sched_ule.c revision 208982
1109864Sjeff/*- 2165762Sjeff * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org> 3109864Sjeff * All rights reserved. 4109864Sjeff * 5109864Sjeff * Redistribution and use in source and binary forms, with or without 6109864Sjeff * modification, are permitted provided that the following conditions 7109864Sjeff * are met: 8109864Sjeff * 1. Redistributions of source code must retain the above copyright 9109864Sjeff * notice unmodified, this list of conditions, and the following 10109864Sjeff * disclaimer. 11109864Sjeff * 2. Redistributions in binary form must reproduce the above copyright 12109864Sjeff * notice, this list of conditions and the following disclaimer in the 13109864Sjeff * documentation and/or other materials provided with the distribution. 14109864Sjeff * 15109864Sjeff * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 16109864Sjeff * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 17109864Sjeff * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 18109864Sjeff * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 19109864Sjeff * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 20109864Sjeff * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 21109864Sjeff * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 22109864Sjeff * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 23109864Sjeff * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 24109864Sjeff * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 25109864Sjeff */ 26109864Sjeff 27171482Sjeff/* 28171482Sjeff * This file implements the ULE scheduler. ULE supports independent CPU 29171482Sjeff * run queues and fine grain locking. It has superior interactive 30171482Sjeff * performance under load even on uni-processor systems. 31171482Sjeff * 32171482Sjeff * etymology: 33172293Sjeff * ULE is the last three letters in schedule. It owes its name to a 34171482Sjeff * generic user created for a scheduling system by Paul Mikesell at 35171482Sjeff * Isilon Systems and a general lack of creativity on the part of the author. 36171482Sjeff */ 37171482Sjeff 38116182Sobrien#include <sys/cdefs.h> 39191645Sjeff__FBSDID("$FreeBSD: head/sys/kern/sched_ule.c 208982 2010-06-10 11:01:17Z ivoras $"); 40116182Sobrien 41147565Speter#include "opt_hwpmc_hooks.h" 42179297Sjb#include "opt_kdtrace.h" 43147565Speter#include "opt_sched.h" 44134649Sscottl 45109864Sjeff#include <sys/param.h> 46109864Sjeff#include <sys/systm.h> 47131929Smarcel#include <sys/kdb.h> 48109864Sjeff#include <sys/kernel.h> 49109864Sjeff#include <sys/ktr.h> 50109864Sjeff#include <sys/lock.h> 51109864Sjeff#include <sys/mutex.h> 52109864Sjeff#include <sys/proc.h> 53112966Sjeff#include <sys/resource.h> 54122038Sjeff#include <sys/resourcevar.h> 55109864Sjeff#include <sys/sched.h> 56109864Sjeff#include <sys/smp.h> 57109864Sjeff#include <sys/sx.h> 58109864Sjeff#include <sys/sysctl.h> 59109864Sjeff#include <sys/sysproto.h> 60139453Sjhb#include <sys/turnstile.h> 61161599Sdavidxu#include <sys/umtx.h> 62109864Sjeff#include <sys/vmmeter.h> 63176735Sjeff#include <sys/cpuset.h> 64184439Sivoras#include <sys/sbuf.h> 65109864Sjeff#ifdef KTRACE 66109864Sjeff#include <sys/uio.h> 67109864Sjeff#include <sys/ktrace.h> 68109864Sjeff#endif 69109864Sjeff 70145256Sjkoshy#ifdef HWPMC_HOOKS 71145256Sjkoshy#include <sys/pmckern.h> 72145256Sjkoshy#endif 73145256Sjkoshy 74179297Sjb#ifdef KDTRACE_HOOKS 75179297Sjb#include <sys/dtrace_bsd.h> 76179297Sjbint dtrace_vtime_active; 77179297Sjbdtrace_vtime_switch_func_t dtrace_vtime_switch_func; 78179297Sjb#endif 79179297Sjb 80109864Sjeff#include <machine/cpu.h> 81121790Sjeff#include <machine/smp.h> 82109864Sjeff 83208165Srrs#if defined(__sparc64__) 84172345Sjeff#error "This architecture is not currently compatible with ULE" 85166190Sjeff#endif 86166190Sjeff 87171482Sjeff#define KTR_ULE 0 88166137Sjeff 89187679Sjeff#define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX))) 90187679Sjeff#define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU))) 91187357Sjeff#define TDQ_LOADNAME_LEN (PCPU_NAME_LEN + sizeof(" load")) 92187357Sjeff 93166137Sjeff/* 94171482Sjeff * Thread scheduler specific section. All fields are protected 95171482Sjeff * by the thread lock. 96146954Sjeff */ 97164936Sjulianstruct td_sched { 98171482Sjeff struct runq *ts_runq; /* Run-queue we're queued on. */ 99171482Sjeff short ts_flags; /* TSF_* flags. */ 100164936Sjulian u_char ts_cpu; /* CPU that we have affinity for. */ 101177009Sjeff int ts_rltick; /* Real last tick, for affinity. */ 102171482Sjeff int ts_slice; /* Ticks of slice remaining. */ 103171482Sjeff u_int ts_slptime; /* Number of ticks we vol. slept */ 104171482Sjeff u_int ts_runtime; /* Number of ticks we were running */ 105164936Sjulian int ts_ltick; /* Last tick that we were running on */ 106199764Sivoras int ts_incrtick; /* Last tick that we incremented on */ 107164936Sjulian int ts_ftick; /* First tick that we were running on */ 108164936Sjulian int ts_ticks; /* Tick count */ 109187357Sjeff#ifdef KTR 110187357Sjeff char ts_name[TS_NAME_LEN]; 111187357Sjeff#endif 112134791Sjulian}; 113164936Sjulian/* flags kept in ts_flags */ 114166108Sjeff#define TSF_BOUND 0x0001 /* Thread can not migrate. */ 115166108Sjeff#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 116121790Sjeff 117164936Sjulianstatic struct td_sched td_sched0; 118109864Sjeff 119176735Sjeff#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 120176735Sjeff#define THREAD_CAN_SCHED(td, cpu) \ 121176735Sjeff CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 122176735Sjeff 123109864Sjeff/* 124165762Sjeff * Cpu percentage computation macros and defines. 125111857Sjeff * 126165762Sjeff * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 127165762Sjeff * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 128165796Sjeff * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 129165762Sjeff * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 130165762Sjeff * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 131165762Sjeff * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 132165762Sjeff */ 133165762Sjeff#define SCHED_TICK_SECS 10 134165762Sjeff#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 135165796Sjeff#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 136165762Sjeff#define SCHED_TICK_SHIFT 10 137165762Sjeff#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 138165830Sjeff#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 139165762Sjeff 140165762Sjeff/* 141165762Sjeff * These macros determine priorities for non-interactive threads. They are 142165762Sjeff * assigned a priority based on their recent cpu utilization as expressed 143165762Sjeff * by the ratio of ticks to the tick total. NHALF priorities at the start 144165762Sjeff * and end of the MIN to MAX timeshare range are only reachable with negative 145165762Sjeff * or positive nice respectively. 146165762Sjeff * 147165762Sjeff * PRI_RANGE: Priority range for utilization dependent priorities. 148116642Sjeff * PRI_NRESV: Number of nice values. 149165762Sjeff * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 150165762Sjeff * PRI_NICE: Determines the part of the priority inherited from nice. 151109864Sjeff */ 152165762Sjeff#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 153121869Sjeff#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 154165762Sjeff#define SCHED_PRI_MIN (PRI_MIN_TIMESHARE + SCHED_PRI_NHALF) 155165762Sjeff#define SCHED_PRI_MAX (PRI_MAX_TIMESHARE - SCHED_PRI_NHALF) 156170787Sjeff#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN) 157165762Sjeff#define SCHED_PRI_TICKS(ts) \ 158165762Sjeff (SCHED_TICK_HZ((ts)) / \ 159165827Sjeff (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 160165762Sjeff#define SCHED_PRI_NICE(nice) (nice) 161109864Sjeff 162109864Sjeff/* 163165762Sjeff * These determine the interactivity of a process. Interactivity differs from 164165762Sjeff * cpu utilization in that it expresses the voluntary time slept vs time ran 165165762Sjeff * while cpu utilization includes all time not running. This more accurately 166165762Sjeff * models the intent of the thread. 167109864Sjeff * 168110645Sjeff * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 169110645Sjeff * before throttling back. 170121868Sjeff * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 171116365Sjeff * INTERACT_MAX: Maximum interactivity value. Smaller is better. 172111857Sjeff * INTERACT_THRESH: Threshhold for placement on the current runq. 173109864Sjeff */ 174165762Sjeff#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 175165762Sjeff#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 176116365Sjeff#define SCHED_INTERACT_MAX (100) 177116365Sjeff#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 178121126Sjeff#define SCHED_INTERACT_THRESH (30) 179111857Sjeff 180109864Sjeff/* 181165762Sjeff * tickincr: Converts a stathz tick into a hz domain scaled by 182165762Sjeff * the shift factor. Without the shift the error rate 183165762Sjeff * due to rounding would be unacceptably high. 184165762Sjeff * realstathz: stathz is sometimes 0 and run off of hz. 185165762Sjeff * sched_slice: Runtime of each thread before rescheduling. 186171482Sjeff * preempt_thresh: Priority threshold for preemption and remote IPIs. 187109864Sjeff */ 188165762Sjeffstatic int sched_interact = SCHED_INTERACT_THRESH; 189165762Sjeffstatic int realstathz; 190165762Sjeffstatic int tickincr; 191177009Sjeffstatic int sched_slice = 1; 192172345Sjeff#ifdef PREEMPTION 193172345Sjeff#ifdef FULL_PREEMPTION 194172345Sjeffstatic int preempt_thresh = PRI_MAX_IDLE; 195172345Sjeff#else 196171482Sjeffstatic int preempt_thresh = PRI_MIN_KERN; 197172345Sjeff#endif 198172345Sjeff#else 199172345Sjeffstatic int preempt_thresh = 0; 200172345Sjeff#endif 201177903Sjeffstatic int static_boost = PRI_MIN_TIMESHARE; 202178277Sjeffstatic int sched_idlespins = 10000; 203178277Sjeffstatic int sched_idlespinthresh = 4; 204109864Sjeff 205109864Sjeff/* 206171482Sjeff * tdq - per processor runqs and statistics. All fields are protected by the 207171482Sjeff * tdq_lock. The load and lowpri may be accessed without to avoid excess 208171482Sjeff * locking in sched_pickcpu(); 209109864Sjeff */ 210164936Sjulianstruct tdq { 211177009Sjeff /* Ordered to improve efficiency of cpu_search() and switch(). */ 212177009Sjeff struct mtx tdq_lock; /* run queue lock. */ 213176735Sjeff struct cpu_group *tdq_cg; /* Pointer to cpu topology. */ 214178277Sjeff volatile int tdq_load; /* Aggregate load. */ 215176735Sjeff int tdq_sysload; /* For loadavg, !ITHD load. */ 216177009Sjeff int tdq_transferable; /* Transferable thread count. */ 217178277Sjeff short tdq_switchcnt; /* Switches this tick. */ 218178277Sjeff short tdq_oldswitchcnt; /* Switches last tick. */ 219177009Sjeff u_char tdq_lowpri; /* Lowest priority thread. */ 220177009Sjeff u_char tdq_ipipending; /* IPI pending. */ 221166557Sjeff u_char tdq_idx; /* Current insert index. */ 222166557Sjeff u_char tdq_ridx; /* Current removal index. */ 223177009Sjeff struct runq tdq_realtime; /* real-time run queue. */ 224177009Sjeff struct runq tdq_timeshare; /* timeshare run queue. */ 225177009Sjeff struct runq tdq_idle; /* Queue of IDLE threads. */ 226187357Sjeff char tdq_name[TDQ_NAME_LEN]; 227187357Sjeff#ifdef KTR 228187357Sjeff char tdq_loadname[TDQ_LOADNAME_LEN]; 229187357Sjeff#endif 230171482Sjeff} __aligned(64); 231109864Sjeff 232178277Sjeff/* Idle thread states and config. */ 233178277Sjeff#define TDQ_RUNNING 1 234178277Sjeff#define TDQ_IDLE 2 235166108Sjeff 236123433Sjeff#ifdef SMP 237184439Sivorasstruct cpu_group *cpu_top; /* CPU topology */ 238123433Sjeff 239176735Sjeff#define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000)) 240176735Sjeff#define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity)) 241166108Sjeff 242123433Sjeff/* 243166108Sjeff * Run-time tunables. 244166108Sjeff */ 245171506Sjeffstatic int rebalance = 1; 246172409Sjeffstatic int balance_interval = 128; /* Default set in sched_initticks(). */ 247166108Sjeffstatic int affinity; 248172409Sjeffstatic int steal_htt = 1; 249171506Sjeffstatic int steal_idle = 1; 250171506Sjeffstatic int steal_thresh = 2; 251166108Sjeff 252166108Sjeff/* 253165620Sjeff * One thread queue per processor. 