sched_ule.c revision 242852
1109864Sjeff/*- 2165762Sjeff * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org> 3109864Sjeff * All rights reserved. 4109864Sjeff * 5109864Sjeff * Redistribution and use in source and binary forms, with or without 6109864Sjeff * modification, are permitted provided that the following conditions 7109864Sjeff * are met: 8109864Sjeff * 1. Redistributions of source code must retain the above copyright 9109864Sjeff * notice unmodified, this list of conditions, and the following 10109864Sjeff * disclaimer. 11109864Sjeff * 2. Redistributions in binary form must reproduce the above copyright 12109864Sjeff * notice, this list of conditions and the following disclaimer in the 13109864Sjeff * documentation and/or other materials provided with the distribution. 14109864Sjeff * 15109864Sjeff * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 16109864Sjeff * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 17109864Sjeff * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 18109864Sjeff * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 19109864Sjeff * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 20109864Sjeff * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 21109864Sjeff * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 22109864Sjeff * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 23109864Sjeff * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 24109864Sjeff * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 25109864Sjeff */ 26109864Sjeff 27171482Sjeff/* 28171482Sjeff * This file implements the ULE scheduler. ULE supports independent CPU 29171482Sjeff * run queues and fine grain locking. It has superior interactive 30171482Sjeff * performance under load even on uni-processor systems. 31171482Sjeff * 32171482Sjeff * etymology: 33172293Sjeff * ULE is the last three letters in schedule. It owes its name to a 34171482Sjeff * generic user created for a scheduling system by Paul Mikesell at 35171482Sjeff * Isilon Systems and a general lack of creativity on the part of the author. 36171482Sjeff */ 37171482Sjeff 38116182Sobrien#include <sys/cdefs.h> 39191645Sjeff__FBSDID("$FreeBSD: head/sys/kern/sched_ule.c 242852 2012-11-10 07:02:57Z mav $"); 40116182Sobrien 41147565Speter#include "opt_hwpmc_hooks.h" 42179297Sjb#include "opt_kdtrace.h" 43147565Speter#include "opt_sched.h" 44134649Sscottl 45109864Sjeff#include <sys/param.h> 46109864Sjeff#include <sys/systm.h> 47131929Smarcel#include <sys/kdb.h> 48109864Sjeff#include <sys/kernel.h> 49109864Sjeff#include <sys/ktr.h> 50109864Sjeff#include <sys/lock.h> 51109864Sjeff#include <sys/mutex.h> 52109864Sjeff#include <sys/proc.h> 53112966Sjeff#include <sys/resource.h> 54122038Sjeff#include <sys/resourcevar.h> 55109864Sjeff#include <sys/sched.h> 56235459Srstone#include <sys/sdt.h> 57109864Sjeff#include <sys/smp.h> 58109864Sjeff#include <sys/sx.h> 59109864Sjeff#include <sys/sysctl.h> 60109864Sjeff#include <sys/sysproto.h> 61139453Sjhb#include <sys/turnstile.h> 62161599Sdavidxu#include <sys/umtx.h> 63109864Sjeff#include <sys/vmmeter.h> 64176735Sjeff#include <sys/cpuset.h> 65184439Sivoras#include <sys/sbuf.h> 66109864Sjeff 67145256Sjkoshy#ifdef HWPMC_HOOKS 68145256Sjkoshy#include <sys/pmckern.h> 69145256Sjkoshy#endif 70145256Sjkoshy 71179297Sjb#ifdef KDTRACE_HOOKS 72179297Sjb#include <sys/dtrace_bsd.h> 73179297Sjbint dtrace_vtime_active; 74179297Sjbdtrace_vtime_switch_func_t dtrace_vtime_switch_func; 75179297Sjb#endif 76179297Sjb 77109864Sjeff#include <machine/cpu.h> 78121790Sjeff#include <machine/smp.h> 79109864Sjeff 80236141Sraj#if defined(__powerpc__) && defined(BOOKE_E500) 81172345Sjeff#error "This architecture is not currently compatible with ULE" 82166190Sjeff#endif 83166190Sjeff 84171482Sjeff#define KTR_ULE 0 85166137Sjeff 86187679Sjeff#define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX))) 87187679Sjeff#define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU))) 88224221Sattilio#define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load")) 89187357Sjeff 90166137Sjeff/* 91171482Sjeff * Thread scheduler specific section. All fields are protected 92171482Sjeff * by the thread lock. 93146954Sjeff */ 94164936Sjulianstruct td_sched { 95171482Sjeff struct runq *ts_runq; /* Run-queue we're queued on. */ 96171482Sjeff short ts_flags; /* TSF_* flags. */ 97164936Sjulian u_char ts_cpu; /* CPU that we have affinity for. */ 98177009Sjeff int ts_rltick; /* Real last tick, for affinity. */ 99171482Sjeff int ts_slice; /* Ticks of slice remaining. */ 100171482Sjeff u_int ts_slptime; /* Number of ticks we vol. slept */ 101171482Sjeff u_int ts_runtime; /* Number of ticks we were running */ 102164936Sjulian int ts_ltick; /* Last tick that we were running on */ 103164936Sjulian int ts_ftick; /* First tick that we were running on */ 104164936Sjulian int ts_ticks; /* Tick count */ 105187357Sjeff#ifdef KTR 106187357Sjeff char ts_name[TS_NAME_LEN]; 107187357Sjeff#endif 108134791Sjulian}; 109164936Sjulian/* flags kept in ts_flags */ 110166108Sjeff#define TSF_BOUND 0x0001 /* Thread can not migrate. */ 111166108Sjeff#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 112121790Sjeff 113164936Sjulianstatic struct td_sched td_sched0; 114109864Sjeff 115176735Sjeff#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 116176735Sjeff#define THREAD_CAN_SCHED(td, cpu) \ 117176735Sjeff CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 118176735Sjeff 119109864Sjeff/* 120217351Sjhb * Priority ranges used for interactive and non-interactive timeshare 121217410Sjhb * threads. The timeshare priorities are split up into four ranges. 122217410Sjhb * The first range handles interactive threads. The last three ranges 123217410Sjhb * (NHALF, x, and NHALF) handle non-interactive threads with the outer 124217410Sjhb * ranges supporting nice values. 125217351Sjhb */ 126217410Sjhb#define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1) 127217410Sjhb#define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2) 128228718Savg#define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE) 129217410Sjhb 130217410Sjhb#define PRI_MIN_INTERACT PRI_MIN_TIMESHARE 131217410Sjhb#define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1) 132217410Sjhb#define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE) 133217351Sjhb#define PRI_MAX_BATCH PRI_MAX_TIMESHARE 134217351Sjhb 135217351Sjhb/* 136165762Sjeff * Cpu percentage computation macros and defines. 137111857Sjeff * 138165762Sjeff * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 139165762Sjeff * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 140165796Sjeff * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 141165762Sjeff * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 142165762Sjeff * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 143165762Sjeff * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 144165762Sjeff */ 145165762Sjeff#define SCHED_TICK_SECS 10 146165762Sjeff#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 147165796Sjeff#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 148165762Sjeff#define SCHED_TICK_SHIFT 10 149165762Sjeff#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 150165830Sjeff#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 151165762Sjeff 152165762Sjeff/* 153165762Sjeff * These macros determine priorities for non-interactive threads. They are 154165762Sjeff * assigned a priority based on their recent cpu utilization as expressed 155165762Sjeff * by the ratio of ticks to the tick total. NHALF priorities at the start 156165762Sjeff * and end of the MIN to MAX timeshare range are only reachable with negative 157165762Sjeff * or positive nice respectively. 158165762Sjeff * 159165762Sjeff * PRI_RANGE: Priority range for utilization dependent priorities. 160116642Sjeff * PRI_NRESV: Number of nice values. 161165762Sjeff * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 162165762Sjeff * PRI_NICE: Determines the part of the priority inherited from nice. 163109864Sjeff */ 164165762Sjeff#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 165121869Sjeff#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 166217351Sjhb#define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF) 167217351Sjhb#define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF) 168217237Sjhb#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1) 169165762Sjeff#define SCHED_PRI_TICKS(ts) \ 170165762Sjeff (SCHED_TICK_HZ((ts)) / \ 171165827Sjeff (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 172165762Sjeff#define SCHED_PRI_NICE(nice) (nice) 173109864Sjeff 174109864Sjeff/* 175165762Sjeff * These determine the interactivity of a process. Interactivity differs from 176165762Sjeff * cpu utilization in that it expresses the voluntary time slept vs time ran 177165762Sjeff * while cpu utilization includes all time not running. This more accurately 178165762Sjeff * models the intent of the thread. 179109864Sjeff * 180110645Sjeff * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 181110645Sjeff * before throttling back. 182121868Sjeff * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 183116365Sjeff * INTERACT_MAX: Maximum interactivity value. Smaller is better. 184215102Sattilio * INTERACT_THRESH: Threshold for placement on the current runq. 185109864Sjeff */ 186165762Sjeff#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 187165762Sjeff#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 188116365Sjeff#define SCHED_INTERACT_MAX (100) 189116365Sjeff#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 190121126Sjeff#define SCHED_INTERACT_THRESH (30) 191111857Sjeff 192242736Sjeff/* 193242736Sjeff * These parameters determine the slice behavior for batch work. 194242736Sjeff */ 195242736Sjeff#define SCHED_SLICE_DEFAULT_DIVISOR 10 /* ~94 ms, 12 stathz ticks. */ 196242736Sjeff#define SCHED_SLICE_MIN_DIVISOR 6 /* DEFAULT/MIN = ~16 ms. */ 197242736Sjeff 198239157Smav/* Flags kept in td_flags. */ 199239157Smav#define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */ 200239157Smav 201109864Sjeff/* 202165762Sjeff * tickincr: Converts a stathz tick into a hz domain scaled by 203165762Sjeff * the shift factor. Without the shift the error rate 204165762Sjeff * due to rounding would be unacceptably high. 205165762Sjeff * realstathz: stathz is sometimes 0 and run off of hz. 206165762Sjeff * sched_slice: Runtime of each thread before rescheduling. 207171482Sjeff * preempt_thresh: Priority threshold for preemption and remote IPIs. 208109864Sjeff */ 209165762Sjeffstatic int sched_interact = SCHED_INTERACT_THRESH; 210241844Seadlerstatic int tickincr = 8 << SCHED_TICK_SHIFT; 211242736Sjeffstatic int realstathz = 127; /* reset during boot. */ 212242736Sjeffstatic int sched_slice = 10; /* reset during boot. */ 213242736Sjeffstatic int sched_slice_min = 1; /* reset during boot. */ 214172345Sjeff#ifdef PREEMPTION 215172345Sjeff#ifdef FULL_PREEMPTION 216172345Sjeffstatic int preempt_thresh = PRI_MAX_IDLE; 217172345Sjeff#else 218171482Sjeffstatic int preempt_thresh = PRI_MIN_KERN; 219172345Sjeff#endif 220172345Sjeff#else 221172345Sjeffstatic int preempt_thresh = 0; 222172345Sjeff#endif 223217351Sjhbstatic int static_boost = PRI_MIN_BATCH; 224178277Sjeffstatic int sched_idlespins = 10000; 225232740Smavstatic int sched_idlespinthresh = -1; 226109864Sjeff 227109864Sjeff/* 228171482Sjeff * tdq - per processor runqs and statistics. All fields are protected by the 229171482Sjeff * tdq_lock. The load and lowpri may be accessed without to avoid excess 230171482Sjeff * locking in sched_pickcpu(); 231109864Sjeff */ 232164936Sjulianstruct tdq { 233242014Sjimharris /* 234242014Sjimharris * Ordered to improve efficiency of cpu_search() and switch(). 235242014Sjimharris * tdq_lock is padded to avoid false sharing with tdq_load and 236242014Sjimharris * tdq_cpu_idle. 237242014Sjimharris */ 238242402Sattilio struct mtx_padalign tdq_lock; /* run queue lock. */ 239176735Sjeff struct cpu_group *tdq_cg; /* Pointer to cpu topology. */ 240178277Sjeff volatile int tdq_load; /* Aggregate load. */ 241212416Smav volatile int tdq_cpu_idle; /* cpu_idle() is active. */ 242176735Sjeff int tdq_sysload; /* For loadavg, !ITHD load. */ 243177009Sjeff int tdq_transferable; /* Transferable thread count. */ 244178277Sjeff short tdq_switchcnt; /* Switches this tick. */ 245178277Sjeff short tdq_oldswitchcnt; /* Switches last tick. */ 246177009Sjeff u_char tdq_lowpri; /* Lowest priority thread. */ 247177009Sjeff u_char tdq_ipipending; /* IPI pending. */ 248166557Sjeff u_char tdq_idx; /* Current insert index. */ 249166557Sjeff u_char tdq_ridx; /* Current removal index. */ 250177009Sjeff struct runq tdq_realtime; /* real-time run queue. */ 251177009Sjeff struct runq tdq_timeshare; /* timeshare run queue. */ 252177009Sjeff struct runq tdq_idle; /* Queue of IDLE threads. */ 253187357Sjeff char tdq_name[TDQ_NAME_LEN]; 254187357Sjeff#ifdef KTR 255187357Sjeff char tdq_loadname[TDQ_LOADNAME_LEN]; 256187357Sjeff#endif 257171482Sjeff} __aligned(64); 258109864Sjeff 259178277Sjeff/* Idle thread states and config. */ 260178277Sjeff#define TDQ_RUNNING 1 261178277Sjeff#define TDQ_IDLE 2 262166108Sjeff 263123433Sjeff#ifdef SMP 264184439Sivorasstruct cpu_group *cpu_top; /* CPU topology */ 265123433Sjeff 266176735Sjeff#define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000)) 267176735Sjeff#define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity)) 268166108Sjeff 269123433Sjeff/* 270166108Sjeff * Run-time tunables. 271166108Sjeff */ 272171506Sjeffstatic int rebalance = 1; 273172409Sjeffstatic int balance_interval = 128; /* Default set in sched_initticks(). */ 274166108Sjeffstatic int affinity; 275171506Sjeffstatic int steal_idle = 1; 276171506Sjeffstatic int steal_thresh = 2; 277166108Sjeff 278166108Sjeff/* 279165620Sjeff * One thread queue per processor. 