sched_ule.c revision 177435
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 177435 2008-03-20 05:51:16Z 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(__i386__) && !defined(__amd64__) && !defined(__powerpc__) && !defined(__arm__) 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 = 1; 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 struct td_sched *ts; 464 int class; 465 466 ts = td->td_sched; 467 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 468 THREAD_LOCK_ASSERT(td, MA_OWNED); 469 class = PRI_BASE(td->td_pri_class); 470 tdq->tdq_load++; 471 CTR2(KTR_SCHED, "cpu %d load: %d", TDQ_ID(tdq), tdq->tdq_load); 472 if (class != PRI_ITHD && (td->td_proc->p_flag & P_NOLOAD) == 0) 473 tdq->tdq_sysload++; 474} 475 476/* 477 * Remove the load from a thread that is transitioning to a sleep state or 478 * exiting. 479 */ 480static void 481tdq_load_rem(struct tdq *tdq, struct thread *td) 482{ 483 struct td_sched *ts; 484 int class; 485 486 ts = td->td_sched; 487 THREAD_LOCK_ASSERT(td, MA_OWNED); 488 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 489 class = PRI_BASE(td->td_pri_class); 490 if (class != PRI_ITHD && (td->td_proc->p_flag & P_NOLOAD) == 0) 491 tdq->tdq_sysload--; 492 KASSERT(tdq->tdq_load != 0, 493 ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 494 tdq->tdq_load--; 495 CTR1(KTR_SCHED, "load: %d", tdq->tdq_load); 496} 497 498/* 499 * Set lowpri to its exact value by searching the run-queue and 500 * evaluating curthread. curthread may be passed as an optimization. 501 */ 502static void 503tdq_setlowpri(struct tdq *tdq, struct thread *ctd) 504{ 505 struct thread *td; 506 507 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 508 if (ctd == NULL) 509 ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 510 td = tdq_choose(tdq); 511 if (td == NULL || td->td_priority > ctd->td_priority) 512 tdq->tdq_lowpri = ctd->td_priority; 513 else 514 tdq->tdq_lowpri = td->td_priority; 515} 516 517#ifdef SMP 518struct cpu_search { 519 cpumask_t cs_mask; /* Mask of valid cpus. */ 520 u_int cs_load; 521 u_int cs_cpu; 522 int cs_limit; /* Min priority for low min load for high. */ 523}; 524 525#define CPU_SEARCH_LOWEST 0x1 526#define CPU_SEARCH_HIGHEST 0x2 527#define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 528 529#define CPUMASK_FOREACH(cpu, mask) \ 530 for ((cpu) = 0; (cpu) < sizeof((mask)) * 8; (cpu)++) \ 531 if ((mask) & 1 << (cpu)) 532 533static __inline int cpu_search(struct cpu_group *cg, struct cpu_search *low, 534 struct cpu_search *high, const int match); 535int cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low); 536int cpu_search_highest(struct cpu_group *cg, struct cpu_search *high); 537int cpu_search_both(struct cpu_group *cg, struct cpu_search *low, 538 struct cpu_search *high); 539 540/* 541 * This routine compares according to the match argument and should be 542 * reduced in actual instantiations via constant propagation and dead code 543 * elimination. 544 */ 545static __inline int 546cpu_compare(int cpu, struct cpu_search *low, struct cpu_search *high, 547 const int match) 548{ 549 struct tdq *tdq; 550 551 tdq = TDQ_CPU(cpu); 552 if (match & CPU_SEARCH_LOWEST) 553 if (low->cs_mask & (1 << cpu) && 554 tdq->tdq_load < low->cs_load && 555 tdq->tdq_lowpri > low->cs_limit) { 556 low->cs_cpu = cpu; 557 low->cs_load = tdq->tdq_load; 558 } 559 if (match & CPU_SEARCH_HIGHEST) 560 if (high->cs_mask & (1 << cpu) && 561 tdq->tdq_load >= high->cs_limit && 562 tdq->tdq_load > high->cs_load && 563 tdq->tdq_transferable) { 564 high->cs_cpu = cpu; 565 high->cs_load = tdq->tdq_load; 566 } 567 return (tdq->tdq_load); 568} 569 570/* 571 * Search the tree of cpu_groups for the lowest or highest loaded cpu 572 * according to the match argument. This routine actually compares the 573 * load on all paths through the tree and finds the least loaded cpu on 574 * the least loaded path, which may differ from the least loaded cpu in 575 * the system. This balances work among caches and busses. 576 * 577 * This inline is instantiated in three forms below using constants for the 578 * match argument. It is reduced to the minimum set for each case. It is 579 * also recursive to the depth of the tree. 580 */ 581static __inline int 582cpu_search(struct cpu_group *cg, struct cpu_search *low, 583 struct cpu_search *high, const int match) 584{ 585 int total; 586 587 total = 0; 588 if (cg->cg_children) { 589 struct cpu_search lgroup; 590 struct cpu_search hgroup; 591 struct cpu_group *child; 592 u_int lload; 593 int hload; 594 int load; 595 int i; 596 597 lload = -1; 598 hload = -1; 599 for (i = 0; i < cg->cg_children; i++) { 600 child = &cg->cg_child[i]; 601 if (match & CPU_SEARCH_LOWEST) { 602 lgroup = *low; 603 lgroup.cs_load = -1; 604 } 605 if (match & CPU_SEARCH_HIGHEST) { 606 hgroup = *high; 607 lgroup.cs_load = 0; 608 } 609 switch (match) { 610 case CPU_SEARCH_LOWEST: 611 load = cpu_search_lowest(child, &lgroup); 612 break; 613 case CPU_SEARCH_HIGHEST: 614 load = cpu_search_highest(child, &hgroup); 615 break; 616 case CPU_SEARCH_BOTH: 617 load = cpu_search_both(child, &lgroup, &hgroup); 618 break; 619 } 620 total += load; 621 if (match & CPU_SEARCH_LOWEST) 622 if (load < lload || low->cs_cpu == -1) { 623 *low = lgroup; 624 lload = load; 625 } 626 if (match & CPU_SEARCH_HIGHEST) 627 if (load > hload || high->cs_cpu == -1) { 628 hload = load; 629 *high = hgroup; 630 } 631 } 632 } else { 633 int cpu; 634 635 CPUMASK_FOREACH(cpu, cg->cg_mask) 636 total += cpu_compare(cpu, low, high, match); 637 } 638 return (total); 639} 640 641/* 642 * cpu_search instantiations must pass constants to maintain the inline 643 * optimization. 644 */ 645int 646cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low) 647{ 648 return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 649} 650 651int 652cpu_search_highest(struct cpu_group *cg, struct cpu_search *high) 653{ 654 return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 655} 656 657int 658cpu_search_both(struct cpu_group *cg, struct cpu_search *low, 659 struct cpu_search *high) 660{ 661 return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 662} 663 664/* 665 * Find the cpu with the least load via the least loaded path that has a 666 * lowpri greater than pri pri. A pri of -1 indicates any priority is 667 * acceptable. 