254109864Sjeff */ 255164936Sjulianstatic struct tdq tdq_cpu[MAXCPU]; 256172409Sjeffstatic struct tdq *balance_tdq; 257172409Sjeffstatic int balance_ticks; 258129982Sjeff 259164936Sjulian#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 260164936Sjulian#define TDQ_CPU(x) (&tdq_cpu[(x)]) 261171713Sjeff#define TDQ_ID(x) ((int)((x) - tdq_cpu)) 262123433Sjeff#else /* !SMP */ 263164936Sjulianstatic struct tdq tdq_cpu; 264129982Sjeff 265170315Sjeff#define TDQ_ID(x) (0) 266164936Sjulian#define TDQ_SELF() (&tdq_cpu) 267164936Sjulian#define TDQ_CPU(x) (&tdq_cpu) 268110028Sjeff#endif 269109864Sjeff 270171482Sjeff#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 271171482Sjeff#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 272171482Sjeff#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 273171482Sjeff#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 274176735Sjeff#define TDQ_LOCKPTR(t) (&(t)->tdq_lock) 275171482Sjeff 276163709Sjbstatic void sched_priority(struct thread *); 277146954Sjeffstatic void sched_thread_priority(struct thread *, u_char); 278163709Sjbstatic int sched_interact_score(struct thread *); 279163709Sjbstatic void sched_interact_update(struct thread *); 280163709Sjbstatic void sched_interact_fork(struct thread *); 281164936Sjulianstatic void sched_pctcpu_update(struct td_sched *); 282109864Sjeff 283110267Sjeff/* Operations on per processor queues */ 284177435Sjeffstatic struct thread *tdq_choose(struct tdq *); 285164936Sjulianstatic void tdq_setup(struct tdq *); 286177435Sjeffstatic void tdq_load_add(struct tdq *, struct thread *); 287177435Sjeffstatic void tdq_load_rem(struct tdq *, struct thread *); 288177435Sjeffstatic __inline void tdq_runq_add(struct tdq *, struct thread *, int); 289177435Sjeffstatic __inline void tdq_runq_rem(struct tdq *, struct thread *); 290177005Sjeffstatic inline int sched_shouldpreempt(int, int, int); 291164936Sjulianvoid tdq_print(int cpu); 292165762Sjeffstatic void runq_print(struct runq *rq); 293171482Sjeffstatic void tdq_add(struct tdq *, struct thread *, int); 294110267Sjeff#ifdef SMP 295176735Sjeffstatic int tdq_move(struct tdq *, struct tdq *); 296171482Sjeffstatic int tdq_idled(struct tdq *); 297177435Sjeffstatic void tdq_notify(struct tdq *, struct thread *); 298177435Sjeffstatic struct thread *tdq_steal(struct tdq *, int); 299177435Sjeffstatic struct thread *runq_steal(struct runq *, int); 300177435Sjeffstatic int sched_pickcpu(struct thread *, int); 301172409Sjeffstatic void sched_balance(void); 302176735Sjeffstatic int sched_balance_pair(struct tdq *, struct tdq *); 303177435Sjeffstatic inline struct tdq *sched_setcpu(struct thread *, int, int); 304171482Sjeffstatic inline void thread_unblock_switch(struct thread *, struct mtx *); 305171713Sjeffstatic struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 306184439Sivorasstatic int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS); 307184439Sivorasstatic int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, 308184439Sivoras struct cpu_group *cg, int indent); 309121790Sjeff#endif 310110028Sjeff 311165762Sjeffstatic void sched_setup(void *dummy); 312177253SrwatsonSYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 313165762Sjeff 314165762Sjeffstatic void sched_initticks(void *dummy); 315177253SrwatsonSYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 316177253Srwatson NULL); 317165762Sjeff 318171482Sjeff/* 319171482Sjeff * Print the threads waiting on a run-queue. 320171482Sjeff */ 321165762Sjeffstatic void 322165762Sjeffrunq_print(struct runq *rq) 323165762Sjeff{ 324165762Sjeff struct rqhead *rqh; 325177435Sjeff struct thread *td; 326165762Sjeff int pri; 327165762Sjeff int j; 328165762Sjeff int i; 329165762Sjeff 330165762Sjeff for (i = 0; i < RQB_LEN; i++) { 331165762Sjeff printf("\t\trunq bits %d 0x%zx\n", 332165762Sjeff i, rq->rq_status.rqb_bits[i]); 333165762Sjeff for (j = 0; j < RQB_BPW; j++) 334165762Sjeff if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 335165762Sjeff pri = j + (i << RQB_L2BPW); 336165762Sjeff rqh = &rq->rq_queues[pri]; 337177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) { 338165762Sjeff printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 339177435Sjeff td, td->td_name, td->td_priority, 340177435Sjeff td->td_rqindex, pri); 341165762Sjeff } 342165762Sjeff } 343165762Sjeff } 344165762Sjeff} 345165762Sjeff 346171482Sjeff/* 347171482Sjeff * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 348171482Sjeff */ 349113357Sjeffvoid 350164936Sjuliantdq_print(int cpu) 351110267Sjeff{ 352164936Sjulian struct tdq *tdq; 353112994Sjeff 354164936Sjulian tdq = TDQ_CPU(cpu); 355112994Sjeff 356171713Sjeff printf("tdq %d:\n", TDQ_ID(tdq)); 357176735Sjeff printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 358176735Sjeff printf("\tLock name: %s\n", tdq->tdq_name); 359165620Sjeff printf("\tload: %d\n", tdq->tdq_load); 360178277Sjeff printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt); 361178277Sjeff printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt); 362171482Sjeff printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 363165766Sjeff printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 364178277Sjeff printf("\tload transferable: %d\n", tdq->tdq_transferable); 365178277Sjeff printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 366165762Sjeff printf("\trealtime runq:\n"); 367165762Sjeff runq_print(&tdq->tdq_realtime); 368165762Sjeff printf("\ttimeshare runq:\n"); 369165762Sjeff runq_print(&tdq->tdq_timeshare); 370165762Sjeff printf("\tidle runq:\n"); 371165762Sjeff runq_print(&tdq->tdq_idle); 372113357Sjeff} 373112994Sjeff 374177005Sjeffstatic inline int 375177005Sjeffsched_shouldpreempt(int pri, int cpri, int remote) 376177005Sjeff{ 377177005Sjeff /* 378177005Sjeff * If the new priority is not better than the current priority there is 379177005Sjeff * nothing to do. 380177005Sjeff */ 381177005Sjeff if (pri >= cpri) 382177005Sjeff return (0); 383177005Sjeff /* 384177005Sjeff * Always preempt idle. 385177005Sjeff */ 386177005Sjeff if (cpri >= PRI_MIN_IDLE) 387177005Sjeff return (1); 388177005Sjeff /* 389177005Sjeff * If preemption is disabled don't preempt others. 390177005Sjeff */ 391177005Sjeff if (preempt_thresh == 0) 392177005Sjeff return (0); 393177005Sjeff /* 394177005Sjeff * Preempt if we exceed the threshold. 395177005Sjeff */ 396177005Sjeff if (pri <= preempt_thresh) 397177005Sjeff return (1); 398177005Sjeff /* 399177005Sjeff * If we're realtime or better and there is timeshare or worse running 400177005Sjeff * preempt only remote processors. 401177005Sjeff */ 402177005Sjeff if (remote && pri <= PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME) 403177005Sjeff return (1); 404177005Sjeff return (0); 405177005Sjeff} 406177005Sjeff 407171482Sjeff#define TS_RQ_PPQ (((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS) 408171482Sjeff/* 409171482Sjeff * Add a thread to the actual run-queue. Keeps transferable counts up to 410171482Sjeff * date with what is actually on the run-queue. Selects the correct 411171482Sjeff * queue position for timeshare threads. 412171482Sjeff */ 413122744Sjeffstatic __inline void 414177435Sjefftdq_runq_add(struct tdq *tdq, struct thread *td, int flags) 415122744Sjeff{ 416177435Sjeff struct td_sched *ts; 417177042Sjeff u_char pri; 418177042Sjeff 419171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 420177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 421177009Sjeff 422177435Sjeff pri = td->td_priority; 423177435Sjeff ts = td->td_sched; 424177435Sjeff TD_SET_RUNQ(td); 425177435Sjeff if (THREAD_CAN_MIGRATE(td)) { 426165620Sjeff tdq->tdq_transferable++; 427164936Sjulian ts->ts_flags |= TSF_XFERABLE; 428123433Sjeff } 429177042Sjeff if (pri <= PRI_MAX_REALTIME) { 430177042Sjeff ts->ts_runq = &tdq->tdq_realtime; 431177042Sjeff } else if (pri <= PRI_MAX_TIMESHARE) { 432177042Sjeff ts->ts_runq = &tdq->tdq_timeshare; 433165762Sjeff KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE, 434165762Sjeff ("Invalid priority %d on timeshare runq", pri)); 435165762Sjeff /* 436165762Sjeff * This queue contains only priorities between MIN and MAX 437165762Sjeff * realtime. Use the whole queue to represent these values. 438165762Sjeff */ 439171713Sjeff if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 440165762Sjeff pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ; 441165762Sjeff pri = (pri + tdq->tdq_idx) % RQ_NQS; 442165766Sjeff /* 443165766Sjeff * This effectively shortens the queue by one so we 444165766Sjeff * can have a one slot difference between idx and 445165766Sjeff * ridx while we wait for threads to drain. 446165766Sjeff */ 447165766Sjeff if (tdq->tdq_ridx != tdq->tdq_idx && 448165766Sjeff pri == tdq->tdq_ridx) 449167664Sjeff pri = (unsigned char)(pri - 1) % RQ_NQS; 450165762Sjeff } else 451165766Sjeff pri = tdq->tdq_ridx; 452177435Sjeff runq_add_pri(ts->ts_runq, td, pri, flags); 453177042Sjeff return; 454165762Sjeff } else 455177009Sjeff ts->ts_runq = &tdq->tdq_idle; 456177435Sjeff runq_add(ts->ts_runq, td, flags); 457177009Sjeff} 458177009Sjeff 459171482Sjeff/* 460171482Sjeff * Remove a thread from a run-queue. This typically happens when a thread 461171482Sjeff * is selected to run. Running threads are not on the queue and the 462171482Sjeff * transferable count does not reflect them. 463171482Sjeff */ 464122744Sjeffstatic __inline void 465177435Sjefftdq_runq_rem(struct tdq *tdq, struct thread *td) 466122744Sjeff{ 467177435Sjeff struct td_sched *ts; 468177435Sjeff 469177435Sjeff ts = td->td_sched; 470171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 471171482Sjeff KASSERT(ts->ts_runq != NULL, 472177435Sjeff ("tdq_runq_remove: thread %p null ts_runq", td)); 473164936Sjulian if (ts->ts_flags & TSF_XFERABLE) { 474165620Sjeff tdq->tdq_transferable--; 475164936Sjulian ts->ts_flags &= ~TSF_XFERABLE; 476123433Sjeff } 477165766Sjeff if (ts->ts_runq == &tdq->tdq_timeshare) { 478165766Sjeff if (tdq->tdq_idx != tdq->tdq_ridx) 479177435Sjeff runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx); 480165766Sjeff else 481177435Sjeff runq_remove_idx(ts->ts_runq, td, NULL); 482165766Sjeff } else 483177435Sjeff runq_remove(ts->ts_runq, td); 484122744Sjeff} 485122744Sjeff 486171482Sjeff/* 487171482Sjeff * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 488171482Sjeff * for this thread to the referenced thread queue. 489171482Sjeff */ 490113357Sjeffstatic void 491177435Sjefftdq_load_add(struct tdq *tdq, struct thread *td) 492113357Sjeff{ 493171482Sjeff 494171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 495177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 496177902Sjeff 497165620Sjeff tdq->tdq_load++; 498198854Sattilio if ((td->td_flags & TDF_NOLOAD) == 0) 499177902Sjeff tdq->tdq_sysload++; 500187357Sjeff KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 501110267Sjeff} 502113357Sjeff 503171482Sjeff/* 504171482Sjeff * Remove the load from a thread that is transitioning to a sleep state or 505171482Sjeff * exiting. 506171482Sjeff */ 507112994Sjeffstatic void 508177435Sjefftdq_load_rem(struct tdq *tdq, struct thread *td) 509110267Sjeff{ 510171482Sjeff 511177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 512171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 513171482Sjeff KASSERT(tdq->tdq_load != 0, 514171713Sjeff ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 515177902Sjeff 516165620Sjeff tdq->tdq_load--; 517198854Sattilio if ((td->td_flags & TDF_NOLOAD) == 0) 518177902Sjeff tdq->tdq_sysload--; 519187357Sjeff KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 520110267Sjeff} 521110267Sjeff 522176735Sjeff/* 523176735Sjeff * Set lowpri to its exact value by searching the run-queue and 524176735Sjeff * evaluating curthread. curthread may be passed as an optimization. 525176735Sjeff */ 526176735Sjeffstatic void 527176735Sjefftdq_setlowpri(struct tdq *tdq, struct thread *ctd) 528176735Sjeff{ 529176735Sjeff struct thread *td; 530176735Sjeff 531176735Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 532176735Sjeff if (ctd == NULL) 533176735Sjeff ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 534177435Sjeff td = tdq_choose(tdq); 535177435Sjeff if (td == NULL || td->td_priority > ctd->td_priority) 536176735Sjeff tdq->tdq_lowpri = ctd->td_priority; 537176735Sjeff else 538176735Sjeff tdq->tdq_lowpri = td->td_priority; 539176735Sjeff} 540176735Sjeff 541113357Sjeff#ifdef SMP 542176735Sjeffstruct cpu_search { 543194779Sjeff cpuset_t cs_mask; 544176735Sjeff u_int cs_load; 545176735Sjeff u_int cs_cpu; 546176735Sjeff int cs_limit; /* Min priority for low min load for high. */ 547176735Sjeff}; 548176735Sjeff 549176735Sjeff#define CPU_SEARCH_LOWEST 0x1 550176735Sjeff#define CPU_SEARCH_HIGHEST 0x2 551176735Sjeff#define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 552176735Sjeff 553194779Sjeff#define CPUSET_FOREACH(cpu, mask) \ 554194779Sjeff for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \ 555176735Sjeff if ((mask) & 1 << (cpu)) 556176735Sjeff 557177169Sjhbstatic __inline int cpu_search(struct cpu_group *cg, struct cpu_search *low, 558176735Sjeff struct cpu_search *high, const int match); 559176735Sjeffint cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low); 560176735Sjeffint cpu_search_highest(struct cpu_group *cg, struct cpu_search *high); 561176735Sjeffint cpu_search_both(struct cpu_group *cg, struct cpu_search *low, 562176735Sjeff struct cpu_search *high); 563176735Sjeff 564116069Sjeff/* 565176735Sjeff * This routine compares according to the match argument and should be 566176735Sjeff * reduced in actual instantiations via constant propagation and dead code 567176735Sjeff * elimination. 568176735Sjeff */ 569176735Sjeffstatic __inline int 570176735Sjeffcpu_compare(int cpu, struct cpu_search *low, struct cpu_search *high, 571176735Sjeff const int match) 572176735Sjeff{ 573176735Sjeff struct tdq *tdq; 574176735Sjeff 575176735Sjeff tdq = TDQ_CPU(cpu); 576176735Sjeff if (match & CPU_SEARCH_LOWEST) 577194779Sjeff if (CPU_ISSET(cpu, &low->cs_mask) && 578176735Sjeff tdq->tdq_load < low->cs_load && 579176735Sjeff tdq->tdq_lowpri > low->cs_limit) { 580176735Sjeff low->cs_cpu = cpu; 581176735Sjeff low->cs_load = tdq->tdq_load; 582176735Sjeff } 583176735Sjeff if (match & CPU_SEARCH_HIGHEST) 584194779Sjeff if (CPU_ISSET(cpu, &high->cs_mask) && 585176735Sjeff tdq->tdq_load >= high->cs_limit && 586176735Sjeff tdq->tdq_load > high->cs_load && 587176735Sjeff tdq->tdq_transferable) { 588176735Sjeff high->cs_cpu = cpu; 589176735Sjeff high->cs_load = tdq->tdq_load; 590176735Sjeff } 591176735Sjeff return (tdq->tdq_load); 592176735Sjeff} 593176735Sjeff 594176735Sjeff/* 595176735Sjeff * Search the tree of cpu_groups for the lowest or highest loaded cpu 596176735Sjeff * according to the match argument. This routine actually compares the 597176735Sjeff * load on all paths through the tree and finds the least loaded cpu on 598176735Sjeff * the least loaded path, which may differ from the least loaded cpu in 599176735Sjeff * the system. This balances work among caches and busses. 