280109864Sjeff */ 281164936Sjulianstatic struct tdq tdq_cpu[MAXCPU]; 282172409Sjeffstatic struct tdq *balance_tdq; 283172409Sjeffstatic int balance_ticks; 284232207Smavstatic DPCPU_DEFINE(uint32_t, randomval); 285129982Sjeff 286164936Sjulian#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 287164936Sjulian#define TDQ_CPU(x) (&tdq_cpu[(x)]) 288171713Sjeff#define TDQ_ID(x) ((int)((x) - tdq_cpu)) 289123433Sjeff#else /* !SMP */ 290164936Sjulianstatic struct tdq tdq_cpu; 291129982Sjeff 292170315Sjeff#define TDQ_ID(x) (0) 293164936Sjulian#define TDQ_SELF() (&tdq_cpu) 294164936Sjulian#define TDQ_CPU(x) (&tdq_cpu) 295110028Sjeff#endif 296109864Sjeff 297171482Sjeff#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 298171482Sjeff#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 299171482Sjeff#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 300171482Sjeff#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 301242402Sattilio#define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock)) 302171482Sjeff 303163709Sjbstatic void sched_priority(struct thread *); 304146954Sjeffstatic void sched_thread_priority(struct thread *, u_char); 305163709Sjbstatic int sched_interact_score(struct thread *); 306163709Sjbstatic void sched_interact_update(struct thread *); 307163709Sjbstatic void sched_interact_fork(struct thread *); 308232917Smavstatic void sched_pctcpu_update(struct td_sched *, int); 309109864Sjeff 310110267Sjeff/* Operations on per processor queues */ 311177435Sjeffstatic struct thread *tdq_choose(struct tdq *); 312164936Sjulianstatic void tdq_setup(struct tdq *); 313177435Sjeffstatic void tdq_load_add(struct tdq *, struct thread *); 314177435Sjeffstatic void tdq_load_rem(struct tdq *, struct thread *); 315177435Sjeffstatic __inline void tdq_runq_add(struct tdq *, struct thread *, int); 316177435Sjeffstatic __inline void tdq_runq_rem(struct tdq *, struct thread *); 317177005Sjeffstatic inline int sched_shouldpreempt(int, int, int); 318164936Sjulianvoid tdq_print(int cpu); 319165762Sjeffstatic void runq_print(struct runq *rq); 320171482Sjeffstatic void tdq_add(struct tdq *, struct thread *, int); 321110267Sjeff#ifdef SMP 322176735Sjeffstatic int tdq_move(struct tdq *, struct tdq *); 323171482Sjeffstatic int tdq_idled(struct tdq *); 324177435Sjeffstatic void tdq_notify(struct tdq *, struct thread *); 325177435Sjeffstatic struct thread *tdq_steal(struct tdq *, int); 326177435Sjeffstatic struct thread *runq_steal(struct runq *, int); 327177435Sjeffstatic int sched_pickcpu(struct thread *, int); 328172409Sjeffstatic void sched_balance(void); 329176735Sjeffstatic int sched_balance_pair(struct tdq *, struct tdq *); 330177435Sjeffstatic inline struct tdq *sched_setcpu(struct thread *, int, int); 331171482Sjeffstatic inline void thread_unblock_switch(struct thread *, struct mtx *); 332171713Sjeffstatic struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 333184439Sivorasstatic int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS); 334184439Sivorasstatic int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, 335184439Sivoras struct cpu_group *cg, int indent); 336121790Sjeff#endif 337110028Sjeff 338165762Sjeffstatic void sched_setup(void *dummy); 339177253SrwatsonSYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 340165762Sjeff 341165762Sjeffstatic void sched_initticks(void *dummy); 342177253SrwatsonSYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 343177253Srwatson NULL); 344165762Sjeff 345235459SrstoneSDT_PROVIDER_DEFINE(sched); 346235459Srstone 347235459SrstoneSDT_PROBE_DEFINE3(sched, , , change_pri, change-pri, "struct thread *", 348235459Srstone "struct proc *", "uint8_t"); 349235459SrstoneSDT_PROBE_DEFINE3(sched, , , dequeue, dequeue, "struct thread *", 350235459Srstone "struct proc *", "void *"); 351235459SrstoneSDT_PROBE_DEFINE4(sched, , , enqueue, enqueue, "struct thread *", 352235459Srstone "struct proc *", "void *", "int"); 353235459SrstoneSDT_PROBE_DEFINE4(sched, , , lend_pri, lend-pri, "struct thread *", 354235459Srstone "struct proc *", "uint8_t", "struct thread *"); 355235459SrstoneSDT_PROBE_DEFINE2(sched, , , load_change, load-change, "int", "int"); 356235459SrstoneSDT_PROBE_DEFINE2(sched, , , off_cpu, off-cpu, "struct thread *", 357235459Srstone "struct proc *"); 358235459SrstoneSDT_PROBE_DEFINE(sched, , , on_cpu, on-cpu); 359235459SrstoneSDT_PROBE_DEFINE(sched, , , remain_cpu, remain-cpu); 360235459SrstoneSDT_PROBE_DEFINE2(sched, , , surrender, surrender, "struct thread *", 361235459Srstone "struct proc *"); 362235459Srstone 363171482Sjeff/* 364171482Sjeff * Print the threads waiting on a run-queue. 365171482Sjeff */ 366165762Sjeffstatic void 367165762Sjeffrunq_print(struct runq *rq) 368165762Sjeff{ 369165762Sjeff struct rqhead *rqh; 370177435Sjeff struct thread *td; 371165762Sjeff int pri; 372165762Sjeff int j; 373165762Sjeff int i; 374165762Sjeff 375165762Sjeff for (i = 0; i < RQB_LEN; i++) { 376165762Sjeff printf("\t\trunq bits %d 0x%zx\n", 377165762Sjeff i, rq->rq_status.rqb_bits[i]); 378165762Sjeff for (j = 0; j < RQB_BPW; j++) 379165762Sjeff if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 380165762Sjeff pri = j + (i << RQB_L2BPW); 381165762Sjeff rqh = &rq->rq_queues[pri]; 382177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) { 383165762Sjeff printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 384177435Sjeff td, td->td_name, td->td_priority, 385177435Sjeff td->td_rqindex, pri); 386165762Sjeff } 387165762Sjeff } 388165762Sjeff } 389165762Sjeff} 390165762Sjeff 391171482Sjeff/* 392171482Sjeff * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 393171482Sjeff */ 394113357Sjeffvoid 395164936Sjuliantdq_print(int cpu) 396110267Sjeff{ 397164936Sjulian struct tdq *tdq; 398112994Sjeff 399164936Sjulian tdq = TDQ_CPU(cpu); 400112994Sjeff 401171713Sjeff printf("tdq %d:\n", TDQ_ID(tdq)); 402176735Sjeff printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 403176735Sjeff printf("\tLock name: %s\n", tdq->tdq_name); 404165620Sjeff printf("\tload: %d\n", tdq->tdq_load); 405178277Sjeff printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt); 406178277Sjeff printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt); 407171482Sjeff printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 408165766Sjeff printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 409178277Sjeff printf("\tload transferable: %d\n", tdq->tdq_transferable); 410178277Sjeff printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 411165762Sjeff printf("\trealtime runq:\n"); 412165762Sjeff runq_print(&tdq->tdq_realtime); 413165762Sjeff printf("\ttimeshare runq:\n"); 414165762Sjeff runq_print(&tdq->tdq_timeshare); 415165762Sjeff printf("\tidle runq:\n"); 416165762Sjeff runq_print(&tdq->tdq_idle); 417113357Sjeff} 418112994Sjeff 419177005Sjeffstatic inline int 420177005Sjeffsched_shouldpreempt(int pri, int cpri, int remote) 421177005Sjeff{ 422177005Sjeff /* 423177005Sjeff * If the new priority is not better than the current priority there is 424177005Sjeff * nothing to do. 425177005Sjeff */ 426177005Sjeff if (pri >= cpri) 427177005Sjeff return (0); 428177005Sjeff /* 429177005Sjeff * Always preempt idle. 430177005Sjeff */ 431177005Sjeff if (cpri >= PRI_MIN_IDLE) 432177005Sjeff return (1); 433177005Sjeff /* 434177005Sjeff * If preemption is disabled don't preempt others. 435177005Sjeff */ 436177005Sjeff if (preempt_thresh == 0) 437177005Sjeff return (0); 438177005Sjeff /* 439177005Sjeff * Preempt if we exceed the threshold. 440177005Sjeff */ 441177005Sjeff if (pri <= preempt_thresh) 442177005Sjeff return (1); 443177005Sjeff /* 444217351Sjhb * If we're interactive or better and there is non-interactive 445217351Sjhb * or worse running preempt only remote processors. 446177005Sjeff */ 447217351Sjhb if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT) 448177005Sjeff return (1); 449177005Sjeff return (0); 450177005Sjeff} 451177005Sjeff 452171482Sjeff/* 453171482Sjeff * Add a thread to the actual run-queue. Keeps transferable counts up to 454171482Sjeff * date with what is actually on the run-queue. Selects the correct 455171482Sjeff * queue position for timeshare threads. 456171482Sjeff */ 457122744Sjeffstatic __inline void 458177435Sjefftdq_runq_add(struct tdq *tdq, struct thread *td, int flags) 459122744Sjeff{ 460177435Sjeff struct td_sched *ts; 461177042Sjeff u_char pri; 462177042Sjeff 463171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 464177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 465177009Sjeff 466177435Sjeff pri = td->td_priority; 467177435Sjeff ts = td->td_sched; 468177435Sjeff TD_SET_RUNQ(td); 469177435Sjeff if (THREAD_CAN_MIGRATE(td)) { 470165620Sjeff tdq->tdq_transferable++; 471164936Sjulian ts->ts_flags |= TSF_XFERABLE; 472123433Sjeff } 473217351Sjhb if (pri < PRI_MIN_BATCH) { 474177042Sjeff ts->ts_runq = &tdq->tdq_realtime; 475217351Sjhb } else if (pri <= PRI_MAX_BATCH) { 476177042Sjeff ts->ts_runq = &tdq->tdq_timeshare; 477217351Sjhb KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH, 478165762Sjeff ("Invalid priority %d on timeshare runq", pri)); 479165762Sjeff /* 480165762Sjeff * This queue contains only priorities between MIN and MAX 481165762Sjeff * realtime. Use the whole queue to represent these values. 482165762Sjeff */ 483171713Sjeff if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 484228718Savg pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE; 485165762Sjeff pri = (pri + tdq->tdq_idx) % RQ_NQS; 486165766Sjeff /* 487165766Sjeff * This effectively shortens the queue by one so we 488165766Sjeff * can have a one slot difference between idx and 489165766Sjeff * ridx while we wait for threads to drain. 490165766Sjeff */ 491165766Sjeff if (tdq->tdq_ridx != tdq->tdq_idx && 492165766Sjeff pri == tdq->tdq_ridx) 493167664Sjeff pri = (unsigned char)(pri - 1) % RQ_NQS; 494165762Sjeff } else 495165766Sjeff pri = tdq->tdq_ridx; 496177435Sjeff runq_add_pri(ts->ts_runq, td, pri, flags); 497177042Sjeff return; 498165762Sjeff } else 499177009Sjeff ts->ts_runq = &tdq->tdq_idle; 500177435Sjeff runq_add(ts->ts_runq, td, flags); 501177009Sjeff} 502177009Sjeff 503171482Sjeff/* 504171482Sjeff * Remove a thread from a run-queue. This typically happens when a thread 505171482Sjeff * is selected to run. Running threads are not on the queue and the 506171482Sjeff * transferable count does not reflect them. 507171482Sjeff */ 508122744Sjeffstatic __inline void 509177435Sjefftdq_runq_rem(struct tdq *tdq, struct thread *td) 510122744Sjeff{ 511177435Sjeff struct td_sched *ts; 512177435Sjeff 513177435Sjeff ts = td->td_sched; 514171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 515171482Sjeff KASSERT(ts->ts_runq != NULL, 516177435Sjeff ("tdq_runq_remove: thread %p null ts_runq", td)); 517164936Sjulian if (ts->ts_flags & TSF_XFERABLE) { 518165620Sjeff tdq->tdq_transferable--; 519164936Sjulian ts->ts_flags &= ~TSF_XFERABLE; 520123433Sjeff } 521165766Sjeff if (ts->ts_runq == &tdq->tdq_timeshare) { 522165766Sjeff if (tdq->tdq_idx != tdq->tdq_ridx) 523177435Sjeff runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx); 524165766Sjeff else 525177435Sjeff runq_remove_idx(ts->ts_runq, td, NULL); 526165766Sjeff } else 527177435Sjeff runq_remove(ts->ts_runq, td); 528122744Sjeff} 529122744Sjeff 530171482Sjeff/* 531171482Sjeff * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 532171482Sjeff * for this thread to the referenced thread queue. 533171482Sjeff */ 534113357Sjeffstatic void 535177435Sjefftdq_load_add(struct tdq *tdq, struct thread *td) 536113357Sjeff{ 537171482Sjeff 538171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 539177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 540177902Sjeff 541165620Sjeff tdq->tdq_load++; 542198854Sattilio if ((td->td_flags & TDF_NOLOAD) == 0) 543177902Sjeff tdq->tdq_sysload++; 544187357Sjeff KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 545235459Srstone SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load); 546110267Sjeff} 547113357Sjeff 548171482Sjeff/* 549171482Sjeff * Remove the load from a thread that is transitioning to a sleep state or 550171482Sjeff * exiting. 551171482Sjeff */ 552112994Sjeffstatic void 553177435Sjefftdq_load_rem(struct tdq *tdq, struct thread *td) 554110267Sjeff{ 555171482Sjeff 556177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 557171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 558171482Sjeff KASSERT(tdq->tdq_load != 0, 559171713Sjeff ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 560177902Sjeff 561165620Sjeff tdq->tdq_load--; 562198854Sattilio if ((td->td_flags & TDF_NOLOAD) == 0) 563177902Sjeff tdq->tdq_sysload--; 564187357Sjeff KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 565235459Srstone SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load); 566110267Sjeff} 567110267Sjeff 568176735Sjeff/* 569242736Sjeff * Bound timeshare latency by decreasing slice size as load increases. We 570242736Sjeff * consider the maximum latency as the sum of the threads waiting to run 571242736Sjeff * aside from curthread and target no more than sched_slice latency but 572242736Sjeff * no less than sched_slice_min runtime. 573242736Sjeff */ 574242736Sjeffstatic inline int 575242736Sjefftdq_slice(struct tdq *tdq) 576242736Sjeff{ 577242736Sjeff int load; 578242736Sjeff 579242736Sjeff /* 580242736Sjeff * It is safe to use sys_load here because this is called from 581242736Sjeff * contexts where timeshare threads are running and so there 582242736Sjeff * cannot be higher priority load in the system. 583242736Sjeff */ 584242736Sjeff load = tdq->tdq_sysload - 1; 585242736Sjeff if (load >= SCHED_SLICE_MIN_DIVISOR) 586242736Sjeff return (sched_slice_min); 587242736Sjeff if (load <= 1) 588242736Sjeff return (sched_slice); 589242736Sjeff return (sched_slice / load); 590242736Sjeff} 591242736Sjeff 592242736Sjeff/* 593176735Sjeff * Set lowpri to its exact value by searching the run-queue and 594176735Sjeff * evaluating curthread. curthread may be passed as an optimization. 595176735Sjeff */ 596176735Sjeffstatic void 597176735Sjefftdq_setlowpri(struct tdq *tdq, struct thread *ctd) 598176735Sjeff{ 599176735Sjeff struct thread *td; 600176735Sjeff 601176735Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 602176735Sjeff if (ctd == NULL) 603176735Sjeff ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 604177435Sjeff td = tdq_choose(tdq); 605177435Sjeff if (td == NULL || td->td_priority > ctd->td_priority) 606176735Sjeff tdq->tdq_lowpri = ctd->td_priority; 607176735Sjeff else 608176735Sjeff tdq->tdq_lowpri = td->td_priority; 609176735Sjeff} 610176735Sjeff 611113357Sjeff#ifdef SMP 612176735Sjeffstruct cpu_search { 613194779Sjeff cpuset_t cs_mask; 614232207Smav u_int cs_prefer; 615232207Smav int cs_pri; /* Min priority for low. */ 616232207Smav int cs_limit; /* Max load for low, min load for high. */ 617232207Smav int cs_cpu; 618232207Smav int cs_load; 619176735Sjeff}; 620176735Sjeff 621176735Sjeff#define CPU_SEARCH_LOWEST 0x1 622176735Sjeff#define CPU_SEARCH_HIGHEST 0x2 623176735Sjeff#define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 624176735Sjeff 625194779Sjeff#define CPUSET_FOREACH(cpu, mask) \ 626194779Sjeff for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \ 627222813Sattilio if (CPU_ISSET(cpu, &mask)) 628176735Sjeff 629232207Smavstatic __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low, 630176735Sjeff struct cpu_search *high, const int match); 631232207Smavint cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low); 632232207Smavint cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high); 633232207Smavint cpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 634176735Sjeff struct cpu_search *high); 635176735Sjeff 636116069Sjeff/* 637176735Sjeff * Search the tree of cpu_groups for the lowest or highest loaded cpu 638176735Sjeff * according to the match argument. This routine actually compares the 639176735Sjeff * load on all paths through the tree and finds the least loaded cpu on 640176735Sjeff * the least loaded path, which may differ from the least loaded cpu in 641176735Sjeff * the system. This balances work among caches and busses. 