668 */ 669static inline int 670sched_lowest(struct cpu_group *cg, cpumask_t mask, int pri) 671{ 672 struct cpu_search low; 673 674 low.cs_cpu = -1; 675 low.cs_load = -1; 676 low.cs_mask = mask; 677 low.cs_limit = pri; 678 cpu_search_lowest(cg, &low); 679 return low.cs_cpu; 680} 681 682/* 683 * Find the cpu with the highest load via the highest loaded path. 684 */ 685static inline int 686sched_highest(struct cpu_group *cg, cpumask_t mask, int minload) 687{ 688 struct cpu_search high; 689 690 high.cs_cpu = -1; 691 high.cs_load = 0; 692 high.cs_mask = mask; 693 high.cs_limit = minload; 694 cpu_search_highest(cg, &high); 695 return high.cs_cpu; 696} 697 698/* 699 * Simultaneously find the highest and lowest loaded cpu reachable via 700 * cg. 701 */ 702static inline void 703sched_both(struct cpu_group *cg, cpumask_t mask, int *lowcpu, int *highcpu) 704{ 705 struct cpu_search high; 706 struct cpu_search low; 707 708 low.cs_cpu = -1; 709 low.cs_limit = -1; 710 low.cs_load = -1; 711 low.cs_mask = mask; 712 high.cs_load = 0; 713 high.cs_cpu = -1; 714 high.cs_limit = -1; 715 high.cs_mask = mask; 716 cpu_search_both(cg, &low, &high); 717 *lowcpu = low.cs_cpu; 718 *highcpu = high.cs_cpu; 719 return; 720} 721 722static void 723sched_balance_group(struct cpu_group *cg) 724{ 725 cpumask_t mask; 726 int high; 727 int low; 728 int i; 729 730 mask = -1; 731 for (;;) { 732 sched_both(cg, mask, &low, &high); 733 if (low == high || low == -1 || high == -1) 734 break; 735 if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) 736 break; 737 /* 738 * If we failed to move any threads determine which cpu 739 * to kick out of the set and try again. 740 */ 741 if (TDQ_CPU(high)->tdq_transferable == 0) 742 mask &= ~(1 << high); 743 else 744 mask &= ~(1 << low); 745 } 746 747 for (i = 0; i < cg->cg_children; i++) 748 sched_balance_group(&cg->cg_child[i]); 749} 750 751static void 752sched_balance() 753{ 754 struct tdq *tdq; 755 756 /* 757 * Select a random time between .5 * balance_interval and 758 * 1.5 * balance_interval. 759 */ 760 balance_ticks = max(balance_interval / 2, 1); 761 balance_ticks += random() % balance_interval; 762 if (smp_started == 0 || rebalance == 0) 763 return; 764 tdq = TDQ_SELF(); 765 TDQ_UNLOCK(tdq); 766 sched_balance_group(cpu_top); 767 TDQ_LOCK(tdq); 768} 769 770/* 771 * Lock two thread queues using their address to maintain lock order. 772 */ 773static void 774tdq_lock_pair(struct tdq *one, struct tdq *two) 775{ 776 if (one < two) { 777 TDQ_LOCK(one); 778 TDQ_LOCK_FLAGS(two, MTX_DUPOK); 779 } else { 780 TDQ_LOCK(two); 781 TDQ_LOCK_FLAGS(one, MTX_DUPOK); 782 } 783} 784 785/* 786 * Unlock two thread queues. Order is not important here. 787 */ 788static void 789tdq_unlock_pair(struct tdq *one, struct tdq *two) 790{ 791 TDQ_UNLOCK(one); 792 TDQ_UNLOCK(two); 793} 794 795/* 796 * Transfer load between two imbalanced thread queues. 797 */ 798static int 799sched_balance_pair(struct tdq *high, struct tdq *low) 800{ 801 int transferable; 802 int high_load; 803 int low_load; 804 int moved; 805 int move; 806 int diff; 807 int i; 808 809 tdq_lock_pair(high, low); 810 transferable = high->tdq_transferable; 811 high_load = high->tdq_load; 812 low_load = low->tdq_load; 813 moved = 0; 814 /* 815 * Determine what the imbalance is and then adjust that to how many 816 * threads we actually have to give up (transferable). 817 */ 818 if (transferable != 0) { 819 diff = high_load - low_load; 820 move = diff / 2; 821 if (diff & 0x1) 822 move++; 823 move = min(move, transferable); 824 for (i = 0; i < move; i++) 825 moved += tdq_move(high, low); 826 /* 827 * IPI the target cpu to force it to reschedule with the new 828 * workload. 829 */ 830 ipi_selected(1 << TDQ_ID(low), IPI_PREEMPT); 831 } 832 tdq_unlock_pair(high, low); 833 return (moved); 834} 835 836/* 837 * Move a thread from one thread queue to another. 838 */ 839static int 840tdq_move(struct tdq *from, struct tdq *to) 841{ 842 struct td_sched *ts; 843 struct thread *td; 844 struct tdq *tdq; 845 int cpu; 846 847 TDQ_LOCK_ASSERT(from, MA_OWNED); 848 TDQ_LOCK_ASSERT(to, MA_OWNED); 849 850 tdq = from; 851 cpu = TDQ_ID(to); 852 td = tdq_steal(tdq, cpu); 853 if (td == NULL) 854 return (0); 855 ts = td->td_sched; 856 /* 857 * Although the run queue is locked the thread may be blocked. Lock 858 * it to clear this and acquire the run-queue lock. 859 */ 860 thread_lock(td); 861 /* Drop recursive lock on from acquired via thread_lock(). */ 862 TDQ_UNLOCK(from); 863 sched_rem(td); 864 ts->ts_cpu = cpu; 865 td->td_lock = TDQ_LOCKPTR(to); 866 tdq_add(to, td, SRQ_YIELDING); 867 return (1); 868} 869 870/* 871 * This tdq has idled. Try to steal a thread from another cpu and switch 872 * to it. 873 */ 874static int 875tdq_idled(struct tdq *tdq) 876{ 877 struct cpu_group *cg; 878 struct tdq *steal; 879 cpumask_t mask; 880 int thresh; 881 int cpu; 882 883 if (smp_started == 0 || steal_idle == 0) 884 return (1); 885 mask = -1; 886 mask &= ~PCPU_GET(cpumask); 887 /* We don't want to be preempted while we're iterating. */ 888 spinlock_enter(); 889 for (cg = tdq->tdq_cg; cg != NULL; ) { 890 if ((cg->cg_flags & (CG_FLAG_HTT | CG_FLAG_THREAD)) == 0) 891 thresh = steal_thresh; 892 else 893 thresh = 1; 894 cpu = sched_highest(cg, mask, thresh); 895 if (cpu == -1) { 896 cg = cg->cg_parent; 897 continue; 898 } 899 steal = TDQ_CPU(cpu); 900 mask &= ~(1 << cpu); 901 tdq_lock_pair(tdq, steal); 902 if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 903 tdq_unlock_pair(tdq, steal); 904 continue; 905 } 906 /* 907 * If a thread was added while interrupts were disabled don't 908 * steal one here. If we fail to acquire one due to affinity 909 * restrictions loop again with this cpu removed from the 910 * set. 911 */ 912 if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 913 tdq_unlock_pair(tdq, steal); 914 continue; 915 } 916 spinlock_exit(); 917 TDQ_UNLOCK(steal); 918 mi_switch(SW_VOL, NULL); 919 thread_unlock(curthread); 920 921 return (0); 922 } 923 spinlock_exit(); 924 return (1); 925} 926 927/* 928 * Notify a remote cpu of new work. Sends an IPI if criteria are met. 929 */ 930static void 931tdq_notify(struct tdq *tdq, struct thread *td) 932{ 933 int cpri; 934 int pri; 935 int cpu; 936 937 if (tdq->tdq_ipipending) 938 return; 939 cpu = td->td_sched->ts_cpu; 940 pri = td->td_priority; 941 cpri = pcpu_find(cpu)->pc_curthread->td_priority; 942 if (!sched_shouldpreempt(pri, cpri, 1)) 943 return; 944 tdq->tdq_ipipending = 1; 945 ipi_selected(1 << cpu, IPI_PREEMPT); 946} 947 948/* 949 * Steals load from a timeshare queue. Honors the rotating queue head 950 * index. 951 */ 952static struct thread * 953runq_steal_from(struct runq *rq, int cpu, u_char start) 954{ 955 struct rqbits *rqb; 956 struct rqhead *rqh; 957 struct thread *td; 958 int first; 959 int bit; 960 int pri; 961 int i; 962 963 rqb = &rq->rq_status; 964 bit = start & (RQB_BPW -1); 965 pri = 0; 966 first = 0; 967again: 968 for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 969 if (rqb->rqb_bits[i] == 0) 970 continue; 971 if (bit != 0) { 972 for (pri = bit; pri < RQB_BPW; pri++) 973 if (rqb->rqb_bits[i] & (1ul << pri)) 974 break; 975 if (pri >= RQB_BPW) 976 continue; 977 } else 978 pri = RQB_FFS(rqb->rqb_bits[i]); 979 pri += (i << RQB_L2BPW); 980 rqh = &rq->rq_queues[pri]; 981 TAILQ_FOREACH(td, rqh, td_runq) { 982 if (first && THREAD_CAN_MIGRATE(td) && 983 THREAD_CAN_SCHED(td, cpu)) 984 return (td); 985 first = 1; 986 } 987 } 988 if (start != 0) { 989 start = 0; 990 goto again; 991 } 992 993 return (NULL); 994} 995 996/* 997 * Steals load from a standard linear queue. 