sched_ule.c revision 172409
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 172409 2007-10-02 00:36:06Z 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 74172345Sjeff#if !defined(__i386__) && !defined(__amd64__) 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", 342165762Sjeff ts->ts_thread, ts->ts_thread->td_proc->p_comm, 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 load; 744171482Sjeff int cpu; 745123433Sjeff 746171482Sjeff /* We don't want to be preempted while we're iterating over tdqs */ 747171482Sjeff spinlock_enter(); 748165620Sjeff tdg = tdq->tdq_group; 749123433Sjeff /* 750165620Sjeff * If we're in a cpu group, try and steal threads from another cpu in 751172409Sjeff * the group before idling. In a HTT group all cpus share the same 752172409Sjeff * run-queue lock, however, we still need a recursive lock to 753172409Sjeff * call tdq_move(). 754123433Sjeff */ 755166108Sjeff if (steal_htt && tdg->tdg_cpus > 1 && tdg->tdg_transferable) { 756172409Sjeff TDQ_LOCK(tdq); 757165620Sjeff LIST_FOREACH(steal, &tdg->tdg_members, tdq_siblings) { 758165620Sjeff if (steal == tdq || steal->tdq_transferable == 0) 759123433Sjeff continue; 760171482Sjeff TDQ_LOCK(steal); 761172409Sjeff goto steal; 762166108Sjeff } 763172409Sjeff TDQ_UNLOCK(tdq); 764166108Sjeff } 765171482Sjeff for (;;) { 766171482Sjeff if (steal_idle == 0) 767171482Sjeff break; 768171482Sjeff highcpu = 0; 769171482Sjeff highload = 0; 770171482Sjeff for (cpu = 0; cpu <= mp_maxid; cpu++) { 771171482Sjeff if (CPU_ABSENT(cpu)) 772171482Sjeff continue; 773166108Sjeff steal = TDQ_CPU(cpu); 774171482Sjeff load = TDQ_CPU(cpu)->tdq_transferable; 775171482Sjeff if (load < highload) 776166108Sjeff continue; 777171482Sjeff highload = load; 778171482Sjeff highcpu = cpu; 779171482Sjeff } 780171506Sjeff if (highload < steal_thresh) 781171482Sjeff break; 782171482Sjeff steal = TDQ_CPU(highcpu); 783172409Sjeff tdq_lock_pair(tdq, steal); 784172409Sjeff if (steal->tdq_transferable >= steal_thresh) 785166108Sjeff goto steal; 786172409Sjeff tdq_unlock_pair(tdq, steal); 787171482Sjeff break; 788123433Sjeff } 789171482Sjeff spinlock_exit(); 790123433Sjeff return (1); 791166108Sjeffsteal: 792171482Sjeff spinlock_exit(); 793172409Sjeff tdq_move(steal, tdq); 794171482Sjeff TDQ_UNLOCK(steal); 795171482Sjeff mi_switch(SW_VOL, NULL); 796171482Sjeff thread_unlock(curthread); 797121790Sjeff 798166108Sjeff return (0); 799121790Sjeff} 800121790Sjeff 801171482Sjeff/* 802171482Sjeff * Notify a remote cpu of new work. Sends an IPI if criteria are met. 803171482Sjeff */ 804121790Sjeffstatic void 805166108Sjefftdq_notify(struct td_sched *ts) 806121790Sjeff{ 807166247Sjeff struct thread *ctd; 808121790Sjeff struct pcpu *pcpu; 809166247Sjeff int cpri; 810166247Sjeff int pri; 811166108Sjeff int cpu; 812121790Sjeff 813166108Sjeff cpu = ts->ts_cpu; 814166247Sjeff pri = ts->ts_thread->td_priority; 815166108Sjeff pcpu = pcpu_find(cpu); 816166247Sjeff ctd = pcpu->pc_curthread; 817166247Sjeff cpri = ctd->td_priority; 818166137Sjeff 819121790Sjeff /* 820166137Sjeff * If our priority is not better than the current priority there is 821166137Sjeff * nothing to do. 822166137Sjeff */ 823166247Sjeff if (pri > cpri) 824166137Sjeff return; 825166137Sjeff /* 826166247Sjeff * Always IPI idle. 827121790Sjeff */ 828166247Sjeff if (cpri > PRI_MIN_IDLE) 829166247Sjeff goto sendipi; 830166247Sjeff /* 831166247Sjeff * If we're realtime or better and there is timeshare or worse running 832166247Sjeff * send an IPI. 833166247Sjeff */ 834166247Sjeff if (pri < PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME) 835166247Sjeff goto sendipi; 836166247Sjeff /* 837166247Sjeff * Otherwise only IPI if we exceed the threshold. 838166247Sjeff */ 839171482Sjeff if (pri > preempt_thresh) 840165819Sjeff return; 841166247Sjeffsendipi: 842166247Sjeff ctd->td_flags |= TDF_NEEDRESCHED; 843171482Sjeff ipi_selected(1 << cpu, IPI_PREEMPT); 844121790Sjeff} 845121790Sjeff 846171482Sjeff/* 847171482Sjeff * Steals load from a timeshare queue. Honors the rotating queue head 848171482Sjeff * index. 849171482Sjeff */ 850164936Sjulianstatic struct td_sched * 851171482Sjeffrunq_steal_from(struct runq *rq, u_char start) 852171482Sjeff{ 853171482Sjeff struct td_sched *ts; 854171482Sjeff struct rqbits *rqb; 855171482Sjeff struct rqhead *rqh; 856171482Sjeff int first; 857171482Sjeff int bit; 858171482Sjeff int pri; 859171482Sjeff int i; 860171482Sjeff 861171482Sjeff rqb = &rq->rq_status; 862171482Sjeff bit = start & (RQB_BPW -1); 863171482Sjeff pri = 0; 864171482Sjeff first = 0; 865171482Sjeffagain: 866171482Sjeff for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 867171482Sjeff if (rqb->rqb_bits[i] == 0) 868171482Sjeff continue; 869171482Sjeff if (bit != 0) { 870171482Sjeff for (pri = bit; pri < RQB_BPW; pri++) 871171482Sjeff if (rqb->rqb_bits[i] & (1ul << pri)) 872171482Sjeff break; 873171482Sjeff if (pri >= RQB_BPW) 874171482Sjeff continue; 875171482Sjeff } else 876171482Sjeff pri = RQB_FFS(rqb->rqb_bits[i]); 877171482Sjeff pri += (i << RQB_L2BPW); 878171482Sjeff rqh = &rq->rq_queues[pri]; 879171482Sjeff TAILQ_FOREACH(ts, rqh, ts_procq) { 880171482Sjeff if (first && THREAD_CAN_MIGRATE(ts->ts_thread)) 881171482Sjeff return (ts); 882171482Sjeff first = 1; 883171482Sjeff } 884171482Sjeff } 885171482Sjeff if (start != 0) { 886171482Sjeff start = 0; 887171482Sjeff goto again; 888171482Sjeff } 889171482Sjeff 890171482Sjeff return (NULL); 891171482Sjeff} 892171482Sjeff 893171482Sjeff/* 894171482Sjeff * Steals load from a standard linear queue. 895171482Sjeff */ 896171482Sjeffstatic struct td_sched * 897121790Sjeffrunq_steal(struct runq *rq) 898121790Sjeff{ 899121790Sjeff struct rqhead *rqh; 900121790Sjeff struct rqbits *rqb; 901164936Sjulian struct td_sched *ts; 902121790Sjeff int word; 903121790Sjeff int bit; 904121790Sjeff 905121790Sjeff rqb = &rq->rq_status; 906121790Sjeff for (word = 0; word < RQB_LEN; word++) { 907121790Sjeff if (rqb->rqb_bits[word] == 0) 908121790Sjeff continue; 909121790Sjeff for (bit = 0; bit < RQB_BPW; bit++) { 910123231Speter if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 911121790Sjeff continue; 912121790Sjeff rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 913171506Sjeff TAILQ_FOREACH(ts, rqh, ts_procq) 914171506Sjeff if (THREAD_CAN_MIGRATE(ts->ts_thread)) 915164936Sjulian return (ts); 916121790Sjeff } 917121790Sjeff } 918121790Sjeff return (NULL); 919121790Sjeff} 920121790Sjeff 921171482Sjeff/* 922171482Sjeff * Attempt to steal a thread in priority order from a thread queue. 923171482Sjeff */ 924164936Sjulianstatic struct td_sched * 925172409Sjefftdq_steal(struct tdq *tdq) 926121790Sjeff{ 927164936Sjulian struct td_sched *ts; 928121790Sjeff 929171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 930165762Sjeff if ((ts = runq_steal(&tdq->tdq_realtime)) != NULL) 931164936Sjulian return (ts); 932171482Sjeff if ((ts = runq_steal_from(&tdq->tdq_timeshare, tdq->tdq_ridx)) != NULL) 933164936Sjulian return (ts); 934172409Sjeff return (runq_steal(&tdq->tdq_idle)); 935121790Sjeff} 936123433Sjeff 937171482Sjeff/* 938171482Sjeff * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 939172409Sjeff * current lock and returns with the assigned queue locked. 940171482Sjeff */ 941171482Sjeffstatic inline struct tdq * 942171482Sjeffsched_setcpu(struct td_sched *ts, int cpu, int flags) 943123433Sjeff{ 944171482Sjeff struct thread *td; 945171482Sjeff struct tdq *tdq; 946123433Sjeff 947171482Sjeff THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 948171482Sjeff 949171482Sjeff tdq = TDQ_CPU(cpu); 950171482Sjeff td = ts->ts_thread; 951171482Sjeff ts->ts_cpu = cpu; 952171713Sjeff 953171713Sjeff /* If the lock matches just return the queue. */ 954171482Sjeff if (td->td_lock == TDQ_LOCKPTR(tdq)) 955171482Sjeff return (tdq); 956171482Sjeff#ifdef notyet 957123433Sjeff /* 958172293Sjeff * If the thread isn't running its lockptr is a 959171482Sjeff * turnstile or a sleepqueue. We can just lock_set without 960171482Sjeff * blocking. 961123685Sjeff */ 962171482Sjeff if (TD_CAN_RUN(td)) { 963171482Sjeff TDQ_LOCK(tdq); 964171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 965171482Sjeff return (tdq); 966171482Sjeff } 967171482Sjeff#endif 968166108Sjeff /* 969171482Sjeff * The hard case, migration, we need to block the thread first to 970171482Sjeff * prevent order reversals with other cpus locks. 971166108Sjeff */ 972171482Sjeff thread_lock_block(td); 973171482Sjeff TDQ_LOCK(tdq); 974171713Sjeff thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 975171482Sjeff return (tdq); 976166108Sjeff} 977166108Sjeff 978171482Sjeff/* 979171482Sjeff * Find the thread queue running the lowest priority thread. 