sched_ule.c revision 171713
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: 33171482Sjeff * ULE is the last three letters in schedule. It owes it's 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 171713 2007-08-03 23:38:46Z 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 74166190Sjeff#ifndef PREEMPTION 75166190Sjeff#error "SCHED_ULE requires options PREEMPTION" 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_slptick; /* Tick when we went to sleep. */ 92171482Sjeff int ts_slice; /* Ticks of slice remaining. */ 93171482Sjeff u_int ts_slptime; /* Number of ticks we vol. slept */ 94171482Sjeff u_int ts_runtime; /* Number of ticks we were running */ 95134791Sjulian /* The following variables are only used for pctcpu calculation */ 96164936Sjulian int ts_ltick; /* Last tick that we were running on */ 97164936Sjulian int ts_ftick; /* First tick that we were running on */ 98164936Sjulian int ts_ticks; /* Tick count */ 99166108Sjeff#ifdef SMP 100166108Sjeff int ts_rltick; /* Real last tick, for affinity. */ 101166108Sjeff#endif 102134791Sjulian}; 103164936Sjulian/* flags kept in ts_flags */ 104166108Sjeff#define TSF_BOUND 0x0001 /* Thread can not migrate. */ 105166108Sjeff#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 106121790Sjeff 107164936Sjulianstatic struct td_sched td_sched0; 108109864Sjeff 109109864Sjeff/* 110165762Sjeff * Cpu percentage computation macros and defines. 111111857Sjeff * 112165762Sjeff * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 113165762Sjeff * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 114165796Sjeff * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 115165762Sjeff * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 116165762Sjeff * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 117165762Sjeff * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 118165762Sjeff */ 119165762Sjeff#define SCHED_TICK_SECS 10 120165762Sjeff#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 121165796Sjeff#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 122165762Sjeff#define SCHED_TICK_SHIFT 10 123165762Sjeff#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 124165830Sjeff#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 125165762Sjeff 126165762Sjeff/* 127165762Sjeff * These macros determine priorities for non-interactive threads. They are 128165762Sjeff * assigned a priority based on their recent cpu utilization as expressed 129165762Sjeff * by the ratio of ticks to the tick total. NHALF priorities at the start 130165762Sjeff * and end of the MIN to MAX timeshare range are only reachable with negative 131165762Sjeff * or positive nice respectively. 132165762Sjeff * 133165762Sjeff * PRI_RANGE: Priority range for utilization dependent priorities. 134116642Sjeff * PRI_NRESV: Number of nice values. 135165762Sjeff * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 136165762Sjeff * PRI_NICE: Determines the part of the priority inherited from nice. 137109864Sjeff */ 138165762Sjeff#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 139121869Sjeff#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 140165762Sjeff#define SCHED_PRI_MIN (PRI_MIN_TIMESHARE + SCHED_PRI_NHALF) 141165762Sjeff#define SCHED_PRI_MAX (PRI_MAX_TIMESHARE - SCHED_PRI_NHALF) 142170787Sjeff#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN) 143165762Sjeff#define SCHED_PRI_TICKS(ts) \ 144165762Sjeff (SCHED_TICK_HZ((ts)) / \ 145165827Sjeff (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 146165762Sjeff#define SCHED_PRI_NICE(nice) (nice) 147109864Sjeff 148109864Sjeff/* 149165762Sjeff * These determine the interactivity of a process. Interactivity differs from 150165762Sjeff * cpu utilization in that it expresses the voluntary time slept vs time ran 151165762Sjeff * while cpu utilization includes all time not running. This more accurately 152165762Sjeff * models the intent of the thread. 153109864Sjeff * 154110645Sjeff * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 155110645Sjeff * before throttling back. 156121868Sjeff * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 157116365Sjeff * INTERACT_MAX: Maximum interactivity value. Smaller is better. 158111857Sjeff * INTERACT_THRESH: Threshhold for placement on the current runq. 159109864Sjeff */ 160165762Sjeff#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 161165762Sjeff#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 162116365Sjeff#define SCHED_INTERACT_MAX (100) 163116365Sjeff#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 164121126Sjeff#define SCHED_INTERACT_THRESH (30) 165111857Sjeff 166109864Sjeff/* 167165762Sjeff * tickincr: Converts a stathz tick into a hz domain scaled by 168165762Sjeff * the shift factor. Without the shift the error rate 169165762Sjeff * due to rounding would be unacceptably high. 170165762Sjeff * realstathz: stathz is sometimes 0 and run off of hz. 171165762Sjeff * sched_slice: Runtime of each thread before rescheduling. 172171482Sjeff * preempt_thresh: Priority threshold for preemption and remote IPIs. 173109864Sjeff */ 174165762Sjeffstatic int sched_interact = SCHED_INTERACT_THRESH; 175165762Sjeffstatic int realstathz; 176165762Sjeffstatic int tickincr; 177165762Sjeffstatic int sched_slice; 178171482Sjeffstatic int preempt_thresh = PRI_MIN_KERN; 179109864Sjeff 180109864Sjeff/* 181171482Sjeff * tdq - per processor runqs and statistics. All fields are protected by the 182171482Sjeff * tdq_lock. The load and lowpri may be accessed without to avoid excess 183171482Sjeff * locking in sched_pickcpu(); 184109864Sjeff */ 185164936Sjulianstruct tdq { 186171713Sjeff struct mtx *tdq_lock; /* Pointer to group lock. */ 187171482Sjeff struct runq tdq_realtime; /* real-time run queue. */ 188171482Sjeff struct runq tdq_timeshare; /* timeshare run queue. */ 189165620Sjeff struct runq tdq_idle; /* Queue of IDLE threads. */ 190171482Sjeff int tdq_load; /* Aggregate load. */ 191166557Sjeff u_char tdq_idx; /* Current insert index. */ 192166557Sjeff u_char tdq_ridx; /* Current removal index. */ 193110267Sjeff#ifdef SMP 194171482Sjeff u_char tdq_lowpri; /* Lowest priority thread. */ 195171482Sjeff int tdq_transferable; /* Transferable thread count. */ 196165620Sjeff LIST_ENTRY(tdq) tdq_siblings; /* Next in tdq group. */ 197165620Sjeff struct tdq_group *tdq_group; /* Our processor group. */ 198125289Sjeff#else 199165620Sjeff int tdq_sysload; /* For loadavg, !ITHD load. */ 200110267Sjeff#endif 201171482Sjeff} __aligned(64); 202109864Sjeff 203166108Sjeff 204123433Sjeff#ifdef SMP 205109864Sjeff/* 206164936Sjulian * tdq groups are groups of processors which can cheaply share threads. When 207123433Sjeff * one processor in the group goes idle it will check the runqs of the other 208123433Sjeff * processors in its group prior to halting and waiting for an interrupt. 209123433Sjeff * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA. 210123433Sjeff * In a numa environment we'd want an idle bitmap per group and a two tiered 211123433Sjeff * load balancer. 212123433Sjeff */ 213164936Sjulianstruct tdq_group { 214171713Sjeff struct mtx tdg_lock; /* Protects all fields below. */ 215171713Sjeff int tdg_cpus; /* Count of CPUs in this tdq group. */ 216171713Sjeff cpumask_t tdg_cpumask; /* Mask of cpus in this group. */ 217171713Sjeff cpumask_t tdg_idlemask; /* Idle cpus in this group. */ 218171713Sjeff cpumask_t tdg_mask; /* Bit mask for first cpu. */ 219171713Sjeff int tdg_load; /* Total load of this group. */ 220165620Sjeff int tdg_transferable; /* Transferable load of this group. */ 221165620Sjeff LIST_HEAD(, tdq) tdg_members; /* Linked list of all members. */ 222171713Sjeff char tdg_name[16]; /* lock name. */ 223171482Sjeff} __aligned(64); 224123433Sjeff 225171482Sjeff#define SCHED_AFFINITY_DEFAULT (max(1, hz / 300)) 226166108Sjeff#define SCHED_AFFINITY(ts) ((ts)->ts_rltick > ticks - affinity) 227166108Sjeff 228123433Sjeff/* 229166108Sjeff * Run-time tunables. 230166108Sjeff */ 231171506Sjeffstatic int rebalance = 1; 232171506Sjeffstatic int balance_secs = 1; 233171506Sjeffstatic int pick_pri = 1; 234166108Sjeffstatic int affinity; 235166108Sjeffstatic int tryself = 1; 236171482Sjeffstatic int steal_htt = 0; 237171506Sjeffstatic int steal_idle = 1; 238171506Sjeffstatic int steal_thresh = 2; 239170293Sjeffstatic int topology = 0; 240166108Sjeff 241166108Sjeff/* 242165620Sjeff * One thread queue per processor. 243109864Sjeff */ 244166108Sjeffstatic volatile cpumask_t tdq_idle; 245165620Sjeffstatic int tdg_maxid; 246164936Sjulianstatic struct tdq tdq_cpu[MAXCPU]; 247164936Sjulianstatic struct tdq_group tdq_groups[MAXCPU]; 248171482Sjeffstatic struct callout balco; 249171482Sjeffstatic struct callout gbalco; 250129982Sjeff 251164936Sjulian#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 252164936Sjulian#define TDQ_CPU(x) (&tdq_cpu[(x)]) 253171713Sjeff#define TDQ_ID(x) ((int)((x) - tdq_cpu)) 254164936Sjulian#define TDQ_GROUP(x) (&tdq_groups[(x)]) 255171713Sjeff#define TDG_ID(x) ((int)((x) - tdq_groups)) 256123433Sjeff#else /* !SMP */ 257164936Sjulianstatic struct tdq tdq_cpu; 258171713Sjeffstatic struct mtx tdq_lock; 259129982Sjeff 260170315Sjeff#define TDQ_ID(x) (0) 261164936Sjulian#define TDQ_SELF() (&tdq_cpu) 262164936Sjulian#define TDQ_CPU(x) (&tdq_cpu) 263110028Sjeff#endif 264109864Sjeff 265171482Sjeff#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 266171482Sjeff#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 267171482Sjeff#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 268171482Sjeff#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 269171713Sjeff#define TDQ_LOCKPTR(t) ((t)->tdq_lock) 270171482Sjeff 271163709Sjbstatic void sched_priority(struct thread *); 272146954Sjeffstatic void sched_thread_priority(struct thread *, u_char); 273163709Sjbstatic int sched_interact_score(struct thread *); 274163709Sjbstatic void sched_interact_update(struct thread *); 275163709Sjbstatic void sched_interact_fork(struct thread *); 276164936Sjulianstatic void sched_pctcpu_update(struct td_sched *); 277109864Sjeff 278110267Sjeff/* Operations on per processor queues */ 279164936Sjulianstatic struct td_sched * tdq_choose(struct tdq *); 280164936Sjulianstatic void tdq_setup(struct tdq *); 281164936Sjulianstatic void tdq_load_add(struct tdq *, struct td_sched *); 282164936Sjulianstatic void tdq_load_rem(struct tdq *, struct td_sched *); 283164936Sjulianstatic __inline void tdq_runq_add(struct tdq *, struct td_sched *, int); 284164936Sjulianstatic __inline void tdq_runq_rem(struct tdq *, struct td_sched *); 285164936Sjulianvoid tdq_print(int cpu); 286165762Sjeffstatic void runq_print(struct runq *rq); 287171482Sjeffstatic void tdq_add(struct tdq *, struct thread *, int); 288110267Sjeff#ifdef SMP 289171482Sjeffstatic void tdq_move(struct tdq *, struct tdq *); 290171482Sjeffstatic int tdq_idled(struct tdq *); 291171482Sjeffstatic void tdq_notify(struct td_sched *); 292171482Sjeffstatic struct td_sched *tdq_steal(struct tdq *, int); 293164936Sjulianstatic struct td_sched *runq_steal(struct runq *); 294171482Sjeffstatic int sched_pickcpu(struct td_sched *, int); 295171482Sjeffstatic void sched_balance(void *); 296171482Sjeffstatic void sched_balance_groups(void *); 297164936Sjulianstatic void sched_balance_group(struct tdq_group *); 298164936Sjulianstatic void sched_balance_pair(struct tdq *, struct tdq *); 299171482Sjeffstatic inline struct tdq *sched_setcpu(struct td_sched *, int, int); 300171713Sjeffstatic inline struct tdq *sched_switchcpu(struct td_sched *, int, int); 301171482Sjeffstatic inline struct mtx *thread_block_switch(struct thread *); 302171482Sjeffstatic inline void thread_unblock_switch(struct thread *, struct mtx *); 303171713Sjeffstatic struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 304165827Sjeff 305166108Sjeff#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 306121790Sjeff#endif 307110028Sjeff 308165762Sjeffstatic void sched_setup(void *dummy); 309165762SjeffSYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL) 310165762Sjeff 311165762Sjeffstatic void sched_initticks(void *dummy); 312165762SjeffSYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, NULL) 313165762Sjeff 314171482Sjeff/* 315171482Sjeff * Print the threads waiting on a run-queue. 