642116069Sjeff * 643176735Sjeff * This inline is instantiated in three forms below using constants for the 644176735Sjeff * match argument. It is reduced to the minimum set for each case. It is 645176735Sjeff * also recursive to the depth of the tree. 646116069Sjeff */ 647177169Sjhbstatic __inline int 648232207Smavcpu_search(const struct cpu_group *cg, struct cpu_search *low, 649176735Sjeff struct cpu_search *high, const int match) 650176735Sjeff{ 651232207Smav struct cpu_search lgroup; 652232207Smav struct cpu_search hgroup; 653232207Smav cpuset_t cpumask; 654232207Smav struct cpu_group *child; 655232207Smav struct tdq *tdq; 656234066Smav int cpu, i, hload, lload, load, total, rnd, *rndptr; 657176735Sjeff 658176735Sjeff total = 0; 659232207Smav cpumask = cg->cg_mask; 660232207Smav if (match & CPU_SEARCH_LOWEST) { 661232207Smav lload = INT_MAX; 662232207Smav lgroup = *low; 663232207Smav } 664232207Smav if (match & CPU_SEARCH_HIGHEST) { 665234066Smav hload = INT_MIN; 666232207Smav hgroup = *high; 667232207Smav } 668176735Sjeff 669232207Smav /* Iterate through the child CPU groups and then remaining CPUs. */ 670234066Smav for (i = cg->cg_children, cpu = mp_maxid; i >= 0; ) { 671234066Smav if (i == 0) { 672234066Smav while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask)) 673234066Smav cpu--; 674234066Smav if (cpu < 0) 675232207Smav break; 676232207Smav child = NULL; 677232207Smav } else 678234066Smav child = &cg->cg_child[i - 1]; 679232207Smav 680234066Smav if (match & CPU_SEARCH_LOWEST) 681234066Smav lgroup.cs_cpu = -1; 682234066Smav if (match & CPU_SEARCH_HIGHEST) 683234066Smav hgroup.cs_cpu = -1; 684232207Smav if (child) { /* Handle child CPU group. */ 685232207Smav CPU_NAND(&cpumask, &child->cg_mask); 686176735Sjeff switch (match) { 687176735Sjeff case CPU_SEARCH_LOWEST: 688176735Sjeff load = cpu_search_lowest(child, &lgroup); 689176735Sjeff break; 690176735Sjeff case CPU_SEARCH_HIGHEST: 691176735Sjeff load = cpu_search_highest(child, &hgroup); 692176735Sjeff break; 693176735Sjeff case CPU_SEARCH_BOTH: 694176735Sjeff load = cpu_search_both(child, &lgroup, &hgroup); 695176735Sjeff break; 696176735Sjeff } 697232207Smav } else { /* Handle child CPU. */ 698232207Smav tdq = TDQ_CPU(cpu); 699232207Smav load = tdq->tdq_load * 256; 700234066Smav rndptr = DPCPU_PTR(randomval); 701234066Smav rnd = (*rndptr = *rndptr * 69069 + 5) >> 26; 702232207Smav if (match & CPU_SEARCH_LOWEST) { 703232207Smav if (cpu == low->cs_prefer) 704232207Smav load -= 64; 705232207Smav /* If that CPU is allowed and get data. */ 706234066Smav if (tdq->tdq_lowpri > lgroup.cs_pri && 707234066Smav tdq->tdq_load <= lgroup.cs_limit && 708234066Smav CPU_ISSET(cpu, &lgroup.cs_mask)) { 709232207Smav lgroup.cs_cpu = cpu; 710232207Smav lgroup.cs_load = load - rnd; 711176735Sjeff } 712232207Smav } 713232207Smav if (match & CPU_SEARCH_HIGHEST) 714234066Smav if (tdq->tdq_load >= hgroup.cs_limit && 715234066Smav tdq->tdq_transferable && 716234066Smav CPU_ISSET(cpu, &hgroup.cs_mask)) { 717232207Smav hgroup.cs_cpu = cpu; 718232207Smav hgroup.cs_load = load - rnd; 719176735Sjeff } 720176735Sjeff } 721232207Smav total += load; 722176735Sjeff 723232207Smav /* We have info about child item. Compare it. */ 724232207Smav if (match & CPU_SEARCH_LOWEST) { 725234066Smav if (lgroup.cs_cpu >= 0 && 726232454Smav (load < lload || 727232454Smav (load == lload && lgroup.cs_load < low->cs_load))) { 728232207Smav lload = load; 729232207Smav low->cs_cpu = lgroup.cs_cpu; 730232207Smav low->cs_load = lgroup.cs_load; 731232207Smav } 732232207Smav } 733232207Smav if (match & CPU_SEARCH_HIGHEST) 734234066Smav if (hgroup.cs_cpu >= 0 && 735232454Smav (load > hload || 736232454Smav (load == hload && hgroup.cs_load > high->cs_load))) { 737232207Smav hload = load; 738232207Smav high->cs_cpu = hgroup.cs_cpu; 739232207Smav high->cs_load = hgroup.cs_load; 740232207Smav } 741234066Smav if (child) { 742234066Smav i--; 743234066Smav if (i == 0 && CPU_EMPTY(&cpumask)) 744234066Smav break; 745234066Smav } else 746234066Smav cpu--; 747176735Sjeff } 748176735Sjeff return (total); 749176735Sjeff} 750176735Sjeff 751176735Sjeff/* 752176735Sjeff * cpu_search instantiations must pass constants to maintain the inline 753176735Sjeff * optimization. 754176735Sjeff */ 755176735Sjeffint 756232207Smavcpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low) 757176735Sjeff{ 758176735Sjeff return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 759176735Sjeff} 760176735Sjeff 761176735Sjeffint 762232207Smavcpu_search_highest(const struct cpu_group *cg, struct cpu_search *high) 763176735Sjeff{ 764176735Sjeff return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 765176735Sjeff} 766176735Sjeff 767176735Sjeffint 768232207Smavcpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 769176735Sjeff struct cpu_search *high) 770176735Sjeff{ 771176735Sjeff return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 772176735Sjeff} 773176735Sjeff 774176735Sjeff/* 775176735Sjeff * Find the cpu with the least load via the least loaded path that has a 776176735Sjeff * lowpri greater than pri pri. A pri of -1 indicates any priority is 777176735Sjeff * acceptable. 778176735Sjeff */ 779176735Sjeffstatic inline int 780232207Smavsched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload, 781232207Smav int prefer) 782176735Sjeff{ 783176735Sjeff struct cpu_search low; 784176735Sjeff 785176735Sjeff low.cs_cpu = -1; 786232207Smav low.cs_prefer = prefer; 787176735Sjeff low.cs_mask = mask; 788232207Smav low.cs_pri = pri; 789232207Smav low.cs_limit = maxload; 790176735Sjeff cpu_search_lowest(cg, &low); 791176735Sjeff return low.cs_cpu; 792176735Sjeff} 793176735Sjeff 794176735Sjeff/* 795176735Sjeff * Find the cpu with the highest load via the highest loaded path. 796176735Sjeff */ 797176735Sjeffstatic inline int 798232207Smavsched_highest(const struct cpu_group *cg, cpuset_t mask, int minload) 799176735Sjeff{ 800176735Sjeff struct cpu_search high; 801176735Sjeff 802176735Sjeff high.cs_cpu = -1; 803176735Sjeff high.cs_mask = mask; 804176735Sjeff high.cs_limit = minload; 805176735Sjeff cpu_search_highest(cg, &high); 806176735Sjeff return high.cs_cpu; 807176735Sjeff} 808176735Sjeff 809176735Sjeff/* 810176735Sjeff * Simultaneously find the highest and lowest loaded cpu reachable via 811176735Sjeff * cg. 812176735Sjeff */ 813232207Smavstatic inline void 814232207Smavsched_both(const struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu) 815176735Sjeff{ 816176735Sjeff struct cpu_search high; 817176735Sjeff struct cpu_search low; 818176735Sjeff 819176735Sjeff low.cs_cpu = -1; 820232207Smav low.cs_prefer = -1; 821232207Smav low.cs_pri = -1; 822232207Smav low.cs_limit = INT_MAX; 823176735Sjeff low.cs_mask = mask; 824176735Sjeff high.cs_cpu = -1; 825176735Sjeff high.cs_limit = -1; 826176735Sjeff high.cs_mask = mask; 827176735Sjeff cpu_search_both(cg, &low, &high); 828176735Sjeff *lowcpu = low.cs_cpu; 829176735Sjeff *highcpu = high.cs_cpu; 830176735Sjeff return; 831176735Sjeff} 832176735Sjeff 833121790Sjeffstatic void 834176735Sjeffsched_balance_group(struct cpu_group *cg) 835116069Sjeff{ 836232207Smav cpuset_t hmask, lmask; 837232207Smav int high, low, anylow; 838123487Sjeff 839232207Smav CPU_FILL(&hmask); 840176735Sjeff for (;;) { 841232207Smav high = sched_highest(cg, hmask, 1); 842232207Smav /* Stop if there is no more CPU with transferrable threads. */ 843232207Smav if (high == -1) 844176735Sjeff break; 845232207Smav CPU_CLR(high, &hmask); 846232207Smav CPU_COPY(&hmask, &lmask); 847232207Smav /* Stop if there is no more CPU left for low. */ 848232207Smav if (CPU_EMPTY(&lmask)) 849176735Sjeff break; 850232207Smav anylow = 1; 851232207Smavnextlow: 852232207Smav low = sched_lowest(cg, lmask, -1, 853232207Smav TDQ_CPU(high)->tdq_load - 1, high); 854232207Smav /* Stop if we looked well and found no less loaded CPU. */ 855232207Smav if (anylow && low == -1) 856232207Smav break; 857232207Smav /* Go to next high if we found no less loaded CPU. */ 858232207Smav if (low == -1) 859232207Smav continue; 860232207Smav /* Transfer thread from high to low. */ 861232207Smav if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) { 862232207Smav /* CPU that got thread can no longer be a donor. */ 863232207Smav CPU_CLR(low, &hmask); 864232207Smav } else { 865232207Smav /* 866232207Smav * If failed, then there is no threads on high 867232207Smav * that can run on this low. Drop low from low 868232207Smav * mask and look for different one. 869232207Smav */ 870232207Smav CPU_CLR(low, &lmask); 871232207Smav anylow = 0; 872232207Smav goto nextlow; 873232207Smav } 874123487Sjeff } 875123487Sjeff} 876123487Sjeff 877123487Sjeffstatic void 878201148Sedsched_balance(void) 879123487Sjeff{ 880172409Sjeff struct tdq *tdq; 881123487Sjeff 882172409Sjeff /* 883172409Sjeff * Select a random time between .5 * balance_interval and 884172409Sjeff * 1.5 * balance_interval. 885172409Sjeff */ 886176735Sjeff balance_ticks = max(balance_interval / 2, 1); 887176735Sjeff balance_ticks += random() % balance_interval; 888171482Sjeff if (smp_started == 0 || rebalance == 0) 889171482Sjeff return; 890172409Sjeff tdq = TDQ_SELF(); 891172409Sjeff TDQ_UNLOCK(tdq); 892176735Sjeff sched_balance_group(cpu_top); 893172409Sjeff TDQ_LOCK(tdq); 894123487Sjeff} 895123487Sjeff 896171482Sjeff/* 897171482Sjeff * Lock two thread queues using their address to maintain lock order. 898171482Sjeff */ 899123487Sjeffstatic void 900171482Sjefftdq_lock_pair(struct tdq *one, struct tdq *two) 901171482Sjeff{ 902171482Sjeff if (one < two) { 903171482Sjeff TDQ_LOCK(one); 904171482Sjeff TDQ_LOCK_FLAGS(two, MTX_DUPOK); 905171482Sjeff } else { 906171482Sjeff TDQ_LOCK(two); 907171482Sjeff TDQ_LOCK_FLAGS(one, MTX_DUPOK); 908171482Sjeff } 909171482Sjeff} 910171482Sjeff 911171482Sjeff/* 912172409Sjeff * Unlock two thread queues. Order is not important here. 913172409Sjeff */ 914172409Sjeffstatic void 915172409Sjefftdq_unlock_pair(struct tdq *one, struct tdq *two) 916172409Sjeff{ 917172409Sjeff TDQ_UNLOCK(one); 918172409Sjeff TDQ_UNLOCK(two); 919172409Sjeff} 920172409Sjeff 921172409Sjeff/* 922171482Sjeff * Transfer load between two imbalanced thread queues. 923171482Sjeff */ 924176735Sjeffstatic int 925164936Sjuliansched_balance_pair(struct tdq *high, struct tdq *low) 926123487Sjeff{ 927176735Sjeff int moved; 928226057Smarius int cpu; 929116069Sjeff 930171482Sjeff tdq_lock_pair(high, low); 931176735Sjeff moved = 0; 932116069Sjeff /* 933122744Sjeff * Determine what the imbalance is and then adjust that to how many 934165620Sjeff * threads we actually have to give up (transferable). 935122744Sjeff */ 936232207Smav if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load && 937232207Smav (moved = tdq_move(high, low)) > 0) { 938172293Sjeff /* 939226057Smarius * In case the target isn't the current cpu IPI it to force a 940226057Smarius * reschedule with the new workload. 941172293Sjeff */ 942226057Smarius cpu = TDQ_ID(low); 943226057Smarius if (cpu != PCPU_GET(cpuid)) 944226057Smarius ipi_cpu(cpu, IPI_PREEMPT); 945171482Sjeff } 946172409Sjeff tdq_unlock_pair(high, low); 947176735Sjeff return (moved); 948116069Sjeff} 949116069Sjeff 950171482Sjeff/* 951171482Sjeff * Move a thread from one thread queue to another. 952171482Sjeff */ 953176735Sjeffstatic int 954171482Sjefftdq_move(struct tdq *from, struct tdq *to) 955116069Sjeff{ 956171482Sjeff struct td_sched *ts; 957171482Sjeff struct thread *td; 958164936Sjulian struct tdq *tdq; 959171482Sjeff int cpu; 960116069Sjeff 961172409Sjeff TDQ_LOCK_ASSERT(from, MA_OWNED); 962172409Sjeff TDQ_LOCK_ASSERT(to, MA_OWNED); 963172409Sjeff 964164936Sjulian tdq = from; 965171482Sjeff cpu = TDQ_ID(to); 966177435Sjeff td = tdq_steal(tdq, cpu); 967177435Sjeff if (td == NULL) 968176735Sjeff return (0); 969177435Sjeff ts = td->td_sched; 970171482Sjeff /* 971171482Sjeff * Although the run queue is locked the thread may be blocked. Lock 972172409Sjeff * it to clear this and acquire the run-queue lock. 973171482Sjeff */ 974171482Sjeff thread_lock(td); 975172409Sjeff /* Drop recursive lock on from acquired via thread_lock(). */ 976171482Sjeff TDQ_UNLOCK(from); 977171482Sjeff sched_rem(td); 978166108Sjeff ts->ts_cpu = cpu; 979171482Sjeff td->td_lock = TDQ_LOCKPTR(to); 980171482Sjeff tdq_add(to, td, SRQ_YIELDING); 981176735Sjeff return (1); 982116069Sjeff} 983110267Sjeff 984171482Sjeff/* 985171482Sjeff * This tdq has idled. Try to steal a thread from another cpu and switch 986171482Sjeff * to it. 987171482Sjeff */ 988123433Sjeffstatic int 989164936Sjuliantdq_idled(struct tdq *tdq) 990121790Sjeff{ 991176735Sjeff struct cpu_group *cg; 992164936Sjulian struct tdq *steal; 993194779Sjeff cpuset_t mask; 994176735Sjeff int thresh; 995171482Sjeff int cpu; 996123433Sjeff 997172484Sjeff if (smp_started == 0 || steal_idle == 0) 998172484Sjeff return (1); 999194779Sjeff CPU_FILL(&mask); 1000194779Sjeff CPU_CLR(PCPU_GET(cpuid), &mask); 1001176735Sjeff /* We don't want to be preempted while we're iterating. */ 1002171482Sjeff spinlock_enter(); 1003176735Sjeff for (cg = tdq->tdq_cg; cg != NULL; ) { 1004191643Sjeff if ((cg->cg_flags & CG_FLAG_THREAD) == 0) 1005176735Sjeff thresh = steal_thresh; 1006176735Sjeff else 1007176735Sjeff thresh = 1; 1008176735Sjeff cpu = sched_highest(cg, mask, thresh); 1009176735Sjeff if (cpu == -1) { 1010176735Sjeff cg = cg->cg_parent; 1011176735Sjeff continue; 1012166108Sjeff } 1013176735Sjeff steal = TDQ_CPU(cpu); 1014194779Sjeff CPU_CLR(cpu, &mask); 1015176735Sjeff tdq_lock_pair(tdq, steal); 1016176735Sjeff if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 1017176735Sjeff tdq_unlock_pair(tdq, steal); 1018176735Sjeff continue; 1019171482Sjeff } 1020176735Sjeff /* 1021176735Sjeff * If a thread was added while interrupts were disabled don't 1022176735Sjeff * steal one here. If we fail to acquire one due to affinity 1023176735Sjeff * restrictions loop again with this cpu removed from the 1024176735Sjeff * set. 1025176735Sjeff */ 1026176735Sjeff if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 1027176735Sjeff tdq_unlock_pair(tdq, steal); 1028176735Sjeff continue; 1029176735Sjeff } 1030176735Sjeff spinlock_exit(); 1031176735Sjeff TDQ_UNLOCK(steal); 1032178272Sjeff mi_switch(SW_VOL | SWT_IDLE, NULL); 1033176735Sjeff thread_unlock(curthread); 1034176735Sjeff 1035176735Sjeff return (0); 1036123433Sjeff } 1037171482Sjeff spinlock_exit(); 1038123433Sjeff return (1); 1039121790Sjeff} 1040121790Sjeff 1041171482Sjeff/* 1042171482Sjeff * Notify a remote cpu of new work. Sends an IPI if criteria are met. 1043171482Sjeff */ 1044121790Sjeffstatic void 1045177435Sjefftdq_notify(struct tdq *tdq, struct thread *td) 1046121790Sjeff{ 1047185047Sjhb struct thread *ctd; 1048166247Sjeff int pri; 1049166108Sjeff int cpu; 1050121790Sjeff 1051177005Sjeff if (tdq->tdq_ipipending) 1052177005Sjeff return; 1053177435Sjeff cpu = td->td_sched->ts_cpu; 1054177435Sjeff pri = td->td_priority; 1055185047Sjhb ctd = pcpu_find(cpu)->pc_curthread; 1056185047Sjhb if (!sched_shouldpreempt(pri, ctd->td_priority, 1)) 1057166137Sjeff return; 1058185047Sjhb if (TD_IS_IDLETHREAD(ctd)) { 1059178277Sjeff /* 1060178471Sjeff * If the MD code has an idle wakeup routine try that before 1061178471Sjeff * falling back to IPI. 1062178471Sjeff */ 1063212416Smav if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu)) 1064178471Sjeff return; 1065178277Sjeff } 1066177005Sjeff tdq->tdq_ipipending = 1; 1067210939Sjhb ipi_cpu(cpu, IPI_PREEMPT); 1068121790Sjeff} 1069121790Sjeff 1070171482Sjeff/* 1071171482Sjeff * Steals load from a timeshare queue. Honors the rotating queue head 1072171482Sjeff * index. 