980171482Sjeff */ 981166108Sjeffstatic int 982171482Sjefftdq_lowestpri(void) 983166108Sjeff{ 984171482Sjeff struct tdq *tdq; 985166108Sjeff int lowpri; 986166108Sjeff int lowcpu; 987166108Sjeff int lowload; 988166108Sjeff int load; 989171482Sjeff int cpu; 990171482Sjeff int pri; 991171482Sjeff 992171482Sjeff lowload = 0; 993171482Sjeff lowpri = lowcpu = 0; 994171482Sjeff for (cpu = 0; cpu <= mp_maxid; cpu++) { 995171482Sjeff if (CPU_ABSENT(cpu)) 996171482Sjeff continue; 997171482Sjeff tdq = TDQ_CPU(cpu); 998171482Sjeff pri = tdq->tdq_lowpri; 999171482Sjeff load = TDQ_CPU(cpu)->tdq_load; 1000171482Sjeff CTR4(KTR_ULE, 1001171482Sjeff "cpu %d pri %d lowcpu %d lowpri %d", 1002171482Sjeff cpu, pri, lowcpu, lowpri); 1003171482Sjeff if (pri < lowpri) 1004171482Sjeff continue; 1005171482Sjeff if (lowpri && lowpri == pri && load > lowload) 1006171482Sjeff continue; 1007171482Sjeff lowpri = pri; 1008171482Sjeff lowcpu = cpu; 1009171482Sjeff lowload = load; 1010171482Sjeff } 1011171482Sjeff 1012171482Sjeff return (lowcpu); 1013171482Sjeff} 1014171482Sjeff 1015171482Sjeff/* 1016171482Sjeff * Find the thread queue with the least load. 1017171482Sjeff */ 1018171482Sjeffstatic int 1019171482Sjefftdq_lowestload(void) 1020171482Sjeff{ 1021171482Sjeff struct tdq *tdq; 1022171482Sjeff int lowload; 1023171482Sjeff int lowpri; 1024171482Sjeff int lowcpu; 1025171482Sjeff int load; 1026171482Sjeff int cpu; 1027171482Sjeff int pri; 1028171482Sjeff 1029171482Sjeff lowcpu = 0; 1030171482Sjeff lowload = TDQ_CPU(0)->tdq_load; 1031171482Sjeff lowpri = TDQ_CPU(0)->tdq_lowpri; 1032171482Sjeff for (cpu = 1; cpu <= mp_maxid; cpu++) { 1033171482Sjeff if (CPU_ABSENT(cpu)) 1034171482Sjeff continue; 1035171482Sjeff tdq = TDQ_CPU(cpu); 1036171482Sjeff load = tdq->tdq_load; 1037171482Sjeff pri = tdq->tdq_lowpri; 1038171482Sjeff CTR4(KTR_ULE, "cpu %d load %d lowcpu %d lowload %d", 1039171482Sjeff cpu, load, lowcpu, lowload); 1040171482Sjeff if (load > lowload) 1041171482Sjeff continue; 1042171482Sjeff if (load == lowload && pri < lowpri) 1043171482Sjeff continue; 1044171482Sjeff lowcpu = cpu; 1045171482Sjeff lowload = load; 1046171482Sjeff lowpri = pri; 1047171482Sjeff } 1048171482Sjeff 1049171482Sjeff return (lowcpu); 1050171482Sjeff} 1051171482Sjeff 1052171482Sjeff/* 1053171482Sjeff * Pick the destination cpu for sched_add(). Respects affinity and makes 1054171482Sjeff * a determination based on load or priority of available processors. 1055171482Sjeff */ 1056171482Sjeffstatic int 1057171482Sjeffsched_pickcpu(struct td_sched *ts, int flags) 1058171482Sjeff{ 1059171482Sjeff struct tdq *tdq; 1060166108Sjeff int self; 1061166108Sjeff int pri; 1062166108Sjeff int cpu; 1063166108Sjeff 1064171482Sjeff cpu = self = PCPU_GET(cpuid); 1065166108Sjeff if (smp_started == 0) 1066166108Sjeff return (self); 1067171506Sjeff /* 1068171506Sjeff * Don't migrate a running thread from sched_switch(). 1069171506Sjeff */ 1070171506Sjeff if (flags & SRQ_OURSELF) { 1071171506Sjeff CTR1(KTR_ULE, "YIELDING %d", 1072171506Sjeff curthread->td_priority); 1073171506Sjeff return (self); 1074171506Sjeff } 1075166108Sjeff pri = ts->ts_thread->td_priority; 1076171482Sjeff cpu = ts->ts_cpu; 1077166108Sjeff /* 1078166108Sjeff * Regardless of affinity, if the last cpu is idle send it there. 1079166108Sjeff */ 1080171482Sjeff tdq = TDQ_CPU(cpu); 1081171482Sjeff if (tdq->tdq_lowpri > PRI_MIN_IDLE) { 1082166229Sjeff CTR5(KTR_ULE, 1083166108Sjeff "ts_cpu %d idle, ltick %d ticks %d pri %d curthread %d", 1084166108Sjeff ts->ts_cpu, ts->ts_rltick, ticks, pri, 1085171482Sjeff tdq->tdq_lowpri); 1086166108Sjeff return (ts->ts_cpu); 1087123433Sjeff } 1088166108Sjeff /* 1089166108Sjeff * If we have affinity, try to place it on the cpu we last ran on. 1090166108Sjeff */ 1091171482Sjeff if (SCHED_AFFINITY(ts) && tdq->tdq_lowpri > pri) { 1092166229Sjeff CTR5(KTR_ULE, 1093166108Sjeff "affinity for %d, ltick %d ticks %d pri %d curthread %d", 1094166108Sjeff ts->ts_cpu, ts->ts_rltick, ticks, pri, 1095171482Sjeff tdq->tdq_lowpri); 1096166108Sjeff return (ts->ts_cpu); 1097139334Sjeff } 1098123433Sjeff /* 1099166108Sjeff * Look for an idle group. 1100123433Sjeff */ 1101166229Sjeff CTR1(KTR_ULE, "tdq_idle %X", tdq_idle); 1102166108Sjeff cpu = ffs(tdq_idle); 1103166108Sjeff if (cpu) 1104171482Sjeff return (--cpu); 1105171506Sjeff /* 1106172409Sjeff * If there are no idle cores see if we can run the thread locally. 1107172409Sjeff * This may improve locality among sleepers and wakers when there 1108172409Sjeff * is shared data. 1109171506Sjeff */ 1110171506Sjeff if (tryself && pri < curthread->td_priority) { 1111171506Sjeff CTR1(KTR_ULE, "tryself %d", 1112166108Sjeff curthread->td_priority); 1113166108Sjeff return (self); 1114123433Sjeff } 1115133427Sjeff /* 1116166108Sjeff * Now search for the cpu running the lowest priority thread with 1117166108Sjeff * the least load. 1118123433Sjeff */ 1119171482Sjeff if (pick_pri) 1120171482Sjeff cpu = tdq_lowestpri(); 1121171482Sjeff else 1122171482Sjeff cpu = tdq_lowestload(); 1123171482Sjeff return (cpu); 1124123433Sjeff} 1125123433Sjeff 1126121790Sjeff#endif /* SMP */ 1127121790Sjeff 1128117326Sjeff/* 1129121790Sjeff * Pick the highest priority task we have and return it. 1130117326Sjeff */ 1131164936Sjulianstatic struct td_sched * 1132164936Sjuliantdq_choose(struct tdq *tdq) 1133110267Sjeff{ 1134164936Sjulian struct td_sched *ts; 1135110267Sjeff 1136171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1137165762Sjeff ts = runq_choose(&tdq->tdq_realtime); 1138170787Sjeff if (ts != NULL) 1139164936Sjulian return (ts); 1140165766Sjeff ts = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1141165762Sjeff if (ts != NULL) { 1142170787Sjeff KASSERT(ts->ts_thread->td_priority >= PRI_MIN_TIMESHARE, 1143165762Sjeff ("tdq_choose: Invalid priority on timeshare queue %d", 1144165762Sjeff ts->ts_thread->td_priority)); 1145165762Sjeff return (ts); 1146165762Sjeff } 1147110267Sjeff 1148165762Sjeff ts = runq_choose(&tdq->tdq_idle); 1149165762Sjeff if (ts != NULL) { 1150165762Sjeff KASSERT(ts->ts_thread->td_priority >= PRI_MIN_IDLE, 1151165762Sjeff ("tdq_choose: Invalid priority on idle queue %d", 1152165762Sjeff ts->ts_thread->td_priority)); 1153165762Sjeff return (ts); 1154165762Sjeff } 1155165762Sjeff 1156165762Sjeff return (NULL); 1157110267Sjeff} 1158110267Sjeff 1159171482Sjeff/* 1160171482Sjeff * Initialize a thread queue. 1161171482Sjeff */ 1162109864Sjeffstatic void 1163164936Sjuliantdq_setup(struct tdq *tdq) 1164110028Sjeff{ 1165171482Sjeff 1166171713Sjeff if (bootverbose) 1167171713Sjeff printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1168165762Sjeff runq_init(&tdq->tdq_realtime); 1169165762Sjeff runq_init(&tdq->tdq_timeshare); 1170165620Sjeff runq_init(&tdq->tdq_idle); 1171165620Sjeff tdq->tdq_load = 0; 1172110028Sjeff} 1173110028Sjeff 1174171713Sjeff#ifdef SMP 1175110028Sjeffstatic void 1176171713Sjefftdg_setup(struct tdq_group *tdg) 1177109864Sjeff{ 1178171713Sjeff if (bootverbose) 1179171713Sjeff printf("ULE: setup cpu group %d\n", TDG_ID(tdg)); 1180171713Sjeff snprintf(tdg->tdg_name, sizeof(tdg->tdg_name), 1181171713Sjeff "sched lock %d", (int)TDG_ID(tdg)); 1182171713Sjeff mtx_init(&tdg->tdg_lock, tdg->tdg_name, "sched lock", 1183171713Sjeff MTX_SPIN | MTX_RECURSE); 1184171713Sjeff LIST_INIT(&tdg->tdg_members); 1185171713Sjeff tdg->tdg_load = 0; 1186171713Sjeff tdg->tdg_transferable = 0; 1187171713Sjeff tdg->tdg_cpus = 0; 1188171713Sjeff tdg->tdg_mask = 0; 1189171713Sjeff tdg->tdg_cpumask = 0; 1190171713Sjeff tdg->tdg_idlemask = 0; 1191171713Sjeff} 1192171713Sjeff 1193171713Sjeffstatic void 1194171713Sjefftdg_add(struct tdq_group *tdg, struct tdq *tdq) 1195171713Sjeff{ 1196171713Sjeff if (tdg->tdg_mask == 0) 1197171713Sjeff tdg->tdg_mask |= 1 << TDQ_ID(tdq); 1198171713Sjeff tdg->tdg_cpumask |= 1 << TDQ_ID(tdq); 1199171713Sjeff tdg->tdg_cpus++; 1200171713Sjeff tdq->tdq_group = tdg; 1201171713Sjeff tdq->tdq_lock = &tdg->tdg_lock; 1202171713Sjeff LIST_INSERT_HEAD(&tdg->tdg_members, tdq, tdq_siblings); 1203171713Sjeff if (bootverbose) 1204171713Sjeff printf("ULE: adding cpu %d to group %d: cpus %d mask 0x%X\n", 1205171713Sjeff TDQ_ID(tdq), TDG_ID(tdg), tdg->tdg_cpus, tdg->tdg_cpumask); 1206171713Sjeff} 1207171713Sjeff 1208171713Sjeffstatic void 1209171713Sjeffsched_setup_topology(void) 1210171713Sjeff{ 1211171713Sjeff struct tdq_group *tdg; 1212171713Sjeff struct cpu_group *cg; 1213171713Sjeff int balance_groups; 1214171482Sjeff struct tdq *tdq; 1215109864Sjeff int i; 1216171713Sjeff int j; 1217109864Sjeff 1218171713Sjeff topology = 1; 1219123487Sjeff balance_groups = 0; 1220171713Sjeff for (i = 0; i < smp_topology->ct_count; i++) { 1221171713Sjeff cg = &smp_topology->ct_group[i]; 1222171713Sjeff tdg = &tdq_groups[i]; 1223171713Sjeff /* 1224171713Sjeff * Initialize the group. 1225171713Sjeff */ 1226171713Sjeff tdg_setup(tdg); 1227171713Sjeff /* 1228171713Sjeff * Find all of the group members and add them. 1229171713Sjeff */ 1230171713Sjeff for (j = 0; j < MAXCPU; j++) { 1231171713Sjeff if ((cg->cg_mask & (1 << j)) != 0) { 1232171713Sjeff tdq = TDQ_CPU(j); 1233171713Sjeff tdq_setup(tdq); 1234171713Sjeff tdg_add(tdg, tdq); 1235171713Sjeff } 1236171713Sjeff } 1237171713Sjeff if (tdg->tdg_cpus > 1) 1238171713Sjeff balance_groups = 1; 1239171713Sjeff } 1240171713Sjeff tdg_maxid = smp_topology->ct_count - 1; 1241171713Sjeff if (balance_groups) 1242172409Sjeff sched_balance_groups(); 1243171713Sjeff} 1244171713Sjeff 1245171713Sjeffstatic void 1246171713Sjeffsched_setup_smp(void) 1247171713Sjeff{ 1248171713Sjeff struct tdq_group *tdg; 1249171713Sjeff struct tdq *tdq; 1250171713Sjeff int cpus; 1251171713Sjeff int i; 1252171713Sjeff 1253171713Sjeff for (cpus = 0, i = 0; i < MAXCPU; i++) { 1254171713Sjeff if (CPU_ABSENT(i)) 1255171713Sjeff continue; 1256165627Sjeff tdq = &tdq_cpu[i]; 1257171713Sjeff tdg = &tdq_groups[i]; 1258171713Sjeff /* 1259171713Sjeff * Setup a tdq group with one member. 