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("$FreeBSD$"); 40116182Sobrien 41147565Speter#include "opt_hwpmc_hooks.h" 42179297Sjb#include "opt_kdtrace.h" 43147565Speter#include "opt_sched.h" 44134649Sscottl 45109864Sjeff#include <sys/param.h> 46109864Sjeff#include <sys/systm.h> 47131929Smarcel#include <sys/kdb.h> 48109864Sjeff#include <sys/kernel.h> 49109864Sjeff#include <sys/ktr.h> 50109864Sjeff#include <sys/lock.h> 51109864Sjeff#include <sys/mutex.h> 52109864Sjeff#include <sys/proc.h> 53112966Sjeff#include <sys/resource.h> 54122038Sjeff#include <sys/resourcevar.h> 55109864Sjeff#include <sys/sched.h> 56236344Srstone#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 80230691Smarius#if defined(__powerpc__) && defined(E500) 81172345Sjeff#error "This architecture is not currently compatible with ULE" 82166190Sjeff#endif 83166190Sjeff 84171482Sjeff#define KTR_ULE 0 85166137Sjeff 86187679Sjeff#define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX))) 87187679Sjeff#define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU))) 88224221Sattilio#define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load")) 89187357Sjeff 90166137Sjeff/* 91171482Sjeff * Thread scheduler specific section. All fields are protected 92171482Sjeff * by the thread lock. 93146954Sjeff */ 94164936Sjulianstruct td_sched { 95171482Sjeff struct runq *ts_runq; /* Run-queue we're queued on. */ 96171482Sjeff short ts_flags; /* TSF_* flags. */ 97164936Sjulian u_char ts_cpu; /* CPU that we have affinity for. */ 98177009Sjeff int ts_rltick; /* Real last tick, for affinity. */ 99171482Sjeff int ts_slice; /* Ticks of slice remaining. */ 100171482Sjeff u_int ts_slptime; /* Number of ticks we vol. slept */ 101171482Sjeff u_int ts_runtime; /* Number of ticks we were running */ 102164936Sjulian int ts_ltick; /* Last tick that we were running on */ 103164936Sjulian int ts_ftick; /* First tick that we were running on */ 104164936Sjulian int ts_ticks; /* Tick count */ 105187357Sjeff#ifdef KTR 106187357Sjeff char ts_name[TS_NAME_LEN]; 107187357Sjeff#endif 108134791Sjulian}; 109164936Sjulian/* flags kept in ts_flags */ 110166108Sjeff#define TSF_BOUND 0x0001 /* Thread can not migrate. */ 111166108Sjeff#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 112121790Sjeff 113164936Sjulianstatic struct td_sched td_sched0; 114109864Sjeff 115176735Sjeff#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 116176735Sjeff#define THREAD_CAN_SCHED(td, cpu) \ 117176735Sjeff CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 118176735Sjeff 119109864Sjeff/* 120217351Sjhb * Priority ranges used for interactive and non-interactive timeshare 121217410Sjhb * threads. The timeshare priorities are split up into four ranges. 122217410Sjhb * The first range handles interactive threads. The last three ranges 123217410Sjhb * (NHALF, x, and NHALF) handle non-interactive threads with the outer 124217410Sjhb * ranges supporting nice values. 125217351Sjhb */ 126217410Sjhb#define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1) 127217410Sjhb#define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2) 128230173Savg#define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE) 129217410Sjhb 130217410Sjhb#define PRI_MIN_INTERACT PRI_MIN_TIMESHARE 131217410Sjhb#define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1) 132217410Sjhb#define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE) 133217351Sjhb#define PRI_MAX_BATCH PRI_MAX_TIMESHARE 134217351Sjhb 135217351Sjhb/* 136165762Sjeff * Cpu percentage computation macros and defines. 137111857Sjeff * 138165762Sjeff * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 139165762Sjeff * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 140165796Sjeff * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 141165762Sjeff * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 142165762Sjeff * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 143165762Sjeff * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 144165762Sjeff */ 145165762Sjeff#define SCHED_TICK_SECS 10 146165762Sjeff#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 147165796Sjeff#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 148165762Sjeff#define SCHED_TICK_SHIFT 10 149165762Sjeff#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 150165830Sjeff#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 151165762Sjeff 152165762Sjeff/* 153165762Sjeff * These macros determine priorities for non-interactive threads. They are 154165762Sjeff * assigned a priority based on their recent cpu utilization as expressed 155165762Sjeff * by the ratio of ticks to the tick total. NHALF priorities at the start 156165762Sjeff * and end of the MIN to MAX timeshare range are only reachable with negative 157165762Sjeff * or positive nice respectively. 158165762Sjeff * 159165762Sjeff * PRI_RANGE: Priority range for utilization dependent priorities. 160116642Sjeff * PRI_NRESV: Number of nice values. 161165762Sjeff * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 162165762Sjeff * PRI_NICE: Determines the part of the priority inherited from nice. 163109864Sjeff */ 164165762Sjeff#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 165121869Sjeff#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 166217351Sjhb#define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF) 167217351Sjhb#define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF) 168217237Sjhb#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1) 169165762Sjeff#define SCHED_PRI_TICKS(ts) \ 170165762Sjeff (SCHED_TICK_HZ((ts)) / \ 171165827Sjeff (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 172165762Sjeff#define SCHED_PRI_NICE(nice) (nice) 173109864Sjeff 174109864Sjeff/* 175165762Sjeff * These determine the interactivity of a process. Interactivity differs from 176165762Sjeff * cpu utilization in that it expresses the voluntary time slept vs time ran 177165762Sjeff * while cpu utilization includes all time not running. This more accurately 178165762Sjeff * models the intent of the thread. 179109864Sjeff * 180110645Sjeff * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 181110645Sjeff * before throttling back. 182121868Sjeff * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 183116365Sjeff * INTERACT_MAX: Maximum interactivity value. Smaller is better. 184215102Sattilio * INTERACT_THRESH: Threshold for placement on the current runq. 185109864Sjeff */ 186165762Sjeff#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 187165762Sjeff#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 188116365Sjeff#define SCHED_INTERACT_MAX (100) 189116365Sjeff#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 190121126Sjeff#define SCHED_INTERACT_THRESH (30) 191111857Sjeff 192241248Smav/* Flags kept in td_flags. */ 193241248Smav#define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */ 194241248Smav 195109864Sjeff/* 196165762Sjeff * tickincr: Converts a stathz tick into a hz domain scaled by 197165762Sjeff * the shift factor. Without the shift the error rate 198165762Sjeff * due to rounding would be unacceptably high. 199165762Sjeff * realstathz: stathz is sometimes 0 and run off of hz. 200165762Sjeff * sched_slice: Runtime of each thread before rescheduling. 201171482Sjeff * preempt_thresh: Priority threshold for preemption and remote IPIs. 202109864Sjeff */ 203165762Sjeffstatic int sched_interact = SCHED_INTERACT_THRESH; 204241249Smavstatic int realstathz = 127; 205242544Seadlerstatic int tickincr = 8 << SCHED_TICK_SHIFT; 206241249Smavstatic int sched_slice = 12; 207172345Sjeff#ifdef PREEMPTION 208172345Sjeff#ifdef FULL_PREEMPTION 209172345Sjeffstatic int preempt_thresh = PRI_MAX_IDLE; 210172345Sjeff#else 211171482Sjeffstatic int preempt_thresh = PRI_MIN_KERN; 212172345Sjeff#endif 213172345Sjeff#else 214172345Sjeffstatic int preempt_thresh = 0; 215172345Sjeff#endif 216217351Sjhbstatic int static_boost = PRI_MIN_BATCH; 217178277Sjeffstatic int sched_idlespins = 10000; 218236546Smavstatic int sched_idlespinthresh = -1; 219109864Sjeff 220109864Sjeff/* 221171482Sjeff * tdq - per processor runqs and statistics. All fields are protected by the 222171482Sjeff * tdq_lock. The load and lowpri may be accessed without to avoid excess 223171482Sjeff * locking in sched_pickcpu(); 224109864Sjeff */ 225164936Sjulianstruct tdq { 226177009Sjeff /* Ordered to improve efficiency of cpu_search() and switch(). */ 227177009Sjeff struct mtx tdq_lock; /* run queue lock. */ 228176735Sjeff struct cpu_group *tdq_cg; /* Pointer to cpu topology. */ 229178277Sjeff volatile int tdq_load; /* Aggregate load. */ 230212416Smav volatile int tdq_cpu_idle; /* cpu_idle() is active. */ 231176735Sjeff int tdq_sysload; /* For loadavg, !ITHD load. */ 232177009Sjeff int tdq_transferable; /* Transferable thread count. */ 233178277Sjeff short tdq_switchcnt; /* Switches this tick. */ 234178277Sjeff short tdq_oldswitchcnt; /* Switches last tick. */ 235177009Sjeff u_char tdq_lowpri; /* Lowest priority thread. */ 236177009Sjeff u_char tdq_ipipending; /* IPI pending. */ 237166557Sjeff u_char tdq_idx; /* Current insert index. */ 238166557Sjeff u_char tdq_ridx; /* Current removal index. */ 239177009Sjeff struct runq tdq_realtime; /* real-time run queue. */ 240177009Sjeff struct runq tdq_timeshare; /* timeshare run queue. */ 241177009Sjeff struct runq tdq_idle; /* Queue of IDLE threads. */ 242187357Sjeff char tdq_name[TDQ_NAME_LEN]; 243187357Sjeff#ifdef KTR 244187357Sjeff char tdq_loadname[TDQ_LOADNAME_LEN]; 245187357Sjeff#endif 246171482Sjeff} __aligned(64); 247109864Sjeff 248178277Sjeff/* Idle thread states and config. */ 249178277Sjeff#define TDQ_RUNNING 1 250178277Sjeff#define TDQ_IDLE 2 251166108Sjeff 252123433Sjeff#ifdef SMP 253184439Sivorasstruct cpu_group *cpu_top; /* CPU topology */ 254123433Sjeff 255176735Sjeff#define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000)) 256176735Sjeff#define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity)) 257166108Sjeff 258123433Sjeff/* 259166108Sjeff * Run-time tunables. 260166108Sjeff */ 261171506Sjeffstatic int rebalance = 1; 262172409Sjeffstatic int balance_interval = 128; /* Default set in sched_initticks(). */ 263166108Sjeffstatic int affinity; 264171506Sjeffstatic int steal_idle = 1; 265171506Sjeffstatic int steal_thresh = 2; 266166108Sjeff 267166108Sjeff/* 268165620Sjeff * One thread queue per processor. 269109864Sjeff */ 270164936Sjulianstatic struct tdq tdq_cpu[MAXCPU]; 271172409Sjeffstatic struct tdq *balance_tdq; 272172409Sjeffstatic int balance_ticks; 273233599Smavstatic DPCPU_DEFINE(uint32_t, randomval); 274129982Sjeff 275164936Sjulian#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 276164936Sjulian#define TDQ_CPU(x) (&tdq_cpu[(x)]) 277171713Sjeff#define TDQ_ID(x) ((int)((x) - tdq_cpu)) 278123433Sjeff#else /* !SMP */ 279164936Sjulianstatic struct tdq tdq_cpu; 280129982Sjeff 281170315Sjeff#define TDQ_ID(x) (0) 282164936Sjulian#define TDQ_SELF() (&tdq_cpu) 283164936Sjulian#define TDQ_CPU(x) (&tdq_cpu) 284110028Sjeff#endif 285109864Sjeff 286171482Sjeff#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 287171482Sjeff#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 288171482Sjeff#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 289171482Sjeff#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 290176735Sjeff#define TDQ_LOCKPTR(t) (&(t)->tdq_lock) 291171482Sjeff 292163709Sjbstatic void sched_priority(struct thread *); 293146954Sjeffstatic void sched_thread_priority(struct thread *, u_char); 294163709Sjbstatic int sched_interact_score(struct thread *); 295163709Sjbstatic void sched_interact_update(struct thread *); 296163709Sjbstatic void sched_interact_fork(struct thread *); 297234166Smavstatic void sched_pctcpu_update(struct td_sched *, int); 298109864Sjeff 299110267Sjeff/* Operations on per processor queues */ 300177435Sjeffstatic struct thread *tdq_choose(struct tdq *); 301164936Sjulianstatic void tdq_setup(struct tdq *); 302177435Sjeffstatic void tdq_load_add(struct tdq *, struct thread *); 303177435Sjeffstatic void tdq_load_rem(struct tdq *, struct thread *); 304177435Sjeffstatic __inline void tdq_runq_add(struct tdq *, struct thread *, int); 305177435Sjeffstatic __inline void tdq_runq_rem(struct tdq *, struct thread *); 306177005Sjeffstatic inline int sched_shouldpreempt(int, int, int); 307164936Sjulianvoid tdq_print(int cpu); 308165762Sjeffstatic void runq_print(struct runq *rq); 309171482Sjeffstatic void tdq_add(struct tdq *, struct thread *, int); 310110267Sjeff#ifdef SMP 311176735Sjeffstatic int tdq_move(struct tdq *, struct tdq *); 312171482Sjeffstatic int tdq_idled(struct tdq *); 313177435Sjeffstatic void tdq_notify(struct tdq *, struct thread *); 314177435Sjeffstatic struct thread *tdq_steal(struct tdq *, int); 315177435Sjeffstatic struct thread *runq_steal(struct runq *, int); 316177435Sjeffstatic int sched_pickcpu(struct thread *, int); 317172409Sjeffstatic void sched_balance(void); 318176735Sjeffstatic int sched_balance_pair(struct tdq *, struct tdq *); 319177435Sjeffstatic inline struct tdq *sched_setcpu(struct thread *, int, int); 320171482Sjeffstatic inline void thread_unblock_switch(struct thread *, struct mtx *); 321171713Sjeffstatic struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 322184439Sivorasstatic int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS); 323184439Sivorasstatic int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, 324184439Sivoras struct cpu_group *cg, int indent); 325121790Sjeff#endif 326110028Sjeff 327165762Sjeffstatic void sched_setup(void *dummy); 328177253SrwatsonSYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 329165762Sjeff 330165762Sjeffstatic void sched_initticks(void *dummy); 331177253SrwatsonSYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 332177253Srwatson NULL); 333165762Sjeff 334236344SrstoneSDT_PROVIDER_DEFINE(sched); 335236344Srstone 336262057SavgSDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *", 337236344Srstone "struct proc *", "uint8_t"); 338262057SavgSDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *", 339236344Srstone "struct proc *", "void *"); 340262057SavgSDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *", 341236344Srstone "struct proc *", "void *", "int"); 342262057SavgSDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *", 343236344Srstone "struct proc *", "uint8_t", "struct thread *"); 344262057SavgSDT_PROBE_DEFINE2(sched, , , load__change, "int", "int"); 345262057SavgSDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *", 346236344Srstone "struct proc *"); 347262057SavgSDT_PROBE_DEFINE(sched, , , on__cpu); 348262057SavgSDT_PROBE_DEFINE(sched, , , remain__cpu); 349262057SavgSDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *", 350236344Srstone "struct proc *"); 351236344Srstone 352171482Sjeff/* 353171482Sjeff * Print the threads waiting on a run-queue. 