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: releng/11.0/sys/kern/sched_ule.c 301456 2016-06-05 17:04:03Z kib $"); 40116182Sobrien 41147565Speter#include "opt_hwpmc_hooks.h" 42147565Speter#include "opt_sched.h" 43134649Sscottl 44109864Sjeff#include <sys/param.h> 45109864Sjeff#include <sys/systm.h> 46131929Smarcel#include <sys/kdb.h> 47109864Sjeff#include <sys/kernel.h> 48109864Sjeff#include <sys/ktr.h> 49263324Sattilio#include <sys/limits.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. */ 93273266Sadrian int 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 109176735Sjeff#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 110176735Sjeff#define THREAD_CAN_SCHED(td, cpu) \ 111176735Sjeff CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 112176735Sjeff 113301456Skib_Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <= 114301456Skib sizeof(struct thread0_storage), 115301456Skib "increase struct thread0_storage.t0st_sched size"); 116301456Skib 117109864Sjeff/* 118217351Sjhb * Priority ranges used for interactive and non-interactive timeshare 119217410Sjhb * threads. The timeshare priorities are split up into four ranges. 120217410Sjhb * The first range handles interactive threads. The last three ranges 121217410Sjhb * (NHALF, x, and NHALF) handle non-interactive threads with the outer 122217410Sjhb * ranges supporting nice values. 123217351Sjhb */ 124217410Sjhb#define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1) 125217410Sjhb#define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2) 126228718Savg#define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE) 127217410Sjhb 128217410Sjhb#define PRI_MIN_INTERACT PRI_MIN_TIMESHARE 129217410Sjhb#define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1) 130217410Sjhb#define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE) 131217351Sjhb#define PRI_MAX_BATCH PRI_MAX_TIMESHARE 132217351Sjhb 133217351Sjhb/* 134165762Sjeff * Cpu percentage computation macros and defines. 135111857Sjeff * 136165762Sjeff * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 137165762Sjeff * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 138165796Sjeff * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 139165762Sjeff * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 140165762Sjeff * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 141165762Sjeff * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 142165762Sjeff */ 143165762Sjeff#define SCHED_TICK_SECS 10 144165762Sjeff#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 145165796Sjeff#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 146165762Sjeff#define SCHED_TICK_SHIFT 10 147165762Sjeff#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 148165830Sjeff#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 149165762Sjeff 150165762Sjeff/* 151165762Sjeff * These macros determine priorities for non-interactive threads. They are 152165762Sjeff * assigned a priority based on their recent cpu utilization as expressed 153165762Sjeff * by the ratio of ticks to the tick total. NHALF priorities at the start 154165762Sjeff * and end of the MIN to MAX timeshare range are only reachable with negative 155165762Sjeff * or positive nice respectively. 156165762Sjeff * 157165762Sjeff * PRI_RANGE: Priority range for utilization dependent priorities. 158116642Sjeff * PRI_NRESV: Number of nice values. 159165762Sjeff * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 160165762Sjeff * PRI_NICE: Determines the part of the priority inherited from nice. 161109864Sjeff */ 162165762Sjeff#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 163121869Sjeff#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 164217351Sjhb#define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF) 165217351Sjhb#define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF) 166217237Sjhb#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1) 167165762Sjeff#define SCHED_PRI_TICKS(ts) \ 168165762Sjeff (SCHED_TICK_HZ((ts)) / \ 169165827Sjeff (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 170165762Sjeff#define SCHED_PRI_NICE(nice) (nice) 171109864Sjeff 172109864Sjeff/* 173165762Sjeff * These determine the interactivity of a process. Interactivity differs from 174165762Sjeff * cpu utilization in that it expresses the voluntary time slept vs time ran 175165762Sjeff * while cpu utilization includes all time not running. This more accurately 176165762Sjeff * models the intent of the thread. 177109864Sjeff * 178110645Sjeff * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 179110645Sjeff * before throttling back. 180121868Sjeff * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 181116365Sjeff * INTERACT_MAX: Maximum interactivity value. Smaller is better. 182215102Sattilio * INTERACT_THRESH: Threshold for placement on the current runq. 183109864Sjeff */ 184165762Sjeff#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 185165762Sjeff#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 186116365Sjeff#define SCHED_INTERACT_MAX (100) 187116365Sjeff#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 188121126Sjeff#define SCHED_INTERACT_THRESH (30) 189111857Sjeff 190242736Sjeff/* 191242736Sjeff * These parameters determine the slice behavior for batch work. 192242736Sjeff */ 193242736Sjeff#define SCHED_SLICE_DEFAULT_DIVISOR 10 /* ~94 ms, 12 stathz ticks. */ 194242736Sjeff#define SCHED_SLICE_MIN_DIVISOR 6 /* DEFAULT/MIN = ~16 ms. */ 195242736Sjeff 196239157Smav/* Flags kept in td_flags. */ 197239157Smav#define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */ 198239157Smav 199109864Sjeff/* 200165762Sjeff * tickincr: Converts a stathz tick into a hz domain scaled by 201165762Sjeff * the shift factor. Without the shift the error rate 202165762Sjeff * due to rounding would be unacceptably high. 203165762Sjeff * realstathz: stathz is sometimes 0 and run off of hz. 204165762Sjeff * sched_slice: Runtime of each thread before rescheduling. 205171482Sjeff * preempt_thresh: Priority threshold for preemption and remote IPIs. 206109864Sjeff */ 207165762Sjeffstatic int sched_interact = SCHED_INTERACT_THRESH; 208241844Seadlerstatic int tickincr = 8 << SCHED_TICK_SHIFT; 209242736Sjeffstatic int realstathz = 127; /* reset during boot. */ 210242736Sjeffstatic int sched_slice = 10; /* reset during boot. */ 211242736Sjeffstatic int sched_slice_min = 1; /* reset during boot. */ 212172345Sjeff#ifdef PREEMPTION 213172345Sjeff#ifdef FULL_PREEMPTION 214172345Sjeffstatic int preempt_thresh = PRI_MAX_IDLE; 215172345Sjeff#else 216171482Sjeffstatic int preempt_thresh = PRI_MIN_KERN; 217172345Sjeff#endif 218172345Sjeff#else 219172345Sjeffstatic int preempt_thresh = 0; 220172345Sjeff#endif 221217351Sjhbstatic int static_boost = PRI_MIN_BATCH; 222178277Sjeffstatic int sched_idlespins = 10000; 223232740Smavstatic int sched_idlespinthresh = -1; 224109864Sjeff 225109864Sjeff/* 226171482Sjeff * tdq - per processor runqs and statistics. All fields are protected by the 227171482Sjeff * tdq_lock. The load and lowpri may be accessed without to avoid excess 228171482Sjeff * locking in sched_pickcpu(); 229109864Sjeff */ 230164936Sjulianstruct tdq { 231242014Sjimharris /* 232242014Sjimharris * Ordered to improve efficiency of cpu_search() and switch(). 233242014Sjimharris * tdq_lock is padded to avoid false sharing with tdq_load and 234242014Sjimharris * tdq_cpu_idle. 235242014Sjimharris */ 236242402Sattilio struct mtx_padalign tdq_lock; /* run queue lock. */ 237176735Sjeff struct cpu_group *tdq_cg; /* Pointer to cpu topology. */ 238178277Sjeff volatile int tdq_load; /* Aggregate load. */ 239212416Smav volatile int tdq_cpu_idle; /* cpu_idle() is active. */ 240176735Sjeff int tdq_sysload; /* For loadavg, !ITHD load. */ 241177009Sjeff int tdq_transferable; /* Transferable thread count. */ 242178277Sjeff short tdq_switchcnt; /* Switches this tick. */ 243178277Sjeff short tdq_oldswitchcnt; /* Switches last tick. */ 244177009Sjeff u_char tdq_lowpri; /* Lowest priority thread. */ 245177009Sjeff u_char tdq_ipipending; /* IPI pending. */ 246166557Sjeff u_char tdq_idx; /* Current insert index. */ 247166557Sjeff u_char tdq_ridx; /* Current removal index. */ 248177009Sjeff struct runq tdq_realtime; /* real-time run queue. */ 249177009Sjeff struct runq tdq_timeshare; /* timeshare run queue. */ 250177009Sjeff struct runq tdq_idle; /* Queue of IDLE threads. */ 251187357Sjeff char tdq_name[TDQ_NAME_LEN]; 252187357Sjeff#ifdef KTR 253187357Sjeff char tdq_loadname[TDQ_LOADNAME_LEN]; 254187357Sjeff#endif 255171482Sjeff} __aligned(64); 256109864Sjeff 257178277Sjeff/* Idle thread states and config. */ 258178277Sjeff#define TDQ_RUNNING 1 259178277Sjeff#define TDQ_IDLE 2 260166108Sjeff 261123433Sjeff#ifdef SMP 262184439Sivorasstruct cpu_group *cpu_top; /* CPU topology */ 263123433Sjeff 264176735Sjeff#define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000)) 265176735Sjeff#define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity)) 266166108Sjeff 267123433Sjeff/* 268166108Sjeff * Run-time tunables. 269166108Sjeff */ 270171506Sjeffstatic int rebalance = 1; 271172409Sjeffstatic int balance_interval = 128; /* Default set in sched_initticks(). */ 272166108Sjeffstatic int affinity; 273171506Sjeffstatic int steal_idle = 1; 274171506Sjeffstatic int steal_thresh = 2; 275166108Sjeff 276166108Sjeff/* 277165620Sjeff * One thread queue per processor. 278109864Sjeff */ 279164936Sjulianstatic struct tdq tdq_cpu[MAXCPU]; 280172409Sjeffstatic struct tdq *balance_tdq; 281172409Sjeffstatic int balance_ticks; 282232207Smavstatic DPCPU_DEFINE(uint32_t, randomval); 283129982Sjeff 284164936Sjulian#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 285164936Sjulian#define TDQ_CPU(x) (&tdq_cpu[(x)]) 286171713Sjeff#define TDQ_ID(x) ((int)((x) - tdq_cpu)) 287123433Sjeff#else /* !SMP */ 288164936Sjulianstatic struct tdq tdq_cpu; 289129982Sjeff 290170315Sjeff#define TDQ_ID(x) (0) 291164936Sjulian#define TDQ_SELF() (&tdq_cpu) 292164936Sjulian#define TDQ_CPU(x) (&tdq_cpu) 293110028Sjeff#endif 294109864Sjeff 295171482Sjeff#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 296171482Sjeff#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 297171482Sjeff#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 298171482Sjeff#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 299242402Sattilio#define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock)) 300171482Sjeff 301163709Sjbstatic void sched_priority(struct thread *); 302146954Sjeffstatic void sched_thread_priority(struct thread *, u_char); 303163709Sjbstatic int sched_interact_score(struct thread *); 304163709Sjbstatic void sched_interact_update(struct thread *); 305163709Sjbstatic void sched_interact_fork(struct thread *); 306232917Smavstatic void sched_pctcpu_update(struct td_sched *, int); 307109864Sjeff 308110267Sjeff/* Operations on per processor queues */ 309177435Sjeffstatic struct thread *tdq_choose(struct tdq *); 310164936Sjulianstatic void tdq_setup(struct tdq *); 311177435Sjeffstatic void tdq_load_add(struct tdq *, struct thread *); 312177435Sjeffstatic void tdq_load_rem(struct tdq *, struct thread *); 313177435Sjeffstatic __inline void tdq_runq_add(struct tdq *, struct thread *, int); 314177435Sjeffstatic __inline void tdq_runq_rem(struct tdq *, struct thread *); 315177005Sjeffstatic inline int sched_shouldpreempt(int, int, int); 316164936Sjulianvoid tdq_print(int cpu); 317165762Sjeffstatic void runq_print(struct runq *rq); 318171482Sjeffstatic void tdq_add(struct tdq *, struct thread *, int); 319110267Sjeff#ifdef SMP 320176735Sjeffstatic int tdq_move(struct tdq *, struct tdq *); 321171482Sjeffstatic int tdq_idled(struct tdq *); 322177435Sjeffstatic void tdq_notify(struct tdq *, struct thread *); 323177435Sjeffstatic struct thread *tdq_steal(struct tdq *, int); 324177435Sjeffstatic struct thread *runq_steal(struct runq *, int); 325177435Sjeffstatic int sched_pickcpu(struct thread *, int); 326172409Sjeffstatic void sched_balance(void); 327176735Sjeffstatic int sched_balance_pair(struct tdq *, struct tdq *); 328177435Sjeffstatic inline struct tdq *sched_setcpu(struct thread *, int, int); 329171482Sjeffstatic inline void thread_unblock_switch(struct thread *, struct mtx *); 330171713Sjeffstatic struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 331184439Sivorasstatic int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS); 332184439Sivorasstatic int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, 333184439Sivoras struct cpu_group *cg, int indent); 334121790Sjeff#endif 335110028Sjeff 336165762Sjeffstatic void sched_setup(void *dummy); 337177253SrwatsonSYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 338165762Sjeff 339165762Sjeffstatic void sched_initticks(void *dummy); 340177253SrwatsonSYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 341177253Srwatson NULL); 342165762Sjeff 343235459SrstoneSDT_PROVIDER_DEFINE(sched); 344235459Srstone 345258622SavgSDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *", 346235459Srstone "struct proc *", "uint8_t"); 347258622SavgSDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *", 348235459Srstone "struct proc *", "void *"); 349258622SavgSDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *", 350235459Srstone "struct proc *", "void *", "int"); 351258622SavgSDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *", 352235459Srstone "struct proc *", "uint8_t", "struct thread *"); 353258622SavgSDT_PROBE_DEFINE2(sched, , , load__change, "int", "int"); 354258622SavgSDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *", 355235459Srstone "struct proc *"); 356258622SavgSDT_PROBE_DEFINE(sched, , , on__cpu); 357258622SavgSDT_PROBE_DEFINE(sched, , , remain__cpu); 358258622SavgSDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *", 359235459Srstone "struct proc *"); 360235459Srstone 361171482Sjeff/* 362171482Sjeff * Print the threads waiting on a run-queue. 