sched_ule.c revision 178272
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: head/sys/kern/sched_ule.c 178272 2008-04-17 04:20:10Z jeff $"); 40 41#include "opt_hwpmc_hooks.h" 42#include "opt_sched.h" 43 44#include <sys/param.h> 45#include <sys/systm.h> 46#include <sys/kdb.h> 47#include <sys/kernel.h> 48#include <sys/ktr.h> 49#include <sys/lock.h> 50#include <sys/mutex.h> 51#include <sys/proc.h> 52#include <sys/resource.h> 53#include <sys/resourcevar.h> 54#include <sys/sched.h> 55#include <sys/smp.h> 56#include <sys/sx.h> 57#include <sys/sysctl.h> 58#include <sys/sysproto.h> 59#include <sys/turnstile.h> 60#include <sys/umtx.h> 61#include <sys/vmmeter.h> 62#include <sys/cpuset.h> 63#ifdef KTRACE 64#include <sys/uio.h> 65#include <sys/ktrace.h> 66#endif 67 68#ifdef HWPMC_HOOKS 69#include <sys/pmckern.h> 70#endif 71 72#include <machine/cpu.h> 73#include <machine/smp.h> 74 75#if defined(__sparc64__) || defined(__mips__) 76#error "This architecture is not currently compatible with ULE" 77#endif 78 79#define KTR_ULE 0 80 81/* 82 * Thread scheduler specific section. All fields are protected 83 * by the thread lock. 84 */ 85struct td_sched { 86 struct runq *ts_runq; /* Run-queue we're queued on. */ 87 short ts_flags; /* TSF_* flags. */ 88 u_char ts_cpu; /* CPU that we have affinity for. */ 89 int ts_rltick; /* Real last tick, for affinity. */ 90 int ts_slice; /* Ticks of slice remaining. */ 91 u_int ts_slptime; /* Number of ticks we vol. slept */ 92 u_int ts_runtime; /* Number of ticks we were running */ 93 int ts_ltick; /* Last tick that we were running on */ 94 int ts_ftick; /* First tick that we were running on */ 95 int ts_ticks; /* Tick count */ 96}; 97/* flags kept in ts_flags */ 98#define TSF_BOUND 0x0001 /* Thread can not migrate. */ 99#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 100 101static struct td_sched td_sched0; 102 103#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 104#define THREAD_CAN_SCHED(td, cpu) \ 105 CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 106 107/* 108 * Cpu percentage computation macros and defines. 109 * 110 * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 111 * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 112 * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 113 * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 114 * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 115 * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 116 */ 117#define SCHED_TICK_SECS 10 118#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 119#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 120#define SCHED_TICK_SHIFT 10 121#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 122#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 123 124/* 125 * These macros determine priorities for non-interactive threads. They are 126 * assigned a priority based on their recent cpu utilization as expressed 127 * by the ratio of ticks to the tick total. NHALF priorities at the start 128 * and end of the MIN to MAX timeshare range are only reachable with negative 129 * or positive nice respectively. 130 * 131 * PRI_RANGE: Priority range for utilization dependent priorities. 132 * PRI_NRESV: Number of nice values. 133 * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 134 * PRI_NICE: Determines the part of the priority inherited from nice. 135 */ 136#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 137#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 138#define SCHED_PRI_MIN (PRI_MIN_TIMESHARE + SCHED_PRI_NHALF) 139#define SCHED_PRI_MAX (PRI_MAX_TIMESHARE - SCHED_PRI_NHALF) 140#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN) 141#define SCHED_PRI_TICKS(ts) \ 142 (SCHED_TICK_HZ((ts)) / \ 143 (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 144#define SCHED_PRI_NICE(nice) (nice) 145 146/* 147 * These determine the interactivity of a process. Interactivity differs from 148 * cpu utilization in that it expresses the voluntary time slept vs time ran 149 * while cpu utilization includes all time not running. This more accurately 150 * models the intent of the thread. 151 * 152 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 153 * before throttling back. 154 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 155 * INTERACT_MAX: Maximum interactivity value. Smaller is better. 156 * INTERACT_THRESH: Threshhold for placement on the current runq. 157 */ 158#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 159#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 160#define SCHED_INTERACT_MAX (100) 161#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 162#define SCHED_INTERACT_THRESH (30) 163 164/* 165 * tickincr: Converts a stathz tick into a hz domain scaled by 166 * the shift factor. Without the shift the error rate 167 * due to rounding would be unacceptably high. 168 * realstathz: stathz is sometimes 0 and run off of hz. 169 * sched_slice: Runtime of each thread before rescheduling. 170 * preempt_thresh: Priority threshold for preemption and remote IPIs. 171 */ 172static int sched_interact = SCHED_INTERACT_THRESH; 173static int realstathz; 174static int tickincr; 175static int sched_slice = 1; 176#ifdef PREEMPTION 177#ifdef FULL_PREEMPTION 178static int preempt_thresh = PRI_MAX_IDLE; 179#else 180static int preempt_thresh = PRI_MIN_KERN; 181#endif 182#else 183static int preempt_thresh = 0; 184#endif 185static int static_boost = PRI_MIN_TIMESHARE; 186 187/* 188 * tdq - per processor runqs and statistics. All fields are protected by the 189 * tdq_lock. The load and lowpri may be accessed without to avoid excess 190 * locking in sched_pickcpu(); 191 */ 192struct tdq { 193 /* Ordered to improve efficiency of cpu_search() and switch(). */ 194 struct mtx tdq_lock; /* run queue lock. */ 195 struct cpu_group *tdq_cg; /* Pointer to cpu topology. */ 196 int tdq_load; /* Aggregate load. */ 197 int tdq_sysload; /* For loadavg, !ITHD load. */ 198 int tdq_transferable; /* Transferable thread count. */ 199 u_char tdq_lowpri; /* Lowest priority thread. */ 200 u_char tdq_ipipending; /* IPI pending. */ 201 u_char tdq_idx; /* Current insert index. */ 202 u_char tdq_ridx; /* Current removal index. */ 203 struct runq tdq_realtime; /* real-time run queue. */ 204 struct runq tdq_timeshare; /* timeshare run queue. */ 205 struct runq tdq_idle; /* Queue of IDLE threads. */ 206 char tdq_name[sizeof("sched lock") + 6]; 207} __aligned(64); 208 209 210#ifdef SMP 211struct cpu_group *cpu_top; 212 213#define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000)) 214#define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity)) 215 216/* 217 * Run-time tunables. 218 */ 219static int rebalance = 1; 220static int balance_interval = 128; /* Default set in sched_initticks(). */ 221static int affinity; 222static int steal_htt = 1; 223static int steal_idle = 1; 224static int steal_thresh = 2; 225 226/* 227 * One thread queue per processor. 228 */ 229static struct tdq tdq_cpu[MAXCPU]; 230static struct tdq *balance_tdq; 231static int balance_ticks; 232 233#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 234#define TDQ_CPU(x) (&tdq_cpu[(x)]) 235#define TDQ_ID(x) ((int)((x) - tdq_cpu)) 236#else /* !SMP */ 237static struct tdq tdq_cpu; 238 239#define TDQ_ID(x) (0) 240#define TDQ_SELF() (&tdq_cpu) 241#define TDQ_CPU(x) (&tdq_cpu) 242#endif 243 244#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 245#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 246#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 247#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 248#define TDQ_LOCKPTR(t) (&(t)->tdq_lock) 249 250static void sched_priority(struct thread *); 251static void sched_thread_priority(struct thread *, u_char); 252static int sched_interact_score(struct thread *); 253static void sched_interact_update(struct thread *); 254static void sched_interact_fork(struct thread *); 255static void sched_pctcpu_update(struct td_sched *); 256 257/* Operations on per processor queues */ 258static struct thread *tdq_choose(struct tdq *); 259static void tdq_setup(struct tdq *); 260static void tdq_load_add(struct tdq *, struct thread *); 261static void tdq_load_rem(struct tdq *, struct thread *); 262static __inline void tdq_runq_add(struct tdq *, struct thread *, int); 263static __inline void tdq_runq_rem(struct tdq *, struct thread *); 264static inline int sched_shouldpreempt(int, int, int); 265void tdq_print(int cpu); 266static void runq_print(struct runq *rq); 267static void tdq_add(struct tdq *, struct thread *, int); 268#ifdef SMP 269static int tdq_move(struct tdq *, struct tdq *); 270static int tdq_idled(struct tdq *); 271static void tdq_notify(struct tdq *, struct thread *); 272static struct thread *tdq_steal(struct tdq *, int); 273static struct thread *runq_steal(struct runq *, int); 274static int sched_pickcpu(struct thread *, int); 275static void sched_balance(void); 276static int sched_balance_pair(struct tdq *, struct tdq *); 277static inline struct tdq *sched_setcpu(struct thread *, int, int); 278static inline struct mtx *thread_block_switch(struct thread *); 279static inline void thread_unblock_switch(struct thread *, struct mtx *); 280static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 281#endif 282 283static void sched_setup(void *dummy); 284SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 285 286static void sched_initticks(void *dummy); 287SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 288 NULL); 289 290/* 291 * Print the threads waiting on a run-queue. 