sched_ule.c revision 172207
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 172207 2007-09-17 05:31:39Z 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); 300171482Sjeffstatic inline struct mtx *thread_block_switch(struct thread *); 301171482Sjeffstatic inline void thread_unblock_switch(struct thread *, struct mtx *); 302171713Sjeffstatic struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 303165827Sjeff 304166108Sjeff#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 305121790Sjeff#endif 306110028Sjeff 307165762Sjeffstatic void sched_setup(void *dummy); 308165762SjeffSYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL) 309165762Sjeff 310165762Sjeffstatic void sched_initticks(void *dummy); 311165762SjeffSYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, NULL) 312165762Sjeff 313171482Sjeff/* 314171482Sjeff * Print the threads waiting on a run-queue. 315171482Sjeff */ 316165762Sjeffstatic void 317165762Sjeffrunq_print(struct runq *rq) 318165762Sjeff{ 319165762Sjeff struct rqhead *rqh; 320165762Sjeff struct td_sched *ts; 321165762Sjeff int pri; 322165762Sjeff int j; 323165762Sjeff int i; 324165762Sjeff 325165762Sjeff for (i = 0; i < RQB_LEN; i++) { 326165762Sjeff printf("\t\trunq bits %d 0x%zx\n", 327165762Sjeff i, rq->rq_status.rqb_bits[i]); 328165762Sjeff for (j = 0; j < RQB_BPW; j++) 329165762Sjeff if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 330165762Sjeff pri = j + (i << RQB_L2BPW); 331165762Sjeff rqh = &rq->rq_queues[pri]; 332165762Sjeff TAILQ_FOREACH(ts, rqh, ts_procq) { 333165762Sjeff printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 334165762Sjeff ts->ts_thread, ts->ts_thread->td_proc->p_comm, ts->ts_thread->td_priority, ts->ts_rqindex, pri); 335165762Sjeff } 336165762Sjeff } 337165762Sjeff } 338165762Sjeff} 339165762Sjeff 340171482Sjeff/* 341171482Sjeff * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 342171482Sjeff */ 343113357Sjeffvoid 344164936Sjuliantdq_print(int cpu) 345110267Sjeff{ 346164936Sjulian struct tdq *tdq; 347112994Sjeff 348164936Sjulian tdq = TDQ_CPU(cpu); 349112994Sjeff 350171713Sjeff printf("tdq %d:\n", TDQ_ID(tdq)); 351171482Sjeff printf("\tlockptr %p\n", TDQ_LOCKPTR(tdq)); 352165620Sjeff printf("\tload: %d\n", tdq->tdq_load); 353171482Sjeff printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 354165766Sjeff printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 355165762Sjeff printf("\trealtime runq:\n"); 356165762Sjeff runq_print(&tdq->tdq_realtime); 357165762Sjeff printf("\ttimeshare runq:\n"); 358165762Sjeff runq_print(&tdq->tdq_timeshare); 359165762Sjeff printf("\tidle runq:\n"); 360165762Sjeff runq_print(&tdq->tdq_idle); 361121896Sjeff#ifdef SMP 362165620Sjeff printf("\tload transferable: %d\n", tdq->tdq_transferable); 363171713Sjeff printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 364171713Sjeff printf("\tgroup: %d\n", TDG_ID(tdq->tdq_group)); 365171713Sjeff printf("\tLock name: %s\n", tdq->tdq_group->tdg_name); 366121896Sjeff#endif 367113357Sjeff} 368112994Sjeff 369171482Sjeff#define TS_RQ_PPQ (((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS) 370171482Sjeff/* 371171482Sjeff * Add a thread to the actual run-queue. Keeps transferable counts up to 372171482Sjeff * date with what is actually on the run-queue. Selects the correct 373171482Sjeff * queue position for timeshare threads. 374171482Sjeff */ 375122744Sjeffstatic __inline void 376164936Sjuliantdq_runq_add(struct tdq *tdq, struct td_sched *ts, int flags) 377122744Sjeff{ 378171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 379171482Sjeff THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 380122744Sjeff#ifdef SMP 381165762Sjeff if (THREAD_CAN_MIGRATE(ts->ts_thread)) { 382165620Sjeff tdq->tdq_transferable++; 383165620Sjeff tdq->tdq_group->tdg_transferable++; 384164936Sjulian ts->ts_flags |= TSF_XFERABLE; 385123433Sjeff } 386122744Sjeff#endif 387165762Sjeff if (ts->ts_runq == &tdq->tdq_timeshare) { 388166557Sjeff u_char pri; 389165762Sjeff 390165762Sjeff pri = ts->ts_thread->td_priority; 391165762Sjeff KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE, 392165762Sjeff ("Invalid priority %d on timeshare runq", pri)); 393165762Sjeff /* 394165762Sjeff * This queue contains only priorities between MIN and MAX 395165762Sjeff * realtime. Use the whole queue to represent these values. 396165762Sjeff */ 397171713Sjeff if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 398165762Sjeff pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ; 399165762Sjeff pri = (pri + tdq->tdq_idx) % RQ_NQS; 400165766Sjeff /* 401165766Sjeff * This effectively shortens the queue by one so we 402165766Sjeff * can have a one slot difference between idx and 403165766Sjeff * ridx while we wait for threads to drain. 404165766Sjeff */ 405165766Sjeff if (tdq->tdq_ridx != tdq->tdq_idx && 406165766Sjeff pri == tdq->tdq_ridx) 407167664Sjeff pri = (unsigned char)(pri - 1) % RQ_NQS; 408165762Sjeff } else 409165766Sjeff pri = tdq->tdq_ridx; 410165762Sjeff runq_add_pri(ts->ts_runq, ts, pri, flags); 411165762Sjeff } else 412165762Sjeff runq_add(ts->ts_runq, ts, flags); 413122744Sjeff} 414122744Sjeff 415171482Sjeff/* 416171482Sjeff * Remove a thread from a run-queue. This typically happens when a thread 417171482Sjeff * is selected to run. Running threads are not on the queue and the 418171482Sjeff * transferable count does not reflect them. 419171482Sjeff */ 420122744Sjeffstatic __inline void 421164936Sjuliantdq_runq_rem(struct tdq *tdq, struct td_sched *ts) 422122744Sjeff{ 423171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 424171482Sjeff KASSERT(ts->ts_runq != NULL, 425171482Sjeff ("tdq_runq_remove: thread %p null ts_runq", ts->ts_thread)); 426122744Sjeff#ifdef SMP 427164936Sjulian if (ts->ts_flags & TSF_XFERABLE) { 428165620Sjeff tdq->tdq_transferable--; 429165620Sjeff tdq->tdq_group->tdg_transferable--; 430164936Sjulian ts->ts_flags &= ~TSF_XFERABLE; 431123433Sjeff } 432122744Sjeff#endif 433165766Sjeff if (ts->ts_runq == &tdq->tdq_timeshare) { 434165766Sjeff if (tdq->tdq_idx != tdq->tdq_ridx) 435165766Sjeff runq_remove_idx(ts->ts_runq, ts, &tdq->tdq_ridx); 436165766Sjeff else 437165766Sjeff runq_remove_idx(ts->ts_runq, ts, NULL); 438165796Sjeff /* 439165796Sjeff * For timeshare threads we update the priority here so 440165796Sjeff * the priority reflects the time we've been sleeping. 441165796Sjeff */ 442165796Sjeff ts->ts_ltick = ticks; 443165796Sjeff sched_pctcpu_update(ts); 444165796Sjeff sched_priority(ts->ts_thread); 445165766Sjeff } else 446165762Sjeff runq_remove(ts->ts_runq, ts); 447122744Sjeff} 448122744Sjeff 449171482Sjeff/* 450171482Sjeff * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 451171482Sjeff * for this thread to the referenced thread queue. 452171482Sjeff */ 453113357Sjeffstatic void 454164936Sjuliantdq_load_add(struct tdq *tdq, struct td_sched *ts) 455113357Sjeff{ 456121896Sjeff int class; 457171482Sjeff 458171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 459171482Sjeff THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 460164936Sjulian class = PRI_BASE(ts->ts_thread->td_pri_class); 461165620Sjeff tdq->tdq_load++; 462171713Sjeff CTR2(KTR_SCHED, "cpu %d load: %d", TDQ_ID(tdq), tdq->tdq_load); 463166108Sjeff if (class != PRI_ITHD && 464166108Sjeff (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 465123487Sjeff#ifdef SMP 466165620Sjeff tdq->tdq_group->tdg_load++; 467125289Sjeff#else 468165620Sjeff tdq->tdq_sysload++; 469123487Sjeff#endif 470110267Sjeff} 471113357Sjeff 472171482Sjeff/* 473171482Sjeff * Remove the load from a thread that is transitioning to a sleep state or 474171482Sjeff * exiting. 475171482Sjeff */ 476112994Sjeffstatic void 477164936Sjuliantdq_load_rem(struct tdq *tdq, struct td_sched *ts) 478110267Sjeff{ 479121896Sjeff int class; 480171482Sjeff 481171482Sjeff THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 482171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 483164936Sjulian class = PRI_BASE(ts->ts_thread->td_pri_class); 484166108Sjeff if (class != PRI_ITHD && 485166108Sjeff (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 486123487Sjeff#ifdef SMP 487165620Sjeff tdq->tdq_group->tdg_load--; 488125289Sjeff#else 489165620Sjeff tdq->tdq_sysload--; 490123487Sjeff#endif 491171482Sjeff KASSERT(tdq->tdq_load != 0, 492171713Sjeff ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 493165620Sjeff tdq->tdq_load--; 494165620Sjeff CTR1(KTR_SCHED, "load: %d", tdq->tdq_load); 495164936Sjulian ts->ts_runq = NULL; 496110267Sjeff} 497110267Sjeff 498113357Sjeff#ifdef SMP 499116069Sjeff/* 500122744Sjeff * sched_balance is a simple CPU load balancing algorithm. It operates by 501116069Sjeff * finding the least loaded and most loaded cpu and equalizing their load 502116069Sjeff * by migrating some processes. 503116069Sjeff * 504116069Sjeff * Dealing only with two CPUs at a time has two advantages. Firstly, most 505116069Sjeff * installations will only have 2 cpus. Secondly, load balancing too much at 506116069Sjeff * once can have an unpleasant effect on the system. The scheduler rarely has 507116069Sjeff * enough information to make perfect decisions. So this algorithm chooses 508171482Sjeff * simplicity and more gradual effects on load in larger systems. 509116069Sjeff * 510116069Sjeff */ 511121790Sjeffstatic void 512171482Sjeffsched_balance(void *arg) 513116069Sjeff{ 514164936Sjulian struct tdq_group *high; 515164936Sjulian struct tdq_group *low; 516165620Sjeff struct tdq_group *tdg; 517123487Sjeff int cnt; 518123487Sjeff int i; 519123487Sjeff 520171506Sjeff callout_reset(&balco, max(hz / 2, random() % (hz * balance_secs)), 521171482Sjeff sched_balance, NULL); 522171482Sjeff if (smp_started == 0 || rebalance == 0) 523139334Sjeff return; 524123487Sjeff low = high = NULL; 525165620Sjeff i = random() % (tdg_maxid + 1); 526165620Sjeff for (cnt = 0; cnt <= tdg_maxid; cnt++) { 527165620Sjeff tdg = TDQ_GROUP(i); 528123487Sjeff /* 529123487Sjeff * Find the CPU with the highest load that has some 530123487Sjeff * threads to transfer. 531123487Sjeff */ 532165620Sjeff if ((high == NULL || tdg->tdg_load > high->tdg_load) 533165620Sjeff && tdg->tdg_transferable) 534165620Sjeff high = tdg; 535165620Sjeff if (low == NULL || tdg->tdg_load < low->tdg_load) 536165620Sjeff low = tdg; 537165620Sjeff if (++i > tdg_maxid) 538123487Sjeff i = 0; 539123487Sjeff } 540123487Sjeff if (low != NULL && high != NULL && high != low) 541165620Sjeff sched_balance_pair(LIST_FIRST(&high->tdg_members), 542165620Sjeff LIST_FIRST(&low->tdg_members)); 543123487Sjeff} 544123487Sjeff 545171482Sjeff/* 546171482Sjeff * Balance load between CPUs in a group. Will only migrate within the group. 547171482Sjeff */ 548123487Sjeffstatic void 549171482Sjeffsched_balance_groups(void *arg) 550123487Sjeff{ 551123487Sjeff int i; 552123487Sjeff 553171506Sjeff callout_reset(&gbalco, max(hz / 2, random() % (hz * balance_secs)), 554171482Sjeff sched_balance_groups, NULL); 555171482Sjeff if (smp_started == 0 || rebalance == 0) 556171482Sjeff return; 557171482Sjeff for (i = 0; i <= tdg_maxid; i++) 558171482Sjeff sched_balance_group(TDQ_GROUP(i)); 559123487Sjeff} 560123487Sjeff 561171482Sjeff/* 562171482Sjeff * Finds the greatest imbalance between two tdqs in a group. 