sched_ule.c revision 259835
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 259835 2013-12-24 19:02:04Z 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/sdt.h> 57#include <sys/smp.h> 58#include <sys/sx.h> 59#include <sys/sysctl.h> 60#include <sys/sysproto.h> 61#include <sys/turnstile.h> 62#include <sys/umtx.h> 63#include <sys/vmmeter.h> 64#include <sys/cpuset.h> 65#include <sys/sbuf.h> 66 67#ifdef HWPMC_HOOKS 68#include <sys/pmckern.h> 69#endif 70 71#ifdef KDTRACE_HOOKS 72#include <sys/dtrace_bsd.h> 73int dtrace_vtime_active; 74dtrace_vtime_switch_func_t dtrace_vtime_switch_func; 75#endif 76 77#include <machine/cpu.h> 78#include <machine/smp.h> 79 80#if defined(__powerpc__) && defined(E500) 81#error "This architecture is not currently compatible with ULE" 82#endif 83 84#define KTR_ULE 0 85 86#define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX))) 87#define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU))) 88#define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load")) 89 90/* 91 * Thread scheduler specific section. All fields are protected 92 * by the thread lock. 93 */ 94struct td_sched { 95 struct runq *ts_runq; /* Run-queue we're queued on. */ 96 short ts_flags; /* TSF_* flags. */ 97 u_char ts_cpu; /* CPU that we have affinity for. */ 98 int ts_rltick; /* Real last tick, for affinity. */ 99 int ts_slice; /* Ticks of slice remaining. */ 100 u_int ts_slptime; /* Number of ticks we vol. slept */ 101 u_int ts_runtime; /* Number of ticks we were running */ 102 int ts_ltick; /* Last tick that we were running on */ 103 int ts_ftick; /* First tick that we were running on */ 104 int ts_ticks; /* Tick count */ 105#ifdef KTR 106 char ts_name[TS_NAME_LEN]; 107#endif 108}; 109/* flags kept in ts_flags */ 110#define TSF_BOUND 0x0001 /* Thread can not migrate. */ 111#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 112 113static struct td_sched td_sched0; 114 115#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 116#define THREAD_CAN_SCHED(td, cpu) \ 117 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 118 119/* 120 * Priority ranges used for interactive and non-interactive timeshare 121 * threads. The timeshare priorities are split up into four ranges. 122 * The first range handles interactive threads. The last three ranges 123 * (NHALF, x, and NHALF) handle non-interactive threads with the outer 124 * ranges supporting nice values. 125 */ 126#define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1) 127#define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2) 128#define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE) 129 130#define PRI_MIN_INTERACT PRI_MIN_TIMESHARE 131#define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1) 132#define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE) 133#define PRI_MAX_BATCH PRI_MAX_TIMESHARE 134 135/* 136 * Cpu percentage computation macros and defines. 137 * 138 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 139 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 140 * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 141 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 142 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 143 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 144 */ 145#define SCHED_TICK_SECS 10 146#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 147#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 148#define SCHED_TICK_SHIFT 10 149#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 150#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 151 152/* 153 * These macros determine priorities for non-interactive threads. They are 154 * assigned a priority based on their recent cpu utilization as expressed 155 * by the ratio of ticks to the tick total. NHALF priorities at the start 156 * and end of the MIN to MAX timeshare range are only reachable with negative 157 * or positive nice respectively. 158 * 159 * PRI_RANGE: Priority range for utilization dependent priorities. 160 * PRI_NRESV: Number of nice values. 161 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 162 * PRI_NICE: Determines the part of the priority inherited from nice. 163 */ 164#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 165#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 166#define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF) 167#define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF) 168#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1) 169#define SCHED_PRI_TICKS(ts) \ 170 (SCHED_TICK_HZ((ts)) / \ 171 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 172#define SCHED_PRI_NICE(nice) (nice) 173 174/* 175 * These determine the interactivity of a process. Interactivity differs from 176 * cpu utilization in that it expresses the voluntary time slept vs time ran 177 * while cpu utilization includes all time not running. This more accurately 178 * models the intent of the thread. 179 * 180 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 181 * before throttling back. 182 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 183 * INTERACT_MAX: Maximum interactivity value. Smaller is better. 184 * INTERACT_THRESH: Threshold for placement on the current runq. 185 */ 186#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 187#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 188#define SCHED_INTERACT_MAX (100) 189#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 190#define SCHED_INTERACT_THRESH (30) 191 192/* Flags kept in td_flags. */ 193#define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */ 194 195/* 196 * tickincr: Converts a stathz tick into a hz domain scaled by 197 * the shift factor. Without the shift the error rate 198 * due to rounding would be unacceptably high. 199 * realstathz: stathz is sometimes 0 and run off of hz. 200 * sched_slice: Runtime of each thread before rescheduling. 201 * preempt_thresh: Priority threshold for preemption and remote IPIs. 202 */ 203static int sched_interact = SCHED_INTERACT_THRESH; 204static int realstathz = 127; 205static int tickincr = 8 << SCHED_TICK_SHIFT; 206static int sched_slice = 12; 207#ifdef PREEMPTION 208#ifdef FULL_PREEMPTION 209static int preempt_thresh = PRI_MAX_IDLE; 210#else 211static int preempt_thresh = PRI_MIN_KERN; 212#endif 213#else 214static int preempt_thresh = 0; 215#endif 216static int static_boost = PRI_MIN_BATCH; 217static int sched_idlespins = 10000; 218static int sched_idlespinthresh = -1; 219 220/* 221 * tdq - per processor runqs and statistics. All fields are protected by the 222 * tdq_lock. The load and lowpri may be accessed without to avoid excess 223 * locking in sched_pickcpu(); 224 */ 225struct tdq { 226 /* Ordered to improve efficiency of cpu_search() and switch(). */ 227 struct mtx tdq_lock; /* run queue lock. */ 228 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */ 229 volatile int tdq_load; /* Aggregate load. */ 230 volatile int tdq_cpu_idle; /* cpu_idle() is active. */ 231 int tdq_sysload; /* For loadavg, !ITHD load. */ 232 int tdq_transferable; /* Transferable thread count. */ 233 short tdq_switchcnt; /* Switches this tick. */ 234 short tdq_oldswitchcnt; /* Switches last tick. */ 235 u_char tdq_lowpri; /* Lowest priority thread. */ 236 u_char tdq_ipipending; /* IPI pending. */ 237 u_char tdq_idx; /* Current insert index. */ 238 u_char tdq_ridx; /* Current removal index. */ 239 struct runq tdq_realtime; /* real-time run queue. */ 240 struct runq tdq_timeshare; /* timeshare run queue. */ 241 struct runq tdq_idle; /* Queue of IDLE threads. */ 242 char tdq_name[TDQ_NAME_LEN]; 243#ifdef KTR 244 char tdq_loadname[TDQ_LOADNAME_LEN]; 245#endif 246} __aligned(64); 247 248/* Idle thread states and config. */ 249#define TDQ_RUNNING 1 250#define TDQ_IDLE 2 251 252#ifdef SMP 253struct cpu_group *cpu_top; /* CPU topology */ 254 255#define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000)) 256#define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity)) 257 258/* 259 * Run-time tunables. 260 */ 261static int rebalance = 1; 262static int balance_interval = 128; /* Default set in sched_initticks(). */ 263static int affinity; 264static int steal_idle = 1; 265static int steal_thresh = 2; 266 267/* 268 * One thread queue per processor. 