292 */ 293static void 294runq_print(struct runq *rq) 295{ 296 struct rqhead *rqh; 297 struct thread *td; 298 int pri; 299 int j; 300 int i; 301 302 for (i = 0; i < RQB_LEN; i++) { 303 printf("\t\trunq bits %d 0x%zx\n", 304 i, rq->rq_status.rqb_bits[i]); 305 for (j = 0; j < RQB_BPW; j++) 306 if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 307 pri = j + (i << RQB_L2BPW); 308 rqh = &rq->rq_queues[pri]; 309 TAILQ_FOREACH(td, rqh, td_runq) { 310 printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 311 td, td->td_name, td->td_priority, 312 td->td_rqindex, pri); 313 } 314 } 315 } 316} 317 318/* 319 * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 320 */ 321void 322tdq_print(int cpu) 323{ 324 struct tdq *tdq; 325 326 tdq = TDQ_CPU(cpu); 327 328 printf("tdq %d:\n", TDQ_ID(tdq)); 329 printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 330 printf("\tLock name: %s\n", tdq->tdq_name); 331 printf("\tload: %d\n", tdq->tdq_load); 332 printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 333 printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 334 printf("\trealtime runq:\n"); 335 runq_print(&tdq->tdq_realtime); 336 printf("\ttimeshare runq:\n"); 337 runq_print(&tdq->tdq_timeshare); 338 printf("\tidle runq:\n"); 339 runq_print(&tdq->tdq_idle); 340 printf("\tload transferable: %d\n", tdq->tdq_transferable); 341 printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 342} 343 344static inline int 345sched_shouldpreempt(int pri, int cpri, int remote) 346{ 347 /* 348 * If the new priority is not better than the current priority there is 349 * nothing to do. 350 */ 351 if (pri >= cpri) 352 return (0); 353 /* 354 * Always preempt idle. 355 */ 356 if (cpri >= PRI_MIN_IDLE) 357 return (1); 358 /* 359 * If preemption is disabled don't preempt others. 360 */ 361 if (preempt_thresh == 0) 362 return (0); 363 /* 364 * Preempt if we exceed the threshold. 365 */ 366 if (pri <= preempt_thresh) 367 return (1); 368 /* 369 * If we're realtime or better and there is timeshare or worse running 370 * preempt only remote processors. 371 */ 372 if (remote && pri <= PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME) 373 return (1); 374 return (0); 375} 376 377#define TS_RQ_PPQ (((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS) 378/* 379 * Add a thread to the actual run-queue. Keeps transferable counts up to 380 * date with what is actually on the run-queue. Selects the correct 381 * queue position for timeshare threads. 382 */ 383static __inline void 384tdq_runq_add(struct tdq *tdq, struct thread *td, int flags) 385{ 386 struct td_sched *ts; 387 u_char pri; 388 389 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 390 THREAD_LOCK_ASSERT(td, MA_OWNED); 391 392 pri = td->td_priority; 393 ts = td->td_sched; 394 TD_SET_RUNQ(td); 395 if (THREAD_CAN_MIGRATE(td)) { 396 tdq->tdq_transferable++; 397 ts->ts_flags |= TSF_XFERABLE; 398 } 399 if (pri <= PRI_MAX_REALTIME) { 400 ts->ts_runq = &tdq->tdq_realtime; 401 } else if (pri <= PRI_MAX_TIMESHARE) { 402 ts->ts_runq = &tdq->tdq_timeshare; 403 KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE, 404 ("Invalid priority %d on timeshare runq", pri)); 405 /* 406 * This queue contains only priorities between MIN and MAX 407 * realtime. Use the whole queue to represent these values. 408 */ 409 if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 410 pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ; 411 pri = (pri + tdq->tdq_idx) % RQ_NQS; 412 /* 413 * This effectively shortens the queue by one so we 414 * can have a one slot difference between idx and 415 * ridx while we wait for threads to drain. 416 */ 417 if (tdq->tdq_ridx != tdq->tdq_idx && 418 pri == tdq->tdq_ridx) 419 pri = (unsigned char)(pri - 1) % RQ_NQS; 420 } else 421 pri = tdq->tdq_ridx; 422 runq_add_pri(ts->ts_runq, td, pri, flags); 423 return; 424 } else 425 ts->ts_runq = &tdq->tdq_idle; 426 runq_add(ts->ts_runq, td, flags); 427} 428 429/* 430 * Remove a thread from a run-queue. This typically happens when a thread 431 * is selected to run. Running threads are not on the queue and the 432 * transferable count does not reflect them. 433 */ 434static __inline void 435tdq_runq_rem(struct tdq *tdq, struct thread *td) 436{ 437 struct td_sched *ts; 438 439 ts = td->td_sched; 440 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 441 KASSERT(ts->ts_runq != NULL, 442 ("tdq_runq_remove: thread %p null ts_runq", td)); 443 if (ts->ts_flags & TSF_XFERABLE) { 444 tdq->tdq_transferable--; 445 ts->ts_flags &= ~TSF_XFERABLE; 446 } 447 if (ts->ts_runq == &tdq->tdq_timeshare) { 448 if (tdq->tdq_idx != tdq->tdq_ridx) 449 runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx); 450 else 451 runq_remove_idx(ts->ts_runq, td, NULL); 452 } else 453 runq_remove(ts->ts_runq, td); 454} 455 456/* 457 * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 458 * for this thread to the referenced thread queue. 459 */ 460static void 461tdq_load_add(struct tdq *tdq, struct thread *td) 462{ 463 464 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 465 THREAD_LOCK_ASSERT(td, MA_OWNED); 466 467 tdq->tdq_load++; 468 if ((td->td_proc->p_flag & P_NOLOAD) == 0) 469 tdq->tdq_sysload++; 470 CTR2(KTR_SCHED, "cpu %d load: %d", TDQ_ID(tdq), tdq->tdq_load); 471} 472 473/* 474 * Remove the load from a thread that is transitioning to a sleep state or 475 * exiting. 476 */ 477static void 478tdq_load_rem(struct tdq *tdq, struct thread *td) 479{ 480 481 THREAD_LOCK_ASSERT(td, MA_OWNED); 482 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 483 KASSERT(tdq->tdq_load != 0, 484 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 485 486 tdq->tdq_load--; 487 if ((td->td_proc->p_flag & P_NOLOAD) == 0) 488 tdq->tdq_sysload--; 489 CTR1(KTR_SCHED, "load: %d", tdq->tdq_load); 490} 491 492/* 493 * Set lowpri to its exact value by searching the run-queue and 494 * evaluating curthread. curthread may be passed as an optimization. 495 */ 496static void 497tdq_setlowpri(struct tdq *tdq, struct thread *ctd) 498{ 499 struct thread *td; 500 501 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 502 if (ctd == NULL) 503 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 504 td = tdq_choose(tdq); 505 if (td == NULL || td->td_priority > ctd->td_priority) 506 tdq->tdq_lowpri = ctd->td_priority; 507 else 508 tdq->tdq_lowpri = td->td_priority; 509} 510 511#ifdef SMP 512struct cpu_search { 513 cpumask_t cs_mask; /* Mask of valid cpus. */ 514 u_int cs_load; 515 u_int cs_cpu; 516 int cs_limit; /* Min priority for low min load for high. */ 517}; 518 519#define CPU_SEARCH_LOWEST 0x1 520#define CPU_SEARCH_HIGHEST 0x2 521#define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 522 523#define CPUMASK_FOREACH(cpu, mask) \ 524 for ((cpu) = 0; (cpu) < sizeof((mask)) * 8; (cpu)++) \ 525 if ((mask) & 1 << (cpu)) 526 527static __inline int cpu_search(struct cpu_group *cg, struct cpu_search *low, 528 struct cpu_search *high, const int match); 529int cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low); 530int cpu_search_highest(struct cpu_group *cg, struct cpu_search *high); 531int cpu_search_both(struct cpu_group *cg, struct cpu_search *low, 532 struct cpu_search *high); 533 534/* 535 * This routine compares according to the match argument and should be 536 * reduced in actual instantiations via constant propagation and dead code 537 * elimination. 538 */ 539static __inline int 540cpu_compare(int cpu, struct cpu_search *low, struct cpu_search *high, 541 const int match) 542{ 543 struct tdq *tdq; 544 545 tdq = TDQ_CPU(cpu); 546 if (match & CPU_SEARCH_LOWEST) 547 if (low->cs_mask & (1 << cpu) && 548 tdq->tdq_load < low->cs_load && 549 tdq->tdq_lowpri > low->cs_limit) { 550 low->cs_cpu = cpu; 551 low->cs_load = tdq->tdq_load; 552 } 553 if (match & CPU_SEARCH_HIGHEST) 554 if (high->cs_mask & (1 << cpu) && 555 tdq->tdq_load >= high->cs_limit && 556 tdq->tdq_load > high->cs_load && 557 tdq->tdq_transferable) { 558 high->cs_cpu = cpu; 559 high->cs_load = tdq->tdq_load; 560 } 561 return (tdq->tdq_load); 562} 563 564/* 565 * Search the tree of cpu_groups for the lowest or highest loaded cpu 566 * according to the match argument. This routine actually compares the 567 * load on all paths through the tree and finds the least loaded cpu on 568 * the least loaded path, which may differ from the least loaded cpu in 569 * the system. This balances work among caches and busses. 