/* 2720 * Build the CPU topology dump string. Is recursively called to collect 2721 * the topology tree. 2722 */ 2723static int 2724sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg, 2725 int indent) 2726{ 2727 char cpusetbuf[CPUSETBUFSIZ]; 2728 int i, first; 2729 2730 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent, 2731 "", 1 + indent / 2, cg->cg_level); 2732 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "", 2733 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask)); 2734 first = TRUE; 2735 for (i = 0; i < MAXCPU; i++) { 2736 if (CPU_ISSET(i, &cg->cg_mask)) { 2737 if (!first) 2738 sbuf_printf(sb, ", "); 2739 else 2740 first = FALSE; 2741 sbuf_printf(sb, "%d", i); 2742 } 2743 } 2744 sbuf_printf(sb, "</cpu>\n"); 2745 2746 if (cg->cg_flags != 0) { 2747 sbuf_printf(sb, "%*s <flags>", indent, ""); 2748 if ((cg->cg_flags & CG_FLAG_HTT) != 0) 2749 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>"); 2750 if ((cg->cg_flags & CG_FLAG_THREAD) != 0) 2751 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>"); 2752 if ((cg->cg_flags & CG_FLAG_SMT) != 0) 2753 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>"); 2754 sbuf_printf(sb, "</flags>\n"); 2755 } 2756 2757 if (cg->cg_children > 0) { 2758 sbuf_printf(sb, "%*s <children>\n", indent, ""); 2759 for (i = 0; i < cg->cg_children; i++) 2760 sysctl_kern_sched_topology_spec_internal(sb, 2761 &cg->cg_child[i], indent+2); 2762 sbuf_printf(sb, "%*s </children>\n", indent, ""); 2763 } 2764 sbuf_printf(sb, "%*s</group>\n", indent, ""); 2765 return (0); 2766} 2767 2768/* 2769 * Sysctl handler for retrieving topology dump. It's a wrapper for 2770 * the recursive sysctl_kern_smp_topology_spec_internal(). 2771 */ 2772static int 2773sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS) 2774{ 2775 struct sbuf *topo; 2776 int err; 2777 2778 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized")); 2779 2780 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND); 2781 if (topo == NULL) 2782 return (ENOMEM); 2783 2784 sbuf_printf(topo, "<groups>\n"); 2785 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1); 2786 sbuf_printf(topo, "</groups>\n"); 2787 2788 if (err == 0) { 2789 sbuf_finish(topo); 2790 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo)); 2791 } 2792 sbuf_delete(topo); 2793 return (err); 2794} 2795 2796#endif 2797 2798SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2799SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2800 "Scheduler name"); 2801SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2802 "Slice size for timeshare threads"); 2803SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2804 "Interactivity score threshold"); 2805SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2806 0,"Min priority for preemption, lower priorities have greater precedence"); 2807SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 2808 0,"Controls whether static kernel priorities are assigned to sleeping threads."); 2809SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 2810 0,"Number of times idle will spin waiting for new work."); 2811SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, &sched_idlespinthresh, 2812 0,"Threshold before we will permit idle spinning."); 2813#ifdef SMP 2814SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2815 "Number of hz ticks to keep thread affinity for"); 2816SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2817 "Enables the long-term load balancer"); 2818SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2819 &balance_interval, 0, 2820 "Average frequency in stathz ticks to run the long-term balancer"); 2821SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2822 "Attempts to steal work from other cores before idling"); 2823SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2824 "Minimum load on remote cpu before we'll steal"); 2825 2826/* Retrieve SMP topology */ 2827SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING | 2828 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A", 2829 "XML dump of detected CPU topology"); 2830 2831#endif 2832 2833/* ps compat. All cpu percentages from ULE are weighted. */ 2834static int ccpu = 0; 2835SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2836