363171482Sjeff */ 364165762Sjeffstatic void 365165762Sjeffrunq_print(struct runq *rq) 366165762Sjeff{ 367165762Sjeff struct rqhead *rqh; 368177435Sjeff struct thread *td; 369165762Sjeff int pri; 370165762Sjeff int j; 371165762Sjeff int i; 372165762Sjeff 373165762Sjeff for (i = 0; i < RQB_LEN; i++) { 374165762Sjeff printf("\t\trunq bits %d 0x%zx\n", 375165762Sjeff i, rq->rq_status.rqb_bits[i]); 376165762Sjeff for (j = 0; j < RQB_BPW; j++) 377165762Sjeff if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 378165762Sjeff pri = j + (i << RQB_L2BPW); 379165762Sjeff rqh = &rq->rq_queues[pri]; 380177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) { 381165762Sjeff printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 382177435Sjeff td, td->td_name, td->td_priority, 383177435Sjeff td->td_rqindex, pri); 384165762Sjeff } 385165762Sjeff } 386165762Sjeff } 387165762Sjeff} 388165762Sjeff 389171482Sjeff/* 390171482Sjeff * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 391171482Sjeff */ 392113357Sjeffvoid 393164936Sjuliantdq_print(int cpu) 394110267Sjeff{ 395164936Sjulian struct tdq *tdq; 396112994Sjeff 397164936Sjulian tdq = TDQ_CPU(cpu); 398112994Sjeff 399171713Sjeff printf("tdq %d:\n", TDQ_ID(tdq)); 400176735Sjeff printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 401176735Sjeff printf("\tLock name: %s\n", tdq->tdq_name); 402165620Sjeff printf("\tload: %d\n", tdq->tdq_load); 403178277Sjeff printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt); 404178277Sjeff printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt); 405171482Sjeff printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 406165766Sjeff printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 407178277Sjeff printf("\tload transferable: %d\n", tdq->tdq_transferable); 408178277Sjeff printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 409165762Sjeff printf("\trealtime runq:\n"); 410165762Sjeff runq_print(&tdq->tdq_realtime); 411165762Sjeff printf("\ttimeshare runq:\n"); 412165762Sjeff runq_print(&tdq->tdq_timeshare); 413165762Sjeff printf("\tidle runq:\n"); 414165762Sjeff runq_print(&tdq->tdq_idle); 415113357Sjeff} 416112994Sjeff 417177005Sjeffstatic inline int 418177005Sjeffsched_shouldpreempt(int pri, int cpri, int remote) 419177005Sjeff{ 420177005Sjeff /* 421177005Sjeff * If the new priority is not better than the current priority there is 422177005Sjeff * nothing to do. 423177005Sjeff */ 424177005Sjeff if (pri >= cpri) 425177005Sjeff return (0); 426177005Sjeff /* 427177005Sjeff * Always preempt idle. 428177005Sjeff */ 429177005Sjeff if (cpri >= PRI_MIN_IDLE) 430177005Sjeff return (1); 431177005Sjeff /* 432177005Sjeff * If preemption is disabled don't preempt others. 433177005Sjeff */ 434177005Sjeff if (preempt_thresh == 0) 435177005Sjeff return (0); 436177005Sjeff /* 437177005Sjeff * Preempt if we exceed the threshold. 438177005Sjeff */ 439177005Sjeff if (pri <= preempt_thresh) 440177005Sjeff return (1); 441177005Sjeff /* 442217351Sjhb * If we're interactive or better and there is non-interactive 443217351Sjhb * or worse running preempt only remote processors. 444177005Sjeff */ 445217351Sjhb if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT) 446177005Sjeff return (1); 447177005Sjeff return (0); 448177005Sjeff} 449177005Sjeff 450171482Sjeff/* 451171482Sjeff * Add a thread to the actual run-queue. Keeps transferable counts up to 452171482Sjeff * date with what is actually on the run-queue. Selects the correct 453171482Sjeff * queue position for timeshare threads. 454171482Sjeff */ 455122744Sjeffstatic __inline void 456177435Sjefftdq_runq_add(struct tdq *tdq, struct thread *td, int flags) 457122744Sjeff{ 458177435Sjeff struct td_sched *ts; 459177042Sjeff u_char pri; 460177042Sjeff 461171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 462177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 463177009Sjeff 464177435Sjeff pri = td->td_priority; 465301456Skib ts = td_get_sched(td); 466177435Sjeff TD_SET_RUNQ(td); 467177435Sjeff if (THREAD_CAN_MIGRATE(td)) { 468165620Sjeff tdq->tdq_transferable++; 469164936Sjulian ts->ts_flags |= TSF_XFERABLE; 470123433Sjeff } 471217351Sjhb if (pri < PRI_MIN_BATCH) { 472177042Sjeff ts->ts_runq = &tdq->tdq_realtime; 473217351Sjhb } else if (pri <= PRI_MAX_BATCH) { 474177042Sjeff ts->ts_runq = &tdq->tdq_timeshare; 475217351Sjhb KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH, 476165762Sjeff ("Invalid priority %d on timeshare runq", pri)); 477165762Sjeff /* 478165762Sjeff * This queue contains only priorities between MIN and MAX 479165762Sjeff * realtime. Use the whole queue to represent these values. 480165762Sjeff */ 481171713Sjeff if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 482228718Savg pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE; 483165762Sjeff pri = (pri + tdq->tdq_idx) % RQ_NQS; 484165766Sjeff /* 485165766Sjeff * This effectively shortens the queue by one so we 486165766Sjeff * can have a one slot difference between idx and 487165766Sjeff * ridx while we wait for threads to drain. 488165766Sjeff */ 489165766Sjeff if (tdq->tdq_ridx != tdq->tdq_idx && 490165766Sjeff pri == tdq->tdq_ridx) 491167664Sjeff pri = (unsigned char)(pri - 1) % RQ_NQS; 492165762Sjeff } else 493165766Sjeff pri = tdq->tdq_ridx; 494177435Sjeff runq_add_pri(ts->ts_runq, td, pri, flags); 495177042Sjeff return; 496165762Sjeff } else 497177009Sjeff ts->ts_runq = &tdq->tdq_idle; 498177435Sjeff runq_add(ts->ts_runq, td, flags); 499177009Sjeff} 500177009Sjeff 501171482Sjeff/* 502171482Sjeff * Remove a thread from a run-queue. This typically happens when a thread 503171482Sjeff * is selected to run. Running threads are not on the queue and the 504171482Sjeff * transferable count does not reflect them. 505171482Sjeff */ 506122744Sjeffstatic __inline void 507177435Sjefftdq_runq_rem(struct tdq *tdq, struct thread *td) 508122744Sjeff{ 509177435Sjeff struct td_sched *ts; 510177435Sjeff 511301456Skib ts = td_get_sched(td); 512171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 513171482Sjeff KASSERT(ts->ts_runq != NULL, 514177435Sjeff ("tdq_runq_remove: thread %p null ts_runq", td)); 515164936Sjulian if (ts->ts_flags & TSF_XFERABLE) { 516165620Sjeff tdq->tdq_transferable--; 517164936Sjulian ts->ts_flags &= ~TSF_XFERABLE; 518123433Sjeff } 519165766Sjeff if (ts->ts_runq == &tdq->tdq_timeshare) { 520165766Sjeff if (tdq->tdq_idx != tdq->tdq_ridx) 521177435Sjeff runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx); 522165766Sjeff else 523177435Sjeff runq_remove_idx(ts->ts_runq, td, NULL); 524165766Sjeff } else 525177435Sjeff runq_remove(ts->ts_runq, td); 526122744Sjeff} 527122744Sjeff 528171482Sjeff/* 529171482Sjeff * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 530171482Sjeff * for this thread to the referenced thread queue. 531171482Sjeff */ 532113357Sjeffstatic void 533177435Sjefftdq_load_add(struct tdq *tdq, struct thread *td) 534113357Sjeff{ 535171482Sjeff 536171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 537177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 538177902Sjeff 539165620Sjeff tdq->tdq_load++; 540198854Sattilio if ((td->td_flags & TDF_NOLOAD) == 0) 541177902Sjeff tdq->tdq_sysload++; 542187357Sjeff KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 543258622Savg SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load); 544110267Sjeff} 545113357Sjeff 546171482Sjeff/* 547171482Sjeff * Remove the load from a thread that is transitioning to a sleep state or 548171482Sjeff * exiting. 549171482Sjeff */ 550112994Sjeffstatic void 551177435Sjefftdq_load_rem(struct tdq *tdq, struct thread *td) 552110267Sjeff{ 553171482Sjeff 554177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 555171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 556171482Sjeff KASSERT(tdq->tdq_load != 0, 557171713Sjeff ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 558177902Sjeff 559165620Sjeff tdq->tdq_load--; 560198854Sattilio if ((td->td_flags & TDF_NOLOAD) == 0) 561177902Sjeff tdq->tdq_sysload--; 562187357Sjeff KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 563258622Savg SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load); 564110267Sjeff} 565110267Sjeff 566176735Sjeff/* 567242736Sjeff * Bound timeshare latency by decreasing slice size as load increases. We 568242736Sjeff * consider the maximum latency as the sum of the threads waiting to run 569242736Sjeff * aside from curthread and target no more than sched_slice latency but 570242736Sjeff * no less than sched_slice_min runtime. 571242736Sjeff */ 572242736Sjeffstatic inline int 573242736Sjefftdq_slice(struct tdq *tdq) 574242736Sjeff{ 575242736Sjeff int load; 576242736Sjeff 577242736Sjeff /* 578242736Sjeff * It is safe to use sys_load here because this is called from 579242736Sjeff * contexts where timeshare threads are running and so there 580242736Sjeff * cannot be higher priority load in the system. 581242736Sjeff */ 582242736Sjeff load = tdq->tdq_sysload - 1; 583242736Sjeff if (load >= SCHED_SLICE_MIN_DIVISOR) 584242736Sjeff return (sched_slice_min); 585242736Sjeff if (load <= 1) 586242736Sjeff return (sched_slice); 587242736Sjeff return (sched_slice / load); 588242736Sjeff} 589242736Sjeff 590242736Sjeff/* 591176735Sjeff * Set lowpri to its exact value by searching the run-queue and 592176735Sjeff * evaluating curthread. curthread may be passed as an optimization. 593176735Sjeff */ 594176735Sjeffstatic void 595176735Sjefftdq_setlowpri(struct tdq *tdq, struct thread *ctd) 596176735Sjeff{ 597176735Sjeff struct thread *td; 598176735Sjeff 599176735Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 600176735Sjeff if (ctd == NULL) 601176735Sjeff ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 602177435Sjeff td = tdq_choose(tdq); 603177435Sjeff if (td == NULL || td->td_priority > ctd->td_priority) 604176735Sjeff tdq->tdq_lowpri = ctd->td_priority; 605176735Sjeff else 606176735Sjeff tdq->tdq_lowpri = td->td_priority; 607176735Sjeff} 608176735Sjeff 609113357Sjeff#ifdef SMP 610285243Spfg/* 611285243Spfg * We need some randomness. Implement a classic Linear Congruential 612285243Spfg * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for 613285243Spfg * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits 614285243Spfg * of the random state (in the low bits of our answer) to keep 615285243Spfg * the maximum randomness. 616285243Spfg */ 617285243Spfgstatic uint32_t 618285243Spfgsched_random(void) 619285243Spfg{ 620285243Spfg uint32_t *rndptr; 621285243Spfg 622285243Spfg rndptr = DPCPU_PTR(randomval); 623285243Spfg *rndptr = *rndptr * 69069 + 5; 624285243Spfg 625285243Spfg return (*rndptr >> 16); 626285243Spfg} 627285243Spfg 628176735Sjeffstruct cpu_search { 629194779Sjeff cpuset_t cs_mask; 630232207Smav u_int cs_prefer; 631232207Smav int cs_pri; /* Min priority for low. */ 632232207Smav int cs_limit; /* Max load for low, min load for high. */ 633232207Smav int cs_cpu; 634232207Smav int cs_load; 635176735Sjeff}; 636176735Sjeff 637176735Sjeff#define CPU_SEARCH_LOWEST 0x1 638176735Sjeff#define CPU_SEARCH_HIGHEST 0x2 639176735Sjeff#define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 640176735Sjeff 641194779Sjeff#define CPUSET_FOREACH(cpu, mask) \ 642194779Sjeff for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \ 643222813Sattilio if (CPU_ISSET(cpu, &mask)) 644176735Sjeff 645268211Skibstatic __always_inline int cpu_search(const struct cpu_group *cg, 646268211Skib struct cpu_search *low, struct cpu_search *high, const int match); 647268211Skibint __noinline cpu_search_lowest(const struct cpu_group *cg, 648268211Skib struct cpu_search *low); 649268211Skibint __noinline cpu_search_highest(const struct cpu_group *cg, 650176735Sjeff struct cpu_search *high); 651268211Skibint __noinline cpu_search_both(const struct cpu_group *cg, 652268211Skib struct cpu_search *low, struct cpu_search *high); 653176735Sjeff 654116069Sjeff/* 655176735Sjeff * Search the tree of cpu_groups for the lowest or highest loaded cpu 656176735Sjeff * according to the match argument. This routine actually compares the 657176735Sjeff * load on all paths through the tree and finds the least loaded cpu on 658176735Sjeff * the least loaded path, which may differ from the least loaded cpu in 659176735Sjeff * the system. This balances work among caches and busses. 660116069Sjeff * 661176735Sjeff * This inline is instantiated in three forms below using constants for the 662176735Sjeff * match argument. It is reduced to the minimum set for each case. It is 663176735Sjeff * also recursive to the depth of the tree. 664116069Sjeff */ 665268211Skibstatic __always_inline int 666232207Smavcpu_search(const struct cpu_group *cg, struct cpu_search *low, 667176735Sjeff struct cpu_search *high, const int match) 668176735Sjeff{ 669232207Smav struct cpu_search lgroup; 670232207Smav struct cpu_search hgroup; 671232207Smav cpuset_t cpumask; 672232207Smav struct cpu_group *child; 673232207Smav struct tdq *tdq; 674279349Simp int cpu, i, hload, lload, load, total, rnd; 675176735Sjeff 676176735Sjeff total = 0; 677232207Smav cpumask = cg->cg_mask; 678232207Smav if (match & CPU_SEARCH_LOWEST) { 679232207Smav lload = INT_MAX; 680232207Smav lgroup = *low; 681232207Smav } 682232207Smav if (match & CPU_SEARCH_HIGHEST) { 683234066Smav hload = INT_MIN; 684232207Smav hgroup = *high; 685232207Smav } 686176735Sjeff 687232207Smav /* Iterate through the child CPU groups and then remaining CPUs. */ 688255363Smav for (i = cg->cg_children, cpu = mp_maxid; ; ) { 689234066Smav if (i == 0) { 690255363Smav#ifdef HAVE_INLINE_FFSL 691255363Smav cpu = CPU_FFS(&cpumask) - 1; 692255363Smav#else 693234066Smav while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask)) 694234066Smav cpu--; 695255363Smav#endif 696234066Smav if (cpu < 0) 697232207Smav break; 698232207Smav child = NULL; 699232207Smav } else 700234066Smav child = &cg->cg_child[i - 1]; 701232207Smav 702234066Smav if (match & CPU_SEARCH_LOWEST) 703234066Smav lgroup.cs_cpu = -1; 704234066Smav if (match & CPU_SEARCH_HIGHEST) 705234066Smav hgroup.cs_cpu = -1; 706232207Smav if (child) { /* Handle child CPU group. */ 707232207Smav CPU_NAND(&cpumask, &child->cg_mask); 708176735Sjeff switch (match) { 709176735Sjeff case CPU_SEARCH_LOWEST: 710176735Sjeff load = cpu_search_lowest(child, &lgroup); 711176735Sjeff break; 712176735Sjeff case CPU_SEARCH_HIGHEST: 713176735Sjeff load = cpu_search_highest(child, &hgroup); 714176735Sjeff break; 715176735Sjeff case CPU_SEARCH_BOTH: 716176735Sjeff load = cpu_search_both(child, &lgroup, &hgroup); 717176735Sjeff break; 718176735Sjeff } 719232207Smav } else { /* Handle child CPU. */ 720255363Smav CPU_CLR(cpu, &cpumask); 721232207Smav tdq = TDQ_CPU(cpu); 722232207Smav load = tdq->tdq_load * 256; 723279373Simp rnd = sched_random() % 32; 724232207Smav if (match & CPU_SEARCH_LOWEST) { 725232207Smav if (cpu == low->cs_prefer) 726232207Smav load -= 64; 727232207Smav /* If that CPU is allowed and get data. */ 728234066Smav if (tdq->tdq_lowpri > lgroup.cs_pri && 729234066Smav tdq->tdq_load <= lgroup.cs_limit && 730234066Smav CPU_ISSET(cpu, &lgroup.cs_mask)) { 731232207Smav lgroup.cs_cpu = cpu; 732232207Smav lgroup.cs_load = load - rnd; 733176735Sjeff } 734232207Smav } 735232207Smav if (match & CPU_SEARCH_HIGHEST) 736234066Smav if (tdq->tdq_load >= hgroup.cs_limit && 737234066Smav tdq->tdq_transferable && 738234066Smav CPU_ISSET(cpu, &hgroup.cs_mask)) { 739232207Smav hgroup.cs_cpu = cpu; 740232207Smav hgroup.cs_load = load - rnd; 741176735Sjeff } 742176735Sjeff } 743232207Smav total += load; 744176735Sjeff 745232207Smav /* We have info about child item. Compare it. */ 746232207Smav if (match & CPU_SEARCH_LOWEST) { 747234066Smav if (lgroup.cs_cpu >= 0 && 748232454Smav (load < lload || 749232454Smav (load == lload && lgroup.cs_load < low->cs_load))) { 750232207Smav lload = load; 751232207Smav low->cs_cpu = lgroup.cs_cpu; 752232207Smav low->cs_load = lgroup.cs_load; 753232207Smav } 754232207Smav } 755232207Smav if (match & CPU_SEARCH_HIGHEST) 756234066Smav if (hgroup.cs_cpu >= 0 && 757232454Smav (load > hload || 758232454Smav (load == hload && hgroup.cs_load > high->cs_load))) { 759232207Smav hload = load; 760232207Smav high->cs_cpu = hgroup.cs_cpu; 761232207Smav high->cs_load = hgroup.cs_load; 762232207Smav } 763234066Smav if (child) { 764234066Smav i--; 765234066Smav if (i == 0 && CPU_EMPTY(&cpumask)) 766234066Smav break; 767255363Smav } 768255363Smav#ifndef HAVE_INLINE_FFSL 769255363Smav else 770234066Smav cpu--; 771255363Smav#endif 772176735Sjeff } 773176735Sjeff return (total); 774176735Sjeff} 775176735Sjeff 776176735Sjeff/* 777176735Sjeff * cpu_search instantiations must pass constants to maintain the inline 778176735Sjeff * optimization. 