1073171482Sjeff */ 1074177435Sjeffstatic struct thread * 1075176735Sjeffrunq_steal_from(struct runq *rq, int cpu, u_char start) 1076171482Sjeff{ 1077171482Sjeff struct rqbits *rqb; 1078171482Sjeff struct rqhead *rqh; 1079232207Smav struct thread *td, *first; 1080171482Sjeff int bit; 1081171482Sjeff int pri; 1082171482Sjeff int i; 1083171482Sjeff 1084171482Sjeff rqb = &rq->rq_status; 1085171482Sjeff bit = start & (RQB_BPW -1); 1086171482Sjeff pri = 0; 1087232207Smav first = NULL; 1088171482Sjeffagain: 1089171482Sjeff for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 1090171482Sjeff if (rqb->rqb_bits[i] == 0) 1091171482Sjeff continue; 1092171482Sjeff if (bit != 0) { 1093171482Sjeff for (pri = bit; pri < RQB_BPW; pri++) 1094171482Sjeff if (rqb->rqb_bits[i] & (1ul << pri)) 1095171482Sjeff break; 1096171482Sjeff if (pri >= RQB_BPW) 1097171482Sjeff continue; 1098171482Sjeff } else 1099171482Sjeff pri = RQB_FFS(rqb->rqb_bits[i]); 1100171482Sjeff pri += (i << RQB_L2BPW); 1101171482Sjeff rqh = &rq->rq_queues[pri]; 1102177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) { 1103177435Sjeff if (first && THREAD_CAN_MIGRATE(td) && 1104177435Sjeff THREAD_CAN_SCHED(td, cpu)) 1105177435Sjeff return (td); 1106232207Smav first = td; 1107171482Sjeff } 1108171482Sjeff } 1109171482Sjeff if (start != 0) { 1110171482Sjeff start = 0; 1111171482Sjeff goto again; 1112171482Sjeff } 1113171482Sjeff 1114232207Smav if (first && THREAD_CAN_MIGRATE(first) && 1115232207Smav THREAD_CAN_SCHED(first, cpu)) 1116232207Smav return (first); 1117171482Sjeff return (NULL); 1118171482Sjeff} 1119171482Sjeff 1120171482Sjeff/* 1121171482Sjeff * Steals load from a standard linear queue. 1122171482Sjeff */ 1123177435Sjeffstatic struct thread * 1124176735Sjeffrunq_steal(struct runq *rq, int cpu) 1125121790Sjeff{ 1126121790Sjeff struct rqhead *rqh; 1127121790Sjeff struct rqbits *rqb; 1128177435Sjeff struct thread *td; 1129121790Sjeff int word; 1130121790Sjeff int bit; 1131121790Sjeff 1132121790Sjeff rqb = &rq->rq_status; 1133121790Sjeff for (word = 0; word < RQB_LEN; word++) { 1134121790Sjeff if (rqb->rqb_bits[word] == 0) 1135121790Sjeff continue; 1136121790Sjeff for (bit = 0; bit < RQB_BPW; bit++) { 1137123231Speter if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1138121790Sjeff continue; 1139121790Sjeff rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1140177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) 1141177435Sjeff if (THREAD_CAN_MIGRATE(td) && 1142177435Sjeff THREAD_CAN_SCHED(td, cpu)) 1143177435Sjeff return (td); 1144121790Sjeff } 1145121790Sjeff } 1146121790Sjeff return (NULL); 1147121790Sjeff} 1148121790Sjeff 1149171482Sjeff/* 1150171482Sjeff * Attempt to steal a thread in priority order from a thread queue. 1151171482Sjeff */ 1152177435Sjeffstatic struct thread * 1153176735Sjefftdq_steal(struct tdq *tdq, int cpu) 1154121790Sjeff{ 1155177435Sjeff struct thread *td; 1156121790Sjeff 1157171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1158177435Sjeff if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1159177435Sjeff return (td); 1160177435Sjeff if ((td = runq_steal_from(&tdq->tdq_timeshare, 1161177435Sjeff cpu, tdq->tdq_ridx)) != NULL) 1162177435Sjeff return (td); 1163176735Sjeff return (runq_steal(&tdq->tdq_idle, cpu)); 1164121790Sjeff} 1165123433Sjeff 1166171482Sjeff/* 1167171482Sjeff * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1168172409Sjeff * current lock and returns with the assigned queue locked. 1169171482Sjeff */ 1170171482Sjeffstatic inline struct tdq * 1171177435Sjeffsched_setcpu(struct thread *td, int cpu, int flags) 1172123433Sjeff{ 1173177435Sjeff 1174171482Sjeff struct tdq *tdq; 1175123433Sjeff 1176177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1177171482Sjeff tdq = TDQ_CPU(cpu); 1178177435Sjeff td->td_sched->ts_cpu = cpu; 1179177435Sjeff /* 1180177435Sjeff * If the lock matches just return the queue. 1181177435Sjeff */ 1182171482Sjeff if (td->td_lock == TDQ_LOCKPTR(tdq)) 1183171482Sjeff return (tdq); 1184171482Sjeff#ifdef notyet 1185123433Sjeff /* 1186172293Sjeff * If the thread isn't running its lockptr is a 1187171482Sjeff * turnstile or a sleepqueue. We can just lock_set without 1188171482Sjeff * blocking. 1189123685Sjeff */ 1190171482Sjeff if (TD_CAN_RUN(td)) { 1191171482Sjeff TDQ_LOCK(tdq); 1192171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1193171482Sjeff return (tdq); 1194171482Sjeff } 1195171482Sjeff#endif 1196166108Sjeff /* 1197171482Sjeff * The hard case, migration, we need to block the thread first to 1198171482Sjeff * prevent order reversals with other cpus locks. 1199166108Sjeff */ 1200202889Sattilio spinlock_enter(); 1201171482Sjeff thread_lock_block(td); 1202171482Sjeff TDQ_LOCK(tdq); 1203171713Sjeff thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1204202889Sattilio spinlock_exit(); 1205171482Sjeff return (tdq); 1206166108Sjeff} 1207166108Sjeff 1208178272SjeffSCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1209178272SjeffSCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1210178272SjeffSCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1211178272SjeffSCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1212178272SjeffSCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1213178272SjeffSCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1214178272Sjeff 1215166108Sjeffstatic int 1216177435Sjeffsched_pickcpu(struct thread *td, int flags) 1217171482Sjeff{ 1218232207Smav struct cpu_group *cg, *ccg; 1219177435Sjeff struct td_sched *ts; 1220171482Sjeff struct tdq *tdq; 1221194779Sjeff cpuset_t mask; 1222232207Smav int cpu, pri, self; 1223166108Sjeff 1224176735Sjeff self = PCPU_GET(cpuid); 1225177435Sjeff ts = td->td_sched; 1226166108Sjeff if (smp_started == 0) 1227166108Sjeff return (self); 1228171506Sjeff /* 1229171506Sjeff * Don't migrate a running thread from sched_switch(). 1230171506Sjeff */ 1231176735Sjeff if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1232176735Sjeff return (ts->ts_cpu); 1233166108Sjeff /* 1234176735Sjeff * Prefer to run interrupt threads on the processors that generate 1235176735Sjeff * the interrupt. 1236166108Sjeff */ 1237232207Smav pri = td->td_priority; 1238176735Sjeff if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1239178272Sjeff curthread->td_intr_nesting_level && ts->ts_cpu != self) { 1240178272Sjeff SCHED_STAT_INC(pickcpu_intrbind); 1241176735Sjeff ts->ts_cpu = self; 1242232207Smav if (TDQ_CPU(self)->tdq_lowpri > pri) { 1243232207Smav SCHED_STAT_INC(pickcpu_affinity); 1244232207Smav return (ts->ts_cpu); 1245232207Smav } 1246178272Sjeff } 1247166108Sjeff /* 1248176735Sjeff * If the thread can run on the last cpu and the affinity has not 1249176735Sjeff * expired or it is idle run it there. 1250166108Sjeff */ 1251176735Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1252232207Smav cg = tdq->tdq_cg; 1253232207Smav if (THREAD_CAN_SCHED(td, ts->ts_cpu) && 1254232207Smav tdq->tdq_lowpri >= PRI_MIN_IDLE && 1255232207Smav SCHED_AFFINITY(ts, CG_SHARE_L2)) { 1256232207Smav if (cg->cg_flags & CG_FLAG_THREAD) { 1257232207Smav CPUSET_FOREACH(cpu, cg->cg_mask) { 1258232207Smav if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) 1259232207Smav break; 1260232207Smav } 1261232207Smav } else 1262232207Smav cpu = INT_MAX; 1263232207Smav if (cpu > mp_maxid) { 1264178272Sjeff SCHED_STAT_INC(pickcpu_idle_affinity); 1265176735Sjeff return (ts->ts_cpu); 1266178272Sjeff } 1267139334Sjeff } 1268123433Sjeff /* 1269232207Smav * Search for the last level cache CPU group in the tree. 1270232207Smav * Skip caches with expired affinity time and SMT groups. 1271232207Smav * Affinity to higher level caches will be handled less aggressively. 1272123433Sjeff */ 1273232207Smav for (ccg = NULL; cg != NULL; cg = cg->cg_parent) { 1274232207Smav if (cg->cg_flags & CG_FLAG_THREAD) 1275232207Smav continue; 1276232207Smav if (!SCHED_AFFINITY(ts, cg->cg_level)) 1277232207Smav continue; 1278232207Smav ccg = cg; 1279232207Smav } 1280232207Smav if (ccg != NULL) 1281232207Smav cg = ccg; 1282176735Sjeff cpu = -1; 1283232207Smav /* Search the group for the less loaded idle CPU we can run now. */ 1284194779Sjeff mask = td->td_cpuset->cs_mask; 1285232207Smav if (cg != NULL && cg != cpu_top && 1286232207Smav CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0) 1287232207Smav cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE), 1288232207Smav INT_MAX, ts->ts_cpu); 1289232207Smav /* Search globally for the less loaded CPU we can run now. */ 1290176735Sjeff if (cpu == -1) 1291232207Smav cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu); 1292232207Smav /* Search globally for the less loaded CPU. */ 1293232207Smav if (cpu == -1) 1294232207Smav cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu); 1295232454Smav KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1296171506Sjeff /* 1297176735Sjeff * Compare the lowest loaded cpu to current cpu. 1298171506Sjeff */ 1299177005Sjeff if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1300232207Smav TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE && 1301232207Smav TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) { 1302178272Sjeff SCHED_STAT_INC(pickcpu_local); 1303177005Sjeff cpu = self; 1304178272Sjeff } else 1305178272Sjeff SCHED_STAT_INC(pickcpu_lowest); 1306178272Sjeff if (cpu != ts->ts_cpu) 1307178272Sjeff SCHED_STAT_INC(pickcpu_migration); 1308171482Sjeff return (cpu); 1309123433Sjeff} 1310176735Sjeff#endif 1311123433Sjeff 1312117326Sjeff/* 1313121790Sjeff * Pick the highest priority task we have and return it. 1314117326Sjeff */ 1315177435Sjeffstatic struct thread * 1316164936Sjuliantdq_choose(struct tdq *tdq) 1317110267Sjeff{ 1318177435Sjeff struct thread *td; 1319110267Sjeff 1320171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1321177435Sjeff td = runq_choose(&tdq->tdq_realtime); 1322177435Sjeff if (td != NULL) 1323177435Sjeff return (td); 1324177435Sjeff td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1325177435Sjeff if (td != NULL) { 1326217351Sjhb KASSERT(td->td_priority >= PRI_MIN_BATCH, 1327165762Sjeff ("tdq_choose: Invalid priority on timeshare queue %d", 1328177435Sjeff td->td_priority)); 1329177435Sjeff return (td); 1330165762Sjeff } 1331177435Sjeff td = runq_choose(&tdq->tdq_idle); 1332177435Sjeff if (td != NULL) { 1333177435Sjeff KASSERT(td->td_priority >= PRI_MIN_IDLE, 1334165762Sjeff ("tdq_choose: Invalid priority on idle queue %d", 1335177435Sjeff td->td_priority)); 1336177435Sjeff return (td); 1337165762Sjeff } 1338165762Sjeff 1339165762Sjeff return (NULL); 1340110267Sjeff} 1341110267Sjeff 1342171482Sjeff/* 1343171482Sjeff * Initialize a thread queue. 1344171482Sjeff */ 1345109864Sjeffstatic void 1346164936Sjuliantdq_setup(struct tdq *tdq) 1347110028Sjeff{ 1348171482Sjeff 1349171713Sjeff if (bootverbose) 1350171713Sjeff printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1351165762Sjeff runq_init(&tdq->tdq_realtime); 1352165762Sjeff runq_init(&tdq->tdq_timeshare); 1353165620Sjeff runq_init(&tdq->tdq_idle); 1354176735Sjeff snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1355176735Sjeff "sched lock %d", (int)TDQ_ID(tdq)); 1356176735Sjeff mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1357176735Sjeff MTX_SPIN | MTX_RECURSE); 1358187357Sjeff#ifdef KTR 1359187357Sjeff snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname), 1360187357Sjeff "CPU %d load", (int)TDQ_ID(tdq)); 1361187357Sjeff#endif 1362110028Sjeff} 1363110028Sjeff 1364171713Sjeff#ifdef SMP 1365110028Sjeffstatic void 1366171713Sjeffsched_setup_smp(void) 1367171713Sjeff{ 1368171713Sjeff struct tdq *tdq; 1369171713Sjeff int i; 1370171713Sjeff 1371176735Sjeff cpu_top = smp_topo(); 1372209059Sjhb CPU_FOREACH(i) { 1373176735Sjeff tdq = TDQ_CPU(i); 1374171713Sjeff tdq_setup(tdq); 1375176735Sjeff tdq->tdq_cg = smp_topo_find(cpu_top, i); 1376176735Sjeff if (tdq->tdq_cg == NULL) 1377176735Sjeff panic("Can't find cpu group for %d\n", i); 1378123433Sjeff } 1379176735Sjeff balance_tdq = TDQ_SELF(); 1380176735Sjeff sched_balance(); 1381171713Sjeff} 1382171713Sjeff#endif 1383171713Sjeff 1384171713Sjeff/* 1385171713Sjeff * Setup the thread queues and initialize the topology based on MD 1386171713Sjeff * information. 1387171713Sjeff */ 1388171713Sjeffstatic void 1389171713Sjeffsched_setup(void *dummy) 1390171713Sjeff{ 1391171713Sjeff struct tdq *tdq; 1392171713Sjeff 1393171713Sjeff tdq = TDQ_SELF(); 1394171713Sjeff#ifdef SMP 1395176734Sjeff sched_setup_smp(); 1396117237Sjeff#else 1397171713Sjeff tdq_setup(tdq); 1398116069Sjeff#endif 1399171482Sjeff 1400171482Sjeff /* Add thread0's load since it's running. */ 1401171482Sjeff TDQ_LOCK(tdq); 1402171713Sjeff thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1403177435Sjeff tdq_load_add(tdq, &thread0); 1404176735Sjeff tdq->tdq_lowpri = thread0.td_priority; 1405171482Sjeff TDQ_UNLOCK(tdq); 1406109864Sjeff} 1407109864Sjeff 1408171482Sjeff/* 1409239185Smav * This routine determines time constants after stathz and hz are setup. 1410171482Sjeff */ 1411153533Sdavidxu/* ARGSUSED */ 1412153533Sdavidxustatic void 1413153533Sdavidxusched_initticks(void *dummy) 1414153533Sdavidxu{ 1415171482Sjeff int incr; 1416171482Sjeff 1417153533Sdavidxu realstathz = stathz ? stathz : hz; 1418242736Sjeff sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR; 1419242736Sjeff sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR; 1420239196Smav hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 1421239196Smav realstathz); 1422153533Sdavidxu 1423153533Sdavidxu /* 1424165762Sjeff * tickincr is shifted out by 10 to avoid rounding errors due to 1425165766Sjeff * hz not being evenly divisible by stathz on all platforms. 