563171482Sjeff */ 564123487Sjeffstatic void 565165620Sjeffsched_balance_group(struct tdq_group *tdg) 566123487Sjeff{ 567164936Sjulian struct tdq *tdq; 568164936Sjulian struct tdq *high; 569164936Sjulian struct tdq *low; 570123487Sjeff int load; 571123487Sjeff 572165620Sjeff if (tdg->tdg_transferable == 0) 573123487Sjeff return; 574123487Sjeff low = NULL; 575123487Sjeff high = NULL; 576165620Sjeff LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) { 577165620Sjeff load = tdq->tdq_load; 578165620Sjeff if (high == NULL || load > high->tdq_load) 579164936Sjulian high = tdq; 580165620Sjeff if (low == NULL || load < low->tdq_load) 581164936Sjulian low = tdq; 582123487Sjeff } 583123487Sjeff if (high != NULL && low != NULL && high != low) 584123487Sjeff sched_balance_pair(high, low); 585123487Sjeff} 586123487Sjeff 587171482Sjeff/* 588171482Sjeff * Lock two thread queues using their address to maintain lock order. 589171482Sjeff */ 590123487Sjeffstatic void 591171482Sjefftdq_lock_pair(struct tdq *one, struct tdq *two) 592171482Sjeff{ 593171482Sjeff if (one < two) { 594171482Sjeff TDQ_LOCK(one); 595171482Sjeff TDQ_LOCK_FLAGS(two, MTX_DUPOK); 596171482Sjeff } else { 597171482Sjeff TDQ_LOCK(two); 598171482Sjeff TDQ_LOCK_FLAGS(one, MTX_DUPOK); 599171482Sjeff } 600171482Sjeff} 601171482Sjeff 602171482Sjeff/* 603171482Sjeff * Transfer load between two imbalanced thread queues. 604171482Sjeff */ 605171482Sjeffstatic void 606164936Sjuliansched_balance_pair(struct tdq *high, struct tdq *low) 607123487Sjeff{ 608123433Sjeff int transferable; 609116069Sjeff int high_load; 610116069Sjeff int low_load; 611116069Sjeff int move; 612116069Sjeff int diff; 613116069Sjeff int i; 614116069Sjeff 615171482Sjeff tdq_lock_pair(high, low); 616116069Sjeff /* 617123433Sjeff * If we're transfering within a group we have to use this specific 618164936Sjulian * tdq's transferable count, otherwise we can steal from other members 619123433Sjeff * of the group. 620123433Sjeff */ 621165620Sjeff if (high->tdq_group == low->tdq_group) { 622165620Sjeff transferable = high->tdq_transferable; 623165620Sjeff high_load = high->tdq_load; 624165620Sjeff low_load = low->tdq_load; 625123487Sjeff } else { 626165620Sjeff transferable = high->tdq_group->tdg_transferable; 627165620Sjeff high_load = high->tdq_group->tdg_load; 628165620Sjeff low_load = low->tdq_group->tdg_load; 629123487Sjeff } 630123433Sjeff /* 631122744Sjeff * Determine what the imbalance is and then adjust that to how many 632165620Sjeff * threads we actually have to give up (transferable). 633122744Sjeff */ 634171482Sjeff if (transferable != 0) { 635171482Sjeff diff = high_load - low_load; 636171482Sjeff move = diff / 2; 637171482Sjeff if (diff & 0x1) 638171482Sjeff move++; 639171482Sjeff move = min(move, transferable); 640171482Sjeff for (i = 0; i < move; i++) 641171482Sjeff tdq_move(high, low); 642171482Sjeff } 643171482Sjeff TDQ_UNLOCK(high); 644171482Sjeff TDQ_UNLOCK(low); 645116069Sjeff return; 646116069Sjeff} 647116069Sjeff 648171482Sjeff/* 649171482Sjeff * Move a thread from one thread queue to another. 650171482Sjeff */ 651121790Sjeffstatic void 652171482Sjefftdq_move(struct tdq *from, struct tdq *to) 653116069Sjeff{ 654171482Sjeff struct td_sched *ts; 655171482Sjeff struct thread *td; 656164936Sjulian struct tdq *tdq; 657171482Sjeff int cpu; 658116069Sjeff 659164936Sjulian tdq = from; 660171482Sjeff cpu = TDQ_ID(to); 661164936Sjulian ts = tdq_steal(tdq, 1); 662164936Sjulian if (ts == NULL) { 663165620Sjeff struct tdq_group *tdg; 664123433Sjeff 665165620Sjeff tdg = tdq->tdq_group; 666165620Sjeff LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) { 667165620Sjeff if (tdq == from || tdq->tdq_transferable == 0) 668123433Sjeff continue; 669164936Sjulian ts = tdq_steal(tdq, 1); 670123433Sjeff break; 671123433Sjeff } 672164936Sjulian if (ts == NULL) 673171482Sjeff return; 674123433Sjeff } 675164936Sjulian if (tdq == to) 676123433Sjeff return; 677171482Sjeff td = ts->ts_thread; 678171482Sjeff /* 679171482Sjeff * Although the run queue is locked the thread may be blocked. Lock 680171482Sjeff * it to clear this. 681171482Sjeff */ 682171482Sjeff thread_lock(td); 683171482Sjeff /* Drop recursive lock on from. */ 684171482Sjeff TDQ_UNLOCK(from); 685171482Sjeff sched_rem(td); 686166108Sjeff ts->ts_cpu = cpu; 687171482Sjeff td->td_lock = TDQ_LOCKPTR(to); 688171482Sjeff tdq_add(to, td, SRQ_YIELDING); 689171505Sjeff tdq_notify(ts); 690116069Sjeff} 691110267Sjeff 692171482Sjeff/* 693171482Sjeff * This tdq has idled. Try to steal a thread from another cpu and switch 694171482Sjeff * to it. 695171482Sjeff */ 696123433Sjeffstatic int 697164936Sjuliantdq_idled(struct tdq *tdq) 698121790Sjeff{ 699165620Sjeff struct tdq_group *tdg; 700164936Sjulian struct tdq *steal; 701164936Sjulian struct td_sched *ts; 702171482Sjeff struct thread *td; 703171482Sjeff int highload; 704171482Sjeff int highcpu; 705171482Sjeff int load; 706171482Sjeff int cpu; 707123433Sjeff 708171482Sjeff /* We don't want to be preempted while we're iterating over tdqs */ 709171482Sjeff spinlock_enter(); 710165620Sjeff tdg = tdq->tdq_group; 711123433Sjeff /* 712165620Sjeff * If we're in a cpu group, try and steal threads from another cpu in 713123433Sjeff * the group before idling. 714123433Sjeff */ 715166108Sjeff if (steal_htt && tdg->tdg_cpus > 1 && tdg->tdg_transferable) { 716165620Sjeff LIST_FOREACH(steal, &tdg->tdg_members, tdq_siblings) { 717165620Sjeff if (steal == tdq || steal->tdq_transferable == 0) 718123433Sjeff continue; 719171482Sjeff TDQ_LOCK(steal); 720164936Sjulian ts = tdq_steal(steal, 0); 721166108Sjeff if (ts) 722166108Sjeff goto steal; 723171482Sjeff TDQ_UNLOCK(steal); 724166108Sjeff } 725166108Sjeff } 726171482Sjeff for (;;) { 727171482Sjeff if (steal_idle == 0) 728171482Sjeff break; 729171482Sjeff highcpu = 0; 730171482Sjeff highload = 0; 731171482Sjeff for (cpu = 0; cpu <= mp_maxid; cpu++) { 732171482Sjeff if (CPU_ABSENT(cpu)) 733171482Sjeff continue; 734166108Sjeff steal = TDQ_CPU(cpu); 735171482Sjeff load = TDQ_CPU(cpu)->tdq_transferable; 736171482Sjeff if (load < highload) 737166108Sjeff continue; 738171482Sjeff highload = load; 739171482Sjeff highcpu = cpu; 740171482Sjeff } 741171506Sjeff if (highload < steal_thresh) 742171482Sjeff break; 743171482Sjeff steal = TDQ_CPU(highcpu); 744171482Sjeff TDQ_LOCK(steal); 745171506Sjeff if (steal->tdq_transferable >= steal_thresh && 746171482Sjeff (ts = tdq_steal(steal, 1)) != NULL) 747166108Sjeff goto steal; 748171482Sjeff TDQ_UNLOCK(steal); 749171482Sjeff break; 750123433Sjeff } 751171482Sjeff spinlock_exit(); 752123433Sjeff return (1); 753166108Sjeffsteal: 754171482Sjeff td = ts->ts_thread; 755171482Sjeff thread_lock(td); 756171482Sjeff spinlock_exit(); 757171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(steal)); 758171482Sjeff TDQ_UNLOCK(steal); 759171482Sjeff sched_rem(td); 760171482Sjeff sched_setcpu(ts, PCPU_GET(cpuid), SRQ_YIELDING); 761171482Sjeff tdq_add(tdq, td, SRQ_YIELDING); 762171482Sjeff MPASS(td->td_lock == curthread->td_lock); 763171482Sjeff mi_switch(SW_VOL, NULL); 764171482Sjeff thread_unlock(curthread); 765121790Sjeff 766166108Sjeff return (0); 767121790Sjeff} 768121790Sjeff 769171482Sjeff/* 770171482Sjeff * Notify a remote cpu of new work. Sends an IPI if criteria are met. 771171482Sjeff */ 772121790Sjeffstatic void 773166108Sjefftdq_notify(struct td_sched *ts) 774121790Sjeff{ 775166247Sjeff struct thread *ctd; 776121790Sjeff struct pcpu *pcpu; 777166247Sjeff int cpri; 778166247Sjeff int pri; 779166108Sjeff int cpu; 780121790Sjeff 781166108Sjeff cpu = ts->ts_cpu; 782166247Sjeff pri = ts->ts_thread->td_priority; 783166108Sjeff pcpu = pcpu_find(cpu); 784166247Sjeff ctd = pcpu->pc_curthread; 785166247Sjeff cpri = ctd->td_priority; 786166137Sjeff 787121790Sjeff /* 788166137Sjeff * If our priority is not better than the current priority there is 789166137Sjeff * nothing to do. 790166137Sjeff */ 791166247Sjeff if (pri > cpri) 792166137Sjeff return; 793166137Sjeff /* 794166247Sjeff * Always IPI idle. 795121790Sjeff */ 796166247Sjeff if (cpri > PRI_MIN_IDLE) 797166247Sjeff goto sendipi; 798166247Sjeff /* 799166247Sjeff * If we're realtime or better and there is timeshare or worse running 800166247Sjeff * send an IPI. 801166247Sjeff */ 802166247Sjeff if (pri < PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME) 803166247Sjeff goto sendipi; 804166247Sjeff /* 805166247Sjeff * Otherwise only IPI if we exceed the threshold. 806166247Sjeff */ 807171482Sjeff if (pri > preempt_thresh) 808165819Sjeff return; 809166247Sjeffsendipi: 810166247Sjeff ctd->td_flags |= TDF_NEEDRESCHED; 811171482Sjeff ipi_selected(1 << cpu, IPI_PREEMPT); 812121790Sjeff} 813121790Sjeff 814171482Sjeff/* 815171482Sjeff * Steals load from a timeshare queue. Honors the rotating queue head 816171482Sjeff * index. 817171482Sjeff */ 818164936Sjulianstatic struct td_sched * 819171482Sjeffrunq_steal_from(struct runq *rq, u_char start) 820171482Sjeff{ 821171482Sjeff struct td_sched *ts; 822171482Sjeff struct rqbits *rqb; 823171482Sjeff struct rqhead *rqh; 824171482Sjeff int first; 825171482Sjeff int bit; 826171482Sjeff int pri; 827171482Sjeff int i; 828171482Sjeff 829171482Sjeff rqb = &rq->rq_status; 830171482Sjeff bit = start & (RQB_BPW -1); 831171482Sjeff pri = 0; 832171482Sjeff first = 0; 833171482Sjeffagain: 834171482Sjeff for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 835171482Sjeff if (rqb->rqb_bits[i] == 0) 836171482Sjeff continue; 837171482Sjeff if (bit != 0) { 838171482Sjeff for (pri = bit; pri < RQB_BPW; pri++) 839171482Sjeff if (rqb->rqb_bits[i] & (1ul << pri)) 840171482Sjeff break; 841171482Sjeff if (pri >= RQB_BPW) 842171482Sjeff continue; 843171482Sjeff } else 844171482Sjeff pri = RQB_FFS(rqb->rqb_bits[i]); 845171482Sjeff pri += (i << RQB_L2BPW); 846171482Sjeff rqh = &rq->rq_queues[pri]; 847171482Sjeff TAILQ_FOREACH(ts, rqh, ts_procq) { 848171482Sjeff if (first && THREAD_CAN_MIGRATE(ts->ts_thread)) 849171482Sjeff return (ts); 850171482Sjeff first = 1; 851171482Sjeff } 852171482Sjeff } 853171482Sjeff if (start != 0) { 854171482Sjeff start = 0; 855171482Sjeff goto again; 856171482Sjeff } 857171482Sjeff 858171482Sjeff return (NULL); 859171482Sjeff} 860171482Sjeff 861171482Sjeff/* 862171482Sjeff * Steals load from a standard linear queue. 863171482Sjeff */ 864171482Sjeffstatic struct td_sched * 865121790Sjeffrunq_steal(struct runq *rq) 866121790Sjeff{ 867121790Sjeff struct rqhead *rqh; 868121790Sjeff struct rqbits *rqb; 869164936Sjulian struct td_sched *ts; 870121790Sjeff int word; 871121790Sjeff int bit; 872121790Sjeff 873121790Sjeff rqb = &rq->rq_status; 874121790Sjeff for (word = 0; word < RQB_LEN; word++) { 875121790Sjeff if (rqb->rqb_bits[word] == 0) 876121790Sjeff continue; 877121790Sjeff for (bit = 0; bit < RQB_BPW; bit++) { 878123231Speter if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 879121790Sjeff continue; 880121790Sjeff rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 881171506Sjeff TAILQ_FOREACH(ts, rqh, ts_procq) 882171506Sjeff if (THREAD_CAN_MIGRATE(ts->ts_thread)) 883164936Sjulian return (ts); 884121790Sjeff } 885121790Sjeff } 886121790Sjeff return (NULL); 887121790Sjeff} 888121790Sjeff 889171482Sjeff/* 890171482Sjeff * Attempt to steal a thread in priority order from a thread queue. 