1260171713Sjeff */ 1261171713Sjeff tdg_setup(tdg); 1262171713Sjeff tdq_setup(tdq); 1263171713Sjeff tdg_add(tdg, tdq); 1264171713Sjeff cpus++; 1265123433Sjeff } 1266171713Sjeff tdg_maxid = cpus - 1; 1267171713Sjeff} 1268123433Sjeff 1269171713Sjeff/* 1270171713Sjeff * Fake a topology with one group containing all CPUs. 1271171713Sjeff */ 1272171713Sjeffstatic void 1273171713Sjeffsched_fake_topo(void) 1274171713Sjeff{ 1275171713Sjeff#ifdef SCHED_FAKE_TOPOLOGY 1276171713Sjeff static struct cpu_top top; 1277171713Sjeff static struct cpu_group group; 1278113357Sjeff 1279171713Sjeff top.ct_count = 1; 1280171713Sjeff top.ct_group = &group; 1281171713Sjeff group.cg_mask = all_cpus; 1282171713Sjeff group.cg_count = mp_ncpus; 1283171713Sjeff group.cg_children = 0; 1284171713Sjeff smp_topology = ⊤ 1285171713Sjeff#endif 1286171713Sjeff} 1287171713Sjeff#endif 1288171713Sjeff 1289171713Sjeff/* 1290171713Sjeff * Setup the thread queues and initialize the topology based on MD 1291171713Sjeff * information. 1292171713Sjeff */ 1293171713Sjeffstatic void 1294171713Sjeffsched_setup(void *dummy) 1295171713Sjeff{ 1296171713Sjeff struct tdq *tdq; 1297171713Sjeff 1298171713Sjeff tdq = TDQ_SELF(); 1299171713Sjeff#ifdef SMP 1300171713Sjeff sched_fake_topo(); 1301171713Sjeff /* 1302171713Sjeff * Setup tdqs based on a topology configuration or vanilla SMP based 1303171713Sjeff * on mp_maxid. 1304171713Sjeff */ 1305171713Sjeff if (smp_topology == NULL) 1306171713Sjeff sched_setup_smp(); 1307171713Sjeff else 1308171713Sjeff sched_setup_topology(); 1309172409Sjeff balance_tdq = tdq; 1310172409Sjeff sched_balance(); 1311117237Sjeff#else 1312171713Sjeff tdq_setup(tdq); 1313171713Sjeff mtx_init(&tdq_lock, "sched lock", "sched lock", MTX_SPIN | MTX_RECURSE); 1314171713Sjeff tdq->tdq_lock = &tdq_lock; 1315116069Sjeff#endif 1316171482Sjeff /* 1317171482Sjeff * To avoid divide-by-zero, we set realstathz a dummy value 1318171482Sjeff * in case which sched_clock() called before sched_initticks(). 1319171482Sjeff */ 1320171482Sjeff realstathz = hz; 1321171482Sjeff sched_slice = (realstathz/10); /* ~100ms */ 1322171482Sjeff tickincr = 1 << SCHED_TICK_SHIFT; 1323171482Sjeff 1324171482Sjeff /* Add thread0's load since it's running. */ 1325171482Sjeff TDQ_LOCK(tdq); 1326171713Sjeff thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1327171482Sjeff tdq_load_add(tdq, &td_sched0); 1328171482Sjeff TDQ_UNLOCK(tdq); 1329109864Sjeff} 1330109864Sjeff 1331171482Sjeff/* 1332171482Sjeff * This routine determines the tickincr after stathz and hz are setup. 1333171482Sjeff */ 1334153533Sdavidxu/* ARGSUSED */ 1335153533Sdavidxustatic void 1336153533Sdavidxusched_initticks(void *dummy) 1337153533Sdavidxu{ 1338171482Sjeff int incr; 1339171482Sjeff 1340153533Sdavidxu realstathz = stathz ? stathz : hz; 1341166229Sjeff sched_slice = (realstathz/10); /* ~100ms */ 1342153533Sdavidxu 1343153533Sdavidxu /* 1344165762Sjeff * tickincr is shifted out by 10 to avoid rounding errors due to 1345165766Sjeff * hz not being evenly divisible by stathz on all platforms. 1346153533Sdavidxu */ 1347171482Sjeff incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1348165762Sjeff /* 1349165762Sjeff * This does not work for values of stathz that are more than 1350165762Sjeff * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1351165762Sjeff */ 1352171482Sjeff if (incr == 0) 1353171482Sjeff incr = 1; 1354171482Sjeff tickincr = incr; 1355166108Sjeff#ifdef SMP 1356171899Sjeff /* 1357172409Sjeff * Set the default balance interval now that we know 1358172409Sjeff * what realstathz is. 1359172409Sjeff */ 1360172409Sjeff balance_interval = realstathz; 1361172409Sjeff /* 1362171899Sjeff * Set steal thresh to log2(mp_ncpu) but no greater than 4. This 1363171899Sjeff * prevents excess thrashing on large machines and excess idle on 1364171899Sjeff * smaller machines. 1365171899Sjeff */ 1366171899Sjeff steal_thresh = min(ffs(mp_ncpus) - 1, 4); 1367166108Sjeff affinity = SCHED_AFFINITY_DEFAULT; 1368166108Sjeff#endif 1369153533Sdavidxu} 1370153533Sdavidxu 1371153533Sdavidxu 1372109864Sjeff/* 1373171482Sjeff * This is the core of the interactivity algorithm. Determines a score based 1374171482Sjeff * on past behavior. It is the ratio of sleep time to run time scaled to 1375171482Sjeff * a [0, 100] integer. This is the voluntary sleep time of a process, which 1376171482Sjeff * differs from the cpu usage because it does not account for time spent 1377171482Sjeff * waiting on a run-queue. Would be prettier if we had floating point. 1378171482Sjeff */ 1379171482Sjeffstatic int 1380171482Sjeffsched_interact_score(struct thread *td) 1381171482Sjeff{ 1382171482Sjeff struct td_sched *ts; 1383171482Sjeff int div; 1384171482Sjeff 1385171482Sjeff ts = td->td_sched; 1386171482Sjeff /* 1387171482Sjeff * The score is only needed if this is likely to be an interactive 1388171482Sjeff * task. Don't go through the expense of computing it if there's 1389171482Sjeff * no chance. 1390171482Sjeff */ 1391171482Sjeff if (sched_interact <= SCHED_INTERACT_HALF && 1392171482Sjeff ts->ts_runtime >= ts->ts_slptime) 1393171482Sjeff return (SCHED_INTERACT_HALF); 1394171482Sjeff 1395171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1396171482Sjeff div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1397171482Sjeff return (SCHED_INTERACT_HALF + 1398171482Sjeff (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1399171482Sjeff } 1400171482Sjeff if (ts->ts_slptime > ts->ts_runtime) { 1401171482Sjeff div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1402171482Sjeff return (ts->ts_runtime / div); 1403171482Sjeff } 1404171482Sjeff /* runtime == slptime */ 1405171482Sjeff if (ts->ts_runtime) 1406171482Sjeff return (SCHED_INTERACT_HALF); 1407171482Sjeff 1408171482Sjeff /* 1409171482Sjeff * This can happen if slptime and runtime are 0. 1410171482Sjeff */ 1411171482Sjeff return (0); 1412171482Sjeff 1413171482Sjeff} 1414171482Sjeff 1415171482Sjeff/* 1416109864Sjeff * Scale the scheduling priority according to the "interactivity" of this 1417109864Sjeff * process. 1418109864Sjeff */ 1419113357Sjeffstatic void 1420163709Sjbsched_priority(struct thread *td) 1421109864Sjeff{ 1422165762Sjeff int score; 1423109864Sjeff int pri; 1424109864Sjeff 1425163709Sjb if (td->td_pri_class != PRI_TIMESHARE) 1426113357Sjeff return; 1427112966Sjeff /* 1428165762Sjeff * If the score is interactive we place the thread in the realtime 1429165762Sjeff * queue with a priority that is less than kernel and interrupt 1430165762Sjeff * priorities. These threads are not subject to nice restrictions. 1431112966Sjeff * 1432171482Sjeff * Scores greater than this are placed on the normal timeshare queue 1433165762Sjeff * where the priority is partially decided by the most recent cpu 1434165762Sjeff * utilization and the rest is decided by nice value. 1435172293Sjeff * 1436172293Sjeff * The nice value of the process has a linear effect on the calculated 1437172293Sjeff * score. Negative nice values make it easier for a thread to be 1438172293Sjeff * considered interactive. 1439112966Sjeff */ 1440172308Sjeff score = imax(0, sched_interact_score(td) - td->td_proc->p_nice); 1441165762Sjeff if (score < sched_interact) { 1442165762Sjeff pri = PRI_MIN_REALTIME; 1443165762Sjeff pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact) 1444165762Sjeff * score; 1445165762Sjeff KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME, 1446166208Sjeff ("sched_priority: invalid interactive priority %d score %d", 1447166208Sjeff pri, score)); 1448165762Sjeff } else { 1449165762Sjeff pri = SCHED_PRI_MIN; 1450165762Sjeff if (td->td_sched->ts_ticks) 1451165762Sjeff pri += SCHED_PRI_TICKS(td->td_sched); 1452165762Sjeff pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1453171482Sjeff KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE, 1454171482Sjeff ("sched_priority: invalid priority %d: nice %d, " 1455171482Sjeff "ticks %d ftick %d ltick %d tick pri %d", 1456171482Sjeff pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1457171482Sjeff td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1458171482Sjeff SCHED_PRI_TICKS(td->td_sched))); 1459165762Sjeff } 1460165762Sjeff sched_user_prio(td, pri); 1461112966Sjeff 1462112966Sjeff return; 1463109864Sjeff} 1464109864Sjeff 1465121868Sjeff/* 1466121868Sjeff * This routine enforces a maximum limit on the amount of scheduling history 1467171482Sjeff * kept. It is called after either the slptime or runtime is adjusted. This 1468171482Sjeff * function is ugly due to integer math. 1469121868Sjeff */ 1470116463Sjeffstatic void 1471163709Sjbsched_interact_update(struct thread *td) 1472116463Sjeff{ 1473165819Sjeff struct td_sched *ts; 1474166208Sjeff u_int sum; 1475121605Sjeff 1476165819Sjeff ts = td->td_sched; 1477171482Sjeff sum = ts->ts_runtime + ts->ts_slptime; 1478121868Sjeff if (sum < SCHED_SLP_RUN_MAX) 1479121868Sjeff return; 1480121868Sjeff /* 1481165819Sjeff * This only happens from two places: 1482165819Sjeff * 1) We have added an unusual amount of run time from fork_exit. 1483165819Sjeff * 2) We have added an unusual amount of sleep time from sched_sleep(). 