998 */ 999static struct thread * 1000runq_steal(struct runq *rq, int cpu) 1001{ 1002 struct rqhead *rqh; 1003 struct rqbits *rqb; 1004 struct thread *td; 1005 int word; 1006 int bit; 1007 1008 rqb = &rq->rq_status; 1009 for (word = 0; word < RQB_LEN; word++) { 1010 if (rqb->rqb_bits[word] == 0) 1011 continue; 1012 for (bit = 0; bit < RQB_BPW; bit++) { 1013 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1014 continue; 1015 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1016 TAILQ_FOREACH(td, rqh, td_runq) 1017 if (THREAD_CAN_MIGRATE(td) && 1018 THREAD_CAN_SCHED(td, cpu)) 1019 return (td); 1020 } 1021 } 1022 return (NULL); 1023} 1024 1025/* 1026 * Attempt to steal a thread in priority order from a thread queue. 1027 */ 1028static struct thread * 1029tdq_steal(struct tdq *tdq, int cpu) 1030{ 1031 struct thread *td; 1032 1033 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1034 if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1035 return (td); 1036 if ((td = runq_steal_from(&tdq->tdq_timeshare, 1037 cpu, tdq->tdq_ridx)) != NULL) 1038 return (td); 1039 return (runq_steal(&tdq->tdq_idle, cpu)); 1040} 1041 1042/* 1043 * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1044 * current lock and returns with the assigned queue locked. 1045 */ 1046static inline struct tdq * 1047sched_setcpu(struct thread *td, int cpu, int flags) 1048{ 1049 1050 struct tdq *tdq; 1051 1052 THREAD_LOCK_ASSERT(td, MA_OWNED); 1053 tdq = TDQ_CPU(cpu); 1054 td->td_sched->ts_cpu = cpu; 1055 /* 1056 * If the lock matches just return the queue. 1057 */ 1058 if (td->td_lock == TDQ_LOCKPTR(tdq)) 1059 return (tdq); 1060#ifdef notyet 1061 /* 1062 * If the thread isn't running its lockptr is a 1063 * turnstile or a sleepqueue. We can just lock_set without 1064 * blocking. 1065 */ 1066 if (TD_CAN_RUN(td)) { 1067 TDQ_LOCK(tdq); 1068 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1069 return (tdq); 1070 } 1071#endif 1072 /* 1073 * The hard case, migration, we need to block the thread first to 1074 * prevent order reversals with other cpus locks. 1075 */ 1076 thread_lock_block(td); 1077 TDQ_LOCK(tdq); 1078 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1079 return (tdq); 1080} 1081 1082static int 1083sched_pickcpu(struct thread *td, int flags) 1084{ 1085 struct cpu_group *cg; 1086 struct td_sched *ts; 1087 struct tdq *tdq; 1088 cpumask_t mask; 1089 int self; 1090 int pri; 1091 int cpu; 1092 1093 self = PCPU_GET(cpuid); 1094 ts = td->td_sched; 1095 if (smp_started == 0) 1096 return (self); 1097 /* 1098 * Don't migrate a running thread from sched_switch(). 1099 */ 1100 if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1101 return (ts->ts_cpu); 1102 /* 1103 * Prefer to run interrupt threads on the processors that generate 1104 * the interrupt. 1105 */ 1106 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1107 curthread->td_intr_nesting_level) 1108 ts->ts_cpu = self; 1109 /* 1110 * If the thread can run on the last cpu and the affinity has not 1111 * expired or it is idle run it there. 1112 */ 1113 pri = td->td_priority; 1114 tdq = TDQ_CPU(ts->ts_cpu); 1115 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) { 1116 if (tdq->tdq_lowpri > PRI_MIN_IDLE) 1117 return (ts->ts_cpu); 1118 if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri) 1119 return (ts->ts_cpu); 1120 } 1121 /* 1122 * Search for the highest level in the tree that still has affinity. 1123 */ 1124 cg = NULL; 1125 for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent) 1126 if (SCHED_AFFINITY(ts, cg->cg_level)) 1127 break; 1128 cpu = -1; 1129 mask = td->td_cpuset->cs_mask.__bits[0]; 1130 if (cg) 1131 cpu = sched_lowest(cg, mask, pri); 1132 if (cpu == -1) 1133 cpu = sched_lowest(cpu_top, mask, -1); 1134 /* 1135 * Compare the lowest loaded cpu to current cpu. 1136 */ 1137 if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1138 TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) 1139 cpu = self; 1140 KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1141 return (cpu); 1142} 1143#endif 1144 1145/* 1146 * Pick the highest priority task we have and return it. 1147 */ 1148static struct thread * 1149tdq_choose(struct tdq *tdq) 1150{ 1151 struct thread *td; 1152 1153 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1154 td = runq_choose(&tdq->tdq_realtime); 1155 if (td != NULL) 1156 return (td); 1157 td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1158 if (td != NULL) { 1159 KASSERT(td->td_priority >= PRI_MIN_TIMESHARE, 1160 ("tdq_choose: Invalid priority on timeshare queue %d", 1161 td->td_priority)); 1162 return (td); 1163 } 1164 td = runq_choose(&tdq->tdq_idle); 1165 if (td != NULL) { 1166 KASSERT(td->td_priority >= PRI_MIN_IDLE, 1167 ("tdq_choose: Invalid priority on idle queue %d", 1168 td->td_priority)); 1169 return (td); 1170 } 1171 1172 return (NULL); 1173} 1174 1175/* 1176 * Initialize a thread queue. 1177 */ 1178static void 1179tdq_setup(struct tdq *tdq) 1180{ 1181 1182 if (bootverbose) 1183 printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1184 runq_init(&tdq->tdq_realtime); 1185 runq_init(&tdq->tdq_timeshare); 1186 runq_init(&tdq->tdq_idle); 1187 snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1188 "sched lock %d", (int)TDQ_ID(tdq)); 1189 mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1190 MTX_SPIN | MTX_RECURSE); 1191} 1192 1193#ifdef SMP 1194static void 1195sched_setup_smp(void) 1196{ 1197 struct tdq *tdq; 1198 int i; 1199 1200 cpu_top = smp_topo(); 1201 for (i = 0; i < MAXCPU; i++) { 1202 if (CPU_ABSENT(i)) 1203 continue; 1204 tdq = TDQ_CPU(i); 1205 tdq_setup(tdq); 1206 tdq->tdq_cg = smp_topo_find(cpu_top, i); 1207 if (tdq->tdq_cg == NULL) 1208 panic("Can't find cpu group for %d\n", i); 1209 } 1210 balance_tdq = TDQ_SELF(); 1211 sched_balance(); 1212} 1213#endif 1214 1215/* 1216 * Setup the thread queues and initialize the topology based on MD 1217 * information. 1218 */ 1219static void 1220sched_setup(void *dummy) 1221{ 1222 struct tdq *tdq; 1223 1224 tdq = TDQ_SELF(); 1225#ifdef SMP 1226 sched_setup_smp(); 1227#else 1228 tdq_setup(tdq); 1229#endif 1230 /* 1231 * To avoid divide-by-zero, we set realstathz a dummy value 1232 * in case which sched_clock() called before sched_initticks(). 1233 */ 1234 realstathz = hz; 1235 sched_slice = (realstathz/10); /* ~100ms */ 1236 tickincr = 1 << SCHED_TICK_SHIFT; 1237 1238 /* Add thread0's load since it's running. */ 1239 TDQ_LOCK(tdq); 1240 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1241 tdq_load_add(tdq, &thread0); 1242 tdq->tdq_lowpri = thread0.td_priority; 1243 TDQ_UNLOCK(tdq); 1244} 1245 1246/* 1247 * This routine determines the tickincr after stathz and hz are setup. 1248 */ 1249/* ARGSUSED */ 1250static void 1251sched_initticks(void *dummy) 1252{ 1253 int incr; 1254 1255 realstathz = stathz ? stathz : hz; 1256 sched_slice = (realstathz/10); /* ~100ms */ 1257 1258 /* 1259 * tickincr is shifted out by 10 to avoid rounding errors due to 1260 * hz not being evenly divisible by stathz on all platforms. 1261 */ 1262 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1263 /* 1264 * This does not work for values of stathz that are more than 1265 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1266 */ 1267 if (incr == 0) 1268 incr = 1; 1269 tickincr = incr; 1270#ifdef SMP 1271 /* 1272 * Set the default balance interval now that we know 1273 * what realstathz is. 1274 */ 1275 balance_interval = realstathz; 1276 /* 1277 * Set steal thresh to log2(mp_ncpu) but no greater than 4. This 1278 * prevents excess thrashing on large machines and excess idle on 1279 * smaller machines. 