sched_ule.c revision 176734
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 176734 2008-03-02 07:58:42Z jeff $"); 40116182Sobrien 41147565Speter#include "opt_hwpmc_hooks.h" 42147565Speter#include "opt_sched.h" 43134649Sscottl 44109864Sjeff#include <sys/param.h> 45109864Sjeff#include <sys/systm.h> 46131929Smarcel#include <sys/kdb.h> 47109864Sjeff#include <sys/kernel.h> 48109864Sjeff#include <sys/ktr.h> 49109864Sjeff#include <sys/lock.h> 50109864Sjeff#include <sys/mutex.h> 51109864Sjeff#include <sys/proc.h> 52112966Sjeff#include <sys/resource.h> 53122038Sjeff#include <sys/resourcevar.h> 54109864Sjeff#include <sys/sched.h> 55109864Sjeff#include <sys/smp.h> 56109864Sjeff#include <sys/sx.h> 57109864Sjeff#include <sys/sysctl.h> 58109864Sjeff#include <sys/sysproto.h> 59139453Sjhb#include <sys/turnstile.h> 60161599Sdavidxu#include <sys/umtx.h> 61109864Sjeff#include <sys/vmmeter.h> 62109864Sjeff#ifdef KTRACE 63109864Sjeff#include <sys/uio.h> 64109864Sjeff#include <sys/ktrace.h> 65109864Sjeff#endif 66109864Sjeff 67145256Sjkoshy#ifdef HWPMC_HOOKS 68145256Sjkoshy#include <sys/pmckern.h> 69145256Sjkoshy#endif 70145256Sjkoshy 71109864Sjeff#include <machine/cpu.h> 72121790Sjeff#include <machine/smp.h> 73109864Sjeff 74172887Sgrehan#if !defined(__i386__) && !defined(__amd64__) && !defined(__powerpc__) && !defined(__arm__) 75172345Sjeff#error "This architecture is not currently compatible with ULE" 76166190Sjeff#endif 77166190Sjeff 78171482Sjeff#define KTR_ULE 0 79166137Sjeff 80166137Sjeff/* 81171482Sjeff * Thread scheduler specific section. All fields are protected 82171482Sjeff * by the thread lock. 83146954Sjeff */ 84164936Sjulianstruct td_sched { 85171482Sjeff TAILQ_ENTRY(td_sched) ts_procq; /* Run queue. */ 86171482Sjeff struct thread *ts_thread; /* Active associated thread. */ 87171482Sjeff struct runq *ts_runq; /* Run-queue we're queued on. */ 88171482Sjeff short ts_flags; /* TSF_* flags. */ 89171482Sjeff u_char ts_rqindex; /* Run queue index. */ 90164936Sjulian u_char ts_cpu; /* CPU that we have affinity for. */ 91171482Sjeff int ts_slice; /* Ticks of slice remaining. */ 92171482Sjeff u_int ts_slptime; /* Number of ticks we vol. slept */ 93171482Sjeff u_int ts_runtime; /* Number of ticks we were running */ 94134791Sjulian /* The following variables are only used for pctcpu calculation */ 95164936Sjulian int ts_ltick; /* Last tick that we were running on */ 96164936Sjulian int ts_ftick; /* First tick that we were running on */ 97164936Sjulian int ts_ticks; /* Tick count */ 98166108Sjeff#ifdef SMP 99166108Sjeff int ts_rltick; /* Real last tick, for affinity. */ 100166108Sjeff#endif 101134791Sjulian}; 102164936Sjulian/* flags kept in ts_flags */ 103166108Sjeff#define TSF_BOUND 0x0001 /* Thread can not migrate. */ 104166108Sjeff#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 105121790Sjeff 106164936Sjulianstatic struct td_sched td_sched0; 107109864Sjeff 108109864Sjeff/* 109165762Sjeff * Cpu percentage computation macros and defines. 110111857Sjeff * 111165762Sjeff * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 112165762Sjeff * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 113165796Sjeff * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 114165762Sjeff * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 115165762Sjeff * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 116165762Sjeff * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 117165762Sjeff */ 118165762Sjeff#define SCHED_TICK_SECS 10 119165762Sjeff#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 120165796Sjeff#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 121165762Sjeff#define SCHED_TICK_SHIFT 10 122165762Sjeff#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 123165830Sjeff#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 124165762Sjeff 125165762Sjeff/* 126165762Sjeff * These macros determine priorities for non-interactive threads. They are 127165762Sjeff * assigned a priority based on their recent cpu utilization as expressed 128165762Sjeff * by the ratio of ticks to the tick total. NHALF priorities at the start 129165762Sjeff * and end of the MIN to MAX timeshare range are only reachable with negative 130165762Sjeff * or positive nice respectively. 131165762Sjeff * 132165762Sjeff * PRI_RANGE: Priority range for utilization dependent priorities. 133116642Sjeff * PRI_NRESV: Number of nice values. 134165762Sjeff * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 135165762Sjeff * PRI_NICE: Determines the part of the priority inherited from nice. 136109864Sjeff */ 137165762Sjeff#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 138121869Sjeff#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 139165762Sjeff#define SCHED_PRI_MIN (PRI_MIN_TIMESHARE + SCHED_PRI_NHALF) 140165762Sjeff#define SCHED_PRI_MAX (PRI_MAX_TIMESHARE - SCHED_PRI_NHALF) 141170787Sjeff#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN) 142165762Sjeff#define SCHED_PRI_TICKS(ts) \ 143165762Sjeff (SCHED_TICK_HZ((ts)) / \ 144165827Sjeff (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 145165762Sjeff#define SCHED_PRI_NICE(nice) (nice) 146109864Sjeff 147109864Sjeff/* 148165762Sjeff * These determine the interactivity of a process. Interactivity differs from 149165762Sjeff * cpu utilization in that it expresses the voluntary time slept vs time ran 150165762Sjeff * while cpu utilization includes all time not running. This more accurately 151165762Sjeff * models the intent of the thread. 152109864Sjeff * 153110645Sjeff * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 154110645Sjeff * before throttling back. 155121868Sjeff * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 156116365Sjeff * INTERACT_MAX: Maximum interactivity value. Smaller is better. 157111857Sjeff * INTERACT_THRESH: Threshhold for placement on the current runq. 158109864Sjeff */ 159165762Sjeff#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 160165762Sjeff#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 161116365Sjeff#define SCHED_INTERACT_MAX (100) 162116365Sjeff#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 163121126Sjeff#define SCHED_INTERACT_THRESH (30) 164111857Sjeff 165109864Sjeff/* 166165762Sjeff * tickincr: Converts a stathz tick into a hz domain scaled by 167165762Sjeff * the shift factor. Without the shift the error rate 168165762Sjeff * due to rounding would be unacceptably high. 169165762Sjeff * realstathz: stathz is sometimes 0 and run off of hz. 170165762Sjeff * sched_slice: Runtime of each thread before rescheduling. 171171482Sjeff * preempt_thresh: Priority threshold for preemption and remote IPIs. 172109864Sjeff */ 173165762Sjeffstatic int sched_interact = SCHED_INTERACT_THRESH; 174165762Sjeffstatic int realstathz; 175165762Sjeffstatic int tickincr; 176165762Sjeffstatic int sched_slice; 177172345Sjeff#ifdef PREEMPTION 178172345Sjeff#ifdef FULL_PREEMPTION 179172345Sjeffstatic int preempt_thresh = PRI_MAX_IDLE; 180172345Sjeff#else 181171482Sjeffstatic int preempt_thresh = PRI_MIN_KERN; 182172345Sjeff#endif 183172345Sjeff#else 184172345Sjeffstatic int preempt_thresh = 0; 185172345Sjeff#endif 186109864Sjeff 187109864Sjeff/* 188171482Sjeff * tdq - per processor runqs and statistics. All fields are protected by the 189171482Sjeff * tdq_lock. The load and lowpri may be accessed without to avoid excess 190171482Sjeff * locking in sched_pickcpu(); 191109864Sjeff */ 192164936Sjulianstruct tdq { 193171713Sjeff struct mtx *tdq_lock; /* Pointer to group lock. */ 194171482Sjeff struct runq tdq_realtime; /* real-time run queue. */ 195171482Sjeff struct runq tdq_timeshare; /* timeshare run queue. */ 196165620Sjeff struct runq tdq_idle; /* Queue of IDLE threads. */ 197171482Sjeff int tdq_load; /* Aggregate load. */ 198166557Sjeff u_char tdq_idx; /* Current insert index. */ 199166557Sjeff u_char tdq_ridx; /* Current removal index. */ 200110267Sjeff#ifdef SMP 201171482Sjeff u_char tdq_lowpri; /* Lowest priority thread. */ 202171482Sjeff int tdq_transferable; /* Transferable thread count. */ 203165620Sjeff LIST_ENTRY(tdq) tdq_siblings; /* Next in tdq group. */ 204165620Sjeff struct tdq_group *tdq_group; /* Our processor group. */ 205125289Sjeff#else 206165620Sjeff int tdq_sysload; /* For loadavg, !ITHD load. */ 207110267Sjeff#endif 208171482Sjeff} __aligned(64); 209109864Sjeff 210166108Sjeff 211123433Sjeff#ifdef SMP 212109864Sjeff/* 213164936Sjulian * tdq groups are groups of processors which can cheaply share threads. When 214123433Sjeff * one processor in the group goes idle it will check the runqs of the other 215123433Sjeff * processors in its group prior to halting and waiting for an interrupt. 216123433Sjeff * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA. 217123433Sjeff * In a numa environment we'd want an idle bitmap per group and a two tiered 218123433Sjeff * load balancer. 219123433Sjeff */ 220164936Sjulianstruct tdq_group { 221171713Sjeff struct mtx tdg_lock; /* Protects all fields below. */ 222171713Sjeff int tdg_cpus; /* Count of CPUs in this tdq group. */ 223171713Sjeff cpumask_t tdg_cpumask; /* Mask of cpus in this group. */ 224171713Sjeff cpumask_t tdg_idlemask; /* Idle cpus in this group. */ 225171713Sjeff cpumask_t tdg_mask; /* Bit mask for first cpu. */ 226171713Sjeff int tdg_load; /* Total load of this group. */ 227165620Sjeff int tdg_transferable; /* Transferable load of this group. */ 228165620Sjeff LIST_HEAD(, tdq) tdg_members; /* Linked list of all members. */ 229171713Sjeff char tdg_name[16]; /* lock name. */ 230171482Sjeff} __aligned(64); 231123433Sjeff 232171482Sjeff#define SCHED_AFFINITY_DEFAULT (max(1, hz / 300)) 233166108Sjeff#define SCHED_AFFINITY(ts) ((ts)->ts_rltick > ticks - affinity) 234166108Sjeff 235123433Sjeff/* 236166108Sjeff * Run-time tunables. 237166108Sjeff */ 238171506Sjeffstatic int rebalance = 1; 239172409Sjeffstatic int balance_interval = 128; /* Default set in sched_initticks(). */ 240171506Sjeffstatic int pick_pri = 1; 241166108Sjeffstatic int affinity; 242166108Sjeffstatic int tryself = 1; 243172409Sjeffstatic int steal_htt = 1; 244171506Sjeffstatic int steal_idle = 1; 245171506Sjeffstatic int steal_thresh = 2; 246166108Sjeff 247166108Sjeff/* 248165620Sjeff * One thread queue per processor. 249109864Sjeff */ 250166108Sjeffstatic volatile cpumask_t tdq_idle; 251165620Sjeffstatic int tdg_maxid; 252164936Sjulianstatic struct tdq tdq_cpu[MAXCPU]; 253164936Sjulianstatic struct tdq_group tdq_groups[MAXCPU]; 254172409Sjeffstatic struct tdq *balance_tdq; 255172409Sjeffstatic int balance_group_ticks; 256172409Sjeffstatic int balance_ticks; 257129982Sjeff 258164936Sjulian#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 259164936Sjulian#define TDQ_CPU(x) (&tdq_cpu[(x)]) 260171713Sjeff#define TDQ_ID(x) ((int)((x) - tdq_cpu)) 261164936Sjulian#define TDQ_GROUP(x) (&tdq_groups[(x)]) 262171713Sjeff#define TDG_ID(x) ((int)((x) - tdq_groups)) 263123433Sjeff#else /* !SMP */ 264164936Sjulianstatic struct tdq tdq_cpu; 265171713Sjeffstatic struct mtx tdq_lock; 266129982Sjeff 267170315Sjeff#define TDQ_ID(x) (0) 268164936Sjulian#define TDQ_SELF() (&tdq_cpu) 269164936Sjulian#define TDQ_CPU(x) (&tdq_cpu) 270110028Sjeff#endif 271109864Sjeff 272171482Sjeff#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 273171482Sjeff#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 274171482Sjeff#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 275171482Sjeff#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 276171713Sjeff#define TDQ_LOCKPTR(t) ((t)->tdq_lock) 277171482Sjeff 278163709Sjbstatic void sched_priority(struct thread *); 279146954Sjeffstatic void sched_thread_priority(struct thread *, u_char); 280163709Sjbstatic int sched_interact_score(struct thread *); 281163709Sjbstatic void sched_interact_update(struct thread *); 282163709Sjbstatic void sched_interact_fork(struct thread *); 283164936Sjulianstatic void sched_pctcpu_update(struct td_sched *); 284109864Sjeff 285110267Sjeff/* Operations on per processor queues */ 286164936Sjulianstatic struct td_sched * tdq_choose(struct tdq *); 287164936Sjulianstatic void tdq_setup(struct tdq *); 288164936Sjulianstatic void tdq_load_add(struct tdq *, struct td_sched *); 289164936Sjulianstatic void tdq_load_rem(struct tdq *, struct td_sched *); 290164936Sjulianstatic __inline void tdq_runq_add(struct tdq *, struct td_sched *, int); 291164936Sjulianstatic __inline void tdq_runq_rem(struct tdq *, struct td_sched *); 292164936Sjulianvoid tdq_print(int cpu); 293165762Sjeffstatic void runq_print(struct runq *rq); 294171482Sjeffstatic void tdq_add(struct tdq *, struct thread *, int); 295110267Sjeff#ifdef SMP 296171482Sjeffstatic void tdq_move(struct tdq *, struct tdq *); 297171482Sjeffstatic int tdq_idled(struct tdq *); 298171482Sjeffstatic void tdq_notify(struct td_sched *); 299172409Sjeffstatic struct td_sched *tdq_steal(struct tdq *); 300164936Sjulianstatic struct td_sched *runq_steal(struct runq *); 301171482Sjeffstatic int sched_pickcpu(struct td_sched *, int); 302172409Sjeffstatic void sched_balance(void); 303172409Sjeffstatic void sched_balance_groups(void); 304164936Sjulianstatic void sched_balance_group(struct tdq_group *); 305164936Sjulianstatic void sched_balance_pair(struct tdq *, struct tdq *); 306171482Sjeffstatic inline struct tdq *sched_setcpu(struct td_sched *, int, int); 307171482Sjeffstatic inline struct mtx *thread_block_switch(struct thread *); 308171482Sjeffstatic inline void thread_unblock_switch(struct thread *, struct mtx *); 309171713Sjeffstatic struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 310165827Sjeff 311166108Sjeff#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 312121790Sjeff#endif 313110028Sjeff 314165762Sjeffstatic void sched_setup(void *dummy); 315165762SjeffSYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL) 316165762Sjeff 317165762Sjeffstatic void sched_initticks(void *dummy); 318165762SjeffSYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, NULL) 319165762Sjeff 320171482Sjeff/* 321171482Sjeff * Print the threads waiting on a run-queue. 