891171482Sjeff */ 892164936Sjulianstatic struct td_sched * 893164936Sjuliantdq_steal(struct tdq *tdq, int stealidle) 894121790Sjeff{ 895164936Sjulian struct td_sched *ts; 896121790Sjeff 897171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 898165762Sjeff if ((ts = runq_steal(&tdq->tdq_realtime)) != NULL) 899164936Sjulian return (ts); 900171482Sjeff if ((ts = runq_steal_from(&tdq->tdq_timeshare, tdq->tdq_ridx)) != NULL) 901164936Sjulian return (ts); 902123433Sjeff if (stealidle) 903165620Sjeff return (runq_steal(&tdq->tdq_idle)); 904123433Sjeff return (NULL); 905121790Sjeff} 906123433Sjeff 907171482Sjeff/* 908171482Sjeff * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 909171482Sjeff * current lock and returns with the assigned queue locked. If this is 910171482Sjeff * via sched_switch() we leave the thread in a blocked state as an 911171482Sjeff * optimization. 912171482Sjeff */ 913171482Sjeffstatic inline struct tdq * 914171482Sjeffsched_setcpu(struct td_sched *ts, int cpu, int flags) 915123433Sjeff{ 916171482Sjeff struct thread *td; 917171482Sjeff struct tdq *tdq; 918123433Sjeff 919171482Sjeff THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 920171482Sjeff 921171482Sjeff tdq = TDQ_CPU(cpu); 922171482Sjeff td = ts->ts_thread; 923171482Sjeff ts->ts_cpu = cpu; 924171713Sjeff 925171713Sjeff /* If the lock matches just return the queue. */ 926171482Sjeff if (td->td_lock == TDQ_LOCKPTR(tdq)) 927171482Sjeff return (tdq); 928171482Sjeff#ifdef notyet 929123433Sjeff /* 930171482Sjeff * If the thread isn't running it's lockptr is a 931171482Sjeff * turnstile or a sleepqueue. We can just lock_set without 932171482Sjeff * blocking. 933123685Sjeff */ 934171482Sjeff if (TD_CAN_RUN(td)) { 935171482Sjeff TDQ_LOCK(tdq); 936171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 937171482Sjeff return (tdq); 938171482Sjeff } 939171482Sjeff#endif 940166108Sjeff /* 941171482Sjeff * The hard case, migration, we need to block the thread first to 942171482Sjeff * prevent order reversals with other cpus locks. 943166108Sjeff */ 944171482Sjeff thread_lock_block(td); 945171482Sjeff TDQ_LOCK(tdq); 946171713Sjeff thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 947171482Sjeff return (tdq); 948166108Sjeff} 949166108Sjeff 950171482Sjeff/* 951171482Sjeff * Find the thread queue running the lowest priority thread. 952171482Sjeff */ 953166108Sjeffstatic int 954171482Sjefftdq_lowestpri(void) 955166108Sjeff{ 956171482Sjeff struct tdq *tdq; 957166108Sjeff int lowpri; 958166108Sjeff int lowcpu; 959166108Sjeff int lowload; 960166108Sjeff int load; 961171482Sjeff int cpu; 962171482Sjeff int pri; 963171482Sjeff 964171482Sjeff lowload = 0; 965171482Sjeff lowpri = lowcpu = 0; 966171482Sjeff for (cpu = 0; cpu <= mp_maxid; cpu++) { 967171482Sjeff if (CPU_ABSENT(cpu)) 968171482Sjeff continue; 969171482Sjeff tdq = TDQ_CPU(cpu); 970171482Sjeff pri = tdq->tdq_lowpri; 971171482Sjeff load = TDQ_CPU(cpu)->tdq_load; 972171482Sjeff CTR4(KTR_ULE, 973171482Sjeff "cpu %d pri %d lowcpu %d lowpri %d", 974171482Sjeff cpu, pri, lowcpu, lowpri); 975171482Sjeff if (pri < lowpri) 976171482Sjeff continue; 977171482Sjeff if (lowpri && lowpri == pri && load > lowload) 978171482Sjeff continue; 979171482Sjeff lowpri = pri; 980171482Sjeff lowcpu = cpu; 981171482Sjeff lowload = load; 982171482Sjeff } 983171482Sjeff 984171482Sjeff return (lowcpu); 985171482Sjeff} 986171482Sjeff 987171482Sjeff/* 988171482Sjeff * Find the thread queue with the least load. 989171482Sjeff */ 990171482Sjeffstatic int 991171482Sjefftdq_lowestload(void) 992171482Sjeff{ 993171482Sjeff struct tdq *tdq; 994171482Sjeff int lowload; 995171482Sjeff int lowpri; 996171482Sjeff int lowcpu; 997171482Sjeff int load; 998171482Sjeff int cpu; 999171482Sjeff int pri; 1000171482Sjeff 1001171482Sjeff lowcpu = 0; 1002171482Sjeff lowload = TDQ_CPU(0)->tdq_load; 1003171482Sjeff lowpri = TDQ_CPU(0)->tdq_lowpri; 1004171482Sjeff for (cpu = 1; cpu <= mp_maxid; cpu++) { 1005171482Sjeff if (CPU_ABSENT(cpu)) 1006171482Sjeff continue; 1007171482Sjeff tdq = TDQ_CPU(cpu); 1008171482Sjeff load = tdq->tdq_load; 1009171482Sjeff pri = tdq->tdq_lowpri; 1010171482Sjeff CTR4(KTR_ULE, "cpu %d load %d lowcpu %d lowload %d", 1011171482Sjeff cpu, load, lowcpu, lowload); 1012171482Sjeff if (load > lowload) 1013171482Sjeff continue; 1014171482Sjeff if (load == lowload && pri < lowpri) 1015171482Sjeff continue; 1016171482Sjeff lowcpu = cpu; 1017171482Sjeff lowload = load; 1018171482Sjeff lowpri = pri; 1019171482Sjeff } 1020171482Sjeff 1021171482Sjeff return (lowcpu); 1022171482Sjeff} 1023171482Sjeff 1024171482Sjeff/* 1025171482Sjeff * Pick the destination cpu for sched_add(). Respects affinity and makes 1026171482Sjeff * a determination based on load or priority of available processors. 1027171482Sjeff */ 1028171482Sjeffstatic int 1029171482Sjeffsched_pickcpu(struct td_sched *ts, int flags) 1030171482Sjeff{ 1031171482Sjeff struct tdq *tdq; 1032166108Sjeff int self; 1033166108Sjeff int pri; 1034166108Sjeff int cpu; 1035166108Sjeff 1036171482Sjeff cpu = self = PCPU_GET(cpuid); 1037166108Sjeff if (smp_started == 0) 1038166108Sjeff return (self); 1039171506Sjeff /* 1040171506Sjeff * Don't migrate a running thread from sched_switch(). 1041171506Sjeff */ 1042171506Sjeff if (flags & SRQ_OURSELF) { 1043171506Sjeff CTR1(KTR_ULE, "YIELDING %d", 1044171506Sjeff curthread->td_priority); 1045171506Sjeff return (self); 1046171506Sjeff } 1047166108Sjeff pri = ts->ts_thread->td_priority; 1048171482Sjeff cpu = ts->ts_cpu; 1049166108Sjeff /* 1050166108Sjeff * Regardless of affinity, if the last cpu is idle send it there. 1051166108Sjeff */ 1052171482Sjeff tdq = TDQ_CPU(cpu); 1053171482Sjeff if (tdq->tdq_lowpri > PRI_MIN_IDLE) { 1054166229Sjeff CTR5(KTR_ULE, 1055166108Sjeff "ts_cpu %d idle, ltick %d ticks %d pri %d curthread %d", 1056166108Sjeff ts->ts_cpu, ts->ts_rltick, ticks, pri, 1057171482Sjeff tdq->tdq_lowpri); 1058166108Sjeff return (ts->ts_cpu); 1059123433Sjeff } 1060166108Sjeff /* 1061166108Sjeff * If we have affinity, try to place it on the cpu we last ran on. 1062166108Sjeff */ 1063171482Sjeff if (SCHED_AFFINITY(ts) && tdq->tdq_lowpri > pri) { 1064166229Sjeff CTR5(KTR_ULE, 1065166108Sjeff "affinity for %d, ltick %d ticks %d pri %d curthread %d", 1066166108Sjeff ts->ts_cpu, ts->ts_rltick, ticks, pri, 1067171482Sjeff tdq->tdq_lowpri); 1068166108Sjeff return (ts->ts_cpu); 1069139334Sjeff } 1070123433Sjeff /* 1071166108Sjeff * Look for an idle group. 1072123433Sjeff */ 1073166229Sjeff CTR1(KTR_ULE, "tdq_idle %X", tdq_idle); 1074166108Sjeff cpu = ffs(tdq_idle); 1075166108Sjeff if (cpu) 1076171482Sjeff return (--cpu); 1077171506Sjeff /* 1078171506Sjeff * If there are no idle cores see if we can run the thread locally. This may 1079171506Sjeff * improve locality among sleepers and wakers when there is shared data. 1080171506Sjeff */ 1081171506Sjeff if (tryself && pri < curthread->td_priority) { 1082171506Sjeff CTR1(KTR_ULE, "tryself %d", 1083166108Sjeff curthread->td_priority); 1084166108Sjeff return (self); 1085123433Sjeff } 1086133427Sjeff /* 1087166108Sjeff * Now search for the cpu running the lowest priority thread with 1088166108Sjeff * the least load. 1089123433Sjeff */ 1090171482Sjeff if (pick_pri) 1091171482Sjeff cpu = tdq_lowestpri(); 1092171482Sjeff else 1093171482Sjeff cpu = tdq_lowestload(); 1094171482Sjeff return (cpu); 1095123433Sjeff} 1096123433Sjeff 1097121790Sjeff#endif /* SMP */ 1098121790Sjeff 1099117326Sjeff/* 1100121790Sjeff * Pick the highest priority task we have and return it. 1101117326Sjeff */ 1102164936Sjulianstatic struct td_sched * 1103164936Sjuliantdq_choose(struct tdq *tdq) 1104110267Sjeff{ 1105164936Sjulian struct td_sched *ts; 1106110267Sjeff 1107171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1108165762Sjeff ts = runq_choose(&tdq->tdq_realtime); 1109170787Sjeff if (ts != NULL) 1110164936Sjulian return (ts); 1111165766Sjeff ts = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1112165762Sjeff if (ts != NULL) { 1113170787Sjeff KASSERT(ts->ts_thread->td_priority >= PRI_MIN_TIMESHARE, 1114165762Sjeff ("tdq_choose: Invalid priority on timeshare queue %d", 1115165762Sjeff ts->ts_thread->td_priority)); 1116165762Sjeff return (ts); 1117165762Sjeff } 1118110267Sjeff 1119165762Sjeff ts = runq_choose(&tdq->tdq_idle); 1120165762Sjeff if (ts != NULL) { 1121165762Sjeff KASSERT(ts->ts_thread->td_priority >= PRI_MIN_IDLE, 1122165762Sjeff ("tdq_choose: Invalid priority on idle queue %d", 1123165762Sjeff ts->ts_thread->td_priority)); 1124165762Sjeff return (ts); 1125165762Sjeff } 1126165762Sjeff 1127165762Sjeff return (NULL); 1128110267Sjeff} 1129110267Sjeff 1130171482Sjeff/* 1131171482Sjeff * Initialize a thread queue. 1132171482Sjeff */ 1133109864Sjeffstatic void 1134164936Sjuliantdq_setup(struct tdq *tdq) 1135110028Sjeff{ 1136171482Sjeff 1137171713Sjeff if (bootverbose) 1138171713Sjeff printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1139165762Sjeff runq_init(&tdq->tdq_realtime); 1140165762Sjeff runq_init(&tdq->tdq_timeshare); 1141165620Sjeff runq_init(&tdq->tdq_idle); 1142165620Sjeff tdq->tdq_load = 0; 1143110028Sjeff} 1144110028Sjeff 1145171713Sjeff#ifdef SMP 1146110028Sjeffstatic void 1147171713Sjefftdg_setup(struct tdq_group *tdg) 1148109864Sjeff{ 1149171713Sjeff if (bootverbose) 1150171713Sjeff printf("ULE: setup cpu group %d\n", TDG_ID(tdg)); 1151171713Sjeff snprintf(tdg->tdg_name, sizeof(tdg->tdg_name), 1152171713Sjeff "sched lock %d", (int)TDG_ID(tdg)); 1153171713Sjeff mtx_init(&tdg->tdg_lock, tdg->tdg_name, "sched lock", 1154171713Sjeff MTX_SPIN | MTX_RECURSE); 1155171713Sjeff LIST_INIT(&tdg->tdg_members); 1156171713Sjeff tdg->tdg_load = 0; 1157171713Sjeff tdg->tdg_transferable = 0; 1158171713Sjeff tdg->tdg_cpus = 0; 1159171713Sjeff tdg->tdg_mask = 0; 1160171713Sjeff tdg->tdg_cpumask = 0; 1161171713Sjeff tdg->tdg_idlemask = 0; 1162171713Sjeff} 1163171713Sjeff 1164171713Sjeffstatic void 1165171713Sjefftdg_add(struct tdq_group *tdg, struct tdq *tdq) 1166171713Sjeff{ 1167171713Sjeff if (tdg->tdg_mask == 0) 1168171713Sjeff tdg->tdg_mask |= 1 << TDQ_ID(tdq); 1169171713Sjeff tdg->tdg_cpumask |= 1 << TDQ_ID(tdq); 1170171713Sjeff tdg->tdg_cpus++; 1171171713Sjeff tdq->tdq_group = tdg; 1172171713Sjeff tdq->tdq_lock = &tdg->tdg_lock; 1173171713Sjeff LIST_INSERT_HEAD(&tdg->tdg_members, tdq, tdq_siblings); 1174171713Sjeff if (bootverbose) 1175171713Sjeff printf("ULE: adding cpu %d to group %d: cpus %d mask 0x%X\n", 1176171713Sjeff TDQ_ID(tdq), TDG_ID(tdg), tdg->tdg_cpus, tdg->tdg_cpumask); 1177171713Sjeff} 1178171713Sjeff 1179171713Sjeffstatic void 1180171713Sjeffsched_setup_topology(void) 1181171713Sjeff{ 1182171713Sjeff struct tdq_group *tdg; 1183171713Sjeff struct cpu_group *cg; 1184171713Sjeff int balance_groups; 1185171482Sjeff struct tdq *tdq; 1186109864Sjeff int i; 1187171713Sjeff int j; 1188109864Sjeff 1189171713Sjeff topology = 1; 1190123487Sjeff balance_groups = 0; 1191171713Sjeff for (i = 0; i < smp_topology->ct_count; i++) { 1192171713Sjeff cg = &smp_topology->ct_group[i]; 1193171713Sjeff tdg = &tdq_groups[i]; 1194171713Sjeff /* 1195171713Sjeff * Initialize the group. 