sched_ule.c revision 174629
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> 39116182Sobrien__FBSDID("$FreeBSD: head/sys/kern/sched_ule.c 174629 2007-12-15 23:13:31Z jeff $"); 40116182Sobrien 41147565Speter#include "opt_hwpmc_hooks.h" 42147565Speter#include "opt_sched.h" 43134649Sscottl 44109864Sjeff#include <sys/param.h> 45109864Sjeff#include <sys/systm.h> 46131929Smarcel#include <sys/kdb.h> 47109864Sjeff#include <sys/kernel.h> 48109864Sjeff#include <sys/ktr.h> 49109864Sjeff#include <sys/lock.h> 50109864Sjeff#include <sys/mutex.h> 51109864Sjeff#include <sys/proc.h> 52112966Sjeff#include <sys/resource.h> 53122038Sjeff#include <sys/resourcevar.h> 54109864Sjeff#include <sys/sched.h> 55109864Sjeff#include <sys/smp.h> 56109864Sjeff#include <sys/sx.h> 57109864Sjeff#include <sys/sysctl.h> 58109864Sjeff#include <sys/sysproto.h> 59139453Sjhb#include <sys/turnstile.h> 60161599Sdavidxu#include <sys/umtx.h> 61109864Sjeff#include <sys/vmmeter.h> 62109864Sjeff#ifdef KTRACE 63109864Sjeff#include <sys/uio.h> 64109864Sjeff#include <sys/ktrace.h> 65109864Sjeff#endif 66109864Sjeff 67145256Sjkoshy#ifdef HWPMC_HOOKS 68145256Sjkoshy#include <sys/pmckern.h> 69145256Sjkoshy#endif 70145256Sjkoshy 71109864Sjeff#include <machine/cpu.h> 72121790Sjeff#include <machine/smp.h> 73109864Sjeff 74172887Sgrehan#if !defined(__i386__) && !defined(__amd64__) && !defined(__powerpc__) && !defined(__arm__) 75172345Sjeff#error "This architecture is not currently compatible with ULE" 76166190Sjeff#endif 77166190Sjeff 78171482Sjeff#define KTR_ULE 0 79166137Sjeff 80166137Sjeff/* 81171482Sjeff * Thread scheduler specific section. All fields are protected 82171482Sjeff * by the thread lock. 83146954Sjeff */ 84164936Sjulianstruct td_sched { 85171482Sjeff TAILQ_ENTRY(td_sched) ts_procq; /* Run queue. */ 86171482Sjeff struct thread *ts_thread; /* Active associated thread. */ 87171482Sjeff struct runq *ts_runq; /* Run-queue we're queued on. */ 88171482Sjeff short ts_flags; /* TSF_* flags. */ 89171482Sjeff u_char ts_rqindex; /* Run queue index. */ 90164936Sjulian u_char ts_cpu; /* CPU that we have affinity for. */ 91171482Sjeff int ts_slice; /* Ticks of slice remaining. */ 92171482Sjeff u_int ts_slptime; /* Number of ticks we vol. slept */ 93171482Sjeff u_int ts_runtime; /* Number of ticks we were running */ 94134791Sjulian /* The following variables are only used for pctcpu calculation */ 95164936Sjulian int ts_ltick; /* Last tick that we were running on */ 96164936Sjulian int ts_ftick; /* First tick that we were running on */ 97164936Sjulian int ts_ticks; /* Tick count */ 98166108Sjeff#ifdef SMP 99166108Sjeff int ts_rltick; /* Real last tick, for affinity. */ 100166108Sjeff#endif 101134791Sjulian}; 102164936Sjulian/* flags kept in ts_flags */ 103166108Sjeff#define TSF_BOUND 0x0001 /* Thread can not migrate. */ 104166108Sjeff#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 105121790Sjeff 106164936Sjulianstatic struct td_sched td_sched0; 107109864Sjeff 108109864Sjeff/* 109165762Sjeff * Cpu percentage computation macros and defines. 110111857Sjeff * 111165762Sjeff * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 112165762Sjeff * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 113165796Sjeff * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 114165762Sjeff * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 115165762Sjeff * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 116165762Sjeff * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 117165762Sjeff */ 118165762Sjeff#define SCHED_TICK_SECS 10 119165762Sjeff#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 120165796Sjeff#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 121165762Sjeff#define SCHED_TICK_SHIFT 10 122165762Sjeff#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 123165830Sjeff#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 124165762Sjeff 125165762Sjeff/* 126165762Sjeff * These macros determine priorities for non-interactive threads. They are 127165762Sjeff * assigned a priority based on their recent cpu utilization as expressed 128165762Sjeff * by the ratio of ticks to the tick total. NHALF priorities at the start 129165762Sjeff * and end of the MIN to MAX timeshare range are only reachable with negative 130165762Sjeff * or positive nice respectively. 131165762Sjeff * 132165762Sjeff * PRI_RANGE: Priority range for utilization dependent priorities. 133116642Sjeff * PRI_NRESV: Number of nice values. 134165762Sjeff * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 135165762Sjeff * PRI_NICE: Determines the part of the priority inherited from nice. 136109864Sjeff */ 137165762Sjeff#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 138121869Sjeff#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 139165762Sjeff#define SCHED_PRI_MIN (PRI_MIN_TIMESHARE + SCHED_PRI_NHALF) 140165762Sjeff#define SCHED_PRI_MAX (PRI_MAX_TIMESHARE - SCHED_PRI_NHALF) 141170787Sjeff#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN) 142165762Sjeff#define SCHED_PRI_TICKS(ts) \ 143165762Sjeff (SCHED_TICK_HZ((ts)) / \ 144165827Sjeff (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 145165762Sjeff#define SCHED_PRI_NICE(nice) (nice) 146109864Sjeff 147109864Sjeff/* 148165762Sjeff * These determine the interactivity of a process. Interactivity differs from 149165762Sjeff * cpu utilization in that it expresses the voluntary time slept vs time ran 150165762Sjeff * while cpu utilization includes all time not running. This more accurately 151165762Sjeff * models the intent of the thread. 152109864Sjeff * 153110645Sjeff * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 154110645Sjeff * before throttling back. 155121868Sjeff * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 156116365Sjeff * INTERACT_MAX: Maximum interactivity value. Smaller is better. 157111857Sjeff * INTERACT_THRESH: Threshhold for placement on the current runq. 158109864Sjeff */ 159165762Sjeff#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 160165762Sjeff#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 161116365Sjeff#define SCHED_INTERACT_MAX (100) 162116365Sjeff#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 163121126Sjeff#define SCHED_INTERACT_THRESH (30) 164111857Sjeff 165109864Sjeff/* 166165762Sjeff * tickincr: Converts a stathz tick into a hz domain scaled by 167165762Sjeff * the shift factor. Without the shift the error rate 168165762Sjeff * due to rounding would be unacceptably high. 169165762Sjeff * realstathz: stathz is sometimes 0 and run off of hz. 170165762Sjeff * sched_slice: Runtime of each thread before rescheduling. 171171482Sjeff * preempt_thresh: Priority threshold for preemption and remote IPIs. 172109864Sjeff */ 173165762Sjeffstatic int sched_interact = SCHED_INTERACT_THRESH; 174165762Sjeffstatic int realstathz; 175165762Sjeffstatic int tickincr; 176165762Sjeffstatic int sched_slice; 177172345Sjeff#ifdef PREEMPTION 178172345Sjeff#ifdef FULL_PREEMPTION 179172345Sjeffstatic int preempt_thresh = PRI_MAX_IDLE; 180172345Sjeff#else 181171482Sjeffstatic int preempt_thresh = PRI_MIN_KERN; 182172345Sjeff#endif 183172345Sjeff#else 184172345Sjeffstatic int preempt_thresh = 0; 185172345Sjeff#endif 186109864Sjeff 187109864Sjeff/* 188171482Sjeff * tdq - per processor runqs and statistics. All fields are protected by the 189171482Sjeff * tdq_lock. The load and lowpri may be accessed without to avoid excess 190171482Sjeff * locking in sched_pickcpu(); 191109864Sjeff */ 192164936Sjulianstruct tdq { 193171713Sjeff struct mtx *tdq_lock; /* Pointer to group lock. */ 194171482Sjeff struct runq tdq_realtime; /* real-time run queue. */ 195171482Sjeff struct runq tdq_timeshare; /* timeshare run queue. */ 196165620Sjeff struct runq tdq_idle; /* Queue of IDLE threads. */ 197171482Sjeff int tdq_load; /* Aggregate load. */ 198166557Sjeff u_char tdq_idx; /* Current insert index. */ 199166557Sjeff u_char tdq_ridx; /* Current removal index. */ 200110267Sjeff#ifdef SMP 201171482Sjeff u_char tdq_lowpri; /* Lowest priority thread. */ 202171482Sjeff int tdq_transferable; /* Transferable thread count. */ 203165620Sjeff LIST_ENTRY(tdq) tdq_siblings; /* Next in tdq group. */ 204165620Sjeff struct tdq_group *tdq_group; /* Our processor group. */ 205125289Sjeff#else 206165620Sjeff int tdq_sysload; /* For loadavg, !ITHD load. */ 207110267Sjeff#endif 208171482Sjeff} __aligned(64); 209109864Sjeff 210166108Sjeff 211123433Sjeff#ifdef SMP 212109864Sjeff/* 213164936Sjulian * tdq groups are groups of processors which can cheaply share threads. When 214123433Sjeff * one processor in the group goes idle it will check the runqs of the other 215123433Sjeff * processors in its group prior to halting and waiting for an interrupt. 216123433Sjeff * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA. 217123433Sjeff * In a numa environment we'd want an idle bitmap per group and a two tiered 218123433Sjeff * load balancer. 219123433Sjeff */ 220164936Sjulianstruct tdq_group { 221171713Sjeff struct mtx tdg_lock; /* Protects all fields below. */ 222171713Sjeff int tdg_cpus; /* Count of CPUs in this tdq group. */ 223171713Sjeff cpumask_t tdg_cpumask; /* Mask of cpus in this group. */ 224171713Sjeff cpumask_t tdg_idlemask; /* Idle cpus in this group. */ 225171713Sjeff cpumask_t tdg_mask; /* Bit mask for first cpu. */ 226171713Sjeff int tdg_load; /* Total load of this group. */ 227165620Sjeff int tdg_transferable; /* Transferable load of this group. */ 228165620Sjeff LIST_HEAD(, tdq) tdg_members; /* Linked list of all members. */ 229171713Sjeff char tdg_name[16]; /* lock name. */ 230171482Sjeff} __aligned(64); 231123433Sjeff 232171482Sjeff#define SCHED_AFFINITY_DEFAULT (max(1, hz / 300)) 233166108Sjeff#define SCHED_AFFINITY(ts) ((ts)->ts_rltick > ticks - affinity) 234166108Sjeff 235123433Sjeff/* 236166108Sjeff * Run-time tunables. 237166108Sjeff */ 238171506Sjeffstatic int rebalance = 1; 239172409Sjeffstatic int balance_interval = 128; /* Default set in sched_initticks(). */ 240171506Sjeffstatic int pick_pri = 1; 241166108Sjeffstatic int affinity; 242166108Sjeffstatic int tryself = 1; 243172409Sjeffstatic int steal_htt = 1; 244171506Sjeffstatic int steal_idle = 1; 245171506Sjeffstatic int steal_thresh = 2; 246170293Sjeffstatic int topology = 0; 247166108Sjeff 248166108Sjeff/* 249165620Sjeff * One thread queue per processor. 250109864Sjeff */ 251166108Sjeffstatic volatile cpumask_t tdq_idle; 252165620Sjeffstatic int tdg_maxid; 253164936Sjulianstatic struct tdq tdq_cpu[MAXCPU]; 254164936Sjulianstatic struct tdq_group tdq_groups[MAXCPU]; 255172409Sjeffstatic struct tdq *balance_tdq; 256172409Sjeffstatic int balance_group_ticks; 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)) 262164936Sjulian#define TDQ_GROUP(x) (&tdq_groups[(x)]) 263171713Sjeff#define TDG_ID(x) ((int)((x) - tdq_groups)) 264123433Sjeff#else /* !SMP */ 265164936Sjulianstatic struct tdq tdq_cpu; 266171713Sjeffstatic struct mtx tdq_lock; 267129982Sjeff 268170315Sjeff#define TDQ_ID(x) (0) 269164936Sjulian#define TDQ_SELF() (&tdq_cpu) 270164936Sjulian#define TDQ_CPU(x) (&tdq_cpu) 271110028Sjeff#endif 272109864Sjeff 273171482Sjeff#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 274171482Sjeff#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 275171482Sjeff#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 276171482Sjeff#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 277171713Sjeff#define TDQ_LOCKPTR(t) ((t)->tdq_lock) 278171482Sjeff 279163709Sjbstatic void sched_priority(struct thread *); 280146954Sjeffstatic void sched_thread_priority(struct thread *, u_char); 281163709Sjbstatic int sched_interact_score(struct thread *); 282163709Sjbstatic void sched_interact_update(struct thread *); 283163709Sjbstatic void sched_interact_fork(struct thread *); 284164936Sjulianstatic void sched_pctcpu_update(struct td_sched *); 285109864Sjeff 286110267Sjeff/* Operations on per processor queues */ 287164936Sjulianstatic struct td_sched * tdq_choose(struct tdq *); 288164936Sjulianstatic void tdq_setup(struct tdq *); 289164936Sjulianstatic void tdq_load_add(struct tdq *, struct td_sched *); 290164936Sjulianstatic void tdq_load_rem(struct tdq *, struct td_sched *); 291164936Sjulianstatic __inline void tdq_runq_add(struct tdq *, struct td_sched *, int); 292164936Sjulianstatic __inline void tdq_runq_rem(struct tdq *, struct td_sched *); 293164936Sjulianvoid tdq_print(int cpu); 294165762Sjeffstatic void runq_print(struct runq *rq); 295171482Sjeffstatic void tdq_add(struct tdq *, struct thread *, int); 296110267Sjeff#ifdef SMP 297171482Sjeffstatic void tdq_move(struct tdq *, struct tdq *); 298171482Sjeffstatic int tdq_idled(struct tdq *); 299171482Sjeffstatic void tdq_notify(struct td_sched *); 300172409Sjeffstatic struct td_sched *tdq_steal(struct tdq *); 301164936Sjulianstatic struct td_sched *runq_steal(struct runq *); 302171482Sjeffstatic int sched_pickcpu(struct td_sched *, int); 303172409Sjeffstatic void sched_balance(void); 304172409Sjeffstatic void sched_balance_groups(void); 305164936Sjulianstatic void sched_balance_group(struct tdq_group *); 306164936Sjulianstatic void sched_balance_pair(struct tdq *, struct tdq *); 307171482Sjeffstatic inline struct tdq *sched_setcpu(struct td_sched *, int, int); 308171482Sjeffstatic inline struct mtx *thread_block_switch(struct thread *); 309171482Sjeffstatic inline void thread_unblock_switch(struct thread *, struct mtx *); 310171713Sjeffstatic struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 311165827Sjeff 312166108Sjeff#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 313121790Sjeff#endif 314110028Sjeff 315165762Sjeffstatic void sched_setup(void *dummy); 316165762SjeffSYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL) 317165762Sjeff 318165762Sjeffstatic void sched_initticks(void *dummy); 319165762SjeffSYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, NULL) 320165762Sjeff 321171482Sjeff/* 322171482Sjeff * Print the threads waiting on a run-queue. 