600116069Sjeff * 601176735Sjeff * This inline is instantiated in three forms below using constants for the 602176735Sjeff * match argument. It is reduced to the minimum set for each case. It is 603176735Sjeff * also recursive to the depth of the tree. 604116069Sjeff */ 605177169Sjhbstatic __inline int 606176735Sjeffcpu_search(struct cpu_group *cg, struct cpu_search *low, 607176735Sjeff struct cpu_search *high, const int match) 608176735Sjeff{ 609176735Sjeff int total; 610176735Sjeff 611176735Sjeff total = 0; 612176735Sjeff if (cg->cg_children) { 613176735Sjeff struct cpu_search lgroup; 614176735Sjeff struct cpu_search hgroup; 615176735Sjeff struct cpu_group *child; 616176735Sjeff u_int lload; 617176735Sjeff int hload; 618176735Sjeff int load; 619176735Sjeff int i; 620176735Sjeff 621176735Sjeff lload = -1; 622176735Sjeff hload = -1; 623176735Sjeff for (i = 0; i < cg->cg_children; i++) { 624176735Sjeff child = &cg->cg_child[i]; 625176735Sjeff if (match & CPU_SEARCH_LOWEST) { 626176735Sjeff lgroup = *low; 627176735Sjeff lgroup.cs_load = -1; 628176735Sjeff } 629176735Sjeff if (match & CPU_SEARCH_HIGHEST) { 630176735Sjeff hgroup = *high; 631176735Sjeff lgroup.cs_load = 0; 632176735Sjeff } 633176735Sjeff switch (match) { 634176735Sjeff case CPU_SEARCH_LOWEST: 635176735Sjeff load = cpu_search_lowest(child, &lgroup); 636176735Sjeff break; 637176735Sjeff case CPU_SEARCH_HIGHEST: 638176735Sjeff load = cpu_search_highest(child, &hgroup); 639176735Sjeff break; 640176735Sjeff case CPU_SEARCH_BOTH: 641176735Sjeff load = cpu_search_both(child, &lgroup, &hgroup); 642176735Sjeff break; 643176735Sjeff } 644176735Sjeff total += load; 645176735Sjeff if (match & CPU_SEARCH_LOWEST) 646176735Sjeff if (load < lload || low->cs_cpu == -1) { 647176735Sjeff *low = lgroup; 648176735Sjeff lload = load; 649176735Sjeff } 650176735Sjeff if (match & CPU_SEARCH_HIGHEST) 651176735Sjeff if (load > hload || high->cs_cpu == -1) { 652176735Sjeff hload = load; 653176735Sjeff *high = hgroup; 654176735Sjeff } 655176735Sjeff } 656176735Sjeff } else { 657176735Sjeff int cpu; 658176735Sjeff 659194779Sjeff CPUSET_FOREACH(cpu, cg->cg_mask) 660176735Sjeff total += cpu_compare(cpu, low, high, match); 661176735Sjeff } 662176735Sjeff return (total); 663176735Sjeff} 664176735Sjeff 665176735Sjeff/* 666176735Sjeff * cpu_search instantiations must pass constants to maintain the inline 667176735Sjeff * optimization. 668176735Sjeff */ 669176735Sjeffint 670176735Sjeffcpu_search_lowest(struct cpu_group *cg, struct cpu_search *low) 671176735Sjeff{ 672176735Sjeff return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 673176735Sjeff} 674176735Sjeff 675176735Sjeffint 676176735Sjeffcpu_search_highest(struct cpu_group *cg, struct cpu_search *high) 677176735Sjeff{ 678176735Sjeff return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 679176735Sjeff} 680176735Sjeff 681176735Sjeffint 682176735Sjeffcpu_search_both(struct cpu_group *cg, struct cpu_search *low, 683176735Sjeff struct cpu_search *high) 684176735Sjeff{ 685176735Sjeff return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 686176735Sjeff} 687176735Sjeff 688176735Sjeff/* 689176735Sjeff * Find the cpu with the least load via the least loaded path that has a 690176735Sjeff * lowpri greater than pri pri. A pri of -1 indicates any priority is 691176735Sjeff * acceptable. 692176735Sjeff */ 693176735Sjeffstatic inline int 694194779Sjeffsched_lowest(struct cpu_group *cg, cpuset_t mask, int pri) 695176735Sjeff{ 696176735Sjeff struct cpu_search low; 697176735Sjeff 698176735Sjeff low.cs_cpu = -1; 699176735Sjeff low.cs_load = -1; 700176735Sjeff low.cs_mask = mask; 701176735Sjeff low.cs_limit = pri; 702176735Sjeff cpu_search_lowest(cg, &low); 703176735Sjeff return low.cs_cpu; 704176735Sjeff} 705176735Sjeff 706176735Sjeff/* 707176735Sjeff * Find the cpu with the highest load via the highest loaded path. 708176735Sjeff */ 709176735Sjeffstatic inline int 710194779Sjeffsched_highest(struct cpu_group *cg, cpuset_t mask, int minload) 711176735Sjeff{ 712176735Sjeff struct cpu_search high; 713176735Sjeff 714176735Sjeff high.cs_cpu = -1; 715176735Sjeff high.cs_load = 0; 716176735Sjeff high.cs_mask = mask; 717176735Sjeff high.cs_limit = minload; 718176735Sjeff cpu_search_highest(cg, &high); 719176735Sjeff return high.cs_cpu; 720176735Sjeff} 721176735Sjeff 722176735Sjeff/* 723176735Sjeff * Simultaneously find the highest and lowest loaded cpu reachable via 724176735Sjeff * cg. 725176735Sjeff */ 726176735Sjeffstatic inline void 727194779Sjeffsched_both(struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu) 728176735Sjeff{ 729176735Sjeff struct cpu_search high; 730176735Sjeff struct cpu_search low; 731176735Sjeff 732176735Sjeff low.cs_cpu = -1; 733176735Sjeff low.cs_limit = -1; 734176735Sjeff low.cs_load = -1; 735176735Sjeff low.cs_mask = mask; 736176735Sjeff high.cs_load = 0; 737176735Sjeff high.cs_cpu = -1; 738176735Sjeff high.cs_limit = -1; 739176735Sjeff high.cs_mask = mask; 740176735Sjeff cpu_search_both(cg, &low, &high); 741176735Sjeff *lowcpu = low.cs_cpu; 742176735Sjeff *highcpu = high.cs_cpu; 743176735Sjeff return; 744176735Sjeff} 745176735Sjeff 746121790Sjeffstatic void 747176735Sjeffsched_balance_group(struct cpu_group *cg) 748116069Sjeff{ 749194779Sjeff cpuset_t mask; 750176735Sjeff int high; 751176735Sjeff int low; 752123487Sjeff int i; 753123487Sjeff 754194779Sjeff CPU_FILL(&mask); 755176735Sjeff for (;;) { 756176735Sjeff sched_both(cg, mask, &low, &high); 757176735Sjeff if (low == high || low == -1 || high == -1) 758176735Sjeff break; 759176735Sjeff if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) 760176735Sjeff break; 761123487Sjeff /* 762176735Sjeff * If we failed to move any threads determine which cpu 763176735Sjeff * to kick out of the set and try again. 764176735Sjeff */ 765176735Sjeff if (TDQ_CPU(high)->tdq_transferable == 0) 766194779Sjeff CPU_CLR(high, &mask); 767176735Sjeff else 768194779Sjeff CPU_CLR(low, &mask); 769123487Sjeff } 770176735Sjeff 771176735Sjeff for (i = 0; i < cg->cg_children; i++) 772176735Sjeff sched_balance_group(&cg->cg_child[i]); 773123487Sjeff} 774123487Sjeff 775123487Sjeffstatic void 776201148Sedsched_balance(void) 777123487Sjeff{ 778172409Sjeff struct tdq *tdq; 779123487Sjeff 780172409Sjeff /* 781172409Sjeff * Select a random time between .5 * balance_interval and 782172409Sjeff * 1.5 * balance_interval. 783172409Sjeff */ 784176735Sjeff balance_ticks = max(balance_interval / 2, 1); 785176735Sjeff balance_ticks += random() % balance_interval; 786171482Sjeff if (smp_started == 0 || rebalance == 0) 787171482Sjeff return; 788172409Sjeff tdq = TDQ_SELF(); 789172409Sjeff TDQ_UNLOCK(tdq); 790176735Sjeff sched_balance_group(cpu_top); 791172409Sjeff TDQ_LOCK(tdq); 792123487Sjeff} 793123487Sjeff 794171482Sjeff/* 795171482Sjeff * Lock two thread queues using their address to maintain lock order. 796171482Sjeff */ 797123487Sjeffstatic void 798171482Sjefftdq_lock_pair(struct tdq *one, struct tdq *two) 799171482Sjeff{ 800171482Sjeff if (one < two) { 801171482Sjeff TDQ_LOCK(one); 802171482Sjeff TDQ_LOCK_FLAGS(two, MTX_DUPOK); 803171482Sjeff } else { 804171482Sjeff TDQ_LOCK(two); 805171482Sjeff TDQ_LOCK_FLAGS(one, MTX_DUPOK); 806171482Sjeff } 807171482Sjeff} 808171482Sjeff 809171482Sjeff/* 810172409Sjeff * Unlock two thread queues. Order is not important here. 811172409Sjeff */ 812172409Sjeffstatic void 813172409Sjefftdq_unlock_pair(struct tdq *one, struct tdq *two) 814172409Sjeff{ 815172409Sjeff TDQ_UNLOCK(one); 816172409Sjeff TDQ_UNLOCK(two); 817172409Sjeff} 818172409Sjeff 819172409Sjeff/* 820171482Sjeff * Transfer load between two imbalanced thread queues. 821171482Sjeff */ 822176735Sjeffstatic int 823164936Sjuliansched_balance_pair(struct tdq *high, struct tdq *low) 824123487Sjeff{ 825123433Sjeff int transferable; 826116069Sjeff int high_load; 827116069Sjeff int low_load; 828176735Sjeff int moved; 829116069Sjeff int move; 830116069Sjeff int diff; 831116069Sjeff int i; 832116069Sjeff 833171482Sjeff tdq_lock_pair(high, low); 834176735Sjeff transferable = high->tdq_transferable; 835176735Sjeff high_load = high->tdq_load; 836176735Sjeff low_load = low->tdq_load; 837176735Sjeff moved = 0; 838116069Sjeff /* 839122744Sjeff * Determine what the imbalance is and then adjust that to how many 840165620Sjeff * threads we actually have to give up (transferable). 841122744Sjeff */ 842171482Sjeff if (transferable != 0) { 843171482Sjeff diff = high_load - low_load; 844171482Sjeff move = diff / 2; 845171482Sjeff if (diff & 0x1) 846171482Sjeff move++; 847171482Sjeff move = min(move, transferable); 848171482Sjeff for (i = 0; i < move; i++) 849176735Sjeff moved += tdq_move(high, low); 850172293Sjeff /* 851172293Sjeff * IPI the target cpu to force it to reschedule with the new 852172293Sjeff * workload. 853172293Sjeff */ 854172293Sjeff ipi_selected(1 << TDQ_ID(low), IPI_PREEMPT); 855171482Sjeff } 856172409Sjeff tdq_unlock_pair(high, low); 857176735Sjeff return (moved); 858116069Sjeff} 859116069Sjeff 860171482Sjeff/* 861171482Sjeff * Move a thread from one thread queue to another. 862171482Sjeff */ 863176735Sjeffstatic int 864171482Sjefftdq_move(struct tdq *from, struct tdq *to) 865116069Sjeff{ 866171482Sjeff struct td_sched *ts; 867171482Sjeff struct thread *td; 868164936Sjulian struct tdq *tdq; 869171482Sjeff int cpu; 870116069Sjeff 871172409Sjeff TDQ_LOCK_ASSERT(from, MA_OWNED); 872172409Sjeff TDQ_LOCK_ASSERT(to, MA_OWNED); 873172409Sjeff 874164936Sjulian tdq = from; 875171482Sjeff cpu = TDQ_ID(to); 876177435Sjeff td = tdq_steal(tdq, cpu); 877177435Sjeff if (td == NULL) 878176735Sjeff return (0); 879177435Sjeff ts = td->td_sched; 880171482Sjeff /* 881171482Sjeff * Although the run queue is locked the thread may be blocked. Lock 882172409Sjeff * it to clear this and acquire the run-queue lock. 883171482Sjeff */ 884171482Sjeff thread_lock(td); 885172409Sjeff /* Drop recursive lock on from acquired via thread_lock(). */ 886171482Sjeff TDQ_UNLOCK(from); 887171482Sjeff sched_rem(td); 888166108Sjeff ts->ts_cpu = cpu; 889171482Sjeff td->td_lock = TDQ_LOCKPTR(to); 890171482Sjeff tdq_add(to, td, SRQ_YIELDING); 891176735Sjeff return (1); 892116069Sjeff} 893110267Sjeff 894171482Sjeff/* 895171482Sjeff * This tdq has idled. Try to steal a thread from another cpu and switch 896171482Sjeff * to it. 897171482Sjeff */ 898123433Sjeffstatic int 899164936Sjuliantdq_idled(struct tdq *tdq) 900121790Sjeff{ 901176735Sjeff struct cpu_group *cg; 902164936Sjulian struct tdq *steal; 903194779Sjeff cpuset_t mask; 904176735Sjeff int thresh; 905171482Sjeff int cpu; 906123433Sjeff 907172484Sjeff if (smp_started == 0 || steal_idle == 0) 908172484Sjeff return (1); 909194779Sjeff CPU_FILL(&mask); 910194779Sjeff CPU_CLR(PCPU_GET(cpuid), &mask); 911176735Sjeff /* We don't want to be preempted while we're iterating. */ 912171482Sjeff spinlock_enter(); 913176735Sjeff for (cg = tdq->tdq_cg; cg != NULL; ) { 914191643Sjeff if ((cg->cg_flags & CG_FLAG_THREAD) == 0) 915176735Sjeff thresh = steal_thresh; 916176735Sjeff else 917176735Sjeff thresh = 1; 918176735Sjeff cpu = sched_highest(cg, mask, thresh); 919176735Sjeff if (cpu == -1) { 920176735Sjeff cg = cg->cg_parent; 921176735Sjeff continue; 922166108Sjeff } 923176735Sjeff steal = TDQ_CPU(cpu); 924194779Sjeff CPU_CLR(cpu, &mask); 925176735Sjeff tdq_lock_pair(tdq, steal); 926176735Sjeff if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 927176735Sjeff tdq_unlock_pair(tdq, steal); 928176735Sjeff continue; 929171482Sjeff } 930176735Sjeff /* 931176735Sjeff * If a thread was added while interrupts were disabled don't 932176735Sjeff * steal one here. If we fail to acquire one due to affinity 933176735Sjeff * restrictions loop again with this cpu removed from the 934176735Sjeff * set. 935176735Sjeff */ 936176735Sjeff if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 937176735Sjeff tdq_unlock_pair(tdq, steal); 938176735Sjeff continue; 939176735Sjeff } 940176735Sjeff spinlock_exit(); 941176735Sjeff TDQ_UNLOCK(steal); 942178272Sjeff mi_switch(SW_VOL | SWT_IDLE, NULL); 943176735Sjeff thread_unlock(curthread); 944176735Sjeff 945176735Sjeff return (0); 946123433Sjeff } 947171482Sjeff spinlock_exit(); 948123433Sjeff return (1); 949121790Sjeff} 950121790Sjeff 951171482Sjeff/* 952171482Sjeff * Notify a remote cpu of new work. Sends an IPI if criteria are met. 953171482Sjeff */ 954121790Sjeffstatic void 955177435Sjefftdq_notify(struct tdq *tdq, struct thread *td) 956121790Sjeff{ 957185047Sjhb struct thread *ctd; 958166247Sjeff int pri; 959166108Sjeff int cpu; 960121790Sjeff 961177005Sjeff if (tdq->tdq_ipipending) 962177005Sjeff return; 963177435Sjeff cpu = td->td_sched->ts_cpu; 964177435Sjeff pri = td->td_priority; 965185047Sjhb ctd = pcpu_find(cpu)->pc_curthread; 966185047Sjhb if (!sched_shouldpreempt(pri, ctd->td_priority, 1)) 967166137Sjeff return; 968185047Sjhb if (TD_IS_IDLETHREAD(ctd)) { 969178277Sjeff /* 970178471Sjeff * If the MD code has an idle wakeup routine try that before 971178471Sjeff * falling back to IPI. 972178471Sjeff */ 973178471Sjeff if (cpu_idle_wakeup(cpu)) 974178471Sjeff return; 975178277Sjeff } 976177005Sjeff tdq->tdq_ipipending = 1; 977171482Sjeff ipi_selected(1 << cpu, IPI_PREEMPT); 978121790Sjeff} 979121790Sjeff 980171482Sjeff/* 981171482Sjeff * Steals load from a timeshare queue. Honors the rotating queue head 982171482Sjeff * index. 983171482Sjeff */ 984177435Sjeffstatic struct thread * 985176735Sjeffrunq_steal_from(struct runq *rq, int cpu, u_char start) 986171482Sjeff{ 987171482Sjeff struct rqbits *rqb; 988171482Sjeff struct rqhead *rqh; 989177435Sjeff struct thread *td; 990171482Sjeff int first; 991171482Sjeff int bit; 992171482Sjeff int pri; 993171482Sjeff int i; 994171482Sjeff 995171482Sjeff rqb = &rq->rq_status; 996171482Sjeff bit = start & (RQB_BPW -1); 997171482Sjeff pri = 0; 998171482Sjeff first = 0; 999171482Sjeffagain: 1000171482Sjeff for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 1001171482Sjeff if (rqb->rqb_bits[i] == 0) 1002171482Sjeff continue; 1003171482Sjeff if (bit != 0) { 1004171482Sjeff for (pri = bit; pri < RQB_BPW; pri++) 1005171482Sjeff if (rqb->rqb_bits[i] & (1ul << pri)) 1006171482Sjeff break; 1007171482Sjeff if (pri >= RQB_BPW) 1008171482Sjeff continue; 1009171482Sjeff } else 1010171482Sjeff pri = RQB_FFS(rqb->rqb_bits[i]); 1011171482Sjeff pri += (i << RQB_L2BPW); 1012171482Sjeff rqh = &rq->rq_queues[pri]; 1013177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) { 1014177435Sjeff if (first && THREAD_CAN_MIGRATE(td) && 1015177435Sjeff THREAD_CAN_SCHED(td, cpu)) 1016177435Sjeff return (td); 1017171482Sjeff first = 1; 1018171482Sjeff } 1019171482Sjeff } 1020171482Sjeff if (start != 0) { 1021171482Sjeff start = 0; 1022171482Sjeff goto again; 1023171482Sjeff } 1024171482Sjeff 1025171482Sjeff return (NULL); 1026171482Sjeff} 1027171482Sjeff 1028171482Sjeff/* 1029171482Sjeff * Steals load from a standard linear queue. 