1280 */ 1281 steal_thresh = min(ffs(mp_ncpus) - 1, 3); 1282 affinity = SCHED_AFFINITY_DEFAULT; 1283#endif 1284} 1285 1286 1287/* 1288 * This is the core of the interactivity algorithm. Determines a score based 1289 * on past behavior. It is the ratio of sleep time to run time scaled to 1290 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1291 * differs from the cpu usage because it does not account for time spent 1292 * waiting on a run-queue. Would be prettier if we had floating point. 1293 */ 1294static int 1295sched_interact_score(struct thread *td) 1296{ 1297 struct td_sched *ts; 1298 int div; 1299 1300 ts = td->td_sched; 1301 /* 1302 * The score is only needed if this is likely to be an interactive 1303 * task. Don't go through the expense of computing it if there's 1304 * no chance. 1305 */ 1306 if (sched_interact <= SCHED_INTERACT_HALF && 1307 ts->ts_runtime >= ts->ts_slptime) 1308 return (SCHED_INTERACT_HALF); 1309 1310 if (ts->ts_runtime > ts->ts_slptime) { 1311 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1312 return (SCHED_INTERACT_HALF + 1313 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1314 } 1315 if (ts->ts_slptime > ts->ts_runtime) { 1316 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1317 return (ts->ts_runtime / div); 1318 } 1319 /* runtime == slptime */ 1320 if (ts->ts_runtime) 1321 return (SCHED_INTERACT_HALF); 1322 1323 /* 1324 * This can happen if slptime and runtime are 0. 1325 */ 1326 return (0); 1327 1328} 1329 1330/* 1331 * Scale the scheduling priority according to the "interactivity" of this 1332 * process. 1333 */ 1334static void 1335sched_priority(struct thread *td) 1336{ 1337 int score; 1338 int pri; 1339 1340 if (td->td_pri_class != PRI_TIMESHARE) 1341 return; 1342 /* 1343 * If the score is interactive we place the thread in the realtime 1344 * queue with a priority that is less than kernel and interrupt 1345 * priorities. These threads are not subject to nice restrictions. 1346 * 1347 * Scores greater than this are placed on the normal timeshare queue 1348 * where the priority is partially decided by the most recent cpu 1349 * utilization and the rest is decided by nice value. 1350 * 1351 * The nice value of the process has a linear effect on the calculated 1352 * score. Negative nice values make it easier for a thread to be 1353 * considered interactive. 1354 */ 1355 score = imax(0, sched_interact_score(td) - td->td_proc->p_nice); 1356 if (score < sched_interact) { 1357 pri = PRI_MIN_REALTIME; 1358 pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact) 1359 * score; 1360 KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME, 1361 ("sched_priority: invalid interactive priority %d score %d", 1362 pri, score)); 1363 } else { 1364 pri = SCHED_PRI_MIN; 1365 if (td->td_sched->ts_ticks) 1366 pri += SCHED_PRI_TICKS(td->td_sched); 1367 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1368 KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE, 1369 ("sched_priority: invalid priority %d: nice %d, " 1370 "ticks %d ftick %d ltick %d tick pri %d", 1371 pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1372 td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1373 SCHED_PRI_TICKS(td->td_sched))); 1374 } 1375 sched_user_prio(td, pri); 1376 1377 return; 1378} 1379 1380/* 1381 * This routine enforces a maximum limit on the amount of scheduling history 1382 * kept. It is called after either the slptime or runtime is adjusted. This 1383 * function is ugly due to integer math. 1384 */ 1385static void 1386sched_interact_update(struct thread *td) 1387{ 1388 struct td_sched *ts; 1389 u_int sum; 1390 1391 ts = td->td_sched; 1392 sum = ts->ts_runtime + ts->ts_slptime; 1393 if (sum < SCHED_SLP_RUN_MAX) 1394 return; 1395 /* 1396 * This only happens from two places: 1397 * 1) We have added an unusual amount of run time from fork_exit. 1398 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1399 */ 1400 if (sum > SCHED_SLP_RUN_MAX * 2) { 1401 if (ts->ts_runtime > ts->ts_slptime) { 1402 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1403 ts->ts_slptime = 1; 1404 } else { 1405 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1406 ts->ts_runtime = 1; 1407 } 1408 return; 1409 } 1410 /* 1411 * If we have exceeded by more than 1/5th then the algorithm below 1412 * will not bring us back into range. Dividing by two here forces 1413 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1414 */ 1415 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1416 ts->ts_runtime /= 2; 1417 ts->ts_slptime /= 2; 1418 return; 1419 } 1420 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1421 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1422} 1423 1424/* 1425 * Scale back the interactivity history when a child thread is created. The 1426 * history is inherited from the parent but the thread may behave totally 1427 * differently. For example, a shell spawning a compiler process. We want 1428 * to learn that the compiler is behaving badly very quickly. 1429 */ 1430static void 1431sched_interact_fork(struct thread *td) 1432{ 1433 int ratio; 1434 int sum; 1435 1436 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1437 if (sum > SCHED_SLP_RUN_FORK) { 1438 ratio = sum / SCHED_SLP_RUN_FORK; 1439 td->td_sched->ts_runtime /= ratio; 1440 td->td_sched->ts_slptime /= ratio; 1441 } 1442} 1443 1444/* 1445 * Called from proc0_init() to setup the scheduler fields. 1446 */ 1447void 1448schedinit(void) 1449{ 1450 1451 /* 1452 * Set up the scheduler specific parts of proc0. 1453 */ 1454 proc0.p_sched = NULL; /* XXX */ 1455 thread0.td_sched = &td_sched0; 1456 td_sched0.ts_ltick = ticks; 1457 td_sched0.ts_ftick = ticks; 1458 td_sched0.ts_slice = sched_slice; 1459} 1460 1461/* 1462 * This is only somewhat accurate since given many processes of the same 1463 * priority they will switch when their slices run out, which will be 1464 * at most sched_slice stathz ticks. 1465 */ 1466int 1467sched_rr_interval(void) 1468{ 1469 1470 /* Convert sched_slice to hz */ 1471 return (hz/(realstathz/sched_slice)); 1472} 1473 1474/* 1475 * Update the percent cpu tracking information when it is requested or 1476 * the total history exceeds the maximum. We keep a sliding history of 1477 * tick counts that slowly decays. This is less precise than the 4BSD 1478 * mechanism since it happens with less regular and frequent events. 1479 */ 1480static void 1481sched_pctcpu_update(struct td_sched *ts) 1482{ 1483 1484 if (ts->ts_ticks == 0) 1485 return; 1486 if (ticks - (hz / 10) < ts->ts_ltick && 1487 SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1488 return; 1489 /* 1490 * Adjust counters and watermark for pctcpu calc. 1491 */ 1492 if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1493 ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1494 SCHED_TICK_TARG; 1495 else 1496 ts->ts_ticks = 0; 1497 ts->ts_ltick = ticks; 1498 ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1499} 1500 1501/* 1502 * Adjust the priority of a thread. Move it to the appropriate run-queue 1503 * if necessary. This is the back-end for several priority related 1504 * functions. 