sched_ule.c revision 242402
1109864Sjeff/*- 2165762Sjeff * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org> 3109864Sjeff * All rights reserved. 4109864Sjeff * 5109864Sjeff * Redistribution and use in source and binary forms, with or without 6109864Sjeff * modification, are permitted provided that the following conditions 7109864Sjeff * are met: 8109864Sjeff * 1. Redistributions of source code must retain the above copyright 9109864Sjeff * notice unmodified, this list of conditions, and the following 10109864Sjeff * disclaimer. 11109864Sjeff * 2. Redistributions in binary form must reproduce the above copyright 12109864Sjeff * notice, this list of conditions and the following disclaimer in the 13109864Sjeff * documentation and/or other materials provided with the distribution. 14109864Sjeff * 15109864Sjeff * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 16109864Sjeff * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 17109864Sjeff * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 18109864Sjeff * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 19109864Sjeff * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 20109864Sjeff * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 21109864Sjeff * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 22109864Sjeff * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 23109864Sjeff * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 24109864Sjeff * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 25109864Sjeff */ 26109864Sjeff 27171482Sjeff/* 28171482Sjeff * This file implements the ULE scheduler. ULE supports independent CPU 29171482Sjeff * run queues and fine grain locking. It has superior interactive 30171482Sjeff * performance under load even on uni-processor systems. 31171482Sjeff * 32171482Sjeff * etymology: 33172293Sjeff * ULE is the last three letters in schedule. It owes its name to a 34171482Sjeff * generic user created for a scheduling system by Paul Mikesell at 35171482Sjeff * Isilon Systems and a general lack of creativity on the part of the author. 36171482Sjeff */ 37171482Sjeff 38116182Sobrien#include <sys/cdefs.h> 39191645Sjeff__FBSDID("$FreeBSD: head/sys/kern/sched_ule.c 242402 2012-10-31 18:07:18Z attilio $"); 40116182Sobrien 41147565Speter#include "opt_hwpmc_hooks.h" 42179297Sjb#include "opt_kdtrace.h" 43147565Speter#include "opt_sched.h" 44134649Sscottl 45109864Sjeff#include <sys/param.h> 46109864Sjeff#include <sys/systm.h> 47131929Smarcel#include <sys/kdb.h> 48109864Sjeff#include <sys/kernel.h> 49109864Sjeff#include <sys/ktr.h> 50109864Sjeff#include <sys/lock.h> 51109864Sjeff#include <sys/mutex.h> 52109864Sjeff#include <sys/proc.h> 53112966Sjeff#include <sys/resource.h> 54122038Sjeff#include <sys/resourcevar.h> 55109864Sjeff#include <sys/sched.h> 56235459Srstone#include <sys/sdt.h> 57109864Sjeff#include <sys/smp.h> 58109864Sjeff#include <sys/sx.h> 59109864Sjeff#include <sys/sysctl.h> 60109864Sjeff#include <sys/sysproto.h> 61139453Sjhb#include <sys/turnstile.h> 62161599Sdavidxu#include <sys/umtx.h> 63109864Sjeff#include <sys/vmmeter.h> 64176735Sjeff#include <sys/cpuset.h> 65184439Sivoras#include <sys/sbuf.h> 66109864Sjeff 67145256Sjkoshy#ifdef HWPMC_HOOKS 68145256Sjkoshy#include <sys/pmckern.h> 69145256Sjkoshy#endif 70145256Sjkoshy 71179297Sjb#ifdef KDTRACE_HOOKS 72179297Sjb#include <sys/dtrace_bsd.h> 73179297Sjbint dtrace_vtime_active; 74179297Sjbdtrace_vtime_switch_func_t dtrace_vtime_switch_func; 75179297Sjb#endif 76179297Sjb 77109864Sjeff#include <machine/cpu.h> 78121790Sjeff#include <machine/smp.h> 79109864Sjeff 80236141Sraj#if defined(__powerpc__) && defined(BOOKE_E500) 81172345Sjeff#error "This architecture is not currently compatible with ULE" 82166190Sjeff#endif 83166190Sjeff 84171482Sjeff#define KTR_ULE 0 85166137Sjeff 86187679Sjeff#define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX))) 87187679Sjeff#define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU))) 88224221Sattilio#define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load")) 89187357Sjeff 90166137Sjeff/* 91171482Sjeff * Thread scheduler specific section. All fields are protected 92171482Sjeff * by the thread lock. 93146954Sjeff */ 94164936Sjulianstruct td_sched { 95171482Sjeff struct runq *ts_runq; /* Run-queue we're queued on. */ 96171482Sjeff short ts_flags; /* TSF_* flags. */ 97164936Sjulian u_char ts_cpu; /* CPU that we have affinity for. */ 98177009Sjeff int ts_rltick; /* Real last tick, for affinity. */ 99171482Sjeff int ts_slice; /* Ticks of slice remaining. */ 100171482Sjeff u_int ts_slptime; /* Number of ticks we vol. slept */ 101171482Sjeff u_int ts_runtime; /* Number of ticks we were running */ 102164936Sjulian int ts_ltick; /* Last tick that we were running on */ 103164936Sjulian int ts_ftick; /* First tick that we were running on */ 104164936Sjulian int ts_ticks; /* Tick count */ 105187357Sjeff#ifdef KTR 106187357Sjeff char ts_name[TS_NAME_LEN]; 107187357Sjeff#endif 108134791Sjulian}; 109164936Sjulian/* flags kept in ts_flags */ 110166108Sjeff#define TSF_BOUND 0x0001 /* Thread can not migrate. */ 111166108Sjeff#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 112121790Sjeff 113164936Sjulianstatic struct td_sched td_sched0; 114109864Sjeff 115176735Sjeff#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 116176735Sjeff#define THREAD_CAN_SCHED(td, cpu) \ 117176735Sjeff CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) 118176735Sjeff 119109864Sjeff/* 120217351Sjhb * Priority ranges used for interactive and non-interactive timeshare 121217410Sjhb * threads. The timeshare priorities are split up into four ranges. 122217410Sjhb * The first range handles interactive threads. The last three ranges 123217410Sjhb * (NHALF, x, and NHALF) handle non-interactive threads with the outer 124217410Sjhb * ranges supporting nice values. 125217351Sjhb */ 126217410Sjhb#define PRI_TIMESHARE_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1) 127217410Sjhb#define PRI_INTERACT_RANGE ((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2) 128228718Savg#define PRI_BATCH_RANGE (PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE) 129217410Sjhb 130217410Sjhb#define PRI_MIN_INTERACT PRI_MIN_TIMESHARE 131217410Sjhb#define PRI_MAX_INTERACT (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1) 132217410Sjhb#define PRI_MIN_BATCH (PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE) 133217351Sjhb#define PRI_MAX_BATCH PRI_MAX_TIMESHARE 134217351Sjhb 135217351Sjhb/* 136165762Sjeff * Cpu percentage computation macros and defines. 137111857Sjeff * 138165762Sjeff * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 139165762Sjeff * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 140165796Sjeff * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 141165762Sjeff * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 142165762Sjeff * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 143165762Sjeff * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 144165762Sjeff */ 145165762Sjeff#define SCHED_TICK_SECS 10 146165762Sjeff#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 147165796Sjeff#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 148165762Sjeff#define SCHED_TICK_SHIFT 10 149165762Sjeff#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 150165830Sjeff#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 151165762Sjeff 152165762Sjeff/* 153165762Sjeff * These macros determine priorities for non-interactive threads. They are 154165762Sjeff * assigned a priority based on their recent cpu utilization as expressed 155165762Sjeff * by the ratio of ticks to the tick total. NHALF priorities at the start 156165762Sjeff * and end of the MIN to MAX timeshare range are only reachable with negative 157165762Sjeff * or positive nice respectively. 158165762Sjeff * 159165762Sjeff * PRI_RANGE: Priority range for utilization dependent priorities. 160116642Sjeff * PRI_NRESV: Number of nice values. 161165762Sjeff * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 162165762Sjeff * PRI_NICE: Determines the part of the priority inherited from nice. 163109864Sjeff */ 164165762Sjeff#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 165121869Sjeff#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 166217351Sjhb#define SCHED_PRI_MIN (PRI_MIN_BATCH + SCHED_PRI_NHALF) 167217351Sjhb#define SCHED_PRI_MAX (PRI_MAX_BATCH - SCHED_PRI_NHALF) 168217237Sjhb#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN + 1) 169165762Sjeff#define SCHED_PRI_TICKS(ts) \ 170165762Sjeff (SCHED_TICK_HZ((ts)) / \ 171165827Sjeff (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 172165762Sjeff#define SCHED_PRI_NICE(nice) (nice) 173109864Sjeff 174109864Sjeff/* 175165762Sjeff * These determine the interactivity of a process. Interactivity differs from 176165762Sjeff * cpu utilization in that it expresses the voluntary time slept vs time ran 177165762Sjeff * while cpu utilization includes all time not running. This more accurately 178165762Sjeff * models the intent of the thread. 179109864Sjeff * 180110645Sjeff * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 181110645Sjeff * before throttling back. 182121868Sjeff * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 183116365Sjeff * INTERACT_MAX: Maximum interactivity value. Smaller is better. 184215102Sattilio * INTERACT_THRESH: Threshold for placement on the current runq. 185109864Sjeff */ 186165762Sjeff#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 187165762Sjeff#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 188116365Sjeff#define SCHED_INTERACT_MAX (100) 189116365Sjeff#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 190121126Sjeff#define SCHED_INTERACT_THRESH (30) 191111857Sjeff 192239157Smav/* Flags kept in td_flags. */ 193239157Smav#define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */ 194239157Smav 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; 204239185Smavstatic int realstathz = 127; 205241844Seadlerstatic int tickincr = 8 << SCHED_TICK_SHIFT; 206239185Smavstatic 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; 218232740Smavstatic 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 { 226242014Sjimharris /* 227242014Sjimharris * Ordered to improve efficiency of cpu_search() and switch(). 228242014Sjimharris * tdq_lock is padded to avoid false sharing with tdq_load and 229242014Sjimharris * tdq_cpu_idle. 230242014Sjimharris */ 231242402Sattilio struct mtx_padalign tdq_lock; /* run queue lock. */ 232176735Sjeff struct cpu_group *tdq_cg; /* Pointer to cpu topology. */ 233178277Sjeff volatile int tdq_load; /* Aggregate load. */ 234212416Smav volatile int tdq_cpu_idle; /* cpu_idle() is active. */ 235176735Sjeff int tdq_sysload; /* For loadavg, !ITHD load. */ 236177009Sjeff int tdq_transferable; /* Transferable thread count. */ 237178277Sjeff short tdq_switchcnt; /* Switches this tick. */ 238178277Sjeff short tdq_oldswitchcnt; /* Switches last tick. */ 239177009Sjeff u_char tdq_lowpri; /* Lowest priority thread. */ 240177009Sjeff u_char tdq_ipipending; /* IPI pending. */ 241166557Sjeff u_char tdq_idx; /* Current insert index. */ 242166557Sjeff u_char tdq_ridx; /* Current removal index. */ 243177009Sjeff struct runq tdq_realtime; /* real-time run queue. */ 244177009Sjeff struct runq tdq_timeshare; /* timeshare run queue. */ 245177009Sjeff struct runq tdq_idle; /* Queue of IDLE threads. */ 246187357Sjeff char tdq_name[TDQ_NAME_LEN]; 247187357Sjeff#ifdef KTR 248187357Sjeff char tdq_loadname[TDQ_LOADNAME_LEN]; 249187357Sjeff#endif 250171482Sjeff} __aligned(64); 251109864Sjeff 252178277Sjeff/* Idle thread states and config. */ 253178277Sjeff#define TDQ_RUNNING 1 254178277Sjeff#define TDQ_IDLE 2 255166108Sjeff 256123433Sjeff#ifdef SMP 257184439Sivorasstruct cpu_group *cpu_top; /* CPU topology */ 258123433Sjeff 259176735Sjeff#define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000)) 260176735Sjeff#define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity)) 261166108Sjeff 262123433Sjeff/* 263166108Sjeff * Run-time tunables. 264166108Sjeff */ 265171506Sjeffstatic int rebalance = 1; 266172409Sjeffstatic int balance_interval = 128; /* Default set in sched_initticks(). */ 267166108Sjeffstatic int affinity; 268171506Sjeffstatic int steal_idle = 1; 269171506Sjeffstatic int steal_thresh = 2; 270166108Sjeff 271166108Sjeff/* 272165620Sjeff * One thread queue per processor. 