570 * 571 * This inline is instantiated in three forms below using constants for the 572 * match argument. It is reduced to the minimum set for each case. It is 573 * also recursive to the depth of the tree. 574 */ 575static __inline int 576cpu_search(struct cpu_group *cg, struct cpu_search *low, 577 struct cpu_search *high, const int match) 578{ 579 int total; 580 581 total = 0; 582 if (cg->cg_children) { 583 struct cpu_search lgroup; 584 struct cpu_search hgroup; 585 struct cpu_group *child; 586 u_int lload; 587 int hload; 588 int load; 589 int i; 590 591 lload = -1; 592 hload = -1; 593 for (i = 0; i < cg->cg_children; i++) { 594 child = &cg->cg_child[i]; 595 if (match & CPU_SEARCH_LOWEST) { 596 lgroup = *low; 597 lgroup.cs_load = -1; 598 } 599 if (match & CPU_SEARCH_HIGHEST) { 600 hgroup = *high; 601 lgroup.cs_load = 0; 602 } 603 switch (match) { 604 case CPU_SEARCH_LOWEST: 605 load = cpu_search_lowest(child, &lgroup); 606 break; 607 case CPU_SEARCH_HIGHEST: 608 load = cpu_search_highest(child, &hgroup); 609 break; 610 case CPU_SEARCH_BOTH: 611 load = cpu_search_both(child, &lgroup, &hgroup); 612 break; 613 } 614 total += load; 615 if (match & CPU_SEARCH_LOWEST) 616 if (load < lload || low->cs_cpu == -1) { 617 *low = lgroup; 618 lload = load; 619 } 620 if (match & CPU_SEARCH_HIGHEST) 621 if (load > hload || high->cs_cpu == -1) { 622 hload = load; 623 *high = hgroup; 624 } 625 } 626 } else { 627 int cpu; 628 629 CPUMASK_FOREACH(cpu, cg->cg_mask) 630 total += cpu_compare(cpu, low, high, match); 631 } 632 return (total); 633} 634 635/* 636 * cpu_search instantiations must pass constants to maintain the inline 637 * optimization. 638 */ 639int 640cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low) 641{ 642 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 643} 644 645int 646cpu_search_highest(struct cpu_group *cg, struct cpu_search *high) 647{ 648 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 649} 650 651int 652cpu_search_both(struct cpu_group *cg, struct cpu_search *low, 653 struct cpu_search *high) 654{ 655 return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 656} 657 658/* 659 * Find the cpu with the least load via the least loaded path that has a 660 * lowpri greater than pri pri. A pri of -1 indicates any priority is 661 * acceptable. 662 */ 663static inline int 664sched_lowest(struct cpu_group *cg, cpumask_t mask, int pri) 665{ 666 struct cpu_search low; 667 668 low.cs_cpu = -1; 669 low.cs_load = -1; 670 low.cs_mask = mask; 671 low.cs_limit = pri; 672 cpu_search_lowest(cg, &low); 673 return low.cs_cpu; 674} 675 676/* 677 * Find the cpu with the highest load via the highest loaded path. 678 */ 679static inline int 680sched_highest(struct cpu_group *cg, cpumask_t mask, int minload) 681{ 682 struct cpu_search high; 683 684 high.cs_cpu = -1; 685 high.cs_load = 0; 686 high.cs_mask = mask; 687 high.cs_limit = minload; 688 cpu_search_highest(cg, &high); 689 return high.cs_cpu; 690} 691 692/* 693 * Simultaneously find the highest and lowest loaded cpu reachable via 694 * cg. 695 */ 696static inline void 697sched_both(struct cpu_group *cg, cpumask_t mask, int *lowcpu, int *highcpu) 698{ 699 struct cpu_search high; 700 struct cpu_search low; 701 702 low.cs_cpu = -1; 703 low.cs_limit = -1; 704 low.cs_load = -1; 705 low.cs_mask = mask; 706 high.cs_load = 0; 707 high.cs_cpu = -1; 708 high.cs_limit = -1; 709 high.cs_mask = mask; 710 cpu_search_both(cg, &low, &high); 711 *lowcpu = low.cs_cpu; 712 *highcpu = high.cs_cpu; 713 return; 714} 715 716static void 717sched_balance_group(struct cpu_group *cg) 718{ 719 cpumask_t mask; 720 int high; 721 int low; 722 int i; 723 724 mask = -1; 725 for (;;) { 726 sched_both(cg, mask, &low, &high); 727 if (low == high || low == -1 || high == -1) 728 break; 729 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) 730 break; 731 /* 732 * If we failed to move any threads determine which cpu 733 * to kick out of the set and try again. 734 */ 735 if (TDQ_CPU(high)->tdq_transferable == 0) 736 mask &= ~(1 << high); 737 else 738 mask &= ~(1 << low); 739 } 740 741 for (i = 0; i < cg->cg_children; i++) 742 sched_balance_group(&cg->cg_child[i]); 743} 744 745static void 746sched_balance() 747{ 748 struct tdq *tdq; 749 750 /* 751 * Select a random time between .5 * balance_interval and 752 * 1.5 * balance_interval. 753 */ 754 balance_ticks = max(balance_interval / 2, 1); 755 balance_ticks += random() % balance_interval; 756 if (smp_started == 0 || rebalance == 0) 757 return; 758 tdq = TDQ_SELF(); 759 TDQ_UNLOCK(tdq); 760 sched_balance_group(cpu_top); 761 TDQ_LOCK(tdq); 762} 763 764/* 765 * Lock two thread queues using their address to maintain lock order. 766 */ 767static void 768tdq_lock_pair(struct tdq *one, struct tdq *two) 769{ 770 if (one < two) { 771 TDQ_LOCK(one); 772 TDQ_LOCK_FLAGS(two, MTX_DUPOK); 773 } else { 774 TDQ_LOCK(two); 775 TDQ_LOCK_FLAGS(one, MTX_DUPOK); 776 } 777} 778 779/* 780 * Unlock two thread queues. Order is not important here. 781 */ 782static void 783tdq_unlock_pair(struct tdq *one, struct tdq *two) 784{ 785 TDQ_UNLOCK(one); 786 TDQ_UNLOCK(two); 787} 788 789/* 790 * Transfer load between two imbalanced thread queues. 791 */ 792static int 793sched_balance_pair(struct tdq *high, struct tdq *low) 794{ 795 int transferable; 796 int high_load; 797 int low_load; 798 int moved; 799 int move; 800 int diff; 801 int i; 802 803 tdq_lock_pair(high, low); 804 transferable = high->tdq_transferable; 805 high_load = high->tdq_load; 806 low_load = low->tdq_load; 807 moved = 0; 808 /* 809 * Determine what the imbalance is and then adjust that to how many 810 * threads we actually have to give up (transferable). 811 */ 812 if (transferable != 0) { 813 diff = high_load - low_load; 814 move = diff / 2; 815 if (diff & 0x1) 816 move++; 817 move = min(move, transferable); 818 for (i = 0; i < move; i++) 819 moved += tdq_move(high, low); 820 /* 821 * IPI the target cpu to force it to reschedule with the new 822 * workload. 823 */ 824 ipi_selected(1 << TDQ_ID(low), IPI_PREEMPT); 825 } 826 tdq_unlock_pair(high, low); 827 return (moved); 828} 829 830/* 831 * Move a thread from one thread queue to another. 832 */ 833static int 834tdq_move(struct tdq *from, struct tdq *to) 835{ 836 struct td_sched *ts; 837 struct thread *td; 838 struct tdq *tdq; 839 int cpu; 840 841 TDQ_LOCK_ASSERT(from, MA_OWNED); 842 TDQ_LOCK_ASSERT(to, MA_OWNED); 843 844 tdq = from; 845 cpu = TDQ_ID(to); 846 td = tdq_steal(tdq, cpu); 847 if (td == NULL) 848 return (0); 849 ts = td->td_sched; 850 /* 851 * Although the run queue is locked the thread may be blocked. Lock 852 * it to clear this and acquire the run-queue lock. 853 */ 854 thread_lock(td); 855 /* Drop recursive lock on from acquired via thread_lock(). */ 856 TDQ_UNLOCK(from); 857 sched_rem(td); 858 ts->ts_cpu = cpu; 859 td->td_lock = TDQ_LOCKPTR(to); 860 tdq_add(to, td, SRQ_YIELDING); 861 return (1); 862} 863 864/* 865 * This tdq has idled. Try to steal a thread from another cpu and switch 866 * to it. 867 */ 868static int 869tdq_idled(struct tdq *tdq) 870{ 871 struct cpu_group *cg; 872 struct tdq *steal; 873 cpumask_t mask; 874 int thresh; 875 int cpu; 876 877 if (smp_started == 0 || steal_idle == 0) 878 return (1); 879 mask = -1; 880 mask &= ~PCPU_GET(cpumask); 881 /* We don't want to be preempted while we're iterating. */ 882 spinlock_enter(); 883 for (cg = tdq->tdq_cg; cg != NULL; ) { 884 if ((cg->cg_flags & (CG_FLAG_HTT | CG_FLAG_THREAD)) == 0) 885 thresh = steal_thresh; 886 else 887 thresh = 1; 888 cpu = sched_highest(cg, mask, thresh); 889 if (cpu == -1) { 890 cg = cg->cg_parent; 891 continue; 892 } 893 steal = TDQ_CPU(cpu); 894 mask &= ~(1 << cpu); 895 tdq_lock_pair(tdq, steal); 896 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 897 tdq_unlock_pair(tdq, steal); 898 continue; 899 } 900 /* 901 * If a thread was added while interrupts were disabled don't 902 * steal one here. If we fail to acquire one due to affinity 903 * restrictions loop again with this cpu removed from the 904 * set. 905 */ 906 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 907 tdq_unlock_pair(tdq, steal); 908 continue; 909 } 910 spinlock_exit(); 911 TDQ_UNLOCK(steal); 912 mi_switch(SW_VOL | SWT_IDLE, NULL); 913 thread_unlock(curthread); 914 915 return (0); 916 } 917 spinlock_exit(); 918 return (1); 919} 920 921/* 922 * Notify a remote cpu of new work. Sends an IPI if criteria are met. 923 */ 924static void 925tdq_notify(struct tdq *tdq, struct thread *td) 926{ 927 int cpri; 928 int pri; 929 int cpu; 930 931 if (tdq->tdq_ipipending) 932 return; 933 cpu = td->td_sched->ts_cpu; 934 pri = td->td_priority; 935 cpri = pcpu_find(cpu)->pc_curthread->td_priority; 936 if (!sched_shouldpreempt(pri, cpri, 1)) 937 return; 938 tdq->tdq_ipipending = 1; 939 ipi_selected(1 << cpu, IPI_PREEMPT); 940} 941 942/* 943 * Steals load from a timeshare queue. Honors the rotating queue head 944 * index. 