1484165819Sjeff */ 1485165819Sjeff if (sum > SCHED_SLP_RUN_MAX * 2) { 1486171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1487171482Sjeff ts->ts_runtime = SCHED_SLP_RUN_MAX; 1488171482Sjeff ts->ts_slptime = 1; 1489165819Sjeff } else { 1490171482Sjeff ts->ts_slptime = SCHED_SLP_RUN_MAX; 1491171482Sjeff ts->ts_runtime = 1; 1492165819Sjeff } 1493165819Sjeff return; 1494165819Sjeff } 1495165819Sjeff /* 1496121868Sjeff * If we have exceeded by more than 1/5th then the algorithm below 1497121868Sjeff * will not bring us back into range. Dividing by two here forces 1498133427Sjeff * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1499121868Sjeff */ 1500127850Sjeff if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1501171482Sjeff ts->ts_runtime /= 2; 1502171482Sjeff ts->ts_slptime /= 2; 1503121868Sjeff return; 1504116463Sjeff } 1505171482Sjeff ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1506171482Sjeff ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1507116463Sjeff} 1508116463Sjeff 1509171482Sjeff/* 1510171482Sjeff * Scale back the interactivity history when a child thread is created. The 1511171482Sjeff * history is inherited from the parent but the thread may behave totally 1512171482Sjeff * differently. For example, a shell spawning a compiler process. We want 1513171482Sjeff * to learn that the compiler is behaving badly very quickly. 1514171482Sjeff */ 1515121868Sjeffstatic void 1516163709Sjbsched_interact_fork(struct thread *td) 1517121868Sjeff{ 1518121868Sjeff int ratio; 1519121868Sjeff int sum; 1520121868Sjeff 1521171482Sjeff sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1522121868Sjeff if (sum > SCHED_SLP_RUN_FORK) { 1523121868Sjeff ratio = sum / SCHED_SLP_RUN_FORK; 1524171482Sjeff td->td_sched->ts_runtime /= ratio; 1525171482Sjeff td->td_sched->ts_slptime /= ratio; 1526121868Sjeff } 1527121868Sjeff} 1528121868Sjeff 1529113357Sjeff/* 1530171482Sjeff * Called from proc0_init() to setup the scheduler fields. 1531134791Sjulian */ 1532134791Sjulianvoid 1533134791Sjulianschedinit(void) 1534134791Sjulian{ 1535165762Sjeff 1536134791Sjulian /* 1537134791Sjulian * Set up the scheduler specific parts of proc0. 1538134791Sjulian */ 1539136167Sjulian proc0.p_sched = NULL; /* XXX */ 1540164936Sjulian thread0.td_sched = &td_sched0; 1541165762Sjeff td_sched0.ts_ltick = ticks; 1542165796Sjeff td_sched0.ts_ftick = ticks; 1543164936Sjulian td_sched0.ts_thread = &thread0; 1544134791Sjulian} 1545134791Sjulian 1546134791Sjulian/* 1547113357Sjeff * This is only somewhat accurate since given many processes of the same 1548113357Sjeff * priority they will switch when their slices run out, which will be 1549165762Sjeff * at most sched_slice stathz ticks. 1550113357Sjeff */ 1551109864Sjeffint 1552109864Sjeffsched_rr_interval(void) 1553109864Sjeff{ 1554165762Sjeff 1555165762Sjeff /* Convert sched_slice to hz */ 1556165762Sjeff return (hz/(realstathz/sched_slice)); 1557109864Sjeff} 1558109864Sjeff 1559171482Sjeff/* 1560171482Sjeff * Update the percent cpu tracking information when it is requested or 1561171482Sjeff * the total history exceeds the maximum. We keep a sliding history of 1562171482Sjeff * tick counts that slowly decays. This is less precise than the 4BSD 1563171482Sjeff * mechanism since it happens with less regular and frequent events. 1564171482Sjeff */ 1565121790Sjeffstatic void 1566164936Sjuliansched_pctcpu_update(struct td_sched *ts) 1567109864Sjeff{ 1568165762Sjeff 1569165762Sjeff if (ts->ts_ticks == 0) 1570165762Sjeff return; 1571165796Sjeff if (ticks - (hz / 10) < ts->ts_ltick && 1572165796Sjeff SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1573165796Sjeff return; 1574109864Sjeff /* 1575109864Sjeff * Adjust counters and watermark for pctcpu calc. 1576116365Sjeff */ 1577165762Sjeff if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1578164936Sjulian ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1579165762Sjeff SCHED_TICK_TARG; 1580165762Sjeff else 1581164936Sjulian ts->ts_ticks = 0; 1582164936Sjulian ts->ts_ltick = ticks; 1583165762Sjeff ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1584109864Sjeff} 1585109864Sjeff 1586171482Sjeff/* 1587171482Sjeff * Adjust the priority of a thread. Move it to the appropriate run-queue 1588171482Sjeff * if necessary. This is the back-end for several priority related 1589171482Sjeff * functions. 1590171482Sjeff */ 1591165762Sjeffstatic void 1592139453Sjhbsched_thread_priority(struct thread *td, u_char prio) 1593109864Sjeff{ 1594164936Sjulian struct td_sched *ts; 1595109864Sjeff 1596139316Sjeff CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 1597139316Sjeff td, td->td_proc->p_comm, td->td_priority, prio, curthread, 1598139316Sjeff curthread->td_proc->p_comm); 1599164936Sjulian ts = td->td_sched; 1600170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1601139453Sjhb if (td->td_priority == prio) 1602139453Sjhb return; 1603165762Sjeff 1604165766Sjeff if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1605121605Sjeff /* 1606121605Sjeff * If the priority has been elevated due to priority 1607121605Sjeff * propagation, we may have to move ourselves to a new 1608165762Sjeff * queue. This could be optimized to not re-add in some 1609165762Sjeff * cases. 1610133555Sjeff */ 1611165762Sjeff sched_rem(td); 1612165762Sjeff td->td_priority = prio; 1613171482Sjeff sched_add(td, SRQ_BORROWING); 1614171482Sjeff } else { 1615171482Sjeff#ifdef SMP 1616171482Sjeff struct tdq *tdq; 1617171482Sjeff 1618171482Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1619171482Sjeff if (prio < tdq->tdq_lowpri) 1620171482Sjeff tdq->tdq_lowpri = prio; 1621171482Sjeff#endif 1622119488Sdavidxu td->td_priority = prio; 1623171482Sjeff } 1624109864Sjeff} 1625109864Sjeff 1626139453Sjhb/* 1627139453Sjhb * Update a thread's priority when it is lent another thread's 1628139453Sjhb * priority. 1629139453Sjhb */ 1630109864Sjeffvoid 1631139453Sjhbsched_lend_prio(struct thread *td, u_char prio) 1632139453Sjhb{ 1633139453Sjhb 1634139453Sjhb td->td_flags |= TDF_BORROWING; 1635139453Sjhb sched_thread_priority(td, prio); 1636139453Sjhb} 1637139453Sjhb 1638139453Sjhb/* 1639139453Sjhb * Restore a thread's priority when priority propagation is 1640139453Sjhb * over. The prio argument is the minimum priority the thread 1641139453Sjhb * needs to have to satisfy other possible priority lending 1642139453Sjhb * requests. If the thread's regular priority is less 1643139453Sjhb * important than prio, the thread will keep a priority boost 1644139453Sjhb * of prio. 1645139453Sjhb */ 1646139453Sjhbvoid 1647139453Sjhbsched_unlend_prio(struct thread *td, u_char prio) 1648139453Sjhb{ 1649139453Sjhb u_char base_pri; 1650139453Sjhb 1651139453Sjhb if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1652139453Sjhb td->td_base_pri <= PRI_MAX_TIMESHARE) 1653163709Sjb base_pri = td->td_user_pri; 1654139453Sjhb else 1655139453Sjhb base_pri = td->td_base_pri; 1656139453Sjhb if (prio >= base_pri) { 1657139455Sjhb td->td_flags &= ~TDF_BORROWING; 1658139453Sjhb sched_thread_priority(td, base_pri); 1659139453Sjhb } else 1660139453Sjhb sched_lend_prio(td, prio); 1661139453Sjhb} 1662139453Sjhb 1663171482Sjeff/* 1664171482Sjeff * Standard entry for setting the priority to an absolute value. 1665171482Sjeff */ 1666139453Sjhbvoid 1667139453Sjhbsched_prio(struct thread *td, u_char prio) 1668139453Sjhb{ 1669139453Sjhb u_char oldprio; 1670139453Sjhb 1671139453Sjhb /* First, update the base priority. */ 1672139453Sjhb td->td_base_pri = prio; 1673139453Sjhb 1674139453Sjhb /* 1675139455Sjhb * If the thread is borrowing another thread's priority, don't 1676139453Sjhb * ever lower the priority. 1677139453Sjhb */ 1678139453Sjhb if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1679139453Sjhb return; 1680139453Sjhb 1681139453Sjhb /* Change the real priority. */ 1682139453Sjhb oldprio = td->td_priority; 1683139453Sjhb sched_thread_priority(td, prio); 1684139453Sjhb 1685139453Sjhb /* 1686139453Sjhb * If the thread is on a turnstile, then let the turnstile update 1687139453Sjhb * its state. 1688139453Sjhb */ 1689139453Sjhb if (TD_ON_LOCK(td) && oldprio != prio) 1690139453Sjhb turnstile_adjust(td, oldprio); 1691139453Sjhb} 1692139455Sjhb 1693171482Sjeff/* 1694171482Sjeff * Set the base user priority, does not effect current running priority. 1695171482Sjeff */ 1696139453Sjhbvoid 1697163709Sjbsched_user_prio(struct thread *td, u_char prio) 1698161599Sdavidxu{ 1699161599Sdavidxu u_char oldprio; 1700161599Sdavidxu 1701163709Sjb td->td_base_user_pri = prio; 1702164939Sjulian if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio) 1703164939Sjulian return; 1704163709Sjb oldprio = td->td_user_pri; 1705163709Sjb td->td_user_pri = prio; 1706163709Sjb 1707161599Sdavidxu if (TD_ON_UPILOCK(td) && oldprio != prio) 1708161599Sdavidxu umtx_pi_adjust(td, oldprio); 1709161599Sdavidxu} 1710161599Sdavidxu 1711161599Sdavidxuvoid 1712161599Sdavidxusched_lend_user_prio(struct thread *td, u_char prio) 1713161599Sdavidxu{ 1714161599Sdavidxu u_char oldprio; 1715161599Sdavidxu 1716161599Sdavidxu td->td_flags |= TDF_UBORROWING; 1717161599Sdavidxu 1718164091Smaxim oldprio = td->td_user_pri; 1719163709Sjb td->td_user_pri = prio; 1720161599Sdavidxu 1721161599Sdavidxu if (TD_ON_UPILOCK(td) && oldprio != prio) 1722161599Sdavidxu umtx_pi_adjust(td, oldprio); 1723161599Sdavidxu} 1724161599Sdavidxu 1725161599Sdavidxuvoid 1726161599Sdavidxusched_unlend_user_prio(struct thread *td, u_char prio) 1727161599Sdavidxu{ 1728161599Sdavidxu u_char base_pri; 1729161599Sdavidxu 1730163709Sjb base_pri = td->td_base_user_pri; 1731161599Sdavidxu if (prio >= base_pri) { 1732161599Sdavidxu td->td_flags &= ~TDF_UBORROWING; 1733163709Sjb sched_user_prio(td, base_pri); 1734161599Sdavidxu } else 1735161599Sdavidxu sched_lend_user_prio(td, prio); 1736161599Sdavidxu} 1737161599Sdavidxu 1738171482Sjeff/* 1739171505Sjeff * Add the thread passed as 'newtd' to the run queue before selecting 1740171505Sjeff * the next thread to run. This is only used for KSE. 1741171505Sjeff */ 1742171505Sjeffstatic void 1743171505Sjeffsched_switchin(struct tdq *tdq, struct thread *td) 1744171505Sjeff{ 1745171505Sjeff#ifdef SMP 1746171505Sjeff spinlock_enter(); 1747171505Sjeff TDQ_UNLOCK(tdq); 1748171505Sjeff thread_lock(td); 1749171505Sjeff spinlock_exit(); 1750171505Sjeff sched_setcpu(td->td_sched, TDQ_ID(tdq), SRQ_YIELDING); 1751171505Sjeff#else 1752171505Sjeff td->td_lock = TDQ_LOCKPTR(tdq); 1753171505Sjeff#endif 1754171505Sjeff tdq_add(tdq, td, SRQ_YIELDING); 1755171505Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1756171505Sjeff} 1757171505Sjeff 1758171505Sjeff/* 1759171713Sjeff * Handle migration from sched_switch(). This happens only for 1760171713Sjeff * cpu binding. 