269 */ 270static struct tdq tdq_cpu[MAXCPU]; 271static struct tdq *balance_tdq; 272static int balance_ticks; 273static DPCPU_DEFINE(uint32_t, randomval); 274 275#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 276#define TDQ_CPU(x) (&tdq_cpu[(x)]) 277#define TDQ_ID(x) ((int)((x) - tdq_cpu)) 278#else /* !SMP */ 279static struct tdq tdq_cpu; 280 281#define TDQ_ID(x) (0) 282#define TDQ_SELF() (&tdq_cpu) 283#define TDQ_CPU(x) (&tdq_cpu) 284#endif 285 286#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 287#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 288#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 289#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 290#define TDQ_LOCKPTR(t) (&(t)->tdq_lock) 291 292static void sched_priority(struct thread *); 293static void sched_thread_priority(struct thread *, u_char); 294static int sched_interact_score(struct thread *); 295static void sched_interact_update(struct thread *); 296static void sched_interact_fork(struct thread *); 297static void sched_pctcpu_update(struct td_sched *, int); 298 299/* Operations on per processor queues */ 300static struct thread *tdq_choose(struct tdq *); 301static void tdq_setup(struct tdq *); 302static void tdq_load_add(struct tdq *, struct thread *); 303static void tdq_load_rem(struct tdq *, struct thread *); 304static __inline void tdq_runq_add(struct tdq *, struct thread *, int); 305static __inline void tdq_runq_rem(struct tdq *, struct thread *); 306static inline int sched_shouldpreempt(int, int, int); 307void tdq_print(int cpu); 308static void runq_print(struct runq *rq); 309static void tdq_add(struct tdq *, struct thread *, int); 310#ifdef SMP 311static int tdq_move(struct tdq *, struct tdq *); 312static int tdq_idled(struct tdq *); 313static void tdq_notify(struct tdq *, struct thread *); 314static struct thread *tdq_steal(struct tdq *, int); 315static struct thread *runq_steal(struct runq *, int); 316static int sched_pickcpu(struct thread *, int); 317static void sched_balance(void); 318static int sched_balance_pair(struct tdq *, struct tdq *); 319static inline struct tdq *sched_setcpu(struct thread *, int, int); 320static inline void thread_unblock_switch(struct thread *, struct mtx *); 321static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 322static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS); 323static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, 324 struct cpu_group *cg, int indent); 325#endif 326 327static void sched_setup(void *dummy); 328SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 329 330static void sched_initticks(void *dummy); 331SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 332 NULL); 333 334SDT_PROVIDER_DEFINE(sched); 335 336SDT_PROBE_DEFINE3(sched, , , change_pri, 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; ; ) { 636 if (i == 0) { 637#ifdef HAVE_INLINE_FFSL 638 cpu = CPU_FFS(&cpumask) - 1; 639#else 640 while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask)) 641 cpu--; 642#endif 643 if (cpu < 0) 644 break; 645 child = NULL; 646 } else 647 child = &cg->cg_child[i - 1]; 648 649 if (match & CPU_SEARCH_LOWEST) 650 lgroup.cs_cpu = -1; 651 if (match & CPU_SEARCH_HIGHEST) 652 hgroup.cs_cpu = -1; 653 if (child) { /* Handle child CPU group. */ 654 CPU_NAND(&cpumask, &child->cg_mask); 655 switch (match) { 656 case CPU_SEARCH_LOWEST: 657 load = cpu_search_lowest(child, &lgroup); 658 break; 659 case CPU_SEARCH_HIGHEST: 660 load = cpu_search_highest(child, &hgroup); 661 break; 662 case CPU_SEARCH_BOTH: 663 load = cpu_search_both(child, &lgroup, &hgroup); 664 break; 665 } 666 } else { /* Handle child CPU. */ 667 CPU_CLR(cpu, &cpumask); 668 tdq = TDQ_CPU(cpu); 669 load = tdq->tdq_load * 256; 670 rndptr = DPCPU_PTR(randomval); 671 rnd = (*rndptr = *rndptr * 69069 + 5) >> 26; 672 if (match & CPU_SEARCH_LOWEST) { 673 if (cpu == low->cs_prefer) 674 load -= 64; 675 /* If that CPU is allowed and get data. */ 676 if (tdq->tdq_lowpri > lgroup.cs_pri && 677 tdq->tdq_load <= lgroup.cs_limit && 678 CPU_ISSET(cpu, &lgroup.cs_mask)) { 679 lgroup.cs_cpu = cpu; 680 lgroup.cs_load = load - rnd; 681 } 682 } 683 if (match & CPU_SEARCH_HIGHEST) 684 if (tdq->tdq_load >= hgroup.cs_limit && 685 tdq->tdq_transferable && 686 CPU_ISSET(cpu, &hgroup.cs_mask)) { 687 hgroup.cs_cpu = cpu; 688 hgroup.cs_load = load - rnd; 689 } 690 } 691 total += load; 692 693 /* We have info about child item. Compare it. */ 694 if (match & CPU_SEARCH_LOWEST) { 695 if (lgroup.cs_cpu >= 0 && 696 (load < lload || 697 (load == lload && lgroup.cs_load < low->cs_load))) { 698 lload = load; 699 low->cs_cpu = lgroup.cs_cpu; 700 low->cs_load = lgroup.cs_load; 701 } 702 } 703 if (match & CPU_SEARCH_HIGHEST) 704 if (hgroup.cs_cpu >= 0 && 705 (load > hload || 706 (load == hload && hgroup.cs_load > high->cs_load))) { 707 hload = load; 708 high->cs_cpu = hgroup.cs_cpu; 709 high->cs_load = hgroup.cs_load; 710 } 711 if (child) { 712 i--; 713 if (i == 0 && CPU_EMPTY(&cpumask)) 714 break; 715 } 716#ifndef HAVE_INLINE_FFSL 717 else 718 cpu--; 719#endif 720 } 721 return (total); 722} 723 724/* 725 * cpu_search instantiations must pass constants to maintain the inline 726 * optimization. 727 */ 728int 729cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low) 730{ 731 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 732} 733 734int 735cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high) 736{ 737 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 738} 739 740int 741cpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 742 struct cpu_search *high) 743{ 744 return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 745} 746 747/* 748 * Find the cpu with the least load via the least loaded path that has a 749 * lowpri greater than pri pri. A pri of -1 indicates any priority is 750 * acceptable. 751 */ 752static inline int 753sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload, 754 int prefer) 755{ 756 struct cpu_search low; 757 758 low.cs_cpu = -1; 759 low.cs_prefer = prefer; 760 low.cs_mask = mask; 761 low.cs_pri = pri; 762 low.cs_limit = maxload; 763 cpu_search_lowest(cg, &low); 764 return low.cs_cpu; 765} 766 767/* 768 * Find the cpu with the highest load via the highest loaded path. 769 */ 770static inline int 771sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload) 772{ 773 struct cpu_search high; 774 775 high.cs_cpu = -1; 776 high.cs_mask = mask; 777 high.cs_limit = minload; 778 cpu_search_highest(cg, &high); 779 return high.cs_cpu; 780} 781 782/* 783 * Simultaneously find the highest and lowest loaded cpu reachable via 784 * cg. 785 */ 786static inline void 787sched_both(const struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu) 788{ 789 struct cpu_search high; 790 struct cpu_search low; 791 792 low.cs_cpu = -1; 793 low.cs_prefer = -1; 794 low.cs_pri = -1; 795 low.cs_limit = INT_MAX; 796 low.cs_mask = mask; 797 high.cs_cpu = -1; 798 high.cs_limit = -1; 799 high.cs_mask = mask; 800 cpu_search_both(cg, &low, &high); 801 *lowcpu = low.cs_cpu; 802 *highcpu = high.cs_cpu; 803 return; 804} 805 806static void 807sched_balance_group(struct cpu_group *cg) 808{ 809 cpuset_t hmask, lmask; 810 int high, low, anylow; 811 812 CPU_FILL(&hmask); 813 for (;;) { 814 high = sched_highest(cg, hmask, 1); 815 /* Stop if there is no more CPU with transferrable threads. */ 816 if (high == -1) 817 break; 818 CPU_CLR(high, &hmask); 819 CPU_COPY(&hmask, &lmask); 820 /* Stop if there is no more CPU left for low. */ 821 if (CPU_EMPTY(&lmask)) 822 break; 823 anylow = 1; 824nextlow: 825 low = sched_lowest(cg, lmask, -1, 826 TDQ_CPU(high)->tdq_load - 1, high); 827 /* Stop if we looked well and found no less loaded CPU. */ 828 if (anylow && low == -1) 829 break; 830 /* Go to next high if we found no less loaded CPU. */ 831 if (low == -1) 832 continue; 833 /* Transfer thread from high to low. */ 834 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) { 835 /* CPU that got thread can no longer be a donor. */ 836 CPU_CLR(low, &hmask); 837 } else { 838 /* 839 * If failed, then there is no threads on high 840 * that can run on this low. Drop low from low 841 * mask and look for different one. 842 */ 843 CPU_CLR(low, &lmask); 844 anylow = 0; 845 goto nextlow; 846 } 847 } 848} 849 850static void 851sched_balance(void) 852{ 853 struct tdq *tdq; 854 855 /* 856 * Select a random time between .5 * balance_interval and 857 * 1.5 * balance_interval. 858 */ 859 balance_ticks = max(balance_interval / 2, 1); 860 balance_ticks += random() % balance_interval; 861 if (smp_started == 0 || rebalance == 0) 862 return; 863 tdq = TDQ_SELF(); 864 TDQ_UNLOCK(tdq); 865 sched_balance_group(cpu_top); 866 TDQ_LOCK(tdq); 867} 868 869/* 870 * Lock two thread queues using their address to maintain lock order. 871 */ 872static void 873tdq_lock_pair(struct tdq *one, struct tdq *two) 874{ 875 if (one < two) { 876 TDQ_LOCK(one); 877 TDQ_LOCK_FLAGS(two, MTX_DUPOK); 878 } else { 879 TDQ_LOCK(two); 880 TDQ_LOCK_FLAGS(one, MTX_DUPOK); 881 } 882} 883 884/* 885 * Unlock two thread queues. Order is not important here. 886 */ 887static void 888tdq_unlock_pair(struct tdq *one, struct tdq *two) 889{ 890 TDQ_UNLOCK(one); 891 TDQ_UNLOCK(two); 892} 893 894/* 895 * Transfer load between two imbalanced thread queues. 896 */ 897static int 898sched_balance_pair(struct tdq *high, struct tdq *low) 899{ 900 int moved; 901 int cpu; 902 903 tdq_lock_pair(high, low); 904 moved = 0; 905 /* 906 * Determine what the imbalance is and then adjust that to how many 907 * threads we actually have to give up (transferable). 908 */ 909 if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load && 910 (moved = tdq_move(high, low)) > 0) { 911 /* 912 * In case the target isn't the current cpu IPI it to force a 913 * reschedule with the new workload. 914 */ 915 cpu = TDQ_ID(low); 916 sched_pin(); 917 if (cpu != PCPU_GET(cpuid)) 918 ipi_cpu(cpu, IPI_PREEMPT); 919 sched_unpin(); 920 } 921 tdq_unlock_pair(high, low); 922 return (moved); 923} 924 925/* 926 * Move a thread from one thread queue to another. 