779176735Sjeff */ 780176735Sjeffint 781232207Smavcpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low) 782176735Sjeff{ 783176735Sjeff return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 784176735Sjeff} 785176735Sjeff 786176735Sjeffint 787232207Smavcpu_search_highest(const struct cpu_group *cg, struct cpu_search *high) 788176735Sjeff{ 789176735Sjeff return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 790176735Sjeff} 791176735Sjeff 792176735Sjeffint 793232207Smavcpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 794176735Sjeff struct cpu_search *high) 795176735Sjeff{ 796176735Sjeff return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 797176735Sjeff} 798176735Sjeff 799176735Sjeff/* 800176735Sjeff * Find the cpu with the least load via the least loaded path that has a 801176735Sjeff * lowpri greater than pri pri. A pri of -1 indicates any priority is 802176735Sjeff * acceptable. 803176735Sjeff */ 804176735Sjeffstatic inline int 805232207Smavsched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload, 806232207Smav int prefer) 807176735Sjeff{ 808176735Sjeff struct cpu_search low; 809176735Sjeff 810176735Sjeff low.cs_cpu = -1; 811232207Smav low.cs_prefer = prefer; 812176735Sjeff low.cs_mask = mask; 813232207Smav low.cs_pri = pri; 814232207Smav low.cs_limit = maxload; 815176735Sjeff cpu_search_lowest(cg, &low); 816176735Sjeff return low.cs_cpu; 817176735Sjeff} 818176735Sjeff 819176735Sjeff/* 820176735Sjeff * Find the cpu with the highest load via the highest loaded path. 821176735Sjeff */ 822176735Sjeffstatic inline int 823232207Smavsched_highest(const struct cpu_group *cg, cpuset_t mask, int minload) 824176735Sjeff{ 825176735Sjeff struct cpu_search high; 826176735Sjeff 827176735Sjeff high.cs_cpu = -1; 828176735Sjeff high.cs_mask = mask; 829176735Sjeff high.cs_limit = minload; 830176735Sjeff cpu_search_highest(cg, &high); 831176735Sjeff return high.cs_cpu; 832176735Sjeff} 833176735Sjeff 834121790Sjeffstatic void 835176735Sjeffsched_balance_group(struct cpu_group *cg) 836116069Sjeff{ 837232207Smav cpuset_t hmask, lmask; 838232207Smav int high, low, anylow; 839123487Sjeff 840232207Smav CPU_FILL(&hmask); 841176735Sjeff for (;;) { 842232207Smav high = sched_highest(cg, hmask, 1); 843232207Smav /* Stop if there is no more CPU with transferrable threads. */ 844232207Smav if (high == -1) 845176735Sjeff break; 846232207Smav CPU_CLR(high, &hmask); 847232207Smav CPU_COPY(&hmask, &lmask); 848232207Smav /* Stop if there is no more CPU left for low. */ 849232207Smav if (CPU_EMPTY(&lmask)) 850176735Sjeff break; 851232207Smav anylow = 1; 852232207Smavnextlow: 853232207Smav low = sched_lowest(cg, lmask, -1, 854232207Smav TDQ_CPU(high)->tdq_load - 1, high); 855232207Smav /* Stop if we looked well and found no less loaded CPU. */ 856232207Smav if (anylow && low == -1) 857232207Smav break; 858232207Smav /* Go to next high if we found no less loaded CPU. */ 859232207Smav if (low == -1) 860232207Smav continue; 861232207Smav /* Transfer thread from high to low. */ 862232207Smav if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) { 863232207Smav /* CPU that got thread can no longer be a donor. */ 864232207Smav CPU_CLR(low, &hmask); 865232207Smav } else { 866232207Smav /* 867232207Smav * If failed, then there is no threads on high 868232207Smav * that can run on this low. Drop low from low 869232207Smav * mask and look for different one. 870232207Smav */ 871232207Smav CPU_CLR(low, &lmask); 872232207Smav anylow = 0; 873232207Smav goto nextlow; 874232207Smav } 875123487Sjeff } 876123487Sjeff} 877123487Sjeff 878123487Sjeffstatic void 879201148Sedsched_balance(void) 880123487Sjeff{ 881172409Sjeff struct tdq *tdq; 882123487Sjeff 883171482Sjeff if (smp_started == 0 || rebalance == 0) 884171482Sjeff return; 885279349Simp 886279349Simp balance_ticks = max(balance_interval / 2, 1) + 887279373Simp (sched_random() % balance_interval); 888172409Sjeff tdq = TDQ_SELF(); 889172409Sjeff TDQ_UNLOCK(tdq); 890176735Sjeff sched_balance_group(cpu_top); 891172409Sjeff TDQ_LOCK(tdq); 892123487Sjeff} 893123487Sjeff 894171482Sjeff/* 895171482Sjeff * Lock two thread queues using their address to maintain lock order. 896171482Sjeff */ 897123487Sjeffstatic void 898171482Sjefftdq_lock_pair(struct tdq *one, struct tdq *two) 899171482Sjeff{ 900171482Sjeff if (one < two) { 901171482Sjeff TDQ_LOCK(one); 902171482Sjeff TDQ_LOCK_FLAGS(two, MTX_DUPOK); 903171482Sjeff } else { 904171482Sjeff TDQ_LOCK(two); 905171482Sjeff TDQ_LOCK_FLAGS(one, MTX_DUPOK); 906171482Sjeff } 907171482Sjeff} 908171482Sjeff 909171482Sjeff/* 910172409Sjeff * Unlock two thread queues. Order is not important here. 911172409Sjeff */ 912172409Sjeffstatic void 913172409Sjefftdq_unlock_pair(struct tdq *one, struct tdq *two) 914172409Sjeff{ 915172409Sjeff TDQ_UNLOCK(one); 916172409Sjeff TDQ_UNLOCK(two); 917172409Sjeff} 918172409Sjeff 919172409Sjeff/* 920171482Sjeff * Transfer load between two imbalanced thread queues. 921171482Sjeff */ 922176735Sjeffstatic int 923164936Sjuliansched_balance_pair(struct tdq *high, struct tdq *low) 924123487Sjeff{ 925176735Sjeff int moved; 926226057Smarius int cpu; 927116069Sjeff 928171482Sjeff tdq_lock_pair(high, low); 929176735Sjeff moved = 0; 930116069Sjeff /* 931122744Sjeff * Determine what the imbalance is and then adjust that to how many 932165620Sjeff * threads we actually have to give up (transferable). 933122744Sjeff */ 934232207Smav if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load && 935232207Smav (moved = tdq_move(high, low)) > 0) { 936172293Sjeff /* 937226057Smarius * In case the target isn't the current cpu IPI it to force a 938226057Smarius * reschedule with the new workload. 939172293Sjeff */ 940226057Smarius cpu = TDQ_ID(low); 941226057Smarius if (cpu != PCPU_GET(cpuid)) 942226057Smarius ipi_cpu(cpu, IPI_PREEMPT); 943171482Sjeff } 944172409Sjeff tdq_unlock_pair(high, low); 945176735Sjeff return (moved); 946116069Sjeff} 947116069Sjeff 948171482Sjeff/* 949171482Sjeff * Move a thread from one thread queue to another. 950171482Sjeff */ 951176735Sjeffstatic int 952171482Sjefftdq_move(struct tdq *from, struct tdq *to) 953116069Sjeff{ 954171482Sjeff struct td_sched *ts; 955171482Sjeff struct thread *td; 956164936Sjulian struct tdq *tdq; 957171482Sjeff int cpu; 958116069Sjeff 959172409Sjeff TDQ_LOCK_ASSERT(from, MA_OWNED); 960172409Sjeff TDQ_LOCK_ASSERT(to, MA_OWNED); 961172409Sjeff 962164936Sjulian tdq = from; 963171482Sjeff cpu = TDQ_ID(to); 964177435Sjeff td = tdq_steal(tdq, cpu); 965177435Sjeff if (td == NULL) 966176735Sjeff return (0); 967301456Skib ts = td_get_sched(td); 968171482Sjeff /* 969171482Sjeff * Although the run queue is locked the thread may be blocked. Lock 970172409Sjeff * it to clear this and acquire the run-queue lock. 971171482Sjeff */ 972171482Sjeff thread_lock(td); 973172409Sjeff /* Drop recursive lock on from acquired via thread_lock(). */ 974171482Sjeff TDQ_UNLOCK(from); 975171482Sjeff sched_rem(td); 976166108Sjeff ts->ts_cpu = cpu; 977171482Sjeff td->td_lock = TDQ_LOCKPTR(to); 978171482Sjeff tdq_add(to, td, SRQ_YIELDING); 979176735Sjeff return (1); 980116069Sjeff} 981110267Sjeff 982171482Sjeff/* 983171482Sjeff * This tdq has idled. Try to steal a thread from another cpu and switch 984171482Sjeff * to it. 985171482Sjeff */ 986123433Sjeffstatic int 987164936Sjuliantdq_idled(struct tdq *tdq) 988121790Sjeff{ 989176735Sjeff struct cpu_group *cg; 990164936Sjulian struct tdq *steal; 991194779Sjeff cpuset_t mask; 992176735Sjeff int thresh; 993171482Sjeff int cpu; 994123433Sjeff 995172484Sjeff if (smp_started == 0 || steal_idle == 0) 996172484Sjeff return (1); 997194779Sjeff CPU_FILL(&mask); 998194779Sjeff CPU_CLR(PCPU_GET(cpuid), &mask); 999176735Sjeff /* We don't want to be preempted while we're iterating. */ 1000171482Sjeff spinlock_enter(); 1001176735Sjeff for (cg = tdq->tdq_cg; cg != NULL; ) { 1002191643Sjeff if ((cg->cg_flags & CG_FLAG_THREAD) == 0) 1003176735Sjeff thresh = steal_thresh; 1004176735Sjeff else 1005176735Sjeff thresh = 1; 1006176735Sjeff cpu = sched_highest(cg, mask, thresh); 1007176735Sjeff if (cpu == -1) { 1008176735Sjeff cg = cg->cg_parent; 1009176735Sjeff continue; 1010166108Sjeff } 1011176735Sjeff steal = TDQ_CPU(cpu); 1012194779Sjeff CPU_CLR(cpu, &mask); 1013176735Sjeff tdq_lock_pair(tdq, steal); 1014176735Sjeff if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 1015176735Sjeff tdq_unlock_pair(tdq, steal); 1016176735Sjeff continue; 1017171482Sjeff } 1018176735Sjeff /* 1019176735Sjeff * If a thread was added while interrupts were disabled don't 1020176735Sjeff * steal one here. If we fail to acquire one due to affinity 1021176735Sjeff * restrictions loop again with this cpu removed from the 1022176735Sjeff * set. 1023176735Sjeff */ 1024176735Sjeff if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 1025176735Sjeff tdq_unlock_pair(tdq, steal); 1026176735Sjeff continue; 1027176735Sjeff } 1028176735Sjeff spinlock_exit(); 1029176735Sjeff TDQ_UNLOCK(steal); 1030178272Sjeff mi_switch(SW_VOL | SWT_IDLE, NULL); 1031176735Sjeff thread_unlock(curthread); 1032176735Sjeff 1033176735Sjeff return (0); 1034123433Sjeff } 1035171482Sjeff spinlock_exit(); 1036123433Sjeff return (1); 1037121790Sjeff} 1038121790Sjeff 1039171482Sjeff/* 1040171482Sjeff * Notify a remote cpu of new work. Sends an IPI if criteria are met. 1041171482Sjeff */ 1042121790Sjeffstatic void 1043177435Sjefftdq_notify(struct tdq *tdq, struct thread *td) 1044121790Sjeff{ 1045185047Sjhb struct thread *ctd; 1046166247Sjeff int pri; 1047166108Sjeff int cpu; 1048121790Sjeff 1049177005Sjeff if (tdq->tdq_ipipending) 1050177005Sjeff return; 1051301456Skib cpu = td_get_sched(td)->ts_cpu; 1052177435Sjeff pri = td->td_priority; 1053185047Sjhb ctd = pcpu_find(cpu)->pc_curthread; 1054185047Sjhb if (!sched_shouldpreempt(pri, ctd->td_priority, 1)) 1055166137Sjeff return; 1056271616Smav 1057271616Smav /* 1058271713Smav * Make sure that our caller's earlier update to tdq_load is 1059271713Smav * globally visible before we read tdq_cpu_idle. Idle thread 1060271616Smav * accesses both of them without locks, and the order is important. 1061271616Smav */ 1062285353Skib atomic_thread_fence_seq_cst(); 1063271616Smav 1064185047Sjhb if (TD_IS_IDLETHREAD(ctd)) { 1065178277Sjeff /* 1066178471Sjeff * If the MD code has an idle wakeup routine try that before 1067178471Sjeff * falling back to IPI. 1068178471Sjeff */ 1069212416Smav if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu)) 1070178471Sjeff return; 1071178277Sjeff } 1072177005Sjeff tdq->tdq_ipipending = 1; 1073210939Sjhb ipi_cpu(cpu, IPI_PREEMPT); 1074121790Sjeff} 1075121790Sjeff 1076171482Sjeff/* 1077171482Sjeff * Steals load from a timeshare queue. Honors the rotating queue head 1078171482Sjeff * index. 1079171482Sjeff */ 1080177435Sjeffstatic struct thread * 1081176735Sjeffrunq_steal_from(struct runq *rq, int cpu, u_char start) 1082171482Sjeff{ 1083171482Sjeff struct rqbits *rqb; 1084171482Sjeff struct rqhead *rqh; 1085232207Smav struct thread *td, *first; 1086171482Sjeff int bit; 1087171482Sjeff int i; 1088171482Sjeff 1089171482Sjeff rqb = &rq->rq_status; 1090171482Sjeff bit = start & (RQB_BPW -1); 1091232207Smav first = NULL; 1092171482Sjeffagain: 1093171482Sjeff for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 1094171482Sjeff if (rqb->rqb_bits[i] == 0) 1095171482Sjeff continue; 1096262917Sjeff if (bit == 0) 1097262917Sjeff bit = RQB_FFS(rqb->rqb_bits[i]); 1098262917Sjeff for (; bit < RQB_BPW; bit++) { 1099262917Sjeff if ((rqb->rqb_bits[i] & (1ul << bit)) == 0) 1100171482Sjeff continue; 1101262917Sjeff rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)]; 1102262917Sjeff TAILQ_FOREACH(td, rqh, td_runq) { 1103262917Sjeff if (first && THREAD_CAN_MIGRATE(td) && 1104262917Sjeff THREAD_CAN_SCHED(td, cpu)) 1105262917Sjeff return (td); 1106262917Sjeff first = td; 1107262917Sjeff } 1108171482Sjeff } 1109171482Sjeff } 1110171482Sjeff if (start != 0) { 1111171482Sjeff start = 0; 1112171482Sjeff goto again; 1113171482Sjeff } 1114171482Sjeff 1115232207Smav if (first && THREAD_CAN_MIGRATE(first) && 1116232207Smav THREAD_CAN_SCHED(first, cpu)) 1117232207Smav return (first); 1118171482Sjeff return (NULL); 1119171482Sjeff} 1120171482Sjeff 1121171482Sjeff/* 1122171482Sjeff * Steals load from a standard linear queue. 