1030171482Sjeff */ 1031177435Sjeffstatic struct thread * 1032176735Sjeffrunq_steal(struct runq *rq, int cpu) 1033121790Sjeff{ 1034121790Sjeff struct rqhead *rqh; 1035121790Sjeff struct rqbits *rqb; 1036177435Sjeff struct thread *td; 1037121790Sjeff int word; 1038121790Sjeff int bit; 1039121790Sjeff 1040121790Sjeff rqb = &rq->rq_status; 1041121790Sjeff for (word = 0; word < RQB_LEN; word++) { 1042121790Sjeff if (rqb->rqb_bits[word] == 0) 1043121790Sjeff continue; 1044121790Sjeff for (bit = 0; bit < RQB_BPW; bit++) { 1045123231Speter if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1046121790Sjeff continue; 1047121790Sjeff rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1048177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) 1049177435Sjeff if (THREAD_CAN_MIGRATE(td) && 1050177435Sjeff THREAD_CAN_SCHED(td, cpu)) 1051177435Sjeff return (td); 1052121790Sjeff } 1053121790Sjeff } 1054121790Sjeff return (NULL); 1055121790Sjeff} 1056121790Sjeff 1057171482Sjeff/* 1058171482Sjeff * Attempt to steal a thread in priority order from a thread queue. 1059171482Sjeff */ 1060177435Sjeffstatic struct thread * 1061176735Sjefftdq_steal(struct tdq *tdq, int cpu) 1062121790Sjeff{ 1063177435Sjeff struct thread *td; 1064121790Sjeff 1065171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1066177435Sjeff if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1067177435Sjeff return (td); 1068177435Sjeff if ((td = runq_steal_from(&tdq->tdq_timeshare, 1069177435Sjeff cpu, tdq->tdq_ridx)) != NULL) 1070177435Sjeff return (td); 1071176735Sjeff return (runq_steal(&tdq->tdq_idle, cpu)); 1072121790Sjeff} 1073123433Sjeff 1074171482Sjeff/* 1075171482Sjeff * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1076172409Sjeff * current lock and returns with the assigned queue locked. 1077171482Sjeff */ 1078171482Sjeffstatic inline struct tdq * 1079177435Sjeffsched_setcpu(struct thread *td, int cpu, int flags) 1080123433Sjeff{ 1081177435Sjeff 1082171482Sjeff struct tdq *tdq; 1083123433Sjeff 1084177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1085171482Sjeff tdq = TDQ_CPU(cpu); 1086177435Sjeff td->td_sched->ts_cpu = cpu; 1087177435Sjeff /* 1088177435Sjeff * If the lock matches just return the queue. 1089177435Sjeff */ 1090171482Sjeff if (td->td_lock == TDQ_LOCKPTR(tdq)) 1091171482Sjeff return (tdq); 1092171482Sjeff#ifdef notyet 1093123433Sjeff /* 1094172293Sjeff * If the thread isn't running its lockptr is a 1095171482Sjeff * turnstile or a sleepqueue. We can just lock_set without 1096171482Sjeff * blocking. 1097123685Sjeff */ 1098171482Sjeff if (TD_CAN_RUN(td)) { 1099171482Sjeff TDQ_LOCK(tdq); 1100171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1101171482Sjeff return (tdq); 1102171482Sjeff } 1103171482Sjeff#endif 1104166108Sjeff /* 1105171482Sjeff * The hard case, migration, we need to block the thread first to 1106171482Sjeff * prevent order reversals with other cpus locks. 1107166108Sjeff */ 1108202889Sattilio spinlock_enter(); 1109171482Sjeff thread_lock_block(td); 1110171482Sjeff TDQ_LOCK(tdq); 1111171713Sjeff thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1112202889Sattilio spinlock_exit(); 1113171482Sjeff return (tdq); 1114166108Sjeff} 1115166108Sjeff 1116178272SjeffSCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1117178272SjeffSCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1118178272SjeffSCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1119178272SjeffSCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1120178272SjeffSCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1121178272SjeffSCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1122178272Sjeff 1123166108Sjeffstatic int 1124177435Sjeffsched_pickcpu(struct thread *td, int flags) 1125171482Sjeff{ 1126176735Sjeff struct cpu_group *cg; 1127177435Sjeff struct td_sched *ts; 1128171482Sjeff struct tdq *tdq; 1129194779Sjeff cpuset_t mask; 1130166108Sjeff int self; 1131166108Sjeff int pri; 1132166108Sjeff int cpu; 1133166108Sjeff 1134176735Sjeff self = PCPU_GET(cpuid); 1135177435Sjeff ts = td->td_sched; 1136166108Sjeff if (smp_started == 0) 1137166108Sjeff return (self); 1138171506Sjeff /* 1139171506Sjeff * Don't migrate a running thread from sched_switch(). 1140171506Sjeff */ 1141176735Sjeff if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1142176735Sjeff return (ts->ts_cpu); 1143166108Sjeff /* 1144176735Sjeff * Prefer to run interrupt threads on the processors that generate 1145176735Sjeff * the interrupt. 1146166108Sjeff */ 1147176735Sjeff if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1148178272Sjeff curthread->td_intr_nesting_level && ts->ts_cpu != self) { 1149178272Sjeff SCHED_STAT_INC(pickcpu_intrbind); 1150176735Sjeff ts->ts_cpu = self; 1151178272Sjeff } 1152166108Sjeff /* 1153176735Sjeff * If the thread can run on the last cpu and the affinity has not 1154176735Sjeff * expired or it is idle run it there. 1155166108Sjeff */ 1156176735Sjeff pri = td->td_priority; 1157176735Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1158176735Sjeff if (THREAD_CAN_SCHED(td, ts->ts_cpu)) { 1159178272Sjeff if (tdq->tdq_lowpri > PRI_MIN_IDLE) { 1160178272Sjeff SCHED_STAT_INC(pickcpu_idle_affinity); 1161176735Sjeff return (ts->ts_cpu); 1162178272Sjeff } 1163178272Sjeff if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri) { 1164178272Sjeff SCHED_STAT_INC(pickcpu_affinity); 1165176735Sjeff return (ts->ts_cpu); 1166178272Sjeff } 1167139334Sjeff } 1168123433Sjeff /* 1169176735Sjeff * Search for the highest level in the tree that still has affinity. 1170123433Sjeff */ 1171176735Sjeff cg = NULL; 1172176735Sjeff for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent) 1173176735Sjeff if (SCHED_AFFINITY(ts, cg->cg_level)) 1174176735Sjeff break; 1175176735Sjeff cpu = -1; 1176194779Sjeff mask = td->td_cpuset->cs_mask; 1177176735Sjeff if (cg) 1178176735Sjeff cpu = sched_lowest(cg, mask, pri); 1179176735Sjeff if (cpu == -1) 1180176735Sjeff cpu = sched_lowest(cpu_top, mask, -1); 1181171506Sjeff /* 1182176735Sjeff * Compare the lowest loaded cpu to current cpu. 1183171506Sjeff */ 1184177005Sjeff if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1185178272Sjeff TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) { 1186178272Sjeff SCHED_STAT_INC(pickcpu_local); 1187177005Sjeff cpu = self; 1188178272Sjeff } else 1189178272Sjeff SCHED_STAT_INC(pickcpu_lowest); 1190178272Sjeff if (cpu != ts->ts_cpu) 1191178272Sjeff SCHED_STAT_INC(pickcpu_migration); 1192177005Sjeff KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1193171482Sjeff return (cpu); 1194123433Sjeff} 1195176735Sjeff#endif 1196123433Sjeff 1197117326Sjeff/* 1198121790Sjeff * Pick the highest priority task we have and return it. 1199117326Sjeff */ 1200177435Sjeffstatic struct thread * 1201164936Sjuliantdq_choose(struct tdq *tdq) 1202110267Sjeff{ 1203177435Sjeff struct thread *td; 1204110267Sjeff 1205171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1206177435Sjeff td = runq_choose(&tdq->tdq_realtime); 1207177435Sjeff if (td != NULL) 1208177435Sjeff return (td); 1209177435Sjeff td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1210177435Sjeff if (td != NULL) { 1211177435Sjeff KASSERT(td->td_priority >= PRI_MIN_TIMESHARE, 1212165762Sjeff ("tdq_choose: Invalid priority on timeshare queue %d", 1213177435Sjeff td->td_priority)); 1214177435Sjeff return (td); 1215165762Sjeff } 1216177435Sjeff td = runq_choose(&tdq->tdq_idle); 1217177435Sjeff if (td != NULL) { 1218177435Sjeff KASSERT(td->td_priority >= PRI_MIN_IDLE, 1219165762Sjeff ("tdq_choose: Invalid priority on idle queue %d", 1220177435Sjeff td->td_priority)); 1221177435Sjeff return (td); 1222165762Sjeff } 1223165762Sjeff 1224165762Sjeff return (NULL); 1225110267Sjeff} 1226110267Sjeff 1227171482Sjeff/* 1228171482Sjeff * Initialize a thread queue. 1229171482Sjeff */ 1230109864Sjeffstatic void 1231164936Sjuliantdq_setup(struct tdq *tdq) 1232110028Sjeff{ 1233171482Sjeff 1234171713Sjeff if (bootverbose) 1235171713Sjeff printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1236165762Sjeff runq_init(&tdq->tdq_realtime); 1237165762Sjeff runq_init(&tdq->tdq_timeshare); 1238165620Sjeff runq_init(&tdq->tdq_idle); 1239176735Sjeff snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1240176735Sjeff "sched lock %d", (int)TDQ_ID(tdq)); 1241176735Sjeff mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1242176735Sjeff MTX_SPIN | MTX_RECURSE); 1243187357Sjeff#ifdef KTR 1244187357Sjeff snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname), 1245187357Sjeff "CPU %d load", (int)TDQ_ID(tdq)); 1246187357Sjeff#endif 1247110028Sjeff} 1248110028Sjeff 1249171713Sjeff#ifdef SMP 1250110028Sjeffstatic void 1251171713Sjeffsched_setup_smp(void) 1252171713Sjeff{ 1253171713Sjeff struct tdq *tdq; 1254171713Sjeff int i; 1255171713Sjeff 1256176735Sjeff cpu_top = smp_topo(); 1257176735Sjeff for (i = 0; i < MAXCPU; i++) { 1258171713Sjeff if (CPU_ABSENT(i)) 1259171713Sjeff continue; 1260176735Sjeff tdq = TDQ_CPU(i); 1261171713Sjeff tdq_setup(tdq); 1262176735Sjeff tdq->tdq_cg = smp_topo_find(cpu_top, i); 1263176735Sjeff if (tdq->tdq_cg == NULL) 1264176735Sjeff panic("Can't find cpu group for %d\n", i); 1265123433Sjeff } 1266176735Sjeff balance_tdq = TDQ_SELF(); 1267176735Sjeff sched_balance(); 1268171713Sjeff} 1269171713Sjeff#endif 1270171713Sjeff 1271171713Sjeff/* 1272171713Sjeff * Setup the thread queues and initialize the topology based on MD 1273171713Sjeff * information. 1274171713Sjeff */ 1275171713Sjeffstatic void 1276171713Sjeffsched_setup(void *dummy) 1277171713Sjeff{ 1278171713Sjeff struct tdq *tdq; 1279171713Sjeff 1280171713Sjeff tdq = TDQ_SELF(); 1281171713Sjeff#ifdef SMP 1282176734Sjeff sched_setup_smp(); 1283117237Sjeff#else 1284171713Sjeff tdq_setup(tdq); 1285116069Sjeff#endif 1286171482Sjeff /* 1287171482Sjeff * To avoid divide-by-zero, we set realstathz a dummy value 1288171482Sjeff * in case which sched_clock() called before sched_initticks(). 1289171482Sjeff */ 1290171482Sjeff realstathz = hz; 1291171482Sjeff sched_slice = (realstathz/10); /* ~100ms */ 1292171482Sjeff tickincr = 1 << SCHED_TICK_SHIFT; 1293171482Sjeff 1294171482Sjeff /* Add thread0's load since it's running. */ 1295171482Sjeff TDQ_LOCK(tdq); 1296171713Sjeff thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1297177435Sjeff tdq_load_add(tdq, &thread0); 1298176735Sjeff tdq->tdq_lowpri = thread0.td_priority; 1299171482Sjeff TDQ_UNLOCK(tdq); 1300109864Sjeff} 1301109864Sjeff 1302171482Sjeff/* 1303171482Sjeff * This routine determines the tickincr after stathz and hz are setup. 1304171482Sjeff */ 1305153533Sdavidxu/* ARGSUSED */ 1306153533Sdavidxustatic void 1307153533Sdavidxusched_initticks(void *dummy) 1308153533Sdavidxu{ 1309171482Sjeff int incr; 1310171482Sjeff 1311153533Sdavidxu realstathz = stathz ? stathz : hz; 1312166229Sjeff sched_slice = (realstathz/10); /* ~100ms */ 1313153533Sdavidxu 1314153533Sdavidxu /* 1315165762Sjeff * tickincr is shifted out by 10 to avoid rounding errors due to 1316165766Sjeff * hz not being evenly divisible by stathz on all platforms. 1317153533Sdavidxu */ 1318171482Sjeff incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1319165762Sjeff /* 1320165762Sjeff * This does not work for values of stathz that are more than 1321165762Sjeff * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1322165762Sjeff */ 1323171482Sjeff if (incr == 0) 1324171482Sjeff incr = 1; 1325171482Sjeff tickincr = incr; 1326166108Sjeff#ifdef SMP 1327171899Sjeff /* 1328172409Sjeff * Set the default balance interval now that we know 1329172409Sjeff * what realstathz is. 1330172409Sjeff */ 1331172409Sjeff balance_interval = realstathz; 1332172409Sjeff /* 1333189787Sjeff * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4. 1334189787Sjeff * This prevents excess thrashing on large machines and excess idle 1335189787Sjeff * on smaller machines. 1336171899Sjeff */ 1337189787Sjeff steal_thresh = min(fls(mp_ncpus) - 1, 3); 1338166108Sjeff affinity = SCHED_AFFINITY_DEFAULT; 1339166108Sjeff#endif 1340153533Sdavidxu} 1341153533Sdavidxu 1342153533Sdavidxu 1343109864Sjeff/* 1344171482Sjeff * This is the core of the interactivity algorithm. Determines a score based 1345171482Sjeff * on past behavior. It is the ratio of sleep time to run time scaled to 1346171482Sjeff * a [0, 100] integer. This is the voluntary sleep time of a process, which 1347171482Sjeff * differs from the cpu usage because it does not account for time spent 1348171482Sjeff * waiting on a run-queue. Would be prettier if we had floating point. 1349171482Sjeff */ 1350171482Sjeffstatic int 1351171482Sjeffsched_interact_score(struct thread *td) 1352171482Sjeff{ 1353171482Sjeff struct td_sched *ts; 1354171482Sjeff int div; 1355171482Sjeff 1356171482Sjeff ts = td->td_sched; 1357171482Sjeff /* 1358171482Sjeff * The score is only needed if this is likely to be an interactive 1359171482Sjeff * task. Don't go through the expense of computing it if there's 1360171482Sjeff * no chance. 