316171482Sjeff */ 317165762Sjeffstatic void 318165762Sjeffrunq_print(struct runq *rq) 319165762Sjeff{ 320165762Sjeff struct rqhead *rqh; 321165762Sjeff struct td_sched *ts; 322165762Sjeff int pri; 323165762Sjeff int j; 324165762Sjeff int i; 325165762Sjeff 326165762Sjeff for (i = 0; i < RQB_LEN; i++) { 327165762Sjeff printf("\t\trunq bits %d 0x%zx\n", 328165762Sjeff i, rq->rq_status.rqb_bits[i]); 329165762Sjeff for (j = 0; j < RQB_BPW; j++) 330165762Sjeff if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 331165762Sjeff pri = j + (i << RQB_L2BPW); 332165762Sjeff rqh = &rq->rq_queues[pri]; 333165762Sjeff TAILQ_FOREACH(ts, rqh, ts_procq) { 334165762Sjeff printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 335165762Sjeff ts->ts_thread, ts->ts_thread->td_proc->p_comm, ts->ts_thread->td_priority, ts->ts_rqindex, pri); 336165762Sjeff } 337165762Sjeff } 338165762Sjeff } 339165762Sjeff} 340165762Sjeff 341171482Sjeff/* 342171482Sjeff * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 343171482Sjeff */ 344113357Sjeffvoid 345164936Sjuliantdq_print(int cpu) 346110267Sjeff{ 347164936Sjulian struct tdq *tdq; 348112994Sjeff 349164936Sjulian tdq = TDQ_CPU(cpu); 350112994Sjeff 351171713Sjeff printf("tdq %d:\n", TDQ_ID(tdq)); 352171482Sjeff printf("\tlockptr %p\n", TDQ_LOCKPTR(tdq)); 353165620Sjeff printf("\tload: %d\n", tdq->tdq_load); 354171482Sjeff printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 355165766Sjeff printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 356165762Sjeff printf("\trealtime runq:\n"); 357165762Sjeff runq_print(&tdq->tdq_realtime); 358165762Sjeff printf("\ttimeshare runq:\n"); 359165762Sjeff runq_print(&tdq->tdq_timeshare); 360165762Sjeff printf("\tidle runq:\n"); 361165762Sjeff runq_print(&tdq->tdq_idle); 362121896Sjeff#ifdef SMP 363165620Sjeff printf("\tload transferable: %d\n", tdq->tdq_transferable); 364171713Sjeff printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 365171713Sjeff printf("\tgroup: %d\n", TDG_ID(tdq->tdq_group)); 366171713Sjeff printf("\tLock name: %s\n", tdq->tdq_group->tdg_name); 367121896Sjeff#endif 368113357Sjeff} 369112994Sjeff 370171482Sjeff#define TS_RQ_PPQ (((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS) 371171482Sjeff/* 372171482Sjeff * Add a thread to the actual run-queue. Keeps transferable counts up to 373171482Sjeff * date with what is actually on the run-queue. Selects the correct 374171482Sjeff * queue position for timeshare threads. 375171482Sjeff */ 376122744Sjeffstatic __inline void 377164936Sjuliantdq_runq_add(struct tdq *tdq, struct td_sched *ts, int flags) 378122744Sjeff{ 379171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 380171482Sjeff THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 381122744Sjeff#ifdef SMP 382165762Sjeff if (THREAD_CAN_MIGRATE(ts->ts_thread)) { 383165620Sjeff tdq->tdq_transferable++; 384165620Sjeff tdq->tdq_group->tdg_transferable++; 385164936Sjulian ts->ts_flags |= TSF_XFERABLE; 386123433Sjeff } 387122744Sjeff#endif 388165762Sjeff if (ts->ts_runq == &tdq->tdq_timeshare) { 389166557Sjeff u_char pri; 390165762Sjeff 391165762Sjeff pri = ts->ts_thread->td_priority; 392165762Sjeff KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE, 393165762Sjeff ("Invalid priority %d on timeshare runq", pri)); 394165762Sjeff /* 395165762Sjeff * This queue contains only priorities between MIN and MAX 396165762Sjeff * realtime. Use the whole queue to represent these values. 397165762Sjeff */ 398171713Sjeff if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 399165762Sjeff pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ; 400165762Sjeff pri = (pri + tdq->tdq_idx) % RQ_NQS; 401165766Sjeff /* 402165766Sjeff * This effectively shortens the queue by one so we 403165766Sjeff * can have a one slot difference between idx and 404165766Sjeff * ridx while we wait for threads to drain. 405165766Sjeff */ 406165766Sjeff if (tdq->tdq_ridx != tdq->tdq_idx && 407165766Sjeff pri == tdq->tdq_ridx) 408167664Sjeff pri = (unsigned char)(pri - 1) % RQ_NQS; 409165762Sjeff } else 410165766Sjeff pri = tdq->tdq_ridx; 411165762Sjeff runq_add_pri(ts->ts_runq, ts, pri, flags); 412165762Sjeff } else 413165762Sjeff runq_add(ts->ts_runq, ts, flags); 414122744Sjeff} 415122744Sjeff 416171482Sjeff/* 417171482Sjeff * Remove a thread from a run-queue. This typically happens when a thread 418171482Sjeff * is selected to run. Running threads are not on the queue and the 419171482Sjeff * transferable count does not reflect them. 420171482Sjeff */ 421122744Sjeffstatic __inline void 422164936Sjuliantdq_runq_rem(struct tdq *tdq, struct td_sched *ts) 423122744Sjeff{ 424171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 425171482Sjeff KASSERT(ts->ts_runq != NULL, 426171482Sjeff ("tdq_runq_remove: thread %p null ts_runq", ts->ts_thread)); 427122744Sjeff#ifdef SMP 428164936Sjulian if (ts->ts_flags & TSF_XFERABLE) { 429165620Sjeff tdq->tdq_transferable--; 430165620Sjeff tdq->tdq_group->tdg_transferable--; 431164936Sjulian ts->ts_flags &= ~TSF_XFERABLE; 432123433Sjeff } 433122744Sjeff#endif 434165766Sjeff if (ts->ts_runq == &tdq->tdq_timeshare) { 435165766Sjeff if (tdq->tdq_idx != tdq->tdq_ridx) 436165766Sjeff runq_remove_idx(ts->ts_runq, ts, &tdq->tdq_ridx); 437165766Sjeff else 438165766Sjeff runq_remove_idx(ts->ts_runq, ts, NULL); 439165796Sjeff /* 440165796Sjeff * For timeshare threads we update the priority here so 441165796Sjeff * the priority reflects the time we've been sleeping. 442165796Sjeff */ 443165796Sjeff ts->ts_ltick = ticks; 444165796Sjeff sched_pctcpu_update(ts); 445165796Sjeff sched_priority(ts->ts_thread); 446165766Sjeff } else 447165762Sjeff runq_remove(ts->ts_runq, ts); 448122744Sjeff} 449122744Sjeff 450171482Sjeff/* 451171482Sjeff * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 452171482Sjeff * for this thread to the referenced thread queue. 453171482Sjeff */ 454113357Sjeffstatic void 455164936Sjuliantdq_load_add(struct tdq *tdq, struct td_sched *ts) 456113357Sjeff{ 457121896Sjeff int class; 458171482Sjeff 459171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 460171482Sjeff THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 461164936Sjulian class = PRI_BASE(ts->ts_thread->td_pri_class); 462165620Sjeff tdq->tdq_load++; 463171713Sjeff CTR2(KTR_SCHED, "cpu %d load: %d", TDQ_ID(tdq), tdq->tdq_load); 464166108Sjeff if (class != PRI_ITHD && 465166108Sjeff (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 466123487Sjeff#ifdef SMP 467165620Sjeff tdq->tdq_group->tdg_load++; 468125289Sjeff#else 469165620Sjeff tdq->tdq_sysload++; 470123487Sjeff#endif 471110267Sjeff} 472113357Sjeff 473171482Sjeff/* 474171482Sjeff * Remove the load from a thread that is transitioning to a sleep state or 475171482Sjeff * exiting. 476171482Sjeff */ 477112994Sjeffstatic void 478164936Sjuliantdq_load_rem(struct tdq *tdq, struct td_sched *ts) 479110267Sjeff{ 480121896Sjeff int class; 481171482Sjeff 482171482Sjeff THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 483171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 484164936Sjulian class = PRI_BASE(ts->ts_thread->td_pri_class); 485166108Sjeff if (class != PRI_ITHD && 486166108Sjeff (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 487123487Sjeff#ifdef SMP 488165620Sjeff tdq->tdq_group->tdg_load--; 489125289Sjeff#else 490165620Sjeff tdq->tdq_sysload--; 491123487Sjeff#endif 492171482Sjeff KASSERT(tdq->tdq_load != 0, 493171713Sjeff ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 494165620Sjeff tdq->tdq_load--; 495165620Sjeff CTR1(KTR_SCHED, "load: %d", tdq->tdq_load); 496164936Sjulian ts->ts_runq = NULL; 497110267Sjeff} 498110267Sjeff 499113357Sjeff#ifdef SMP 500116069Sjeff/* 501122744Sjeff * sched_balance is a simple CPU load balancing algorithm. It operates by 502116069Sjeff * finding the least loaded and most loaded cpu and equalizing their load 503116069Sjeff * by migrating some processes. 504116069Sjeff * 505116069Sjeff * Dealing only with two CPUs at a time has two advantages. Firstly, most 506116069Sjeff * installations will only have 2 cpus. Secondly, load balancing too much at 507116069Sjeff * once can have an unpleasant effect on the system. The scheduler rarely has 508116069Sjeff * enough information to make perfect decisions. So this algorithm chooses 509171482Sjeff * simplicity and more gradual effects on load in larger systems. 510116069Sjeff * 511116069Sjeff */ 512121790Sjeffstatic void 513171482Sjeffsched_balance(void *arg) 514116069Sjeff{ 515164936Sjulian struct tdq_group *high; 516164936Sjulian struct tdq_group *low; 517165620Sjeff struct tdq_group *tdg; 518123487Sjeff int cnt; 519123487Sjeff int i; 520123487Sjeff 521171506Sjeff callout_reset(&balco, max(hz / 2, random() % (hz * balance_secs)), 522171482Sjeff sched_balance, NULL); 523171482Sjeff if (smp_started == 0 || rebalance == 0) 524139334Sjeff return; 525123487Sjeff low = high = NULL; 526165620Sjeff i = random() % (tdg_maxid + 1); 527165620Sjeff for (cnt = 0; cnt <= tdg_maxid; cnt++) { 528165620Sjeff tdg = TDQ_GROUP(i); 529123487Sjeff /* 530123487Sjeff * Find the CPU with the highest load that has some 531123487Sjeff * threads to transfer. 532123487Sjeff */ 533165620Sjeff if ((high == NULL || tdg->tdg_load > high->tdg_load) 534165620Sjeff && tdg->tdg_transferable) 535165620Sjeff high = tdg; 536165620Sjeff if (low == NULL || tdg->tdg_load < low->tdg_load) 537165620Sjeff low = tdg; 538165620Sjeff if (++i > tdg_maxid) 539123487Sjeff i = 0; 540123487Sjeff } 541123487Sjeff if (low != NULL && high != NULL && high != low) 542165620Sjeff sched_balance_pair(LIST_FIRST(&high->tdg_members), 543165620Sjeff LIST_FIRST(&low->tdg_members)); 544123487Sjeff} 545123487Sjeff 546171482Sjeff/* 547171482Sjeff * Balance load between CPUs in a group. Will only migrate within the group. 548171482Sjeff */ 549123487Sjeffstatic void 550171482Sjeffsched_balance_groups(void *arg) 551123487Sjeff{ 552123487Sjeff int i; 553123487Sjeff 554171506Sjeff callout_reset(&gbalco, max(hz / 2, random() % (hz * balance_secs)), 555171482Sjeff sched_balance_groups, NULL); 556171482Sjeff if (smp_started == 0 || rebalance == 0) 557171482Sjeff return; 558171482Sjeff for (i = 0; i <= tdg_maxid; i++) 559171482Sjeff sched_balance_group(TDQ_GROUP(i)); 560123487Sjeff} 561123487Sjeff 562171482Sjeff/* 563171482Sjeff * Finds the greatest imbalance between two tdqs in a group. 564171482Sjeff */ 565123487Sjeffstatic void 566165620Sjeffsched_balance_group(struct tdq_group *tdg) 567123487Sjeff{ 568164936Sjulian struct tdq *tdq; 569164936Sjulian struct tdq *high; 570164936Sjulian struct tdq *low; 571123487Sjeff int load; 572123487Sjeff 573165620Sjeff if (tdg->tdg_transferable == 0) 574123487Sjeff return; 575123487Sjeff low = NULL; 576123487Sjeff high = NULL; 577165620Sjeff LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) { 578165620Sjeff load = tdq->tdq_load; 579165620Sjeff if (high == NULL || load > high->tdq_load) 580164936Sjulian high = tdq; 581165620Sjeff if (low == NULL || load < low->tdq_load) 582164936Sjulian low = tdq; 583123487Sjeff } 584123487Sjeff if (high != NULL && low != NULL && high != low) 585123487Sjeff sched_balance_pair(high, low); 586123487Sjeff} 587123487Sjeff 588171482Sjeff/* 589171482Sjeff * Lock two thread queues using their address to maintain lock order. 590171482Sjeff */ 591123487Sjeffstatic void 592171482Sjefftdq_lock_pair(struct tdq *one, struct tdq *two) 593171482Sjeff{ 594171482Sjeff if (one < two) { 595171482Sjeff TDQ_LOCK(one); 596171482Sjeff TDQ_LOCK_FLAGS(two, MTX_DUPOK); 597171482Sjeff } else { 598171482Sjeff TDQ_LOCK(two); 599171482Sjeff TDQ_LOCK_FLAGS(one, MTX_DUPOK); 600171482Sjeff } 601171482Sjeff} 602171482Sjeff 603171482Sjeff/* 604171482Sjeff * Transfer load between two imbalanced thread queues. 605171482Sjeff */ 606171482Sjeffstatic void 607164936Sjuliansched_balance_pair(struct tdq *high, struct tdq *low) 608123487Sjeff{ 609123433Sjeff int transferable; 610116069Sjeff int high_load; 611116069Sjeff int low_load; 612116069Sjeff int move; 613116069Sjeff int diff; 614116069Sjeff int i; 615116069Sjeff 616171482Sjeff tdq_lock_pair(high, low); 617116069Sjeff /* 618123433Sjeff * If we're transfering within a group we have to use this specific 619164936Sjulian * tdq's transferable count, otherwise we can steal from other members 620123433Sjeff * of the group. 621123433Sjeff */ 622165620Sjeff if (high->tdq_group == low->tdq_group) { 623165620Sjeff transferable = high->tdq_transferable; 624165620Sjeff high_load = high->tdq_load; 625165620Sjeff low_load = low->tdq_load; 626123487Sjeff } else { 627165620Sjeff transferable = high->tdq_group->tdg_transferable; 628165620Sjeff high_load = high->tdq_group->tdg_load; 629165620Sjeff low_load = low->tdq_group->tdg_load; 630123487Sjeff } 631123433Sjeff /* 632122744Sjeff * Determine what the imbalance is and then adjust that to how many 633165620Sjeff * threads we actually have to give up (transferable). 