323171482Sjeff */ 324165762Sjeffstatic void 325165762Sjeffrunq_print(struct runq *rq) 326165762Sjeff{ 327165762Sjeff struct rqhead *rqh; 328165762Sjeff struct td_sched *ts; 329165762Sjeff int pri; 330165762Sjeff int j; 331165762Sjeff int i; 332165762Sjeff 333165762Sjeff for (i = 0; i < RQB_LEN; i++) { 334165762Sjeff printf("\t\trunq bits %d 0x%zx\n", 335165762Sjeff i, rq->rq_status.rqb_bits[i]); 336165762Sjeff for (j = 0; j < RQB_BPW; j++) 337165762Sjeff if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 338165762Sjeff pri = j + (i << RQB_L2BPW); 339165762Sjeff rqh = &rq->rq_queues[pri]; 340165762Sjeff TAILQ_FOREACH(ts, rqh, ts_procq) { 341165762Sjeff printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 342173600Sjulian ts->ts_thread, ts->ts_thread->td_name, ts->ts_thread->td_priority, ts->ts_rqindex, pri); 343165762Sjeff } 344165762Sjeff } 345165762Sjeff } 346165762Sjeff} 347165762Sjeff 348171482Sjeff/* 349171482Sjeff * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 350171482Sjeff */ 351113357Sjeffvoid 352164936Sjuliantdq_print(int cpu) 353110267Sjeff{ 354164936Sjulian struct tdq *tdq; 355112994Sjeff 356164936Sjulian tdq = TDQ_CPU(cpu); 357112994Sjeff 358171713Sjeff printf("tdq %d:\n", TDQ_ID(tdq)); 359171482Sjeff printf("\tlockptr %p\n", TDQ_LOCKPTR(tdq)); 360165620Sjeff printf("\tload: %d\n", tdq->tdq_load); 361171482Sjeff printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 362165766Sjeff printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 363165762Sjeff printf("\trealtime runq:\n"); 364165762Sjeff runq_print(&tdq->tdq_realtime); 365165762Sjeff printf("\ttimeshare runq:\n"); 366165762Sjeff runq_print(&tdq->tdq_timeshare); 367165762Sjeff printf("\tidle runq:\n"); 368165762Sjeff runq_print(&tdq->tdq_idle); 369121896Sjeff#ifdef SMP 370165620Sjeff printf("\tload transferable: %d\n", tdq->tdq_transferable); 371171713Sjeff printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 372171713Sjeff printf("\tgroup: %d\n", TDG_ID(tdq->tdq_group)); 373171713Sjeff printf("\tLock name: %s\n", tdq->tdq_group->tdg_name); 374121896Sjeff#endif 375113357Sjeff} 376112994Sjeff 377171482Sjeff#define TS_RQ_PPQ (((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS) 378171482Sjeff/* 379171482Sjeff * Add a thread to the actual run-queue. Keeps transferable counts up to 380171482Sjeff * date with what is actually on the run-queue. Selects the correct 381171482Sjeff * queue position for timeshare threads. 382171482Sjeff */ 383122744Sjeffstatic __inline void 384164936Sjuliantdq_runq_add(struct tdq *tdq, struct td_sched *ts, int flags) 385122744Sjeff{ 386171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 387171482Sjeff THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 388122744Sjeff#ifdef SMP 389165762Sjeff if (THREAD_CAN_MIGRATE(ts->ts_thread)) { 390165620Sjeff tdq->tdq_transferable++; 391165620Sjeff tdq->tdq_group->tdg_transferable++; 392164936Sjulian ts->ts_flags |= TSF_XFERABLE; 393123433Sjeff } 394122744Sjeff#endif 395165762Sjeff if (ts->ts_runq == &tdq->tdq_timeshare) { 396166557Sjeff u_char pri; 397165762Sjeff 398165762Sjeff pri = ts->ts_thread->td_priority; 399165762Sjeff KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE, 400165762Sjeff ("Invalid priority %d on timeshare runq", pri)); 401165762Sjeff /* 402165762Sjeff * This queue contains only priorities between MIN and MAX 403165762Sjeff * realtime. Use the whole queue to represent these values. 404165762Sjeff */ 405171713Sjeff if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 406165762Sjeff pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ; 407165762Sjeff pri = (pri + tdq->tdq_idx) % RQ_NQS; 408165766Sjeff /* 409165766Sjeff * This effectively shortens the queue by one so we 410165766Sjeff * can have a one slot difference between idx and 411165766Sjeff * ridx while we wait for threads to drain. 412165766Sjeff */ 413165766Sjeff if (tdq->tdq_ridx != tdq->tdq_idx && 414165766Sjeff pri == tdq->tdq_ridx) 415167664Sjeff pri = (unsigned char)(pri - 1) % RQ_NQS; 416165762Sjeff } else 417165766Sjeff pri = tdq->tdq_ridx; 418165762Sjeff runq_add_pri(ts->ts_runq, ts, pri, flags); 419165762Sjeff } else 420165762Sjeff runq_add(ts->ts_runq, ts, flags); 421122744Sjeff} 422122744Sjeff 423171482Sjeff/* 424171482Sjeff * Remove a thread from a run-queue. This typically happens when a thread 425171482Sjeff * is selected to run. Running threads are not on the queue and the 426171482Sjeff * transferable count does not reflect them. 427171482Sjeff */ 428122744Sjeffstatic __inline void 429164936Sjuliantdq_runq_rem(struct tdq *tdq, struct td_sched *ts) 430122744Sjeff{ 431171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 432171482Sjeff KASSERT(ts->ts_runq != NULL, 433171482Sjeff ("tdq_runq_remove: thread %p null ts_runq", ts->ts_thread)); 434122744Sjeff#ifdef SMP 435164936Sjulian if (ts->ts_flags & TSF_XFERABLE) { 436165620Sjeff tdq->tdq_transferable--; 437165620Sjeff tdq->tdq_group->tdg_transferable--; 438164936Sjulian ts->ts_flags &= ~TSF_XFERABLE; 439123433Sjeff } 440122744Sjeff#endif 441165766Sjeff if (ts->ts_runq == &tdq->tdq_timeshare) { 442165766Sjeff if (tdq->tdq_idx != tdq->tdq_ridx) 443165766Sjeff runq_remove_idx(ts->ts_runq, ts, &tdq->tdq_ridx); 444165766Sjeff else 445165766Sjeff runq_remove_idx(ts->ts_runq, ts, NULL); 446165796Sjeff /* 447165796Sjeff * For timeshare threads we update the priority here so 448165796Sjeff * the priority reflects the time we've been sleeping. 449165796Sjeff */ 450165796Sjeff ts->ts_ltick = ticks; 451165796Sjeff sched_pctcpu_update(ts); 452165796Sjeff sched_priority(ts->ts_thread); 453165766Sjeff } else 454165762Sjeff runq_remove(ts->ts_runq, ts); 455122744Sjeff} 456122744Sjeff 457171482Sjeff/* 458171482Sjeff * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 459171482Sjeff * for this thread to the referenced thread queue. 460171482Sjeff */ 461113357Sjeffstatic void 462164936Sjuliantdq_load_add(struct tdq *tdq, struct td_sched *ts) 463113357Sjeff{ 464121896Sjeff int class; 465171482Sjeff 466171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 467171482Sjeff THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 468164936Sjulian class = PRI_BASE(ts->ts_thread->td_pri_class); 469165620Sjeff tdq->tdq_load++; 470171713Sjeff CTR2(KTR_SCHED, "cpu %d load: %d", TDQ_ID(tdq), tdq->tdq_load); 471166108Sjeff if (class != PRI_ITHD && 472166108Sjeff (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 473123487Sjeff#ifdef SMP 474165620Sjeff tdq->tdq_group->tdg_load++; 475125289Sjeff#else 476165620Sjeff tdq->tdq_sysload++; 477123487Sjeff#endif 478110267Sjeff} 479113357Sjeff 480171482Sjeff/* 481171482Sjeff * Remove the load from a thread that is transitioning to a sleep state or 482171482Sjeff * exiting. 483171482Sjeff */ 484112994Sjeffstatic void 485164936Sjuliantdq_load_rem(struct tdq *tdq, struct td_sched *ts) 486110267Sjeff{ 487121896Sjeff int class; 488171482Sjeff 489171482Sjeff THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 490171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 491164936Sjulian class = PRI_BASE(ts->ts_thread->td_pri_class); 492166108Sjeff if (class != PRI_ITHD && 493166108Sjeff (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 494123487Sjeff#ifdef SMP 495165620Sjeff tdq->tdq_group->tdg_load--; 496125289Sjeff#else 497165620Sjeff tdq->tdq_sysload--; 498123487Sjeff#endif 499171482Sjeff KASSERT(tdq->tdq_load != 0, 500171713Sjeff ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 501165620Sjeff tdq->tdq_load--; 502165620Sjeff CTR1(KTR_SCHED, "load: %d", tdq->tdq_load); 503164936Sjulian ts->ts_runq = NULL; 504110267Sjeff} 505110267Sjeff 506113357Sjeff#ifdef SMP 507116069Sjeff/* 508122744Sjeff * sched_balance is a simple CPU load balancing algorithm. It operates by 509116069Sjeff * finding the least loaded and most loaded cpu and equalizing their load 510116069Sjeff * by migrating some processes. 511116069Sjeff * 512116069Sjeff * Dealing only with two CPUs at a time has two advantages. Firstly, most 513116069Sjeff * installations will only have 2 cpus. Secondly, load balancing too much at 514116069Sjeff * once can have an unpleasant effect on the system. The scheduler rarely has 515116069Sjeff * enough information to make perfect decisions. So this algorithm chooses 516171482Sjeff * simplicity and more gradual effects on load in larger systems. 517116069Sjeff * 518116069Sjeff */ 519121790Sjeffstatic void 520172409Sjeffsched_balance() 521116069Sjeff{ 522164936Sjulian struct tdq_group *high; 523164936Sjulian struct tdq_group *low; 524165620Sjeff struct tdq_group *tdg; 525172409Sjeff struct tdq *tdq; 526123487Sjeff int cnt; 527123487Sjeff int i; 528123487Sjeff 529172409Sjeff /* 530172409Sjeff * Select a random time between .5 * balance_interval and 531172409Sjeff * 1.5 * balance_interval. 532172409Sjeff */ 533172409Sjeff balance_ticks = max(balance_interval / 2, 1); 534172409Sjeff balance_ticks += random() % balance_interval; 535171482Sjeff if (smp_started == 0 || rebalance == 0) 536139334Sjeff return; 537172409Sjeff tdq = TDQ_SELF(); 538172409Sjeff TDQ_UNLOCK(tdq); 539123487Sjeff low = high = NULL; 540165620Sjeff i = random() % (tdg_maxid + 1); 541165620Sjeff for (cnt = 0; cnt <= tdg_maxid; cnt++) { 542165620Sjeff tdg = TDQ_GROUP(i); 543123487Sjeff /* 544123487Sjeff * Find the CPU with the highest load that has some 545123487Sjeff * threads to transfer. 546123487Sjeff */ 547165620Sjeff if ((high == NULL || tdg->tdg_load > high->tdg_load) 548165620Sjeff && tdg->tdg_transferable) 549165620Sjeff high = tdg; 550165620Sjeff if (low == NULL || tdg->tdg_load < low->tdg_load) 551165620Sjeff low = tdg; 552165620Sjeff if (++i > tdg_maxid) 553123487Sjeff i = 0; 554123487Sjeff } 555123487Sjeff if (low != NULL && high != NULL && high != low) 556165620Sjeff sched_balance_pair(LIST_FIRST(&high->tdg_members), 557165620Sjeff LIST_FIRST(&low->tdg_members)); 558172409Sjeff TDQ_LOCK(tdq); 559123487Sjeff} 560123487Sjeff 561171482Sjeff/* 562171482Sjeff * Balance load between CPUs in a group. Will only migrate within the group. 563171482Sjeff */ 564123487Sjeffstatic void 565172409Sjeffsched_balance_groups() 566123487Sjeff{ 567172409Sjeff struct tdq *tdq; 568123487Sjeff int i; 569123487Sjeff 570172409Sjeff /* 571172409Sjeff * Select a random time between .5 * balance_interval and 572172409Sjeff * 1.5 * balance_interval. 573172409Sjeff */ 574172409Sjeff balance_group_ticks = max(balance_interval / 2, 1); 575172409Sjeff balance_group_ticks += random() % balance_interval; 576171482Sjeff if (smp_started == 0 || rebalance == 0) 577171482Sjeff return; 578172409Sjeff tdq = TDQ_SELF(); 579172409Sjeff TDQ_UNLOCK(tdq); 580171482Sjeff for (i = 0; i <= tdg_maxid; i++) 581171482Sjeff sched_balance_group(TDQ_GROUP(i)); 582172409Sjeff TDQ_LOCK(tdq); 583123487Sjeff} 584123487Sjeff 585171482Sjeff/* 586171482Sjeff * Finds the greatest imbalance between two tdqs in a group. 587171482Sjeff */ 588123487Sjeffstatic void 589165620Sjeffsched_balance_group(struct tdq_group *tdg) 590123487Sjeff{ 591164936Sjulian struct tdq *tdq; 592164936Sjulian struct tdq *high; 593164936Sjulian struct tdq *low; 594123487Sjeff int load; 595123487Sjeff 596165620Sjeff if (tdg->tdg_transferable == 0) 597123487Sjeff return; 598123487Sjeff low = NULL; 599123487Sjeff high = NULL; 600165620Sjeff LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) { 601165620Sjeff load = tdq->tdq_load; 602165620Sjeff if (high == NULL || load > high->tdq_load) 603164936Sjulian high = tdq; 604165620Sjeff if (low == NULL || load < low->tdq_load) 605164936Sjulian low = tdq; 606123487Sjeff } 607123487Sjeff if (high != NULL && low != NULL && high != low) 608123487Sjeff sched_balance_pair(high, low); 609123487Sjeff} 610123487Sjeff 611171482Sjeff/* 612171482Sjeff * Lock two thread queues using their address to maintain lock order. 613171482Sjeff */ 614123487Sjeffstatic void 615171482Sjefftdq_lock_pair(struct tdq *one, struct tdq *two) 616171482Sjeff{ 617171482Sjeff if (one < two) { 618171482Sjeff TDQ_LOCK(one); 619171482Sjeff TDQ_LOCK_FLAGS(two, MTX_DUPOK); 620171482Sjeff } else { 621171482Sjeff TDQ_LOCK(two); 622171482Sjeff TDQ_LOCK_FLAGS(one, MTX_DUPOK); 623171482Sjeff } 624171482Sjeff} 625171482Sjeff 626171482Sjeff/* 627172409Sjeff * Unlock two thread queues. Order is not important here. 628172409Sjeff */ 629172409Sjeffstatic void 630172409Sjefftdq_unlock_pair(struct tdq *one, struct tdq *two) 631172409Sjeff{ 632172409Sjeff TDQ_UNLOCK(one); 633172409Sjeff TDQ_UNLOCK(two); 634172409Sjeff} 635172409Sjeff 636172409Sjeff/* 637171482Sjeff * Transfer load between two imbalanced thread queues. 