1361171482Sjeff */ 1362171482Sjeff if (sched_interact <= SCHED_INTERACT_HALF && 1363171482Sjeff ts->ts_runtime >= ts->ts_slptime) 1364171482Sjeff return (SCHED_INTERACT_HALF); 1365171482Sjeff 1366171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1367171482Sjeff div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1368171482Sjeff return (SCHED_INTERACT_HALF + 1369171482Sjeff (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1370171482Sjeff } 1371171482Sjeff if (ts->ts_slptime > ts->ts_runtime) { 1372171482Sjeff div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1373171482Sjeff return (ts->ts_runtime / div); 1374171482Sjeff } 1375171482Sjeff /* runtime == slptime */ 1376171482Sjeff if (ts->ts_runtime) 1377171482Sjeff return (SCHED_INTERACT_HALF); 1378171482Sjeff 1379171482Sjeff /* 1380171482Sjeff * This can happen if slptime and runtime are 0. 1381171482Sjeff */ 1382171482Sjeff return (0); 1383171482Sjeff 1384171482Sjeff} 1385171482Sjeff 1386171482Sjeff/* 1387109864Sjeff * Scale the scheduling priority according to the "interactivity" of this 1388109864Sjeff * process. 1389109864Sjeff */ 1390113357Sjeffstatic void 1391163709Sjbsched_priority(struct thread *td) 1392109864Sjeff{ 1393165762Sjeff int score; 1394109864Sjeff int pri; 1395109864Sjeff 1396163709Sjb if (td->td_pri_class != PRI_TIMESHARE) 1397113357Sjeff return; 1398112966Sjeff /* 1399165762Sjeff * If the score is interactive we place the thread in the realtime 1400165762Sjeff * queue with a priority that is less than kernel and interrupt 1401165762Sjeff * priorities. These threads are not subject to nice restrictions. 1402112966Sjeff * 1403171482Sjeff * Scores greater than this are placed on the normal timeshare queue 1404165762Sjeff * where the priority is partially decided by the most recent cpu 1405165762Sjeff * utilization and the rest is decided by nice value. 1406172293Sjeff * 1407172293Sjeff * The nice value of the process has a linear effect on the calculated 1408172293Sjeff * score. Negative nice values make it easier for a thread to be 1409172293Sjeff * considered interactive. 1410112966Sjeff */ 1411198126Sjhb score = imax(0, sched_interact_score(td) + td->td_proc->p_nice); 1412165762Sjeff if (score < sched_interact) { 1413165762Sjeff pri = PRI_MIN_REALTIME; 1414165762Sjeff pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact) 1415165762Sjeff * score; 1416165762Sjeff KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME, 1417166208Sjeff ("sched_priority: invalid interactive priority %d score %d", 1418166208Sjeff pri, score)); 1419165762Sjeff } else { 1420165762Sjeff pri = SCHED_PRI_MIN; 1421165762Sjeff if (td->td_sched->ts_ticks) 1422165762Sjeff pri += SCHED_PRI_TICKS(td->td_sched); 1423165762Sjeff pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1424171482Sjeff KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE, 1425171482Sjeff ("sched_priority: invalid priority %d: nice %d, " 1426171482Sjeff "ticks %d ftick %d ltick %d tick pri %d", 1427171482Sjeff pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1428171482Sjeff td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1429171482Sjeff SCHED_PRI_TICKS(td->td_sched))); 1430165762Sjeff } 1431165762Sjeff sched_user_prio(td, pri); 1432112966Sjeff 1433112966Sjeff return; 1434109864Sjeff} 1435109864Sjeff 1436121868Sjeff/* 1437121868Sjeff * This routine enforces a maximum limit on the amount of scheduling history 1438171482Sjeff * kept. It is called after either the slptime or runtime is adjusted. This 1439171482Sjeff * function is ugly due to integer math. 1440121868Sjeff */ 1441116463Sjeffstatic void 1442163709Sjbsched_interact_update(struct thread *td) 1443116463Sjeff{ 1444165819Sjeff struct td_sched *ts; 1445166208Sjeff u_int sum; 1446121605Sjeff 1447165819Sjeff ts = td->td_sched; 1448171482Sjeff sum = ts->ts_runtime + ts->ts_slptime; 1449121868Sjeff if (sum < SCHED_SLP_RUN_MAX) 1450121868Sjeff return; 1451121868Sjeff /* 1452165819Sjeff * This only happens from two places: 1453165819Sjeff * 1) We have added an unusual amount of run time from fork_exit. 1454165819Sjeff * 2) We have added an unusual amount of sleep time from sched_sleep(). 1455165819Sjeff */ 1456165819Sjeff if (sum > SCHED_SLP_RUN_MAX * 2) { 1457171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1458171482Sjeff ts->ts_runtime = SCHED_SLP_RUN_MAX; 1459171482Sjeff ts->ts_slptime = 1; 1460165819Sjeff } else { 1461171482Sjeff ts->ts_slptime = SCHED_SLP_RUN_MAX; 1462171482Sjeff ts->ts_runtime = 1; 1463165819Sjeff } 1464165819Sjeff return; 1465165819Sjeff } 1466165819Sjeff /* 1467121868Sjeff * If we have exceeded by more than 1/5th then the algorithm below 1468121868Sjeff * will not bring us back into range. Dividing by two here forces 1469133427Sjeff * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1470121868Sjeff */ 1471127850Sjeff if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1472171482Sjeff ts->ts_runtime /= 2; 1473171482Sjeff ts->ts_slptime /= 2; 1474121868Sjeff return; 1475116463Sjeff } 1476171482Sjeff ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1477171482Sjeff ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1478116463Sjeff} 1479116463Sjeff 1480171482Sjeff/* 1481171482Sjeff * Scale back the interactivity history when a child thread is created. The 1482171482Sjeff * history is inherited from the parent but the thread may behave totally 1483171482Sjeff * differently. For example, a shell spawning a compiler process. We want 1484171482Sjeff * to learn that the compiler is behaving badly very quickly. 1485171482Sjeff */ 1486121868Sjeffstatic void 1487163709Sjbsched_interact_fork(struct thread *td) 1488121868Sjeff{ 1489121868Sjeff int ratio; 1490121868Sjeff int sum; 1491121868Sjeff 1492171482Sjeff sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1493121868Sjeff if (sum > SCHED_SLP_RUN_FORK) { 1494121868Sjeff ratio = sum / SCHED_SLP_RUN_FORK; 1495171482Sjeff td->td_sched->ts_runtime /= ratio; 1496171482Sjeff td->td_sched->ts_slptime /= ratio; 1497121868Sjeff } 1498121868Sjeff} 1499121868Sjeff 1500113357Sjeff/* 1501171482Sjeff * Called from proc0_init() to setup the scheduler fields. 1502134791Sjulian */ 1503134791Sjulianvoid 1504134791Sjulianschedinit(void) 1505134791Sjulian{ 1506165762Sjeff 1507134791Sjulian /* 1508134791Sjulian * Set up the scheduler specific parts of proc0. 1509134791Sjulian */ 1510136167Sjulian proc0.p_sched = NULL; /* XXX */ 1511164936Sjulian thread0.td_sched = &td_sched0; 1512165762Sjeff td_sched0.ts_ltick = ticks; 1513165796Sjeff td_sched0.ts_ftick = ticks; 1514177009Sjeff td_sched0.ts_slice = sched_slice; 1515134791Sjulian} 1516134791Sjulian 1517134791Sjulian/* 1518113357Sjeff * This is only somewhat accurate since given many processes of the same 1519113357Sjeff * priority they will switch when their slices run out, which will be 1520165762Sjeff * at most sched_slice stathz ticks. 1521113357Sjeff */ 1522109864Sjeffint 1523109864Sjeffsched_rr_interval(void) 1524109864Sjeff{ 1525165762Sjeff 1526165762Sjeff /* Convert sched_slice to hz */ 1527165762Sjeff return (hz/(realstathz/sched_slice)); 1528109864Sjeff} 1529109864Sjeff 1530171482Sjeff/* 1531171482Sjeff * Update the percent cpu tracking information when it is requested or 1532171482Sjeff * the total history exceeds the maximum. We keep a sliding history of 1533171482Sjeff * tick counts that slowly decays. This is less precise than the 4BSD 1534171482Sjeff * mechanism since it happens with less regular and frequent events. 1535171482Sjeff */ 1536121790Sjeffstatic void 1537164936Sjuliansched_pctcpu_update(struct td_sched *ts) 1538109864Sjeff{ 1539165762Sjeff 1540165762Sjeff if (ts->ts_ticks == 0) 1541165762Sjeff return; 1542165796Sjeff if (ticks - (hz / 10) < ts->ts_ltick && 1543165796Sjeff SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1544165796Sjeff return; 1545109864Sjeff /* 1546109864Sjeff * Adjust counters and watermark for pctcpu calc. 1547116365Sjeff */ 1548165762Sjeff if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1549164936Sjulian ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1550165762Sjeff SCHED_TICK_TARG; 1551165762Sjeff else 1552164936Sjulian ts->ts_ticks = 0; 1553164936Sjulian ts->ts_ltick = ticks; 1554165762Sjeff ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1555109864Sjeff} 1556109864Sjeff 1557171482Sjeff/* 1558171482Sjeff * Adjust the priority of a thread. Move it to the appropriate run-queue 1559171482Sjeff * if necessary. This is the back-end for several priority related 1560171482Sjeff * functions. 1561171482Sjeff */ 1562165762Sjeffstatic void 1563139453Sjhbsched_thread_priority(struct thread *td, u_char prio) 1564109864Sjeff{ 1565164936Sjulian struct td_sched *ts; 1566177009Sjeff struct tdq *tdq; 1567177009Sjeff int oldpri; 1568109864Sjeff 1569187357Sjeff KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio", 1570187357Sjeff "prio:%d", td->td_priority, "new prio:%d", prio, 1571187357Sjeff KTR_ATTR_LINKED, sched_tdname(curthread)); 1572187357Sjeff if (td != curthread && prio > td->td_priority) { 1573187357Sjeff KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 1574187357Sjeff "lend prio", "prio:%d", td->td_priority, "new prio:%d", 1575187357Sjeff prio, KTR_ATTR_LINKED, sched_tdname(td)); 1576187357Sjeff } 1577164936Sjulian ts = td->td_sched; 1578170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1579139453Sjhb if (td->td_priority == prio) 1580139453Sjhb return; 1581177376Sjeff /* 1582177376Sjeff * If the priority has been elevated due to priority 1583177376Sjeff * propagation, we may have to move ourselves to a new 1584177376Sjeff * queue. This could be optimized to not re-add in some 1585177376Sjeff * cases. 1586177376Sjeff */ 1587165766Sjeff if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1588165762Sjeff sched_rem(td); 1589165762Sjeff td->td_priority = prio; 1590171482Sjeff sched_add(td, SRQ_BORROWING); 1591177009Sjeff return; 1592177009Sjeff } 1593177376Sjeff /* 1594177376Sjeff * If the thread is currently running we may have to adjust the lowpri 1595177376Sjeff * information so other cpus are aware of our current priority. 1596177376Sjeff */ 1597177009Sjeff if (TD_IS_RUNNING(td)) { 1598177376Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1599177376Sjeff oldpri = td->td_priority; 1600177376Sjeff td->td_priority = prio; 1601176735Sjeff if (prio < tdq->tdq_lowpri) 1602171482Sjeff tdq->tdq_lowpri = prio; 1603176735Sjeff else if (tdq->tdq_lowpri == oldpri) 1604176735Sjeff tdq_setlowpri(tdq, td); 1605177376Sjeff return; 1606177009Sjeff } 1607177376Sjeff td->td_priority = prio; 1608109864Sjeff} 1609109864Sjeff 1610139453Sjhb/* 1611139453Sjhb * Update a thread's priority when it is lent another thread's 1612139453Sjhb * priority. 1613139453Sjhb */ 1614109864Sjeffvoid 1615139453Sjhbsched_lend_prio(struct thread *td, u_char prio) 1616139453Sjhb{ 1617139453Sjhb 1618139453Sjhb td->td_flags |= TDF_BORROWING; 1619139453Sjhb sched_thread_priority(td, prio); 1620139453Sjhb} 1621139453Sjhb 1622139453Sjhb/* 1623139453Sjhb * Restore a thread's priority when priority propagation is 1624139453Sjhb * over. The prio argument is the minimum priority the thread 1625139453Sjhb * needs to have to satisfy other possible priority lending 1626139453Sjhb * requests. If the thread's regular priority is less 1627139453Sjhb * important than prio, the thread will keep a priority boost 1628139453Sjhb * of prio. 1629139453Sjhb */ 1630139453Sjhbvoid 1631139453Sjhbsched_unlend_prio(struct thread *td, u_char prio) 1632139453Sjhb{ 1633139453Sjhb u_char base_pri; 1634139453Sjhb 1635139453Sjhb if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1636139453Sjhb td->td_base_pri <= PRI_MAX_TIMESHARE) 1637163709Sjb base_pri = td->td_user_pri; 1638139453Sjhb else 1639139453Sjhb base_pri = td->td_base_pri; 1640139453Sjhb if (prio >= base_pri) { 1641139455Sjhb td->td_flags &= ~TDF_BORROWING; 1642139453Sjhb sched_thread_priority(td, base_pri); 1643139453Sjhb } else 1644139453Sjhb sched_lend_prio(td, prio); 1645139453Sjhb} 1646139453Sjhb 1647171482Sjeff/* 1648171482Sjeff * Standard entry for setting the priority to an absolute value. 1649171482Sjeff */ 1650139453Sjhbvoid 1651139453Sjhbsched_prio(struct thread *td, u_char prio) 1652139453Sjhb{ 1653139453Sjhb u_char oldprio; 1654139453Sjhb 1655139453Sjhb /* First, update the base priority. */ 1656139453Sjhb td->td_base_pri = prio; 1657139453Sjhb 1658139453Sjhb /* 1659139455Sjhb * If the thread is borrowing another thread's priority, don't 1660139453Sjhb * ever lower the priority. 1661139453Sjhb */ 1662139453Sjhb if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1663139453Sjhb return; 1664139453Sjhb 1665139453Sjhb /* Change the real priority. */ 1666139453Sjhb oldprio = td->td_priority; 1667139453Sjhb sched_thread_priority(td, prio); 1668139453Sjhb 1669139453Sjhb /* 1670139453Sjhb * If the thread is on a turnstile, then let the turnstile update 1671139453Sjhb * its state. 1672139453Sjhb */ 1673139453Sjhb if (TD_ON_LOCK(td) && oldprio != prio) 1674139453Sjhb turnstile_adjust(td, oldprio); 1675139453Sjhb} 1676139455Sjhb 1677171482Sjeff/* 1678171482Sjeff * Set the base user priority, does not effect current running priority. 1679171482Sjeff */ 1680139453Sjhbvoid 1681163709Sjbsched_user_prio(struct thread *td, u_char prio) 1682161599Sdavidxu{ 1683161599Sdavidxu u_char oldprio; 1684161599Sdavidxu 1685163709Sjb td->td_base_user_pri = prio; 1686164939Sjulian if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio) 1687164939Sjulian return; 1688163709Sjb oldprio = td->td_user_pri; 1689163709Sjb td->td_user_pri = prio; 1690161599Sdavidxu} 1691161599Sdavidxu 1692161599Sdavidxuvoid 1693161599Sdavidxusched_lend_user_prio(struct thread *td, u_char prio) 1694161599Sdavidxu{ 1695161599Sdavidxu u_char oldprio; 1696161599Sdavidxu 1697174536Sdavidxu THREAD_LOCK_ASSERT(td, MA_OWNED); 1698161599Sdavidxu td->td_flags |= TDF_UBORROWING; 1699164091Smaxim oldprio = td->td_user_pri; 1700163709Sjb td->td_user_pri = prio; 1701161599Sdavidxu} 1702161599Sdavidxu 1703161599Sdavidxuvoid 1704161599Sdavidxusched_unlend_user_prio(struct thread *td, u_char prio) 1705161599Sdavidxu{ 1706161599Sdavidxu u_char base_pri; 1707161599Sdavidxu 1708174536Sdavidxu THREAD_LOCK_ASSERT(td, MA_OWNED); 1709163709Sjb base_pri = td->td_base_user_pri; 1710161599Sdavidxu if (prio >= base_pri) { 1711161599Sdavidxu td->td_flags &= ~TDF_UBORROWING; 1712163709Sjb sched_user_prio(td, base_pri); 1713174536Sdavidxu } else { 1714161599Sdavidxu sched_lend_user_prio(td, prio); 1715174536Sdavidxu } 1716161599Sdavidxu} 1717161599Sdavidxu 1718171482Sjeff/* 1719171713Sjeff * Handle migration from sched_switch(). This happens only for 1720171713Sjeff * cpu binding. 