927 */ 928static int 929tdq_move(struct tdq *from, struct tdq *to) 930{ 931 struct td_sched *ts; 932 struct thread *td; 933 struct tdq *tdq; 934 int cpu; 935 936 TDQ_LOCK_ASSERT(from, MA_OWNED); 937 TDQ_LOCK_ASSERT(to, MA_OWNED); 938 939 tdq = from; 940 cpu = TDQ_ID(to); 941 td = tdq_steal(tdq, cpu); 942 if (td == NULL) 943 return (0); 944 ts = td->td_sched; 945 /* 946 * Although the run queue is locked the thread may be blocked. Lock 947 * it to clear this and acquire the run-queue lock. 948 */ 949 thread_lock(td); 950 /* Drop recursive lock on from acquired via thread_lock(). */ 951 TDQ_UNLOCK(from); 952 sched_rem(td); 953 ts->ts_cpu = cpu; 954 td->td_lock = TDQ_LOCKPTR(to); 955 tdq_add(to, td, SRQ_YIELDING); 956 return (1); 957} 958 959/* 960 * This tdq has idled. Try to steal a thread from another cpu and switch 961 * to it. 962 */ 963static int 964tdq_idled(struct tdq *tdq) 965{ 966 struct cpu_group *cg; 967 struct tdq *steal; 968 cpuset_t mask; 969 int thresh; 970 int cpu; 971 972 if (smp_started == 0 || steal_idle == 0) 973 return (1); 974 CPU_FILL(&mask); 975 CPU_CLR(PCPU_GET(cpuid), &mask); 976 /* We don't want to be preempted while we're iterating. */ 977 spinlock_enter(); 978 for (cg = tdq->tdq_cg; cg != NULL; ) { 979 if ((cg->cg_flags & CG_FLAG_THREAD) == 0) 980 thresh = steal_thresh; 981 else 982 thresh = 1; 983 cpu = sched_highest(cg, mask, thresh); 984 if (cpu == -1) { 985 cg = cg->cg_parent; 986 continue; 987 } 988 steal = TDQ_CPU(cpu); 989 CPU_CLR(cpu, &mask); 990 tdq_lock_pair(tdq, steal); 991 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 992 tdq_unlock_pair(tdq, steal); 993 continue; 994 } 995 /* 996 * If a thread was added while interrupts were disabled don't 997 * steal one here. If we fail to acquire one due to affinity 998 * restrictions loop again with this cpu removed from the 999 * set. 1000 */ 1001 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 1002 tdq_unlock_pair(tdq, steal); 1003 continue; 1004 } 1005 spinlock_exit(); 1006 TDQ_UNLOCK(steal); 1007 mi_switch(SW_VOL | SWT_IDLE, NULL); 1008 thread_unlock(curthread); 1009 1010 return (0); 1011 } 1012 spinlock_exit(); 1013 return (1); 1014} 1015 1016/* 1017 * Notify a remote cpu of new work. Sends an IPI if criteria are met. 1018 */ 1019static void 1020tdq_notify(struct tdq *tdq, struct thread *td) 1021{ 1022 struct thread *ctd; 1023 int pri; 1024 int cpu; 1025 1026 if (tdq->tdq_ipipending) 1027 return; 1028 cpu = td->td_sched->ts_cpu; 1029 pri = td->td_priority; 1030 ctd = pcpu_find(cpu)->pc_curthread; 1031 if (!sched_shouldpreempt(pri, ctd->td_priority, 1)) 1032 return; 1033 if (TD_IS_IDLETHREAD(ctd)) { 1034 /* 1035 * If the MD code has an idle wakeup routine try that before 1036 * falling back to IPI. 1037 */ 1038 if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu)) 1039 return; 1040 } 1041 tdq->tdq_ipipending = 1; 1042 ipi_cpu(cpu, IPI_PREEMPT); 1043} 1044 1045/* 1046 * Steals load from a timeshare queue. Honors the rotating queue head 1047 * index. 1048 */ 1049static struct thread * 1050runq_steal_from(struct runq *rq, int cpu, u_char start) 1051{ 1052 struct rqbits *rqb; 1053 struct rqhead *rqh; 1054 struct thread *td, *first; 1055 int bit; 1056 int pri; 1057 int i; 1058 1059 rqb = &rq->rq_status; 1060 bit = start & (RQB_BPW -1); 1061 pri = 0; 1062 first = NULL; 1063again: 1064 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 1065 if (rqb->rqb_bits[i] == 0) 1066 continue; 1067 if (bit != 0) { 1068 for (pri = bit; pri < RQB_BPW; pri++) 1069 if (rqb->rqb_bits[i] & (1ul << pri)) 1070 break; 1071 if (pri >= RQB_BPW) 1072 continue; 1073 } else 1074 pri = RQB_FFS(rqb->rqb_bits[i]); 1075 pri += (i << RQB_L2BPW); 1076 rqh = &rq->rq_queues[pri]; 1077 TAILQ_FOREACH(td, rqh, td_runq) { 1078 if (first && THREAD_CAN_MIGRATE(td) && 1079 THREAD_CAN_SCHED(td, cpu)) 1080 return (td); 1081 first = td; 1082 } 1083 } 1084 if (start != 0) { 1085 start = 0; 1086 goto again; 1087 } 1088 1089 if (first && THREAD_CAN_MIGRATE(first) && 1090 THREAD_CAN_SCHED(first, cpu)) 1091 return (first); 1092 return (NULL); 1093} 1094 1095/* 1096 * Steals load from a standard linear queue. 1097 */ 1098static struct thread * 1099runq_steal(struct runq *rq, int cpu) 1100{ 1101 struct rqhead *rqh; 1102 struct rqbits *rqb; 1103 struct thread *td; 1104 int word; 1105 int bit; 1106 1107 rqb = &rq->rq_status; 1108 for (word = 0; word < RQB_LEN; word++) { 1109 if (rqb->rqb_bits[word] == 0) 1110 continue; 1111 for (bit = 0; bit < RQB_BPW; bit++) { 1112 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1113 continue; 1114 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1115 TAILQ_FOREACH(td, rqh, td_runq) 1116 if (THREAD_CAN_MIGRATE(td) && 1117 THREAD_CAN_SCHED(td, cpu)) 1118 return (td); 1119 } 1120 } 1121 return (NULL); 1122} 1123 1124/* 1125 * Attempt to steal a thread in priority order from a thread queue. 1126 */ 1127static struct thread * 1128tdq_steal(struct tdq *tdq, int cpu) 1129{ 1130 struct thread *td; 1131 1132 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1133 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1134 return (td); 1135 if ((td = runq_steal_from(&tdq->tdq_timeshare, 1136 cpu, tdq->tdq_ridx)) != NULL) 1137 return (td); 1138 return (runq_steal(&tdq->tdq_idle, cpu)); 1139} 1140 1141/* 1142 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1143 * current lock and returns with the assigned queue locked. 1144 */ 1145static inline struct tdq * 1146sched_setcpu(struct thread *td, int cpu, int flags) 1147{ 1148 1149 struct tdq *tdq; 1150 1151 THREAD_LOCK_ASSERT(td, MA_OWNED); 1152 tdq = TDQ_CPU(cpu); 1153 td->td_sched->ts_cpu = cpu; 1154 /* 1155 * If the lock matches just return the queue. 1156 */ 1157 if (td->td_lock == TDQ_LOCKPTR(tdq)) 1158 return (tdq); 1159#ifdef notyet 1160 /* 1161 * If the thread isn't running its lockptr is a 1162 * turnstile or a sleepqueue. We can just lock_set without 1163 * blocking. 1164 */ 1165 if (TD_CAN_RUN(td)) { 1166 TDQ_LOCK(tdq); 1167 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1168 return (tdq); 1169 } 1170#endif 1171 /* 1172 * The hard case, migration, we need to block the thread first to 1173 * prevent order reversals with other cpus locks. 1174 */ 1175 spinlock_enter(); 1176 thread_lock_block(td); 1177 TDQ_LOCK(tdq); 1178 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1179 spinlock_exit(); 1180 return (tdq); 1181} 1182 1183SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1184SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1185SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1186SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1187SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1188SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1189 1190static int 1191sched_pickcpu(struct thread *td, int flags) 1192{ 1193 struct cpu_group *cg, *ccg; 1194 struct td_sched *ts; 1195 struct tdq *tdq; 1196 cpuset_t mask; 1197 int cpu, pri, self; 1198 1199 self = PCPU_GET(cpuid); 1200 ts = td->td_sched; 1201 if (smp_started == 0) 1202 return (self); 1203 /* 1204 * Don't migrate a running thread from sched_switch(). 1205 */ 1206 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1207 return (ts->ts_cpu); 1208 /* 1209 * Prefer to run interrupt threads on the processors that generate 1210 * the interrupt. 1211 */ 1212 pri = td->td_priority; 1213 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1214 curthread->td_intr_nesting_level && ts->ts_cpu != self) { 1215 SCHED_STAT_INC(pickcpu_intrbind); 1216 ts->ts_cpu = self; 1217 if (TDQ_CPU(self)->tdq_lowpri > pri) { 1218 SCHED_STAT_INC(pickcpu_affinity); 1219 return (ts->ts_cpu); 1220 } 1221 } 1222 /* 1223 * If the thread can run on the last cpu and the affinity has not 1224 * expired or it is idle run it there. 1225 */ 1226 tdq = TDQ_CPU(ts->ts_cpu); 1227 cg = tdq->tdq_cg; 1228 if (THREAD_CAN_SCHED(td, ts->ts_cpu) && 1229 tdq->tdq_lowpri >= PRI_MIN_IDLE && 1230 SCHED_AFFINITY(ts, CG_SHARE_L2)) { 1231 if (cg->cg_flags & CG_FLAG_THREAD) { 1232 CPUSET_FOREACH(cpu, cg->cg_mask) { 1233 if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) 1234 break; 1235 } 1236 } else 1237 cpu = INT_MAX; 1238 if (cpu > mp_maxid) { 1239 SCHED_STAT_INC(pickcpu_idle_affinity); 1240 return (ts->ts_cpu); 1241 } 1242 } 1243 /* 1244 * Search for the last level cache CPU group in the tree. 1245 * Skip caches with expired affinity time and SMT groups. 1246 * Affinity to higher level caches will be handled less aggressively. 