945 */ 946static struct thread * 947runq_steal_from(struct runq *rq, int cpu, u_char start) 948{ 949 struct rqbits *rqb; 950 struct rqhead *rqh; 951 struct thread *td; 952 int first; 953 int bit; 954 int pri; 955 int i; 956 957 rqb = &rq->rq_status; 958 bit = start & (RQB_BPW -1); 959 pri = 0; 960 first = 0; 961again: 962 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 963 if (rqb->rqb_bits[i] == 0) 964 continue; 965 if (bit != 0) { 966 for (pri = bit; pri < RQB_BPW; pri++) 967 if (rqb->rqb_bits[i] & (1ul << pri)) 968 break; 969 if (pri >= RQB_BPW) 970 continue; 971 } else 972 pri = RQB_FFS(rqb->rqb_bits[i]); 973 pri += (i << RQB_L2BPW); 974 rqh = &rq->rq_queues[pri]; 975 TAILQ_FOREACH(td, rqh, td_runq) { 976 if (first && THREAD_CAN_MIGRATE(td) && 977 THREAD_CAN_SCHED(td, cpu)) 978 return (td); 979 first = 1; 980 } 981 } 982 if (start != 0) { 983 start = 0; 984 goto again; 985 } 986 987 return (NULL); 988} 989 990/* 991 * Steals load from a standard linear queue. 992 */ 993static struct thread * 994runq_steal(struct runq *rq, int cpu) 995{ 996 struct rqhead *rqh; 997 struct rqbits *rqb; 998 struct thread *td; 999 int word; 1000 int bit; 1001 1002 rqb = &rq->rq_status; 1003 for (word = 0; word < RQB_LEN; word++) { 1004 if (rqb->rqb_bits[word] == 0) 1005 continue; 1006 for (bit = 0; bit < RQB_BPW; bit++) { 1007 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1008 continue; 1009 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1010 TAILQ_FOREACH(td, rqh, td_runq) 1011 if (THREAD_CAN_MIGRATE(td) && 1012 THREAD_CAN_SCHED(td, cpu)) 1013 return (td); 1014 } 1015 } 1016 return (NULL); 1017} 1018 1019/* 1020 * Attempt to steal a thread in priority order from a thread queue. 1021 */ 1022static struct thread * 1023tdq_steal(struct tdq *tdq, int cpu) 1024{ 1025 struct thread *td; 1026 1027 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1028 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1029 return (td); 1030 if ((td = runq_steal_from(&tdq->tdq_timeshare, 1031 cpu, tdq->tdq_ridx)) != NULL) 1032 return (td); 1033 return (runq_steal(&tdq->tdq_idle, cpu)); 1034} 1035 1036/* 1037 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1038 * current lock and returns with the assigned queue locked. 1039 */ 1040static inline struct tdq * 1041sched_setcpu(struct thread *td, int cpu, int flags) 1042{ 1043 1044 struct tdq *tdq; 1045 1046 THREAD_LOCK_ASSERT(td, MA_OWNED); 1047 tdq = TDQ_CPU(cpu); 1048 td->td_sched->ts_cpu = cpu; 1049 /* 1050 * If the lock matches just return the queue. 1051 */ 1052 if (td->td_lock == TDQ_LOCKPTR(tdq)) 1053 return (tdq); 1054#ifdef notyet 1055 /* 1056 * If the thread isn't running its lockptr is a 1057 * turnstile or a sleepqueue. We can just lock_set without 1058 * blocking. 1059 */ 1060 if (TD_CAN_RUN(td)) { 1061 TDQ_LOCK(tdq); 1062 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1063 return (tdq); 1064 } 1065#endif 1066 /* 1067 * The hard case, migration, we need to block the thread first to 1068 * prevent order reversals with other cpus locks. 1069 */ 1070 thread_lock_block(td); 1071 TDQ_LOCK(tdq); 1072 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1073 return (tdq); 1074} 1075 1076SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1077SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1078SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1079SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1080SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1081SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1082 1083static int 1084sched_pickcpu(struct thread *td, int flags) 1085{ 1086 struct cpu_group *cg; 1087 struct td_sched *ts; 1088 struct tdq *tdq; 1089 cpumask_t mask; 1090 int self; 1091 int pri; 1092 int cpu; 1093 1094 self = PCPU_GET(cpuid); 1095 ts = td->td_sched; 1096 if (smp_started == 0) 1097 return (self); 1098 /* 1099 * Don't migrate a running thread from sched_switch(). 1100 */ 1101 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1102 return (ts->ts_cpu); 1103 /* 1104 * Prefer to run interrupt threads on the processors that generate 1105 * the interrupt. 1106 */ 1107 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1108 curthread->td_intr_nesting_level && ts->ts_cpu != self) { 1109 SCHED_STAT_INC(pickcpu_intrbind); 1110 ts->ts_cpu = self; 1111 } 1112 /* 1113 * If the thread can run on the last cpu and the affinity has not 1114 * expired or it is idle run it there. 1115 */ 1116 pri = td->td_priority; 1117 tdq = TDQ_CPU(ts->ts_cpu); 1118 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) { 1119 if (tdq->tdq_lowpri > PRI_MIN_IDLE) { 1120 SCHED_STAT_INC(pickcpu_idle_affinity); 1121 return (ts->ts_cpu); 1122 } 1123 if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri) { 1124 SCHED_STAT_INC(pickcpu_affinity); 1125 return (ts->ts_cpu); 1126 } 1127 } 1128 /* 1129 * Search for the highest level in the tree that still has affinity. 1130 */ 1131 cg = NULL; 1132 for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent) 1133 if (SCHED_AFFINITY(ts, cg->cg_level)) 1134 break; 1135 cpu = -1; 1136 mask = td->td_cpuset->cs_mask.__bits[0]; 1137 if (cg) 1138 cpu = sched_lowest(cg, mask, pri); 1139 if (cpu == -1) 1140 cpu = sched_lowest(cpu_top, mask, -1); 1141 /* 1142 * Compare the lowest loaded cpu to current cpu. 1143 */ 1144 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1145 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) { 1146 SCHED_STAT_INC(pickcpu_local); 1147 cpu = self; 1148 } else 1149 SCHED_STAT_INC(pickcpu_lowest); 1150 if (cpu != ts->ts_cpu) 1151 SCHED_STAT_INC(pickcpu_migration); 1152 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1153 return (cpu); 1154} 1155#endif 1156 1157/* 1158 * Pick the highest priority task we have and return it. 1159 */ 1160static struct thread * 1161tdq_choose(struct tdq *tdq) 1162{ 1163 struct thread *td; 1164 1165 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1166 td = runq_choose(&tdq->tdq_realtime); 1167 if (td != NULL) 1168 return (td); 1169 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1170 if (td != NULL) { 1171 KASSERT(td->td_priority >= PRI_MIN_TIMESHARE, 1172 ("tdq_choose: Invalid priority on timeshare queue %d", 1173 td->td_priority)); 1174 return (td); 1175 } 1176 td = runq_choose(&tdq->tdq_idle); 1177 if (td != NULL) { 1178 KASSERT(td->td_priority >= PRI_MIN_IDLE, 1179 ("tdq_choose: Invalid priority on idle queue %d", 1180 td->td_priority)); 1181 return (td); 1182 } 1183 1184 return (NULL); 1185} 1186 1187/* 1188 * Initialize a thread queue. 1189 */ 1190static void 1191tdq_setup(struct tdq *tdq) 1192{ 1193 1194 if (bootverbose) 1195 printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1196 runq_init(&tdq->tdq_realtime); 1197 runq_init(&tdq->tdq_timeshare); 1198 runq_init(&tdq->tdq_idle); 1199 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1200 "sched lock %d", (int)TDQ_ID(tdq)); 1201 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1202 MTX_SPIN | MTX_RECURSE); 1203} 1204 1205#ifdef SMP 1206static void 1207sched_setup_smp(void) 1208{ 1209 struct tdq *tdq; 1210 int i; 1211 1212 cpu_top = smp_topo(); 1213 for (i = 0; i < MAXCPU; i++) { 1214 if (CPU_ABSENT(i)) 1215 continue; 1216 tdq = TDQ_CPU(i); 1217 tdq_setup(tdq); 1218 tdq->tdq_cg = smp_topo_find(cpu_top, i); 1219 if (tdq->tdq_cg == NULL) 1220 panic("Can't find cpu group for %d\n", i); 1221 } 1222 balance_tdq = TDQ_SELF(); 1223 sched_balance(); 1224} 1225#endif 1226 1227/* 1228 * Setup the thread queues and initialize the topology based on MD 1229 * information. 1230 */ 1231static void 1232sched_setup(void *dummy) 1233{ 1234 struct tdq *tdq; 1235 1236 tdq = TDQ_SELF(); 1237#ifdef SMP 1238 sched_setup_smp(); 1239#else 1240 tdq_setup(tdq); 1241#endif 1242 /* 1243 * To avoid divide-by-zero, we set realstathz a dummy value 1244 * in case which sched_clock() called before sched_initticks(). 1245 */ 1246 realstathz = hz; 1247 sched_slice = (realstathz/10); /* ~100ms */ 1248 tickincr = 1 << SCHED_TICK_SHIFT; 1249 1250 /* Add thread0's load since it's running. */ 1251 TDQ_LOCK(tdq); 1252 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1253 tdq_load_add(tdq, &thread0); 1254 tdq->tdq_lowpri = thread0.td_priority; 1255 TDQ_UNLOCK(tdq); 1256} 1257 1258/* 1259 * This routine determines the tickincr after stathz and hz are setup. 1260 */ 1261/* ARGSUSED */ 1262static void 1263sched_initticks(void *dummy) 1264{ 1265 int incr; 1266 1267 realstathz = stathz ? stathz : hz; 1268 sched_slice = (realstathz/10); /* ~100ms */ 1269 1270 /* 1271 * tickincr is shifted out by 10 to avoid rounding errors due to 1272 * hz not being evenly divisible by stathz on all platforms. 1273 */ 1274 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1275 /* 1276 * This does not work for values of stathz that are more than 1277 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1278 */ 1279 if (incr == 0) 1280 incr = 1; 1281 tickincr = incr; 1282#ifdef SMP 1283 /* 1284 * Set the default balance interval now that we know 1285 * what realstathz is. 1286 */ 1287 balance_interval = realstathz; 1288 /* 1289 * Set steal thresh to log2(mp_ncpu) but no greater than 4. This 1290 * prevents excess thrashing on large machines and excess idle on 1291 * smaller machines. 