322171482Sjeff */ 323165762Sjeffstatic void 324165762Sjeffrunq_print(struct runq *rq) 325165762Sjeff{ 326165762Sjeff struct rqhead *rqh; 327165762Sjeff struct td_sched *ts; 328165762Sjeff int pri; 329165762Sjeff int j; 330165762Sjeff int i; 331165762Sjeff 332165762Sjeff for (i = 0; i < RQB_LEN; i++) { 333165762Sjeff printf("\t\trunq bits %d 0x%zx\n", 334165762Sjeff i, rq->rq_status.rqb_bits[i]); 335165762Sjeff for (j = 0; j < RQB_BPW; j++) 336165762Sjeff if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 337165762Sjeff pri = j + (i << RQB_L2BPW); 338165762Sjeff rqh = &rq->rq_queues[pri]; 339165762Sjeff TAILQ_FOREACH(ts, rqh, ts_procq) { 340165762Sjeff printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 341173600Sjulian ts->ts_thread, ts->ts_thread->td_name, ts->ts_thread->td_priority, ts->ts_rqindex, pri); 342165762Sjeff } 343165762Sjeff } 344165762Sjeff } 345165762Sjeff} 346165762Sjeff 347171482Sjeff/* 348171482Sjeff * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 349171482Sjeff */ 350113357Sjeffvoid 351164936Sjuliantdq_print(int cpu) 352110267Sjeff{ 353164936Sjulian struct tdq *tdq; 354112994Sjeff 355164936Sjulian tdq = TDQ_CPU(cpu); 356112994Sjeff 357171713Sjeff printf("tdq %d:\n", TDQ_ID(tdq)); 358171482Sjeff printf("\tlockptr %p\n", TDQ_LOCKPTR(tdq)); 359165620Sjeff printf("\tload: %d\n", tdq->tdq_load); 360171482Sjeff printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 361165766Sjeff printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 362165762Sjeff printf("\trealtime runq:\n"); 363165762Sjeff runq_print(&tdq->tdq_realtime); 364165762Sjeff printf("\ttimeshare runq:\n"); 365165762Sjeff runq_print(&tdq->tdq_timeshare); 366165762Sjeff printf("\tidle runq:\n"); 367165762Sjeff runq_print(&tdq->tdq_idle); 368121896Sjeff#ifdef SMP 369165620Sjeff printf("\tload transferable: %d\n", tdq->tdq_transferable); 370171713Sjeff printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 371171713Sjeff printf("\tgroup: %d\n", TDG_ID(tdq->tdq_group)); 372171713Sjeff printf("\tLock name: %s\n", tdq->tdq_group->tdg_name); 373121896Sjeff#endif 374113357Sjeff} 375112994Sjeff 376171482Sjeff#define TS_RQ_PPQ (((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS) 377171482Sjeff/* 378171482Sjeff * Add a thread to the actual run-queue. Keeps transferable counts up to 379171482Sjeff * date with what is actually on the run-queue. Selects the correct 380171482Sjeff * queue position for timeshare threads. 381171482Sjeff */ 382122744Sjeffstatic __inline void 383164936Sjuliantdq_runq_add(struct tdq *tdq, struct td_sched *ts, int flags) 384122744Sjeff{ 385171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 386171482Sjeff THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 387122744Sjeff#ifdef SMP 388165762Sjeff if (THREAD_CAN_MIGRATE(ts->ts_thread)) { 389165620Sjeff tdq->tdq_transferable++; 390165620Sjeff tdq->tdq_group->tdg_transferable++; 391164936Sjulian ts->ts_flags |= TSF_XFERABLE; 392123433Sjeff } 393122744Sjeff#endif 394165762Sjeff if (ts->ts_runq == &tdq->tdq_timeshare) { 395166557Sjeff u_char pri; 396165762Sjeff 397165762Sjeff pri = ts->ts_thread->td_priority; 398165762Sjeff KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE, 399165762Sjeff ("Invalid priority %d on timeshare runq", pri)); 400165762Sjeff /* 401165762Sjeff * This queue contains only priorities between MIN and MAX 402165762Sjeff * realtime. Use the whole queue to represent these values. 403165762Sjeff */ 404171713Sjeff if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 405165762Sjeff pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ; 406165762Sjeff pri = (pri + tdq->tdq_idx) % RQ_NQS; 407165766Sjeff /* 408165766Sjeff * This effectively shortens the queue by one so we 409165766Sjeff * can have a one slot difference between idx and 410165766Sjeff * ridx while we wait for threads to drain. 411165766Sjeff */ 412165766Sjeff if (tdq->tdq_ridx != tdq->tdq_idx && 413165766Sjeff pri == tdq->tdq_ridx) 414167664Sjeff pri = (unsigned char)(pri - 1) % RQ_NQS; 415165762Sjeff } else 416165766Sjeff pri = tdq->tdq_ridx; 417165762Sjeff runq_add_pri(ts->ts_runq, ts, pri, flags); 418165762Sjeff } else 419165762Sjeff runq_add(ts->ts_runq, ts, flags); 420122744Sjeff} 421122744Sjeff 422171482Sjeff/* 423171482Sjeff * Remove a thread from a run-queue. This typically happens when a thread 424171482Sjeff * is selected to run. Running threads are not on the queue and the 425171482Sjeff * transferable count does not reflect them. 426171482Sjeff */ 427122744Sjeffstatic __inline void 428164936Sjuliantdq_runq_rem(struct tdq *tdq, struct td_sched *ts) 429122744Sjeff{ 430171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 431171482Sjeff KASSERT(ts->ts_runq != NULL, 432171482Sjeff ("tdq_runq_remove: thread %p null ts_runq", ts->ts_thread)); 433122744Sjeff#ifdef SMP 434164936Sjulian if (ts->ts_flags & TSF_XFERABLE) { 435165620Sjeff tdq->tdq_transferable--; 436165620Sjeff tdq->tdq_group->tdg_transferable--; 437164936Sjulian ts->ts_flags &= ~TSF_XFERABLE; 438123433Sjeff } 439122744Sjeff#endif 440165766Sjeff if (ts->ts_runq == &tdq->tdq_timeshare) { 441165766Sjeff if (tdq->tdq_idx != tdq->tdq_ridx) 442165766Sjeff runq_remove_idx(ts->ts_runq, ts, &tdq->tdq_ridx); 443165766Sjeff else 444165766Sjeff runq_remove_idx(ts->ts_runq, ts, NULL); 445165796Sjeff /* 446165796Sjeff * For timeshare threads we update the priority here so 447165796Sjeff * the priority reflects the time we've been sleeping. 448165796Sjeff */ 449165796Sjeff ts->ts_ltick = ticks; 450165796Sjeff sched_pctcpu_update(ts); 451165796Sjeff sched_priority(ts->ts_thread); 452165766Sjeff } else 453165762Sjeff runq_remove(ts->ts_runq, ts); 454122744Sjeff} 455122744Sjeff 456171482Sjeff/* 457171482Sjeff * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 458171482Sjeff * for this thread to the referenced thread queue. 459171482Sjeff */ 460113357Sjeffstatic void 461164936Sjuliantdq_load_add(struct tdq *tdq, struct td_sched *ts) 462113357Sjeff{ 463121896Sjeff int class; 464171482Sjeff 465171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 466171482Sjeff THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 467164936Sjulian class = PRI_BASE(ts->ts_thread->td_pri_class); 468165620Sjeff tdq->tdq_load++; 469171713Sjeff CTR2(KTR_SCHED, "cpu %d load: %d", TDQ_ID(tdq), tdq->tdq_load); 470166108Sjeff if (class != PRI_ITHD && 471166108Sjeff (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 472123487Sjeff#ifdef SMP 473165620Sjeff tdq->tdq_group->tdg_load++; 474125289Sjeff#else 475165620Sjeff tdq->tdq_sysload++; 476123487Sjeff#endif 477110267Sjeff} 478113357Sjeff 479171482Sjeff/* 480171482Sjeff * Remove the load from a thread that is transitioning to a sleep state or 481171482Sjeff * exiting. 482171482Sjeff */ 483112994Sjeffstatic void 484164936Sjuliantdq_load_rem(struct tdq *tdq, struct td_sched *ts) 485110267Sjeff{ 486121896Sjeff int class; 487171482Sjeff 488171482Sjeff THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 489171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 490164936Sjulian class = PRI_BASE(ts->ts_thread->td_pri_class); 491166108Sjeff if (class != PRI_ITHD && 492166108Sjeff (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 493123487Sjeff#ifdef SMP 494165620Sjeff tdq->tdq_group->tdg_load--; 495125289Sjeff#else 496165620Sjeff tdq->tdq_sysload--; 497123487Sjeff#endif 498171482Sjeff KASSERT(tdq->tdq_load != 0, 499171713Sjeff ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 500165620Sjeff tdq->tdq_load--; 501165620Sjeff CTR1(KTR_SCHED, "load: %d", tdq->tdq_load); 502164936Sjulian ts->ts_runq = NULL; 503110267Sjeff} 504110267Sjeff 505113357Sjeff#ifdef SMP 506116069Sjeff/* 507122744Sjeff * sched_balance is a simple CPU load balancing algorithm. It operates by 508116069Sjeff * finding the least loaded and most loaded cpu and equalizing their load 509116069Sjeff * by migrating some processes. 510116069Sjeff * 511116069Sjeff * Dealing only with two CPUs at a time has two advantages. Firstly, most 512116069Sjeff * installations will only have 2 cpus. Secondly, load balancing too much at 513116069Sjeff * once can have an unpleasant effect on the system. The scheduler rarely has 514116069Sjeff * enough information to make perfect decisions. So this algorithm chooses 515171482Sjeff * simplicity and more gradual effects on load in larger systems. 516116069Sjeff * 517116069Sjeff */ 518121790Sjeffstatic void 519172409Sjeffsched_balance() 520116069Sjeff{ 521164936Sjulian struct tdq_group *high; 522164936Sjulian struct tdq_group *low; 523165620Sjeff struct tdq_group *tdg; 524172409Sjeff struct tdq *tdq; 525123487Sjeff int cnt; 526123487Sjeff int i; 527123487Sjeff 528172409Sjeff /* 529172409Sjeff * Select a random time between .5 * balance_interval and 530172409Sjeff * 1.5 * balance_interval. 531172409Sjeff */ 532172409Sjeff balance_ticks = max(balance_interval / 2, 1); 533172409Sjeff balance_ticks += random() % balance_interval; 534171482Sjeff if (smp_started == 0 || rebalance == 0) 535139334Sjeff return; 536172409Sjeff tdq = TDQ_SELF(); 537172409Sjeff TDQ_UNLOCK(tdq); 538123487Sjeff low = high = NULL; 539165620Sjeff i = random() % (tdg_maxid + 1); 540165620Sjeff for (cnt = 0; cnt <= tdg_maxid; cnt++) { 541165620Sjeff tdg = TDQ_GROUP(i); 542123487Sjeff /* 543123487Sjeff * Find the CPU with the highest load that has some 544123487Sjeff * threads to transfer. 545123487Sjeff */ 546165620Sjeff if ((high == NULL || tdg->tdg_load > high->tdg_load) 547165620Sjeff && tdg->tdg_transferable) 548165620Sjeff high = tdg; 549165620Sjeff if (low == NULL || tdg->tdg_load < low->tdg_load) 550165620Sjeff low = tdg; 551165620Sjeff if (++i > tdg_maxid) 552123487Sjeff i = 0; 553123487Sjeff } 554123487Sjeff if (low != NULL && high != NULL && high != low) 555165620Sjeff sched_balance_pair(LIST_FIRST(&high->tdg_members), 556165620Sjeff LIST_FIRST(&low->tdg_members)); 557172409Sjeff TDQ_LOCK(tdq); 558123487Sjeff} 559123487Sjeff 560171482Sjeff/* 561171482Sjeff * Balance load between CPUs in a group. Will only migrate within the group. 562171482Sjeff */ 563123487Sjeffstatic void 564172409Sjeffsched_balance_groups() 565123487Sjeff{ 566172409Sjeff struct tdq *tdq; 567123487Sjeff int i; 568123487Sjeff 569172409Sjeff /* 570172409Sjeff * Select a random time between .5 * balance_interval and 571172409Sjeff * 1.5 * balance_interval. 572172409Sjeff */ 573172409Sjeff balance_group_ticks = max(balance_interval / 2, 1); 574172409Sjeff balance_group_ticks += random() % balance_interval; 575171482Sjeff if (smp_started == 0 || rebalance == 0) 576171482Sjeff return; 577172409Sjeff tdq = TDQ_SELF(); 578172409Sjeff TDQ_UNLOCK(tdq); 579171482Sjeff for (i = 0; i <= tdg_maxid; i++) 580171482Sjeff sched_balance_group(TDQ_GROUP(i)); 581172409Sjeff TDQ_LOCK(tdq); 582123487Sjeff} 583123487Sjeff 584171482Sjeff/* 585171482Sjeff * Finds the greatest imbalance between two tdqs in a group. 586171482Sjeff */ 587123487Sjeffstatic void 588165620Sjeffsched_balance_group(struct tdq_group *tdg) 589123487Sjeff{ 590164936Sjulian struct tdq *tdq; 591164936Sjulian struct tdq *high; 592164936Sjulian struct tdq *low; 593123487Sjeff int load; 594123487Sjeff 595165620Sjeff if (tdg->tdg_transferable == 0) 596123487Sjeff return; 597123487Sjeff low = NULL; 598123487Sjeff high = NULL; 599165620Sjeff LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) { 600165620Sjeff load = tdq->tdq_load; 601165620Sjeff if (high == NULL || load > high->tdq_load) 602164936Sjulian high = tdq; 603165620Sjeff if (low == NULL || load < low->tdq_load) 604164936Sjulian low = tdq; 605123487Sjeff } 606123487Sjeff if (high != NULL && low != NULL && high != low) 607123487Sjeff sched_balance_pair(high, low); 608123487Sjeff} 609123487Sjeff 610171482Sjeff/* 611171482Sjeff * Lock two thread queues using their address to maintain lock order. 