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: stable/10/sys/kern/sched_ule.c 316841 2017-04-14 14:44:06Z avg $"); 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 80171482Sjeff#define KTR_ULE 0 81166137Sjeff 82187679Sjeff#define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX))) 83187679Sjeff#define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU))) 84224221Sattilio#define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load")) 85187357Sjeff 86166137Sjeff/* 87171482Sjeff * Thread scheduler specific section. All fields are protected 88171482Sjeff * by the thread lock. 89146954Sjeff */ 90164936Sjulianstruct td_sched { 91171482Sjeff struct runq *ts_runq; /* Run-queue we're queued on. */ 92171482Sjeff short ts_flags; /* TSF_* flags. */ 93164936Sjulian u_char ts_cpu; /* CPU that we have affinity for. */ 94177009Sjeff int ts_rltick; /* Real last tick, for affinity. */ 95171482Sjeff int ts_slice; /* Ticks of slice remaining. */ 96171482Sjeff u_int ts_slptime; /* Number of ticks we vol. slept */ 97171482Sjeff u_int ts_runtime; /* Number of ticks we were running */ 98164936Sjulian int ts_ltick; /* Last tick that we were running on */ 99164936Sjulian int ts_ftick; /* First tick that we were running on */ 100164936Sjulian int ts_ticks; /* Tick count */ 101187357Sjeff#ifdef KTR 102187357Sjeff char ts_name[TS_NAME_LEN]; 103187357Sjeff#endif 104134791Sjulian}; 105164936Sjulian/* flags kept in ts_flags */ 106166108Sjeff#define TSF_BOUND 0x0001 /* Thread can not migrate. */ 107166108Sjeff#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 108121790Sjeff 109164936Sjulianstatic struct td_sched td_sched0; 110109864Sjeff 111176735Sjeff#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 112176735Sjeff#define THREAD_CAN_SCHED(td, cpu) \ 113176735Sjeff CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 114176735Sjeff 115109864Sjeff/* 116217351Sjhb * Priority ranges used for interactive and non-interactive timeshare 117217410Sjhb * threads. The timeshare priorities are split up into four ranges. 118217410Sjhb * The first range handles interactive threads. The last three ranges 119217410Sjhb * (NHALF, x, and NHALF) handle non-interactive threads with the outer 120217410Sjhb * ranges supporting nice values. 121217351Sjhb */ 122217410Sjhb#define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1) 123217410Sjhb#define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2) 124228718Savg#define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE) 125217410Sjhb 126217410Sjhb#define PRI_MIN_INTERACT PRI_MIN_TIMESHARE 127217410Sjhb#define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1) 128217410Sjhb#define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE) 129217351Sjhb#define PRI_MAX_BATCH PRI_MAX_TIMESHARE 130217351Sjhb 131217351Sjhb/* 132165762Sjeff * Cpu percentage computation macros and defines. 133111857Sjeff * 134165762Sjeff * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 135165762Sjeff * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 136165796Sjeff * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 137165762Sjeff * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 138165762Sjeff * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 139165762Sjeff * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 140165762Sjeff */ 141165762Sjeff#define SCHED_TICK_SECS 10 142165762Sjeff#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 143165796Sjeff#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 144165762Sjeff#define SCHED_TICK_SHIFT 10 145165762Sjeff#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 146165830Sjeff#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 147165762Sjeff 148165762Sjeff/* 149165762Sjeff * These macros determine priorities for non-interactive threads. They are 150165762Sjeff * assigned a priority based on their recent cpu utilization as expressed 151165762Sjeff * by the ratio of ticks to the tick total. NHALF priorities at the start 152165762Sjeff * and end of the MIN to MAX timeshare range are only reachable with negative 153165762Sjeff * or positive nice respectively. 154165762Sjeff * 155165762Sjeff * PRI_RANGE: Priority range for utilization dependent priorities. 156116642Sjeff * PRI_NRESV: Number of nice values. 157165762Sjeff * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 158165762Sjeff * PRI_NICE: Determines the part of the priority inherited from nice. 159109864Sjeff */ 160165762Sjeff#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 161121869Sjeff#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 162217351Sjhb#define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF) 163217351Sjhb#define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF) 164217237Sjhb#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1) 165165762Sjeff#define SCHED_PRI_TICKS(ts) \ 166165762Sjeff (SCHED_TICK_HZ((ts)) / \ 167165827Sjeff (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 168165762Sjeff#define SCHED_PRI_NICE(nice) (nice) 169109864Sjeff 170109864Sjeff/* 171165762Sjeff * These determine the interactivity of a process. Interactivity differs from 172165762Sjeff * cpu utilization in that it expresses the voluntary time slept vs time ran 173165762Sjeff * while cpu utilization includes all time not running. This more accurately 174165762Sjeff * models the intent of the thread. 175109864Sjeff * 176110645Sjeff * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 177110645Sjeff * before throttling back. 178121868Sjeff * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 179116365Sjeff * INTERACT_MAX: Maximum interactivity value. Smaller is better. 180215102Sattilio * INTERACT_THRESH: Threshold for placement on the current runq. 181109864Sjeff */ 182165762Sjeff#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 183165762Sjeff#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 184116365Sjeff#define SCHED_INTERACT_MAX (100) 185116365Sjeff#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 186121126Sjeff#define SCHED_INTERACT_THRESH (30) 187111857Sjeff 188242736Sjeff/* 189242736Sjeff * These parameters determine the slice behavior for batch work. 190242736Sjeff */ 191242736Sjeff#define SCHED_SLICE_DEFAULT_DIVISOR 10 /* ~94 ms, 12 stathz ticks. */ 192242736Sjeff#define SCHED_SLICE_MIN_DIVISOR 6 /* DEFAULT/MIN = ~16 ms. */ 193242736Sjeff 194239157Smav/* Flags kept in td_flags. */ 195239157Smav#define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */ 196239157Smav 197109864Sjeff/* 198165762Sjeff * tickincr: Converts a stathz tick into a hz domain scaled by 199165762Sjeff * the shift factor. Without the shift the error rate 200165762Sjeff * due to rounding would be unacceptably high. 201165762Sjeff * realstathz: stathz is sometimes 0 and run off of hz. 202165762Sjeff * sched_slice: Runtime of each thread before rescheduling. 203171482Sjeff * preempt_thresh: Priority threshold for preemption and remote IPIs. 204109864Sjeff */ 205165762Sjeffstatic int sched_interact = SCHED_INTERACT_THRESH; 206241844Seadlerstatic int tickincr = 8 << SCHED_TICK_SHIFT; 207242736Sjeffstatic int realstathz = 127; /* reset during boot. */ 208242736Sjeffstatic int sched_slice = 10; /* reset during boot. */ 209242736Sjeffstatic int sched_slice_min = 1; /* reset during boot. */ 210172345Sjeff#ifdef PREEMPTION 211172345Sjeff#ifdef FULL_PREEMPTION 212172345Sjeffstatic int preempt_thresh = PRI_MAX_IDLE; 213172345Sjeff#else 214171482Sjeffstatic int preempt_thresh = PRI_MIN_KERN; 215172345Sjeff#endif 216172345Sjeff#else 217172345Sjeffstatic int preempt_thresh = 0; 218172345Sjeff#endif 219217351Sjhbstatic int static_boost = PRI_MIN_BATCH; 220178277Sjeffstatic int sched_idlespins = 10000; 221232740Smavstatic int sched_idlespinthresh = -1; 222109864Sjeff 223109864Sjeff/* 224171482Sjeff * tdq - per processor runqs and statistics. All fields are protected by the 225171482Sjeff * tdq_lock. The load and lowpri may be accessed without to avoid excess 226171482Sjeff * locking in sched_pickcpu(); 227109864Sjeff */ 228164936Sjulianstruct tdq { 229242014Sjimharris /* 230242014Sjimharris * Ordered to improve efficiency of cpu_search() and switch(). 231242014Sjimharris * tdq_lock is padded to avoid false sharing with tdq_load and 232242014Sjimharris * tdq_cpu_idle. 233242014Sjimharris */ 234242402Sattilio struct mtx_padalign tdq_lock; /* run queue lock. */ 235176735Sjeff struct cpu_group *tdq_cg; /* Pointer to cpu topology. */ 236178277Sjeff volatile int tdq_load; /* Aggregate load. */ 237212416Smav volatile int tdq_cpu_idle; /* cpu_idle() is active. */ 238176735Sjeff int tdq_sysload; /* For loadavg, !ITHD load. */ 239177009Sjeff int tdq_transferable; /* Transferable thread count. */ 240178277Sjeff short tdq_switchcnt; /* Switches this tick. */ 241178277Sjeff short tdq_oldswitchcnt; /* Switches last tick. */ 242177009Sjeff u_char tdq_lowpri; /* Lowest priority thread. */ 243177009Sjeff u_char tdq_ipipending; /* IPI pending. */ 244166557Sjeff u_char tdq_idx; /* Current insert index. */ 245166557Sjeff u_char tdq_ridx; /* Current removal index. */ 246177009Sjeff struct runq tdq_realtime; /* real-time run queue. */ 247177009Sjeff struct runq tdq_timeshare; /* timeshare run queue. */ 248177009Sjeff struct runq tdq_idle; /* Queue of IDLE threads. */ 249187357Sjeff char tdq_name[TDQ_NAME_LEN]; 250187357Sjeff#ifdef KTR 251187357Sjeff char tdq_loadname[TDQ_LOADNAME_LEN]; 252187357Sjeff#endif 253171482Sjeff} __aligned(64); 254109864Sjeff 255178277Sjeff/* Idle thread states and config. */ 256178277Sjeff#define TDQ_RUNNING 1 257178277Sjeff#define TDQ_IDLE 2 258166108Sjeff 259123433Sjeff#ifdef SMP 260184439Sivorasstruct cpu_group *cpu_top; /* CPU topology */ 261123433Sjeff 262176735Sjeff#define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000)) 263176735Sjeff#define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity)) 264166108Sjeff 265123433Sjeff/* 266166108Sjeff * Run-time tunables. 267166108Sjeff */ 268171506Sjeffstatic int rebalance = 1; 269172409Sjeffstatic int balance_interval = 128; /* Default set in sched_initticks(). */ 270166108Sjeffstatic int affinity; 271171506Sjeffstatic int steal_idle = 1; 272171506Sjeffstatic int steal_thresh = 2; 273166108Sjeff 274166108Sjeff/* 275165620Sjeff * One thread queue per processor. 276109864Sjeff */ 277164936Sjulianstatic struct tdq tdq_cpu[MAXCPU]; 278172409Sjeffstatic struct tdq *balance_tdq; 279172409Sjeffstatic int balance_ticks; 280232207Smavstatic DPCPU_DEFINE(uint32_t, randomval); 281129982Sjeff 282164936Sjulian#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 283164936Sjulian#define TDQ_CPU(x) (&tdq_cpu[(x)]) 284171713Sjeff#define TDQ_ID(x) ((int)((x) - tdq_cpu)) 285123433Sjeff#else /* !SMP */ 286164936Sjulianstatic struct tdq tdq_cpu; 287129982Sjeff 288170315Sjeff#define TDQ_ID(x) (0) 289164936Sjulian#define TDQ_SELF() (&tdq_cpu) 290164936Sjulian#define TDQ_CPU(x) (&tdq_cpu) 291110028Sjeff#endif 292109864Sjeff 293171482Sjeff#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 294171482Sjeff#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 295171482Sjeff#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 296171482Sjeff#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 297242402Sattilio#define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock)) 298171482Sjeff 299163709Sjbstatic void sched_priority(struct thread *); 300146954Sjeffstatic void sched_thread_priority(struct thread *, u_char); 301163709Sjbstatic int sched_interact_score(struct thread *); 302163709Sjbstatic void sched_interact_update(struct thread *); 303163709Sjbstatic void sched_interact_fork(struct thread *); 304232917Smavstatic void sched_pctcpu_update(struct td_sched *, int); 305109864Sjeff 306110267Sjeff/* Operations on per processor queues */ 307177435Sjeffstatic struct thread *tdq_choose(struct tdq *); 308164936Sjulianstatic void tdq_setup(struct tdq *); 309177435Sjeffstatic void tdq_load_add(struct tdq *, struct thread *); 310177435Sjeffstatic void tdq_load_rem(struct tdq *, struct thread *); 311177435Sjeffstatic __inline void tdq_runq_add(struct tdq *, struct thread *, int); 312177435Sjeffstatic __inline void tdq_runq_rem(struct tdq *, struct thread *); 313177005Sjeffstatic inline int sched_shouldpreempt(int, int, int); 314164936Sjulianvoid tdq_print(int cpu); 315165762Sjeffstatic void runq_print(struct runq *rq); 316171482Sjeffstatic void tdq_add(struct tdq *, struct thread *, int); 317110267Sjeff#ifdef SMP 318176735Sjeffstatic int tdq_move(struct tdq *, struct tdq *); 319171482Sjeffstatic int tdq_idled(struct tdq *); 320177435Sjeffstatic void tdq_notify(struct tdq *, struct thread *); 321177435Sjeffstatic struct thread *tdq_steal(struct tdq *, int); 322177435Sjeffstatic struct thread *runq_steal(struct runq *, int); 323177435Sjeffstatic int sched_pickcpu(struct thread *, int); 324172409Sjeffstatic void sched_balance(void); 325176735Sjeffstatic int sched_balance_pair(struct tdq *, struct tdq *); 326177435Sjeffstatic inline struct tdq *sched_setcpu(struct thread *, int, int); 327171482Sjeffstatic inline void thread_unblock_switch(struct thread *, struct mtx *); 328171713Sjeffstatic struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 329184439Sivorasstatic int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS); 330184439Sivorasstatic int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, 331184439Sivoras struct cpu_group *cg, int indent); 332121790Sjeff#endif 333110028Sjeff 334165762Sjeffstatic void sched_setup(void *dummy); 335177253SrwatsonSYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 336165762Sjeff 337165762Sjeffstatic void sched_initticks(void *dummy); 338177253SrwatsonSYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 339177253Srwatson NULL); 340165762Sjeff 341235459SrstoneSDT_PROVIDER_DEFINE(sched); 342235459Srstone 343260817SavgSDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *", 344235459Srstone "struct proc *", "uint8_t"); 345260817SavgSDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *", 346235459Srstone "struct proc *", "void *"); 347260817SavgSDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *", 348235459Srstone "struct proc *", "void *", "int"); 349260817SavgSDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *", 350235459Srstone "struct proc *", "uint8_t", "struct thread *"); 351260817SavgSDT_PROBE_DEFINE2(sched, , , load__change, "int", "int"); 352260817SavgSDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *", 353235459Srstone "struct proc *"); 354260817SavgSDT_PROBE_DEFINE(sched, , , on__cpu); 355260817SavgSDT_PROBE_DEFINE(sched, , , remain__cpu); 356260817SavgSDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *", 357235459Srstone "struct proc *"); 358235459Srstone 359171482Sjeff/* 360171482Sjeff * Print the threads waiting on a run-queue. 