1426153533Sdavidxu */ 1427171482Sjeff incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1428165762Sjeff /* 1429165762Sjeff * This does not work for values of stathz that are more than 1430165762Sjeff * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1431165762Sjeff */ 1432171482Sjeff if (incr == 0) 1433171482Sjeff incr = 1; 1434171482Sjeff tickincr = incr; 1435166108Sjeff#ifdef SMP 1436171899Sjeff /* 1437172409Sjeff * Set the default balance interval now that we know 1438172409Sjeff * what realstathz is. 1439172409Sjeff */ 1440172409Sjeff balance_interval = realstathz; 1441166108Sjeff affinity = SCHED_AFFINITY_DEFAULT; 1442166108Sjeff#endif 1443232740Smav if (sched_idlespinthresh < 0) 1444242852Smav sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz; 1445153533Sdavidxu} 1446153533Sdavidxu 1447153533Sdavidxu 1448109864Sjeff/* 1449171482Sjeff * This is the core of the interactivity algorithm. Determines a score based 1450171482Sjeff * on past behavior. It is the ratio of sleep time to run time scaled to 1451171482Sjeff * a [0, 100] integer. This is the voluntary sleep time of a process, which 1452171482Sjeff * differs from the cpu usage because it does not account for time spent 1453171482Sjeff * waiting on a run-queue. Would be prettier if we had floating point. 1454171482Sjeff */ 1455171482Sjeffstatic int 1456171482Sjeffsched_interact_score(struct thread *td) 1457171482Sjeff{ 1458171482Sjeff struct td_sched *ts; 1459171482Sjeff int div; 1460171482Sjeff 1461171482Sjeff ts = td->td_sched; 1462171482Sjeff /* 1463171482Sjeff * The score is only needed if this is likely to be an interactive 1464171482Sjeff * task. Don't go through the expense of computing it if there's 1465171482Sjeff * no chance. 1466171482Sjeff */ 1467171482Sjeff if (sched_interact <= SCHED_INTERACT_HALF && 1468171482Sjeff ts->ts_runtime >= ts->ts_slptime) 1469171482Sjeff return (SCHED_INTERACT_HALF); 1470171482Sjeff 1471171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1472171482Sjeff div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1473171482Sjeff return (SCHED_INTERACT_HALF + 1474171482Sjeff (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1475171482Sjeff } 1476171482Sjeff if (ts->ts_slptime > ts->ts_runtime) { 1477171482Sjeff div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1478171482Sjeff return (ts->ts_runtime / div); 1479171482Sjeff } 1480171482Sjeff /* runtime == slptime */ 1481171482Sjeff if (ts->ts_runtime) 1482171482Sjeff return (SCHED_INTERACT_HALF); 1483171482Sjeff 1484171482Sjeff /* 1485171482Sjeff * This can happen if slptime and runtime are 0. 1486171482Sjeff */ 1487171482Sjeff return (0); 1488171482Sjeff 1489171482Sjeff} 1490171482Sjeff 1491171482Sjeff/* 1492109864Sjeff * Scale the scheduling priority according to the "interactivity" of this 1493109864Sjeff * process. 1494109864Sjeff */ 1495113357Sjeffstatic void 1496163709Sjbsched_priority(struct thread *td) 1497109864Sjeff{ 1498165762Sjeff int score; 1499109864Sjeff int pri; 1500109864Sjeff 1501217291Sjhb if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1502113357Sjeff return; 1503112966Sjeff /* 1504165762Sjeff * If the score is interactive we place the thread in the realtime 1505165762Sjeff * queue with a priority that is less than kernel and interrupt 1506165762Sjeff * priorities. These threads are not subject to nice restrictions. 1507112966Sjeff * 1508171482Sjeff * Scores greater than this are placed on the normal timeshare queue 1509165762Sjeff * where the priority is partially decided by the most recent cpu 1510165762Sjeff * utilization and the rest is decided by nice value. 1511172293Sjeff * 1512172293Sjeff * The nice value of the process has a linear effect on the calculated 1513172293Sjeff * score. Negative nice values make it easier for a thread to be 1514172293Sjeff * considered interactive. 1515112966Sjeff */ 1516198126Sjhb score = imax(0, sched_interact_score(td) + td->td_proc->p_nice); 1517165762Sjeff if (score < sched_interact) { 1518217351Sjhb pri = PRI_MIN_INTERACT; 1519217351Sjhb pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) / 1520217237Sjhb sched_interact) * score; 1521217351Sjhb KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT, 1522166208Sjeff ("sched_priority: invalid interactive priority %d score %d", 1523166208Sjeff pri, score)); 1524165762Sjeff } else { 1525165762Sjeff pri = SCHED_PRI_MIN; 1526165762Sjeff if (td->td_sched->ts_ticks) 1527228960Sjhb pri += min(SCHED_PRI_TICKS(td->td_sched), 1528228960Sjhb SCHED_PRI_RANGE); 1529165762Sjeff pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1530217351Sjhb KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH, 1531171482Sjeff ("sched_priority: invalid priority %d: nice %d, " 1532171482Sjeff "ticks %d ftick %d ltick %d tick pri %d", 1533171482Sjeff pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1534171482Sjeff td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1535171482Sjeff SCHED_PRI_TICKS(td->td_sched))); 1536165762Sjeff } 1537165762Sjeff sched_user_prio(td, pri); 1538112966Sjeff 1539112966Sjeff return; 1540109864Sjeff} 1541109864Sjeff 1542121868Sjeff/* 1543121868Sjeff * This routine enforces a maximum limit on the amount of scheduling history 1544171482Sjeff * kept. It is called after either the slptime or runtime is adjusted. This 1545171482Sjeff * function is ugly due to integer math. 1546121868Sjeff */ 1547116463Sjeffstatic void 1548163709Sjbsched_interact_update(struct thread *td) 1549116463Sjeff{ 1550165819Sjeff struct td_sched *ts; 1551166208Sjeff u_int sum; 1552121605Sjeff 1553165819Sjeff ts = td->td_sched; 1554171482Sjeff sum = ts->ts_runtime + ts->ts_slptime; 1555121868Sjeff if (sum < SCHED_SLP_RUN_MAX) 1556121868Sjeff return; 1557121868Sjeff /* 1558165819Sjeff * This only happens from two places: 1559165819Sjeff * 1) We have added an unusual amount of run time from fork_exit. 1560165819Sjeff * 2) We have added an unusual amount of sleep time from sched_sleep(). 1561165819Sjeff */ 1562165819Sjeff if (sum > SCHED_SLP_RUN_MAX * 2) { 1563171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1564171482Sjeff ts->ts_runtime = SCHED_SLP_RUN_MAX; 1565171482Sjeff ts->ts_slptime = 1; 1566165819Sjeff } else { 1567171482Sjeff ts->ts_slptime = SCHED_SLP_RUN_MAX; 1568171482Sjeff ts->ts_runtime = 1; 1569165819Sjeff } 1570165819Sjeff return; 1571165819Sjeff } 1572165819Sjeff /* 1573121868Sjeff * If we have exceeded by more than 1/5th then the algorithm below 1574121868Sjeff * will not bring us back into range. Dividing by two here forces 1575133427Sjeff * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1576121868Sjeff */ 1577127850Sjeff if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1578171482Sjeff ts->ts_runtime /= 2; 1579171482Sjeff ts->ts_slptime /= 2; 1580121868Sjeff return; 1581116463Sjeff } 1582171482Sjeff ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1583171482Sjeff ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1584116463Sjeff} 1585116463Sjeff 1586171482Sjeff/* 1587171482Sjeff * Scale back the interactivity history when a child thread is created. The 1588171482Sjeff * history is inherited from the parent but the thread may behave totally 1589171482Sjeff * differently. For example, a shell spawning a compiler process. We want 1590171482Sjeff * to learn that the compiler is behaving badly very quickly. 1591171482Sjeff */ 1592121868Sjeffstatic void 1593163709Sjbsched_interact_fork(struct thread *td) 1594121868Sjeff{ 1595121868Sjeff int ratio; 1596121868Sjeff int sum; 1597121868Sjeff 1598171482Sjeff sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1599121868Sjeff if (sum > SCHED_SLP_RUN_FORK) { 1600121868Sjeff ratio = sum / SCHED_SLP_RUN_FORK; 1601171482Sjeff td->td_sched->ts_runtime /= ratio; 1602171482Sjeff td->td_sched->ts_slptime /= ratio; 1603121868Sjeff } 1604121868Sjeff} 1605121868Sjeff 1606113357Sjeff/* 1607171482Sjeff * Called from proc0_init() to setup the scheduler fields. 1608134791Sjulian */ 1609134791Sjulianvoid 1610134791Sjulianschedinit(void) 1611134791Sjulian{ 1612165762Sjeff 1613134791Sjulian /* 1614134791Sjulian * Set up the scheduler specific parts of proc0. 1615134791Sjulian */ 1616136167Sjulian proc0.p_sched = NULL; /* XXX */ 1617164936Sjulian thread0.td_sched = &td_sched0; 1618165762Sjeff td_sched0.ts_ltick = ticks; 1619165796Sjeff td_sched0.ts_ftick = ticks; 1620242736Sjeff td_sched0.ts_slice = 0; 1621134791Sjulian} 1622134791Sjulian 1623134791Sjulian/* 1624113357Sjeff * This is only somewhat accurate since given many processes of the same 1625113357Sjeff * priority they will switch when their slices run out, which will be 1626165762Sjeff * at most sched_slice stathz ticks. 1627113357Sjeff */ 1628109864Sjeffint 1629109864Sjeffsched_rr_interval(void) 1630109864Sjeff{ 1631165762Sjeff 1632239185Smav /* Convert sched_slice from stathz to hz. */ 1633239196Smav return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz)); 1634109864Sjeff} 1635109864Sjeff 1636171482Sjeff/* 1637171482Sjeff * Update the percent cpu tracking information when it is requested or 1638171482Sjeff * the total history exceeds the maximum. We keep a sliding history of 1639171482Sjeff * tick counts that slowly decays. This is less precise than the 4BSD 1640171482Sjeff * mechanism since it happens with less regular and frequent events. 1641171482Sjeff */ 1642121790Sjeffstatic void 1643232917Smavsched_pctcpu_update(struct td_sched *ts, int run) 1644109864Sjeff{ 1645232917Smav int t = ticks; 1646165762Sjeff 1647232917Smav if (t - ts->ts_ltick >= SCHED_TICK_TARG) { 1648164936Sjulian ts->ts_ticks = 0; 1649232917Smav ts->ts_ftick = t - SCHED_TICK_TARG; 1650232917Smav } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) { 1651232917Smav ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) * 1652232917Smav (ts->ts_ltick - (t - SCHED_TICK_TARG)); 1653232917Smav ts->ts_ftick = t - SCHED_TICK_TARG; 1654232917Smav } 1655232917Smav if (run) 1656232917Smav ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT; 1657232917Smav ts->ts_ltick = t; 1658109864Sjeff} 1659109864Sjeff 1660171482Sjeff/* 1661171482Sjeff * Adjust the priority of a thread. Move it to the appropriate run-queue 1662171482Sjeff * if necessary. This is the back-end for several priority related 1663171482Sjeff * functions. 1664171482Sjeff */ 1665165762Sjeffstatic void 1666139453Sjhbsched_thread_priority(struct thread *td, u_char prio) 1667109864Sjeff{ 1668164936Sjulian struct td_sched *ts; 1669177009Sjeff struct tdq *tdq; 1670177009Sjeff int oldpri; 1671109864Sjeff 1672187357Sjeff KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio", 1673187357Sjeff "prio:%d", td->td_priority, "new prio:%d", prio, 1674187357Sjeff KTR_ATTR_LINKED, sched_tdname(curthread)); 1675235459Srstone SDT_PROBE3(sched, , , change_pri, td, td->td_proc, prio); 1676240513Savg if (td != curthread && prio < td->td_priority) { 1677187357Sjeff KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 1678187357Sjeff "lend prio", "prio:%d", td->td_priority, "new prio:%d", 1679187357Sjeff prio, KTR_ATTR_LINKED, sched_tdname(td)); 1680235459Srstone SDT_PROBE4(sched, , , lend_pri, td, td->td_proc, prio, 1681235459Srstone curthread); 1682187357Sjeff } 1683164936Sjulian ts = td->td_sched; 1684170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1685139453Sjhb if (td->td_priority == prio) 1686139453Sjhb return; 1687177376Sjeff /* 1688177376Sjeff * If the priority has been elevated due to priority 1689177376Sjeff * propagation, we may have to move ourselves to a new 1690177376Sjeff * queue. This could be optimized to not re-add in some 1691177376Sjeff * cases. 1692177376Sjeff */ 1693165766Sjeff if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1694165762Sjeff sched_rem(td); 1695165762Sjeff td->td_priority = prio; 1696171482Sjeff sched_add(td, SRQ_BORROWING); 1697177009Sjeff return; 1698177009Sjeff } 1699177376Sjeff /* 1700177376Sjeff * If the thread is currently running we may have to adjust the lowpri 1701177376Sjeff * information so other cpus are aware of our current priority. 1702177376Sjeff */ 1703177009Sjeff if (TD_IS_RUNNING(td)) { 1704177376Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1705177376Sjeff oldpri = td->td_priority; 1706177376Sjeff td->td_priority = prio; 1707176735Sjeff if (prio < tdq->tdq_lowpri) 1708171482Sjeff tdq->tdq_lowpri = prio; 1709176735Sjeff else if (tdq->tdq_lowpri == oldpri) 1710176735Sjeff tdq_setlowpri(tdq, td); 1711177376Sjeff return; 1712177009Sjeff } 1713177376Sjeff td->td_priority = prio; 1714109864Sjeff} 1715109864Sjeff 1716139453Sjhb/* 1717139453Sjhb * Update a thread's priority when it is lent another thread's 1718139453Sjhb * priority. 1719139453Sjhb */ 1720109864Sjeffvoid 1721139453Sjhbsched_lend_prio(struct thread *td, u_char prio) 1722139453Sjhb{ 1723139453Sjhb 1724139453Sjhb td->td_flags |= TDF_BORROWING; 1725139453Sjhb sched_thread_priority(td, prio); 1726139453Sjhb} 1727139453Sjhb 1728139453Sjhb/* 1729139453Sjhb * Restore a thread's priority when priority propagation is 1730139453Sjhb * over. The prio argument is the minimum priority the thread 1731139453Sjhb * needs to have to satisfy other possible priority lending 1732139453Sjhb * requests. If the thread's regular priority is less 1733139453Sjhb * important than prio, the thread will keep a priority boost 1734139453Sjhb * of prio. 1735139453Sjhb */ 1736139453Sjhbvoid 1737139453Sjhbsched_unlend_prio(struct thread *td, u_char prio) 1738139453Sjhb{ 1739139453Sjhb u_char base_pri; 1740139453Sjhb 1741139453Sjhb if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1742139453Sjhb td->td_base_pri <= PRI_MAX_TIMESHARE) 1743163709Sjb base_pri = td->td_user_pri; 1744139453Sjhb else 1745139453Sjhb base_pri = td->td_base_pri; 1746139453Sjhb if (prio >= base_pri) { 1747139455Sjhb td->td_flags &= ~TDF_BORROWING; 1748139453Sjhb sched_thread_priority(td, base_pri); 1749139453Sjhb } else 1750139453Sjhb sched_lend_prio(td, prio); 1751139453Sjhb} 1752139453Sjhb 1753171482Sjeff/* 1754171482Sjeff * Standard entry for setting the priority to an absolute value. 1755171482Sjeff */ 1756139453Sjhbvoid 1757139453Sjhbsched_prio(struct thread *td, u_char prio) 1758139453Sjhb{ 1759139453Sjhb u_char oldprio; 1760139453Sjhb 1761139453Sjhb /* First, update the base priority. */ 1762139453Sjhb td->td_base_pri = prio; 1763139453Sjhb 1764139453Sjhb /* 1765139455Sjhb * If the thread is borrowing another thread's priority, don't 1766139453Sjhb * ever lower the priority. 1767139453Sjhb */ 1768139453Sjhb if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1769139453Sjhb return; 1770139453Sjhb 1771139453Sjhb /* Change the real priority. */ 1772139453Sjhb oldprio = td->td_priority; 1773139453Sjhb sched_thread_priority(td, prio); 1774139453Sjhb 1775139453Sjhb /* 1776139453Sjhb * If the thread is on a turnstile, then let the turnstile update 1777139453Sjhb * its state. 