638171482Sjeff */ 639171482Sjeffstatic void 640164936Sjuliansched_balance_pair(struct tdq *high, struct tdq *low) 641123487Sjeff{ 642123433Sjeff int transferable; 643116069Sjeff int high_load; 644116069Sjeff int low_load; 645116069Sjeff int move; 646116069Sjeff int diff; 647116069Sjeff int i; 648116069Sjeff 649171482Sjeff tdq_lock_pair(high, low); 650116069Sjeff /* 651123433Sjeff * If we're transfering within a group we have to use this specific 652164936Sjulian * tdq's transferable count, otherwise we can steal from other members 653123433Sjeff * of the group. 654123433Sjeff */ 655165620Sjeff if (high->tdq_group == low->tdq_group) { 656165620Sjeff transferable = high->tdq_transferable; 657165620Sjeff high_load = high->tdq_load; 658165620Sjeff low_load = low->tdq_load; 659123487Sjeff } else { 660165620Sjeff transferable = high->tdq_group->tdg_transferable; 661165620Sjeff high_load = high->tdq_group->tdg_load; 662165620Sjeff low_load = low->tdq_group->tdg_load; 663123487Sjeff } 664123433Sjeff /* 665122744Sjeff * Determine what the imbalance is and then adjust that to how many 666165620Sjeff * threads we actually have to give up (transferable). 667122744Sjeff */ 668171482Sjeff if (transferable != 0) { 669171482Sjeff diff = high_load - low_load; 670171482Sjeff move = diff / 2; 671171482Sjeff if (diff & 0x1) 672171482Sjeff move++; 673171482Sjeff move = min(move, transferable); 674171482Sjeff for (i = 0; i < move; i++) 675171482Sjeff tdq_move(high, low); 676172293Sjeff /* 677172293Sjeff * IPI the target cpu to force it to reschedule with the new 678172293Sjeff * workload. 679172293Sjeff */ 680172293Sjeff ipi_selected(1 << TDQ_ID(low), IPI_PREEMPT); 681171482Sjeff } 682172409Sjeff tdq_unlock_pair(high, low); 683116069Sjeff return; 684116069Sjeff} 685116069Sjeff 686171482Sjeff/* 687171482Sjeff * Move a thread from one thread queue to another. 688171482Sjeff */ 689121790Sjeffstatic void 690171482Sjefftdq_move(struct tdq *from, struct tdq *to) 691116069Sjeff{ 692171482Sjeff struct td_sched *ts; 693171482Sjeff struct thread *td; 694164936Sjulian struct tdq *tdq; 695171482Sjeff int cpu; 696116069Sjeff 697172409Sjeff TDQ_LOCK_ASSERT(from, MA_OWNED); 698172409Sjeff TDQ_LOCK_ASSERT(to, MA_OWNED); 699172409Sjeff 700164936Sjulian tdq = from; 701171482Sjeff cpu = TDQ_ID(to); 702172409Sjeff ts = tdq_steal(tdq); 703164936Sjulian if (ts == NULL) { 704165620Sjeff struct tdq_group *tdg; 705123433Sjeff 706165620Sjeff tdg = tdq->tdq_group; 707165620Sjeff LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) { 708165620Sjeff if (tdq == from || tdq->tdq_transferable == 0) 709123433Sjeff continue; 710172409Sjeff ts = tdq_steal(tdq); 711123433Sjeff break; 712123433Sjeff } 713164936Sjulian if (ts == NULL) 714171482Sjeff return; 715123433Sjeff } 716164936Sjulian if (tdq == to) 717123433Sjeff return; 718171482Sjeff td = ts->ts_thread; 719171482Sjeff /* 720171482Sjeff * Although the run queue is locked the thread may be blocked. Lock 721172409Sjeff * it to clear this and acquire the run-queue lock. 722171482Sjeff */ 723171482Sjeff thread_lock(td); 724172409Sjeff /* Drop recursive lock on from acquired via thread_lock(). */ 725171482Sjeff TDQ_UNLOCK(from); 726171482Sjeff sched_rem(td); 727166108Sjeff ts->ts_cpu = cpu; 728171482Sjeff td->td_lock = TDQ_LOCKPTR(to); 729171482Sjeff tdq_add(to, td, SRQ_YIELDING); 730116069Sjeff} 731110267Sjeff 732171482Sjeff/* 733171482Sjeff * This tdq has idled. Try to steal a thread from another cpu and switch 734171482Sjeff * to it. 735171482Sjeff */ 736123433Sjeffstatic int 737164936Sjuliantdq_idled(struct tdq *tdq) 738121790Sjeff{ 739165620Sjeff struct tdq_group *tdg; 740164936Sjulian struct tdq *steal; 741171482Sjeff int highload; 742171482Sjeff int highcpu; 743171482Sjeff int cpu; 744123433Sjeff 745172484Sjeff if (smp_started == 0 || steal_idle == 0) 746172484Sjeff return (1); 747171482Sjeff /* We don't want to be preempted while we're iterating over tdqs */ 748171482Sjeff spinlock_enter(); 749165620Sjeff tdg = tdq->tdq_group; 750123433Sjeff /* 751165620Sjeff * If we're in a cpu group, try and steal threads from another cpu in 752172409Sjeff * the group before idling. In a HTT group all cpus share the same 753172409Sjeff * run-queue lock, however, we still need a recursive lock to 754172409Sjeff * call tdq_move(). 755123433Sjeff */ 756166108Sjeff if (steal_htt && tdg->tdg_cpus > 1 && tdg->tdg_transferable) { 757172409Sjeff TDQ_LOCK(tdq); 758165620Sjeff LIST_FOREACH(steal, &tdg->tdg_members, tdq_siblings) { 759165620Sjeff if (steal == tdq || steal->tdq_transferable == 0) 760123433Sjeff continue; 761171482Sjeff TDQ_LOCK(steal); 762172409Sjeff goto steal; 763166108Sjeff } 764172409Sjeff TDQ_UNLOCK(tdq); 765166108Sjeff } 766172484Sjeff /* 767172484Sjeff * Find the least loaded CPU with a transferable thread and attempt 768172484Sjeff * to steal it. We make a lockless pass and then verify that the 769172484Sjeff * thread is still available after locking. 770172484Sjeff */ 771171482Sjeff for (;;) { 772171482Sjeff highcpu = 0; 773171482Sjeff highload = 0; 774171482Sjeff for (cpu = 0; cpu <= mp_maxid; cpu++) { 775171482Sjeff if (CPU_ABSENT(cpu)) 776171482Sjeff continue; 777166108Sjeff steal = TDQ_CPU(cpu); 778172484Sjeff if (steal->tdq_transferable == 0) 779166108Sjeff continue; 780172484Sjeff if (steal->tdq_load < highload) 781172484Sjeff continue; 782172484Sjeff highload = steal->tdq_load; 783171482Sjeff highcpu = cpu; 784171482Sjeff } 785171506Sjeff if (highload < steal_thresh) 786171482Sjeff break; 787171482Sjeff steal = TDQ_CPU(highcpu); 788172484Sjeff if (steal == tdq) 789172484Sjeff break; 790172409Sjeff tdq_lock_pair(tdq, steal); 791172484Sjeff if (steal->tdq_load >= steal_thresh && steal->tdq_transferable) 792166108Sjeff goto steal; 793172409Sjeff tdq_unlock_pair(tdq, steal); 794123433Sjeff } 795171482Sjeff spinlock_exit(); 796123433Sjeff return (1); 797166108Sjeffsteal: 798171482Sjeff spinlock_exit(); 799172409Sjeff tdq_move(steal, tdq); 800171482Sjeff TDQ_UNLOCK(steal); 801171482Sjeff mi_switch(SW_VOL, NULL); 802171482Sjeff thread_unlock(curthread); 803121790Sjeff 804166108Sjeff return (0); 805121790Sjeff} 806121790Sjeff 807171482Sjeff/* 808171482Sjeff * Notify a remote cpu of new work. Sends an IPI if criteria are met. 809171482Sjeff */ 810121790Sjeffstatic void 811166108Sjefftdq_notify(struct td_sched *ts) 812121790Sjeff{ 813166247Sjeff struct thread *ctd; 814121790Sjeff struct pcpu *pcpu; 815166247Sjeff int cpri; 816166247Sjeff int pri; 817166108Sjeff int cpu; 818121790Sjeff 819166108Sjeff cpu = ts->ts_cpu; 820166247Sjeff pri = ts->ts_thread->td_priority; 821166108Sjeff pcpu = pcpu_find(cpu); 822166247Sjeff ctd = pcpu->pc_curthread; 823166247Sjeff cpri = ctd->td_priority; 824166137Sjeff 825121790Sjeff /* 826166137Sjeff * If our priority is not better than the current priority there is 827166137Sjeff * nothing to do. 828166137Sjeff */ 829166247Sjeff if (pri > cpri) 830166137Sjeff return; 831166137Sjeff /* 832166247Sjeff * Always IPI idle. 833121790Sjeff */ 834166247Sjeff if (cpri > PRI_MIN_IDLE) 835166247Sjeff goto sendipi; 836166247Sjeff /* 837166247Sjeff * If we're realtime or better and there is timeshare or worse running 838166247Sjeff * send an IPI. 839166247Sjeff */ 840166247Sjeff if (pri < PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME) 841166247Sjeff goto sendipi; 842166247Sjeff /* 843166247Sjeff * Otherwise only IPI if we exceed the threshold. 844166247Sjeff */ 845171482Sjeff if (pri > preempt_thresh) 846165819Sjeff return; 847166247Sjeffsendipi: 848166247Sjeff ctd->td_flags |= TDF_NEEDRESCHED; 849171482Sjeff ipi_selected(1 << cpu, IPI_PREEMPT); 850121790Sjeff} 851121790Sjeff 852171482Sjeff/* 853171482Sjeff * Steals load from a timeshare queue. Honors the rotating queue head 854171482Sjeff * index. 855171482Sjeff */ 856164936Sjulianstatic struct td_sched * 857171482Sjeffrunq_steal_from(struct runq *rq, u_char start) 858171482Sjeff{ 859171482Sjeff struct td_sched *ts; 860171482Sjeff struct rqbits *rqb; 861171482Sjeff struct rqhead *rqh; 862171482Sjeff int first; 863171482Sjeff int bit; 864171482Sjeff int pri; 865171482Sjeff int i; 866171482Sjeff 867171482Sjeff rqb = &rq->rq_status; 868171482Sjeff bit = start & (RQB_BPW -1); 869171482Sjeff pri = 0; 870171482Sjeff first = 0; 871171482Sjeffagain: 872171482Sjeff for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 873171482Sjeff if (rqb->rqb_bits[i] == 0) 874171482Sjeff continue; 875171482Sjeff if (bit != 0) { 876171482Sjeff for (pri = bit; pri < RQB_BPW; pri++) 877171482Sjeff if (rqb->rqb_bits[i] & (1ul << pri)) 878171482Sjeff break; 879171482Sjeff if (pri >= RQB_BPW) 880171482Sjeff continue; 881171482Sjeff } else 882171482Sjeff pri = RQB_FFS(rqb->rqb_bits[i]); 883171482Sjeff pri += (i << RQB_L2BPW); 884171482Sjeff rqh = &rq->rq_queues[pri]; 885171482Sjeff TAILQ_FOREACH(ts, rqh, ts_procq) { 886171482Sjeff if (first && THREAD_CAN_MIGRATE(ts->ts_thread)) 887171482Sjeff return (ts); 888171482Sjeff first = 1; 889171482Sjeff } 890171482Sjeff } 891171482Sjeff if (start != 0) { 892171482Sjeff start = 0; 893171482Sjeff goto again; 894171482Sjeff } 895171482Sjeff 896171482Sjeff return (NULL); 897171482Sjeff} 898171482Sjeff 899171482Sjeff/* 900171482Sjeff * Steals load from a standard linear queue. 901171482Sjeff */ 902171482Sjeffstatic struct td_sched * 903121790Sjeffrunq_steal(struct runq *rq) 904121790Sjeff{ 905121790Sjeff struct rqhead *rqh; 906121790Sjeff struct rqbits *rqb; 907164936Sjulian struct td_sched *ts; 908121790Sjeff int word; 909121790Sjeff int bit; 910121790Sjeff 911121790Sjeff rqb = &rq->rq_status; 912121790Sjeff for (word = 0; word < RQB_LEN; word++) { 913121790Sjeff if (rqb->rqb_bits[word] == 0) 914121790Sjeff continue; 915121790Sjeff for (bit = 0; bit < RQB_BPW; bit++) { 916123231Speter if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 917121790Sjeff continue; 918121790Sjeff rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 919171506Sjeff TAILQ_FOREACH(ts, rqh, ts_procq) 920171506Sjeff if (THREAD_CAN_MIGRATE(ts->ts_thread)) 921164936Sjulian return (ts); 922121790Sjeff } 923121790Sjeff } 924121790Sjeff return (NULL); 925121790Sjeff} 926121790Sjeff 927171482Sjeff/* 928171482Sjeff * Attempt to steal a thread in priority order from a thread queue. 929171482Sjeff */ 930164936Sjulianstatic struct td_sched * 931172409Sjefftdq_steal(struct tdq *tdq) 932121790Sjeff{ 933164936Sjulian struct td_sched *ts; 934121790Sjeff 935171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 936165762Sjeff if ((ts = runq_steal(&tdq->tdq_realtime)) != NULL) 937164936Sjulian return (ts); 938171482Sjeff if ((ts = runq_steal_from(&tdq->tdq_timeshare, tdq->tdq_ridx)) != NULL) 939164936Sjulian return (ts); 940172409Sjeff return (runq_steal(&tdq->tdq_idle)); 941121790Sjeff} 942123433Sjeff 943171482Sjeff/* 944171482Sjeff * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 945172409Sjeff * current lock and returns with the assigned queue locked. 946171482Sjeff */ 947171482Sjeffstatic inline struct tdq * 948171482Sjeffsched_setcpu(struct td_sched *ts, int cpu, int flags) 949123433Sjeff{ 950171482Sjeff struct thread *td; 951171482Sjeff struct tdq *tdq; 952123433Sjeff 953171482Sjeff THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 954171482Sjeff 955171482Sjeff tdq = TDQ_CPU(cpu); 956171482Sjeff td = ts->ts_thread; 957171482Sjeff ts->ts_cpu = cpu; 958171713Sjeff 959171713Sjeff /* If the lock matches just return the queue. */ 960171482Sjeff if (td->td_lock == TDQ_LOCKPTR(tdq)) 961171482Sjeff return (tdq); 962171482Sjeff#ifdef notyet 963123433Sjeff /* 964172293Sjeff * If the thread isn't running its lockptr is a 965171482Sjeff * turnstile or a sleepqueue. We can just lock_set without 966171482Sjeff * blocking. 967123685Sjeff */ 968171482Sjeff if (TD_CAN_RUN(td)) { 969171482Sjeff TDQ_LOCK(tdq); 970171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 971171482Sjeff return (tdq); 972171482Sjeff } 973171482Sjeff#endif 974166108Sjeff /* 975171482Sjeff * The hard case, migration, we need to block the thread first to 976171482Sjeff * prevent order reversals with other cpus locks. 977166108Sjeff */ 978171482Sjeff thread_lock_block(td); 979171482Sjeff TDQ_LOCK(tdq); 980171713Sjeff thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 981171482Sjeff return (tdq); 982166108Sjeff} 983166108Sjeff 984171482Sjeff/* 985171482Sjeff * Find the thread queue running the lowest priority thread. 