354171482Sjeff */ 355165762Sjeffstatic void 356165762Sjeffrunq_print(struct runq *rq) 357165762Sjeff{ 358165762Sjeff struct rqhead *rqh; 359177435Sjeff struct thread *td; 360165762Sjeff int pri; 361165762Sjeff int j; 362165762Sjeff int i; 363165762Sjeff 364165762Sjeff for (i = 0; i < RQB_LEN; i++) { 365165762Sjeff printf("\t\trunq bits %d 0x%zx\n", 366165762Sjeff i, rq->rq_status.rqb_bits[i]); 367165762Sjeff for (j = 0; j < RQB_BPW; j++) 368165762Sjeff if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 369165762Sjeff pri = j + (i << RQB_L2BPW); 370165762Sjeff rqh = &rq->rq_queues[pri]; 371177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) { 372165762Sjeff printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 373177435Sjeff td, td->td_name, td->td_priority, 374177435Sjeff td->td_rqindex, pri); 375165762Sjeff } 376165762Sjeff } 377165762Sjeff } 378165762Sjeff} 379165762Sjeff 380171482Sjeff/* 381171482Sjeff * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 382171482Sjeff */ 383113357Sjeffvoid 384164936Sjuliantdq_print(int cpu) 385110267Sjeff{ 386164936Sjulian struct tdq *tdq; 387112994Sjeff 388164936Sjulian tdq = TDQ_CPU(cpu); 389112994Sjeff 390171713Sjeff printf("tdq %d:\n", TDQ_ID(tdq)); 391176735Sjeff printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 392176735Sjeff printf("\tLock name: %s\n", tdq->tdq_name); 393165620Sjeff printf("\tload: %d\n", tdq->tdq_load); 394178277Sjeff printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt); 395178277Sjeff printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt); 396171482Sjeff printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 397165766Sjeff printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 398178277Sjeff printf("\tload transferable: %d\n", tdq->tdq_transferable); 399178277Sjeff printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 400165762Sjeff printf("\trealtime runq:\n"); 401165762Sjeff runq_print(&tdq->tdq_realtime); 402165762Sjeff printf("\ttimeshare runq:\n"); 403165762Sjeff runq_print(&tdq->tdq_timeshare); 404165762Sjeff printf("\tidle runq:\n"); 405165762Sjeff runq_print(&tdq->tdq_idle); 406113357Sjeff} 407112994Sjeff 408177005Sjeffstatic inline int 409177005Sjeffsched_shouldpreempt(int pri, int cpri, int remote) 410177005Sjeff{ 411177005Sjeff /* 412177005Sjeff * If the new priority is not better than the current priority there is 413177005Sjeff * nothing to do. 414177005Sjeff */ 415177005Sjeff if (pri >= cpri) 416177005Sjeff return (0); 417177005Sjeff /* 418177005Sjeff * Always preempt idle. 419177005Sjeff */ 420177005Sjeff if (cpri >= PRI_MIN_IDLE) 421177005Sjeff return (1); 422177005Sjeff /* 423177005Sjeff * If preemption is disabled don't preempt others. 424177005Sjeff */ 425177005Sjeff if (preempt_thresh == 0) 426177005Sjeff return (0); 427177005Sjeff /* 428177005Sjeff * Preempt if we exceed the threshold. 429177005Sjeff */ 430177005Sjeff if (pri <= preempt_thresh) 431177005Sjeff return (1); 432177005Sjeff /* 433217351Sjhb * If we're interactive or better and there is non-interactive 434217351Sjhb * or worse running preempt only remote processors. 435177005Sjeff */ 436217351Sjhb if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT) 437177005Sjeff return (1); 438177005Sjeff return (0); 439177005Sjeff} 440177005Sjeff 441171482Sjeff/* 442171482Sjeff * Add a thread to the actual run-queue. Keeps transferable counts up to 443171482Sjeff * date with what is actually on the run-queue. Selects the correct 444171482Sjeff * queue position for timeshare threads. 445171482Sjeff */ 446122744Sjeffstatic __inline void 447177435Sjefftdq_runq_add(struct tdq *tdq, struct thread *td, int flags) 448122744Sjeff{ 449177435Sjeff struct td_sched *ts; 450177042Sjeff u_char pri; 451177042Sjeff 452171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 453177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 454177009Sjeff 455177435Sjeff pri = td->td_priority; 456177435Sjeff ts = td->td_sched; 457177435Sjeff TD_SET_RUNQ(td); 458177435Sjeff if (THREAD_CAN_MIGRATE(td)) { 459165620Sjeff tdq->tdq_transferable++; 460164936Sjulian ts->ts_flags |= TSF_XFERABLE; 461123433Sjeff } 462217351Sjhb if (pri < PRI_MIN_BATCH) { 463177042Sjeff ts->ts_runq = &tdq->tdq_realtime; 464217351Sjhb } else if (pri <= PRI_MAX_BATCH) { 465177042Sjeff ts->ts_runq = &tdq->tdq_timeshare; 466217351Sjhb KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH, 467165762Sjeff ("Invalid priority %d on timeshare runq", pri)); 468165762Sjeff /* 469165762Sjeff * This queue contains only priorities between MIN and MAX 470165762Sjeff * realtime. Use the whole queue to represent these values. 471165762Sjeff */ 472171713Sjeff if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 473230173Savg pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE; 474165762Sjeff pri = (pri + tdq->tdq_idx) % RQ_NQS; 475165766Sjeff /* 476165766Sjeff * This effectively shortens the queue by one so we 477165766Sjeff * can have a one slot difference between idx and 478165766Sjeff * ridx while we wait for threads to drain. 479165766Sjeff */ 480165766Sjeff if (tdq->tdq_ridx != tdq->tdq_idx && 481165766Sjeff pri == tdq->tdq_ridx) 482167664Sjeff pri = (unsigned char)(pri - 1) % RQ_NQS; 483165762Sjeff } else 484165766Sjeff pri = tdq->tdq_ridx; 485177435Sjeff runq_add_pri(ts->ts_runq, td, pri, flags); 486177042Sjeff return; 487165762Sjeff } else 488177009Sjeff ts->ts_runq = &tdq->tdq_idle; 489177435Sjeff runq_add(ts->ts_runq, td, flags); 490177009Sjeff} 491177009Sjeff 492171482Sjeff/* 493171482Sjeff * Remove a thread from a run-queue. This typically happens when a thread 494171482Sjeff * is selected to run. Running threads are not on the queue and the 495171482Sjeff * transferable count does not reflect them. 496171482Sjeff */ 497122744Sjeffstatic __inline void 498177435Sjefftdq_runq_rem(struct tdq *tdq, struct thread *td) 499122744Sjeff{ 500177435Sjeff struct td_sched *ts; 501177435Sjeff 502177435Sjeff ts = td->td_sched; 503171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 504171482Sjeff KASSERT(ts->ts_runq != NULL, 505177435Sjeff ("tdq_runq_remove: thread %p null ts_runq", td)); 506164936Sjulian if (ts->ts_flags & TSF_XFERABLE) { 507165620Sjeff tdq->tdq_transferable--; 508164936Sjulian ts->ts_flags &= ~TSF_XFERABLE; 509123433Sjeff } 510165766Sjeff if (ts->ts_runq == &tdq->tdq_timeshare) { 511165766Sjeff if (tdq->tdq_idx != tdq->tdq_ridx) 512177435Sjeff runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx); 513165766Sjeff else 514177435Sjeff runq_remove_idx(ts->ts_runq, td, NULL); 515165766Sjeff } else 516177435Sjeff runq_remove(ts->ts_runq, td); 517122744Sjeff} 518122744Sjeff 519171482Sjeff/* 520171482Sjeff * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 521171482Sjeff * for this thread to the referenced thread queue. 522171482Sjeff */ 523113357Sjeffstatic void 524177435Sjefftdq_load_add(struct tdq *tdq, struct thread *td) 525113357Sjeff{ 526171482Sjeff 527171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 528177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 529177902Sjeff 530165620Sjeff tdq->tdq_load++; 531198854Sattilio if ((td->td_flags & TDF_NOLOAD) == 0) 532177902Sjeff tdq->tdq_sysload++; 533187357Sjeff KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 534262057Savg SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load); 535110267Sjeff} 536113357Sjeff 537171482Sjeff/* 538171482Sjeff * Remove the load from a thread that is transitioning to a sleep state or 539171482Sjeff * exiting. 540171482Sjeff */ 541112994Sjeffstatic void 542177435Sjefftdq_load_rem(struct tdq *tdq, struct thread *td) 543110267Sjeff{ 544171482Sjeff 545177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 546171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 547171482Sjeff KASSERT(tdq->tdq_load != 0, 548171713Sjeff ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 549177902Sjeff 550165620Sjeff tdq->tdq_load--; 551198854Sattilio if ((td->td_flags & TDF_NOLOAD) == 0) 552177902Sjeff tdq->tdq_sysload--; 553187357Sjeff KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 554262057Savg SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load); 555110267Sjeff} 556110267Sjeff 557176735Sjeff/* 558176735Sjeff * Set lowpri to its exact value by searching the run-queue and 559176735Sjeff * evaluating curthread. curthread may be passed as an optimization. 560176735Sjeff */ 561176735Sjeffstatic void 562176735Sjefftdq_setlowpri(struct tdq *tdq, struct thread *ctd) 563176735Sjeff{ 564176735Sjeff struct thread *td; 565176735Sjeff 566176735Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 567176735Sjeff if (ctd == NULL) 568176735Sjeff ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 569177435Sjeff td = tdq_choose(tdq); 570177435Sjeff if (td == NULL || td->td_priority > ctd->td_priority) 571176735Sjeff tdq->tdq_lowpri = ctd->td_priority; 572176735Sjeff else 573176735Sjeff tdq->tdq_lowpri = td->td_priority; 574176735Sjeff} 575176735Sjeff 576113357Sjeff#ifdef SMP 577176735Sjeffstruct cpu_search { 578194779Sjeff cpuset_t cs_mask; 579233599Smav u_int cs_prefer; 580233599Smav int cs_pri; /* Min priority for low. */ 581233599Smav int cs_limit; /* Max load for low, min load for high. */ 582233599Smav int cs_cpu; 583233599Smav int cs_load; 584176735Sjeff}; 585176735Sjeff 586176735Sjeff#define CPU_SEARCH_LOWEST 0x1 587176735Sjeff#define CPU_SEARCH_HIGHEST 0x2 588176735Sjeff#define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 589176735Sjeff 590194779Sjeff#define CPUSET_FOREACH(cpu, mask) \ 591194779Sjeff for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \ 592222813Sattilio if (CPU_ISSET(cpu, &mask)) 593176735Sjeff 594233599Smavstatic __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low, 595176735Sjeff struct cpu_search *high, const int match); 596233599Smavint cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low); 597233599Smavint cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high); 598233599Smavint cpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 599176735Sjeff struct cpu_search *high); 600176735Sjeff 601116069Sjeff/* 602176735Sjeff * Search the tree of cpu_groups for the lowest or highest loaded cpu 603176735Sjeff * according to the match argument. This routine actually compares the 604176735Sjeff * load on all paths through the tree and finds the least loaded cpu on 605176735Sjeff * the least loaded path, which may differ from the least loaded cpu in 606176735Sjeff * the system. This balances work among caches and busses. 607116069Sjeff * 608176735Sjeff * This inline is instantiated in three forms below using constants for the 609176735Sjeff * match argument. It is reduced to the minimum set for each case. It is 610176735Sjeff * also recursive to the depth of the tree. 611116069Sjeff */ 612177169Sjhbstatic __inline int 613233599Smavcpu_search(const struct cpu_group *cg, struct cpu_search *low, 614176735Sjeff struct cpu_search *high, const int match) 615176735Sjeff{ 616233599Smav struct cpu_search lgroup; 617233599Smav struct cpu_search hgroup; 618233599Smav cpuset_t cpumask; 619233599Smav struct cpu_group *child; 620233599Smav struct tdq *tdq; 621236547Smav int cpu, i, hload, lload, load, total, rnd, *rndptr; 622176735Sjeff 623176735Sjeff total = 0; 624233599Smav cpumask = cg->cg_mask; 625233599Smav if (match & CPU_SEARCH_LOWEST) { 626233599Smav lload = INT_MAX; 627233599Smav lgroup = *low; 628233599Smav } 629233599Smav if (match & CPU_SEARCH_HIGHEST) { 630236547Smav hload = INT_MIN; 631233599Smav hgroup = *high; 632233599Smav } 633176735Sjeff 634233599Smav /* Iterate through the child CPU groups and then remaining CPUs. */ 635256208Smav for (i = cg->cg_children, cpu = mp_maxid; ; ) { 636236547Smav if (i == 0) { 637256208Smav#ifdef HAVE_INLINE_FFSL 638256208Smav cpu = CPU_FFS(&cpumask) - 1; 639256208Smav#else 640236547Smav while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask)) 641236547Smav cpu--; 642256208Smav#endif 643236547Smav if (cpu < 0) 644233599Smav break; 645233599Smav child = NULL; 646233599Smav } else 647236547Smav child = &cg->cg_child[i - 1]; 648233599Smav 649236547Smav if (match & CPU_SEARCH_LOWEST) 650236547Smav lgroup.cs_cpu = -1; 651236547Smav if (match & CPU_SEARCH_HIGHEST) 652236547Smav hgroup.cs_cpu = -1; 653233599Smav if (child) { /* Handle child CPU group. */ 654233599Smav CPU_NAND(&cpumask, &child->cg_mask); 655176735Sjeff switch (match) { 656176735Sjeff case CPU_SEARCH_LOWEST: 657176735Sjeff load = cpu_search_lowest(child, &lgroup); 658176735Sjeff break; 659176735Sjeff case CPU_SEARCH_HIGHEST: 660176735Sjeff load = cpu_search_highest(child, &hgroup); 661176735Sjeff break; 662176735Sjeff case CPU_SEARCH_BOTH: 663176735Sjeff load = cpu_search_both(child, &lgroup, &hgroup); 664176735Sjeff break; 665176735Sjeff } 666233599Smav } else { /* Handle child CPU. */ 667256208Smav CPU_CLR(cpu, &cpumask); 668233599Smav tdq = TDQ_CPU(cpu); 669233599Smav load = tdq->tdq_load * 256; 670236547Smav rndptr = DPCPU_PTR(randomval); 671236547Smav rnd = (*rndptr = *rndptr * 69069 + 5) >> 26; 672233599Smav if (match & CPU_SEARCH_LOWEST) { 673233599Smav if (cpu == low->cs_prefer) 674233599Smav load -= 64; 675233599Smav /* If that CPU is allowed and get data. */ 676236547Smav if (tdq->tdq_lowpri > lgroup.cs_pri && 677236547Smav tdq->tdq_load <= lgroup.cs_limit && 678236547Smav CPU_ISSET(cpu, &lgroup.cs_mask)) { 679233599Smav lgroup.cs_cpu = cpu; 680233599Smav lgroup.cs_load = load - rnd; 681176735Sjeff } 682233599Smav } 683233599Smav if (match & CPU_SEARCH_HIGHEST) 684236547Smav if (tdq->tdq_load >= hgroup.cs_limit && 685236547Smav tdq->tdq_transferable && 686236547Smav CPU_ISSET(cpu, &hgroup.cs_mask)) { 687233599Smav hgroup.cs_cpu = cpu; 688233599Smav hgroup.cs_load = load - rnd; 689176735Sjeff } 690176735Sjeff } 691233599Smav total += load; 692176735Sjeff 693233599Smav /* We have info about child item. Compare it. */ 694233599Smav if (match & CPU_SEARCH_LOWEST) { 695236547Smav if (lgroup.cs_cpu >= 0 && 696233599Smav (load < lload || 697233599Smav (load == lload && lgroup.cs_load < low->cs_load))) { 698233599Smav lload = load; 699233599Smav low->cs_cpu = lgroup.cs_cpu; 700233599Smav low->cs_load = lgroup.cs_load; 701233599Smav } 702233599Smav } 703233599Smav if (match & CPU_SEARCH_HIGHEST) 704236547Smav if (hgroup.cs_cpu >= 0 && 705233599Smav (load > hload || 706233599Smav (load == hload && hgroup.cs_load > high->cs_load))) { 707233599Smav hload = load; 708233599Smav high->cs_cpu = hgroup.cs_cpu; 709233599Smav high->cs_load = hgroup.cs_load; 710233599Smav } 711236547Smav if (child) { 712236547Smav i--; 713236547Smav if (i == 0 && CPU_EMPTY(&cpumask)) 714236547Smav break; 715256208Smav } 716256208Smav#ifndef HAVE_INLINE_FFSL 717256208Smav else 718236547Smav cpu--; 719256208Smav#endif 720176735Sjeff } 721176735Sjeff return (total); 722176735Sjeff} 723176735Sjeff 724176735Sjeff/* 725176735Sjeff * cpu_search instantiations must pass constants to maintain the inline 726176735Sjeff * optimization. 