361171482Sjeff */ 362165762Sjeffstatic void 363165762Sjeffrunq_print(struct runq *rq) 364165762Sjeff{ 365165762Sjeff struct rqhead *rqh; 366177435Sjeff struct thread *td; 367165762Sjeff int pri; 368165762Sjeff int j; 369165762Sjeff int i; 370165762Sjeff 371165762Sjeff for (i = 0; i < RQB_LEN; i++) { 372165762Sjeff printf("\t\trunq bits %d 0x%zx\n", 373165762Sjeff i, rq->rq_status.rqb_bits[i]); 374165762Sjeff for (j = 0; j < RQB_BPW; j++) 375165762Sjeff if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 376165762Sjeff pri = j + (i << RQB_L2BPW); 377165762Sjeff rqh = &rq->rq_queues[pri]; 378177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) { 379165762Sjeff printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 380177435Sjeff td, td->td_name, td->td_priority, 381177435Sjeff td->td_rqindex, pri); 382165762Sjeff } 383165762Sjeff } 384165762Sjeff } 385165762Sjeff} 386165762Sjeff 387171482Sjeff/* 388171482Sjeff * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 389171482Sjeff */ 390113357Sjeffvoid 391164936Sjuliantdq_print(int cpu) 392110267Sjeff{ 393164936Sjulian struct tdq *tdq; 394112994Sjeff 395164936Sjulian tdq = TDQ_CPU(cpu); 396112994Sjeff 397171713Sjeff printf("tdq %d:\n", TDQ_ID(tdq)); 398176735Sjeff printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 399176735Sjeff printf("\tLock name: %s\n", tdq->tdq_name); 400165620Sjeff printf("\tload: %d\n", tdq->tdq_load); 401178277Sjeff printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt); 402178277Sjeff printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt); 403171482Sjeff printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 404165766Sjeff printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 405178277Sjeff printf("\tload transferable: %d\n", tdq->tdq_transferable); 406178277Sjeff printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 407165762Sjeff printf("\trealtime runq:\n"); 408165762Sjeff runq_print(&tdq->tdq_realtime); 409165762Sjeff printf("\ttimeshare runq:\n"); 410165762Sjeff runq_print(&tdq->tdq_timeshare); 411165762Sjeff printf("\tidle runq:\n"); 412165762Sjeff runq_print(&tdq->tdq_idle); 413113357Sjeff} 414112994Sjeff 415177005Sjeffstatic inline int 416177005Sjeffsched_shouldpreempt(int pri, int cpri, int remote) 417177005Sjeff{ 418177005Sjeff /* 419177005Sjeff * If the new priority is not better than the current priority there is 420177005Sjeff * nothing to do. 421177005Sjeff */ 422177005Sjeff if (pri >= cpri) 423177005Sjeff return (0); 424177005Sjeff /* 425177005Sjeff * Always preempt idle. 426177005Sjeff */ 427177005Sjeff if (cpri >= PRI_MIN_IDLE) 428177005Sjeff return (1); 429177005Sjeff /* 430177005Sjeff * If preemption is disabled don't preempt others. 431177005Sjeff */ 432177005Sjeff if (preempt_thresh == 0) 433177005Sjeff return (0); 434177005Sjeff /* 435177005Sjeff * Preempt if we exceed the threshold. 436177005Sjeff */ 437177005Sjeff if (pri <= preempt_thresh) 438177005Sjeff return (1); 439177005Sjeff /* 440217351Sjhb * If we're interactive or better and there is non-interactive 441217351Sjhb * or worse running preempt only remote processors. 442177005Sjeff */ 443217351Sjhb if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT) 444177005Sjeff return (1); 445177005Sjeff return (0); 446177005Sjeff} 447177005Sjeff 448171482Sjeff/* 449171482Sjeff * Add a thread to the actual run-queue. Keeps transferable counts up to 450171482Sjeff * date with what is actually on the run-queue. Selects the correct 451171482Sjeff * queue position for timeshare threads. 452171482Sjeff */ 453122744Sjeffstatic __inline void 454177435Sjefftdq_runq_add(struct tdq *tdq, struct thread *td, int flags) 455122744Sjeff{ 456177435Sjeff struct td_sched *ts; 457177042Sjeff u_char pri; 458177042Sjeff 459171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 460177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 461177009Sjeff 462177435Sjeff pri = td->td_priority; 463177435Sjeff ts = td->td_sched; 464177435Sjeff TD_SET_RUNQ(td); 465177435Sjeff if (THREAD_CAN_MIGRATE(td)) { 466165620Sjeff tdq->tdq_transferable++; 467164936Sjulian ts->ts_flags |= TSF_XFERABLE; 468123433Sjeff } 469217351Sjhb if (pri < PRI_MIN_BATCH) { 470177042Sjeff ts->ts_runq = &tdq->tdq_realtime; 471217351Sjhb } else if (pri <= PRI_MAX_BATCH) { 472177042Sjeff ts->ts_runq = &tdq->tdq_timeshare; 473217351Sjhb KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH, 474165762Sjeff ("Invalid priority %d on timeshare runq", pri)); 475165762Sjeff /* 476165762Sjeff * This queue contains only priorities between MIN and MAX 477165762Sjeff * realtime. Use the whole queue to represent these values. 478165762Sjeff */ 479171713Sjeff if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 480228718Savg pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE; 481165762Sjeff pri = (pri + tdq->tdq_idx) % RQ_NQS; 482165766Sjeff /* 483165766Sjeff * This effectively shortens the queue by one so we 484165766Sjeff * can have a one slot difference between idx and 485165766Sjeff * ridx while we wait for threads to drain. 486165766Sjeff */ 487165766Sjeff if (tdq->tdq_ridx != tdq->tdq_idx && 488165766Sjeff pri == tdq->tdq_ridx) 489167664Sjeff pri = (unsigned char)(pri - 1) % RQ_NQS; 490165762Sjeff } else 491165766Sjeff pri = tdq->tdq_ridx; 492177435Sjeff runq_add_pri(ts->ts_runq, td, pri, flags); 493177042Sjeff return; 494165762Sjeff } else 495177009Sjeff ts->ts_runq = &tdq->tdq_idle; 496177435Sjeff runq_add(ts->ts_runq, td, flags); 497177009Sjeff} 498177009Sjeff 499171482Sjeff/* 500171482Sjeff * Remove a thread from a run-queue. This typically happens when a thread 501171482Sjeff * is selected to run. Running threads are not on the queue and the 502171482Sjeff * transferable count does not reflect them. 503171482Sjeff */ 504122744Sjeffstatic __inline void 505177435Sjefftdq_runq_rem(struct tdq *tdq, struct thread *td) 506122744Sjeff{ 507177435Sjeff struct td_sched *ts; 508177435Sjeff 509177435Sjeff ts = td->td_sched; 510171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 511171482Sjeff KASSERT(ts->ts_runq != NULL, 512177435Sjeff ("tdq_runq_remove: thread %p null ts_runq", td)); 513164936Sjulian if (ts->ts_flags & TSF_XFERABLE) { 514165620Sjeff tdq->tdq_transferable--; 515164936Sjulian ts->ts_flags &= ~TSF_XFERABLE; 516123433Sjeff } 517165766Sjeff if (ts->ts_runq == &tdq->tdq_timeshare) { 518165766Sjeff if (tdq->tdq_idx != tdq->tdq_ridx) 519177435Sjeff runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx); 520165766Sjeff else 521177435Sjeff runq_remove_idx(ts->ts_runq, td, NULL); 522165766Sjeff } else 523177435Sjeff runq_remove(ts->ts_runq, td); 524122744Sjeff} 525122744Sjeff 526171482Sjeff/* 527171482Sjeff * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 528171482Sjeff * for this thread to the referenced thread queue. 529171482Sjeff */ 530113357Sjeffstatic void 531177435Sjefftdq_load_add(struct tdq *tdq, struct thread *td) 532113357Sjeff{ 533171482Sjeff 534171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 535177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 536177902Sjeff 537165620Sjeff tdq->tdq_load++; 538198854Sattilio if ((td->td_flags & TDF_NOLOAD) == 0) 539177902Sjeff tdq->tdq_sysload++; 540187357Sjeff KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 541260817Savg SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load); 542110267Sjeff} 543113357Sjeff 544171482Sjeff/* 545171482Sjeff * Remove the load from a thread that is transitioning to a sleep state or 546171482Sjeff * exiting. 547171482Sjeff */ 548112994Sjeffstatic void 549177435Sjefftdq_load_rem(struct tdq *tdq, struct thread *td) 550110267Sjeff{ 551171482Sjeff 552177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 553171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 554171482Sjeff KASSERT(tdq->tdq_load != 0, 555171713Sjeff ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 556177902Sjeff 557165620Sjeff tdq->tdq_load--; 558198854Sattilio if ((td->td_flags & TDF_NOLOAD) == 0) 559177902Sjeff tdq->tdq_sysload--; 560187357Sjeff KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 561260817Savg SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load); 562110267Sjeff} 563110267Sjeff 564176735Sjeff/* 565242736Sjeff * Bound timeshare latency by decreasing slice size as load increases. We 566242736Sjeff * consider the maximum latency as the sum of the threads waiting to run 567242736Sjeff * aside from curthread and target no more than sched_slice latency but 568242736Sjeff * no less than sched_slice_min runtime. 569242736Sjeff */ 570242736Sjeffstatic inline int 571242736Sjefftdq_slice(struct tdq *tdq) 572242736Sjeff{ 573242736Sjeff int load; 574242736Sjeff 575242736Sjeff /* 576242736Sjeff * It is safe to use sys_load here because this is called from 577242736Sjeff * contexts where timeshare threads are running and so there 578242736Sjeff * cannot be higher priority load in the system. 579242736Sjeff */ 580242736Sjeff load = tdq->tdq_sysload - 1; 581242736Sjeff if (load >= SCHED_SLICE_MIN_DIVISOR) 582242736Sjeff return (sched_slice_min); 583242736Sjeff if (load <= 1) 584242736Sjeff return (sched_slice); 585242736Sjeff return (sched_slice / load); 586242736Sjeff} 587242736Sjeff 588242736Sjeff/* 589176735Sjeff * Set lowpri to its exact value by searching the run-queue and 590176735Sjeff * evaluating curthread. curthread may be passed as an optimization. 591176735Sjeff */ 592176735Sjeffstatic void 593176735Sjefftdq_setlowpri(struct tdq *tdq, struct thread *ctd) 594176735Sjeff{ 595176735Sjeff struct thread *td; 596176735Sjeff 597176735Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 598176735Sjeff if (ctd == NULL) 599176735Sjeff ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 600177435Sjeff td = tdq_choose(tdq); 601177435Sjeff if (td == NULL || td->td_priority > ctd->td_priority) 602176735Sjeff tdq->tdq_lowpri = ctd->td_priority; 603176735Sjeff else 604176735Sjeff tdq->tdq_lowpri = td->td_priority; 605176735Sjeff} 606176735Sjeff 607113357Sjeff#ifdef SMP 608176735Sjeffstruct cpu_search { 609194779Sjeff cpuset_t cs_mask; 610232207Smav u_int cs_prefer; 611232207Smav int cs_pri; /* Min priority for low. */ 612232207Smav int cs_limit; /* Max load for low, min load for high. */ 613232207Smav int cs_cpu; 614232207Smav int cs_load; 615176735Sjeff}; 616176735Sjeff 617176735Sjeff#define CPU_SEARCH_LOWEST 0x1 618176735Sjeff#define CPU_SEARCH_HIGHEST 0x2 619176735Sjeff#define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 620176735Sjeff 621194779Sjeff#define CPUSET_FOREACH(cpu, mask) \ 622194779Sjeff for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \ 623222813Sattilio if (CPU_ISSET(cpu, &mask)) 624176735Sjeff 625268483Skibstatic __always_inline int cpu_search(const struct cpu_group *cg, 626268483Skib struct cpu_search *low, struct cpu_search *high, const int match); 627268483Skibint __noinline cpu_search_lowest(const struct cpu_group *cg, 628268483Skib struct cpu_search *low); 629268483Skibint __noinline cpu_search_highest(const struct cpu_group *cg, 630176735Sjeff struct cpu_search *high); 631268483Skibint __noinline cpu_search_both(const struct cpu_group *cg, 632268483Skib struct cpu_search *low, struct cpu_search *high); 633176735Sjeff 634116069Sjeff/* 635176735Sjeff * Search the tree of cpu_groups for the lowest or highest loaded cpu 636176735Sjeff * according to the match argument. This routine actually compares the 637176735Sjeff * load on all paths through the tree and finds the least loaded cpu on 638176735Sjeff * the least loaded path, which may differ from the least loaded cpu in 639176735Sjeff * the system. This balances work among caches and busses. 640116069Sjeff * 641176735Sjeff * This inline is instantiated in three forms below using constants for the 642176735Sjeff * match argument. It is reduced to the minimum set for each case. It is 643176735Sjeff * also recursive to the depth of the tree. 644116069Sjeff */ 645268483Skibstatic __always_inline int 646232207Smavcpu_search(const struct cpu_group *cg, struct cpu_search *low, 647176735Sjeff struct cpu_search *high, const int match) 648176735Sjeff{ 649232207Smav struct cpu_search lgroup; 650232207Smav struct cpu_search hgroup; 651232207Smav cpuset_t cpumask; 652232207Smav struct cpu_group *child; 653232207Smav struct tdq *tdq; 654234066Smav int cpu, i, hload, lload, load, total, rnd, *rndptr; 655176735Sjeff 656176735Sjeff total = 0; 657232207Smav cpumask = cg->cg_mask; 658232207Smav if (match & CPU_SEARCH_LOWEST) { 659232207Smav lload = INT_MAX; 660232207Smav lgroup = *low; 661232207Smav } 662232207Smav if (match & CPU_SEARCH_HIGHEST) { 663234066Smav hload = INT_MIN; 664232207Smav hgroup = *high; 665232207Smav } 666176735Sjeff 667232207Smav /* Iterate through the child CPU groups and then remaining CPUs. */ 668255363Smav for (i = cg->cg_children, cpu = mp_maxid; ; ) { 669234066Smav if (i == 0) { 670255363Smav#ifdef HAVE_INLINE_FFSL 671255363Smav cpu = CPU_FFS(&cpumask) - 1; 672255363Smav#else 673234066Smav while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask)) 674234066Smav cpu--; 675255363Smav#endif 676234066Smav if (cpu < 0) 677232207Smav break; 678232207Smav child = NULL; 679232207Smav } else 680234066Smav child = &cg->cg_child[i - 1]; 681232207Smav 682234066Smav if (match & CPU_SEARCH_LOWEST) 683234066Smav lgroup.cs_cpu = -1; 684234066Smav if (match & CPU_SEARCH_HIGHEST) 685234066Smav hgroup.cs_cpu = -1; 686232207Smav if (child) { /* Handle child CPU group. */ 687232207Smav CPU_NAND(&cpumask, &child->cg_mask); 688176735Sjeff switch (match) { 689176735Sjeff case CPU_SEARCH_LOWEST: 690176735Sjeff load = cpu_search_lowest(child, &lgroup); 691176735Sjeff break; 692176735Sjeff case CPU_SEARCH_HIGHEST: 693176735Sjeff load = cpu_search_highest(child, &hgroup); 694176735Sjeff break; 695176735Sjeff case CPU_SEARCH_BOTH: 696176735Sjeff load = cpu_search_both(child, &lgroup, &hgroup); 697176735Sjeff break; 698176735Sjeff } 699232207Smav } else { /* Handle child CPU. */ 700255363Smav CPU_CLR(cpu, &cpumask); 701232207Smav tdq = TDQ_CPU(cpu); 702232207Smav load = tdq->tdq_load * 256; 703234066Smav rndptr = DPCPU_PTR(randomval); 704234066Smav rnd = (*rndptr = *rndptr * 69069 + 5) >> 26; 705232207Smav if (match & CPU_SEARCH_LOWEST) { 706232207Smav if (cpu == low->cs_prefer) 707232207Smav load -= 64; 708232207Smav /* If that CPU is allowed and get data. */ 709234066Smav if (tdq->tdq_lowpri > lgroup.cs_pri && 710234066Smav tdq->tdq_load <= lgroup.cs_limit && 711234066Smav CPU_ISSET(cpu, &lgroup.cs_mask)) { 712232207Smav lgroup.cs_cpu = cpu; 713232207Smav lgroup.cs_load = load - rnd; 714176735Sjeff } 715232207Smav } 716232207Smav if (match & CPU_SEARCH_HIGHEST) 717234066Smav if (tdq->tdq_load >= hgroup.cs_limit && 718234066Smav tdq->tdq_transferable && 719234066Smav CPU_ISSET(cpu, &hgroup.cs_mask)) { 720232207Smav hgroup.cs_cpu = cpu; 721232207Smav hgroup.cs_load = load - rnd; 722176735Sjeff } 723176735Sjeff } 724232207Smav total += load; 725176735Sjeff 726232207Smav /* We have info about child item. Compare it. */ 727232207Smav if (match & CPU_SEARCH_LOWEST) { 728234066Smav if (lgroup.cs_cpu >= 0 && 729232454Smav (load < lload || 730232454Smav (load == lload && lgroup.cs_load < low->cs_load))) { 731232207Smav lload = load; 732232207Smav low->cs_cpu = lgroup.cs_cpu; 733232207Smav low->cs_load = lgroup.cs_load; 734232207Smav } 735232207Smav } 736232207Smav if (match & CPU_SEARCH_HIGHEST) 737234066Smav if (hgroup.cs_cpu >= 0 && 738232454Smav (load > hload || 739232454Smav (load == hload && hgroup.cs_load > high->cs_load))) { 740232207Smav hload = load; 741232207Smav high->cs_cpu = hgroup.cs_cpu; 742232207Smav high->cs_load = hgroup.cs_load; 743232207Smav } 744234066Smav if (child) { 745234066Smav i--; 746234066Smav if (i == 0 && CPU_EMPTY(&cpumask)) 747234066Smav break; 748255363Smav } 749255363Smav#ifndef HAVE_INLINE_FFSL 750255363Smav else 751234066Smav cpu--; 752255363Smav#endif 753176735Sjeff } 754176735Sjeff return (total); 755176735Sjeff} 756176735Sjeff 757176735Sjeff/* 758176735Sjeff * cpu_search instantiations must pass constants to maintain the inline 759176735Sjeff * optimization. 