986171482Sjeff */ 987166108Sjeffstatic int 988171482Sjefftdq_lowestpri(void) 989166108Sjeff{ 990171482Sjeff struct tdq *tdq; 991166108Sjeff int lowpri; 992166108Sjeff int lowcpu; 993166108Sjeff int lowload; 994166108Sjeff int load; 995171482Sjeff int cpu; 996171482Sjeff int pri; 997171482Sjeff 998171482Sjeff lowload = 0; 999171482Sjeff lowpri = lowcpu = 0; 1000171482Sjeff for (cpu = 0; cpu <= mp_maxid; cpu++) { 1001171482Sjeff if (CPU_ABSENT(cpu)) 1002171482Sjeff continue; 1003171482Sjeff tdq = TDQ_CPU(cpu); 1004171482Sjeff pri = tdq->tdq_lowpri; 1005171482Sjeff load = TDQ_CPU(cpu)->tdq_load; 1006171482Sjeff CTR4(KTR_ULE, 1007171482Sjeff "cpu %d pri %d lowcpu %d lowpri %d", 1008171482Sjeff cpu, pri, lowcpu, lowpri); 1009171482Sjeff if (pri < lowpri) 1010171482Sjeff continue; 1011171482Sjeff if (lowpri && lowpri == pri && load > lowload) 1012171482Sjeff continue; 1013171482Sjeff lowpri = pri; 1014171482Sjeff lowcpu = cpu; 1015171482Sjeff lowload = load; 1016171482Sjeff } 1017171482Sjeff 1018171482Sjeff return (lowcpu); 1019171482Sjeff} 1020171482Sjeff 1021171482Sjeff/* 1022171482Sjeff * Find the thread queue with the least load. 1023171482Sjeff */ 1024171482Sjeffstatic int 1025171482Sjefftdq_lowestload(void) 1026171482Sjeff{ 1027171482Sjeff struct tdq *tdq; 1028171482Sjeff int lowload; 1029171482Sjeff int lowpri; 1030171482Sjeff int lowcpu; 1031171482Sjeff int load; 1032171482Sjeff int cpu; 1033171482Sjeff int pri; 1034171482Sjeff 1035171482Sjeff lowcpu = 0; 1036171482Sjeff lowload = TDQ_CPU(0)->tdq_load; 1037171482Sjeff lowpri = TDQ_CPU(0)->tdq_lowpri; 1038171482Sjeff for (cpu = 1; cpu <= mp_maxid; cpu++) { 1039171482Sjeff if (CPU_ABSENT(cpu)) 1040171482Sjeff continue; 1041171482Sjeff tdq = TDQ_CPU(cpu); 1042171482Sjeff load = tdq->tdq_load; 1043171482Sjeff pri = tdq->tdq_lowpri; 1044171482Sjeff CTR4(KTR_ULE, "cpu %d load %d lowcpu %d lowload %d", 1045171482Sjeff cpu, load, lowcpu, lowload); 1046171482Sjeff if (load > lowload) 1047171482Sjeff continue; 1048171482Sjeff if (load == lowload && pri < lowpri) 1049171482Sjeff continue; 1050171482Sjeff lowcpu = cpu; 1051171482Sjeff lowload = load; 1052171482Sjeff lowpri = pri; 1053171482Sjeff } 1054171482Sjeff 1055171482Sjeff return (lowcpu); 1056171482Sjeff} 1057171482Sjeff 1058171482Sjeff/* 1059171482Sjeff * Pick the destination cpu for sched_add(). Respects affinity and makes 1060171482Sjeff * a determination based on load or priority of available processors. 1061171482Sjeff */ 1062171482Sjeffstatic int 1063171482Sjeffsched_pickcpu(struct td_sched *ts, int flags) 1064171482Sjeff{ 1065171482Sjeff struct tdq *tdq; 1066166108Sjeff int self; 1067166108Sjeff int pri; 1068166108Sjeff int cpu; 1069166108Sjeff 1070171482Sjeff cpu = self = PCPU_GET(cpuid); 1071166108Sjeff if (smp_started == 0) 1072166108Sjeff return (self); 1073171506Sjeff /* 1074171506Sjeff * Don't migrate a running thread from sched_switch(). 1075171506Sjeff */ 1076171506Sjeff if (flags & SRQ_OURSELF) { 1077171506Sjeff CTR1(KTR_ULE, "YIELDING %d", 1078171506Sjeff curthread->td_priority); 1079171506Sjeff return (self); 1080171506Sjeff } 1081166108Sjeff pri = ts->ts_thread->td_priority; 1082171482Sjeff cpu = ts->ts_cpu; 1083166108Sjeff /* 1084166108Sjeff * Regardless of affinity, if the last cpu is idle send it there. 1085166108Sjeff */ 1086171482Sjeff tdq = TDQ_CPU(cpu); 1087171482Sjeff if (tdq->tdq_lowpri > PRI_MIN_IDLE) { 1088166229Sjeff CTR5(KTR_ULE, 1089166108Sjeff "ts_cpu %d idle, ltick %d ticks %d pri %d curthread %d", 1090166108Sjeff ts->ts_cpu, ts->ts_rltick, ticks, pri, 1091171482Sjeff tdq->tdq_lowpri); 1092166108Sjeff return (ts->ts_cpu); 1093123433Sjeff } 1094166108Sjeff /* 1095166108Sjeff * If we have affinity, try to place it on the cpu we last ran on. 1096166108Sjeff */ 1097171482Sjeff if (SCHED_AFFINITY(ts) && tdq->tdq_lowpri > pri) { 1098166229Sjeff CTR5(KTR_ULE, 1099166108Sjeff "affinity for %d, ltick %d ticks %d pri %d curthread %d", 1100166108Sjeff ts->ts_cpu, ts->ts_rltick, ticks, pri, 1101171482Sjeff tdq->tdq_lowpri); 1102166108Sjeff return (ts->ts_cpu); 1103139334Sjeff } 1104123433Sjeff /* 1105166108Sjeff * Look for an idle group. 1106123433Sjeff */ 1107166229Sjeff CTR1(KTR_ULE, "tdq_idle %X", tdq_idle); 1108166108Sjeff cpu = ffs(tdq_idle); 1109166108Sjeff if (cpu) 1110171482Sjeff return (--cpu); 1111171506Sjeff /* 1112172409Sjeff * If there are no idle cores see if we can run the thread locally. 1113172409Sjeff * This may improve locality among sleepers and wakers when there 1114172409Sjeff * is shared data. 1115171506Sjeff */ 1116171506Sjeff if (tryself && pri < curthread->td_priority) { 1117171506Sjeff CTR1(KTR_ULE, "tryself %d", 1118166108Sjeff curthread->td_priority); 1119166108Sjeff return (self); 1120123433Sjeff } 1121133427Sjeff /* 1122166108Sjeff * Now search for the cpu running the lowest priority thread with 1123166108Sjeff * the least load. 1124123433Sjeff */ 1125171482Sjeff if (pick_pri) 1126171482Sjeff cpu = tdq_lowestpri(); 1127171482Sjeff else 1128171482Sjeff cpu = tdq_lowestload(); 1129171482Sjeff return (cpu); 1130123433Sjeff} 1131123433Sjeff 1132121790Sjeff#endif /* SMP */ 1133121790Sjeff 1134117326Sjeff/* 1135121790Sjeff * Pick the highest priority task we have and return it. 1136117326Sjeff */ 1137164936Sjulianstatic struct td_sched * 1138164936Sjuliantdq_choose(struct tdq *tdq) 1139110267Sjeff{ 1140164936Sjulian struct td_sched *ts; 1141110267Sjeff 1142171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1143165762Sjeff ts = runq_choose(&tdq->tdq_realtime); 1144170787Sjeff if (ts != NULL) 1145164936Sjulian return (ts); 1146165766Sjeff ts = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1147165762Sjeff if (ts != NULL) { 1148170787Sjeff KASSERT(ts->ts_thread->td_priority >= PRI_MIN_TIMESHARE, 1149165762Sjeff ("tdq_choose: Invalid priority on timeshare queue %d", 1150165762Sjeff ts->ts_thread->td_priority)); 1151165762Sjeff return (ts); 1152165762Sjeff } 1153110267Sjeff 1154165762Sjeff ts = runq_choose(&tdq->tdq_idle); 1155165762Sjeff if (ts != NULL) { 1156165762Sjeff KASSERT(ts->ts_thread->td_priority >= PRI_MIN_IDLE, 1157165762Sjeff ("tdq_choose: Invalid priority on idle queue %d", 1158165762Sjeff ts->ts_thread->td_priority)); 1159165762Sjeff return (ts); 1160165762Sjeff } 1161165762Sjeff 1162165762Sjeff return (NULL); 1163110267Sjeff} 1164110267Sjeff 1165171482Sjeff/* 1166171482Sjeff * Initialize a thread queue. 1167171482Sjeff */ 1168109864Sjeffstatic void 1169164936Sjuliantdq_setup(struct tdq *tdq) 1170110028Sjeff{ 1171171482Sjeff 1172171713Sjeff if (bootverbose) 1173171713Sjeff printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1174165762Sjeff runq_init(&tdq->tdq_realtime); 1175165762Sjeff runq_init(&tdq->tdq_timeshare); 1176165620Sjeff runq_init(&tdq->tdq_idle); 1177165620Sjeff tdq->tdq_load = 0; 1178110028Sjeff} 1179110028Sjeff 1180171713Sjeff#ifdef SMP 1181110028Sjeffstatic void 1182171713Sjefftdg_setup(struct tdq_group *tdg) 1183109864Sjeff{ 1184171713Sjeff if (bootverbose) 1185171713Sjeff printf("ULE: setup cpu group %d\n", TDG_ID(tdg)); 1186171713Sjeff snprintf(tdg->tdg_name, sizeof(tdg->tdg_name), 1187171713Sjeff "sched lock %d", (int)TDG_ID(tdg)); 1188171713Sjeff mtx_init(&tdg->tdg_lock, tdg->tdg_name, "sched lock", 1189171713Sjeff MTX_SPIN | MTX_RECURSE); 1190171713Sjeff LIST_INIT(&tdg->tdg_members); 1191171713Sjeff tdg->tdg_load = 0; 1192171713Sjeff tdg->tdg_transferable = 0; 1193171713Sjeff tdg->tdg_cpus = 0; 1194171713Sjeff tdg->tdg_mask = 0; 1195171713Sjeff tdg->tdg_cpumask = 0; 1196171713Sjeff tdg->tdg_idlemask = 0; 1197171713Sjeff} 1198171713Sjeff 1199171713Sjeffstatic void 1200171713Sjefftdg_add(struct tdq_group *tdg, struct tdq *tdq) 1201171713Sjeff{ 1202171713Sjeff if (tdg->tdg_mask == 0) 1203171713Sjeff tdg->tdg_mask |= 1 << TDQ_ID(tdq); 1204171713Sjeff tdg->tdg_cpumask |= 1 << TDQ_ID(tdq); 1205171713Sjeff tdg->tdg_cpus++; 1206171713Sjeff tdq->tdq_group = tdg; 1207171713Sjeff tdq->tdq_lock = &tdg->tdg_lock; 1208171713Sjeff LIST_INSERT_HEAD(&tdg->tdg_members, tdq, tdq_siblings); 1209171713Sjeff if (bootverbose) 1210171713Sjeff printf("ULE: adding cpu %d to group %d: cpus %d mask 0x%X\n", 1211171713Sjeff TDQ_ID(tdq), TDG_ID(tdg), tdg->tdg_cpus, tdg->tdg_cpumask); 1212171713Sjeff} 1213171713Sjeff 1214171713Sjeffstatic void 1215171713Sjeffsched_setup_topology(void) 1216171713Sjeff{ 1217171713Sjeff struct tdq_group *tdg; 1218171713Sjeff struct cpu_group *cg; 1219171713Sjeff int balance_groups; 1220171482Sjeff struct tdq *tdq; 1221109864Sjeff int i; 1222171713Sjeff int j; 1223109864Sjeff 1224171713Sjeff topology = 1; 1225123487Sjeff balance_groups = 0; 1226171713Sjeff for (i = 0; i < smp_topology->ct_count; i++) { 1227171713Sjeff cg = &smp_topology->ct_group[i]; 1228171713Sjeff tdg = &tdq_groups[i]; 1229171713Sjeff /* 1230171713Sjeff * Initialize the group. 1231171713Sjeff */ 1232171713Sjeff tdg_setup(tdg); 1233171713Sjeff /* 1234171713Sjeff * Find all of the group members and add them. 1235171713Sjeff */ 1236171713Sjeff for (j = 0; j < MAXCPU; j++) { 1237171713Sjeff if ((cg->cg_mask & (1 << j)) != 0) { 1238171713Sjeff tdq = TDQ_CPU(j); 1239171713Sjeff tdq_setup(tdq); 1240171713Sjeff tdg_add(tdg, tdq); 1241171713Sjeff } 1242171713Sjeff } 1243171713Sjeff if (tdg->tdg_cpus > 1) 1244171713Sjeff balance_groups = 1; 1245171713Sjeff } 1246171713Sjeff tdg_maxid = smp_topology->ct_count - 1; 1247171713Sjeff if (balance_groups) 1248172409Sjeff sched_balance_groups(); 1249171713Sjeff} 1250171713Sjeff 1251171713Sjeffstatic void 1252171713Sjeffsched_setup_smp(void) 1253171713Sjeff{ 1254171713Sjeff struct tdq_group *tdg; 1255171713Sjeff struct tdq *tdq; 1256171713Sjeff int cpus; 1257171713Sjeff int i; 1258171713Sjeff 1259171713Sjeff for (cpus = 0, i = 0; i < MAXCPU; i++) { 1260171713Sjeff if (CPU_ABSENT(i)) 1261171713Sjeff continue; 1262165627Sjeff tdq = &tdq_cpu[i]; 1263171713Sjeff tdg = &tdq_groups[i]; 1264171713Sjeff /* 1265171713Sjeff * Setup a tdq group with one member. 1266171713Sjeff */ 1267171713Sjeff tdg_setup(tdg); 1268171713Sjeff tdq_setup(tdq); 1269171713Sjeff tdg_add(tdg, tdq); 1270171713Sjeff cpus++; 1271123433Sjeff } 1272171713Sjeff tdg_maxid = cpus - 1; 1273171713Sjeff} 1274123433Sjeff 1275171713Sjeff/* 1276171713Sjeff * Fake a topology with one group containing all CPUs. 1277171713Sjeff */ 1278171713Sjeffstatic void 1279171713Sjeffsched_fake_topo(void) 1280171713Sjeff{ 1281171713Sjeff#ifdef SCHED_FAKE_TOPOLOGY 1282171713Sjeff static struct cpu_top top; 1283171713Sjeff static struct cpu_group group; 1284113357Sjeff 1285171713Sjeff top.ct_count = 1; 1286171713Sjeff top.ct_group = &group; 1287171713Sjeff group.cg_mask = all_cpus; 1288171713Sjeff group.cg_count = mp_ncpus; 1289171713Sjeff group.cg_children = 0; 1290171713Sjeff smp_topology = ⊤ 1291171713Sjeff#endif 1292171713Sjeff} 1293171713Sjeff#endif 1294171713Sjeff 1295171713Sjeff/* 1296171713Sjeff * Setup the thread queues and initialize the topology based on MD 1297171713Sjeff * information. 1298171713Sjeff */ 1299171713Sjeffstatic void 1300171713Sjeffsched_setup(void *dummy) 1301171713Sjeff{ 1302171713Sjeff struct tdq *tdq; 1303171713Sjeff 1304171713Sjeff tdq = TDQ_SELF(); 1305171713Sjeff#ifdef SMP 1306171713Sjeff sched_fake_topo(); 1307171713Sjeff /* 1308171713Sjeff * Setup tdqs based on a topology configuration or vanilla SMP based 1309171713Sjeff * on mp_maxid. 1310171713Sjeff */ 1311171713Sjeff if (smp_topology == NULL) 1312171713Sjeff sched_setup_smp(); 1313171713Sjeff else 1314171713Sjeff sched_setup_topology(); 1315172409Sjeff balance_tdq = tdq; 1316172409Sjeff sched_balance(); 1317117237Sjeff#else 1318171713Sjeff tdq_setup(tdq); 1319171713Sjeff mtx_init(&tdq_lock, "sched lock", "sched lock", MTX_SPIN | MTX_RECURSE); 1320171713Sjeff tdq->tdq_lock = &tdq_lock; 1321116069Sjeff#endif 1322171482Sjeff /* 1323171482Sjeff * To avoid divide-by-zero, we set realstathz a dummy value 1324171482Sjeff * in case which sched_clock() called before sched_initticks(). 1325171482Sjeff */ 1326171482Sjeff realstathz = hz; 1327171482Sjeff sched_slice = (realstathz/10); /* ~100ms */ 1328171482Sjeff tickincr = 1 << SCHED_TICK_SHIFT; 1329171482Sjeff 1330171482Sjeff /* Add thread0's load since it's running. */ 1331171482Sjeff TDQ_LOCK(tdq); 1332171713Sjeff thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1333171482Sjeff tdq_load_add(tdq, &td_sched0); 1334171482Sjeff TDQ_UNLOCK(tdq); 1335109864Sjeff} 1336109864Sjeff 1337171482Sjeff/* 1338171482Sjeff * This routine determines the tickincr after stathz and hz are setup. 1339171482Sjeff */ 1340153533Sdavidxu/* ARGSUSED */ 1341153533Sdavidxustatic void 1342153533Sdavidxusched_initticks(void *dummy) 1343153533Sdavidxu{ 1344171482Sjeff int incr; 1345171482Sjeff 1346153533Sdavidxu realstathz = stathz ? stathz : hz; 1347166229Sjeff sched_slice = (realstathz/10); /* ~100ms */ 1348153533Sdavidxu 1349153533Sdavidxu /* 1350165762Sjeff * tickincr is shifted out by 10 to avoid rounding errors due to 1351165766Sjeff * hz not being evenly divisible by stathz on all platforms. 1352153533Sdavidxu */ 1353171482Sjeff incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1354165762Sjeff /* 1355165762Sjeff * This does not work for values of stathz that are more than 1356165762Sjeff * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1357165762Sjeff */ 1358171482Sjeff if (incr == 0) 1359171482Sjeff incr = 1; 1360171482Sjeff tickincr = incr; 1361166108Sjeff#ifdef SMP 1362171899Sjeff /* 1363172409Sjeff * Set the default balance interval now that we know 1364172409Sjeff * what realstathz is. 1365172409Sjeff */ 1366172409Sjeff balance_interval = realstathz; 1367172409Sjeff /* 1368171899Sjeff * Set steal thresh to log2(mp_ncpu) but no greater than 4. This 1369171899Sjeff * prevents excess thrashing on large machines and excess idle on 1370171899Sjeff * smaller machines. 