1721171713Sjeff */ 1722171713Sjeffstatic struct mtx * 1723171713Sjeffsched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1724171713Sjeff{ 1725171713Sjeff struct tdq *tdn; 1726171713Sjeff 1727171713Sjeff tdn = TDQ_CPU(td->td_sched->ts_cpu); 1728171713Sjeff#ifdef SMP 1729177435Sjeff tdq_load_rem(tdq, td); 1730171713Sjeff /* 1731171713Sjeff * Do the lock dance required to avoid LOR. We grab an extra 1732171713Sjeff * spinlock nesting to prevent preemption while we're 1733171713Sjeff * not holding either run-queue lock. 1734171713Sjeff */ 1735171713Sjeff spinlock_enter(); 1736202889Sattilio thread_lock_block(td); /* This releases the lock on tdq. */ 1737197223Sattilio 1738197223Sattilio /* 1739197223Sattilio * Acquire both run-queue locks before placing the thread on the new 1740197223Sattilio * run-queue to avoid deadlocks created by placing a thread with a 1741197223Sattilio * blocked lock on the run-queue of a remote processor. The deadlock 1742197223Sattilio * occurs when a third processor attempts to lock the two queues in 1743197223Sattilio * question while the target processor is spinning with its own 1744197223Sattilio * run-queue lock held while waiting for the blocked lock to clear. 1745197223Sattilio */ 1746197223Sattilio tdq_lock_pair(tdn, tdq); 1747171713Sjeff tdq_add(tdn, td, flags); 1748177435Sjeff tdq_notify(tdn, td); 1749197223Sattilio TDQ_UNLOCK(tdn); 1750171713Sjeff spinlock_exit(); 1751171713Sjeff#endif 1752171713Sjeff return (TDQ_LOCKPTR(tdn)); 1753171713Sjeff} 1754171713Sjeff 1755171713Sjeff/* 1756202889Sattilio * Variadic version of thread_lock_unblock() that does not assume td_lock 1757202889Sattilio * is blocked. 1758171482Sjeff */ 1759171482Sjeffstatic inline void 1760171482Sjeffthread_unblock_switch(struct thread *td, struct mtx *mtx) 1761171482Sjeff{ 1762171482Sjeff atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1763171482Sjeff (uintptr_t)mtx); 1764171482Sjeff} 1765171482Sjeff 1766171482Sjeff/* 1767171482Sjeff * Switch threads. This function has to handle threads coming in while 1768171482Sjeff * blocked for some reason, running, or idle. It also must deal with 1769171482Sjeff * migrating a thread from one queue to another as running threads may 1770171482Sjeff * be assigned elsewhere via binding. 1771171482Sjeff */ 1772161599Sdavidxuvoid 1773135051Sjuliansched_switch(struct thread *td, struct thread *newtd, int flags) 1774109864Sjeff{ 1775165627Sjeff struct tdq *tdq; 1776164936Sjulian struct td_sched *ts; 1777171482Sjeff struct mtx *mtx; 1778171713Sjeff int srqflag; 1779171482Sjeff int cpuid; 1780109864Sjeff 1781170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1782177376Sjeff KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1783109864Sjeff 1784171482Sjeff cpuid = PCPU_GET(cpuid); 1785171482Sjeff tdq = TDQ_CPU(cpuid); 1786164936Sjulian ts = td->td_sched; 1787171713Sjeff mtx = td->td_lock; 1788171482Sjeff ts->ts_rltick = ticks; 1789133555Sjeff td->td_lastcpu = td->td_oncpu; 1790113339Sjulian td->td_oncpu = NOCPU; 1791132266Sjhb td->td_flags &= ~TDF_NEEDRESCHED; 1792144777Sups td->td_owepreempt = 0; 1793178277Sjeff tdq->tdq_switchcnt++; 1794123434Sjeff /* 1795171482Sjeff * The lock pointer in an idle thread should never change. Reset it 1796171482Sjeff * to CAN_RUN as well. 1797123434Sjeff */ 1798167327Sjulian if (TD_IS_IDLETHREAD(td)) { 1799171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1800139334Sjeff TD_SET_CAN_RUN(td); 1801170293Sjeff } else if (TD_IS_RUNNING(td)) { 1802171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1803171713Sjeff srqflag = (flags & SW_PREEMPT) ? 1804170293Sjeff SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1805171713Sjeff SRQ_OURSELF|SRQ_YIELDING; 1806171713Sjeff if (ts->ts_cpu == cpuid) 1807177435Sjeff tdq_runq_add(tdq, td, srqflag); 1808171713Sjeff else 1809171713Sjeff mtx = sched_switch_migrate(tdq, td, srqflag); 1810171482Sjeff } else { 1811171482Sjeff /* This thread must be going to sleep. */ 1812171482Sjeff TDQ_LOCK(tdq); 1813202889Sattilio mtx = thread_lock_block(td); 1814177435Sjeff tdq_load_rem(tdq, td); 1815171482Sjeff } 1816171482Sjeff /* 1817171482Sjeff * We enter here with the thread blocked and assigned to the 1818171482Sjeff * appropriate cpu run-queue or sleep-queue and with the current 1819171482Sjeff * thread-queue locked. 1820171482Sjeff */ 1821171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1822171482Sjeff newtd = choosethread(); 1823171482Sjeff /* 1824171482Sjeff * Call the MD code to switch contexts if necessary. 1825171482Sjeff */ 1826145256Sjkoshy if (td != newtd) { 1827145256Sjkoshy#ifdef HWPMC_HOOKS 1828145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1829145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1830145256Sjkoshy#endif 1831174629Sjeff lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1832172411Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1833179297Sjb 1834179297Sjb#ifdef KDTRACE_HOOKS 1835179297Sjb /* 1836179297Sjb * If DTrace has set the active vtime enum to anything 1837179297Sjb * other than INACTIVE (0), then it should have set the 1838179297Sjb * function to call. 1839179297Sjb */ 1840179297Sjb if (dtrace_vtime_active) 1841179297Sjb (*dtrace_vtime_switch_func)(newtd); 1842179297Sjb#endif 1843179297Sjb 1844171482Sjeff cpu_switch(td, newtd, mtx); 1845171482Sjeff /* 1846171482Sjeff * We may return from cpu_switch on a different cpu. However, 1847171482Sjeff * we always return with td_lock pointing to the current cpu's 1848171482Sjeff * run queue lock. 1849171482Sjeff */ 1850171482Sjeff cpuid = PCPU_GET(cpuid); 1851171482Sjeff tdq = TDQ_CPU(cpuid); 1852174629Sjeff lock_profile_obtain_lock_success( 1853174629Sjeff &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1854145256Sjkoshy#ifdef HWPMC_HOOKS 1855145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1856145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1857145256Sjkoshy#endif 1858171482Sjeff } else 1859171482Sjeff thread_unblock_switch(td, mtx); 1860171482Sjeff /* 1861171482Sjeff * Assert that all went well and return. 1862171482Sjeff */ 1863171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1864171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1865171482Sjeff td->td_oncpu = cpuid; 1866109864Sjeff} 1867109864Sjeff 1868171482Sjeff/* 1869171482Sjeff * Adjust thread priorities as a result of a nice request. 1870171482Sjeff */ 1871109864Sjeffvoid 1872130551Sjuliansched_nice(struct proc *p, int nice) 1873109864Sjeff{ 1874109864Sjeff struct thread *td; 1875109864Sjeff 1876130551Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 1877165762Sjeff 1878130551Sjulian p->p_nice = nice; 1879163709Sjb FOREACH_THREAD_IN_PROC(p, td) { 1880170293Sjeff thread_lock(td); 1881163709Sjb sched_priority(td); 1882165762Sjeff sched_prio(td, td->td_base_user_pri); 1883170293Sjeff thread_unlock(td); 1884130551Sjulian } 1885109864Sjeff} 1886109864Sjeff 1887171482Sjeff/* 1888171482Sjeff * Record the sleep time for the interactivity scorer. 1889171482Sjeff */ 1890109864Sjeffvoid 1891177085Sjeffsched_sleep(struct thread *td, int prio) 1892109864Sjeff{ 1893165762Sjeff 1894170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1895109864Sjeff 1896172264Sjeff td->td_slptick = ticks; 1897201347Skib if (TD_IS_SUSPENDED(td) || prio >= PSOCK) 1898177085Sjeff td->td_flags |= TDF_CANSWAP; 1899177903Sjeff if (static_boost == 1 && prio) 1900177085Sjeff sched_prio(td, prio); 1901177903Sjeff else if (static_boost && td->td_priority > static_boost) 1902177903Sjeff sched_prio(td, static_boost); 1903109864Sjeff} 1904109864Sjeff 1905171482Sjeff/* 1906171482Sjeff * Schedule a thread to resume execution and record how long it voluntarily 1907171482Sjeff * slept. We also update the pctcpu, interactivity, and priority. 1908171482Sjeff */ 1909109864Sjeffvoid 1910109864Sjeffsched_wakeup(struct thread *td) 1911109864Sjeff{ 1912166229Sjeff struct td_sched *ts; 1913171482Sjeff int slptick; 1914165762Sjeff 1915170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1916166229Sjeff ts = td->td_sched; 1917177085Sjeff td->td_flags &= ~TDF_CANSWAP; 1918109864Sjeff /* 1919165762Sjeff * If we slept for more than a tick update our interactivity and 1920165762Sjeff * priority. 1921109864Sjeff */ 1922172264Sjeff slptick = td->td_slptick; 1923172264Sjeff td->td_slptick = 0; 1924171482Sjeff if (slptick && slptick != ticks) { 1925166208Sjeff u_int hzticks; 1926109864Sjeff 1927171482Sjeff hzticks = (ticks - slptick) << SCHED_TICK_SHIFT; 1928171482Sjeff ts->ts_slptime += hzticks; 1929165819Sjeff sched_interact_update(td); 1930166229Sjeff sched_pctcpu_update(ts); 1931109864Sjeff } 1932166229Sjeff /* Reset the slice value after we sleep. */ 1933166229Sjeff ts->ts_slice = sched_slice; 1934166190Sjeff sched_add(td, SRQ_BORING); 1935109864Sjeff} 1936109864Sjeff 1937109864Sjeff/* 1938109864Sjeff * Penalize the parent for creating a new child and initialize the child's 1939109864Sjeff * priority. 1940109864Sjeff */ 1941109864Sjeffvoid 1942163709Sjbsched_fork(struct thread *td, struct thread *child) 1943109864Sjeff{ 1944170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1945164936Sjulian sched_fork_thread(td, child); 1946165762Sjeff /* 1947165762Sjeff * Penalize the parent and child for forking. 1948165762Sjeff */ 1949165762Sjeff sched_interact_fork(child); 1950165762Sjeff sched_priority(child); 1951171482Sjeff td->td_sched->ts_runtime += tickincr; 1952165762Sjeff sched_interact_update(td); 1953165762Sjeff sched_priority(td); 1954164936Sjulian} 1955109864Sjeff 1956171482Sjeff/* 1957171482Sjeff * Fork a new thread, may be within the same process. 1958171482Sjeff */ 1959164936Sjulianvoid 1960164936Sjuliansched_fork_thread(struct thread *td, struct thread *child) 1961164936Sjulian{ 1962164936Sjulian struct td_sched *ts; 1963164936Sjulian struct td_sched *ts2; 1964164936Sjulian 1965177426Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1966165762Sjeff /* 1967165762Sjeff * Initialize child. 1968165762Sjeff */ 1969177426Sjeff ts = td->td_sched; 1970177426Sjeff ts2 = child->td_sched; 1971171482Sjeff child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1972176735Sjeff child->td_cpuset = cpuset_ref(td->td_cpuset); 1973164936Sjulian ts2->ts_cpu = ts->ts_cpu; 1974177426Sjeff ts2->ts_flags = 0; 1975165762Sjeff /* 1976165762Sjeff * Grab our parents cpu estimation information and priority. 1977165762Sjeff */ 1978164936Sjulian ts2->ts_ticks = ts->ts_ticks; 1979164936Sjulian ts2->ts_ltick = ts->ts_ltick; 1980199764Sivoras ts2->ts_incrtick = ts->ts_incrtick; 1981164936Sjulian ts2->ts_ftick = ts->ts_ftick; 1982165762Sjeff child->td_user_pri = td->td_user_pri; 1983165762Sjeff child->td_base_user_pri = td->td_base_user_pri; 1984165762Sjeff /* 1985165762Sjeff * And update interactivity score. 1986165762Sjeff */ 1987171482Sjeff ts2->ts_slptime = ts->ts_slptime; 1988171482Sjeff ts2->ts_runtime = ts->ts_runtime; 1989165762Sjeff ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 1990187357Sjeff#ifdef KTR 1991187357Sjeff bzero(ts2->ts_name, sizeof(ts2->ts_name)); 1992187357Sjeff#endif 1993113357Sjeff} 1994113357Sjeff 1995171482Sjeff/* 1996171482Sjeff * Adjust the priority class of a thread. 1997171482Sjeff */ 1998113357Sjeffvoid 1999163709Sjbsched_class(struct thread *td, int class) 2000113357Sjeff{ 2001113357Sjeff 2002170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2003163709Sjb if (td->td_pri_class == class) 2004113357Sjeff return; 2005163709Sjb td->td_pri_class = class; 2006109864Sjeff} 2007109864Sjeff 2008109864Sjeff/* 2009109864Sjeff * Return some of the child's priority and interactivity to the parent. 2010109864Sjeff */ 2011109864Sjeffvoid 2012164939Sjuliansched_exit(struct proc *p, struct thread *child) 2013109864Sjeff{ 2014165762Sjeff struct thread *td; 2015113372Sjeff 2016187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit", 2017187357Sjeff "prio:td", child->td_priority); 2018177368Sjeff PROC_LOCK_ASSERT(p, MA_OWNED); 2019165762Sjeff td = FIRST_THREAD_IN_PROC(p); 2020165762Sjeff sched_exit_thread(td, child); 2021113372Sjeff} 2022113372Sjeff 2023171482Sjeff/* 2024171482Sjeff * Penalize another thread for the time spent on this one. This helps to 2025171482Sjeff * worsen the priority and interactivity of processes which schedule batch 2026171482Sjeff * jobs such as make. This has little effect on the make process itself but 2027171482Sjeff * causes new processes spawned by it to receive worse scores immediately. 2028171482Sjeff */ 2029113372Sjeffvoid 2030164939Sjuliansched_exit_thread(struct thread *td, struct thread *child) 2031164936Sjulian{ 2032165762Sjeff 2033187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit", 2034187357Sjeff "prio:td", child->td_priority); 2035165762Sjeff /* 2036165762Sjeff * Give the child's runtime to the parent without returning the 2037165762Sjeff * sleep time as a penalty to the parent. This causes shells that 2038165762Sjeff * launch expensive things to mark their children as expensive. 