612171482Sjeff */ 613123487Sjeffstatic void 614171482Sjefftdq_lock_pair(struct tdq *one, struct tdq *two) 615171482Sjeff{ 616171482Sjeff if (one < two) { 617171482Sjeff TDQ_LOCK(one); 618171482Sjeff TDQ_LOCK_FLAGS(two, MTX_DUPOK); 619171482Sjeff } else { 620171482Sjeff TDQ_LOCK(two); 621171482Sjeff TDQ_LOCK_FLAGS(one, MTX_DUPOK); 622171482Sjeff } 623171482Sjeff} 624171482Sjeff 625171482Sjeff/* 626172409Sjeff * Unlock two thread queues. Order is not important here. 627172409Sjeff */ 628172409Sjeffstatic void 629172409Sjefftdq_unlock_pair(struct tdq *one, struct tdq *two) 630172409Sjeff{ 631172409Sjeff TDQ_UNLOCK(one); 632172409Sjeff TDQ_UNLOCK(two); 633172409Sjeff} 634172409Sjeff 635172409Sjeff/* 636171482Sjeff * Transfer load between two imbalanced thread queues. 637171482Sjeff */ 638171482Sjeffstatic void 639164936Sjuliansched_balance_pair(struct tdq *high, struct tdq *low) 640123487Sjeff{ 641123433Sjeff int transferable; 642116069Sjeff int high_load; 643116069Sjeff int low_load; 644116069Sjeff int move; 645116069Sjeff int diff; 646116069Sjeff int i; 647116069Sjeff 648171482Sjeff tdq_lock_pair(high, low); 649116069Sjeff /* 650123433Sjeff * If we're transfering within a group we have to use this specific 651164936Sjulian * tdq's transferable count, otherwise we can steal from other members 652123433Sjeff * of the group. 653123433Sjeff */ 654165620Sjeff if (high->tdq_group == low->tdq_group) { 655165620Sjeff transferable = high->tdq_transferable; 656165620Sjeff high_load = high->tdq_load; 657165620Sjeff low_load = low->tdq_load; 658123487Sjeff } else { 659165620Sjeff transferable = high->tdq_group->tdg_transferable; 660165620Sjeff high_load = high->tdq_group->tdg_load; 661165620Sjeff low_load = low->tdq_group->tdg_load; 662123487Sjeff } 663123433Sjeff /* 664122744Sjeff * Determine what the imbalance is and then adjust that to how many 665165620Sjeff * threads we actually have to give up (transferable). 666122744Sjeff */ 667171482Sjeff if (transferable != 0) { 668171482Sjeff diff = high_load - low_load; 669171482Sjeff move = diff / 2; 670171482Sjeff if (diff & 0x1) 671171482Sjeff move++; 672171482Sjeff move = min(move, transferable); 673171482Sjeff for (i = 0; i < move; i++) 674171482Sjeff tdq_move(high, low); 675172293Sjeff /* 676172293Sjeff * IPI the target cpu to force it to reschedule with the new 677172293Sjeff * workload. 678172293Sjeff */ 679172293Sjeff ipi_selected(1 << TDQ_ID(low), IPI_PREEMPT); 680171482Sjeff } 681172409Sjeff tdq_unlock_pair(high, low); 682116069Sjeff return; 683116069Sjeff} 684116069Sjeff 685171482Sjeff/* 686171482Sjeff * Move a thread from one thread queue to another. 687171482Sjeff */ 688121790Sjeffstatic void 689171482Sjefftdq_move(struct tdq *from, struct tdq *to) 690116069Sjeff{ 691171482Sjeff struct td_sched *ts; 692171482Sjeff struct thread *td; 693164936Sjulian struct tdq *tdq; 694171482Sjeff int cpu; 695116069Sjeff 696172409Sjeff TDQ_LOCK_ASSERT(from, MA_OWNED); 697172409Sjeff TDQ_LOCK_ASSERT(to, MA_OWNED); 698172409Sjeff 699164936Sjulian tdq = from; 700171482Sjeff cpu = TDQ_ID(to); 701172409Sjeff ts = tdq_steal(tdq); 702164936Sjulian if (ts == NULL) { 703165620Sjeff struct tdq_group *tdg; 704123433Sjeff 705165620Sjeff tdg = tdq->tdq_group; 706165620Sjeff LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) { 707165620Sjeff if (tdq == from || tdq->tdq_transferable == 0) 708123433Sjeff continue; 709172409Sjeff ts = tdq_steal(tdq); 710123433Sjeff break; 711123433Sjeff } 712164936Sjulian if (ts == NULL) 713171482Sjeff return; 714123433Sjeff } 715164936Sjulian if (tdq == to) 716123433Sjeff return; 717171482Sjeff td = ts->ts_thread; 718171482Sjeff /* 719171482Sjeff * Although the run queue is locked the thread may be blocked. Lock 720172409Sjeff * it to clear this and acquire the run-queue lock. 721171482Sjeff */ 722171482Sjeff thread_lock(td); 723172409Sjeff /* Drop recursive lock on from acquired via thread_lock(). */ 724171482Sjeff TDQ_UNLOCK(from); 725171482Sjeff sched_rem(td); 726166108Sjeff ts->ts_cpu = cpu; 727171482Sjeff td->td_lock = TDQ_LOCKPTR(to); 728171482Sjeff tdq_add(to, td, SRQ_YIELDING); 729116069Sjeff} 730110267Sjeff 731171482Sjeff/* 732171482Sjeff * This tdq has idled. Try to steal a thread from another cpu and switch 733171482Sjeff * to it. 734171482Sjeff */ 735123433Sjeffstatic int 736164936Sjuliantdq_idled(struct tdq *tdq) 737121790Sjeff{ 738165620Sjeff struct tdq_group *tdg; 739164936Sjulian struct tdq *steal; 740171482Sjeff int highload; 741171482Sjeff int highcpu; 742171482Sjeff int cpu; 743123433Sjeff 744172484Sjeff if (smp_started == 0 || steal_idle == 0) 745172484Sjeff return (1); 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 } 765172484Sjeff /* 766172484Sjeff * Find the least loaded CPU with a transferable thread and attempt 767172484Sjeff * to steal it. We make a lockless pass and then verify that the 768172484Sjeff * thread is still available after locking. 769172484Sjeff */ 770171482Sjeff for (;;) { 771171482Sjeff highcpu = 0; 772171482Sjeff highload = 0; 773171482Sjeff for (cpu = 0; cpu <= mp_maxid; cpu++) { 774171482Sjeff if (CPU_ABSENT(cpu)) 775171482Sjeff continue; 776166108Sjeff steal = TDQ_CPU(cpu); 777172484Sjeff if (steal->tdq_transferable == 0) 778166108Sjeff continue; 779172484Sjeff if (steal->tdq_load < highload) 780172484Sjeff continue; 781172484Sjeff highload = steal->tdq_load; 782171482Sjeff highcpu = cpu; 783171482Sjeff } 784171506Sjeff if (highload < steal_thresh) 785171482Sjeff break; 786171482Sjeff steal = TDQ_CPU(highcpu); 787172484Sjeff if (steal == tdq) 788172484Sjeff break; 789172409Sjeff tdq_lock_pair(tdq, steal); 790172484Sjeff if (steal->tdq_load >= steal_thresh && steal->tdq_transferable) 791166108Sjeff goto steal; 792172409Sjeff tdq_unlock_pair(tdq, steal); 793123433Sjeff } 794171482Sjeff spinlock_exit(); 795123433Sjeff return (1); 796166108Sjeffsteal: 797171482Sjeff spinlock_exit(); 798172409Sjeff tdq_move(steal, tdq); 799171482Sjeff TDQ_UNLOCK(steal); 800171482Sjeff mi_switch(SW_VOL, NULL); 801171482Sjeff thread_unlock(curthread); 802121790Sjeff 803166108Sjeff return (0); 804121790Sjeff} 805121790Sjeff 806171482Sjeff/* 807171482Sjeff * Notify a remote cpu of new work. Sends an IPI if criteria are met. 808171482Sjeff */ 809121790Sjeffstatic void 810166108Sjefftdq_notify(struct td_sched *ts) 811121790Sjeff{ 812166247Sjeff struct thread *ctd; 813121790Sjeff struct pcpu *pcpu; 814166247Sjeff int cpri; 815166247Sjeff int pri; 816166108Sjeff int cpu; 817121790Sjeff 818166108Sjeff cpu = ts->ts_cpu; 819166247Sjeff pri = ts->ts_thread->td_priority; 820166108Sjeff pcpu = pcpu_find(cpu); 821166247Sjeff ctd = pcpu->pc_curthread; 822166247Sjeff cpri = ctd->td_priority; 823166137Sjeff 824121790Sjeff /* 825166137Sjeff * If our priority is not better than the current priority there is 826166137Sjeff * nothing to do. 827166137Sjeff */ 828166247Sjeff if (pri > cpri) 829166137Sjeff return; 830166137Sjeff /* 831166247Sjeff * Always IPI idle. 832121790Sjeff */ 833166247Sjeff if (cpri > PRI_MIN_IDLE) 834166247Sjeff goto sendipi; 835166247Sjeff /* 836166247Sjeff * If we're realtime or better and there is timeshare or worse running 837166247Sjeff * send an IPI. 838166247Sjeff */ 839166247Sjeff if (pri < PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME) 840166247Sjeff goto sendipi; 841166247Sjeff /* 842166247Sjeff * Otherwise only IPI if we exceed the threshold. 843166247Sjeff */ 844171482Sjeff if (pri > preempt_thresh) 845165819Sjeff return; 846166247Sjeffsendipi: 847166247Sjeff ctd->td_flags |= TDF_NEEDRESCHED; 848171482Sjeff ipi_selected(1 << cpu, IPI_PREEMPT); 849121790Sjeff} 850121790Sjeff 851171482Sjeff/* 852171482Sjeff * Steals load from a timeshare queue. Honors the rotating queue head 853171482Sjeff * index. 854171482Sjeff */ 855164936Sjulianstatic struct td_sched * 856171482Sjeffrunq_steal_from(struct runq *rq, u_char start) 857171482Sjeff{ 858171482Sjeff struct td_sched *ts; 859171482Sjeff struct rqbits *rqb; 860171482Sjeff struct rqhead *rqh; 861171482Sjeff int first; 862171482Sjeff int bit; 863171482Sjeff int pri; 864171482Sjeff int i; 865171482Sjeff 866171482Sjeff rqb = &rq->rq_status; 867171482Sjeff bit = start & (RQB_BPW -1); 868171482Sjeff pri = 0; 869171482Sjeff first = 0; 870171482Sjeffagain: 871171482Sjeff for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 872171482Sjeff if (rqb->rqb_bits[i] == 0) 873171482Sjeff continue; 874171482Sjeff if (bit != 0) { 875171482Sjeff for (pri = bit; pri < RQB_BPW; pri++) 876171482Sjeff if (rqb->rqb_bits[i] & (1ul << pri)) 877171482Sjeff break; 878171482Sjeff if (pri >= RQB_BPW) 879171482Sjeff continue; 880171482Sjeff } else 881171482Sjeff pri = RQB_FFS(rqb->rqb_bits[i]); 882171482Sjeff pri += (i << RQB_L2BPW); 883171482Sjeff rqh = &rq->rq_queues[pri]; 884171482Sjeff TAILQ_FOREACH(ts, rqh, ts_procq) { 885171482Sjeff if (first && THREAD_CAN_MIGRATE(ts->ts_thread)) 886171482Sjeff return (ts); 887171482Sjeff first = 1; 888171482Sjeff } 889171482Sjeff } 890171482Sjeff if (start != 0) { 891171482Sjeff start = 0; 892171482Sjeff goto again; 893171482Sjeff } 894171482Sjeff 895171482Sjeff return (NULL); 896171482Sjeff} 897171482Sjeff 898171482Sjeff/* 899171482Sjeff * Steals load from a standard linear queue. 900171482Sjeff */ 901171482Sjeffstatic struct td_sched * 902121790Sjeffrunq_steal(struct runq *rq) 903121790Sjeff{ 904121790Sjeff struct rqhead *rqh; 905121790Sjeff struct rqbits *rqb; 906164936Sjulian struct td_sched *ts; 907121790Sjeff int word; 908121790Sjeff int bit; 909121790Sjeff 910121790Sjeff rqb = &rq->rq_status; 911121790Sjeff for (word = 0; word < RQB_LEN; word++) { 912121790Sjeff if (rqb->rqb_bits[word] == 0) 913121790Sjeff continue; 914121790Sjeff for (bit = 0; bit < RQB_BPW; bit++) { 915123231Speter if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 916121790Sjeff continue; 917121790Sjeff rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 918171506Sjeff TAILQ_FOREACH(ts, rqh, ts_procq) 919171506Sjeff if (THREAD_CAN_MIGRATE(ts->ts_thread)) 920164936Sjulian return (ts); 921121790Sjeff } 922121790Sjeff } 923121790Sjeff return (NULL); 924121790Sjeff} 925121790Sjeff 926171482Sjeff/* 927171482Sjeff * Attempt to steal a thread in priority order from a thread queue. 928171482Sjeff */ 929164936Sjulianstatic struct td_sched * 930172409Sjefftdq_steal(struct tdq *tdq) 931121790Sjeff{ 932164936Sjulian struct td_sched *ts; 933121790Sjeff 934171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 935165762Sjeff if ((ts = runq_steal(&tdq->tdq_realtime)) != NULL) 936164936Sjulian return (ts); 937171482Sjeff if ((ts = runq_steal_from(&tdq->tdq_timeshare, tdq->tdq_ridx)) != NULL) 938164936Sjulian return (ts); 939172409Sjeff return (runq_steal(&tdq->tdq_idle)); 940121790Sjeff} 941123433Sjeff 942171482Sjeff/* 943171482Sjeff * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 944172409Sjeff * current lock and returns with the assigned queue locked. 945171482Sjeff */ 946171482Sjeffstatic inline struct tdq * 947171482Sjeffsched_setcpu(struct td_sched *ts, int cpu, int flags) 948123433Sjeff{ 949171482Sjeff struct thread *td; 950171482Sjeff struct tdq *tdq; 951123433Sjeff 952171482Sjeff THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 953171482Sjeff 954171482Sjeff tdq = TDQ_CPU(cpu); 955171482Sjeff td = ts->ts_thread; 956171482Sjeff ts->ts_cpu = cpu; 957171713Sjeff 958171713Sjeff /* If the lock matches just return the queue. */ 959171482Sjeff if (td->td_lock == TDQ_LOCKPTR(tdq)) 960171482Sjeff return (tdq); 961171482Sjeff#ifdef notyet 962123433Sjeff /* 963172293Sjeff * If the thread isn't running its lockptr is a 964171482Sjeff * turnstile or a sleepqueue. We can just lock_set without 965171482Sjeff * blocking. 966123685Sjeff */ 967171482Sjeff if (TD_CAN_RUN(td)) { 968171482Sjeff TDQ_LOCK(tdq); 969171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 970171482Sjeff return (tdq); 971171482Sjeff } 972171482Sjeff#endif 973166108Sjeff /* 974171482Sjeff * The hard case, migration, we need to block the thread first to 975171482Sjeff * prevent order reversals with other cpus locks. 976166108Sjeff */ 977171482Sjeff thread_lock_block(td); 978171482Sjeff TDQ_LOCK(tdq); 979171713Sjeff thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 980171482Sjeff return (tdq); 981166108Sjeff} 982166108Sjeff 983171482Sjeff/* 984171482Sjeff * Find the thread queue running the lowest priority thread. 