1371171899Sjeff */ 1372171899Sjeff steal_thresh = min(ffs(mp_ncpus) - 1, 4); 1373166108Sjeff affinity = SCHED_AFFINITY_DEFAULT; 1374166108Sjeff#endif 1375153533Sdavidxu} 1376153533Sdavidxu 1377153533Sdavidxu 1378109864Sjeff/* 1379171482Sjeff * This is the core of the interactivity algorithm. Determines a score based 1380171482Sjeff * on past behavior. It is the ratio of sleep time to run time scaled to 1381171482Sjeff * a [0, 100] integer. This is the voluntary sleep time of a process, which 1382171482Sjeff * differs from the cpu usage because it does not account for time spent 1383171482Sjeff * waiting on a run-queue. Would be prettier if we had floating point. 1384171482Sjeff */ 1385171482Sjeffstatic int 1386171482Sjeffsched_interact_score(struct thread *td) 1387171482Sjeff{ 1388171482Sjeff struct td_sched *ts; 1389171482Sjeff int div; 1390171482Sjeff 1391171482Sjeff ts = td->td_sched; 1392171482Sjeff /* 1393171482Sjeff * The score is only needed if this is likely to be an interactive 1394171482Sjeff * task. Don't go through the expense of computing it if there's 1395171482Sjeff * no chance. 1396171482Sjeff */ 1397171482Sjeff if (sched_interact <= SCHED_INTERACT_HALF && 1398171482Sjeff ts->ts_runtime >= ts->ts_slptime) 1399171482Sjeff return (SCHED_INTERACT_HALF); 1400171482Sjeff 1401171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1402171482Sjeff div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1403171482Sjeff return (SCHED_INTERACT_HALF + 1404171482Sjeff (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1405171482Sjeff } 1406171482Sjeff if (ts->ts_slptime > ts->ts_runtime) { 1407171482Sjeff div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1408171482Sjeff return (ts->ts_runtime / div); 1409171482Sjeff } 1410171482Sjeff /* runtime == slptime */ 1411171482Sjeff if (ts->ts_runtime) 1412171482Sjeff return (SCHED_INTERACT_HALF); 1413171482Sjeff 1414171482Sjeff /* 1415171482Sjeff * This can happen if slptime and runtime are 0. 1416171482Sjeff */ 1417171482Sjeff return (0); 1418171482Sjeff 1419171482Sjeff} 1420171482Sjeff 1421171482Sjeff/* 1422109864Sjeff * Scale the scheduling priority according to the "interactivity" of this 1423109864Sjeff * process. 1424109864Sjeff */ 1425113357Sjeffstatic void 1426163709Sjbsched_priority(struct thread *td) 1427109864Sjeff{ 1428165762Sjeff int score; 1429109864Sjeff int pri; 1430109864Sjeff 1431163709Sjb if (td->td_pri_class != PRI_TIMESHARE) 1432113357Sjeff return; 1433112966Sjeff /* 1434165762Sjeff * If the score is interactive we place the thread in the realtime 1435165762Sjeff * queue with a priority that is less than kernel and interrupt 1436165762Sjeff * priorities. These threads are not subject to nice restrictions. 1437112966Sjeff * 1438171482Sjeff * Scores greater than this are placed on the normal timeshare queue 1439165762Sjeff * where the priority is partially decided by the most recent cpu 1440165762Sjeff * utilization and the rest is decided by nice value. 1441172293Sjeff * 1442172293Sjeff * The nice value of the process has a linear effect on the calculated 1443172293Sjeff * score. Negative nice values make it easier for a thread to be 1444172293Sjeff * considered interactive. 1445112966Sjeff */ 1446172308Sjeff score = imax(0, sched_interact_score(td) - td->td_proc->p_nice); 1447165762Sjeff if (score < sched_interact) { 1448165762Sjeff pri = PRI_MIN_REALTIME; 1449165762Sjeff pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact) 1450165762Sjeff * score; 1451165762Sjeff KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME, 1452166208Sjeff ("sched_priority: invalid interactive priority %d score %d", 1453166208Sjeff pri, score)); 1454165762Sjeff } else { 1455165762Sjeff pri = SCHED_PRI_MIN; 1456165762Sjeff if (td->td_sched->ts_ticks) 1457165762Sjeff pri += SCHED_PRI_TICKS(td->td_sched); 1458165762Sjeff pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1459171482Sjeff KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE, 1460171482Sjeff ("sched_priority: invalid priority %d: nice %d, " 1461171482Sjeff "ticks %d ftick %d ltick %d tick pri %d", 1462171482Sjeff pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1463171482Sjeff td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1464171482Sjeff SCHED_PRI_TICKS(td->td_sched))); 1465165762Sjeff } 1466165762Sjeff sched_user_prio(td, pri); 1467112966Sjeff 1468112966Sjeff return; 1469109864Sjeff} 1470109864Sjeff 1471121868Sjeff/* 1472121868Sjeff * This routine enforces a maximum limit on the amount of scheduling history 1473171482Sjeff * kept. It is called after either the slptime or runtime is adjusted. This 1474171482Sjeff * function is ugly due to integer math. 1475121868Sjeff */ 1476116463Sjeffstatic void 1477163709Sjbsched_interact_update(struct thread *td) 1478116463Sjeff{ 1479165819Sjeff struct td_sched *ts; 1480166208Sjeff u_int sum; 1481121605Sjeff 1482165819Sjeff ts = td->td_sched; 1483171482Sjeff sum = ts->ts_runtime + ts->ts_slptime; 1484121868Sjeff if (sum < SCHED_SLP_RUN_MAX) 1485121868Sjeff return; 1486121868Sjeff /* 1487165819Sjeff * This only happens from two places: 1488165819Sjeff * 1) We have added an unusual amount of run time from fork_exit. 1489165819Sjeff * 2) We have added an unusual amount of sleep time from sched_sleep(). 1490165819Sjeff */ 1491165819Sjeff if (sum > SCHED_SLP_RUN_MAX * 2) { 1492171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1493171482Sjeff ts->ts_runtime = SCHED_SLP_RUN_MAX; 1494171482Sjeff ts->ts_slptime = 1; 1495165819Sjeff } else { 1496171482Sjeff ts->ts_slptime = SCHED_SLP_RUN_MAX; 1497171482Sjeff ts->ts_runtime = 1; 1498165819Sjeff } 1499165819Sjeff return; 1500165819Sjeff } 1501165819Sjeff /* 1502121868Sjeff * If we have exceeded by more than 1/5th then the algorithm below 1503121868Sjeff * will not bring us back into range. Dividing by two here forces 1504133427Sjeff * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1505121868Sjeff */ 1506127850Sjeff if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1507171482Sjeff ts->ts_runtime /= 2; 1508171482Sjeff ts->ts_slptime /= 2; 1509121868Sjeff return; 1510116463Sjeff } 1511171482Sjeff ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1512171482Sjeff ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1513116463Sjeff} 1514116463Sjeff 1515171482Sjeff/* 1516171482Sjeff * Scale back the interactivity history when a child thread is created. The 1517171482Sjeff * history is inherited from the parent but the thread may behave totally 1518171482Sjeff * differently. For example, a shell spawning a compiler process. We want 1519171482Sjeff * to learn that the compiler is behaving badly very quickly. 1520171482Sjeff */ 1521121868Sjeffstatic void 1522163709Sjbsched_interact_fork(struct thread *td) 1523121868Sjeff{ 1524121868Sjeff int ratio; 1525121868Sjeff int sum; 1526121868Sjeff 1527171482Sjeff sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1528121868Sjeff if (sum > SCHED_SLP_RUN_FORK) { 1529121868Sjeff ratio = sum / SCHED_SLP_RUN_FORK; 1530171482Sjeff td->td_sched->ts_runtime /= ratio; 1531171482Sjeff td->td_sched->ts_slptime /= ratio; 1532121868Sjeff } 1533121868Sjeff} 1534121868Sjeff 1535113357Sjeff/* 1536171482Sjeff * Called from proc0_init() to setup the scheduler fields. 1537134791Sjulian */ 1538134791Sjulianvoid 1539134791Sjulianschedinit(void) 1540134791Sjulian{ 1541165762Sjeff 1542134791Sjulian /* 1543134791Sjulian * Set up the scheduler specific parts of proc0. 1544134791Sjulian */ 1545136167Sjulian proc0.p_sched = NULL; /* XXX */ 1546164936Sjulian thread0.td_sched = &td_sched0; 1547165762Sjeff td_sched0.ts_ltick = ticks; 1548165796Sjeff td_sched0.ts_ftick = ticks; 1549164936Sjulian td_sched0.ts_thread = &thread0; 1550134791Sjulian} 1551134791Sjulian 1552134791Sjulian/* 1553113357Sjeff * This is only somewhat accurate since given many processes of the same 1554113357Sjeff * priority they will switch when their slices run out, which will be 1555165762Sjeff * at most sched_slice stathz ticks. 1556113357Sjeff */ 1557109864Sjeffint 1558109864Sjeffsched_rr_interval(void) 1559109864Sjeff{ 1560165762Sjeff 1561165762Sjeff /* Convert sched_slice to hz */ 1562165762Sjeff return (hz/(realstathz/sched_slice)); 1563109864Sjeff} 1564109864Sjeff 1565171482Sjeff/* 1566171482Sjeff * Update the percent cpu tracking information when it is requested or 1567171482Sjeff * the total history exceeds the maximum. We keep a sliding history of 1568171482Sjeff * tick counts that slowly decays. This is less precise than the 4BSD 1569171482Sjeff * mechanism since it happens with less regular and frequent events. 1570171482Sjeff */ 1571121790Sjeffstatic void 1572164936Sjuliansched_pctcpu_update(struct td_sched *ts) 1573109864Sjeff{ 1574165762Sjeff 1575165762Sjeff if (ts->ts_ticks == 0) 1576165762Sjeff return; 1577165796Sjeff if (ticks - (hz / 10) < ts->ts_ltick && 1578165796Sjeff SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1579165796Sjeff return; 1580109864Sjeff /* 1581109864Sjeff * Adjust counters and watermark for pctcpu calc. 1582116365Sjeff */ 1583165762Sjeff if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1584164936Sjulian ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1585165762Sjeff SCHED_TICK_TARG; 1586165762Sjeff else 1587164936Sjulian ts->ts_ticks = 0; 1588164936Sjulian ts->ts_ltick = ticks; 1589165762Sjeff ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1590109864Sjeff} 1591109864Sjeff 1592171482Sjeff/* 1593171482Sjeff * Adjust the priority of a thread. Move it to the appropriate run-queue 1594171482Sjeff * if necessary. This is the back-end for several priority related 1595171482Sjeff * functions. 1596171482Sjeff */ 1597165762Sjeffstatic void 1598139453Sjhbsched_thread_priority(struct thread *td, u_char prio) 1599109864Sjeff{ 1600164936Sjulian struct td_sched *ts; 1601109864Sjeff 1602139316Sjeff CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 1603173600Sjulian td, td->td_name, td->td_priority, prio, curthread, 1604173600Sjulian curthread->td_name); 1605164936Sjulian ts = td->td_sched; 1606170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1607139453Sjhb if (td->td_priority == prio) 1608139453Sjhb return; 1609165762Sjeff 1610165766Sjeff if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1611121605Sjeff /* 1612121605Sjeff * If the priority has been elevated due to priority 1613121605Sjeff * propagation, we may have to move ourselves to a new 1614165762Sjeff * queue. This could be optimized to not re-add in some 1615165762Sjeff * cases. 1616133555Sjeff */ 1617165762Sjeff sched_rem(td); 1618165762Sjeff td->td_priority = prio; 1619171482Sjeff sched_add(td, SRQ_BORROWING); 1620171482Sjeff } else { 1621171482Sjeff#ifdef SMP 1622171482Sjeff struct tdq *tdq; 1623171482Sjeff 1624171482Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1625171482Sjeff if (prio < tdq->tdq_lowpri) 1626171482Sjeff tdq->tdq_lowpri = prio; 1627171482Sjeff#endif 1628119488Sdavidxu td->td_priority = prio; 1629171482Sjeff } 1630109864Sjeff} 1631109864Sjeff 1632139453Sjhb/* 1633139453Sjhb * Update a thread's priority when it is lent another thread's 1634139453Sjhb * priority. 1635139453Sjhb */ 1636109864Sjeffvoid 1637139453Sjhbsched_lend_prio(struct thread *td, u_char prio) 1638139453Sjhb{ 1639139453Sjhb 1640139453Sjhb td->td_flags |= TDF_BORROWING; 1641139453Sjhb sched_thread_priority(td, prio); 1642139453Sjhb} 1643139453Sjhb 1644139453Sjhb/* 1645139453Sjhb * Restore a thread's priority when priority propagation is 1646139453Sjhb * over. The prio argument is the minimum priority the thread 1647139453Sjhb * needs to have to satisfy other possible priority lending 1648139453Sjhb * requests. If the thread's regular priority is less 1649139453Sjhb * important than prio, the thread will keep a priority boost 1650139453Sjhb * of prio. 1651139453Sjhb */ 1652139453Sjhbvoid 1653139453Sjhbsched_unlend_prio(struct thread *td, u_char prio) 1654139453Sjhb{ 1655139453Sjhb u_char base_pri; 1656139453Sjhb 1657139453Sjhb if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1658139453Sjhb td->td_base_pri <= PRI_MAX_TIMESHARE) 1659163709Sjb base_pri = td->td_user_pri; 1660139453Sjhb else 1661139453Sjhb base_pri = td->td_base_pri; 1662139453Sjhb if (prio >= base_pri) { 1663139455Sjhb td->td_flags &= ~TDF_BORROWING; 1664139453Sjhb sched_thread_priority(td, base_pri); 1665139453Sjhb } else 1666139453Sjhb sched_lend_prio(td, prio); 1667139453Sjhb} 1668139453Sjhb 1669171482Sjeff/* 1670171482Sjeff * Standard entry for setting the priority to an absolute value. 1671171482Sjeff */ 1672139453Sjhbvoid 1673139453Sjhbsched_prio(struct thread *td, u_char prio) 1674139453Sjhb{ 1675139453Sjhb u_char oldprio; 1676139453Sjhb 1677139453Sjhb /* First, update the base priority. */ 1678139453Sjhb td->td_base_pri = prio; 1679139453Sjhb 1680139453Sjhb /* 1681139455Sjhb * If the thread is borrowing another thread's priority, don't 1682139453Sjhb * ever lower the priority. 