1247 */ 1248 for (ccg = NULL; cg != NULL; cg = cg->cg_parent) { 1249 if (cg->cg_flags & CG_FLAG_THREAD) 1250 continue; 1251 if (!SCHED_AFFINITY(ts, cg->cg_level)) 1252 continue; 1253 ccg = cg; 1254 } 1255 if (ccg != NULL) 1256 cg = ccg; 1257 cpu = -1; 1258 /* Search the group for the less loaded idle CPU we can run now. */ 1259 mask = td->td_cpuset->cs_mask; 1260 if (cg != NULL && cg != cpu_top && 1261 CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0) 1262 cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE), 1263 INT_MAX, ts->ts_cpu); 1264 /* Search globally for the less loaded CPU we can run now. */ 1265 if (cpu == -1) 1266 cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu); 1267 /* Search globally for the less loaded CPU. */ 1268 if (cpu == -1) 1269 cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu); 1270 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1271 /* 1272 * Compare the lowest loaded cpu to current cpu. 1273 */ 1274 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1275 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE && 1276 TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) { 1277 SCHED_STAT_INC(pickcpu_local); 1278 cpu = self; 1279 } else 1280 SCHED_STAT_INC(pickcpu_lowest); 1281 if (cpu != ts->ts_cpu) 1282 SCHED_STAT_INC(pickcpu_migration); 1283 return (cpu); 1284} 1285#endif 1286 1287/* 1288 * Pick the highest priority task we have and return it. 1289 */ 1290static struct thread * 1291tdq_choose(struct tdq *tdq) 1292{ 1293 struct thread *td; 1294 1295 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1296 td = runq_choose(&tdq->tdq_realtime); 1297 if (td != NULL) 1298 return (td); 1299 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1300 if (td != NULL) { 1301 KASSERT(td->td_priority >= PRI_MIN_BATCH, 1302 ("tdq_choose: Invalid priority on timeshare queue %d", 1303 td->td_priority)); 1304 return (td); 1305 } 1306 td = runq_choose(&tdq->tdq_idle); 1307 if (td != NULL) { 1308 KASSERT(td->td_priority >= PRI_MIN_IDLE, 1309 ("tdq_choose: Invalid priority on idle queue %d", 1310 td->td_priority)); 1311 return (td); 1312 } 1313 1314 return (NULL); 1315} 1316 1317/* 1318 * Initialize a thread queue. 1319 */ 1320static void 1321tdq_setup(struct tdq *tdq) 1322{ 1323 1324 if (bootverbose) 1325 printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1326 runq_init(&tdq->tdq_realtime); 1327 runq_init(&tdq->tdq_timeshare); 1328 runq_init(&tdq->tdq_idle); 1329 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1330 "sched lock %d", (int)TDQ_ID(tdq)); 1331 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1332 MTX_SPIN | MTX_RECURSE); 1333#ifdef KTR 1334 snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname), 1335 "CPU %d load", (int)TDQ_ID(tdq)); 1336#endif 1337} 1338 1339#ifdef SMP 1340static void 1341sched_setup_smp(void) 1342{ 1343 struct tdq *tdq; 1344 int i; 1345 1346 cpu_top = smp_topo(); 1347 CPU_FOREACH(i) { 1348 tdq = TDQ_CPU(i); 1349 tdq_setup(tdq); 1350 tdq->tdq_cg = smp_topo_find(cpu_top, i); 1351 if (tdq->tdq_cg == NULL) 1352 panic("Can't find cpu group for %d\n", i); 1353 } 1354 balance_tdq = TDQ_SELF(); 1355 sched_balance(); 1356} 1357#endif 1358 1359/* 1360 * Setup the thread queues and initialize the topology based on MD 1361 * information. 1362 */ 1363static void 1364sched_setup(void *dummy) 1365{ 1366 struct tdq *tdq; 1367 1368 tdq = TDQ_SELF(); 1369#ifdef SMP 1370 sched_setup_smp(); 1371#else 1372 tdq_setup(tdq); 1373#endif 1374 1375 /* Add thread0's load since it's running. */ 1376 TDQ_LOCK(tdq); 1377 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1378 tdq_load_add(tdq, &thread0); 1379 tdq->tdq_lowpri = thread0.td_priority; 1380 TDQ_UNLOCK(tdq); 1381} 1382 1383/* 1384 * This routine determines time constants after stathz and hz are setup. 1385 */ 1386/* ARGSUSED */ 1387static void 1388sched_initticks(void *dummy) 1389{ 1390 int incr; 1391 1392 realstathz = stathz ? stathz : hz; 1393 sched_slice = realstathz / 10; /* ~100ms */ 1394 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 1395 realstathz); 1396 1397 /* 1398 * tickincr is shifted out by 10 to avoid rounding errors due to 1399 * hz not being evenly divisible by stathz on all platforms. 1400 */ 1401 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1402 /* 1403 * This does not work for values of stathz that are more than 1404 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1405 */ 1406 if (incr == 0) 1407 incr = 1; 1408 tickincr = incr; 1409#ifdef SMP 1410 /* 1411 * Set the default balance interval now that we know 1412 * what realstathz is. 1413 */ 1414 balance_interval = realstathz; 1415 affinity = SCHED_AFFINITY_DEFAULT; 1416#endif 1417 if (sched_idlespinthresh < 0) 1418 sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz; 1419} 1420 1421 1422/* 1423 * This is the core of the interactivity algorithm. Determines a score based 1424 * on past behavior. It is the ratio of sleep time to run time scaled to 1425 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1426 * differs from the cpu usage because it does not account for time spent 1427 * waiting on a run-queue. Would be prettier if we had floating point. 1428 */ 1429static int 1430sched_interact_score(struct thread *td) 1431{ 1432 struct td_sched *ts; 1433 int div; 1434 1435 ts = td->td_sched; 1436 /* 1437 * The score is only needed if this is likely to be an interactive 1438 * task. Don't go through the expense of computing it if there's 1439 * no chance. 1440 */ 1441 if (sched_interact <= SCHED_INTERACT_HALF && 1442 ts->ts_runtime >= ts->ts_slptime) 1443 return (SCHED_INTERACT_HALF); 1444 1445 if (ts->ts_runtime > ts->ts_slptime) { 1446 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1447 return (SCHED_INTERACT_HALF + 1448 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1449 } 1450 if (ts->ts_slptime > ts->ts_runtime) { 1451 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1452 return (ts->ts_runtime / div); 1453 } 1454 /* runtime == slptime */ 1455 if (ts->ts_runtime) 1456 return (SCHED_INTERACT_HALF); 1457 1458 /* 1459 * This can happen if slptime and runtime are 0. 1460 */ 1461 return (0); 1462 1463} 1464 1465/* 1466 * Scale the scheduling priority according to the "interactivity" of this 1467 * process. 1468 */ 1469static void 1470sched_priority(struct thread *td) 1471{ 1472 int score; 1473 int pri; 1474 1475 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1476 return; 1477 /* 1478 * If the score is interactive we place the thread in the realtime 1479 * queue with a priority that is less than kernel and interrupt 1480 * priorities. These threads are not subject to nice restrictions. 1481 * 1482 * Scores greater than this are placed on the normal timeshare queue 1483 * where the priority is partially decided by the most recent cpu 1484 * utilization and the rest is decided by nice value. 1485 * 1486 * The nice value of the process has a linear effect on the calculated 1487 * score. Negative nice values make it easier for a thread to be 1488 * considered interactive. 1489 */ 1490 score = imax(0, sched_interact_score(td) + td->td_proc->p_nice); 1491 if (score < sched_interact) { 1492 pri = PRI_MIN_INTERACT; 1493 pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) / 1494 sched_interact) * score; 1495 KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT, 1496 ("sched_priority: invalid interactive priority %d score %d", 1497 pri, score)); 1498 } else { 1499 pri = SCHED_PRI_MIN; 1500 if (td->td_sched->ts_ticks) 1501 pri += min(SCHED_PRI_TICKS(td->td_sched), 1502 SCHED_PRI_RANGE - 1); 1503 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1504 KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH, 1505 ("sched_priority: invalid priority %d: nice %d, " 1506 "ticks %d ftick %d ltick %d tick pri %d", 1507 pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1508 td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1509 SCHED_PRI_TICKS(td->td_sched))); 1510 } 1511 sched_user_prio(td, pri); 1512 1513 return; 1514} 1515 1516/* 1517 * This routine enforces a maximum limit on the amount of scheduling history 1518 * kept. It is called after either the slptime or runtime is adjusted. This 1519 * function is ugly due to integer math. 1520 */ 1521static void 1522sched_interact_update(struct thread *td) 1523{ 1524 struct td_sched *ts; 1525 u_int sum; 1526 1527 ts = td->td_sched; 1528 sum = ts->ts_runtime + ts->ts_slptime; 1529 if (sum < SCHED_SLP_RUN_MAX) 1530 return; 1531 /* 1532 * This only happens from two places: 1533 * 1) We have added an unusual amount of run time from fork_exit. 1534 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1535 */ 1536 if (sum > SCHED_SLP_RUN_MAX * 2) { 1537 if (ts->ts_runtime > ts->ts_slptime) { 1538 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1539 ts->ts_slptime = 1; 1540 } else { 1541 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1542 ts->ts_runtime = 1; 1543 } 1544 return; 1545 } 1546 /* 1547 * If we have exceeded by more than 1/5th then the algorithm below 1548 * will not bring us back into range. Dividing by two here forces 1549 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1550 */ 1551 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1552 ts->ts_runtime /= 2; 1553 ts->ts_slptime /= 2; 1554 return; 1555 } 1556 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1557 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1558} 1559 1560/* 1561 * Scale back the interactivity history when a child thread is created. The 1562 * history is inherited from the parent but the thread may behave totally 1563 * differently. For example, a shell spawning a compiler process. We want 1564 * to learn that the compiler is behaving badly very quickly. 1565 */ 1566static void 1567sched_interact_fork(struct thread *td) 1568{ 1569 int ratio; 1570 int sum; 1571 1572 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1573 if (sum > SCHED_SLP_RUN_FORK) { 1574 ratio = sum / SCHED_SLP_RUN_FORK; 1575 td->td_sched->ts_runtime /= ratio; 1576 td->td_sched->ts_slptime /= ratio; 1577 } 1578} 1579 1580/* 1581 * Called from proc0_init() to setup the scheduler fields. 