1123171482Sjeff */ 1124177435Sjeffstatic struct thread * 1125176735Sjeffrunq_steal(struct runq *rq, int cpu) 1126121790Sjeff{ 1127121790Sjeff struct rqhead *rqh; 1128121790Sjeff struct rqbits *rqb; 1129177435Sjeff struct thread *td; 1130121790Sjeff int word; 1131121790Sjeff int bit; 1132121790Sjeff 1133121790Sjeff rqb = &rq->rq_status; 1134121790Sjeff for (word = 0; word < RQB_LEN; word++) { 1135121790Sjeff if (rqb->rqb_bits[word] == 0) 1136121790Sjeff continue; 1137121790Sjeff for (bit = 0; bit < RQB_BPW; bit++) { 1138123231Speter if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1139121790Sjeff continue; 1140121790Sjeff rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1141177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) 1142177435Sjeff if (THREAD_CAN_MIGRATE(td) && 1143177435Sjeff THREAD_CAN_SCHED(td, cpu)) 1144177435Sjeff return (td); 1145121790Sjeff } 1146121790Sjeff } 1147121790Sjeff return (NULL); 1148121790Sjeff} 1149121790Sjeff 1150171482Sjeff/* 1151171482Sjeff * Attempt to steal a thread in priority order from a thread queue. 1152171482Sjeff */ 1153177435Sjeffstatic struct thread * 1154176735Sjefftdq_steal(struct tdq *tdq, int cpu) 1155121790Sjeff{ 1156177435Sjeff struct thread *td; 1157121790Sjeff 1158171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1159177435Sjeff if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1160177435Sjeff return (td); 1161177435Sjeff if ((td = runq_steal_from(&tdq->tdq_timeshare, 1162177435Sjeff cpu, tdq->tdq_ridx)) != NULL) 1163177435Sjeff return (td); 1164176735Sjeff return (runq_steal(&tdq->tdq_idle, cpu)); 1165121790Sjeff} 1166123433Sjeff 1167171482Sjeff/* 1168171482Sjeff * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1169172409Sjeff * current lock and returns with the assigned queue locked. 1170171482Sjeff */ 1171171482Sjeffstatic inline struct tdq * 1172177435Sjeffsched_setcpu(struct thread *td, int cpu, int flags) 1173123433Sjeff{ 1174177435Sjeff 1175171482Sjeff struct tdq *tdq; 1176123433Sjeff 1177177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1178171482Sjeff tdq = TDQ_CPU(cpu); 1179301456Skib td_get_sched(td)->ts_cpu = cpu; 1180177435Sjeff /* 1181177435Sjeff * If the lock matches just return the queue. 1182177435Sjeff */ 1183171482Sjeff if (td->td_lock == TDQ_LOCKPTR(tdq)) 1184171482Sjeff return (tdq); 1185171482Sjeff#ifdef notyet 1186123433Sjeff /* 1187172293Sjeff * If the thread isn't running its lockptr is a 1188171482Sjeff * turnstile or a sleepqueue. We can just lock_set without 1189171482Sjeff * blocking. 1190123685Sjeff */ 1191171482Sjeff if (TD_CAN_RUN(td)) { 1192171482Sjeff TDQ_LOCK(tdq); 1193171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1194171482Sjeff return (tdq); 1195171482Sjeff } 1196171482Sjeff#endif 1197166108Sjeff /* 1198171482Sjeff * The hard case, migration, we need to block the thread first to 1199171482Sjeff * prevent order reversals with other cpus locks. 1200166108Sjeff */ 1201202889Sattilio spinlock_enter(); 1202171482Sjeff thread_lock_block(td); 1203171482Sjeff TDQ_LOCK(tdq); 1204171713Sjeff thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1205202889Sattilio spinlock_exit(); 1206171482Sjeff return (tdq); 1207166108Sjeff} 1208166108Sjeff 1209178272SjeffSCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1210178272SjeffSCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1211178272SjeffSCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1212178272SjeffSCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1213178272SjeffSCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1214178272SjeffSCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1215178272Sjeff 1216166108Sjeffstatic int 1217177435Sjeffsched_pickcpu(struct thread *td, int flags) 1218171482Sjeff{ 1219232207Smav struct cpu_group *cg, *ccg; 1220177435Sjeff struct td_sched *ts; 1221171482Sjeff struct tdq *tdq; 1222194779Sjeff cpuset_t mask; 1223232207Smav int cpu, pri, self; 1224166108Sjeff 1225176735Sjeff self = PCPU_GET(cpuid); 1226301456Skib ts = td_get_sched(td); 1227166108Sjeff if (smp_started == 0) 1228166108Sjeff return (self); 1229171506Sjeff /* 1230171506Sjeff * Don't migrate a running thread from sched_switch(). 1231171506Sjeff */ 1232176735Sjeff if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1233176735Sjeff return (ts->ts_cpu); 1234166108Sjeff /* 1235176735Sjeff * Prefer to run interrupt threads on the processors that generate 1236176735Sjeff * the interrupt. 1237166108Sjeff */ 1238232207Smav pri = td->td_priority; 1239176735Sjeff if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1240178272Sjeff curthread->td_intr_nesting_level && ts->ts_cpu != self) { 1241178272Sjeff SCHED_STAT_INC(pickcpu_intrbind); 1242176735Sjeff ts->ts_cpu = self; 1243232207Smav if (TDQ_CPU(self)->tdq_lowpri > pri) { 1244232207Smav SCHED_STAT_INC(pickcpu_affinity); 1245232207Smav return (ts->ts_cpu); 1246232207Smav } 1247178272Sjeff } 1248166108Sjeff /* 1249176735Sjeff * If the thread can run on the last cpu and the affinity has not 1250176735Sjeff * expired or it is idle run it there. 1251166108Sjeff */ 1252176735Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1253232207Smav cg = tdq->tdq_cg; 1254232207Smav if (THREAD_CAN_SCHED(td, ts->ts_cpu) && 1255232207Smav tdq->tdq_lowpri >= PRI_MIN_IDLE && 1256232207Smav SCHED_AFFINITY(ts, CG_SHARE_L2)) { 1257232207Smav if (cg->cg_flags & CG_FLAG_THREAD) { 1258232207Smav CPUSET_FOREACH(cpu, cg->cg_mask) { 1259232207Smav if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) 1260232207Smav break; 1261232207Smav } 1262232207Smav } else 1263232207Smav cpu = INT_MAX; 1264232207Smav if (cpu > mp_maxid) { 1265178272Sjeff SCHED_STAT_INC(pickcpu_idle_affinity); 1266176735Sjeff return (ts->ts_cpu); 1267178272Sjeff } 1268139334Sjeff } 1269123433Sjeff /* 1270232207Smav * Search for the last level cache CPU group in the tree. 1271232207Smav * Skip caches with expired affinity time and SMT groups. 1272232207Smav * Affinity to higher level caches will be handled less aggressively. 1273123433Sjeff */ 1274232207Smav for (ccg = NULL; cg != NULL; cg = cg->cg_parent) { 1275232207Smav if (cg->cg_flags & CG_FLAG_THREAD) 1276232207Smav continue; 1277232207Smav if (!SCHED_AFFINITY(ts, cg->cg_level)) 1278232207Smav continue; 1279232207Smav ccg = cg; 1280232207Smav } 1281232207Smav if (ccg != NULL) 1282232207Smav cg = ccg; 1283176735Sjeff cpu = -1; 1284232207Smav /* Search the group for the less loaded idle CPU we can run now. */ 1285194779Sjeff mask = td->td_cpuset->cs_mask; 1286232207Smav if (cg != NULL && cg != cpu_top && 1287232207Smav CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0) 1288232207Smav cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE), 1289232207Smav INT_MAX, ts->ts_cpu); 1290232207Smav /* Search globally for the less loaded CPU we can run now. */ 1291176735Sjeff if (cpu == -1) 1292232207Smav cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu); 1293232207Smav /* Search globally for the less loaded CPU. */ 1294232207Smav if (cpu == -1) 1295232207Smav cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu); 1296232454Smav KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1297171506Sjeff /* 1298176735Sjeff * Compare the lowest loaded cpu to current cpu. 1299171506Sjeff */ 1300177005Sjeff if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1301232207Smav TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE && 1302232207Smav TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) { 1303178272Sjeff SCHED_STAT_INC(pickcpu_local); 1304177005Sjeff cpu = self; 1305178272Sjeff } else 1306178272Sjeff SCHED_STAT_INC(pickcpu_lowest); 1307178272Sjeff if (cpu != ts->ts_cpu) 1308178272Sjeff SCHED_STAT_INC(pickcpu_migration); 1309171482Sjeff return (cpu); 1310123433Sjeff} 1311176735Sjeff#endif 1312123433Sjeff 1313117326Sjeff/* 1314121790Sjeff * Pick the highest priority task we have and return it. 1315117326Sjeff */ 1316177435Sjeffstatic struct thread * 1317164936Sjuliantdq_choose(struct tdq *tdq) 1318110267Sjeff{ 1319177435Sjeff struct thread *td; 1320110267Sjeff 1321171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1322177435Sjeff td = runq_choose(&tdq->tdq_realtime); 1323177435Sjeff if (td != NULL) 1324177435Sjeff return (td); 1325177435Sjeff td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1326177435Sjeff if (td != NULL) { 1327217351Sjhb KASSERT(td->td_priority >= PRI_MIN_BATCH, 1328165762Sjeff ("tdq_choose: Invalid priority on timeshare queue %d", 1329177435Sjeff td->td_priority)); 1330177435Sjeff return (td); 1331165762Sjeff } 1332177435Sjeff td = runq_choose(&tdq->tdq_idle); 1333177435Sjeff if (td != NULL) { 1334177435Sjeff KASSERT(td->td_priority >= PRI_MIN_IDLE, 1335165762Sjeff ("tdq_choose: Invalid priority on idle queue %d", 1336177435Sjeff td->td_priority)); 1337177435Sjeff return (td); 1338165762Sjeff } 1339165762Sjeff 1340165762Sjeff return (NULL); 1341110267Sjeff} 1342110267Sjeff 1343171482Sjeff/* 1344171482Sjeff * Initialize a thread queue. 1345171482Sjeff */ 1346109864Sjeffstatic void 1347164936Sjuliantdq_setup(struct tdq *tdq) 1348110028Sjeff{ 1349171482Sjeff 1350171713Sjeff if (bootverbose) 1351171713Sjeff printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1352165762Sjeff runq_init(&tdq->tdq_realtime); 1353165762Sjeff runq_init(&tdq->tdq_timeshare); 1354165620Sjeff runq_init(&tdq->tdq_idle); 1355176735Sjeff snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1356176735Sjeff "sched lock %d", (int)TDQ_ID(tdq)); 1357176735Sjeff mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1358176735Sjeff MTX_SPIN | MTX_RECURSE); 1359187357Sjeff#ifdef KTR 1360187357Sjeff snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname), 1361187357Sjeff "CPU %d load", (int)TDQ_ID(tdq)); 1362187357Sjeff#endif 1363110028Sjeff} 1364110028Sjeff 1365171713Sjeff#ifdef SMP 1366110028Sjeffstatic void 1367171713Sjeffsched_setup_smp(void) 1368171713Sjeff{ 1369171713Sjeff struct tdq *tdq; 1370171713Sjeff int i; 1371171713Sjeff 1372176735Sjeff cpu_top = smp_topo(); 1373209059Sjhb CPU_FOREACH(i) { 1374176735Sjeff tdq = TDQ_CPU(i); 1375171713Sjeff tdq_setup(tdq); 1376176735Sjeff tdq->tdq_cg = smp_topo_find(cpu_top, i); 1377176735Sjeff if (tdq->tdq_cg == NULL) 1378176735Sjeff panic("Can't find cpu group for %d\n", i); 1379123433Sjeff } 1380176735Sjeff balance_tdq = TDQ_SELF(); 1381176735Sjeff sched_balance(); 1382171713Sjeff} 1383171713Sjeff#endif 1384171713Sjeff 1385171713Sjeff/* 1386171713Sjeff * Setup the thread queues and initialize the topology based on MD 1387171713Sjeff * information. 1388171713Sjeff */ 1389171713Sjeffstatic void 1390171713Sjeffsched_setup(void *dummy) 1391171713Sjeff{ 1392171713Sjeff struct tdq *tdq; 1393171713Sjeff 1394171713Sjeff tdq = TDQ_SELF(); 1395171713Sjeff#ifdef SMP 1396176734Sjeff sched_setup_smp(); 1397117237Sjeff#else 1398171713Sjeff tdq_setup(tdq); 1399116069Sjeff#endif 1400171482Sjeff 1401171482Sjeff /* Add thread0's load since it's running. */ 1402171482Sjeff TDQ_LOCK(tdq); 1403171713Sjeff thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1404177435Sjeff tdq_load_add(tdq, &thread0); 1405176735Sjeff tdq->tdq_lowpri = thread0.td_priority; 1406171482Sjeff TDQ_UNLOCK(tdq); 1407109864Sjeff} 1408109864Sjeff 1409171482Sjeff/* 1410239185Smav * This routine determines time constants after stathz and hz are setup. 1411171482Sjeff */ 1412153533Sdavidxu/* ARGSUSED */ 1413153533Sdavidxustatic void 1414153533Sdavidxusched_initticks(void *dummy) 1415153533Sdavidxu{ 1416171482Sjeff int incr; 1417171482Sjeff 1418153533Sdavidxu realstathz = stathz ? stathz : hz; 1419242736Sjeff sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR; 1420242736Sjeff sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR; 1421239196Smav hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 1422239196Smav realstathz); 1423153533Sdavidxu 1424153533Sdavidxu /* 1425165762Sjeff * tickincr is shifted out by 10 to avoid rounding errors due to 1426165766Sjeff * hz not being evenly divisible by stathz on all platforms. 1427153533Sdavidxu */ 1428171482Sjeff incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1429165762Sjeff /* 1430165762Sjeff * This does not work for values of stathz that are more than 1431165762Sjeff * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1432165762Sjeff */ 1433171482Sjeff if (incr == 0) 1434171482Sjeff incr = 1; 1435171482Sjeff tickincr = incr; 1436166108Sjeff#ifdef SMP 1437171899Sjeff /* 1438172409Sjeff * Set the default balance interval now that we know 1439172409Sjeff * what realstathz is. 1440172409Sjeff */ 1441172409Sjeff balance_interval = realstathz; 1442166108Sjeff affinity = SCHED_AFFINITY_DEFAULT; 1443166108Sjeff#endif 1444232740Smav if (sched_idlespinthresh < 0) 1445242852Smav sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz; 1446153533Sdavidxu} 1447153533Sdavidxu 1448153533Sdavidxu 1449109864Sjeff/* 1450171482Sjeff * This is the core of the interactivity algorithm. Determines a score based 1451171482Sjeff * on past behavior. It is the ratio of sleep time to run time scaled to 1452171482Sjeff * a [0, 100] integer. This is the voluntary sleep time of a process, which 1453171482Sjeff * differs from the cpu usage because it does not account for time spent 1454171482Sjeff * waiting on a run-queue. Would be prettier if we had floating point. 1455287166Sgnn * 1456287166Sgnn * When a thread's sleep time is greater than its run time the 1457287166Sgnn * calculation is: 1458287166Sgnn * 1459287166Sgnn * scaling factor 1460287166Sgnn * interactivity score = --------------------- 1461287166Sgnn * sleep time / run time 1462287166Sgnn * 1463287166Sgnn * 1464287166Sgnn * When a thread's run time is greater than its sleep time the 1465287166Sgnn * calculation is: 1466287166Sgnn * 1467287166Sgnn * scaling factor 1468287166Sgnn * interactivity score = --------------------- + scaling factor 1469287166Sgnn * run time / sleep time 1470171482Sjeff */ 1471171482Sjeffstatic int 1472171482Sjeffsched_interact_score(struct thread *td) 1473171482Sjeff{ 1474171482Sjeff struct td_sched *ts; 1475171482Sjeff int div; 1476171482Sjeff 1477301456Skib ts = td_get_sched(td); 1478171482Sjeff /* 1479171482Sjeff * The score is only needed if this is likely to be an interactive 1480171482Sjeff * task. Don't go through the expense of computing it if there's 1481171482Sjeff * no chance. 