760176735Sjeff */ 761176735Sjeffint 762232207Smavcpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low) 763176735Sjeff{ 764176735Sjeff return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 765176735Sjeff} 766176735Sjeff 767176735Sjeffint 768232207Smavcpu_search_highest(const struct cpu_group *cg, struct cpu_search *high) 769176735Sjeff{ 770176735Sjeff return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 771176735Sjeff} 772176735Sjeff 773176735Sjeffint 774232207Smavcpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 775176735Sjeff struct cpu_search *high) 776176735Sjeff{ 777176735Sjeff return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 778176735Sjeff} 779176735Sjeff 780176735Sjeff/* 781176735Sjeff * Find the cpu with the least load via the least loaded path that has a 782176735Sjeff * lowpri greater than pri pri. A pri of -1 indicates any priority is 783176735Sjeff * acceptable. 784176735Sjeff */ 785176735Sjeffstatic inline int 786232207Smavsched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload, 787232207Smav int prefer) 788176735Sjeff{ 789176735Sjeff struct cpu_search low; 790176735Sjeff 791176735Sjeff low.cs_cpu = -1; 792232207Smav low.cs_prefer = prefer; 793176735Sjeff low.cs_mask = mask; 794232207Smav low.cs_pri = pri; 795232207Smav low.cs_limit = maxload; 796176735Sjeff cpu_search_lowest(cg, &low); 797176735Sjeff return low.cs_cpu; 798176735Sjeff} 799176735Sjeff 800176735Sjeff/* 801176735Sjeff * Find the cpu with the highest load via the highest loaded path. 802176735Sjeff */ 803176735Sjeffstatic inline int 804232207Smavsched_highest(const struct cpu_group *cg, cpuset_t mask, int minload) 805176735Sjeff{ 806176735Sjeff struct cpu_search high; 807176735Sjeff 808176735Sjeff high.cs_cpu = -1; 809176735Sjeff high.cs_mask = mask; 810176735Sjeff high.cs_limit = minload; 811176735Sjeff cpu_search_highest(cg, &high); 812176735Sjeff return high.cs_cpu; 813176735Sjeff} 814176735Sjeff 815121790Sjeffstatic void 816176735Sjeffsched_balance_group(struct cpu_group *cg) 817116069Sjeff{ 818232207Smav cpuset_t hmask, lmask; 819232207Smav int high, low, anylow; 820123487Sjeff 821232207Smav CPU_FILL(&hmask); 822176735Sjeff for (;;) { 823232207Smav high = sched_highest(cg, hmask, 1); 824232207Smav /* Stop if there is no more CPU with transferrable threads. */ 825232207Smav if (high == -1) 826176735Sjeff break; 827232207Smav CPU_CLR(high, &hmask); 828232207Smav CPU_COPY(&hmask, &lmask); 829232207Smav /* Stop if there is no more CPU left for low. */ 830232207Smav if (CPU_EMPTY(&lmask)) 831176735Sjeff break; 832232207Smav anylow = 1; 833232207Smavnextlow: 834232207Smav low = sched_lowest(cg, lmask, -1, 835232207Smav TDQ_CPU(high)->tdq_load - 1, high); 836232207Smav /* Stop if we looked well and found no less loaded CPU. */ 837232207Smav if (anylow && low == -1) 838232207Smav break; 839232207Smav /* Go to next high if we found no less loaded CPU. */ 840232207Smav if (low == -1) 841232207Smav continue; 842232207Smav /* Transfer thread from high to low. */ 843232207Smav if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) { 844232207Smav /* CPU that got thread can no longer be a donor. */ 845232207Smav CPU_CLR(low, &hmask); 846232207Smav } else { 847232207Smav /* 848232207Smav * If failed, then there is no threads on high 849232207Smav * that can run on this low. Drop low from low 850232207Smav * mask and look for different one. 851232207Smav */ 852232207Smav CPU_CLR(low, &lmask); 853232207Smav anylow = 0; 854232207Smav goto nextlow; 855232207Smav } 856123487Sjeff } 857123487Sjeff} 858123487Sjeff 859123487Sjeffstatic void 860201148Sedsched_balance(void) 861123487Sjeff{ 862172409Sjeff struct tdq *tdq; 863123487Sjeff 864172409Sjeff /* 865172409Sjeff * Select a random time between .5 * balance_interval and 866172409Sjeff * 1.5 * balance_interval. 867172409Sjeff */ 868176735Sjeff balance_ticks = max(balance_interval / 2, 1); 869176735Sjeff balance_ticks += random() % balance_interval; 870171482Sjeff if (smp_started == 0 || rebalance == 0) 871171482Sjeff return; 872172409Sjeff tdq = TDQ_SELF(); 873172409Sjeff TDQ_UNLOCK(tdq); 874176735Sjeff sched_balance_group(cpu_top); 875172409Sjeff TDQ_LOCK(tdq); 876123487Sjeff} 877123487Sjeff 878171482Sjeff/* 879171482Sjeff * Lock two thread queues using their address to maintain lock order. 880171482Sjeff */ 881123487Sjeffstatic void 882171482Sjefftdq_lock_pair(struct tdq *one, struct tdq *two) 883171482Sjeff{ 884171482Sjeff if (one < two) { 885171482Sjeff TDQ_LOCK(one); 886171482Sjeff TDQ_LOCK_FLAGS(two, MTX_DUPOK); 887171482Sjeff } else { 888171482Sjeff TDQ_LOCK(two); 889171482Sjeff TDQ_LOCK_FLAGS(one, MTX_DUPOK); 890171482Sjeff } 891171482Sjeff} 892171482Sjeff 893171482Sjeff/* 894172409Sjeff * Unlock two thread queues. Order is not important here. 895172409Sjeff */ 896172409Sjeffstatic void 897172409Sjefftdq_unlock_pair(struct tdq *one, struct tdq *two) 898172409Sjeff{ 899172409Sjeff TDQ_UNLOCK(one); 900172409Sjeff TDQ_UNLOCK(two); 901172409Sjeff} 902172409Sjeff 903172409Sjeff/* 904171482Sjeff * Transfer load between two imbalanced thread queues. 905171482Sjeff */ 906176735Sjeffstatic int 907164936Sjuliansched_balance_pair(struct tdq *high, struct tdq *low) 908123487Sjeff{ 909176735Sjeff int moved; 910226057Smarius int cpu; 911116069Sjeff 912171482Sjeff tdq_lock_pair(high, low); 913176735Sjeff moved = 0; 914116069Sjeff /* 915122744Sjeff * Determine what the imbalance is and then adjust that to how many 916165620Sjeff * threads we actually have to give up (transferable). 917122744Sjeff */ 918232207Smav if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load && 919232207Smav (moved = tdq_move(high, low)) > 0) { 920172293Sjeff /* 921226057Smarius * In case the target isn't the current cpu IPI it to force a 922226057Smarius * reschedule with the new workload. 923172293Sjeff */ 924226057Smarius cpu = TDQ_ID(low); 925226057Smarius if (cpu != PCPU_GET(cpuid)) 926226057Smarius ipi_cpu(cpu, IPI_PREEMPT); 927171482Sjeff } 928172409Sjeff tdq_unlock_pair(high, low); 929176735Sjeff return (moved); 930116069Sjeff} 931116069Sjeff 932171482Sjeff/* 933171482Sjeff * Move a thread from one thread queue to another. 934171482Sjeff */ 935176735Sjeffstatic int 936171482Sjefftdq_move(struct tdq *from, struct tdq *to) 937116069Sjeff{ 938171482Sjeff struct td_sched *ts; 939171482Sjeff struct thread *td; 940164936Sjulian struct tdq *tdq; 941171482Sjeff int cpu; 942116069Sjeff 943172409Sjeff TDQ_LOCK_ASSERT(from, MA_OWNED); 944172409Sjeff TDQ_LOCK_ASSERT(to, MA_OWNED); 945172409Sjeff 946164936Sjulian tdq = from; 947171482Sjeff cpu = TDQ_ID(to); 948177435Sjeff td = tdq_steal(tdq, cpu); 949177435Sjeff if (td == NULL) 950176735Sjeff return (0); 951177435Sjeff ts = td->td_sched; 952171482Sjeff /* 953171482Sjeff * Although the run queue is locked the thread may be blocked. Lock 954172409Sjeff * it to clear this and acquire the run-queue lock. 955171482Sjeff */ 956171482Sjeff thread_lock(td); 957172409Sjeff /* Drop recursive lock on from acquired via thread_lock(). */ 958171482Sjeff TDQ_UNLOCK(from); 959171482Sjeff sched_rem(td); 960166108Sjeff ts->ts_cpu = cpu; 961171482Sjeff td->td_lock = TDQ_LOCKPTR(to); 962171482Sjeff tdq_add(to, td, SRQ_YIELDING); 963176735Sjeff return (1); 964116069Sjeff} 965110267Sjeff 966171482Sjeff/* 967171482Sjeff * This tdq has idled. Try to steal a thread from another cpu and switch 968171482Sjeff * to it. 969171482Sjeff */ 970123433Sjeffstatic int 971164936Sjuliantdq_idled(struct tdq *tdq) 972121790Sjeff{ 973176735Sjeff struct cpu_group *cg; 974164936Sjulian struct tdq *steal; 975194779Sjeff cpuset_t mask; 976176735Sjeff int thresh; 977171482Sjeff int cpu; 978123433Sjeff 979172484Sjeff if (smp_started == 0 || steal_idle == 0) 980172484Sjeff return (1); 981194779Sjeff CPU_FILL(&mask); 982194779Sjeff CPU_CLR(PCPU_GET(cpuid), &mask); 983176735Sjeff /* We don't want to be preempted while we're iterating. */ 984171482Sjeff spinlock_enter(); 985176735Sjeff for (cg = tdq->tdq_cg; cg != NULL; ) { 986191643Sjeff if ((cg->cg_flags & CG_FLAG_THREAD) == 0) 987176735Sjeff thresh = steal_thresh; 988176735Sjeff else 989176735Sjeff thresh = 1; 990176735Sjeff cpu = sched_highest(cg, mask, thresh); 991176735Sjeff if (cpu == -1) { 992176735Sjeff cg = cg->cg_parent; 993176735Sjeff continue; 994166108Sjeff } 995176735Sjeff steal = TDQ_CPU(cpu); 996194779Sjeff CPU_CLR(cpu, &mask); 997176735Sjeff tdq_lock_pair(tdq, steal); 998176735Sjeff if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 999176735Sjeff tdq_unlock_pair(tdq, steal); 1000176735Sjeff continue; 1001171482Sjeff } 1002176735Sjeff /* 1003176735Sjeff * If a thread was added while interrupts were disabled don't 1004176735Sjeff * steal one here. If we fail to acquire one due to affinity 1005176735Sjeff * restrictions loop again with this cpu removed from the 1006176735Sjeff * set. 1007176735Sjeff */ 1008176735Sjeff if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 1009176735Sjeff tdq_unlock_pair(tdq, steal); 1010176735Sjeff continue; 1011176735Sjeff } 1012176735Sjeff spinlock_exit(); 1013176735Sjeff TDQ_UNLOCK(steal); 1014178272Sjeff mi_switch(SW_VOL | SWT_IDLE, NULL); 1015176735Sjeff thread_unlock(curthread); 1016176735Sjeff 1017176735Sjeff return (0); 1018123433Sjeff } 1019171482Sjeff spinlock_exit(); 1020123433Sjeff return (1); 1021121790Sjeff} 1022121790Sjeff 1023171482Sjeff/* 1024171482Sjeff * Notify a remote cpu of new work. Sends an IPI if criteria are met. 1025171482Sjeff */ 1026121790Sjeffstatic void 1027177435Sjefftdq_notify(struct tdq *tdq, struct thread *td) 1028121790Sjeff{ 1029185047Sjhb struct thread *ctd; 1030166247Sjeff int pri; 1031166108Sjeff int cpu; 1032121790Sjeff 1033177005Sjeff if (tdq->tdq_ipipending) 1034177005Sjeff return; 1035177435Sjeff cpu = td->td_sched->ts_cpu; 1036177435Sjeff pri = td->td_priority; 1037185047Sjhb ctd = pcpu_find(cpu)->pc_curthread; 1038185047Sjhb if (!sched_shouldpreempt(pri, ctd->td_priority, 1)) 1039166137Sjeff return; 1040271707Smav 1041271707Smav /* 1042271707Smav * Make sure that tdq_load updated before calling this function 1043271707Smav * is globally visible before we read tdq_cpu_idle. Idle thread 1044271707Smav * accesses both of them without locks, and the order is important. 1045271707Smav */ 1046271707Smav mb(); 1047271707Smav 1048185047Sjhb if (TD_IS_IDLETHREAD(ctd)) { 1049178277Sjeff /* 1050178471Sjeff * If the MD code has an idle wakeup routine try that before 1051178471Sjeff * falling back to IPI. 1052178471Sjeff */ 1053212416Smav if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu)) 1054178471Sjeff return; 1055178277Sjeff } 1056177005Sjeff tdq->tdq_ipipending = 1; 1057210939Sjhb ipi_cpu(cpu, IPI_PREEMPT); 1058121790Sjeff} 1059121790Sjeff 1060171482Sjeff/* 1061171482Sjeff * Steals load from a timeshare queue. Honors the rotating queue head 1062171482Sjeff * index. 1063171482Sjeff */ 1064177435Sjeffstatic struct thread * 1065176735Sjeffrunq_steal_from(struct runq *rq, int cpu, u_char start) 1066171482Sjeff{ 1067171482Sjeff struct rqbits *rqb; 1068171482Sjeff struct rqhead *rqh; 1069232207Smav struct thread *td, *first; 1070171482Sjeff int bit; 1071171482Sjeff int pri; 1072171482Sjeff int i; 1073171482Sjeff 1074171482Sjeff rqb = &rq->rq_status; 1075171482Sjeff bit = start & (RQB_BPW -1); 1076171482Sjeff pri = 0; 1077232207Smav first = NULL; 1078171482Sjeffagain: 1079171482Sjeff for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 1080171482Sjeff if (rqb->rqb_bits[i] == 0) 1081171482Sjeff continue; 1082171482Sjeff if (bit != 0) { 1083171482Sjeff for (pri = bit; pri < RQB_BPW; pri++) 1084171482Sjeff if (rqb->rqb_bits[i] & (1ul << pri)) 1085171482Sjeff break; 1086171482Sjeff if (pri >= RQB_BPW) 1087171482Sjeff continue; 1088171482Sjeff } else 1089171482Sjeff pri = RQB_FFS(rqb->rqb_bits[i]); 1090171482Sjeff pri += (i << RQB_L2BPW); 1091171482Sjeff rqh = &rq->rq_queues[pri]; 1092177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) { 1093177435Sjeff if (first && THREAD_CAN_MIGRATE(td) && 1094177435Sjeff THREAD_CAN_SCHED(td, cpu)) 1095177435Sjeff return (td); 1096232207Smav first = td; 1097171482Sjeff } 1098171482Sjeff } 1099171482Sjeff if (start != 0) { 1100171482Sjeff start = 0; 1101171482Sjeff goto again; 1102171482Sjeff } 1103171482Sjeff 1104232207Smav if (first && THREAD_CAN_MIGRATE(first) && 1105232207Smav THREAD_CAN_SCHED(first, cpu)) 1106232207Smav return (first); 1107171482Sjeff return (NULL); 1108171482Sjeff} 1109171482Sjeff 1110171482Sjeff/* 1111171482Sjeff * Steals load from a standard linear queue. 