634122744Sjeff */ 635171482Sjeff if (transferable != 0) { 636171482Sjeff diff = high_load - low_load; 637171482Sjeff move = diff / 2; 638171482Sjeff if (diff & 0x1) 639171482Sjeff move++; 640171482Sjeff move = min(move, transferable); 641171482Sjeff for (i = 0; i < move; i++) 642171482Sjeff tdq_move(high, low); 643171482Sjeff } 644171482Sjeff TDQ_UNLOCK(high); 645171482Sjeff TDQ_UNLOCK(low); 646116069Sjeff return; 647116069Sjeff} 648116069Sjeff 649171482Sjeff/* 650171482Sjeff * Move a thread from one thread queue to another. 651171482Sjeff */ 652121790Sjeffstatic void 653171482Sjefftdq_move(struct tdq *from, struct tdq *to) 654116069Sjeff{ 655171482Sjeff struct td_sched *ts; 656171482Sjeff struct thread *td; 657164936Sjulian struct tdq *tdq; 658171482Sjeff int cpu; 659116069Sjeff 660164936Sjulian tdq = from; 661171482Sjeff cpu = TDQ_ID(to); 662164936Sjulian ts = tdq_steal(tdq, 1); 663164936Sjulian if (ts == NULL) { 664165620Sjeff struct tdq_group *tdg; 665123433Sjeff 666165620Sjeff tdg = tdq->tdq_group; 667165620Sjeff LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) { 668165620Sjeff if (tdq == from || tdq->tdq_transferable == 0) 669123433Sjeff continue; 670164936Sjulian ts = tdq_steal(tdq, 1); 671123433Sjeff break; 672123433Sjeff } 673164936Sjulian if (ts == NULL) 674171482Sjeff return; 675123433Sjeff } 676164936Sjulian if (tdq == to) 677123433Sjeff return; 678171482Sjeff td = ts->ts_thread; 679171482Sjeff /* 680171482Sjeff * Although the run queue is locked the thread may be blocked. Lock 681171482Sjeff * it to clear this. 682171482Sjeff */ 683171482Sjeff thread_lock(td); 684171482Sjeff /* Drop recursive lock on from. */ 685171482Sjeff TDQ_UNLOCK(from); 686171482Sjeff sched_rem(td); 687166108Sjeff ts->ts_cpu = cpu; 688171482Sjeff td->td_lock = TDQ_LOCKPTR(to); 689171482Sjeff tdq_add(to, td, SRQ_YIELDING); 690171505Sjeff tdq_notify(ts); 691116069Sjeff} 692110267Sjeff 693171482Sjeff/* 694171482Sjeff * This tdq has idled. Try to steal a thread from another cpu and switch 695171482Sjeff * to it. 696171482Sjeff */ 697123433Sjeffstatic int 698164936Sjuliantdq_idled(struct tdq *tdq) 699121790Sjeff{ 700165620Sjeff struct tdq_group *tdg; 701164936Sjulian struct tdq *steal; 702164936Sjulian struct td_sched *ts; 703171482Sjeff struct thread *td; 704171482Sjeff int highload; 705171482Sjeff int highcpu; 706171482Sjeff int load; 707171482Sjeff int cpu; 708123433Sjeff 709171482Sjeff /* We don't want to be preempted while we're iterating over tdqs */ 710171482Sjeff spinlock_enter(); 711165620Sjeff tdg = tdq->tdq_group; 712123433Sjeff /* 713165620Sjeff * If we're in a cpu group, try and steal threads from another cpu in 714123433Sjeff * the group before idling. 715123433Sjeff */ 716166108Sjeff if (steal_htt && tdg->tdg_cpus > 1 && tdg->tdg_transferable) { 717165620Sjeff LIST_FOREACH(steal, &tdg->tdg_members, tdq_siblings) { 718165620Sjeff if (steal == tdq || steal->tdq_transferable == 0) 719123433Sjeff continue; 720171482Sjeff TDQ_LOCK(steal); 721164936Sjulian ts = tdq_steal(steal, 0); 722166108Sjeff if (ts) 723166108Sjeff goto steal; 724171482Sjeff TDQ_UNLOCK(steal); 725166108Sjeff } 726166108Sjeff } 727171482Sjeff for (;;) { 728171482Sjeff if (steal_idle == 0) 729171482Sjeff break; 730171482Sjeff highcpu = 0; 731171482Sjeff highload = 0; 732171482Sjeff for (cpu = 0; cpu <= mp_maxid; cpu++) { 733171482Sjeff if (CPU_ABSENT(cpu)) 734171482Sjeff continue; 735166108Sjeff steal = TDQ_CPU(cpu); 736171482Sjeff load = TDQ_CPU(cpu)->tdq_transferable; 737171482Sjeff if (load < highload) 738166108Sjeff continue; 739171482Sjeff highload = load; 740171482Sjeff highcpu = cpu; 741171482Sjeff } 742171506Sjeff if (highload < steal_thresh) 743171482Sjeff break; 744171482Sjeff steal = TDQ_CPU(highcpu); 745171482Sjeff TDQ_LOCK(steal); 746171506Sjeff if (steal->tdq_transferable >= steal_thresh && 747171482Sjeff (ts = tdq_steal(steal, 1)) != NULL) 748166108Sjeff goto steal; 749171482Sjeff TDQ_UNLOCK(steal); 750171482Sjeff break; 751123433Sjeff } 752171482Sjeff spinlock_exit(); 753123433Sjeff return (1); 754166108Sjeffsteal: 755171482Sjeff td = ts->ts_thread; 756171482Sjeff thread_lock(td); 757171482Sjeff spinlock_exit(); 758171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(steal)); 759171482Sjeff TDQ_UNLOCK(steal); 760171482Sjeff sched_rem(td); 761171482Sjeff sched_setcpu(ts, PCPU_GET(cpuid), SRQ_YIELDING); 762171482Sjeff tdq_add(tdq, td, SRQ_YIELDING); 763171482Sjeff MPASS(td->td_lock == curthread->td_lock); 764171482Sjeff mi_switch(SW_VOL, NULL); 765171482Sjeff thread_unlock(curthread); 766121790Sjeff 767166108Sjeff return (0); 768121790Sjeff} 769121790Sjeff 770171482Sjeff/* 771171482Sjeff * Notify a remote cpu of new work. Sends an IPI if criteria are met. 772171482Sjeff */ 773121790Sjeffstatic void 774166108Sjefftdq_notify(struct td_sched *ts) 775121790Sjeff{ 776166247Sjeff struct thread *ctd; 777121790Sjeff struct pcpu *pcpu; 778166247Sjeff int cpri; 779166247Sjeff int pri; 780166108Sjeff int cpu; 781121790Sjeff 782166108Sjeff cpu = ts->ts_cpu; 783166247Sjeff pri = ts->ts_thread->td_priority; 784166108Sjeff pcpu = pcpu_find(cpu); 785166247Sjeff ctd = pcpu->pc_curthread; 786166247Sjeff cpri = ctd->td_priority; 787166137Sjeff 788121790Sjeff /* 789166137Sjeff * If our priority is not better than the current priority there is 790166137Sjeff * nothing to do. 791166137Sjeff */ 792166247Sjeff if (pri > cpri) 793166137Sjeff return; 794166137Sjeff /* 795166247Sjeff * Always IPI idle. 796121790Sjeff */ 797166247Sjeff if (cpri > PRI_MIN_IDLE) 798166247Sjeff goto sendipi; 799166247Sjeff /* 800166247Sjeff * If we're realtime or better and there is timeshare or worse running 801166247Sjeff * send an IPI. 802166247Sjeff */ 803166247Sjeff if (pri < PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME) 804166247Sjeff goto sendipi; 805166247Sjeff /* 806166247Sjeff * Otherwise only IPI if we exceed the threshold. 807166247Sjeff */ 808171482Sjeff if (pri > preempt_thresh) 809165819Sjeff return; 810166247Sjeffsendipi: 811166247Sjeff ctd->td_flags |= TDF_NEEDRESCHED; 812171482Sjeff ipi_selected(1 << cpu, IPI_PREEMPT); 813121790Sjeff} 814121790Sjeff 815171482Sjeff/* 816171482Sjeff * Steals load from a timeshare queue. Honors the rotating queue head 817171482Sjeff * index. 818171482Sjeff */ 819164936Sjulianstatic struct td_sched * 820171482Sjeffrunq_steal_from(struct runq *rq, u_char start) 821171482Sjeff{ 822171482Sjeff struct td_sched *ts; 823171482Sjeff struct rqbits *rqb; 824171482Sjeff struct rqhead *rqh; 825171482Sjeff int first; 826171482Sjeff int bit; 827171482Sjeff int pri; 828171482Sjeff int i; 829171482Sjeff 830171482Sjeff rqb = &rq->rq_status; 831171482Sjeff bit = start & (RQB_BPW -1); 832171482Sjeff pri = 0; 833171482Sjeff first = 0; 834171482Sjeffagain: 835171482Sjeff for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 836171482Sjeff if (rqb->rqb_bits[i] == 0) 837171482Sjeff continue; 838171482Sjeff if (bit != 0) { 839171482Sjeff for (pri = bit; pri < RQB_BPW; pri++) 840171482Sjeff if (rqb->rqb_bits[i] & (1ul << pri)) 841171482Sjeff break; 842171482Sjeff if (pri >= RQB_BPW) 843171482Sjeff continue; 844171482Sjeff } else 845171482Sjeff pri = RQB_FFS(rqb->rqb_bits[i]); 846171482Sjeff pri += (i << RQB_L2BPW); 847171482Sjeff rqh = &rq->rq_queues[pri]; 848171482Sjeff TAILQ_FOREACH(ts, rqh, ts_procq) { 849171482Sjeff if (first && THREAD_CAN_MIGRATE(ts->ts_thread)) 850171482Sjeff return (ts); 851171482Sjeff first = 1; 852171482Sjeff } 853171482Sjeff } 854171482Sjeff if (start != 0) { 855171482Sjeff start = 0; 856171482Sjeff goto again; 857171482Sjeff } 858171482Sjeff 859171482Sjeff return (NULL); 860171482Sjeff} 861171482Sjeff 862171482Sjeff/* 863171482Sjeff * Steals load from a standard linear queue. 864171482Sjeff */ 865171482Sjeffstatic struct td_sched * 866121790Sjeffrunq_steal(struct runq *rq) 867121790Sjeff{ 868121790Sjeff struct rqhead *rqh; 869121790Sjeff struct rqbits *rqb; 870164936Sjulian struct td_sched *ts; 871121790Sjeff int word; 872121790Sjeff int bit; 873121790Sjeff 874121790Sjeff rqb = &rq->rq_status; 875121790Sjeff for (word = 0; word < RQB_LEN; word++) { 876121790Sjeff if (rqb->rqb_bits[word] == 0) 877121790Sjeff continue; 878121790Sjeff for (bit = 0; bit < RQB_BPW; bit++) { 879123231Speter if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 880121790Sjeff continue; 881121790Sjeff rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 882171506Sjeff TAILQ_FOREACH(ts, rqh, ts_procq) 883171506Sjeff if (THREAD_CAN_MIGRATE(ts->ts_thread)) 884164936Sjulian return (ts); 885121790Sjeff } 886121790Sjeff } 887121790Sjeff return (NULL); 888121790Sjeff} 889121790Sjeff 890171482Sjeff/* 891171482Sjeff * Attempt to steal a thread in priority order from a thread queue. 892171482Sjeff */ 893164936Sjulianstatic struct td_sched * 894164936Sjuliantdq_steal(struct tdq *tdq, int stealidle) 895121790Sjeff{ 896164936Sjulian struct td_sched *ts; 897121790Sjeff 898171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 899165762Sjeff if ((ts = runq_steal(&tdq->tdq_realtime)) != NULL) 900164936Sjulian return (ts); 901171482Sjeff if ((ts = runq_steal_from(&tdq->tdq_timeshare, tdq->tdq_ridx)) != NULL) 902164936Sjulian return (ts); 903123433Sjeff if (stealidle) 904165620Sjeff return (runq_steal(&tdq->tdq_idle)); 905123433Sjeff return (NULL); 906121790Sjeff} 907123433Sjeff 908171482Sjeff/* 909171482Sjeff * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 910171482Sjeff * current lock and returns with the assigned queue locked. If this is 911171482Sjeff * via sched_switch() we leave the thread in a blocked state as an 912171482Sjeff * optimization. 913171482Sjeff */ 914171482Sjeffstatic inline struct tdq * 915171482Sjeffsched_setcpu(struct td_sched *ts, int cpu, int flags) 916123433Sjeff{ 917171482Sjeff struct thread *td; 918171482Sjeff struct tdq *tdq; 919123433Sjeff 920171482Sjeff THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 921171482Sjeff 922171482Sjeff tdq = TDQ_CPU(cpu); 923171482Sjeff td = ts->ts_thread; 924171482Sjeff ts->ts_cpu = cpu; 925171713Sjeff 926171713Sjeff /* If the lock matches just return the queue. */ 927171482Sjeff if (td->td_lock == TDQ_LOCKPTR(tdq)) 928171482Sjeff return (tdq); 929171482Sjeff#ifdef notyet 930123433Sjeff /* 931171482Sjeff * If the thread isn't running it's lockptr is a 932171482Sjeff * turnstile or a sleepqueue. We can just lock_set without 933171482Sjeff * blocking. 934123685Sjeff */ 935171482Sjeff if (TD_CAN_RUN(td)) { 936171482Sjeff TDQ_LOCK(tdq); 937171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 938171482Sjeff return (tdq); 939171482Sjeff } 940171482Sjeff#endif 941166108Sjeff /* 942171482Sjeff * The hard case, migration, we need to block the thread first to 943171482Sjeff * prevent order reversals with other cpus locks. 944166108Sjeff */ 945171482Sjeff thread_lock_block(td); 946171482Sjeff TDQ_LOCK(tdq); 947171713Sjeff thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 948171482Sjeff return (tdq); 949166108Sjeff} 950166108Sjeff 951171482Sjeff/* 952171482Sjeff * Find the thread queue running the lowest priority thread. 