1196171713Sjeff */ 1197171713Sjeff tdg_setup(tdg); 1198171713Sjeff /* 1199171713Sjeff * Find all of the group members and add them. 1200171713Sjeff */ 1201171713Sjeff for (j = 0; j < MAXCPU; j++) { 1202171713Sjeff if ((cg->cg_mask & (1 << j)) != 0) { 1203171713Sjeff tdq = TDQ_CPU(j); 1204171713Sjeff tdq_setup(tdq); 1205171713Sjeff tdg_add(tdg, tdq); 1206171713Sjeff } 1207171713Sjeff } 1208171713Sjeff if (tdg->tdg_cpus > 1) 1209171713Sjeff balance_groups = 1; 1210171713Sjeff } 1211171713Sjeff tdg_maxid = smp_topology->ct_count - 1; 1212171713Sjeff if (balance_groups) 1213171713Sjeff sched_balance_groups(NULL); 1214171713Sjeff} 1215171713Sjeff 1216171713Sjeffstatic void 1217171713Sjeffsched_setup_smp(void) 1218171713Sjeff{ 1219171713Sjeff struct tdq_group *tdg; 1220171713Sjeff struct tdq *tdq; 1221171713Sjeff int cpus; 1222171713Sjeff int i; 1223171713Sjeff 1224171713Sjeff for (cpus = 0, i = 0; i < MAXCPU; i++) { 1225171713Sjeff if (CPU_ABSENT(i)) 1226171713Sjeff continue; 1227165627Sjeff tdq = &tdq_cpu[i]; 1228171713Sjeff tdg = &tdq_groups[i]; 1229171713Sjeff /* 1230171713Sjeff * Setup a tdq group with one member. 1231171713Sjeff */ 1232171713Sjeff tdg_setup(tdg); 1233171713Sjeff tdq_setup(tdq); 1234171713Sjeff tdg_add(tdg, tdq); 1235171713Sjeff cpus++; 1236123433Sjeff } 1237171713Sjeff tdg_maxid = cpus - 1; 1238171713Sjeff} 1239123433Sjeff 1240171713Sjeff/* 1241171713Sjeff * Fake a topology with one group containing all CPUs. 1242171713Sjeff */ 1243171713Sjeffstatic void 1244171713Sjeffsched_fake_topo(void) 1245171713Sjeff{ 1246171713Sjeff#ifdef SCHED_FAKE_TOPOLOGY 1247171713Sjeff static struct cpu_top top; 1248171713Sjeff static struct cpu_group group; 1249113357Sjeff 1250171713Sjeff top.ct_count = 1; 1251171713Sjeff top.ct_group = &group; 1252171713Sjeff group.cg_mask = all_cpus; 1253171713Sjeff group.cg_count = mp_ncpus; 1254171713Sjeff group.cg_children = 0; 1255171713Sjeff smp_topology = ⊤ 1256171713Sjeff#endif 1257171713Sjeff} 1258171713Sjeff#endif 1259171713Sjeff 1260171713Sjeff/* 1261171713Sjeff * Setup the thread queues and initialize the topology based on MD 1262171713Sjeff * information. 1263171713Sjeff */ 1264171713Sjeffstatic void 1265171713Sjeffsched_setup(void *dummy) 1266171713Sjeff{ 1267171713Sjeff struct tdq *tdq; 1268171713Sjeff 1269171713Sjeff tdq = TDQ_SELF(); 1270171713Sjeff#ifdef SMP 1271123487Sjeff /* 1272171482Sjeff * Initialize long-term cpu balancing algorithm. 1273123487Sjeff */ 1274171482Sjeff callout_init(&balco, CALLOUT_MPSAFE); 1275171482Sjeff callout_init(&gbalco, CALLOUT_MPSAFE); 1276171713Sjeff sched_fake_topo(); 1277171713Sjeff /* 1278171713Sjeff * Setup tdqs based on a topology configuration or vanilla SMP based 1279171713Sjeff * on mp_maxid. 1280171713Sjeff */ 1281171713Sjeff if (smp_topology == NULL) 1282171713Sjeff sched_setup_smp(); 1283171713Sjeff else 1284171713Sjeff sched_setup_topology(); 1285171482Sjeff sched_balance(NULL); 1286117237Sjeff#else 1287171713Sjeff tdq_setup(tdq); 1288171713Sjeff mtx_init(&tdq_lock, "sched lock", "sched lock", MTX_SPIN | MTX_RECURSE); 1289171713Sjeff tdq->tdq_lock = &tdq_lock; 1290116069Sjeff#endif 1291171482Sjeff /* 1292171482Sjeff * To avoid divide-by-zero, we set realstathz a dummy value 1293171482Sjeff * in case which sched_clock() called before sched_initticks(). 1294171482Sjeff */ 1295171482Sjeff realstathz = hz; 1296171482Sjeff sched_slice = (realstathz/10); /* ~100ms */ 1297171482Sjeff tickincr = 1 << SCHED_TICK_SHIFT; 1298171482Sjeff 1299171482Sjeff /* Add thread0's load since it's running. */ 1300171482Sjeff TDQ_LOCK(tdq); 1301171713Sjeff thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1302171482Sjeff tdq_load_add(tdq, &td_sched0); 1303171482Sjeff TDQ_UNLOCK(tdq); 1304109864Sjeff} 1305109864Sjeff 1306171482Sjeff/* 1307171482Sjeff * This routine determines the tickincr after stathz and hz are setup. 1308171482Sjeff */ 1309153533Sdavidxu/* ARGSUSED */ 1310153533Sdavidxustatic void 1311153533Sdavidxusched_initticks(void *dummy) 1312153533Sdavidxu{ 1313171482Sjeff int incr; 1314171482Sjeff 1315153533Sdavidxu realstathz = stathz ? stathz : hz; 1316166229Sjeff sched_slice = (realstathz/10); /* ~100ms */ 1317153533Sdavidxu 1318153533Sdavidxu /* 1319165762Sjeff * tickincr is shifted out by 10 to avoid rounding errors due to 1320165766Sjeff * hz not being evenly divisible by stathz on all platforms. 1321153533Sdavidxu */ 1322171482Sjeff incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1323165762Sjeff /* 1324165762Sjeff * This does not work for values of stathz that are more than 1325165762Sjeff * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1326165762Sjeff */ 1327171482Sjeff if (incr == 0) 1328171482Sjeff incr = 1; 1329171482Sjeff tickincr = incr; 1330166108Sjeff#ifdef SMP 1331171899Sjeff /* 1332171899Sjeff * Set steal thresh to log2(mp_ncpu) but no greater than 4. This 1333171899Sjeff * prevents excess thrashing on large machines and excess idle on 1334171899Sjeff * smaller machines. 1335171899Sjeff */ 1336171899Sjeff steal_thresh = min(ffs(mp_ncpus) - 1, 4); 1337166108Sjeff affinity = SCHED_AFFINITY_DEFAULT; 1338166108Sjeff#endif 1339153533Sdavidxu} 1340153533Sdavidxu 1341153533Sdavidxu 1342109864Sjeff/* 1343171482Sjeff * This is the core of the interactivity algorithm. Determines a score based 1344171482Sjeff * on past behavior. It is the ratio of sleep time to run time scaled to 1345171482Sjeff * a [0, 100] integer. This is the voluntary sleep time of a process, which 1346171482Sjeff * differs from the cpu usage because it does not account for time spent 1347171482Sjeff * waiting on a run-queue. Would be prettier if we had floating point. 1348171482Sjeff */ 1349171482Sjeffstatic int 1350171482Sjeffsched_interact_score(struct thread *td) 1351171482Sjeff{ 1352171482Sjeff struct td_sched *ts; 1353171482Sjeff int div; 1354171482Sjeff 1355171482Sjeff ts = td->td_sched; 1356171482Sjeff /* 1357171482Sjeff * The score is only needed if this is likely to be an interactive 1358171482Sjeff * task. Don't go through the expense of computing it if there's 1359171482Sjeff * no chance. 1360171482Sjeff */ 1361171482Sjeff if (sched_interact <= SCHED_INTERACT_HALF && 1362171482Sjeff ts->ts_runtime >= ts->ts_slptime) 1363171482Sjeff return (SCHED_INTERACT_HALF); 1364171482Sjeff 1365171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1366171482Sjeff div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1367171482Sjeff return (SCHED_INTERACT_HALF + 1368171482Sjeff (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1369171482Sjeff } 1370171482Sjeff if (ts->ts_slptime > ts->ts_runtime) { 1371171482Sjeff div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1372171482Sjeff return (ts->ts_runtime / div); 1373171482Sjeff } 1374171482Sjeff /* runtime == slptime */ 1375171482Sjeff if (ts->ts_runtime) 1376171482Sjeff return (SCHED_INTERACT_HALF); 1377171482Sjeff 1378171482Sjeff /* 1379171482Sjeff * This can happen if slptime and runtime are 0. 1380171482Sjeff */ 1381171482Sjeff return (0); 1382171482Sjeff 1383171482Sjeff} 1384171482Sjeff 1385171482Sjeff/* 1386109864Sjeff * Scale the scheduling priority according to the "interactivity" of this 1387109864Sjeff * process. 1388109864Sjeff */ 1389113357Sjeffstatic void 1390163709Sjbsched_priority(struct thread *td) 1391109864Sjeff{ 1392165762Sjeff int score; 1393109864Sjeff int pri; 1394109864Sjeff 1395163709Sjb if (td->td_pri_class != PRI_TIMESHARE) 1396113357Sjeff return; 1397112966Sjeff /* 1398165762Sjeff * If the score is interactive we place the thread in the realtime 1399165762Sjeff * queue with a priority that is less than kernel and interrupt 1400165762Sjeff * priorities. These threads are not subject to nice restrictions. 1401112966Sjeff * 1402171482Sjeff * Scores greater than this are placed on the normal timeshare queue 1403165762Sjeff * where the priority is partially decided by the most recent cpu 1404165762Sjeff * utilization and the rest is decided by nice value. 1405112966Sjeff */ 1406165762Sjeff score = sched_interact_score(td); 1407165762Sjeff if (score < sched_interact) { 1408165762Sjeff pri = PRI_MIN_REALTIME; 1409165762Sjeff pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact) 1410165762Sjeff * score; 1411165762Sjeff KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME, 1412166208Sjeff ("sched_priority: invalid interactive priority %d score %d", 1413166208Sjeff pri, score)); 1414165762Sjeff } else { 1415165762Sjeff pri = SCHED_PRI_MIN; 1416165762Sjeff if (td->td_sched->ts_ticks) 1417165762Sjeff pri += SCHED_PRI_TICKS(td->td_sched); 1418165762Sjeff pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1419171482Sjeff KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE, 1420171482Sjeff ("sched_priority: invalid priority %d: nice %d, " 1421171482Sjeff "ticks %d ftick %d ltick %d tick pri %d", 1422171482Sjeff pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1423171482Sjeff td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1424171482Sjeff SCHED_PRI_TICKS(td->td_sched))); 1425165762Sjeff } 1426165762Sjeff sched_user_prio(td, pri); 1427112966Sjeff 1428112966Sjeff return; 1429109864Sjeff} 1430109864Sjeff 1431121868Sjeff/* 1432121868Sjeff * This routine enforces a maximum limit on the amount of scheduling history 1433171482Sjeff * kept. It is called after either the slptime or runtime is adjusted. This 1434171482Sjeff * function is ugly due to integer math. 1435121868Sjeff */ 1436116463Sjeffstatic void 1437163709Sjbsched_interact_update(struct thread *td) 1438116463Sjeff{ 1439165819Sjeff struct td_sched *ts; 1440166208Sjeff u_int sum; 1441121605Sjeff 1442165819Sjeff ts = td->td_sched; 1443171482Sjeff sum = ts->ts_runtime + ts->ts_slptime; 1444121868Sjeff if (sum < SCHED_SLP_RUN_MAX) 1445121868Sjeff return; 1446121868Sjeff /* 1447165819Sjeff * This only happens from two places: 1448165819Sjeff * 1) We have added an unusual amount of run time from fork_exit. 1449165819Sjeff * 2) We have added an unusual amount of sleep time from sched_sleep(). 1450165819Sjeff */ 1451165819Sjeff if (sum > SCHED_SLP_RUN_MAX * 2) { 1452171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1453171482Sjeff ts->ts_runtime = SCHED_SLP_RUN_MAX; 1454171482Sjeff ts->ts_slptime = 1; 1455165819Sjeff } else { 1456171482Sjeff ts->ts_slptime = SCHED_SLP_RUN_MAX; 1457171482Sjeff ts->ts_runtime = 1; 1458165819Sjeff } 1459165819Sjeff return; 1460165819Sjeff } 1461165819Sjeff /* 1462121868Sjeff * If we have exceeded by more than 1/5th then the algorithm below 1463121868Sjeff * will not bring us back into range. Dividing by two here forces 1464133427Sjeff * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1465121868Sjeff */ 1466127850Sjeff if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1467171482Sjeff ts->ts_runtime /= 2; 1468171482Sjeff ts->ts_slptime /= 2; 1469121868Sjeff return; 1470116463Sjeff } 1471171482Sjeff ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1472171482Sjeff ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1473116463Sjeff} 1474116463Sjeff 1475171482Sjeff/* 1476171482Sjeff * Scale back the interactivity history when a child thread is created. The 1477171482Sjeff * history is inherited from the parent but the thread may behave totally 1478171482Sjeff * differently. For example, a shell spawning a compiler process. We want 1479171482Sjeff * to learn that the compiler is behaving badly very quickly. 