727176735Sjeff */ 728176735Sjeffint 729233599Smavcpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low) 730176735Sjeff{ 731176735Sjeff return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 732176735Sjeff} 733176735Sjeff 734176735Sjeffint 735233599Smavcpu_search_highest(const struct cpu_group *cg, struct cpu_search *high) 736176735Sjeff{ 737176735Sjeff return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 738176735Sjeff} 739176735Sjeff 740176735Sjeffint 741233599Smavcpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 742176735Sjeff struct cpu_search *high) 743176735Sjeff{ 744176735Sjeff return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 745176735Sjeff} 746176735Sjeff 747176735Sjeff/* 748176735Sjeff * Find the cpu with the least load via the least loaded path that has a 749176735Sjeff * lowpri greater than pri pri. A pri of -1 indicates any priority is 750176735Sjeff * acceptable. 751176735Sjeff */ 752176735Sjeffstatic inline int 753233599Smavsched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload, 754233599Smav int prefer) 755176735Sjeff{ 756176735Sjeff struct cpu_search low; 757176735Sjeff 758176735Sjeff low.cs_cpu = -1; 759233599Smav low.cs_prefer = prefer; 760176735Sjeff low.cs_mask = mask; 761233599Smav low.cs_pri = pri; 762233599Smav low.cs_limit = maxload; 763176735Sjeff cpu_search_lowest(cg, &low); 764176735Sjeff return low.cs_cpu; 765176735Sjeff} 766176735Sjeff 767176735Sjeff/* 768176735Sjeff * Find the cpu with the highest load via the highest loaded path. 769176735Sjeff */ 770176735Sjeffstatic inline int 771233599Smavsched_highest(const struct cpu_group *cg, cpuset_t mask, int minload) 772176735Sjeff{ 773176735Sjeff struct cpu_search high; 774176735Sjeff 775176735Sjeff high.cs_cpu = -1; 776176735Sjeff high.cs_mask = mask; 777176735Sjeff high.cs_limit = minload; 778176735Sjeff cpu_search_highest(cg, &high); 779176735Sjeff return high.cs_cpu; 780176735Sjeff} 781176735Sjeff 782121790Sjeffstatic void 783176735Sjeffsched_balance_group(struct cpu_group *cg) 784116069Sjeff{ 785233599Smav cpuset_t hmask, lmask; 786233599Smav int high, low, anylow; 787123487Sjeff 788233599Smav CPU_FILL(&hmask); 789176735Sjeff for (;;) { 790233599Smav high = sched_highest(cg, hmask, 1); 791233599Smav /* Stop if there is no more CPU with transferrable threads. */ 792233599Smav if (high == -1) 793176735Sjeff break; 794233599Smav CPU_CLR(high, &hmask); 795233599Smav CPU_COPY(&hmask, &lmask); 796233599Smav /* Stop if there is no more CPU left for low. */ 797233599Smav if (CPU_EMPTY(&lmask)) 798176735Sjeff break; 799233599Smav anylow = 1; 800233599Smavnextlow: 801233599Smav low = sched_lowest(cg, lmask, -1, 802233599Smav TDQ_CPU(high)->tdq_load - 1, high); 803233599Smav /* Stop if we looked well and found no less loaded CPU. */ 804233599Smav if (anylow && low == -1) 805233599Smav break; 806233599Smav /* Go to next high if we found no less loaded CPU. */ 807233599Smav if (low == -1) 808233599Smav continue; 809233599Smav /* Transfer thread from high to low. */ 810233599Smav if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) { 811233599Smav /* CPU that got thread can no longer be a donor. */ 812233599Smav CPU_CLR(low, &hmask); 813233599Smav } else { 814233599Smav /* 815233599Smav * If failed, then there is no threads on high 816233599Smav * that can run on this low. Drop low from low 817233599Smav * mask and look for different one. 818233599Smav */ 819233599Smav CPU_CLR(low, &lmask); 820233599Smav anylow = 0; 821233599Smav goto nextlow; 822233599Smav } 823123487Sjeff } 824123487Sjeff} 825123487Sjeff 826123487Sjeffstatic void 827201148Sedsched_balance(void) 828123487Sjeff{ 829172409Sjeff struct tdq *tdq; 830123487Sjeff 831172409Sjeff /* 832172409Sjeff * Select a random time between .5 * balance_interval and 833172409Sjeff * 1.5 * balance_interval. 834172409Sjeff */ 835176735Sjeff balance_ticks = max(balance_interval / 2, 1); 836176735Sjeff balance_ticks += random() % balance_interval; 837171482Sjeff if (smp_started == 0 || rebalance == 0) 838171482Sjeff return; 839172409Sjeff tdq = TDQ_SELF(); 840172409Sjeff TDQ_UNLOCK(tdq); 841176735Sjeff sched_balance_group(cpu_top); 842172409Sjeff TDQ_LOCK(tdq); 843123487Sjeff} 844123487Sjeff 845171482Sjeff/* 846171482Sjeff * Lock two thread queues using their address to maintain lock order. 847171482Sjeff */ 848123487Sjeffstatic void 849171482Sjefftdq_lock_pair(struct tdq *one, struct tdq *two) 850171482Sjeff{ 851171482Sjeff if (one < two) { 852171482Sjeff TDQ_LOCK(one); 853171482Sjeff TDQ_LOCK_FLAGS(two, MTX_DUPOK); 854171482Sjeff } else { 855171482Sjeff TDQ_LOCK(two); 856171482Sjeff TDQ_LOCK_FLAGS(one, MTX_DUPOK); 857171482Sjeff } 858171482Sjeff} 859171482Sjeff 860171482Sjeff/* 861172409Sjeff * Unlock two thread queues. Order is not important here. 862172409Sjeff */ 863172409Sjeffstatic void 864172409Sjefftdq_unlock_pair(struct tdq *one, struct tdq *two) 865172409Sjeff{ 866172409Sjeff TDQ_UNLOCK(one); 867172409Sjeff TDQ_UNLOCK(two); 868172409Sjeff} 869172409Sjeff 870172409Sjeff/* 871171482Sjeff * Transfer load between two imbalanced thread queues. 872171482Sjeff */ 873176735Sjeffstatic int 874164936Sjuliansched_balance_pair(struct tdq *high, struct tdq *low) 875123487Sjeff{ 876176735Sjeff int moved; 877230691Smarius int cpu; 878116069Sjeff 879171482Sjeff tdq_lock_pair(high, low); 880176735Sjeff moved = 0; 881116069Sjeff /* 882122744Sjeff * Determine what the imbalance is and then adjust that to how many 883165620Sjeff * threads we actually have to give up (transferable). 884122744Sjeff */ 885233599Smav if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load && 886233599Smav (moved = tdq_move(high, low)) > 0) { 887172293Sjeff /* 888230691Smarius * In case the target isn't the current cpu IPI it to force a 889230691Smarius * reschedule with the new workload. 890172293Sjeff */ 891230691Smarius cpu = TDQ_ID(low); 892230691Smarius sched_pin(); 893230691Smarius if (cpu != PCPU_GET(cpuid)) 894230691Smarius ipi_cpu(cpu, IPI_PREEMPT); 895230691Smarius sched_unpin(); 896171482Sjeff } 897172409Sjeff tdq_unlock_pair(high, low); 898176735Sjeff return (moved); 899116069Sjeff} 900116069Sjeff 901171482Sjeff/* 902171482Sjeff * Move a thread from one thread queue to another. 903171482Sjeff */ 904176735Sjeffstatic int 905171482Sjefftdq_move(struct tdq *from, struct tdq *to) 906116069Sjeff{ 907171482Sjeff struct td_sched *ts; 908171482Sjeff struct thread *td; 909164936Sjulian struct tdq *tdq; 910171482Sjeff int cpu; 911116069Sjeff 912172409Sjeff TDQ_LOCK_ASSERT(from, MA_OWNED); 913172409Sjeff TDQ_LOCK_ASSERT(to, MA_OWNED); 914172409Sjeff 915164936Sjulian tdq = from; 916171482Sjeff cpu = TDQ_ID(to); 917177435Sjeff td = tdq_steal(tdq, cpu); 918177435Sjeff if (td == NULL) 919176735Sjeff return (0); 920177435Sjeff ts = td->td_sched; 921171482Sjeff /* 922171482Sjeff * Although the run queue is locked the thread may be blocked. Lock 923172409Sjeff * it to clear this and acquire the run-queue lock. 924171482Sjeff */ 925171482Sjeff thread_lock(td); 926172409Sjeff /* Drop recursive lock on from acquired via thread_lock(). */ 927171482Sjeff TDQ_UNLOCK(from); 928171482Sjeff sched_rem(td); 929166108Sjeff ts->ts_cpu = cpu; 930171482Sjeff td->td_lock = TDQ_LOCKPTR(to); 931171482Sjeff tdq_add(to, td, SRQ_YIELDING); 932176735Sjeff return (1); 933116069Sjeff} 934110267Sjeff 935171482Sjeff/* 936171482Sjeff * This tdq has idled. Try to steal a thread from another cpu and switch 937171482Sjeff * to it. 938171482Sjeff */ 939123433Sjeffstatic int 940164936Sjuliantdq_idled(struct tdq *tdq) 941121790Sjeff{ 942176735Sjeff struct cpu_group *cg; 943164936Sjulian struct tdq *steal; 944194779Sjeff cpuset_t mask; 945176735Sjeff int thresh; 946171482Sjeff int cpu; 947123433Sjeff 948172484Sjeff if (smp_started == 0 || steal_idle == 0) 949172484Sjeff return (1); 950194779Sjeff CPU_FILL(&mask); 951194779Sjeff CPU_CLR(PCPU_GET(cpuid), &mask); 952176735Sjeff /* We don't want to be preempted while we're iterating. */ 953171482Sjeff spinlock_enter(); 954176735Sjeff for (cg = tdq->tdq_cg; cg != NULL; ) { 955191643Sjeff if ((cg->cg_flags & CG_FLAG_THREAD) == 0) 956176735Sjeff thresh = steal_thresh; 957176735Sjeff else 958176735Sjeff thresh = 1; 959176735Sjeff cpu = sched_highest(cg, mask, thresh); 960176735Sjeff if (cpu == -1) { 961176735Sjeff cg = cg->cg_parent; 962176735Sjeff continue; 963166108Sjeff } 964176735Sjeff steal = TDQ_CPU(cpu); 965194779Sjeff CPU_CLR(cpu, &mask); 966176735Sjeff tdq_lock_pair(tdq, steal); 967176735Sjeff if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 968176735Sjeff tdq_unlock_pair(tdq, steal); 969176735Sjeff continue; 970171482Sjeff } 971176735Sjeff /* 972176735Sjeff * If a thread was added while interrupts were disabled don't 973176735Sjeff * steal one here. If we fail to acquire one due to affinity 974176735Sjeff * restrictions loop again with this cpu removed from the 975176735Sjeff * set. 976176735Sjeff */ 977176735Sjeff if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 978176735Sjeff tdq_unlock_pair(tdq, steal); 979176735Sjeff continue; 980176735Sjeff } 981176735Sjeff spinlock_exit(); 982176735Sjeff TDQ_UNLOCK(steal); 983178272Sjeff mi_switch(SW_VOL | SWT_IDLE, NULL); 984176735Sjeff thread_unlock(curthread); 985176735Sjeff 986176735Sjeff return (0); 987123433Sjeff } 988171482Sjeff spinlock_exit(); 989123433Sjeff return (1); 990121790Sjeff} 991121790Sjeff 992171482Sjeff/* 993171482Sjeff * Notify a remote cpu of new work. Sends an IPI if criteria are met. 994171482Sjeff */ 995121790Sjeffstatic void 996177435Sjefftdq_notify(struct tdq *tdq, struct thread *td) 997121790Sjeff{ 998185047Sjhb struct thread *ctd; 999166247Sjeff int pri; 1000166108Sjeff int cpu; 1001121790Sjeff 1002177005Sjeff if (tdq->tdq_ipipending) 1003177005Sjeff return; 1004177435Sjeff cpu = td->td_sched->ts_cpu; 1005177435Sjeff pri = td->td_priority; 1006185047Sjhb ctd = pcpu_find(cpu)->pc_curthread; 1007185047Sjhb if (!sched_shouldpreempt(pri, ctd->td_priority, 1)) 1008166137Sjeff return; 1009185047Sjhb if (TD_IS_IDLETHREAD(ctd)) { 1010178277Sjeff /* 1011178471Sjeff * If the MD code has an idle wakeup routine try that before 1012178471Sjeff * falling back to IPI. 1013178471Sjeff */ 1014212416Smav if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu)) 1015178471Sjeff return; 1016178277Sjeff } 1017177005Sjeff tdq->tdq_ipipending = 1; 1018210939Sjhb ipi_cpu(cpu, IPI_PREEMPT); 1019121790Sjeff} 1020121790Sjeff 1021171482Sjeff/* 1022171482Sjeff * Steals load from a timeshare queue. Honors the rotating queue head 1023171482Sjeff * index. 1024171482Sjeff */ 1025177435Sjeffstatic struct thread * 1026176735Sjeffrunq_steal_from(struct runq *rq, int cpu, u_char start) 1027171482Sjeff{ 1028171482Sjeff struct rqbits *rqb; 1029171482Sjeff struct rqhead *rqh; 1030233599Smav struct thread *td, *first; 1031171482Sjeff int bit; 1032171482Sjeff int pri; 1033171482Sjeff int i; 1034171482Sjeff 1035171482Sjeff rqb = &rq->rq_status; 1036171482Sjeff bit = start & (RQB_BPW -1); 1037171482Sjeff pri = 0; 1038233599Smav first = NULL; 1039171482Sjeffagain: 1040171482Sjeff for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 1041171482Sjeff if (rqb->rqb_bits[i] == 0) 1042171482Sjeff continue; 1043171482Sjeff if (bit != 0) { 1044171482Sjeff for (pri = bit; pri < RQB_BPW; pri++) 1045171482Sjeff if (rqb->rqb_bits[i] & (1ul << pri)) 1046171482Sjeff break; 1047171482Sjeff if (pri >= RQB_BPW) 1048171482Sjeff continue; 1049171482Sjeff } else 1050171482Sjeff pri = RQB_FFS(rqb->rqb_bits[i]); 1051171482Sjeff pri += (i << RQB_L2BPW); 1052171482Sjeff rqh = &rq->rq_queues[pri]; 1053177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) { 1054177435Sjeff if (first && THREAD_CAN_MIGRATE(td) && 1055177435Sjeff THREAD_CAN_SCHED(td, cpu)) 1056177435Sjeff return (td); 1057233599Smav first = td; 1058171482Sjeff } 1059171482Sjeff } 1060171482Sjeff if (start != 0) { 1061171482Sjeff start = 0; 1062171482Sjeff goto again; 1063171482Sjeff } 1064171482Sjeff 1065233599Smav if (first && THREAD_CAN_MIGRATE(first) && 1066233599Smav THREAD_CAN_SCHED(first, cpu)) 1067233599Smav return (first); 1068171482Sjeff return (NULL); 1069171482Sjeff} 1070171482Sjeff 1071171482Sjeff/* 1072171482Sjeff * Steals load from a standard linear queue. 