1505 */ 1506static void 1507sched_thread_priority(struct thread *td, u_char prio) 1508{ 1509 struct td_sched *ts; 1510 struct tdq *tdq; 1511 int oldpri; 1512 1513 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 1514 td, td->td_name, td->td_priority, prio, curthread, 1515 curthread->td_name); 1516 ts = td->td_sched; 1517 THREAD_LOCK_ASSERT(td, MA_OWNED); 1518 if (td->td_priority == prio) 1519 return; 1520 /* 1521 * If the priority has been elevated due to priority 1522 * propagation, we may have to move ourselves to a new 1523 * queue. This could be optimized to not re-add in some 1524 * cases. 1525 */ 1526 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1527 sched_rem(td); 1528 td->td_priority = prio; 1529 sched_add(td, SRQ_BORROWING); 1530 return; 1531 } 1532 /* 1533 * If the thread is currently running we may have to adjust the lowpri 1534 * information so other cpus are aware of our current priority. 1535 */ 1536 if (TD_IS_RUNNING(td)) { 1537 tdq = TDQ_CPU(ts->ts_cpu); 1538 oldpri = td->td_priority; 1539 td->td_priority = prio; 1540 if (prio < tdq->tdq_lowpri) 1541 tdq->tdq_lowpri = prio; 1542 else if (tdq->tdq_lowpri == oldpri) 1543 tdq_setlowpri(tdq, td); 1544 return; 1545 } 1546 td->td_priority = prio; 1547} 1548 1549/* 1550 * Update a thread's priority when it is lent another thread's 1551 * priority. 1552 */ 1553void 1554sched_lend_prio(struct thread *td, u_char prio) 1555{ 1556 1557 td->td_flags |= TDF_BORROWING; 1558 sched_thread_priority(td, prio); 1559} 1560 1561/* 1562 * Restore a thread's priority when priority propagation is 1563 * over. The prio argument is the minimum priority the thread 1564 * needs to have to satisfy other possible priority lending 1565 * requests. If the thread's regular priority is less 1566 * important than prio, the thread will keep a priority boost 1567 * of prio. 1568 */ 1569void 1570sched_unlend_prio(struct thread *td, u_char prio) 1571{ 1572 u_char base_pri; 1573 1574 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1575 td->td_base_pri <= PRI_MAX_TIMESHARE) 1576 base_pri = td->td_user_pri; 1577 else 1578 base_pri = td->td_base_pri; 1579 if (prio >= base_pri) { 1580 td->td_flags &= ~TDF_BORROWING; 1581 sched_thread_priority(td, base_pri); 1582 } else 1583 sched_lend_prio(td, prio); 1584} 1585 1586/* 1587 * Standard entry for setting the priority to an absolute value. 1588 */ 1589void 1590sched_prio(struct thread *td, u_char prio) 1591{ 1592 u_char oldprio; 1593 1594 /* First, update the base priority. */ 1595 td->td_base_pri = prio; 1596 1597 /* 1598 * If the thread is borrowing another thread's priority, don't 1599 * ever lower the priority. 1600 */ 1601 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1602 return; 1603 1604 /* Change the real priority. */ 1605 oldprio = td->td_priority; 1606 sched_thread_priority(td, prio); 1607 1608 /* 1609 * If the thread is on a turnstile, then let the turnstile update 1610 * its state. 1611 */ 1612 if (TD_ON_LOCK(td) && oldprio != prio) 1613 turnstile_adjust(td, oldprio); 1614} 1615 1616/* 1617 * Set the base user priority, does not effect current running priority. 1618 */ 1619void 1620sched_user_prio(struct thread *td, u_char prio) 1621{ 1622 u_char oldprio; 1623 1624 td->td_base_user_pri = prio; 1625 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio) 1626 return; 1627 oldprio = td->td_user_pri; 1628 td->td_user_pri = prio; 1629} 1630 1631void 1632sched_lend_user_prio(struct thread *td, u_char prio) 1633{ 1634 u_char oldprio; 1635 1636 THREAD_LOCK_ASSERT(td, MA_OWNED); 1637 td->td_flags |= TDF_UBORROWING; 1638 oldprio = td->td_user_pri; 1639 td->td_user_pri = prio; 1640} 1641 1642void 1643sched_unlend_user_prio(struct thread *td, u_char prio) 1644{ 1645 u_char base_pri; 1646 1647 THREAD_LOCK_ASSERT(td, MA_OWNED); 1648 base_pri = td->td_base_user_pri; 1649 if (prio >= base_pri) { 1650 td->td_flags &= ~TDF_UBORROWING; 1651 sched_user_prio(td, base_pri); 1652 } else { 1653 sched_lend_user_prio(td, prio); 1654 } 1655} 1656 1657/* 1658 * Block a thread for switching. Similar to thread_block() but does not 1659 * bump the spin count. 1660 */ 1661static inline struct mtx * 1662thread_block_switch(struct thread *td) 1663{ 1664 struct mtx *lock; 1665 1666 THREAD_LOCK_ASSERT(td, MA_OWNED); 1667 lock = td->td_lock; 1668 td->td_lock = &blocked_lock; 1669 mtx_unlock_spin(lock); 1670 1671 return (lock); 1672} 1673 1674/* 1675 * Handle migration from sched_switch(). This happens only for 1676 * cpu binding. 1677 */ 1678static struct mtx * 1679sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1680{ 1681 struct tdq *tdn; 1682 1683 tdn = TDQ_CPU(td->td_sched->ts_cpu); 1684#ifdef SMP 1685 tdq_load_rem(tdq, td); 1686 /* 1687 * Do the lock dance required to avoid LOR. We grab an extra 1688 * spinlock nesting to prevent preemption while we're 1689 * not holding either run-queue lock. 1690 */ 1691 spinlock_enter(); 1692 thread_block_switch(td); /* This releases the lock on tdq. */ 1693 TDQ_LOCK(tdn); 1694 tdq_add(tdn, td, flags); 1695 tdq_notify(tdn, td); 1696 /* 1697 * After we unlock tdn the new cpu still can't switch into this 1698 * thread until we've unblocked it in cpu_switch(). The lock 1699 * pointers may match in the case of HTT cores. Don't unlock here 1700 * or we can deadlock when the other CPU runs the IPI handler. 1701 */ 1702 if (TDQ_LOCKPTR(tdn) != TDQ_LOCKPTR(tdq)) { 1703 TDQ_UNLOCK(tdn); 1704 TDQ_LOCK(tdq); 1705 } 1706 spinlock_exit(); 1707#endif 1708 return (TDQ_LOCKPTR(tdn)); 1709} 1710 1711/* 1712 * Release a thread that was blocked with thread_block_switch(). 1713 */ 1714static inline void 1715thread_unblock_switch(struct thread *td, struct mtx *mtx) 1716{ 1717 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1718 (uintptr_t)mtx); 1719} 1720 1721/* 1722 * Switch threads. This function has to handle threads coming in while 1723 * blocked for some reason, running, or idle. It also must deal with 1724 * migrating a thread from one queue to another as running threads may 1725 * be assigned elsewhere via binding. 1726 */ 1727void 1728sched_switch(struct thread *td, struct thread *newtd, int flags) 1729{ 1730 struct tdq *tdq; 1731 struct td_sched *ts; 1732 struct mtx *mtx; 1733 int srqflag; 1734 int cpuid; 1735 1736 THREAD_LOCK_ASSERT(td, MA_OWNED); 1737 KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1738 1739 cpuid = PCPU_GET(cpuid); 1740 tdq = TDQ_CPU(cpuid); 1741 ts = td->td_sched; 1742 mtx = td->td_lock; 1743 ts->ts_rltick = ticks; 1744 td->td_lastcpu = td->td_oncpu; 1745 td->td_oncpu = NOCPU; 1746 td->td_flags &= ~TDF_NEEDRESCHED; 1747 td->td_owepreempt = 0; 1748 /* 1749 * The lock pointer in an idle thread should never change. Reset it 1750 * to CAN_RUN as well. 1751 */ 1752 if (TD_IS_IDLETHREAD(td)) { 1753 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1754 TD_SET_CAN_RUN(td); 1755 } else if (TD_IS_RUNNING(td)) { 1756 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1757 srqflag = (flags & SW_PREEMPT) ? 