1292 */ 1293 steal_thresh = min(ffs(mp_ncpus) - 1, 3); 1294 affinity = SCHED_AFFINITY_DEFAULT; 1295#endif 1296} 1297 1298 1299/* 1300 * This is the core of the interactivity algorithm. Determines a score based 1301 * on past behavior. It is the ratio of sleep time to run time scaled to 1302 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1303 * differs from the cpu usage because it does not account for time spent 1304 * waiting on a run-queue. Would be prettier if we had floating point. 1305 */ 1306static int 1307sched_interact_score(struct thread *td) 1308{ 1309 struct td_sched *ts; 1310 int div; 1311 1312 ts = td->td_sched; 1313 /* 1314 * The score is only needed if this is likely to be an interactive 1315 * task. Don't go through the expense of computing it if there's 1316 * no chance. 1317 */ 1318 if (sched_interact <= SCHED_INTERACT_HALF && 1319 ts->ts_runtime >= ts->ts_slptime) 1320 return (SCHED_INTERACT_HALF); 1321 1322 if (ts->ts_runtime > ts->ts_slptime) { 1323 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1324 return (SCHED_INTERACT_HALF + 1325 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1326 } 1327 if (ts->ts_slptime > ts->ts_runtime) { 1328 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1329 return (ts->ts_runtime / div); 1330 } 1331 /* runtime == slptime */ 1332 if (ts->ts_runtime) 1333 return (SCHED_INTERACT_HALF); 1334 1335 /* 1336 * This can happen if slptime and runtime are 0. 1337 */ 1338 return (0); 1339 1340} 1341 1342/* 1343 * Scale the scheduling priority according to the "interactivity" of this 1344 * process. 1345 */ 1346static void 1347sched_priority(struct thread *td) 1348{ 1349 int score; 1350 int pri; 1351 1352 if (td->td_pri_class != PRI_TIMESHARE) 1353 return; 1354 /* 1355 * If the score is interactive we place the thread in the realtime 1356 * queue with a priority that is less than kernel and interrupt 1357 * priorities. These threads are not subject to nice restrictions. 1358 * 1359 * Scores greater than this are placed on the normal timeshare queue 1360 * where the priority is partially decided by the most recent cpu 1361 * utilization and the rest is decided by nice value. 1362 * 1363 * The nice value of the process has a linear effect on the calculated 1364 * score. Negative nice values make it easier for a thread to be 1365 * considered interactive. 1366 */ 1367 score = imax(0, sched_interact_score(td) - td->td_proc->p_nice); 1368 if (score < sched_interact) { 1369 pri = PRI_MIN_REALTIME; 1370 pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact) 1371 * score; 1372 KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME, 1373 ("sched_priority: invalid interactive priority %d score %d", 1374 pri, score)); 1375 } else { 1376 pri = SCHED_PRI_MIN; 1377 if (td->td_sched->ts_ticks) 1378 pri += SCHED_PRI_TICKS(td->td_sched); 1379 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1380 KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE, 1381 ("sched_priority: invalid priority %d: nice %d, " 1382 "ticks %d ftick %d ltick %d tick pri %d", 1383 pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1384 td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1385 SCHED_PRI_TICKS(td->td_sched))); 1386 } 1387 sched_user_prio(td, pri); 1388 1389 return; 1390} 1391 1392/* 1393 * This routine enforces a maximum limit on the amount of scheduling history 1394 * kept. It is called after either the slptime or runtime is adjusted. This 1395 * function is ugly due to integer math. 1396 */ 1397static void 1398sched_interact_update(struct thread *td) 1399{ 1400 struct td_sched *ts; 1401 u_int sum; 1402 1403 ts = td->td_sched; 1404 sum = ts->ts_runtime + ts->ts_slptime; 1405 if (sum < SCHED_SLP_RUN_MAX) 1406 return; 1407 /* 1408 * This only happens from two places: 1409 * 1) We have added an unusual amount of run time from fork_exit. 1410 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1411 */ 1412 if (sum > SCHED_SLP_RUN_MAX * 2) { 1413 if (ts->ts_runtime > ts->ts_slptime) { 1414 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1415 ts->ts_slptime = 1; 1416 } else { 1417 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1418 ts->ts_runtime = 1; 1419 } 1420 return; 1421 } 1422 /* 1423 * If we have exceeded by more than 1/5th then the algorithm below 1424 * will not bring us back into range. Dividing by two here forces 1425 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1426 */ 1427 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1428 ts->ts_runtime /= 2; 1429 ts->ts_slptime /= 2; 1430 return; 1431 } 1432 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1433 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1434} 1435 1436/* 1437 * Scale back the interactivity history when a child thread is created. The 1438 * history is inherited from the parent but the thread may behave totally 1439 * differently. For example, a shell spawning a compiler process. We want 1440 * to learn that the compiler is behaving badly very quickly. 1441 */ 1442static void 1443sched_interact_fork(struct thread *td) 1444{ 1445 int ratio; 1446 int sum; 1447 1448 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1449 if (sum > SCHED_SLP_RUN_FORK) { 1450 ratio = sum / SCHED_SLP_RUN_FORK; 1451 td->td_sched->ts_runtime /= ratio; 1452 td->td_sched->ts_slptime /= ratio; 1453 } 1454} 1455 1456/* 1457 * Called from proc0_init() to setup the scheduler fields. 1458 */ 1459void 1460schedinit(void) 1461{ 1462 1463 /* 1464 * Set up the scheduler specific parts of proc0. 1465 */ 1466 proc0.p_sched = NULL; /* XXX */ 1467 thread0.td_sched = &td_sched0; 1468 td_sched0.ts_ltick = ticks; 1469 td_sched0.ts_ftick = ticks; 1470 td_sched0.ts_slice = sched_slice; 1471} 1472 1473/* 1474 * This is only somewhat accurate since given many processes of the same 1475 * priority they will switch when their slices run out, which will be 1476 * at most sched_slice stathz ticks. 1477 */ 1478int 1479sched_rr_interval(void) 1480{ 1481 1482 /* Convert sched_slice to hz */ 1483 return (hz/(realstathz/sched_slice)); 1484} 1485 1486/* 1487 * Update the percent cpu tracking information when it is requested or 1488 * the total history exceeds the maximum. We keep a sliding history of 1489 * tick counts that slowly decays. This is less precise than the 4BSD 1490 * mechanism since it happens with less regular and frequent events. 1491 */ 1492static void 1493sched_pctcpu_update(struct td_sched *ts) 1494{ 1495 1496 if (ts->ts_ticks == 0) 1497 return; 1498 if (ticks - (hz / 10) < ts->ts_ltick && 1499 SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1500 return; 1501 /* 1502 * Adjust counters and watermark for pctcpu calc. 1503 */ 1504 if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1505 ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1506 SCHED_TICK_TARG; 1507 else 1508 ts->ts_ticks = 0; 1509 ts->ts_ltick = ticks; 1510 ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1511} 1512 1513/* 1514 * Adjust the priority of a thread. Move it to the appropriate run-queue 1515 * if necessary. This is the back-end for several priority related 1516 * functions. 1517 */ 1518static void 1519sched_thread_priority(struct thread *td, u_char prio) 1520{ 1521 struct td_sched *ts; 1522 struct tdq *tdq; 1523 int oldpri; 1524 1525 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 1526 td, td->td_name, td->td_priority, prio, curthread, 1527 curthread->td_name); 1528 ts = td->td_sched; 1529 THREAD_LOCK_ASSERT(td, MA_OWNED); 1530 if (td->td_priority == prio) 1531 return; 1532 /* 1533 * If the priority has been elevated due to priority 1534 * propagation, we may have to move ourselves to a new 1535 * queue. This could be optimized to not re-add in some 1536 * cases. 1537 */ 1538 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1539 sched_rem(td); 1540 td->td_priority = prio; 1541 sched_add(td, SRQ_BORROWING); 1542 return; 1543 } 1544 /* 1545 * If the thread is currently running we may have to adjust the lowpri 1546 * information so other cpus are aware of our current priority. 1547 */ 1548 if (TD_IS_RUNNING(td)) { 1549 tdq = TDQ_CPU(ts->ts_cpu); 1550 oldpri = td->td_priority; 1551 td->td_priority = prio; 1552 if (prio < tdq->tdq_lowpri) 1553 tdq->tdq_lowpri = prio; 1554 else if (tdq->tdq_lowpri == oldpri) 1555 tdq_setlowpri(tdq, td); 1556 return; 1557 } 1558 td->td_priority = prio; 1559} 1560 1561/* 1562 * Update a thread's priority when it is lent another thread's 1563 * priority. 