1482171482Sjeff */ 1483171482Sjeff if (sched_interact <= SCHED_INTERACT_HALF && 1484171482Sjeff ts->ts_runtime >= ts->ts_slptime) 1485171482Sjeff return (SCHED_INTERACT_HALF); 1486171482Sjeff 1487171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1488171482Sjeff div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1489171482Sjeff return (SCHED_INTERACT_HALF + 1490171482Sjeff (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1491171482Sjeff } 1492171482Sjeff if (ts->ts_slptime > ts->ts_runtime) { 1493171482Sjeff div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1494171482Sjeff return (ts->ts_runtime / div); 1495171482Sjeff } 1496171482Sjeff /* runtime == slptime */ 1497171482Sjeff if (ts->ts_runtime) 1498171482Sjeff return (SCHED_INTERACT_HALF); 1499171482Sjeff 1500171482Sjeff /* 1501171482Sjeff * This can happen if slptime and runtime are 0. 1502171482Sjeff */ 1503171482Sjeff return (0); 1504171482Sjeff 1505171482Sjeff} 1506171482Sjeff 1507171482Sjeff/* 1508109864Sjeff * Scale the scheduling priority according to the "interactivity" of this 1509109864Sjeff * process. 1510109864Sjeff */ 1511113357Sjeffstatic void 1512163709Sjbsched_priority(struct thread *td) 1513109864Sjeff{ 1514165762Sjeff int score; 1515109864Sjeff int pri; 1516109864Sjeff 1517217291Sjhb if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1518113357Sjeff return; 1519112966Sjeff /* 1520165762Sjeff * If the score is interactive we place the thread in the realtime 1521165762Sjeff * queue with a priority that is less than kernel and interrupt 1522165762Sjeff * priorities. These threads are not subject to nice restrictions. 1523112966Sjeff * 1524171482Sjeff * Scores greater than this are placed on the normal timeshare queue 1525165762Sjeff * where the priority is partially decided by the most recent cpu 1526165762Sjeff * utilization and the rest is decided by nice value. 1527172293Sjeff * 1528172293Sjeff * The nice value of the process has a linear effect on the calculated 1529172293Sjeff * score. Negative nice values make it easier for a thread to be 1530172293Sjeff * considered interactive. 1531112966Sjeff */ 1532198126Sjhb score = imax(0, sched_interact_score(td) + td->td_proc->p_nice); 1533165762Sjeff if (score < sched_interact) { 1534217351Sjhb pri = PRI_MIN_INTERACT; 1535217351Sjhb pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) / 1536217237Sjhb sched_interact) * score; 1537217351Sjhb KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT, 1538166208Sjeff ("sched_priority: invalid interactive priority %d score %d", 1539166208Sjeff pri, score)); 1540165762Sjeff } else { 1541165762Sjeff pri = SCHED_PRI_MIN; 1542301456Skib if (td_get_sched(td)->ts_ticks) 1543301456Skib pri += min(SCHED_PRI_TICKS(td_get_sched(td)), 1544258869Sjhb SCHED_PRI_RANGE - 1); 1545165762Sjeff pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1546217351Sjhb KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH, 1547171482Sjeff ("sched_priority: invalid priority %d: nice %d, " 1548171482Sjeff "ticks %d ftick %d ltick %d tick pri %d", 1549301456Skib pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks, 1550301456Skib td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick, 1551301456Skib SCHED_PRI_TICKS(td_get_sched(td)))); 1552165762Sjeff } 1553165762Sjeff sched_user_prio(td, pri); 1554112966Sjeff 1555112966Sjeff return; 1556109864Sjeff} 1557109864Sjeff 1558121868Sjeff/* 1559121868Sjeff * This routine enforces a maximum limit on the amount of scheduling history 1560171482Sjeff * kept. It is called after either the slptime or runtime is adjusted. This 1561171482Sjeff * function is ugly due to integer math. 1562121868Sjeff */ 1563116463Sjeffstatic void 1564163709Sjbsched_interact_update(struct thread *td) 1565116463Sjeff{ 1566165819Sjeff struct td_sched *ts; 1567166208Sjeff u_int sum; 1568121605Sjeff 1569301456Skib ts = td_get_sched(td); 1570171482Sjeff sum = ts->ts_runtime + ts->ts_slptime; 1571121868Sjeff if (sum < SCHED_SLP_RUN_MAX) 1572121868Sjeff return; 1573121868Sjeff /* 1574165819Sjeff * This only happens from two places: 1575165819Sjeff * 1) We have added an unusual amount of run time from fork_exit. 1576165819Sjeff * 2) We have added an unusual amount of sleep time from sched_sleep(). 1577165819Sjeff */ 1578165819Sjeff if (sum > SCHED_SLP_RUN_MAX * 2) { 1579171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1580171482Sjeff ts->ts_runtime = SCHED_SLP_RUN_MAX; 1581171482Sjeff ts->ts_slptime = 1; 1582165819Sjeff } else { 1583171482Sjeff ts->ts_slptime = SCHED_SLP_RUN_MAX; 1584171482Sjeff ts->ts_runtime = 1; 1585165819Sjeff } 1586165819Sjeff return; 1587165819Sjeff } 1588165819Sjeff /* 1589121868Sjeff * If we have exceeded by more than 1/5th then the algorithm below 1590121868Sjeff * will not bring us back into range. Dividing by two here forces 1591133427Sjeff * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1592121868Sjeff */ 1593127850Sjeff if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1594171482Sjeff ts->ts_runtime /= 2; 1595171482Sjeff ts->ts_slptime /= 2; 1596121868Sjeff return; 1597116463Sjeff } 1598171482Sjeff ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1599171482Sjeff ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1600116463Sjeff} 1601116463Sjeff 1602171482Sjeff/* 1603171482Sjeff * Scale back the interactivity history when a child thread is created. The 1604171482Sjeff * history is inherited from the parent but the thread may behave totally 1605171482Sjeff * differently. For example, a shell spawning a compiler process. We want 1606171482Sjeff * to learn that the compiler is behaving badly very quickly. 1607171482Sjeff */ 1608121868Sjeffstatic void 1609163709Sjbsched_interact_fork(struct thread *td) 1610121868Sjeff{ 1611301456Skib struct td_sched *ts; 1612121868Sjeff int ratio; 1613121868Sjeff int sum; 1614121868Sjeff 1615301456Skib ts = td_get_sched(td); 1616301456Skib sum = ts->ts_runtime + ts->ts_slptime; 1617121868Sjeff if (sum > SCHED_SLP_RUN_FORK) { 1618121868Sjeff ratio = sum / SCHED_SLP_RUN_FORK; 1619301456Skib ts->ts_runtime /= ratio; 1620301456Skib ts->ts_slptime /= ratio; 1621121868Sjeff } 1622121868Sjeff} 1623121868Sjeff 1624113357Sjeff/* 1625171482Sjeff * Called from proc0_init() to setup the scheduler fields. 1626134791Sjulian */ 1627134791Sjulianvoid 1628134791Sjulianschedinit(void) 1629134791Sjulian{ 1630301456Skib struct td_sched *ts0; 1631165762Sjeff 1632134791Sjulian /* 1633301456Skib * Set up the scheduler specific parts of thread0. 1634134791Sjulian */ 1635301456Skib ts0 = td_get_sched(&thread0); 1636301456Skib ts0->ts_ltick = ticks; 1637301456Skib ts0->ts_ftick = ticks; 1638301456Skib ts0->ts_slice = 0; 1639134791Sjulian} 1640134791Sjulian 1641134791Sjulian/* 1642113357Sjeff * This is only somewhat accurate since given many processes of the same 1643113357Sjeff * priority they will switch when their slices run out, which will be 1644165762Sjeff * at most sched_slice stathz ticks. 1645113357Sjeff */ 1646109864Sjeffint 1647109864Sjeffsched_rr_interval(void) 1648109864Sjeff{ 1649165762Sjeff 1650239185Smav /* Convert sched_slice from stathz to hz. */ 1651239196Smav return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz)); 1652109864Sjeff} 1653109864Sjeff 1654171482Sjeff/* 1655171482Sjeff * Update the percent cpu tracking information when it is requested or 1656171482Sjeff * the total history exceeds the maximum. We keep a sliding history of 1657171482Sjeff * tick counts that slowly decays. This is less precise than the 4BSD 1658171482Sjeff * mechanism since it happens with less regular and frequent events. 1659171482Sjeff */ 1660121790Sjeffstatic void 1661232917Smavsched_pctcpu_update(struct td_sched *ts, int run) 1662109864Sjeff{ 1663232917Smav int t = ticks; 1664165762Sjeff 1665232917Smav if (t - ts->ts_ltick >= SCHED_TICK_TARG) { 1666164936Sjulian ts->ts_ticks = 0; 1667232917Smav ts->ts_ftick = t - SCHED_TICK_TARG; 1668232917Smav } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) { 1669232917Smav ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) * 1670232917Smav (ts->ts_ltick - (t - SCHED_TICK_TARG)); 1671232917Smav ts->ts_ftick = t - SCHED_TICK_TARG; 1672232917Smav } 1673232917Smav if (run) 1674232917Smav ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT; 1675232917Smav ts->ts_ltick = t; 1676109864Sjeff} 1677109864Sjeff 1678171482Sjeff/* 1679171482Sjeff * Adjust the priority of a thread. Move it to the appropriate run-queue 1680171482Sjeff * if necessary. This is the back-end for several priority related 1681171482Sjeff * functions. 1682171482Sjeff */ 1683165762Sjeffstatic void 1684139453Sjhbsched_thread_priority(struct thread *td, u_char prio) 1685109864Sjeff{ 1686164936Sjulian struct td_sched *ts; 1687177009Sjeff struct tdq *tdq; 1688177009Sjeff int oldpri; 1689109864Sjeff 1690187357Sjeff KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio", 1691187357Sjeff "prio:%d", td->td_priority, "new prio:%d", prio, 1692187357Sjeff KTR_ATTR_LINKED, sched_tdname(curthread)); 1693258622Savg SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio); 1694240513Savg if (td != curthread && prio < td->td_priority) { 1695187357Sjeff KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 1696187357Sjeff "lend prio", "prio:%d", td->td_priority, "new prio:%d", 1697187357Sjeff prio, KTR_ATTR_LINKED, sched_tdname(td)); 1698258622Savg SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio, 1699235459Srstone curthread); 1700187357Sjeff } 1701301456Skib ts = td_get_sched(td); 1702170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1703139453Sjhb if (td->td_priority == prio) 1704139453Sjhb return; 1705177376Sjeff /* 1706177376Sjeff * If the priority has been elevated due to priority 1707177376Sjeff * propagation, we may have to move ourselves to a new 1708177376Sjeff * queue. This could be optimized to not re-add in some 1709177376Sjeff * cases. 1710177376Sjeff */ 1711165766Sjeff if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1712165762Sjeff sched_rem(td); 1713165762Sjeff td->td_priority = prio; 1714171482Sjeff sched_add(td, SRQ_BORROWING); 1715177009Sjeff return; 1716177009Sjeff } 1717177376Sjeff /* 1718177376Sjeff * If the thread is currently running we may have to adjust the lowpri 1719177376Sjeff * information so other cpus are aware of our current priority. 1720177376Sjeff */ 1721177009Sjeff if (TD_IS_RUNNING(td)) { 1722177376Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1723177376Sjeff oldpri = td->td_priority; 1724177376Sjeff td->td_priority = prio; 1725176735Sjeff if (prio < tdq->tdq_lowpri) 1726171482Sjeff tdq->tdq_lowpri = prio; 1727176735Sjeff else if (tdq->tdq_lowpri == oldpri) 1728176735Sjeff tdq_setlowpri(tdq, td); 1729177376Sjeff return; 1730177009Sjeff } 1731177376Sjeff td->td_priority = prio; 1732109864Sjeff} 1733109864Sjeff 1734139453Sjhb/* 1735139453Sjhb * Update a thread's priority when it is lent another thread's 1736139453Sjhb * priority. 1737139453Sjhb */ 1738109864Sjeffvoid 1739139453Sjhbsched_lend_prio(struct thread *td, u_char prio) 1740139453Sjhb{ 1741139453Sjhb 1742139453Sjhb td->td_flags |= TDF_BORROWING; 1743139453Sjhb sched_thread_priority(td, prio); 1744139453Sjhb} 1745139453Sjhb 1746139453Sjhb/* 1747139453Sjhb * Restore a thread's priority when priority propagation is 1748139453Sjhb * over. The prio argument is the minimum priority the thread 1749139453Sjhb * needs to have to satisfy other possible priority lending 1750139453Sjhb * requests. If the thread's regular priority is less 1751139453Sjhb * important than prio, the thread will keep a priority boost 1752139453Sjhb * of prio. 1753139453Sjhb */ 1754139453Sjhbvoid 1755139453Sjhbsched_unlend_prio(struct thread *td, u_char prio) 1756139453Sjhb{ 1757139453Sjhb u_char base_pri; 1758139453Sjhb 1759139453Sjhb if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1760139453Sjhb td->td_base_pri <= PRI_MAX_TIMESHARE) 1761163709Sjb base_pri = td->td_user_pri; 1762139453Sjhb else 1763139453Sjhb base_pri = td->td_base_pri; 1764139453Sjhb if (prio >= base_pri) { 1765139455Sjhb td->td_flags &= ~TDF_BORROWING; 1766139453Sjhb sched_thread_priority(td, base_pri); 1767139453Sjhb } else 1768139453Sjhb sched_lend_prio(td, prio); 1769139453Sjhb} 1770139453Sjhb 1771171482Sjeff/* 1772171482Sjeff * Standard entry for setting the priority to an absolute value. 1773171482Sjeff */ 1774139453Sjhbvoid 1775139453Sjhbsched_prio(struct thread *td, u_char prio) 1776139453Sjhb{ 1777139453Sjhb u_char oldprio; 1778139453Sjhb 1779139453Sjhb /* First, update the base priority. */ 1780139453Sjhb td->td_base_pri = prio; 1781139453Sjhb 1782139453Sjhb /* 1783139455Sjhb * If the thread is borrowing another thread's priority, don't 1784139453Sjhb * ever lower the priority. 1785139453Sjhb */ 1786139453Sjhb if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1787139453Sjhb return; 1788139453Sjhb 1789139453Sjhb /* Change the real priority. */ 1790139453Sjhb oldprio = td->td_priority; 1791139453Sjhb sched_thread_priority(td, prio); 1792139453Sjhb 1793139453Sjhb /* 1794139453Sjhb * If the thread is on a turnstile, then let the turnstile update 1795139453Sjhb * its state. 1796139453Sjhb */ 1797139453Sjhb if (TD_ON_LOCK(td) && oldprio != prio) 1798139453Sjhb turnstile_adjust(td, oldprio); 1799139453Sjhb} 1800139455Sjhb 1801171482Sjeff/* 1802171482Sjeff * Set the base user priority, does not effect current running priority. 1803171482Sjeff */ 1804139453Sjhbvoid 1805163709Sjbsched_user_prio(struct thread *td, u_char prio) 1806161599Sdavidxu{ 1807161599Sdavidxu 1808163709Sjb td->td_base_user_pri = prio; 1809216313Sdavidxu if (td->td_lend_user_pri <= prio) 1810216313Sdavidxu return; 1811163709Sjb td->td_user_pri = prio; 1812161599Sdavidxu} 1813161599Sdavidxu 1814161599Sdavidxuvoid 1815161599Sdavidxusched_lend_user_prio(struct thread *td, u_char prio) 1816161599Sdavidxu{ 1817161599Sdavidxu 1818174536Sdavidxu THREAD_LOCK_ASSERT(td, MA_OWNED); 1819216313Sdavidxu td->td_lend_user_pri = prio; 1820216791Sdavidxu td->td_user_pri = min(prio, td->td_base_user_pri); 1821216791Sdavidxu if (td->td_priority > td->td_user_pri) 1822216791Sdavidxu sched_prio(td, td->td_user_pri); 1823216791Sdavidxu else if (td->td_priority != td->td_user_pri) 1824216791Sdavidxu td->td_flags |= TDF_NEEDRESCHED; 1825161599Sdavidxu} 1826161599Sdavidxu 1827171482Sjeff/* 1828171713Sjeff * Handle migration from sched_switch(). This happens only for 1829171713Sjeff * cpu binding. 