1112171482Sjeff */ 1113177435Sjeffstatic struct thread * 1114176735Sjeffrunq_steal(struct runq *rq, int cpu) 1115121790Sjeff{ 1116121790Sjeff struct rqhead *rqh; 1117121790Sjeff struct rqbits *rqb; 1118177435Sjeff struct thread *td; 1119121790Sjeff int word; 1120121790Sjeff int bit; 1121121790Sjeff 1122121790Sjeff rqb = &rq->rq_status; 1123121790Sjeff for (word = 0; word < RQB_LEN; word++) { 1124121790Sjeff if (rqb->rqb_bits[word] == 0) 1125121790Sjeff continue; 1126121790Sjeff for (bit = 0; bit < RQB_BPW; bit++) { 1127123231Speter if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1128121790Sjeff continue; 1129121790Sjeff rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1130177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) 1131177435Sjeff if (THREAD_CAN_MIGRATE(td) && 1132177435Sjeff THREAD_CAN_SCHED(td, cpu)) 1133177435Sjeff return (td); 1134121790Sjeff } 1135121790Sjeff } 1136121790Sjeff return (NULL); 1137121790Sjeff} 1138121790Sjeff 1139171482Sjeff/* 1140171482Sjeff * Attempt to steal a thread in priority order from a thread queue. 1141171482Sjeff */ 1142177435Sjeffstatic struct thread * 1143176735Sjefftdq_steal(struct tdq *tdq, int cpu) 1144121790Sjeff{ 1145177435Sjeff struct thread *td; 1146121790Sjeff 1147171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1148177435Sjeff if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1149177435Sjeff return (td); 1150177435Sjeff if ((td = runq_steal_from(&tdq->tdq_timeshare, 1151177435Sjeff cpu, tdq->tdq_ridx)) != NULL) 1152177435Sjeff return (td); 1153176735Sjeff return (runq_steal(&tdq->tdq_idle, cpu)); 1154121790Sjeff} 1155123433Sjeff 1156171482Sjeff/* 1157171482Sjeff * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1158172409Sjeff * current lock and returns with the assigned queue locked. 1159171482Sjeff */ 1160171482Sjeffstatic inline struct tdq * 1161177435Sjeffsched_setcpu(struct thread *td, int cpu, int flags) 1162123433Sjeff{ 1163177435Sjeff 1164171482Sjeff struct tdq *tdq; 1165123433Sjeff 1166177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1167171482Sjeff tdq = TDQ_CPU(cpu); 1168177435Sjeff td->td_sched->ts_cpu = cpu; 1169177435Sjeff /* 1170177435Sjeff * If the lock matches just return the queue. 1171177435Sjeff */ 1172171482Sjeff if (td->td_lock == TDQ_LOCKPTR(tdq)) 1173171482Sjeff return (tdq); 1174171482Sjeff#ifdef notyet 1175123433Sjeff /* 1176172293Sjeff * If the thread isn't running its lockptr is a 1177171482Sjeff * turnstile or a sleepqueue. We can just lock_set without 1178171482Sjeff * blocking. 1179123685Sjeff */ 1180171482Sjeff if (TD_CAN_RUN(td)) { 1181171482Sjeff TDQ_LOCK(tdq); 1182171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1183171482Sjeff return (tdq); 1184171482Sjeff } 1185171482Sjeff#endif 1186166108Sjeff /* 1187171482Sjeff * The hard case, migration, we need to block the thread first to 1188171482Sjeff * prevent order reversals with other cpus locks. 1189166108Sjeff */ 1190202889Sattilio spinlock_enter(); 1191171482Sjeff thread_lock_block(td); 1192171482Sjeff TDQ_LOCK(tdq); 1193171713Sjeff thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1194202889Sattilio spinlock_exit(); 1195171482Sjeff return (tdq); 1196166108Sjeff} 1197166108Sjeff 1198178272SjeffSCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1199178272SjeffSCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1200178272SjeffSCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1201178272SjeffSCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1202178272SjeffSCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1203178272SjeffSCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1204178272Sjeff 1205166108Sjeffstatic int 1206177435Sjeffsched_pickcpu(struct thread *td, int flags) 1207171482Sjeff{ 1208232207Smav struct cpu_group *cg, *ccg; 1209177435Sjeff struct td_sched *ts; 1210171482Sjeff struct tdq *tdq; 1211194779Sjeff cpuset_t mask; 1212232207Smav int cpu, pri, self; 1213166108Sjeff 1214176735Sjeff self = PCPU_GET(cpuid); 1215177435Sjeff ts = td->td_sched; 1216166108Sjeff if (smp_started == 0) 1217166108Sjeff return (self); 1218171506Sjeff /* 1219171506Sjeff * Don't migrate a running thread from sched_switch(). 1220171506Sjeff */ 1221176735Sjeff if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1222176735Sjeff return (ts->ts_cpu); 1223166108Sjeff /* 1224176735Sjeff * Prefer to run interrupt threads on the processors that generate 1225176735Sjeff * the interrupt. 1226166108Sjeff */ 1227232207Smav pri = td->td_priority; 1228176735Sjeff if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1229178272Sjeff curthread->td_intr_nesting_level && ts->ts_cpu != self) { 1230178272Sjeff SCHED_STAT_INC(pickcpu_intrbind); 1231176735Sjeff ts->ts_cpu = self; 1232232207Smav if (TDQ_CPU(self)->tdq_lowpri > pri) { 1233232207Smav SCHED_STAT_INC(pickcpu_affinity); 1234232207Smav return (ts->ts_cpu); 1235232207Smav } 1236178272Sjeff } 1237166108Sjeff /* 1238176735Sjeff * If the thread can run on the last cpu and the affinity has not 1239176735Sjeff * expired or it is idle run it there. 1240166108Sjeff */ 1241176735Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1242232207Smav cg = tdq->tdq_cg; 1243232207Smav if (THREAD_CAN_SCHED(td, ts->ts_cpu) && 1244232207Smav tdq->tdq_lowpri >= PRI_MIN_IDLE && 1245232207Smav SCHED_AFFINITY(ts, CG_SHARE_L2)) { 1246232207Smav if (cg->cg_flags & CG_FLAG_THREAD) { 1247232207Smav CPUSET_FOREACH(cpu, cg->cg_mask) { 1248232207Smav if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) 1249232207Smav break; 1250232207Smav } 1251232207Smav } else 1252232207Smav cpu = INT_MAX; 1253232207Smav if (cpu > mp_maxid) { 1254178272Sjeff SCHED_STAT_INC(pickcpu_idle_affinity); 1255176735Sjeff return (ts->ts_cpu); 1256178272Sjeff } 1257139334Sjeff } 1258123433Sjeff /* 1259232207Smav * Search for the last level cache CPU group in the tree. 1260232207Smav * Skip caches with expired affinity time and SMT groups. 1261232207Smav * Affinity to higher level caches will be handled less aggressively. 1262123433Sjeff */ 1263232207Smav for (ccg = NULL; cg != NULL; cg = cg->cg_parent) { 1264232207Smav if (cg->cg_flags & CG_FLAG_THREAD) 1265232207Smav continue; 1266232207Smav if (!SCHED_AFFINITY(ts, cg->cg_level)) 1267232207Smav continue; 1268232207Smav ccg = cg; 1269232207Smav } 1270232207Smav if (ccg != NULL) 1271232207Smav cg = ccg; 1272176735Sjeff cpu = -1; 1273232207Smav /* Search the group for the less loaded idle CPU we can run now. */ 1274194779Sjeff mask = td->td_cpuset->cs_mask; 1275232207Smav if (cg != NULL && cg != cpu_top && 1276232207Smav CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0) 1277232207Smav cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE), 1278232207Smav INT_MAX, ts->ts_cpu); 1279232207Smav /* Search globally for the less loaded CPU we can run now. */ 1280176735Sjeff if (cpu == -1) 1281232207Smav cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu); 1282232207Smav /* Search globally for the less loaded CPU. */ 1283232207Smav if (cpu == -1) 1284232207Smav cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu); 1285232454Smav KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1286171506Sjeff /* 1287176735Sjeff * Compare the lowest loaded cpu to current cpu. 1288171506Sjeff */ 1289177005Sjeff if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1290232207Smav TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE && 1291232207Smav TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) { 1292178272Sjeff SCHED_STAT_INC(pickcpu_local); 1293177005Sjeff cpu = self; 1294178272Sjeff } else 1295178272Sjeff SCHED_STAT_INC(pickcpu_lowest); 1296178272Sjeff if (cpu != ts->ts_cpu) 1297178272Sjeff SCHED_STAT_INC(pickcpu_migration); 1298171482Sjeff return (cpu); 1299123433Sjeff} 1300176735Sjeff#endif 1301123433Sjeff 1302117326Sjeff/* 1303121790Sjeff * Pick the highest priority task we have and return it. 1304117326Sjeff */ 1305177435Sjeffstatic struct thread * 1306164936Sjuliantdq_choose(struct tdq *tdq) 1307110267Sjeff{ 1308177435Sjeff struct thread *td; 1309110267Sjeff 1310171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1311177435Sjeff td = runq_choose(&tdq->tdq_realtime); 1312177435Sjeff if (td != NULL) 1313177435Sjeff return (td); 1314177435Sjeff td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1315177435Sjeff if (td != NULL) { 1316217351Sjhb KASSERT(td->td_priority >= PRI_MIN_BATCH, 1317165762Sjeff ("tdq_choose: Invalid priority on timeshare queue %d", 1318177435Sjeff td->td_priority)); 1319177435Sjeff return (td); 1320165762Sjeff } 1321177435Sjeff td = runq_choose(&tdq->tdq_idle); 1322177435Sjeff if (td != NULL) { 1323177435Sjeff KASSERT(td->td_priority >= PRI_MIN_IDLE, 1324165762Sjeff ("tdq_choose: Invalid priority on idle queue %d", 1325177435Sjeff td->td_priority)); 1326177435Sjeff return (td); 1327165762Sjeff } 1328165762Sjeff 1329165762Sjeff return (NULL); 1330110267Sjeff} 1331110267Sjeff 1332171482Sjeff/* 1333171482Sjeff * Initialize a thread queue. 1334171482Sjeff */ 1335109864Sjeffstatic void 1336164936Sjuliantdq_setup(struct tdq *tdq) 1337110028Sjeff{ 1338171482Sjeff 1339171713Sjeff if (bootverbose) 1340171713Sjeff printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1341165762Sjeff runq_init(&tdq->tdq_realtime); 1342165762Sjeff runq_init(&tdq->tdq_timeshare); 1343165620Sjeff runq_init(&tdq->tdq_idle); 1344176735Sjeff snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1345176735Sjeff "sched lock %d", (int)TDQ_ID(tdq)); 1346176735Sjeff mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1347176735Sjeff MTX_SPIN | MTX_RECURSE); 1348187357Sjeff#ifdef KTR 1349187357Sjeff snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname), 1350187357Sjeff "CPU %d load", (int)TDQ_ID(tdq)); 1351187357Sjeff#endif 1352110028Sjeff} 1353110028Sjeff 1354171713Sjeff#ifdef SMP 1355110028Sjeffstatic void 1356171713Sjeffsched_setup_smp(void) 1357171713Sjeff{ 1358171713Sjeff struct tdq *tdq; 1359171713Sjeff int i; 1360171713Sjeff 1361176735Sjeff cpu_top = smp_topo(); 1362209059Sjhb CPU_FOREACH(i) { 1363176735Sjeff tdq = TDQ_CPU(i); 1364171713Sjeff tdq_setup(tdq); 1365176735Sjeff tdq->tdq_cg = smp_topo_find(cpu_top, i); 1366176735Sjeff if (tdq->tdq_cg == NULL) 1367176735Sjeff panic("Can't find cpu group for %d\n", i); 1368123433Sjeff } 1369176735Sjeff balance_tdq = TDQ_SELF(); 1370176735Sjeff sched_balance(); 1371171713Sjeff} 1372171713Sjeff#endif 1373171713Sjeff 1374171713Sjeff/* 1375171713Sjeff * Setup the thread queues and initialize the topology based on MD 1376171713Sjeff * information. 1377171713Sjeff */ 1378171713Sjeffstatic void 1379171713Sjeffsched_setup(void *dummy) 1380171713Sjeff{ 1381171713Sjeff struct tdq *tdq; 1382171713Sjeff 1383171713Sjeff tdq = TDQ_SELF(); 1384171713Sjeff#ifdef SMP 1385176734Sjeff sched_setup_smp(); 1386117237Sjeff#else 1387171713Sjeff tdq_setup(tdq); 1388116069Sjeff#endif 1389171482Sjeff 1390171482Sjeff /* Add thread0's load since it's running. */ 1391171482Sjeff TDQ_LOCK(tdq); 1392171713Sjeff thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1393177435Sjeff tdq_load_add(tdq, &thread0); 1394176735Sjeff tdq->tdq_lowpri = thread0.td_priority; 1395171482Sjeff TDQ_UNLOCK(tdq); 1396109864Sjeff} 1397109864Sjeff 1398171482Sjeff/* 1399239185Smav * This routine determines time constants after stathz and hz are setup. 1400171482Sjeff */ 1401153533Sdavidxu/* ARGSUSED */ 1402153533Sdavidxustatic void 1403153533Sdavidxusched_initticks(void *dummy) 1404153533Sdavidxu{ 1405171482Sjeff int incr; 1406171482Sjeff 1407153533Sdavidxu realstathz = stathz ? stathz : hz; 1408242736Sjeff sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR; 1409242736Sjeff sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR; 1410239196Smav hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 1411239196Smav realstathz); 1412153533Sdavidxu 1413153533Sdavidxu /* 1414165762Sjeff * tickincr is shifted out by 10 to avoid rounding errors due to 1415165766Sjeff * hz not being evenly divisible by stathz on all platforms. 1416153533Sdavidxu */ 1417171482Sjeff incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1418165762Sjeff /* 1419165762Sjeff * This does not work for values of stathz that are more than 1420165762Sjeff * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1421165762Sjeff */ 1422171482Sjeff if (incr == 0) 1423171482Sjeff incr = 1; 1424171482Sjeff tickincr = incr; 1425166108Sjeff#ifdef SMP 1426171899Sjeff /* 1427172409Sjeff * Set the default balance interval now that we know 1428172409Sjeff * what realstathz is. 1429172409Sjeff */ 1430172409Sjeff balance_interval = realstathz; 1431166108Sjeff affinity = SCHED_AFFINITY_DEFAULT; 1432166108Sjeff#endif 1433232740Smav if (sched_idlespinthresh < 0) 1434242852Smav sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz; 1435153533Sdavidxu} 1436153533Sdavidxu 1437153533Sdavidxu 1438109864Sjeff/* 1439171482Sjeff * This is the core of the interactivity algorithm. Determines a score based 1440171482Sjeff * on past behavior. It is the ratio of sleep time to run time scaled to 1441171482Sjeff * a [0, 100] integer. This is the voluntary sleep time of a process, which 1442171482Sjeff * differs from the cpu usage because it does not account for time spent 1443171482Sjeff * waiting on a run-queue. Would be prettier if we had floating point. 1444171482Sjeff */ 1445171482Sjeffstatic int 1446171482Sjeffsched_interact_score(struct thread *td) 1447171482Sjeff{ 1448171482Sjeff struct td_sched *ts; 1449171482Sjeff int div; 1450171482Sjeff 1451171482Sjeff ts = td->td_sched; 1452171482Sjeff /* 1453171482Sjeff * The score is only needed if this is likely to be an interactive 1454171482Sjeff * task. Don't go through the expense of computing it if there's 1455171482Sjeff * no chance. 1456171482Sjeff */ 1457171482Sjeff if (sched_interact <= SCHED_INTERACT_HALF && 1458171482Sjeff ts->ts_runtime >= ts->ts_slptime) 1459171482Sjeff return (SCHED_INTERACT_HALF); 1460171482Sjeff 1461171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1462171482Sjeff div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1463171482Sjeff return (SCHED_INTERACT_HALF + 1464171482Sjeff (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1465171482Sjeff } 1466171482Sjeff if (ts->ts_slptime > ts->ts_runtime) { 1467171482Sjeff div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1468171482Sjeff return (ts->ts_runtime / div); 1469171482Sjeff } 1470171482Sjeff /* runtime == slptime */ 1471171482Sjeff if (ts->ts_runtime) 1472171482Sjeff return (SCHED_INTERACT_HALF); 1473171482Sjeff 1474171482Sjeff /* 1475171482Sjeff * This can happen if slptime and runtime are 0. 