1582 */ 1583void 1584schedinit(void) 1585{ 1586 1587 /* 1588 * Set up the scheduler specific parts of proc0. 1589 */ 1590 proc0.p_sched = NULL; /* XXX */ 1591 thread0.td_sched = &td_sched0; 1592 td_sched0.ts_ltick = ticks; 1593 td_sched0.ts_ftick = ticks; 1594 td_sched0.ts_slice = sched_slice; 1595} 1596 1597/* 1598 * This is only somewhat accurate since given many processes of the same 1599 * priority they will switch when their slices run out, which will be 1600 * at most sched_slice stathz ticks. 1601 */ 1602int 1603sched_rr_interval(void) 1604{ 1605 1606 /* Convert sched_slice from stathz to hz. */ 1607 return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz)); 1608} 1609 1610/* 1611 * Update the percent cpu tracking information when it is requested or 1612 * the total history exceeds the maximum. We keep a sliding history of 1613 * tick counts that slowly decays. This is less precise than the 4BSD 1614 * mechanism since it happens with less regular and frequent events. 1615 */ 1616static void 1617sched_pctcpu_update(struct td_sched *ts, int run) 1618{ 1619 int t = ticks; 1620 1621 if (t - ts->ts_ltick >= SCHED_TICK_TARG) { 1622 ts->ts_ticks = 0; 1623 ts->ts_ftick = t - SCHED_TICK_TARG; 1624 } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) { 1625 ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) * 1626 (ts->ts_ltick - (t - SCHED_TICK_TARG)); 1627 ts->ts_ftick = t - SCHED_TICK_TARG; 1628 } 1629 if (run) 1630 ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT; 1631 ts->ts_ltick = t; 1632} 1633 1634/* 1635 * Adjust the priority of a thread. Move it to the appropriate run-queue 1636 * if necessary. This is the back-end for several priority related 1637 * functions. 1638 */ 1639static void 1640sched_thread_priority(struct thread *td, u_char prio) 1641{ 1642 struct td_sched *ts; 1643 struct tdq *tdq; 1644 int oldpri; 1645 1646 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio", 1647 "prio:%d", td->td_priority, "new prio:%d", prio, 1648 KTR_ATTR_LINKED, sched_tdname(curthread)); 1649 SDT_PROBE3(sched, , , change_pri, td, td->td_proc, prio); 1650 if (td != curthread && prio < td->td_priority) { 1651 KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 1652 "lend prio", "prio:%d", td->td_priority, "new prio:%d", 1653 prio, KTR_ATTR_LINKED, sched_tdname(td)); 1654 SDT_PROBE4(sched, , , lend_pri, td, td->td_proc, prio, 1655 curthread); 1656 } 1657 ts = td->td_sched; 1658 THREAD_LOCK_ASSERT(td, MA_OWNED); 1659 if (td->td_priority == prio) 1660 return; 1661 /* 1662 * If the priority has been elevated due to priority 1663 * propagation, we may have to move ourselves to a new 1664 * queue. This could be optimized to not re-add in some 1665 * cases. 1666 */ 1667 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1668 sched_rem(td); 1669 td->td_priority = prio; 1670 sched_add(td, SRQ_BORROWING); 1671 return; 1672 } 1673 /* 1674 * If the thread is currently running we may have to adjust the lowpri 1675 * information so other cpus are aware of our current priority. 1676 */ 1677 if (TD_IS_RUNNING(td)) { 1678 tdq = TDQ_CPU(ts->ts_cpu); 1679 oldpri = td->td_priority; 1680 td->td_priority = prio; 1681 if (prio < tdq->tdq_lowpri) 1682 tdq->tdq_lowpri = prio; 1683 else if (tdq->tdq_lowpri == oldpri) 1684 tdq_setlowpri(tdq, td); 1685 return; 1686 } 1687 td->td_priority = prio; 1688} 1689 1690/* 1691 * Update a thread's priority when it is lent another thread's 1692 * priority. 1693 */ 1694void 1695sched_lend_prio(struct thread *td, u_char prio) 1696{ 1697 1698 td->td_flags |= TDF_BORROWING; 1699 sched_thread_priority(td, prio); 1700} 1701 1702/* 1703 * Restore a thread's priority when priority propagation is 1704 * over. The prio argument is the minimum priority the thread 1705 * needs to have to satisfy other possible priority lending 1706 * requests. If the thread's regular priority is less 1707 * important than prio, the thread will keep a priority boost 1708 * of prio. 1709 */ 1710void 1711sched_unlend_prio(struct thread *td, u_char prio) 1712{ 1713 u_char base_pri; 1714 1715 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1716 td->td_base_pri <= PRI_MAX_TIMESHARE) 1717 base_pri = td->td_user_pri; 1718 else 1719 base_pri = td->td_base_pri; 1720 if (prio >= base_pri) { 1721 td->td_flags &= ~TDF_BORROWING; 1722 sched_thread_priority(td, base_pri); 1723 } else 1724 sched_lend_prio(td, prio); 1725} 1726 1727/* 1728 * Standard entry for setting the priority to an absolute value. 1729 */ 1730void 1731sched_prio(struct thread *td, u_char prio) 1732{ 1733 u_char oldprio; 1734 1735 /* First, update the base priority. */ 1736 td->td_base_pri = prio; 1737 1738 /* 1739 * If the thread is borrowing another thread's priority, don't 1740 * ever lower the priority. 1741 */ 1742 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1743 return; 1744 1745 /* Change the real priority. */ 1746 oldprio = td->td_priority; 1747 sched_thread_priority(td, prio); 1748 1749 /* 1750 * If the thread is on a turnstile, then let the turnstile update 1751 * its state. 1752 */ 1753 if (TD_ON_LOCK(td) && oldprio != prio) 1754 turnstile_adjust(td, oldprio); 1755} 1756 1757/* 1758 * Set the base user priority, does not effect current running priority. 1759 */ 1760void 1761sched_user_prio(struct thread *td, u_char prio) 1762{ 1763 1764 td->td_base_user_pri = prio; 1765 if (td->td_lend_user_pri <= prio) 1766 return; 1767 td->td_user_pri = prio; 1768} 1769 1770void 1771sched_lend_user_prio(struct thread *td, u_char prio) 1772{ 1773 1774 THREAD_LOCK_ASSERT(td, MA_OWNED); 1775 td->td_lend_user_pri = prio; 1776 td->td_user_pri = min(prio, td->td_base_user_pri); 1777 if (td->td_priority > td->td_user_pri) 1778 sched_prio(td, td->td_user_pri); 1779 else if (td->td_priority != td->td_user_pri) 1780 td->td_flags |= TDF_NEEDRESCHED; 1781} 1782 1783/* 1784 * Handle migration from sched_switch(). This happens only for 1785 * cpu binding. 1786 */ 1787static struct mtx * 1788sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1789{ 1790 struct tdq *tdn; 1791 1792 tdn = TDQ_CPU(td->td_sched->ts_cpu); 1793#ifdef SMP 1794 tdq_load_rem(tdq, td); 1795 /* 1796 * Do the lock dance required to avoid LOR. We grab an extra 1797 * spinlock nesting to prevent preemption while we're 1798 * not holding either run-queue lock. 1799 */ 1800 spinlock_enter(); 1801 thread_lock_block(td); /* This releases the lock on tdq. */ 1802 1803 /* 1804 * Acquire both run-queue locks before placing the thread on the new 1805 * run-queue to avoid deadlocks created by placing a thread with a 1806 * blocked lock on the run-queue of a remote processor. The deadlock 1807 * occurs when a third processor attempts to lock the two queues in 1808 * question while the target processor is spinning with its own 1809 * run-queue lock held while waiting for the blocked lock to clear. 1810 */ 1811 tdq_lock_pair(tdn, tdq); 1812 tdq_add(tdn, td, flags); 1813 tdq_notify(tdn, td); 1814 TDQ_UNLOCK(tdn); 1815 spinlock_exit(); 1816#endif 1817 return (TDQ_LOCKPTR(tdn)); 1818} 1819 1820/* 1821 * Variadic version of thread_lock_unblock() that does not assume td_lock 1822 * is blocked. 1823 */ 1824static inline void 1825thread_unblock_switch(struct thread *td, struct mtx *mtx) 1826{ 1827 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1828 (uintptr_t)mtx); 1829} 1830 1831/* 1832 * Switch threads. This function has to handle threads coming in while 1833 * blocked for some reason, running, or idle. It also must deal with 1834 * migrating a thread from one queue to another as running threads may 1835 * be assigned elsewhere via binding. 1836 */ 1837void 1838sched_switch(struct thread *td, struct thread *newtd, int flags) 1839{ 1840 struct tdq *tdq; 1841 struct td_sched *ts; 1842 struct mtx *mtx; 1843 int srqflag; 1844 int cpuid, preempted; 1845 1846 THREAD_LOCK_ASSERT(td, MA_OWNED); 1847 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1848 1849 cpuid = PCPU_GET(cpuid); 1850 tdq = TDQ_CPU(cpuid); 1851 ts = td->td_sched; 1852 mtx = td->td_lock; 1853 sched_pctcpu_update(ts, 1); 1854 ts->ts_rltick = ticks; 1855 td->td_lastcpu = td->td_oncpu; 1856 td->td_oncpu = NOCPU; 1857 preempted = !(td->td_flags & TDF_SLICEEND); 1858 td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND); 1859 td->td_owepreempt = 0; 1860 if (!TD_IS_IDLETHREAD(td)) 1861 tdq->tdq_switchcnt++; 1862 /* 1863 * The lock pointer in an idle thread should never change. Reset it 1864 * to CAN_RUN as well. 1865 */ 1866 if (TD_IS_IDLETHREAD(td)) { 1867 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1868 TD_SET_CAN_RUN(td); 1869 } else if (TD_IS_RUNNING(td)) { 1870 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1871 srqflag = preempted ? 1872 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1873 SRQ_OURSELF|SRQ_YIELDING; 1874#ifdef SMP 1875 if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu)) 1876 ts->ts_cpu = sched_pickcpu(td, 0); 1877#endif 1878 if (ts->ts_cpu == cpuid) 1879 tdq_runq_add(tdq, td, srqflag); 1880 else { 1881 KASSERT(THREAD_CAN_MIGRATE(td) || 1882 (ts->ts_flags & TSF_BOUND) != 0, 1883 ("Thread %p shouldn't migrate", td)); 1884 mtx = sched_switch_migrate(tdq, td, srqflag); 1885 } 1886 } else { 1887 /* This thread must be going to sleep. */ 1888 TDQ_LOCK(tdq); 1889 mtx = thread_lock_block(td); 1890 tdq_load_rem(tdq, td); 1891 } 1892 /* 1893 * We enter here with the thread blocked and assigned to the 1894 * appropriate cpu run-queue or sleep-queue and with the current 1895 * thread-queue locked. 