1073171482Sjeff */ 1074177435Sjeffstatic struct thread * 1075176735Sjeffrunq_steal(struct runq *rq, int cpu) 1076121790Sjeff{ 1077121790Sjeff struct rqhead *rqh; 1078121790Sjeff struct rqbits *rqb; 1079177435Sjeff struct thread *td; 1080121790Sjeff int word; 1081121790Sjeff int bit; 1082121790Sjeff 1083121790Sjeff rqb = &rq->rq_status; 1084121790Sjeff for (word = 0; word < RQB_LEN; word++) { 1085121790Sjeff if (rqb->rqb_bits[word] == 0) 1086121790Sjeff continue; 1087121790Sjeff for (bit = 0; bit < RQB_BPW; bit++) { 1088123231Speter if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1089121790Sjeff continue; 1090121790Sjeff rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1091177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) 1092177435Sjeff if (THREAD_CAN_MIGRATE(td) && 1093177435Sjeff THREAD_CAN_SCHED(td, cpu)) 1094177435Sjeff return (td); 1095121790Sjeff } 1096121790Sjeff } 1097121790Sjeff return (NULL); 1098121790Sjeff} 1099121790Sjeff 1100171482Sjeff/* 1101171482Sjeff * Attempt to steal a thread in priority order from a thread queue. 1102171482Sjeff */ 1103177435Sjeffstatic struct thread * 1104176735Sjefftdq_steal(struct tdq *tdq, int cpu) 1105121790Sjeff{ 1106177435Sjeff struct thread *td; 1107121790Sjeff 1108171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1109177435Sjeff if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1110177435Sjeff return (td); 1111177435Sjeff if ((td = runq_steal_from(&tdq->tdq_timeshare, 1112177435Sjeff cpu, tdq->tdq_ridx)) != NULL) 1113177435Sjeff return (td); 1114176735Sjeff return (runq_steal(&tdq->tdq_idle, cpu)); 1115121790Sjeff} 1116123433Sjeff 1117171482Sjeff/* 1118171482Sjeff * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1119172409Sjeff * current lock and returns with the assigned queue locked. 1120171482Sjeff */ 1121171482Sjeffstatic inline struct tdq * 1122177435Sjeffsched_setcpu(struct thread *td, int cpu, int flags) 1123123433Sjeff{ 1124177435Sjeff 1125171482Sjeff struct tdq *tdq; 1126123433Sjeff 1127177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1128171482Sjeff tdq = TDQ_CPU(cpu); 1129177435Sjeff td->td_sched->ts_cpu = cpu; 1130177435Sjeff /* 1131177435Sjeff * If the lock matches just return the queue. 1132177435Sjeff */ 1133171482Sjeff if (td->td_lock == TDQ_LOCKPTR(tdq)) 1134171482Sjeff return (tdq); 1135171482Sjeff#ifdef notyet 1136123433Sjeff /* 1137172293Sjeff * If the thread isn't running its lockptr is a 1138171482Sjeff * turnstile or a sleepqueue. We can just lock_set without 1139171482Sjeff * blocking. 1140123685Sjeff */ 1141171482Sjeff if (TD_CAN_RUN(td)) { 1142171482Sjeff TDQ_LOCK(tdq); 1143171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1144171482Sjeff return (tdq); 1145171482Sjeff } 1146171482Sjeff#endif 1147166108Sjeff /* 1148171482Sjeff * The hard case, migration, we need to block the thread first to 1149171482Sjeff * prevent order reversals with other cpus locks. 1150166108Sjeff */ 1151202889Sattilio spinlock_enter(); 1152171482Sjeff thread_lock_block(td); 1153171482Sjeff TDQ_LOCK(tdq); 1154171713Sjeff thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1155202889Sattilio spinlock_exit(); 1156171482Sjeff return (tdq); 1157166108Sjeff} 1158166108Sjeff 1159178272SjeffSCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1160178272SjeffSCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1161178272SjeffSCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1162178272SjeffSCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1163178272SjeffSCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1164178272SjeffSCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1165178272Sjeff 1166166108Sjeffstatic int 1167177435Sjeffsched_pickcpu(struct thread *td, int flags) 1168171482Sjeff{ 1169233599Smav struct cpu_group *cg, *ccg; 1170177435Sjeff struct td_sched *ts; 1171171482Sjeff struct tdq *tdq; 1172194779Sjeff cpuset_t mask; 1173233599Smav int cpu, pri, self; 1174166108Sjeff 1175176735Sjeff self = PCPU_GET(cpuid); 1176177435Sjeff ts = td->td_sched; 1177166108Sjeff if (smp_started == 0) 1178166108Sjeff return (self); 1179171506Sjeff /* 1180171506Sjeff * Don't migrate a running thread from sched_switch(). 1181171506Sjeff */ 1182176735Sjeff if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1183176735Sjeff return (ts->ts_cpu); 1184166108Sjeff /* 1185176735Sjeff * Prefer to run interrupt threads on the processors that generate 1186176735Sjeff * the interrupt. 1187166108Sjeff */ 1188233599Smav pri = td->td_priority; 1189176735Sjeff if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1190178272Sjeff curthread->td_intr_nesting_level && ts->ts_cpu != self) { 1191178272Sjeff SCHED_STAT_INC(pickcpu_intrbind); 1192176735Sjeff ts->ts_cpu = self; 1193233599Smav if (TDQ_CPU(self)->tdq_lowpri > pri) { 1194233599Smav SCHED_STAT_INC(pickcpu_affinity); 1195233599Smav return (ts->ts_cpu); 1196233599Smav } 1197178272Sjeff } 1198166108Sjeff /* 1199176735Sjeff * If the thread can run on the last cpu and the affinity has not 1200176735Sjeff * expired or it is idle run it there. 1201166108Sjeff */ 1202176735Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1203233599Smav cg = tdq->tdq_cg; 1204233599Smav if (THREAD_CAN_SCHED(td, ts->ts_cpu) && 1205233599Smav tdq->tdq_lowpri >= PRI_MIN_IDLE && 1206233599Smav SCHED_AFFINITY(ts, CG_SHARE_L2)) { 1207233599Smav if (cg->cg_flags & CG_FLAG_THREAD) { 1208233599Smav CPUSET_FOREACH(cpu, cg->cg_mask) { 1209233599Smav if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) 1210233599Smav break; 1211233599Smav } 1212233599Smav } else 1213233599Smav cpu = INT_MAX; 1214233599Smav if (cpu > mp_maxid) { 1215178272Sjeff SCHED_STAT_INC(pickcpu_idle_affinity); 1216176735Sjeff return (ts->ts_cpu); 1217178272Sjeff } 1218139334Sjeff } 1219123433Sjeff /* 1220233599Smav * Search for the last level cache CPU group in the tree. 1221233599Smav * Skip caches with expired affinity time and SMT groups. 1222233599Smav * Affinity to higher level caches will be handled less aggressively. 1223123433Sjeff */ 1224233599Smav for (ccg = NULL; cg != NULL; cg = cg->cg_parent) { 1225233599Smav if (cg->cg_flags & CG_FLAG_THREAD) 1226233599Smav continue; 1227233599Smav if (!SCHED_AFFINITY(ts, cg->cg_level)) 1228233599Smav continue; 1229233599Smav ccg = cg; 1230233599Smav } 1231233599Smav if (ccg != NULL) 1232233599Smav cg = ccg; 1233176735Sjeff cpu = -1; 1234233599Smav /* Search the group for the less loaded idle CPU we can run now. */ 1235194779Sjeff mask = td->td_cpuset->cs_mask; 1236233599Smav if (cg != NULL && cg != cpu_top && 1237233599Smav CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0) 1238233599Smav cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE), 1239233599Smav INT_MAX, ts->ts_cpu); 1240233599Smav /* Search globally for the less loaded CPU we can run now. */ 1241176735Sjeff if (cpu == -1) 1242233599Smav cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu); 1243233599Smav /* Search globally for the less loaded CPU. */ 1244233599Smav if (cpu == -1) 1245233599Smav cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu); 1246233599Smav KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1247171506Sjeff /* 1248176735Sjeff * Compare the lowest loaded cpu to current cpu. 1249171506Sjeff */ 1250177005Sjeff if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1251233599Smav TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE && 1252233599Smav TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) { 1253178272Sjeff SCHED_STAT_INC(pickcpu_local); 1254177005Sjeff cpu = self; 1255178272Sjeff } else 1256178272Sjeff SCHED_STAT_INC(pickcpu_lowest); 1257178272Sjeff if (cpu != ts->ts_cpu) 1258178272Sjeff SCHED_STAT_INC(pickcpu_migration); 1259171482Sjeff return (cpu); 1260123433Sjeff} 1261176735Sjeff#endif 1262123433Sjeff 1263117326Sjeff/* 1264121790Sjeff * Pick the highest priority task we have and return it. 1265117326Sjeff */ 1266177435Sjeffstatic struct thread * 1267164936Sjuliantdq_choose(struct tdq *tdq) 1268110267Sjeff{ 1269177435Sjeff struct thread *td; 1270110267Sjeff 1271171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1272177435Sjeff td = runq_choose(&tdq->tdq_realtime); 1273177435Sjeff if (td != NULL) 1274177435Sjeff return (td); 1275177435Sjeff td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1276177435Sjeff if (td != NULL) { 1277217351Sjhb KASSERT(td->td_priority >= PRI_MIN_BATCH, 1278165762Sjeff ("tdq_choose: Invalid priority on timeshare queue %d", 1279177435Sjeff td->td_priority)); 1280177435Sjeff return (td); 1281165762Sjeff } 1282177435Sjeff td = runq_choose(&tdq->tdq_idle); 1283177435Sjeff if (td != NULL) { 1284177435Sjeff KASSERT(td->td_priority >= PRI_MIN_IDLE, 1285165762Sjeff ("tdq_choose: Invalid priority on idle queue %d", 1286177435Sjeff td->td_priority)); 1287177435Sjeff return (td); 1288165762Sjeff } 1289165762Sjeff 1290165762Sjeff return (NULL); 1291110267Sjeff} 1292110267Sjeff 1293171482Sjeff/* 1294171482Sjeff * Initialize a thread queue. 1295171482Sjeff */ 1296109864Sjeffstatic void 1297164936Sjuliantdq_setup(struct tdq *tdq) 1298110028Sjeff{ 1299171482Sjeff 1300171713Sjeff if (bootverbose) 1301171713Sjeff printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1302165762Sjeff runq_init(&tdq->tdq_realtime); 1303165762Sjeff runq_init(&tdq->tdq_timeshare); 1304165620Sjeff runq_init(&tdq->tdq_idle); 1305176735Sjeff snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1306176735Sjeff "sched lock %d", (int)TDQ_ID(tdq)); 1307176735Sjeff mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1308176735Sjeff MTX_SPIN | MTX_RECURSE); 1309187357Sjeff#ifdef KTR 1310187357Sjeff snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname), 1311187357Sjeff "CPU %d load", (int)TDQ_ID(tdq)); 1312187357Sjeff#endif 1313110028Sjeff} 1314110028Sjeff 1315171713Sjeff#ifdef SMP 1316110028Sjeffstatic void 1317171713Sjeffsched_setup_smp(void) 1318171713Sjeff{ 1319171713Sjeff struct tdq *tdq; 1320171713Sjeff int i; 1321171713Sjeff 1322176735Sjeff cpu_top = smp_topo(); 1323209059Sjhb CPU_FOREACH(i) { 1324176735Sjeff tdq = TDQ_CPU(i); 1325171713Sjeff tdq_setup(tdq); 1326176735Sjeff tdq->tdq_cg = smp_topo_find(cpu_top, i); 1327176735Sjeff if (tdq->tdq_cg == NULL) 1328176735Sjeff panic("Can't find cpu group for %d\n", i); 1329123433Sjeff } 1330176735Sjeff balance_tdq = TDQ_SELF(); 1331176735Sjeff sched_balance(); 1332171713Sjeff} 1333171713Sjeff#endif 1334171713Sjeff 1335171713Sjeff/* 1336171713Sjeff * Setup the thread queues and initialize the topology based on MD 1337171713Sjeff * information. 1338171713Sjeff */ 1339171713Sjeffstatic void 1340171713Sjeffsched_setup(void *dummy) 1341171713Sjeff{ 1342171713Sjeff struct tdq *tdq; 1343171713Sjeff 1344171713Sjeff tdq = TDQ_SELF(); 1345171713Sjeff#ifdef SMP 1346176734Sjeff sched_setup_smp(); 1347117237Sjeff#else 1348171713Sjeff tdq_setup(tdq); 1349116069Sjeff#endif 1350171482Sjeff 1351171482Sjeff /* Add thread0's load since it's running. */ 1352171482Sjeff TDQ_LOCK(tdq); 1353171713Sjeff thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1354177435Sjeff tdq_load_add(tdq, &thread0); 1355176735Sjeff tdq->tdq_lowpri = thread0.td_priority; 1356171482Sjeff TDQ_UNLOCK(tdq); 1357109864Sjeff} 1358109864Sjeff 1359171482Sjeff/* 1360241249Smav * This routine determines time constants after stathz and hz are setup. 1361171482Sjeff */ 1362153533Sdavidxu/* ARGSUSED */ 1363153533Sdavidxustatic void 1364153533Sdavidxusched_initticks(void *dummy) 1365153533Sdavidxu{ 1366171482Sjeff int incr; 1367171482Sjeff 1368153533Sdavidxu realstathz = stathz ? stathz : hz; 1369241249Smav sched_slice = realstathz / 10; /* ~100ms */ 1370241249Smav hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 1371241249Smav realstathz); 1372153533Sdavidxu 1373153533Sdavidxu /* 1374165762Sjeff * tickincr is shifted out by 10 to avoid rounding errors due to 1375165766Sjeff * hz not being evenly divisible by stathz on all platforms. 1376153533Sdavidxu */ 1377171482Sjeff incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1378165762Sjeff /* 1379165762Sjeff * This does not work for values of stathz that are more than 1380165762Sjeff * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1381165762Sjeff */ 1382171482Sjeff if (incr == 0) 1383171482Sjeff incr = 1; 1384171482Sjeff tickincr = incr; 1385166108Sjeff#ifdef SMP 1386171899Sjeff /* 1387172409Sjeff * Set the default balance interval now that we know 1388172409Sjeff * what realstathz is. 1389172409Sjeff */ 1390172409Sjeff balance_interval = realstathz; 1391166108Sjeff affinity = SCHED_AFFINITY_DEFAULT; 1392166108Sjeff#endif 1393236546Smav if (sched_idlespinthresh < 0) 1394247150Smav sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz; 1395153533Sdavidxu} 1396153533Sdavidxu 1397153533Sdavidxu 1398109864Sjeff/* 1399171482Sjeff * This is the core of the interactivity algorithm. Determines a score based 1400171482Sjeff * on past behavior. It is the ratio of sleep time to run time scaled to 1401171482Sjeff * a [0, 100] integer. This is the voluntary sleep time of a process, which 1402171482Sjeff * differs from the cpu usage because it does not account for time spent 1403171482Sjeff * waiting on a run-queue. Would be prettier if we had floating point. 1404171482Sjeff */ 1405171482Sjeffstatic int 1406171482Sjeffsched_interact_score(struct thread *td) 1407171482Sjeff{ 1408171482Sjeff struct td_sched *ts; 1409171482Sjeff int div; 1410171482Sjeff 1411171482Sjeff ts = td->td_sched; 1412171482Sjeff /* 1413171482Sjeff * The score is only needed if this is likely to be an interactive 1414171482Sjeff * task. Don't go through the expense of computing it if there's 1415171482Sjeff * no chance. 1416171482Sjeff */ 1417171482Sjeff if (sched_interact <= SCHED_INTERACT_HALF && 1418171482Sjeff ts->ts_runtime >= ts->ts_slptime) 1419171482Sjeff return (SCHED_INTERACT_HALF); 1420171482Sjeff 1421171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1422171482Sjeff div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1423171482Sjeff return (SCHED_INTERACT_HALF + 1424171482Sjeff (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1425171482Sjeff } 1426171482Sjeff if (ts->ts_slptime > ts->ts_runtime) { 1427171482Sjeff div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1428171482Sjeff return (ts->ts_runtime / div); 1429171482Sjeff } 1430171482Sjeff /* runtime == slptime */ 1431171482Sjeff if (ts->ts_runtime) 1432171482Sjeff return (SCHED_INTERACT_HALF); 1433171482Sjeff 1434171482Sjeff /* 1435171482Sjeff * This can happen if slptime and runtime are 0. 