1480171482Sjeff */ 1481121868Sjeffstatic void 1482163709Sjbsched_interact_fork(struct thread *td) 1483121868Sjeff{ 1484121868Sjeff int ratio; 1485121868Sjeff int sum; 1486121868Sjeff 1487171482Sjeff sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1488121868Sjeff if (sum > SCHED_SLP_RUN_FORK) { 1489121868Sjeff ratio = sum / SCHED_SLP_RUN_FORK; 1490171482Sjeff td->td_sched->ts_runtime /= ratio; 1491171482Sjeff td->td_sched->ts_slptime /= ratio; 1492121868Sjeff } 1493121868Sjeff} 1494121868Sjeff 1495113357Sjeff/* 1496171482Sjeff * Called from proc0_init() to setup the scheduler fields. 1497134791Sjulian */ 1498134791Sjulianvoid 1499134791Sjulianschedinit(void) 1500134791Sjulian{ 1501165762Sjeff 1502134791Sjulian /* 1503134791Sjulian * Set up the scheduler specific parts of proc0. 1504134791Sjulian */ 1505136167Sjulian proc0.p_sched = NULL; /* XXX */ 1506164936Sjulian thread0.td_sched = &td_sched0; 1507165762Sjeff td_sched0.ts_ltick = ticks; 1508165796Sjeff td_sched0.ts_ftick = ticks; 1509164936Sjulian td_sched0.ts_thread = &thread0; 1510134791Sjulian} 1511134791Sjulian 1512134791Sjulian/* 1513113357Sjeff * This is only somewhat accurate since given many processes of the same 1514113357Sjeff * priority they will switch when their slices run out, which will be 1515165762Sjeff * at most sched_slice stathz ticks. 1516113357Sjeff */ 1517109864Sjeffint 1518109864Sjeffsched_rr_interval(void) 1519109864Sjeff{ 1520165762Sjeff 1521165762Sjeff /* Convert sched_slice to hz */ 1522165762Sjeff return (hz/(realstathz/sched_slice)); 1523109864Sjeff} 1524109864Sjeff 1525171482Sjeff/* 1526171482Sjeff * Update the percent cpu tracking information when it is requested or 1527171482Sjeff * the total history exceeds the maximum. We keep a sliding history of 1528171482Sjeff * tick counts that slowly decays. This is less precise than the 4BSD 1529171482Sjeff * mechanism since it happens with less regular and frequent events. 1530171482Sjeff */ 1531121790Sjeffstatic void 1532164936Sjuliansched_pctcpu_update(struct td_sched *ts) 1533109864Sjeff{ 1534165762Sjeff 1535165762Sjeff if (ts->ts_ticks == 0) 1536165762Sjeff return; 1537165796Sjeff if (ticks - (hz / 10) < ts->ts_ltick && 1538165796Sjeff SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1539165796Sjeff return; 1540109864Sjeff /* 1541109864Sjeff * Adjust counters and watermark for pctcpu calc. 1542116365Sjeff */ 1543165762Sjeff if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1544164936Sjulian ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1545165762Sjeff SCHED_TICK_TARG; 1546165762Sjeff else 1547164936Sjulian ts->ts_ticks = 0; 1548164936Sjulian ts->ts_ltick = ticks; 1549165762Sjeff ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1550109864Sjeff} 1551109864Sjeff 1552171482Sjeff/* 1553171482Sjeff * Adjust the priority of a thread. Move it to the appropriate run-queue 1554171482Sjeff * if necessary. This is the back-end for several priority related 1555171482Sjeff * functions. 1556171482Sjeff */ 1557165762Sjeffstatic void 1558139453Sjhbsched_thread_priority(struct thread *td, u_char prio) 1559109864Sjeff{ 1560164936Sjulian struct td_sched *ts; 1561109864Sjeff 1562139316Sjeff CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 1563139316Sjeff td, td->td_proc->p_comm, td->td_priority, prio, curthread, 1564139316Sjeff curthread->td_proc->p_comm); 1565164936Sjulian ts = td->td_sched; 1566170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1567139453Sjhb if (td->td_priority == prio) 1568139453Sjhb return; 1569165762Sjeff 1570165766Sjeff if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1571121605Sjeff /* 1572121605Sjeff * If the priority has been elevated due to priority 1573121605Sjeff * propagation, we may have to move ourselves to a new 1574165762Sjeff * queue. This could be optimized to not re-add in some 1575165762Sjeff * cases. 1576133555Sjeff */ 1577165762Sjeff sched_rem(td); 1578165762Sjeff td->td_priority = prio; 1579171482Sjeff sched_add(td, SRQ_BORROWING); 1580171482Sjeff } else { 1581171482Sjeff#ifdef SMP 1582171482Sjeff struct tdq *tdq; 1583171482Sjeff 1584171482Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1585171482Sjeff if (prio < tdq->tdq_lowpri) 1586171482Sjeff tdq->tdq_lowpri = prio; 1587171482Sjeff#endif 1588119488Sdavidxu td->td_priority = prio; 1589171482Sjeff } 1590109864Sjeff} 1591109864Sjeff 1592139453Sjhb/* 1593139453Sjhb * Update a thread's priority when it is lent another thread's 1594139453Sjhb * priority. 1595139453Sjhb */ 1596109864Sjeffvoid 1597139453Sjhbsched_lend_prio(struct thread *td, u_char prio) 1598139453Sjhb{ 1599139453Sjhb 1600139453Sjhb td->td_flags |= TDF_BORROWING; 1601139453Sjhb sched_thread_priority(td, prio); 1602139453Sjhb} 1603139453Sjhb 1604139453Sjhb/* 1605139453Sjhb * Restore a thread's priority when priority propagation is 1606139453Sjhb * over. The prio argument is the minimum priority the thread 1607139453Sjhb * needs to have to satisfy other possible priority lending 1608139453Sjhb * requests. If the thread's regular priority is less 1609139453Sjhb * important than prio, the thread will keep a priority boost 1610139453Sjhb * of prio. 1611139453Sjhb */ 1612139453Sjhbvoid 1613139453Sjhbsched_unlend_prio(struct thread *td, u_char prio) 1614139453Sjhb{ 1615139453Sjhb u_char base_pri; 1616139453Sjhb 1617139453Sjhb if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1618139453Sjhb td->td_base_pri <= PRI_MAX_TIMESHARE) 1619163709Sjb base_pri = td->td_user_pri; 1620139453Sjhb else 1621139453Sjhb base_pri = td->td_base_pri; 1622139453Sjhb if (prio >= base_pri) { 1623139455Sjhb td->td_flags &= ~TDF_BORROWING; 1624139453Sjhb sched_thread_priority(td, base_pri); 1625139453Sjhb } else 1626139453Sjhb sched_lend_prio(td, prio); 1627139453Sjhb} 1628139453Sjhb 1629171482Sjeff/* 1630171482Sjeff * Standard entry for setting the priority to an absolute value. 1631171482Sjeff */ 1632139453Sjhbvoid 1633139453Sjhbsched_prio(struct thread *td, u_char prio) 1634139453Sjhb{ 1635139453Sjhb u_char oldprio; 1636139453Sjhb 1637139453Sjhb /* First, update the base priority. */ 1638139453Sjhb td->td_base_pri = prio; 1639139453Sjhb 1640139453Sjhb /* 1641139455Sjhb * If the thread is borrowing another thread's priority, don't 1642139453Sjhb * ever lower the priority. 1643139453Sjhb */ 1644139453Sjhb if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1645139453Sjhb return; 1646139453Sjhb 1647139453Sjhb /* Change the real priority. */ 1648139453Sjhb oldprio = td->td_priority; 1649139453Sjhb sched_thread_priority(td, prio); 1650139453Sjhb 1651139453Sjhb /* 1652139453Sjhb * If the thread is on a turnstile, then let the turnstile update 1653139453Sjhb * its state. 1654139453Sjhb */ 1655139453Sjhb if (TD_ON_LOCK(td) && oldprio != prio) 1656139453Sjhb turnstile_adjust(td, oldprio); 1657139453Sjhb} 1658139455Sjhb 1659171482Sjeff/* 1660171482Sjeff * Set the base user priority, does not effect current running priority. 1661171482Sjeff */ 1662139453Sjhbvoid 1663163709Sjbsched_user_prio(struct thread *td, u_char prio) 1664161599Sdavidxu{ 1665161599Sdavidxu u_char oldprio; 1666161599Sdavidxu 1667163709Sjb td->td_base_user_pri = prio; 1668164939Sjulian if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio) 1669164939Sjulian return; 1670163709Sjb oldprio = td->td_user_pri; 1671163709Sjb td->td_user_pri = prio; 1672163709Sjb 1673161599Sdavidxu if (TD_ON_UPILOCK(td) && oldprio != prio) 1674161599Sdavidxu umtx_pi_adjust(td, oldprio); 1675161599Sdavidxu} 1676161599Sdavidxu 1677161599Sdavidxuvoid 1678161599Sdavidxusched_lend_user_prio(struct thread *td, u_char prio) 1679161599Sdavidxu{ 1680161599Sdavidxu u_char oldprio; 1681161599Sdavidxu 1682161599Sdavidxu td->td_flags |= TDF_UBORROWING; 1683161599Sdavidxu 1684164091Smaxim oldprio = td->td_user_pri; 1685163709Sjb td->td_user_pri = prio; 1686161599Sdavidxu 1687161599Sdavidxu if (TD_ON_UPILOCK(td) && oldprio != prio) 1688161599Sdavidxu umtx_pi_adjust(td, oldprio); 1689161599Sdavidxu} 1690161599Sdavidxu 1691161599Sdavidxuvoid 1692161599Sdavidxusched_unlend_user_prio(struct thread *td, u_char prio) 1693161599Sdavidxu{ 1694161599Sdavidxu u_char base_pri; 1695161599Sdavidxu 1696163709Sjb base_pri = td->td_base_user_pri; 1697161599Sdavidxu if (prio >= base_pri) { 1698161599Sdavidxu td->td_flags &= ~TDF_UBORROWING; 1699163709Sjb sched_user_prio(td, base_pri); 1700161599Sdavidxu } else 1701161599Sdavidxu sched_lend_user_prio(td, prio); 1702161599Sdavidxu} 1703161599Sdavidxu 1704171482Sjeff/* 1705171505Sjeff * Add the thread passed as 'newtd' to the run queue before selecting 1706171505Sjeff * the next thread to run. This is only used for KSE. 1707171505Sjeff */ 1708171505Sjeffstatic void 1709171505Sjeffsched_switchin(struct tdq *tdq, struct thread *td) 1710171505Sjeff{ 1711171505Sjeff#ifdef SMP 1712171505Sjeff spinlock_enter(); 1713171505Sjeff TDQ_UNLOCK(tdq); 1714171505Sjeff thread_lock(td); 1715171505Sjeff spinlock_exit(); 1716171505Sjeff sched_setcpu(td->td_sched, TDQ_ID(tdq), SRQ_YIELDING); 1717171505Sjeff#else 1718171505Sjeff td->td_lock = TDQ_LOCKPTR(tdq); 1719171505Sjeff#endif 1720171505Sjeff tdq_add(tdq, td, SRQ_YIELDING); 1721171505Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1722171505Sjeff} 1723171505Sjeff 1724171505Sjeff/* 1725171713Sjeff * Handle migration from sched_switch(). This happens only for 1726171713Sjeff * cpu binding. 1727171713Sjeff */ 1728171713Sjeffstatic struct mtx * 1729171713Sjeffsched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1730171713Sjeff{ 1731171713Sjeff struct tdq *tdn; 1732171713Sjeff 1733171713Sjeff tdn = TDQ_CPU(td->td_sched->ts_cpu); 1734171713Sjeff#ifdef SMP 1735171713Sjeff /* 1736171713Sjeff * Do the lock dance required to avoid LOR. We grab an extra 1737171713Sjeff * spinlock nesting to prevent preemption while we're 1738171713Sjeff * not holding either run-queue lock. 1739171713Sjeff */ 1740171713Sjeff spinlock_enter(); 1741171713Sjeff thread_block_switch(td); /* This releases the lock on tdq. */ 1742171713Sjeff TDQ_LOCK(tdn); 1743171713Sjeff tdq_add(tdn, td, flags); 1744171713Sjeff tdq_notify(td->td_sched); 1745171713Sjeff /* 1746171713Sjeff * After we unlock tdn the new cpu still can't switch into this 1747171713Sjeff * thread until we've unblocked it in cpu_switch(). The lock 1748171713Sjeff * pointers may match in the case of HTT cores. Don't unlock here 1749171713Sjeff * or we can deadlock when the other CPU runs the IPI handler. 1750171713Sjeff */ 1751171713Sjeff if (TDQ_LOCKPTR(tdn) != TDQ_LOCKPTR(tdq)) { 1752171713Sjeff TDQ_UNLOCK(tdn); 1753171713Sjeff TDQ_LOCK(tdq); 1754171713Sjeff } 1755171713Sjeff spinlock_exit(); 1756171713Sjeff#endif 1757171713Sjeff return (TDQ_LOCKPTR(tdn)); 1758171713Sjeff} 1759171713Sjeff 1760171713Sjeff/* 1761171482Sjeff * Block a thread for switching. Similar to thread_block() but does not 1762171482Sjeff * bump the spin count. 1763171482Sjeff */ 1764171482Sjeffstatic inline struct mtx * 1765171482Sjeffthread_block_switch(struct thread *td) 1766171482Sjeff{ 1767171482Sjeff struct mtx *lock; 1768171482Sjeff 1769171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1770171482Sjeff lock = td->td_lock; 1771171482Sjeff td->td_lock = &blocked_lock; 1772171482Sjeff mtx_unlock_spin(lock); 1773171482Sjeff 1774171482Sjeff return (lock); 1775171482Sjeff} 1776171482Sjeff 1777171482Sjeff/* 1778171482Sjeff * Release a thread that was blocked with thread_block_switch(). 