273109864Sjeff */ 274164936Sjulianstatic struct tdq tdq_cpu[MAXCPU]; 275172409Sjeffstatic struct tdq *balance_tdq; 276172409Sjeffstatic int balance_ticks; 277232207Smavstatic DPCPU_DEFINE(uint32_t, randomval); 278129982Sjeff 279164936Sjulian#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 280164936Sjulian#define TDQ_CPU(x) (&tdq_cpu[(x)]) 281171713Sjeff#define TDQ_ID(x) ((int)((x) - tdq_cpu)) 282123433Sjeff#else /* !SMP */ 283164936Sjulianstatic struct tdq tdq_cpu; 284129982Sjeff 285170315Sjeff#define TDQ_ID(x) (0) 286164936Sjulian#define TDQ_SELF() (&tdq_cpu) 287164936Sjulian#define TDQ_CPU(x) (&tdq_cpu) 288110028Sjeff#endif 289109864Sjeff 290171482Sjeff#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 291171482Sjeff#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 292171482Sjeff#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 293171482Sjeff#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 294242402Sattilio#define TDQ_LOCKPTR(t) ((struct mtx *)(&(t)->tdq_lock)) 295171482Sjeff 296163709Sjbstatic void sched_priority(struct thread *); 297146954Sjeffstatic void sched_thread_priority(struct thread *, u_char); 298163709Sjbstatic int sched_interact_score(struct thread *); 299163709Sjbstatic void sched_interact_update(struct thread *); 300163709Sjbstatic void sched_interact_fork(struct thread *); 301232917Smavstatic void sched_pctcpu_update(struct td_sched *, int); 302109864Sjeff 303110267Sjeff/* Operations on per processor queues */ 304177435Sjeffstatic struct thread *tdq_choose(struct tdq *); 305164936Sjulianstatic void tdq_setup(struct tdq *); 306177435Sjeffstatic void tdq_load_add(struct tdq *, struct thread *); 307177435Sjeffstatic void tdq_load_rem(struct tdq *, struct thread *); 308177435Sjeffstatic __inline void tdq_runq_add(struct tdq *, struct thread *, int); 309177435Sjeffstatic __inline void tdq_runq_rem(struct tdq *, struct thread *); 310177005Sjeffstatic inline int sched_shouldpreempt(int, int, int); 311164936Sjulianvoid tdq_print(int cpu); 312165762Sjeffstatic void runq_print(struct runq *rq); 313171482Sjeffstatic void tdq_add(struct tdq *, struct thread *, int); 314110267Sjeff#ifdef SMP 315176735Sjeffstatic int tdq_move(struct tdq *, struct tdq *); 316171482Sjeffstatic int tdq_idled(struct tdq *); 317177435Sjeffstatic void tdq_notify(struct tdq *, struct thread *); 318177435Sjeffstatic struct thread *tdq_steal(struct tdq *, int); 319177435Sjeffstatic struct thread *runq_steal(struct runq *, int); 320177435Sjeffstatic int sched_pickcpu(struct thread *, int); 321172409Sjeffstatic void sched_balance(void); 322176735Sjeffstatic int sched_balance_pair(struct tdq *, struct tdq *); 323177435Sjeffstatic inline struct tdq *sched_setcpu(struct thread *, int, int); 324171482Sjeffstatic inline void thread_unblock_switch(struct thread *, struct mtx *); 325171713Sjeffstatic struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 326184439Sivorasstatic int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS); 327184439Sivorasstatic int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, 328184439Sivoras struct cpu_group *cg, int indent); 329121790Sjeff#endif 330110028Sjeff 331165762Sjeffstatic void sched_setup(void *dummy); 332177253SrwatsonSYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 333165762Sjeff 334165762Sjeffstatic void sched_initticks(void *dummy); 335177253SrwatsonSYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 336177253Srwatson NULL); 337165762Sjeff 338235459SrstoneSDT_PROVIDER_DEFINE(sched); 339235459Srstone 340235459SrstoneSDT_PROBE_DEFINE3(sched, , , change_pri, change-pri, "struct thread *", 341235459Srstone "struct proc *", "uint8_t"); 342235459SrstoneSDT_PROBE_DEFINE3(sched, , , dequeue, dequeue, "struct thread *", 343235459Srstone "struct proc *", "void *"); 344235459SrstoneSDT_PROBE_DEFINE4(sched, , , enqueue, enqueue, "struct thread *", 345235459Srstone "struct proc *", "void *", "int"); 346235459SrstoneSDT_PROBE_DEFINE4(sched, , , lend_pri, lend-pri, "struct thread *", 347235459Srstone "struct proc *", "uint8_t", "struct thread *"); 348235459SrstoneSDT_PROBE_DEFINE2(sched, , , load_change, load-change, "int", "int"); 349235459SrstoneSDT_PROBE_DEFINE2(sched, , , off_cpu, off-cpu, "struct thread *", 350235459Srstone "struct proc *"); 351235459SrstoneSDT_PROBE_DEFINE(sched, , , on_cpu, on-cpu); 352235459SrstoneSDT_PROBE_DEFINE(sched, , , remain_cpu, remain-cpu); 353235459SrstoneSDT_PROBE_DEFINE2(sched, , , surrender, surrender, "struct thread *", 354235459Srstone "struct proc *"); 355235459Srstone 356171482Sjeff/* 357171482Sjeff * Print the threads waiting on a run-queue. 358171482Sjeff */ 359165762Sjeffstatic void 360165762Sjeffrunq_print(struct runq *rq) 361165762Sjeff{ 362165762Sjeff struct rqhead *rqh; 363177435Sjeff struct thread *td; 364165762Sjeff int pri; 365165762Sjeff int j; 366165762Sjeff int i; 367165762Sjeff 368165762Sjeff for (i = 0; i < RQB_LEN; i++) { 369165762Sjeff printf("\t\trunq bits %d 0x%zx\n", 370165762Sjeff i, rq->rq_status.rqb_bits[i]); 371165762Sjeff for (j = 0; j < RQB_BPW; j++) 372165762Sjeff if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 373165762Sjeff pri = j + (i << RQB_L2BPW); 374165762Sjeff rqh = &rq->rq_queues[pri]; 375177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) { 376165762Sjeff printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 377177435Sjeff td, td->td_name, td->td_priority, 378177435Sjeff td->td_rqindex, pri); 379165762Sjeff } 380165762Sjeff } 381165762Sjeff } 382165762Sjeff} 383165762Sjeff 384171482Sjeff/* 385171482Sjeff * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 386171482Sjeff */ 387113357Sjeffvoid 388164936Sjuliantdq_print(int cpu) 389110267Sjeff{ 390164936Sjulian struct tdq *tdq; 391112994Sjeff 392164936Sjulian tdq = TDQ_CPU(cpu); 393112994Sjeff 394171713Sjeff printf("tdq %d:\n", TDQ_ID(tdq)); 395176735Sjeff printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 396176735Sjeff printf("\tLock name: %s\n", tdq->tdq_name); 397165620Sjeff printf("\tload: %d\n", tdq->tdq_load); 398178277Sjeff printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt); 399178277Sjeff printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt); 400171482Sjeff printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 401165766Sjeff printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 402178277Sjeff printf("\tload transferable: %d\n", tdq->tdq_transferable); 403178277Sjeff printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 404165762Sjeff printf("\trealtime runq:\n"); 405165762Sjeff runq_print(&tdq->tdq_realtime); 406165762Sjeff printf("\ttimeshare runq:\n"); 407165762Sjeff runq_print(&tdq->tdq_timeshare); 408165762Sjeff printf("\tidle runq:\n"); 409165762Sjeff runq_print(&tdq->tdq_idle); 410113357Sjeff} 411112994Sjeff 412177005Sjeffstatic inline int 413177005Sjeffsched_shouldpreempt(int pri, int cpri, int remote) 414177005Sjeff{ 415177005Sjeff /* 416177005Sjeff * If the new priority is not better than the current priority there is 417177005Sjeff * nothing to do. 418177005Sjeff */ 419177005Sjeff if (pri >= cpri) 420177005Sjeff return (0); 421177005Sjeff /* 422177005Sjeff * Always preempt idle. 423177005Sjeff */ 424177005Sjeff if (cpri >= PRI_MIN_IDLE) 425177005Sjeff return (1); 426177005Sjeff /* 427177005Sjeff * If preemption is disabled don't preempt others. 428177005Sjeff */ 429177005Sjeff if (preempt_thresh == 0) 430177005Sjeff return (0); 431177005Sjeff /* 432177005Sjeff * Preempt if we exceed the threshold. 433177005Sjeff */ 434177005Sjeff if (pri <= preempt_thresh) 435177005Sjeff return (1); 436177005Sjeff /* 437217351Sjhb * If we're interactive or better and there is non-interactive 438217351Sjhb * or worse running preempt only remote processors. 439177005Sjeff */ 440217351Sjhb if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT) 441177005Sjeff return (1); 442177005Sjeff return (0); 443177005Sjeff} 444177005Sjeff 445171482Sjeff/* 446171482Sjeff * Add a thread to the actual run-queue. Keeps transferable counts up to 447171482Sjeff * date with what is actually on the run-queue. Selects the correct 448171482Sjeff * queue position for timeshare threads. 449171482Sjeff */ 450122744Sjeffstatic __inline void 451177435Sjefftdq_runq_add(struct tdq *tdq, struct thread *td, int flags) 452122744Sjeff{ 453177435Sjeff struct td_sched *ts; 454177042Sjeff u_char pri; 455177042Sjeff 456171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 457177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 458177009Sjeff 459177435Sjeff pri = td->td_priority; 460177435Sjeff ts = td->td_sched; 461177435Sjeff TD_SET_RUNQ(td); 462177435Sjeff if (THREAD_CAN_MIGRATE(td)) { 463165620Sjeff tdq->tdq_transferable++; 464164936Sjulian ts->ts_flags |= TSF_XFERABLE; 465123433Sjeff } 466217351Sjhb if (pri < PRI_MIN_BATCH) { 467177042Sjeff ts->ts_runq = &tdq->tdq_realtime; 468217351Sjhb } else if (pri <= PRI_MAX_BATCH) { 469177042Sjeff ts->ts_runq = &tdq->tdq_timeshare; 470217351Sjhb KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH, 471165762Sjeff ("Invalid priority %d on timeshare runq", pri)); 472165762Sjeff /* 473165762Sjeff * This queue contains only priorities between MIN and MAX 474165762Sjeff * realtime. Use the whole queue to represent these values. 475165762Sjeff */ 476171713Sjeff if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 477228718Savg pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE; 478165762Sjeff pri = (pri + tdq->tdq_idx) % RQ_NQS; 479165766Sjeff /* 480165766Sjeff * This effectively shortens the queue by one so we 481165766Sjeff * can have a one slot difference between idx and 482165766Sjeff * ridx while we wait for threads to drain. 483165766Sjeff */ 484165766Sjeff if (tdq->tdq_ridx != tdq->tdq_idx && 485165766Sjeff pri == tdq->tdq_ridx) 486167664Sjeff pri = (unsigned char)(pri - 1) % RQ_NQS; 487165762Sjeff } else 488165766Sjeff pri = tdq->tdq_ridx; 489177435Sjeff runq_add_pri(ts->ts_runq, td, pri, flags); 490177042Sjeff return; 491165762Sjeff } else 492177009Sjeff ts->ts_runq = &tdq->tdq_idle; 493177435Sjeff runq_add(ts->ts_runq, td, flags); 494177009Sjeff} 495177009Sjeff 496171482Sjeff/* 497171482Sjeff * Remove a thread from a run-queue. This typically happens when a thread 498171482Sjeff * is selected to run. Running threads are not on the queue and the 499171482Sjeff * transferable count does not reflect them. 500171482Sjeff */ 501122744Sjeffstatic __inline void 502177435Sjefftdq_runq_rem(struct tdq *tdq, struct thread *td) 503122744Sjeff{ 504177435Sjeff struct td_sched *ts; 505177435Sjeff 506177435Sjeff ts = td->td_sched; 507171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 508171482Sjeff KASSERT(ts->ts_runq != NULL, 509177435Sjeff ("tdq_runq_remove: thread %p null ts_runq", td)); 510164936Sjulian if (ts->ts_flags & TSF_XFERABLE) { 511165620Sjeff tdq->tdq_transferable--; 512164936Sjulian ts->ts_flags &= ~TSF_XFERABLE; 513123433Sjeff } 514165766Sjeff if (ts->ts_runq == &tdq->tdq_timeshare) { 515165766Sjeff if (tdq->tdq_idx != tdq->tdq_ridx) 516177435Sjeff runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx); 517165766Sjeff else 518177435Sjeff runq_remove_idx(ts->ts_runq, td, NULL); 519165766Sjeff } else 520177435Sjeff runq_remove(ts->ts_runq, td); 521122744Sjeff} 522122744Sjeff 523171482Sjeff/* 524171482Sjeff * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 525171482Sjeff * for this thread to the referenced thread queue. 526171482Sjeff */ 527113357Sjeffstatic void 528177435Sjefftdq_load_add(struct tdq *tdq, struct thread *td) 529113357Sjeff{ 530171482Sjeff 531171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 532177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 533177902Sjeff 534165620Sjeff tdq->tdq_load++; 535198854Sattilio if ((td->td_flags & TDF_NOLOAD) == 0) 536177902Sjeff tdq->tdq_sysload++; 537187357Sjeff KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 538235459Srstone SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load); 539110267Sjeff} 540113357Sjeff 541171482Sjeff/* 542171482Sjeff * Remove the load from a thread that is transitioning to a sleep state or 543171482Sjeff * exiting. 544171482Sjeff */ 545112994Sjeffstatic void 546177435Sjefftdq_load_rem(struct tdq *tdq, struct thread *td) 547110267Sjeff{ 548171482Sjeff 549177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 550171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 551171482Sjeff KASSERT(tdq->tdq_load != 0, 552171713Sjeff ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 553177902Sjeff 554165620Sjeff tdq->tdq_load--; 555198854Sattilio if ((td->td_flags & TDF_NOLOAD) == 0) 556177902Sjeff tdq->tdq_sysload--; 557187357Sjeff KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 558235459Srstone SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load); 559110267Sjeff} 560110267Sjeff 561176735Sjeff/* 562176735Sjeff * Set lowpri to its exact value by searching the run-queue and 563176735Sjeff * evaluating curthread. curthread may be passed as an optimization. 564176735Sjeff */ 565176735Sjeffstatic void 566176735Sjefftdq_setlowpri(struct tdq *tdq, struct thread *ctd) 567176735Sjeff{ 568176735Sjeff struct thread *td; 569176735Sjeff 570176735Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 571176735Sjeff if (ctd == NULL) 572176735Sjeff ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 573177435Sjeff td = tdq_choose(tdq); 574177435Sjeff if (td == NULL || td->td_priority > ctd->td_priority) 575176735Sjeff tdq->tdq_lowpri = ctd->td_priority; 576176735Sjeff else 577176735Sjeff tdq->tdq_lowpri = td->td_priority; 578176735Sjeff} 579176735Sjeff 580113357Sjeff#ifdef SMP 581176735Sjeffstruct cpu_search { 582194779Sjeff cpuset_t cs_mask; 583232207Smav u_int cs_prefer; 584232207Smav int cs_pri; /* Min priority for low. */ 585232207Smav int cs_limit; /* Max load for low, min load for high. */ 586232207Smav int cs_cpu; 587232207Smav int cs_load; 588176735Sjeff}; 589176735Sjeff 590176735Sjeff#define CPU_SEARCH_LOWEST 0x1 591176735Sjeff#define CPU_SEARCH_HIGHEST 0x2 592176735Sjeff#define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 593176735Sjeff 594194779Sjeff#define CPUSET_FOREACH(cpu, mask) \ 595194779Sjeff for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \ 596222813Sattilio if (CPU_ISSET(cpu, &mask)) 597176735Sjeff 598232207Smavstatic __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low, 599176735Sjeff struct cpu_search *high, const int match); 600232207Smavint cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low); 601232207Smavint cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high); 602232207Smavint cpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 603176735Sjeff struct cpu_search *high); 604176735Sjeff 605116069Sjeff/* 606176735Sjeff * Search the tree of cpu_groups for the lowest or highest loaded cpu 607176735Sjeff * according to the match argument. This routine actually compares the 608176735Sjeff * load on all paths through the tree and finds the least loaded cpu on 609176735Sjeff * the least loaded path, which may differ from the least loaded cpu in 610176735Sjeff * the system. This balances work among caches and busses. 