1436171482Sjeff */ 1437171482Sjeff return (0); 1438171482Sjeff 1439171482Sjeff} 1440171482Sjeff 1441171482Sjeff/* 1442109864Sjeff * Scale the scheduling priority according to the "interactivity" of this 1443109864Sjeff * process. 1444109864Sjeff */ 1445113357Sjeffstatic void 1446163709Sjbsched_priority(struct thread *td) 1447109864Sjeff{ 1448165762Sjeff int score; 1449109864Sjeff int pri; 1450109864Sjeff 1451217291Sjhb if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1452113357Sjeff return; 1453112966Sjeff /* 1454165762Sjeff * If the score is interactive we place the thread in the realtime 1455165762Sjeff * queue with a priority that is less than kernel and interrupt 1456165762Sjeff * priorities. These threads are not subject to nice restrictions. 1457112966Sjeff * 1458171482Sjeff * Scores greater than this are placed on the normal timeshare queue 1459165762Sjeff * where the priority is partially decided by the most recent cpu 1460165762Sjeff * utilization and the rest is decided by nice value. 1461172293Sjeff * 1462172293Sjeff * The nice value of the process has a linear effect on the calculated 1463172293Sjeff * score. Negative nice values make it easier for a thread to be 1464172293Sjeff * considered interactive. 1465112966Sjeff */ 1466198126Sjhb score = imax(0, sched_interact_score(td) + td->td_proc->p_nice); 1467165762Sjeff if (score < sched_interact) { 1468217351Sjhb pri = PRI_MIN_INTERACT; 1469217351Sjhb pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) / 1470217237Sjhb sched_interact) * score; 1471217351Sjhb KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT, 1472166208Sjeff ("sched_priority: invalid interactive priority %d score %d", 1473166208Sjeff pri, score)); 1474165762Sjeff } else { 1475165762Sjeff pri = SCHED_PRI_MIN; 1476165762Sjeff if (td->td_sched->ts_ticks) 1477230083Sjhb pri += min(SCHED_PRI_TICKS(td->td_sched), 1478259835Sjhb SCHED_PRI_RANGE - 1); 1479165762Sjeff pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1480217351Sjhb KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH, 1481171482Sjeff ("sched_priority: invalid priority %d: nice %d, " 1482171482Sjeff "ticks %d ftick %d ltick %d tick pri %d", 1483171482Sjeff pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1484171482Sjeff td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1485171482Sjeff SCHED_PRI_TICKS(td->td_sched))); 1486165762Sjeff } 1487165762Sjeff sched_user_prio(td, pri); 1488112966Sjeff 1489112966Sjeff return; 1490109864Sjeff} 1491109864Sjeff 1492121868Sjeff/* 1493121868Sjeff * This routine enforces a maximum limit on the amount of scheduling history 1494171482Sjeff * kept. It is called after either the slptime or runtime is adjusted. This 1495171482Sjeff * function is ugly due to integer math. 1496121868Sjeff */ 1497116463Sjeffstatic void 1498163709Sjbsched_interact_update(struct thread *td) 1499116463Sjeff{ 1500165819Sjeff struct td_sched *ts; 1501166208Sjeff u_int sum; 1502121605Sjeff 1503165819Sjeff ts = td->td_sched; 1504171482Sjeff sum = ts->ts_runtime + ts->ts_slptime; 1505121868Sjeff if (sum < SCHED_SLP_RUN_MAX) 1506121868Sjeff return; 1507121868Sjeff /* 1508165819Sjeff * This only happens from two places: 1509165819Sjeff * 1) We have added an unusual amount of run time from fork_exit. 1510165819Sjeff * 2) We have added an unusual amount of sleep time from sched_sleep(). 1511165819Sjeff */ 1512165819Sjeff if (sum > SCHED_SLP_RUN_MAX * 2) { 1513171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1514171482Sjeff ts->ts_runtime = SCHED_SLP_RUN_MAX; 1515171482Sjeff ts->ts_slptime = 1; 1516165819Sjeff } else { 1517171482Sjeff ts->ts_slptime = SCHED_SLP_RUN_MAX; 1518171482Sjeff ts->ts_runtime = 1; 1519165819Sjeff } 1520165819Sjeff return; 1521165819Sjeff } 1522165819Sjeff /* 1523121868Sjeff * If we have exceeded by more than 1/5th then the algorithm below 1524121868Sjeff * will not bring us back into range. Dividing by two here forces 1525133427Sjeff * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1526121868Sjeff */ 1527127850Sjeff if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1528171482Sjeff ts->ts_runtime /= 2; 1529171482Sjeff ts->ts_slptime /= 2; 1530121868Sjeff return; 1531116463Sjeff } 1532171482Sjeff ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1533171482Sjeff ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1534116463Sjeff} 1535116463Sjeff 1536171482Sjeff/* 1537171482Sjeff * Scale back the interactivity history when a child thread is created. The 1538171482Sjeff * history is inherited from the parent but the thread may behave totally 1539171482Sjeff * differently. For example, a shell spawning a compiler process. We want 1540171482Sjeff * to learn that the compiler is behaving badly very quickly. 1541171482Sjeff */ 1542121868Sjeffstatic void 1543163709Sjbsched_interact_fork(struct thread *td) 1544121868Sjeff{ 1545121868Sjeff int ratio; 1546121868Sjeff int sum; 1547121868Sjeff 1548171482Sjeff sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1549121868Sjeff if (sum > SCHED_SLP_RUN_FORK) { 1550121868Sjeff ratio = sum / SCHED_SLP_RUN_FORK; 1551171482Sjeff td->td_sched->ts_runtime /= ratio; 1552171482Sjeff td->td_sched->ts_slptime /= ratio; 1553121868Sjeff } 1554121868Sjeff} 1555121868Sjeff 1556113357Sjeff/* 1557171482Sjeff * Called from proc0_init() to setup the scheduler fields. 1558134791Sjulian */ 1559134791Sjulianvoid 1560134791Sjulianschedinit(void) 1561134791Sjulian{ 1562165762Sjeff 1563134791Sjulian /* 1564134791Sjulian * Set up the scheduler specific parts of proc0. 1565134791Sjulian */ 1566136167Sjulian proc0.p_sched = NULL; /* XXX */ 1567164936Sjulian thread0.td_sched = &td_sched0; 1568165762Sjeff td_sched0.ts_ltick = ticks; 1569165796Sjeff td_sched0.ts_ftick = ticks; 1570177009Sjeff td_sched0.ts_slice = sched_slice; 1571134791Sjulian} 1572134791Sjulian 1573134791Sjulian/* 1574113357Sjeff * This is only somewhat accurate since given many processes of the same 1575113357Sjeff * priority they will switch when their slices run out, which will be 1576165762Sjeff * at most sched_slice stathz ticks. 1577113357Sjeff */ 1578109864Sjeffint 1579109864Sjeffsched_rr_interval(void) 1580109864Sjeff{ 1581165762Sjeff 1582241249Smav /* Convert sched_slice from stathz to hz. */ 1583241249Smav return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz)); 1584109864Sjeff} 1585109864Sjeff 1586171482Sjeff/* 1587171482Sjeff * Update the percent cpu tracking information when it is requested or 1588171482Sjeff * the total history exceeds the maximum. We keep a sliding history of 1589171482Sjeff * tick counts that slowly decays. This is less precise than the 4BSD 1590171482Sjeff * mechanism since it happens with less regular and frequent events. 1591171482Sjeff */ 1592121790Sjeffstatic void 1593234166Smavsched_pctcpu_update(struct td_sched *ts, int run) 1594109864Sjeff{ 1595234166Smav int t = ticks; 1596165762Sjeff 1597234166Smav if (t - ts->ts_ltick >= SCHED_TICK_TARG) { 1598164936Sjulian ts->ts_ticks = 0; 1599234166Smav ts->ts_ftick = t - SCHED_TICK_TARG; 1600234166Smav } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) { 1601234166Smav ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) * 1602234166Smav (ts->ts_ltick - (t - SCHED_TICK_TARG)); 1603234166Smav ts->ts_ftick = t - SCHED_TICK_TARG; 1604234166Smav } 1605234166Smav if (run) 1606234166Smav ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT; 1607234166Smav ts->ts_ltick = t; 1608109864Sjeff} 1609109864Sjeff 1610171482Sjeff/* 1611171482Sjeff * Adjust the priority of a thread. Move it to the appropriate run-queue 1612171482Sjeff * if necessary. This is the back-end for several priority related 1613171482Sjeff * functions. 1614171482Sjeff */ 1615165762Sjeffstatic void 1616139453Sjhbsched_thread_priority(struct thread *td, u_char prio) 1617109864Sjeff{ 1618164936Sjulian struct td_sched *ts; 1619177009Sjeff struct tdq *tdq; 1620177009Sjeff int oldpri; 1621109864Sjeff 1622187357Sjeff KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio", 1623187357Sjeff "prio:%d", td->td_priority, "new prio:%d", prio, 1624187357Sjeff KTR_ATTR_LINKED, sched_tdname(curthread)); 1625262057Savg SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio); 1626240839Savg if (td != curthread && prio < td->td_priority) { 1627187357Sjeff KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 1628187357Sjeff "lend prio", "prio:%d", td->td_priority, "new prio:%d", 1629187357Sjeff prio, KTR_ATTR_LINKED, sched_tdname(td)); 1630262057Savg SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio, 1631236344Srstone curthread); 1632187357Sjeff } 1633164936Sjulian ts = td->td_sched; 1634170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1635139453Sjhb if (td->td_priority == prio) 1636139453Sjhb return; 1637177376Sjeff /* 1638177376Sjeff * If the priority has been elevated due to priority 1639177376Sjeff * propagation, we may have to move ourselves to a new 1640177376Sjeff * queue. This could be optimized to not re-add in some 1641177376Sjeff * cases. 1642177376Sjeff */ 1643165766Sjeff if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1644165762Sjeff sched_rem(td); 1645165762Sjeff td->td_priority = prio; 1646171482Sjeff sched_add(td, SRQ_BORROWING); 1647177009Sjeff return; 1648177009Sjeff } 1649177376Sjeff /* 1650177376Sjeff * If the thread is currently running we may have to adjust the lowpri 1651177376Sjeff * information so other cpus are aware of our current priority. 1652177376Sjeff */ 1653177009Sjeff if (TD_IS_RUNNING(td)) { 1654177376Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1655177376Sjeff oldpri = td->td_priority; 1656177376Sjeff td->td_priority = prio; 1657176735Sjeff if (prio < tdq->tdq_lowpri) 1658171482Sjeff tdq->tdq_lowpri = prio; 1659176735Sjeff else if (tdq->tdq_lowpri == oldpri) 1660176735Sjeff tdq_setlowpri(tdq, td); 1661177376Sjeff return; 1662177009Sjeff } 1663177376Sjeff td->td_priority = prio; 1664109864Sjeff} 1665109864Sjeff 1666139453Sjhb/* 1667139453Sjhb * Update a thread's priority when it is lent another thread's 1668139453Sjhb * priority. 1669139453Sjhb */ 1670109864Sjeffvoid 1671139453Sjhbsched_lend_prio(struct thread *td, u_char prio) 1672139453Sjhb{ 1673139453Sjhb 1674139453Sjhb td->td_flags |= TDF_BORROWING; 1675139453Sjhb sched_thread_priority(td, prio); 1676139453Sjhb} 1677139453Sjhb 1678139453Sjhb/* 1679139453Sjhb * Restore a thread's priority when priority propagation is 1680139453Sjhb * over. The prio argument is the minimum priority the thread 1681139453Sjhb * needs to have to satisfy other possible priority lending 1682139453Sjhb * requests. If the thread's regular priority is less 1683139453Sjhb * important than prio, the thread will keep a priority boost 1684139453Sjhb * of prio. 1685139453Sjhb */ 1686139453Sjhbvoid 1687139453Sjhbsched_unlend_prio(struct thread *td, u_char prio) 1688139453Sjhb{ 1689139453Sjhb u_char base_pri; 1690139453Sjhb 1691139453Sjhb if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1692139453Sjhb td->td_base_pri <= PRI_MAX_TIMESHARE) 1693163709Sjb base_pri = td->td_user_pri; 1694139453Sjhb else 1695139453Sjhb base_pri = td->td_base_pri; 1696139453Sjhb if (prio >= base_pri) { 1697139455Sjhb td->td_flags &= ~TDF_BORROWING; 1698139453Sjhb sched_thread_priority(td, base_pri); 1699139453Sjhb } else 1700139453Sjhb sched_lend_prio(td, prio); 1701139453Sjhb} 1702139453Sjhb 1703171482Sjeff/* 1704171482Sjeff * Standard entry for setting the priority to an absolute value. 1705171482Sjeff */ 1706139453Sjhbvoid 1707139453Sjhbsched_prio(struct thread *td, u_char prio) 1708139453Sjhb{ 1709139453Sjhb u_char oldprio; 1710139453Sjhb 1711139453Sjhb /* First, update the base priority. */ 1712139453Sjhb td->td_base_pri = prio; 1713139453Sjhb 1714139453Sjhb /* 1715139455Sjhb * If the thread is borrowing another thread's priority, don't 1716139453Sjhb * ever lower the priority. 1717139453Sjhb */ 1718139453Sjhb if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1719139453Sjhb return; 1720139453Sjhb 1721139453Sjhb /* Change the real priority. */ 1722139453Sjhb oldprio = td->td_priority; 1723139453Sjhb sched_thread_priority(td, prio); 1724139453Sjhb 1725139453Sjhb /* 1726139453Sjhb * If the thread is on a turnstile, then let the turnstile update 1727139453Sjhb * its state. 1728139453Sjhb */ 1729139453Sjhb if (TD_ON_LOCK(td) && oldprio != prio) 1730139453Sjhb turnstile_adjust(td, oldprio); 1731139453Sjhb} 1732139455Sjhb 1733171482Sjeff/* 1734171482Sjeff * Set the base user priority, does not effect current running priority. 1735171482Sjeff */ 1736139453Sjhbvoid 1737163709Sjbsched_user_prio(struct thread *td, u_char prio) 1738161599Sdavidxu{ 1739161599Sdavidxu 1740163709Sjb td->td_base_user_pri = prio; 1741216313Sdavidxu if (td->td_lend_user_pri <= prio) 1742216313Sdavidxu return; 1743163709Sjb td->td_user_pri = prio; 1744161599Sdavidxu} 1745161599Sdavidxu 1746161599Sdavidxuvoid 1747161599Sdavidxusched_lend_user_prio(struct thread *td, u_char prio) 1748161599Sdavidxu{ 1749161599Sdavidxu 1750174536Sdavidxu THREAD_LOCK_ASSERT(td, MA_OWNED); 1751216313Sdavidxu td->td_lend_user_pri = prio; 1752216791Sdavidxu td->td_user_pri = min(prio, td->td_base_user_pri); 1753216791Sdavidxu if (td->td_priority > td->td_user_pri) 1754216791Sdavidxu sched_prio(td, td->td_user_pri); 1755216791Sdavidxu else if (td->td_priority != td->td_user_pri) 1756216791Sdavidxu td->td_flags |= TDF_NEEDRESCHED; 1757161599Sdavidxu} 1758161599Sdavidxu 1759171482Sjeff/* 1760171713Sjeff * Handle migration from sched_switch(). This happens only for 1761171713Sjeff * cpu binding. 1762171713Sjeff */ 1763171713Sjeffstatic struct mtx * 1764171713Sjeffsched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1765171713Sjeff{ 1766171713Sjeff struct tdq *tdn; 1767171713Sjeff 1768171713Sjeff tdn = TDQ_CPU(td->td_sched->ts_cpu); 1769171713Sjeff#ifdef SMP 1770177435Sjeff tdq_load_rem(tdq, td); 1771171713Sjeff /* 1772171713Sjeff * Do the lock dance required to avoid LOR. We grab an extra 1773171713Sjeff * spinlock nesting to prevent preemption while we're 1774171713Sjeff * not holding either run-queue lock. 1775171713Sjeff */ 1776171713Sjeff spinlock_enter(); 1777202889Sattilio thread_lock_block(td); /* This releases the lock on tdq. */ 1778197223Sattilio 1779197223Sattilio /* 1780197223Sattilio * Acquire both run-queue locks before placing the thread on the new 1781197223Sattilio * run-queue to avoid deadlocks created by placing a thread with a 1782197223Sattilio * blocked lock on the run-queue of a remote processor. The deadlock 1783197223Sattilio * occurs when a third processor attempts to lock the two queues in 1784197223Sattilio * question while the target processor is spinning with its own 1785197223Sattilio * run-queue lock held while waiting for the blocked lock to clear. 