1830171713Sjeff */ 1831171713Sjeffstatic struct mtx * 1832171713Sjeffsched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1833171713Sjeff{ 1834171713Sjeff struct tdq *tdn; 1835171713Sjeff 1836301456Skib tdn = TDQ_CPU(td_get_sched(td)->ts_cpu); 1837171713Sjeff#ifdef SMP 1838177435Sjeff tdq_load_rem(tdq, td); 1839171713Sjeff /* 1840171713Sjeff * Do the lock dance required to avoid LOR. We grab an extra 1841171713Sjeff * spinlock nesting to prevent preemption while we're 1842171713Sjeff * not holding either run-queue lock. 1843171713Sjeff */ 1844171713Sjeff spinlock_enter(); 1845202889Sattilio thread_lock_block(td); /* This releases the lock on tdq. */ 1846197223Sattilio 1847197223Sattilio /* 1848197223Sattilio * Acquire both run-queue locks before placing the thread on the new 1849197223Sattilio * run-queue to avoid deadlocks created by placing a thread with a 1850197223Sattilio * blocked lock on the run-queue of a remote processor. The deadlock 1851197223Sattilio * occurs when a third processor attempts to lock the two queues in 1852197223Sattilio * question while the target processor is spinning with its own 1853197223Sattilio * run-queue lock held while waiting for the blocked lock to clear. 1854197223Sattilio */ 1855197223Sattilio tdq_lock_pair(tdn, tdq); 1856171713Sjeff tdq_add(tdn, td, flags); 1857177435Sjeff tdq_notify(tdn, td); 1858197223Sattilio TDQ_UNLOCK(tdn); 1859171713Sjeff spinlock_exit(); 1860171713Sjeff#endif 1861171713Sjeff return (TDQ_LOCKPTR(tdn)); 1862171713Sjeff} 1863171713Sjeff 1864171713Sjeff/* 1865202889Sattilio * Variadic version of thread_lock_unblock() that does not assume td_lock 1866202889Sattilio * is blocked. 1867171482Sjeff */ 1868171482Sjeffstatic inline void 1869171482Sjeffthread_unblock_switch(struct thread *td, struct mtx *mtx) 1870171482Sjeff{ 1871171482Sjeff atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1872171482Sjeff (uintptr_t)mtx); 1873171482Sjeff} 1874171482Sjeff 1875171482Sjeff/* 1876171482Sjeff * Switch threads. This function has to handle threads coming in while 1877171482Sjeff * blocked for some reason, running, or idle. It also must deal with 1878171482Sjeff * migrating a thread from one queue to another as running threads may 1879171482Sjeff * be assigned elsewhere via binding. 1880171482Sjeff */ 1881161599Sdavidxuvoid 1882135051Sjuliansched_switch(struct thread *td, struct thread *newtd, int flags) 1883109864Sjeff{ 1884165627Sjeff struct tdq *tdq; 1885164936Sjulian struct td_sched *ts; 1886171482Sjeff struct mtx *mtx; 1887171713Sjeff int srqflag; 1888239157Smav int cpuid, preempted; 1889109864Sjeff 1890170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1891177376Sjeff KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1892109864Sjeff 1893171482Sjeff cpuid = PCPU_GET(cpuid); 1894171482Sjeff tdq = TDQ_CPU(cpuid); 1895301456Skib ts = td_get_sched(td); 1896171713Sjeff mtx = td->td_lock; 1897232917Smav sched_pctcpu_update(ts, 1); 1898171482Sjeff ts->ts_rltick = ticks; 1899133555Sjeff td->td_lastcpu = td->td_oncpu; 1900113339Sjulian td->td_oncpu = NOCPU; 1901270423Smav preempted = !((td->td_flags & TDF_SLICEEND) || 1902270423Smav (flags & SWT_RELINQUISH)); 1903239157Smav td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND); 1904144777Sups td->td_owepreempt = 0; 1905242852Smav if (!TD_IS_IDLETHREAD(td)) 1906242852Smav tdq->tdq_switchcnt++; 1907123434Sjeff /* 1908171482Sjeff * The lock pointer in an idle thread should never change. Reset it 1909171482Sjeff * to CAN_RUN as well. 1910123434Sjeff */ 1911167327Sjulian if (TD_IS_IDLETHREAD(td)) { 1912171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1913139334Sjeff TD_SET_CAN_RUN(td); 1914170293Sjeff } else if (TD_IS_RUNNING(td)) { 1915171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1916239157Smav srqflag = preempted ? 1917170293Sjeff SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1918171713Sjeff SRQ_OURSELF|SRQ_YIELDING; 1919212153Smdf#ifdef SMP 1920212115Smdf if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu)) 1921212115Smdf ts->ts_cpu = sched_pickcpu(td, 0); 1922212153Smdf#endif 1923171713Sjeff if (ts->ts_cpu == cpuid) 1924177435Sjeff tdq_runq_add(tdq, td, srqflag); 1925212115Smdf else { 1926212115Smdf KASSERT(THREAD_CAN_MIGRATE(td) || 1927212115Smdf (ts->ts_flags & TSF_BOUND) != 0, 1928212115Smdf ("Thread %p shouldn't migrate", td)); 1929171713Sjeff mtx = sched_switch_migrate(tdq, td, srqflag); 1930212115Smdf } 1931171482Sjeff } else { 1932171482Sjeff /* This thread must be going to sleep. */ 1933171482Sjeff TDQ_LOCK(tdq); 1934202889Sattilio mtx = thread_lock_block(td); 1935177435Sjeff tdq_load_rem(tdq, td); 1936171482Sjeff } 1937171482Sjeff /* 1938171482Sjeff * We enter here with the thread blocked and assigned to the 1939171482Sjeff * appropriate cpu run-queue or sleep-queue and with the current 1940171482Sjeff * thread-queue locked. 1941171482Sjeff */ 1942171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1943171482Sjeff newtd = choosethread(); 1944171482Sjeff /* 1945171482Sjeff * Call the MD code to switch contexts if necessary. 1946171482Sjeff */ 1947145256Sjkoshy if (td != newtd) { 1948145256Sjkoshy#ifdef HWPMC_HOOKS 1949145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1950145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1951145256Sjkoshy#endif 1952258622Savg SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc); 1953174629Sjeff lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1954172411Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1955301456Skib sched_pctcpu_update(td_get_sched(newtd), 0); 1956179297Sjb 1957179297Sjb#ifdef KDTRACE_HOOKS 1958179297Sjb /* 1959179297Sjb * If DTrace has set the active vtime enum to anything 1960179297Sjb * other than INACTIVE (0), then it should have set the 1961179297Sjb * function to call. 1962179297Sjb */ 1963179297Sjb if (dtrace_vtime_active) 1964179297Sjb (*dtrace_vtime_switch_func)(newtd); 1965179297Sjb#endif 1966179297Sjb 1967171482Sjeff cpu_switch(td, newtd, mtx); 1968171482Sjeff /* 1969171482Sjeff * We may return from cpu_switch on a different cpu. However, 1970171482Sjeff * we always return with td_lock pointing to the current cpu's 1971171482Sjeff * run queue lock. 1972171482Sjeff */ 1973171482Sjeff cpuid = PCPU_GET(cpuid); 1974171482Sjeff tdq = TDQ_CPU(cpuid); 1975174629Sjeff lock_profile_obtain_lock_success( 1976174629Sjeff &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1977235459Srstone 1978258622Savg SDT_PROBE0(sched, , , on__cpu); 1979145256Sjkoshy#ifdef HWPMC_HOOKS 1980145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1981145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1982145256Sjkoshy#endif 1983235459Srstone } else { 1984171482Sjeff thread_unblock_switch(td, mtx); 1985258622Savg SDT_PROBE0(sched, , , remain__cpu); 1986235459Srstone } 1987171482Sjeff /* 1988171482Sjeff * Assert that all went well and return. 1989171482Sjeff */ 1990171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1991171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1992171482Sjeff td->td_oncpu = cpuid; 1993109864Sjeff} 1994109864Sjeff 1995171482Sjeff/* 1996171482Sjeff * Adjust thread priorities as a result of a nice request. 1997171482Sjeff */ 1998109864Sjeffvoid 1999130551Sjuliansched_nice(struct proc *p, int nice) 2000109864Sjeff{ 2001109864Sjeff struct thread *td; 2002109864Sjeff 2003130551Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 2004165762Sjeff 2005130551Sjulian p->p_nice = nice; 2006163709Sjb FOREACH_THREAD_IN_PROC(p, td) { 2007170293Sjeff thread_lock(td); 2008163709Sjb sched_priority(td); 2009165762Sjeff sched_prio(td, td->td_base_user_pri); 2010170293Sjeff thread_unlock(td); 2011130551Sjulian } 2012109864Sjeff} 2013109864Sjeff 2014171482Sjeff/* 2015171482Sjeff * Record the sleep time for the interactivity scorer. 2016171482Sjeff */ 2017109864Sjeffvoid 2018177085Sjeffsched_sleep(struct thread *td, int prio) 2019109864Sjeff{ 2020165762Sjeff 2021170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2022109864Sjeff 2023172264Sjeff td->td_slptick = ticks; 2024201347Skib if (TD_IS_SUSPENDED(td) || prio >= PSOCK) 2025177085Sjeff td->td_flags |= TDF_CANSWAP; 2026217410Sjhb if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 2027217410Sjhb return; 2028177903Sjeff if (static_boost == 1 && prio) 2029177085Sjeff sched_prio(td, prio); 2030177903Sjeff else if (static_boost && td->td_priority > static_boost) 2031177903Sjeff sched_prio(td, static_boost); 2032109864Sjeff} 2033109864Sjeff 2034171482Sjeff/* 2035171482Sjeff * Schedule a thread to resume execution and record how long it voluntarily 2036171482Sjeff * slept. We also update the pctcpu, interactivity, and priority. 2037171482Sjeff */ 2038109864Sjeffvoid 2039109864Sjeffsched_wakeup(struct thread *td) 2040109864Sjeff{ 2041166229Sjeff struct td_sched *ts; 2042171482Sjeff int slptick; 2043165762Sjeff 2044170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2045301456Skib ts = td_get_sched(td); 2046177085Sjeff td->td_flags &= ~TDF_CANSWAP; 2047109864Sjeff /* 2048165762Sjeff * If we slept for more than a tick update our interactivity and 2049165762Sjeff * priority. 2050109864Sjeff */ 2051172264Sjeff slptick = td->td_slptick; 2052172264Sjeff td->td_slptick = 0; 2053171482Sjeff if (slptick && slptick != ticks) { 2054232917Smav ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT; 2055165819Sjeff sched_interact_update(td); 2056232917Smav sched_pctcpu_update(ts, 0); 2057109864Sjeff } 2058242736Sjeff /* 2059242736Sjeff * Reset the slice value since we slept and advanced the round-robin. 2060242736Sjeff */ 2061242736Sjeff ts->ts_slice = 0; 2062166190Sjeff sched_add(td, SRQ_BORING); 2063109864Sjeff} 2064109864Sjeff 2065109864Sjeff/* 2066109864Sjeff * Penalize the parent for creating a new child and initialize the child's 2067109864Sjeff * priority. 2068109864Sjeff */ 2069109864Sjeffvoid 2070163709Sjbsched_fork(struct thread *td, struct thread *child) 2071109864Sjeff{ 2072170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2073301456Skib sched_pctcpu_update(td_get_sched(td), 1); 2074164936Sjulian sched_fork_thread(td, child); 2075165762Sjeff /* 2076165762Sjeff * Penalize the parent and child for forking. 2077165762Sjeff */ 2078165762Sjeff sched_interact_fork(child); 2079165762Sjeff sched_priority(child); 2080301456Skib td_get_sched(td)->ts_runtime += tickincr; 2081165762Sjeff sched_interact_update(td); 2082165762Sjeff sched_priority(td); 2083164936Sjulian} 2084109864Sjeff 2085171482Sjeff/* 2086171482Sjeff * Fork a new thread, may be within the same process. 2087171482Sjeff */ 2088164936Sjulianvoid 2089164936Sjuliansched_fork_thread(struct thread *td, struct thread *child) 2090164936Sjulian{ 2091164936Sjulian struct td_sched *ts; 2092164936Sjulian struct td_sched *ts2; 2093242736Sjeff struct tdq *tdq; 2094164936Sjulian 2095242736Sjeff tdq = TDQ_SELF(); 2096177426Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2097165762Sjeff /* 2098165762Sjeff * Initialize child. 2099165762Sjeff */ 2100301456Skib ts = td_get_sched(td); 2101301456Skib ts2 = td_get_sched(child); 2102286256Sjhb child->td_oncpu = NOCPU; 2103286256Sjhb child->td_lastcpu = NOCPU; 2104242736Sjeff child->td_lock = TDQ_LOCKPTR(tdq); 2105176735Sjeff child->td_cpuset = cpuset_ref(td->td_cpuset); 2106164936Sjulian ts2->ts_cpu = ts->ts_cpu; 2107177426Sjeff ts2->ts_flags = 0; 2108165762Sjeff /* 2109217078Sjhb * Grab our parents cpu estimation information. 2110165762Sjeff */ 2111164936Sjulian ts2->ts_ticks = ts->ts_ticks; 2112164936Sjulian ts2->ts_ltick = ts->ts_ltick; 2113164936Sjulian ts2->ts_ftick = ts->ts_ftick; 2114165762Sjeff /* 2115217078Sjhb * Do not inherit any borrowed priority from the parent. 2116217078Sjhb */ 2117217078Sjhb child->td_priority = child->td_base_pri; 2118217078Sjhb /* 2119165762Sjeff * And update interactivity score. 2120165762Sjeff */ 2121171482Sjeff ts2->ts_slptime = ts->ts_slptime; 2122171482Sjeff ts2->ts_runtime = ts->ts_runtime; 2123242736Sjeff /* Attempt to quickly learn interactivity. */ 2124242736Sjeff ts2->ts_slice = tdq_slice(tdq) - sched_slice_min; 2125187357Sjeff#ifdef KTR 2126187357Sjeff bzero(ts2->ts_name, sizeof(ts2->ts_name)); 2127187357Sjeff#endif 2128113357Sjeff} 2129113357Sjeff 2130171482Sjeff/* 2131171482Sjeff * Adjust the priority class of a thread. 2132171482Sjeff */ 2133113357Sjeffvoid 2134163709Sjbsched_class(struct thread *td, int class) 2135113357Sjeff{ 2136113357Sjeff 2137170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2138163709Sjb if (td->td_pri_class == class) 2139113357Sjeff return; 2140163709Sjb td->td_pri_class = class; 2141109864Sjeff} 2142109864Sjeff 2143109864Sjeff/* 2144109864Sjeff * Return some of the child's priority and interactivity to the parent. 2145109864Sjeff */ 2146109864Sjeffvoid 2147164939Sjuliansched_exit(struct proc *p, struct thread *child) 2148109864Sjeff{ 2149165762Sjeff struct thread *td; 2150113372Sjeff 2151187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit", 2152225199Sdelphij "prio:%d", child->td_priority); 2153177368Sjeff PROC_LOCK_ASSERT(p, MA_OWNED); 2154165762Sjeff td = FIRST_THREAD_IN_PROC(p); 2155165762Sjeff sched_exit_thread(td, child); 2156113372Sjeff} 2157113372Sjeff 2158171482Sjeff/* 2159171482Sjeff * Penalize another thread for the time spent on this one. This helps to 2160171482Sjeff * worsen the priority and interactivity of processes which schedule batch 2161171482Sjeff * jobs such as make. This has little effect on the make process itself but 2162171482Sjeff * causes new processes spawned by it to receive worse scores immediately. 2163171482Sjeff */ 2164113372Sjeffvoid 2165164939Sjuliansched_exit_thread(struct thread *td, struct thread *child) 2166164936Sjulian{ 2167165762Sjeff 2168187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit", 2169225199Sdelphij "prio:%d", child->td_priority); 2170165762Sjeff /* 2171165762Sjeff * Give the child's runtime to the parent without returning the 2172165762Sjeff * sleep time as a penalty to the parent. This causes shells that 2173165762Sjeff * launch expensive things to mark their children as expensive. 2174165762Sjeff */ 2175170293Sjeff thread_lock(td); 2176301456Skib td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime; 2177164939Sjulian sched_interact_update(td); 2178165762Sjeff sched_priority(td); 2179170293Sjeff thread_unlock(td); 2180164936Sjulian} 2181164936Sjulian 2182177005Sjeffvoid 2183177005Sjeffsched_preempt(struct thread *td) 2184177005Sjeff{ 2185177005Sjeff struct tdq *tdq; 2186177005Sjeff 2187235459Srstone SDT_PROBE2(sched, , , surrender, td, td->td_proc); 2188235459Srstone 2189177005Sjeff thread_lock(td); 2190177005Sjeff tdq = TDQ_SELF(); 2191177005Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2192177005Sjeff tdq->tdq_ipipending = 0; 2193177005Sjeff if (td->td_priority > tdq->tdq_lowpri) { 2194178272Sjeff int flags; 2195178272Sjeff 2196178272Sjeff flags = SW_INVOL | SW_PREEMPT; 2197177005Sjeff if (td->td_critnest > 1) 2198177005Sjeff td->td_owepreempt = 1; 2199178272Sjeff else if (TD_IS_IDLETHREAD(td)) 2200178272Sjeff mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL); 2201177005Sjeff else 2202178272Sjeff mi_switch(flags | SWT_REMOTEPREEMPT, NULL); 2203177005Sjeff } 2204177005Sjeff thread_unlock(td); 2205177005Sjeff} 2206177005Sjeff 2207171482Sjeff/* 2208171482Sjeff * Fix priorities on return to user-space. Priorities may be elevated due 2209171482Sjeff * to static priorities in msleep() or similar. 