1564 */ 1565void 1566sched_lend_prio(struct thread *td, u_char prio) 1567{ 1568 1569 td->td_flags |= TDF_BORROWING; 1570 sched_thread_priority(td, prio); 1571} 1572 1573/* 1574 * Restore a thread's priority when priority propagation is 1575 * over. The prio argument is the minimum priority the thread 1576 * needs to have to satisfy other possible priority lending 1577 * requests. If the thread's regular priority is less 1578 * important than prio, the thread will keep a priority boost 1579 * of prio. 1580 */ 1581void 1582sched_unlend_prio(struct thread *td, u_char prio) 1583{ 1584 u_char base_pri; 1585 1586 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1587 td->td_base_pri <= PRI_MAX_TIMESHARE) 1588 base_pri = td->td_user_pri; 1589 else 1590 base_pri = td->td_base_pri; 1591 if (prio >= base_pri) { 1592 td->td_flags &= ~TDF_BORROWING; 1593 sched_thread_priority(td, base_pri); 1594 } else 1595 sched_lend_prio(td, prio); 1596} 1597 1598/* 1599 * Standard entry for setting the priority to an absolute value. 1600 */ 1601void 1602sched_prio(struct thread *td, u_char prio) 1603{ 1604 u_char oldprio; 1605 1606 /* First, update the base priority. */ 1607 td->td_base_pri = prio; 1608 1609 /* 1610 * If the thread is borrowing another thread's priority, don't 1611 * ever lower the priority. 1612 */ 1613 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1614 return; 1615 1616 /* Change the real priority. */ 1617 oldprio = td->td_priority; 1618 sched_thread_priority(td, prio); 1619 1620 /* 1621 * If the thread is on a turnstile, then let the turnstile update 1622 * its state. 1623 */ 1624 if (TD_ON_LOCK(td) && oldprio != prio) 1625 turnstile_adjust(td, oldprio); 1626} 1627 1628/* 1629 * Set the base user priority, does not effect current running priority. 1630 */ 1631void 1632sched_user_prio(struct thread *td, u_char prio) 1633{ 1634 u_char oldprio; 1635 1636 td->td_base_user_pri = prio; 1637 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio) 1638 return; 1639 oldprio = td->td_user_pri; 1640 td->td_user_pri = prio; 1641} 1642 1643void 1644sched_lend_user_prio(struct thread *td, u_char prio) 1645{ 1646 u_char oldprio; 1647 1648 THREAD_LOCK_ASSERT(td, MA_OWNED); 1649 td->td_flags |= TDF_UBORROWING; 1650 oldprio = td->td_user_pri; 1651 td->td_user_pri = prio; 1652} 1653 1654void 1655sched_unlend_user_prio(struct thread *td, u_char prio) 1656{ 1657 u_char base_pri; 1658 1659 THREAD_LOCK_ASSERT(td, MA_OWNED); 1660 base_pri = td->td_base_user_pri; 1661 if (prio >= base_pri) { 1662 td->td_flags &= ~TDF_UBORROWING; 1663 sched_user_prio(td, base_pri); 1664 } else { 1665 sched_lend_user_prio(td, prio); 1666 } 1667} 1668 1669/* 1670 * Block a thread for switching. Similar to thread_block() but does not 1671 * bump the spin count. 1672 */ 1673static inline struct mtx * 1674thread_block_switch(struct thread *td) 1675{ 1676 struct mtx *lock; 1677 1678 THREAD_LOCK_ASSERT(td, MA_OWNED); 1679 lock = td->td_lock; 1680 td->td_lock = &blocked_lock; 1681 mtx_unlock_spin(lock); 1682 1683 return (lock); 1684} 1685 1686/* 1687 * Handle migration from sched_switch(). This happens only for 1688 * cpu binding. 1689 */ 1690static struct mtx * 1691sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1692{ 1693 struct tdq *tdn; 1694 1695 tdn = TDQ_CPU(td->td_sched->ts_cpu); 1696#ifdef SMP 1697 tdq_load_rem(tdq, td); 1698 /* 1699 * Do the lock dance required to avoid LOR. We grab an extra 1700 * spinlock nesting to prevent preemption while we're 1701 * not holding either run-queue lock. 1702 */ 1703 spinlock_enter(); 1704 thread_block_switch(td); /* This releases the lock on tdq. */ 1705 TDQ_LOCK(tdn); 1706 tdq_add(tdn, td, flags); 1707 tdq_notify(tdn, td); 1708 /* 1709 * After we unlock tdn the new cpu still can't switch into this 1710 * thread until we've unblocked it in cpu_switch(). The lock 1711 * pointers may match in the case of HTT cores. Don't unlock here 1712 * or we can deadlock when the other CPU runs the IPI handler. 1713 */ 1714 if (TDQ_LOCKPTR(tdn) != TDQ_LOCKPTR(tdq)) { 1715 TDQ_UNLOCK(tdn); 1716 TDQ_LOCK(tdq); 1717 } 1718 spinlock_exit(); 1719#endif 1720 return (TDQ_LOCKPTR(tdn)); 1721} 1722 1723/* 1724 * Release a thread that was blocked with thread_block_switch(). 1725 */ 1726static inline void 1727thread_unblock_switch(struct thread *td, struct mtx *mtx) 1728{ 1729 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1730 (uintptr_t)mtx); 1731} 1732 1733/* 1734 * Switch threads. This function has to handle threads coming in while 1735 * blocked for some reason, running, or idle. It also must deal with 1736 * migrating a thread from one queue to another as running threads may 1737 * be assigned elsewhere via binding. 1738 */ 1739void 1740sched_switch(struct thread *td, struct thread *newtd, int flags) 1741{ 1742 struct tdq *tdq; 1743 struct td_sched *ts; 1744 struct mtx *mtx; 1745 int srqflag; 1746 int cpuid; 1747 1748 THREAD_LOCK_ASSERT(td, MA_OWNED); 1749 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1750 1751 cpuid = PCPU_GET(cpuid); 1752 tdq = TDQ_CPU(cpuid); 1753 ts = td->td_sched; 1754 mtx = td->td_lock; 1755 ts->ts_rltick = ticks; 1756 td->td_lastcpu = td->td_oncpu; 1757 td->td_oncpu = NOCPU; 1758 td->td_flags &= ~TDF_NEEDRESCHED; 1759 td->td_owepreempt = 0; 1760 /* 1761 * The lock pointer in an idle thread should never change. Reset it 1762 * to CAN_RUN as well. 1763 */ 1764 if (TD_IS_IDLETHREAD(td)) { 1765 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1766 TD_SET_CAN_RUN(td); 1767 } else if (TD_IS_RUNNING(td)) { 1768 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1769 srqflag = (flags & SW_PREEMPT) ? 1770 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1771 SRQ_OURSELF|SRQ_YIELDING; 1772 if (ts->ts_cpu == cpuid) 1773 tdq_runq_add(tdq, td, srqflag); 1774 else 1775 mtx = sched_switch_migrate(tdq, td, srqflag); 1776 } else { 1777 /* This thread must be going to sleep. */ 1778 TDQ_LOCK(tdq); 1779 mtx = thread_block_switch(td); 1780 tdq_load_rem(tdq, td); 1781 } 1782 /* 1783 * We enter here with the thread blocked and assigned to the 1784 * appropriate cpu run-queue or sleep-queue and with the current 1785 * thread-queue locked. 1786 */ 1787 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1788 newtd = choosethread(); 1789 /* 1790 * Call the MD code to switch contexts if necessary. 1791 */ 1792 if (td != newtd) { 1793#ifdef HWPMC_HOOKS 1794 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1795 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1796#endif 1797 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1798 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1799 cpu_switch(td, newtd, mtx); 1800 /* 1801 * We may return from cpu_switch on a different cpu. However, 1802 * we always return with td_lock pointing to the current cpu's 1803 * run queue lock. 1804 */ 1805 cpuid = PCPU_GET(cpuid); 1806 tdq = TDQ_CPU(cpuid); 1807 lock_profile_obtain_lock_success( 1808 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1809#ifdef HWPMC_HOOKS 1810 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1811 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1812#endif 1813 } else 1814 thread_unblock_switch(td, mtx); 1815 /* 1816 * Assert that all went well and return. 1817 */ 1818 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1819 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1820 td->td_oncpu = cpuid; 1821} 1822 1823/* 1824 * Adjust thread priorities as a result of a nice request. 1825 */ 1826void 1827sched_nice(struct proc *p, int nice) 1828{ 1829 struct thread *td; 1830 1831 PROC_LOCK_ASSERT(p, MA_OWNED); 1832 1833 p->p_nice = nice; 1834 FOREACH_THREAD_IN_PROC(p, td) { 1835 thread_lock(td); 1836 sched_priority(td); 1837 sched_prio(td, td->td_base_user_pri); 1838 thread_unlock(td); 1839 } 1840} 1841 1842/* 1843 * Record the sleep time for the interactivity scorer. 1844 */ 1845void 1846sched_sleep(struct thread *td, int prio) 1847{ 1848 1849 THREAD_LOCK_ASSERT(td, MA_OWNED); 1850 1851 td->td_slptick = ticks; 1852 if (TD_IS_SUSPENDED(td) || prio <= PSOCK) 1853 td->td_flags |= TDF_CANSWAP; 1854 if (static_boost == 1 && prio) 1855 sched_prio(td, prio); 1856 else if (static_boost && td->td_priority > static_boost) 1857 sched_prio(td, static_boost); 1858} 1859 1860/* 1861 * Schedule a thread to resume execution and record how long it voluntarily 1862 * slept. We also update the pctcpu, interactivity, and priority. 1863 */ 1864void 1865sched_wakeup(struct thread *td) 1866{ 1867 struct td_sched *ts; 1868 int slptick; 1869 1870 THREAD_LOCK_ASSERT(td, MA_OWNED); 1871 ts = td->td_sched; 1872 td->td_flags &= ~TDF_CANSWAP; 1873 /* 1874 * If we slept for more than a tick update our interactivity and 1875 * priority. 1876 */ 1877 slptick = td->td_slptick; 1878 td->td_slptick = 0; 1879 if (slptick && slptick != ticks) { 1880 u_int hzticks; 1881 1882 hzticks = (ticks - slptick) << SCHED_TICK_SHIFT; 1883 ts->ts_slptime += hzticks; 1884 sched_interact_update(td); 1885 sched_pctcpu_update(ts); 1886 } 1887 /* Reset the slice value after we sleep. */ 1888 ts->ts_slice = sched_slice; 1889 sched_add(td, SRQ_BORING); 1890} 1891 1892/* 1893 * Penalize the parent for creating a new child and initialize the child's 1894 * priority. 