1476171482Sjeff */ 1477171482Sjeff return (0); 1478171482Sjeff 1479171482Sjeff} 1480171482Sjeff 1481171482Sjeff/* 1482109864Sjeff * Scale the scheduling priority according to the "interactivity" of this 1483109864Sjeff * process. 1484109864Sjeff */ 1485113357Sjeffstatic void 1486163709Sjbsched_priority(struct thread *td) 1487109864Sjeff{ 1488165762Sjeff int score; 1489109864Sjeff int pri; 1490109864Sjeff 1491217291Sjhb if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1492113357Sjeff return; 1493112966Sjeff /* 1494165762Sjeff * If the score is interactive we place the thread in the realtime 1495165762Sjeff * queue with a priority that is less than kernel and interrupt 1496165762Sjeff * priorities. These threads are not subject to nice restrictions. 1497112966Sjeff * 1498171482Sjeff * Scores greater than this are placed on the normal timeshare queue 1499165762Sjeff * where the priority is partially decided by the most recent cpu 1500165762Sjeff * utilization and the rest is decided by nice value. 1501172293Sjeff * 1502172293Sjeff * The nice value of the process has a linear effect on the calculated 1503172293Sjeff * score. Negative nice values make it easier for a thread to be 1504172293Sjeff * considered interactive. 1505112966Sjeff */ 1506198126Sjhb score = imax(0, sched_interact_score(td) + td->td_proc->p_nice); 1507165762Sjeff if (score < sched_interact) { 1508217351Sjhb pri = PRI_MIN_INTERACT; 1509217351Sjhb pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) / 1510217237Sjhb sched_interact) * score; 1511217351Sjhb KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT, 1512166208Sjeff ("sched_priority: invalid interactive priority %d score %d", 1513166208Sjeff pri, score)); 1514165762Sjeff } else { 1515165762Sjeff pri = SCHED_PRI_MIN; 1516165762Sjeff if (td->td_sched->ts_ticks) 1517228960Sjhb pri += min(SCHED_PRI_TICKS(td->td_sched), 1518259834Sjhb SCHED_PRI_RANGE - 1); 1519165762Sjeff pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1520217351Sjhb KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH, 1521171482Sjeff ("sched_priority: invalid priority %d: nice %d, " 1522171482Sjeff "ticks %d ftick %d ltick %d tick pri %d", 1523171482Sjeff pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1524171482Sjeff td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1525171482Sjeff SCHED_PRI_TICKS(td->td_sched))); 1526165762Sjeff } 1527165762Sjeff sched_user_prio(td, pri); 1528112966Sjeff 1529112966Sjeff return; 1530109864Sjeff} 1531109864Sjeff 1532121868Sjeff/* 1533121868Sjeff * This routine enforces a maximum limit on the amount of scheduling history 1534171482Sjeff * kept. It is called after either the slptime or runtime is adjusted. This 1535171482Sjeff * function is ugly due to integer math. 1536121868Sjeff */ 1537116463Sjeffstatic void 1538163709Sjbsched_interact_update(struct thread *td) 1539116463Sjeff{ 1540165819Sjeff struct td_sched *ts; 1541166208Sjeff u_int sum; 1542121605Sjeff 1543165819Sjeff ts = td->td_sched; 1544171482Sjeff sum = ts->ts_runtime + ts->ts_slptime; 1545121868Sjeff if (sum < SCHED_SLP_RUN_MAX) 1546121868Sjeff return; 1547121868Sjeff /* 1548165819Sjeff * This only happens from two places: 1549165819Sjeff * 1) We have added an unusual amount of run time from fork_exit. 1550165819Sjeff * 2) We have added an unusual amount of sleep time from sched_sleep(). 1551165819Sjeff */ 1552165819Sjeff if (sum > SCHED_SLP_RUN_MAX * 2) { 1553171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1554171482Sjeff ts->ts_runtime = SCHED_SLP_RUN_MAX; 1555171482Sjeff ts->ts_slptime = 1; 1556165819Sjeff } else { 1557171482Sjeff ts->ts_slptime = SCHED_SLP_RUN_MAX; 1558171482Sjeff ts->ts_runtime = 1; 1559165819Sjeff } 1560165819Sjeff return; 1561165819Sjeff } 1562165819Sjeff /* 1563121868Sjeff * If we have exceeded by more than 1/5th then the algorithm below 1564121868Sjeff * will not bring us back into range. Dividing by two here forces 1565133427Sjeff * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1566121868Sjeff */ 1567127850Sjeff if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1568171482Sjeff ts->ts_runtime /= 2; 1569171482Sjeff ts->ts_slptime /= 2; 1570121868Sjeff return; 1571116463Sjeff } 1572171482Sjeff ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1573171482Sjeff ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1574116463Sjeff} 1575116463Sjeff 1576171482Sjeff/* 1577171482Sjeff * Scale back the interactivity history when a child thread is created. The 1578171482Sjeff * history is inherited from the parent but the thread may behave totally 1579171482Sjeff * differently. For example, a shell spawning a compiler process. We want 1580171482Sjeff * to learn that the compiler is behaving badly very quickly. 1581171482Sjeff */ 1582121868Sjeffstatic void 1583163709Sjbsched_interact_fork(struct thread *td) 1584121868Sjeff{ 1585121868Sjeff int ratio; 1586121868Sjeff int sum; 1587121868Sjeff 1588171482Sjeff sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1589121868Sjeff if (sum > SCHED_SLP_RUN_FORK) { 1590121868Sjeff ratio = sum / SCHED_SLP_RUN_FORK; 1591171482Sjeff td->td_sched->ts_runtime /= ratio; 1592171482Sjeff td->td_sched->ts_slptime /= ratio; 1593121868Sjeff } 1594121868Sjeff} 1595121868Sjeff 1596113357Sjeff/* 1597171482Sjeff * Called from proc0_init() to setup the scheduler fields. 1598134791Sjulian */ 1599134791Sjulianvoid 1600134791Sjulianschedinit(void) 1601134791Sjulian{ 1602165762Sjeff 1603134791Sjulian /* 1604134791Sjulian * Set up the scheduler specific parts of proc0. 1605134791Sjulian */ 1606136167Sjulian proc0.p_sched = NULL; /* XXX */ 1607164936Sjulian thread0.td_sched = &td_sched0; 1608165762Sjeff td_sched0.ts_ltick = ticks; 1609165796Sjeff td_sched0.ts_ftick = ticks; 1610242736Sjeff td_sched0.ts_slice = 0; 1611134791Sjulian} 1612134791Sjulian 1613134791Sjulian/* 1614113357Sjeff * This is only somewhat accurate since given many processes of the same 1615113357Sjeff * priority they will switch when their slices run out, which will be 1616165762Sjeff * at most sched_slice stathz ticks. 1617113357Sjeff */ 1618109864Sjeffint 1619109864Sjeffsched_rr_interval(void) 1620109864Sjeff{ 1621165762Sjeff 1622239185Smav /* Convert sched_slice from stathz to hz. */ 1623239196Smav return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz)); 1624109864Sjeff} 1625109864Sjeff 1626171482Sjeff/* 1627171482Sjeff * Update the percent cpu tracking information when it is requested or 1628171482Sjeff * the total history exceeds the maximum. We keep a sliding history of 1629171482Sjeff * tick counts that slowly decays. This is less precise than the 4BSD 1630171482Sjeff * mechanism since it happens with less regular and frequent events. 1631171482Sjeff */ 1632121790Sjeffstatic void 1633232917Smavsched_pctcpu_update(struct td_sched *ts, int run) 1634109864Sjeff{ 1635232917Smav int t = ticks; 1636165762Sjeff 1637232917Smav if (t - ts->ts_ltick >= SCHED_TICK_TARG) { 1638164936Sjulian ts->ts_ticks = 0; 1639232917Smav ts->ts_ftick = t - SCHED_TICK_TARG; 1640232917Smav } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) { 1641232917Smav ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) * 1642232917Smav (ts->ts_ltick - (t - SCHED_TICK_TARG)); 1643232917Smav ts->ts_ftick = t - SCHED_TICK_TARG; 1644232917Smav } 1645232917Smav if (run) 1646232917Smav ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT; 1647232917Smav ts->ts_ltick = t; 1648109864Sjeff} 1649109864Sjeff 1650171482Sjeff/* 1651171482Sjeff * Adjust the priority of a thread. Move it to the appropriate run-queue 1652171482Sjeff * if necessary. This is the back-end for several priority related 1653171482Sjeff * functions. 1654171482Sjeff */ 1655165762Sjeffstatic void 1656139453Sjhbsched_thread_priority(struct thread *td, u_char prio) 1657109864Sjeff{ 1658164936Sjulian struct td_sched *ts; 1659177009Sjeff struct tdq *tdq; 1660177009Sjeff int oldpri; 1661109864Sjeff 1662187357Sjeff KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio", 1663187357Sjeff "prio:%d", td->td_priority, "new prio:%d", prio, 1664187357Sjeff KTR_ATTR_LINKED, sched_tdname(curthread)); 1665260817Savg SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio); 1666240513Savg if (td != curthread && prio < td->td_priority) { 1667187357Sjeff KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 1668187357Sjeff "lend prio", "prio:%d", td->td_priority, "new prio:%d", 1669187357Sjeff prio, KTR_ATTR_LINKED, sched_tdname(td)); 1670260817Savg SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio, 1671235459Srstone curthread); 1672187357Sjeff } 1673164936Sjulian ts = td->td_sched; 1674170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1675139453Sjhb if (td->td_priority == prio) 1676139453Sjhb return; 1677177376Sjeff /* 1678177376Sjeff * If the priority has been elevated due to priority 1679177376Sjeff * propagation, we may have to move ourselves to a new 1680177376Sjeff * queue. This could be optimized to not re-add in some 1681177376Sjeff * cases. 1682177376Sjeff */ 1683165766Sjeff if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1684165762Sjeff sched_rem(td); 1685165762Sjeff td->td_priority = prio; 1686171482Sjeff sched_add(td, SRQ_BORROWING); 1687177009Sjeff return; 1688177009Sjeff } 1689177376Sjeff /* 1690177376Sjeff * If the thread is currently running we may have to adjust the lowpri 1691177376Sjeff * information so other cpus are aware of our current priority. 1692177376Sjeff */ 1693177009Sjeff if (TD_IS_RUNNING(td)) { 1694177376Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1695177376Sjeff oldpri = td->td_priority; 1696177376Sjeff td->td_priority = prio; 1697176735Sjeff if (prio < tdq->tdq_lowpri) 1698171482Sjeff tdq->tdq_lowpri = prio; 1699176735Sjeff else if (tdq->tdq_lowpri == oldpri) 1700176735Sjeff tdq_setlowpri(tdq, td); 1701177376Sjeff return; 1702177009Sjeff } 1703177376Sjeff td->td_priority = prio; 1704109864Sjeff} 1705109864Sjeff 1706139453Sjhb/* 1707139453Sjhb * Update a thread's priority when it is lent another thread's 1708139453Sjhb * priority. 1709139453Sjhb */ 1710109864Sjeffvoid 1711139453Sjhbsched_lend_prio(struct thread *td, u_char prio) 1712139453Sjhb{ 1713139453Sjhb 1714139453Sjhb td->td_flags |= TDF_BORROWING; 1715139453Sjhb sched_thread_priority(td, prio); 1716139453Sjhb} 1717139453Sjhb 1718139453Sjhb/* 1719139453Sjhb * Restore a thread's priority when priority propagation is 1720139453Sjhb * over. The prio argument is the minimum priority the thread 1721139453Sjhb * needs to have to satisfy other possible priority lending 1722139453Sjhb * requests. If the thread's regular priority is less 1723139453Sjhb * important than prio, the thread will keep a priority boost 1724139453Sjhb * of prio. 1725139453Sjhb */ 1726139453Sjhbvoid 1727139453Sjhbsched_unlend_prio(struct thread *td, u_char prio) 1728139453Sjhb{ 1729139453Sjhb u_char base_pri; 1730139453Sjhb 1731139453Sjhb if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1732139453Sjhb td->td_base_pri <= PRI_MAX_TIMESHARE) 1733163709Sjb base_pri = td->td_user_pri; 1734139453Sjhb else 1735139453Sjhb base_pri = td->td_base_pri; 1736139453Sjhb if (prio >= base_pri) { 1737139455Sjhb td->td_flags &= ~TDF_BORROWING; 1738139453Sjhb sched_thread_priority(td, base_pri); 1739139453Sjhb } else 1740139453Sjhb sched_lend_prio(td, prio); 1741139453Sjhb} 1742139453Sjhb 1743171482Sjeff/* 1744171482Sjeff * Standard entry for setting the priority to an absolute value. 1745171482Sjeff */ 1746139453Sjhbvoid 1747139453Sjhbsched_prio(struct thread *td, u_char prio) 1748139453Sjhb{ 1749139453Sjhb u_char oldprio; 1750139453Sjhb 1751139453Sjhb /* First, update the base priority. */ 1752139453Sjhb td->td_base_pri = prio; 1753139453Sjhb 1754139453Sjhb /* 1755139455Sjhb * If the thread is borrowing another thread's priority, don't 1756139453Sjhb * ever lower the priority. 1757139453Sjhb */ 1758139453Sjhb if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1759139453Sjhb return; 1760139453Sjhb 1761139453Sjhb /* Change the real priority. */ 1762139453Sjhb oldprio = td->td_priority; 1763139453Sjhb sched_thread_priority(td, prio); 1764139453Sjhb 1765139453Sjhb /* 1766139453Sjhb * If the thread is on a turnstile, then let the turnstile update 1767139453Sjhb * its state. 1768139453Sjhb */ 1769139453Sjhb if (TD_ON_LOCK(td) && oldprio != prio) 1770139453Sjhb turnstile_adjust(td, oldprio); 1771139453Sjhb} 1772139455Sjhb 1773171482Sjeff/* 1774171482Sjeff * Set the base user priority, does not effect current running priority. 1775171482Sjeff */ 1776139453Sjhbvoid 1777163709Sjbsched_user_prio(struct thread *td, u_char prio) 1778161599Sdavidxu{ 1779161599Sdavidxu 1780163709Sjb td->td_base_user_pri = prio; 1781216313Sdavidxu if (td->td_lend_user_pri <= prio) 1782216313Sdavidxu return; 1783163709Sjb td->td_user_pri = prio; 1784161599Sdavidxu} 1785161599Sdavidxu 1786161599Sdavidxuvoid 1787161599Sdavidxusched_lend_user_prio(struct thread *td, u_char prio) 1788161599Sdavidxu{ 1789161599Sdavidxu 1790174536Sdavidxu THREAD_LOCK_ASSERT(td, MA_OWNED); 1791216313Sdavidxu td->td_lend_user_pri = prio; 1792216791Sdavidxu td->td_user_pri = min(prio, td->td_base_user_pri); 1793216791Sdavidxu if (td->td_priority > td->td_user_pri) 1794216791Sdavidxu sched_prio(td, td->td_user_pri); 1795216791Sdavidxu else if (td->td_priority != td->td_user_pri) 1796216791Sdavidxu td->td_flags |= TDF_NEEDRESCHED; 1797161599Sdavidxu} 1798161599Sdavidxu 1799171482Sjeff/* 1800171713Sjeff * Handle migration from sched_switch(). This happens only for 1801171713Sjeff * cpu binding. 1802171713Sjeff */ 1803171713Sjeffstatic struct mtx * 1804171713Sjeffsched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1805171713Sjeff{ 1806171713Sjeff struct tdq *tdn; 1807171713Sjeff 1808171713Sjeff tdn = TDQ_CPU(td->td_sched->ts_cpu); 1809171713Sjeff#ifdef SMP 1810177435Sjeff tdq_load_rem(tdq, td); 1811171713Sjeff /* 1812171713Sjeff * Do the lock dance required to avoid LOR. We grab an extra 1813171713Sjeff * spinlock nesting to prevent preemption while we're 1814171713Sjeff * not holding either run-queue lock. 1815171713Sjeff */ 1816171713Sjeff spinlock_enter(); 1817202889Sattilio thread_lock_block(td); /* This releases the lock on tdq. */ 1818197223Sattilio 1819197223Sattilio /* 1820197223Sattilio * Acquire both run-queue locks before placing the thread on the new 1821197223Sattilio * run-queue to avoid deadlocks created by placing a thread with a 1822197223Sattilio * blocked lock on the run-queue of a remote processor. The deadlock 1823197223Sattilio * occurs when a third processor attempts to lock the two queues in 1824197223Sattilio * question while the target processor is spinning with its own 1825197223Sattilio * run-queue lock held while waiting for the blocked lock to clear. 