1683139453Sjhb */ 1684139453Sjhb if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1685139453Sjhb return; 1686139453Sjhb 1687139453Sjhb /* Change the real priority. */ 1688139453Sjhb oldprio = td->td_priority; 1689139453Sjhb sched_thread_priority(td, prio); 1690139453Sjhb 1691139453Sjhb /* 1692139453Sjhb * If the thread is on a turnstile, then let the turnstile update 1693139453Sjhb * its state. 1694139453Sjhb */ 1695139453Sjhb if (TD_ON_LOCK(td) && oldprio != prio) 1696139453Sjhb turnstile_adjust(td, oldprio); 1697139453Sjhb} 1698139455Sjhb 1699171482Sjeff/* 1700171482Sjeff * Set the base user priority, does not effect current running priority. 1701171482Sjeff */ 1702139453Sjhbvoid 1703163709Sjbsched_user_prio(struct thread *td, u_char prio) 1704161599Sdavidxu{ 1705161599Sdavidxu u_char oldprio; 1706161599Sdavidxu 1707163709Sjb td->td_base_user_pri = prio; 1708164939Sjulian if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio) 1709164939Sjulian return; 1710163709Sjb oldprio = td->td_user_pri; 1711163709Sjb td->td_user_pri = prio; 1712161599Sdavidxu} 1713161599Sdavidxu 1714161599Sdavidxuvoid 1715161599Sdavidxusched_lend_user_prio(struct thread *td, u_char prio) 1716161599Sdavidxu{ 1717161599Sdavidxu u_char oldprio; 1718161599Sdavidxu 1719174536Sdavidxu THREAD_LOCK_ASSERT(td, MA_OWNED); 1720161599Sdavidxu td->td_flags |= TDF_UBORROWING; 1721164091Smaxim oldprio = td->td_user_pri; 1722163709Sjb td->td_user_pri = prio; 1723161599Sdavidxu} 1724161599Sdavidxu 1725161599Sdavidxuvoid 1726161599Sdavidxusched_unlend_user_prio(struct thread *td, u_char prio) 1727161599Sdavidxu{ 1728161599Sdavidxu u_char base_pri; 1729161599Sdavidxu 1730174536Sdavidxu THREAD_LOCK_ASSERT(td, MA_OWNED); 1731163709Sjb base_pri = td->td_base_user_pri; 1732161599Sdavidxu if (prio >= base_pri) { 1733161599Sdavidxu td->td_flags &= ~TDF_UBORROWING; 1734163709Sjb sched_user_prio(td, base_pri); 1735174536Sdavidxu } else { 1736161599Sdavidxu sched_lend_user_prio(td, prio); 1737174536Sdavidxu } 1738161599Sdavidxu} 1739161599Sdavidxu 1740171482Sjeff/* 1741171505Sjeff * Add the thread passed as 'newtd' to the run queue before selecting 1742171505Sjeff * the next thread to run. This is only used for KSE. 1743171505Sjeff */ 1744171505Sjeffstatic void 1745171505Sjeffsched_switchin(struct tdq *tdq, struct thread *td) 1746171505Sjeff{ 1747171505Sjeff#ifdef SMP 1748171505Sjeff spinlock_enter(); 1749171505Sjeff TDQ_UNLOCK(tdq); 1750171505Sjeff thread_lock(td); 1751171505Sjeff spinlock_exit(); 1752171505Sjeff sched_setcpu(td->td_sched, TDQ_ID(tdq), SRQ_YIELDING); 1753171505Sjeff#else 1754171505Sjeff td->td_lock = TDQ_LOCKPTR(tdq); 1755171505Sjeff#endif 1756171505Sjeff tdq_add(tdq, td, SRQ_YIELDING); 1757171505Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1758171505Sjeff} 1759171505Sjeff 1760171505Sjeff/* 1761171713Sjeff * Handle migration from sched_switch(). This happens only for 1762171713Sjeff * cpu binding. 1763171713Sjeff */ 1764171713Sjeffstatic struct mtx * 1765171713Sjeffsched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1766171713Sjeff{ 1767171713Sjeff struct tdq *tdn; 1768171713Sjeff 1769171713Sjeff tdn = TDQ_CPU(td->td_sched->ts_cpu); 1770171713Sjeff#ifdef SMP 1771171713Sjeff /* 1772171713Sjeff * Do the lock dance required to avoid LOR. We grab an extra 1773171713Sjeff * spinlock nesting to prevent preemption while we're 1774171713Sjeff * not holding either run-queue lock. 1775171713Sjeff */ 1776171713Sjeff spinlock_enter(); 1777171713Sjeff thread_block_switch(td); /* This releases the lock on tdq. */ 1778171713Sjeff TDQ_LOCK(tdn); 1779171713Sjeff tdq_add(tdn, td, flags); 1780171713Sjeff tdq_notify(td->td_sched); 1781171713Sjeff /* 1782171713Sjeff * After we unlock tdn the new cpu still can't switch into this 1783171713Sjeff * thread until we've unblocked it in cpu_switch(). The lock 1784171713Sjeff * pointers may match in the case of HTT cores. Don't unlock here 1785171713Sjeff * or we can deadlock when the other CPU runs the IPI handler. 1786171713Sjeff */ 1787171713Sjeff if (TDQ_LOCKPTR(tdn) != TDQ_LOCKPTR(tdq)) { 1788171713Sjeff TDQ_UNLOCK(tdn); 1789171713Sjeff TDQ_LOCK(tdq); 1790171713Sjeff } 1791171713Sjeff spinlock_exit(); 1792171713Sjeff#endif 1793171713Sjeff return (TDQ_LOCKPTR(tdn)); 1794171713Sjeff} 1795171713Sjeff 1796171713Sjeff/* 1797171482Sjeff * Block a thread for switching. Similar to thread_block() but does not 1798171482Sjeff * bump the spin count. 1799171482Sjeff */ 1800171482Sjeffstatic inline struct mtx * 1801171482Sjeffthread_block_switch(struct thread *td) 1802171482Sjeff{ 1803171482Sjeff struct mtx *lock; 1804171482Sjeff 1805171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1806171482Sjeff lock = td->td_lock; 1807171482Sjeff td->td_lock = &blocked_lock; 1808171482Sjeff mtx_unlock_spin(lock); 1809171482Sjeff 1810171482Sjeff return (lock); 1811171482Sjeff} 1812171482Sjeff 1813171482Sjeff/* 1814171482Sjeff * Release a thread that was blocked with thread_block_switch(). 1815171482Sjeff */ 1816171482Sjeffstatic inline void 1817171482Sjeffthread_unblock_switch(struct thread *td, struct mtx *mtx) 1818171482Sjeff{ 1819171482Sjeff atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1820171482Sjeff (uintptr_t)mtx); 1821171482Sjeff} 1822171482Sjeff 1823171482Sjeff/* 1824171482Sjeff * Switch threads. This function has to handle threads coming in while 1825171482Sjeff * blocked for some reason, running, or idle. It also must deal with 1826171482Sjeff * migrating a thread from one queue to another as running threads may 1827171482Sjeff * be assigned elsewhere via binding. 1828171482Sjeff */ 1829161599Sdavidxuvoid 1830135051Sjuliansched_switch(struct thread *td, struct thread *newtd, int flags) 1831109864Sjeff{ 1832165627Sjeff struct tdq *tdq; 1833164936Sjulian struct td_sched *ts; 1834171482Sjeff struct mtx *mtx; 1835171713Sjeff int srqflag; 1836171482Sjeff int cpuid; 1837109864Sjeff 1838170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1839109864Sjeff 1840171482Sjeff cpuid = PCPU_GET(cpuid); 1841171482Sjeff tdq = TDQ_CPU(cpuid); 1842164936Sjulian ts = td->td_sched; 1843171713Sjeff mtx = td->td_lock; 1844171482Sjeff#ifdef SMP 1845171482Sjeff ts->ts_rltick = ticks; 1846171482Sjeff if (newtd && newtd->td_priority < tdq->tdq_lowpri) 1847171482Sjeff tdq->tdq_lowpri = newtd->td_priority; 1848171482Sjeff#endif 1849133555Sjeff td->td_lastcpu = td->td_oncpu; 1850113339Sjulian td->td_oncpu = NOCPU; 1851132266Sjhb td->td_flags &= ~TDF_NEEDRESCHED; 1852144777Sups td->td_owepreempt = 0; 1853123434Sjeff /* 1854171482Sjeff * The lock pointer in an idle thread should never change. Reset it 1855171482Sjeff * to CAN_RUN as well. 1856123434Sjeff */ 1857167327Sjulian if (TD_IS_IDLETHREAD(td)) { 1858171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1859139334Sjeff TD_SET_CAN_RUN(td); 1860170293Sjeff } else if (TD_IS_RUNNING(td)) { 1861171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1862165627Sjeff tdq_load_rem(tdq, ts); 1863171713Sjeff srqflag = (flags & SW_PREEMPT) ? 1864170293Sjeff SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1865171713Sjeff SRQ_OURSELF|SRQ_YIELDING; 1866171713Sjeff if (ts->ts_cpu == cpuid) 1867171713Sjeff tdq_add(tdq, td, srqflag); 1868171713Sjeff else 1869171713Sjeff mtx = sched_switch_migrate(tdq, td, srqflag); 1870171482Sjeff } else { 1871171482Sjeff /* This thread must be going to sleep. */ 1872171482Sjeff TDQ_LOCK(tdq); 1873171482Sjeff mtx = thread_block_switch(td); 1874170293Sjeff tdq_load_rem(tdq, ts); 1875171482Sjeff } 1876171482Sjeff /* 1877171482Sjeff * We enter here with the thread blocked and assigned to the 1878171482Sjeff * appropriate cpu run-queue or sleep-queue and with the current 1879171482Sjeff * thread-queue locked. 1880171482Sjeff */ 1881171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1882171482Sjeff /* 1883171505Sjeff * If KSE assigned a new thread just add it here and let choosethread 1884171505Sjeff * select the best one. 1885171482Sjeff */ 1886171505Sjeff if (newtd != NULL) 1887171505Sjeff sched_switchin(tdq, newtd); 1888171482Sjeff newtd = choosethread(); 1889171482Sjeff /* 1890171482Sjeff * Call the MD code to switch contexts if necessary. 1891171482Sjeff */ 1892145256Sjkoshy if (td != newtd) { 1893145256Sjkoshy#ifdef HWPMC_HOOKS 1894145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1895145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1896145256Sjkoshy#endif 1897174629Sjeff lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1898172411Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1899171482Sjeff cpu_switch(td, newtd, mtx); 1900171482Sjeff /* 1901171482Sjeff * We may return from cpu_switch on a different cpu. However, 1902171482Sjeff * we always return with td_lock pointing to the current cpu's 1903171482Sjeff * run queue lock. 1904171482Sjeff */ 1905171482Sjeff cpuid = PCPU_GET(cpuid); 1906171482Sjeff tdq = TDQ_CPU(cpuid); 1907174629Sjeff lock_profile_obtain_lock_success( 1908174629Sjeff &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1909145256Sjkoshy#ifdef HWPMC_HOOKS 1910145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1911145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1912145256Sjkoshy#endif 1913171482Sjeff } else 1914171482Sjeff thread_unblock_switch(td, mtx); 1915171482Sjeff /* 1916171482Sjeff * Assert that all went well and return. 1917171482Sjeff */ 1918171482Sjeff#ifdef SMP 1919171482Sjeff /* We should always get here with the lowest priority td possible */ 1920171482Sjeff tdq->tdq_lowpri = td->td_priority; 1921171482Sjeff#endif 1922171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1923171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1924171482Sjeff td->td_oncpu = cpuid; 1925109864Sjeff} 1926109864Sjeff 1927171482Sjeff/* 1928171482Sjeff * Adjust thread priorities as a result of a nice request. 1929171482Sjeff */ 1930109864Sjeffvoid 1931130551Sjuliansched_nice(struct proc *p, int nice) 1932109864Sjeff{ 1933109864Sjeff struct thread *td; 1934109864Sjeff 1935130551Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 1936170293Sjeff PROC_SLOCK_ASSERT(p, MA_OWNED); 1937165762Sjeff 1938130551Sjulian p->p_nice = nice; 1939163709Sjb FOREACH_THREAD_IN_PROC(p, td) { 1940170293Sjeff thread_lock(td); 1941163709Sjb sched_priority(td); 1942165762Sjeff sched_prio(td, td->td_base_user_pri); 1943170293Sjeff thread_unlock(td); 1944130551Sjulian } 1945109864Sjeff} 1946109864Sjeff 1947171482Sjeff/* 1948171482Sjeff * Record the sleep time for the interactivity scorer. 1949171482Sjeff */ 1950109864Sjeffvoid 1951126326Sjhbsched_sleep(struct thread *td) 1952109864Sjeff{ 1953165762Sjeff 1954170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1955109864Sjeff 1956172264Sjeff td->td_slptick = ticks; 1957109864Sjeff} 1958109864Sjeff 1959171482Sjeff/* 1960171482Sjeff * Schedule a thread to resume execution and record how long it voluntarily 1961171482Sjeff * slept. We also update the pctcpu, interactivity, and priority. 1962171482Sjeff */ 1963109864Sjeffvoid 1964109864Sjeffsched_wakeup(struct thread *td) 1965109864Sjeff{ 1966166229Sjeff struct td_sched *ts; 1967171482Sjeff int slptick; 1968165762Sjeff 1969170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1970166229Sjeff ts = td->td_sched; 1971109864Sjeff /* 1972165762Sjeff * If we slept for more than a tick update our interactivity and 1973165762Sjeff * priority. 1974109864Sjeff */ 1975172264Sjeff slptick = td->td_slptick; 1976172264Sjeff td->td_slptick = 0; 1977171482Sjeff if (slptick && slptick != ticks) { 1978166208Sjeff u_int hzticks; 1979109864Sjeff 1980171482Sjeff hzticks = (ticks - slptick) << SCHED_TICK_SHIFT; 1981171482Sjeff ts->ts_slptime += hzticks; 1982165819Sjeff sched_interact_update(td); 1983166229Sjeff sched_pctcpu_update(ts); 1984163709Sjb sched_priority(td); 1985109864Sjeff } 1986166229Sjeff /* Reset the slice value after we sleep. */ 1987166229Sjeff ts->ts_slice = sched_slice; 1988166190Sjeff sched_add(td, SRQ_BORING); 1989109864Sjeff} 1990109864Sjeff 1991109864Sjeff/* 1992109864Sjeff * Penalize the parent for creating a new child and initialize the child's 1993109864Sjeff * priority. 1994109864Sjeff */ 1995109864Sjeffvoid 1996163709Sjbsched_fork(struct thread *td, struct thread *child) 1997109864Sjeff{ 1998170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1999164936Sjulian sched_fork_thread(td, child); 2000165762Sjeff /* 2001165762Sjeff * Penalize the parent and child for forking. 2002165762Sjeff */ 2003165762Sjeff sched_interact_fork(child); 2004165762Sjeff sched_priority(child); 2005171482Sjeff td->td_sched->ts_runtime += tickincr; 2006165762Sjeff sched_interact_update(td); 2007165762Sjeff sched_priority(td); 2008164936Sjulian} 2009109864Sjeff 2010171482Sjeff/* 2011171482Sjeff * Fork a new thread, may be within the same process. 2012171482Sjeff */ 2013164936Sjulianvoid 2014164936Sjuliansched_fork_thread(struct thread *td, struct thread *child) 2015164936Sjulian{ 2016164936Sjulian struct td_sched *ts; 2017164936Sjulian struct td_sched *ts2; 2018164936Sjulian 2019165762Sjeff /* 2020165762Sjeff * Initialize child. 2021165762Sjeff */ 2022170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2023163709Sjb sched_newthread(child); 2024171482Sjeff child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 2025164936Sjulian ts = td->td_sched; 2026164936Sjulian ts2 = child->td_sched; 2027164936Sjulian ts2->ts_cpu = ts->ts_cpu; 2028164936Sjulian ts2->ts_runq = NULL; 2029165762Sjeff /* 2030165762Sjeff * Grab our parents cpu estimation information and priority. 