611116069Sjeff * 612176735Sjeff * This inline is instantiated in three forms below using constants for the 613176735Sjeff * match argument. It is reduced to the minimum set for each case. It is 614176735Sjeff * also recursive to the depth of the tree. 615116069Sjeff */ 616177169Sjhbstatic __inline int 617232207Smavcpu_search(const struct cpu_group *cg, struct cpu_search *low, 618176735Sjeff struct cpu_search *high, const int match) 619176735Sjeff{ 620232207Smav struct cpu_search lgroup; 621232207Smav struct cpu_search hgroup; 622232207Smav cpuset_t cpumask; 623232207Smav struct cpu_group *child; 624232207Smav struct tdq *tdq; 625234066Smav int cpu, i, hload, lload, load, total, rnd, *rndptr; 626176735Sjeff 627176735Sjeff total = 0; 628232207Smav cpumask = cg->cg_mask; 629232207Smav if (match & CPU_SEARCH_LOWEST) { 630232207Smav lload = INT_MAX; 631232207Smav lgroup = *low; 632232207Smav } 633232207Smav if (match & CPU_SEARCH_HIGHEST) { 634234066Smav hload = INT_MIN; 635232207Smav hgroup = *high; 636232207Smav } 637176735Sjeff 638232207Smav /* Iterate through the child CPU groups and then remaining CPUs. */ 639234066Smav for (i = cg->cg_children, cpu = mp_maxid; i >= 0; ) { 640234066Smav if (i == 0) { 641234066Smav while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask)) 642234066Smav cpu--; 643234066Smav if (cpu < 0) 644232207Smav break; 645232207Smav child = NULL; 646232207Smav } else 647234066Smav child = &cg->cg_child[i - 1]; 648232207Smav 649234066Smav if (match & CPU_SEARCH_LOWEST) 650234066Smav lgroup.cs_cpu = -1; 651234066Smav if (match & CPU_SEARCH_HIGHEST) 652234066Smav hgroup.cs_cpu = -1; 653232207Smav if (child) { /* Handle child CPU group. */ 654232207Smav 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 } 666232207Smav } else { /* Handle child CPU. */ 667232207Smav tdq = TDQ_CPU(cpu); 668232207Smav load = tdq->tdq_load * 256; 669234066Smav rndptr = DPCPU_PTR(randomval); 670234066Smav rnd = (*rndptr = *rndptr * 69069 + 5) >> 26; 671232207Smav if (match & CPU_SEARCH_LOWEST) { 672232207Smav if (cpu == low->cs_prefer) 673232207Smav load -= 64; 674232207Smav /* If that CPU is allowed and get data. */ 675234066Smav if (tdq->tdq_lowpri > lgroup.cs_pri && 676234066Smav tdq->tdq_load <= lgroup.cs_limit && 677234066Smav CPU_ISSET(cpu, &lgroup.cs_mask)) { 678232207Smav lgroup.cs_cpu = cpu; 679232207Smav lgroup.cs_load = load - rnd; 680176735Sjeff } 681232207Smav } 682232207Smav if (match & CPU_SEARCH_HIGHEST) 683234066Smav if (tdq->tdq_load >= hgroup.cs_limit && 684234066Smav tdq->tdq_transferable && 685234066Smav CPU_ISSET(cpu, &hgroup.cs_mask)) { 686232207Smav hgroup.cs_cpu = cpu; 687232207Smav hgroup.cs_load = load - rnd; 688176735Sjeff } 689176735Sjeff } 690232207Smav total += load; 691176735Sjeff 692232207Smav /* We have info about child item. Compare it. */ 693232207Smav if (match & CPU_SEARCH_LOWEST) { 694234066Smav if (lgroup.cs_cpu >= 0 && 695232454Smav (load < lload || 696232454Smav (load == lload && lgroup.cs_load < low->cs_load))) { 697232207Smav lload = load; 698232207Smav low->cs_cpu = lgroup.cs_cpu; 699232207Smav low->cs_load = lgroup.cs_load; 700232207Smav } 701232207Smav } 702232207Smav if (match & CPU_SEARCH_HIGHEST) 703234066Smav if (hgroup.cs_cpu >= 0 && 704232454Smav (load > hload || 705232454Smav (load == hload && hgroup.cs_load > high->cs_load))) { 706232207Smav hload = load; 707232207Smav high->cs_cpu = hgroup.cs_cpu; 708232207Smav high->cs_load = hgroup.cs_load; 709232207Smav } 710234066Smav if (child) { 711234066Smav i--; 712234066Smav if (i == 0 && CPU_EMPTY(&cpumask)) 713234066Smav break; 714234066Smav } else 715234066Smav cpu--; 716176735Sjeff } 717176735Sjeff return (total); 718176735Sjeff} 719176735Sjeff 720176735Sjeff/* 721176735Sjeff * cpu_search instantiations must pass constants to maintain the inline 722176735Sjeff * optimization. 723176735Sjeff */ 724176735Sjeffint 725232207Smavcpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low) 726176735Sjeff{ 727176735Sjeff return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 728176735Sjeff} 729176735Sjeff 730176735Sjeffint 731232207Smavcpu_search_highest(const struct cpu_group *cg, struct cpu_search *high) 732176735Sjeff{ 733176735Sjeff return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 734176735Sjeff} 735176735Sjeff 736176735Sjeffint 737232207Smavcpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 738176735Sjeff struct cpu_search *high) 739176735Sjeff{ 740176735Sjeff return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 741176735Sjeff} 742176735Sjeff 743176735Sjeff/* 744176735Sjeff * Find the cpu with the least load via the least loaded path that has a 745176735Sjeff * lowpri greater than pri pri. A pri of -1 indicates any priority is 746176735Sjeff * acceptable. 747176735Sjeff */ 748176735Sjeffstatic inline int 749232207Smavsched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload, 750232207Smav int prefer) 751176735Sjeff{ 752176735Sjeff struct cpu_search low; 753176735Sjeff 754176735Sjeff low.cs_cpu = -1; 755232207Smav low.cs_prefer = prefer; 756176735Sjeff low.cs_mask = mask; 757232207Smav low.cs_pri = pri; 758232207Smav low.cs_limit = maxload; 759176735Sjeff cpu_search_lowest(cg, &low); 760176735Sjeff return low.cs_cpu; 761176735Sjeff} 762176735Sjeff 763176735Sjeff/* 764176735Sjeff * Find the cpu with the highest load via the highest loaded path. 765176735Sjeff */ 766176735Sjeffstatic inline int 767232207Smavsched_highest(const struct cpu_group *cg, cpuset_t mask, int minload) 768176735Sjeff{ 769176735Sjeff struct cpu_search high; 770176735Sjeff 771176735Sjeff high.cs_cpu = -1; 772176735Sjeff high.cs_mask = mask; 773176735Sjeff high.cs_limit = minload; 774176735Sjeff cpu_search_highest(cg, &high); 775176735Sjeff return high.cs_cpu; 776176735Sjeff} 777176735Sjeff 778176735Sjeff/* 779176735Sjeff * Simultaneously find the highest and lowest loaded cpu reachable via 780176735Sjeff * cg. 781176735Sjeff */ 782232207Smavstatic inline void 783232207Smavsched_both(const struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu) 784176735Sjeff{ 785176735Sjeff struct cpu_search high; 786176735Sjeff struct cpu_search low; 787176735Sjeff 788176735Sjeff low.cs_cpu = -1; 789232207Smav low.cs_prefer = -1; 790232207Smav low.cs_pri = -1; 791232207Smav low.cs_limit = INT_MAX; 792176735Sjeff low.cs_mask = mask; 793176735Sjeff high.cs_cpu = -1; 794176735Sjeff high.cs_limit = -1; 795176735Sjeff high.cs_mask = mask; 796176735Sjeff cpu_search_both(cg, &low, &high); 797176735Sjeff *lowcpu = low.cs_cpu; 798176735Sjeff *highcpu = high.cs_cpu; 799176735Sjeff return; 800176735Sjeff} 801176735Sjeff 802121790Sjeffstatic void 803176735Sjeffsched_balance_group(struct cpu_group *cg) 804116069Sjeff{ 805232207Smav cpuset_t hmask, lmask; 806232207Smav int high, low, anylow; 807123487Sjeff 808232207Smav CPU_FILL(&hmask); 809176735Sjeff for (;;) { 810232207Smav high = sched_highest(cg, hmask, 1); 811232207Smav /* Stop if there is no more CPU with transferrable threads. */ 812232207Smav if (high == -1) 813176735Sjeff break; 814232207Smav CPU_CLR(high, &hmask); 815232207Smav CPU_COPY(&hmask, &lmask); 816232207Smav /* Stop if there is no more CPU left for low. */ 817232207Smav if (CPU_EMPTY(&lmask)) 818176735Sjeff break; 819232207Smav anylow = 1; 820232207Smavnextlow: 821232207Smav low = sched_lowest(cg, lmask, -1, 822232207Smav TDQ_CPU(high)->tdq_load - 1, high); 823232207Smav /* Stop if we looked well and found no less loaded CPU. */ 824232207Smav if (anylow && low == -1) 825232207Smav break; 826232207Smav /* Go to next high if we found no less loaded CPU. */ 827232207Smav if (low == -1) 828232207Smav continue; 829232207Smav /* Transfer thread from high to low. */ 830232207Smav if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) { 831232207Smav /* CPU that got thread can no longer be a donor. */ 832232207Smav CPU_CLR(low, &hmask); 833232207Smav } else { 834232207Smav /* 835232207Smav * If failed, then there is no threads on high 836232207Smav * that can run on this low. Drop low from low 837232207Smav * mask and look for different one. 838232207Smav */ 839232207Smav CPU_CLR(low, &lmask); 840232207Smav anylow = 0; 841232207Smav goto nextlow; 842232207Smav } 843123487Sjeff } 844123487Sjeff} 845123487Sjeff 846123487Sjeffstatic void 847201148Sedsched_balance(void) 848123487Sjeff{ 849172409Sjeff struct tdq *tdq; 850123487Sjeff 851172409Sjeff /* 852172409Sjeff * Select a random time between .5 * balance_interval and 853172409Sjeff * 1.5 * balance_interval. 854172409Sjeff */ 855176735Sjeff balance_ticks = max(balance_interval / 2, 1); 856176735Sjeff balance_ticks += random() % balance_interval; 857171482Sjeff if (smp_started == 0 || rebalance == 0) 858171482Sjeff return; 859172409Sjeff tdq = TDQ_SELF(); 860172409Sjeff TDQ_UNLOCK(tdq); 861176735Sjeff sched_balance_group(cpu_top); 862172409Sjeff TDQ_LOCK(tdq); 863123487Sjeff} 864123487Sjeff 865171482Sjeff/* 866171482Sjeff * Lock two thread queues using their address to maintain lock order. 867171482Sjeff */ 868123487Sjeffstatic void 869171482Sjefftdq_lock_pair(struct tdq *one, struct tdq *two) 870171482Sjeff{ 871171482Sjeff if (one < two) { 872171482Sjeff TDQ_LOCK(one); 873171482Sjeff TDQ_LOCK_FLAGS(two, MTX_DUPOK); 874171482Sjeff } else { 875171482Sjeff TDQ_LOCK(two); 876171482Sjeff TDQ_LOCK_FLAGS(one, MTX_DUPOK); 877171482Sjeff } 878171482Sjeff} 879171482Sjeff 880171482Sjeff/* 881172409Sjeff * Unlock two thread queues. Order is not important here. 882172409Sjeff */ 883172409Sjeffstatic void 884172409Sjefftdq_unlock_pair(struct tdq *one, struct tdq *two) 885172409Sjeff{ 886172409Sjeff TDQ_UNLOCK(one); 887172409Sjeff TDQ_UNLOCK(two); 888172409Sjeff} 889172409Sjeff 890172409Sjeff/* 891171482Sjeff * Transfer load between two imbalanced thread queues. 892171482Sjeff */ 893176735Sjeffstatic int 894164936Sjuliansched_balance_pair(struct tdq *high, struct tdq *low) 895123487Sjeff{ 896176735Sjeff int moved; 897226057Smarius int cpu; 898116069Sjeff 899171482Sjeff tdq_lock_pair(high, low); 900176735Sjeff moved = 0; 901116069Sjeff /* 902122744Sjeff * Determine what the imbalance is and then adjust that to how many 903165620Sjeff * threads we actually have to give up (transferable). 904122744Sjeff */ 905232207Smav if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load && 906232207Smav (moved = tdq_move(high, low)) > 0) { 907172293Sjeff /* 908226057Smarius * In case the target isn't the current cpu IPI it to force a 909226057Smarius * reschedule with the new workload. 910172293Sjeff */ 911226057Smarius cpu = TDQ_ID(low); 912226057Smarius if (cpu != PCPU_GET(cpuid)) 913226057Smarius ipi_cpu(cpu, IPI_PREEMPT); 914171482Sjeff } 915172409Sjeff tdq_unlock_pair(high, low); 916176735Sjeff return (moved); 917116069Sjeff} 918116069Sjeff 919171482Sjeff/* 920171482Sjeff * Move a thread from one thread queue to another. 921171482Sjeff */ 922176735Sjeffstatic int 923171482Sjefftdq_move(struct tdq *from, struct tdq *to) 924116069Sjeff{ 925171482Sjeff struct td_sched *ts; 926171482Sjeff struct thread *td; 927164936Sjulian struct tdq *tdq; 928171482Sjeff int cpu; 929116069Sjeff 930172409Sjeff TDQ_LOCK_ASSERT(from, MA_OWNED); 931172409Sjeff TDQ_LOCK_ASSERT(to, MA_OWNED); 932172409Sjeff 933164936Sjulian tdq = from; 934171482Sjeff cpu = TDQ_ID(to); 935177435Sjeff td = tdq_steal(tdq, cpu); 936177435Sjeff if (td == NULL) 937176735Sjeff return (0); 938177435Sjeff ts = td->td_sched; 939171482Sjeff /* 940171482Sjeff * Although the run queue is locked the thread may be blocked. Lock 941172409Sjeff * it to clear this and acquire the run-queue lock. 942171482Sjeff */ 943171482Sjeff thread_lock(td); 944172409Sjeff /* Drop recursive lock on from acquired via thread_lock(). */ 945171482Sjeff TDQ_UNLOCK(from); 946171482Sjeff sched_rem(td); 947166108Sjeff ts->ts_cpu = cpu; 948171482Sjeff td->td_lock = TDQ_LOCKPTR(to); 949171482Sjeff tdq_add(to, td, SRQ_YIELDING); 950176735Sjeff return (1); 951116069Sjeff} 952110267Sjeff 953171482Sjeff/* 954171482Sjeff * This tdq has idled. Try to steal a thread from another cpu and switch 955171482Sjeff * to it. 956171482Sjeff */ 957123433Sjeffstatic int 958164936Sjuliantdq_idled(struct tdq *tdq) 959121790Sjeff{ 960176735Sjeff struct cpu_group *cg; 961164936Sjulian struct tdq *steal; 962194779Sjeff cpuset_t mask; 963176735Sjeff int thresh; 964171482Sjeff int cpu; 965123433Sjeff 966172484Sjeff if (smp_started == 0 || steal_idle == 0) 967172484Sjeff return (1); 968194779Sjeff CPU_FILL(&mask); 969194779Sjeff CPU_CLR(PCPU_GET(cpuid), &mask); 970176735Sjeff /* We don't want to be preempted while we're iterating. */ 971171482Sjeff spinlock_enter(); 972176735Sjeff for (cg = tdq->tdq_cg; cg != NULL; ) { 973191643Sjeff if ((cg->cg_flags & CG_FLAG_THREAD) == 0) 974176735Sjeff thresh = steal_thresh; 975176735Sjeff else 976176735Sjeff thresh = 1; 977176735Sjeff cpu = sched_highest(cg, mask, thresh); 978176735Sjeff if (cpu == -1) { 979176735Sjeff cg = cg->cg_parent; 980176735Sjeff continue; 981166108Sjeff } 982176735Sjeff steal = TDQ_CPU(cpu); 983194779Sjeff CPU_CLR(cpu, &mask); 984176735Sjeff tdq_lock_pair(tdq, steal); 985176735Sjeff if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 986176735Sjeff tdq_unlock_pair(tdq, steal); 987176735Sjeff continue; 988171482Sjeff } 989176735Sjeff /* 990176735Sjeff * If a thread was added while interrupts were disabled don't 991176735Sjeff * steal one here. If we fail to acquire one due to affinity 992176735Sjeff * restrictions loop again with this cpu removed from the 993176735Sjeff * set. 994176735Sjeff */ 995176735Sjeff if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 996176735Sjeff tdq_unlock_pair(tdq, steal); 997176735Sjeff continue; 998176735Sjeff } 999176735Sjeff spinlock_exit(); 1000176735Sjeff TDQ_UNLOCK(steal); 1001178272Sjeff mi_switch(SW_VOL | SWT_IDLE, NULL); 1002176735Sjeff thread_unlock(curthread); 1003176735Sjeff 1004176735Sjeff return (0); 1005123433Sjeff } 1006171482Sjeff spinlock_exit(); 1007123433Sjeff return (1); 1008121790Sjeff} 1009121790Sjeff 1010171482Sjeff/* 1011171482Sjeff * Notify a remote cpu of new work. Sends an IPI if criteria are met. 1012171482Sjeff */ 1013121790Sjeffstatic void 1014177435Sjefftdq_notify(struct tdq *tdq, struct thread *td) 1015121790Sjeff{ 1016185047Sjhb struct thread *ctd; 1017166247Sjeff int pri; 1018166108Sjeff int cpu; 1019121790Sjeff 1020177005Sjeff if (tdq->tdq_ipipending) 1021177005Sjeff return; 1022177435Sjeff cpu = td->td_sched->ts_cpu; 1023177435Sjeff pri = td->td_priority; 1024185047Sjhb ctd = pcpu_find(cpu)->pc_curthread; 1025185047Sjhb if (!sched_shouldpreempt(pri, ctd->td_priority, 1)) 1026166137Sjeff return; 1027185047Sjhb if (TD_IS_IDLETHREAD(ctd)) { 1028178277Sjeff /* 1029178471Sjeff * If the MD code has an idle wakeup routine try that before 1030178471Sjeff * falling back to IPI. 1031178471Sjeff */ 1032212416Smav if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu)) 1033178471Sjeff return; 1034178277Sjeff } 1035177005Sjeff tdq->tdq_ipipending = 1; 1036210939Sjhb ipi_cpu(cpu, IPI_PREEMPT); 1037121790Sjeff} 1038121790Sjeff 1039171482Sjeff/* 1040171482Sjeff * Steals load from a timeshare queue. Honors the rotating queue head 1041171482Sjeff * index. 1042171482Sjeff */ 1043177435Sjeffstatic struct thread * 1044176735Sjeffrunq_steal_from(struct runq *rq, int cpu, u_char start) 1045171482Sjeff{ 1046171482Sjeff struct rqbits *rqb; 1047171482Sjeff struct rqhead *rqh; 1048232207Smav struct thread *td, *first; 1049171482Sjeff int bit; 1050171482Sjeff int pri; 1051171482Sjeff int i; 1052171482Sjeff 1053171482Sjeff rqb = &rq->rq_status; 1054171482Sjeff bit = start & (RQB_BPW -1); 1055171482Sjeff pri = 0; 1056232207Smav first = NULL; 1057171482Sjeffagain: 1058171482Sjeff for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 1059171482Sjeff if (rqb->rqb_bits[i] == 0) 1060171482Sjeff continue; 1061171482Sjeff if (bit != 0) { 1062171482Sjeff for (pri = bit; pri < RQB_BPW; pri++) 1063171482Sjeff if (rqb->rqb_bits[i] & (1ul << pri)) 1064171482Sjeff break; 1065171482Sjeff if (pri >= RQB_BPW) 1066171482Sjeff continue; 1067171482Sjeff } else 1068171482Sjeff pri = RQB_FFS(rqb->rqb_bits[i]); 1069171482Sjeff pri += (i << RQB_L2BPW); 1070171482Sjeff rqh = &rq->rq_queues[pri]; 1071177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) { 1072177435Sjeff if (first && THREAD_CAN_MIGRATE(td) && 1073177435Sjeff THREAD_CAN_SCHED(td, cpu)) 1074177435Sjeff return (td); 1075232207Smav first = td; 1076171482Sjeff } 1077171482Sjeff } 1078171482Sjeff if (start != 0) { 1079171482Sjeff start = 0; 1080171482Sjeff goto again; 1081171482Sjeff } 1082171482Sjeff 1083232207Smav if (first && THREAD_CAN_MIGRATE(first) && 1084232207Smav THREAD_CAN_SCHED(first, cpu)) 1085232207Smav return (first); 1086171482Sjeff return (NULL); 1087171482Sjeff} 1088171482Sjeff 1089171482Sjeff/* 1090171482Sjeff * Steals load from a standard linear queue. 1091171482Sjeff */ 1092177435Sjeffstatic struct thread * 1093176735Sjeffrunq_steal(struct runq *rq, int cpu) 1094121790Sjeff{ 1095121790Sjeff struct rqhead *rqh; 1096121790Sjeff struct rqbits *rqb; 1097177435Sjeff struct thread *td; 1098121790Sjeff int word; 1099121790Sjeff int bit; 1100121790Sjeff 1101121790Sjeff rqb = &rq->rq_status; 1102121790Sjeff for (word = 0; word < RQB_LEN; word++) { 1103121790Sjeff if (rqb->rqb_bits[word] == 0) 1104121790Sjeff continue; 1105121790Sjeff for (bit = 0; bit < RQB_BPW; bit++) { 1106123231Speter if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1107121790Sjeff continue; 1108121790Sjeff rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1109177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) 1110177435Sjeff if (THREAD_CAN_MIGRATE(td) && 1111177435Sjeff THREAD_CAN_SCHED(td, cpu)) 1112177435Sjeff return (td); 1113121790Sjeff } 1114121790Sjeff } 1115121790Sjeff return (NULL); 1116121790Sjeff} 1117121790Sjeff 1118171482Sjeff/* 1119171482Sjeff * Attempt to steal a thread in priority order from a thread queue. 1120171482Sjeff */ 1121177435Sjeffstatic struct thread * 1122176735Sjefftdq_steal(struct tdq *tdq, int cpu) 1123121790Sjeff{ 1124177435Sjeff struct thread *td; 1125121790Sjeff 1126171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1127177435Sjeff if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1128177435Sjeff return (td); 1129177435Sjeff if ((td = runq_steal_from(&tdq->tdq_timeshare, 1130177435Sjeff cpu, tdq->tdq_ridx)) != NULL) 1131177435Sjeff return (td); 1132176735Sjeff return (runq_steal(&tdq->tdq_idle, cpu)); 1133121790Sjeff} 1134123433Sjeff 1135171482Sjeff/* 1136171482Sjeff * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1137172409Sjeff * current lock and returns with the assigned queue locked. 1138171482Sjeff */ 1139171482Sjeffstatic inline struct tdq * 1140177435Sjeffsched_setcpu(struct thread *td, int cpu, int flags) 1141123433Sjeff{ 1142177435Sjeff 1143171482Sjeff struct tdq *tdq; 1144123433Sjeff 1145177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1146171482Sjeff tdq = TDQ_CPU(cpu); 1147177435Sjeff td->td_sched->ts_cpu = cpu; 1148177435Sjeff /* 1149177435Sjeff * If the lock matches just return the queue. 1150177435Sjeff */ 1151171482Sjeff if (td->td_lock == TDQ_LOCKPTR(tdq)) 1152171482Sjeff return (tdq); 1153171482Sjeff#ifdef notyet 1154123433Sjeff /* 1155172293Sjeff * If the thread isn't running its lockptr is a 1156171482Sjeff * turnstile or a sleepqueue. We can just lock_set without 1157171482Sjeff * blocking. 1158123685Sjeff */ 1159171482Sjeff if (TD_CAN_RUN(td)) { 1160171482Sjeff TDQ_LOCK(tdq); 1161171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1162171482Sjeff return (tdq); 1163171482Sjeff } 1164171482Sjeff#endif 1165166108Sjeff /* 1166171482Sjeff * The hard case, migration, we need to block the thread first to 1167171482Sjeff * prevent order reversals with other cpus locks. 1168166108Sjeff */ 1169202889Sattilio spinlock_enter(); 1170171482Sjeff thread_lock_block(td); 1171171482Sjeff TDQ_LOCK(tdq); 1172171713Sjeff thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1173202889Sattilio spinlock_exit(); 1174171482Sjeff return (tdq); 1175166108Sjeff} 1176166108Sjeff 1177178272SjeffSCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1178178272SjeffSCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1179178272SjeffSCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1180178272SjeffSCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1181178272SjeffSCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1182178272SjeffSCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1183178272Sjeff 1184166108Sjeffstatic int 1185177435Sjeffsched_pickcpu(struct thread *td, int flags) 1186171482Sjeff{ 1187232207Smav struct cpu_group *cg, *ccg; 1188177435Sjeff struct td_sched *ts; 1189171482Sjeff struct tdq *tdq; 1190194779Sjeff cpuset_t mask; 1191232207Smav int cpu, pri, self; 1192166108Sjeff 1193176735Sjeff self = PCPU_GET(cpuid); 1194177435Sjeff ts = td->td_sched; 1195166108Sjeff if (smp_started == 0) 1196166108Sjeff return (self); 1197171506Sjeff /* 1198171506Sjeff * Don't migrate a running thread from sched_switch(). 1199171506Sjeff */ 1200176735Sjeff if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1201176735Sjeff return (ts->ts_cpu); 1202166108Sjeff /* 1203176735Sjeff * Prefer to run interrupt threads on the processors that generate 1204176735Sjeff * the interrupt. 1205166108Sjeff */ 1206232207Smav pri = td->td_priority; 1207176735Sjeff if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1208178272Sjeff curthread->td_intr_nesting_level && ts->ts_cpu != self) { 1209178272Sjeff SCHED_STAT_INC(pickcpu_intrbind); 1210176735Sjeff ts->ts_cpu = self; 1211232207Smav if (TDQ_CPU(self)->tdq_lowpri > pri) { 1212232207Smav SCHED_STAT_INC(pickcpu_affinity); 1213232207Smav return (ts->ts_cpu); 1214232207Smav } 1215178272Sjeff } 1216166108Sjeff /* 1217176735Sjeff * If the thread can run on the last cpu and the affinity has not 1218176735Sjeff * expired or it is idle run it there. 1219166108Sjeff */ 1220176735Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1221232207Smav cg = tdq->tdq_cg; 1222232207Smav if (THREAD_CAN_SCHED(td, ts->ts_cpu) && 1223232207Smav tdq->tdq_lowpri >= PRI_MIN_IDLE && 1224232207Smav SCHED_AFFINITY(ts, CG_SHARE_L2)) { 1225232207Smav if (cg->cg_flags & CG_FLAG_THREAD) { 1226232207Smav CPUSET_FOREACH(cpu, cg->cg_mask) { 1227232207Smav if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) 1228232207Smav break; 1229232207Smav } 1230232207Smav } else 1231232207Smav cpu = INT_MAX; 1232232207Smav if (cpu > mp_maxid) { 1233178272Sjeff SCHED_STAT_INC(pickcpu_idle_affinity); 1234176735Sjeff return (ts->ts_cpu); 1235178272Sjeff } 1236139334Sjeff } 1237123433Sjeff /* 1238232207Smav * Search for the last level cache CPU group in the tree. 1239232207Smav * Skip caches with expired affinity time and SMT groups. 1240232207Smav * Affinity to higher level caches will be handled less aggressively. 1241123433Sjeff */ 1242232207Smav for (ccg = NULL; cg != NULL; cg = cg->cg_parent) { 1243232207Smav if (cg->cg_flags & CG_FLAG_THREAD) 1244232207Smav continue; 1245232207Smav if (!SCHED_AFFINITY(ts, cg->cg_level)) 1246232207Smav continue; 1247232207Smav ccg = cg; 1248232207Smav } 1249232207Smav if (ccg != NULL) 1250232207Smav cg = ccg; 1251176735Sjeff cpu = -1; 1252232207Smav /* Search the group for the less loaded idle CPU we can run now. */ 1253194779Sjeff mask = td->td_cpuset->cs_mask; 1254232207Smav if (cg != NULL && cg != cpu_top && 1255232207Smav CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0) 1256232207Smav cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE), 1257232207Smav INT_MAX, ts->ts_cpu); 1258232207Smav /* Search globally for the less loaded CPU we can run now. */ 1259176735Sjeff if (cpu == -1) 1260232207Smav cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu); 1261232207Smav /* Search globally for the less loaded CPU. */ 1262232207Smav if (cpu == -1) 1263232207Smav cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu); 1264232454Smav KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1265171506Sjeff /* 1266176735Sjeff * Compare the lowest loaded cpu to current cpu. 1267171506Sjeff */ 1268177005Sjeff if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1269232207Smav TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE && 1270232207Smav TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) { 1271178272Sjeff SCHED_STAT_INC(pickcpu_local); 1272177005Sjeff cpu = self; 1273178272Sjeff } else 1274178272Sjeff SCHED_STAT_INC(pickcpu_lowest); 1275178272Sjeff if (cpu != ts->ts_cpu) 1276178272Sjeff SCHED_STAT_INC(pickcpu_migration); 1277171482Sjeff return (cpu); 1278123433Sjeff} 1279176735Sjeff#endif 1280123433Sjeff 1281117326Sjeff/* 1282121790Sjeff * Pick the highest priority task we have and return it. 1283117326Sjeff */ 1284177435Sjeffstatic struct thread * 1285164936Sjuliantdq_choose(struct tdq *tdq) 1286110267Sjeff{ 1287177435Sjeff struct thread *td; 1288110267Sjeff 1289171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1290177435Sjeff td = runq_choose(&tdq->tdq_realtime); 1291177435Sjeff if (td != NULL) 1292177435Sjeff return (td); 1293177435Sjeff td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1294177435Sjeff if (td != NULL) { 1295217351Sjhb KASSERT(td->td_priority >= PRI_MIN_BATCH, 1296165762Sjeff ("tdq_choose: Invalid priority on timeshare queue %d", 1297177435Sjeff td->td_priority)); 1298177435Sjeff return (td); 1299165762Sjeff } 1300177435Sjeff td = runq_choose(&tdq->tdq_idle); 1301177435Sjeff if (td != NULL) { 1302177435Sjeff KASSERT(td->td_priority >= PRI_MIN_IDLE, 1303165762Sjeff ("tdq_choose: Invalid priority on idle queue %d", 1304177435Sjeff td->td_priority)); 1305177435Sjeff return (td); 1306165762Sjeff } 1307165762Sjeff 1308165762Sjeff return (NULL); 1309110267Sjeff} 1310110267Sjeff 1311171482Sjeff/* 1312171482Sjeff * Initialize a thread queue. 