953171482Sjeff */ 954166108Sjeffstatic int 955171482Sjefftdq_lowestpri(void) 956166108Sjeff{ 957171482Sjeff struct tdq *tdq; 958166108Sjeff int lowpri; 959166108Sjeff int lowcpu; 960166108Sjeff int lowload; 961166108Sjeff int load; 962171482Sjeff int cpu; 963171482Sjeff int pri; 964171482Sjeff 965171482Sjeff lowload = 0; 966171482Sjeff lowpri = lowcpu = 0; 967171482Sjeff for (cpu = 0; cpu <= mp_maxid; cpu++) { 968171482Sjeff if (CPU_ABSENT(cpu)) 969171482Sjeff continue; 970171482Sjeff tdq = TDQ_CPU(cpu); 971171482Sjeff pri = tdq->tdq_lowpri; 972171482Sjeff load = TDQ_CPU(cpu)->tdq_load; 973171482Sjeff CTR4(KTR_ULE, 974171482Sjeff "cpu %d pri %d lowcpu %d lowpri %d", 975171482Sjeff cpu, pri, lowcpu, lowpri); 976171482Sjeff if (pri < lowpri) 977171482Sjeff continue; 978171482Sjeff if (lowpri && lowpri == pri && load > lowload) 979171482Sjeff continue; 980171482Sjeff lowpri = pri; 981171482Sjeff lowcpu = cpu; 982171482Sjeff lowload = load; 983171482Sjeff } 984171482Sjeff 985171482Sjeff return (lowcpu); 986171482Sjeff} 987171482Sjeff 988171482Sjeff/* 989171482Sjeff * Find the thread queue with the least load. 990171482Sjeff */ 991171482Sjeffstatic int 992171482Sjefftdq_lowestload(void) 993171482Sjeff{ 994171482Sjeff struct tdq *tdq; 995171482Sjeff int lowload; 996171482Sjeff int lowpri; 997171482Sjeff int lowcpu; 998171482Sjeff int load; 999171482Sjeff int cpu; 1000171482Sjeff int pri; 1001171482Sjeff 1002171482Sjeff lowcpu = 0; 1003171482Sjeff lowload = TDQ_CPU(0)->tdq_load; 1004171482Sjeff lowpri = TDQ_CPU(0)->tdq_lowpri; 1005171482Sjeff for (cpu = 1; cpu <= mp_maxid; cpu++) { 1006171482Sjeff if (CPU_ABSENT(cpu)) 1007171482Sjeff continue; 1008171482Sjeff tdq = TDQ_CPU(cpu); 1009171482Sjeff load = tdq->tdq_load; 1010171482Sjeff pri = tdq->tdq_lowpri; 1011171482Sjeff CTR4(KTR_ULE, "cpu %d load %d lowcpu %d lowload %d", 1012171482Sjeff cpu, load, lowcpu, lowload); 1013171482Sjeff if (load > lowload) 1014171482Sjeff continue; 1015171482Sjeff if (load == lowload && pri < lowpri) 1016171482Sjeff continue; 1017171482Sjeff lowcpu = cpu; 1018171482Sjeff lowload = load; 1019171482Sjeff lowpri = pri; 1020171482Sjeff } 1021171482Sjeff 1022171482Sjeff return (lowcpu); 1023171482Sjeff} 1024171482Sjeff 1025171482Sjeff/* 1026171482Sjeff * Pick the destination cpu for sched_add(). Respects affinity and makes 1027171482Sjeff * a determination based on load or priority of available processors. 1028171482Sjeff */ 1029171482Sjeffstatic int 1030171482Sjeffsched_pickcpu(struct td_sched *ts, int flags) 1031171482Sjeff{ 1032171482Sjeff struct tdq *tdq; 1033166108Sjeff int self; 1034166108Sjeff int pri; 1035166108Sjeff int cpu; 1036166108Sjeff 1037171482Sjeff cpu = self = PCPU_GET(cpuid); 1038166108Sjeff if (smp_started == 0) 1039166108Sjeff return (self); 1040171506Sjeff /* 1041171506Sjeff * Don't migrate a running thread from sched_switch(). 1042171506Sjeff */ 1043171506Sjeff if (flags & SRQ_OURSELF) { 1044171506Sjeff CTR1(KTR_ULE, "YIELDING %d", 1045171506Sjeff curthread->td_priority); 1046171506Sjeff return (self); 1047171506Sjeff } 1048166108Sjeff pri = ts->ts_thread->td_priority; 1049171482Sjeff cpu = ts->ts_cpu; 1050166108Sjeff /* 1051166108Sjeff * Regardless of affinity, if the last cpu is idle send it there. 1052166108Sjeff */ 1053171482Sjeff tdq = TDQ_CPU(cpu); 1054171482Sjeff if (tdq->tdq_lowpri > PRI_MIN_IDLE) { 1055166229Sjeff CTR5(KTR_ULE, 1056166108Sjeff "ts_cpu %d idle, ltick %d ticks %d pri %d curthread %d", 1057166108Sjeff ts->ts_cpu, ts->ts_rltick, ticks, pri, 1058171482Sjeff tdq->tdq_lowpri); 1059166108Sjeff return (ts->ts_cpu); 1060123433Sjeff } 1061166108Sjeff /* 1062166108Sjeff * If we have affinity, try to place it on the cpu we last ran on. 1063166108Sjeff */ 1064171482Sjeff if (SCHED_AFFINITY(ts) && tdq->tdq_lowpri > pri) { 1065166229Sjeff CTR5(KTR_ULE, 1066166108Sjeff "affinity for %d, ltick %d ticks %d pri %d curthread %d", 1067166108Sjeff ts->ts_cpu, ts->ts_rltick, ticks, pri, 1068171482Sjeff tdq->tdq_lowpri); 1069166108Sjeff return (ts->ts_cpu); 1070139334Sjeff } 1071123433Sjeff /* 1072166108Sjeff * Look for an idle group. 1073123433Sjeff */ 1074166229Sjeff CTR1(KTR_ULE, "tdq_idle %X", tdq_idle); 1075166108Sjeff cpu = ffs(tdq_idle); 1076166108Sjeff if (cpu) 1077171482Sjeff return (--cpu); 1078171506Sjeff /* 1079171506Sjeff * If there are no idle cores see if we can run the thread locally. This may 1080171506Sjeff * improve locality among sleepers and wakers when there is shared data. 1081171506Sjeff */ 1082171506Sjeff if (tryself && pri < curthread->td_priority) { 1083171506Sjeff CTR1(KTR_ULE, "tryself %d", 1084166108Sjeff curthread->td_priority); 1085166108Sjeff return (self); 1086123433Sjeff } 1087133427Sjeff /* 1088166108Sjeff * Now search for the cpu running the lowest priority thread with 1089166108Sjeff * the least load. 1090123433Sjeff */ 1091171482Sjeff if (pick_pri) 1092171482Sjeff cpu = tdq_lowestpri(); 1093171482Sjeff else 1094171482Sjeff cpu = tdq_lowestload(); 1095171482Sjeff return (cpu); 1096123433Sjeff} 1097123433Sjeff 1098121790Sjeff#endif /* SMP */ 1099121790Sjeff 1100117326Sjeff/* 1101121790Sjeff * Pick the highest priority task we have and return it. 1102117326Sjeff */ 1103164936Sjulianstatic struct td_sched * 1104164936Sjuliantdq_choose(struct tdq *tdq) 1105110267Sjeff{ 1106164936Sjulian struct td_sched *ts; 1107110267Sjeff 1108171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1109165762Sjeff ts = runq_choose(&tdq->tdq_realtime); 1110170787Sjeff if (ts != NULL) 1111164936Sjulian return (ts); 1112165766Sjeff ts = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1113165762Sjeff if (ts != NULL) { 1114170787Sjeff KASSERT(ts->ts_thread->td_priority >= PRI_MIN_TIMESHARE, 1115165762Sjeff ("tdq_choose: Invalid priority on timeshare queue %d", 1116165762Sjeff ts->ts_thread->td_priority)); 1117165762Sjeff return (ts); 1118165762Sjeff } 1119110267Sjeff 1120165762Sjeff ts = runq_choose(&tdq->tdq_idle); 1121165762Sjeff if (ts != NULL) { 1122165762Sjeff KASSERT(ts->ts_thread->td_priority >= PRI_MIN_IDLE, 1123165762Sjeff ("tdq_choose: Invalid priority on idle queue %d", 1124165762Sjeff ts->ts_thread->td_priority)); 1125165762Sjeff return (ts); 1126165762Sjeff } 1127165762Sjeff 1128165762Sjeff return (NULL); 1129110267Sjeff} 1130110267Sjeff 1131171482Sjeff/* 1132171482Sjeff * Initialize a thread queue. 1133171482Sjeff */ 1134109864Sjeffstatic void 1135164936Sjuliantdq_setup(struct tdq *tdq) 1136110028Sjeff{ 1137171482Sjeff 1138171713Sjeff if (bootverbose) 1139171713Sjeff printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1140165762Sjeff runq_init(&tdq->tdq_realtime); 1141165762Sjeff runq_init(&tdq->tdq_timeshare); 1142165620Sjeff runq_init(&tdq->tdq_idle); 1143165620Sjeff tdq->tdq_load = 0; 1144110028Sjeff} 1145110028Sjeff 1146171713Sjeff#ifdef SMP 1147110028Sjeffstatic void 1148171713Sjefftdg_setup(struct tdq_group *tdg) 1149109864Sjeff{ 1150171713Sjeff if (bootverbose) 1151171713Sjeff printf("ULE: setup cpu group %d\n", TDG_ID(tdg)); 1152171713Sjeff snprintf(tdg->tdg_name, sizeof(tdg->tdg_name), 1153171713Sjeff "sched lock %d", (int)TDG_ID(tdg)); 1154171713Sjeff mtx_init(&tdg->tdg_lock, tdg->tdg_name, "sched lock", 1155171713Sjeff MTX_SPIN | MTX_RECURSE); 1156171713Sjeff LIST_INIT(&tdg->tdg_members); 1157171713Sjeff tdg->tdg_load = 0; 1158171713Sjeff tdg->tdg_transferable = 0; 1159171713Sjeff tdg->tdg_cpus = 0; 1160171713Sjeff tdg->tdg_mask = 0; 1161171713Sjeff tdg->tdg_cpumask = 0; 1162171713Sjeff tdg->tdg_idlemask = 0; 1163171713Sjeff} 1164171713Sjeff 1165171713Sjeffstatic void 1166171713Sjefftdg_add(struct tdq_group *tdg, struct tdq *tdq) 1167171713Sjeff{ 1168171713Sjeff if (tdg->tdg_mask == 0) 1169171713Sjeff tdg->tdg_mask |= 1 << TDQ_ID(tdq); 1170171713Sjeff tdg->tdg_cpumask |= 1 << TDQ_ID(tdq); 1171171713Sjeff tdg->tdg_cpus++; 1172171713Sjeff tdq->tdq_group = tdg; 1173171713Sjeff tdq->tdq_lock = &tdg->tdg_lock; 1174171713Sjeff LIST_INSERT_HEAD(&tdg->tdg_members, tdq, tdq_siblings); 1175171713Sjeff if (bootverbose) 1176171713Sjeff printf("ULE: adding cpu %d to group %d: cpus %d mask 0x%X\n", 1177171713Sjeff TDQ_ID(tdq), TDG_ID(tdg), tdg->tdg_cpus, tdg->tdg_cpumask); 1178171713Sjeff} 1179171713Sjeff 1180171713Sjeffstatic void 1181171713Sjeffsched_setup_topology(void) 1182171713Sjeff{ 1183171713Sjeff struct tdq_group *tdg; 1184171713Sjeff struct cpu_group *cg; 1185171713Sjeff int balance_groups; 1186171482Sjeff struct tdq *tdq; 1187109864Sjeff int i; 1188171713Sjeff int j; 1189109864Sjeff 1190171713Sjeff topology = 1; 1191123487Sjeff balance_groups = 0; 1192171713Sjeff for (i = 0; i < smp_topology->ct_count; i++) { 1193171713Sjeff cg = &smp_topology->ct_group[i]; 1194171713Sjeff tdg = &tdq_groups[i]; 1195171713Sjeff /* 1196171713Sjeff * Initialize the group. 1197171713Sjeff */ 1198171713Sjeff tdg_setup(tdg); 1199171713Sjeff /* 1200171713Sjeff * Find all of the group members and add them. 1201171713Sjeff */ 1202171713Sjeff for (j = 0; j < MAXCPU; j++) { 1203171713Sjeff if ((cg->cg_mask & (1 << j)) != 0) { 1204171713Sjeff tdq = TDQ_CPU(j); 1205171713Sjeff tdq_setup(tdq); 1206171713Sjeff tdg_add(tdg, tdq); 1207171713Sjeff } 1208171713Sjeff } 1209171713Sjeff if (tdg->tdg_cpus > 1) 1210171713Sjeff balance_groups = 1; 1211171713Sjeff } 1212171713Sjeff tdg_maxid = smp_topology->ct_count - 1; 1213171713Sjeff if (balance_groups) 1214171713Sjeff sched_balance_groups(NULL); 1215171713Sjeff} 1216171713Sjeff 1217171713Sjeffstatic void 1218171713Sjeffsched_setup_smp(void) 1219171713Sjeff{ 1220171713Sjeff struct tdq_group *tdg; 1221171713Sjeff struct tdq *tdq; 1222171713Sjeff int cpus; 1223171713Sjeff int i; 1224171713Sjeff 1225171713Sjeff for (cpus = 0, i = 0; i < MAXCPU; i++) { 1226171713Sjeff if (CPU_ABSENT(i)) 1227171713Sjeff continue; 1228165627Sjeff tdq = &tdq_cpu[i]; 1229171713Sjeff tdg = &tdq_groups[i]; 1230171713Sjeff /* 1231171713Sjeff * Setup a tdq group with one member. 1232171713Sjeff */ 1233171713Sjeff tdg_setup(tdg); 1234171713Sjeff tdq_setup(tdq); 1235171713Sjeff tdg_add(tdg, tdq); 1236171713Sjeff cpus++; 1237123433Sjeff } 1238171713Sjeff tdg_maxid = cpus - 1; 1239171713Sjeff} 1240123433Sjeff 1241171713Sjeff/* 1242171713Sjeff * Fake a topology with one group containing all CPUs. 1243171713Sjeff */ 1244171713Sjeffstatic void 1245171713Sjeffsched_fake_topo(void) 1246171713Sjeff{ 1247171713Sjeff#ifdef SCHED_FAKE_TOPOLOGY 1248171713Sjeff static struct cpu_top top; 1249171713Sjeff static struct cpu_group group; 1250113357Sjeff 1251171713Sjeff top.ct_count = 1; 1252171713Sjeff top.ct_group = &group; 1253171713Sjeff group.cg_mask = all_cpus; 1254171713Sjeff group.cg_count = mp_ncpus; 1255171713Sjeff group.cg_children = 0; 1256171713Sjeff smp_topology = ⊤ 1257171713Sjeff#endif 1258171713Sjeff} 1259171713Sjeff#endif 1260171713Sjeff 1261171713Sjeff/* 1262171713Sjeff * Setup the thread queues and initialize the topology based on MD 1263171713Sjeff * information. 1264171713Sjeff */ 1265171713Sjeffstatic void 1266171713Sjeffsched_setup(void *dummy) 1267171713Sjeff{ 1268171713Sjeff struct tdq *tdq; 1269171713Sjeff 1270171713Sjeff tdq = TDQ_SELF(); 1271171713Sjeff#ifdef SMP 1272123487Sjeff /* 1273171482Sjeff * Initialize long-term cpu balancing algorithm. 1274123487Sjeff */ 1275171482Sjeff callout_init(&balco, CALLOUT_MPSAFE); 1276171482Sjeff callout_init(&gbalco, CALLOUT_MPSAFE); 1277171713Sjeff sched_fake_topo(); 1278171713Sjeff /* 1279171713Sjeff * Setup tdqs based on a topology configuration or vanilla SMP based 1280171713Sjeff * on mp_maxid. 1281171713Sjeff */ 1282171713Sjeff if (smp_topology == NULL) 1283171713Sjeff sched_setup_smp(); 1284171713Sjeff else 1285171713Sjeff sched_setup_topology(); 1286171482Sjeff sched_balance(NULL); 1287117237Sjeff#else 1288171713Sjeff tdq_setup(tdq); 1289171713Sjeff mtx_init(&tdq_lock, "sched lock", "sched lock", MTX_SPIN | MTX_RECURSE); 1290171713Sjeff tdq->tdq_lock = &tdq_lock; 1291116069Sjeff#endif 1292171482Sjeff /* 1293171482Sjeff * To avoid divide-by-zero, we set realstathz a dummy value 1294171482Sjeff * in case which sched_clock() called before sched_initticks(). 1295171482Sjeff */ 1296171482Sjeff realstathz = hz; 1297171482Sjeff sched_slice = (realstathz/10); /* ~100ms */ 1298171482Sjeff tickincr = 1 << SCHED_TICK_SHIFT; 1299171482Sjeff 1300171482Sjeff /* Add thread0's load since it's running. */ 1301171482Sjeff TDQ_LOCK(tdq); 1302171713Sjeff thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1303171482Sjeff tdq_load_add(tdq, &td_sched0); 1304171482Sjeff TDQ_UNLOCK(tdq); 1305109864Sjeff} 1306109864Sjeff 1307171482Sjeff/* 1308171482Sjeff * This routine determines the tickincr after stathz and hz are setup. 1309171482Sjeff */ 1310153533Sdavidxu/* ARGSUSED */ 1311153533Sdavidxustatic void 1312153533Sdavidxusched_initticks(void *dummy) 1313153533Sdavidxu{ 1314171482Sjeff int incr; 1315171482Sjeff 1316153533Sdavidxu realstathz = stathz ? stathz : hz; 1317166229Sjeff sched_slice = (realstathz/10); /* ~100ms */ 1318153533Sdavidxu 1319153533Sdavidxu /* 1320165762Sjeff * tickincr is shifted out by 10 to avoid rounding errors due to 1321165766Sjeff * hz not being evenly divisible by stathz on all platforms. 1322153533Sdavidxu */ 1323171482Sjeff incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1324165762Sjeff /* 1325165762Sjeff * This does not work for values of stathz that are more than 1326165762Sjeff * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1327165762Sjeff */ 1328171482Sjeff if (incr == 0) 1329171482Sjeff incr = 1; 1330171482Sjeff tickincr = incr; 1331166108Sjeff#ifdef SMP 1332166108Sjeff affinity = SCHED_AFFINITY_DEFAULT; 1333166108Sjeff#endif 1334153533Sdavidxu} 1335153533Sdavidxu 1336153533Sdavidxu 1337109864Sjeff/* 1338171482Sjeff * This is the core of the interactivity algorithm. Determines a score based 1339171482Sjeff * on past behavior. It is the ratio of sleep time to run time scaled to 1340171482Sjeff * a [0, 100] integer. This is the voluntary sleep time of a process, which 1341171482Sjeff * differs from the cpu usage because it does not account for time spent 1342171482Sjeff * waiting on a run-queue. Would be prettier if we had floating point. 