1826197223Sattilio */ 1827197223Sattilio tdq_lock_pair(tdn, tdq); 1828171713Sjeff tdq_add(tdn, td, flags); 1829177435Sjeff tdq_notify(tdn, td); 1830197223Sattilio TDQ_UNLOCK(tdn); 1831171713Sjeff spinlock_exit(); 1832171713Sjeff#endif 1833171713Sjeff return (TDQ_LOCKPTR(tdn)); 1834171713Sjeff} 1835171713Sjeff 1836171713Sjeff/* 1837202889Sattilio * Variadic version of thread_lock_unblock() that does not assume td_lock 1838202889Sattilio * is blocked. 1839171482Sjeff */ 1840171482Sjeffstatic inline void 1841171482Sjeffthread_unblock_switch(struct thread *td, struct mtx *mtx) 1842171482Sjeff{ 1843171482Sjeff atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1844171482Sjeff (uintptr_t)mtx); 1845171482Sjeff} 1846171482Sjeff 1847171482Sjeff/* 1848171482Sjeff * Switch threads. This function has to handle threads coming in while 1849171482Sjeff * blocked for some reason, running, or idle. It also must deal with 1850171482Sjeff * migrating a thread from one queue to another as running threads may 1851171482Sjeff * be assigned elsewhere via binding. 1852171482Sjeff */ 1853161599Sdavidxuvoid 1854135051Sjuliansched_switch(struct thread *td, struct thread *newtd, int flags) 1855109864Sjeff{ 1856165627Sjeff struct tdq *tdq; 1857164936Sjulian struct td_sched *ts; 1858171482Sjeff struct mtx *mtx; 1859171713Sjeff int srqflag; 1860239157Smav int cpuid, preempted; 1861109864Sjeff 1862170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1863177376Sjeff KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1864109864Sjeff 1865171482Sjeff cpuid = PCPU_GET(cpuid); 1866171482Sjeff tdq = TDQ_CPU(cpuid); 1867164936Sjulian ts = td->td_sched; 1868171713Sjeff mtx = td->td_lock; 1869232917Smav sched_pctcpu_update(ts, 1); 1870171482Sjeff ts->ts_rltick = ticks; 1871133555Sjeff td->td_lastcpu = td->td_oncpu; 1872113339Sjulian td->td_oncpu = NOCPU; 1873312666Savg preempted = (td->td_flags & TDF_SLICEEND) == 0 && 1874312666Savg (flags & SW_PREEMPT) != 0; 1875239157Smav td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND); 1876144777Sups td->td_owepreempt = 0; 1877242852Smav if (!TD_IS_IDLETHREAD(td)) 1878242852Smav tdq->tdq_switchcnt++; 1879123434Sjeff /* 1880171482Sjeff * The lock pointer in an idle thread should never change. Reset it 1881171482Sjeff * to CAN_RUN as well. 1882123434Sjeff */ 1883167327Sjulian if (TD_IS_IDLETHREAD(td)) { 1884171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1885139334Sjeff TD_SET_CAN_RUN(td); 1886170293Sjeff } else if (TD_IS_RUNNING(td)) { 1887171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1888239157Smav srqflag = preempted ? 1889170293Sjeff SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1890171713Sjeff SRQ_OURSELF|SRQ_YIELDING; 1891212153Smdf#ifdef SMP 1892212115Smdf if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu)) 1893212115Smdf ts->ts_cpu = sched_pickcpu(td, 0); 1894212153Smdf#endif 1895171713Sjeff if (ts->ts_cpu == cpuid) 1896177435Sjeff tdq_runq_add(tdq, td, srqflag); 1897212115Smdf else { 1898212115Smdf KASSERT(THREAD_CAN_MIGRATE(td) || 1899212115Smdf (ts->ts_flags & TSF_BOUND) != 0, 1900212115Smdf ("Thread %p shouldn't migrate", td)); 1901171713Sjeff mtx = sched_switch_migrate(tdq, td, srqflag); 1902212115Smdf } 1903171482Sjeff } else { 1904171482Sjeff /* This thread must be going to sleep. */ 1905171482Sjeff TDQ_LOCK(tdq); 1906202889Sattilio mtx = thread_lock_block(td); 1907177435Sjeff tdq_load_rem(tdq, td); 1908171482Sjeff } 1909316841Savg 1910316841Savg#if (KTR_COMPILE & KTR_SCHED) != 0 1911316841Savg if (TD_IS_IDLETHREAD(td)) 1912316841Savg KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle", 1913316841Savg "prio:%d", td->td_priority); 1914316841Savg else 1915316841Savg KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td), 1916316841Savg "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg, 1917316841Savg "lockname:\"%s\"", td->td_lockname); 1918316841Savg#endif 1919316841Savg 1920171482Sjeff /* 1921171482Sjeff * We enter here with the thread blocked and assigned to the 1922171482Sjeff * appropriate cpu run-queue or sleep-queue and with the current 1923171482Sjeff * thread-queue locked. 1924171482Sjeff */ 1925171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1926171482Sjeff newtd = choosethread(); 1927171482Sjeff /* 1928171482Sjeff * Call the MD code to switch contexts if necessary. 1929171482Sjeff */ 1930145256Sjkoshy if (td != newtd) { 1931145256Sjkoshy#ifdef HWPMC_HOOKS 1932145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1933145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1934145256Sjkoshy#endif 1935260817Savg SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc); 1936174629Sjeff lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1937172411Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1938232917Smav sched_pctcpu_update(newtd->td_sched, 0); 1939179297Sjb 1940179297Sjb#ifdef KDTRACE_HOOKS 1941179297Sjb /* 1942179297Sjb * If DTrace has set the active vtime enum to anything 1943179297Sjb * other than INACTIVE (0), then it should have set the 1944179297Sjb * function to call. 1945179297Sjb */ 1946179297Sjb if (dtrace_vtime_active) 1947179297Sjb (*dtrace_vtime_switch_func)(newtd); 1948179297Sjb#endif 1949179297Sjb 1950171482Sjeff cpu_switch(td, newtd, mtx); 1951171482Sjeff /* 1952171482Sjeff * We may return from cpu_switch on a different cpu. However, 1953171482Sjeff * we always return with td_lock pointing to the current cpu's 1954171482Sjeff * run queue lock. 1955171482Sjeff */ 1956171482Sjeff cpuid = PCPU_GET(cpuid); 1957171482Sjeff tdq = TDQ_CPU(cpuid); 1958174629Sjeff lock_profile_obtain_lock_success( 1959174629Sjeff &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1960235459Srstone 1961260817Savg SDT_PROBE0(sched, , , on__cpu); 1962145256Sjkoshy#ifdef HWPMC_HOOKS 1963145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1964145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1965145256Sjkoshy#endif 1966235459Srstone } else { 1967171482Sjeff thread_unblock_switch(td, mtx); 1968260817Savg SDT_PROBE0(sched, , , remain__cpu); 1969235459Srstone } 1970316841Savg 1971316841Savg KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running", 1972316841Savg "prio:%d", td->td_priority); 1973316841Savg 1974171482Sjeff /* 1975171482Sjeff * Assert that all went well and return. 1976171482Sjeff */ 1977171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1978171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1979171482Sjeff td->td_oncpu = cpuid; 1980109864Sjeff} 1981109864Sjeff 1982171482Sjeff/* 1983171482Sjeff * Adjust thread priorities as a result of a nice request. 1984171482Sjeff */ 1985109864Sjeffvoid 1986130551Sjuliansched_nice(struct proc *p, int nice) 1987109864Sjeff{ 1988109864Sjeff struct thread *td; 1989109864Sjeff 1990130551Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 1991165762Sjeff 1992130551Sjulian p->p_nice = nice; 1993163709Sjb FOREACH_THREAD_IN_PROC(p, td) { 1994170293Sjeff thread_lock(td); 1995163709Sjb sched_priority(td); 1996165762Sjeff sched_prio(td, td->td_base_user_pri); 1997170293Sjeff thread_unlock(td); 1998130551Sjulian } 1999109864Sjeff} 2000109864Sjeff 2001171482Sjeff/* 2002171482Sjeff * Record the sleep time for the interactivity scorer. 2003171482Sjeff */ 2004109864Sjeffvoid 2005177085Sjeffsched_sleep(struct thread *td, int prio) 2006109864Sjeff{ 2007165762Sjeff 2008170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2009109864Sjeff 2010172264Sjeff td->td_slptick = ticks; 2011201347Skib if (TD_IS_SUSPENDED(td) || prio >= PSOCK) 2012177085Sjeff td->td_flags |= TDF_CANSWAP; 2013217410Sjhb if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 2014217410Sjhb return; 2015177903Sjeff if (static_boost == 1 && prio) 2016177085Sjeff sched_prio(td, prio); 2017177903Sjeff else if (static_boost && td->td_priority > static_boost) 2018177903Sjeff sched_prio(td, static_boost); 2019109864Sjeff} 2020109864Sjeff 2021171482Sjeff/* 2022171482Sjeff * Schedule a thread to resume execution and record how long it voluntarily 2023171482Sjeff * slept. We also update the pctcpu, interactivity, and priority. 2024171482Sjeff */ 2025109864Sjeffvoid 2026109864Sjeffsched_wakeup(struct thread *td) 2027109864Sjeff{ 2028166229Sjeff struct td_sched *ts; 2029171482Sjeff int slptick; 2030165762Sjeff 2031170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2032166229Sjeff ts = td->td_sched; 2033177085Sjeff td->td_flags &= ~TDF_CANSWAP; 2034109864Sjeff /* 2035165762Sjeff * If we slept for more than a tick update our interactivity and 2036165762Sjeff * priority. 2037109864Sjeff */ 2038172264Sjeff slptick = td->td_slptick; 2039172264Sjeff td->td_slptick = 0; 2040171482Sjeff if (slptick && slptick != ticks) { 2041232917Smav ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT; 2042165819Sjeff sched_interact_update(td); 2043232917Smav sched_pctcpu_update(ts, 0); 2044109864Sjeff } 2045242736Sjeff /* 2046242736Sjeff * Reset the slice value since we slept and advanced the round-robin. 2047242736Sjeff */ 2048242736Sjeff ts->ts_slice = 0; 2049166190Sjeff sched_add(td, SRQ_BORING); 2050109864Sjeff} 2051109864Sjeff 2052109864Sjeff/* 2053109864Sjeff * Penalize the parent for creating a new child and initialize the child's 2054109864Sjeff * priority. 2055109864Sjeff */ 2056109864Sjeffvoid 2057163709Sjbsched_fork(struct thread *td, struct thread *child) 2058109864Sjeff{ 2059170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2060232917Smav sched_pctcpu_update(td->td_sched, 1); 2061164936Sjulian sched_fork_thread(td, child); 2062165762Sjeff /* 2063165762Sjeff * Penalize the parent and child for forking. 2064165762Sjeff */ 2065165762Sjeff sched_interact_fork(child); 2066165762Sjeff sched_priority(child); 2067171482Sjeff td->td_sched->ts_runtime += tickincr; 2068165762Sjeff sched_interact_update(td); 2069165762Sjeff sched_priority(td); 2070164936Sjulian} 2071109864Sjeff 2072171482Sjeff/* 2073171482Sjeff * Fork a new thread, may be within the same process. 2074171482Sjeff */ 2075164936Sjulianvoid 2076164936Sjuliansched_fork_thread(struct thread *td, struct thread *child) 2077164936Sjulian{ 2078164936Sjulian struct td_sched *ts; 2079164936Sjulian struct td_sched *ts2; 2080242736Sjeff struct tdq *tdq; 2081164936Sjulian 2082242736Sjeff tdq = TDQ_SELF(); 2083177426Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2084165762Sjeff /* 2085165762Sjeff * Initialize child. 2086165762Sjeff */ 2087177426Sjeff ts = td->td_sched; 2088177426Sjeff ts2 = child->td_sched; 2089288463Sjhb child->td_oncpu = NOCPU; 2090288463Sjhb child->td_lastcpu = NOCPU; 2091242736Sjeff child->td_lock = TDQ_LOCKPTR(tdq); 2092176735Sjeff child->td_cpuset = cpuset_ref(td->td_cpuset); 2093164936Sjulian ts2->ts_cpu = ts->ts_cpu; 2094177426Sjeff ts2->ts_flags = 0; 2095165762Sjeff /* 2096217078Sjhb * Grab our parents cpu estimation information. 2097165762Sjeff */ 2098164936Sjulian ts2->ts_ticks = ts->ts_ticks; 2099164936Sjulian ts2->ts_ltick = ts->ts_ltick; 2100164936Sjulian ts2->ts_ftick = ts->ts_ftick; 2101165762Sjeff /* 2102217078Sjhb * Do not inherit any borrowed priority from the parent. 2103217078Sjhb */ 2104217078Sjhb child->td_priority = child->td_base_pri; 2105217078Sjhb /* 2106165762Sjeff * And update interactivity score. 2107165762Sjeff */ 2108171482Sjeff ts2->ts_slptime = ts->ts_slptime; 2109171482Sjeff ts2->ts_runtime = ts->ts_runtime; 2110242736Sjeff /* Attempt to quickly learn interactivity. */ 2111242736Sjeff ts2->ts_slice = tdq_slice(tdq) - sched_slice_min; 2112187357Sjeff#ifdef KTR 2113187357Sjeff bzero(ts2->ts_name, sizeof(ts2->ts_name)); 2114187357Sjeff#endif 2115113357Sjeff} 2116113357Sjeff 2117171482Sjeff/* 2118171482Sjeff * Adjust the priority class of a thread. 2119171482Sjeff */ 2120113357Sjeffvoid 2121163709Sjbsched_class(struct thread *td, int class) 2122113357Sjeff{ 2123113357Sjeff 2124170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2125163709Sjb if (td->td_pri_class == class) 2126113357Sjeff return; 2127163709Sjb td->td_pri_class = class; 2128109864Sjeff} 2129109864Sjeff 2130109864Sjeff/* 2131109864Sjeff * Return some of the child's priority and interactivity to the parent. 2132109864Sjeff */ 2133109864Sjeffvoid 2134164939Sjuliansched_exit(struct proc *p, struct thread *child) 2135109864Sjeff{ 2136165762Sjeff struct thread *td; 2137113372Sjeff 2138187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit", 2139225199Sdelphij "prio:%d", child->td_priority); 2140177368Sjeff PROC_LOCK_ASSERT(p, MA_OWNED); 2141165762Sjeff td = FIRST_THREAD_IN_PROC(p); 2142165762Sjeff sched_exit_thread(td, child); 2143113372Sjeff} 2144113372Sjeff 2145171482Sjeff/* 2146171482Sjeff * Penalize another thread for the time spent on this one. This helps to 2147171482Sjeff * worsen the priority and interactivity of processes which schedule batch 2148171482Sjeff * jobs such as make. This has little effect on the make process itself but 2149171482Sjeff * causes new processes spawned by it to receive worse scores immediately. 2150171482Sjeff */ 2151113372Sjeffvoid 2152164939Sjuliansched_exit_thread(struct thread *td, struct thread *child) 2153164936Sjulian{ 2154165762Sjeff 2155187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit", 2156225199Sdelphij "prio:%d", child->td_priority); 2157165762Sjeff /* 2158165762Sjeff * Give the child's runtime to the parent without returning the 2159165762Sjeff * sleep time as a penalty to the parent. This causes shells that 2160165762Sjeff * launch expensive things to mark their children as expensive. 2161165762Sjeff */ 2162170293Sjeff thread_lock(td); 2163171482Sjeff td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2164164939Sjulian sched_interact_update(td); 2165165762Sjeff sched_priority(td); 2166170293Sjeff thread_unlock(td); 2167164936Sjulian} 2168164936Sjulian 2169177005Sjeffvoid 2170177005Sjeffsched_preempt(struct thread *td) 2171177005Sjeff{ 2172177005Sjeff struct tdq *tdq; 2173177005Sjeff 2174235459Srstone SDT_PROBE2(sched, , , surrender, td, td->td_proc); 2175235459Srstone 2176177005Sjeff thread_lock(td); 2177177005Sjeff tdq = TDQ_SELF(); 2178177005Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2179177005Sjeff tdq->tdq_ipipending = 0; 2180177005Sjeff if (td->td_priority > tdq->tdq_lowpri) { 2181178272Sjeff int flags; 2182178272Sjeff 2183178272Sjeff flags = SW_INVOL | SW_PREEMPT; 2184177005Sjeff if (td->td_critnest > 1) 2185177005Sjeff td->td_owepreempt = 1; 2186178272Sjeff else if (TD_IS_IDLETHREAD(td)) 2187178272Sjeff mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL); 2188177005Sjeff else 2189178272Sjeff mi_switch(flags | SWT_REMOTEPREEMPT, NULL); 2190177005Sjeff } 2191177005Sjeff thread_unlock(td); 2192177005Sjeff} 2193177005Sjeff 2194171482Sjeff/* 2195171482Sjeff * Fix priorities on return to user-space. Priorities may be elevated due 2196171482Sjeff * to static priorities in msleep() or similar. 