1778139453Sjhb */ 1779139453Sjhb if (TD_ON_LOCK(td) && oldprio != prio) 1780139453Sjhb turnstile_adjust(td, oldprio); 1781139453Sjhb} 1782139455Sjhb 1783171482Sjeff/* 1784171482Sjeff * Set the base user priority, does not effect current running priority. 1785171482Sjeff */ 1786139453Sjhbvoid 1787163709Sjbsched_user_prio(struct thread *td, u_char prio) 1788161599Sdavidxu{ 1789161599Sdavidxu 1790163709Sjb td->td_base_user_pri = prio; 1791216313Sdavidxu if (td->td_lend_user_pri <= prio) 1792216313Sdavidxu return; 1793163709Sjb td->td_user_pri = prio; 1794161599Sdavidxu} 1795161599Sdavidxu 1796161599Sdavidxuvoid 1797161599Sdavidxusched_lend_user_prio(struct thread *td, u_char prio) 1798161599Sdavidxu{ 1799161599Sdavidxu 1800174536Sdavidxu THREAD_LOCK_ASSERT(td, MA_OWNED); 1801216313Sdavidxu td->td_lend_user_pri = prio; 1802216791Sdavidxu td->td_user_pri = min(prio, td->td_base_user_pri); 1803216791Sdavidxu if (td->td_priority > td->td_user_pri) 1804216791Sdavidxu sched_prio(td, td->td_user_pri); 1805216791Sdavidxu else if (td->td_priority != td->td_user_pri) 1806216791Sdavidxu td->td_flags |= TDF_NEEDRESCHED; 1807161599Sdavidxu} 1808161599Sdavidxu 1809171482Sjeff/* 1810171713Sjeff * Handle migration from sched_switch(). This happens only for 1811171713Sjeff * cpu binding. 1812171713Sjeff */ 1813171713Sjeffstatic struct mtx * 1814171713Sjeffsched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1815171713Sjeff{ 1816171713Sjeff struct tdq *tdn; 1817171713Sjeff 1818171713Sjeff tdn = TDQ_CPU(td->td_sched->ts_cpu); 1819171713Sjeff#ifdef SMP 1820177435Sjeff tdq_load_rem(tdq, td); 1821171713Sjeff /* 1822171713Sjeff * Do the lock dance required to avoid LOR. We grab an extra 1823171713Sjeff * spinlock nesting to prevent preemption while we're 1824171713Sjeff * not holding either run-queue lock. 1825171713Sjeff */ 1826171713Sjeff spinlock_enter(); 1827202889Sattilio thread_lock_block(td); /* This releases the lock on tdq. */ 1828197223Sattilio 1829197223Sattilio /* 1830197223Sattilio * Acquire both run-queue locks before placing the thread on the new 1831197223Sattilio * run-queue to avoid deadlocks created by placing a thread with a 1832197223Sattilio * blocked lock on the run-queue of a remote processor. The deadlock 1833197223Sattilio * occurs when a third processor attempts to lock the two queues in 1834197223Sattilio * question while the target processor is spinning with its own 1835197223Sattilio * run-queue lock held while waiting for the blocked lock to clear. 1836197223Sattilio */ 1837197223Sattilio tdq_lock_pair(tdn, tdq); 1838171713Sjeff tdq_add(tdn, td, flags); 1839177435Sjeff tdq_notify(tdn, td); 1840197223Sattilio TDQ_UNLOCK(tdn); 1841171713Sjeff spinlock_exit(); 1842171713Sjeff#endif 1843171713Sjeff return (TDQ_LOCKPTR(tdn)); 1844171713Sjeff} 1845171713Sjeff 1846171713Sjeff/* 1847202889Sattilio * Variadic version of thread_lock_unblock() that does not assume td_lock 1848202889Sattilio * is blocked. 1849171482Sjeff */ 1850171482Sjeffstatic inline void 1851171482Sjeffthread_unblock_switch(struct thread *td, struct mtx *mtx) 1852171482Sjeff{ 1853171482Sjeff atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1854171482Sjeff (uintptr_t)mtx); 1855171482Sjeff} 1856171482Sjeff 1857171482Sjeff/* 1858171482Sjeff * Switch threads. This function has to handle threads coming in while 1859171482Sjeff * blocked for some reason, running, or idle. It also must deal with 1860171482Sjeff * migrating a thread from one queue to another as running threads may 1861171482Sjeff * be assigned elsewhere via binding. 1862171482Sjeff */ 1863161599Sdavidxuvoid 1864135051Sjuliansched_switch(struct thread *td, struct thread *newtd, int flags) 1865109864Sjeff{ 1866165627Sjeff struct tdq *tdq; 1867164936Sjulian struct td_sched *ts; 1868171482Sjeff struct mtx *mtx; 1869171713Sjeff int srqflag; 1870239157Smav int cpuid, preempted; 1871109864Sjeff 1872170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1873177376Sjeff KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1874109864Sjeff 1875171482Sjeff cpuid = PCPU_GET(cpuid); 1876171482Sjeff tdq = TDQ_CPU(cpuid); 1877164936Sjulian ts = td->td_sched; 1878171713Sjeff mtx = td->td_lock; 1879232917Smav sched_pctcpu_update(ts, 1); 1880171482Sjeff ts->ts_rltick = ticks; 1881133555Sjeff td->td_lastcpu = td->td_oncpu; 1882113339Sjulian td->td_oncpu = NOCPU; 1883239157Smav preempted = !(td->td_flags & TDF_SLICEEND); 1884239157Smav td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND); 1885144777Sups td->td_owepreempt = 0; 1886242852Smav if (!TD_IS_IDLETHREAD(td)) 1887242852Smav tdq->tdq_switchcnt++; 1888123434Sjeff /* 1889171482Sjeff * The lock pointer in an idle thread should never change. Reset it 1890171482Sjeff * to CAN_RUN as well. 1891123434Sjeff */ 1892167327Sjulian if (TD_IS_IDLETHREAD(td)) { 1893171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1894139334Sjeff TD_SET_CAN_RUN(td); 1895170293Sjeff } else if (TD_IS_RUNNING(td)) { 1896171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1897239157Smav srqflag = preempted ? 1898170293Sjeff SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1899171713Sjeff SRQ_OURSELF|SRQ_YIELDING; 1900212153Smdf#ifdef SMP 1901212115Smdf if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu)) 1902212115Smdf ts->ts_cpu = sched_pickcpu(td, 0); 1903212153Smdf#endif 1904171713Sjeff if (ts->ts_cpu == cpuid) 1905177435Sjeff tdq_runq_add(tdq, td, srqflag); 1906212115Smdf else { 1907212115Smdf KASSERT(THREAD_CAN_MIGRATE(td) || 1908212115Smdf (ts->ts_flags & TSF_BOUND) != 0, 1909212115Smdf ("Thread %p shouldn't migrate", td)); 1910171713Sjeff mtx = sched_switch_migrate(tdq, td, srqflag); 1911212115Smdf } 1912171482Sjeff } else { 1913171482Sjeff /* This thread must be going to sleep. */ 1914171482Sjeff TDQ_LOCK(tdq); 1915202889Sattilio mtx = thread_lock_block(td); 1916177435Sjeff tdq_load_rem(tdq, td); 1917171482Sjeff } 1918171482Sjeff /* 1919171482Sjeff * We enter here with the thread blocked and assigned to the 1920171482Sjeff * appropriate cpu run-queue or sleep-queue and with the current 1921171482Sjeff * thread-queue locked. 1922171482Sjeff */ 1923171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1924171482Sjeff newtd = choosethread(); 1925171482Sjeff /* 1926171482Sjeff * Call the MD code to switch contexts if necessary. 1927171482Sjeff */ 1928145256Sjkoshy if (td != newtd) { 1929145256Sjkoshy#ifdef HWPMC_HOOKS 1930145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1931145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1932145256Sjkoshy#endif 1933235459Srstone SDT_PROBE2(sched, , , off_cpu, td, td->td_proc); 1934174629Sjeff lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1935172411Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1936232917Smav sched_pctcpu_update(newtd->td_sched, 0); 1937179297Sjb 1938179297Sjb#ifdef KDTRACE_HOOKS 1939179297Sjb /* 1940179297Sjb * If DTrace has set the active vtime enum to anything 1941179297Sjb * other than INACTIVE (0), then it should have set the 1942179297Sjb * function to call. 1943179297Sjb */ 1944179297Sjb if (dtrace_vtime_active) 1945179297Sjb (*dtrace_vtime_switch_func)(newtd); 1946179297Sjb#endif 1947179297Sjb 1948171482Sjeff cpu_switch(td, newtd, mtx); 1949171482Sjeff /* 1950171482Sjeff * We may return from cpu_switch on a different cpu. However, 1951171482Sjeff * we always return with td_lock pointing to the current cpu's 1952171482Sjeff * run queue lock. 1953171482Sjeff */ 1954171482Sjeff cpuid = PCPU_GET(cpuid); 1955171482Sjeff tdq = TDQ_CPU(cpuid); 1956174629Sjeff lock_profile_obtain_lock_success( 1957174629Sjeff &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1958235459Srstone 1959235459Srstone SDT_PROBE0(sched, , , on_cpu); 1960145256Sjkoshy#ifdef HWPMC_HOOKS 1961145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1962145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1963145256Sjkoshy#endif 1964235459Srstone } else { 1965171482Sjeff thread_unblock_switch(td, mtx); 1966235459Srstone SDT_PROBE0(sched, , , remain_cpu); 1967235459Srstone } 1968171482Sjeff /* 1969171482Sjeff * Assert that all went well and return. 1970171482Sjeff */ 1971171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1972171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1973171482Sjeff td->td_oncpu = cpuid; 1974109864Sjeff} 1975109864Sjeff 1976171482Sjeff/* 1977171482Sjeff * Adjust thread priorities as a result of a nice request. 1978171482Sjeff */ 1979109864Sjeffvoid 1980130551Sjuliansched_nice(struct proc *p, int nice) 1981109864Sjeff{ 1982109864Sjeff struct thread *td; 1983109864Sjeff 1984130551Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 1985165762Sjeff 1986130551Sjulian p->p_nice = nice; 1987163709Sjb FOREACH_THREAD_IN_PROC(p, td) { 1988170293Sjeff thread_lock(td); 1989163709Sjb sched_priority(td); 1990165762Sjeff sched_prio(td, td->td_base_user_pri); 1991170293Sjeff thread_unlock(td); 1992130551Sjulian } 1993109864Sjeff} 1994109864Sjeff 1995171482Sjeff/* 1996171482Sjeff * Record the sleep time for the interactivity scorer. 1997171482Sjeff */ 1998109864Sjeffvoid 1999177085Sjeffsched_sleep(struct thread *td, int prio) 2000109864Sjeff{ 2001165762Sjeff 2002170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2003109864Sjeff 2004172264Sjeff td->td_slptick = ticks; 2005201347Skib if (TD_IS_SUSPENDED(td) || prio >= PSOCK) 2006177085Sjeff td->td_flags |= TDF_CANSWAP; 2007217410Sjhb if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 2008217410Sjhb return; 2009177903Sjeff if (static_boost == 1 && prio) 2010177085Sjeff sched_prio(td, prio); 2011177903Sjeff else if (static_boost && td->td_priority > static_boost) 2012177903Sjeff sched_prio(td, static_boost); 2013109864Sjeff} 2014109864Sjeff 2015171482Sjeff/* 2016171482Sjeff * Schedule a thread to resume execution and record how long it voluntarily 2017171482Sjeff * slept. We also update the pctcpu, interactivity, and priority. 2018171482Sjeff */ 2019109864Sjeffvoid 2020109864Sjeffsched_wakeup(struct thread *td) 2021109864Sjeff{ 2022166229Sjeff struct td_sched *ts; 2023171482Sjeff int slptick; 2024165762Sjeff 2025170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2026166229Sjeff ts = td->td_sched; 2027177085Sjeff td->td_flags &= ~TDF_CANSWAP; 2028109864Sjeff /* 2029165762Sjeff * If we slept for more than a tick update our interactivity and 2030165762Sjeff * priority. 2031109864Sjeff */ 2032172264Sjeff slptick = td->td_slptick; 2033172264Sjeff td->td_slptick = 0; 2034171482Sjeff if (slptick && slptick != ticks) { 2035232917Smav ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT; 2036165819Sjeff sched_interact_update(td); 2037232917Smav sched_pctcpu_update(ts, 0); 2038109864Sjeff } 2039242736Sjeff /* 2040242736Sjeff * Reset the slice value since we slept and advanced the round-robin. 2041242736Sjeff */ 2042242736Sjeff ts->ts_slice = 0; 2043166190Sjeff sched_add(td, SRQ_BORING); 2044109864Sjeff} 2045109864Sjeff 2046109864Sjeff/* 2047109864Sjeff * Penalize the parent for creating a new child and initialize the child's 2048109864Sjeff * priority. 2049109864Sjeff */ 2050109864Sjeffvoid 2051163709Sjbsched_fork(struct thread *td, struct thread *child) 2052109864Sjeff{ 2053170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2054232917Smav sched_pctcpu_update(td->td_sched, 1); 2055164936Sjulian sched_fork_thread(td, child); 2056165762Sjeff /* 2057165762Sjeff * Penalize the parent and child for forking. 2058165762Sjeff */ 2059165762Sjeff sched_interact_fork(child); 2060165762Sjeff sched_priority(child); 2061171482Sjeff td->td_sched->ts_runtime += tickincr; 2062165762Sjeff sched_interact_update(td); 2063165762Sjeff sched_priority(td); 2064164936Sjulian} 2065109864Sjeff 2066171482Sjeff/* 2067171482Sjeff * Fork a new thread, may be within the same process. 2068171482Sjeff */ 2069164936Sjulianvoid 2070164936Sjuliansched_fork_thread(struct thread *td, struct thread *child) 2071164936Sjulian{ 2072164936Sjulian struct td_sched *ts; 2073164936Sjulian struct td_sched *ts2; 2074242736Sjeff struct tdq *tdq; 2075164936Sjulian 2076242736Sjeff tdq = TDQ_SELF(); 2077177426Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2078165762Sjeff /* 2079165762Sjeff * Initialize child. 2080165762Sjeff */ 2081177426Sjeff ts = td->td_sched; 2082177426Sjeff ts2 = child->td_sched; 2083242736Sjeff child->td_lock = TDQ_LOCKPTR(tdq); 2084176735Sjeff child->td_cpuset = cpuset_ref(td->td_cpuset); 2085164936Sjulian ts2->ts_cpu = ts->ts_cpu; 2086177426Sjeff ts2->ts_flags = 0; 2087165762Sjeff /* 2088217078Sjhb * Grab our parents cpu estimation information. 2089165762Sjeff */ 2090164936Sjulian ts2->ts_ticks = ts->ts_ticks; 2091164936Sjulian ts2->ts_ltick = ts->ts_ltick; 2092164936Sjulian ts2->ts_ftick = ts->ts_ftick; 2093165762Sjeff /* 2094217078Sjhb * Do not inherit any borrowed priority from the parent. 2095217078Sjhb */ 2096217078Sjhb child->td_priority = child->td_base_pri; 2097217078Sjhb /* 2098165762Sjeff * And update interactivity score. 2099165762Sjeff */ 2100171482Sjeff ts2->ts_slptime = ts->ts_slptime; 2101171482Sjeff ts2->ts_runtime = ts->ts_runtime; 2102242736Sjeff /* Attempt to quickly learn interactivity. */ 2103242736Sjeff ts2->ts_slice = tdq_slice(tdq) - sched_slice_min; 2104187357Sjeff#ifdef KTR 2105187357Sjeff bzero(ts2->ts_name, sizeof(ts2->ts_name)); 2106187357Sjeff#endif 2107113357Sjeff} 2108113357Sjeff 2109171482Sjeff/* 2110171482Sjeff * Adjust the priority class of a thread. 2111171482Sjeff */ 2112113357Sjeffvoid 2113163709Sjbsched_class(struct thread *td, int class) 2114113357Sjeff{ 2115113357Sjeff 2116170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2117163709Sjb if (td->td_pri_class == class) 2118113357Sjeff return; 2119163709Sjb td->td_pri_class = class; 2120109864Sjeff} 2121109864Sjeff 2122109864Sjeff/* 2123109864Sjeff * Return some of the child's priority and interactivity to the parent. 2124109864Sjeff */ 2125109864Sjeffvoid 2126164939Sjuliansched_exit(struct proc *p, struct thread *child) 2127109864Sjeff{ 2128165762Sjeff struct thread *td; 2129113372Sjeff 2130187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit", 2131225199Sdelphij "prio:%d", child->td_priority); 2132177368Sjeff PROC_LOCK_ASSERT(p, MA_OWNED); 2133165762Sjeff td = FIRST_THREAD_IN_PROC(p); 2134165762Sjeff sched_exit_thread(td, child); 2135113372Sjeff} 2136113372Sjeff 2137171482Sjeff/* 2138171482Sjeff * Penalize another thread for the time spent on this one. This helps to 2139171482Sjeff * worsen the priority and interactivity of processes which schedule batch 2140171482Sjeff * jobs such as make. This has little effect on the make process itself but 2141171482Sjeff * causes new processes spawned by it to receive worse scores immediately. 