1343171482Sjeff */ 1344171482Sjeffstatic int 1345171482Sjeffsched_interact_score(struct thread *td) 1346171482Sjeff{ 1347171482Sjeff struct td_sched *ts; 1348171482Sjeff int div; 1349171482Sjeff 1350171482Sjeff ts = td->td_sched; 1351171482Sjeff /* 1352171482Sjeff * The score is only needed if this is likely to be an interactive 1353171482Sjeff * task. Don't go through the expense of computing it if there's 1354171482Sjeff * no chance. 1355171482Sjeff */ 1356171482Sjeff if (sched_interact <= SCHED_INTERACT_HALF && 1357171482Sjeff ts->ts_runtime >= ts->ts_slptime) 1358171482Sjeff return (SCHED_INTERACT_HALF); 1359171482Sjeff 1360171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1361171482Sjeff div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1362171482Sjeff return (SCHED_INTERACT_HALF + 1363171482Sjeff (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1364171482Sjeff } 1365171482Sjeff if (ts->ts_slptime > ts->ts_runtime) { 1366171482Sjeff div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1367171482Sjeff return (ts->ts_runtime / div); 1368171482Sjeff } 1369171482Sjeff /* runtime == slptime */ 1370171482Sjeff if (ts->ts_runtime) 1371171482Sjeff return (SCHED_INTERACT_HALF); 1372171482Sjeff 1373171482Sjeff /* 1374171482Sjeff * This can happen if slptime and runtime are 0. 1375171482Sjeff */ 1376171482Sjeff return (0); 1377171482Sjeff 1378171482Sjeff} 1379171482Sjeff 1380171482Sjeff/* 1381109864Sjeff * Scale the scheduling priority according to the "interactivity" of this 1382109864Sjeff * process. 1383109864Sjeff */ 1384113357Sjeffstatic void 1385163709Sjbsched_priority(struct thread *td) 1386109864Sjeff{ 1387165762Sjeff int score; 1388109864Sjeff int pri; 1389109864Sjeff 1390163709Sjb if (td->td_pri_class != PRI_TIMESHARE) 1391113357Sjeff return; 1392112966Sjeff /* 1393165762Sjeff * If the score is interactive we place the thread in the realtime 1394165762Sjeff * queue with a priority that is less than kernel and interrupt 1395165762Sjeff * priorities. These threads are not subject to nice restrictions. 1396112966Sjeff * 1397171482Sjeff * Scores greater than this are placed on the normal timeshare queue 1398165762Sjeff * where the priority is partially decided by the most recent cpu 1399165762Sjeff * utilization and the rest is decided by nice value. 1400112966Sjeff */ 1401165762Sjeff score = sched_interact_score(td); 1402165762Sjeff if (score < sched_interact) { 1403165762Sjeff pri = PRI_MIN_REALTIME; 1404165762Sjeff pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact) 1405165762Sjeff * score; 1406165762Sjeff KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME, 1407166208Sjeff ("sched_priority: invalid interactive priority %d score %d", 1408166208Sjeff pri, score)); 1409165762Sjeff } else { 1410165762Sjeff pri = SCHED_PRI_MIN; 1411165762Sjeff if (td->td_sched->ts_ticks) 1412165762Sjeff pri += SCHED_PRI_TICKS(td->td_sched); 1413165762Sjeff pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1414171482Sjeff KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE, 1415171482Sjeff ("sched_priority: invalid priority %d: nice %d, " 1416171482Sjeff "ticks %d ftick %d ltick %d tick pri %d", 1417171482Sjeff pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1418171482Sjeff td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1419171482Sjeff SCHED_PRI_TICKS(td->td_sched))); 1420165762Sjeff } 1421165762Sjeff sched_user_prio(td, pri); 1422112966Sjeff 1423112966Sjeff return; 1424109864Sjeff} 1425109864Sjeff 1426121868Sjeff/* 1427121868Sjeff * This routine enforces a maximum limit on the amount of scheduling history 1428171482Sjeff * kept. It is called after either the slptime or runtime is adjusted. This 1429171482Sjeff * function is ugly due to integer math. 1430121868Sjeff */ 1431116463Sjeffstatic void 1432163709Sjbsched_interact_update(struct thread *td) 1433116463Sjeff{ 1434165819Sjeff struct td_sched *ts; 1435166208Sjeff u_int sum; 1436121605Sjeff 1437165819Sjeff ts = td->td_sched; 1438171482Sjeff sum = ts->ts_runtime + ts->ts_slptime; 1439121868Sjeff if (sum < SCHED_SLP_RUN_MAX) 1440121868Sjeff return; 1441121868Sjeff /* 1442165819Sjeff * This only happens from two places: 1443165819Sjeff * 1) We have added an unusual amount of run time from fork_exit. 1444165819Sjeff * 2) We have added an unusual amount of sleep time from sched_sleep(). 1445165819Sjeff */ 1446165819Sjeff if (sum > SCHED_SLP_RUN_MAX * 2) { 1447171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1448171482Sjeff ts->ts_runtime = SCHED_SLP_RUN_MAX; 1449171482Sjeff ts->ts_slptime = 1; 1450165819Sjeff } else { 1451171482Sjeff ts->ts_slptime = SCHED_SLP_RUN_MAX; 1452171482Sjeff ts->ts_runtime = 1; 1453165819Sjeff } 1454165819Sjeff return; 1455165819Sjeff } 1456165819Sjeff /* 1457121868Sjeff * If we have exceeded by more than 1/5th then the algorithm below 1458121868Sjeff * will not bring us back into range. Dividing by two here forces 1459133427Sjeff * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1460121868Sjeff */ 1461127850Sjeff if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1462171482Sjeff ts->ts_runtime /= 2; 1463171482Sjeff ts->ts_slptime /= 2; 1464121868Sjeff return; 1465116463Sjeff } 1466171482Sjeff ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1467171482Sjeff ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1468116463Sjeff} 1469116463Sjeff 1470171482Sjeff/* 1471171482Sjeff * Scale back the interactivity history when a child thread is created. The 1472171482Sjeff * history is inherited from the parent but the thread may behave totally 1473171482Sjeff * differently. For example, a shell spawning a compiler process. We want 1474171482Sjeff * to learn that the compiler is behaving badly very quickly. 1475171482Sjeff */ 1476121868Sjeffstatic void 1477163709Sjbsched_interact_fork(struct thread *td) 1478121868Sjeff{ 1479121868Sjeff int ratio; 1480121868Sjeff int sum; 1481121868Sjeff 1482171482Sjeff sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1483121868Sjeff if (sum > SCHED_SLP_RUN_FORK) { 1484121868Sjeff ratio = sum / SCHED_SLP_RUN_FORK; 1485171482Sjeff td->td_sched->ts_runtime /= ratio; 1486171482Sjeff td->td_sched->ts_slptime /= ratio; 1487121868Sjeff } 1488121868Sjeff} 1489121868Sjeff 1490113357Sjeff/* 1491171482Sjeff * Called from proc0_init() to setup the scheduler fields. 1492134791Sjulian */ 1493134791Sjulianvoid 1494134791Sjulianschedinit(void) 1495134791Sjulian{ 1496165762Sjeff 1497134791Sjulian /* 1498134791Sjulian * Set up the scheduler specific parts of proc0. 1499134791Sjulian */ 1500136167Sjulian proc0.p_sched = NULL; /* XXX */ 1501164936Sjulian thread0.td_sched = &td_sched0; 1502165762Sjeff td_sched0.ts_ltick = ticks; 1503165796Sjeff td_sched0.ts_ftick = ticks; 1504164936Sjulian td_sched0.ts_thread = &thread0; 1505134791Sjulian} 1506134791Sjulian 1507134791Sjulian/* 1508113357Sjeff * This is only somewhat accurate since given many processes of the same 1509113357Sjeff * priority they will switch when their slices run out, which will be 1510165762Sjeff * at most sched_slice stathz ticks. 1511113357Sjeff */ 1512109864Sjeffint 1513109864Sjeffsched_rr_interval(void) 1514109864Sjeff{ 1515165762Sjeff 1516165762Sjeff /* Convert sched_slice to hz */ 1517165762Sjeff return (hz/(realstathz/sched_slice)); 1518109864Sjeff} 1519109864Sjeff 1520171482Sjeff/* 1521171482Sjeff * Update the percent cpu tracking information when it is requested or 1522171482Sjeff * the total history exceeds the maximum. We keep a sliding history of 1523171482Sjeff * tick counts that slowly decays. This is less precise than the 4BSD 1524171482Sjeff * mechanism since it happens with less regular and frequent events. 1525171482Sjeff */ 1526121790Sjeffstatic void 1527164936Sjuliansched_pctcpu_update(struct td_sched *ts) 1528109864Sjeff{ 1529165762Sjeff 1530165762Sjeff if (ts->ts_ticks == 0) 1531165762Sjeff return; 1532165796Sjeff if (ticks - (hz / 10) < ts->ts_ltick && 1533165796Sjeff SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1534165796Sjeff return; 1535109864Sjeff /* 1536109864Sjeff * Adjust counters and watermark for pctcpu calc. 1537116365Sjeff */ 1538165762Sjeff if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1539164936Sjulian ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1540165762Sjeff SCHED_TICK_TARG; 1541165762Sjeff else 1542164936Sjulian ts->ts_ticks = 0; 1543164936Sjulian ts->ts_ltick = ticks; 1544165762Sjeff ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1545109864Sjeff} 1546109864Sjeff 1547171482Sjeff/* 1548171482Sjeff * Adjust the priority of a thread. Move it to the appropriate run-queue 1549171482Sjeff * if necessary. This is the back-end for several priority related 1550171482Sjeff * functions. 1551171482Sjeff */ 1552165762Sjeffstatic void 1553139453Sjhbsched_thread_priority(struct thread *td, u_char prio) 1554109864Sjeff{ 1555164936Sjulian struct td_sched *ts; 1556109864Sjeff 1557139316Sjeff CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 1558139316Sjeff td, td->td_proc->p_comm, td->td_priority, prio, curthread, 1559139316Sjeff curthread->td_proc->p_comm); 1560164936Sjulian ts = td->td_sched; 1561170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1562139453Sjhb if (td->td_priority == prio) 1563139453Sjhb return; 1564165762Sjeff 1565165766Sjeff if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1566121605Sjeff /* 1567121605Sjeff * If the priority has been elevated due to priority 1568121605Sjeff * propagation, we may have to move ourselves to a new 1569165762Sjeff * queue. This could be optimized to not re-add in some 1570165762Sjeff * cases. 1571133555Sjeff */ 1572165762Sjeff sched_rem(td); 1573165762Sjeff td->td_priority = prio; 1574171482Sjeff sched_add(td, SRQ_BORROWING); 1575171482Sjeff } else { 1576171482Sjeff#ifdef SMP 1577171482Sjeff struct tdq *tdq; 1578171482Sjeff 1579171482Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1580171482Sjeff if (prio < tdq->tdq_lowpri) 1581171482Sjeff tdq->tdq_lowpri = prio; 1582171482Sjeff#endif 1583119488Sdavidxu td->td_priority = prio; 1584171482Sjeff } 1585109864Sjeff} 1586109864Sjeff 1587139453Sjhb/* 1588139453Sjhb * Update a thread's priority when it is lent another thread's 1589139453Sjhb * priority. 1590139453Sjhb */ 1591109864Sjeffvoid 1592139453Sjhbsched_lend_prio(struct thread *td, u_char prio) 1593139453Sjhb{ 1594139453Sjhb 1595139453Sjhb td->td_flags |= TDF_BORROWING; 1596139453Sjhb sched_thread_priority(td, prio); 1597139453Sjhb} 1598139453Sjhb 1599139453Sjhb/* 1600139453Sjhb * Restore a thread's priority when priority propagation is 1601139453Sjhb * over. The prio argument is the minimum priority the thread 1602139453Sjhb * needs to have to satisfy other possible priority lending 1603139453Sjhb * requests. If the thread's regular priority is less 1604139453Sjhb * important than prio, the thread will keep a priority boost 1605139453Sjhb * of prio. 1606139453Sjhb */ 1607139453Sjhbvoid 1608139453Sjhbsched_unlend_prio(struct thread *td, u_char prio) 1609139453Sjhb{ 1610139453Sjhb u_char base_pri; 1611139453Sjhb 1612139453Sjhb if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1613139453Sjhb td->td_base_pri <= PRI_MAX_TIMESHARE) 1614163709Sjb base_pri = td->td_user_pri; 1615139453Sjhb else 1616139453Sjhb base_pri = td->td_base_pri; 1617139453Sjhb if (prio >= base_pri) { 1618139455Sjhb td->td_flags &= ~TDF_BORROWING; 1619139453Sjhb sched_thread_priority(td, base_pri); 1620139453Sjhb } else 1621139453Sjhb sched_lend_prio(td, prio); 1622139453Sjhb} 1623139453Sjhb 1624171482Sjeff/* 1625171482Sjeff * Standard entry for setting the priority to an absolute value. 1626171482Sjeff */ 1627139453Sjhbvoid 1628139453Sjhbsched_prio(struct thread *td, u_char prio) 1629139453Sjhb{ 1630139453Sjhb u_char oldprio; 1631139453Sjhb 1632139453Sjhb /* First, update the base priority. */ 1633139453Sjhb td->td_base_pri = prio; 1634139453Sjhb 1635139453Sjhb /* 1636139455Sjhb * If the thread is borrowing another thread's priority, don't 1637139453Sjhb * ever lower the priority. 