1313171482Sjeff */ 1314109864Sjeffstatic void 1315164936Sjuliantdq_setup(struct tdq *tdq) 1316110028Sjeff{ 1317171482Sjeff 1318171713Sjeff if (bootverbose) 1319171713Sjeff printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1320165762Sjeff runq_init(&tdq->tdq_realtime); 1321165762Sjeff runq_init(&tdq->tdq_timeshare); 1322165620Sjeff runq_init(&tdq->tdq_idle); 1323176735Sjeff snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1324176735Sjeff "sched lock %d", (int)TDQ_ID(tdq)); 1325176735Sjeff mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1326176735Sjeff MTX_SPIN | MTX_RECURSE); 1327187357Sjeff#ifdef KTR 1328187357Sjeff snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname), 1329187357Sjeff "CPU %d load", (int)TDQ_ID(tdq)); 1330187357Sjeff#endif 1331110028Sjeff} 1332110028Sjeff 1333171713Sjeff#ifdef SMP 1334110028Sjeffstatic void 1335171713Sjeffsched_setup_smp(void) 1336171713Sjeff{ 1337171713Sjeff struct tdq *tdq; 1338171713Sjeff int i; 1339171713Sjeff 1340176735Sjeff cpu_top = smp_topo(); 1341209059Sjhb CPU_FOREACH(i) { 1342176735Sjeff tdq = TDQ_CPU(i); 1343171713Sjeff tdq_setup(tdq); 1344176735Sjeff tdq->tdq_cg = smp_topo_find(cpu_top, i); 1345176735Sjeff if (tdq->tdq_cg == NULL) 1346176735Sjeff panic("Can't find cpu group for %d\n", i); 1347123433Sjeff } 1348176735Sjeff balance_tdq = TDQ_SELF(); 1349176735Sjeff sched_balance(); 1350171713Sjeff} 1351171713Sjeff#endif 1352171713Sjeff 1353171713Sjeff/* 1354171713Sjeff * Setup the thread queues and initialize the topology based on MD 1355171713Sjeff * information. 1356171713Sjeff */ 1357171713Sjeffstatic void 1358171713Sjeffsched_setup(void *dummy) 1359171713Sjeff{ 1360171713Sjeff struct tdq *tdq; 1361171713Sjeff 1362171713Sjeff tdq = TDQ_SELF(); 1363171713Sjeff#ifdef SMP 1364176734Sjeff sched_setup_smp(); 1365117237Sjeff#else 1366171713Sjeff tdq_setup(tdq); 1367116069Sjeff#endif 1368171482Sjeff 1369171482Sjeff /* Add thread0's load since it's running. */ 1370171482Sjeff TDQ_LOCK(tdq); 1371171713Sjeff thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1372177435Sjeff tdq_load_add(tdq, &thread0); 1373176735Sjeff tdq->tdq_lowpri = thread0.td_priority; 1374171482Sjeff TDQ_UNLOCK(tdq); 1375109864Sjeff} 1376109864Sjeff 1377171482Sjeff/* 1378239185Smav * This routine determines time constants after stathz and hz are setup. 1379171482Sjeff */ 1380153533Sdavidxu/* ARGSUSED */ 1381153533Sdavidxustatic void 1382153533Sdavidxusched_initticks(void *dummy) 1383153533Sdavidxu{ 1384171482Sjeff int incr; 1385171482Sjeff 1386153533Sdavidxu realstathz = stathz ? stathz : hz; 1387239185Smav sched_slice = realstathz / 10; /* ~100ms */ 1388239196Smav hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 1389239196Smav realstathz); 1390153533Sdavidxu 1391153533Sdavidxu /* 1392165762Sjeff * tickincr is shifted out by 10 to avoid rounding errors due to 1393165766Sjeff * hz not being evenly divisible by stathz on all platforms. 1394153533Sdavidxu */ 1395171482Sjeff incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1396165762Sjeff /* 1397165762Sjeff * This does not work for values of stathz that are more than 1398165762Sjeff * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1399165762Sjeff */ 1400171482Sjeff if (incr == 0) 1401171482Sjeff incr = 1; 1402171482Sjeff tickincr = incr; 1403166108Sjeff#ifdef SMP 1404171899Sjeff /* 1405172409Sjeff * Set the default balance interval now that we know 1406172409Sjeff * what realstathz is. 1407172409Sjeff */ 1408172409Sjeff balance_interval = realstathz; 1409166108Sjeff affinity = SCHED_AFFINITY_DEFAULT; 1410166108Sjeff#endif 1411232740Smav if (sched_idlespinthresh < 0) 1412239196Smav sched_idlespinthresh = imax(16, 2 * hz / realstathz); 1413153533Sdavidxu} 1414153533Sdavidxu 1415153533Sdavidxu 1416109864Sjeff/* 1417171482Sjeff * This is the core of the interactivity algorithm. Determines a score based 1418171482Sjeff * on past behavior. It is the ratio of sleep time to run time scaled to 1419171482Sjeff * a [0, 100] integer. This is the voluntary sleep time of a process, which 1420171482Sjeff * differs from the cpu usage because it does not account for time spent 1421171482Sjeff * waiting on a run-queue. Would be prettier if we had floating point. 1422171482Sjeff */ 1423171482Sjeffstatic int 1424171482Sjeffsched_interact_score(struct thread *td) 1425171482Sjeff{ 1426171482Sjeff struct td_sched *ts; 1427171482Sjeff int div; 1428171482Sjeff 1429171482Sjeff ts = td->td_sched; 1430171482Sjeff /* 1431171482Sjeff * The score is only needed if this is likely to be an interactive 1432171482Sjeff * task. Don't go through the expense of computing it if there's 1433171482Sjeff * no chance. 1434171482Sjeff */ 1435171482Sjeff if (sched_interact <= SCHED_INTERACT_HALF && 1436171482Sjeff ts->ts_runtime >= ts->ts_slptime) 1437171482Sjeff return (SCHED_INTERACT_HALF); 1438171482Sjeff 1439171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1440171482Sjeff div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1441171482Sjeff return (SCHED_INTERACT_HALF + 1442171482Sjeff (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1443171482Sjeff } 1444171482Sjeff if (ts->ts_slptime > ts->ts_runtime) { 1445171482Sjeff div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1446171482Sjeff return (ts->ts_runtime / div); 1447171482Sjeff } 1448171482Sjeff /* runtime == slptime */ 1449171482Sjeff if (ts->ts_runtime) 1450171482Sjeff return (SCHED_INTERACT_HALF); 1451171482Sjeff 1452171482Sjeff /* 1453171482Sjeff * This can happen if slptime and runtime are 0. 1454171482Sjeff */ 1455171482Sjeff return (0); 1456171482Sjeff 1457171482Sjeff} 1458171482Sjeff 1459171482Sjeff/* 1460109864Sjeff * Scale the scheduling priority according to the "interactivity" of this 1461109864Sjeff * process. 1462109864Sjeff */ 1463113357Sjeffstatic void 1464163709Sjbsched_priority(struct thread *td) 1465109864Sjeff{ 1466165762Sjeff int score; 1467109864Sjeff int pri; 1468109864Sjeff 1469217291Sjhb if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1470113357Sjeff return; 1471112966Sjeff /* 1472165762Sjeff * If the score is interactive we place the thread in the realtime 1473165762Sjeff * queue with a priority that is less than kernel and interrupt 1474165762Sjeff * priorities. These threads are not subject to nice restrictions. 1475112966Sjeff * 1476171482Sjeff * Scores greater than this are placed on the normal timeshare queue 1477165762Sjeff * where the priority is partially decided by the most recent cpu 1478165762Sjeff * utilization and the rest is decided by nice value. 1479172293Sjeff * 1480172293Sjeff * The nice value of the process has a linear effect on the calculated 1481172293Sjeff * score. Negative nice values make it easier for a thread to be 1482172293Sjeff * considered interactive. 1483112966Sjeff */ 1484198126Sjhb score = imax(0, sched_interact_score(td) + td->td_proc->p_nice); 1485165762Sjeff if (score < sched_interact) { 1486217351Sjhb pri = PRI_MIN_INTERACT; 1487217351Sjhb pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) / 1488217237Sjhb sched_interact) * score; 1489217351Sjhb KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT, 1490166208Sjeff ("sched_priority: invalid interactive priority %d score %d", 1491166208Sjeff pri, score)); 1492165762Sjeff } else { 1493165762Sjeff pri = SCHED_PRI_MIN; 1494165762Sjeff if (td->td_sched->ts_ticks) 1495228960Sjhb pri += min(SCHED_PRI_TICKS(td->td_sched), 1496228960Sjhb SCHED_PRI_RANGE); 1497165762Sjeff pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1498217351Sjhb KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH, 1499171482Sjeff ("sched_priority: invalid priority %d: nice %d, " 1500171482Sjeff "ticks %d ftick %d ltick %d tick pri %d", 1501171482Sjeff pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1502171482Sjeff td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1503171482Sjeff SCHED_PRI_TICKS(td->td_sched))); 1504165762Sjeff } 1505165762Sjeff sched_user_prio(td, pri); 1506112966Sjeff 1507112966Sjeff return; 1508109864Sjeff} 1509109864Sjeff 1510121868Sjeff/* 1511121868Sjeff * This routine enforces a maximum limit on the amount of scheduling history 1512171482Sjeff * kept. It is called after either the slptime or runtime is adjusted. This 1513171482Sjeff * function is ugly due to integer math. 1514121868Sjeff */ 1515116463Sjeffstatic void 1516163709Sjbsched_interact_update(struct thread *td) 1517116463Sjeff{ 1518165819Sjeff struct td_sched *ts; 1519166208Sjeff u_int sum; 1520121605Sjeff 1521165819Sjeff ts = td->td_sched; 1522171482Sjeff sum = ts->ts_runtime + ts->ts_slptime; 1523121868Sjeff if (sum < SCHED_SLP_RUN_MAX) 1524121868Sjeff return; 1525121868Sjeff /* 1526165819Sjeff * This only happens from two places: 1527165819Sjeff * 1) We have added an unusual amount of run time from fork_exit. 1528165819Sjeff * 2) We have added an unusual amount of sleep time from sched_sleep(). 1529165819Sjeff */ 1530165819Sjeff if (sum > SCHED_SLP_RUN_MAX * 2) { 1531171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1532171482Sjeff ts->ts_runtime = SCHED_SLP_RUN_MAX; 1533171482Sjeff ts->ts_slptime = 1; 1534165819Sjeff } else { 1535171482Sjeff ts->ts_slptime = SCHED_SLP_RUN_MAX; 1536171482Sjeff ts->ts_runtime = 1; 1537165819Sjeff } 1538165819Sjeff return; 1539165819Sjeff } 1540165819Sjeff /* 1541121868Sjeff * If we have exceeded by more than 1/5th then the algorithm below 1542121868Sjeff * will not bring us back into range. Dividing by two here forces 1543133427Sjeff * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1544121868Sjeff */ 1545127850Sjeff if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1546171482Sjeff ts->ts_runtime /= 2; 1547171482Sjeff ts->ts_slptime /= 2; 1548121868Sjeff return; 1549116463Sjeff } 1550171482Sjeff ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1551171482Sjeff ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1552116463Sjeff} 1553116463Sjeff 1554171482Sjeff/* 1555171482Sjeff * Scale back the interactivity history when a child thread is created. The 1556171482Sjeff * history is inherited from the parent but the thread may behave totally 1557171482Sjeff * differently. For example, a shell spawning a compiler process. We want 1558171482Sjeff * to learn that the compiler is behaving badly very quickly. 1559171482Sjeff */ 1560121868Sjeffstatic void 1561163709Sjbsched_interact_fork(struct thread *td) 1562121868Sjeff{ 1563121868Sjeff int ratio; 1564121868Sjeff int sum; 1565121868Sjeff 1566171482Sjeff sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1567121868Sjeff if (sum > SCHED_SLP_RUN_FORK) { 1568121868Sjeff ratio = sum / SCHED_SLP_RUN_FORK; 1569171482Sjeff td->td_sched->ts_runtime /= ratio; 1570171482Sjeff td->td_sched->ts_slptime /= ratio; 1571121868Sjeff } 1572121868Sjeff} 1573121868Sjeff 1574113357Sjeff/* 1575171482Sjeff * Called from proc0_init() to setup the scheduler fields. 1576134791Sjulian */ 1577134791Sjulianvoid 1578134791Sjulianschedinit(void) 1579134791Sjulian{ 1580165762Sjeff 1581134791Sjulian /* 1582134791Sjulian * Set up the scheduler specific parts of proc0. 1583134791Sjulian */ 1584136167Sjulian proc0.p_sched = NULL; /* XXX */ 1585164936Sjulian thread0.td_sched = &td_sched0; 1586165762Sjeff td_sched0.ts_ltick = ticks; 1587165796Sjeff td_sched0.ts_ftick = ticks; 1588177009Sjeff td_sched0.ts_slice = sched_slice; 1589134791Sjulian} 1590134791Sjulian 1591134791Sjulian/* 1592113357Sjeff * This is only somewhat accurate since given many processes of the same 1593113357Sjeff * priority they will switch when their slices run out, which will be 1594165762Sjeff * at most sched_slice stathz ticks. 1595113357Sjeff */ 1596109864Sjeffint 1597109864Sjeffsched_rr_interval(void) 1598109864Sjeff{ 1599165762Sjeff 1600239185Smav /* Convert sched_slice from stathz to hz. */ 1601239196Smav return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz)); 1602109864Sjeff} 1603109864Sjeff 1604171482Sjeff/* 1605171482Sjeff * Update the percent cpu tracking information when it is requested or 1606171482Sjeff * the total history exceeds the maximum. We keep a sliding history of 1607171482Sjeff * tick counts that slowly decays. This is less precise than the 4BSD 1608171482Sjeff * mechanism since it happens with less regular and frequent events. 1609171482Sjeff */ 1610121790Sjeffstatic void 1611232917Smavsched_pctcpu_update(struct td_sched *ts, int run) 1612109864Sjeff{ 1613232917Smav int t = ticks; 1614165762Sjeff 1615232917Smav if (t - ts->ts_ltick >= SCHED_TICK_TARG) { 1616164936Sjulian ts->ts_ticks = 0; 1617232917Smav ts->ts_ftick = t - SCHED_TICK_TARG; 1618232917Smav } else if (t - ts->ts_ftick >= SCHED_TICK_MAX) { 1619232917Smav ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) * 1620232917Smav (ts->ts_ltick - (t - SCHED_TICK_TARG)); 1621232917Smav ts->ts_ftick = t - SCHED_TICK_TARG; 1622232917Smav } 1623232917Smav if (run) 1624232917Smav ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT; 1625232917Smav ts->ts_ltick = t; 1626109864Sjeff} 1627109864Sjeff 1628171482Sjeff/* 1629171482Sjeff * Adjust the priority of a thread. Move it to the appropriate run-queue 1630171482Sjeff * if necessary. This is the back-end for several priority related 1631171482Sjeff * functions. 1632171482Sjeff */ 1633165762Sjeffstatic void 1634139453Sjhbsched_thread_priority(struct thread *td, u_char prio) 1635109864Sjeff{ 1636164936Sjulian struct td_sched *ts; 1637177009Sjeff struct tdq *tdq; 1638177009Sjeff int oldpri; 1639109864Sjeff 1640187357Sjeff KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio", 1641187357Sjeff "prio:%d", td->td_priority, "new prio:%d", prio, 1642187357Sjeff KTR_ATTR_LINKED, sched_tdname(curthread)); 1643235459Srstone SDT_PROBE3(sched, , , change_pri, td, td->td_proc, prio); 1644240513Savg if (td != curthread && prio < td->td_priority) { 1645187357Sjeff KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 1646187357Sjeff "lend prio", "prio:%d", td->td_priority, "new prio:%d", 1647187357Sjeff prio, KTR_ATTR_LINKED, sched_tdname(td)); 1648235459Srstone SDT_PROBE4(sched, , , lend_pri, td, td->td_proc, prio, 1649235459Srstone curthread); 1650187357Sjeff } 1651164936Sjulian ts = td->td_sched; 1652170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1653139453Sjhb if (td->td_priority == prio) 1654139453Sjhb return; 1655177376Sjeff /* 1656177376Sjeff * If the priority has been elevated due to priority 1657177376Sjeff * propagation, we may have to move ourselves to a new 1658177376Sjeff * queue. This could be optimized to not re-add in some 1659177376Sjeff * cases. 