1786197223Sattilio */ 1787197223Sattilio tdq_lock_pair(tdn, tdq); 1788171713Sjeff tdq_add(tdn, td, flags); 1789177435Sjeff tdq_notify(tdn, td); 1790197223Sattilio TDQ_UNLOCK(tdn); 1791171713Sjeff spinlock_exit(); 1792171713Sjeff#endif 1793171713Sjeff return (TDQ_LOCKPTR(tdn)); 1794171713Sjeff} 1795171713Sjeff 1796171713Sjeff/* 1797202889Sattilio * Variadic version of thread_lock_unblock() that does not assume td_lock 1798202889Sattilio * is blocked. 1799171482Sjeff */ 1800171482Sjeffstatic inline void 1801171482Sjeffthread_unblock_switch(struct thread *td, struct mtx *mtx) 1802171482Sjeff{ 1803171482Sjeff atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1804171482Sjeff (uintptr_t)mtx); 1805171482Sjeff} 1806171482Sjeff 1807171482Sjeff/* 1808171482Sjeff * Switch threads. This function has to handle threads coming in while 1809171482Sjeff * blocked for some reason, running, or idle. It also must deal with 1810171482Sjeff * migrating a thread from one queue to another as running threads may 1811171482Sjeff * be assigned elsewhere via binding. 1812171482Sjeff */ 1813161599Sdavidxuvoid 1814135051Sjuliansched_switch(struct thread *td, struct thread *newtd, int flags) 1815109864Sjeff{ 1816165627Sjeff struct tdq *tdq; 1817164936Sjulian struct td_sched *ts; 1818171482Sjeff struct mtx *mtx; 1819171713Sjeff int srqflag; 1820241248Smav int cpuid, preempted; 1821109864Sjeff 1822170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1823177376Sjeff KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1824109864Sjeff 1825171482Sjeff cpuid = PCPU_GET(cpuid); 1826171482Sjeff tdq = TDQ_CPU(cpuid); 1827164936Sjulian ts = td->td_sched; 1828171713Sjeff mtx = td->td_lock; 1829234166Smav sched_pctcpu_update(ts, 1); 1830171482Sjeff ts->ts_rltick = ticks; 1831133555Sjeff td->td_lastcpu = td->td_oncpu; 1832113339Sjulian td->td_oncpu = NOCPU; 1833241248Smav preempted = !(td->td_flags & TDF_SLICEEND); 1834241248Smav td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND); 1835144777Sups td->td_owepreempt = 0; 1836247150Smav if (!TD_IS_IDLETHREAD(td)) 1837247150Smav tdq->tdq_switchcnt++; 1838123434Sjeff /* 1839171482Sjeff * The lock pointer in an idle thread should never change. Reset it 1840171482Sjeff * to CAN_RUN as well. 1841123434Sjeff */ 1842167327Sjulian if (TD_IS_IDLETHREAD(td)) { 1843171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1844139334Sjeff TD_SET_CAN_RUN(td); 1845170293Sjeff } else if (TD_IS_RUNNING(td)) { 1846171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1847241248Smav srqflag = preempted ? 1848170293Sjeff SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1849171713Sjeff SRQ_OURSELF|SRQ_YIELDING; 1850212153Smdf#ifdef SMP 1851212115Smdf if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu)) 1852212115Smdf ts->ts_cpu = sched_pickcpu(td, 0); 1853212153Smdf#endif 1854171713Sjeff if (ts->ts_cpu == cpuid) 1855177435Sjeff tdq_runq_add(tdq, td, srqflag); 1856212115Smdf else { 1857212115Smdf KASSERT(THREAD_CAN_MIGRATE(td) || 1858212115Smdf (ts->ts_flags & TSF_BOUND) != 0, 1859212115Smdf ("Thread %p shouldn't migrate", td)); 1860171713Sjeff mtx = sched_switch_migrate(tdq, td, srqflag); 1861212115Smdf } 1862171482Sjeff } else { 1863171482Sjeff /* This thread must be going to sleep. */ 1864171482Sjeff TDQ_LOCK(tdq); 1865202889Sattilio mtx = thread_lock_block(td); 1866177435Sjeff tdq_load_rem(tdq, td); 1867171482Sjeff } 1868171482Sjeff /* 1869171482Sjeff * We enter here with the thread blocked and assigned to the 1870171482Sjeff * appropriate cpu run-queue or sleep-queue and with the current 1871171482Sjeff * thread-queue locked. 1872171482Sjeff */ 1873171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1874171482Sjeff newtd = choosethread(); 1875171482Sjeff /* 1876171482Sjeff * Call the MD code to switch contexts if necessary. 1877171482Sjeff */ 1878145256Sjkoshy if (td != newtd) { 1879145256Sjkoshy#ifdef HWPMC_HOOKS 1880145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1881145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1882145256Sjkoshy#endif 1883262057Savg SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc); 1884174629Sjeff lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1885172411Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1886234166Smav sched_pctcpu_update(newtd->td_sched, 0); 1887179297Sjb 1888179297Sjb#ifdef KDTRACE_HOOKS 1889179297Sjb /* 1890179297Sjb * If DTrace has set the active vtime enum to anything 1891179297Sjb * other than INACTIVE (0), then it should have set the 1892179297Sjb * function to call. 1893179297Sjb */ 1894179297Sjb if (dtrace_vtime_active) 1895179297Sjb (*dtrace_vtime_switch_func)(newtd); 1896179297Sjb#endif 1897179297Sjb 1898171482Sjeff cpu_switch(td, newtd, mtx); 1899171482Sjeff /* 1900171482Sjeff * We may return from cpu_switch on a different cpu. However, 1901171482Sjeff * we always return with td_lock pointing to the current cpu's 1902171482Sjeff * run queue lock. 1903171482Sjeff */ 1904171482Sjeff cpuid = PCPU_GET(cpuid); 1905171482Sjeff tdq = TDQ_CPU(cpuid); 1906174629Sjeff lock_profile_obtain_lock_success( 1907174629Sjeff &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1908236344Srstone 1909262057Savg SDT_PROBE0(sched, , , on__cpu); 1910145256Sjkoshy#ifdef HWPMC_HOOKS 1911145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1912145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1913145256Sjkoshy#endif 1914236344Srstone } else { 1915171482Sjeff thread_unblock_switch(td, mtx); 1916262057Savg SDT_PROBE0(sched, , , remain__cpu); 1917236344Srstone } 1918171482Sjeff /* 1919171482Sjeff * Assert that all went well and return. 1920171482Sjeff */ 1921171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1922171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1923171482Sjeff td->td_oncpu = cpuid; 1924109864Sjeff} 1925109864Sjeff 1926171482Sjeff/* 1927171482Sjeff * Adjust thread priorities as a result of a nice request. 1928171482Sjeff */ 1929109864Sjeffvoid 1930130551Sjuliansched_nice(struct proc *p, int nice) 1931109864Sjeff{ 1932109864Sjeff struct thread *td; 1933109864Sjeff 1934130551Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 1935165762Sjeff 1936130551Sjulian p->p_nice = nice; 1937163709Sjb FOREACH_THREAD_IN_PROC(p, td) { 1938170293Sjeff thread_lock(td); 1939163709Sjb sched_priority(td); 1940165762Sjeff sched_prio(td, td->td_base_user_pri); 1941170293Sjeff thread_unlock(td); 1942130551Sjulian } 1943109864Sjeff} 1944109864Sjeff 1945171482Sjeff/* 1946171482Sjeff * Record the sleep time for the interactivity scorer. 1947171482Sjeff */ 1948109864Sjeffvoid 1949177085Sjeffsched_sleep(struct thread *td, int prio) 1950109864Sjeff{ 1951165762Sjeff 1952170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1953109864Sjeff 1954172264Sjeff td->td_slptick = ticks; 1955201347Skib if (TD_IS_SUSPENDED(td) || prio >= PSOCK) 1956177085Sjeff td->td_flags |= TDF_CANSWAP; 1957217410Sjhb if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1958217410Sjhb return; 1959177903Sjeff if (static_boost == 1 && prio) 1960177085Sjeff sched_prio(td, prio); 1961177903Sjeff else if (static_boost && td->td_priority > static_boost) 1962177903Sjeff sched_prio(td, static_boost); 1963109864Sjeff} 1964109864Sjeff 1965171482Sjeff/* 1966171482Sjeff * Schedule a thread to resume execution and record how long it voluntarily 1967171482Sjeff * slept. We also update the pctcpu, interactivity, and priority. 1968171482Sjeff */ 1969109864Sjeffvoid 1970109864Sjeffsched_wakeup(struct thread *td) 1971109864Sjeff{ 1972166229Sjeff struct td_sched *ts; 1973171482Sjeff int slptick; 1974165762Sjeff 1975170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1976166229Sjeff ts = td->td_sched; 1977177085Sjeff td->td_flags &= ~TDF_CANSWAP; 1978109864Sjeff /* 1979165762Sjeff * If we slept for more than a tick update our interactivity and 1980165762Sjeff * priority. 1981109864Sjeff */ 1982172264Sjeff slptick = td->td_slptick; 1983172264Sjeff td->td_slptick = 0; 1984171482Sjeff if (slptick && slptick != ticks) { 1985234166Smav ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT; 1986165819Sjeff sched_interact_update(td); 1987234166Smav sched_pctcpu_update(ts, 0); 1988109864Sjeff } 1989166229Sjeff /* Reset the slice value after we sleep. */ 1990166229Sjeff ts->ts_slice = sched_slice; 1991166190Sjeff sched_add(td, SRQ_BORING); 1992109864Sjeff} 1993109864Sjeff 1994109864Sjeff/* 1995109864Sjeff * Penalize the parent for creating a new child and initialize the child's 1996109864Sjeff * priority. 1997109864Sjeff */ 1998109864Sjeffvoid 1999163709Sjbsched_fork(struct thread *td, struct thread *child) 2000109864Sjeff{ 2001170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2002234166Smav sched_pctcpu_update(td->td_sched, 1); 2003164936Sjulian sched_fork_thread(td, child); 2004165762Sjeff /* 2005165762Sjeff * Penalize the parent and child for forking. 2006165762Sjeff */ 2007165762Sjeff sched_interact_fork(child); 2008165762Sjeff sched_priority(child); 2009171482Sjeff td->td_sched->ts_runtime += tickincr; 2010165762Sjeff sched_interact_update(td); 2011165762Sjeff sched_priority(td); 2012164936Sjulian} 2013109864Sjeff 2014171482Sjeff/* 2015171482Sjeff * Fork a new thread, may be within the same process. 2016171482Sjeff */ 2017164936Sjulianvoid 2018164936Sjuliansched_fork_thread(struct thread *td, struct thread *child) 2019164936Sjulian{ 2020164936Sjulian struct td_sched *ts; 2021164936Sjulian struct td_sched *ts2; 2022164936Sjulian 2023177426Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2024165762Sjeff /* 2025165762Sjeff * Initialize child. 2026165762Sjeff */ 2027177426Sjeff ts = td->td_sched; 2028177426Sjeff ts2 = child->td_sched; 2029171482Sjeff child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 2030176735Sjeff child->td_cpuset = cpuset_ref(td->td_cpuset); 2031164936Sjulian ts2->ts_cpu = ts->ts_cpu; 2032177426Sjeff ts2->ts_flags = 0; 2033165762Sjeff /* 2034217078Sjhb * Grab our parents cpu estimation information. 2035165762Sjeff */ 2036164936Sjulian ts2->ts_ticks = ts->ts_ticks; 2037164936Sjulian ts2->ts_ltick = ts->ts_ltick; 2038164936Sjulian ts2->ts_ftick = ts->ts_ftick; 2039165762Sjeff /* 2040217078Sjhb * Do not inherit any borrowed priority from the parent. 2041217078Sjhb */ 2042217078Sjhb child->td_priority = child->td_base_pri; 2043217078Sjhb /* 2044165762Sjeff * And update interactivity score. 2045165762Sjeff */ 2046171482Sjeff ts2->ts_slptime = ts->ts_slptime; 2047171482Sjeff ts2->ts_runtime = ts->ts_runtime; 2048165762Sjeff ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 2049187357Sjeff#ifdef KTR 2050187357Sjeff bzero(ts2->ts_name, sizeof(ts2->ts_name)); 2051187357Sjeff#endif 2052113357Sjeff} 2053113357Sjeff 2054171482Sjeff/* 2055171482Sjeff * Adjust the priority class of a thread. 2056171482Sjeff */ 2057113357Sjeffvoid 2058163709Sjbsched_class(struct thread *td, int class) 2059113357Sjeff{ 2060113357Sjeff 2061170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2062163709Sjb if (td->td_pri_class == class) 2063113357Sjeff return; 2064163709Sjb td->td_pri_class = class; 2065109864Sjeff} 2066109864Sjeff 2067109864Sjeff/* 2068109864Sjeff * Return some of the child's priority and interactivity to the parent. 2069109864Sjeff */ 2070109864Sjeffvoid 2071164939Sjuliansched_exit(struct proc *p, struct thread *child) 2072109864Sjeff{ 2073165762Sjeff struct thread *td; 2074113372Sjeff 2075187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit", 2076225199Sdelphij "prio:%d", child->td_priority); 2077177368Sjeff PROC_LOCK_ASSERT(p, MA_OWNED); 2078165762Sjeff td = FIRST_THREAD_IN_PROC(p); 2079165762Sjeff sched_exit_thread(td, child); 2080113372Sjeff} 2081113372Sjeff 2082171482Sjeff/* 2083171482Sjeff * Penalize another thread for the time spent on this one. This helps to 2084171482Sjeff * worsen the priority and interactivity of processes which schedule batch 2085171482Sjeff * jobs such as make. This has little effect on the make process itself but 2086171482Sjeff * causes new processes spawned by it to receive worse scores immediately. 2087171482Sjeff */ 2088113372Sjeffvoid 2089164939Sjuliansched_exit_thread(struct thread *td, struct thread *child) 2090164936Sjulian{ 2091165762Sjeff 2092187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit", 2093225199Sdelphij "prio:%d", child->td_priority); 2094165762Sjeff /* 2095165762Sjeff * Give the child's runtime to the parent without returning the 2096165762Sjeff * sleep time as a penalty to the parent. This causes shells that 2097165762Sjeff * launch expensive things to mark their children as expensive. 2098165762Sjeff */ 2099170293Sjeff thread_lock(td); 2100171482Sjeff td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2101164939Sjulian sched_interact_update(td); 2102165762Sjeff sched_priority(td); 2103170293Sjeff thread_unlock(td); 2104164936Sjulian} 2105164936Sjulian 2106177005Sjeffvoid 2107177005Sjeffsched_preempt(struct thread *td) 2108177005Sjeff{ 2109177005Sjeff struct tdq *tdq; 2110177005Sjeff 2111236344Srstone SDT_PROBE2(sched, , , surrender, td, td->td_proc); 2112236344Srstone 2113177005Sjeff thread_lock(td); 2114177005Sjeff tdq = TDQ_SELF(); 2115177005Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2116177005Sjeff tdq->tdq_ipipending = 0; 2117177005Sjeff if (td->td_priority > tdq->tdq_lowpri) { 2118178272Sjeff int flags; 2119178272Sjeff 2120178272Sjeff flags = SW_INVOL | SW_PREEMPT; 2121177005Sjeff if (td->td_critnest > 1) 2122177005Sjeff td->td_owepreempt = 1; 2123178272Sjeff else if (TD_IS_IDLETHREAD(td)) 2124178272Sjeff mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL); 2125177005Sjeff else 2126178272Sjeff mi_switch(flags | SWT_REMOTEPREEMPT, NULL); 2127177005Sjeff } 2128177005Sjeff thread_unlock(td); 2129177005Sjeff} 2130177005Sjeff 2131171482Sjeff/* 2132171482Sjeff * Fix priorities on return to user-space. Priorities may be elevated due 2133171482Sjeff * to static priorities in msleep() or similar. 