1779171482Sjeff */ 1780171482Sjeffstatic inline void 1781171482Sjeffthread_unblock_switch(struct thread *td, struct mtx *mtx) 1782171482Sjeff{ 1783171482Sjeff atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1784171482Sjeff (uintptr_t)mtx); 1785171482Sjeff} 1786171482Sjeff 1787171482Sjeff/* 1788171482Sjeff * Switch threads. This function has to handle threads coming in while 1789171482Sjeff * blocked for some reason, running, or idle. It also must deal with 1790171482Sjeff * migrating a thread from one queue to another as running threads may 1791171482Sjeff * be assigned elsewhere via binding. 1792171482Sjeff */ 1793161599Sdavidxuvoid 1794135051Sjuliansched_switch(struct thread *td, struct thread *newtd, int flags) 1795109864Sjeff{ 1796165627Sjeff struct tdq *tdq; 1797164936Sjulian struct td_sched *ts; 1798171482Sjeff struct mtx *mtx; 1799171713Sjeff int srqflag; 1800171482Sjeff int cpuid; 1801109864Sjeff 1802170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1803109864Sjeff 1804171482Sjeff cpuid = PCPU_GET(cpuid); 1805171482Sjeff tdq = TDQ_CPU(cpuid); 1806164936Sjulian ts = td->td_sched; 1807171713Sjeff mtx = td->td_lock; 1808171482Sjeff#ifdef SMP 1809171482Sjeff ts->ts_rltick = ticks; 1810171482Sjeff if (newtd && newtd->td_priority < tdq->tdq_lowpri) 1811171482Sjeff tdq->tdq_lowpri = newtd->td_priority; 1812171482Sjeff#endif 1813133555Sjeff td->td_lastcpu = td->td_oncpu; 1814113339Sjulian td->td_oncpu = NOCPU; 1815132266Sjhb td->td_flags &= ~TDF_NEEDRESCHED; 1816144777Sups td->td_owepreempt = 0; 1817123434Sjeff /* 1818171482Sjeff * The lock pointer in an idle thread should never change. Reset it 1819171482Sjeff * to CAN_RUN as well. 1820123434Sjeff */ 1821167327Sjulian if (TD_IS_IDLETHREAD(td)) { 1822171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1823139334Sjeff TD_SET_CAN_RUN(td); 1824170293Sjeff } else if (TD_IS_RUNNING(td)) { 1825171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1826165627Sjeff tdq_load_rem(tdq, ts); 1827171713Sjeff srqflag = (flags & SW_PREEMPT) ? 1828170293Sjeff SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1829171713Sjeff SRQ_OURSELF|SRQ_YIELDING; 1830171713Sjeff if (ts->ts_cpu == cpuid) 1831171713Sjeff tdq_add(tdq, td, srqflag); 1832171713Sjeff else 1833171713Sjeff mtx = sched_switch_migrate(tdq, td, srqflag); 1834171482Sjeff } else { 1835171482Sjeff /* This thread must be going to sleep. */ 1836171482Sjeff TDQ_LOCK(tdq); 1837171482Sjeff mtx = thread_block_switch(td); 1838170293Sjeff tdq_load_rem(tdq, ts); 1839171482Sjeff } 1840171482Sjeff /* 1841171482Sjeff * We enter here with the thread blocked and assigned to the 1842171482Sjeff * appropriate cpu run-queue or sleep-queue and with the current 1843171482Sjeff * thread-queue locked. 1844171482Sjeff */ 1845171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1846171482Sjeff /* 1847171505Sjeff * If KSE assigned a new thread just add it here and let choosethread 1848171505Sjeff * select the best one. 1849171482Sjeff */ 1850171505Sjeff if (newtd != NULL) 1851171505Sjeff sched_switchin(tdq, newtd); 1852171482Sjeff newtd = choosethread(); 1853171482Sjeff /* 1854171482Sjeff * Call the MD code to switch contexts if necessary. 1855171482Sjeff */ 1856145256Sjkoshy if (td != newtd) { 1857145256Sjkoshy#ifdef HWPMC_HOOKS 1858145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1859145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1860145256Sjkoshy#endif 1861171482Sjeff cpu_switch(td, newtd, mtx); 1862171482Sjeff /* 1863171482Sjeff * We may return from cpu_switch on a different cpu. However, 1864171482Sjeff * we always return with td_lock pointing to the current cpu's 1865171482Sjeff * run queue lock. 1866171482Sjeff */ 1867171482Sjeff cpuid = PCPU_GET(cpuid); 1868171482Sjeff tdq = TDQ_CPU(cpuid); 1869171482Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)td; 1870145256Sjkoshy#ifdef HWPMC_HOOKS 1871145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1872145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1873145256Sjkoshy#endif 1874171482Sjeff } else 1875171482Sjeff thread_unblock_switch(td, mtx); 1876171482Sjeff /* 1877171482Sjeff * Assert that all went well and return. 1878171482Sjeff */ 1879171482Sjeff#ifdef SMP 1880171482Sjeff /* We should always get here with the lowest priority td possible */ 1881171482Sjeff tdq->tdq_lowpri = td->td_priority; 1882171482Sjeff#endif 1883171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1884171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1885171482Sjeff td->td_oncpu = cpuid; 1886109864Sjeff} 1887109864Sjeff 1888171482Sjeff/* 1889171482Sjeff * Adjust thread priorities as a result of a nice request. 1890171482Sjeff */ 1891109864Sjeffvoid 1892130551Sjuliansched_nice(struct proc *p, int nice) 1893109864Sjeff{ 1894109864Sjeff struct thread *td; 1895109864Sjeff 1896130551Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 1897170293Sjeff PROC_SLOCK_ASSERT(p, MA_OWNED); 1898165762Sjeff 1899130551Sjulian p->p_nice = nice; 1900163709Sjb FOREACH_THREAD_IN_PROC(p, td) { 1901170293Sjeff thread_lock(td); 1902163709Sjb sched_priority(td); 1903165762Sjeff sched_prio(td, td->td_base_user_pri); 1904170293Sjeff thread_unlock(td); 1905130551Sjulian } 1906109864Sjeff} 1907109864Sjeff 1908171482Sjeff/* 1909171482Sjeff * Record the sleep time for the interactivity scorer. 1910171482Sjeff */ 1911109864Sjeffvoid 1912126326Sjhbsched_sleep(struct thread *td) 1913109864Sjeff{ 1914165762Sjeff 1915170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1916109864Sjeff 1917171482Sjeff td->td_sched->ts_slptick = ticks; 1918109864Sjeff} 1919109864Sjeff 1920171482Sjeff/* 1921171482Sjeff * Schedule a thread to resume execution and record how long it voluntarily 1922171482Sjeff * slept. We also update the pctcpu, interactivity, and priority. 1923171482Sjeff */ 1924109864Sjeffvoid 1925109864Sjeffsched_wakeup(struct thread *td) 1926109864Sjeff{ 1927166229Sjeff struct td_sched *ts; 1928171482Sjeff int slptick; 1929165762Sjeff 1930170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1931166229Sjeff ts = td->td_sched; 1932109864Sjeff /* 1933165762Sjeff * If we slept for more than a tick update our interactivity and 1934165762Sjeff * priority. 1935109864Sjeff */ 1936171482Sjeff slptick = ts->ts_slptick; 1937171482Sjeff ts->ts_slptick = 0; 1938171482Sjeff if (slptick && slptick != ticks) { 1939166208Sjeff u_int hzticks; 1940109864Sjeff 1941171482Sjeff hzticks = (ticks - slptick) << SCHED_TICK_SHIFT; 1942171482Sjeff ts->ts_slptime += hzticks; 1943165819Sjeff sched_interact_update(td); 1944166229Sjeff sched_pctcpu_update(ts); 1945163709Sjb sched_priority(td); 1946109864Sjeff } 1947166229Sjeff /* Reset the slice value after we sleep. */ 1948166229Sjeff ts->ts_slice = sched_slice; 1949166190Sjeff sched_add(td, SRQ_BORING); 1950109864Sjeff} 1951109864Sjeff 1952109864Sjeff/* 1953109864Sjeff * Penalize the parent for creating a new child and initialize the child's 1954109864Sjeff * priority. 1955109864Sjeff */ 1956109864Sjeffvoid 1957163709Sjbsched_fork(struct thread *td, struct thread *child) 1958109864Sjeff{ 1959170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1960164936Sjulian sched_fork_thread(td, child); 1961165762Sjeff /* 1962165762Sjeff * Penalize the parent and child for forking. 1963165762Sjeff */ 1964165762Sjeff sched_interact_fork(child); 1965165762Sjeff sched_priority(child); 1966171482Sjeff td->td_sched->ts_runtime += tickincr; 1967165762Sjeff sched_interact_update(td); 1968165762Sjeff sched_priority(td); 1969164936Sjulian} 1970109864Sjeff 1971171482Sjeff/* 1972171482Sjeff * Fork a new thread, may be within the same process. 1973171482Sjeff */ 1974164936Sjulianvoid 1975164936Sjuliansched_fork_thread(struct thread *td, struct thread *child) 1976164936Sjulian{ 1977164936Sjulian struct td_sched *ts; 1978164936Sjulian struct td_sched *ts2; 1979164936Sjulian 1980165762Sjeff /* 1981165762Sjeff * Initialize child. 1982165762Sjeff */ 1983170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1984163709Sjb sched_newthread(child); 1985171482Sjeff child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1986164936Sjulian ts = td->td_sched; 1987164936Sjulian ts2 = child->td_sched; 1988164936Sjulian ts2->ts_cpu = ts->ts_cpu; 1989164936Sjulian ts2->ts_runq = NULL; 1990165762Sjeff /* 1991165762Sjeff * Grab our parents cpu estimation information and priority. 1992165762Sjeff */ 1993164936Sjulian ts2->ts_ticks = ts->ts_ticks; 1994164936Sjulian ts2->ts_ltick = ts->ts_ltick; 1995164936Sjulian ts2->ts_ftick = ts->ts_ftick; 1996165762Sjeff child->td_user_pri = td->td_user_pri; 1997165762Sjeff child->td_base_user_pri = td->td_base_user_pri; 1998165762Sjeff /* 1999165762Sjeff * And update interactivity score. 2000165762Sjeff */ 2001171482Sjeff ts2->ts_slptime = ts->ts_slptime; 2002171482Sjeff ts2->ts_runtime = ts->ts_runtime; 2003165762Sjeff ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 2004113357Sjeff} 2005113357Sjeff 2006171482Sjeff/* 2007171482Sjeff * Adjust the priority class of a thread. 2008171482Sjeff */ 2009113357Sjeffvoid 2010163709Sjbsched_class(struct thread *td, int class) 2011113357Sjeff{ 2012113357Sjeff 2013170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2014163709Sjb if (td->td_pri_class == class) 2015113357Sjeff return; 2016113357Sjeff 2017121896Sjeff#ifdef SMP 2018165827Sjeff /* 2019165827Sjeff * On SMP if we're on the RUNQ we must adjust the transferable 2020165827Sjeff * count because could be changing to or from an interrupt 2021165827Sjeff * class. 2022165827Sjeff */ 2023166190Sjeff if (TD_ON_RUNQ(td)) { 2024165827Sjeff struct tdq *tdq; 2025165827Sjeff 2026165827Sjeff tdq = TDQ_CPU(td->td_sched->ts_cpu); 2027165827Sjeff if (THREAD_CAN_MIGRATE(td)) { 2028165827Sjeff tdq->tdq_transferable--; 2029165827Sjeff tdq->tdq_group->tdg_transferable--; 2030122744Sjeff } 2031165827Sjeff td->td_pri_class = class; 2032165827Sjeff if (THREAD_CAN_MIGRATE(td)) { 2033165827Sjeff tdq->tdq_transferable++; 2034165827Sjeff tdq->tdq_group->tdg_transferable++; 2035165827Sjeff } 2036165827Sjeff } 2037164936Sjulian#endif 2038163709Sjb td->td_pri_class = class; 2039109864Sjeff} 2040109864Sjeff 2041109864Sjeff/* 2042109864Sjeff * Return some of the child's priority and interactivity to the parent. 2043109864Sjeff */ 2044109864Sjeffvoid 2045164939Sjuliansched_exit(struct proc *p, struct thread *child) 2046109864Sjeff{ 2047165762Sjeff struct thread *td; 2048164939Sjulian 2049163709Sjb CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d", 2050164939Sjulian child, child->td_proc->p_comm, child->td_priority); 2051113372Sjeff 2052170293Sjeff PROC_SLOCK_ASSERT(p, MA_OWNED); 2053165762Sjeff td = FIRST_THREAD_IN_PROC(p); 2054165762Sjeff sched_exit_thread(td, child); 2055113372Sjeff} 2056113372Sjeff 2057171482Sjeff/* 2058171482Sjeff * Penalize another thread for the time spent on this one. This helps to 2059171482Sjeff * worsen the priority and interactivity of processes which schedule batch 2060171482Sjeff * jobs such as make. This has little effect on the make process itself but 2061171482Sjeff * causes new processes spawned by it to receive worse scores immediately. 