2210171482Sjeff */ 2211164936Sjulianvoid 2212164936Sjuliansched_userret(struct thread *td) 2213164936Sjulian{ 2214164936Sjulian /* 2215164936Sjulian * XXX we cheat slightly on the locking here to avoid locking in 2216164936Sjulian * the usual case. Setting td_priority here is essentially an 2217164936Sjulian * incomplete workaround for not setting it properly elsewhere. 2218164936Sjulian * Now that some interrupt handlers are threads, not setting it 2219164936Sjulian * properly elsewhere can clobber it in the window between setting 2220164936Sjulian * it here and returning to user mode, so don't waste time setting 2221164936Sjulian * it perfectly here. 2222164936Sjulian */ 2223164936Sjulian KASSERT((td->td_flags & TDF_BORROWING) == 0, 2224164936Sjulian ("thread with borrowed priority returning to userland")); 2225164936Sjulian if (td->td_priority != td->td_user_pri) { 2226170293Sjeff thread_lock(td); 2227164936Sjulian td->td_priority = td->td_user_pri; 2228164936Sjulian td->td_base_pri = td->td_user_pri; 2229177005Sjeff tdq_setlowpri(TDQ_SELF(), td); 2230170293Sjeff thread_unlock(td); 2231164936Sjulian } 2232164936Sjulian} 2233164936Sjulian 2234171482Sjeff/* 2235171482Sjeff * Handle a stathz tick. This is really only relevant for timeshare 2236171482Sjeff * threads. 2237171482Sjeff */ 2238164936Sjulianvoid 2239121127Sjeffsched_clock(struct thread *td) 2240109864Sjeff{ 2241164936Sjulian struct tdq *tdq; 2242164936Sjulian struct td_sched *ts; 2243109864Sjeff 2244171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2245164936Sjulian tdq = TDQ_SELF(); 2246172409Sjeff#ifdef SMP 2247133427Sjeff /* 2248172409Sjeff * We run the long term load balancer infrequently on the first cpu. 2249172409Sjeff */ 2250172409Sjeff if (balance_tdq == tdq) { 2251172409Sjeff if (balance_ticks && --balance_ticks == 0) 2252172409Sjeff sched_balance(); 2253172409Sjeff } 2254172409Sjeff#endif 2255172409Sjeff /* 2256178277Sjeff * Save the old switch count so we have a record of the last ticks 2257178277Sjeff * activity. Initialize the new switch count based on our load. 2258178277Sjeff * If there is some activity seed it to reflect that. 2259178277Sjeff */ 2260178277Sjeff tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2261178471Sjeff tdq->tdq_switchcnt = tdq->tdq_load; 2262178277Sjeff /* 2263165766Sjeff * Advance the insert index once for each tick to ensure that all 2264165766Sjeff * threads get a chance to run. 2265133427Sjeff */ 2266165766Sjeff if (tdq->tdq_idx == tdq->tdq_ridx) { 2267165766Sjeff tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2268165766Sjeff if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2269165766Sjeff tdq->tdq_ridx = tdq->tdq_idx; 2270165766Sjeff } 2271301456Skib ts = td_get_sched(td); 2272232917Smav sched_pctcpu_update(ts, 1); 2273175104Sjeff if (td->td_pri_class & PRI_FIFO_BIT) 2274113357Sjeff return; 2275217291Sjhb if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) { 2276175104Sjeff /* 2277175104Sjeff * We used a tick; charge it to the thread so 2278175104Sjeff * that we can compute our interactivity. 2279175104Sjeff */ 2280301456Skib td_get_sched(td)->ts_runtime += tickincr; 2281175104Sjeff sched_interact_update(td); 2282177009Sjeff sched_priority(td); 2283175104Sjeff } 2284239185Smav 2285113357Sjeff /* 2286239185Smav * Force a context switch if the current thread has used up a full 2287239185Smav * time slice (default is 100ms). 2288109864Sjeff */ 2289242736Sjeff if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) { 2290242736Sjeff ts->ts_slice = 0; 2291239185Smav td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND; 2292239185Smav } 2293109864Sjeff} 2294109864Sjeff 2295298145Skibu_int 2296298145Skibsched_estcpu(struct thread *td __unused) 2297171482Sjeff{ 2298171482Sjeff 2299298145Skib return (0); 2300171482Sjeff} 2301171482Sjeff 2302171482Sjeff/* 2303171482Sjeff * Return whether the current CPU has runnable tasks. Used for in-kernel 2304171482Sjeff * cooperative idle threads. 2305171482Sjeff */ 2306109864Sjeffint 2307109864Sjeffsched_runnable(void) 2308109864Sjeff{ 2309164936Sjulian struct tdq *tdq; 2310115998Sjeff int load; 2311109864Sjeff 2312115998Sjeff load = 1; 2313115998Sjeff 2314164936Sjulian tdq = TDQ_SELF(); 2315121605Sjeff if ((curthread->td_flags & TDF_IDLETD) != 0) { 2316165620Sjeff if (tdq->tdq_load > 0) 2317121605Sjeff goto out; 2318121605Sjeff } else 2319165620Sjeff if (tdq->tdq_load - 1 > 0) 2320121605Sjeff goto out; 2321115998Sjeff load = 0; 2322115998Sjeffout: 2323115998Sjeff return (load); 2324109864Sjeff} 2325109864Sjeff 2326171482Sjeff/* 2327171482Sjeff * Choose the highest priority thread to run. The thread is removed from 2328171482Sjeff * the run-queue while running however the load remains. For SMP we set 2329171482Sjeff * the tdq in the global idle bitmask if it idles here. 2330171482Sjeff */ 2331166190Sjeffstruct thread * 2332109970Sjeffsched_choose(void) 2333109970Sjeff{ 2334177435Sjeff struct thread *td; 2335164936Sjulian struct tdq *tdq; 2336109970Sjeff 2337164936Sjulian tdq = TDQ_SELF(); 2338171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2339177435Sjeff td = tdq_choose(tdq); 2340177435Sjeff if (td) { 2341177435Sjeff tdq_runq_rem(tdq, td); 2342177903Sjeff tdq->tdq_lowpri = td->td_priority; 2343177435Sjeff return (td); 2344109864Sjeff } 2345177903Sjeff tdq->tdq_lowpri = PRI_MAX_IDLE; 2346176735Sjeff return (PCPU_GET(idlethread)); 2347109864Sjeff} 2348109864Sjeff 2349171482Sjeff/* 2350171482Sjeff * Set owepreempt if necessary. Preemption never happens directly in ULE, 2351171482Sjeff * we always request it once we exit a critical section. 2352171482Sjeff */ 2353171482Sjeffstatic inline void 2354171482Sjeffsched_setpreempt(struct thread *td) 2355166190Sjeff{ 2356166190Sjeff struct thread *ctd; 2357166190Sjeff int cpri; 2358166190Sjeff int pri; 2359166190Sjeff 2360177005Sjeff THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2361177005Sjeff 2362166190Sjeff ctd = curthread; 2363166190Sjeff pri = td->td_priority; 2364166190Sjeff cpri = ctd->td_priority; 2365177005Sjeff if (pri < cpri) 2366177005Sjeff ctd->td_flags |= TDF_NEEDRESCHED; 2367166190Sjeff if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2368171482Sjeff return; 2369177005Sjeff if (!sched_shouldpreempt(pri, cpri, 0)) 2370171482Sjeff return; 2371171482Sjeff ctd->td_owepreempt = 1; 2372166190Sjeff} 2373166190Sjeff 2374171482Sjeff/* 2375177009Sjeff * Add a thread to a thread queue. Select the appropriate runq and add the 2376177009Sjeff * thread to it. This is the internal function called when the tdq is 2377177009Sjeff * predetermined. 2378171482Sjeff */ 2379109864Sjeffvoid 2380171482Sjefftdq_add(struct tdq *tdq, struct thread *td, int flags) 2381109864Sjeff{ 2382109864Sjeff 2383171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2384166190Sjeff KASSERT((td->td_inhibitors == 0), 2385166190Sjeff ("sched_add: trying to run inhibited thread")); 2386166190Sjeff KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2387166190Sjeff ("sched_add: bad thread state")); 2388172207Sjeff KASSERT(td->td_flags & TDF_INMEM, 2389172207Sjeff ("sched_add: thread swapped out")); 2390171482Sjeff 2391171482Sjeff if (td->td_priority < tdq->tdq_lowpri) 2392171482Sjeff tdq->tdq_lowpri = td->td_priority; 2393177435Sjeff tdq_runq_add(tdq, td, flags); 2394177435Sjeff tdq_load_add(tdq, td); 2395171482Sjeff} 2396171482Sjeff 2397171482Sjeff/* 2398171482Sjeff * Select the target thread queue and add a thread to it. Request 2399171482Sjeff * preemption or IPI a remote processor if required. 2400171482Sjeff */ 2401171482Sjeffvoid 2402171482Sjeffsched_add(struct thread *td, int flags) 2403171482Sjeff{ 2404171482Sjeff struct tdq *tdq; 2405171482Sjeff#ifdef SMP 2406171482Sjeff int cpu; 2407171482Sjeff#endif 2408187357Sjeff 2409187357Sjeff KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 2410187357Sjeff "prio:%d", td->td_priority, KTR_ATTR_LINKED, 2411187357Sjeff sched_tdname(curthread)); 2412187357Sjeff KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 2413187357Sjeff KTR_ATTR_LINKED, sched_tdname(td)); 2414235459Srstone SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 2415235459Srstone flags & SRQ_PREEMPTED); 2416171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2417166108Sjeff /* 2418171482Sjeff * Recalculate the priority before we select the target cpu or 2419171482Sjeff * run-queue. 2420166108Sjeff */ 2421171482Sjeff if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2422171482Sjeff sched_priority(td); 2423171482Sjeff#ifdef SMP 2424171482Sjeff /* 2425171482Sjeff * Pick the destination cpu and if it isn't ours transfer to the 2426171482Sjeff * target cpu. 2427171482Sjeff */ 2428177435Sjeff cpu = sched_pickcpu(td, flags); 2429177435Sjeff tdq = sched_setcpu(td, cpu, flags); 2430171482Sjeff tdq_add(tdq, td, flags); 2431177009Sjeff if (cpu != PCPU_GET(cpuid)) { 2432177435Sjeff tdq_notify(tdq, td); 2433166108Sjeff return; 2434166108Sjeff } 2435171482Sjeff#else 2436171482Sjeff tdq = TDQ_SELF(); 2437171482Sjeff TDQ_LOCK(tdq); 2438171482Sjeff /* 2439171482Sjeff * Now that the thread is moving to the run-queue, set the lock 2440171482Sjeff * to the scheduler's lock. 2441171482Sjeff */ 2442171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2443171482Sjeff tdq_add(tdq, td, flags); 2444166108Sjeff#endif 2445171482Sjeff if (!(flags & SRQ_YIELDING)) 2446171482Sjeff sched_setpreempt(td); 2447109864Sjeff} 2448109864Sjeff 2449171482Sjeff/* 2450171482Sjeff * Remove a thread from a run-queue without running it. This is used 2451171482Sjeff * when we're stealing a thread from a remote queue. Otherwise all threads 2452171482Sjeff * exit by calling sched_exit_thread() and sched_throw() themselves. 2453171482Sjeff */ 2454109864Sjeffvoid 2455121127Sjeffsched_rem(struct thread *td) 2456109864Sjeff{ 2457164936Sjulian struct tdq *tdq; 2458113357Sjeff 2459187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 2460187357Sjeff "prio:%d", td->td_priority); 2461235459Srstone SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); 2462301456Skib tdq = TDQ_CPU(td_get_sched(td)->ts_cpu); 2463171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2464171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2465166190Sjeff KASSERT(TD_ON_RUNQ(td), 2466164936Sjulian ("sched_rem: thread not on run queue")); 2467177435Sjeff tdq_runq_rem(tdq, td); 2468177435Sjeff tdq_load_rem(tdq, td); 2469166190Sjeff TD_SET_CAN_RUN(td); 2470176735Sjeff if (td->td_priority == tdq->tdq_lowpri) 2471176735Sjeff tdq_setlowpri(tdq, NULL); 2472109864Sjeff} 2473109864Sjeff 2474171482Sjeff/* 2475171482Sjeff * Fetch cpu utilization information. Updates on demand. 2476171482Sjeff */ 2477109864Sjefffixpt_t 2478121127Sjeffsched_pctcpu(struct thread *td) 2479109864Sjeff{ 2480109864Sjeff fixpt_t pctcpu; 2481164936Sjulian struct td_sched *ts; 2482109864Sjeff 2483109864Sjeff pctcpu = 0; 2484301456Skib ts = td_get_sched(td); 2485109864Sjeff 2486208787Sjhb THREAD_LOCK_ASSERT(td, MA_OWNED); 2487232917Smav sched_pctcpu_update(ts, TD_IS_RUNNING(td)); 2488164936Sjulian if (ts->ts_ticks) { 2489109864Sjeff int rtick; 2490109864Sjeff 2491109864Sjeff /* How many rtick per second ? */ 2492165762Sjeff rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2493165762Sjeff pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2494109864Sjeff } 2495109864Sjeff 2496109864Sjeff return (pctcpu); 2497109864Sjeff} 2498109864Sjeff 2499176735Sjeff/* 2500176735Sjeff * Enforce affinity settings for a thread. Called after adjustments to 2501176735Sjeff * cpumask. 2502176735Sjeff */ 2503176729Sjeffvoid 2504176729Sjeffsched_affinity(struct thread *td) 2505176729Sjeff{ 2506176735Sjeff#ifdef SMP 2507176735Sjeff struct td_sched *ts; 2508176735Sjeff 2509176735Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2510301456Skib ts = td_get_sched(td); 2511176735Sjeff if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2512176735Sjeff return; 2513189787Sjeff if (TD_ON_RUNQ(td)) { 2514189787Sjeff sched_rem(td); 2515189787Sjeff sched_add(td, SRQ_BORING); 2516189787Sjeff return; 2517189787Sjeff } 2518176735Sjeff if (!TD_IS_RUNNING(td)) 2519176735Sjeff return; 2520176735Sjeff /* 2521212115Smdf * Force a switch before returning to userspace. If the 2522212115Smdf * target thread is not running locally send an ipi to force 2523212115Smdf * the issue. 2524176735Sjeff */ 2525212974Sjhb td->td_flags |= TDF_NEEDRESCHED; 2526212115Smdf if (td != curthread) 2527212115Smdf ipi_cpu(ts->ts_cpu, IPI_PREEMPT); 2528176735Sjeff#endif 2529176729Sjeff} 2530176729Sjeff 2531171482Sjeff/* 2532171482Sjeff * Bind a thread to a target cpu. 2533171482Sjeff */ 2534122038Sjeffvoid 2535122038Sjeffsched_bind(struct thread *td, int cpu) 2536122038Sjeff{ 2537164936Sjulian struct td_sched *ts; 2538122038Sjeff 2539171713Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2540208391Sjhb KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 2541301456Skib ts = td_get_sched(td); 2542166137Sjeff if (ts->ts_flags & TSF_BOUND) 2543166152Sjeff sched_unbind(td); 2544212115Smdf KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td)); 2545164936Sjulian ts->ts_flags |= TSF_BOUND; 2546166137Sjeff sched_pin(); 2547123433Sjeff if (PCPU_GET(cpuid) == cpu) 2548122038Sjeff return; 2549166137Sjeff ts->ts_cpu = cpu; 2550122038Sjeff /* When we return from mi_switch we'll be on the correct cpu. */ 2551131527Sphk mi_switch(SW_VOL, NULL); 2552122038Sjeff} 2553122038Sjeff 2554171482Sjeff/* 2555171482Sjeff * Release a bound thread. 