985171482Sjeff */ 986166108Sjeffstatic int 987171482Sjefftdq_lowestpri(void) 988166108Sjeff{ 989171482Sjeff struct tdq *tdq; 990166108Sjeff int lowpri; 991166108Sjeff int lowcpu; 992166108Sjeff int lowload; 993166108Sjeff int load; 994171482Sjeff int cpu; 995171482Sjeff int pri; 996171482Sjeff 997171482Sjeff lowload = 0; 998171482Sjeff lowpri = lowcpu = 0; 999171482Sjeff for (cpu = 0; cpu <= mp_maxid; cpu++) { 1000171482Sjeff if (CPU_ABSENT(cpu)) 1001171482Sjeff continue; 1002171482Sjeff tdq = TDQ_CPU(cpu); 1003171482Sjeff pri = tdq->tdq_lowpri; 1004171482Sjeff load = TDQ_CPU(cpu)->tdq_load; 1005171482Sjeff CTR4(KTR_ULE, 1006171482Sjeff "cpu %d pri %d lowcpu %d lowpri %d", 1007171482Sjeff cpu, pri, lowcpu, lowpri); 1008171482Sjeff if (pri < lowpri) 1009171482Sjeff continue; 1010171482Sjeff if (lowpri && lowpri == pri && load > lowload) 1011171482Sjeff continue; 1012171482Sjeff lowpri = pri; 1013171482Sjeff lowcpu = cpu; 1014171482Sjeff lowload = load; 1015171482Sjeff } 1016171482Sjeff 1017171482Sjeff return (lowcpu); 1018171482Sjeff} 1019171482Sjeff 1020171482Sjeff/* 1021171482Sjeff * Find the thread queue with the least load. 1022171482Sjeff */ 1023171482Sjeffstatic int 1024171482Sjefftdq_lowestload(void) 1025171482Sjeff{ 1026171482Sjeff struct tdq *tdq; 1027171482Sjeff int lowload; 1028171482Sjeff int lowpri; 1029171482Sjeff int lowcpu; 1030171482Sjeff int load; 1031171482Sjeff int cpu; 1032171482Sjeff int pri; 1033171482Sjeff 1034171482Sjeff lowcpu = 0; 1035171482Sjeff lowload = TDQ_CPU(0)->tdq_load; 1036171482Sjeff lowpri = TDQ_CPU(0)->tdq_lowpri; 1037171482Sjeff for (cpu = 1; cpu <= mp_maxid; cpu++) { 1038171482Sjeff if (CPU_ABSENT(cpu)) 1039171482Sjeff continue; 1040171482Sjeff tdq = TDQ_CPU(cpu); 1041171482Sjeff load = tdq->tdq_load; 1042171482Sjeff pri = tdq->tdq_lowpri; 1043171482Sjeff CTR4(KTR_ULE, "cpu %d load %d lowcpu %d lowload %d", 1044171482Sjeff cpu, load, lowcpu, lowload); 1045171482Sjeff if (load > lowload) 1046171482Sjeff continue; 1047171482Sjeff if (load == lowload && pri < lowpri) 1048171482Sjeff continue; 1049171482Sjeff lowcpu = cpu; 1050171482Sjeff lowload = load; 1051171482Sjeff lowpri = pri; 1052171482Sjeff } 1053171482Sjeff 1054171482Sjeff return (lowcpu); 1055171482Sjeff} 1056171482Sjeff 1057171482Sjeff/* 1058171482Sjeff * Pick the destination cpu for sched_add(). Respects affinity and makes 1059171482Sjeff * a determination based on load or priority of available processors. 1060171482Sjeff */ 1061171482Sjeffstatic int 1062171482Sjeffsched_pickcpu(struct td_sched *ts, int flags) 1063171482Sjeff{ 1064171482Sjeff struct tdq *tdq; 1065166108Sjeff int self; 1066166108Sjeff int pri; 1067166108Sjeff int cpu; 1068166108Sjeff 1069171482Sjeff cpu = self = PCPU_GET(cpuid); 1070166108Sjeff if (smp_started == 0) 1071166108Sjeff return (self); 1072171506Sjeff /* 1073171506Sjeff * Don't migrate a running thread from sched_switch(). 1074171506Sjeff */ 1075171506Sjeff if (flags & SRQ_OURSELF) { 1076171506Sjeff CTR1(KTR_ULE, "YIELDING %d", 1077171506Sjeff curthread->td_priority); 1078171506Sjeff return (self); 1079171506Sjeff } 1080166108Sjeff pri = ts->ts_thread->td_priority; 1081171482Sjeff cpu = ts->ts_cpu; 1082166108Sjeff /* 1083166108Sjeff * Regardless of affinity, if the last cpu is idle send it there. 1084166108Sjeff */ 1085171482Sjeff tdq = TDQ_CPU(cpu); 1086171482Sjeff if (tdq->tdq_lowpri > PRI_MIN_IDLE) { 1087166229Sjeff CTR5(KTR_ULE, 1088166108Sjeff "ts_cpu %d idle, ltick %d ticks %d pri %d curthread %d", 1089166108Sjeff ts->ts_cpu, ts->ts_rltick, ticks, pri, 1090171482Sjeff tdq->tdq_lowpri); 1091166108Sjeff return (ts->ts_cpu); 1092123433Sjeff } 1093166108Sjeff /* 1094166108Sjeff * If we have affinity, try to place it on the cpu we last ran on. 1095166108Sjeff */ 1096171482Sjeff if (SCHED_AFFINITY(ts) && tdq->tdq_lowpri > pri) { 1097166229Sjeff CTR5(KTR_ULE, 1098166108Sjeff "affinity for %d, ltick %d ticks %d pri %d curthread %d", 1099166108Sjeff ts->ts_cpu, ts->ts_rltick, ticks, pri, 1100171482Sjeff tdq->tdq_lowpri); 1101166108Sjeff return (ts->ts_cpu); 1102139334Sjeff } 1103123433Sjeff /* 1104166108Sjeff * Look for an idle group. 1105123433Sjeff */ 1106166229Sjeff CTR1(KTR_ULE, "tdq_idle %X", tdq_idle); 1107166108Sjeff cpu = ffs(tdq_idle); 1108166108Sjeff if (cpu) 1109171482Sjeff return (--cpu); 1110171506Sjeff /* 1111172409Sjeff * If there are no idle cores see if we can run the thread locally. 1112172409Sjeff * This may improve locality among sleepers and wakers when there 1113172409Sjeff * is shared data. 1114171506Sjeff */ 1115175348Sjeff if (tryself && pri < TDQ_CPU(self)->tdq_lowpri) { 1116171506Sjeff CTR1(KTR_ULE, "tryself %d", 1117166108Sjeff curthread->td_priority); 1118166108Sjeff return (self); 1119123433Sjeff } 1120133427Sjeff /* 1121166108Sjeff * Now search for the cpu running the lowest priority thread with 1122166108Sjeff * the least load. 1123123433Sjeff */ 1124171482Sjeff if (pick_pri) 1125171482Sjeff cpu = tdq_lowestpri(); 1126171482Sjeff else 1127171482Sjeff cpu = tdq_lowestload(); 1128171482Sjeff return (cpu); 1129123433Sjeff} 1130123433Sjeff 1131121790Sjeff#endif /* SMP */ 1132121790Sjeff 1133117326Sjeff/* 1134121790Sjeff * Pick the highest priority task we have and return it. 1135117326Sjeff */ 1136164936Sjulianstatic struct td_sched * 1137164936Sjuliantdq_choose(struct tdq *tdq) 1138110267Sjeff{ 1139164936Sjulian struct td_sched *ts; 1140110267Sjeff 1141171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1142165762Sjeff ts = runq_choose(&tdq->tdq_realtime); 1143170787Sjeff if (ts != NULL) 1144164936Sjulian return (ts); 1145165766Sjeff ts = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1146165762Sjeff if (ts != NULL) { 1147170787Sjeff KASSERT(ts->ts_thread->td_priority >= PRI_MIN_TIMESHARE, 1148165762Sjeff ("tdq_choose: Invalid priority on timeshare queue %d", 1149165762Sjeff ts->ts_thread->td_priority)); 1150165762Sjeff return (ts); 1151165762Sjeff } 1152110267Sjeff 1153165762Sjeff ts = runq_choose(&tdq->tdq_idle); 1154165762Sjeff if (ts != NULL) { 1155165762Sjeff KASSERT(ts->ts_thread->td_priority >= PRI_MIN_IDLE, 1156165762Sjeff ("tdq_choose: Invalid priority on idle queue %d", 1157165762Sjeff ts->ts_thread->td_priority)); 1158165762Sjeff return (ts); 1159165762Sjeff } 1160165762Sjeff 1161165762Sjeff return (NULL); 1162110267Sjeff} 1163110267Sjeff 1164171482Sjeff/* 1165171482Sjeff * Initialize a thread queue. 1166171482Sjeff */ 1167109864Sjeffstatic void 1168164936Sjuliantdq_setup(struct tdq *tdq) 1169110028Sjeff{ 1170171482Sjeff 1171171713Sjeff if (bootverbose) 1172171713Sjeff printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1173165762Sjeff runq_init(&tdq->tdq_realtime); 1174165762Sjeff runq_init(&tdq->tdq_timeshare); 1175165620Sjeff runq_init(&tdq->tdq_idle); 1176165620Sjeff tdq->tdq_load = 0; 1177110028Sjeff} 1178110028Sjeff 1179171713Sjeff#ifdef SMP 1180110028Sjeffstatic void 1181171713Sjefftdg_setup(struct tdq_group *tdg) 1182109864Sjeff{ 1183171713Sjeff if (bootverbose) 1184171713Sjeff printf("ULE: setup cpu group %d\n", TDG_ID(tdg)); 1185171713Sjeff snprintf(tdg->tdg_name, sizeof(tdg->tdg_name), 1186171713Sjeff "sched lock %d", (int)TDG_ID(tdg)); 1187171713Sjeff mtx_init(&tdg->tdg_lock, tdg->tdg_name, "sched lock", 1188171713Sjeff MTX_SPIN | MTX_RECURSE); 1189171713Sjeff LIST_INIT(&tdg->tdg_members); 1190171713Sjeff tdg->tdg_load = 0; 1191171713Sjeff tdg->tdg_transferable = 0; 1192171713Sjeff tdg->tdg_cpus = 0; 1193171713Sjeff tdg->tdg_mask = 0; 1194171713Sjeff tdg->tdg_cpumask = 0; 1195171713Sjeff tdg->tdg_idlemask = 0; 1196171713Sjeff} 1197171713Sjeff 1198171713Sjeffstatic void 1199171713Sjefftdg_add(struct tdq_group *tdg, struct tdq *tdq) 1200171713Sjeff{ 1201171713Sjeff if (tdg->tdg_mask == 0) 1202171713Sjeff tdg->tdg_mask |= 1 << TDQ_ID(tdq); 1203171713Sjeff tdg->tdg_cpumask |= 1 << TDQ_ID(tdq); 1204171713Sjeff tdg->tdg_cpus++; 1205171713Sjeff tdq->tdq_group = tdg; 1206171713Sjeff tdq->tdq_lock = &tdg->tdg_lock; 1207171713Sjeff LIST_INSERT_HEAD(&tdg->tdg_members, tdq, tdq_siblings); 1208171713Sjeff if (bootverbose) 1209171713Sjeff printf("ULE: adding cpu %d to group %d: cpus %d mask 0x%X\n", 1210171713Sjeff TDQ_ID(tdq), TDG_ID(tdg), tdg->tdg_cpus, tdg->tdg_cpumask); 1211171713Sjeff} 1212171713Sjeff 1213171713Sjeffstatic void 1214171713Sjeffsched_setup_smp(void) 1215171713Sjeff{ 1216171713Sjeff struct tdq_group *tdg; 1217171713Sjeff struct tdq *tdq; 1218171713Sjeff int cpus; 1219171713Sjeff int i; 1220171713Sjeff 1221171713Sjeff for (cpus = 0, i = 0; i < MAXCPU; i++) { 1222171713Sjeff if (CPU_ABSENT(i)) 1223171713Sjeff continue; 1224165627Sjeff tdq = &tdq_cpu[i]; 1225171713Sjeff tdg = &tdq_groups[i]; 1226171713Sjeff /* 1227171713Sjeff * Setup a tdq group with one member. 1228171713Sjeff */ 1229171713Sjeff tdg_setup(tdg); 1230171713Sjeff tdq_setup(tdq); 1231171713Sjeff tdg_add(tdg, tdq); 1232171713Sjeff cpus++; 1233123433Sjeff } 1234171713Sjeff tdg_maxid = cpus - 1; 1235171713Sjeff} 1236171713Sjeff#endif 1237171713Sjeff 1238171713Sjeff/* 1239171713Sjeff * Setup the thread queues and initialize the topology based on MD 1240171713Sjeff * information. 1241171713Sjeff */ 1242171713Sjeffstatic void 1243171713Sjeffsched_setup(void *dummy) 1244171713Sjeff{ 1245171713Sjeff struct tdq *tdq; 1246171713Sjeff 1247171713Sjeff tdq = TDQ_SELF(); 1248171713Sjeff#ifdef SMP 1249171713Sjeff /* 1250171713Sjeff * Setup tdqs based on a topology configuration or vanilla SMP based 1251171713Sjeff * on mp_maxid. 1252171713Sjeff */ 1253176734Sjeff sched_setup_smp(); 1254172409Sjeff balance_tdq = tdq; 1255172409Sjeff sched_balance(); 1256117237Sjeff#else 1257171713Sjeff tdq_setup(tdq); 1258171713Sjeff mtx_init(&tdq_lock, "sched lock", "sched lock", MTX_SPIN | MTX_RECURSE); 1259171713Sjeff tdq->tdq_lock = &tdq_lock; 1260116069Sjeff#endif 1261171482Sjeff /* 1262171482Sjeff * To avoid divide-by-zero, we set realstathz a dummy value 1263171482Sjeff * in case which sched_clock() called before sched_initticks(). 1264171482Sjeff */ 1265171482Sjeff realstathz = hz; 1266171482Sjeff sched_slice = (realstathz/10); /* ~100ms */ 1267171482Sjeff tickincr = 1 << SCHED_TICK_SHIFT; 1268171482Sjeff 1269171482Sjeff /* Add thread0's load since it's running. */ 1270171482Sjeff TDQ_LOCK(tdq); 1271171713Sjeff thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1272171482Sjeff tdq_load_add(tdq, &td_sched0); 1273171482Sjeff TDQ_UNLOCK(tdq); 1274109864Sjeff} 1275109864Sjeff 1276171482Sjeff/* 1277171482Sjeff * This routine determines the tickincr after stathz and hz are setup. 1278171482Sjeff */ 1279153533Sdavidxu/* ARGSUSED */ 1280153533Sdavidxustatic void 1281153533Sdavidxusched_initticks(void *dummy) 1282153533Sdavidxu{ 1283171482Sjeff int incr; 1284171482Sjeff 1285153533Sdavidxu realstathz = stathz ? stathz : hz; 1286166229Sjeff sched_slice = (realstathz/10); /* ~100ms */ 1287153533Sdavidxu 1288153533Sdavidxu /* 1289165762Sjeff * tickincr is shifted out by 10 to avoid rounding errors due to 1290165766Sjeff * hz not being evenly divisible by stathz on all platforms. 1291153533Sdavidxu */ 1292171482Sjeff incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1293165762Sjeff /* 1294165762Sjeff * This does not work for values of stathz that are more than 1295165762Sjeff * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1296165762Sjeff */ 1297171482Sjeff if (incr == 0) 1298171482Sjeff incr = 1; 1299171482Sjeff tickincr = incr; 1300166108Sjeff#ifdef SMP 1301171899Sjeff /* 1302172409Sjeff * Set the default balance interval now that we know 1303172409Sjeff * what realstathz is. 1304172409Sjeff */ 1305172409Sjeff balance_interval = realstathz; 1306172409Sjeff /* 1307171899Sjeff * Set steal thresh to log2(mp_ncpu) but no greater than 4. This 1308171899Sjeff * prevents excess thrashing on large machines and excess idle on 1309171899Sjeff * smaller machines. 1310171899Sjeff */ 1311171899Sjeff steal_thresh = min(ffs(mp_ncpus) - 1, 4); 1312166108Sjeff affinity = SCHED_AFFINITY_DEFAULT; 1313166108Sjeff#endif 1314153533Sdavidxu} 1315153533Sdavidxu 1316153533Sdavidxu 1317109864Sjeff/* 1318171482Sjeff * This is the core of the interactivity algorithm. Determines a score based 1319171482Sjeff * on past behavior. It is the ratio of sleep time to run time scaled to 1320171482Sjeff * a [0, 100] integer. This is the voluntary sleep time of a process, which 1321171482Sjeff * differs from the cpu usage because it does not account for time spent 1322171482Sjeff * waiting on a run-queue. Would be prettier if we had floating point. 1323171482Sjeff */ 1324171482Sjeffstatic int 1325171482Sjeffsched_interact_score(struct thread *td) 1326171482Sjeff{ 1327171482Sjeff struct td_sched *ts; 1328171482Sjeff int div; 1329171482Sjeff 1330171482Sjeff ts = td->td_sched; 1331171482Sjeff /* 1332171482Sjeff * The score is only needed if this is likely to be an interactive 1333171482Sjeff * task. Don't go through the expense of computing it if there's 1334171482Sjeff * no chance. 