1758 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1759 SRQ_OURSELF|SRQ_YIELDING; 1760 if (ts->ts_cpu == cpuid) 1761 tdq_runq_add(tdq, td, srqflag); 1762 else 1763 mtx = sched_switch_migrate(tdq, td, srqflag); 1764 } else { 1765 /* This thread must be going to sleep. */ 1766 TDQ_LOCK(tdq); 1767 mtx = thread_block_switch(td); 1768 tdq_load_rem(tdq, td); 1769 } 1770 /* 1771 * We enter here with the thread blocked and assigned to the 1772 * appropriate cpu run-queue or sleep-queue and with the current 1773 * thread-queue locked. 1774 */ 1775 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1776 newtd = choosethread(); 1777 /* 1778 * Call the MD code to switch contexts if necessary. 1779 */ 1780 if (td != newtd) { 1781#ifdef HWPMC_HOOKS 1782 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1783 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1784#endif 1785 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1786 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1787 cpu_switch(td, newtd, mtx); 1788 /* 1789 * We may return from cpu_switch on a different cpu. However, 1790 * we always return with td_lock pointing to the current cpu's 1791 * run queue lock. 1792 */ 1793 cpuid = PCPU_GET(cpuid); 1794 tdq = TDQ_CPU(cpuid); 1795 lock_profile_obtain_lock_success( 1796 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1797#ifdef HWPMC_HOOKS 1798 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1799 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1800#endif 1801 } else 1802 thread_unblock_switch(td, mtx); 1803 /* 1804 * Assert that all went well and return. 1805 */ 1806 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1807 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1808 td->td_oncpu = cpuid; 1809} 1810 1811/* 1812 * Adjust thread priorities as a result of a nice request. 1813 */ 1814void 1815sched_nice(struct proc *p, int nice) 1816{ 1817 struct thread *td; 1818 1819 PROC_LOCK_ASSERT(p, MA_OWNED); 1820 1821 p->p_nice = nice; 1822 FOREACH_THREAD_IN_PROC(p, td) { 1823 thread_lock(td); 1824 sched_priority(td); 1825 sched_prio(td, td->td_base_user_pri); 1826 thread_unlock(td); 1827 } 1828} 1829 1830/* 1831 * Record the sleep time for the interactivity scorer. 1832 */ 1833void 1834sched_sleep(struct thread *td, int prio) 1835{ 1836 1837 THREAD_LOCK_ASSERT(td, MA_OWNED); 1838 1839 td->td_slptick = ticks; 1840 if (TD_IS_SUSPENDED(td) || prio <= PSOCK) 1841 td->td_flags |= TDF_CANSWAP; 1842 if (static_boost && prio) 1843 sched_prio(td, prio); 1844} 1845 1846/* 1847 * Schedule a thread to resume execution and record how long it voluntarily 1848 * slept. We also update the pctcpu, interactivity, and priority. 1849 */ 1850void 1851sched_wakeup(struct thread *td) 1852{ 1853 struct td_sched *ts; 1854 int slptick; 1855 1856 THREAD_LOCK_ASSERT(td, MA_OWNED); 1857 ts = td->td_sched; 1858 td->td_flags &= ~TDF_CANSWAP; 1859 /* 1860 * If we slept for more than a tick update our interactivity and 1861 * priority. 1862 */ 1863 slptick = td->td_slptick; 1864 td->td_slptick = 0; 1865 if (slptick && slptick != ticks) { 1866 u_int hzticks; 1867 1868 hzticks = (ticks - slptick) << SCHED_TICK_SHIFT; 1869 ts->ts_slptime += hzticks; 1870 sched_interact_update(td); 1871 sched_pctcpu_update(ts); 1872 } 1873 /* Reset the slice value after we sleep. */ 1874 ts->ts_slice = sched_slice; 1875 sched_add(td, SRQ_BORING); 1876} 1877 1878/* 1879 * Penalize the parent for creating a new child and initialize the child's 1880 * priority. 1881 */ 1882void 1883sched_fork(struct thread *td, struct thread *child) 1884{ 1885 THREAD_LOCK_ASSERT(td, MA_OWNED); 1886 sched_fork_thread(td, child); 1887 /* 1888 * Penalize the parent and child for forking. 1889 */ 1890 sched_interact_fork(child); 1891 sched_priority(child); 1892 td->td_sched->ts_runtime += tickincr; 1893 sched_interact_update(td); 1894 sched_priority(td); 1895} 1896 1897/* 1898 * Fork a new thread, may be within the same process. 1899 */ 1900void 1901sched_fork_thread(struct thread *td, struct thread *child) 1902{ 1903 struct td_sched *ts; 1904 struct td_sched *ts2; 1905 1906 THREAD_LOCK_ASSERT(td, MA_OWNED); 1907 /* 1908 * Initialize child. 1909 */ 1910 ts = td->td_sched; 1911 ts2 = child->td_sched; 1912 child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1913 child->td_cpuset = cpuset_ref(td->td_cpuset); 1914 ts2->ts_cpu = ts->ts_cpu; 1915 ts2->ts_flags = 0; 1916 /* 1917 * Grab our parents cpu estimation information and priority. 1918 */ 1919 ts2->ts_ticks = ts->ts_ticks; 1920 ts2->ts_ltick = ts->ts_ltick; 1921 ts2->ts_ftick = ts->ts_ftick; 1922 child->td_user_pri = td->td_user_pri; 1923 child->td_base_user_pri = td->td_base_user_pri; 1924 /* 1925 * And update interactivity score. 1926 */ 1927 ts2->ts_slptime = ts->ts_slptime; 1928 ts2->ts_runtime = ts->ts_runtime; 1929 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 1930} 1931 1932/* 1933 * Adjust the priority class of a thread. 1934 */ 1935void 1936sched_class(struct thread *td, int class) 1937{ 1938 1939 THREAD_LOCK_ASSERT(td, MA_OWNED); 1940 if (td->td_pri_class == class) 1941 return; 1942 td->td_pri_class = class; 1943} 1944 1945/* 1946 * Return some of the child's priority and interactivity to the parent. 1947 */ 1948void 1949sched_exit(struct proc *p, struct thread *child) 1950{ 1951 struct thread *td; 1952 1953 CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d", 1954 child, child->td_name, child->td_priority); 1955 1956 PROC_LOCK_ASSERT(p, MA_OWNED); 1957 td = FIRST_THREAD_IN_PROC(p); 1958 sched_exit_thread(td, child); 1959} 1960 1961/* 1962 * Penalize another thread for the time spent on this one. This helps to 1963 * worsen the priority and interactivity of processes which schedule batch 1964 * jobs such as make. This has little effect on the make process itself but 1965 * causes new processes spawned by it to receive worse scores immediately. 1966 */ 1967void 1968sched_exit_thread(struct thread *td, struct thread *child) 1969{ 1970 1971 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d", 1972 child, child->td_name, child->td_priority); 1973 1974 /* 1975 * Give the child's runtime to the parent without returning the 1976 * sleep time as a penalty to the parent. This causes shells that 1977 * launch expensive things to mark their children as expensive. 1978 */ 1979 thread_lock(td); 1980 td->td_sched->ts_runtime += child->td_sched->ts_runtime; 1981 sched_interact_update(td); 1982 sched_priority(td); 1983 thread_unlock(td); 1984} 1985 1986void 1987sched_preempt(struct thread *td) 1988{ 1989 struct tdq *tdq; 1990 1991 thread_lock(td); 1992 tdq = TDQ_SELF(); 1993 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1994 tdq->tdq_ipipending = 0; 1995 if (td->td_priority > tdq->tdq_lowpri) { 1996 if (td->td_critnest > 1) 1997 td->td_owepreempt = 1; 1998 else 1999 mi_switch(SW_INVOL | SW_PREEMPT, NULL); 2000 } 2001 thread_unlock(td); 2002} 2003 2004/* 2005 * Fix priorities on return to user-space. Priorities may be elevated due 2006 * to static priorities in msleep() or similar. 