1761171713Sjeff */ 1762171713Sjeffstatic struct mtx * 1763171713Sjeffsched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1764171713Sjeff{ 1765171713Sjeff struct tdq *tdn; 1766171713Sjeff 1767171713Sjeff tdn = TDQ_CPU(td->td_sched->ts_cpu); 1768171713Sjeff#ifdef SMP 1769171713Sjeff /* 1770171713Sjeff * Do the lock dance required to avoid LOR. We grab an extra 1771171713Sjeff * spinlock nesting to prevent preemption while we're 1772171713Sjeff * not holding either run-queue lock. 1773171713Sjeff */ 1774171713Sjeff spinlock_enter(); 1775171713Sjeff thread_block_switch(td); /* This releases the lock on tdq. */ 1776171713Sjeff TDQ_LOCK(tdn); 1777171713Sjeff tdq_add(tdn, td, flags); 1778171713Sjeff tdq_notify(td->td_sched); 1779171713Sjeff /* 1780171713Sjeff * After we unlock tdn the new cpu still can't switch into this 1781171713Sjeff * thread until we've unblocked it in cpu_switch(). The lock 1782171713Sjeff * pointers may match in the case of HTT cores. Don't unlock here 1783171713Sjeff * or we can deadlock when the other CPU runs the IPI handler. 1784171713Sjeff */ 1785171713Sjeff if (TDQ_LOCKPTR(tdn) != TDQ_LOCKPTR(tdq)) { 1786171713Sjeff TDQ_UNLOCK(tdn); 1787171713Sjeff TDQ_LOCK(tdq); 1788171713Sjeff } 1789171713Sjeff spinlock_exit(); 1790171713Sjeff#endif 1791171713Sjeff return (TDQ_LOCKPTR(tdn)); 1792171713Sjeff} 1793171713Sjeff 1794171713Sjeff/* 1795171482Sjeff * Block a thread for switching. Similar to thread_block() but does not 1796171482Sjeff * bump the spin count. 1797171482Sjeff */ 1798171482Sjeffstatic inline struct mtx * 1799171482Sjeffthread_block_switch(struct thread *td) 1800171482Sjeff{ 1801171482Sjeff struct mtx *lock; 1802171482Sjeff 1803171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1804171482Sjeff lock = td->td_lock; 1805171482Sjeff td->td_lock = &blocked_lock; 1806171482Sjeff mtx_unlock_spin(lock); 1807171482Sjeff 1808171482Sjeff return (lock); 1809171482Sjeff} 1810171482Sjeff 1811171482Sjeff/* 1812171482Sjeff * Release a thread that was blocked with thread_block_switch(). 1813171482Sjeff */ 1814171482Sjeffstatic inline void 1815171482Sjeffthread_unblock_switch(struct thread *td, struct mtx *mtx) 1816171482Sjeff{ 1817171482Sjeff atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1818171482Sjeff (uintptr_t)mtx); 1819171482Sjeff} 1820171482Sjeff 1821171482Sjeff/* 1822171482Sjeff * Switch threads. This function has to handle threads coming in while 1823171482Sjeff * blocked for some reason, running, or idle. It also must deal with 1824171482Sjeff * migrating a thread from one queue to another as running threads may 1825171482Sjeff * be assigned elsewhere via binding. 1826171482Sjeff */ 1827161599Sdavidxuvoid 1828135051Sjuliansched_switch(struct thread *td, struct thread *newtd, int flags) 1829109864Sjeff{ 1830165627Sjeff struct tdq *tdq; 1831164936Sjulian struct td_sched *ts; 1832171482Sjeff struct mtx *mtx; 1833171713Sjeff int srqflag; 1834171482Sjeff int cpuid; 1835109864Sjeff 1836170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1837109864Sjeff 1838171482Sjeff cpuid = PCPU_GET(cpuid); 1839171482Sjeff tdq = TDQ_CPU(cpuid); 1840164936Sjulian ts = td->td_sched; 1841171713Sjeff mtx = td->td_lock; 1842171482Sjeff#ifdef SMP 1843171482Sjeff ts->ts_rltick = ticks; 1844171482Sjeff if (newtd && newtd->td_priority < tdq->tdq_lowpri) 1845171482Sjeff tdq->tdq_lowpri = newtd->td_priority; 1846171482Sjeff#endif 1847133555Sjeff td->td_lastcpu = td->td_oncpu; 1848113339Sjulian td->td_oncpu = NOCPU; 1849132266Sjhb td->td_flags &= ~TDF_NEEDRESCHED; 1850144777Sups td->td_owepreempt = 0; 1851123434Sjeff /* 1852171482Sjeff * The lock pointer in an idle thread should never change. Reset it 1853171482Sjeff * to CAN_RUN as well. 1854123434Sjeff */ 1855167327Sjulian if (TD_IS_IDLETHREAD(td)) { 1856171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1857139334Sjeff TD_SET_CAN_RUN(td); 1858170293Sjeff } else if (TD_IS_RUNNING(td)) { 1859171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1860165627Sjeff tdq_load_rem(tdq, ts); 1861171713Sjeff srqflag = (flags & SW_PREEMPT) ? 1862170293Sjeff SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1863171713Sjeff SRQ_OURSELF|SRQ_YIELDING; 1864171713Sjeff if (ts->ts_cpu == cpuid) 1865171713Sjeff tdq_add(tdq, td, srqflag); 1866171713Sjeff else 1867171713Sjeff mtx = sched_switch_migrate(tdq, td, srqflag); 1868171482Sjeff } else { 1869171482Sjeff /* This thread must be going to sleep. */ 1870171482Sjeff TDQ_LOCK(tdq); 1871171482Sjeff mtx = thread_block_switch(td); 1872170293Sjeff tdq_load_rem(tdq, ts); 1873171482Sjeff } 1874171482Sjeff /* 1875171482Sjeff * We enter here with the thread blocked and assigned to the 1876171482Sjeff * appropriate cpu run-queue or sleep-queue and with the current 1877171482Sjeff * thread-queue locked. 1878171482Sjeff */ 1879171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1880171482Sjeff /* 1881171505Sjeff * If KSE assigned a new thread just add it here and let choosethread 1882171505Sjeff * select the best one. 1883171482Sjeff */ 1884171505Sjeff if (newtd != NULL) 1885171505Sjeff sched_switchin(tdq, newtd); 1886171482Sjeff newtd = choosethread(); 1887171482Sjeff /* 1888171482Sjeff * Call the MD code to switch contexts if necessary. 1889171482Sjeff */ 1890145256Sjkoshy if (td != newtd) { 1891145256Sjkoshy#ifdef HWPMC_HOOKS 1892145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1893145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1894145256Sjkoshy#endif 1895171482Sjeff cpu_switch(td, newtd, mtx); 1896171482Sjeff /* 1897171482Sjeff * We may return from cpu_switch on a different cpu. However, 1898171482Sjeff * we always return with td_lock pointing to the current cpu's 1899171482Sjeff * run queue lock. 1900171482Sjeff */ 1901171482Sjeff cpuid = PCPU_GET(cpuid); 1902171482Sjeff tdq = TDQ_CPU(cpuid); 1903171482Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)td; 1904145256Sjkoshy#ifdef HWPMC_HOOKS 1905145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1906145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1907145256Sjkoshy#endif 1908171482Sjeff } else 1909171482Sjeff thread_unblock_switch(td, mtx); 1910171482Sjeff /* 1911171482Sjeff * Assert that all went well and return. 1912171482Sjeff */ 1913171482Sjeff#ifdef SMP 1914171482Sjeff /* We should always get here with the lowest priority td possible */ 1915171482Sjeff tdq->tdq_lowpri = td->td_priority; 1916171482Sjeff#endif 1917171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1918171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1919171482Sjeff td->td_oncpu = cpuid; 1920109864Sjeff} 1921109864Sjeff 1922171482Sjeff/* 1923171482Sjeff * Adjust thread priorities as a result of a nice request. 1924171482Sjeff */ 1925109864Sjeffvoid 1926130551Sjuliansched_nice(struct proc *p, int nice) 1927109864Sjeff{ 1928109864Sjeff struct thread *td; 1929109864Sjeff 1930130551Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 1931170293Sjeff PROC_SLOCK_ASSERT(p, MA_OWNED); 1932165762Sjeff 1933130551Sjulian p->p_nice = nice; 1934163709Sjb FOREACH_THREAD_IN_PROC(p, td) { 1935170293Sjeff thread_lock(td); 1936163709Sjb sched_priority(td); 1937165762Sjeff sched_prio(td, td->td_base_user_pri); 1938170293Sjeff thread_unlock(td); 1939130551Sjulian } 1940109864Sjeff} 1941109864Sjeff 1942171482Sjeff/* 1943171482Sjeff * Record the sleep time for the interactivity scorer. 1944171482Sjeff */ 1945109864Sjeffvoid 1946126326Sjhbsched_sleep(struct thread *td) 1947109864Sjeff{ 1948165762Sjeff 1949170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1950109864Sjeff 1951172264Sjeff td->td_slptick = ticks; 1952109864Sjeff} 1953109864Sjeff 1954171482Sjeff/* 1955171482Sjeff * Schedule a thread to resume execution and record how long it voluntarily 1956171482Sjeff * slept. We also update the pctcpu, interactivity, and priority. 1957171482Sjeff */ 1958109864Sjeffvoid 1959109864Sjeffsched_wakeup(struct thread *td) 1960109864Sjeff{ 1961166229Sjeff struct td_sched *ts; 1962171482Sjeff int slptick; 1963165762Sjeff 1964170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1965166229Sjeff ts = td->td_sched; 1966109864Sjeff /* 1967165762Sjeff * If we slept for more than a tick update our interactivity and 1968165762Sjeff * priority. 