2134171482Sjeff */ 2135164936Sjulianvoid 2136164936Sjuliansched_userret(struct thread *td) 2137164936Sjulian{ 2138164936Sjulian /* 2139164936Sjulian * XXX we cheat slightly on the locking here to avoid locking in 2140164936Sjulian * the usual case. Setting td_priority here is essentially an 2141164936Sjulian * incomplete workaround for not setting it properly elsewhere. 2142164936Sjulian * Now that some interrupt handlers are threads, not setting it 2143164936Sjulian * properly elsewhere can clobber it in the window between setting 2144164936Sjulian * it here and returning to user mode, so don't waste time setting 2145164936Sjulian * it perfectly here. 2146164936Sjulian */ 2147164936Sjulian KASSERT((td->td_flags & TDF_BORROWING) == 0, 2148164936Sjulian ("thread with borrowed priority returning to userland")); 2149164936Sjulian if (td->td_priority != td->td_user_pri) { 2150170293Sjeff thread_lock(td); 2151164936Sjulian td->td_priority = td->td_user_pri; 2152164936Sjulian td->td_base_pri = td->td_user_pri; 2153177005Sjeff tdq_setlowpri(TDQ_SELF(), td); 2154170293Sjeff thread_unlock(td); 2155164936Sjulian } 2156164936Sjulian} 2157164936Sjulian 2158171482Sjeff/* 2159171482Sjeff * Handle a stathz tick. This is really only relevant for timeshare 2160171482Sjeff * threads. 2161171482Sjeff */ 2162164936Sjulianvoid 2163121127Sjeffsched_clock(struct thread *td) 2164109864Sjeff{ 2165164936Sjulian struct tdq *tdq; 2166164936Sjulian struct td_sched *ts; 2167109864Sjeff 2168171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2169164936Sjulian tdq = TDQ_SELF(); 2170172409Sjeff#ifdef SMP 2171133427Sjeff /* 2172172409Sjeff * We run the long term load balancer infrequently on the first cpu. 2173172409Sjeff */ 2174172409Sjeff if (balance_tdq == tdq) { 2175172409Sjeff if (balance_ticks && --balance_ticks == 0) 2176172409Sjeff sched_balance(); 2177172409Sjeff } 2178172409Sjeff#endif 2179172409Sjeff /* 2180178277Sjeff * Save the old switch count so we have a record of the last ticks 2181178277Sjeff * activity. Initialize the new switch count based on our load. 2182178277Sjeff * If there is some activity seed it to reflect that. 2183178277Sjeff */ 2184178277Sjeff tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2185178471Sjeff tdq->tdq_switchcnt = tdq->tdq_load; 2186178277Sjeff /* 2187165766Sjeff * Advance the insert index once for each tick to ensure that all 2188165766Sjeff * threads get a chance to run. 2189133427Sjeff */ 2190165766Sjeff if (tdq->tdq_idx == tdq->tdq_ridx) { 2191165766Sjeff tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2192165766Sjeff if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2193165766Sjeff tdq->tdq_ridx = tdq->tdq_idx; 2194165766Sjeff } 2195165766Sjeff ts = td->td_sched; 2196234166Smav sched_pctcpu_update(ts, 1); 2197175104Sjeff if (td->td_pri_class & PRI_FIFO_BIT) 2198113357Sjeff return; 2199217291Sjhb if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) { 2200175104Sjeff /* 2201175104Sjeff * We used a tick; charge it to the thread so 2202175104Sjeff * that we can compute our interactivity. 2203175104Sjeff */ 2204175104Sjeff td->td_sched->ts_runtime += tickincr; 2205175104Sjeff sched_interact_update(td); 2206177009Sjeff sched_priority(td); 2207175104Sjeff } 2208241249Smav 2209113357Sjeff /* 2210241249Smav * Force a context switch if the current thread has used up a full 2211241249Smav * time slice (default is 100ms). 2212109864Sjeff */ 2213241249Smav if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) { 2214241249Smav ts->ts_slice = sched_slice; 2215241249Smav td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND; 2216241249Smav } 2217109864Sjeff} 2218109864Sjeff 2219171482Sjeff/* 2220234166Smav * Called once per hz tick. 2221171482Sjeff */ 2222171482Sjeffvoid 2223212541Smavsched_tick(int cnt) 2224171482Sjeff{ 2225171482Sjeff 2226171482Sjeff} 2227171482Sjeff 2228171482Sjeff/* 2229171482Sjeff * Return whether the current CPU has runnable tasks. Used for in-kernel 2230171482Sjeff * cooperative idle threads. 2231171482Sjeff */ 2232109864Sjeffint 2233109864Sjeffsched_runnable(void) 2234109864Sjeff{ 2235164936Sjulian struct tdq *tdq; 2236115998Sjeff int load; 2237109864Sjeff 2238115998Sjeff load = 1; 2239115998Sjeff 2240164936Sjulian tdq = TDQ_SELF(); 2241121605Sjeff if ((curthread->td_flags & TDF_IDLETD) != 0) { 2242165620Sjeff if (tdq->tdq_load > 0) 2243121605Sjeff goto out; 2244121605Sjeff } else 2245165620Sjeff if (tdq->tdq_load - 1 > 0) 2246121605Sjeff goto out; 2247115998Sjeff load = 0; 2248115998Sjeffout: 2249115998Sjeff return (load); 2250109864Sjeff} 2251109864Sjeff 2252171482Sjeff/* 2253171482Sjeff * Choose the highest priority thread to run. The thread is removed from 2254171482Sjeff * the run-queue while running however the load remains. For SMP we set 2255171482Sjeff * the tdq in the global idle bitmask if it idles here. 2256171482Sjeff */ 2257166190Sjeffstruct thread * 2258109970Sjeffsched_choose(void) 2259109970Sjeff{ 2260177435Sjeff struct thread *td; 2261164936Sjulian struct tdq *tdq; 2262109970Sjeff 2263164936Sjulian tdq = TDQ_SELF(); 2264171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2265177435Sjeff td = tdq_choose(tdq); 2266177435Sjeff if (td) { 2267177435Sjeff tdq_runq_rem(tdq, td); 2268177903Sjeff tdq->tdq_lowpri = td->td_priority; 2269177435Sjeff return (td); 2270109864Sjeff } 2271177903Sjeff tdq->tdq_lowpri = PRI_MAX_IDLE; 2272176735Sjeff return (PCPU_GET(idlethread)); 2273109864Sjeff} 2274109864Sjeff 2275171482Sjeff/* 2276171482Sjeff * Set owepreempt if necessary. Preemption never happens directly in ULE, 2277171482Sjeff * we always request it once we exit a critical section. 2278171482Sjeff */ 2279171482Sjeffstatic inline void 2280171482Sjeffsched_setpreempt(struct thread *td) 2281166190Sjeff{ 2282166190Sjeff struct thread *ctd; 2283166190Sjeff int cpri; 2284166190Sjeff int pri; 2285166190Sjeff 2286177005Sjeff THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2287177005Sjeff 2288166190Sjeff ctd = curthread; 2289166190Sjeff pri = td->td_priority; 2290166190Sjeff cpri = ctd->td_priority; 2291177005Sjeff if (pri < cpri) 2292177005Sjeff ctd->td_flags |= TDF_NEEDRESCHED; 2293166190Sjeff if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2294171482Sjeff return; 2295177005Sjeff if (!sched_shouldpreempt(pri, cpri, 0)) 2296171482Sjeff return; 2297171482Sjeff ctd->td_owepreempt = 1; 2298166190Sjeff} 2299166190Sjeff 2300171482Sjeff/* 2301177009Sjeff * Add a thread to a thread queue. Select the appropriate runq and add the 2302177009Sjeff * thread to it. This is the internal function called when the tdq is 2303177009Sjeff * predetermined. 2304171482Sjeff */ 2305109864Sjeffvoid 2306171482Sjefftdq_add(struct tdq *tdq, struct thread *td, int flags) 2307109864Sjeff{ 2308109864Sjeff 2309171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2310166190Sjeff KASSERT((td->td_inhibitors == 0), 2311166190Sjeff ("sched_add: trying to run inhibited thread")); 2312166190Sjeff KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2313166190Sjeff ("sched_add: bad thread state")); 2314172207Sjeff KASSERT(td->td_flags & TDF_INMEM, 2315172207Sjeff ("sched_add: thread swapped out")); 2316171482Sjeff 2317171482Sjeff if (td->td_priority < tdq->tdq_lowpri) 2318171482Sjeff tdq->tdq_lowpri = td->td_priority; 2319177435Sjeff tdq_runq_add(tdq, td, flags); 2320177435Sjeff tdq_load_add(tdq, td); 2321171482Sjeff} 2322171482Sjeff 2323171482Sjeff/* 2324171482Sjeff * Select the target thread queue and add a thread to it. Request 2325171482Sjeff * preemption or IPI a remote processor if required. 2326171482Sjeff */ 2327171482Sjeffvoid 2328171482Sjeffsched_add(struct thread *td, int flags) 2329171482Sjeff{ 2330171482Sjeff struct tdq *tdq; 2331171482Sjeff#ifdef SMP 2332171482Sjeff int cpu; 2333171482Sjeff#endif 2334187357Sjeff 2335187357Sjeff KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 2336187357Sjeff "prio:%d", td->td_priority, KTR_ATTR_LINKED, 2337187357Sjeff sched_tdname(curthread)); 2338187357Sjeff KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 2339187357Sjeff KTR_ATTR_LINKED, sched_tdname(td)); 2340236344Srstone SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 2341236344Srstone flags & SRQ_PREEMPTED); 2342171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2343166108Sjeff /* 2344171482Sjeff * Recalculate the priority before we select the target cpu or 2345171482Sjeff * run-queue. 2346166108Sjeff */ 2347171482Sjeff if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2348171482Sjeff sched_priority(td); 2349171482Sjeff#ifdef SMP 2350171482Sjeff /* 2351171482Sjeff * Pick the destination cpu and if it isn't ours transfer to the 2352171482Sjeff * target cpu. 2353171482Sjeff */ 2354177435Sjeff cpu = sched_pickcpu(td, flags); 2355177435Sjeff tdq = sched_setcpu(td, cpu, flags); 2356171482Sjeff tdq_add(tdq, td, flags); 2357177009Sjeff if (cpu != PCPU_GET(cpuid)) { 2358177435Sjeff tdq_notify(tdq, td); 2359166108Sjeff return; 2360166108Sjeff } 2361171482Sjeff#else 2362171482Sjeff tdq = TDQ_SELF(); 2363171482Sjeff TDQ_LOCK(tdq); 2364171482Sjeff /* 2365171482Sjeff * Now that the thread is moving to the run-queue, set the lock 2366171482Sjeff * to the scheduler's lock. 2367171482Sjeff */ 2368171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2369171482Sjeff tdq_add(tdq, td, flags); 2370166108Sjeff#endif 2371171482Sjeff if (!(flags & SRQ_YIELDING)) 2372171482Sjeff sched_setpreempt(td); 2373109864Sjeff} 2374109864Sjeff 2375171482Sjeff/* 2376171482Sjeff * Remove a thread from a run-queue without running it. This is used 2377171482Sjeff * when we're stealing a thread from a remote queue. Otherwise all threads 2378171482Sjeff * exit by calling sched_exit_thread() and sched_throw() themselves. 2379171482Sjeff */ 2380109864Sjeffvoid 2381121127Sjeffsched_rem(struct thread *td) 2382109864Sjeff{ 2383164936Sjulian struct tdq *tdq; 2384113357Sjeff 2385187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 2386187357Sjeff "prio:%d", td->td_priority); 2387236344Srstone SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); 2388177435Sjeff tdq = TDQ_CPU(td->td_sched->ts_cpu); 2389171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2390171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2391166190Sjeff KASSERT(TD_ON_RUNQ(td), 2392164936Sjulian ("sched_rem: thread not on run queue")); 2393177435Sjeff tdq_runq_rem(tdq, td); 2394177435Sjeff tdq_load_rem(tdq, td); 2395166190Sjeff TD_SET_CAN_RUN(td); 2396176735Sjeff if (td->td_priority == tdq->tdq_lowpri) 2397176735Sjeff tdq_setlowpri(tdq, NULL); 2398109864Sjeff} 2399109864Sjeff 2400171482Sjeff/* 2401171482Sjeff * Fetch cpu utilization information. Updates on demand. 2402171482Sjeff */ 2403109864Sjefffixpt_t 2404121127Sjeffsched_pctcpu(struct thread *td) 2405109864Sjeff{ 2406109864Sjeff fixpt_t pctcpu; 2407164936Sjulian struct td_sched *ts; 2408109864Sjeff 2409109864Sjeff pctcpu = 0; 2410164936Sjulian ts = td->td_sched; 2411164936Sjulian if (ts == NULL) 2412121290Sjeff return (0); 2413109864Sjeff 2414208787Sjhb THREAD_LOCK_ASSERT(td, MA_OWNED); 2415234166Smav sched_pctcpu_update(ts, TD_IS_RUNNING(td)); 2416164936Sjulian if (ts->ts_ticks) { 2417109864Sjeff int rtick; 2418109864Sjeff 2419109864Sjeff /* How many rtick per second ? */ 2420165762Sjeff rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2421165762Sjeff pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2422109864Sjeff } 2423109864Sjeff 2424109864Sjeff return (pctcpu); 2425109864Sjeff} 2426109864Sjeff 2427176735Sjeff/* 2428176735Sjeff * Enforce affinity settings for a thread. Called after adjustments to 2429176735Sjeff * cpumask. 2430176735Sjeff */ 2431176729Sjeffvoid 2432176729Sjeffsched_affinity(struct thread *td) 2433176729Sjeff{ 2434176735Sjeff#ifdef SMP 2435176735Sjeff struct td_sched *ts; 2436176735Sjeff 2437176735Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2438176735Sjeff ts = td->td_sched; 2439176735Sjeff if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2440176735Sjeff return; 2441189787Sjeff if (TD_ON_RUNQ(td)) { 2442189787Sjeff sched_rem(td); 2443189787Sjeff sched_add(td, SRQ_BORING); 2444189787Sjeff return; 2445189787Sjeff } 2446176735Sjeff if (!TD_IS_RUNNING(td)) 2447176735Sjeff return; 2448176735Sjeff /* 2449212115Smdf * Force a switch before returning to userspace. If the 2450212115Smdf * target thread is not running locally send an ipi to force 2451212115Smdf * the issue. 2452176735Sjeff */ 2453212974Sjhb td->td_flags |= TDF_NEEDRESCHED; 2454212115Smdf if (td != curthread) 2455212115Smdf ipi_cpu(ts->ts_cpu, IPI_PREEMPT); 2456176735Sjeff#endif 2457176729Sjeff} 2458176729Sjeff 2459171482Sjeff/* 2460171482Sjeff * Bind a thread to a target cpu. 2461171482Sjeff */ 2462122038Sjeffvoid 2463122038Sjeffsched_bind(struct thread *td, int cpu) 2464122038Sjeff{ 2465164936Sjulian struct td_sched *ts; 2466122038Sjeff 2467171713Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2468208391Sjhb KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 2469164936Sjulian ts = td->td_sched; 2470166137Sjeff if (ts->ts_flags & TSF_BOUND) 2471166152Sjeff sched_unbind(td); 2472212115Smdf KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td)); 2473164936Sjulian ts->ts_flags |= TSF_BOUND; 2474166137Sjeff sched_pin(); 2475123433Sjeff if (PCPU_GET(cpuid) == cpu) 2476122038Sjeff return; 2477166137Sjeff ts->ts_cpu = cpu; 2478122038Sjeff /* When we return from mi_switch we'll be on the correct cpu. */ 2479131527Sphk mi_switch(SW_VOL, NULL); 2480122038Sjeff} 2481122038Sjeff 2482171482Sjeff/* 2483171482Sjeff * Release a bound thread. 