2039165762Sjeff */ 2040170293Sjeff thread_lock(td); 2041171482Sjeff td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2042164939Sjulian sched_interact_update(td); 2043165762Sjeff sched_priority(td); 2044170293Sjeff thread_unlock(td); 2045164936Sjulian} 2046164936Sjulian 2047177005Sjeffvoid 2048177005Sjeffsched_preempt(struct thread *td) 2049177005Sjeff{ 2050177005Sjeff struct tdq *tdq; 2051177005Sjeff 2052177005Sjeff thread_lock(td); 2053177005Sjeff tdq = TDQ_SELF(); 2054177005Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2055177005Sjeff tdq->tdq_ipipending = 0; 2056177005Sjeff if (td->td_priority > tdq->tdq_lowpri) { 2057178272Sjeff int flags; 2058178272Sjeff 2059178272Sjeff flags = SW_INVOL | SW_PREEMPT; 2060177005Sjeff if (td->td_critnest > 1) 2061177005Sjeff td->td_owepreempt = 1; 2062178272Sjeff else if (TD_IS_IDLETHREAD(td)) 2063178272Sjeff mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL); 2064177005Sjeff else 2065178272Sjeff mi_switch(flags | SWT_REMOTEPREEMPT, NULL); 2066177005Sjeff } 2067177005Sjeff thread_unlock(td); 2068177005Sjeff} 2069177005Sjeff 2070171482Sjeff/* 2071171482Sjeff * Fix priorities on return to user-space. Priorities may be elevated due 2072171482Sjeff * to static priorities in msleep() or similar. 2073171482Sjeff */ 2074164936Sjulianvoid 2075164936Sjuliansched_userret(struct thread *td) 2076164936Sjulian{ 2077164936Sjulian /* 2078164936Sjulian * XXX we cheat slightly on the locking here to avoid locking in 2079164936Sjulian * the usual case. Setting td_priority here is essentially an 2080164936Sjulian * incomplete workaround for not setting it properly elsewhere. 2081164936Sjulian * Now that some interrupt handlers are threads, not setting it 2082164936Sjulian * properly elsewhere can clobber it in the window between setting 2083164936Sjulian * it here and returning to user mode, so don't waste time setting 2084164936Sjulian * it perfectly here. 2085164936Sjulian */ 2086164936Sjulian KASSERT((td->td_flags & TDF_BORROWING) == 0, 2087164936Sjulian ("thread with borrowed priority returning to userland")); 2088164936Sjulian if (td->td_priority != td->td_user_pri) { 2089170293Sjeff thread_lock(td); 2090164936Sjulian td->td_priority = td->td_user_pri; 2091164936Sjulian td->td_base_pri = td->td_user_pri; 2092177005Sjeff tdq_setlowpri(TDQ_SELF(), td); 2093170293Sjeff thread_unlock(td); 2094164936Sjulian } 2095164936Sjulian} 2096164936Sjulian 2097171482Sjeff/* 2098171482Sjeff * Handle a stathz tick. This is really only relevant for timeshare 2099171482Sjeff * threads. 2100171482Sjeff */ 2101164936Sjulianvoid 2102121127Sjeffsched_clock(struct thread *td) 2103109864Sjeff{ 2104164936Sjulian struct tdq *tdq; 2105164936Sjulian struct td_sched *ts; 2106109864Sjeff 2107171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2108164936Sjulian tdq = TDQ_SELF(); 2109172409Sjeff#ifdef SMP 2110133427Sjeff /* 2111172409Sjeff * We run the long term load balancer infrequently on the first cpu. 2112172409Sjeff */ 2113172409Sjeff if (balance_tdq == tdq) { 2114172409Sjeff if (balance_ticks && --balance_ticks == 0) 2115172409Sjeff sched_balance(); 2116172409Sjeff } 2117172409Sjeff#endif 2118172409Sjeff /* 2119178277Sjeff * Save the old switch count so we have a record of the last ticks 2120178277Sjeff * activity. Initialize the new switch count based on our load. 2121178277Sjeff * If there is some activity seed it to reflect that. 2122178277Sjeff */ 2123178277Sjeff tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2124178471Sjeff tdq->tdq_switchcnt = tdq->tdq_load; 2125178277Sjeff /* 2126165766Sjeff * Advance the insert index once for each tick to ensure that all 2127165766Sjeff * threads get a chance to run. 2128133427Sjeff */ 2129165766Sjeff if (tdq->tdq_idx == tdq->tdq_ridx) { 2130165766Sjeff tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2131165766Sjeff if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2132165766Sjeff tdq->tdq_ridx = tdq->tdq_idx; 2133165766Sjeff } 2134165766Sjeff ts = td->td_sched; 2135175104Sjeff if (td->td_pri_class & PRI_FIFO_BIT) 2136113357Sjeff return; 2137175104Sjeff if (td->td_pri_class == PRI_TIMESHARE) { 2138175104Sjeff /* 2139175104Sjeff * We used a tick; charge it to the thread so 2140175104Sjeff * that we can compute our interactivity. 2141175104Sjeff */ 2142175104Sjeff td->td_sched->ts_runtime += tickincr; 2143175104Sjeff sched_interact_update(td); 2144177009Sjeff sched_priority(td); 2145175104Sjeff } 2146113357Sjeff /* 2147109864Sjeff * We used up one time slice. 2148109864Sjeff */ 2149164936Sjulian if (--ts->ts_slice > 0) 2150113357Sjeff return; 2151109864Sjeff /* 2152177009Sjeff * We're out of time, force a requeue at userret(). 2153109864Sjeff */ 2154177009Sjeff ts->ts_slice = sched_slice; 2155113357Sjeff td->td_flags |= TDF_NEEDRESCHED; 2156109864Sjeff} 2157109864Sjeff 2158171482Sjeff/* 2159171482Sjeff * Called once per hz tick. Used for cpu utilization information. This 2160171482Sjeff * is easier than trying to scale based on stathz. 2161171482Sjeff */ 2162171482Sjeffvoid 2163171482Sjeffsched_tick(void) 2164171482Sjeff{ 2165171482Sjeff struct td_sched *ts; 2166171482Sjeff 2167171482Sjeff ts = curthread->td_sched; 2168180607Sjeff /* 2169180607Sjeff * Ticks is updated asynchronously on a single cpu. Check here to 2170180607Sjeff * avoid incrementing ts_ticks multiple times in a single tick. 2171180607Sjeff */ 2172199764Sivoras if (ts->ts_incrtick == ticks) 2173180607Sjeff return; 2174171482Sjeff /* Adjust ticks for pctcpu */ 2175171482Sjeff ts->ts_ticks += 1 << SCHED_TICK_SHIFT; 2176171482Sjeff ts->ts_ltick = ticks; 2177199764Sivoras ts->ts_incrtick = ticks; 2178171482Sjeff /* 2179171482Sjeff * Update if we've exceeded our desired tick threshhold by over one 2180171482Sjeff * second. 2181171482Sjeff */ 2182171482Sjeff if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2183171482Sjeff sched_pctcpu_update(ts); 2184171482Sjeff} 2185171482Sjeff 2186171482Sjeff/* 2187171482Sjeff * Return whether the current CPU has runnable tasks. Used for in-kernel 2188171482Sjeff * cooperative idle threads. 2189171482Sjeff */ 2190109864Sjeffint 2191109864Sjeffsched_runnable(void) 2192109864Sjeff{ 2193164936Sjulian struct tdq *tdq; 2194115998Sjeff int load; 2195109864Sjeff 2196115998Sjeff load = 1; 2197115998Sjeff 2198164936Sjulian tdq = TDQ_SELF(); 2199121605Sjeff if ((curthread->td_flags & TDF_IDLETD) != 0) { 2200165620Sjeff if (tdq->tdq_load > 0) 2201121605Sjeff goto out; 2202121605Sjeff } else 2203165620Sjeff if (tdq->tdq_load - 1 > 0) 2204121605Sjeff goto out; 2205115998Sjeff load = 0; 2206115998Sjeffout: 2207115998Sjeff return (load); 2208109864Sjeff} 2209109864Sjeff 2210171482Sjeff/* 2211171482Sjeff * Choose the highest priority thread to run. The thread is removed from 2212171482Sjeff * the run-queue while running however the load remains. For SMP we set 2213171482Sjeff * the tdq in the global idle bitmask if it idles here. 2214171482Sjeff */ 2215166190Sjeffstruct thread * 2216109970Sjeffsched_choose(void) 2217109970Sjeff{ 2218177435Sjeff struct thread *td; 2219164936Sjulian struct tdq *tdq; 2220109970Sjeff 2221164936Sjulian tdq = TDQ_SELF(); 2222171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2223177435Sjeff td = tdq_choose(tdq); 2224177435Sjeff if (td) { 2225177435Sjeff td->td_sched->ts_ltick = ticks; 2226177435Sjeff tdq_runq_rem(tdq, td); 2227177903Sjeff tdq->tdq_lowpri = td->td_priority; 2228177435Sjeff return (td); 2229109864Sjeff } 2230177903Sjeff tdq->tdq_lowpri = PRI_MAX_IDLE; 2231176735Sjeff return (PCPU_GET(idlethread)); 2232109864Sjeff} 2233109864Sjeff 2234171482Sjeff/* 2235171482Sjeff * Set owepreempt if necessary. Preemption never happens directly in ULE, 2236171482Sjeff * we always request it once we exit a critical section. 2237171482Sjeff */ 2238171482Sjeffstatic inline void 2239171482Sjeffsched_setpreempt(struct thread *td) 2240166190Sjeff{ 2241166190Sjeff struct thread *ctd; 2242166190Sjeff int cpri; 2243166190Sjeff int pri; 2244166190Sjeff 2245177005Sjeff THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2246177005Sjeff 2247166190Sjeff ctd = curthread; 2248166190Sjeff pri = td->td_priority; 2249166190Sjeff cpri = ctd->td_priority; 2250177005Sjeff if (pri < cpri) 2251177005Sjeff ctd->td_flags |= TDF_NEEDRESCHED; 2252166190Sjeff if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2253171482Sjeff return; 2254177005Sjeff if (!sched_shouldpreempt(pri, cpri, 0)) 2255171482Sjeff return; 2256171482Sjeff ctd->td_owepreempt = 1; 2257166190Sjeff} 2258166190Sjeff 2259171482Sjeff/* 2260177009Sjeff * Add a thread to a thread queue. Select the appropriate runq and add the 2261177009Sjeff * thread to it. This is the internal function called when the tdq is 2262177009Sjeff * predetermined. 2263171482Sjeff */ 2264109864Sjeffvoid 2265171482Sjefftdq_add(struct tdq *tdq, struct thread *td, int flags) 2266109864Sjeff{ 2267109864Sjeff 2268171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2269166190Sjeff KASSERT((td->td_inhibitors == 0), 2270166190Sjeff ("sched_add: trying to run inhibited thread")); 2271166190Sjeff KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2272166190Sjeff ("sched_add: bad thread state")); 2273172207Sjeff KASSERT(td->td_flags & TDF_INMEM, 2274172207Sjeff ("sched_add: thread swapped out")); 2275171482Sjeff 2276171482Sjeff if (td->td_priority < tdq->tdq_lowpri) 2277171482Sjeff tdq->tdq_lowpri = td->td_priority; 2278177435Sjeff tdq_runq_add(tdq, td, flags); 2279177435Sjeff tdq_load_add(tdq, td); 2280171482Sjeff} 2281171482Sjeff 2282171482Sjeff/* 2283171482Sjeff * Select the target thread queue and add a thread to it. Request 2284171482Sjeff * preemption or IPI a remote processor if required. 2285171482Sjeff */ 2286171482Sjeffvoid 2287171482Sjeffsched_add(struct thread *td, int flags) 2288171482Sjeff{ 2289171482Sjeff struct tdq *tdq; 2290171482Sjeff#ifdef SMP 2291171482Sjeff int cpu; 2292171482Sjeff#endif 2293187357Sjeff 2294187357Sjeff KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 2295187357Sjeff "prio:%d", td->td_priority, KTR_ATTR_LINKED, 2296187357Sjeff sched_tdname(curthread)); 2297187357Sjeff KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 2298187357Sjeff KTR_ATTR_LINKED, sched_tdname(td)); 2299171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2300166108Sjeff /* 2301171482Sjeff * Recalculate the priority before we select the target cpu or 2302171482Sjeff * run-queue. 2303166108Sjeff */ 2304171482Sjeff if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2305171482Sjeff sched_priority(td); 2306171482Sjeff#ifdef SMP 2307171482Sjeff /* 2308171482Sjeff * Pick the destination cpu and if it isn't ours transfer to the 2309171482Sjeff * target cpu. 2310171482Sjeff */ 2311177435Sjeff cpu = sched_pickcpu(td, flags); 2312177435Sjeff tdq = sched_setcpu(td, cpu, flags); 2313171482Sjeff tdq_add(tdq, td, flags); 2314177009Sjeff if (cpu != PCPU_GET(cpuid)) { 2315177435Sjeff tdq_notify(tdq, td); 2316166108Sjeff return; 2317166108Sjeff } 2318171482Sjeff#else 2319171482Sjeff tdq = TDQ_SELF(); 2320171482Sjeff TDQ_LOCK(tdq); 2321171482Sjeff /* 2322171482Sjeff * Now that the thread is moving to the run-queue, set the lock 2323171482Sjeff * to the scheduler's lock. 2324171482Sjeff */ 2325171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2326171482Sjeff tdq_add(tdq, td, flags); 2327166108Sjeff#endif 2328171482Sjeff if (!(flags & SRQ_YIELDING)) 2329171482Sjeff sched_setpreempt(td); 2330109864Sjeff} 2331109864Sjeff 2332171482Sjeff/* 2333171482Sjeff * Remove a thread from a run-queue without running it. This is used 2334171482Sjeff * when we're stealing a thread from a remote queue. Otherwise all threads 2335171482Sjeff * exit by calling sched_exit_thread() and sched_throw() themselves. 2336171482Sjeff */ 2337109864Sjeffvoid 2338121127Sjeffsched_rem(struct thread *td) 2339109864Sjeff{ 2340164936Sjulian struct tdq *tdq; 2341113357Sjeff 2342187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 2343187357Sjeff "prio:%d", td->td_priority); 2344177435Sjeff tdq = TDQ_CPU(td->td_sched->ts_cpu); 2345171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2346171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2347166190Sjeff KASSERT(TD_ON_RUNQ(td), 2348164936Sjulian ("sched_rem: thread not on run queue")); 2349177435Sjeff tdq_runq_rem(tdq, td); 2350177435Sjeff tdq_load_rem(tdq, td); 2351166190Sjeff TD_SET_CAN_RUN(td); 2352176735Sjeff if (td->td_priority == tdq->tdq_lowpri) 2353176735Sjeff tdq_setlowpri(tdq, NULL); 2354109864Sjeff} 2355109864Sjeff 2356171482Sjeff/* 2357171482Sjeff * Fetch cpu utilization information. Updates on demand. 2358171482Sjeff */ 2359109864Sjefffixpt_t 2360121127Sjeffsched_pctcpu(struct thread *td) 2361109864Sjeff{ 2362109864Sjeff fixpt_t pctcpu; 2363164936Sjulian struct td_sched *ts; 2364109864Sjeff 2365109864Sjeff pctcpu = 0; 2366164936Sjulian ts = td->td_sched; 2367164936Sjulian if (ts == NULL) 2368121290Sjeff return (0); 2369109864Sjeff 2370208787Sjhb THREAD_LOCK_ASSERT(td, MA_OWNED); 2371164936Sjulian if (ts->ts_ticks) { 2372109864Sjeff int rtick; 2373109864Sjeff 2374165796Sjeff sched_pctcpu_update(ts); 2375109864Sjeff /* How many rtick per second ? */ 2376165762Sjeff rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2377165762Sjeff pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2378109864Sjeff } 2379109864Sjeff 2380109864Sjeff return (pctcpu); 2381109864Sjeff} 2382109864Sjeff 2383176735Sjeff/* 2384176735Sjeff * Enforce affinity settings for a thread. Called after adjustments to 2385176735Sjeff * cpumask. 