1335171482Sjeff */ 1336171482Sjeff if (sched_interact <= SCHED_INTERACT_HALF && 1337171482Sjeff ts->ts_runtime >= ts->ts_slptime) 1338171482Sjeff return (SCHED_INTERACT_HALF); 1339171482Sjeff 1340171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1341171482Sjeff div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1342171482Sjeff return (SCHED_INTERACT_HALF + 1343171482Sjeff (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1344171482Sjeff } 1345171482Sjeff if (ts->ts_slptime > ts->ts_runtime) { 1346171482Sjeff div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1347171482Sjeff return (ts->ts_runtime / div); 1348171482Sjeff } 1349171482Sjeff /* runtime == slptime */ 1350171482Sjeff if (ts->ts_runtime) 1351171482Sjeff return (SCHED_INTERACT_HALF); 1352171482Sjeff 1353171482Sjeff /* 1354171482Sjeff * This can happen if slptime and runtime are 0. 1355171482Sjeff */ 1356171482Sjeff return (0); 1357171482Sjeff 1358171482Sjeff} 1359171482Sjeff 1360171482Sjeff/* 1361109864Sjeff * Scale the scheduling priority according to the "interactivity" of this 1362109864Sjeff * process. 1363109864Sjeff */ 1364113357Sjeffstatic void 1365163709Sjbsched_priority(struct thread *td) 1366109864Sjeff{ 1367165762Sjeff int score; 1368109864Sjeff int pri; 1369109864Sjeff 1370163709Sjb if (td->td_pri_class != PRI_TIMESHARE) 1371113357Sjeff return; 1372112966Sjeff /* 1373165762Sjeff * If the score is interactive we place the thread in the realtime 1374165762Sjeff * queue with a priority that is less than kernel and interrupt 1375165762Sjeff * priorities. These threads are not subject to nice restrictions. 1376112966Sjeff * 1377171482Sjeff * Scores greater than this are placed on the normal timeshare queue 1378165762Sjeff * where the priority is partially decided by the most recent cpu 1379165762Sjeff * utilization and the rest is decided by nice value. 1380172293Sjeff * 1381172293Sjeff * The nice value of the process has a linear effect on the calculated 1382172293Sjeff * score. Negative nice values make it easier for a thread to be 1383172293Sjeff * considered interactive. 1384112966Sjeff */ 1385172308Sjeff score = imax(0, sched_interact_score(td) - td->td_proc->p_nice); 1386165762Sjeff if (score < sched_interact) { 1387165762Sjeff pri = PRI_MIN_REALTIME; 1388165762Sjeff pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact) 1389165762Sjeff * score; 1390165762Sjeff KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME, 1391166208Sjeff ("sched_priority: invalid interactive priority %d score %d", 1392166208Sjeff pri, score)); 1393165762Sjeff } else { 1394165762Sjeff pri = SCHED_PRI_MIN; 1395165762Sjeff if (td->td_sched->ts_ticks) 1396165762Sjeff pri += SCHED_PRI_TICKS(td->td_sched); 1397165762Sjeff pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1398171482Sjeff KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE, 1399171482Sjeff ("sched_priority: invalid priority %d: nice %d, " 1400171482Sjeff "ticks %d ftick %d ltick %d tick pri %d", 1401171482Sjeff pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1402171482Sjeff td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1403171482Sjeff SCHED_PRI_TICKS(td->td_sched))); 1404165762Sjeff } 1405165762Sjeff sched_user_prio(td, pri); 1406112966Sjeff 1407112966Sjeff return; 1408109864Sjeff} 1409109864Sjeff 1410121868Sjeff/* 1411121868Sjeff * This routine enforces a maximum limit on the amount of scheduling history 1412171482Sjeff * kept. It is called after either the slptime or runtime is adjusted. This 1413171482Sjeff * function is ugly due to integer math. 1414121868Sjeff */ 1415116463Sjeffstatic void 1416163709Sjbsched_interact_update(struct thread *td) 1417116463Sjeff{ 1418165819Sjeff struct td_sched *ts; 1419166208Sjeff u_int sum; 1420121605Sjeff 1421165819Sjeff ts = td->td_sched; 1422171482Sjeff sum = ts->ts_runtime + ts->ts_slptime; 1423121868Sjeff if (sum < SCHED_SLP_RUN_MAX) 1424121868Sjeff return; 1425121868Sjeff /* 1426165819Sjeff * This only happens from two places: 1427165819Sjeff * 1) We have added an unusual amount of run time from fork_exit. 1428165819Sjeff * 2) We have added an unusual amount of sleep time from sched_sleep(). 1429165819Sjeff */ 1430165819Sjeff if (sum > SCHED_SLP_RUN_MAX * 2) { 1431171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1432171482Sjeff ts->ts_runtime = SCHED_SLP_RUN_MAX; 1433171482Sjeff ts->ts_slptime = 1; 1434165819Sjeff } else { 1435171482Sjeff ts->ts_slptime = SCHED_SLP_RUN_MAX; 1436171482Sjeff ts->ts_runtime = 1; 1437165819Sjeff } 1438165819Sjeff return; 1439165819Sjeff } 1440165819Sjeff /* 1441121868Sjeff * If we have exceeded by more than 1/5th then the algorithm below 1442121868Sjeff * will not bring us back into range. Dividing by two here forces 1443133427Sjeff * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1444121868Sjeff */ 1445127850Sjeff if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1446171482Sjeff ts->ts_runtime /= 2; 1447171482Sjeff ts->ts_slptime /= 2; 1448121868Sjeff return; 1449116463Sjeff } 1450171482Sjeff ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1451171482Sjeff ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1452116463Sjeff} 1453116463Sjeff 1454171482Sjeff/* 1455171482Sjeff * Scale back the interactivity history when a child thread is created. The 1456171482Sjeff * history is inherited from the parent but the thread may behave totally 1457171482Sjeff * differently. For example, a shell spawning a compiler process. We want 1458171482Sjeff * to learn that the compiler is behaving badly very quickly. 1459171482Sjeff */ 1460121868Sjeffstatic void 1461163709Sjbsched_interact_fork(struct thread *td) 1462121868Sjeff{ 1463121868Sjeff int ratio; 1464121868Sjeff int sum; 1465121868Sjeff 1466171482Sjeff sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1467121868Sjeff if (sum > SCHED_SLP_RUN_FORK) { 1468121868Sjeff ratio = sum / SCHED_SLP_RUN_FORK; 1469171482Sjeff td->td_sched->ts_runtime /= ratio; 1470171482Sjeff td->td_sched->ts_slptime /= ratio; 1471121868Sjeff } 1472121868Sjeff} 1473121868Sjeff 1474113357Sjeff/* 1475171482Sjeff * Called from proc0_init() to setup the scheduler fields. 1476134791Sjulian */ 1477134791Sjulianvoid 1478134791Sjulianschedinit(void) 1479134791Sjulian{ 1480165762Sjeff 1481134791Sjulian /* 1482134791Sjulian * Set up the scheduler specific parts of proc0. 1483134791Sjulian */ 1484136167Sjulian proc0.p_sched = NULL; /* XXX */ 1485164936Sjulian thread0.td_sched = &td_sched0; 1486165762Sjeff td_sched0.ts_ltick = ticks; 1487165796Sjeff td_sched0.ts_ftick = ticks; 1488164936Sjulian td_sched0.ts_thread = &thread0; 1489134791Sjulian} 1490134791Sjulian 1491134791Sjulian/* 1492113357Sjeff * This is only somewhat accurate since given many processes of the same 1493113357Sjeff * priority they will switch when their slices run out, which will be 1494165762Sjeff * at most sched_slice stathz ticks. 1495113357Sjeff */ 1496109864Sjeffint 1497109864Sjeffsched_rr_interval(void) 1498109864Sjeff{ 1499165762Sjeff 1500165762Sjeff /* Convert sched_slice to hz */ 1501165762Sjeff return (hz/(realstathz/sched_slice)); 1502109864Sjeff} 1503109864Sjeff 1504171482Sjeff/* 1505171482Sjeff * Update the percent cpu tracking information when it is requested or 1506171482Sjeff * the total history exceeds the maximum. We keep a sliding history of 1507171482Sjeff * tick counts that slowly decays. This is less precise than the 4BSD 1508171482Sjeff * mechanism since it happens with less regular and frequent events. 1509171482Sjeff */ 1510121790Sjeffstatic void 1511164936Sjuliansched_pctcpu_update(struct td_sched *ts) 1512109864Sjeff{ 1513165762Sjeff 1514165762Sjeff if (ts->ts_ticks == 0) 1515165762Sjeff return; 1516165796Sjeff if (ticks - (hz / 10) < ts->ts_ltick && 1517165796Sjeff SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1518165796Sjeff return; 1519109864Sjeff /* 1520109864Sjeff * Adjust counters and watermark for pctcpu calc. 1521116365Sjeff */ 1522165762Sjeff if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1523164936Sjulian ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1524165762Sjeff SCHED_TICK_TARG; 1525165762Sjeff else 1526164936Sjulian ts->ts_ticks = 0; 1527164936Sjulian ts->ts_ltick = ticks; 1528165762Sjeff ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1529109864Sjeff} 1530109864Sjeff 1531171482Sjeff/* 1532171482Sjeff * Adjust the priority of a thread. Move it to the appropriate run-queue 1533171482Sjeff * if necessary. This is the back-end for several priority related 1534171482Sjeff * functions. 1535171482Sjeff */ 1536165762Sjeffstatic void 1537139453Sjhbsched_thread_priority(struct thread *td, u_char prio) 1538109864Sjeff{ 1539164936Sjulian struct td_sched *ts; 1540109864Sjeff 1541139316Sjeff CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 1542173600Sjulian td, td->td_name, td->td_priority, prio, curthread, 1543173600Sjulian curthread->td_name); 1544164936Sjulian ts = td->td_sched; 1545170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1546139453Sjhb if (td->td_priority == prio) 1547139453Sjhb return; 1548165762Sjeff 1549165766Sjeff if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1550121605Sjeff /* 1551121605Sjeff * If the priority has been elevated due to priority 1552121605Sjeff * propagation, we may have to move ourselves to a new 1553165762Sjeff * queue. This could be optimized to not re-add in some 1554165762Sjeff * cases. 1555133555Sjeff */ 1556165762Sjeff sched_rem(td); 1557165762Sjeff td->td_priority = prio; 1558171482Sjeff sched_add(td, SRQ_BORROWING); 1559171482Sjeff#ifdef SMP 1560175587Sjeff } else if (TD_IS_RUNNING(td)) { 1561171482Sjeff struct tdq *tdq; 1562171482Sjeff 1563171482Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1564175587Sjeff if (prio < tdq->tdq_lowpri || 1565175587Sjeff (td->td_priority == tdq->tdq_lowpri && tdq->tdq_load <= 1)) 1566171482Sjeff tdq->tdq_lowpri = prio; 1567175587Sjeff td->td_priority = prio; 1568171482Sjeff#endif 1569175587Sjeff } else 1570119488Sdavidxu td->td_priority = prio; 1571109864Sjeff} 1572109864Sjeff 1573139453Sjhb/* 1574139453Sjhb * Update a thread's priority when it is lent another thread's 1575139453Sjhb * priority. 1576139453Sjhb */ 1577109864Sjeffvoid 1578139453Sjhbsched_lend_prio(struct thread *td, u_char prio) 1579139453Sjhb{ 1580139453Sjhb 1581139453Sjhb td->td_flags |= TDF_BORROWING; 1582139453Sjhb sched_thread_priority(td, prio); 1583139453Sjhb} 1584139453Sjhb 1585139453Sjhb/* 1586139453Sjhb * Restore a thread's priority when priority propagation is 1587139453Sjhb * over. The prio argument is the minimum priority the thread 1588139453Sjhb * needs to have to satisfy other possible priority lending 1589139453Sjhb * requests. If the thread's regular priority is less 1590139453Sjhb * important than prio, the thread will keep a priority boost 1591139453Sjhb * of prio. 1592139453Sjhb */ 1593139453Sjhbvoid 1594139453Sjhbsched_unlend_prio(struct thread *td, u_char prio) 1595139453Sjhb{ 1596139453Sjhb u_char base_pri; 1597139453Sjhb 1598139453Sjhb if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1599139453Sjhb td->td_base_pri <= PRI_MAX_TIMESHARE) 1600163709Sjb base_pri = td->td_user_pri; 1601139453Sjhb else 1602139453Sjhb base_pri = td->td_base_pri; 1603139453Sjhb if (prio >= base_pri) { 1604139455Sjhb td->td_flags &= ~TDF_BORROWING; 1605139453Sjhb sched_thread_priority(td, base_pri); 1606139453Sjhb } else 1607139453Sjhb sched_lend_prio(td, prio); 1608139453Sjhb} 1609139453Sjhb 1610171482Sjeff/* 1611171482Sjeff * Standard entry for setting the priority to an absolute value. 1612171482Sjeff */ 1613139453Sjhbvoid 1614139453Sjhbsched_prio(struct thread *td, u_char prio) 1615139453Sjhb{ 1616139453Sjhb u_char oldprio; 1617139453Sjhb 1618139453Sjhb /* First, update the base priority. */ 1619139453Sjhb td->td_base_pri = prio; 1620139453Sjhb 1621139453Sjhb /* 1622139455Sjhb * If the thread is borrowing another thread's priority, don't 1623139453Sjhb * ever lower the priority. 1624139453Sjhb */ 1625139453Sjhb if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1626139453Sjhb return; 1627139453Sjhb 1628139453Sjhb /* Change the real priority. */ 1629139453Sjhb oldprio = td->td_priority; 1630139453Sjhb sched_thread_priority(td, prio); 1631139453Sjhb 1632139453Sjhb /* 1633139453Sjhb * If the thread is on a turnstile, then let the turnstile update 1634139453Sjhb * its state. 1635139453Sjhb */ 1636139453Sjhb if (TD_ON_LOCK(td) && oldprio != prio) 1637139453Sjhb turnstile_adjust(td, oldprio); 1638139453Sjhb} 1639139455Sjhb 1640171482Sjeff/* 1641171482Sjeff * Set the base user priority, does not effect current running priority. 