1660177376Sjeff */ 1661165766Sjeff if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1662165762Sjeff sched_rem(td); 1663165762Sjeff td->td_priority = prio; 1664171482Sjeff sched_add(td, SRQ_BORROWING); 1665177009Sjeff return; 1666177009Sjeff } 1667177376Sjeff /* 1668177376Sjeff * If the thread is currently running we may have to adjust the lowpri 1669177376Sjeff * information so other cpus are aware of our current priority. 1670177376Sjeff */ 1671177009Sjeff if (TD_IS_RUNNING(td)) { 1672177376Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1673177376Sjeff oldpri = td->td_priority; 1674177376Sjeff td->td_priority = prio; 1675176735Sjeff if (prio < tdq->tdq_lowpri) 1676171482Sjeff tdq->tdq_lowpri = prio; 1677176735Sjeff else if (tdq->tdq_lowpri == oldpri) 1678176735Sjeff tdq_setlowpri(tdq, td); 1679177376Sjeff return; 1680177009Sjeff } 1681177376Sjeff td->td_priority = prio; 1682109864Sjeff} 1683109864Sjeff 1684139453Sjhb/* 1685139453Sjhb * Update a thread's priority when it is lent another thread's 1686139453Sjhb * priority. 1687139453Sjhb */ 1688109864Sjeffvoid 1689139453Sjhbsched_lend_prio(struct thread *td, u_char prio) 1690139453Sjhb{ 1691139453Sjhb 1692139453Sjhb td->td_flags |= TDF_BORROWING; 1693139453Sjhb sched_thread_priority(td, prio); 1694139453Sjhb} 1695139453Sjhb 1696139453Sjhb/* 1697139453Sjhb * Restore a thread's priority when priority propagation is 1698139453Sjhb * over. The prio argument is the minimum priority the thread 1699139453Sjhb * needs to have to satisfy other possible priority lending 1700139453Sjhb * requests. If the thread's regular priority is less 1701139453Sjhb * important than prio, the thread will keep a priority boost 1702139453Sjhb * of prio. 1703139453Sjhb */ 1704139453Sjhbvoid 1705139453Sjhbsched_unlend_prio(struct thread *td, u_char prio) 1706139453Sjhb{ 1707139453Sjhb u_char base_pri; 1708139453Sjhb 1709139453Sjhb if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1710139453Sjhb td->td_base_pri <= PRI_MAX_TIMESHARE) 1711163709Sjb base_pri = td->td_user_pri; 1712139453Sjhb else 1713139453Sjhb base_pri = td->td_base_pri; 1714139453Sjhb if (prio >= base_pri) { 1715139455Sjhb td->td_flags &= ~TDF_BORROWING; 1716139453Sjhb sched_thread_priority(td, base_pri); 1717139453Sjhb } else 1718139453Sjhb sched_lend_prio(td, prio); 1719139453Sjhb} 1720139453Sjhb 1721171482Sjeff/* 1722171482Sjeff * Standard entry for setting the priority to an absolute value. 1723171482Sjeff */ 1724139453Sjhbvoid 1725139453Sjhbsched_prio(struct thread *td, u_char prio) 1726139453Sjhb{ 1727139453Sjhb u_char oldprio; 1728139453Sjhb 1729139453Sjhb /* First, update the base priority. */ 1730139453Sjhb td->td_base_pri = prio; 1731139453Sjhb 1732139453Sjhb /* 1733139455Sjhb * If the thread is borrowing another thread's priority, don't 1734139453Sjhb * ever lower the priority. 1735139453Sjhb */ 1736139453Sjhb if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1737139453Sjhb return; 1738139453Sjhb 1739139453Sjhb /* Change the real priority. */ 1740139453Sjhb oldprio = td->td_priority; 1741139453Sjhb sched_thread_priority(td, prio); 1742139453Sjhb 1743139453Sjhb /* 1744139453Sjhb * If the thread is on a turnstile, then let the turnstile update 1745139453Sjhb * its state. 1746139453Sjhb */ 1747139453Sjhb if (TD_ON_LOCK(td) && oldprio != prio) 1748139453Sjhb turnstile_adjust(td, oldprio); 1749139453Sjhb} 1750139455Sjhb 1751171482Sjeff/* 1752171482Sjeff * Set the base user priority, does not effect current running priority. 1753171482Sjeff */ 1754139453Sjhbvoid 1755163709Sjbsched_user_prio(struct thread *td, u_char prio) 1756161599Sdavidxu{ 1757161599Sdavidxu 1758163709Sjb td->td_base_user_pri = prio; 1759216313Sdavidxu if (td->td_lend_user_pri <= prio) 1760216313Sdavidxu return; 1761163709Sjb td->td_user_pri = prio; 1762161599Sdavidxu} 1763161599Sdavidxu 1764161599Sdavidxuvoid 1765161599Sdavidxusched_lend_user_prio(struct thread *td, u_char prio) 1766161599Sdavidxu{ 1767161599Sdavidxu 1768174536Sdavidxu THREAD_LOCK_ASSERT(td, MA_OWNED); 1769216313Sdavidxu td->td_lend_user_pri = prio; 1770216791Sdavidxu td->td_user_pri = min(prio, td->td_base_user_pri); 1771216791Sdavidxu if (td->td_priority > td->td_user_pri) 1772216791Sdavidxu sched_prio(td, td->td_user_pri); 1773216791Sdavidxu else if (td->td_priority != td->td_user_pri) 1774216791Sdavidxu td->td_flags |= TDF_NEEDRESCHED; 1775161599Sdavidxu} 1776161599Sdavidxu 1777171482Sjeff/* 1778171713Sjeff * Handle migration from sched_switch(). This happens only for 1779171713Sjeff * cpu binding. 1780171713Sjeff */ 1781171713Sjeffstatic struct mtx * 1782171713Sjeffsched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1783171713Sjeff{ 1784171713Sjeff struct tdq *tdn; 1785171713Sjeff 1786171713Sjeff tdn = TDQ_CPU(td->td_sched->ts_cpu); 1787171713Sjeff#ifdef SMP 1788177435Sjeff tdq_load_rem(tdq, td); 1789171713Sjeff /* 1790171713Sjeff * Do the lock dance required to avoid LOR. We grab an extra 1791171713Sjeff * spinlock nesting to prevent preemption while we're 1792171713Sjeff * not holding either run-queue lock. 1793171713Sjeff */ 1794171713Sjeff spinlock_enter(); 1795202889Sattilio thread_lock_block(td); /* This releases the lock on tdq. */ 1796197223Sattilio 1797197223Sattilio /* 1798197223Sattilio * Acquire both run-queue locks before placing the thread on the new 1799197223Sattilio * run-queue to avoid deadlocks created by placing a thread with a 1800197223Sattilio * blocked lock on the run-queue of a remote processor. The deadlock 1801197223Sattilio * occurs when a third processor attempts to lock the two queues in 1802197223Sattilio * question while the target processor is spinning with its own 1803197223Sattilio * run-queue lock held while waiting for the blocked lock to clear. 1804197223Sattilio */ 1805197223Sattilio tdq_lock_pair(tdn, tdq); 1806171713Sjeff tdq_add(tdn, td, flags); 1807177435Sjeff tdq_notify(tdn, td); 1808197223Sattilio TDQ_UNLOCK(tdn); 1809171713Sjeff spinlock_exit(); 1810171713Sjeff#endif 1811171713Sjeff return (TDQ_LOCKPTR(tdn)); 1812171713Sjeff} 1813171713Sjeff 1814171713Sjeff/* 1815202889Sattilio * Variadic version of thread_lock_unblock() that does not assume td_lock 1816202889Sattilio * is blocked. 1817171482Sjeff */ 1818171482Sjeffstatic inline void 1819171482Sjeffthread_unblock_switch(struct thread *td, struct mtx *mtx) 1820171482Sjeff{ 1821171482Sjeff atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1822171482Sjeff (uintptr_t)mtx); 1823171482Sjeff} 1824171482Sjeff 1825171482Sjeff/* 1826171482Sjeff * Switch threads. This function has to handle threads coming in while 1827171482Sjeff * blocked for some reason, running, or idle. It also must deal with 1828171482Sjeff * migrating a thread from one queue to another as running threads may 1829171482Sjeff * be assigned elsewhere via binding. 1830171482Sjeff */ 1831161599Sdavidxuvoid 1832135051Sjuliansched_switch(struct thread *td, struct thread *newtd, int flags) 1833109864Sjeff{ 1834165627Sjeff struct tdq *tdq; 1835164936Sjulian struct td_sched *ts; 1836171482Sjeff struct mtx *mtx; 1837171713Sjeff int srqflag; 1838239157Smav int cpuid, preempted; 1839109864Sjeff 1840170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1841177376Sjeff KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1842109864Sjeff 1843171482Sjeff cpuid = PCPU_GET(cpuid); 1844171482Sjeff tdq = TDQ_CPU(cpuid); 1845164936Sjulian ts = td->td_sched; 1846171713Sjeff mtx = td->td_lock; 1847232917Smav sched_pctcpu_update(ts, 1); 1848171482Sjeff ts->ts_rltick = ticks; 1849133555Sjeff td->td_lastcpu = td->td_oncpu; 1850113339Sjulian td->td_oncpu = NOCPU; 1851239157Smav preempted = !(td->td_flags & TDF_SLICEEND); 1852239157Smav td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND); 1853144777Sups td->td_owepreempt = 0; 1854178277Sjeff tdq->tdq_switchcnt++; 1855123434Sjeff /* 1856171482Sjeff * The lock pointer in an idle thread should never change. Reset it 1857171482Sjeff * to CAN_RUN as well. 1858123434Sjeff */ 1859167327Sjulian if (TD_IS_IDLETHREAD(td)) { 1860171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1861139334Sjeff TD_SET_CAN_RUN(td); 1862170293Sjeff } else if (TD_IS_RUNNING(td)) { 1863171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1864239157Smav srqflag = preempted ? 1865170293Sjeff SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1866171713Sjeff SRQ_OURSELF|SRQ_YIELDING; 1867212153Smdf#ifdef SMP 1868212115Smdf if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu)) 1869212115Smdf ts->ts_cpu = sched_pickcpu(td, 0); 1870212153Smdf#endif 1871171713Sjeff if (ts->ts_cpu == cpuid) 1872177435Sjeff tdq_runq_add(tdq, td, srqflag); 1873212115Smdf else { 1874212115Smdf KASSERT(THREAD_CAN_MIGRATE(td) || 1875212115Smdf (ts->ts_flags & TSF_BOUND) != 0, 1876212115Smdf ("Thread %p shouldn't migrate", td)); 1877171713Sjeff mtx = sched_switch_migrate(tdq, td, srqflag); 1878212115Smdf } 1879171482Sjeff } else { 1880171482Sjeff /* This thread must be going to sleep. */ 1881171482Sjeff TDQ_LOCK(tdq); 1882202889Sattilio mtx = thread_lock_block(td); 1883177435Sjeff tdq_load_rem(tdq, td); 1884171482Sjeff } 1885171482Sjeff /* 1886171482Sjeff * We enter here with the thread blocked and assigned to the 1887171482Sjeff * appropriate cpu run-queue or sleep-queue and with the current 1888171482Sjeff * thread-queue locked. 1889171482Sjeff */ 1890171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1891171482Sjeff newtd = choosethread(); 1892171482Sjeff /* 1893171482Sjeff * Call the MD code to switch contexts if necessary. 1894171482Sjeff */ 1895145256Sjkoshy if (td != newtd) { 1896145256Sjkoshy#ifdef HWPMC_HOOKS 1897145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1898145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1899145256Sjkoshy#endif 1900235459Srstone SDT_PROBE2(sched, , , off_cpu, td, td->td_proc); 1901174629Sjeff lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1902172411Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1903232917Smav sched_pctcpu_update(newtd->td_sched, 0); 1904179297Sjb 1905179297Sjb#ifdef KDTRACE_HOOKS 1906179297Sjb /* 1907179297Sjb * If DTrace has set the active vtime enum to anything 1908179297Sjb * other than INACTIVE (0), then it should have set the 1909179297Sjb * function to call. 1910179297Sjb */ 1911179297Sjb if (dtrace_vtime_active) 1912179297Sjb (*dtrace_vtime_switch_func)(newtd); 1913179297Sjb#endif 1914179297Sjb 1915171482Sjeff cpu_switch(td, newtd, mtx); 1916171482Sjeff /* 1917171482Sjeff * We may return from cpu_switch on a different cpu. However, 1918171482Sjeff * we always return with td_lock pointing to the current cpu's 1919171482Sjeff * run queue lock. 1920171482Sjeff */ 1921171482Sjeff cpuid = PCPU_GET(cpuid); 1922171482Sjeff tdq = TDQ_CPU(cpuid); 1923174629Sjeff lock_profile_obtain_lock_success( 1924174629Sjeff &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1925235459Srstone 1926235459Srstone SDT_PROBE0(sched, , , on_cpu); 1927145256Sjkoshy#ifdef HWPMC_HOOKS 1928145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1929145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1930145256Sjkoshy#endif 1931235459Srstone } else { 1932171482Sjeff thread_unblock_switch(td, mtx); 1933235459Srstone SDT_PROBE0(sched, , , remain_cpu); 1934235459Srstone } 1935171482Sjeff /* 1936171482Sjeff * Assert that all went well and return. 1937171482Sjeff */ 1938171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1939171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1940171482Sjeff td->td_oncpu = cpuid; 1941109864Sjeff} 1942109864Sjeff 1943171482Sjeff/* 1944171482Sjeff * Adjust thread priorities as a result of a nice request. 1945171482Sjeff */ 1946109864Sjeffvoid 1947130551Sjuliansched_nice(struct proc *p, int nice) 1948109864Sjeff{ 1949109864Sjeff struct thread *td; 1950109864Sjeff 1951130551Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 1952165762Sjeff 1953130551Sjulian p->p_nice = nice; 1954163709Sjb FOREACH_THREAD_IN_PROC(p, td) { 1955170293Sjeff thread_lock(td); 1956163709Sjb sched_priority(td); 1957165762Sjeff sched_prio(td, td->td_base_user_pri); 1958170293Sjeff thread_unlock(td); 1959130551Sjulian } 1960109864Sjeff} 1961109864Sjeff 1962171482Sjeff/* 1963171482Sjeff * Record the sleep time for the interactivity scorer. 1964171482Sjeff */ 1965109864Sjeffvoid 1966177085Sjeffsched_sleep(struct thread *td, int prio) 1967109864Sjeff{ 1968165762Sjeff 1969170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1970109864Sjeff 1971172264Sjeff td->td_slptick = ticks; 1972201347Skib if (TD_IS_SUSPENDED(td) || prio >= PSOCK) 1973177085Sjeff td->td_flags |= TDF_CANSWAP; 1974217410Sjhb if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1975217410Sjhb return; 1976177903Sjeff if (static_boost == 1 && prio) 1977177085Sjeff sched_prio(td, prio); 1978177903Sjeff else if (static_boost && td->td_priority > static_boost) 1979177903Sjeff sched_prio(td, static_boost); 1980109864Sjeff} 1981109864Sjeff 1982171482Sjeff/* 1983171482Sjeff * Schedule a thread to resume execution and record how long it voluntarily 1984171482Sjeff * slept. We also update the pctcpu, interactivity, and priority. 