1896 */ 1897 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1898 newtd = choosethread(); 1899 /* 1900 * Call the MD code to switch contexts if necessary. 1901 */ 1902 if (td != newtd) { 1903#ifdef HWPMC_HOOKS 1904 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1905 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1906#endif 1907 SDT_PROBE2(sched, , , off_cpu, newtd, newtd->td_proc); 1908 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1909 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1910 sched_pctcpu_update(newtd->td_sched, 0); 1911 1912#ifdef KDTRACE_HOOKS 1913 /* 1914 * If DTrace has set the active vtime enum to anything 1915 * other than INACTIVE (0), then it should have set the 1916 * function to call. 1917 */ 1918 if (dtrace_vtime_active) 1919 (*dtrace_vtime_switch_func)(newtd); 1920#endif 1921 1922 cpu_switch(td, newtd, mtx); 1923 /* 1924 * We may return from cpu_switch on a different cpu. However, 1925 * we always return with td_lock pointing to the current cpu's 1926 * run queue lock. 1927 */ 1928 cpuid = PCPU_GET(cpuid); 1929 tdq = TDQ_CPU(cpuid); 1930 lock_profile_obtain_lock_success( 1931 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1932 1933 SDT_PROBE0(sched, , , on_cpu); 1934#ifdef HWPMC_HOOKS 1935 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1936 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1937#endif 1938 } else { 1939 thread_unblock_switch(td, mtx); 1940 SDT_PROBE0(sched, , , remain_cpu); 1941 } 1942 /* 1943 * Assert that all went well and return. 1944 */ 1945 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1946 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1947 td->td_oncpu = cpuid; 1948} 1949 1950/* 1951 * Adjust thread priorities as a result of a nice request. 1952 */ 1953void 1954sched_nice(struct proc *p, int nice) 1955{ 1956 struct thread *td; 1957 1958 PROC_LOCK_ASSERT(p, MA_OWNED); 1959 1960 p->p_nice = nice; 1961 FOREACH_THREAD_IN_PROC(p, td) { 1962 thread_lock(td); 1963 sched_priority(td); 1964 sched_prio(td, td->td_base_user_pri); 1965 thread_unlock(td); 1966 } 1967} 1968 1969/* 1970 * Record the sleep time for the interactivity scorer. 1971 */ 1972void 1973sched_sleep(struct thread *td, int prio) 1974{ 1975 1976 THREAD_LOCK_ASSERT(td, MA_OWNED); 1977 1978 td->td_slptick = ticks; 1979 if (TD_IS_SUSPENDED(td) || prio >= PSOCK) 1980 td->td_flags |= TDF_CANSWAP; 1981 if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1982 return; 1983 if (static_boost == 1 && prio) 1984 sched_prio(td, prio); 1985 else if (static_boost && td->td_priority > static_boost) 1986 sched_prio(td, static_boost); 1987} 1988 1989/* 1990 * Schedule a thread to resume execution and record how long it voluntarily 1991 * slept. We also update the pctcpu, interactivity, and priority. 1992 */ 1993void 1994sched_wakeup(struct thread *td) 1995{ 1996 struct td_sched *ts; 1997 int slptick; 1998 1999 THREAD_LOCK_ASSERT(td, MA_OWNED); 2000 ts = td->td_sched; 2001 td->td_flags &= ~TDF_CANSWAP; 2002 /* 2003 * If we slept for more than a tick update our interactivity and 2004 * priority. 2005 */ 2006 slptick = td->td_slptick; 2007 td->td_slptick = 0; 2008 if (slptick && slptick != ticks) { 2009 ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT; 2010 sched_interact_update(td); 2011 sched_pctcpu_update(ts, 0); 2012 } 2013 /* Reset the slice value after we sleep. */ 2014 ts->ts_slice = sched_slice; 2015 sched_add(td, SRQ_BORING); 2016} 2017 2018/* 2019 * Penalize the parent for creating a new child and initialize the child's 2020 * priority. 2021 */ 2022void 2023sched_fork(struct thread *td, struct thread *child) 2024{ 2025 THREAD_LOCK_ASSERT(td, MA_OWNED); 2026 sched_pctcpu_update(td->td_sched, 1); 2027 sched_fork_thread(td, child); 2028 /* 2029 * Penalize the parent and child for forking. 2030 */ 2031 sched_interact_fork(child); 2032 sched_priority(child); 2033 td->td_sched->ts_runtime += tickincr; 2034 sched_interact_update(td); 2035 sched_priority(td); 2036} 2037 2038/* 2039 * Fork a new thread, may be within the same process. 2040 */ 2041void 2042sched_fork_thread(struct thread *td, struct thread *child) 2043{ 2044 struct td_sched *ts; 2045 struct td_sched *ts2; 2046 2047 THREAD_LOCK_ASSERT(td, MA_OWNED); 2048 /* 2049 * Initialize child. 2050 */ 2051 ts = td->td_sched; 2052 ts2 = child->td_sched; 2053 child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 2054 child->td_cpuset = cpuset_ref(td->td_cpuset); 2055 ts2->ts_cpu = ts->ts_cpu; 2056 ts2->ts_flags = 0; 2057 /* 2058 * Grab our parents cpu estimation information. 2059 */ 2060 ts2->ts_ticks = ts->ts_ticks; 2061 ts2->ts_ltick = ts->ts_ltick; 2062 ts2->ts_ftick = ts->ts_ftick; 2063 /* 2064 * Do not inherit any borrowed priority from the parent. 2065 */ 2066 child->td_priority = child->td_base_pri; 2067 /* 2068 * And update interactivity score. 2069 */ 2070 ts2->ts_slptime = ts->ts_slptime; 2071 ts2->ts_runtime = ts->ts_runtime; 2072 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 2073#ifdef KTR 2074 bzero(ts2->ts_name, sizeof(ts2->ts_name)); 2075#endif 2076} 2077 2078/* 2079 * Adjust the priority class of a thread. 2080 */ 2081void 2082sched_class(struct thread *td, int class) 2083{ 2084 2085 THREAD_LOCK_ASSERT(td, MA_OWNED); 2086 if (td->td_pri_class == class) 2087 return; 2088 td->td_pri_class = class; 2089} 2090 2091/* 2092 * Return some of the child's priority and interactivity to the parent. 2093 */ 2094void 2095sched_exit(struct proc *p, struct thread *child) 2096{ 2097 struct thread *td; 2098 2099 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit", 2100 "prio:%d", child->td_priority); 2101 PROC_LOCK_ASSERT(p, MA_OWNED); 2102 td = FIRST_THREAD_IN_PROC(p); 2103 sched_exit_thread(td, child); 2104} 2105 2106/* 2107 * Penalize another thread for the time spent on this one. This helps to 2108 * worsen the priority and interactivity of processes which schedule batch 2109 * jobs such as make. This has little effect on the make process itself but 2110 * causes new processes spawned by it to receive worse scores immediately. 2111 */ 2112void 2113sched_exit_thread(struct thread *td, struct thread *child) 2114{ 2115 2116 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit", 2117 "prio:%d", child->td_priority); 2118 /* 2119 * Give the child's runtime to the parent without returning the 2120 * sleep time as a penalty to the parent. This causes shells that 2121 * launch expensive things to mark their children as expensive. 2122 */ 2123 thread_lock(td); 2124 td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2125 sched_interact_update(td); 2126 sched_priority(td); 2127 thread_unlock(td); 2128} 2129 2130void 2131sched_preempt(struct thread *td) 2132{ 2133 struct tdq *tdq; 2134 2135 SDT_PROBE2(sched, , , surrender, td, td->td_proc); 2136 2137 thread_lock(td); 2138 tdq = TDQ_SELF(); 2139 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2140 tdq->tdq_ipipending = 0; 2141 if (td->td_priority > tdq->tdq_lowpri) { 2142 int flags; 2143 2144 flags = SW_INVOL | SW_PREEMPT; 2145 if (td->td_critnest > 1) 2146 td->td_owepreempt = 1; 2147 else if (TD_IS_IDLETHREAD(td)) 2148 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL); 2149 else 2150 mi_switch(flags | SWT_REMOTEPREEMPT, NULL); 2151 } 2152 thread_unlock(td); 2153} 2154 2155/* 2156 * Fix priorities on return to user-space. Priorities may be elevated due 2157 * to static priorities in msleep() or similar. 2158 */ 2159void 2160sched_userret(struct thread *td) 2161{ 2162 /* 2163 * XXX we cheat slightly on the locking here to avoid locking in 2164 * the usual case. Setting td_priority here is essentially an 2165 * incomplete workaround for not setting it properly elsewhere. 2166 * Now that some interrupt handlers are threads, not setting it 2167 * properly elsewhere can clobber it in the window between setting 2168 * it here and returning to user mode, so don't waste time setting 2169 * it perfectly here. 2170 */ 2171 KASSERT((td->td_flags & TDF_BORROWING) == 0, 2172 ("thread with borrowed priority returning to userland")); 2173 if (td->td_priority != td->td_user_pri) { 2174 thread_lock(td); 2175 td->td_priority = td->td_user_pri; 2176 td->td_base_pri = td->td_user_pri; 2177 tdq_setlowpri(TDQ_SELF(), td); 2178 thread_unlock(td); 2179 } 2180} 2181 2182/* 2183 * Handle a stathz tick. This is really only relevant for timeshare 2184 * threads. 