2031165762Sjeff */ 2032164936Sjulian ts2->ts_ticks = ts->ts_ticks; 2033164936Sjulian ts2->ts_ltick = ts->ts_ltick; 2034164936Sjulian ts2->ts_ftick = ts->ts_ftick; 2035165762Sjeff child->td_user_pri = td->td_user_pri; 2036165762Sjeff child->td_base_user_pri = td->td_base_user_pri; 2037165762Sjeff /* 2038165762Sjeff * And update interactivity score. 2039165762Sjeff */ 2040171482Sjeff ts2->ts_slptime = ts->ts_slptime; 2041171482Sjeff ts2->ts_runtime = ts->ts_runtime; 2042165762Sjeff ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 2043113357Sjeff} 2044113357Sjeff 2045171482Sjeff/* 2046171482Sjeff * Adjust the priority class of a thread. 2047171482Sjeff */ 2048113357Sjeffvoid 2049163709Sjbsched_class(struct thread *td, int class) 2050113357Sjeff{ 2051113357Sjeff 2052170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2053163709Sjb if (td->td_pri_class == class) 2054113357Sjeff return; 2055113357Sjeff 2056121896Sjeff#ifdef SMP 2057165827Sjeff /* 2058165827Sjeff * On SMP if we're on the RUNQ we must adjust the transferable 2059165827Sjeff * count because could be changing to or from an interrupt 2060165827Sjeff * class. 2061165827Sjeff */ 2062166190Sjeff if (TD_ON_RUNQ(td)) { 2063165827Sjeff struct tdq *tdq; 2064165827Sjeff 2065165827Sjeff tdq = TDQ_CPU(td->td_sched->ts_cpu); 2066165827Sjeff if (THREAD_CAN_MIGRATE(td)) { 2067165827Sjeff tdq->tdq_transferable--; 2068165827Sjeff tdq->tdq_group->tdg_transferable--; 2069122744Sjeff } 2070165827Sjeff td->td_pri_class = class; 2071165827Sjeff if (THREAD_CAN_MIGRATE(td)) { 2072165827Sjeff tdq->tdq_transferable++; 2073165827Sjeff tdq->tdq_group->tdg_transferable++; 2074165827Sjeff } 2075165827Sjeff } 2076164936Sjulian#endif 2077163709Sjb td->td_pri_class = class; 2078109864Sjeff} 2079109864Sjeff 2080109864Sjeff/* 2081109864Sjeff * Return some of the child's priority and interactivity to the parent. 2082109864Sjeff */ 2083109864Sjeffvoid 2084164939Sjuliansched_exit(struct proc *p, struct thread *child) 2085109864Sjeff{ 2086165762Sjeff struct thread *td; 2087164939Sjulian 2088163709Sjb CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d", 2089173600Sjulian child, child->td_name, child->td_priority); 2090113372Sjeff 2091170293Sjeff PROC_SLOCK_ASSERT(p, MA_OWNED); 2092165762Sjeff td = FIRST_THREAD_IN_PROC(p); 2093165762Sjeff sched_exit_thread(td, child); 2094113372Sjeff} 2095113372Sjeff 2096171482Sjeff/* 2097171482Sjeff * Penalize another thread for the time spent on this one. This helps to 2098171482Sjeff * worsen the priority and interactivity of processes which schedule batch 2099171482Sjeff * jobs such as make. This has little effect on the make process itself but 2100171482Sjeff * causes new processes spawned by it to receive worse scores immediately. 2101171482Sjeff */ 2102113372Sjeffvoid 2103164939Sjuliansched_exit_thread(struct thread *td, struct thread *child) 2104164936Sjulian{ 2105165762Sjeff 2106164939Sjulian CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d", 2107173600Sjulian child, child->td_name, child->td_priority); 2108164939Sjulian 2109165762Sjeff#ifdef KSE 2110165762Sjeff /* 2111165762Sjeff * KSE forks and exits so often that this penalty causes short-lived 2112165762Sjeff * threads to always be non-interactive. This causes mozilla to 2113165762Sjeff * crawl under load. 2114165762Sjeff */ 2115165762Sjeff if ((td->td_pflags & TDP_SA) && td->td_proc == child->td_proc) 2116165762Sjeff return; 2117165762Sjeff#endif 2118165762Sjeff /* 2119165762Sjeff * Give the child's runtime to the parent without returning the 2120165762Sjeff * sleep time as a penalty to the parent. This causes shells that 2121165762Sjeff * launch expensive things to mark their children as expensive. 2122165762Sjeff */ 2123170293Sjeff thread_lock(td); 2124171482Sjeff td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2125164939Sjulian sched_interact_update(td); 2126165762Sjeff sched_priority(td); 2127170293Sjeff thread_unlock(td); 2128164936Sjulian} 2129164936Sjulian 2130171482Sjeff/* 2131171482Sjeff * Fix priorities on return to user-space. Priorities may be elevated due 2132171482Sjeff * to static priorities in msleep() or similar. 2133171482Sjeff */ 2134164936Sjulianvoid 2135164936Sjuliansched_userret(struct thread *td) 2136164936Sjulian{ 2137164936Sjulian /* 2138164936Sjulian * XXX we cheat slightly on the locking here to avoid locking in 2139164936Sjulian * the usual case. Setting td_priority here is essentially an 2140164936Sjulian * incomplete workaround for not setting it properly elsewhere. 2141164936Sjulian * Now that some interrupt handlers are threads, not setting it 2142164936Sjulian * properly elsewhere can clobber it in the window between setting 2143164936Sjulian * it here and returning to user mode, so don't waste time setting 2144164936Sjulian * it perfectly here. 2145164936Sjulian */ 2146164936Sjulian KASSERT((td->td_flags & TDF_BORROWING) == 0, 2147164936Sjulian ("thread with borrowed priority returning to userland")); 2148164936Sjulian if (td->td_priority != td->td_user_pri) { 2149170293Sjeff thread_lock(td); 2150164936Sjulian td->td_priority = td->td_user_pri; 2151164936Sjulian td->td_base_pri = td->td_user_pri; 2152170293Sjeff thread_unlock(td); 2153164936Sjulian } 2154164936Sjulian} 2155164936Sjulian 2156171482Sjeff/* 2157171482Sjeff * Handle a stathz tick. This is really only relevant for timeshare 2158171482Sjeff * threads. 2159171482Sjeff */ 2160164936Sjulianvoid 2161121127Sjeffsched_clock(struct thread *td) 2162109864Sjeff{ 2163164936Sjulian struct tdq *tdq; 2164164936Sjulian struct td_sched *ts; 2165109864Sjeff 2166171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2167164936Sjulian tdq = TDQ_SELF(); 2168172409Sjeff#ifdef SMP 2169133427Sjeff /* 2170172409Sjeff * We run the long term load balancer infrequently on the first cpu. 2171172409Sjeff */ 2172172409Sjeff if (balance_tdq == tdq) { 2173172409Sjeff if (balance_ticks && --balance_ticks == 0) 2174172409Sjeff sched_balance(); 2175172409Sjeff if (balance_group_ticks && --balance_group_ticks == 0) 2176172409Sjeff sched_balance_groups(); 2177172409Sjeff } 2178172409Sjeff#endif 2179172409Sjeff /* 2180165766Sjeff * Advance the insert index once for each tick to ensure that all 2181165766Sjeff * threads get a chance to run. 2182133427Sjeff */ 2183165766Sjeff if (tdq->tdq_idx == tdq->tdq_ridx) { 2184165766Sjeff tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2185165766Sjeff if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2186165766Sjeff tdq->tdq_ridx = tdq->tdq_idx; 2187165766Sjeff } 2188165766Sjeff ts = td->td_sched; 2189165762Sjeff /* 2190163709Sjb * We only do slicing code for TIMESHARE threads. 2191113357Sjeff */ 2192163709Sjb if (td->td_pri_class != PRI_TIMESHARE) 2193113357Sjeff return; 2194113357Sjeff /* 2195165766Sjeff * We used a tick; charge it to the thread so that we can compute our 2196113357Sjeff * interactivity. 2197109864Sjeff */ 2198171482Sjeff td->td_sched->ts_runtime += tickincr; 2199163709Sjb sched_interact_update(td); 2200109864Sjeff /* 2201109864Sjeff * We used up one time slice. 2202109864Sjeff */ 2203164936Sjulian if (--ts->ts_slice > 0) 2204113357Sjeff return; 2205109864Sjeff /* 2206113357Sjeff * We're out of time, recompute priorities and requeue. 2207109864Sjeff */ 2208165796Sjeff sched_priority(td); 2209113357Sjeff td->td_flags |= TDF_NEEDRESCHED; 2210109864Sjeff} 2211109864Sjeff 2212171482Sjeff/* 2213171482Sjeff * Called once per hz tick. Used for cpu utilization information. This 2214171482Sjeff * is easier than trying to scale based on stathz. 2215171482Sjeff */ 2216171482Sjeffvoid 2217171482Sjeffsched_tick(void) 2218171482Sjeff{ 2219171482Sjeff struct td_sched *ts; 2220171482Sjeff 2221171482Sjeff ts = curthread->td_sched; 2222171482Sjeff /* Adjust ticks for pctcpu */ 2223171482Sjeff ts->ts_ticks += 1 << SCHED_TICK_SHIFT; 2224171482Sjeff ts->ts_ltick = ticks; 2225171482Sjeff /* 2226171482Sjeff * Update if we've exceeded our desired tick threshhold by over one 2227171482Sjeff * second. 2228171482Sjeff */ 2229171482Sjeff if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2230171482Sjeff sched_pctcpu_update(ts); 2231171482Sjeff} 2232171482Sjeff 2233171482Sjeff/* 2234171482Sjeff * Return whether the current CPU has runnable tasks. Used for in-kernel 2235171482Sjeff * cooperative idle threads. 2236171482Sjeff */ 2237109864Sjeffint 2238109864Sjeffsched_runnable(void) 2239109864Sjeff{ 2240164936Sjulian struct tdq *tdq; 2241115998Sjeff int load; 2242109864Sjeff 2243115998Sjeff load = 1; 2244115998Sjeff 2245164936Sjulian tdq = TDQ_SELF(); 2246121605Sjeff if ((curthread->td_flags & TDF_IDLETD) != 0) { 2247165620Sjeff if (tdq->tdq_load > 0) 2248121605Sjeff goto out; 2249121605Sjeff } else 2250165620Sjeff if (tdq->tdq_load - 1 > 0) 2251121605Sjeff goto out; 2252115998Sjeff load = 0; 2253115998Sjeffout: 2254115998Sjeff return (load); 2255109864Sjeff} 2256109864Sjeff 2257171482Sjeff/* 2258171482Sjeff * Choose the highest priority thread to run. The thread is removed from 2259171482Sjeff * the run-queue while running however the load remains. For SMP we set 2260171482Sjeff * the tdq in the global idle bitmask if it idles here. 2261171482Sjeff */ 2262166190Sjeffstruct thread * 2263109970Sjeffsched_choose(void) 2264109970Sjeff{ 2265171482Sjeff#ifdef SMP 2266171482Sjeff struct tdq_group *tdg; 2267171482Sjeff#endif 2268171482Sjeff struct td_sched *ts; 2269164936Sjulian struct tdq *tdq; 2270109970Sjeff 2271164936Sjulian tdq = TDQ_SELF(); 2272171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2273164936Sjulian ts = tdq_choose(tdq); 2274164936Sjulian if (ts) { 2275164936Sjulian tdq_runq_rem(tdq, ts); 2276166190Sjeff return (ts->ts_thread); 2277109864Sjeff } 2278109970Sjeff#ifdef SMP 2279171482Sjeff /* 2280171482Sjeff * We only set the idled bit when all of the cpus in the group are 2281171482Sjeff * idle. Otherwise we could get into a situation where a thread bounces 2282171482Sjeff * back and forth between two idle cores on seperate physical CPUs. 2283171482Sjeff */ 2284171482Sjeff tdg = tdq->tdq_group; 2285171482Sjeff tdg->tdg_idlemask |= PCPU_GET(cpumask); 2286171482Sjeff if (tdg->tdg_idlemask == tdg->tdg_cpumask) 2287171482Sjeff atomic_set_int(&tdq_idle, tdg->tdg_mask); 2288171482Sjeff tdq->tdq_lowpri = PRI_MAX_IDLE; 2289109970Sjeff#endif 2290166190Sjeff return (PCPU_GET(idlethread)); 2291109864Sjeff} 2292109864Sjeff 2293171482Sjeff/* 2294171482Sjeff * Set owepreempt if necessary. Preemption never happens directly in ULE, 2295171482Sjeff * we always request it once we exit a critical section. 2296171482Sjeff */ 2297171482Sjeffstatic inline void 2298171482Sjeffsched_setpreempt(struct thread *td) 2299166190Sjeff{ 2300166190Sjeff struct thread *ctd; 2301166190Sjeff int cpri; 2302166190Sjeff int pri; 2303166190Sjeff 2304166190Sjeff ctd = curthread; 2305166190Sjeff pri = td->td_priority; 2306166190Sjeff cpri = ctd->td_priority; 2307171482Sjeff if (td->td_priority < ctd->td_priority) 2308171482Sjeff curthread->td_flags |= TDF_NEEDRESCHED; 2309166190Sjeff if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2310171482Sjeff return; 2311166190Sjeff /* 2312166190Sjeff * Always preempt IDLE threads. Otherwise only if the preempting 2313166190Sjeff * thread is an ithread. 2314166190Sjeff */ 2315171482Sjeff if (pri > preempt_thresh && cpri < PRI_MIN_IDLE) 2316171482Sjeff return; 2317171482Sjeff ctd->td_owepreempt = 1; 2318171482Sjeff return; 2319166190Sjeff} 2320166190Sjeff 2321171482Sjeff/* 2322171482Sjeff * Add a thread to a thread queue. Initializes priority, slice, runq, and 2323171482Sjeff * add it to the appropriate queue. This is the internal function called 2324171482Sjeff * when the tdq is predetermined. 2325171482Sjeff */ 2326109864Sjeffvoid 2327171482Sjefftdq_add(struct tdq *tdq, struct thread *td, int flags) 2328109864Sjeff{ 2329164936Sjulian struct td_sched *ts; 2330121790Sjeff int class; 2331166108Sjeff#ifdef SMP 2332166108Sjeff int cpumask; 2333166108Sjeff#endif 2334109864Sjeff 2335171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2336166190Sjeff KASSERT((td->td_inhibitors == 0), 2337166190Sjeff ("sched_add: trying to run inhibited thread")); 2338166190Sjeff KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2339166190Sjeff ("sched_add: bad thread state")); 2340172207Sjeff KASSERT(td->td_flags & TDF_INMEM, 2341172207Sjeff ("sched_add: thread swapped out")); 2342171482Sjeff 2343171482Sjeff ts = td->td_sched; 2344171482Sjeff class = PRI_BASE(td->td_pri_class); 2345166190Sjeff TD_SET_RUNQ(td); 2346166190Sjeff if (ts->ts_slice == 0) 2347166190Sjeff ts->ts_slice = sched_slice; 2348133427Sjeff /* 2349171482Sjeff * Pick the run queue based on priority. 2350133427Sjeff */ 2351171482Sjeff if (td->td_priority <= PRI_MAX_REALTIME) 2352171482Sjeff ts->ts_runq = &tdq->tdq_realtime; 2353171482Sjeff else if (td->td_priority <= PRI_MAX_TIMESHARE) 2354171482Sjeff ts->ts_runq = &tdq->tdq_timeshare; 2355171482Sjeff else 2356171482Sjeff ts->ts_runq = &tdq->tdq_idle; 2357171482Sjeff#ifdef SMP 2358166108Sjeff cpumask = 1 << ts->ts_cpu; 2359121790Sjeff /* 2360123685Sjeff * If we had been idle, clear our bit in the group and potentially 2361166108Sjeff * the global bitmap. 