2386176735Sjeff */ 2387176729Sjeffvoid 2388176729Sjeffsched_affinity(struct thread *td) 2389176729Sjeff{ 2390176735Sjeff#ifdef SMP 2391176735Sjeff struct td_sched *ts; 2392176735Sjeff int cpu; 2393176735Sjeff 2394176735Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2395176735Sjeff ts = td->td_sched; 2396176735Sjeff if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2397176735Sjeff return; 2398189787Sjeff if (TD_ON_RUNQ(td)) { 2399189787Sjeff sched_rem(td); 2400189787Sjeff sched_add(td, SRQ_BORING); 2401189787Sjeff return; 2402189787Sjeff } 2403176735Sjeff if (!TD_IS_RUNNING(td)) 2404176735Sjeff return; 2405176735Sjeff td->td_flags |= TDF_NEEDRESCHED; 2406176735Sjeff if (!THREAD_CAN_MIGRATE(td)) 2407176735Sjeff return; 2408176735Sjeff /* 2409176735Sjeff * Assign the new cpu and force a switch before returning to 2410176735Sjeff * userspace. If the target thread is not running locally send 2411176735Sjeff * an ipi to force the issue. 2412176735Sjeff */ 2413176735Sjeff cpu = ts->ts_cpu; 2414177435Sjeff ts->ts_cpu = sched_pickcpu(td, 0); 2415176735Sjeff if (cpu != PCPU_GET(cpuid)) 2416176735Sjeff ipi_selected(1 << cpu, IPI_PREEMPT); 2417176735Sjeff#endif 2418176729Sjeff} 2419176729Sjeff 2420171482Sjeff/* 2421171482Sjeff * Bind a thread to a target cpu. 2422171482Sjeff */ 2423122038Sjeffvoid 2424122038Sjeffsched_bind(struct thread *td, int cpu) 2425122038Sjeff{ 2426164936Sjulian struct td_sched *ts; 2427122038Sjeff 2428171713Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2429208391Sjhb KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 2430164936Sjulian ts = td->td_sched; 2431166137Sjeff if (ts->ts_flags & TSF_BOUND) 2432166152Sjeff sched_unbind(td); 2433164936Sjulian ts->ts_flags |= TSF_BOUND; 2434166137Sjeff sched_pin(); 2435123433Sjeff if (PCPU_GET(cpuid) == cpu) 2436122038Sjeff return; 2437166137Sjeff ts->ts_cpu = cpu; 2438122038Sjeff /* When we return from mi_switch we'll be on the correct cpu. */ 2439131527Sphk mi_switch(SW_VOL, NULL); 2440122038Sjeff} 2441122038Sjeff 2442171482Sjeff/* 2443171482Sjeff * Release a bound thread. 2444171482Sjeff */ 2445122038Sjeffvoid 2446122038Sjeffsched_unbind(struct thread *td) 2447122038Sjeff{ 2448165762Sjeff struct td_sched *ts; 2449165762Sjeff 2450170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2451208391Sjhb KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 2452165762Sjeff ts = td->td_sched; 2453166137Sjeff if ((ts->ts_flags & TSF_BOUND) == 0) 2454166137Sjeff return; 2455165762Sjeff ts->ts_flags &= ~TSF_BOUND; 2456165762Sjeff sched_unpin(); 2457122038Sjeff} 2458122038Sjeff 2459109864Sjeffint 2460145256Sjkoshysched_is_bound(struct thread *td) 2461145256Sjkoshy{ 2462170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2463164936Sjulian return (td->td_sched->ts_flags & TSF_BOUND); 2464145256Sjkoshy} 2465145256Sjkoshy 2466171482Sjeff/* 2467171482Sjeff * Basic yield call. 2468171482Sjeff */ 2469159630Sdavidxuvoid 2470159630Sdavidxusched_relinquish(struct thread *td) 2471159630Sdavidxu{ 2472170293Sjeff thread_lock(td); 2473178272Sjeff mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 2474170293Sjeff thread_unlock(td); 2475159630Sdavidxu} 2476159630Sdavidxu 2477171482Sjeff/* 2478171482Sjeff * Return the total system load. 2479171482Sjeff */ 2480145256Sjkoshyint 2481125289Sjeffsched_load(void) 2482125289Sjeff{ 2483125289Sjeff#ifdef SMP 2484125289Sjeff int total; 2485125289Sjeff int i; 2486125289Sjeff 2487125289Sjeff total = 0; 2488176735Sjeff for (i = 0; i <= mp_maxid; i++) 2489176735Sjeff total += TDQ_CPU(i)->tdq_sysload; 2490125289Sjeff return (total); 2491125289Sjeff#else 2492165620Sjeff return (TDQ_SELF()->tdq_sysload); 2493125289Sjeff#endif 2494125289Sjeff} 2495125289Sjeff 2496125289Sjeffint 2497109864Sjeffsched_sizeof_proc(void) 2498109864Sjeff{ 2499109864Sjeff return (sizeof(struct proc)); 2500109864Sjeff} 2501109864Sjeff 2502109864Sjeffint 2503109864Sjeffsched_sizeof_thread(void) 2504109864Sjeff{ 2505109864Sjeff return (sizeof(struct thread) + sizeof(struct td_sched)); 2506109864Sjeff} 2507159570Sdavidxu 2508191676Sjeff#ifdef SMP 2509191676Sjeff#define TDQ_IDLESPIN(tdq) \ 2510191676Sjeff ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0) 2511191676Sjeff#else 2512191676Sjeff#define TDQ_IDLESPIN(tdq) 1 2513191676Sjeff#endif 2514191676Sjeff 2515166190Sjeff/* 2516166190Sjeff * The actual idle process. 2517166190Sjeff */ 2518166190Sjeffvoid 2519166190Sjeffsched_idletd(void *dummy) 2520166190Sjeff{ 2521166190Sjeff struct thread *td; 2522171482Sjeff struct tdq *tdq; 2523178277Sjeff int switchcnt; 2524178277Sjeff int i; 2525166190Sjeff 2526191643Sjeff mtx_assert(&Giant, MA_NOTOWNED); 2527166190Sjeff td = curthread; 2528171482Sjeff tdq = TDQ_SELF(); 2529171482Sjeff for (;;) { 2530171482Sjeff#ifdef SMP 2531178277Sjeff if (tdq_idled(tdq) == 0) 2532178277Sjeff continue; 2533171482Sjeff#endif 2534178277Sjeff switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2535178277Sjeff /* 2536178277Sjeff * If we're switching very frequently, spin while checking 2537178277Sjeff * for load rather than entering a low power state that 2538191643Sjeff * may require an IPI. However, don't do any busy 2539191643Sjeff * loops while on SMT machines as this simply steals 2540191643Sjeff * cycles from cores doing useful work. 2541178277Sjeff */ 2542191676Sjeff if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) { 2543178277Sjeff for (i = 0; i < sched_idlespins; i++) { 2544178277Sjeff if (tdq->tdq_load) 2545178277Sjeff break; 2546178277Sjeff cpu_spinwait(); 2547178277Sjeff } 2548178277Sjeff } 2549191643Sjeff switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2550191643Sjeff if (tdq->tdq_load == 0) 2551191643Sjeff cpu_idle(switchcnt > 1); 2552178277Sjeff if (tdq->tdq_load) { 2553178277Sjeff thread_lock(td); 2554178277Sjeff mi_switch(SW_VOL | SWT_IDLE, NULL); 2555178277Sjeff thread_unlock(td); 2556178277Sjeff } 2557171482Sjeff } 2558166190Sjeff} 2559166190Sjeff 2560170293Sjeff/* 2561170293Sjeff * A CPU is entering for the first time or a thread is exiting. 2562170293Sjeff */ 2563170293Sjeffvoid 2564170293Sjeffsched_throw(struct thread *td) 2565170293Sjeff{ 2566172411Sjeff struct thread *newtd; 2567171482Sjeff struct tdq *tdq; 2568171482Sjeff 2569171482Sjeff tdq = TDQ_SELF(); 2570170293Sjeff if (td == NULL) { 2571171482Sjeff /* Correct spinlock nesting and acquire the correct lock. */ 2572171482Sjeff TDQ_LOCK(tdq); 2573170293Sjeff spinlock_exit(); 2574170293Sjeff } else { 2575171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2576177435Sjeff tdq_load_rem(tdq, td); 2577174629Sjeff lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2578170293Sjeff } 2579170293Sjeff KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2580172411Sjeff newtd = choosethread(); 2581172411Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2582170293Sjeff PCPU_SET(switchtime, cpu_ticks()); 2583170293Sjeff PCPU_SET(switchticks, ticks); 2584172411Sjeff cpu_throw(td, newtd); /* doesn't return */ 2585170293Sjeff} 2586170293Sjeff 2587171482Sjeff/* 2588171482Sjeff * This is called from fork_exit(). Just acquire the correct locks and 2589171482Sjeff * let fork do the rest of the work. 2590171482Sjeff */ 2591170293Sjeffvoid 2592170600Sjeffsched_fork_exit(struct thread *td) 2593170293Sjeff{ 2594171482Sjeff struct td_sched *ts; 2595171482Sjeff struct tdq *tdq; 2596171482Sjeff int cpuid; 2597170293Sjeff 2598170293Sjeff /* 2599170293Sjeff * Finish setting up thread glue so that it begins execution in a 2600171482Sjeff * non-nested critical section with the scheduler lock held. 2601170293Sjeff */ 2602171482Sjeff cpuid = PCPU_GET(cpuid); 2603171482Sjeff tdq = TDQ_CPU(cpuid); 2604171482Sjeff ts = td->td_sched; 2605171482Sjeff if (TD_IS_IDLETHREAD(td)) 2606171482Sjeff td->td_lock = TDQ_LOCKPTR(tdq); 2607171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2608171482Sjeff td->td_oncpu = cpuid; 2609172411Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2610174629Sjeff lock_profile_obtain_lock_success( 2611174629Sjeff &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2612170293Sjeff} 2613170293Sjeff 2614187357Sjeff/* 2615187357Sjeff * Create on first use to catch odd startup conditons. 2616187357Sjeff */ 2617187357Sjeffchar * 2618187357Sjeffsched_tdname(struct thread *td) 2619187357Sjeff{ 2620187357Sjeff#ifdef KTR 2621187357Sjeff struct td_sched *ts; 2622187357Sjeff 2623187357Sjeff ts = td->td_sched; 2624187357Sjeff if (ts->ts_name[0] == '\0') 2625187357Sjeff snprintf(ts->ts_name, sizeof(ts->ts_name), 2626187357Sjeff "%s tid %d", td->td_name, td->td_tid); 2627187357Sjeff return (ts->ts_name); 2628187357Sjeff#else 2629187357Sjeff return (td->td_name); 2630187357Sjeff#endif 2631187357Sjeff} 2632187357Sjeff 2633184439Sivoras#ifdef SMP 2634184439Sivoras 2635184439Sivoras/* 2636184439Sivoras * Build the CPU topology dump string. Is recursively called to collect 2637184439Sivoras * the topology tree. 2638184439Sivoras */ 2639184439Sivorasstatic int 2640184439Sivorassysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg, 2641184439Sivoras int indent) 2642184439Sivoras{ 2643184439Sivoras int i, first; 2644184439Sivoras 2645184439Sivoras sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent, 2646184439Sivoras "", indent, cg->cg_level); 2647184439Sivoras sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"0x%x\">", indent, "", 2648184439Sivoras cg->cg_count, cg->cg_mask); 2649184439Sivoras first = TRUE; 2650184439Sivoras for (i = 0; i < MAXCPU; i++) { 2651184439Sivoras if ((cg->cg_mask & (1 << i)) != 0) { 2652184439Sivoras if (!first) 2653184439Sivoras sbuf_printf(sb, ", "); 2654184439Sivoras else 2655184439Sivoras first = FALSE; 2656184439Sivoras sbuf_printf(sb, "%d", i); 2657184439Sivoras } 2658184439Sivoras } 2659184439Sivoras sbuf_printf(sb, "</cpu>\n"); 2660184439Sivoras 2661184439Sivoras sbuf_printf(sb, "%*s <flags>", indent, ""); 2662184439Sivoras if (cg->cg_flags != 0) { 2663184439Sivoras if ((cg->cg_flags & CG_FLAG_HTT) != 0) 2664208982Sivoras sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>"); 2665191643Sjeff if ((cg->cg_flags & CG_FLAG_SMT) != 0) 2666208982Sivoras sbuf_printf(sb, "<flag name=\"THREAD\">SMT group</flag>"); 2667184439Sivoras } 2668184439Sivoras sbuf_printf(sb, "</flags>\n"); 2669184439Sivoras 2670184439Sivoras if (cg->cg_children > 0) { 2671184439Sivoras sbuf_printf(sb, "%*s <children>\n", indent, ""); 2672184439Sivoras for (i = 0; i < cg->cg_children; i++) 2673184439Sivoras sysctl_kern_sched_topology_spec_internal(sb, 2674184439Sivoras &cg->cg_child[i], indent+2); 2675184439Sivoras sbuf_printf(sb, "%*s </children>\n", indent, ""); 2676184439Sivoras } 2677184439Sivoras sbuf_printf(sb, "%*s</group>\n", indent, ""); 2678184439Sivoras return (0); 2679184439Sivoras} 2680184439Sivoras 2681184439Sivoras/* 2682184439Sivoras * Sysctl handler for retrieving topology dump. It's a wrapper for 2683184439Sivoras * the recursive sysctl_kern_smp_topology_spec_internal(). 2684184439Sivoras */ 2685184439Sivorasstatic int 2686184439Sivorassysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS) 2687184439Sivoras{ 2688184439Sivoras struct sbuf *topo; 2689184439Sivoras int err; 2690184439Sivoras 2691184439Sivoras KASSERT(cpu_top != NULL, ("cpu_top isn't initialized")); 2692184439Sivoras 2693184570Sivoras topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND); 2694184439Sivoras if (topo == NULL) 2695184439Sivoras return (ENOMEM); 2696184439Sivoras 2697184439Sivoras sbuf_printf(topo, "<groups>\n"); 2698184439Sivoras err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1); 2699184439Sivoras sbuf_printf(topo, "</groups>\n"); 2700184439Sivoras 2701184439Sivoras if (err == 0) { 2702184439Sivoras sbuf_finish(topo); 2703184439Sivoras err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo)); 2704184439Sivoras } 2705184439Sivoras sbuf_delete(topo); 2706184439Sivoras return (err); 2707184439Sivoras} 2708184439Sivoras#endif 2709184439Sivoras 2710177435SjeffSYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2711171482SjeffSYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2712165762Sjeff "Scheduler name"); 2713171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2714171482Sjeff "Slice size for timeshare threads"); 2715171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2716171482Sjeff "Interactivity score threshold"); 2717171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2718171482Sjeff 0,"Min priority for preemption, lower priorities have greater precedence"); 2719177085SjeffSYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 2720177085Sjeff 0,"Controls whether static kernel priorities are assigned to sleeping threads."); 2721178277SjeffSYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 2722178277Sjeff 0,"Number of times idle will spin waiting for new work."); 2723178277SjeffSYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, &sched_idlespinthresh, 2724178277Sjeff 0,"Threshold before we will permit idle spinning."); 2725166108Sjeff#ifdef SMP 2726171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2727171482Sjeff "Number of hz ticks to keep thread affinity for"); 2728171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2729171482Sjeff "Enables the long-term load balancer"); 2730172409SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2731172409Sjeff &balance_interval, 0, 2732172409Sjeff "Average frequency in stathz ticks to run the long-term balancer"); 2733171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, 2734171482Sjeff "Steals work from another hyper-threaded core on idle"); 2735171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2736171482Sjeff "Attempts to steal work from other cores before idling"); 2737171506SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2738171506Sjeff "Minimum load on remote cpu before we'll steal"); 2739184439Sivoras 2740184439Sivoras/* Retrieve SMP topology */ 2741184439SivorasSYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING | 2742184439Sivoras CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A", 2743184439Sivoras "XML dump of detected CPU topology"); 2744166108Sjeff#endif 2745165762Sjeff 2746172264Sjeff/* ps compat. All cpu percentages from ULE are weighted. */ 2747172293Sjeffstatic int ccpu = 0; 2748165762SjeffSYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2749