2062171482Sjeff */ 2063113372Sjeffvoid 2064164939Sjuliansched_exit_thread(struct thread *td, struct thread *child) 2065164936Sjulian{ 2066165762Sjeff 2067164939Sjulian CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d", 2068165762Sjeff child, child->td_proc->p_comm, child->td_priority); 2069164939Sjulian 2070165762Sjeff#ifdef KSE 2071165762Sjeff /* 2072165762Sjeff * KSE forks and exits so often that this penalty causes short-lived 2073165762Sjeff * threads to always be non-interactive. This causes mozilla to 2074165762Sjeff * crawl under load. 2075165762Sjeff */ 2076165762Sjeff if ((td->td_pflags & TDP_SA) && td->td_proc == child->td_proc) 2077165762Sjeff return; 2078165762Sjeff#endif 2079165762Sjeff /* 2080165762Sjeff * Give the child's runtime to the parent without returning the 2081165762Sjeff * sleep time as a penalty to the parent. This causes shells that 2082165762Sjeff * launch expensive things to mark their children as expensive. 2083165762Sjeff */ 2084170293Sjeff thread_lock(td); 2085171482Sjeff td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2086164939Sjulian sched_interact_update(td); 2087165762Sjeff sched_priority(td); 2088170293Sjeff thread_unlock(td); 2089164936Sjulian} 2090164936Sjulian 2091171482Sjeff/* 2092171482Sjeff * Fix priorities on return to user-space. Priorities may be elevated due 2093171482Sjeff * to static priorities in msleep() or similar. 2094171482Sjeff */ 2095164936Sjulianvoid 2096164936Sjuliansched_userret(struct thread *td) 2097164936Sjulian{ 2098164936Sjulian /* 2099164936Sjulian * XXX we cheat slightly on the locking here to avoid locking in 2100164936Sjulian * the usual case. Setting td_priority here is essentially an 2101164936Sjulian * incomplete workaround for not setting it properly elsewhere. 2102164936Sjulian * Now that some interrupt handlers are threads, not setting it 2103164936Sjulian * properly elsewhere can clobber it in the window between setting 2104164936Sjulian * it here and returning to user mode, so don't waste time setting 2105164936Sjulian * it perfectly here. 2106164936Sjulian */ 2107164936Sjulian KASSERT((td->td_flags & TDF_BORROWING) == 0, 2108164936Sjulian ("thread with borrowed priority returning to userland")); 2109164936Sjulian if (td->td_priority != td->td_user_pri) { 2110170293Sjeff thread_lock(td); 2111164936Sjulian td->td_priority = td->td_user_pri; 2112164936Sjulian td->td_base_pri = td->td_user_pri; 2113170293Sjeff thread_unlock(td); 2114164936Sjulian } 2115164936Sjulian} 2116164936Sjulian 2117171482Sjeff/* 2118171482Sjeff * Handle a stathz tick. This is really only relevant for timeshare 2119171482Sjeff * threads. 2120171482Sjeff */ 2121164936Sjulianvoid 2122121127Sjeffsched_clock(struct thread *td) 2123109864Sjeff{ 2124164936Sjulian struct tdq *tdq; 2125164936Sjulian struct td_sched *ts; 2126109864Sjeff 2127171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2128164936Sjulian tdq = TDQ_SELF(); 2129133427Sjeff /* 2130165766Sjeff * Advance the insert index once for each tick to ensure that all 2131165766Sjeff * threads get a chance to run. 2132133427Sjeff */ 2133165766Sjeff if (tdq->tdq_idx == tdq->tdq_ridx) { 2134165766Sjeff tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2135165766Sjeff if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2136165766Sjeff tdq->tdq_ridx = tdq->tdq_idx; 2137165766Sjeff } 2138165766Sjeff ts = td->td_sched; 2139165762Sjeff /* 2140163709Sjb * We only do slicing code for TIMESHARE threads. 2141113357Sjeff */ 2142163709Sjb if (td->td_pri_class != PRI_TIMESHARE) 2143113357Sjeff return; 2144113357Sjeff /* 2145165766Sjeff * We used a tick; charge it to the thread so that we can compute our 2146113357Sjeff * interactivity. 2147109864Sjeff */ 2148171482Sjeff td->td_sched->ts_runtime += tickincr; 2149163709Sjb sched_interact_update(td); 2150109864Sjeff /* 2151109864Sjeff * We used up one time slice. 2152109864Sjeff */ 2153164936Sjulian if (--ts->ts_slice > 0) 2154113357Sjeff return; 2155109864Sjeff /* 2156113357Sjeff * We're out of time, recompute priorities and requeue. 2157109864Sjeff */ 2158165796Sjeff sched_priority(td); 2159113357Sjeff td->td_flags |= TDF_NEEDRESCHED; 2160109864Sjeff} 2161109864Sjeff 2162171482Sjeff/* 2163171482Sjeff * Called once per hz tick. Used for cpu utilization information. This 2164171482Sjeff * is easier than trying to scale based on stathz. 2165171482Sjeff */ 2166171482Sjeffvoid 2167171482Sjeffsched_tick(void) 2168171482Sjeff{ 2169171482Sjeff struct td_sched *ts; 2170171482Sjeff 2171171482Sjeff ts = curthread->td_sched; 2172171482Sjeff /* Adjust ticks for pctcpu */ 2173171482Sjeff ts->ts_ticks += 1 << SCHED_TICK_SHIFT; 2174171482Sjeff ts->ts_ltick = ticks; 2175171482Sjeff /* 2176171482Sjeff * Update if we've exceeded our desired tick threshhold by over one 2177171482Sjeff * second. 2178171482Sjeff */ 2179171482Sjeff if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2180171482Sjeff sched_pctcpu_update(ts); 2181171482Sjeff} 2182171482Sjeff 2183171482Sjeff/* 2184171482Sjeff * Return whether the current CPU has runnable tasks. Used for in-kernel 2185171482Sjeff * cooperative idle threads. 2186171482Sjeff */ 2187109864Sjeffint 2188109864Sjeffsched_runnable(void) 2189109864Sjeff{ 2190164936Sjulian struct tdq *tdq; 2191115998Sjeff int load; 2192109864Sjeff 2193115998Sjeff load = 1; 2194115998Sjeff 2195164936Sjulian tdq = TDQ_SELF(); 2196121605Sjeff if ((curthread->td_flags & TDF_IDLETD) != 0) { 2197165620Sjeff if (tdq->tdq_load > 0) 2198121605Sjeff goto out; 2199121605Sjeff } else 2200165620Sjeff if (tdq->tdq_load - 1 > 0) 2201121605Sjeff goto out; 2202115998Sjeff load = 0; 2203115998Sjeffout: 2204115998Sjeff return (load); 2205109864Sjeff} 2206109864Sjeff 2207171482Sjeff/* 2208171482Sjeff * Choose the highest priority thread to run. The thread is removed from 2209171482Sjeff * the run-queue while running however the load remains. For SMP we set 2210171482Sjeff * the tdq in the global idle bitmask if it idles here. 2211171482Sjeff */ 2212166190Sjeffstruct thread * 2213109970Sjeffsched_choose(void) 2214109970Sjeff{ 2215171482Sjeff#ifdef SMP 2216171482Sjeff struct tdq_group *tdg; 2217171482Sjeff#endif 2218171482Sjeff struct td_sched *ts; 2219164936Sjulian struct tdq *tdq; 2220109970Sjeff 2221164936Sjulian tdq = TDQ_SELF(); 2222171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2223164936Sjulian ts = tdq_choose(tdq); 2224164936Sjulian if (ts) { 2225164936Sjulian tdq_runq_rem(tdq, ts); 2226166190Sjeff return (ts->ts_thread); 2227109864Sjeff } 2228109970Sjeff#ifdef SMP 2229171482Sjeff /* 2230171482Sjeff * We only set the idled bit when all of the cpus in the group are 2231171482Sjeff * idle. Otherwise we could get into a situation where a thread bounces 2232171482Sjeff * back and forth between two idle cores on seperate physical CPUs. 2233171482Sjeff */ 2234171482Sjeff tdg = tdq->tdq_group; 2235171482Sjeff tdg->tdg_idlemask |= PCPU_GET(cpumask); 2236171482Sjeff if (tdg->tdg_idlemask == tdg->tdg_cpumask) 2237171482Sjeff atomic_set_int(&tdq_idle, tdg->tdg_mask); 2238171482Sjeff tdq->tdq_lowpri = PRI_MAX_IDLE; 2239109970Sjeff#endif 2240166190Sjeff return (PCPU_GET(idlethread)); 2241109864Sjeff} 2242109864Sjeff 2243171482Sjeff/* 2244171482Sjeff * Set owepreempt if necessary. Preemption never happens directly in ULE, 2245171482Sjeff * we always request it once we exit a critical section. 2246171482Sjeff */ 2247171482Sjeffstatic inline void 2248171482Sjeffsched_setpreempt(struct thread *td) 2249166190Sjeff{ 2250166190Sjeff struct thread *ctd; 2251166190Sjeff int cpri; 2252166190Sjeff int pri; 2253166190Sjeff 2254166190Sjeff ctd = curthread; 2255166190Sjeff pri = td->td_priority; 2256166190Sjeff cpri = ctd->td_priority; 2257171482Sjeff if (td->td_priority < ctd->td_priority) 2258171482Sjeff curthread->td_flags |= TDF_NEEDRESCHED; 2259166190Sjeff if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2260171482Sjeff return; 2261166190Sjeff /* 2262166190Sjeff * Always preempt IDLE threads. Otherwise only if the preempting 2263166190Sjeff * thread is an ithread. 2264166190Sjeff */ 2265171482Sjeff if (pri > preempt_thresh && cpri < PRI_MIN_IDLE) 2266171482Sjeff return; 2267171482Sjeff ctd->td_owepreempt = 1; 2268171482Sjeff return; 2269166190Sjeff} 2270166190Sjeff 2271171482Sjeff/* 2272171482Sjeff * Add a thread to a thread queue. Initializes priority, slice, runq, and 2273171482Sjeff * add it to the appropriate queue. This is the internal function called 2274171482Sjeff * when the tdq is predetermined. 2275171482Sjeff */ 2276109864Sjeffvoid 2277171482Sjefftdq_add(struct tdq *tdq, struct thread *td, int flags) 2278109864Sjeff{ 2279164936Sjulian struct td_sched *ts; 2280121790Sjeff int class; 2281166108Sjeff#ifdef SMP 2282166108Sjeff int cpumask; 2283166108Sjeff#endif 2284109864Sjeff 2285171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2286166190Sjeff KASSERT((td->td_inhibitors == 0), 2287166190Sjeff ("sched_add: trying to run inhibited thread")); 2288166190Sjeff KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2289166190Sjeff ("sched_add: bad thread state")); 2290172207Sjeff KASSERT(td->td_flags & TDF_INMEM, 2291172207Sjeff ("sched_add: thread swapped out")); 2292171482Sjeff 2293171482Sjeff ts = td->td_sched; 2294171482Sjeff class = PRI_BASE(td->td_pri_class); 2295166190Sjeff TD_SET_RUNQ(td); 2296166190Sjeff if (ts->ts_slice == 0) 2297166190Sjeff ts->ts_slice = sched_slice; 2298133427Sjeff /* 2299171482Sjeff * Pick the run queue based on priority. 2300133427Sjeff */ 2301171482Sjeff if (td->td_priority <= PRI_MAX_REALTIME) 2302171482Sjeff ts->ts_runq = &tdq->tdq_realtime; 2303171482Sjeff else if (td->td_priority <= PRI_MAX_TIMESHARE) 2304171482Sjeff ts->ts_runq = &tdq->tdq_timeshare; 2305171482Sjeff else 2306171482Sjeff ts->ts_runq = &tdq->tdq_idle; 2307171482Sjeff#ifdef SMP 2308166108Sjeff cpumask = 1 << ts->ts_cpu; 2309121790Sjeff /* 2310123685Sjeff * If we had been idle, clear our bit in the group and potentially 2311166108Sjeff * the global bitmap. 2312121790Sjeff */ 2313165762Sjeff if ((class != PRI_IDLE && class != PRI_ITHD) && 2314166108Sjeff (tdq->tdq_group->tdg_idlemask & cpumask) != 0) { 2315121790Sjeff /* 2316123433Sjeff * Check to see if our group is unidling, and if so, remove it 2317123433Sjeff * from the global idle mask. 2318121790Sjeff */ 2319165620Sjeff if (tdq->tdq_group->tdg_idlemask == 2320165620Sjeff tdq->tdq_group->tdg_cpumask) 2321165620Sjeff atomic_clear_int(&tdq_idle, tdq->tdq_group->tdg_mask); 2322123433Sjeff /* 2323123433Sjeff * Now remove ourselves from the group specific idle mask. 2324123433Sjeff */ 2325166108Sjeff tdq->tdq_group->tdg_idlemask &= ~cpumask; 2326166108Sjeff } 2327171482Sjeff if (td->td_priority < tdq->tdq_lowpri) 2328171482Sjeff tdq->tdq_lowpri = td->td_priority; 2329121790Sjeff#endif 2330171482Sjeff tdq_runq_add(tdq, ts, flags); 2331171482Sjeff tdq_load_add(tdq, ts); 2332171482Sjeff} 2333171482Sjeff 2334171482Sjeff/* 2335171482Sjeff * Select the target thread queue and add a thread to it. Request 2336171482Sjeff * preemption or IPI a remote processor if required. 2337171482Sjeff */ 2338171482Sjeffvoid 2339171482Sjeffsched_add(struct thread *td, int flags) 2340171482Sjeff{ 2341171482Sjeff struct td_sched *ts; 2342171482Sjeff struct tdq *tdq; 2343171482Sjeff#ifdef SMP 2344171482Sjeff int cpuid; 2345171482Sjeff int cpu; 2346171482Sjeff#endif 2347171482Sjeff CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 2348171482Sjeff td, td->td_proc->p_comm, td->td_priority, curthread, 2349171482Sjeff curthread->td_proc->p_comm); 2350171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2351171482Sjeff ts = td->td_sched; 2352166108Sjeff /* 2353171482Sjeff * Recalculate the priority before we select the target cpu or 2354171482Sjeff * run-queue. 