1985171482Sjeff */ 1986109864Sjeffvoid 1987109864Sjeffsched_wakeup(struct thread *td) 1988109864Sjeff{ 1989166229Sjeff struct td_sched *ts; 1990171482Sjeff int slptick; 1991165762Sjeff 1992170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1993166229Sjeff ts = td->td_sched; 1994177085Sjeff td->td_flags &= ~TDF_CANSWAP; 1995109864Sjeff /* 1996165762Sjeff * If we slept for more than a tick update our interactivity and 1997165762Sjeff * priority. 1998109864Sjeff */ 1999172264Sjeff slptick = td->td_slptick; 2000172264Sjeff td->td_slptick = 0; 2001171482Sjeff if (slptick && slptick != ticks) { 2002232917Smav ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT; 2003165819Sjeff sched_interact_update(td); 2004232917Smav sched_pctcpu_update(ts, 0); 2005109864Sjeff } 2006166229Sjeff /* Reset the slice value after we sleep. */ 2007166229Sjeff ts->ts_slice = sched_slice; 2008166190Sjeff sched_add(td, SRQ_BORING); 2009109864Sjeff} 2010109864Sjeff 2011109864Sjeff/* 2012109864Sjeff * Penalize the parent for creating a new child and initialize the child's 2013109864Sjeff * priority. 2014109864Sjeff */ 2015109864Sjeffvoid 2016163709Sjbsched_fork(struct thread *td, struct thread *child) 2017109864Sjeff{ 2018170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2019232917Smav sched_pctcpu_update(td->td_sched, 1); 2020164936Sjulian sched_fork_thread(td, child); 2021165762Sjeff /* 2022165762Sjeff * Penalize the parent and child for forking. 2023165762Sjeff */ 2024165762Sjeff sched_interact_fork(child); 2025165762Sjeff sched_priority(child); 2026171482Sjeff td->td_sched->ts_runtime += tickincr; 2027165762Sjeff sched_interact_update(td); 2028165762Sjeff sched_priority(td); 2029164936Sjulian} 2030109864Sjeff 2031171482Sjeff/* 2032171482Sjeff * Fork a new thread, may be within the same process. 2033171482Sjeff */ 2034164936Sjulianvoid 2035164936Sjuliansched_fork_thread(struct thread *td, struct thread *child) 2036164936Sjulian{ 2037164936Sjulian struct td_sched *ts; 2038164936Sjulian struct td_sched *ts2; 2039164936Sjulian 2040177426Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2041165762Sjeff /* 2042165762Sjeff * Initialize child. 2043165762Sjeff */ 2044177426Sjeff ts = td->td_sched; 2045177426Sjeff ts2 = child->td_sched; 2046171482Sjeff child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 2047176735Sjeff child->td_cpuset = cpuset_ref(td->td_cpuset); 2048164936Sjulian ts2->ts_cpu = ts->ts_cpu; 2049177426Sjeff ts2->ts_flags = 0; 2050165762Sjeff /* 2051217078Sjhb * Grab our parents cpu estimation information. 2052165762Sjeff */ 2053164936Sjulian ts2->ts_ticks = ts->ts_ticks; 2054164936Sjulian ts2->ts_ltick = ts->ts_ltick; 2055164936Sjulian ts2->ts_ftick = ts->ts_ftick; 2056165762Sjeff /* 2057217078Sjhb * Do not inherit any borrowed priority from the parent. 2058217078Sjhb */ 2059217078Sjhb child->td_priority = child->td_base_pri; 2060217078Sjhb /* 2061165762Sjeff * And update interactivity score. 2062165762Sjeff */ 2063171482Sjeff ts2->ts_slptime = ts->ts_slptime; 2064171482Sjeff ts2->ts_runtime = ts->ts_runtime; 2065165762Sjeff ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 2066187357Sjeff#ifdef KTR 2067187357Sjeff bzero(ts2->ts_name, sizeof(ts2->ts_name)); 2068187357Sjeff#endif 2069113357Sjeff} 2070113357Sjeff 2071171482Sjeff/* 2072171482Sjeff * Adjust the priority class of a thread. 2073171482Sjeff */ 2074113357Sjeffvoid 2075163709Sjbsched_class(struct thread *td, int class) 2076113357Sjeff{ 2077113357Sjeff 2078170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2079163709Sjb if (td->td_pri_class == class) 2080113357Sjeff return; 2081163709Sjb td->td_pri_class = class; 2082109864Sjeff} 2083109864Sjeff 2084109864Sjeff/* 2085109864Sjeff * Return some of the child's priority and interactivity to the parent. 2086109864Sjeff */ 2087109864Sjeffvoid 2088164939Sjuliansched_exit(struct proc *p, struct thread *child) 2089109864Sjeff{ 2090165762Sjeff struct thread *td; 2091113372Sjeff 2092187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit", 2093225199Sdelphij "prio:%d", child->td_priority); 2094177368Sjeff PROC_LOCK_ASSERT(p, MA_OWNED); 2095165762Sjeff td = FIRST_THREAD_IN_PROC(p); 2096165762Sjeff sched_exit_thread(td, child); 2097113372Sjeff} 2098113372Sjeff 2099171482Sjeff/* 2100171482Sjeff * Penalize another thread for the time spent on this one. This helps to 2101171482Sjeff * worsen the priority and interactivity of processes which schedule batch 2102171482Sjeff * jobs such as make. This has little effect on the make process itself but 2103171482Sjeff * causes new processes spawned by it to receive worse scores immediately. 2104171482Sjeff */ 2105113372Sjeffvoid 2106164939Sjuliansched_exit_thread(struct thread *td, struct thread *child) 2107164936Sjulian{ 2108165762Sjeff 2109187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit", 2110225199Sdelphij "prio:%d", child->td_priority); 2111165762Sjeff /* 2112165762Sjeff * Give the child's runtime to the parent without returning the 2113165762Sjeff * sleep time as a penalty to the parent. This causes shells that 2114165762Sjeff * launch expensive things to mark their children as expensive. 2115165762Sjeff */ 2116170293Sjeff thread_lock(td); 2117171482Sjeff td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2118164939Sjulian sched_interact_update(td); 2119165762Sjeff sched_priority(td); 2120170293Sjeff thread_unlock(td); 2121164936Sjulian} 2122164936Sjulian 2123177005Sjeffvoid 2124177005Sjeffsched_preempt(struct thread *td) 2125177005Sjeff{ 2126177005Sjeff struct tdq *tdq; 2127177005Sjeff 2128235459Srstone SDT_PROBE2(sched, , , surrender, td, td->td_proc); 2129235459Srstone 2130177005Sjeff thread_lock(td); 2131177005Sjeff tdq = TDQ_SELF(); 2132177005Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2133177005Sjeff tdq->tdq_ipipending = 0; 2134177005Sjeff if (td->td_priority > tdq->tdq_lowpri) { 2135178272Sjeff int flags; 2136178272Sjeff 2137178272Sjeff flags = SW_INVOL | SW_PREEMPT; 2138177005Sjeff if (td->td_critnest > 1) 2139177005Sjeff td->td_owepreempt = 1; 2140178272Sjeff else if (TD_IS_IDLETHREAD(td)) 2141178272Sjeff mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL); 2142177005Sjeff else 2143178272Sjeff mi_switch(flags | SWT_REMOTEPREEMPT, NULL); 2144177005Sjeff } 2145177005Sjeff thread_unlock(td); 2146177005Sjeff} 2147177005Sjeff 2148171482Sjeff/* 2149171482Sjeff * Fix priorities on return to user-space. Priorities may be elevated due 2150171482Sjeff * to static priorities in msleep() or similar. 2151171482Sjeff */ 2152164936Sjulianvoid 2153164936Sjuliansched_userret(struct thread *td) 2154164936Sjulian{ 2155164936Sjulian /* 2156164936Sjulian * XXX we cheat slightly on the locking here to avoid locking in 2157164936Sjulian * the usual case. Setting td_priority here is essentially an 2158164936Sjulian * incomplete workaround for not setting it properly elsewhere. 2159164936Sjulian * Now that some interrupt handlers are threads, not setting it 2160164936Sjulian * properly elsewhere can clobber it in the window between setting 2161164936Sjulian * it here and returning to user mode, so don't waste time setting 2162164936Sjulian * it perfectly here. 2163164936Sjulian */ 2164164936Sjulian KASSERT((td->td_flags & TDF_BORROWING) == 0, 2165164936Sjulian ("thread with borrowed priority returning to userland")); 2166164936Sjulian if (td->td_priority != td->td_user_pri) { 2167170293Sjeff thread_lock(td); 2168164936Sjulian td->td_priority = td->td_user_pri; 2169164936Sjulian td->td_base_pri = td->td_user_pri; 2170177005Sjeff tdq_setlowpri(TDQ_SELF(), td); 2171170293Sjeff thread_unlock(td); 2172164936Sjulian } 2173164936Sjulian} 2174164936Sjulian 2175171482Sjeff/* 2176171482Sjeff * Handle a stathz tick. This is really only relevant for timeshare 2177171482Sjeff * threads. 2178171482Sjeff */ 2179164936Sjulianvoid 2180121127Sjeffsched_clock(struct thread *td) 2181109864Sjeff{ 2182164936Sjulian struct tdq *tdq; 2183164936Sjulian struct td_sched *ts; 2184109864Sjeff 2185171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2186164936Sjulian tdq = TDQ_SELF(); 2187172409Sjeff#ifdef SMP 2188133427Sjeff /* 2189172409Sjeff * We run the long term load balancer infrequently on the first cpu. 2190172409Sjeff */ 2191172409Sjeff if (balance_tdq == tdq) { 2192172409Sjeff if (balance_ticks && --balance_ticks == 0) 2193172409Sjeff sched_balance(); 2194172409Sjeff } 2195172409Sjeff#endif 2196172409Sjeff /* 2197178277Sjeff * Save the old switch count so we have a record of the last ticks 2198178277Sjeff * activity. Initialize the new switch count based on our load. 2199178277Sjeff * If there is some activity seed it to reflect that. 2200178277Sjeff */ 2201178277Sjeff tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2202178471Sjeff tdq->tdq_switchcnt = tdq->tdq_load; 2203178277Sjeff /* 2204165766Sjeff * Advance the insert index once for each tick to ensure that all 2205165766Sjeff * threads get a chance to run. 2206133427Sjeff */ 2207165766Sjeff if (tdq->tdq_idx == tdq->tdq_ridx) { 2208165766Sjeff tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2209165766Sjeff if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2210165766Sjeff tdq->tdq_ridx = tdq->tdq_idx; 2211165766Sjeff } 2212165766Sjeff ts = td->td_sched; 2213232917Smav sched_pctcpu_update(ts, 1); 2214175104Sjeff if (td->td_pri_class & PRI_FIFO_BIT) 2215113357Sjeff return; 2216217291Sjhb if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) { 2217175104Sjeff /* 2218175104Sjeff * We used a tick; charge it to the thread so 2219175104Sjeff * that we can compute our interactivity. 2220175104Sjeff */ 2221175104Sjeff td->td_sched->ts_runtime += tickincr; 2222175104Sjeff sched_interact_update(td); 2223177009Sjeff sched_priority(td); 2224175104Sjeff } 2225239185Smav 2226113357Sjeff /* 2227239185Smav * Force a context switch if the current thread has used up a full 2228239185Smav * time slice (default is 100ms). 2229109864Sjeff */ 2230239185Smav if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) { 2231239185Smav ts->ts_slice = sched_slice; 2232239185Smav td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND; 2233239185Smav } 2234109864Sjeff} 2235109864Sjeff 2236171482Sjeff/* 2237232917Smav * Called once per hz tick. 2238171482Sjeff */ 2239171482Sjeffvoid 2240212541Smavsched_tick(int cnt) 2241171482Sjeff{ 2242171482Sjeff 2243171482Sjeff} 2244171482Sjeff 2245171482Sjeff/* 2246171482Sjeff * Return whether the current CPU has runnable tasks. Used for in-kernel 2247171482Sjeff * cooperative idle threads. 2248171482Sjeff */ 2249109864Sjeffint 2250109864Sjeffsched_runnable(void) 2251109864Sjeff{ 2252164936Sjulian struct tdq *tdq; 2253115998Sjeff int load; 2254109864Sjeff 2255115998Sjeff load = 1; 2256115998Sjeff 2257164936Sjulian tdq = TDQ_SELF(); 2258121605Sjeff if ((curthread->td_flags & TDF_IDLETD) != 0) { 2259165620Sjeff if (tdq->tdq_load > 0) 2260121605Sjeff goto out; 2261121605Sjeff } else 2262165620Sjeff if (tdq->tdq_load - 1 > 0) 2263121605Sjeff goto out; 2264115998Sjeff load = 0; 2265115998Sjeffout: 2266115998Sjeff return (load); 2267109864Sjeff} 2268109864Sjeff 2269171482Sjeff/* 2270171482Sjeff * Choose the highest priority thread to run. The thread is removed from 2271171482Sjeff * the run-queue while running however the load remains. For SMP we set 2272171482Sjeff * the tdq in the global idle bitmask if it idles here. 2273171482Sjeff */ 2274166190Sjeffstruct thread * 2275109970Sjeffsched_choose(void) 2276109970Sjeff{ 2277177435Sjeff struct thread *td; 2278164936Sjulian struct tdq *tdq; 2279109970Sjeff 2280164936Sjulian tdq = TDQ_SELF(); 2281171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2282177435Sjeff td = tdq_choose(tdq); 2283177435Sjeff if (td) { 2284177435Sjeff tdq_runq_rem(tdq, td); 2285177903Sjeff tdq->tdq_lowpri = td->td_priority; 2286177435Sjeff return (td); 2287109864Sjeff } 2288177903Sjeff tdq->tdq_lowpri = PRI_MAX_IDLE; 2289176735Sjeff return (PCPU_GET(idlethread)); 2290109864Sjeff} 2291109864Sjeff 2292171482Sjeff/* 2293171482Sjeff * Set owepreempt if necessary. Preemption never happens directly in ULE, 2294171482Sjeff * we always request it once we exit a critical section. 2295171482Sjeff */ 2296171482Sjeffstatic inline void 2297171482Sjeffsched_setpreempt(struct thread *td) 2298166190Sjeff{ 2299166190Sjeff struct thread *ctd; 2300166190Sjeff int cpri; 2301166190Sjeff int pri; 2302166190Sjeff 2303177005Sjeff THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2304177005Sjeff 2305166190Sjeff ctd = curthread; 2306166190Sjeff pri = td->td_priority; 2307166190Sjeff cpri = ctd->td_priority; 2308177005Sjeff if (pri < cpri) 2309177005Sjeff ctd->td_flags |= TDF_NEEDRESCHED; 2310166190Sjeff if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2311171482Sjeff return; 2312177005Sjeff if (!sched_shouldpreempt(pri, cpri, 0)) 2313171482Sjeff return; 2314171482Sjeff ctd->td_owepreempt = 1; 2315166190Sjeff} 2316166190Sjeff 2317171482Sjeff/* 2318177009Sjeff * Add a thread to a thread queue. Select the appropriate runq and add the 2319177009Sjeff * thread to it. This is the internal function called when the tdq is 2320177009Sjeff * predetermined. 