1638139453Sjhb */ 1639139453Sjhb if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1640139453Sjhb return; 1641139453Sjhb 1642139453Sjhb /* Change the real priority. */ 1643139453Sjhb oldprio = td->td_priority; 1644139453Sjhb sched_thread_priority(td, prio); 1645139453Sjhb 1646139453Sjhb /* 1647139453Sjhb * If the thread is on a turnstile, then let the turnstile update 1648139453Sjhb * its state. 1649139453Sjhb */ 1650139453Sjhb if (TD_ON_LOCK(td) && oldprio != prio) 1651139453Sjhb turnstile_adjust(td, oldprio); 1652139453Sjhb} 1653139455Sjhb 1654171482Sjeff/* 1655171482Sjeff * Set the base user priority, does not effect current running priority. 1656171482Sjeff */ 1657139453Sjhbvoid 1658163709Sjbsched_user_prio(struct thread *td, u_char prio) 1659161599Sdavidxu{ 1660161599Sdavidxu u_char oldprio; 1661161599Sdavidxu 1662163709Sjb td->td_base_user_pri = prio; 1663164939Sjulian if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio) 1664164939Sjulian return; 1665163709Sjb oldprio = td->td_user_pri; 1666163709Sjb td->td_user_pri = prio; 1667163709Sjb 1668161599Sdavidxu if (TD_ON_UPILOCK(td) && oldprio != prio) 1669161599Sdavidxu umtx_pi_adjust(td, oldprio); 1670161599Sdavidxu} 1671161599Sdavidxu 1672161599Sdavidxuvoid 1673161599Sdavidxusched_lend_user_prio(struct thread *td, u_char prio) 1674161599Sdavidxu{ 1675161599Sdavidxu u_char oldprio; 1676161599Sdavidxu 1677161599Sdavidxu td->td_flags |= TDF_UBORROWING; 1678161599Sdavidxu 1679164091Smaxim oldprio = td->td_user_pri; 1680163709Sjb td->td_user_pri = prio; 1681161599Sdavidxu 1682161599Sdavidxu if (TD_ON_UPILOCK(td) && oldprio != prio) 1683161599Sdavidxu umtx_pi_adjust(td, oldprio); 1684161599Sdavidxu} 1685161599Sdavidxu 1686161599Sdavidxuvoid 1687161599Sdavidxusched_unlend_user_prio(struct thread *td, u_char prio) 1688161599Sdavidxu{ 1689161599Sdavidxu u_char base_pri; 1690161599Sdavidxu 1691163709Sjb base_pri = td->td_base_user_pri; 1692161599Sdavidxu if (prio >= base_pri) { 1693161599Sdavidxu td->td_flags &= ~TDF_UBORROWING; 1694163709Sjb sched_user_prio(td, base_pri); 1695161599Sdavidxu } else 1696161599Sdavidxu sched_lend_user_prio(td, prio); 1697161599Sdavidxu} 1698161599Sdavidxu 1699171482Sjeff/* 1700171505Sjeff * Add the thread passed as 'newtd' to the run queue before selecting 1701171505Sjeff * the next thread to run. This is only used for KSE. 1702171505Sjeff */ 1703171505Sjeffstatic void 1704171505Sjeffsched_switchin(struct tdq *tdq, struct thread *td) 1705171505Sjeff{ 1706171505Sjeff#ifdef SMP 1707171505Sjeff spinlock_enter(); 1708171505Sjeff TDQ_UNLOCK(tdq); 1709171505Sjeff thread_lock(td); 1710171505Sjeff spinlock_exit(); 1711171505Sjeff sched_setcpu(td->td_sched, TDQ_ID(tdq), SRQ_YIELDING); 1712171505Sjeff#else 1713171505Sjeff td->td_lock = TDQ_LOCKPTR(tdq); 1714171505Sjeff#endif 1715171505Sjeff tdq_add(tdq, td, SRQ_YIELDING); 1716171505Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1717171505Sjeff} 1718171505Sjeff 1719171505Sjeff/* 1720171713Sjeff * Handle migration from sched_switch(). This happens only for 1721171713Sjeff * cpu binding. 1722171713Sjeff */ 1723171713Sjeffstatic struct mtx * 1724171713Sjeffsched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1725171713Sjeff{ 1726171713Sjeff struct tdq *tdn; 1727171713Sjeff 1728171713Sjeff tdn = TDQ_CPU(td->td_sched->ts_cpu); 1729171713Sjeff#ifdef SMP 1730171713Sjeff /* 1731171713Sjeff * Do the lock dance required to avoid LOR. We grab an extra 1732171713Sjeff * spinlock nesting to prevent preemption while we're 1733171713Sjeff * not holding either run-queue lock. 1734171713Sjeff */ 1735171713Sjeff spinlock_enter(); 1736171713Sjeff thread_block_switch(td); /* This releases the lock on tdq. */ 1737171713Sjeff TDQ_LOCK(tdn); 1738171713Sjeff tdq_add(tdn, td, flags); 1739171713Sjeff tdq_notify(td->td_sched); 1740171713Sjeff /* 1741171713Sjeff * After we unlock tdn the new cpu still can't switch into this 1742171713Sjeff * thread until we've unblocked it in cpu_switch(). The lock 1743171713Sjeff * pointers may match in the case of HTT cores. Don't unlock here 1744171713Sjeff * or we can deadlock when the other CPU runs the IPI handler. 1745171713Sjeff */ 1746171713Sjeff if (TDQ_LOCKPTR(tdn) != TDQ_LOCKPTR(tdq)) { 1747171713Sjeff TDQ_UNLOCK(tdn); 1748171713Sjeff TDQ_LOCK(tdq); 1749171713Sjeff } 1750171713Sjeff spinlock_exit(); 1751171713Sjeff#endif 1752171713Sjeff return (TDQ_LOCKPTR(tdn)); 1753171713Sjeff} 1754171713Sjeff 1755171713Sjeff/* 1756171482Sjeff * Block a thread for switching. Similar to thread_block() but does not 1757171482Sjeff * bump the spin count. 1758171482Sjeff */ 1759171482Sjeffstatic inline struct mtx * 1760171482Sjeffthread_block_switch(struct thread *td) 1761171482Sjeff{ 1762171482Sjeff struct mtx *lock; 1763171482Sjeff 1764171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1765171482Sjeff lock = td->td_lock; 1766171482Sjeff td->td_lock = &blocked_lock; 1767171482Sjeff mtx_unlock_spin(lock); 1768171482Sjeff 1769171482Sjeff return (lock); 1770171482Sjeff} 1771171482Sjeff 1772171482Sjeff/* 1773171482Sjeff * Release a thread that was blocked with thread_block_switch(). 1774171482Sjeff */ 1775171482Sjeffstatic inline void 1776171482Sjeffthread_unblock_switch(struct thread *td, struct mtx *mtx) 1777171482Sjeff{ 1778171482Sjeff atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1779171482Sjeff (uintptr_t)mtx); 1780171482Sjeff} 1781171482Sjeff 1782171482Sjeff/* 1783171482Sjeff * Switch threads. This function has to handle threads coming in while 1784171482Sjeff * blocked for some reason, running, or idle. It also must deal with 1785171482Sjeff * migrating a thread from one queue to another as running threads may 1786171482Sjeff * be assigned elsewhere via binding. 1787171482Sjeff */ 1788161599Sdavidxuvoid 1789135051Sjuliansched_switch(struct thread *td, struct thread *newtd, int flags) 1790109864Sjeff{ 1791165627Sjeff struct tdq *tdq; 1792164936Sjulian struct td_sched *ts; 1793171482Sjeff struct mtx *mtx; 1794171713Sjeff int srqflag; 1795171482Sjeff int cpuid; 1796109864Sjeff 1797170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1798109864Sjeff 1799171482Sjeff cpuid = PCPU_GET(cpuid); 1800171482Sjeff tdq = TDQ_CPU(cpuid); 1801164936Sjulian ts = td->td_sched; 1802171713Sjeff mtx = td->td_lock; 1803171482Sjeff#ifdef SMP 1804171482Sjeff ts->ts_rltick = ticks; 1805171482Sjeff if (newtd && newtd->td_priority < tdq->tdq_lowpri) 1806171482Sjeff tdq->tdq_lowpri = newtd->td_priority; 1807171482Sjeff#endif 1808133555Sjeff td->td_lastcpu = td->td_oncpu; 1809113339Sjulian td->td_oncpu = NOCPU; 1810132266Sjhb td->td_flags &= ~TDF_NEEDRESCHED; 1811144777Sups td->td_owepreempt = 0; 1812123434Sjeff /* 1813171482Sjeff * The lock pointer in an idle thread should never change. Reset it 1814171482Sjeff * to CAN_RUN as well. 1815123434Sjeff */ 1816167327Sjulian if (TD_IS_IDLETHREAD(td)) { 1817171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1818139334Sjeff TD_SET_CAN_RUN(td); 1819170293Sjeff } else if (TD_IS_RUNNING(td)) { 1820171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1821165627Sjeff tdq_load_rem(tdq, ts); 1822171713Sjeff srqflag = (flags & SW_PREEMPT) ? 1823170293Sjeff SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1824171713Sjeff SRQ_OURSELF|SRQ_YIELDING; 1825171713Sjeff if (ts->ts_cpu == cpuid) 1826171713Sjeff tdq_add(tdq, td, srqflag); 1827171713Sjeff else 1828171713Sjeff mtx = sched_switch_migrate(tdq, td, srqflag); 1829171482Sjeff } else { 1830171482Sjeff /* This thread must be going to sleep. */ 1831171482Sjeff TDQ_LOCK(tdq); 1832171482Sjeff mtx = thread_block_switch(td); 1833170293Sjeff tdq_load_rem(tdq, ts); 1834171482Sjeff } 1835171482Sjeff /* 1836171482Sjeff * We enter here with the thread blocked and assigned to the 1837171482Sjeff * appropriate cpu run-queue or sleep-queue and with the current 1838171482Sjeff * thread-queue locked. 1839171482Sjeff */ 1840171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1841171482Sjeff /* 1842171505Sjeff * If KSE assigned a new thread just add it here and let choosethread 1843171505Sjeff * select the best one. 1844171482Sjeff */ 1845171505Sjeff if (newtd != NULL) 1846171505Sjeff sched_switchin(tdq, newtd); 1847171482Sjeff newtd = choosethread(); 1848171482Sjeff /* 1849171482Sjeff * Call the MD code to switch contexts if necessary. 1850171482Sjeff */ 1851145256Sjkoshy if (td != newtd) { 1852145256Sjkoshy#ifdef HWPMC_HOOKS 1853145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1854145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1855145256Sjkoshy#endif 1856171482Sjeff cpu_switch(td, newtd, mtx); 1857171482Sjeff /* 1858171482Sjeff * We may return from cpu_switch on a different cpu. However, 1859171482Sjeff * we always return with td_lock pointing to the current cpu's 1860171482Sjeff * run queue lock. 1861171482Sjeff */ 1862171482Sjeff cpuid = PCPU_GET(cpuid); 1863171482Sjeff tdq = TDQ_CPU(cpuid); 1864171482Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)td; 1865145256Sjkoshy#ifdef HWPMC_HOOKS 1866145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1867145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1868145256Sjkoshy#endif 1869171482Sjeff } else 1870171482Sjeff thread_unblock_switch(td, mtx); 1871171482Sjeff /* 1872171482Sjeff * Assert that all went well and return. 1873171482Sjeff */ 1874171482Sjeff#ifdef SMP 1875171482Sjeff /* We should always get here with the lowest priority td possible */ 1876171482Sjeff tdq->tdq_lowpri = td->td_priority; 1877171482Sjeff#endif 1878171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1879171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1880171482Sjeff td->td_oncpu = cpuid; 1881109864Sjeff} 1882109864Sjeff 1883171482Sjeff/* 1884171482Sjeff * Adjust thread priorities as a result of a nice request. 1885171482Sjeff */ 1886109864Sjeffvoid 1887130551Sjuliansched_nice(struct proc *p, int nice) 1888109864Sjeff{ 1889109864Sjeff struct thread *td; 1890109864Sjeff 1891130551Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 1892170293Sjeff PROC_SLOCK_ASSERT(p, MA_OWNED); 1893165762Sjeff 1894130551Sjulian p->p_nice = nice; 1895163709Sjb FOREACH_THREAD_IN_PROC(p, td) { 1896170293Sjeff thread_lock(td); 1897163709Sjb sched_priority(td); 1898165762Sjeff sched_prio(td, td->td_base_user_pri); 1899170293Sjeff thread_unlock(td); 1900130551Sjulian } 1901109864Sjeff} 1902109864Sjeff 1903171482Sjeff/* 1904171482Sjeff * Record the sleep time for the interactivity scorer. 1905171482Sjeff */ 1906109864Sjeffvoid 1907126326Sjhbsched_sleep(struct thread *td) 1908109864Sjeff{ 1909165762Sjeff 1910170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1911109864Sjeff 1912171482Sjeff td->td_sched->ts_slptick = ticks; 1913109864Sjeff} 1914109864Sjeff 1915171482Sjeff/* 1916171482Sjeff * Schedule a thread to resume execution and record how long it voluntarily 1917171482Sjeff * slept. We also update the pctcpu, interactivity, and priority. 1918171482Sjeff */ 1919109864Sjeffvoid 1920109864Sjeffsched_wakeup(struct thread *td) 1921109864Sjeff{ 1922166229Sjeff struct td_sched *ts; 1923171482Sjeff int slptick; 1924165762Sjeff 1925170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1926166229Sjeff ts = td->td_sched; 1927109864Sjeff /* 1928165762Sjeff * If we slept for more than a tick update our interactivity and 1929165762Sjeff * priority. 1930109864Sjeff */ 1931171482Sjeff slptick = ts->ts_slptick; 1932171482Sjeff ts->ts_slptick = 0; 1933171482Sjeff if (slptick && slptick != ticks) { 1934166208Sjeff u_int hzticks; 1935109864Sjeff 1936171482Sjeff hzticks = (ticks - slptick) << SCHED_TICK_SHIFT; 1937171482Sjeff ts->ts_slptime += hzticks; 1938165819Sjeff sched_interact_update(td); 1939166229Sjeff sched_pctcpu_update(ts); 1940163709Sjb sched_priority(td); 1941109864Sjeff } 1942166229Sjeff /* Reset the slice value after we sleep. */ 1943166229Sjeff ts->ts_slice = sched_slice; 1944166190Sjeff sched_add(td, SRQ_BORING); 1945109864Sjeff} 1946109864Sjeff 1947109864Sjeff/* 1948109864Sjeff * Penalize the parent for creating a new child and initialize the child's 1949109864Sjeff * priority. 1950109864Sjeff */ 1951109864Sjeffvoid 1952163709Sjbsched_fork(struct thread *td, struct thread *child) 1953109864Sjeff{ 1954170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1955164936Sjulian sched_fork_thread(td, child); 1956165762Sjeff /* 1957165762Sjeff * Penalize the parent and child for forking. 