2185 */ 2186void 2187sched_clock(struct thread *td) 2188{ 2189 struct tdq *tdq; 2190 struct td_sched *ts; 2191 2192 THREAD_LOCK_ASSERT(td, MA_OWNED); 2193 tdq = TDQ_SELF(); 2194#ifdef SMP 2195 /* 2196 * We run the long term load balancer infrequently on the first cpu. 2197 */ 2198 if (balance_tdq == tdq) { 2199 if (balance_ticks && --balance_ticks == 0) 2200 sched_balance(); 2201 } 2202#endif 2203 /* 2204 * Save the old switch count so we have a record of the last ticks 2205 * activity. Initialize the new switch count based on our load. 2206 * If there is some activity seed it to reflect that. 2207 */ 2208 tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2209 tdq->tdq_switchcnt = tdq->tdq_load; 2210 /* 2211 * Advance the insert index once for each tick to ensure that all 2212 * threads get a chance to run. 2213 */ 2214 if (tdq->tdq_idx == tdq->tdq_ridx) { 2215 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2216 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2217 tdq->tdq_ridx = tdq->tdq_idx; 2218 } 2219 ts = td->td_sched; 2220 sched_pctcpu_update(ts, 1); 2221 if (td->td_pri_class & PRI_FIFO_BIT) 2222 return; 2223 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) { 2224 /* 2225 * We used a tick; charge it to the thread so 2226 * that we can compute our interactivity. 2227 */ 2228 td->td_sched->ts_runtime += tickincr; 2229 sched_interact_update(td); 2230 sched_priority(td); 2231 } 2232 2233 /* 2234 * Force a context switch if the current thread has used up a full 2235 * time slice (default is 100ms). 2236 */ 2237 if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) { 2238 ts->ts_slice = sched_slice; 2239 td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND; 2240 } 2241} 2242 2243/* 2244 * Called once per hz tick. 2245 */ 2246void 2247sched_tick(int cnt) 2248{ 2249 2250} 2251 2252/* 2253 * Return whether the current CPU has runnable tasks. Used for in-kernel 2254 * cooperative idle threads. 2255 */ 2256int 2257sched_runnable(void) 2258{ 2259 struct tdq *tdq; 2260 int load; 2261 2262 load = 1; 2263 2264 tdq = TDQ_SELF(); 2265 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2266 if (tdq->tdq_load > 0) 2267 goto out; 2268 } else 2269 if (tdq->tdq_load - 1 > 0) 2270 goto out; 2271 load = 0; 2272out: 2273 return (load); 2274} 2275 2276/* 2277 * Choose the highest priority thread to run. The thread is removed from 2278 * the run-queue while running however the load remains. For SMP we set 2279 * the tdq in the global idle bitmask if it idles here. 2280 */ 2281struct thread * 2282sched_choose(void) 2283{ 2284 struct thread *td; 2285 struct tdq *tdq; 2286 2287 tdq = TDQ_SELF(); 2288 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2289 td = tdq_choose(tdq); 2290 if (td) { 2291 tdq_runq_rem(tdq, td); 2292 tdq->tdq_lowpri = td->td_priority; 2293 return (td); 2294 } 2295 tdq->tdq_lowpri = PRI_MAX_IDLE; 2296 return (PCPU_GET(idlethread)); 2297} 2298 2299/* 2300 * Set owepreempt if necessary. Preemption never happens directly in ULE, 2301 * we always request it once we exit a critical section. 2302 */ 2303static inline void 2304sched_setpreempt(struct thread *td) 2305{ 2306 struct thread *ctd; 2307 int cpri; 2308 int pri; 2309 2310 THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2311 2312 ctd = curthread; 2313 pri = td->td_priority; 2314 cpri = ctd->td_priority; 2315 if (pri < cpri) 2316 ctd->td_flags |= TDF_NEEDRESCHED; 2317 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2318 return; 2319 if (!sched_shouldpreempt(pri, cpri, 0)) 2320 return; 2321 ctd->td_owepreempt = 1; 2322} 2323 2324/* 2325 * Add a thread to a thread queue. Select the appropriate runq and add the 2326 * thread to it. This is the internal function called when the tdq is 2327 * predetermined. 2328 */ 2329void 2330tdq_add(struct tdq *tdq, struct thread *td, int flags) 2331{ 2332 2333 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2334 KASSERT((td->td_inhibitors == 0), 2335 ("sched_add: trying to run inhibited thread")); 2336 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2337 ("sched_add: bad thread state")); 2338 KASSERT(td->td_flags & TDF_INMEM, 2339 ("sched_add: thread swapped out")); 2340 2341 if (td->td_priority < tdq->tdq_lowpri) 2342 tdq->tdq_lowpri = td->td_priority; 2343 tdq_runq_add(tdq, td, flags); 2344 tdq_load_add(tdq, td); 2345} 2346 2347/* 2348 * Select the target thread queue and add a thread to it. Request 2349 * preemption or IPI a remote processor if required. 2350 */ 2351void 2352sched_add(struct thread *td, int flags) 2353{ 2354 struct tdq *tdq; 2355#ifdef SMP 2356 int cpu; 2357#endif 2358 2359 KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 2360 "prio:%d", td->td_priority, KTR_ATTR_LINKED, 2361 sched_tdname(curthread)); 2362 KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 2363 KTR_ATTR_LINKED, sched_tdname(td)); 2364 SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 2365 flags & SRQ_PREEMPTED); 2366 THREAD_LOCK_ASSERT(td, MA_OWNED); 2367 /* 2368 * Recalculate the priority before we select the target cpu or 2369 * run-queue. 2370 */ 2371 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2372 sched_priority(td); 2373#ifdef SMP 2374 /* 2375 * Pick the destination cpu and if it isn't ours transfer to the 2376 * target cpu. 2377 */ 2378 cpu = sched_pickcpu(td, flags); 2379 tdq = sched_setcpu(td, cpu, flags); 2380 tdq_add(tdq, td, flags); 2381 if (cpu != PCPU_GET(cpuid)) { 2382 tdq_notify(tdq, td); 2383 return; 2384 } 2385#else 2386 tdq = TDQ_SELF(); 2387 TDQ_LOCK(tdq); 2388 /* 2389 * Now that the thread is moving to the run-queue, set the lock 2390 * to the scheduler's lock. 2391 */ 2392 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2393 tdq_add(tdq, td, flags); 2394#endif 2395 if (!(flags & SRQ_YIELDING)) 2396 sched_setpreempt(td); 2397} 2398 2399/* 2400 * Remove a thread from a run-queue without running it. This is used 2401 * when we're stealing a thread from a remote queue. Otherwise all threads 2402 * exit by calling sched_exit_thread() and sched_throw() themselves. 2403 */ 2404void 2405sched_rem(struct thread *td) 2406{ 2407 struct tdq *tdq; 2408 2409 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 2410 "prio:%d", td->td_priority); 2411 SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); 2412 tdq = TDQ_CPU(td->td_sched->ts_cpu); 2413 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2414 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2415 KASSERT(TD_ON_RUNQ(td), 2416 ("sched_rem: thread not on run queue")); 2417 tdq_runq_rem(tdq, td); 2418 tdq_load_rem(tdq, td); 2419 TD_SET_CAN_RUN(td); 2420 if (td->td_priority == tdq->tdq_lowpri) 2421 tdq_setlowpri(tdq, NULL); 2422} 2423 2424/* 2425 * Fetch cpu utilization information. Updates on demand. 2426 */ 2427fixpt_t 2428sched_pctcpu(struct thread *td) 2429{ 2430 fixpt_t pctcpu; 2431 struct td_sched *ts; 2432 2433 pctcpu = 0; 2434 ts = td->td_sched; 2435 if (ts == NULL) 2436 return (0); 2437 2438 THREAD_LOCK_ASSERT(td, MA_OWNED); 2439 sched_pctcpu_update(ts, TD_IS_RUNNING(td)); 2440 if (ts->ts_ticks) { 2441 int rtick; 2442 2443 /* How many rtick per second ? */ 2444 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2445 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2446 } 2447 2448 return (pctcpu); 2449} 2450 2451/* 2452 * Enforce affinity settings for a thread. Called after adjustments to 2453 * cpumask. 2454 */ 2455void 2456sched_affinity(struct thread *td) 2457{ 2458#ifdef SMP 2459 struct td_sched *ts; 2460 2461 THREAD_LOCK_ASSERT(td, MA_OWNED); 2462 ts = td->td_sched; 2463 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2464 return; 2465 if (TD_ON_RUNQ(td)) { 2466 sched_rem(td); 2467 sched_add(td, SRQ_BORING); 2468 return; 2469 } 2470 if (!TD_IS_RUNNING(td)) 2471 return; 2472 /* 2473 * Force a switch before returning to userspace. If the 2474 * target thread is not running locally send an ipi to force 2475 * the issue. 2476 */ 2477 td->td_flags |= TDF_NEEDRESCHED; 2478 if (td != curthread) 2479 ipi_cpu(ts->ts_cpu, IPI_PREEMPT); 2480#endif 2481} 2482 2483/* 2484 * Bind a thread to a target cpu. 2485 */ 2486void 2487sched_bind(struct thread *td, int cpu) 2488{ 2489 struct td_sched *ts; 2490 2491 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2492 KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 2493 ts = td->td_sched; 2494 if (ts->ts_flags & TSF_BOUND) 2495 sched_unbind(td); 2496 KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td)); 2497 ts->ts_flags |= TSF_BOUND; 2498 sched_pin(); 2499 if (PCPU_GET(cpuid) == cpu) 2500 return; 2501 ts->ts_cpu = cpu; 2502 /* When we return from mi_switch we'll be on the correct cpu. */ 2503 mi_switch(SW_VOL, NULL); 2504} 2505 2506/* 2507 * Release a bound thread. 2508 */ 2509void 2510sched_unbind(struct thread *td) 2511{ 2512 struct td_sched *ts; 2513 2514 THREAD_LOCK_ASSERT(td, MA_OWNED); 2515 KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 2516 ts = td->td_sched; 2517 if ((ts->ts_flags & TSF_BOUND) == 0) 2518 return; 2519 ts->ts_flags &= ~TSF_BOUND; 2520 sched_unpin(); 2521} 2522 2523int 2524sched_is_bound(struct thread *td) 2525{ 2526 THREAD_LOCK_ASSERT(td, MA_OWNED); 2527 return (td->td_sched->ts_flags & TSF_BOUND); 2528} 2529 2530/* 2531 * Basic yield call. 2532 */ 2533void 2534sched_relinquish(struct thread *td) 2535{ 2536 thread_lock(td); 2537 mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 2538 thread_unlock(td); 2539} 2540 2541/* 2542 * Return the total system load. 2543 */ 2544int 2545sched_load(void) 2546{ 2547#ifdef SMP 2548 int total; 2549 int i; 2550 2551 total = 0; 2552 CPU_FOREACH(i) 2553 total += TDQ_CPU(i)->tdq_sysload; 2554 return (total); 2555#else 2556 return (TDQ_SELF()->tdq_sysload); 2557#endif 2558} 2559 2560int 2561sched_sizeof_proc(void) 2562{ 2563 return (sizeof(struct proc)); 2564} 2565 2566int 2567sched_sizeof_thread(void) 2568{ 2569 return (sizeof(struct thread) + sizeof(struct td_sched)); 2570} 2571 2572#ifdef SMP 2573#define TDQ_IDLESPIN(tdq) \ 2574 ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0) 2575#else 2576#define TDQ_IDLESPIN(tdq) 1 2577#endif 2578 2579/* 2580 * The actual idle process. 