2197171482Sjeff */ 2198164936Sjulianvoid 2199164936Sjuliansched_userret(struct thread *td) 2200164936Sjulian{ 2201164936Sjulian /* 2202164936Sjulian * XXX we cheat slightly on the locking here to avoid locking in 2203164936Sjulian * the usual case. Setting td_priority here is essentially an 2204164936Sjulian * incomplete workaround for not setting it properly elsewhere. 2205164936Sjulian * Now that some interrupt handlers are threads, not setting it 2206164936Sjulian * properly elsewhere can clobber it in the window between setting 2207164936Sjulian * it here and returning to user mode, so don't waste time setting 2208164936Sjulian * it perfectly here. 2209164936Sjulian */ 2210164936Sjulian KASSERT((td->td_flags & TDF_BORROWING) == 0, 2211164936Sjulian ("thread with borrowed priority returning to userland")); 2212164936Sjulian if (td->td_priority != td->td_user_pri) { 2213170293Sjeff thread_lock(td); 2214164936Sjulian td->td_priority = td->td_user_pri; 2215164936Sjulian td->td_base_pri = td->td_user_pri; 2216177005Sjeff tdq_setlowpri(TDQ_SELF(), td); 2217170293Sjeff thread_unlock(td); 2218164936Sjulian } 2219164936Sjulian} 2220164936Sjulian 2221171482Sjeff/* 2222171482Sjeff * Handle a stathz tick. This is really only relevant for timeshare 2223171482Sjeff * threads. 2224171482Sjeff */ 2225164936Sjulianvoid 2226121127Sjeffsched_clock(struct thread *td) 2227109864Sjeff{ 2228164936Sjulian struct tdq *tdq; 2229164936Sjulian struct td_sched *ts; 2230109864Sjeff 2231171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2232164936Sjulian tdq = TDQ_SELF(); 2233172409Sjeff#ifdef SMP 2234133427Sjeff /* 2235172409Sjeff * We run the long term load balancer infrequently on the first cpu. 2236172409Sjeff */ 2237172409Sjeff if (balance_tdq == tdq) { 2238172409Sjeff if (balance_ticks && --balance_ticks == 0) 2239172409Sjeff sched_balance(); 2240172409Sjeff } 2241172409Sjeff#endif 2242172409Sjeff /* 2243178277Sjeff * Save the old switch count so we have a record of the last ticks 2244178277Sjeff * activity. Initialize the new switch count based on our load. 2245178277Sjeff * If there is some activity seed it to reflect that. 2246178277Sjeff */ 2247178277Sjeff tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2248178471Sjeff tdq->tdq_switchcnt = tdq->tdq_load; 2249178277Sjeff /* 2250165766Sjeff * Advance the insert index once for each tick to ensure that all 2251165766Sjeff * threads get a chance to run. 2252133427Sjeff */ 2253165766Sjeff if (tdq->tdq_idx == tdq->tdq_ridx) { 2254165766Sjeff tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2255165766Sjeff if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2256165766Sjeff tdq->tdq_ridx = tdq->tdq_idx; 2257165766Sjeff } 2258165766Sjeff ts = td->td_sched; 2259232917Smav sched_pctcpu_update(ts, 1); 2260175104Sjeff if (td->td_pri_class & PRI_FIFO_BIT) 2261113357Sjeff return; 2262217291Sjhb if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) { 2263175104Sjeff /* 2264175104Sjeff * We used a tick; charge it to the thread so 2265175104Sjeff * that we can compute our interactivity. 2266175104Sjeff */ 2267175104Sjeff td->td_sched->ts_runtime += tickincr; 2268175104Sjeff sched_interact_update(td); 2269177009Sjeff sched_priority(td); 2270175104Sjeff } 2271239185Smav 2272113357Sjeff /* 2273239185Smav * Force a context switch if the current thread has used up a full 2274239185Smav * time slice (default is 100ms). 2275109864Sjeff */ 2276242736Sjeff if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) { 2277242736Sjeff ts->ts_slice = 0; 2278239185Smav td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND; 2279239185Smav } 2280109864Sjeff} 2281109864Sjeff 2282171482Sjeff/* 2283232917Smav * Called once per hz tick. 2284171482Sjeff */ 2285171482Sjeffvoid 2286212541Smavsched_tick(int cnt) 2287171482Sjeff{ 2288171482Sjeff 2289171482Sjeff} 2290171482Sjeff 2291171482Sjeff/* 2292171482Sjeff * Return whether the current CPU has runnable tasks. Used for in-kernel 2293171482Sjeff * cooperative idle threads. 2294171482Sjeff */ 2295109864Sjeffint 2296109864Sjeffsched_runnable(void) 2297109864Sjeff{ 2298164936Sjulian struct tdq *tdq; 2299115998Sjeff int load; 2300109864Sjeff 2301115998Sjeff load = 1; 2302115998Sjeff 2303164936Sjulian tdq = TDQ_SELF(); 2304121605Sjeff if ((curthread->td_flags & TDF_IDLETD) != 0) { 2305165620Sjeff if (tdq->tdq_load > 0) 2306121605Sjeff goto out; 2307121605Sjeff } else 2308165620Sjeff if (tdq->tdq_load - 1 > 0) 2309121605Sjeff goto out; 2310115998Sjeff load = 0; 2311115998Sjeffout: 2312115998Sjeff return (load); 2313109864Sjeff} 2314109864Sjeff 2315171482Sjeff/* 2316171482Sjeff * Choose the highest priority thread to run. The thread is removed from 2317171482Sjeff * the run-queue while running however the load remains. For SMP we set 2318171482Sjeff * the tdq in the global idle bitmask if it idles here. 2319171482Sjeff */ 2320166190Sjeffstruct thread * 2321109970Sjeffsched_choose(void) 2322109970Sjeff{ 2323177435Sjeff struct thread *td; 2324164936Sjulian struct tdq *tdq; 2325109970Sjeff 2326164936Sjulian tdq = TDQ_SELF(); 2327171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2328177435Sjeff td = tdq_choose(tdq); 2329177435Sjeff if (td) { 2330177435Sjeff tdq_runq_rem(tdq, td); 2331177903Sjeff tdq->tdq_lowpri = td->td_priority; 2332177435Sjeff return (td); 2333109864Sjeff } 2334177903Sjeff tdq->tdq_lowpri = PRI_MAX_IDLE; 2335176735Sjeff return (PCPU_GET(idlethread)); 2336109864Sjeff} 2337109864Sjeff 2338171482Sjeff/* 2339171482Sjeff * Set owepreempt if necessary. Preemption never happens directly in ULE, 2340171482Sjeff * we always request it once we exit a critical section. 2341171482Sjeff */ 2342171482Sjeffstatic inline void 2343171482Sjeffsched_setpreempt(struct thread *td) 2344166190Sjeff{ 2345166190Sjeff struct thread *ctd; 2346166190Sjeff int cpri; 2347166190Sjeff int pri; 2348166190Sjeff 2349177005Sjeff THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2350177005Sjeff 2351166190Sjeff ctd = curthread; 2352166190Sjeff pri = td->td_priority; 2353166190Sjeff cpri = ctd->td_priority; 2354177005Sjeff if (pri < cpri) 2355177005Sjeff ctd->td_flags |= TDF_NEEDRESCHED; 2356166190Sjeff if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2357171482Sjeff return; 2358177005Sjeff if (!sched_shouldpreempt(pri, cpri, 0)) 2359171482Sjeff return; 2360171482Sjeff ctd->td_owepreempt = 1; 2361166190Sjeff} 2362166190Sjeff 2363171482Sjeff/* 2364177009Sjeff * Add a thread to a thread queue. Select the appropriate runq and add the 2365177009Sjeff * thread to it. This is the internal function called when the tdq is 2366177009Sjeff * predetermined. 2367171482Sjeff */ 2368109864Sjeffvoid 2369171482Sjefftdq_add(struct tdq *tdq, struct thread *td, int flags) 2370109864Sjeff{ 2371109864Sjeff 2372171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2373166190Sjeff KASSERT((td->td_inhibitors == 0), 2374166190Sjeff ("sched_add: trying to run inhibited thread")); 2375166190Sjeff KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2376166190Sjeff ("sched_add: bad thread state")); 2377172207Sjeff KASSERT(td->td_flags & TDF_INMEM, 2378172207Sjeff ("sched_add: thread swapped out")); 2379171482Sjeff 2380171482Sjeff if (td->td_priority < tdq->tdq_lowpri) 2381171482Sjeff tdq->tdq_lowpri = td->td_priority; 2382177435Sjeff tdq_runq_add(tdq, td, flags); 2383177435Sjeff tdq_load_add(tdq, td); 2384171482Sjeff} 2385171482Sjeff 2386171482Sjeff/* 2387171482Sjeff * Select the target thread queue and add a thread to it. Request 2388171482Sjeff * preemption or IPI a remote processor if required. 2389171482Sjeff */ 2390171482Sjeffvoid 2391171482Sjeffsched_add(struct thread *td, int flags) 2392171482Sjeff{ 2393171482Sjeff struct tdq *tdq; 2394171482Sjeff#ifdef SMP 2395171482Sjeff int cpu; 2396171482Sjeff#endif 2397187357Sjeff 2398187357Sjeff KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 2399187357Sjeff "prio:%d", td->td_priority, KTR_ATTR_LINKED, 2400187357Sjeff sched_tdname(curthread)); 2401187357Sjeff KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 2402187357Sjeff KTR_ATTR_LINKED, sched_tdname(td)); 2403235459Srstone SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 2404235459Srstone flags & SRQ_PREEMPTED); 2405171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2406166108Sjeff /* 2407171482Sjeff * Recalculate the priority before we select the target cpu or 2408171482Sjeff * run-queue. 2409166108Sjeff */ 2410171482Sjeff if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2411171482Sjeff sched_priority(td); 2412171482Sjeff#ifdef SMP 2413171482Sjeff /* 2414171482Sjeff * Pick the destination cpu and if it isn't ours transfer to the 2415171482Sjeff * target cpu. 2416171482Sjeff */ 2417177435Sjeff cpu = sched_pickcpu(td, flags); 2418177435Sjeff tdq = sched_setcpu(td, cpu, flags); 2419171482Sjeff tdq_add(tdq, td, flags); 2420177009Sjeff if (cpu != PCPU_GET(cpuid)) { 2421177435Sjeff tdq_notify(tdq, td); 2422166108Sjeff return; 2423166108Sjeff } 2424171482Sjeff#else 2425171482Sjeff tdq = TDQ_SELF(); 2426171482Sjeff TDQ_LOCK(tdq); 2427171482Sjeff /* 2428171482Sjeff * Now that the thread is moving to the run-queue, set the lock 2429171482Sjeff * to the scheduler's lock. 2430171482Sjeff */ 2431171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2432171482Sjeff tdq_add(tdq, td, flags); 2433166108Sjeff#endif 2434171482Sjeff if (!(flags & SRQ_YIELDING)) 2435171482Sjeff sched_setpreempt(td); 2436109864Sjeff} 2437109864Sjeff 2438171482Sjeff/* 2439171482Sjeff * Remove a thread from a run-queue without running it. This is used 2440171482Sjeff * when we're stealing a thread from a remote queue. Otherwise all threads 2441171482Sjeff * exit by calling sched_exit_thread() and sched_throw() themselves. 2442171482Sjeff */ 2443109864Sjeffvoid 2444121127Sjeffsched_rem(struct thread *td) 2445109864Sjeff{ 2446164936Sjulian struct tdq *tdq; 2447113357Sjeff 2448187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 2449187357Sjeff "prio:%d", td->td_priority); 2450235459Srstone SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); 2451177435Sjeff tdq = TDQ_CPU(td->td_sched->ts_cpu); 2452171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2453171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2454166190Sjeff KASSERT(TD_ON_RUNQ(td), 2455164936Sjulian ("sched_rem: thread not on run queue")); 2456177435Sjeff tdq_runq_rem(tdq, td); 2457177435Sjeff tdq_load_rem(tdq, td); 2458166190Sjeff TD_SET_CAN_RUN(td); 2459176735Sjeff if (td->td_priority == tdq->tdq_lowpri) 2460176735Sjeff tdq_setlowpri(tdq, NULL); 2461109864Sjeff} 2462109864Sjeff 2463171482Sjeff/* 2464171482Sjeff * Fetch cpu utilization information. Updates on demand. 2465171482Sjeff */ 2466109864Sjefffixpt_t 2467121127Sjeffsched_pctcpu(struct thread *td) 2468109864Sjeff{ 2469109864Sjeff fixpt_t pctcpu; 2470164936Sjulian struct td_sched *ts; 2471109864Sjeff 2472109864Sjeff pctcpu = 0; 2473164936Sjulian ts = td->td_sched; 2474164936Sjulian if (ts == NULL) 2475121290Sjeff return (0); 2476109864Sjeff 2477208787Sjhb THREAD_LOCK_ASSERT(td, MA_OWNED); 2478232917Smav sched_pctcpu_update(ts, TD_IS_RUNNING(td)); 2479164936Sjulian if (ts->ts_ticks) { 2480109864Sjeff int rtick; 2481109864Sjeff 2482109864Sjeff /* How many rtick per second ? */ 2483165762Sjeff rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2484165762Sjeff pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2485109864Sjeff } 2486109864Sjeff 2487109864Sjeff return (pctcpu); 2488109864Sjeff} 2489109864Sjeff 2490176735Sjeff/* 2491176735Sjeff * Enforce affinity settings for a thread. Called after adjustments to 2492176735Sjeff * cpumask. 2493176735Sjeff */ 2494176729Sjeffvoid 2495176729Sjeffsched_affinity(struct thread *td) 2496176729Sjeff{ 2497176735Sjeff#ifdef SMP 2498176735Sjeff struct td_sched *ts; 2499176735Sjeff 2500176735Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2501176735Sjeff ts = td->td_sched; 2502176735Sjeff if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2503176735Sjeff return; 2504189787Sjeff if (TD_ON_RUNQ(td)) { 2505189787Sjeff sched_rem(td); 2506189787Sjeff sched_add(td, SRQ_BORING); 2507189787Sjeff return; 2508189787Sjeff } 2509176735Sjeff if (!TD_IS_RUNNING(td)) 2510176735Sjeff return; 2511176735Sjeff /* 2512212115Smdf * Force a switch before returning to userspace. If the 2513212115Smdf * target thread is not running locally send an ipi to force 2514212115Smdf * the issue. 2515176735Sjeff */ 2516212974Sjhb td->td_flags |= TDF_NEEDRESCHED; 2517212115Smdf if (td != curthread) 2518212115Smdf ipi_cpu(ts->ts_cpu, IPI_PREEMPT); 2519176735Sjeff#endif 2520176729Sjeff} 2521176729Sjeff 2522171482Sjeff/* 2523171482Sjeff * Bind a thread to a target cpu. 2524171482Sjeff */ 2525122038Sjeffvoid 2526122038Sjeffsched_bind(struct thread *td, int cpu) 2527122038Sjeff{ 2528164936Sjulian struct td_sched *ts; 2529122038Sjeff 2530171713Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2531208391Sjhb KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 2532164936Sjulian ts = td->td_sched; 2533166137Sjeff if (ts->ts_flags & TSF_BOUND) 2534166152Sjeff sched_unbind(td); 2535212115Smdf KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td)); 2536164936Sjulian ts->ts_flags |= TSF_BOUND; 2537166137Sjeff sched_pin(); 2538123433Sjeff if (PCPU_GET(cpuid) == cpu) 2539122038Sjeff return; 2540166137Sjeff ts->ts_cpu = cpu; 2541122038Sjeff /* When we return from mi_switch we'll be on the correct cpu. */ 2542131527Sphk mi_switch(SW_VOL, NULL); 2543122038Sjeff} 2544122038Sjeff 2545171482Sjeff/* 2546171482Sjeff * Release a bound thread. 