1958165762Sjeff */ 1959165762Sjeff sched_interact_fork(child); 1960165762Sjeff sched_priority(child); 1961171482Sjeff td->td_sched->ts_runtime += tickincr; 1962165762Sjeff sched_interact_update(td); 1963165762Sjeff sched_priority(td); 1964164936Sjulian} 1965109864Sjeff 1966171482Sjeff/* 1967171482Sjeff * Fork a new thread, may be within the same process. 1968171482Sjeff */ 1969164936Sjulianvoid 1970164936Sjuliansched_fork_thread(struct thread *td, struct thread *child) 1971164936Sjulian{ 1972164936Sjulian struct td_sched *ts; 1973164936Sjulian struct td_sched *ts2; 1974164936Sjulian 1975165762Sjeff /* 1976165762Sjeff * Initialize child. 1977165762Sjeff */ 1978170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1979163709Sjb sched_newthread(child); 1980171482Sjeff child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1981164936Sjulian ts = td->td_sched; 1982164936Sjulian ts2 = child->td_sched; 1983164936Sjulian ts2->ts_cpu = ts->ts_cpu; 1984164936Sjulian ts2->ts_runq = NULL; 1985165762Sjeff /* 1986165762Sjeff * Grab our parents cpu estimation information and priority. 1987165762Sjeff */ 1988164936Sjulian ts2->ts_ticks = ts->ts_ticks; 1989164936Sjulian ts2->ts_ltick = ts->ts_ltick; 1990164936Sjulian ts2->ts_ftick = ts->ts_ftick; 1991165762Sjeff child->td_user_pri = td->td_user_pri; 1992165762Sjeff child->td_base_user_pri = td->td_base_user_pri; 1993165762Sjeff /* 1994165762Sjeff * And update interactivity score. 1995165762Sjeff */ 1996171482Sjeff ts2->ts_slptime = ts->ts_slptime; 1997171482Sjeff ts2->ts_runtime = ts->ts_runtime; 1998165762Sjeff ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 1999113357Sjeff} 2000113357Sjeff 2001171482Sjeff/* 2002171482Sjeff * Adjust the priority class of a thread. 2003171482Sjeff */ 2004113357Sjeffvoid 2005163709Sjbsched_class(struct thread *td, int class) 2006113357Sjeff{ 2007113357Sjeff 2008170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2009163709Sjb if (td->td_pri_class == class) 2010113357Sjeff return; 2011113357Sjeff 2012121896Sjeff#ifdef SMP 2013165827Sjeff /* 2014165827Sjeff * On SMP if we're on the RUNQ we must adjust the transferable 2015165827Sjeff * count because could be changing to or from an interrupt 2016165827Sjeff * class. 2017165827Sjeff */ 2018166190Sjeff if (TD_ON_RUNQ(td)) { 2019165827Sjeff struct tdq *tdq; 2020165827Sjeff 2021165827Sjeff tdq = TDQ_CPU(td->td_sched->ts_cpu); 2022165827Sjeff if (THREAD_CAN_MIGRATE(td)) { 2023165827Sjeff tdq->tdq_transferable--; 2024165827Sjeff tdq->tdq_group->tdg_transferable--; 2025122744Sjeff } 2026165827Sjeff td->td_pri_class = class; 2027165827Sjeff if (THREAD_CAN_MIGRATE(td)) { 2028165827Sjeff tdq->tdq_transferable++; 2029165827Sjeff tdq->tdq_group->tdg_transferable++; 2030165827Sjeff } 2031165827Sjeff } 2032164936Sjulian#endif 2033163709Sjb td->td_pri_class = class; 2034109864Sjeff} 2035109864Sjeff 2036109864Sjeff/* 2037109864Sjeff * Return some of the child's priority and interactivity to the parent. 2038109864Sjeff */ 2039109864Sjeffvoid 2040164939Sjuliansched_exit(struct proc *p, struct thread *child) 2041109864Sjeff{ 2042165762Sjeff struct thread *td; 2043164939Sjulian 2044163709Sjb CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d", 2045164939Sjulian child, child->td_proc->p_comm, child->td_priority); 2046113372Sjeff 2047170293Sjeff PROC_SLOCK_ASSERT(p, MA_OWNED); 2048165762Sjeff td = FIRST_THREAD_IN_PROC(p); 2049165762Sjeff sched_exit_thread(td, child); 2050113372Sjeff} 2051113372Sjeff 2052171482Sjeff/* 2053171482Sjeff * Penalize another thread for the time spent on this one. This helps to 2054171482Sjeff * worsen the priority and interactivity of processes which schedule batch 2055171482Sjeff * jobs such as make. This has little effect on the make process itself but 2056171482Sjeff * causes new processes spawned by it to receive worse scores immediately. 2057171482Sjeff */ 2058113372Sjeffvoid 2059164939Sjuliansched_exit_thread(struct thread *td, struct thread *child) 2060164936Sjulian{ 2061165762Sjeff 2062164939Sjulian CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d", 2063165762Sjeff child, child->td_proc->p_comm, child->td_priority); 2064164939Sjulian 2065165762Sjeff#ifdef KSE 2066165762Sjeff /* 2067165762Sjeff * KSE forks and exits so often that this penalty causes short-lived 2068165762Sjeff * threads to always be non-interactive. This causes mozilla to 2069165762Sjeff * crawl under load. 2070165762Sjeff */ 2071165762Sjeff if ((td->td_pflags & TDP_SA) && td->td_proc == child->td_proc) 2072165762Sjeff return; 2073165762Sjeff#endif 2074165762Sjeff /* 2075165762Sjeff * Give the child's runtime to the parent without returning the 2076165762Sjeff * sleep time as a penalty to the parent. This causes shells that 2077165762Sjeff * launch expensive things to mark their children as expensive. 2078165762Sjeff */ 2079170293Sjeff thread_lock(td); 2080171482Sjeff td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2081164939Sjulian sched_interact_update(td); 2082165762Sjeff sched_priority(td); 2083170293Sjeff thread_unlock(td); 2084164936Sjulian} 2085164936Sjulian 2086171482Sjeff/* 2087171482Sjeff * Fix priorities on return to user-space. Priorities may be elevated due 2088171482Sjeff * to static priorities in msleep() or similar. 2089171482Sjeff */ 2090164936Sjulianvoid 2091164936Sjuliansched_userret(struct thread *td) 2092164936Sjulian{ 2093164936Sjulian /* 2094164936Sjulian * XXX we cheat slightly on the locking here to avoid locking in 2095164936Sjulian * the usual case. Setting td_priority here is essentially an 2096164936Sjulian * incomplete workaround for not setting it properly elsewhere. 2097164936Sjulian * Now that some interrupt handlers are threads, not setting it 2098164936Sjulian * properly elsewhere can clobber it in the window between setting 2099164936Sjulian * it here and returning to user mode, so don't waste time setting 2100164936Sjulian * it perfectly here. 2101164936Sjulian */ 2102164936Sjulian KASSERT((td->td_flags & TDF_BORROWING) == 0, 2103164936Sjulian ("thread with borrowed priority returning to userland")); 2104164936Sjulian if (td->td_priority != td->td_user_pri) { 2105170293Sjeff thread_lock(td); 2106164936Sjulian td->td_priority = td->td_user_pri; 2107164936Sjulian td->td_base_pri = td->td_user_pri; 2108170293Sjeff thread_unlock(td); 2109164936Sjulian } 2110164936Sjulian} 2111164936Sjulian 2112171482Sjeff/* 2113171482Sjeff * Handle a stathz tick. This is really only relevant for timeshare 2114171482Sjeff * threads. 2115171482Sjeff */ 2116164936Sjulianvoid 2117121127Sjeffsched_clock(struct thread *td) 2118109864Sjeff{ 2119164936Sjulian struct tdq *tdq; 2120164936Sjulian struct td_sched *ts; 2121109864Sjeff 2122171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2123164936Sjulian tdq = TDQ_SELF(); 2124133427Sjeff /* 2125165766Sjeff * Advance the insert index once for each tick to ensure that all 2126165766Sjeff * threads get a chance to run. 2127133427Sjeff */ 2128165766Sjeff if (tdq->tdq_idx == tdq->tdq_ridx) { 2129165766Sjeff tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2130165766Sjeff if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2131165766Sjeff tdq->tdq_ridx = tdq->tdq_idx; 2132165766Sjeff } 2133165766Sjeff ts = td->td_sched; 2134165762Sjeff /* 2135163709Sjb * We only do slicing code for TIMESHARE threads. 2136113357Sjeff */ 2137163709Sjb if (td->td_pri_class != PRI_TIMESHARE) 2138113357Sjeff return; 2139113357Sjeff /* 2140165766Sjeff * We used a tick; charge it to the thread so that we can compute our 2141113357Sjeff * interactivity. 2142109864Sjeff */ 2143171482Sjeff td->td_sched->ts_runtime += tickincr; 2144163709Sjb sched_interact_update(td); 2145109864Sjeff /* 2146109864Sjeff * We used up one time slice. 2147109864Sjeff */ 2148164936Sjulian if (--ts->ts_slice > 0) 2149113357Sjeff return; 2150109864Sjeff /* 2151113357Sjeff * We're out of time, recompute priorities and requeue. 2152109864Sjeff */ 2153165796Sjeff sched_priority(td); 2154113357Sjeff td->td_flags |= TDF_NEEDRESCHED; 2155109864Sjeff} 2156109864Sjeff 2157171482Sjeff/* 2158171482Sjeff * Called once per hz tick. Used for cpu utilization information. This 2159171482Sjeff * is easier than trying to scale based on stathz. 2160171482Sjeff */ 2161171482Sjeffvoid 2162171482Sjeffsched_tick(void) 2163171482Sjeff{ 2164171482Sjeff struct td_sched *ts; 2165171482Sjeff 2166171482Sjeff ts = curthread->td_sched; 2167171482Sjeff /* Adjust ticks for pctcpu */ 2168171482Sjeff ts->ts_ticks += 1 << SCHED_TICK_SHIFT; 2169171482Sjeff ts->ts_ltick = ticks; 2170171482Sjeff /* 2171171482Sjeff * Update if we've exceeded our desired tick threshhold by over one 2172171482Sjeff * second. 2173171482Sjeff */ 2174171482Sjeff if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2175171482Sjeff sched_pctcpu_update(ts); 2176171482Sjeff} 2177171482Sjeff 2178171482Sjeff/* 2179171482Sjeff * Return whether the current CPU has runnable tasks. Used for in-kernel 2180171482Sjeff * cooperative idle threads. 2181171482Sjeff */ 2182109864Sjeffint 2183109864Sjeffsched_runnable(void) 2184109864Sjeff{ 2185164936Sjulian struct tdq *tdq; 2186115998Sjeff int load; 2187109864Sjeff 2188115998Sjeff load = 1; 2189115998Sjeff 2190164936Sjulian tdq = TDQ_SELF(); 2191121605Sjeff if ((curthread->td_flags & TDF_IDLETD) != 0) { 2192165620Sjeff if (tdq->tdq_load > 0) 2193121605Sjeff goto out; 2194121605Sjeff } else 2195165620Sjeff if (tdq->tdq_load - 1 > 0) 2196121605Sjeff goto out; 2197115998Sjeff load = 0; 2198115998Sjeffout: 2199115998Sjeff return (load); 2200109864Sjeff} 2201109864Sjeff 2202171482Sjeff/* 2203171482Sjeff * Choose the highest priority thread to run. The thread is removed from 2204171482Sjeff * the run-queue while running however the load remains. For SMP we set 2205171482Sjeff * the tdq in the global idle bitmask if it idles here. 2206171482Sjeff */ 2207166190Sjeffstruct thread * 2208109970Sjeffsched_choose(void) 2209109970Sjeff{ 2210171482Sjeff#ifdef SMP 2211171482Sjeff struct tdq_group *tdg; 2212171482Sjeff#endif 2213171482Sjeff struct td_sched *ts; 2214164936Sjulian struct tdq *tdq; 2215109970Sjeff 2216164936Sjulian tdq = TDQ_SELF(); 2217171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2218164936Sjulian ts = tdq_choose(tdq); 2219164936Sjulian if (ts) { 2220164936Sjulian tdq_runq_rem(tdq, ts); 2221166190Sjeff return (ts->ts_thread); 2222109864Sjeff } 2223109970Sjeff#ifdef SMP 2224171482Sjeff /* 2225171482Sjeff * We only set the idled bit when all of the cpus in the group are 2226171482Sjeff * idle. Otherwise we could get into a situation where a thread bounces 2227171482Sjeff * back and forth between two idle cores on seperate physical CPUs. 2228171482Sjeff */ 2229171482Sjeff tdg = tdq->tdq_group; 2230171482Sjeff tdg->tdg_idlemask |= PCPU_GET(cpumask); 2231171482Sjeff if (tdg->tdg_idlemask == tdg->tdg_cpumask) 2232171482Sjeff atomic_set_int(&tdq_idle, tdg->tdg_mask); 2233171482Sjeff tdq->tdq_lowpri = PRI_MAX_IDLE; 2234109970Sjeff#endif 2235166190Sjeff return (PCPU_GET(idlethread)); 2236109864Sjeff} 2237109864Sjeff 2238171482Sjeff/* 2239171482Sjeff * Set owepreempt if necessary. Preemption never happens directly in ULE, 2240171482Sjeff * we always request it once we exit a critical section. 2241171482Sjeff */ 2242171482Sjeffstatic inline void 2243171482Sjeffsched_setpreempt(struct thread *td) 2244166190Sjeff{ 2245166190Sjeff struct thread *ctd; 2246166190Sjeff int cpri; 2247166190Sjeff int pri; 2248166190Sjeff 2249166190Sjeff ctd = curthread; 2250166190Sjeff pri = td->td_priority; 2251166190Sjeff cpri = ctd->td_priority; 2252171482Sjeff if (td->td_priority < ctd->td_priority) 2253171482Sjeff curthread->td_flags |= TDF_NEEDRESCHED; 2254166190Sjeff if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2255171482Sjeff return; 2256166190Sjeff /* 2257166190Sjeff * Always preempt IDLE threads. Otherwise only if the preempting 2258166190Sjeff * thread is an ithread. 2259166190Sjeff */ 2260171482Sjeff if (pri > preempt_thresh && cpri < PRI_MIN_IDLE) 2261171482Sjeff return; 2262171482Sjeff ctd->td_owepreempt = 1; 2263171482Sjeff return; 2264166190Sjeff} 2265166190Sjeff 2266171482Sjeff/* 2267171482Sjeff * Add a thread to a thread queue. Initializes priority, slice, runq, and 2268171482Sjeff * add it to the appropriate queue. This is the internal function called 2269171482Sjeff * when the tdq is predetermined. 