1969109864Sjeff */ 1970172264Sjeff slptick = td->td_slptick; 1971172264Sjeff td->td_slptick = 0; 1972171482Sjeff if (slptick && slptick != ticks) { 1973166208Sjeff u_int hzticks; 1974109864Sjeff 1975171482Sjeff hzticks = (ticks - slptick) << SCHED_TICK_SHIFT; 1976171482Sjeff ts->ts_slptime += hzticks; 1977165819Sjeff sched_interact_update(td); 1978166229Sjeff sched_pctcpu_update(ts); 1979163709Sjb sched_priority(td); 1980109864Sjeff } 1981166229Sjeff /* Reset the slice value after we sleep. */ 1982166229Sjeff ts->ts_slice = sched_slice; 1983166190Sjeff sched_add(td, SRQ_BORING); 1984109864Sjeff} 1985109864Sjeff 1986109864Sjeff/* 1987109864Sjeff * Penalize the parent for creating a new child and initialize the child's 1988109864Sjeff * priority. 1989109864Sjeff */ 1990109864Sjeffvoid 1991163709Sjbsched_fork(struct thread *td, struct thread *child) 1992109864Sjeff{ 1993170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1994164936Sjulian sched_fork_thread(td, child); 1995165762Sjeff /* 1996165762Sjeff * Penalize the parent and child for forking. 1997165762Sjeff */ 1998165762Sjeff sched_interact_fork(child); 1999165762Sjeff sched_priority(child); 2000171482Sjeff td->td_sched->ts_runtime += tickincr; 2001165762Sjeff sched_interact_update(td); 2002165762Sjeff sched_priority(td); 2003164936Sjulian} 2004109864Sjeff 2005171482Sjeff/* 2006171482Sjeff * Fork a new thread, may be within the same process. 2007171482Sjeff */ 2008164936Sjulianvoid 2009164936Sjuliansched_fork_thread(struct thread *td, struct thread *child) 2010164936Sjulian{ 2011164936Sjulian struct td_sched *ts; 2012164936Sjulian struct td_sched *ts2; 2013164936Sjulian 2014165762Sjeff /* 2015165762Sjeff * Initialize child. 2016165762Sjeff */ 2017170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2018163709Sjb sched_newthread(child); 2019171482Sjeff child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 2020164936Sjulian ts = td->td_sched; 2021164936Sjulian ts2 = child->td_sched; 2022164936Sjulian ts2->ts_cpu = ts->ts_cpu; 2023164936Sjulian ts2->ts_runq = NULL; 2024165762Sjeff /* 2025165762Sjeff * Grab our parents cpu estimation information and priority. 2026165762Sjeff */ 2027164936Sjulian ts2->ts_ticks = ts->ts_ticks; 2028164936Sjulian ts2->ts_ltick = ts->ts_ltick; 2029164936Sjulian ts2->ts_ftick = ts->ts_ftick; 2030165762Sjeff child->td_user_pri = td->td_user_pri; 2031165762Sjeff child->td_base_user_pri = td->td_base_user_pri; 2032165762Sjeff /* 2033165762Sjeff * And update interactivity score. 2034165762Sjeff */ 2035171482Sjeff ts2->ts_slptime = ts->ts_slptime; 2036171482Sjeff ts2->ts_runtime = ts->ts_runtime; 2037165762Sjeff ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 2038113357Sjeff} 2039113357Sjeff 2040171482Sjeff/* 2041171482Sjeff * Adjust the priority class of a thread. 2042171482Sjeff */ 2043113357Sjeffvoid 2044163709Sjbsched_class(struct thread *td, int class) 2045113357Sjeff{ 2046113357Sjeff 2047170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2048163709Sjb if (td->td_pri_class == class) 2049113357Sjeff return; 2050113357Sjeff 2051121896Sjeff#ifdef SMP 2052165827Sjeff /* 2053165827Sjeff * On SMP if we're on the RUNQ we must adjust the transferable 2054165827Sjeff * count because could be changing to or from an interrupt 2055165827Sjeff * class. 2056165827Sjeff */ 2057166190Sjeff if (TD_ON_RUNQ(td)) { 2058165827Sjeff struct tdq *tdq; 2059165827Sjeff 2060165827Sjeff tdq = TDQ_CPU(td->td_sched->ts_cpu); 2061165827Sjeff if (THREAD_CAN_MIGRATE(td)) { 2062165827Sjeff tdq->tdq_transferable--; 2063165827Sjeff tdq->tdq_group->tdg_transferable--; 2064122744Sjeff } 2065165827Sjeff td->td_pri_class = class; 2066165827Sjeff if (THREAD_CAN_MIGRATE(td)) { 2067165827Sjeff tdq->tdq_transferable++; 2068165827Sjeff tdq->tdq_group->tdg_transferable++; 2069165827Sjeff } 2070165827Sjeff } 2071164936Sjulian#endif 2072163709Sjb td->td_pri_class = class; 2073109864Sjeff} 2074109864Sjeff 2075109864Sjeff/* 2076109864Sjeff * Return some of the child's priority and interactivity to the parent. 2077109864Sjeff */ 2078109864Sjeffvoid 2079164939Sjuliansched_exit(struct proc *p, struct thread *child) 2080109864Sjeff{ 2081165762Sjeff struct thread *td; 2082164939Sjulian 2083163709Sjb CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d", 2084164939Sjulian child, child->td_proc->p_comm, child->td_priority); 2085113372Sjeff 2086170293Sjeff PROC_SLOCK_ASSERT(p, MA_OWNED); 2087165762Sjeff td = FIRST_THREAD_IN_PROC(p); 2088165762Sjeff sched_exit_thread(td, child); 2089113372Sjeff} 2090113372Sjeff 2091171482Sjeff/* 2092171482Sjeff * Penalize another thread for the time spent on this one. This helps to 2093171482Sjeff * worsen the priority and interactivity of processes which schedule batch 2094171482Sjeff * jobs such as make. This has little effect on the make process itself but 2095171482Sjeff * causes new processes spawned by it to receive worse scores immediately. 2096171482Sjeff */ 2097113372Sjeffvoid 2098164939Sjuliansched_exit_thread(struct thread *td, struct thread *child) 2099164936Sjulian{ 2100165762Sjeff 2101164939Sjulian CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d", 2102165762Sjeff child, child->td_proc->p_comm, child->td_priority); 2103164939Sjulian 2104165762Sjeff#ifdef KSE 2105165762Sjeff /* 2106165762Sjeff * KSE forks and exits so often that this penalty causes short-lived 2107165762Sjeff * threads to always be non-interactive. This causes mozilla to 2108165762Sjeff * crawl under load. 2109165762Sjeff */ 2110165762Sjeff if ((td->td_pflags & TDP_SA) && td->td_proc == child->td_proc) 2111165762Sjeff return; 2112165762Sjeff#endif 2113165762Sjeff /* 2114165762Sjeff * Give the child's runtime to the parent without returning the 2115165762Sjeff * sleep time as a penalty to the parent. This causes shells that 2116165762Sjeff * launch expensive things to mark their children as expensive. 2117165762Sjeff */ 2118170293Sjeff thread_lock(td); 2119171482Sjeff td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2120164939Sjulian sched_interact_update(td); 2121165762Sjeff sched_priority(td); 2122170293Sjeff thread_unlock(td); 2123164936Sjulian} 2124164936Sjulian 2125171482Sjeff/* 2126171482Sjeff * Fix priorities on return to user-space. Priorities may be elevated due 2127171482Sjeff * to static priorities in msleep() or similar. 2128171482Sjeff */ 2129164936Sjulianvoid 2130164936Sjuliansched_userret(struct thread *td) 2131164936Sjulian{ 2132164936Sjulian /* 2133164936Sjulian * XXX we cheat slightly on the locking here to avoid locking in 2134164936Sjulian * the usual case. Setting td_priority here is essentially an 2135164936Sjulian * incomplete workaround for not setting it properly elsewhere. 2136164936Sjulian * Now that some interrupt handlers are threads, not setting it 2137164936Sjulian * properly elsewhere can clobber it in the window between setting 2138164936Sjulian * it here and returning to user mode, so don't waste time setting 2139164936Sjulian * it perfectly here. 2140164936Sjulian */ 2141164936Sjulian KASSERT((td->td_flags & TDF_BORROWING) == 0, 2142164936Sjulian ("thread with borrowed priority returning to userland")); 2143164936Sjulian if (td->td_priority != td->td_user_pri) { 2144170293Sjeff thread_lock(td); 2145164936Sjulian td->td_priority = td->td_user_pri; 2146164936Sjulian td->td_base_pri = td->td_user_pri; 2147170293Sjeff thread_unlock(td); 2148164936Sjulian } 2149164936Sjulian} 2150164936Sjulian 2151171482Sjeff/* 2152171482Sjeff * Handle a stathz tick. This is really only relevant for timeshare 2153171482Sjeff * threads. 2154171482Sjeff */ 2155164936Sjulianvoid 2156121127Sjeffsched_clock(struct thread *td) 2157109864Sjeff{ 2158164936Sjulian struct tdq *tdq; 2159164936Sjulian struct td_sched *ts; 2160109864Sjeff 2161171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2162164936Sjulian tdq = TDQ_SELF(); 2163172409Sjeff#ifdef SMP 2164133427Sjeff /* 2165172409Sjeff * We run the long term load balancer infrequently on the first cpu. 2166172409Sjeff */ 2167172409Sjeff if (balance_tdq == tdq) { 2168172409Sjeff if (balance_ticks && --balance_ticks == 0) 2169172409Sjeff sched_balance(); 2170172409Sjeff if (balance_group_ticks && --balance_group_ticks == 0) 2171172409Sjeff sched_balance_groups(); 2172172409Sjeff } 2173172409Sjeff#endif 2174172409Sjeff /* 2175165766Sjeff * Advance the insert index once for each tick to ensure that all 2176165766Sjeff * threads get a chance to run. 2177133427Sjeff */ 2178165766Sjeff if (tdq->tdq_idx == tdq->tdq_ridx) { 2179165766Sjeff tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2180165766Sjeff if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2181165766Sjeff tdq->tdq_ridx = tdq->tdq_idx; 2182165766Sjeff } 2183165766Sjeff ts = td->td_sched; 2184165762Sjeff /* 2185163709Sjb * We only do slicing code for TIMESHARE threads. 2186113357Sjeff */ 2187163709Sjb if (td->td_pri_class != PRI_TIMESHARE) 2188113357Sjeff return; 2189113357Sjeff /* 2190165766Sjeff * We used a tick; charge it to the thread so that we can compute our 2191113357Sjeff * interactivity. 2192109864Sjeff */ 2193171482Sjeff td->td_sched->ts_runtime += tickincr; 2194163709Sjb sched_interact_update(td); 2195109864Sjeff /* 2196109864Sjeff * We used up one time slice. 2197109864Sjeff */ 2198164936Sjulian if (--ts->ts_slice > 0) 2199113357Sjeff return; 2200109864Sjeff /* 2201113357Sjeff * We're out of time, recompute priorities and requeue. 