2007 */ 2008void 2009sched_userret(struct thread *td) 2010{ 2011 /* 2012 * XXX we cheat slightly on the locking here to avoid locking in 2013 * the usual case. Setting td_priority here is essentially an 2014 * incomplete workaround for not setting it properly elsewhere. 2015 * Now that some interrupt handlers are threads, not setting it 2016 * properly elsewhere can clobber it in the window between setting 2017 * it here and returning to user mode, so don't waste time setting 2018 * it perfectly here. 2019 */ 2020 KASSERT((td->td_flags & TDF_BORROWING) == 0, 2021 ("thread with borrowed priority returning to userland")); 2022 if (td->td_priority != td->td_user_pri) { 2023 thread_lock(td); 2024 td->td_priority = td->td_user_pri; 2025 td->td_base_pri = td->td_user_pri; 2026 tdq_setlowpri(TDQ_SELF(), td); 2027 thread_unlock(td); 2028 } 2029} 2030 2031/* 2032 * Handle a stathz tick. This is really only relevant for timeshare 2033 * threads. 2034 */ 2035void 2036sched_clock(struct thread *td) 2037{ 2038 struct tdq *tdq; 2039 struct td_sched *ts; 2040 2041 THREAD_LOCK_ASSERT(td, MA_OWNED); 2042 tdq = TDQ_SELF(); 2043#ifdef SMP 2044 /* 2045 * We run the long term load balancer infrequently on the first cpu. 2046 */ 2047 if (balance_tdq == tdq) { 2048 if (balance_ticks && --balance_ticks == 0) 2049 sched_balance(); 2050 } 2051#endif 2052 /* 2053 * Advance the insert index once for each tick to ensure that all 2054 * threads get a chance to run. 2055 */ 2056 if (tdq->tdq_idx == tdq->tdq_ridx) { 2057 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2058 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2059 tdq->tdq_ridx = tdq->tdq_idx; 2060 } 2061 ts = td->td_sched; 2062 if (td->td_pri_class & PRI_FIFO_BIT) 2063 return; 2064 if (td->td_pri_class == PRI_TIMESHARE) { 2065 /* 2066 * We used a tick; charge it to the thread so 2067 * that we can compute our interactivity. 2068 */ 2069 td->td_sched->ts_runtime += tickincr; 2070 sched_interact_update(td); 2071 sched_priority(td); 2072 } 2073 /* 2074 * We used up one time slice. 2075 */ 2076 if (--ts->ts_slice > 0) 2077 return; 2078 /* 2079 * We're out of time, force a requeue at userret(). 2080 */ 2081 ts->ts_slice = sched_slice; 2082 td->td_flags |= TDF_NEEDRESCHED; 2083} 2084 2085/* 2086 * Called once per hz tick. Used for cpu utilization information. This 2087 * is easier than trying to scale based on stathz. 2088 */ 2089void 2090sched_tick(void) 2091{ 2092 struct td_sched *ts; 2093 2094 ts = curthread->td_sched; 2095 /* Adjust ticks for pctcpu */ 2096 ts->ts_ticks += 1 << SCHED_TICK_SHIFT; 2097 ts->ts_ltick = ticks; 2098 /* 2099 * Update if we've exceeded our desired tick threshhold by over one 2100 * second. 2101 */ 2102 if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2103 sched_pctcpu_update(ts); 2104} 2105 2106/* 2107 * Return whether the current CPU has runnable tasks. Used for in-kernel 2108 * cooperative idle threads. 2109 */ 2110int 2111sched_runnable(void) 2112{ 2113 struct tdq *tdq; 2114 int load; 2115 2116 load = 1; 2117 2118 tdq = TDQ_SELF(); 2119 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2120 if (tdq->tdq_load > 0) 2121 goto out; 2122 } else 2123 if (tdq->tdq_load - 1 > 0) 2124 goto out; 2125 load = 0; 2126out: 2127 return (load); 2128} 2129 2130/* 2131 * Choose the highest priority thread to run. The thread is removed from 2132 * the run-queue while running however the load remains. For SMP we set 2133 * the tdq in the global idle bitmask if it idles here. 2134 */ 2135struct thread * 2136sched_choose(void) 2137{ 2138 struct thread *td; 2139 struct tdq *tdq; 2140 2141 tdq = TDQ_SELF(); 2142 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2143 td = tdq_choose(tdq); 2144 if (td) { 2145 td->td_sched->ts_ltick = ticks; 2146 tdq_runq_rem(tdq, td); 2147 return (td); 2148 } 2149 return (PCPU_GET(idlethread)); 2150} 2151 2152/* 2153 * Set owepreempt if necessary. Preemption never happens directly in ULE, 2154 * we always request it once we exit a critical section. 2155 */ 2156static inline void 2157sched_setpreempt(struct thread *td) 2158{ 2159 struct thread *ctd; 2160 int cpri; 2161 int pri; 2162 2163 THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2164 2165 ctd = curthread; 2166 pri = td->td_priority; 2167 cpri = ctd->td_priority; 2168 if (pri < cpri) 2169 ctd->td_flags |= TDF_NEEDRESCHED; 2170 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2171 return; 2172 if (!sched_shouldpreempt(pri, cpri, 0)) 2173 return; 2174 ctd->td_owepreempt = 1; 2175} 2176 2177/* 2178 * Add a thread to a thread queue. Select the appropriate runq and add the 2179 * thread to it. This is the internal function called when the tdq is 2180 * predetermined. 2181 */ 2182void 2183tdq_add(struct tdq *tdq, struct thread *td, int flags) 2184{ 2185 2186 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2187 KASSERT((td->td_inhibitors == 0), 2188 ("sched_add: trying to run inhibited thread")); 2189 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2190 ("sched_add: bad thread state")); 2191 KASSERT(td->td_flags & TDF_INMEM, 2192 ("sched_add: thread swapped out")); 2193 2194 if (td->td_priority < tdq->tdq_lowpri) 2195 tdq->tdq_lowpri = td->td_priority; 2196 tdq_runq_add(tdq, td, flags); 2197 tdq_load_add(tdq, td); 2198} 2199 2200/* 2201 * Select the target thread queue and add a thread to it. Request 2202 * preemption or IPI a remote processor if required. 2203 */ 2204void 2205sched_add(struct thread *td, int flags) 2206{ 2207 struct tdq *tdq; 2208#ifdef SMP 2209 int cpu; 2210#endif 2211 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 2212 td, td->td_name, td->td_priority, curthread, 2213 curthread->td_name); 2214 THREAD_LOCK_ASSERT(td, MA_OWNED); 2215 /* 2216 * Recalculate the priority before we select the target cpu or 2217 * run-queue. 2218 */ 2219 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2220 sched_priority(td); 2221#ifdef SMP 2222 /* 2223 * Pick the destination cpu and if it isn't ours transfer to the 2224 * target cpu. 2225 */ 2226 cpu = sched_pickcpu(td, flags); 2227 tdq = sched_setcpu(td, cpu, flags); 2228 tdq_add(tdq, td, flags); 2229 if (cpu != PCPU_GET(cpuid)) { 2230 tdq_notify(tdq, td); 2231 return; 2232 } 2233#else 2234 tdq = TDQ_SELF(); 2235 TDQ_LOCK(tdq); 2236 /* 2237 * Now that the thread is moving to the run-queue, set the lock 2238 * to the scheduler's lock. 2239 */ 2240 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2241 tdq_add(tdq, td, flags); 2242#endif 2243 if (!(flags & SRQ_YIELDING)) 2244 sched_setpreempt(td); 2245} 2246 2247/* 2248 * Remove a thread from a run-queue without running it. This is used 2249 * when we're stealing a thread from a remote queue. Otherwise all threads 2250 * exit by calling sched_exit_thread() and sched_throw() themselves. 