1642171482Sjeff */ 1643139453Sjhbvoid 1644163709Sjbsched_user_prio(struct thread *td, u_char prio) 1645161599Sdavidxu{ 1646161599Sdavidxu u_char oldprio; 1647161599Sdavidxu 1648163709Sjb td->td_base_user_pri = prio; 1649164939Sjulian if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio) 1650164939Sjulian return; 1651163709Sjb oldprio = td->td_user_pri; 1652163709Sjb td->td_user_pri = prio; 1653161599Sdavidxu} 1654161599Sdavidxu 1655161599Sdavidxuvoid 1656161599Sdavidxusched_lend_user_prio(struct thread *td, u_char prio) 1657161599Sdavidxu{ 1658161599Sdavidxu u_char oldprio; 1659161599Sdavidxu 1660174536Sdavidxu THREAD_LOCK_ASSERT(td, MA_OWNED); 1661161599Sdavidxu td->td_flags |= TDF_UBORROWING; 1662164091Smaxim oldprio = td->td_user_pri; 1663163709Sjb td->td_user_pri = prio; 1664161599Sdavidxu} 1665161599Sdavidxu 1666161599Sdavidxuvoid 1667161599Sdavidxusched_unlend_user_prio(struct thread *td, u_char prio) 1668161599Sdavidxu{ 1669161599Sdavidxu u_char base_pri; 1670161599Sdavidxu 1671174536Sdavidxu THREAD_LOCK_ASSERT(td, MA_OWNED); 1672163709Sjb base_pri = td->td_base_user_pri; 1673161599Sdavidxu if (prio >= base_pri) { 1674161599Sdavidxu td->td_flags &= ~TDF_UBORROWING; 1675163709Sjb sched_user_prio(td, base_pri); 1676174536Sdavidxu } else { 1677161599Sdavidxu sched_lend_user_prio(td, prio); 1678174536Sdavidxu } 1679161599Sdavidxu} 1680161599Sdavidxu 1681171482Sjeff/* 1682171505Sjeff * Add the thread passed as 'newtd' to the run queue before selecting 1683171505Sjeff * the next thread to run. This is only used for KSE. 1684171505Sjeff */ 1685171505Sjeffstatic void 1686171505Sjeffsched_switchin(struct tdq *tdq, struct thread *td) 1687171505Sjeff{ 1688171505Sjeff#ifdef SMP 1689171505Sjeff spinlock_enter(); 1690171505Sjeff TDQ_UNLOCK(tdq); 1691171505Sjeff thread_lock(td); 1692171505Sjeff spinlock_exit(); 1693171505Sjeff sched_setcpu(td->td_sched, TDQ_ID(tdq), SRQ_YIELDING); 1694171505Sjeff#else 1695171505Sjeff td->td_lock = TDQ_LOCKPTR(tdq); 1696171505Sjeff#endif 1697171505Sjeff tdq_add(tdq, td, SRQ_YIELDING); 1698171505Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1699171505Sjeff} 1700171505Sjeff 1701171505Sjeff/* 1702174847Swkoszek * Block a thread for switching. Similar to thread_block() but does not 1703174847Swkoszek * bump the spin count. 1704174847Swkoszek */ 1705174847Swkoszekstatic inline struct mtx * 1706174847Swkoszekthread_block_switch(struct thread *td) 1707174847Swkoszek{ 1708174847Swkoszek struct mtx *lock; 1709174847Swkoszek 1710174847Swkoszek THREAD_LOCK_ASSERT(td, MA_OWNED); 1711174847Swkoszek lock = td->td_lock; 1712174847Swkoszek td->td_lock = &blocked_lock; 1713174847Swkoszek mtx_unlock_spin(lock); 1714174847Swkoszek 1715174847Swkoszek return (lock); 1716174847Swkoszek} 1717174847Swkoszek 1718174847Swkoszek/* 1719171713Sjeff * Handle migration from sched_switch(). This happens only for 1720171713Sjeff * cpu binding. 1721171713Sjeff */ 1722171713Sjeffstatic struct mtx * 1723171713Sjeffsched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1724171713Sjeff{ 1725171713Sjeff struct tdq *tdn; 1726171713Sjeff 1727171713Sjeff tdn = TDQ_CPU(td->td_sched->ts_cpu); 1728171713Sjeff#ifdef SMP 1729171713Sjeff /* 1730171713Sjeff * Do the lock dance required to avoid LOR. We grab an extra 1731171713Sjeff * spinlock nesting to prevent preemption while we're 1732171713Sjeff * not holding either run-queue lock. 1733171713Sjeff */ 1734171713Sjeff spinlock_enter(); 1735171713Sjeff thread_block_switch(td); /* This releases the lock on tdq. */ 1736171713Sjeff TDQ_LOCK(tdn); 1737171713Sjeff tdq_add(tdn, td, flags); 1738171713Sjeff tdq_notify(td->td_sched); 1739171713Sjeff /* 1740171713Sjeff * After we unlock tdn the new cpu still can't switch into this 1741171713Sjeff * thread until we've unblocked it in cpu_switch(). The lock 1742171713Sjeff * pointers may match in the case of HTT cores. Don't unlock here 1743171713Sjeff * or we can deadlock when the other CPU runs the IPI handler. 1744171713Sjeff */ 1745171713Sjeff if (TDQ_LOCKPTR(tdn) != TDQ_LOCKPTR(tdq)) { 1746171713Sjeff TDQ_UNLOCK(tdn); 1747171713Sjeff TDQ_LOCK(tdq); 1748171713Sjeff } 1749171713Sjeff spinlock_exit(); 1750171713Sjeff#endif 1751171713Sjeff return (TDQ_LOCKPTR(tdn)); 1752171713Sjeff} 1753171713Sjeff 1754171713Sjeff/* 1755171482Sjeff * Release a thread that was blocked with thread_block_switch(). 1756171482Sjeff */ 1757171482Sjeffstatic inline void 1758171482Sjeffthread_unblock_switch(struct thread *td, struct mtx *mtx) 1759171482Sjeff{ 1760171482Sjeff atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1761171482Sjeff (uintptr_t)mtx); 1762171482Sjeff} 1763171482Sjeff 1764171482Sjeff/* 1765171482Sjeff * Switch threads. This function has to handle threads coming in while 1766171482Sjeff * blocked for some reason, running, or idle. It also must deal with 1767171482Sjeff * migrating a thread from one queue to another as running threads may 1768171482Sjeff * be assigned elsewhere via binding. 1769171482Sjeff */ 1770161599Sdavidxuvoid 1771135051Sjuliansched_switch(struct thread *td, struct thread *newtd, int flags) 1772109864Sjeff{ 1773165627Sjeff struct tdq *tdq; 1774164936Sjulian struct td_sched *ts; 1775171482Sjeff struct mtx *mtx; 1776171713Sjeff int srqflag; 1777171482Sjeff int cpuid; 1778109864Sjeff 1779170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1780109864Sjeff 1781171482Sjeff cpuid = PCPU_GET(cpuid); 1782171482Sjeff tdq = TDQ_CPU(cpuid); 1783164936Sjulian ts = td->td_sched; 1784171713Sjeff mtx = td->td_lock; 1785171482Sjeff#ifdef SMP 1786171482Sjeff ts->ts_rltick = ticks; 1787171482Sjeff#endif 1788133555Sjeff td->td_lastcpu = td->td_oncpu; 1789113339Sjulian td->td_oncpu = NOCPU; 1790132266Sjhb td->td_flags &= ~TDF_NEEDRESCHED; 1791144777Sups td->td_owepreempt = 0; 1792123434Sjeff /* 1793171482Sjeff * The lock pointer in an idle thread should never change. Reset it 1794171482Sjeff * to CAN_RUN as well. 1795123434Sjeff */ 1796167327Sjulian if (TD_IS_IDLETHREAD(td)) { 1797171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1798139334Sjeff TD_SET_CAN_RUN(td); 1799170293Sjeff } else if (TD_IS_RUNNING(td)) { 1800171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1801165627Sjeff tdq_load_rem(tdq, ts); 1802171713Sjeff srqflag = (flags & SW_PREEMPT) ? 1803170293Sjeff SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1804171713Sjeff SRQ_OURSELF|SRQ_YIELDING; 1805171713Sjeff if (ts->ts_cpu == cpuid) 1806171713Sjeff tdq_add(tdq, td, srqflag); 1807171713Sjeff else 1808171713Sjeff mtx = sched_switch_migrate(tdq, td, srqflag); 1809171482Sjeff } else { 1810171482Sjeff /* This thread must be going to sleep. */ 1811171482Sjeff TDQ_LOCK(tdq); 1812171482Sjeff mtx = thread_block_switch(td); 1813170293Sjeff tdq_load_rem(tdq, ts); 1814171482Sjeff } 1815171482Sjeff /* 1816171482Sjeff * We enter here with the thread blocked and assigned to the 1817171482Sjeff * appropriate cpu run-queue or sleep-queue and with the current 1818171482Sjeff * thread-queue locked. 1819171482Sjeff */ 1820171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1821171482Sjeff /* 1822171505Sjeff * If KSE assigned a new thread just add it here and let choosethread 1823171505Sjeff * select the best one. 1824171482Sjeff */ 1825171505Sjeff if (newtd != NULL) 1826171505Sjeff sched_switchin(tdq, newtd); 1827171482Sjeff newtd = choosethread(); 1828171482Sjeff /* 1829171482Sjeff * Call the MD code to switch contexts if necessary. 1830171482Sjeff */ 1831145256Sjkoshy if (td != newtd) { 1832145256Sjkoshy#ifdef HWPMC_HOOKS 1833145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1834145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1835145256Sjkoshy#endif 1836174629Sjeff lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1837172411Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1838171482Sjeff cpu_switch(td, newtd, mtx); 1839171482Sjeff /* 1840171482Sjeff * We may return from cpu_switch on a different cpu. However, 1841171482Sjeff * we always return with td_lock pointing to the current cpu's 1842171482Sjeff * run queue lock. 1843171482Sjeff */ 1844171482Sjeff cpuid = PCPU_GET(cpuid); 1845171482Sjeff tdq = TDQ_CPU(cpuid); 1846174629Sjeff lock_profile_obtain_lock_success( 1847174629Sjeff &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1848145256Sjkoshy#ifdef HWPMC_HOOKS 1849145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1850145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1851145256Sjkoshy#endif 1852171482Sjeff } else 1853171482Sjeff thread_unblock_switch(td, mtx); 1854171482Sjeff /* 1855171482Sjeff * Assert that all went well and return. 1856171482Sjeff */ 1857171482Sjeff#ifdef SMP 1858171482Sjeff /* We should always get here with the lowest priority td possible */ 1859171482Sjeff tdq->tdq_lowpri = td->td_priority; 1860171482Sjeff#endif 1861171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1862171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1863171482Sjeff td->td_oncpu = cpuid; 1864109864Sjeff} 1865109864Sjeff 1866171482Sjeff/* 1867171482Sjeff * Adjust thread priorities as a result of a nice request. 1868171482Sjeff */ 1869109864Sjeffvoid 1870130551Sjuliansched_nice(struct proc *p, int nice) 1871109864Sjeff{ 1872109864Sjeff struct thread *td; 1873109864Sjeff 1874130551Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 1875170293Sjeff PROC_SLOCK_ASSERT(p, MA_OWNED); 1876165762Sjeff 1877130551Sjulian p->p_nice = nice; 1878163709Sjb FOREACH_THREAD_IN_PROC(p, td) { 1879170293Sjeff thread_lock(td); 1880163709Sjb sched_priority(td); 1881165762Sjeff sched_prio(td, td->td_base_user_pri); 1882170293Sjeff thread_unlock(td); 1883130551Sjulian } 1884109864Sjeff} 1885109864Sjeff 1886171482Sjeff/* 1887171482Sjeff * Record the sleep time for the interactivity scorer. 1888171482Sjeff */ 1889109864Sjeffvoid 1890126326Sjhbsched_sleep(struct thread *td) 1891109864Sjeff{ 1892165762Sjeff 1893170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1894109864Sjeff 1895172264Sjeff td->td_slptick = ticks; 1896109864Sjeff} 1897109864Sjeff 1898171482Sjeff/* 1899171482Sjeff * Schedule a thread to resume execution and record how long it voluntarily 1900171482Sjeff * slept. We also update the pctcpu, interactivity, and priority. 1901171482Sjeff */ 1902109864Sjeffvoid 1903109864Sjeffsched_wakeup(struct thread *td) 1904109864Sjeff{ 1905166229Sjeff struct td_sched *ts; 1906171482Sjeff int slptick; 1907165762Sjeff 1908170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1909166229Sjeff ts = td->td_sched; 1910109864Sjeff /* 1911165762Sjeff * If we slept for more than a tick update our interactivity and 1912165762Sjeff * priority. 1913109864Sjeff */ 1914172264Sjeff slptick = td->td_slptick; 1915172264Sjeff td->td_slptick = 0; 1916171482Sjeff if (slptick && slptick != ticks) { 1917166208Sjeff u_int hzticks; 1918109864Sjeff 1919171482Sjeff hzticks = (ticks - slptick) << SCHED_TICK_SHIFT; 1920171482Sjeff ts->ts_slptime += hzticks; 1921165819Sjeff sched_interact_update(td); 1922166229Sjeff sched_pctcpu_update(ts); 1923163709Sjb sched_priority(td); 1924109864Sjeff } 1925166229Sjeff /* Reset the slice value after we sleep. */ 1926166229Sjeff ts->ts_slice = sched_slice; 1927166190Sjeff sched_add(td, SRQ_BORING); 1928109864Sjeff} 1929109864Sjeff 1930109864Sjeff/* 1931109864Sjeff * Penalize the parent for creating a new child and initialize the child's 1932109864Sjeff * priority. 1933109864Sjeff */ 1934109864Sjeffvoid 1935163709Sjbsched_fork(struct thread *td, struct thread *child) 1936109864Sjeff{ 1937170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1938164936Sjulian sched_fork_thread(td, child); 1939165762Sjeff /* 1940165762Sjeff * Penalize the parent and child for forking. 1941165762Sjeff */ 1942165762Sjeff sched_interact_fork(child); 1943165762Sjeff sched_priority(child); 1944171482Sjeff td->td_sched->ts_runtime += tickincr; 1945165762Sjeff sched_interact_update(td); 1946165762Sjeff sched_priority(td); 1947164936Sjulian} 1948109864Sjeff 1949171482Sjeff/* 1950171482Sjeff * Fork a new thread, may be within the same process. 1951171482Sjeff */ 1952164936Sjulianvoid 1953164936Sjuliansched_fork_thread(struct thread *td, struct thread *child) 1954164936Sjulian{ 1955164936Sjulian struct td_sched *ts; 1956164936Sjulian struct td_sched *ts2; 1957164936Sjulian 1958165762Sjeff /* 1959165762Sjeff * Initialize child. 1960165762Sjeff */ 1961170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1962163709Sjb sched_newthread(child); 1963171482Sjeff child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1964164936Sjulian ts = td->td_sched; 1965164936Sjulian ts2 = child->td_sched; 1966164936Sjulian ts2->ts_cpu = ts->ts_cpu; 1967164936Sjulian ts2->ts_runq = NULL; 1968165762Sjeff /* 1969165762Sjeff * Grab our parents cpu estimation information and priority. 1970165762Sjeff */ 1971164936Sjulian ts2->ts_ticks = ts->ts_ticks; 1972164936Sjulian ts2->ts_ltick = ts->ts_ltick; 1973164936Sjulian ts2->ts_ftick = ts->ts_ftick; 1974165762Sjeff child->td_user_pri = td->td_user_pri; 1975165762Sjeff child->td_base_user_pri = td->td_base_user_pri; 1976165762Sjeff /* 1977165762Sjeff * And update interactivity score. 