2362121790Sjeff */ 2363165762Sjeff if ((class != PRI_IDLE && class != PRI_ITHD) && 2364166108Sjeff (tdq->tdq_group->tdg_idlemask & cpumask) != 0) { 2365121790Sjeff /* 2366123433Sjeff * Check to see if our group is unidling, and if so, remove it 2367123433Sjeff * from the global idle mask. 2368121790Sjeff */ 2369165620Sjeff if (tdq->tdq_group->tdg_idlemask == 2370165620Sjeff tdq->tdq_group->tdg_cpumask) 2371165620Sjeff atomic_clear_int(&tdq_idle, tdq->tdq_group->tdg_mask); 2372123433Sjeff /* 2373123433Sjeff * Now remove ourselves from the group specific idle mask. 2374123433Sjeff */ 2375166108Sjeff tdq->tdq_group->tdg_idlemask &= ~cpumask; 2376166108Sjeff } 2377171482Sjeff if (td->td_priority < tdq->tdq_lowpri) 2378171482Sjeff tdq->tdq_lowpri = td->td_priority; 2379121790Sjeff#endif 2380171482Sjeff tdq_runq_add(tdq, ts, flags); 2381171482Sjeff tdq_load_add(tdq, ts); 2382171482Sjeff} 2383171482Sjeff 2384171482Sjeff/* 2385171482Sjeff * Select the target thread queue and add a thread to it. Request 2386171482Sjeff * preemption or IPI a remote processor if required. 2387171482Sjeff */ 2388171482Sjeffvoid 2389171482Sjeffsched_add(struct thread *td, int flags) 2390171482Sjeff{ 2391171482Sjeff struct td_sched *ts; 2392171482Sjeff struct tdq *tdq; 2393171482Sjeff#ifdef SMP 2394171482Sjeff int cpuid; 2395171482Sjeff int cpu; 2396171482Sjeff#endif 2397171482Sjeff CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 2398173600Sjulian td, td->td_name, td->td_priority, curthread, 2399173600Sjulian curthread->td_name); 2400171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2401171482Sjeff ts = td->td_sched; 2402166108Sjeff /* 2403171482Sjeff * Recalculate the priority before we select the target cpu or 2404171482Sjeff * run-queue. 2405166108Sjeff */ 2406171482Sjeff if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2407171482Sjeff sched_priority(td); 2408171482Sjeff#ifdef SMP 2409171482Sjeff cpuid = PCPU_GET(cpuid); 2410171482Sjeff /* 2411171482Sjeff * Pick the destination cpu and if it isn't ours transfer to the 2412171482Sjeff * target cpu. 2413171482Sjeff */ 2414171482Sjeff if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_MIGRATE(td)) 2415171482Sjeff cpu = cpuid; 2416171482Sjeff else if (!THREAD_CAN_MIGRATE(td)) 2417171482Sjeff cpu = ts->ts_cpu; 2418166108Sjeff else 2419171482Sjeff cpu = sched_pickcpu(ts, flags); 2420171482Sjeff tdq = sched_setcpu(ts, cpu, flags); 2421171482Sjeff tdq_add(tdq, td, flags); 2422171482Sjeff if (cpu != cpuid) { 2423166108Sjeff tdq_notify(ts); 2424166108Sjeff return; 2425166108Sjeff } 2426171482Sjeff#else 2427171482Sjeff tdq = TDQ_SELF(); 2428171482Sjeff TDQ_LOCK(tdq); 2429171482Sjeff /* 2430171482Sjeff * Now that the thread is moving to the run-queue, set the lock 2431171482Sjeff * to the scheduler's lock. 2432171482Sjeff */ 2433171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2434171482Sjeff tdq_add(tdq, td, flags); 2435166108Sjeff#endif 2436171482Sjeff if (!(flags & SRQ_YIELDING)) 2437171482Sjeff sched_setpreempt(td); 2438109864Sjeff} 2439109864Sjeff 2440171482Sjeff/* 2441171482Sjeff * Remove a thread from a run-queue without running it. This is used 2442171482Sjeff * when we're stealing a thread from a remote queue. Otherwise all threads 2443171482Sjeff * exit by calling sched_exit_thread() and sched_throw() themselves. 2444171482Sjeff */ 2445109864Sjeffvoid 2446121127Sjeffsched_rem(struct thread *td) 2447109864Sjeff{ 2448164936Sjulian struct tdq *tdq; 2449164936Sjulian struct td_sched *ts; 2450113357Sjeff 2451139316Sjeff CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 2452173600Sjulian td, td->td_name, td->td_priority, curthread, 2453173600Sjulian curthread->td_name); 2454164936Sjulian ts = td->td_sched; 2455171482Sjeff tdq = TDQ_CPU(ts->ts_cpu); 2456171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2457171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2458166190Sjeff KASSERT(TD_ON_RUNQ(td), 2459164936Sjulian ("sched_rem: thread not on run queue")); 2460164936Sjulian tdq_runq_rem(tdq, ts); 2461164936Sjulian tdq_load_rem(tdq, ts); 2462166190Sjeff TD_SET_CAN_RUN(td); 2463109864Sjeff} 2464109864Sjeff 2465171482Sjeff/* 2466171482Sjeff * Fetch cpu utilization information. Updates on demand. 2467171482Sjeff */ 2468109864Sjefffixpt_t 2469121127Sjeffsched_pctcpu(struct thread *td) 2470109864Sjeff{ 2471109864Sjeff fixpt_t pctcpu; 2472164936Sjulian struct td_sched *ts; 2473109864Sjeff 2474109864Sjeff pctcpu = 0; 2475164936Sjulian ts = td->td_sched; 2476164936Sjulian if (ts == NULL) 2477121290Sjeff return (0); 2478109864Sjeff 2479170293Sjeff thread_lock(td); 2480164936Sjulian if (ts->ts_ticks) { 2481109864Sjeff int rtick; 2482109864Sjeff 2483165796Sjeff sched_pctcpu_update(ts); 2484109864Sjeff /* How many rtick per second ? */ 2485165762Sjeff rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2486165762Sjeff pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2487109864Sjeff } 2488170293Sjeff thread_unlock(td); 2489109864Sjeff 2490109864Sjeff return (pctcpu); 2491109864Sjeff} 2492109864Sjeff 2493171482Sjeff/* 2494171482Sjeff * Bind a thread to a target cpu. 2495171482Sjeff */ 2496122038Sjeffvoid 2497122038Sjeffsched_bind(struct thread *td, int cpu) 2498122038Sjeff{ 2499164936Sjulian struct td_sched *ts; 2500122038Sjeff 2501171713Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2502164936Sjulian ts = td->td_sched; 2503166137Sjeff if (ts->ts_flags & TSF_BOUND) 2504166152Sjeff sched_unbind(td); 2505164936Sjulian ts->ts_flags |= TSF_BOUND; 2506123433Sjeff#ifdef SMP 2507166137Sjeff sched_pin(); 2508123433Sjeff if (PCPU_GET(cpuid) == cpu) 2509122038Sjeff return; 2510166137Sjeff ts->ts_cpu = cpu; 2511122038Sjeff /* When we return from mi_switch we'll be on the correct cpu. */ 2512131527Sphk mi_switch(SW_VOL, NULL); 2513122038Sjeff#endif 2514122038Sjeff} 2515122038Sjeff 2516171482Sjeff/* 2517171482Sjeff * Release a bound thread. 2518171482Sjeff */ 2519122038Sjeffvoid 2520122038Sjeffsched_unbind(struct thread *td) 2521122038Sjeff{ 2522165762Sjeff struct td_sched *ts; 2523165762Sjeff 2524170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2525165762Sjeff ts = td->td_sched; 2526166137Sjeff if ((ts->ts_flags & TSF_BOUND) == 0) 2527166137Sjeff return; 2528165762Sjeff ts->ts_flags &= ~TSF_BOUND; 2529165762Sjeff#ifdef SMP 2530165762Sjeff sched_unpin(); 2531165762Sjeff#endif 2532122038Sjeff} 2533122038Sjeff 2534109864Sjeffint 2535145256Sjkoshysched_is_bound(struct thread *td) 2536145256Sjkoshy{ 2537170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2538164936Sjulian return (td->td_sched->ts_flags & TSF_BOUND); 2539145256Sjkoshy} 2540145256Sjkoshy 2541171482Sjeff/* 2542171482Sjeff * Basic yield call. 2543171482Sjeff */ 2544159630Sdavidxuvoid 2545159630Sdavidxusched_relinquish(struct thread *td) 2546159630Sdavidxu{ 2547170293Sjeff thread_lock(td); 2548170293Sjeff SCHED_STAT_INC(switch_relinquish); 2549159630Sdavidxu mi_switch(SW_VOL, NULL); 2550170293Sjeff thread_unlock(td); 2551159630Sdavidxu} 2552159630Sdavidxu 2553171482Sjeff/* 2554171482Sjeff * Return the total system load. 2555171482Sjeff */ 2556145256Sjkoshyint 2557125289Sjeffsched_load(void) 2558125289Sjeff{ 2559125289Sjeff#ifdef SMP 2560125289Sjeff int total; 2561125289Sjeff int i; 2562125289Sjeff 2563125289Sjeff total = 0; 2564165620Sjeff for (i = 0; i <= tdg_maxid; i++) 2565165620Sjeff total += TDQ_GROUP(i)->tdg_load; 2566125289Sjeff return (total); 2567125289Sjeff#else 2568165620Sjeff return (TDQ_SELF()->tdq_sysload); 2569125289Sjeff#endif 2570125289Sjeff} 2571125289Sjeff 2572125289Sjeffint 2573109864Sjeffsched_sizeof_proc(void) 2574109864Sjeff{ 2575109864Sjeff return (sizeof(struct proc)); 2576109864Sjeff} 2577109864Sjeff 2578109864Sjeffint 2579109864Sjeffsched_sizeof_thread(void) 2580109864Sjeff{ 2581109864Sjeff return (sizeof(struct thread) + sizeof(struct td_sched)); 2582109864Sjeff} 2583159570Sdavidxu 2584166190Sjeff/* 2585166190Sjeff * The actual idle process. 2586166190Sjeff */ 2587166190Sjeffvoid 2588166190Sjeffsched_idletd(void *dummy) 2589166190Sjeff{ 2590166190Sjeff struct thread *td; 2591171482Sjeff struct tdq *tdq; 2592166190Sjeff 2593166190Sjeff td = curthread; 2594171482Sjeff tdq = TDQ_SELF(); 2595166190Sjeff mtx_assert(&Giant, MA_NOTOWNED); 2596171482Sjeff /* ULE relies on preemption for idle interruption. */ 2597171482Sjeff for (;;) { 2598171482Sjeff#ifdef SMP 2599171482Sjeff if (tdq_idled(tdq)) 2600171482Sjeff cpu_idle(); 2601171482Sjeff#else 2602166190Sjeff cpu_idle(); 2603171482Sjeff#endif 2604171482Sjeff } 2605166190Sjeff} 2606166190Sjeff 2607170293Sjeff/* 2608170293Sjeff * A CPU is entering for the first time or a thread is exiting. 2609170293Sjeff */ 2610170293Sjeffvoid 2611170293Sjeffsched_throw(struct thread *td) 2612170293Sjeff{ 2613172411Sjeff struct thread *newtd; 2614171482Sjeff struct tdq *tdq; 2615171482Sjeff 2616171482Sjeff tdq = TDQ_SELF(); 2617170293Sjeff if (td == NULL) { 2618171482Sjeff /* Correct spinlock nesting and acquire the correct lock. */ 2619171482Sjeff TDQ_LOCK(tdq); 2620170293Sjeff spinlock_exit(); 2621170293Sjeff } else { 2622171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2623171482Sjeff tdq_load_rem(tdq, td->td_sched); 2624174629Sjeff lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2625170293Sjeff } 2626170293Sjeff KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2627172411Sjeff newtd = choosethread(); 2628172411Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2629170293Sjeff PCPU_SET(switchtime, cpu_ticks()); 2630170293Sjeff PCPU_SET(switchticks, ticks); 2631172411Sjeff cpu_throw(td, newtd); /* doesn't return */ 2632170293Sjeff} 2633170293Sjeff 2634171482Sjeff/* 2635171482Sjeff * This is called from fork_exit(). Just acquire the correct locks and 2636171482Sjeff * let fork do the rest of the work. 2637171482Sjeff */ 2638170293Sjeffvoid 2639170600Sjeffsched_fork_exit(struct thread *td) 2640170293Sjeff{ 2641171482Sjeff struct td_sched *ts; 2642171482Sjeff struct tdq *tdq; 2643171482Sjeff int cpuid; 2644170293Sjeff 2645170293Sjeff /* 2646170293Sjeff * Finish setting up thread glue so that it begins execution in a 2647171482Sjeff * non-nested critical section with the scheduler lock held. 2648170293Sjeff */ 2649171482Sjeff cpuid = PCPU_GET(cpuid); 2650171482Sjeff tdq = TDQ_CPU(cpuid); 2651171482Sjeff ts = td->td_sched; 2652171482Sjeff if (TD_IS_IDLETHREAD(td)) 2653171482Sjeff td->td_lock = TDQ_LOCKPTR(tdq); 2654171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2655171482Sjeff td->td_oncpu = cpuid; 2656172411Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2657174629Sjeff lock_profile_obtain_lock_success( 2658174629Sjeff &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2659170293Sjeff} 2660170293Sjeff 2661171482Sjeffstatic SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, 2662171482Sjeff "Scheduler"); 2663171482SjeffSYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2664165762Sjeff "Scheduler name"); 2665171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2666171482Sjeff "Slice size for timeshare threads"); 2667171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2668171482Sjeff "Interactivity score threshold"); 2669171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2670171482Sjeff 0,"Min priority for preemption, lower priorities have greater precedence"); 2671166108Sjeff#ifdef SMP 2672171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, pick_pri, CTLFLAG_RW, &pick_pri, 0, 2673171482Sjeff "Pick the target cpu based on priority rather than load."); 2674171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2675171482Sjeff "Number of hz ticks to keep thread affinity for"); 2676171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, tryself, CTLFLAG_RW, &tryself, 0, ""); 2677171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2678171482Sjeff "Enables the long-term load balancer"); 2679172409SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2680172409Sjeff &balance_interval, 0, 2681172409Sjeff "Average frequency in stathz ticks to run the long-term balancer"); 2682171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, 2683171482Sjeff "Steals work from another hyper-threaded core on idle"); 2684171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2685171482Sjeff "Attempts to steal work from other cores before idling"); 2686171506SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2687171506Sjeff "Minimum load on remote cpu before we'll steal"); 2688171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, topology, CTLFLAG_RD, &topology, 0, 2689171482Sjeff "True when a topology has been specified by the MD code."); 2690166108Sjeff#endif 2691165762Sjeff 2692172264Sjeff/* ps compat. All cpu percentages from ULE are weighted. */ 2693172293Sjeffstatic int ccpu = 0; 2694165762SjeffSYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2695165762Sjeff 2696165762Sjeff 2697134791Sjulian#define KERN_SWITCH_INCLUDE 1 2698134791Sjulian#include "kern/kern_switch.c" 2699