2484171482Sjeff */ 2485122038Sjeffvoid 2486122038Sjeffsched_unbind(struct thread *td) 2487122038Sjeff{ 2488165762Sjeff struct td_sched *ts; 2489165762Sjeff 2490170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2491208391Sjhb KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 2492165762Sjeff ts = td->td_sched; 2493166137Sjeff if ((ts->ts_flags & TSF_BOUND) == 0) 2494166137Sjeff return; 2495165762Sjeff ts->ts_flags &= ~TSF_BOUND; 2496165762Sjeff sched_unpin(); 2497122038Sjeff} 2498122038Sjeff 2499109864Sjeffint 2500145256Sjkoshysched_is_bound(struct thread *td) 2501145256Sjkoshy{ 2502170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2503164936Sjulian return (td->td_sched->ts_flags & TSF_BOUND); 2504145256Sjkoshy} 2505145256Sjkoshy 2506171482Sjeff/* 2507171482Sjeff * Basic yield call. 2508171482Sjeff */ 2509159630Sdavidxuvoid 2510159630Sdavidxusched_relinquish(struct thread *td) 2511159630Sdavidxu{ 2512170293Sjeff thread_lock(td); 2513178272Sjeff mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 2514170293Sjeff thread_unlock(td); 2515159630Sdavidxu} 2516159630Sdavidxu 2517171482Sjeff/* 2518171482Sjeff * Return the total system load. 2519171482Sjeff */ 2520145256Sjkoshyint 2521125289Sjeffsched_load(void) 2522125289Sjeff{ 2523125289Sjeff#ifdef SMP 2524125289Sjeff int total; 2525125289Sjeff int i; 2526125289Sjeff 2527125289Sjeff total = 0; 2528209059Sjhb CPU_FOREACH(i) 2529176735Sjeff total += TDQ_CPU(i)->tdq_sysload; 2530125289Sjeff return (total); 2531125289Sjeff#else 2532165620Sjeff return (TDQ_SELF()->tdq_sysload); 2533125289Sjeff#endif 2534125289Sjeff} 2535125289Sjeff 2536125289Sjeffint 2537109864Sjeffsched_sizeof_proc(void) 2538109864Sjeff{ 2539109864Sjeff return (sizeof(struct proc)); 2540109864Sjeff} 2541109864Sjeff 2542109864Sjeffint 2543109864Sjeffsched_sizeof_thread(void) 2544109864Sjeff{ 2545109864Sjeff return (sizeof(struct thread) + sizeof(struct td_sched)); 2546109864Sjeff} 2547159570Sdavidxu 2548191676Sjeff#ifdef SMP 2549191676Sjeff#define TDQ_IDLESPIN(tdq) \ 2550191676Sjeff ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0) 2551191676Sjeff#else 2552191676Sjeff#define TDQ_IDLESPIN(tdq) 1 2553191676Sjeff#endif 2554191676Sjeff 2555166190Sjeff/* 2556166190Sjeff * The actual idle process. 2557166190Sjeff */ 2558166190Sjeffvoid 2559166190Sjeffsched_idletd(void *dummy) 2560166190Sjeff{ 2561166190Sjeff struct thread *td; 2562171482Sjeff struct tdq *tdq; 2563247150Smav int oldswitchcnt, switchcnt; 2564178277Sjeff int i; 2565166190Sjeff 2566191643Sjeff mtx_assert(&Giant, MA_NOTOWNED); 2567166190Sjeff td = curthread; 2568171482Sjeff tdq = TDQ_SELF(); 2569247150Smav oldswitchcnt = -1; 2570171482Sjeff for (;;) { 2571247150Smav if (tdq->tdq_load) { 2572247150Smav thread_lock(td); 2573247150Smav mi_switch(SW_VOL | SWT_IDLE, NULL); 2574247150Smav thread_unlock(td); 2575247150Smav } 2576247150Smav switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2577171482Sjeff#ifdef SMP 2578247150Smav if (switchcnt != oldswitchcnt) { 2579247150Smav oldswitchcnt = switchcnt; 2580247150Smav if (tdq_idled(tdq) == 0) 2581247150Smav continue; 2582247150Smav } 2583247150Smav switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2584247150Smav#else 2585247150Smav oldswitchcnt = switchcnt; 2586171482Sjeff#endif 2587178277Sjeff /* 2588178277Sjeff * If we're switching very frequently, spin while checking 2589178277Sjeff * for load rather than entering a low power state that 2590191643Sjeff * may require an IPI. However, don't do any busy 2591191643Sjeff * loops while on SMT machines as this simply steals 2592191643Sjeff * cycles from cores doing useful work. 2593178277Sjeff */ 2594191676Sjeff if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) { 2595178277Sjeff for (i = 0; i < sched_idlespins; i++) { 2596178277Sjeff if (tdq->tdq_load) 2597178277Sjeff break; 2598178277Sjeff cpu_spinwait(); 2599178277Sjeff } 2600178277Sjeff } 2601247150Smav 2602247150Smav /* If there was context switch during spin, restart it. */ 2603191643Sjeff switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2604247150Smav if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt) 2605247150Smav continue; 2606247150Smav 2607247150Smav /* Run main MD idle handler. */ 2608247150Smav tdq->tdq_cpu_idle = 1; 2609247150Smav cpu_idle(switchcnt * 4 > sched_idlespinthresh); 2610247150Smav tdq->tdq_cpu_idle = 0; 2611247150Smav 2612247150Smav /* 2613247150Smav * Account thread-less hardware interrupts and 2614247150Smav * other wakeup reasons equal to context switches. 2615247150Smav */ 2616247150Smav switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2617247150Smav if (switchcnt != oldswitchcnt) 2618247150Smav continue; 2619247150Smav tdq->tdq_switchcnt++; 2620247150Smav oldswitchcnt++; 2621171482Sjeff } 2622166190Sjeff} 2623166190Sjeff 2624170293Sjeff/* 2625170293Sjeff * A CPU is entering for the first time or a thread is exiting. 2626170293Sjeff */ 2627170293Sjeffvoid 2628170293Sjeffsched_throw(struct thread *td) 2629170293Sjeff{ 2630172411Sjeff struct thread *newtd; 2631171482Sjeff struct tdq *tdq; 2632171482Sjeff 2633171482Sjeff tdq = TDQ_SELF(); 2634170293Sjeff if (td == NULL) { 2635171482Sjeff /* Correct spinlock nesting and acquire the correct lock. */ 2636171482Sjeff TDQ_LOCK(tdq); 2637170293Sjeff spinlock_exit(); 2638230079Sjhb PCPU_SET(switchtime, cpu_ticks()); 2639230079Sjhb PCPU_SET(switchticks, ticks); 2640170293Sjeff } else { 2641171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2642177435Sjeff tdq_load_rem(tdq, td); 2643174629Sjeff lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2644170293Sjeff } 2645170293Sjeff KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2646172411Sjeff newtd = choosethread(); 2647172411Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2648172411Sjeff cpu_throw(td, newtd); /* doesn't return */ 2649170293Sjeff} 2650170293Sjeff 2651171482Sjeff/* 2652171482Sjeff * This is called from fork_exit(). Just acquire the correct locks and 2653171482Sjeff * let fork do the rest of the work. 2654171482Sjeff */ 2655170293Sjeffvoid 2656170600Sjeffsched_fork_exit(struct thread *td) 2657170293Sjeff{ 2658171482Sjeff struct td_sched *ts; 2659171482Sjeff struct tdq *tdq; 2660171482Sjeff int cpuid; 2661170293Sjeff 2662170293Sjeff /* 2663170293Sjeff * Finish setting up thread glue so that it begins execution in a 2664171482Sjeff * non-nested critical section with the scheduler lock held. 2665170293Sjeff */ 2666171482Sjeff cpuid = PCPU_GET(cpuid); 2667171482Sjeff tdq = TDQ_CPU(cpuid); 2668171482Sjeff ts = td->td_sched; 2669171482Sjeff if (TD_IS_IDLETHREAD(td)) 2670171482Sjeff td->td_lock = TDQ_LOCKPTR(tdq); 2671171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2672171482Sjeff td->td_oncpu = cpuid; 2673172411Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2674174629Sjeff lock_profile_obtain_lock_success( 2675174629Sjeff &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2676170293Sjeff} 2677170293Sjeff 2678187357Sjeff/* 2679187357Sjeff * Create on first use to catch odd startup conditons. 2680187357Sjeff */ 2681187357Sjeffchar * 2682187357Sjeffsched_tdname(struct thread *td) 2683187357Sjeff{ 2684187357Sjeff#ifdef KTR 2685187357Sjeff struct td_sched *ts; 2686187357Sjeff 2687187357Sjeff ts = td->td_sched; 2688187357Sjeff if (ts->ts_name[0] == '\0') 2689187357Sjeff snprintf(ts->ts_name, sizeof(ts->ts_name), 2690187357Sjeff "%s tid %d", td->td_name, td->td_tid); 2691187357Sjeff return (ts->ts_name); 2692187357Sjeff#else 2693187357Sjeff return (td->td_name); 2694187357Sjeff#endif 2695187357Sjeff} 2696187357Sjeff 2697233814Sjhb#ifdef KTR 2698233814Sjhbvoid 2699233814Sjhbsched_clear_tdname(struct thread *td) 2700233814Sjhb{ 2701233814Sjhb struct td_sched *ts; 2702233814Sjhb 2703233814Sjhb ts = td->td_sched; 2704233814Sjhb ts->ts_name[0] = '\0'; 2705233814Sjhb} 2706233814Sjhb#endif 2707233814Sjhb 2708184439Sivoras#ifdef SMP 2709184439Sivoras 2710184439Sivoras/* 2711184439Sivoras * Build the CPU topology dump string. Is recursively called to collect 2712184439Sivoras * the topology tree. 2713184439Sivoras */ 2714184439Sivorasstatic int 2715184439Sivorassysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg, 2716184439Sivoras int indent) 2717184439Sivoras{ 2718222813Sattilio char cpusetbuf[CPUSETBUFSIZ]; 2719184439Sivoras int i, first; 2720184439Sivoras 2721184439Sivoras sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent, 2722212821Savg "", 1 + indent / 2, cg->cg_level); 2723222813Sattilio sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "", 2724222813Sattilio cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask)); 2725184439Sivoras first = TRUE; 2726184439Sivoras for (i = 0; i < MAXCPU; i++) { 2727222813Sattilio if (CPU_ISSET(i, &cg->cg_mask)) { 2728184439Sivoras if (!first) 2729184439Sivoras sbuf_printf(sb, ", "); 2730184439Sivoras else 2731184439Sivoras first = FALSE; 2732184439Sivoras sbuf_printf(sb, "%d", i); 2733184439Sivoras } 2734184439Sivoras } 2735184439Sivoras sbuf_printf(sb, "</cpu>\n"); 2736184439Sivoras 2737184439Sivoras if (cg->cg_flags != 0) { 2738210117Sivoras sbuf_printf(sb, "%*s <flags>", indent, ""); 2739184439Sivoras if ((cg->cg_flags & CG_FLAG_HTT) != 0) 2740208982Sivoras sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>"); 2741208983Sivoras if ((cg->cg_flags & CG_FLAG_THREAD) != 0) 2742208983Sivoras sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>"); 2743191643Sjeff if ((cg->cg_flags & CG_FLAG_SMT) != 0) 2744208983Sivoras sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>"); 2745210117Sivoras sbuf_printf(sb, "</flags>\n"); 2746184439Sivoras } 2747184439Sivoras 2748184439Sivoras if (cg->cg_children > 0) { 2749184439Sivoras sbuf_printf(sb, "%*s <children>\n", indent, ""); 2750184439Sivoras for (i = 0; i < cg->cg_children; i++) 2751184439Sivoras sysctl_kern_sched_topology_spec_internal(sb, 2752184439Sivoras &cg->cg_child[i], indent+2); 2753184439Sivoras sbuf_printf(sb, "%*s </children>\n", indent, ""); 2754184439Sivoras } 2755184439Sivoras sbuf_printf(sb, "%*s</group>\n", indent, ""); 2756184439Sivoras return (0); 2757184439Sivoras} 2758184439Sivoras 2759184439Sivoras/* 2760184439Sivoras * Sysctl handler for retrieving topology dump. It's a wrapper for 2761184439Sivoras * the recursive sysctl_kern_smp_topology_spec_internal(). 2762184439Sivoras */ 2763184439Sivorasstatic int 2764184439Sivorassysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS) 2765184439Sivoras{ 2766184439Sivoras struct sbuf *topo; 2767184439Sivoras int err; 2768184439Sivoras 2769184439Sivoras KASSERT(cpu_top != NULL, ("cpu_top isn't initialized")); 2770184439Sivoras 2771184570Sivoras topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND); 2772184439Sivoras if (topo == NULL) 2773184439Sivoras return (ENOMEM); 2774184439Sivoras 2775184439Sivoras sbuf_printf(topo, "<groups>\n"); 2776184439Sivoras err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1); 2777184439Sivoras sbuf_printf(topo, "</groups>\n"); 2778184439Sivoras 2779184439Sivoras if (err == 0) { 2780184439Sivoras sbuf_finish(topo); 2781184439Sivoras err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo)); 2782184439Sivoras } 2783184439Sivoras sbuf_delete(topo); 2784184439Sivoras return (err); 2785184439Sivoras} 2786214510Sdavidxu 2787184439Sivoras#endif 2788184439Sivoras 2789241249Smavstatic int 2790241249Smavsysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 2791241249Smav{ 2792241249Smav int error, new_val, period; 2793241249Smav 2794241249Smav period = 1000000 / realstathz; 2795241249Smav new_val = period * sched_slice; 2796241249Smav error = sysctl_handle_int(oidp, &new_val, 0, req); 2797241249Smav if (error != 0 || req->newptr == NULL) 2798241249Smav return (error); 2799241249Smav if (new_val <= 0) 2800241249Smav return (EINVAL); 2801241249Smav sched_slice = imax(1, (new_val + period / 2) / period); 2802241249Smav hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 2803241249Smav realstathz); 2804241249Smav return (0); 2805241249Smav} 2806241249Smav 2807177435SjeffSYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2808171482SjeffSYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2809165762Sjeff "Scheduler name"); 2810241249SmavSYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW, 2811241249Smav NULL, 0, sysctl_kern_quantum, "I", 2812241249Smav "Quantum for timeshare threads in microseconds"); 2813171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2814241249Smav "Quantum for timeshare threads in stathz ticks"); 2815171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2816241249Smav "Interactivity score threshold"); 2817241249SmavSYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, 2818241249Smav &preempt_thresh, 0, 2819241249Smav "Maximal (lowest) priority for preemption"); 2820241249SmavSYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0, 2821241249Smav "Assign static kernel priorities to sleeping threads"); 2822241249SmavSYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0, 2823241249Smav "Number of times idle thread will spin waiting for new work"); 2824241249SmavSYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, 2825241249Smav &sched_idlespinthresh, 0, 2826241249Smav "Threshold before we will permit idle thread spinning"); 2827166108Sjeff#ifdef SMP 2828171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2829171482Sjeff "Number of hz ticks to keep thread affinity for"); 2830171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2831171482Sjeff "Enables the long-term load balancer"); 2832172409SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2833172409Sjeff &balance_interval, 0, 2834241249Smav "Average period in stathz ticks to run the long-term balancer"); 2835171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2836171482Sjeff "Attempts to steal work from other cores before idling"); 2837171506SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2838241249Smav "Minimum load on remote CPU before we'll steal"); 2839184439SivorasSYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING | 2840241249Smav CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A", 2841184439Sivoras "XML dump of detected CPU topology"); 2842166108Sjeff#endif 2843165762Sjeff 2844172264Sjeff/* ps compat. All cpu percentages from ULE are weighted. */ 2845172293Sjeffstatic int ccpu = 0; 2846165762SjeffSYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2847