1978165762Sjeff */ 1979171482Sjeff ts2->ts_slptime = ts->ts_slptime; 1980171482Sjeff ts2->ts_runtime = ts->ts_runtime; 1981165762Sjeff ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 1982113357Sjeff} 1983113357Sjeff 1984171482Sjeff/* 1985171482Sjeff * Adjust the priority class of a thread. 1986171482Sjeff */ 1987113357Sjeffvoid 1988163709Sjbsched_class(struct thread *td, int class) 1989113357Sjeff{ 1990113357Sjeff 1991170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1992163709Sjb if (td->td_pri_class == class) 1993113357Sjeff return; 1994113357Sjeff 1995121896Sjeff#ifdef SMP 1996165827Sjeff /* 1997165827Sjeff * On SMP if we're on the RUNQ we must adjust the transferable 1998165827Sjeff * count because could be changing to or from an interrupt 1999165827Sjeff * class. 2000165827Sjeff */ 2001166190Sjeff if (TD_ON_RUNQ(td)) { 2002165827Sjeff struct tdq *tdq; 2003165827Sjeff 2004165827Sjeff tdq = TDQ_CPU(td->td_sched->ts_cpu); 2005165827Sjeff if (THREAD_CAN_MIGRATE(td)) { 2006165827Sjeff tdq->tdq_transferable--; 2007165827Sjeff tdq->tdq_group->tdg_transferable--; 2008122744Sjeff } 2009165827Sjeff td->td_pri_class = class; 2010165827Sjeff if (THREAD_CAN_MIGRATE(td)) { 2011165827Sjeff tdq->tdq_transferable++; 2012165827Sjeff tdq->tdq_group->tdg_transferable++; 2013165827Sjeff } 2014165827Sjeff } 2015164936Sjulian#endif 2016163709Sjb td->td_pri_class = class; 2017109864Sjeff} 2018109864Sjeff 2019109864Sjeff/* 2020109864Sjeff * Return some of the child's priority and interactivity to the parent. 2021109864Sjeff */ 2022109864Sjeffvoid 2023164939Sjuliansched_exit(struct proc *p, struct thread *child) 2024109864Sjeff{ 2025165762Sjeff struct thread *td; 2026164939Sjulian 2027163709Sjb CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d", 2028173600Sjulian child, child->td_name, child->td_priority); 2029113372Sjeff 2030170293Sjeff PROC_SLOCK_ASSERT(p, MA_OWNED); 2031165762Sjeff td = FIRST_THREAD_IN_PROC(p); 2032165762Sjeff sched_exit_thread(td, child); 2033113372Sjeff} 2034113372Sjeff 2035171482Sjeff/* 2036171482Sjeff * Penalize another thread for the time spent on this one. This helps to 2037171482Sjeff * worsen the priority and interactivity of processes which schedule batch 2038171482Sjeff * jobs such as make. This has little effect on the make process itself but 2039171482Sjeff * causes new processes spawned by it to receive worse scores immediately. 2040171482Sjeff */ 2041113372Sjeffvoid 2042164939Sjuliansched_exit_thread(struct thread *td, struct thread *child) 2043164936Sjulian{ 2044165762Sjeff 2045164939Sjulian CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d", 2046173600Sjulian child, child->td_name, child->td_priority); 2047164939Sjulian 2048165762Sjeff#ifdef KSE 2049165762Sjeff /* 2050165762Sjeff * KSE forks and exits so often that this penalty causes short-lived 2051165762Sjeff * threads to always be non-interactive. This causes mozilla to 2052165762Sjeff * crawl under load. 2053165762Sjeff */ 2054165762Sjeff if ((td->td_pflags & TDP_SA) && td->td_proc == child->td_proc) 2055165762Sjeff return; 2056165762Sjeff#endif 2057165762Sjeff /* 2058165762Sjeff * Give the child's runtime to the parent without returning the 2059165762Sjeff * sleep time as a penalty to the parent. This causes shells that 2060165762Sjeff * launch expensive things to mark their children as expensive. 2061165762Sjeff */ 2062170293Sjeff thread_lock(td); 2063171482Sjeff td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2064164939Sjulian sched_interact_update(td); 2065165762Sjeff sched_priority(td); 2066170293Sjeff thread_unlock(td); 2067164936Sjulian} 2068164936Sjulian 2069171482Sjeff/* 2070171482Sjeff * Fix priorities on return to user-space. Priorities may be elevated due 2071171482Sjeff * to static priorities in msleep() or similar. 2072171482Sjeff */ 2073164936Sjulianvoid 2074164936Sjuliansched_userret(struct thread *td) 2075164936Sjulian{ 2076164936Sjulian /* 2077164936Sjulian * XXX we cheat slightly on the locking here to avoid locking in 2078164936Sjulian * the usual case. Setting td_priority here is essentially an 2079164936Sjulian * incomplete workaround for not setting it properly elsewhere. 2080164936Sjulian * Now that some interrupt handlers are threads, not setting it 2081164936Sjulian * properly elsewhere can clobber it in the window between setting 2082164936Sjulian * it here and returning to user mode, so don't waste time setting 2083164936Sjulian * it perfectly here. 2084164936Sjulian */ 2085164936Sjulian KASSERT((td->td_flags & TDF_BORROWING) == 0, 2086164936Sjulian ("thread with borrowed priority returning to userland")); 2087164936Sjulian if (td->td_priority != td->td_user_pri) { 2088170293Sjeff thread_lock(td); 2089164936Sjulian td->td_priority = td->td_user_pri; 2090164936Sjulian td->td_base_pri = td->td_user_pri; 2091170293Sjeff thread_unlock(td); 2092164936Sjulian } 2093164936Sjulian} 2094164936Sjulian 2095171482Sjeff/* 2096171482Sjeff * Handle a stathz tick. This is really only relevant for timeshare 2097171482Sjeff * threads. 2098171482Sjeff */ 2099164936Sjulianvoid 2100121127Sjeffsched_clock(struct thread *td) 2101109864Sjeff{ 2102164936Sjulian struct tdq *tdq; 2103164936Sjulian struct td_sched *ts; 2104109864Sjeff 2105171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2106164936Sjulian tdq = TDQ_SELF(); 2107172409Sjeff#ifdef SMP 2108133427Sjeff /* 2109172409Sjeff * We run the long term load balancer infrequently on the first cpu. 2110172409Sjeff */ 2111172409Sjeff if (balance_tdq == tdq) { 2112172409Sjeff if (balance_ticks && --balance_ticks == 0) 2113172409Sjeff sched_balance(); 2114172409Sjeff if (balance_group_ticks && --balance_group_ticks == 0) 2115172409Sjeff sched_balance_groups(); 2116172409Sjeff } 2117172409Sjeff#endif 2118172409Sjeff /* 2119165766Sjeff * Advance the insert index once for each tick to ensure that all 2120165766Sjeff * threads get a chance to run. 2121133427Sjeff */ 2122165766Sjeff if (tdq->tdq_idx == tdq->tdq_ridx) { 2123165766Sjeff tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2124165766Sjeff if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2125165766Sjeff tdq->tdq_ridx = tdq->tdq_idx; 2126165766Sjeff } 2127165766Sjeff ts = td->td_sched; 2128175104Sjeff if (td->td_pri_class & PRI_FIFO_BIT) 2129113357Sjeff return; 2130175104Sjeff if (td->td_pri_class == PRI_TIMESHARE) { 2131175104Sjeff /* 2132175104Sjeff * We used a tick; charge it to the thread so 2133175104Sjeff * that we can compute our interactivity. 2134175104Sjeff */ 2135175104Sjeff td->td_sched->ts_runtime += tickincr; 2136175104Sjeff sched_interact_update(td); 2137175104Sjeff } 2138113357Sjeff /* 2139109864Sjeff * We used up one time slice. 2140109864Sjeff */ 2141164936Sjulian if (--ts->ts_slice > 0) 2142113357Sjeff return; 2143109864Sjeff /* 2144113357Sjeff * We're out of time, recompute priorities and requeue. 2145109864Sjeff */ 2146165796Sjeff sched_priority(td); 2147113357Sjeff td->td_flags |= TDF_NEEDRESCHED; 2148109864Sjeff} 2149109864Sjeff 2150171482Sjeff/* 2151171482Sjeff * Called once per hz tick. Used for cpu utilization information. This 2152171482Sjeff * is easier than trying to scale based on stathz. 2153171482Sjeff */ 2154171482Sjeffvoid 2155171482Sjeffsched_tick(void) 2156171482Sjeff{ 2157171482Sjeff struct td_sched *ts; 2158171482Sjeff 2159171482Sjeff ts = curthread->td_sched; 2160171482Sjeff /* Adjust ticks for pctcpu */ 2161171482Sjeff ts->ts_ticks += 1 << SCHED_TICK_SHIFT; 2162171482Sjeff ts->ts_ltick = ticks; 2163171482Sjeff /* 2164171482Sjeff * Update if we've exceeded our desired tick threshhold by over one 2165171482Sjeff * second. 2166171482Sjeff */ 2167171482Sjeff if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2168171482Sjeff sched_pctcpu_update(ts); 2169171482Sjeff} 2170171482Sjeff 2171171482Sjeff/* 2172171482Sjeff * Return whether the current CPU has runnable tasks. Used for in-kernel 2173171482Sjeff * cooperative idle threads. 2174171482Sjeff */ 2175109864Sjeffint 2176109864Sjeffsched_runnable(void) 2177109864Sjeff{ 2178164936Sjulian struct tdq *tdq; 2179115998Sjeff int load; 2180109864Sjeff 2181115998Sjeff load = 1; 2182115998Sjeff 2183164936Sjulian tdq = TDQ_SELF(); 2184121605Sjeff if ((curthread->td_flags & TDF_IDLETD) != 0) { 2185165620Sjeff if (tdq->tdq_load > 0) 2186121605Sjeff goto out; 2187121605Sjeff } else 2188165620Sjeff if (tdq->tdq_load - 1 > 0) 2189121605Sjeff goto out; 2190115998Sjeff load = 0; 2191115998Sjeffout: 2192115998Sjeff return (load); 2193109864Sjeff} 2194109864Sjeff 2195171482Sjeff/* 2196171482Sjeff * Choose the highest priority thread to run. The thread is removed from 2197171482Sjeff * the run-queue while running however the load remains. For SMP we set 2198171482Sjeff * the tdq in the global idle bitmask if it idles here. 2199171482Sjeff */ 2200166190Sjeffstruct thread * 2201109970Sjeffsched_choose(void) 2202109970Sjeff{ 2203171482Sjeff#ifdef SMP 2204171482Sjeff struct tdq_group *tdg; 2205171482Sjeff#endif 2206171482Sjeff struct td_sched *ts; 2207175587Sjeff struct thread *td; 2208164936Sjulian struct tdq *tdq; 2209109970Sjeff 2210164936Sjulian tdq = TDQ_SELF(); 2211171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2212164936Sjulian ts = tdq_choose(tdq); 2213164936Sjulian if (ts) { 2214164936Sjulian tdq_runq_rem(tdq, ts); 2215166190Sjeff return (ts->ts_thread); 2216109864Sjeff } 2217175587Sjeff td = PCPU_GET(idlethread); 2218109970Sjeff#ifdef SMP 2219171482Sjeff /* 2220171482Sjeff * We only set the idled bit when all of the cpus in the group are 2221171482Sjeff * idle. Otherwise we could get into a situation where a thread bounces 2222171482Sjeff * back and forth between two idle cores on seperate physical CPUs. 2223171482Sjeff */ 2224171482Sjeff tdg = tdq->tdq_group; 2225171482Sjeff tdg->tdg_idlemask |= PCPU_GET(cpumask); 2226171482Sjeff if (tdg->tdg_idlemask == tdg->tdg_cpumask) 2227171482Sjeff atomic_set_int(&tdq_idle, tdg->tdg_mask); 2228175587Sjeff tdq->tdq_lowpri = td->td_priority; 2229109970Sjeff#endif 2230175587Sjeff return (td); 2231109864Sjeff} 2232109864Sjeff 2233171482Sjeff/* 2234171482Sjeff * Set owepreempt if necessary. Preemption never happens directly in ULE, 2235171482Sjeff * we always request it once we exit a critical section. 2236171482Sjeff */ 2237171482Sjeffstatic inline void 2238171482Sjeffsched_setpreempt(struct thread *td) 2239166190Sjeff{ 2240166190Sjeff struct thread *ctd; 2241166190Sjeff int cpri; 2242166190Sjeff int pri; 2243166190Sjeff 2244166190Sjeff ctd = curthread; 2245166190Sjeff pri = td->td_priority; 2246166190Sjeff cpri = ctd->td_priority; 2247171482Sjeff if (td->td_priority < ctd->td_priority) 2248171482Sjeff curthread->td_flags |= TDF_NEEDRESCHED; 2249166190Sjeff if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2250171482Sjeff return; 2251166190Sjeff /* 2252166190Sjeff * Always preempt IDLE threads. Otherwise only if the preempting 2253166190Sjeff * thread is an ithread. 2254166190Sjeff */ 2255171482Sjeff if (pri > preempt_thresh && cpri < PRI_MIN_IDLE) 2256171482Sjeff return; 2257171482Sjeff ctd->td_owepreempt = 1; 2258171482Sjeff return; 2259166190Sjeff} 2260166190Sjeff 2261171482Sjeff/* 2262171482Sjeff * Add a thread to a thread queue. Initializes priority, slice, runq, and 2263171482Sjeff * add it to the appropriate queue. This is the internal function called 2264171482Sjeff * when the tdq is predetermined. 2265171482Sjeff */ 2266109864Sjeffvoid 2267171482Sjefftdq_add(struct tdq *tdq, struct thread *td, int flags) 2268109864Sjeff{ 2269164936Sjulian struct td_sched *ts; 2270121790Sjeff int class; 2271166108Sjeff#ifdef SMP 2272166108Sjeff int cpumask; 2273166108Sjeff#endif 2274109864Sjeff 2275171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2276166190Sjeff KASSERT((td->td_inhibitors == 0), 2277166190Sjeff ("sched_add: trying to run inhibited thread")); 2278166190Sjeff KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2279166190Sjeff ("sched_add: bad thread state")); 2280172207Sjeff KASSERT(td->td_flags & TDF_INMEM, 2281172207Sjeff ("sched_add: thread swapped out")); 2282171482Sjeff 2283171482Sjeff ts = td->td_sched; 2284171482Sjeff class = PRI_BASE(td->td_pri_class); 2285166190Sjeff TD_SET_RUNQ(td); 2286166190Sjeff if (ts->ts_slice == 0) 2287166190Sjeff ts->ts_slice = sched_slice; 2288133427Sjeff /* 2289171482Sjeff * Pick the run queue based on priority. 2290133427Sjeff */ 2291171482Sjeff if (td->td_priority <= PRI_MAX_REALTIME) 2292171482Sjeff ts->ts_runq = &tdq->tdq_realtime; 2293171482Sjeff else if (td->td_priority <= PRI_MAX_TIMESHARE) 2294171482Sjeff ts->ts_runq = &tdq->tdq_timeshare; 2295171482Sjeff else 2296171482Sjeff ts->ts_runq = &tdq->tdq_idle; 2297171482Sjeff#ifdef SMP 2298166108Sjeff cpumask = 1 << ts->ts_cpu; 2299121790Sjeff /* 2300123685Sjeff * If we had been idle, clear our bit in the group and potentially 2301166108Sjeff * the global bitmap. 