2202109864Sjeff */ 2203165796Sjeff sched_priority(td); 2204113357Sjeff td->td_flags |= TDF_NEEDRESCHED; 2205109864Sjeff} 2206109864Sjeff 2207171482Sjeff/* 2208171482Sjeff * Called once per hz tick. Used for cpu utilization information. This 2209171482Sjeff * is easier than trying to scale based on stathz. 2210171482Sjeff */ 2211171482Sjeffvoid 2212171482Sjeffsched_tick(void) 2213171482Sjeff{ 2214171482Sjeff struct td_sched *ts; 2215171482Sjeff 2216171482Sjeff ts = curthread->td_sched; 2217171482Sjeff /* Adjust ticks for pctcpu */ 2218171482Sjeff ts->ts_ticks += 1 << SCHED_TICK_SHIFT; 2219171482Sjeff ts->ts_ltick = ticks; 2220171482Sjeff /* 2221171482Sjeff * Update if we've exceeded our desired tick threshhold by over one 2222171482Sjeff * second. 2223171482Sjeff */ 2224171482Sjeff if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2225171482Sjeff sched_pctcpu_update(ts); 2226171482Sjeff} 2227171482Sjeff 2228171482Sjeff/* 2229171482Sjeff * Return whether the current CPU has runnable tasks. Used for in-kernel 2230171482Sjeff * cooperative idle threads. 2231171482Sjeff */ 2232109864Sjeffint 2233109864Sjeffsched_runnable(void) 2234109864Sjeff{ 2235164936Sjulian struct tdq *tdq; 2236115998Sjeff int load; 2237109864Sjeff 2238115998Sjeff load = 1; 2239115998Sjeff 2240164936Sjulian tdq = TDQ_SELF(); 2241121605Sjeff if ((curthread->td_flags & TDF_IDLETD) != 0) { 2242165620Sjeff if (tdq->tdq_load > 0) 2243121605Sjeff goto out; 2244121605Sjeff } else 2245165620Sjeff if (tdq->tdq_load - 1 > 0) 2246121605Sjeff goto out; 2247115998Sjeff load = 0; 2248115998Sjeffout: 2249115998Sjeff return (load); 2250109864Sjeff} 2251109864Sjeff 2252171482Sjeff/* 2253171482Sjeff * Choose the highest priority thread to run. The thread is removed from 2254171482Sjeff * the run-queue while running however the load remains. For SMP we set 2255171482Sjeff * the tdq in the global idle bitmask if it idles here. 2256171482Sjeff */ 2257166190Sjeffstruct thread * 2258109970Sjeffsched_choose(void) 2259109970Sjeff{ 2260171482Sjeff#ifdef SMP 2261171482Sjeff struct tdq_group *tdg; 2262171482Sjeff#endif 2263171482Sjeff struct td_sched *ts; 2264164936Sjulian struct tdq *tdq; 2265109970Sjeff 2266164936Sjulian tdq = TDQ_SELF(); 2267171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2268164936Sjulian ts = tdq_choose(tdq); 2269164936Sjulian if (ts) { 2270164936Sjulian tdq_runq_rem(tdq, ts); 2271166190Sjeff return (ts->ts_thread); 2272109864Sjeff } 2273109970Sjeff#ifdef SMP 2274171482Sjeff /* 2275171482Sjeff * We only set the idled bit when all of the cpus in the group are 2276171482Sjeff * idle. Otherwise we could get into a situation where a thread bounces 2277171482Sjeff * back and forth between two idle cores on seperate physical CPUs. 2278171482Sjeff */ 2279171482Sjeff tdg = tdq->tdq_group; 2280171482Sjeff tdg->tdg_idlemask |= PCPU_GET(cpumask); 2281171482Sjeff if (tdg->tdg_idlemask == tdg->tdg_cpumask) 2282171482Sjeff atomic_set_int(&tdq_idle, tdg->tdg_mask); 2283171482Sjeff tdq->tdq_lowpri = PRI_MAX_IDLE; 2284109970Sjeff#endif 2285166190Sjeff return (PCPU_GET(idlethread)); 2286109864Sjeff} 2287109864Sjeff 2288171482Sjeff/* 2289171482Sjeff * Set owepreempt if necessary. Preemption never happens directly in ULE, 2290171482Sjeff * we always request it once we exit a critical section. 2291171482Sjeff */ 2292171482Sjeffstatic inline void 2293171482Sjeffsched_setpreempt(struct thread *td) 2294166190Sjeff{ 2295166190Sjeff struct thread *ctd; 2296166190Sjeff int cpri; 2297166190Sjeff int pri; 2298166190Sjeff 2299166190Sjeff ctd = curthread; 2300166190Sjeff pri = td->td_priority; 2301166190Sjeff cpri = ctd->td_priority; 2302171482Sjeff if (td->td_priority < ctd->td_priority) 2303171482Sjeff curthread->td_flags |= TDF_NEEDRESCHED; 2304166190Sjeff if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2305171482Sjeff return; 2306166190Sjeff /* 2307166190Sjeff * Always preempt IDLE threads. Otherwise only if the preempting 2308166190Sjeff * thread is an ithread. 2309166190Sjeff */ 2310171482Sjeff if (pri > preempt_thresh && cpri < PRI_MIN_IDLE) 2311171482Sjeff return; 2312171482Sjeff ctd->td_owepreempt = 1; 2313171482Sjeff return; 2314166190Sjeff} 2315166190Sjeff 2316171482Sjeff/* 2317171482Sjeff * Add a thread to a thread queue. Initializes priority, slice, runq, and 2318171482Sjeff * add it to the appropriate queue. This is the internal function called 2319171482Sjeff * when the tdq is predetermined. 2320171482Sjeff */ 2321109864Sjeffvoid 2322171482Sjefftdq_add(struct tdq *tdq, struct thread *td, int flags) 2323109864Sjeff{ 2324164936Sjulian struct td_sched *ts; 2325121790Sjeff int class; 2326166108Sjeff#ifdef SMP 2327166108Sjeff int cpumask; 2328166108Sjeff#endif 2329109864Sjeff 2330171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2331166190Sjeff KASSERT((td->td_inhibitors == 0), 2332166190Sjeff ("sched_add: trying to run inhibited thread")); 2333166190Sjeff KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2334166190Sjeff ("sched_add: bad thread state")); 2335172207Sjeff KASSERT(td->td_flags & TDF_INMEM, 2336172207Sjeff ("sched_add: thread swapped out")); 2337171482Sjeff 2338171482Sjeff ts = td->td_sched; 2339171482Sjeff class = PRI_BASE(td->td_pri_class); 2340166190Sjeff TD_SET_RUNQ(td); 2341166190Sjeff if (ts->ts_slice == 0) 2342166190Sjeff ts->ts_slice = sched_slice; 2343133427Sjeff /* 2344171482Sjeff * Pick the run queue based on priority. 2345133427Sjeff */ 2346171482Sjeff if (td->td_priority <= PRI_MAX_REALTIME) 2347171482Sjeff ts->ts_runq = &tdq->tdq_realtime; 2348171482Sjeff else if (td->td_priority <= PRI_MAX_TIMESHARE) 2349171482Sjeff ts->ts_runq = &tdq->tdq_timeshare; 2350171482Sjeff else 2351171482Sjeff ts->ts_runq = &tdq->tdq_idle; 2352171482Sjeff#ifdef SMP 2353166108Sjeff cpumask = 1 << ts->ts_cpu; 2354121790Sjeff /* 2355123685Sjeff * If we had been idle, clear our bit in the group and potentially 2356166108Sjeff * the global bitmap. 2357121790Sjeff */ 2358165762Sjeff if ((class != PRI_IDLE && class != PRI_ITHD) && 2359166108Sjeff (tdq->tdq_group->tdg_idlemask & cpumask) != 0) { 2360121790Sjeff /* 2361123433Sjeff * Check to see if our group is unidling, and if so, remove it 2362123433Sjeff * from the global idle mask. 2363121790Sjeff */ 2364165620Sjeff if (tdq->tdq_group->tdg_idlemask == 2365165620Sjeff tdq->tdq_group->tdg_cpumask) 2366165620Sjeff atomic_clear_int(&tdq_idle, tdq->tdq_group->tdg_mask); 2367123433Sjeff /* 2368123433Sjeff * Now remove ourselves from the group specific idle mask. 2369123433Sjeff */ 2370166108Sjeff tdq->tdq_group->tdg_idlemask &= ~cpumask; 2371166108Sjeff } 2372171482Sjeff if (td->td_priority < tdq->tdq_lowpri) 2373171482Sjeff tdq->tdq_lowpri = td->td_priority; 2374121790Sjeff#endif 2375171482Sjeff tdq_runq_add(tdq, ts, flags); 2376171482Sjeff tdq_load_add(tdq, ts); 2377171482Sjeff} 2378171482Sjeff 2379171482Sjeff/* 2380171482Sjeff * Select the target thread queue and add a thread to it. Request 2381171482Sjeff * preemption or IPI a remote processor if required. 2382171482Sjeff */ 2383171482Sjeffvoid 2384171482Sjeffsched_add(struct thread *td, int flags) 2385171482Sjeff{ 2386171482Sjeff struct td_sched *ts; 2387171482Sjeff struct tdq *tdq; 2388171482Sjeff#ifdef SMP 2389171482Sjeff int cpuid; 2390171482Sjeff int cpu; 2391171482Sjeff#endif 2392171482Sjeff CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 2393171482Sjeff td, td->td_proc->p_comm, td->td_priority, curthread, 2394171482Sjeff curthread->td_proc->p_comm); 2395171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2396171482Sjeff ts = td->td_sched; 2397166108Sjeff /* 2398171482Sjeff * Recalculate the priority before we select the target cpu or 2399171482Sjeff * run-queue. 2400166108Sjeff */ 2401171482Sjeff if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2402171482Sjeff sched_priority(td); 2403171482Sjeff#ifdef SMP 2404171482Sjeff cpuid = PCPU_GET(cpuid); 2405171482Sjeff /* 2406171482Sjeff * Pick the destination cpu and if it isn't ours transfer to the 2407171482Sjeff * target cpu. 2408171482Sjeff */ 2409171482Sjeff if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_MIGRATE(td)) 2410171482Sjeff cpu = cpuid; 2411171482Sjeff else if (!THREAD_CAN_MIGRATE(td)) 2412171482Sjeff cpu = ts->ts_cpu; 2413166108Sjeff else 2414171482Sjeff cpu = sched_pickcpu(ts, flags); 2415171482Sjeff tdq = sched_setcpu(ts, cpu, flags); 2416171482Sjeff tdq_add(tdq, td, flags); 2417171482Sjeff if (cpu != cpuid) { 2418166108Sjeff tdq_notify(ts); 2419166108Sjeff return; 2420166108Sjeff } 2421171482Sjeff#else 2422171482Sjeff tdq = TDQ_SELF(); 2423171482Sjeff TDQ_LOCK(tdq); 2424171482Sjeff /* 2425171482Sjeff * Now that the thread is moving to the run-queue, set the lock 2426171482Sjeff * to the scheduler's lock. 2427171482Sjeff */ 2428171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2429171482Sjeff tdq_add(tdq, td, flags); 2430166108Sjeff#endif 2431171482Sjeff if (!(flags & SRQ_YIELDING)) 2432171482Sjeff sched_setpreempt(td); 2433109864Sjeff} 2434109864Sjeff 2435171482Sjeff/* 2436171482Sjeff * Remove a thread from a run-queue without running it. This is used 2437171482Sjeff * when we're stealing a thread from a remote queue. Otherwise all threads 2438171482Sjeff * exit by calling sched_exit_thread() and sched_throw() themselves. 