1895 */ 1896void 1897sched_fork(struct thread *td, struct thread *child) 1898{ 1899 THREAD_LOCK_ASSERT(td, MA_OWNED); 1900 sched_fork_thread(td, child); 1901 /* 1902 * Penalize the parent and child for forking. 1903 */ 1904 sched_interact_fork(child); 1905 sched_priority(child); 1906 td->td_sched->ts_runtime += tickincr; 1907 sched_interact_update(td); 1908 sched_priority(td); 1909} 1910 1911/* 1912 * Fork a new thread, may be within the same process. 1913 */ 1914void 1915sched_fork_thread(struct thread *td, struct thread *child) 1916{ 1917 struct td_sched *ts; 1918 struct td_sched *ts2; 1919 1920 THREAD_LOCK_ASSERT(td, MA_OWNED); 1921 /* 1922 * Initialize child. 1923 */ 1924 ts = td->td_sched; 1925 ts2 = child->td_sched; 1926 child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1927 child->td_cpuset = cpuset_ref(td->td_cpuset); 1928 ts2->ts_cpu = ts->ts_cpu; 1929 ts2->ts_flags = 0; 1930 /* 1931 * Grab our parents cpu estimation information and priority. 1932 */ 1933 ts2->ts_ticks = ts->ts_ticks; 1934 ts2->ts_ltick = ts->ts_ltick; 1935 ts2->ts_ftick = ts->ts_ftick; 1936 child->td_user_pri = td->td_user_pri; 1937 child->td_base_user_pri = td->td_base_user_pri; 1938 /* 1939 * And update interactivity score. 1940 */ 1941 ts2->ts_slptime = ts->ts_slptime; 1942 ts2->ts_runtime = ts->ts_runtime; 1943 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 1944} 1945 1946/* 1947 * Adjust the priority class of a thread. 1948 */ 1949void 1950sched_class(struct thread *td, int class) 1951{ 1952 1953 THREAD_LOCK_ASSERT(td, MA_OWNED); 1954 if (td->td_pri_class == class) 1955 return; 1956 td->td_pri_class = class; 1957} 1958 1959/* 1960 * Return some of the child's priority and interactivity to the parent. 1961 */ 1962void 1963sched_exit(struct proc *p, struct thread *child) 1964{ 1965 struct thread *td; 1966 1967 CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d", 1968 child, child->td_name, child->td_priority); 1969 1970 PROC_LOCK_ASSERT(p, MA_OWNED); 1971 td = FIRST_THREAD_IN_PROC(p); 1972 sched_exit_thread(td, child); 1973} 1974 1975/* 1976 * Penalize another thread for the time spent on this one. This helps to 1977 * worsen the priority and interactivity of processes which schedule batch 1978 * jobs such as make. This has little effect on the make process itself but 1979 * causes new processes spawned by it to receive worse scores immediately. 1980 */ 1981void 1982sched_exit_thread(struct thread *td, struct thread *child) 1983{ 1984 1985 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d", 1986 child, child->td_name, child->td_priority); 1987 1988 /* 1989 * Give the child's runtime to the parent without returning the 1990 * sleep time as a penalty to the parent. This causes shells that 1991 * launch expensive things to mark their children as expensive. 1992 */ 1993 thread_lock(td); 1994 td->td_sched->ts_runtime += child->td_sched->ts_runtime; 1995 sched_interact_update(td); 1996 sched_priority(td); 1997 thread_unlock(td); 1998} 1999 2000void 2001sched_preempt(struct thread *td) 2002{ 2003 struct tdq *tdq; 2004 2005 thread_lock(td); 2006 tdq = TDQ_SELF(); 2007 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2008 tdq->tdq_ipipending = 0; 2009 if (td->td_priority > tdq->tdq_lowpri) { 2010 int flags; 2011 2012 flags = SW_INVOL | SW_PREEMPT; 2013 if (td->td_critnest > 1) 2014 td->td_owepreempt = 1; 2015 else if (TD_IS_IDLETHREAD(td)) 2016 mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL); 2017 else 2018 mi_switch(flags | SWT_REMOTEPREEMPT, NULL); 2019 } 2020 thread_unlock(td); 2021} 2022 2023/* 2024 * Fix priorities on return to user-space. Priorities may be elevated due 2025 * to static priorities in msleep() or similar. 2026 */ 2027void 2028sched_userret(struct thread *td) 2029{ 2030 /* 2031 * XXX we cheat slightly on the locking here to avoid locking in 2032 * the usual case. Setting td_priority here is essentially an 2033 * incomplete workaround for not setting it properly elsewhere. 2034 * Now that some interrupt handlers are threads, not setting it 2035 * properly elsewhere can clobber it in the window between setting 2036 * it here and returning to user mode, so don't waste time setting 2037 * it perfectly here. 2038 */ 2039 KASSERT((td->td_flags & TDF_BORROWING) == 0, 2040 ("thread with borrowed priority returning to userland")); 2041 if (td->td_priority != td->td_user_pri) { 2042 thread_lock(td); 2043 td->td_priority = td->td_user_pri; 2044 td->td_base_pri = td->td_user_pri; 2045 tdq_setlowpri(TDQ_SELF(), td); 2046 thread_unlock(td); 2047 } 2048} 2049 2050/* 2051 * Handle a stathz tick. This is really only relevant for timeshare 2052 * threads. 2053 */ 2054void 2055sched_clock(struct thread *td) 2056{ 2057 struct tdq *tdq; 2058 struct td_sched *ts; 2059 2060 THREAD_LOCK_ASSERT(td, MA_OWNED); 2061 tdq = TDQ_SELF(); 2062#ifdef SMP 2063 /* 2064 * We run the long term load balancer infrequently on the first cpu. 2065 */ 2066 if (balance_tdq == tdq) { 2067 if (balance_ticks && --balance_ticks == 0) 2068 sched_balance(); 2069 } 2070#endif 2071 /* 2072 * Advance the insert index once for each tick to ensure that all 2073 * threads get a chance to run. 2074 */ 2075 if (tdq->tdq_idx == tdq->tdq_ridx) { 2076 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2077 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2078 tdq->tdq_ridx = tdq->tdq_idx; 2079 } 2080 ts = td->td_sched; 2081 if (td->td_pri_class & PRI_FIFO_BIT) 2082 return; 2083 if (td->td_pri_class == PRI_TIMESHARE) { 2084 /* 2085 * We used a tick; charge it to the thread so 2086 * that we can compute our interactivity. 2087 */ 2088 td->td_sched->ts_runtime += tickincr; 2089 sched_interact_update(td); 2090 sched_priority(td); 2091 } 2092 /* 2093 * We used up one time slice. 2094 */ 2095 if (--ts->ts_slice > 0) 2096 return; 2097 /* 2098 * We're out of time, force a requeue at userret(). 2099 */ 2100 ts->ts_slice = sched_slice; 2101 td->td_flags |= TDF_NEEDRESCHED; 2102} 2103 2104/* 2105 * Called once per hz tick. Used for cpu utilization information. This 2106 * is easier than trying to scale based on stathz. 2107 */ 2108void 2109sched_tick(void) 2110{ 2111 struct td_sched *ts; 2112 2113 ts = curthread->td_sched; 2114 /* Adjust ticks for pctcpu */ 2115 ts->ts_ticks += 1 << SCHED_TICK_SHIFT; 2116 ts->ts_ltick = ticks; 2117 /* 2118 * Update if we've exceeded our desired tick threshhold by over one 2119 * second. 2120 */ 2121 if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2122 sched_pctcpu_update(ts); 2123} 2124 2125/* 2126 * Return whether the current CPU has runnable tasks. Used for in-kernel 2127 * cooperative idle threads. 2128 */ 2129int 2130sched_runnable(void) 2131{ 2132 struct tdq *tdq; 2133 int load; 2134 2135 load = 1; 2136 2137 tdq = TDQ_SELF(); 2138 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2139 if (tdq->tdq_load > 0) 2140 goto out; 2141 } else 2142 if (tdq->tdq_load - 1 > 0) 2143 goto out; 2144 load = 0; 2145out: 2146 return (load); 2147} 2148 2149/* 2150 * Choose the highest priority thread to run. The thread is removed from 2151 * the run-queue while running however the load remains. For SMP we set 2152 * the tdq in the global idle bitmask if it idles here. 2153 */ 2154struct thread * 2155sched_choose(void) 2156{ 2157 struct thread *td; 2158 struct tdq *tdq; 2159 2160 tdq = TDQ_SELF(); 2161 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2162 td = tdq_choose(tdq); 2163 if (td) { 2164 td->td_sched->ts_ltick = ticks; 2165 tdq_runq_rem(tdq, td); 2166 tdq->tdq_lowpri = td->td_priority; 2167 return (td); 2168 } 2169 tdq->tdq_lowpri = PRI_MAX_IDLE; 2170 return (PCPU_GET(idlethread)); 2171} 2172 2173/* 2174 * Set owepreempt if necessary. Preemption never happens directly in ULE, 2175 * we always request it once we exit a critical section. 2176 */ 2177static inline void 2178sched_setpreempt(struct thread *td) 2179{ 2180 struct thread *ctd; 2181 int cpri; 2182 int pri; 2183 2184 THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2185 2186 ctd = curthread; 2187 pri = td->td_priority; 2188 cpri = ctd->td_priority; 2189 if (pri < cpri) 2190 ctd->td_flags |= TDF_NEEDRESCHED; 2191 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2192 return; 2193 if (!sched_shouldpreempt(pri, cpri, 0)) 2194 return; 2195 ctd->td_owepreempt = 1; 2196} 2197 2198/* 2199 * Add a thread to a thread queue. Select the appropriate runq and add the 2200 * thread to it. This is the internal function called when the tdq is 2201 * predetermined. 2202 */ 2203void 2204tdq_add(struct tdq *tdq, struct thread *td, int flags) 2205{ 2206 2207 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2208 KASSERT((td->td_inhibitors == 0), 2209 ("sched_add: trying to run inhibited thread")); 2210 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2211 ("sched_add: bad thread state")); 2212 KASSERT(td->td_flags & TDF_INMEM, 2213 ("sched_add: thread swapped out")); 2214 2215 if (td->td_priority < tdq->tdq_lowpri) 2216 tdq->tdq_lowpri = td->td_priority; 2217 tdq_runq_add(tdq, td, flags); 2218 tdq_load_add(tdq, td); 2219} 2220 2221/* 2222 * Select the target thread queue and add a thread to it. Request 2223 * preemption or IPI a remote processor if required. 