2142171482Sjeff */ 2143113372Sjeffvoid 2144164939Sjuliansched_exit_thread(struct thread *td, struct thread *child) 2145164936Sjulian{ 2146165762Sjeff 2147187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit", 2148225199Sdelphij "prio:%d", child->td_priority); 2149165762Sjeff /* 2150165762Sjeff * Give the child's runtime to the parent without returning the 2151165762Sjeff * sleep time as a penalty to the parent. This causes shells that 2152165762Sjeff * launch expensive things to mark their children as expensive. 2153165762Sjeff */ 2154170293Sjeff thread_lock(td); 2155171482Sjeff td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2156164939Sjulian sched_interact_update(td); 2157165762Sjeff sched_priority(td); 2158170293Sjeff thread_unlock(td); 2159164936Sjulian} 2160164936Sjulian 2161177005Sjeffvoid 2162177005Sjeffsched_preempt(struct thread *td) 2163177005Sjeff{ 2164177005Sjeff struct tdq *tdq; 2165177005Sjeff 2166235459Srstone SDT_PROBE2(sched, , , surrender, td, td->td_proc); 2167235459Srstone 2168177005Sjeff thread_lock(td); 2169177005Sjeff tdq = TDQ_SELF(); 2170177005Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2171177005Sjeff tdq->tdq_ipipending = 0; 2172177005Sjeff if (td->td_priority > tdq->tdq_lowpri) { 2173178272Sjeff int flags; 2174178272Sjeff 2175178272Sjeff flags = SW_INVOL | SW_PREEMPT; 2176177005Sjeff if (td->td_critnest > 1) 2177177005Sjeff td->td_owepreempt = 1; 2178178272Sjeff else if (TD_IS_IDLETHREAD(td)) 2179178272Sjeff mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL); 2180177005Sjeff else 2181178272Sjeff mi_switch(flags | SWT_REMOTEPREEMPT, NULL); 2182177005Sjeff } 2183177005Sjeff thread_unlock(td); 2184177005Sjeff} 2185177005Sjeff 2186171482Sjeff/* 2187171482Sjeff * Fix priorities on return to user-space. Priorities may be elevated due 2188171482Sjeff * to static priorities in msleep() or similar. 2189171482Sjeff */ 2190164936Sjulianvoid 2191164936Sjuliansched_userret(struct thread *td) 2192164936Sjulian{ 2193164936Sjulian /* 2194164936Sjulian * XXX we cheat slightly on the locking here to avoid locking in 2195164936Sjulian * the usual case. Setting td_priority here is essentially an 2196164936Sjulian * incomplete workaround for not setting it properly elsewhere. 2197164936Sjulian * Now that some interrupt handlers are threads, not setting it 2198164936Sjulian * properly elsewhere can clobber it in the window between setting 2199164936Sjulian * it here and returning to user mode, so don't waste time setting 2200164936Sjulian * it perfectly here. 2201164936Sjulian */ 2202164936Sjulian KASSERT((td->td_flags & TDF_BORROWING) == 0, 2203164936Sjulian ("thread with borrowed priority returning to userland")); 2204164936Sjulian if (td->td_priority != td->td_user_pri) { 2205170293Sjeff thread_lock(td); 2206164936Sjulian td->td_priority = td->td_user_pri; 2207164936Sjulian td->td_base_pri = td->td_user_pri; 2208177005Sjeff tdq_setlowpri(TDQ_SELF(), td); 2209170293Sjeff thread_unlock(td); 2210164936Sjulian } 2211164936Sjulian} 2212164936Sjulian 2213171482Sjeff/* 2214171482Sjeff * Handle a stathz tick. This is really only relevant for timeshare 2215171482Sjeff * threads. 2216171482Sjeff */ 2217164936Sjulianvoid 2218121127Sjeffsched_clock(struct thread *td) 2219109864Sjeff{ 2220164936Sjulian struct tdq *tdq; 2221164936Sjulian struct td_sched *ts; 2222109864Sjeff 2223171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2224164936Sjulian tdq = TDQ_SELF(); 2225172409Sjeff#ifdef SMP 2226133427Sjeff /* 2227172409Sjeff * We run the long term load balancer infrequently on the first cpu. 2228172409Sjeff */ 2229172409Sjeff if (balance_tdq == tdq) { 2230172409Sjeff if (balance_ticks && --balance_ticks == 0) 2231172409Sjeff sched_balance(); 2232172409Sjeff } 2233172409Sjeff#endif 2234172409Sjeff /* 2235178277Sjeff * Save the old switch count so we have a record of the last ticks 2236178277Sjeff * activity. Initialize the new switch count based on our load. 2237178277Sjeff * If there is some activity seed it to reflect that. 2238178277Sjeff */ 2239178277Sjeff tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2240178471Sjeff tdq->tdq_switchcnt = tdq->tdq_load; 2241178277Sjeff /* 2242165766Sjeff * Advance the insert index once for each tick to ensure that all 2243165766Sjeff * threads get a chance to run. 2244133427Sjeff */ 2245165766Sjeff if (tdq->tdq_idx == tdq->tdq_ridx) { 2246165766Sjeff tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2247165766Sjeff if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2248165766Sjeff tdq->tdq_ridx = tdq->tdq_idx; 2249165766Sjeff } 2250165766Sjeff ts = td->td_sched; 2251232917Smav sched_pctcpu_update(ts, 1); 2252175104Sjeff if (td->td_pri_class & PRI_FIFO_BIT) 2253113357Sjeff return; 2254217291Sjhb if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) { 2255175104Sjeff /* 2256175104Sjeff * We used a tick; charge it to the thread so 2257175104Sjeff * that we can compute our interactivity. 2258175104Sjeff */ 2259175104Sjeff td->td_sched->ts_runtime += tickincr; 2260175104Sjeff sched_interact_update(td); 2261177009Sjeff sched_priority(td); 2262175104Sjeff } 2263239185Smav 2264113357Sjeff /* 2265239185Smav * Force a context switch if the current thread has used up a full 2266239185Smav * time slice (default is 100ms). 2267109864Sjeff */ 2268242736Sjeff if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) { 2269242736Sjeff ts->ts_slice = 0; 2270239185Smav td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND; 2271239185Smav } 2272109864Sjeff} 2273109864Sjeff 2274171482Sjeff/* 2275232917Smav * Called once per hz tick. 2276171482Sjeff */ 2277171482Sjeffvoid 2278212541Smavsched_tick(int cnt) 2279171482Sjeff{ 2280171482Sjeff 2281171482Sjeff} 2282171482Sjeff 2283171482Sjeff/* 2284171482Sjeff * Return whether the current CPU has runnable tasks. Used for in-kernel 2285171482Sjeff * cooperative idle threads. 2286171482Sjeff */ 2287109864Sjeffint 2288109864Sjeffsched_runnable(void) 2289109864Sjeff{ 2290164936Sjulian struct tdq *tdq; 2291115998Sjeff int load; 2292109864Sjeff 2293115998Sjeff load = 1; 2294115998Sjeff 2295164936Sjulian tdq = TDQ_SELF(); 2296121605Sjeff if ((curthread->td_flags & TDF_IDLETD) != 0) { 2297165620Sjeff if (tdq->tdq_load > 0) 2298121605Sjeff goto out; 2299121605Sjeff } else 2300165620Sjeff if (tdq->tdq_load - 1 > 0) 2301121605Sjeff goto out; 2302115998Sjeff load = 0; 2303115998Sjeffout: 2304115998Sjeff return (load); 2305109864Sjeff} 2306109864Sjeff 2307171482Sjeff/* 2308171482Sjeff * Choose the highest priority thread to run. The thread is removed from 2309171482Sjeff * the run-queue while running however the load remains. For SMP we set 2310171482Sjeff * the tdq in the global idle bitmask if it idles here. 2311171482Sjeff */ 2312166190Sjeffstruct thread * 2313109970Sjeffsched_choose(void) 2314109970Sjeff{ 2315177435Sjeff struct thread *td; 2316164936Sjulian struct tdq *tdq; 2317109970Sjeff 2318164936Sjulian tdq = TDQ_SELF(); 2319171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2320177435Sjeff td = tdq_choose(tdq); 2321177435Sjeff if (td) { 2322177435Sjeff tdq_runq_rem(tdq, td); 2323177903Sjeff tdq->tdq_lowpri = td->td_priority; 2324177435Sjeff return (td); 2325109864Sjeff } 2326177903Sjeff tdq->tdq_lowpri = PRI_MAX_IDLE; 2327176735Sjeff return (PCPU_GET(idlethread)); 2328109864Sjeff} 2329109864Sjeff 2330171482Sjeff/* 2331171482Sjeff * Set owepreempt if necessary. Preemption never happens directly in ULE, 2332171482Sjeff * we always request it once we exit a critical section. 2333171482Sjeff */ 2334171482Sjeffstatic inline void 2335171482Sjeffsched_setpreempt(struct thread *td) 2336166190Sjeff{ 2337166190Sjeff struct thread *ctd; 2338166190Sjeff int cpri; 2339166190Sjeff int pri; 2340166190Sjeff 2341177005Sjeff THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2342177005Sjeff 2343166190Sjeff ctd = curthread; 2344166190Sjeff pri = td->td_priority; 2345166190Sjeff cpri = ctd->td_priority; 2346177005Sjeff if (pri < cpri) 2347177005Sjeff ctd->td_flags |= TDF_NEEDRESCHED; 2348166190Sjeff if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2349171482Sjeff return; 2350177005Sjeff if (!sched_shouldpreempt(pri, cpri, 0)) 2351171482Sjeff return; 2352171482Sjeff ctd->td_owepreempt = 1; 2353166190Sjeff} 2354166190Sjeff 2355171482Sjeff/* 2356177009Sjeff * Add a thread to a thread queue. Select the appropriate runq and add the 2357177009Sjeff * thread to it. This is the internal function called when the tdq is 2358177009Sjeff * predetermined. 2359171482Sjeff */ 2360109864Sjeffvoid 2361171482Sjefftdq_add(struct tdq *tdq, struct thread *td, int flags) 2362109864Sjeff{ 2363109864Sjeff 2364171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2365166190Sjeff KASSERT((td->td_inhibitors == 0), 2366166190Sjeff ("sched_add: trying to run inhibited thread")); 2367166190Sjeff KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2368166190Sjeff ("sched_add: bad thread state")); 2369172207Sjeff KASSERT(td->td_flags & TDF_INMEM, 2370172207Sjeff ("sched_add: thread swapped out")); 2371171482Sjeff 2372171482Sjeff if (td->td_priority < tdq->tdq_lowpri) 2373171482Sjeff tdq->tdq_lowpri = td->td_priority; 2374177435Sjeff tdq_runq_add(tdq, td, flags); 2375177435Sjeff tdq_load_add(tdq, td); 2376171482Sjeff} 2377171482Sjeff 2378171482Sjeff/* 2379171482Sjeff * Select the target thread queue and add a thread to it. Request 2380171482Sjeff * preemption or IPI a remote processor if required. 2381171482Sjeff */ 2382171482Sjeffvoid 2383171482Sjeffsched_add(struct thread *td, int flags) 2384171482Sjeff{ 2385171482Sjeff struct tdq *tdq; 2386171482Sjeff#ifdef SMP 2387171482Sjeff int cpu; 2388171482Sjeff#endif 2389187357Sjeff 2390187357Sjeff KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 2391187357Sjeff "prio:%d", td->td_priority, KTR_ATTR_LINKED, 2392187357Sjeff sched_tdname(curthread)); 2393187357Sjeff KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 2394187357Sjeff KTR_ATTR_LINKED, sched_tdname(td)); 2395235459Srstone SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 2396235459Srstone flags & SRQ_PREEMPTED); 2397171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2398166108Sjeff /* 2399171482Sjeff * Recalculate the priority before we select the target cpu or 2400171482Sjeff * run-queue. 2401166108Sjeff */ 2402171482Sjeff if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2403171482Sjeff sched_priority(td); 2404171482Sjeff#ifdef SMP 2405171482Sjeff /* 2406171482Sjeff * Pick the destination cpu and if it isn't ours transfer to the 2407171482Sjeff * target cpu. 2408171482Sjeff */ 2409177435Sjeff cpu = sched_pickcpu(td, flags); 2410177435Sjeff tdq = sched_setcpu(td, cpu, flags); 2411171482Sjeff tdq_add(tdq, td, flags); 2412177009Sjeff if (cpu != PCPU_GET(cpuid)) { 2413177435Sjeff tdq_notify(tdq, td); 2414166108Sjeff return; 2415166108Sjeff } 2416171482Sjeff#else 2417171482Sjeff tdq = TDQ_SELF(); 2418171482Sjeff TDQ_LOCK(tdq); 2419171482Sjeff /* 2420171482Sjeff * Now that the thread is moving to the run-queue, set the lock 2421171482Sjeff * to the scheduler's lock. 2422171482Sjeff */ 2423171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2424171482Sjeff tdq_add(tdq, td, flags); 2425166108Sjeff#endif 2426171482Sjeff if (!(flags & SRQ_YIELDING)) 2427171482Sjeff sched_setpreempt(td); 2428109864Sjeff} 2429109864Sjeff 2430171482Sjeff/* 2431171482Sjeff * Remove a thread from a run-queue without running it. This is used 2432171482Sjeff * when we're stealing a thread from a remote queue. Otherwise all threads 2433171482Sjeff * exit by calling sched_exit_thread() and sched_throw() themselves. 2434171482Sjeff */ 2435109864Sjeffvoid 2436121127Sjeffsched_rem(struct thread *td) 2437109864Sjeff{ 2438164936Sjulian struct tdq *tdq; 2439113357Sjeff 2440187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 2441187357Sjeff "prio:%d", td->td_priority); 2442235459Srstone SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); 2443177435Sjeff tdq = TDQ_CPU(td->td_sched->ts_cpu); 2444171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2445171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2446166190Sjeff KASSERT(TD_ON_RUNQ(td), 2447164936Sjulian ("sched_rem: thread not on run queue")); 2448177435Sjeff tdq_runq_rem(tdq, td); 2449177435Sjeff tdq_load_rem(tdq, td); 2450166190Sjeff TD_SET_CAN_RUN(td); 2451176735Sjeff if (td->td_priority == tdq->tdq_lowpri) 2452176735Sjeff tdq_setlowpri(tdq, NULL); 2453109864Sjeff} 2454109864Sjeff 2455171482Sjeff/* 2456171482Sjeff * Fetch cpu utilization information. Updates on demand. 2457171482Sjeff */ 2458109864Sjefffixpt_t 2459121127Sjeffsched_pctcpu(struct thread *td) 2460109864Sjeff{ 2461109864Sjeff fixpt_t pctcpu; 2462164936Sjulian struct td_sched *ts; 2463109864Sjeff 2464109864Sjeff pctcpu = 0; 2465164936Sjulian ts = td->td_sched; 2466164936Sjulian if (ts == NULL) 2467121290Sjeff return (0); 2468109864Sjeff 2469208787Sjhb THREAD_LOCK_ASSERT(td, MA_OWNED); 2470232917Smav sched_pctcpu_update(ts, TD_IS_RUNNING(td)); 2471164936Sjulian if (ts->ts_ticks) { 2472109864Sjeff int rtick; 2473109864Sjeff 2474109864Sjeff /* How many rtick per second ? */ 2475165762Sjeff rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2476165762Sjeff pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2477109864Sjeff } 2478109864Sjeff 2479109864Sjeff return (pctcpu); 2480109864Sjeff} 2481109864Sjeff 2482176735Sjeff/* 2483176735Sjeff * Enforce affinity settings for a thread. Called after adjustments to 2484176735Sjeff * cpumask. 2485176735Sjeff */ 2486176729Sjeffvoid 2487176729Sjeffsched_affinity(struct thread *td) 2488176729Sjeff{ 2489176735Sjeff#ifdef SMP 2490176735Sjeff struct td_sched *ts; 2491176735Sjeff 2492176735Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2493176735Sjeff ts = td->td_sched; 2494176735Sjeff if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2495176735Sjeff return; 2496189787Sjeff if (TD_ON_RUNQ(td)) { 2497189787Sjeff sched_rem(td); 2498189787Sjeff sched_add(td, SRQ_BORING); 2499189787Sjeff return; 2500189787Sjeff } 2501176735Sjeff if (!TD_IS_RUNNING(td)) 2502176735Sjeff return; 2503176735Sjeff /* 2504212115Smdf * Force a switch before returning to userspace. If the 2505212115Smdf * target thread is not running locally send an ipi to force 2506212115Smdf * the issue. 2507176735Sjeff */ 2508212974Sjhb td->td_flags |= TDF_NEEDRESCHED; 2509212115Smdf if (td != curthread) 2510212115Smdf ipi_cpu(ts->ts_cpu, IPI_PREEMPT); 2511176735Sjeff#endif 2512176729Sjeff} 2513176729Sjeff 2514171482Sjeff/* 2515171482Sjeff * Bind a thread to a target cpu. 2516171482Sjeff */ 2517122038Sjeffvoid 2518122038Sjeffsched_bind(struct thread *td, int cpu) 2519122038Sjeff{ 2520164936Sjulian struct td_sched *ts; 2521122038Sjeff 2522171713Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2523208391Sjhb KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 2524164936Sjulian ts = td->td_sched; 2525166137Sjeff if (ts->ts_flags & TSF_BOUND) 2526166152Sjeff sched_unbind(td); 2527212115Smdf KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td)); 2528164936Sjulian ts->ts_flags |= TSF_BOUND; 2529166137Sjeff sched_pin(); 2530123433Sjeff if (PCPU_GET(cpuid) == cpu) 2531122038Sjeff return; 2532166137Sjeff ts->ts_cpu = cpu; 2533122038Sjeff /* When we return from mi_switch we'll be on the correct cpu. */ 2534131527Sphk mi_switch(SW_VOL, NULL); 2535122038Sjeff} 2536122038Sjeff 2537171482Sjeff/* 2538171482Sjeff * Release a bound thread. 