2581 */ 2582void 2583sched_idletd(void *dummy) 2584{ 2585 struct thread *td; 2586 struct tdq *tdq; 2587 int oldswitchcnt, switchcnt; 2588 int i; 2589 2590 mtx_assert(&Giant, MA_NOTOWNED); 2591 td = curthread; 2592 tdq = TDQ_SELF(); 2593 oldswitchcnt = -1; 2594 for (;;) { 2595 if (tdq->tdq_load) { 2596 thread_lock(td); 2597 mi_switch(SW_VOL | SWT_IDLE, NULL); 2598 thread_unlock(td); 2599 } 2600 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2601#ifdef SMP 2602 if (switchcnt != oldswitchcnt) { 2603 oldswitchcnt = switchcnt; 2604 if (tdq_idled(tdq) == 0) 2605 continue; 2606 } 2607 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2608#else 2609 oldswitchcnt = switchcnt; 2610#endif 2611 /* 2612 * If we're switching very frequently, spin while checking 2613 * for load rather than entering a low power state that 2614 * may require an IPI. However, don't do any busy 2615 * loops while on SMT machines as this simply steals 2616 * cycles from cores doing useful work. 2617 */ 2618 if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) { 2619 for (i = 0; i < sched_idlespins; i++) { 2620 if (tdq->tdq_load) 2621 break; 2622 cpu_spinwait(); 2623 } 2624 } 2625 2626 /* If there was context switch during spin, restart it. */ 2627 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2628 if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt) 2629 continue; 2630 2631 /* Run main MD idle handler. */ 2632 tdq->tdq_cpu_idle = 1; 2633 cpu_idle(switchcnt * 4 > sched_idlespinthresh); 2634 tdq->tdq_cpu_idle = 0; 2635 2636 /* 2637 * Account thread-less hardware interrupts and 2638 * other wakeup reasons equal to context switches. 2639 */ 2640 switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2641 if (switchcnt != oldswitchcnt) 2642 continue; 2643 tdq->tdq_switchcnt++; 2644 oldswitchcnt++; 2645 } 2646} 2647 2648/* 2649 * A CPU is entering for the first time or a thread is exiting. 2650 */ 2651void 2652sched_throw(struct thread *td) 2653{ 2654 struct thread *newtd; 2655 struct tdq *tdq; 2656 2657 tdq = TDQ_SELF(); 2658 if (td == NULL) { 2659 /* Correct spinlock nesting and acquire the correct lock. */ 2660 TDQ_LOCK(tdq); 2661 spinlock_exit(); 2662 PCPU_SET(switchtime, cpu_ticks()); 2663 PCPU_SET(switchticks, ticks); 2664 } else { 2665 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2666 tdq_load_rem(tdq, td); 2667 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2668 } 2669 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2670 newtd = choosethread(); 2671 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2672 cpu_throw(td, newtd); /* doesn't return */ 2673} 2674 2675/* 2676 * This is called from fork_exit(). Just acquire the correct locks and 2677 * let fork do the rest of the work. 2678 */ 2679void 2680sched_fork_exit(struct thread *td) 2681{ 2682 struct td_sched *ts; 2683 struct tdq *tdq; 2684 int cpuid; 2685 2686 /* 2687 * Finish setting up thread glue so that it begins execution in a 2688 * non-nested critical section with the scheduler lock held. 2689 */ 2690 cpuid = PCPU_GET(cpuid); 2691 tdq = TDQ_CPU(cpuid); 2692 ts = td->td_sched; 2693 if (TD_IS_IDLETHREAD(td)) 2694 td->td_lock = TDQ_LOCKPTR(tdq); 2695 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2696 td->td_oncpu = cpuid; 2697 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2698 lock_profile_obtain_lock_success( 2699 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2700} 2701 2702/* 2703 * Create on first use to catch odd startup conditons. 2704 */ 2705char * 2706sched_tdname(struct thread *td) 2707{ 2708#ifdef KTR 2709 struct td_sched *ts; 2710 2711 ts = td->td_sched; 2712 if (ts->ts_name[0] == '\0') 2713 snprintf(ts->ts_name, sizeof(ts->ts_name), 2714 "%s tid %d", td->td_name, td->td_tid); 2715 return (ts->ts_name); 2716#else 2717 return (td->td_name); 2718#endif 2719} 2720 2721#ifdef KTR 2722void 2723sched_clear_tdname(struct thread *td) 2724{ 2725 struct td_sched *ts; 2726 2727 ts = td->td_sched; 2728 ts->ts_name[0] = '\0'; 2729} 2730#endif 2731 2732#ifdef SMP 2733 2734/* 2735 * Build the CPU topology dump string. Is recursively called to collect 2736 * the topology tree. 2737 */ 2738static int 2739sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg, 2740 int indent) 2741{ 2742 char cpusetbuf[CPUSETBUFSIZ]; 2743 int i, first; 2744 2745 sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent, 2746 "", 1 + indent / 2, cg->cg_level); 2747 sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "", 2748 cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask)); 2749 first = TRUE; 2750 for (i = 0; i < MAXCPU; i++) { 2751 if (CPU_ISSET(i, &cg->cg_mask)) { 2752 if (!first) 2753 sbuf_printf(sb, ", "); 2754 else 2755 first = FALSE; 2756 sbuf_printf(sb, "%d", i); 2757 } 2758 } 2759 sbuf_printf(sb, "</cpu>\n"); 2760 2761 if (cg->cg_flags != 0) { 2762 sbuf_printf(sb, "%*s <flags>", indent, ""); 2763 if ((cg->cg_flags & CG_FLAG_HTT) != 0) 2764 sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>"); 2765 if ((cg->cg_flags & CG_FLAG_THREAD) != 0) 2766 sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>"); 2767 if ((cg->cg_flags & CG_FLAG_SMT) != 0) 2768 sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>"); 2769 sbuf_printf(sb, "</flags>\n"); 2770 } 2771 2772 if (cg->cg_children > 0) { 2773 sbuf_printf(sb, "%*s <children>\n", indent, ""); 2774 for (i = 0; i < cg->cg_children; i++) 2775 sysctl_kern_sched_topology_spec_internal(sb, 2776 &cg->cg_child[i], indent+2); 2777 sbuf_printf(sb, "%*s </children>\n", indent, ""); 2778 } 2779 sbuf_printf(sb, "%*s</group>\n", indent, ""); 2780 return (0); 2781} 2782 2783/* 2784 * Sysctl handler for retrieving topology dump. It's a wrapper for 2785 * the recursive sysctl_kern_smp_topology_spec_internal(). 2786 */ 2787static int 2788sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS) 2789{ 2790 struct sbuf *topo; 2791 int err; 2792 2793 KASSERT(cpu_top != NULL, ("cpu_top isn't initialized")); 2794 2795 topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND); 2796 if (topo == NULL) 2797 return (ENOMEM); 2798 2799 sbuf_printf(topo, "<groups>\n"); 2800 err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1); 2801 sbuf_printf(topo, "</groups>\n"); 2802 2803 if (err == 0) { 2804 sbuf_finish(topo); 2805 err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo)); 2806 } 2807 sbuf_delete(topo); 2808 return (err); 2809} 2810 2811#endif 2812 2813static int 2814sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 2815{ 2816 int error, new_val, period; 2817 2818 period = 1000000 / realstathz; 2819 new_val = period * sched_slice; 2820 error = sysctl_handle_int(oidp, &new_val, 0, req); 2821 if (error != 0 || req->newptr == NULL) 2822 return (error); 2823 if (new_val <= 0) 2824 return (EINVAL); 2825 sched_slice = imax(1, (new_val + period / 2) / period); 2826 hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 2827 realstathz); 2828 return (0); 2829} 2830 2831SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2832SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2833 "Scheduler name"); 2834SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW, 2835 NULL, 0, sysctl_kern_quantum, "I", 2836 "Quantum for timeshare threads in microseconds"); 2837SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2838 "Quantum for timeshare threads in stathz ticks"); 2839SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2840 "Interactivity score threshold"); 2841SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, 2842 &preempt_thresh, 0, 2843 "Maximal (lowest) priority for preemption"); 2844SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0, 2845 "Assign static kernel priorities to sleeping threads"); 2846SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0, 2847 "Number of times idle thread will spin waiting for new work"); 2848SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, 2849 &sched_idlespinthresh, 0, 2850 "Threshold before we will permit idle thread spinning"); 2851#ifdef SMP 2852SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2853 "Number of hz ticks to keep thread affinity for"); 2854SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2855 "Enables the long-term load balancer"); 2856SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2857 &balance_interval, 0, 2858 "Average period in stathz ticks to run the long-term balancer"); 2859SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2860 "Attempts to steal work from other cores before idling"); 2861SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2862 "Minimum load on remote CPU before we'll steal"); 2863SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING | 2864 CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A", 2865 "XML dump of detected CPU topology"); 2866#endif 2867 2868/* ps compat. All cpu percentages from ULE are weighted. */ 2869static int ccpu = 0; 2870SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2871