2556171482Sjeff */ 2557122038Sjeffvoid 2558122038Sjeffsched_unbind(struct thread *td) 2559122038Sjeff{ 2560165762Sjeff struct td_sched *ts; 2561165762Sjeff 2562170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2563208391Sjhb KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 2564301456Skib ts = td_get_sched(td); 2565166137Sjeff if ((ts->ts_flags & TSF_BOUND) == 0) 2566166137Sjeff return; 2567165762Sjeff ts->ts_flags &= ~TSF_BOUND; 2568165762Sjeff sched_unpin(); 2569122038Sjeff} 2570122038Sjeff 2571109864Sjeffint 2572145256Sjkoshysched_is_bound(struct thread *td) 2573145256Sjkoshy{ 2574170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2575301456Skib return (td_get_sched(td)->ts_flags & TSF_BOUND); 2576145256Sjkoshy} 2577145256Sjkoshy 2578171482Sjeff/* 2579171482Sjeff * Basic yield call. 2580171482Sjeff */ 2581159630Sdavidxuvoid 2582159630Sdavidxusched_relinquish(struct thread *td) 2583159630Sdavidxu{ 2584170293Sjeff thread_lock(td); 2585178272Sjeff mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 2586170293Sjeff thread_unlock(td); 2587159630Sdavidxu} 2588159630Sdavidxu 2589171482Sjeff/* 2590171482Sjeff * Return the total system load. 2591171482Sjeff */ 2592145256Sjkoshyint 2593125289Sjeffsched_load(void) 2594125289Sjeff{ 2595125289Sjeff#ifdef SMP 2596125289Sjeff int total; 2597125289Sjeff int i; 2598125289Sjeff 2599125289Sjeff total = 0; 2600209059Sjhb CPU_FOREACH(i) 2601176735Sjeff total += TDQ_CPU(i)->tdq_sysload; 2602125289Sjeff return (total); 2603125289Sjeff#else 2604165620Sjeff return (TDQ_SELF()->tdq_sysload); 2605125289Sjeff#endif 2606125289Sjeff} 2607125289Sjeff 2608125289Sjeffint 2609109864Sjeffsched_sizeof_proc(void) 2610109864Sjeff{ 2611109864Sjeff return (sizeof(struct proc)); 2612109864Sjeff} 2613109864Sjeff 2614109864Sjeffint 2615109864Sjeffsched_sizeof_thread(void) 2616109864Sjeff{ 2617109864Sjeff return (sizeof(struct thread) + sizeof(struct td_sched)); 2618109864Sjeff} 2619159570Sdavidxu 2620191676Sjeff#ifdef SMP 2621191676Sjeff#define TDQ_IDLESPIN(tdq) \ 2622191676Sjeff ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0) 2623191676Sjeff#else 2624191676Sjeff#define TDQ_IDLESPIN(tdq) 1 2625191676Sjeff#endif 2626191676Sjeff 2627166190Sjeff/* 2628166190Sjeff * The actual idle process. 2629166190Sjeff */ 2630166190Sjeffvoid 2631166190Sjeffsched_idletd(void *dummy) 2632166190Sjeff{ 2633166190Sjeff struct thread *td; 2634171482Sjeff struct tdq *tdq; 2635242852Smav int oldswitchcnt, switchcnt; 2636178277Sjeff int i; 2637166190Sjeff 2638191643Sjeff mtx_assert(&Giant, MA_NOTOWNED); 2639166190Sjeff td = curthread; 2640171482Sjeff tdq = TDQ_SELF(); 2641239585Sjhb THREAD_NO_SLEEPING(); 2642242852Smav oldswitchcnt = -1; 2643171482Sjeff for (;;) { 2644242852Smav if (tdq->tdq_load) { 2645242852Smav thread_lock(td); 2646242852Smav mi_switch(SW_VOL | SWT_IDLE, NULL); 2647242852Smav thread_unlock(td); 2648242852Smav } 2649242852Smav switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2650171482Sjeff#ifdef SMP 2651242852Smav if (switchcnt != oldswitchcnt) { 2652242852Smav oldswitchcnt = switchcnt; 2653242852Smav if (tdq_idled(tdq) == 0) 2654242852Smav continue; 2655242852Smav } 2656243069Smav switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2657243069Smav#else 2658243069Smav oldswitchcnt = switchcnt; 2659171482Sjeff#endif 2660178277Sjeff /* 2661178277Sjeff * If we're switching very frequently, spin while checking 2662178277Sjeff * for load rather than entering a low power state that 2663191643Sjeff * may require an IPI. However, don't do any busy 2664191643Sjeff * loops while on SMT machines as this simply steals 2665191643Sjeff * cycles from cores doing useful work. 2666178277Sjeff */ 2667191676Sjeff if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) { 2668178277Sjeff for (i = 0; i < sched_idlespins; i++) { 2669178277Sjeff if (tdq->tdq_load) 2670178277Sjeff break; 2671178277Sjeff cpu_spinwait(); 2672178277Sjeff } 2673178277Sjeff } 2674242852Smav 2675242852Smav /* If there was context switch during spin, restart it. */ 2676191643Sjeff switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2677242852Smav if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt) 2678242852Smav continue; 2679242852Smav 2680242852Smav /* Run main MD idle handler. */ 2681242852Smav tdq->tdq_cpu_idle = 1; 2682271616Smav /* 2683271616Smav * Make sure that tdq_cpu_idle update is globally visible 2684271616Smav * before cpu_idle() read tdq_load. The order is important 2685271616Smav * to avoid race with tdq_notify. 2686271616Smav */ 2687285353Skib atomic_thread_fence_seq_cst(); 2688242852Smav cpu_idle(switchcnt * 4 > sched_idlespinthresh); 2689242852Smav tdq->tdq_cpu_idle = 0; 2690242852Smav 2691242852Smav /* 2692242852Smav * Account thread-less hardware interrupts and 2693242852Smav * other wakeup reasons equal to context switches. 2694242852Smav */ 2695242852Smav switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2696242852Smav if (switchcnt != oldswitchcnt) 2697242852Smav continue; 2698242852Smav tdq->tdq_switchcnt++; 2699242852Smav oldswitchcnt++; 2700171482Sjeff } 2701166190Sjeff} 2702166190Sjeff 2703170293Sjeff/* 2704170293Sjeff * A CPU is entering for the first time or a thread is exiting. 2705170293Sjeff */ 2706170293Sjeffvoid 2707170293Sjeffsched_throw(struct thread *td) 2708170293Sjeff{ 2709172411Sjeff struct thread *newtd; 2710171482Sjeff struct tdq *tdq; 2711171482Sjeff 2712171482Sjeff tdq = TDQ_SELF(); 2713170293Sjeff if (td == NULL) { 2714171482Sjeff /* Correct spinlock nesting and acquire the correct lock. */ 2715171482Sjeff TDQ_LOCK(tdq); 2716170293Sjeff spinlock_exit(); 2717229429Sjhb PCPU_SET(switchtime, cpu_ticks()); 2718229429Sjhb PCPU_SET(switchticks, ticks); 2719170293Sjeff } else { 2720171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2721177435Sjeff tdq_load_rem(tdq, td); 2722174629Sjeff lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2723286256Sjhb td->td_lastcpu = td->td_oncpu; 2724286256Sjhb td->td_oncpu = NOCPU; 2725170293Sjeff } 2726170293Sjeff KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2727172411Sjeff newtd = choosethread(); 2728172411Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2729172411Sjeff cpu_throw(td, newtd); /* doesn't return */ 2730170293Sjeff} 2731170293Sjeff 2732171482Sjeff/* 2733171482Sjeff * This is called from fork_exit(). Just acquire the correct locks and 2734171482Sjeff * let fork do the rest of the work. 2735171482Sjeff */ 2736170293Sjeffvoid 2737170600Sjeffsched_fork_exit(struct thread *td) 2738170293Sjeff{ 2739171482Sjeff struct tdq *tdq; 2740171482Sjeff int cpuid; 2741170293Sjeff 2742170293Sjeff /* 2743170293Sjeff * Finish setting up thread glue so that it begins execution in a 2744171482Sjeff * non-nested critical section with the scheduler lock held. 2745170293Sjeff */ 2746171482Sjeff cpuid = PCPU_GET(cpuid); 2747171482Sjeff tdq = TDQ_CPU(cpuid); 2748171482Sjeff if (TD_IS_IDLETHREAD(td)) 2749171482Sjeff td->td_lock = TDQ_LOCKPTR(tdq); 2750171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2751171482Sjeff td->td_oncpu = cpuid; 2752172411Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2753174629Sjeff lock_profile_obtain_lock_success( 2754174629Sjeff &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2755170293Sjeff} 2756170293Sjeff 2757187357Sjeff/* 2758187357Sjeff * Create on first use to catch odd startup conditons. 2759187357Sjeff */ 2760187357Sjeffchar * 2761187357Sjeffsched_tdname(struct thread *td) 2762187357Sjeff{ 2763187357Sjeff#ifdef KTR 2764187357Sjeff struct td_sched *ts; 2765187357Sjeff 2766301456Skib ts = td_get_sched(td); 2767187357Sjeff if (ts->ts_name[0] == '\0') 2768187357Sjeff snprintf(ts->ts_name, sizeof(ts->ts_name), 2769187357Sjeff "%s tid %d", td->td_name, td->td_tid); 2770187357Sjeff return (ts->ts_name); 2771187357Sjeff#else 2772187357Sjeff return (td->td_name); 2773187357Sjeff#endif 2774187357Sjeff} 2775187357Sjeff 2776232700Sjhb#ifdef KTR 2777232700Sjhbvoid 2778232700Sjhbsched_clear_tdname(struct thread *td) 2779232700Sjhb{ 2780232700Sjhb struct td_sched *ts; 2781232700Sjhb 2782301456Skib ts = td_get_sched(td); 2783232700Sjhb ts->ts_name[0] = '\0'; 2784232700Sjhb} 2785232700Sjhb#endif 2786232700Sjhb 2787184439Sivoras#ifdef SMP 2788184439Sivoras 2789184439Sivoras/* 2790184439Sivoras * Build the CPU topology dump string. Is recursively called to collect 2791184439Sivoras * the topology tree. 2792184439Sivoras */ 2793184439Sivorasstatic int 2794184439Sivorassysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg, 2795184439Sivoras int indent) 2796184439Sivoras{ 2797222813Sattilio char cpusetbuf[CPUSETBUFSIZ]; 2798184439Sivoras int i, first; 2799184439Sivoras 2800184439Sivoras sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent, 2801212821Savg "", 1 + indent / 2, cg->cg_level); 2802222813Sattilio sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "", 2803222813Sattilio cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask)); 2804184439Sivoras first = TRUE; 2805184439Sivoras for (i = 0; i < MAXCPU; i++) { 2806222813Sattilio if (CPU_ISSET(i, &cg->cg_mask)) { 2807184439Sivoras if (!first) 2808184439Sivoras sbuf_printf(sb, ", "); 2809184439Sivoras else 2810184439Sivoras first = FALSE; 2811184439Sivoras sbuf_printf(sb, "%d", i); 2812184439Sivoras } 2813184439Sivoras } 2814184439Sivoras sbuf_printf(sb, "</cpu>\n"); 2815184439Sivoras 2816184439Sivoras if (cg->cg_flags != 0) { 2817210117Sivoras sbuf_printf(sb, "%*s <flags>", indent, ""); 2818184439Sivoras if ((cg->cg_flags & CG_FLAG_HTT) != 0) 2819208982Sivoras sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>"); 2820208983Sivoras if ((cg->cg_flags & CG_FLAG_THREAD) != 0) 2821208983Sivoras sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>"); 2822191643Sjeff if ((cg->cg_flags & CG_FLAG_SMT) != 0) 2823208983Sivoras sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>"); 2824210117Sivoras sbuf_printf(sb, "</flags>\n"); 2825184439Sivoras } 2826184439Sivoras 2827184439Sivoras if (cg->cg_children > 0) { 2828184439Sivoras sbuf_printf(sb, "%*s <children>\n", indent, ""); 2829184439Sivoras for (i = 0; i < cg->cg_children; i++) 2830184439Sivoras sysctl_kern_sched_topology_spec_internal(sb, 2831184439Sivoras &cg->cg_child[i], indent+2); 2832184439Sivoras sbuf_printf(sb, "%*s </children>\n", indent, ""); 2833184439Sivoras } 2834184439Sivoras sbuf_printf(sb, "%*s</group>\n", indent, ""); 2835184439Sivoras return (0); 2836184439Sivoras} 2837184439Sivoras 2838184439Sivoras/* 2839184439Sivoras * Sysctl handler for retrieving topology dump. It's a wrapper for 2840184439Sivoras * the recursive sysctl_kern_smp_topology_spec_internal(). 2841184439Sivoras */ 2842184439Sivorasstatic int 2843184439Sivorassysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS) 2844184439Sivoras{ 2845184439Sivoras struct sbuf *topo; 2846184439Sivoras int err; 2847184439Sivoras 2848184439Sivoras KASSERT(cpu_top != NULL, ("cpu_top isn't initialized")); 2849184439Sivoras 2850279999Sian topo = sbuf_new_for_sysctl(NULL, NULL, 512, req); 2851184439Sivoras if (topo == NULL) 2852184439Sivoras return (ENOMEM); 2853184439Sivoras 2854184439Sivoras sbuf_printf(topo, "<groups>\n"); 2855184439Sivoras err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1); 2856184439Sivoras sbuf_printf(topo, "</groups>\n"); 2857184439Sivoras 2858184439Sivoras if (err == 0) { 2859279999Sian err = sbuf_finish(topo); 2860184439Sivoras } 2861184439Sivoras sbuf_delete(topo); 2862184439Sivoras return (err); 2863184439Sivoras} 2864214510Sdavidxu 2865184439Sivoras#endif 2866184439Sivoras 2867239185Smavstatic int 2868239185Smavsysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 2869239185Smav{ 2870239185Smav int error, new_val, period; 2871239185Smav 2872239185Smav period = 1000000 / realstathz; 2873239185Smav new_val = period * sched_slice; 2874239185Smav error = sysctl_handle_int(oidp, &new_val, 0, req); 2875239196Smav if (error != 0 || req->newptr == NULL) 2876239185Smav return (error); 2877239185Smav if (new_val <= 0) 2878239185Smav return (EINVAL); 2879239196Smav sched_slice = imax(1, (new_val + period / 2) / period); 2880242736Sjeff sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR; 2881239196Smav hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 2882239196Smav realstathz); 2883239185Smav return (0); 2884239185Smav} 2885239185Smav 2886177435SjeffSYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2887171482SjeffSYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2888165762Sjeff "Scheduler name"); 2889239185SmavSYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW, 2890239185Smav NULL, 0, sysctl_kern_quantum, "I", 2891239196Smav "Quantum for timeshare threads in microseconds"); 2892171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2893239196Smav "Quantum for timeshare threads in stathz ticks"); 2894171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2895239196Smav "Interactivity score threshold"); 2896239196SmavSYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, 2897239196Smav &preempt_thresh, 0, 2898239196Smav "Maximal (lowest) priority for preemption"); 2899239196SmavSYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0, 2900239196Smav "Assign static kernel priorities to sleeping threads"); 2901239196SmavSYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0, 2902239196Smav "Number of times idle thread will spin waiting for new work"); 2903239196SmavSYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, 2904239196Smav &sched_idlespinthresh, 0, 2905239196Smav "Threshold before we will permit idle thread spinning"); 2906166108Sjeff#ifdef SMP 2907171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2908171482Sjeff "Number of hz ticks to keep thread affinity for"); 2909171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2910171482Sjeff "Enables the long-term load balancer"); 2911172409SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2912172409Sjeff &balance_interval, 0, 2913239185Smav "Average period in stathz ticks to run the long-term balancer"); 2914171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2915171482Sjeff "Attempts to steal work from other cores before idling"); 2916171506SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2917239196Smav "Minimum load on remote CPU before we'll steal"); 2918184439SivorasSYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING | 2919239185Smav CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A", 2920184439Sivoras "XML dump of detected CPU topology"); 2921166108Sjeff#endif 2922165762Sjeff 2923172264Sjeff/* ps compat. All cpu percentages from ULE are weighted. */ 2924172293Sjeffstatic int ccpu = 0; 2925165762SjeffSYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2926