2251 */ 2252void 2253sched_rem(struct thread *td) 2254{ 2255 struct tdq *tdq; 2256 2257 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 2258 td, td->td_name, td->td_priority, curthread, 2259 curthread->td_name); 2260 tdq = TDQ_CPU(td->td_sched->ts_cpu); 2261 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2262 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2263 KASSERT(TD_ON_RUNQ(td), 2264 ("sched_rem: thread not on run queue")); 2265 tdq_runq_rem(tdq, td); 2266 tdq_load_rem(tdq, td); 2267 TD_SET_CAN_RUN(td); 2268 if (td->td_priority == tdq->tdq_lowpri) 2269 tdq_setlowpri(tdq, NULL); 2270} 2271 2272/* 2273 * Fetch cpu utilization information. Updates on demand. 2274 */ 2275fixpt_t 2276sched_pctcpu(struct thread *td) 2277{ 2278 fixpt_t pctcpu; 2279 struct td_sched *ts; 2280 2281 pctcpu = 0; 2282 ts = td->td_sched; 2283 if (ts == NULL) 2284 return (0); 2285 2286 thread_lock(td); 2287 if (ts->ts_ticks) { 2288 int rtick; 2289 2290 sched_pctcpu_update(ts); 2291 /* How many rtick per second ? */ 2292 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2293 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2294 } 2295 thread_unlock(td); 2296 2297 return (pctcpu); 2298} 2299 2300/* 2301 * Enforce affinity settings for a thread. Called after adjustments to 2302 * cpumask. 2303 */ 2304void 2305sched_affinity(struct thread *td) 2306{ 2307#ifdef SMP 2308 struct td_sched *ts; 2309 int cpu; 2310 2311 THREAD_LOCK_ASSERT(td, MA_OWNED); 2312 ts = td->td_sched; 2313 if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2314 return; 2315 if (!TD_IS_RUNNING(td)) 2316 return; 2317 td->td_flags |= TDF_NEEDRESCHED; 2318 if (!THREAD_CAN_MIGRATE(td)) 2319 return; 2320 /* 2321 * Assign the new cpu and force a switch before returning to 2322 * userspace. If the target thread is not running locally send 2323 * an ipi to force the issue. 2324 */ 2325 cpu = ts->ts_cpu; 2326 ts->ts_cpu = sched_pickcpu(td, 0); 2327 if (cpu != PCPU_GET(cpuid)) 2328 ipi_selected(1 << cpu, IPI_PREEMPT); 2329#endif 2330} 2331 2332/* 2333 * Bind a thread to a target cpu. 2334 */ 2335void 2336sched_bind(struct thread *td, int cpu) 2337{ 2338 struct td_sched *ts; 2339 2340 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2341 ts = td->td_sched; 2342 if (ts->ts_flags & TSF_BOUND) 2343 sched_unbind(td); 2344 ts->ts_flags |= TSF_BOUND; 2345 sched_pin(); 2346 if (PCPU_GET(cpuid) == cpu) 2347 return; 2348 ts->ts_cpu = cpu; 2349 /* When we return from mi_switch we'll be on the correct cpu. */ 2350 mi_switch(SW_VOL, NULL); 2351} 2352 2353/* 2354 * Release a bound thread. 2355 */ 2356void 2357sched_unbind(struct thread *td) 2358{ 2359 struct td_sched *ts; 2360 2361 THREAD_LOCK_ASSERT(td, MA_OWNED); 2362 ts = td->td_sched; 2363 if ((ts->ts_flags & TSF_BOUND) == 0) 2364 return; 2365 ts->ts_flags &= ~TSF_BOUND; 2366 sched_unpin(); 2367} 2368 2369int 2370sched_is_bound(struct thread *td) 2371{ 2372 THREAD_LOCK_ASSERT(td, MA_OWNED); 2373 return (td->td_sched->ts_flags & TSF_BOUND); 2374} 2375 2376/* 2377 * Basic yield call. 2378 */ 2379void 2380sched_relinquish(struct thread *td) 2381{ 2382 thread_lock(td); 2383 SCHED_STAT_INC(switch_relinquish); 2384 mi_switch(SW_VOL, NULL); 2385 thread_unlock(td); 2386} 2387 2388/* 2389 * Return the total system load. 2390 */ 2391int 2392sched_load(void) 2393{ 2394#ifdef SMP 2395 int total; 2396 int i; 2397 2398 total = 0; 2399 for (i = 0; i <= mp_maxid; i++) 2400 total += TDQ_CPU(i)->tdq_sysload; 2401 return (total); 2402#else 2403 return (TDQ_SELF()->tdq_sysload); 2404#endif 2405} 2406 2407int 2408sched_sizeof_proc(void) 2409{ 2410 return (sizeof(struct proc)); 2411} 2412 2413int 2414sched_sizeof_thread(void) 2415{ 2416 return (sizeof(struct thread) + sizeof(struct td_sched)); 2417} 2418 2419/* 2420 * The actual idle process. 2421 */ 2422void 2423sched_idletd(void *dummy) 2424{ 2425 struct thread *td; 2426 struct tdq *tdq; 2427 2428 td = curthread; 2429 tdq = TDQ_SELF(); 2430 mtx_assert(&Giant, MA_NOTOWNED); 2431 /* ULE relies on preemption for idle interruption. */ 2432 for (;;) { 2433#ifdef SMP 2434 if (tdq_idled(tdq)) 2435 cpu_idle(); 2436#else 2437 cpu_idle(); 2438#endif 2439 } 2440} 2441 2442/* 2443 * A CPU is entering for the first time or a thread is exiting. 2444 */ 2445void 2446sched_throw(struct thread *td) 2447{ 2448 struct thread *newtd; 2449 struct tdq *tdq; 2450 2451 tdq = TDQ_SELF(); 2452 if (td == NULL) { 2453 /* Correct spinlock nesting and acquire the correct lock. */ 2454 TDQ_LOCK(tdq); 2455 spinlock_exit(); 2456 } else { 2457 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2458 tdq_load_rem(tdq, td); 2459 lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2460 } 2461 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2462 newtd = choosethread(); 2463 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2464 PCPU_SET(switchtime, cpu_ticks()); 2465 PCPU_SET(switchticks, ticks); 2466 cpu_throw(td, newtd); /* doesn't return */ 2467} 2468 2469/* 2470 * This is called from fork_exit(). Just acquire the correct locks and 2471 * let fork do the rest of the work. 2472 */ 2473void 2474sched_fork_exit(struct thread *td) 2475{ 2476 struct td_sched *ts; 2477 struct tdq *tdq; 2478 int cpuid; 2479 2480 /* 2481 * Finish setting up thread glue so that it begins execution in a 2482 * non-nested critical section with the scheduler lock held. 2483 */ 2484 cpuid = PCPU_GET(cpuid); 2485 tdq = TDQ_CPU(cpuid); 2486 ts = td->td_sched; 2487 if (TD_IS_IDLETHREAD(td)) 2488 td->td_lock = TDQ_LOCKPTR(tdq); 2489 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2490 td->td_oncpu = cpuid; 2491 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2492 lock_profile_obtain_lock_success( 2493 &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2494} 2495 2496SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2497SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2498 "Scheduler name"); 2499SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2500 "Slice size for timeshare threads"); 2501SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2502 "Interactivity score threshold"); 2503SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2504 0,"Min priority for preemption, lower priorities have greater precedence"); 2505SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 2506 0,"Controls whether static kernel priorities are assigned to sleeping threads."); 2507#ifdef SMP 2508SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2509 "Number of hz ticks to keep thread affinity for"); 2510SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2511 "Enables the long-term load balancer"); 2512SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2513 &balance_interval, 0, 2514 "Average frequency in stathz ticks to run the long-term balancer"); 2515SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, 2516 "Steals work from another hyper-threaded core on idle"); 2517SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2518 "Attempts to steal work from other cores before idling"); 2519SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2520 "Minimum load on remote cpu before we'll steal"); 2521#endif 2522 2523/* ps compat. All cpu percentages from ULE are weighted. */ 2524static int ccpu = 0; 2525SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2526