2224 */ 2225void 2226sched_add(struct thread *td, int flags) 2227{ 2228 struct tdq *tdq; 2229#ifdef SMP 2230 int cpu; 2231#endif 2232 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 2233 td, td->td_name, td->td_priority, curthread, 2234 curthread->td_name); 2235 THREAD_LOCK_ASSERT(td, MA_OWNED); 2236 /* 2237 * Recalculate the priority before we select the target cpu or 2238 * run-queue. 2239 */ 2240 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2241 sched_priority(td); 2242#ifdef SMP 2243 /* 2244 * Pick the destination cpu and if it isn't ours transfer to the 2245 * target cpu. 2246 */ 2247 cpu = sched_pickcpu(td, flags); 2248 tdq = sched_setcpu(td, cpu, flags); 2249 tdq_add(tdq, td, flags); 2250 if (cpu != PCPU_GET(cpuid)) { 2251 tdq_notify(tdq, td); 2252 return; 2253 } 2254#else 2255 tdq = TDQ_SELF(); 2256 TDQ_LOCK(tdq); 2257 /* 2258 * Now that the thread is moving to the run-queue, set the lock 2259 * to the scheduler's lock. 2260 */ 2261 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2262 tdq_add(tdq, td, flags); 2263#endif 2264 if (!(flags & SRQ_YIELDING)) 2265 sched_setpreempt(td); 2266} 2267 2268/* 2269 * Remove a thread from a run-queue without running it. This is used 2270 * when we're stealing a thread from a remote queue. Otherwise all threads 2271 * exit by calling sched_exit_thread() and sched_throw() themselves. 2272 */ 2273void 2274sched_rem(struct thread *td) 2275{ 2276 struct tdq *tdq; 2277 2278 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 2279 td, td->td_name, td->td_priority, curthread, 2280 curthread->td_name); 2281 tdq = TDQ_CPU(td->td_sched->ts_cpu); 2282 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2283 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2284 KASSERT(TD_ON_RUNQ(td), 2285 ("sched_rem: thread not on run queue")); 2286 tdq_runq_rem(tdq, td); 2287 tdq_load_rem(tdq, td); 2288 TD_SET_CAN_RUN(td); 2289 if (td->td_priority == tdq->tdq_lowpri) 2290 tdq_setlowpri(tdq, NULL); 2291} 2292 2293/* 2294 * Fetch cpu utilization information. Updates on demand. 2295 */ 2296fixpt_t 2297sched_pctcpu(struct thread *td) 2298{ 2299 fixpt_t pctcpu; 2300 struct td_sched *ts; 2301 2302 pctcpu = 0; 2303 ts = td->td_sched; 2304 if (ts == NULL) 2305 return (0); 2306 2307 thread_lock(td); 2308 if (ts->ts_ticks) { 2309 int rtick; 2310 2311 sched_pctcpu_update(ts); 2312 /* How many rtick per second ? */ 2313 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2314 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2315 } 2316 thread_unlock(td); 2317 2318 return (pctcpu); 2319} 2320 2321/* 2322 * Enforce affinity settings for a thread. Called after adjustments to 2323 * cpumask. 2324 */ 2325void 2326sched_affinity(struct thread *td) 2327{ 2328#ifdef SMP 2329 struct td_sched *ts; 2330 int cpu; 2331 2332 THREAD_LOCK_ASSERT(td, MA_OWNED); 2333 ts = td->td_sched; 2334 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2335 return; 2336 if (!TD_IS_RUNNING(td)) 2337 return; 2338 td->td_flags |= TDF_NEEDRESCHED; 2339 if (!THREAD_CAN_MIGRATE(td)) 2340 return; 2341 /* 2342 * Assign the new cpu and force a switch before returning to 2343 * userspace. If the target thread is not running locally send 2344 * an ipi to force the issue. 2345 */ 2346 cpu = ts->ts_cpu; 2347 ts->ts_cpu = sched_pickcpu(td, 0); 2348 if (cpu != PCPU_GET(cpuid)) 2349 ipi_selected(1 << cpu, IPI_PREEMPT); 2350#endif 2351} 2352 2353/* 2354 * Bind a thread to a target cpu. 2355 */ 2356void 2357sched_bind(struct thread *td, int cpu) 2358{ 2359 struct td_sched *ts; 2360 2361 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2362 ts = td->td_sched; 2363 if (ts->ts_flags & TSF_BOUND) 2364 sched_unbind(td); 2365 ts->ts_flags |= TSF_BOUND; 2366 sched_pin(); 2367 if (PCPU_GET(cpuid) == cpu) 2368 return; 2369 ts->ts_cpu = cpu; 2370 /* When we return from mi_switch we'll be on the correct cpu. */ 2371 mi_switch(SW_VOL, NULL); 2372} 2373 2374/* 2375 * Release a bound thread. 2376 */ 2377void 2378sched_unbind(struct thread *td) 2379{ 2380 struct td_sched *ts; 2381 2382 THREAD_LOCK_ASSERT(td, MA_OWNED); 2383 ts = td->td_sched; 2384 if ((ts->ts_flags & TSF_BOUND) == 0) 2385 return; 2386 ts->ts_flags &= ~TSF_BOUND; 2387 sched_unpin(); 2388} 2389 2390int 2391sched_is_bound(struct thread *td) 2392{ 2393 THREAD_LOCK_ASSERT(td, MA_OWNED); 2394 return (td->td_sched->ts_flags & TSF_BOUND); 2395} 2396 2397/* 2398 * Basic yield call. 2399 */ 2400void 2401sched_relinquish(struct thread *td) 2402{ 2403 thread_lock(td); 2404 mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 2405 thread_unlock(td); 2406} 2407 2408/* 2409 * Return the total system load. 2410 */ 2411int 2412sched_load(void) 2413{ 2414#ifdef SMP 2415 int total; 2416 int i; 2417 2418 total = 0; 2419 for (i = 0; i <= mp_maxid; i++) 2420 total += TDQ_CPU(i)->tdq_sysload; 2421 return (total); 2422#else 2423 return (TDQ_SELF()->tdq_sysload); 2424#endif 2425} 2426 2427int 2428sched_sizeof_proc(void) 2429{ 2430 return (sizeof(struct proc)); 2431} 2432 2433int 2434sched_sizeof_thread(void) 2435{ 2436 return (sizeof(struct thread) + sizeof(struct td_sched)); 2437} 2438 2439/* 2440 * The actual idle process. 2441 */ 2442void 2443sched_idletd(void *dummy) 2444{ 2445 struct thread *td; 2446 struct tdq *tdq; 2447 2448 td = curthread; 2449 tdq = TDQ_SELF(); 2450 mtx_assert(&Giant, MA_NOTOWNED); 2451 /* ULE relies on preemption for idle interruption. */ 2452 for (;;) { 2453#ifdef SMP 2454 if (tdq_idled(tdq)) 2455 cpu_idle(); 2456#else 2457 cpu_idle(); 2458#endif 2459 } 2460} 2461 2462/* 2463 * A CPU is entering for the first time or a thread is exiting. 2464 */ 2465void 2466sched_throw(struct thread *td) 2467{ 2468 struct thread *newtd; 2469 struct tdq *tdq; 2470 2471 tdq = TDQ_SELF(); 2472 if (td == NULL) { 2473 /* Correct spinlock nesting and acquire the correct lock. */ 2474 TDQ_LOCK(tdq); 2475 spinlock_exit(); 2476 } else { 2477 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2478 tdq_load_rem(tdq, td); 2479 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2480 } 2481 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2482 newtd = choosethread(); 2483 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2484 PCPU_SET(switchtime, cpu_ticks()); 2485 PCPU_SET(switchticks, ticks); 2486 cpu_throw(td, newtd); /* doesn't return */ 2487} 2488 2489/* 2490 * This is called from fork_exit(). Just acquire the correct locks and 2491 * let fork do the rest of the work. 2492 */ 2493void 2494sched_fork_exit(struct thread *td) 2495{ 2496 struct td_sched *ts; 2497 struct tdq *tdq; 2498 int cpuid; 2499 2500 /* 2501 * Finish setting up thread glue so that it begins execution in a 2502 * non-nested critical section with the scheduler lock held. 2503 */ 2504 cpuid = PCPU_GET(cpuid); 2505 tdq = TDQ_CPU(cpuid); 2506 ts = td->td_sched; 2507 if (TD_IS_IDLETHREAD(td)) 2508 td->td_lock = TDQ_LOCKPTR(tdq); 2509 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2510 td->td_oncpu = cpuid; 2511 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2512 lock_profile_obtain_lock_success( 2513 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2514} 2515 2516SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2517SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2518 "Scheduler name"); 2519SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2520 "Slice size for timeshare threads"); 2521SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2522 "Interactivity score threshold"); 2523SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2524 0,"Min priority for preemption, lower priorities have greater precedence"); 2525SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 2526 0,"Controls whether static kernel priorities are assigned to sleeping threads."); 2527#ifdef SMP 2528SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2529 "Number of hz ticks to keep thread affinity for"); 2530SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2531 "Enables the long-term load balancer"); 2532SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2533 &balance_interval, 0, 2534 "Average frequency in stathz ticks to run the long-term balancer"); 2535SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, 2536 "Steals work from another hyper-threaded core on idle"); 2537SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2538 "Attempts to steal work from other cores before idling"); 2539SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2540 "Minimum load on remote cpu before we'll steal"); 2541#endif 2542 2543/* ps compat. All cpu percentages from ULE are weighted. */ 2544static int ccpu = 0; 2545SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2546