2539171482Sjeff */ 2540122038Sjeffvoid 2541122038Sjeffsched_unbind(struct thread *td) 2542122038Sjeff{ 2543165762Sjeff struct td_sched *ts; 2544165762Sjeff 2545170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2546208391Sjhb KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 2547165762Sjeff ts = td->td_sched; 2548166137Sjeff if ((ts->ts_flags & TSF_BOUND) == 0) 2549166137Sjeff return; 2550165762Sjeff ts->ts_flags &= ~TSF_BOUND; 2551165762Sjeff sched_unpin(); 2552122038Sjeff} 2553122038Sjeff 2554109864Sjeffint 2555145256Sjkoshysched_is_bound(struct thread *td) 2556145256Sjkoshy{ 2557170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2558164936Sjulian return (td->td_sched->ts_flags & TSF_BOUND); 2559145256Sjkoshy} 2560145256Sjkoshy 2561171482Sjeff/* 2562171482Sjeff * Basic yield call. 2563171482Sjeff */ 2564159630Sdavidxuvoid 2565159630Sdavidxusched_relinquish(struct thread *td) 2566159630Sdavidxu{ 2567170293Sjeff thread_lock(td); 2568178272Sjeff mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 2569170293Sjeff thread_unlock(td); 2570159630Sdavidxu} 2571159630Sdavidxu 2572171482Sjeff/* 2573171482Sjeff * Return the total system load. 2574171482Sjeff */ 2575145256Sjkoshyint 2576125289Sjeffsched_load(void) 2577125289Sjeff{ 2578125289Sjeff#ifdef SMP 2579125289Sjeff int total; 2580125289Sjeff int i; 2581125289Sjeff 2582125289Sjeff total = 0; 2583209059Sjhb CPU_FOREACH(i) 2584176735Sjeff total += TDQ_CPU(i)->tdq_sysload; 2585125289Sjeff return (total); 2586125289Sjeff#else 2587165620Sjeff return (TDQ_SELF()->tdq_sysload); 2588125289Sjeff#endif 2589125289Sjeff} 2590125289Sjeff 2591125289Sjeffint 2592109864Sjeffsched_sizeof_proc(void) 2593109864Sjeff{ 2594109864Sjeff return (sizeof(struct proc)); 2595109864Sjeff} 2596109864Sjeff 2597109864Sjeffint 2598109864Sjeffsched_sizeof_thread(void) 2599109864Sjeff{ 2600109864Sjeff return (sizeof(struct thread) + sizeof(struct td_sched)); 2601109864Sjeff} 2602159570Sdavidxu 2603191676Sjeff#ifdef SMP 2604191676Sjeff#define TDQ_IDLESPIN(tdq) \ 2605191676Sjeff ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0) 2606191676Sjeff#else 2607191676Sjeff#define TDQ_IDLESPIN(tdq) 1 2608191676Sjeff#endif 2609191676Sjeff 2610166190Sjeff/* 2611166190Sjeff * The actual idle process. 2612166190Sjeff */ 2613166190Sjeffvoid 2614166190Sjeffsched_idletd(void *dummy) 2615166190Sjeff{ 2616166190Sjeff struct thread *td; 2617171482Sjeff struct tdq *tdq; 2618242852Smav int oldswitchcnt, switchcnt; 2619178277Sjeff int i; 2620166190Sjeff 2621191643Sjeff mtx_assert(&Giant, MA_NOTOWNED); 2622166190Sjeff td = curthread; 2623171482Sjeff tdq = TDQ_SELF(); 2624239585Sjhb THREAD_NO_SLEEPING(); 2625242852Smav oldswitchcnt = -1; 2626171482Sjeff for (;;) { 2627242852Smav if (tdq->tdq_load) { 2628242852Smav thread_lock(td); 2629242852Smav mi_switch(SW_VOL | SWT_IDLE, NULL); 2630242852Smav thread_unlock(td); 2631242852Smav } 2632242852Smav switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2633171482Sjeff#ifdef SMP 2634242852Smav if (switchcnt != oldswitchcnt) { 2635242852Smav oldswitchcnt = switchcnt; 2636242852Smav if (tdq_idled(tdq) == 0) 2637242852Smav continue; 2638242852Smav } 2639171482Sjeff#endif 2640178277Sjeff switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2641178277Sjeff /* 2642178277Sjeff * If we're switching very frequently, spin while checking 2643178277Sjeff * for load rather than entering a low power state that 2644191643Sjeff * may require an IPI. However, don't do any busy 2645191643Sjeff * loops while on SMT machines as this simply steals 2646191643Sjeff * cycles from cores doing useful work. 2647178277Sjeff */ 2648191676Sjeff if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) { 2649178277Sjeff for (i = 0; i < sched_idlespins; i++) { 2650178277Sjeff if (tdq->tdq_load) 2651178277Sjeff break; 2652178277Sjeff cpu_spinwait(); 2653178277Sjeff } 2654178277Sjeff } 2655242852Smav 2656242852Smav /* If there was context switch during spin, restart it. */ 2657191643Sjeff switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2658242852Smav if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt) 2659242852Smav continue; 2660242852Smav 2661242852Smav /* Run main MD idle handler. */ 2662242852Smav tdq->tdq_cpu_idle = 1; 2663242852Smav cpu_idle(switchcnt * 4 > sched_idlespinthresh); 2664242852Smav tdq->tdq_cpu_idle = 0; 2665242852Smav 2666242852Smav /* 2667242852Smav * Account thread-less hardware interrupts and 2668242852Smav * other wakeup reasons equal to context switches. 2669242852Smav */ 2670242852Smav switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2671242852Smav if (switchcnt != oldswitchcnt) 2672242852Smav continue; 2673242852Smav tdq->tdq_switchcnt++; 2674242852Smav oldswitchcnt++; 2675171482Sjeff } 2676166190Sjeff} 2677166190Sjeff 2678170293Sjeff/* 2679170293Sjeff * A CPU is entering for the first time or a thread is exiting. 2680170293Sjeff */ 2681170293Sjeffvoid 2682170293Sjeffsched_throw(struct thread *td) 2683170293Sjeff{ 2684172411Sjeff struct thread *newtd; 2685171482Sjeff struct tdq *tdq; 2686171482Sjeff 2687171482Sjeff tdq = TDQ_SELF(); 2688170293Sjeff if (td == NULL) { 2689171482Sjeff /* Correct spinlock nesting and acquire the correct lock. */ 2690171482Sjeff TDQ_LOCK(tdq); 2691170293Sjeff spinlock_exit(); 2692229429Sjhb PCPU_SET(switchtime, cpu_ticks()); 2693229429Sjhb PCPU_SET(switchticks, ticks); 2694170293Sjeff } else { 2695171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2696177435Sjeff tdq_load_rem(tdq, td); 2697174629Sjeff lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2698170293Sjeff } 2699170293Sjeff KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2700172411Sjeff newtd = choosethread(); 2701172411Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2702172411Sjeff cpu_throw(td, newtd); /* doesn't return */ 2703170293Sjeff} 2704170293Sjeff 2705171482Sjeff/* 2706171482Sjeff * This is called from fork_exit(). Just acquire the correct locks and 2707171482Sjeff * let fork do the rest of the work. 2708171482Sjeff */ 2709170293Sjeffvoid 2710170600Sjeffsched_fork_exit(struct thread *td) 2711170293Sjeff{ 2712171482Sjeff struct td_sched *ts; 2713171482Sjeff struct tdq *tdq; 2714171482Sjeff int cpuid; 2715170293Sjeff 2716170293Sjeff /* 2717170293Sjeff * Finish setting up thread glue so that it begins execution in a 2718171482Sjeff * non-nested critical section with the scheduler lock held. 2719170293Sjeff */ 2720171482Sjeff cpuid = PCPU_GET(cpuid); 2721171482Sjeff tdq = TDQ_CPU(cpuid); 2722171482Sjeff ts = td->td_sched; 2723171482Sjeff if (TD_IS_IDLETHREAD(td)) 2724171482Sjeff td->td_lock = TDQ_LOCKPTR(tdq); 2725171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2726171482Sjeff td->td_oncpu = cpuid; 2727172411Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2728174629Sjeff lock_profile_obtain_lock_success( 2729174629Sjeff &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2730170293Sjeff} 2731170293Sjeff 2732187357Sjeff/* 2733187357Sjeff * Create on first use to catch odd startup conditons. 2734187357Sjeff */ 2735187357Sjeffchar * 2736187357Sjeffsched_tdname(struct thread *td) 2737187357Sjeff{ 2738187357Sjeff#ifdef KTR 2739187357Sjeff struct td_sched *ts; 2740187357Sjeff 2741187357Sjeff ts = td->td_sched; 2742187357Sjeff if (ts->ts_name[0] == '\0') 2743187357Sjeff snprintf(ts->ts_name, sizeof(ts->ts_name), 2744187357Sjeff "%s tid %d", td->td_name, td->td_tid); 2745187357Sjeff return (ts->ts_name); 2746187357Sjeff#else 2747187357Sjeff return (td->td_name); 2748187357Sjeff#endif 2749187357Sjeff} 2750187357Sjeff 2751232700Sjhb#ifdef KTR 2752232700Sjhbvoid 2753232700Sjhbsched_clear_tdname(struct thread *td) 2754232700Sjhb{ 2755232700Sjhb struct td_sched *ts; 2756232700Sjhb 2757232700Sjhb ts = td->td_sched; 2758232700Sjhb ts->ts_name[0] = '\0'; 2759232700Sjhb} 2760232700Sjhb#endif 2761232700Sjhb 2762184439Sivoras#ifdef SMP 2763184439Sivoras 2764184439Sivoras/* 2765184439Sivoras * Build the CPU topology dump string. Is recursively called to collect 2766184439Sivoras * the topology tree. 2767184439Sivoras */ 2768184439Sivorasstatic int 2769184439Sivorassysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg, 2770184439Sivoras int indent) 2771184439Sivoras{ 2772222813Sattilio char cpusetbuf[CPUSETBUFSIZ]; 2773184439Sivoras int i, first; 2774184439Sivoras 2775184439Sivoras sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent, 2776212821Savg "", 1 + indent / 2, cg->cg_level); 2777222813Sattilio sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "", 2778222813Sattilio cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask)); 2779184439Sivoras first = TRUE; 2780184439Sivoras for (i = 0; i < MAXCPU; i++) { 2781222813Sattilio if (CPU_ISSET(i, &cg->cg_mask)) { 2782184439Sivoras if (!first) 2783184439Sivoras sbuf_printf(sb, ", "); 2784184439Sivoras else 2785184439Sivoras first = FALSE; 2786184439Sivoras sbuf_printf(sb, "%d", i); 2787184439Sivoras } 2788184439Sivoras } 2789184439Sivoras sbuf_printf(sb, "</cpu>\n"); 2790184439Sivoras 2791184439Sivoras if (cg->cg_flags != 0) { 2792210117Sivoras sbuf_printf(sb, "%*s <flags>", indent, ""); 2793184439Sivoras if ((cg->cg_flags & CG_FLAG_HTT) != 0) 2794208982Sivoras sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>"); 2795208983Sivoras if ((cg->cg_flags & CG_FLAG_THREAD) != 0) 2796208983Sivoras sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>"); 2797191643Sjeff if ((cg->cg_flags & CG_FLAG_SMT) != 0) 2798208983Sivoras sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>"); 2799210117Sivoras sbuf_printf(sb, "</flags>\n"); 2800184439Sivoras } 2801184439Sivoras 2802184439Sivoras if (cg->cg_children > 0) { 2803184439Sivoras sbuf_printf(sb, "%*s <children>\n", indent, ""); 2804184439Sivoras for (i = 0; i < cg->cg_children; i++) 2805184439Sivoras sysctl_kern_sched_topology_spec_internal(sb, 2806184439Sivoras &cg->cg_child[i], indent+2); 2807184439Sivoras sbuf_printf(sb, "%*s </children>\n", indent, ""); 2808184439Sivoras } 2809184439Sivoras sbuf_printf(sb, "%*s</group>\n", indent, ""); 2810184439Sivoras return (0); 2811184439Sivoras} 2812184439Sivoras 2813184439Sivoras/* 2814184439Sivoras * Sysctl handler for retrieving topology dump. It's a wrapper for 2815184439Sivoras * the recursive sysctl_kern_smp_topology_spec_internal(). 2816184439Sivoras */ 2817184439Sivorasstatic int 2818184439Sivorassysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS) 2819184439Sivoras{ 2820184439Sivoras struct sbuf *topo; 2821184439Sivoras int err; 2822184439Sivoras 2823184439Sivoras KASSERT(cpu_top != NULL, ("cpu_top isn't initialized")); 2824184439Sivoras 2825184570Sivoras topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND); 2826184439Sivoras if (topo == NULL) 2827184439Sivoras return (ENOMEM); 2828184439Sivoras 2829184439Sivoras sbuf_printf(topo, "<groups>\n"); 2830184439Sivoras err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1); 2831184439Sivoras sbuf_printf(topo, "</groups>\n"); 2832184439Sivoras 2833184439Sivoras if (err == 0) { 2834184439Sivoras sbuf_finish(topo); 2835184439Sivoras err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo)); 2836184439Sivoras } 2837184439Sivoras sbuf_delete(topo); 2838184439Sivoras return (err); 2839184439Sivoras} 2840214510Sdavidxu 2841184439Sivoras#endif 2842184439Sivoras 2843239185Smavstatic int 2844239185Smavsysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 2845239185Smav{ 2846239185Smav int error, new_val, period; 2847239185Smav 2848239185Smav period = 1000000 / realstathz; 2849239185Smav new_val = period * sched_slice; 2850239185Smav error = sysctl_handle_int(oidp, &new_val, 0, req); 2851239196Smav if (error != 0 || req->newptr == NULL) 2852239185Smav return (error); 2853239185Smav if (new_val <= 0) 2854239185Smav return (EINVAL); 2855239196Smav sched_slice = imax(1, (new_val + period / 2) / period); 2856242736Sjeff sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR; 2857239196Smav hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 2858239196Smav realstathz); 2859239185Smav return (0); 2860239185Smav} 2861239185Smav 2862177435SjeffSYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2863171482SjeffSYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2864165762Sjeff "Scheduler name"); 2865239185SmavSYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW, 2866239185Smav NULL, 0, sysctl_kern_quantum, "I", 2867239196Smav "Quantum for timeshare threads in microseconds"); 2868171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2869239196Smav "Quantum for timeshare threads in stathz ticks"); 2870171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2871239196Smav "Interactivity score threshold"); 2872239196SmavSYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, 2873239196Smav &preempt_thresh, 0, 2874239196Smav "Maximal (lowest) priority for preemption"); 2875239196SmavSYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0, 2876239196Smav "Assign static kernel priorities to sleeping threads"); 2877239196SmavSYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0, 2878239196Smav "Number of times idle thread will spin waiting for new work"); 2879239196SmavSYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, 2880239196Smav &sched_idlespinthresh, 0, 2881239196Smav "Threshold before we will permit idle thread spinning"); 2882166108Sjeff#ifdef SMP 2883171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2884171482Sjeff "Number of hz ticks to keep thread affinity for"); 2885171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2886171482Sjeff "Enables the long-term load balancer"); 2887172409SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2888172409Sjeff &balance_interval, 0, 2889239185Smav "Average period in stathz ticks to run the long-term balancer"); 2890171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2891171482Sjeff "Attempts to steal work from other cores before idling"); 2892171506SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2893239196Smav "Minimum load on remote CPU before we'll steal"); 2894184439SivorasSYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING | 2895239185Smav CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A", 2896184439Sivoras "XML dump of detected CPU topology"); 2897166108Sjeff#endif 2898165762Sjeff 2899172264Sjeff/* ps compat. All cpu percentages from ULE are weighted. */ 2900172293Sjeffstatic int ccpu = 0; 2901165762SjeffSYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2902