2270171482Sjeff */ 2271109864Sjeffvoid 2272171482Sjefftdq_add(struct tdq *tdq, struct thread *td, int flags) 2273109864Sjeff{ 2274164936Sjulian struct td_sched *ts; 2275121790Sjeff int class; 2276166108Sjeff#ifdef SMP 2277166108Sjeff int cpumask; 2278166108Sjeff#endif 2279109864Sjeff 2280171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2281166190Sjeff KASSERT((td->td_inhibitors == 0), 2282166190Sjeff ("sched_add: trying to run inhibited thread")); 2283166190Sjeff KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2284166190Sjeff ("sched_add: bad thread state")); 2285163709Sjb KASSERT(td->td_proc->p_sflag & PS_INMEM, 2286110267Sjeff ("sched_add: process swapped out")); 2287171482Sjeff 2288171482Sjeff ts = td->td_sched; 2289171482Sjeff class = PRI_BASE(td->td_pri_class); 2290166190Sjeff TD_SET_RUNQ(td); 2291166190Sjeff if (ts->ts_slice == 0) 2292166190Sjeff ts->ts_slice = sched_slice; 2293133427Sjeff /* 2294171482Sjeff * Pick the run queue based on priority. 2295133427Sjeff */ 2296171482Sjeff if (td->td_priority <= PRI_MAX_REALTIME) 2297171482Sjeff ts->ts_runq = &tdq->tdq_realtime; 2298171482Sjeff else if (td->td_priority <= PRI_MAX_TIMESHARE) 2299171482Sjeff ts->ts_runq = &tdq->tdq_timeshare; 2300171482Sjeff else 2301171482Sjeff ts->ts_runq = &tdq->tdq_idle; 2302171482Sjeff#ifdef SMP 2303166108Sjeff cpumask = 1 << ts->ts_cpu; 2304121790Sjeff /* 2305123685Sjeff * If we had been idle, clear our bit in the group and potentially 2306166108Sjeff * the global bitmap. 2307121790Sjeff */ 2308165762Sjeff if ((class != PRI_IDLE && class != PRI_ITHD) && 2309166108Sjeff (tdq->tdq_group->tdg_idlemask & cpumask) != 0) { 2310121790Sjeff /* 2311123433Sjeff * Check to see if our group is unidling, and if so, remove it 2312123433Sjeff * from the global idle mask. 2313121790Sjeff */ 2314165620Sjeff if (tdq->tdq_group->tdg_idlemask == 2315165620Sjeff tdq->tdq_group->tdg_cpumask) 2316165620Sjeff atomic_clear_int(&tdq_idle, tdq->tdq_group->tdg_mask); 2317123433Sjeff /* 2318123433Sjeff * Now remove ourselves from the group specific idle mask. 2319123433Sjeff */ 2320166108Sjeff tdq->tdq_group->tdg_idlemask &= ~cpumask; 2321166108Sjeff } 2322171482Sjeff if (td->td_priority < tdq->tdq_lowpri) 2323171482Sjeff tdq->tdq_lowpri = td->td_priority; 2324121790Sjeff#endif 2325171482Sjeff tdq_runq_add(tdq, ts, flags); 2326171482Sjeff tdq_load_add(tdq, ts); 2327171482Sjeff} 2328171482Sjeff 2329171482Sjeff/* 2330171482Sjeff * Select the target thread queue and add a thread to it. Request 2331171482Sjeff * preemption or IPI a remote processor if required. 2332171482Sjeff */ 2333171482Sjeffvoid 2334171482Sjeffsched_add(struct thread *td, int flags) 2335171482Sjeff{ 2336171482Sjeff struct td_sched *ts; 2337171482Sjeff struct tdq *tdq; 2338171482Sjeff#ifdef SMP 2339171482Sjeff int cpuid; 2340171482Sjeff int cpu; 2341171482Sjeff#endif 2342171482Sjeff CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 2343171482Sjeff td, td->td_proc->p_comm, td->td_priority, curthread, 2344171482Sjeff curthread->td_proc->p_comm); 2345171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2346171482Sjeff ts = td->td_sched; 2347166108Sjeff /* 2348171482Sjeff * Recalculate the priority before we select the target cpu or 2349171482Sjeff * run-queue. 2350166108Sjeff */ 2351171482Sjeff if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2352171482Sjeff sched_priority(td); 2353171482Sjeff#ifdef SMP 2354171482Sjeff cpuid = PCPU_GET(cpuid); 2355171482Sjeff /* 2356171482Sjeff * Pick the destination cpu and if it isn't ours transfer to the 2357171482Sjeff * target cpu. 2358171482Sjeff */ 2359171482Sjeff if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_MIGRATE(td)) 2360171482Sjeff cpu = cpuid; 2361171482Sjeff else if (!THREAD_CAN_MIGRATE(td)) 2362171482Sjeff cpu = ts->ts_cpu; 2363166108Sjeff else 2364171482Sjeff cpu = sched_pickcpu(ts, flags); 2365171482Sjeff tdq = sched_setcpu(ts, cpu, flags); 2366171482Sjeff tdq_add(tdq, td, flags); 2367171482Sjeff if (cpu != cpuid) { 2368166108Sjeff tdq_notify(ts); 2369166108Sjeff return; 2370166108Sjeff } 2371171482Sjeff#else 2372171482Sjeff tdq = TDQ_SELF(); 2373171482Sjeff TDQ_LOCK(tdq); 2374171482Sjeff /* 2375171482Sjeff * Now that the thread is moving to the run-queue, set the lock 2376171482Sjeff * to the scheduler's lock. 2377171482Sjeff */ 2378171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2379171482Sjeff tdq_add(tdq, td, flags); 2380166108Sjeff#endif 2381171482Sjeff if (!(flags & SRQ_YIELDING)) 2382171482Sjeff sched_setpreempt(td); 2383109864Sjeff} 2384109864Sjeff 2385171482Sjeff/* 2386171482Sjeff * Remove a thread from a run-queue without running it. This is used 2387171482Sjeff * when we're stealing a thread from a remote queue. Otherwise all threads 2388171482Sjeff * exit by calling sched_exit_thread() and sched_throw() themselves. 2389171482Sjeff */ 2390109864Sjeffvoid 2391121127Sjeffsched_rem(struct thread *td) 2392109864Sjeff{ 2393164936Sjulian struct tdq *tdq; 2394164936Sjulian struct td_sched *ts; 2395113357Sjeff 2396139316Sjeff CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 2397139316Sjeff td, td->td_proc->p_comm, td->td_priority, curthread, 2398139316Sjeff curthread->td_proc->p_comm); 2399164936Sjulian ts = td->td_sched; 2400171482Sjeff tdq = TDQ_CPU(ts->ts_cpu); 2401171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2402171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2403166190Sjeff KASSERT(TD_ON_RUNQ(td), 2404164936Sjulian ("sched_rem: thread not on run queue")); 2405164936Sjulian tdq_runq_rem(tdq, ts); 2406164936Sjulian tdq_load_rem(tdq, ts); 2407166190Sjeff TD_SET_CAN_RUN(td); 2408109864Sjeff} 2409109864Sjeff 2410171482Sjeff/* 2411171482Sjeff * Fetch cpu utilization information. Updates on demand. 2412171482Sjeff */ 2413109864Sjefffixpt_t 2414121127Sjeffsched_pctcpu(struct thread *td) 2415109864Sjeff{ 2416109864Sjeff fixpt_t pctcpu; 2417164936Sjulian struct td_sched *ts; 2418109864Sjeff 2419109864Sjeff pctcpu = 0; 2420164936Sjulian ts = td->td_sched; 2421164936Sjulian if (ts == NULL) 2422121290Sjeff return (0); 2423109864Sjeff 2424170293Sjeff thread_lock(td); 2425164936Sjulian if (ts->ts_ticks) { 2426109864Sjeff int rtick; 2427109864Sjeff 2428165796Sjeff sched_pctcpu_update(ts); 2429109864Sjeff /* How many rtick per second ? */ 2430165762Sjeff rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2431165762Sjeff pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2432109864Sjeff } 2433164936Sjulian td->td_proc->p_swtime = ts->ts_ltick - ts->ts_ftick; 2434170293Sjeff thread_unlock(td); 2435109864Sjeff 2436109864Sjeff return (pctcpu); 2437109864Sjeff} 2438109864Sjeff 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); 2494163709Sjb if (td->td_pri_class == PRI_TIMESHARE) 2495159630Sdavidxu sched_prio(td, PRI_MAX_TIMESHARE); 2496170293Sjeff SCHED_STAT_INC(switch_relinquish); 2497159630Sdavidxu mi_switch(SW_VOL, NULL); 2498170293Sjeff thread_unlock(td); 2499159630Sdavidxu} 2500159630Sdavidxu 2501171482Sjeff/* 2502171482Sjeff * Return the total system load. 2503171482Sjeff */ 2504145256Sjkoshyint 2505125289Sjeffsched_load(void) 2506125289Sjeff{ 2507125289Sjeff#ifdef SMP 2508125289Sjeff int total; 2509125289Sjeff int i; 2510125289Sjeff 2511125289Sjeff total = 0; 2512165620Sjeff for (i = 0; i <= tdg_maxid; i++) 2513165620Sjeff total += TDQ_GROUP(i)->tdg_load; 2514125289Sjeff return (total); 2515125289Sjeff#else 2516165620Sjeff return (TDQ_SELF()->tdq_sysload); 2517125289Sjeff#endif 2518125289Sjeff} 2519125289Sjeff 2520125289Sjeffint 2521109864Sjeffsched_sizeof_proc(void) 2522109864Sjeff{ 2523109864Sjeff return (sizeof(struct proc)); 2524109864Sjeff} 2525109864Sjeff 2526109864Sjeffint 2527109864Sjeffsched_sizeof_thread(void) 2528109864Sjeff{ 2529109864Sjeff return (sizeof(struct thread) + sizeof(struct td_sched)); 2530109864Sjeff} 2531159570Sdavidxu 2532166190Sjeff/* 2533166190Sjeff * The actual idle process. 2534166190Sjeff */ 2535166190Sjeffvoid 2536166190Sjeffsched_idletd(void *dummy) 2537166190Sjeff{ 2538166190Sjeff struct thread *td; 2539171482Sjeff struct tdq *tdq; 2540166190Sjeff 2541166190Sjeff td = curthread; 2542171482Sjeff tdq = TDQ_SELF(); 2543166190Sjeff mtx_assert(&Giant, MA_NOTOWNED); 2544171482Sjeff /* ULE relies on preemption for idle interruption. */ 2545171482Sjeff for (;;) { 2546171482Sjeff#ifdef SMP 2547171482Sjeff if (tdq_idled(tdq)) 2548171482Sjeff cpu_idle(); 2549171482Sjeff#else 2550166190Sjeff cpu_idle(); 2551171482Sjeff#endif 2552171482Sjeff } 2553166190Sjeff} 2554166190Sjeff 2555170293Sjeff/* 2556170293Sjeff * A CPU is entering for the first time or a thread is exiting. 2557170293Sjeff */ 2558170293Sjeffvoid 2559170293Sjeffsched_throw(struct thread *td) 2560170293Sjeff{ 2561171482Sjeff struct tdq *tdq; 2562171482Sjeff 2563171482Sjeff tdq = TDQ_SELF(); 2564170293Sjeff if (td == NULL) { 2565171482Sjeff /* Correct spinlock nesting and acquire the correct lock. */ 2566171482Sjeff TDQ_LOCK(tdq); 2567170293Sjeff spinlock_exit(); 2568170293Sjeff } else { 2569171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2570171482Sjeff tdq_load_rem(tdq, td->td_sched); 2571170293Sjeff } 2572170293Sjeff KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2573170293Sjeff PCPU_SET(switchtime, cpu_ticks()); 2574170293Sjeff PCPU_SET(switchticks, ticks); 2575170293Sjeff cpu_throw(td, choosethread()); /* doesn't return */ 2576170293Sjeff} 2577170293Sjeff 2578171482Sjeff/* 2579171482Sjeff * This is called from fork_exit(). Just acquire the correct locks and 2580171482Sjeff * let fork do the rest of the work. 2581171482Sjeff */ 2582170293Sjeffvoid 2583170600Sjeffsched_fork_exit(struct thread *td) 2584170293Sjeff{ 2585171482Sjeff struct td_sched *ts; 2586171482Sjeff struct tdq *tdq; 2587171482Sjeff int cpuid; 2588170293Sjeff 2589170293Sjeff /* 2590170293Sjeff * Finish setting up thread glue so that it begins execution in a 2591171482Sjeff * non-nested critical section with the scheduler lock held. 2592170293Sjeff */ 2593171482Sjeff cpuid = PCPU_GET(cpuid); 2594171482Sjeff tdq = TDQ_CPU(cpuid); 2595171482Sjeff ts = td->td_sched; 2596171482Sjeff if (TD_IS_IDLETHREAD(td)) 2597171482Sjeff td->td_lock = TDQ_LOCKPTR(tdq); 2598171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2599171482Sjeff td->td_oncpu = cpuid; 2600171482Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)td; 2601170600Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED); 2602170293Sjeff} 2603170293Sjeff 2604171482Sjeffstatic SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, 2605171482Sjeff "Scheduler"); 2606171482SjeffSYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2607165762Sjeff "Scheduler name"); 2608171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2609171482Sjeff "Slice size for timeshare threads"); 2610171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2611171482Sjeff "Interactivity score threshold"); 2612171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2613171482Sjeff 0,"Min priority for preemption, lower priorities have greater precedence"); 2614166108Sjeff#ifdef SMP 2615171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, pick_pri, CTLFLAG_RW, &pick_pri, 0, 2616171482Sjeff "Pick the target cpu based on priority rather than load."); 2617171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2618171482Sjeff "Number of hz ticks to keep thread affinity for"); 2619171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, tryself, CTLFLAG_RW, &tryself, 0, ""); 2620171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2621171482Sjeff "Enables the long-term load balancer"); 2622171506SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance_secs, CTLFLAG_RW, &balance_secs, 0, 2623171506Sjeff "Average frequence in seconds to run the long-term balancer"); 2624171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, 2625171482Sjeff "Steals work from another hyper-threaded core on idle"); 2626171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2627171482Sjeff "Attempts to steal work from other cores before idling"); 2628171506SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2629171506Sjeff "Minimum load on remote cpu before we'll steal"); 2630171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, topology, CTLFLAG_RD, &topology, 0, 2631171482Sjeff "True when a topology has been specified by the MD code."); 2632166108Sjeff#endif 2633165762Sjeff 2634165762Sjeff/* ps compat */ 2635165762Sjeffstatic fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 2636165762SjeffSYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2637165762Sjeff 2638165762Sjeff 2639134791Sjulian#define KERN_SWITCH_INCLUDE 1 2640134791Sjulian#include "kern/kern_switch.c" 2641