2355166108Sjeff */ 2356171482Sjeff if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2357171482Sjeff sched_priority(td); 2358171482Sjeff#ifdef SMP 2359171482Sjeff cpuid = PCPU_GET(cpuid); 2360171482Sjeff /* 2361171482Sjeff * Pick the destination cpu and if it isn't ours transfer to the 2362171482Sjeff * target cpu. 2363171482Sjeff */ 2364171482Sjeff if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_MIGRATE(td)) 2365171482Sjeff cpu = cpuid; 2366171482Sjeff else if (!THREAD_CAN_MIGRATE(td)) 2367171482Sjeff cpu = ts->ts_cpu; 2368166108Sjeff else 2369171482Sjeff cpu = sched_pickcpu(ts, flags); 2370171482Sjeff tdq = sched_setcpu(ts, cpu, flags); 2371171482Sjeff tdq_add(tdq, td, flags); 2372171482Sjeff if (cpu != cpuid) { 2373166108Sjeff tdq_notify(ts); 2374166108Sjeff return; 2375166108Sjeff } 2376171482Sjeff#else 2377171482Sjeff tdq = TDQ_SELF(); 2378171482Sjeff TDQ_LOCK(tdq); 2379171482Sjeff /* 2380171482Sjeff * Now that the thread is moving to the run-queue, set the lock 2381171482Sjeff * to the scheduler's lock. 2382171482Sjeff */ 2383171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2384171482Sjeff tdq_add(tdq, td, flags); 2385166108Sjeff#endif 2386171482Sjeff if (!(flags & SRQ_YIELDING)) 2387171482Sjeff sched_setpreempt(td); 2388109864Sjeff} 2389109864Sjeff 2390171482Sjeff/* 2391171482Sjeff * Remove a thread from a run-queue without running it. This is used 2392171482Sjeff * when we're stealing a thread from a remote queue. Otherwise all threads 2393171482Sjeff * exit by calling sched_exit_thread() and sched_throw() themselves. 2394171482Sjeff */ 2395109864Sjeffvoid 2396121127Sjeffsched_rem(struct thread *td) 2397109864Sjeff{ 2398164936Sjulian struct tdq *tdq; 2399164936Sjulian struct td_sched *ts; 2400113357Sjeff 2401139316Sjeff CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 2402139316Sjeff td, td->td_proc->p_comm, td->td_priority, curthread, 2403139316Sjeff curthread->td_proc->p_comm); 2404164936Sjulian ts = td->td_sched; 2405171482Sjeff tdq = TDQ_CPU(ts->ts_cpu); 2406171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2407171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2408166190Sjeff KASSERT(TD_ON_RUNQ(td), 2409164936Sjulian ("sched_rem: thread not on run queue")); 2410164936Sjulian tdq_runq_rem(tdq, ts); 2411164936Sjulian tdq_load_rem(tdq, ts); 2412166190Sjeff TD_SET_CAN_RUN(td); 2413109864Sjeff} 2414109864Sjeff 2415171482Sjeff/* 2416171482Sjeff * Fetch cpu utilization information. Updates on demand. 2417171482Sjeff */ 2418109864Sjefffixpt_t 2419121127Sjeffsched_pctcpu(struct thread *td) 2420109864Sjeff{ 2421109864Sjeff fixpt_t pctcpu; 2422164936Sjulian struct td_sched *ts; 2423109864Sjeff 2424109864Sjeff pctcpu = 0; 2425164936Sjulian ts = td->td_sched; 2426164936Sjulian if (ts == NULL) 2427121290Sjeff return (0); 2428109864Sjeff 2429170293Sjeff thread_lock(td); 2430164936Sjulian if (ts->ts_ticks) { 2431109864Sjeff int rtick; 2432109864Sjeff 2433165796Sjeff sched_pctcpu_update(ts); 2434109864Sjeff /* How many rtick per second ? */ 2435165762Sjeff rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2436165762Sjeff pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2437109864Sjeff } 2438164936Sjulian td->td_proc->p_swtime = ts->ts_ltick - ts->ts_ftick; 2439170293Sjeff thread_unlock(td); 2440109864Sjeff 2441109864Sjeff return (pctcpu); 2442109864Sjeff} 2443109864Sjeff 2444171482Sjeff/* 2445171482Sjeff * Bind a thread to a target cpu. 2446171482Sjeff */ 2447122038Sjeffvoid 2448122038Sjeffsched_bind(struct thread *td, int cpu) 2449122038Sjeff{ 2450164936Sjulian struct td_sched *ts; 2451122038Sjeff 2452171713Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2453164936Sjulian ts = td->td_sched; 2454166137Sjeff if (ts->ts_flags & TSF_BOUND) 2455166152Sjeff sched_unbind(td); 2456164936Sjulian ts->ts_flags |= TSF_BOUND; 2457123433Sjeff#ifdef SMP 2458166137Sjeff sched_pin(); 2459123433Sjeff if (PCPU_GET(cpuid) == cpu) 2460122038Sjeff return; 2461166137Sjeff ts->ts_cpu = cpu; 2462122038Sjeff /* When we return from mi_switch we'll be on the correct cpu. */ 2463131527Sphk mi_switch(SW_VOL, NULL); 2464122038Sjeff#endif 2465122038Sjeff} 2466122038Sjeff 2467171482Sjeff/* 2468171482Sjeff * Release a bound thread. 2469171482Sjeff */ 2470122038Sjeffvoid 2471122038Sjeffsched_unbind(struct thread *td) 2472122038Sjeff{ 2473165762Sjeff struct td_sched *ts; 2474165762Sjeff 2475170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2476165762Sjeff ts = td->td_sched; 2477166137Sjeff if ((ts->ts_flags & TSF_BOUND) == 0) 2478166137Sjeff return; 2479165762Sjeff ts->ts_flags &= ~TSF_BOUND; 2480165762Sjeff#ifdef SMP 2481165762Sjeff sched_unpin(); 2482165762Sjeff#endif 2483122038Sjeff} 2484122038Sjeff 2485109864Sjeffint 2486145256Sjkoshysched_is_bound(struct thread *td) 2487145256Sjkoshy{ 2488170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2489164936Sjulian return (td->td_sched->ts_flags & TSF_BOUND); 2490145256Sjkoshy} 2491145256Sjkoshy 2492171482Sjeff/* 2493171482Sjeff * Basic yield call. 2494171482Sjeff */ 2495159630Sdavidxuvoid 2496159630Sdavidxusched_relinquish(struct thread *td) 2497159630Sdavidxu{ 2498170293Sjeff thread_lock(td); 2499163709Sjb if (td->td_pri_class == PRI_TIMESHARE) 2500159630Sdavidxu sched_prio(td, PRI_MAX_TIMESHARE); 2501170293Sjeff SCHED_STAT_INC(switch_relinquish); 2502159630Sdavidxu mi_switch(SW_VOL, NULL); 2503170293Sjeff thread_unlock(td); 2504159630Sdavidxu} 2505159630Sdavidxu 2506171482Sjeff/* 2507171482Sjeff * Return the total system load. 2508171482Sjeff */ 2509145256Sjkoshyint 2510125289Sjeffsched_load(void) 2511125289Sjeff{ 2512125289Sjeff#ifdef SMP 2513125289Sjeff int total; 2514125289Sjeff int i; 2515125289Sjeff 2516125289Sjeff total = 0; 2517165620Sjeff for (i = 0; i <= tdg_maxid; i++) 2518165620Sjeff total += TDQ_GROUP(i)->tdg_load; 2519125289Sjeff return (total); 2520125289Sjeff#else 2521165620Sjeff return (TDQ_SELF()->tdq_sysload); 2522125289Sjeff#endif 2523125289Sjeff} 2524125289Sjeff 2525125289Sjeffint 2526109864Sjeffsched_sizeof_proc(void) 2527109864Sjeff{ 2528109864Sjeff return (sizeof(struct proc)); 2529109864Sjeff} 2530109864Sjeff 2531109864Sjeffint 2532109864Sjeffsched_sizeof_thread(void) 2533109864Sjeff{ 2534109864Sjeff return (sizeof(struct thread) + sizeof(struct td_sched)); 2535109864Sjeff} 2536159570Sdavidxu 2537166190Sjeff/* 2538166190Sjeff * The actual idle process. 2539166190Sjeff */ 2540166190Sjeffvoid 2541166190Sjeffsched_idletd(void *dummy) 2542166190Sjeff{ 2543166190Sjeff struct thread *td; 2544171482Sjeff struct tdq *tdq; 2545166190Sjeff 2546166190Sjeff td = curthread; 2547171482Sjeff tdq = TDQ_SELF(); 2548166190Sjeff mtx_assert(&Giant, MA_NOTOWNED); 2549171482Sjeff /* ULE relies on preemption for idle interruption. */ 2550171482Sjeff for (;;) { 2551171482Sjeff#ifdef SMP 2552171482Sjeff if (tdq_idled(tdq)) 2553171482Sjeff cpu_idle(); 2554171482Sjeff#else 2555166190Sjeff cpu_idle(); 2556171482Sjeff#endif 2557171482Sjeff } 2558166190Sjeff} 2559166190Sjeff 2560170293Sjeff/* 2561170293Sjeff * A CPU is entering for the first time or a thread is exiting. 2562170293Sjeff */ 2563170293Sjeffvoid 2564170293Sjeffsched_throw(struct thread *td) 2565170293Sjeff{ 2566171482Sjeff struct tdq *tdq; 2567171482Sjeff 2568171482Sjeff tdq = TDQ_SELF(); 2569170293Sjeff if (td == NULL) { 2570171482Sjeff /* Correct spinlock nesting and acquire the correct lock. */ 2571171482Sjeff TDQ_LOCK(tdq); 2572170293Sjeff spinlock_exit(); 2573170293Sjeff } else { 2574171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2575171482Sjeff tdq_load_rem(tdq, td->td_sched); 2576170293Sjeff } 2577170293Sjeff KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2578170293Sjeff PCPU_SET(switchtime, cpu_ticks()); 2579170293Sjeff PCPU_SET(switchticks, ticks); 2580170293Sjeff cpu_throw(td, choosethread()); /* doesn't return */ 2581170293Sjeff} 2582170293Sjeff 2583171482Sjeff/* 2584171482Sjeff * This is called from fork_exit(). Just acquire the correct locks and 2585171482Sjeff * let fork do the rest of the work. 2586171482Sjeff */ 2587170293Sjeffvoid 2588170600Sjeffsched_fork_exit(struct thread *td) 2589170293Sjeff{ 2590171482Sjeff struct td_sched *ts; 2591171482Sjeff struct tdq *tdq; 2592171482Sjeff int cpuid; 2593170293Sjeff 2594170293Sjeff /* 2595170293Sjeff * Finish setting up thread glue so that it begins execution in a 2596171482Sjeff * non-nested critical section with the scheduler lock held. 2597170293Sjeff */ 2598171482Sjeff cpuid = PCPU_GET(cpuid); 2599171482Sjeff tdq = TDQ_CPU(cpuid); 2600171482Sjeff ts = td->td_sched; 2601171482Sjeff if (TD_IS_IDLETHREAD(td)) 2602171482Sjeff td->td_lock = TDQ_LOCKPTR(tdq); 2603171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2604171482Sjeff td->td_oncpu = cpuid; 2605171482Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)td; 2606170600Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED); 2607170293Sjeff} 2608170293Sjeff 2609171482Sjeffstatic SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, 2610171482Sjeff "Scheduler"); 2611171482SjeffSYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2612165762Sjeff "Scheduler name"); 2613171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2614171482Sjeff "Slice size for timeshare threads"); 2615171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2616171482Sjeff "Interactivity score threshold"); 2617171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2618171482Sjeff 0,"Min priority for preemption, lower priorities have greater precedence"); 2619166108Sjeff#ifdef SMP 2620171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, pick_pri, CTLFLAG_RW, &pick_pri, 0, 2621171482Sjeff "Pick the target cpu based on priority rather than load."); 2622171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2623171482Sjeff "Number of hz ticks to keep thread affinity for"); 2624171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, tryself, CTLFLAG_RW, &tryself, 0, ""); 2625171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2626171482Sjeff "Enables the long-term load balancer"); 2627171506SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance_secs, CTLFLAG_RW, &balance_secs, 0, 2628171506Sjeff "Average frequence in seconds to run the long-term balancer"); 2629171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, 2630171482Sjeff "Steals work from another hyper-threaded core on idle"); 2631171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2632171482Sjeff "Attempts to steal work from other cores before idling"); 2633171506SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2634171506Sjeff "Minimum load on remote cpu before we'll steal"); 2635171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, topology, CTLFLAG_RD, &topology, 0, 2636171482Sjeff "True when a topology has been specified by the MD code."); 2637166108Sjeff#endif 2638165762Sjeff 2639165762Sjeff/* ps compat */ 2640165762Sjeffstatic fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 2641165762SjeffSYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2642165762Sjeff 2643165762Sjeff 2644134791Sjulian#define KERN_SWITCH_INCLUDE 1 2645134791Sjulian#include "kern/kern_switch.c" 2646