2439171482Sjeff */ 2440109864Sjeffvoid 2441121127Sjeffsched_rem(struct thread *td) 2442109864Sjeff{ 2443164936Sjulian struct tdq *tdq; 2444164936Sjulian struct td_sched *ts; 2445113357Sjeff 2446139316Sjeff CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 2447139316Sjeff td, td->td_proc->p_comm, td->td_priority, curthread, 2448139316Sjeff curthread->td_proc->p_comm); 2449164936Sjulian ts = td->td_sched; 2450171482Sjeff tdq = TDQ_CPU(ts->ts_cpu); 2451171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2452171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2453166190Sjeff KASSERT(TD_ON_RUNQ(td), 2454164936Sjulian ("sched_rem: thread not on run queue")); 2455164936Sjulian tdq_runq_rem(tdq, ts); 2456164936Sjulian tdq_load_rem(tdq, ts); 2457166190Sjeff TD_SET_CAN_RUN(td); 2458109864Sjeff} 2459109864Sjeff 2460171482Sjeff/* 2461171482Sjeff * Fetch cpu utilization information. Updates on demand. 2462171482Sjeff */ 2463109864Sjefffixpt_t 2464121127Sjeffsched_pctcpu(struct thread *td) 2465109864Sjeff{ 2466109864Sjeff fixpt_t pctcpu; 2467164936Sjulian struct td_sched *ts; 2468109864Sjeff 2469109864Sjeff pctcpu = 0; 2470164936Sjulian ts = td->td_sched; 2471164936Sjulian if (ts == NULL) 2472121290Sjeff return (0); 2473109864Sjeff 2474170293Sjeff thread_lock(td); 2475164936Sjulian if (ts->ts_ticks) { 2476109864Sjeff int rtick; 2477109864Sjeff 2478165796Sjeff sched_pctcpu_update(ts); 2479109864Sjeff /* How many rtick per second ? */ 2480165762Sjeff rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2481165762Sjeff pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2482109864Sjeff } 2483170293Sjeff thread_unlock(td); 2484109864Sjeff 2485109864Sjeff return (pctcpu); 2486109864Sjeff} 2487109864Sjeff 2488171482Sjeff/* 2489171482Sjeff * Bind a thread to a target cpu. 2490171482Sjeff */ 2491122038Sjeffvoid 2492122038Sjeffsched_bind(struct thread *td, int cpu) 2493122038Sjeff{ 2494164936Sjulian struct td_sched *ts; 2495122038Sjeff 2496171713Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2497164936Sjulian ts = td->td_sched; 2498166137Sjeff if (ts->ts_flags & TSF_BOUND) 2499166152Sjeff sched_unbind(td); 2500164936Sjulian ts->ts_flags |= TSF_BOUND; 2501123433Sjeff#ifdef SMP 2502166137Sjeff sched_pin(); 2503123433Sjeff if (PCPU_GET(cpuid) == cpu) 2504122038Sjeff return; 2505166137Sjeff ts->ts_cpu = cpu; 2506122038Sjeff /* When we return from mi_switch we'll be on the correct cpu. */ 2507131527Sphk mi_switch(SW_VOL, NULL); 2508122038Sjeff#endif 2509122038Sjeff} 2510122038Sjeff 2511171482Sjeff/* 2512171482Sjeff * Release a bound thread. 2513171482Sjeff */ 2514122038Sjeffvoid 2515122038Sjeffsched_unbind(struct thread *td) 2516122038Sjeff{ 2517165762Sjeff struct td_sched *ts; 2518165762Sjeff 2519170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2520165762Sjeff ts = td->td_sched; 2521166137Sjeff if ((ts->ts_flags & TSF_BOUND) == 0) 2522166137Sjeff return; 2523165762Sjeff ts->ts_flags &= ~TSF_BOUND; 2524165762Sjeff#ifdef SMP 2525165762Sjeff sched_unpin(); 2526165762Sjeff#endif 2527122038Sjeff} 2528122038Sjeff 2529109864Sjeffint 2530145256Sjkoshysched_is_bound(struct thread *td) 2531145256Sjkoshy{ 2532170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2533164936Sjulian return (td->td_sched->ts_flags & TSF_BOUND); 2534145256Sjkoshy} 2535145256Sjkoshy 2536171482Sjeff/* 2537171482Sjeff * Basic yield call. 2538171482Sjeff */ 2539159630Sdavidxuvoid 2540159630Sdavidxusched_relinquish(struct thread *td) 2541159630Sdavidxu{ 2542170293Sjeff thread_lock(td); 2543163709Sjb if (td->td_pri_class == PRI_TIMESHARE) 2544159630Sdavidxu sched_prio(td, PRI_MAX_TIMESHARE); 2545170293Sjeff SCHED_STAT_INC(switch_relinquish); 2546159630Sdavidxu mi_switch(SW_VOL, NULL); 2547170293Sjeff thread_unlock(td); 2548159630Sdavidxu} 2549159630Sdavidxu 2550171482Sjeff/* 2551171482Sjeff * Return the total system load. 2552171482Sjeff */ 2553145256Sjkoshyint 2554125289Sjeffsched_load(void) 2555125289Sjeff{ 2556125289Sjeff#ifdef SMP 2557125289Sjeff int total; 2558125289Sjeff int i; 2559125289Sjeff 2560125289Sjeff total = 0; 2561165620Sjeff for (i = 0; i <= tdg_maxid; i++) 2562165620Sjeff total += TDQ_GROUP(i)->tdg_load; 2563125289Sjeff return (total); 2564125289Sjeff#else 2565165620Sjeff return (TDQ_SELF()->tdq_sysload); 2566125289Sjeff#endif 2567125289Sjeff} 2568125289Sjeff 2569125289Sjeffint 2570109864Sjeffsched_sizeof_proc(void) 2571109864Sjeff{ 2572109864Sjeff return (sizeof(struct proc)); 2573109864Sjeff} 2574109864Sjeff 2575109864Sjeffint 2576109864Sjeffsched_sizeof_thread(void) 2577109864Sjeff{ 2578109864Sjeff return (sizeof(struct thread) + sizeof(struct td_sched)); 2579109864Sjeff} 2580159570Sdavidxu 2581166190Sjeff/* 2582166190Sjeff * The actual idle process. 2583166190Sjeff */ 2584166190Sjeffvoid 2585166190Sjeffsched_idletd(void *dummy) 2586166190Sjeff{ 2587166190Sjeff struct thread *td; 2588171482Sjeff struct tdq *tdq; 2589166190Sjeff 2590166190Sjeff td = curthread; 2591171482Sjeff tdq = TDQ_SELF(); 2592166190Sjeff mtx_assert(&Giant, MA_NOTOWNED); 2593171482Sjeff /* ULE relies on preemption for idle interruption. */ 2594171482Sjeff for (;;) { 2595171482Sjeff#ifdef SMP 2596171482Sjeff if (tdq_idled(tdq)) 2597171482Sjeff cpu_idle(); 2598171482Sjeff#else 2599166190Sjeff cpu_idle(); 2600171482Sjeff#endif 2601171482Sjeff } 2602166190Sjeff} 2603166190Sjeff 2604170293Sjeff/* 2605170293Sjeff * A CPU is entering for the first time or a thread is exiting. 2606170293Sjeff */ 2607170293Sjeffvoid 2608170293Sjeffsched_throw(struct thread *td) 2609170293Sjeff{ 2610171482Sjeff struct tdq *tdq; 2611171482Sjeff 2612171482Sjeff tdq = TDQ_SELF(); 2613170293Sjeff if (td == NULL) { 2614171482Sjeff /* Correct spinlock nesting and acquire the correct lock. */ 2615171482Sjeff TDQ_LOCK(tdq); 2616170293Sjeff spinlock_exit(); 2617170293Sjeff } else { 2618171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2619171482Sjeff tdq_load_rem(tdq, td->td_sched); 2620170293Sjeff } 2621170293Sjeff KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2622170293Sjeff PCPU_SET(switchtime, cpu_ticks()); 2623170293Sjeff PCPU_SET(switchticks, ticks); 2624170293Sjeff cpu_throw(td, choosethread()); /* doesn't return */ 2625170293Sjeff} 2626170293Sjeff 2627171482Sjeff/* 2628171482Sjeff * This is called from fork_exit(). Just acquire the correct locks and 2629171482Sjeff * let fork do the rest of the work. 2630171482Sjeff */ 2631170293Sjeffvoid 2632170600Sjeffsched_fork_exit(struct thread *td) 2633170293Sjeff{ 2634171482Sjeff struct td_sched *ts; 2635171482Sjeff struct tdq *tdq; 2636171482Sjeff int cpuid; 2637170293Sjeff 2638170293Sjeff /* 2639170293Sjeff * Finish setting up thread glue so that it begins execution in a 2640171482Sjeff * non-nested critical section with the scheduler lock held. 2641170293Sjeff */ 2642171482Sjeff cpuid = PCPU_GET(cpuid); 2643171482Sjeff tdq = TDQ_CPU(cpuid); 2644171482Sjeff ts = td->td_sched; 2645171482Sjeff if (TD_IS_IDLETHREAD(td)) 2646171482Sjeff td->td_lock = TDQ_LOCKPTR(tdq); 2647171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2648171482Sjeff td->td_oncpu = cpuid; 2649171482Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)td; 2650170600Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED); 2651170293Sjeff} 2652170293Sjeff 2653171482Sjeffstatic SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, 2654171482Sjeff "Scheduler"); 2655171482SjeffSYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2656165762Sjeff "Scheduler name"); 2657171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2658171482Sjeff "Slice size for timeshare threads"); 2659171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2660171482Sjeff "Interactivity score threshold"); 2661171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2662171482Sjeff 0,"Min priority for preemption, lower priorities have greater precedence"); 2663166108Sjeff#ifdef SMP 2664171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, pick_pri, CTLFLAG_RW, &pick_pri, 0, 2665171482Sjeff "Pick the target cpu based on priority rather than load."); 2666171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2667171482Sjeff "Number of hz ticks to keep thread affinity for"); 2668171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, tryself, CTLFLAG_RW, &tryself, 0, ""); 2669171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2670171482Sjeff "Enables the long-term load balancer"); 2671172409SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2672172409Sjeff &balance_interval, 0, 2673172409Sjeff "Average frequency in stathz ticks to run the long-term balancer"); 2674171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, 2675171482Sjeff "Steals work from another hyper-threaded core on idle"); 2676171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2677171482Sjeff "Attempts to steal work from other cores before idling"); 2678171506SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2679171506Sjeff "Minimum load on remote cpu before we'll steal"); 2680171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, topology, CTLFLAG_RD, &topology, 0, 2681171482Sjeff "True when a topology has been specified by the MD code."); 2682166108Sjeff#endif 2683165762Sjeff 2684172264Sjeff/* ps compat. All cpu percentages from ULE are weighted. */ 2685172293Sjeffstatic int ccpu = 0; 2686165762SjeffSYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2687165762Sjeff 2688165762Sjeff 2689134791Sjulian#define KERN_SWITCH_INCLUDE 1 2690134791Sjulian#include "kern/kern_switch.c" 2691