2302121790Sjeff */ 2303165762Sjeff if ((class != PRI_IDLE && class != PRI_ITHD) && 2304166108Sjeff (tdq->tdq_group->tdg_idlemask & cpumask) != 0) { 2305121790Sjeff /* 2306123433Sjeff * Check to see if our group is unidling, and if so, remove it 2307123433Sjeff * from the global idle mask. 2308121790Sjeff */ 2309165620Sjeff if (tdq->tdq_group->tdg_idlemask == 2310165620Sjeff tdq->tdq_group->tdg_cpumask) 2311165620Sjeff atomic_clear_int(&tdq_idle, tdq->tdq_group->tdg_mask); 2312123433Sjeff /* 2313123433Sjeff * Now remove ourselves from the group specific idle mask. 2314123433Sjeff */ 2315166108Sjeff tdq->tdq_group->tdg_idlemask &= ~cpumask; 2316166108Sjeff } 2317171482Sjeff if (td->td_priority < tdq->tdq_lowpri) 2318171482Sjeff tdq->tdq_lowpri = td->td_priority; 2319121790Sjeff#endif 2320171482Sjeff tdq_runq_add(tdq, ts, flags); 2321171482Sjeff tdq_load_add(tdq, ts); 2322171482Sjeff} 2323171482Sjeff 2324171482Sjeff/* 2325171482Sjeff * Select the target thread queue and add a thread to it. Request 2326171482Sjeff * preemption or IPI a remote processor if required. 2327171482Sjeff */ 2328171482Sjeffvoid 2329171482Sjeffsched_add(struct thread *td, int flags) 2330171482Sjeff{ 2331171482Sjeff struct td_sched *ts; 2332171482Sjeff struct tdq *tdq; 2333171482Sjeff#ifdef SMP 2334171482Sjeff int cpuid; 2335171482Sjeff int cpu; 2336171482Sjeff#endif 2337171482Sjeff CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 2338173600Sjulian td, td->td_name, td->td_priority, curthread, 2339173600Sjulian curthread->td_name); 2340171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2341171482Sjeff ts = td->td_sched; 2342166108Sjeff /* 2343171482Sjeff * Recalculate the priority before we select the target cpu or 2344171482Sjeff * run-queue. 2345166108Sjeff */ 2346171482Sjeff if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2347171482Sjeff sched_priority(td); 2348171482Sjeff#ifdef SMP 2349171482Sjeff cpuid = PCPU_GET(cpuid); 2350171482Sjeff /* 2351171482Sjeff * Pick the destination cpu and if it isn't ours transfer to the 2352171482Sjeff * target cpu. 2353171482Sjeff */ 2354175348Sjeff if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_MIGRATE(td) && 2355175348Sjeff curthread->td_intr_nesting_level) 2356175348Sjeff ts->ts_cpu = cpuid; 2357175348Sjeff if (!THREAD_CAN_MIGRATE(td)) 2358171482Sjeff cpu = ts->ts_cpu; 2359166108Sjeff else 2360171482Sjeff cpu = sched_pickcpu(ts, flags); 2361171482Sjeff tdq = sched_setcpu(ts, cpu, flags); 2362171482Sjeff tdq_add(tdq, td, flags); 2363171482Sjeff if (cpu != cpuid) { 2364166108Sjeff tdq_notify(ts); 2365166108Sjeff return; 2366166108Sjeff } 2367171482Sjeff#else 2368171482Sjeff tdq = TDQ_SELF(); 2369171482Sjeff TDQ_LOCK(tdq); 2370171482Sjeff /* 2371171482Sjeff * Now that the thread is moving to the run-queue, set the lock 2372171482Sjeff * to the scheduler's lock. 2373171482Sjeff */ 2374171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2375171482Sjeff tdq_add(tdq, td, flags); 2376166108Sjeff#endif 2377171482Sjeff if (!(flags & SRQ_YIELDING)) 2378171482Sjeff sched_setpreempt(td); 2379109864Sjeff} 2380109864Sjeff 2381171482Sjeff/* 2382171482Sjeff * Remove a thread from a run-queue without running it. This is used 2383171482Sjeff * when we're stealing a thread from a remote queue. Otherwise all threads 2384171482Sjeff * exit by calling sched_exit_thread() and sched_throw() themselves. 2385171482Sjeff */ 2386109864Sjeffvoid 2387121127Sjeffsched_rem(struct thread *td) 2388109864Sjeff{ 2389164936Sjulian struct tdq *tdq; 2390164936Sjulian struct td_sched *ts; 2391113357Sjeff 2392139316Sjeff CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 2393173600Sjulian td, td->td_name, td->td_priority, curthread, 2394173600Sjulian curthread->td_name); 2395164936Sjulian ts = td->td_sched; 2396171482Sjeff tdq = TDQ_CPU(ts->ts_cpu); 2397171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2398171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2399166190Sjeff KASSERT(TD_ON_RUNQ(td), 2400164936Sjulian ("sched_rem: thread not on run queue")); 2401164936Sjulian tdq_runq_rem(tdq, ts); 2402164936Sjulian tdq_load_rem(tdq, ts); 2403166190Sjeff TD_SET_CAN_RUN(td); 2404109864Sjeff} 2405109864Sjeff 2406171482Sjeff/* 2407171482Sjeff * Fetch cpu utilization information. Updates on demand. 2408171482Sjeff */ 2409109864Sjefffixpt_t 2410121127Sjeffsched_pctcpu(struct thread *td) 2411109864Sjeff{ 2412109864Sjeff fixpt_t pctcpu; 2413164936Sjulian struct td_sched *ts; 2414109864Sjeff 2415109864Sjeff pctcpu = 0; 2416164936Sjulian ts = td->td_sched; 2417164936Sjulian if (ts == NULL) 2418121290Sjeff return (0); 2419109864Sjeff 2420170293Sjeff thread_lock(td); 2421164936Sjulian if (ts->ts_ticks) { 2422109864Sjeff int rtick; 2423109864Sjeff 2424165796Sjeff sched_pctcpu_update(ts); 2425109864Sjeff /* How many rtick per second ? */ 2426165762Sjeff rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2427165762Sjeff pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2428109864Sjeff } 2429170293Sjeff thread_unlock(td); 2430109864Sjeff 2431109864Sjeff return (pctcpu); 2432109864Sjeff} 2433109864Sjeff 2434176729Sjeffvoid 2435176729Sjeffsched_affinity(struct thread *td) 2436176729Sjeff{ 2437176729Sjeff} 2438176729Sjeff 2439171482Sjeff/* 2440171482Sjeff * Bind a thread to a target cpu. 2441171482Sjeff */ 2442122038Sjeffvoid 2443122038Sjeffsched_bind(struct thread *td, int cpu) 2444122038Sjeff{ 2445164936Sjulian struct td_sched *ts; 2446122038Sjeff 2447171713Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2448164936Sjulian ts = td->td_sched; 2449166137Sjeff if (ts->ts_flags & TSF_BOUND) 2450166152Sjeff sched_unbind(td); 2451164936Sjulian ts->ts_flags |= TSF_BOUND; 2452123433Sjeff#ifdef SMP 2453166137Sjeff sched_pin(); 2454123433Sjeff if (PCPU_GET(cpuid) == cpu) 2455122038Sjeff return; 2456166137Sjeff ts->ts_cpu = cpu; 2457122038Sjeff /* When we return from mi_switch we'll be on the correct cpu. */ 2458131527Sphk mi_switch(SW_VOL, NULL); 2459122038Sjeff#endif 2460122038Sjeff} 2461122038Sjeff 2462171482Sjeff/* 2463171482Sjeff * Release a bound thread. 2464171482Sjeff */ 2465122038Sjeffvoid 2466122038Sjeffsched_unbind(struct thread *td) 2467122038Sjeff{ 2468165762Sjeff struct td_sched *ts; 2469165762Sjeff 2470170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2471165762Sjeff ts = td->td_sched; 2472166137Sjeff if ((ts->ts_flags & TSF_BOUND) == 0) 2473166137Sjeff return; 2474165762Sjeff ts->ts_flags &= ~TSF_BOUND; 2475165762Sjeff#ifdef SMP 2476165762Sjeff sched_unpin(); 2477165762Sjeff#endif 2478122038Sjeff} 2479122038Sjeff 2480109864Sjeffint 2481145256Sjkoshysched_is_bound(struct thread *td) 2482145256Sjkoshy{ 2483170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2484164936Sjulian return (td->td_sched->ts_flags & TSF_BOUND); 2485145256Sjkoshy} 2486145256Sjkoshy 2487171482Sjeff/* 2488171482Sjeff * Basic yield call. 2489171482Sjeff */ 2490159630Sdavidxuvoid 2491159630Sdavidxusched_relinquish(struct thread *td) 2492159630Sdavidxu{ 2493170293Sjeff thread_lock(td); 2494170293Sjeff SCHED_STAT_INC(switch_relinquish); 2495159630Sdavidxu mi_switch(SW_VOL, NULL); 2496170293Sjeff thread_unlock(td); 2497159630Sdavidxu} 2498159630Sdavidxu 2499171482Sjeff/* 2500171482Sjeff * Return the total system load. 2501171482Sjeff */ 2502145256Sjkoshyint 2503125289Sjeffsched_load(void) 2504125289Sjeff{ 2505125289Sjeff#ifdef SMP 2506125289Sjeff int total; 2507125289Sjeff int i; 2508125289Sjeff 2509125289Sjeff total = 0; 2510165620Sjeff for (i = 0; i <= tdg_maxid; i++) 2511165620Sjeff total += TDQ_GROUP(i)->tdg_load; 2512125289Sjeff return (total); 2513125289Sjeff#else 2514165620Sjeff return (TDQ_SELF()->tdq_sysload); 2515125289Sjeff#endif 2516125289Sjeff} 2517125289Sjeff 2518125289Sjeffint 2519109864Sjeffsched_sizeof_proc(void) 2520109864Sjeff{ 2521109864Sjeff return (sizeof(struct proc)); 2522109864Sjeff} 2523109864Sjeff 2524109864Sjeffint 2525109864Sjeffsched_sizeof_thread(void) 2526109864Sjeff{ 2527109864Sjeff return (sizeof(struct thread) + sizeof(struct td_sched)); 2528109864Sjeff} 2529159570Sdavidxu 2530166190Sjeff/* 2531166190Sjeff * The actual idle process. 2532166190Sjeff */ 2533166190Sjeffvoid 2534166190Sjeffsched_idletd(void *dummy) 2535166190Sjeff{ 2536166190Sjeff struct thread *td; 2537171482Sjeff struct tdq *tdq; 2538166190Sjeff 2539166190Sjeff td = curthread; 2540171482Sjeff tdq = TDQ_SELF(); 2541166190Sjeff mtx_assert(&Giant, MA_NOTOWNED); 2542171482Sjeff /* ULE relies on preemption for idle interruption. */ 2543171482Sjeff for (;;) { 2544171482Sjeff#ifdef SMP 2545171482Sjeff if (tdq_idled(tdq)) 2546171482Sjeff cpu_idle(); 2547171482Sjeff#else 2548166190Sjeff cpu_idle(); 2549171482Sjeff#endif 2550171482Sjeff } 2551166190Sjeff} 2552166190Sjeff 2553170293Sjeff/* 2554170293Sjeff * A CPU is entering for the first time or a thread is exiting. 2555170293Sjeff */ 2556170293Sjeffvoid 2557170293Sjeffsched_throw(struct thread *td) 2558170293Sjeff{ 2559172411Sjeff struct thread *newtd; 2560171482Sjeff struct tdq *tdq; 2561171482Sjeff 2562171482Sjeff tdq = TDQ_SELF(); 2563170293Sjeff if (td == NULL) { 2564171482Sjeff /* Correct spinlock nesting and acquire the correct lock. */ 2565171482Sjeff TDQ_LOCK(tdq); 2566170293Sjeff spinlock_exit(); 2567170293Sjeff } else { 2568171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2569171482Sjeff tdq_load_rem(tdq, td->td_sched); 2570174629Sjeff lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2571170293Sjeff } 2572170293Sjeff KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2573172411Sjeff newtd = choosethread(); 2574172411Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2575170293Sjeff PCPU_SET(switchtime, cpu_ticks()); 2576170293Sjeff PCPU_SET(switchticks, ticks); 2577172411Sjeff cpu_throw(td, newtd); /* doesn't return */ 2578170293Sjeff} 2579170293Sjeff 2580171482Sjeff/* 2581171482Sjeff * This is called from fork_exit(). Just acquire the correct locks and 2582171482Sjeff * let fork do the rest of the work. 2583171482Sjeff */ 2584170293Sjeffvoid 2585170600Sjeffsched_fork_exit(struct thread *td) 2586170293Sjeff{ 2587171482Sjeff struct td_sched *ts; 2588171482Sjeff struct tdq *tdq; 2589171482Sjeff int cpuid; 2590170293Sjeff 2591170293Sjeff /* 2592170293Sjeff * Finish setting up thread glue so that it begins execution in a 2593171482Sjeff * non-nested critical section with the scheduler lock held. 2594170293Sjeff */ 2595171482Sjeff cpuid = PCPU_GET(cpuid); 2596171482Sjeff tdq = TDQ_CPU(cpuid); 2597171482Sjeff ts = td->td_sched; 2598171482Sjeff if (TD_IS_IDLETHREAD(td)) 2599171482Sjeff td->td_lock = TDQ_LOCKPTR(tdq); 2600171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2601171482Sjeff td->td_oncpu = cpuid; 2602172411Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2603174629Sjeff lock_profile_obtain_lock_success( 2604174629Sjeff &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2605170293Sjeff} 2606170293Sjeff 2607171482Sjeffstatic SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, 2608171482Sjeff "Scheduler"); 2609171482SjeffSYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2610165762Sjeff "Scheduler name"); 2611171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2612171482Sjeff "Slice size for timeshare threads"); 2613171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2614171482Sjeff "Interactivity score threshold"); 2615171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2616171482Sjeff 0,"Min priority for preemption, lower priorities have greater precedence"); 2617166108Sjeff#ifdef SMP 2618171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, pick_pri, CTLFLAG_RW, &pick_pri, 0, 2619171482Sjeff "Pick the target cpu based on priority rather than load."); 2620171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2621171482Sjeff "Number of hz ticks to keep thread affinity for"); 2622171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, tryself, CTLFLAG_RW, &tryself, 0, ""); 2623171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2624171482Sjeff "Enables the long-term load balancer"); 2625172409SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2626172409Sjeff &balance_interval, 0, 2627172409Sjeff "Average frequency in stathz ticks to run the long-term balancer"); 2628171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, 2629171482Sjeff "Steals work from another hyper-threaded core on idle"); 2630171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2631171482Sjeff "Attempts to steal work from other cores before idling"); 2632171506SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2633171506Sjeff "Minimum load on remote cpu before we'll steal"); 2634166108Sjeff#endif 2635165762Sjeff 2636172264Sjeff/* ps compat. All cpu percentages from ULE are weighted. */ 2637172293Sjeffstatic int ccpu = 0; 2638165762SjeffSYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2639165762Sjeff 2640165762Sjeff 2641134791Sjulian#define KERN_SWITCH_INCLUDE 1 2642134791Sjulian#include "kern/kern_switch.c" 2643