2547171482Sjeff */ 2548122038Sjeffvoid 2549122038Sjeffsched_unbind(struct thread *td) 2550122038Sjeff{ 2551165762Sjeff struct td_sched *ts; 2552165762Sjeff 2553170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2554208391Sjhb KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 2555165762Sjeff ts = td->td_sched; 2556166137Sjeff if ((ts->ts_flags & TSF_BOUND) == 0) 2557166137Sjeff return; 2558165762Sjeff ts->ts_flags &= ~TSF_BOUND; 2559165762Sjeff sched_unpin(); 2560122038Sjeff} 2561122038Sjeff 2562109864Sjeffint 2563145256Sjkoshysched_is_bound(struct thread *td) 2564145256Sjkoshy{ 2565170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2566164936Sjulian return (td->td_sched->ts_flags & TSF_BOUND); 2567145256Sjkoshy} 2568145256Sjkoshy 2569171482Sjeff/* 2570171482Sjeff * Basic yield call. 2571171482Sjeff */ 2572159630Sdavidxuvoid 2573159630Sdavidxusched_relinquish(struct thread *td) 2574159630Sdavidxu{ 2575170293Sjeff thread_lock(td); 2576178272Sjeff mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 2577170293Sjeff thread_unlock(td); 2578159630Sdavidxu} 2579159630Sdavidxu 2580171482Sjeff/* 2581171482Sjeff * Return the total system load. 2582171482Sjeff */ 2583145256Sjkoshyint 2584125289Sjeffsched_load(void) 2585125289Sjeff{ 2586125289Sjeff#ifdef SMP 2587125289Sjeff int total; 2588125289Sjeff int i; 2589125289Sjeff 2590125289Sjeff total = 0; 2591209059Sjhb CPU_FOREACH(i) 2592176735Sjeff total += TDQ_CPU(i)->tdq_sysload; 2593125289Sjeff return (total); 2594125289Sjeff#else 2595165620Sjeff return (TDQ_SELF()->tdq_sysload); 2596125289Sjeff#endif 2597125289Sjeff} 2598125289Sjeff 2599125289Sjeffint 2600109864Sjeffsched_sizeof_proc(void) 2601109864Sjeff{ 2602109864Sjeff return (sizeof(struct proc)); 2603109864Sjeff} 2604109864Sjeff 2605109864Sjeffint 2606109864Sjeffsched_sizeof_thread(void) 2607109864Sjeff{ 2608109864Sjeff return (sizeof(struct thread) + sizeof(struct td_sched)); 2609109864Sjeff} 2610159570Sdavidxu 2611191676Sjeff#ifdef SMP 2612191676Sjeff#define TDQ_IDLESPIN(tdq) \ 2613191676Sjeff ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0) 2614191676Sjeff#else 2615191676Sjeff#define TDQ_IDLESPIN(tdq) 1 2616191676Sjeff#endif 2617191676Sjeff 2618166190Sjeff/* 2619166190Sjeff * The actual idle process. 2620166190Sjeff */ 2621166190Sjeffvoid 2622166190Sjeffsched_idletd(void *dummy) 2623166190Sjeff{ 2624166190Sjeff struct thread *td; 2625171482Sjeff struct tdq *tdq; 2626242852Smav int oldswitchcnt, switchcnt; 2627178277Sjeff int i; 2628166190Sjeff 2629191643Sjeff mtx_assert(&Giant, MA_NOTOWNED); 2630166190Sjeff td = curthread; 2631171482Sjeff tdq = TDQ_SELF(); 2632239585Sjhb THREAD_NO_SLEEPING(); 2633242852Smav oldswitchcnt = -1; 2634171482Sjeff for (;;) { 2635242852Smav if (tdq->tdq_load) { 2636242852Smav thread_lock(td); 2637242852Smav mi_switch(SW_VOL | SWT_IDLE, NULL); 2638242852Smav thread_unlock(td); 2639242852Smav } 2640242852Smav switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2641171482Sjeff#ifdef SMP 2642242852Smav if (switchcnt != oldswitchcnt) { 2643242852Smav oldswitchcnt = switchcnt; 2644242852Smav if (tdq_idled(tdq) == 0) 2645242852Smav continue; 2646242852Smav } 2647243069Smav switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2648243069Smav#else 2649243069Smav oldswitchcnt = switchcnt; 2650171482Sjeff#endif 2651178277Sjeff /* 2652178277Sjeff * If we're switching very frequently, spin while checking 2653178277Sjeff * for load rather than entering a low power state that 2654191643Sjeff * may require an IPI. However, don't do any busy 2655191643Sjeff * loops while on SMT machines as this simply steals 2656191643Sjeff * cycles from cores doing useful work. 2657178277Sjeff */ 2658191676Sjeff if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) { 2659178277Sjeff for (i = 0; i < sched_idlespins; i++) { 2660178277Sjeff if (tdq->tdq_load) 2661178277Sjeff break; 2662178277Sjeff cpu_spinwait(); 2663178277Sjeff } 2664178277Sjeff } 2665242852Smav 2666242852Smav /* If there was context switch during spin, restart it. */ 2667191643Sjeff switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2668242852Smav if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt) 2669242852Smav continue; 2670242852Smav 2671242852Smav /* Run main MD idle handler. */ 2672242852Smav tdq->tdq_cpu_idle = 1; 2673271707Smav /* 2674271707Smav * Make sure that tdq_cpu_idle update is globally visible 2675271707Smav * before cpu_idle() read tdq_load. The order is important 2676271707Smav * to avoid race with tdq_notify. 2677271707Smav */ 2678271707Smav mb(); 2679242852Smav cpu_idle(switchcnt * 4 > sched_idlespinthresh); 2680242852Smav tdq->tdq_cpu_idle = 0; 2681242852Smav 2682242852Smav /* 2683242852Smav * Account thread-less hardware interrupts and 2684242852Smav * other wakeup reasons equal to context switches. 2685242852Smav */ 2686242852Smav switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2687242852Smav if (switchcnt != oldswitchcnt) 2688242852Smav continue; 2689242852Smav tdq->tdq_switchcnt++; 2690242852Smav oldswitchcnt++; 2691171482Sjeff } 2692166190Sjeff} 2693166190Sjeff 2694170293Sjeff/* 2695170293Sjeff * A CPU is entering for the first time or a thread is exiting. 2696170293Sjeff */ 2697170293Sjeffvoid 2698170293Sjeffsched_throw(struct thread *td) 2699170293Sjeff{ 2700172411Sjeff struct thread *newtd; 2701171482Sjeff struct tdq *tdq; 2702171482Sjeff 2703171482Sjeff tdq = TDQ_SELF(); 2704170293Sjeff if (td == NULL) { 2705171482Sjeff /* Correct spinlock nesting and acquire the correct lock. */ 2706171482Sjeff TDQ_LOCK(tdq); 2707170293Sjeff spinlock_exit(); 2708229429Sjhb PCPU_SET(switchtime, cpu_ticks()); 2709229429Sjhb PCPU_SET(switchticks, ticks); 2710170293Sjeff } else { 2711171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2712177435Sjeff tdq_load_rem(tdq, td); 2713174629Sjeff lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2714288463Sjhb td->td_lastcpu = td->td_oncpu; 2715288463Sjhb td->td_oncpu = NOCPU; 2716170293Sjeff } 2717170293Sjeff KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2718172411Sjeff newtd = choosethread(); 2719172411Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2720172411Sjeff cpu_throw(td, newtd); /* doesn't return */ 2721170293Sjeff} 2722170293Sjeff 2723171482Sjeff/* 2724171482Sjeff * This is called from fork_exit(). Just acquire the correct locks and 2725171482Sjeff * let fork do the rest of the work. 2726171482Sjeff */ 2727170293Sjeffvoid 2728170600Sjeffsched_fork_exit(struct thread *td) 2729170293Sjeff{ 2730171482Sjeff struct tdq *tdq; 2731171482Sjeff int cpuid; 2732170293Sjeff 2733170293Sjeff /* 2734170293Sjeff * Finish setting up thread glue so that it begins execution in a 2735171482Sjeff * non-nested critical section with the scheduler lock held. 2736170293Sjeff */ 2737171482Sjeff cpuid = PCPU_GET(cpuid); 2738171482Sjeff tdq = TDQ_CPU(cpuid); 2739171482Sjeff if (TD_IS_IDLETHREAD(td)) 2740171482Sjeff td->td_lock = TDQ_LOCKPTR(tdq); 2741171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2742171482Sjeff td->td_oncpu = cpuid; 2743172411Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2744174629Sjeff lock_profile_obtain_lock_success( 2745174629Sjeff &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2746315839Savg 2747315839Savg KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running", 2748315839Savg "prio:%d", td->td_priority); 2749315839Savg SDT_PROBE0(sched, , , on__cpu); 2750170293Sjeff} 2751170293Sjeff 2752187357Sjeff/* 2753187357Sjeff * Create on first use to catch odd startup conditons. 2754187357Sjeff */ 2755187357Sjeffchar * 2756187357Sjeffsched_tdname(struct thread *td) 2757187357Sjeff{ 2758187357Sjeff#ifdef KTR 2759187357Sjeff struct td_sched *ts; 2760187357Sjeff 2761187357Sjeff ts = td->td_sched; 2762187357Sjeff if (ts->ts_name[0] == '\0') 2763187357Sjeff snprintf(ts->ts_name, sizeof(ts->ts_name), 2764187357Sjeff "%s tid %d", td->td_name, td->td_tid); 2765187357Sjeff return (ts->ts_name); 2766187357Sjeff#else 2767187357Sjeff return (td->td_name); 2768187357Sjeff#endif 2769187357Sjeff} 2770187357Sjeff 2771232700Sjhb#ifdef KTR 2772232700Sjhbvoid 2773232700Sjhbsched_clear_tdname(struct thread *td) 2774232700Sjhb{ 2775232700Sjhb struct td_sched *ts; 2776232700Sjhb 2777232700Sjhb ts = td->td_sched; 2778232700Sjhb ts->ts_name[0] = '\0'; 2779232700Sjhb} 2780232700Sjhb#endif 2781232700Sjhb 2782184439Sivoras#ifdef SMP 2783184439Sivoras 2784184439Sivoras/* 2785184439Sivoras * Build the CPU topology dump string. Is recursively called to collect 2786184439Sivoras * the topology tree. 2787184439Sivoras */ 2788184439Sivorasstatic int 2789184439Sivorassysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg, 2790184439Sivoras int indent) 2791184439Sivoras{ 2792222813Sattilio char cpusetbuf[CPUSETBUFSIZ]; 2793184439Sivoras int i, first; 2794184439Sivoras 2795184439Sivoras sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent, 2796212821Savg "", 1 + indent / 2, cg->cg_level); 2797222813Sattilio sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "", 2798222813Sattilio cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask)); 2799184439Sivoras first = TRUE; 2800184439Sivoras for (i = 0; i < MAXCPU; i++) { 2801222813Sattilio if (CPU_ISSET(i, &cg->cg_mask)) { 2802184439Sivoras if (!first) 2803184439Sivoras sbuf_printf(sb, ", "); 2804184439Sivoras else 2805184439Sivoras first = FALSE; 2806184439Sivoras sbuf_printf(sb, "%d", i); 2807184439Sivoras } 2808184439Sivoras } 2809184439Sivoras sbuf_printf(sb, "</cpu>\n"); 2810184439Sivoras 2811184439Sivoras if (cg->cg_flags != 0) { 2812210117Sivoras sbuf_printf(sb, "%*s <flags>", indent, ""); 2813184439Sivoras if ((cg->cg_flags & CG_FLAG_HTT) != 0) 2814208982Sivoras sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>"); 2815208983Sivoras if ((cg->cg_flags & CG_FLAG_THREAD) != 0) 2816208983Sivoras sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>"); 2817191643Sjeff if ((cg->cg_flags & CG_FLAG_SMT) != 0) 2818208983Sivoras sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>"); 2819210117Sivoras sbuf_printf(sb, "</flags>\n"); 2820184439Sivoras } 2821184439Sivoras 2822184439Sivoras if (cg->cg_children > 0) { 2823184439Sivoras sbuf_printf(sb, "%*s <children>\n", indent, ""); 2824184439Sivoras for (i = 0; i < cg->cg_children; i++) 2825184439Sivoras sysctl_kern_sched_topology_spec_internal(sb, 2826184439Sivoras &cg->cg_child[i], indent+2); 2827184439Sivoras sbuf_printf(sb, "%*s </children>\n", indent, ""); 2828184439Sivoras } 2829184439Sivoras sbuf_printf(sb, "%*s</group>\n", indent, ""); 2830184439Sivoras return (0); 2831184439Sivoras} 2832184439Sivoras 2833184439Sivoras/* 2834184439Sivoras * Sysctl handler for retrieving topology dump. It's a wrapper for 2835184439Sivoras * the recursive sysctl_kern_smp_topology_spec_internal(). 2836184439Sivoras */ 2837184439Sivorasstatic int 2838184439Sivorassysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS) 2839184439Sivoras{ 2840184439Sivoras struct sbuf *topo; 2841184439Sivoras int err; 2842184439Sivoras 2843184439Sivoras KASSERT(cpu_top != NULL, ("cpu_top isn't initialized")); 2844184439Sivoras 2845184570Sivoras topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND); 2846184439Sivoras if (topo == NULL) 2847184439Sivoras return (ENOMEM); 2848184439Sivoras 2849184439Sivoras sbuf_printf(topo, "<groups>\n"); 2850184439Sivoras err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1); 2851184439Sivoras sbuf_printf(topo, "</groups>\n"); 2852184439Sivoras 2853184439Sivoras if (err == 0) { 2854184439Sivoras sbuf_finish(topo); 2855184439Sivoras err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo)); 2856184439Sivoras } 2857184439Sivoras sbuf_delete(topo); 2858184439Sivoras return (err); 2859184439Sivoras} 2860214510Sdavidxu 2861184439Sivoras#endif 2862184439Sivoras 2863239185Smavstatic int 2864239185Smavsysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 2865239185Smav{ 2866239185Smav int error, new_val, period; 2867239185Smav 2868239185Smav period = 1000000 / realstathz; 2869239185Smav new_val = period * sched_slice; 2870239185Smav error = sysctl_handle_int(oidp, &new_val, 0, req); 2871239196Smav if (error != 0 || req->newptr == NULL) 2872239185Smav return (error); 2873239185Smav if (new_val <= 0) 2874239185Smav return (EINVAL); 2875239196Smav sched_slice = imax(1, (new_val + period / 2) / period); 2876242736Sjeff sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR; 2877239196Smav hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 2878239196Smav realstathz); 2879239185Smav return (0); 2880239185Smav} 2881239185Smav 2882177435SjeffSYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2883171482SjeffSYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2884165762Sjeff "Scheduler name"); 2885239185SmavSYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW, 2886239185Smav NULL, 0, sysctl_kern_quantum, "I", 2887239196Smav "Quantum for timeshare threads in microseconds"); 2888171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2889239196Smav "Quantum for timeshare threads in stathz ticks"); 2890171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2891239196Smav "Interactivity score threshold"); 2892239196SmavSYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, 2893239196Smav &preempt_thresh, 0, 2894239196Smav "Maximal (lowest) priority for preemption"); 2895239196SmavSYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0, 2896239196Smav "Assign static kernel priorities to sleeping threads"); 2897239196SmavSYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0, 2898239196Smav "Number of times idle thread will spin waiting for new work"); 2899239196SmavSYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, 2900239196Smav &sched_idlespinthresh, 0, 2901239196Smav "Threshold before we will permit idle thread spinning"); 2902166108Sjeff#ifdef SMP 2903171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2904171482Sjeff "Number of hz ticks to keep thread affinity for"); 2905171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2906171482Sjeff "Enables the long-term load balancer"); 2907172409SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2908172409Sjeff &balance_interval, 0, 2909239185Smav "Average period in stathz ticks to run the long-term balancer"); 2910171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2911171482Sjeff "Attempts to steal work from other cores before idling"); 2912171506SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2913239196Smav "Minimum load on remote CPU before we'll steal"); 2914184439SivorasSYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING | 2915239185Smav CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A", 2916184439Sivoras "XML dump of detected CPU topology"); 2917166108Sjeff#endif 2918165762Sjeff 2919172264Sjeff/* ps compat. All cpu percentages from ULE are weighted. */ 2920172293Sjeffstatic int ccpu = 0; 2921165762SjeffSYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2922