2321171482Sjeff */ 2322109864Sjeffvoid 2323171482Sjefftdq_add(struct tdq *tdq, struct thread *td, int flags) 2324109864Sjeff{ 2325109864Sjeff 2326171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2327166190Sjeff KASSERT((td->td_inhibitors == 0), 2328166190Sjeff ("sched_add: trying to run inhibited thread")); 2329166190Sjeff KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2330166190Sjeff ("sched_add: bad thread state")); 2331172207Sjeff KASSERT(td->td_flags & TDF_INMEM, 2332172207Sjeff ("sched_add: thread swapped out")); 2333171482Sjeff 2334171482Sjeff if (td->td_priority < tdq->tdq_lowpri) 2335171482Sjeff tdq->tdq_lowpri = td->td_priority; 2336177435Sjeff tdq_runq_add(tdq, td, flags); 2337177435Sjeff tdq_load_add(tdq, td); 2338171482Sjeff} 2339171482Sjeff 2340171482Sjeff/* 2341171482Sjeff * Select the target thread queue and add a thread to it. Request 2342171482Sjeff * preemption or IPI a remote processor if required. 2343171482Sjeff */ 2344171482Sjeffvoid 2345171482Sjeffsched_add(struct thread *td, int flags) 2346171482Sjeff{ 2347171482Sjeff struct tdq *tdq; 2348171482Sjeff#ifdef SMP 2349171482Sjeff int cpu; 2350171482Sjeff#endif 2351187357Sjeff 2352187357Sjeff KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 2353187357Sjeff "prio:%d", td->td_priority, KTR_ATTR_LINKED, 2354187357Sjeff sched_tdname(curthread)); 2355187357Sjeff KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 2356187357Sjeff KTR_ATTR_LINKED, sched_tdname(td)); 2357235459Srstone SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, 2358235459Srstone flags & SRQ_PREEMPTED); 2359171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2360166108Sjeff /* 2361171482Sjeff * Recalculate the priority before we select the target cpu or 2362171482Sjeff * run-queue. 2363166108Sjeff */ 2364171482Sjeff if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2365171482Sjeff sched_priority(td); 2366171482Sjeff#ifdef SMP 2367171482Sjeff /* 2368171482Sjeff * Pick the destination cpu and if it isn't ours transfer to the 2369171482Sjeff * target cpu. 2370171482Sjeff */ 2371177435Sjeff cpu = sched_pickcpu(td, flags); 2372177435Sjeff tdq = sched_setcpu(td, cpu, flags); 2373171482Sjeff tdq_add(tdq, td, flags); 2374177009Sjeff if (cpu != PCPU_GET(cpuid)) { 2375177435Sjeff tdq_notify(tdq, td); 2376166108Sjeff return; 2377166108Sjeff } 2378171482Sjeff#else 2379171482Sjeff tdq = TDQ_SELF(); 2380171482Sjeff TDQ_LOCK(tdq); 2381171482Sjeff /* 2382171482Sjeff * Now that the thread is moving to the run-queue, set the lock 2383171482Sjeff * to the scheduler's lock. 2384171482Sjeff */ 2385171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2386171482Sjeff tdq_add(tdq, td, flags); 2387166108Sjeff#endif 2388171482Sjeff if (!(flags & SRQ_YIELDING)) 2389171482Sjeff sched_setpreempt(td); 2390109864Sjeff} 2391109864Sjeff 2392171482Sjeff/* 2393171482Sjeff * Remove a thread from a run-queue without running it. This is used 2394171482Sjeff * when we're stealing a thread from a remote queue. Otherwise all threads 2395171482Sjeff * exit by calling sched_exit_thread() and sched_throw() themselves. 2396171482Sjeff */ 2397109864Sjeffvoid 2398121127Sjeffsched_rem(struct thread *td) 2399109864Sjeff{ 2400164936Sjulian struct tdq *tdq; 2401113357Sjeff 2402187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 2403187357Sjeff "prio:%d", td->td_priority); 2404235459Srstone SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); 2405177435Sjeff tdq = TDQ_CPU(td->td_sched->ts_cpu); 2406171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2407171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2408166190Sjeff KASSERT(TD_ON_RUNQ(td), 2409164936Sjulian ("sched_rem: thread not on run queue")); 2410177435Sjeff tdq_runq_rem(tdq, td); 2411177435Sjeff tdq_load_rem(tdq, td); 2412166190Sjeff TD_SET_CAN_RUN(td); 2413176735Sjeff if (td->td_priority == tdq->tdq_lowpri) 2414176735Sjeff tdq_setlowpri(tdq, NULL); 2415109864Sjeff} 2416109864Sjeff 2417171482Sjeff/* 2418171482Sjeff * Fetch cpu utilization information. Updates on demand. 2419171482Sjeff */ 2420109864Sjefffixpt_t 2421121127Sjeffsched_pctcpu(struct thread *td) 2422109864Sjeff{ 2423109864Sjeff fixpt_t pctcpu; 2424164936Sjulian struct td_sched *ts; 2425109864Sjeff 2426109864Sjeff pctcpu = 0; 2427164936Sjulian ts = td->td_sched; 2428164936Sjulian if (ts == NULL) 2429121290Sjeff return (0); 2430109864Sjeff 2431208787Sjhb THREAD_LOCK_ASSERT(td, MA_OWNED); 2432232917Smav sched_pctcpu_update(ts, TD_IS_RUNNING(td)); 2433164936Sjulian if (ts->ts_ticks) { 2434109864Sjeff int rtick; 2435109864Sjeff 2436109864Sjeff /* How many rtick per second ? */ 2437165762Sjeff rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2438165762Sjeff pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2439109864Sjeff } 2440109864Sjeff 2441109864Sjeff return (pctcpu); 2442109864Sjeff} 2443109864Sjeff 2444176735Sjeff/* 2445176735Sjeff * Enforce affinity settings for a thread. Called after adjustments to 2446176735Sjeff * cpumask. 2447176735Sjeff */ 2448176729Sjeffvoid 2449176729Sjeffsched_affinity(struct thread *td) 2450176729Sjeff{ 2451176735Sjeff#ifdef SMP 2452176735Sjeff struct td_sched *ts; 2453176735Sjeff 2454176735Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2455176735Sjeff ts = td->td_sched; 2456176735Sjeff if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2457176735Sjeff return; 2458189787Sjeff if (TD_ON_RUNQ(td)) { 2459189787Sjeff sched_rem(td); 2460189787Sjeff sched_add(td, SRQ_BORING); 2461189787Sjeff return; 2462189787Sjeff } 2463176735Sjeff if (!TD_IS_RUNNING(td)) 2464176735Sjeff return; 2465176735Sjeff /* 2466212115Smdf * Force a switch before returning to userspace. If the 2467212115Smdf * target thread is not running locally send an ipi to force 2468212115Smdf * the issue. 2469176735Sjeff */ 2470212974Sjhb td->td_flags |= TDF_NEEDRESCHED; 2471212115Smdf if (td != curthread) 2472212115Smdf ipi_cpu(ts->ts_cpu, IPI_PREEMPT); 2473176735Sjeff#endif 2474176729Sjeff} 2475176729Sjeff 2476171482Sjeff/* 2477171482Sjeff * Bind a thread to a target cpu. 2478171482Sjeff */ 2479122038Sjeffvoid 2480122038Sjeffsched_bind(struct thread *td, int cpu) 2481122038Sjeff{ 2482164936Sjulian struct td_sched *ts; 2483122038Sjeff 2484171713Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2485208391Sjhb KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 2486164936Sjulian ts = td->td_sched; 2487166137Sjeff if (ts->ts_flags & TSF_BOUND) 2488166152Sjeff sched_unbind(td); 2489212115Smdf KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td)); 2490164936Sjulian ts->ts_flags |= TSF_BOUND; 2491166137Sjeff sched_pin(); 2492123433Sjeff if (PCPU_GET(cpuid) == cpu) 2493122038Sjeff return; 2494166137Sjeff ts->ts_cpu = cpu; 2495122038Sjeff /* When we return from mi_switch we'll be on the correct cpu. */ 2496131527Sphk mi_switch(SW_VOL, NULL); 2497122038Sjeff} 2498122038Sjeff 2499171482Sjeff/* 2500171482Sjeff * Release a bound thread. 2501171482Sjeff */ 2502122038Sjeffvoid 2503122038Sjeffsched_unbind(struct thread *td) 2504122038Sjeff{ 2505165762Sjeff struct td_sched *ts; 2506165762Sjeff 2507170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2508208391Sjhb KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 2509165762Sjeff ts = td->td_sched; 2510166137Sjeff if ((ts->ts_flags & TSF_BOUND) == 0) 2511166137Sjeff return; 2512165762Sjeff ts->ts_flags &= ~TSF_BOUND; 2513165762Sjeff sched_unpin(); 2514122038Sjeff} 2515122038Sjeff 2516109864Sjeffint 2517145256Sjkoshysched_is_bound(struct thread *td) 2518145256Sjkoshy{ 2519170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2520164936Sjulian return (td->td_sched->ts_flags & TSF_BOUND); 2521145256Sjkoshy} 2522145256Sjkoshy 2523171482Sjeff/* 2524171482Sjeff * Basic yield call. 2525171482Sjeff */ 2526159630Sdavidxuvoid 2527159630Sdavidxusched_relinquish(struct thread *td) 2528159630Sdavidxu{ 2529170293Sjeff thread_lock(td); 2530178272Sjeff mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 2531170293Sjeff thread_unlock(td); 2532159630Sdavidxu} 2533159630Sdavidxu 2534171482Sjeff/* 2535171482Sjeff * Return the total system load. 2536171482Sjeff */ 2537145256Sjkoshyint 2538125289Sjeffsched_load(void) 2539125289Sjeff{ 2540125289Sjeff#ifdef SMP 2541125289Sjeff int total; 2542125289Sjeff int i; 2543125289Sjeff 2544125289Sjeff total = 0; 2545209059Sjhb CPU_FOREACH(i) 2546176735Sjeff total += TDQ_CPU(i)->tdq_sysload; 2547125289Sjeff return (total); 2548125289Sjeff#else 2549165620Sjeff return (TDQ_SELF()->tdq_sysload); 2550125289Sjeff#endif 2551125289Sjeff} 2552125289Sjeff 2553125289Sjeffint 2554109864Sjeffsched_sizeof_proc(void) 2555109864Sjeff{ 2556109864Sjeff return (sizeof(struct proc)); 2557109864Sjeff} 2558109864Sjeff 2559109864Sjeffint 2560109864Sjeffsched_sizeof_thread(void) 2561109864Sjeff{ 2562109864Sjeff return (sizeof(struct thread) + sizeof(struct td_sched)); 2563109864Sjeff} 2564159570Sdavidxu 2565191676Sjeff#ifdef SMP 2566191676Sjeff#define TDQ_IDLESPIN(tdq) \ 2567191676Sjeff ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0) 2568191676Sjeff#else 2569191676Sjeff#define TDQ_IDLESPIN(tdq) 1 2570191676Sjeff#endif 2571191676Sjeff 2572166190Sjeff/* 2573166190Sjeff * The actual idle process. 2574166190Sjeff */ 2575166190Sjeffvoid 2576166190Sjeffsched_idletd(void *dummy) 2577166190Sjeff{ 2578166190Sjeff struct thread *td; 2579171482Sjeff struct tdq *tdq; 2580178277Sjeff int switchcnt; 2581178277Sjeff int i; 2582166190Sjeff 2583191643Sjeff mtx_assert(&Giant, MA_NOTOWNED); 2584166190Sjeff td = curthread; 2585171482Sjeff tdq = TDQ_SELF(); 2586239585Sjhb THREAD_NO_SLEEPING(); 2587171482Sjeff for (;;) { 2588171482Sjeff#ifdef SMP 2589178277Sjeff if (tdq_idled(tdq) == 0) 2590178277Sjeff continue; 2591171482Sjeff#endif 2592178277Sjeff switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2593178277Sjeff /* 2594178277Sjeff * If we're switching very frequently, spin while checking 2595178277Sjeff * for load rather than entering a low power state that 2596191643Sjeff * may require an IPI. However, don't do any busy 2597191643Sjeff * loops while on SMT machines as this simply steals 2598191643Sjeff * cycles from cores doing useful work. 2599178277Sjeff */ 2600191676Sjeff if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) { 2601178277Sjeff for (i = 0; i < sched_idlespins; i++) { 2602178277Sjeff if (tdq->tdq_load) 2603178277Sjeff break; 2604178277Sjeff cpu_spinwait(); 2605178277Sjeff } 2606178277Sjeff } 2607191643Sjeff switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2608212416Smav if (tdq->tdq_load == 0) { 2609212416Smav tdq->tdq_cpu_idle = 1; 2610212416Smav if (tdq->tdq_load == 0) { 2611212541Smav cpu_idle(switchcnt > sched_idlespinthresh * 4); 2612212416Smav tdq->tdq_switchcnt++; 2613212416Smav } 2614212416Smav tdq->tdq_cpu_idle = 0; 2615212416Smav } 2616178277Sjeff if (tdq->tdq_load) { 2617178277Sjeff thread_lock(td); 2618178277Sjeff mi_switch(SW_VOL | SWT_IDLE, NULL); 2619178277Sjeff thread_unlock(td); 2620178277Sjeff } 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(); 2638229429Sjhb PCPU_SET(switchtime, cpu_ticks()); 2639229429Sjhb 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 2697232700Sjhb#ifdef KTR 2698232700Sjhbvoid 2699232700Sjhbsched_clear_tdname(struct thread *td) 2700232700Sjhb{ 2701232700Sjhb struct td_sched *ts; 2702232700Sjhb 2703232700Sjhb ts = td->td_sched; 2704232700Sjhb ts->ts_name[0] = '\0'; 2705232700Sjhb} 2706232700Sjhb#endif 2707232700Sjhb 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 2789239185Smavstatic int 2790239185Smavsysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 2791239185Smav{ 2792239185Smav int error, new_val, period; 2793239185Smav 2794239185Smav period = 1000000 / realstathz; 2795239185Smav new_val = period * sched_slice; 2796239185Smav error = sysctl_handle_int(oidp, &new_val, 0, req); 2797239196Smav if (error != 0 || req->newptr == NULL) 2798239185Smav return (error); 2799239185Smav if (new_val <= 0) 2800239185Smav return (EINVAL); 2801239196Smav sched_slice = imax(1, (new_val + period / 2) / period); 2802239196Smav hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / 2803239196Smav realstathz); 2804239185Smav return (0); 2805239185Smav} 2806239185Smav 2807177435SjeffSYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2808171482SjeffSYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2809165762Sjeff "Scheduler name"); 2810239185SmavSYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW, 2811239185Smav NULL, 0, sysctl_kern_quantum, "I", 2812239196Smav "Quantum for timeshare threads in microseconds"); 2813171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2814239196Smav "Quantum for timeshare threads in stathz ticks"); 2815171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2816239196Smav "Interactivity score threshold"); 2817239196SmavSYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, 2818239196Smav &preempt_thresh, 0, 2819239196Smav "Maximal (lowest) priority for preemption"); 2820239196SmavSYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0, 2821239196Smav "Assign static kernel priorities to sleeping threads"); 2822239196SmavSYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0, 2823239196Smav "Number of times idle thread will spin waiting for new work"); 2824239196SmavSYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, 2825239196Smav &sched_idlespinthresh, 0, 2826239196Smav "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, 2834239185Smav "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, 2838239196Smav "Minimum load on remote CPU before we'll steal"); 2839184439SivorasSYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING | 2840239185Smav 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