sched_ule.c revision 232454
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 232454 2012-03-03 11:50:48Z mav $"); 40116182Sobrien 41147565Speter#include "opt_hwpmc_hooks.h" 42179297Sjb#include "opt_kdtrace.h" 43147565Speter#include "opt_sched.h" 44134649Sscottl 45109864Sjeff#include <sys/param.h> 46109864Sjeff#include <sys/systm.h> 47131929Smarcel#include <sys/kdb.h> 48109864Sjeff#include <sys/kernel.h> 49109864Sjeff#include <sys/ktr.h> 50109864Sjeff#include <sys/lock.h> 51109864Sjeff#include <sys/mutex.h> 52109864Sjeff#include <sys/proc.h> 53112966Sjeff#include <sys/resource.h> 54122038Sjeff#include <sys/resourcevar.h> 55109864Sjeff#include <sys/sched.h> 56109864Sjeff#include <sys/smp.h> 57109864Sjeff#include <sys/sx.h> 58109864Sjeff#include <sys/sysctl.h> 59109864Sjeff#include <sys/sysproto.h> 60139453Sjhb#include <sys/turnstile.h> 61161599Sdavidxu#include <sys/umtx.h> 62109864Sjeff#include <sys/vmmeter.h> 63176735Sjeff#include <sys/cpuset.h> 64184439Sivoras#include <sys/sbuf.h> 65109864Sjeff 66145256Sjkoshy#ifdef HWPMC_HOOKS 67145256Sjkoshy#include <sys/pmckern.h> 68145256Sjkoshy#endif 69145256Sjkoshy 70179297Sjb#ifdef KDTRACE_HOOKS 71179297Sjb#include <sys/dtrace_bsd.h> 72179297Sjbint dtrace_vtime_active; 73179297Sjbdtrace_vtime_switch_func_t dtrace_vtime_switch_func; 74179297Sjb#endif 75179297Sjb 76109864Sjeff#include <machine/cpu.h> 77121790Sjeff#include <machine/smp.h> 78109864Sjeff 79226057Smarius#if defined(__powerpc__) && defined(E500) 80172345Sjeff#error "This architecture is not currently compatible with ULE" 81166190Sjeff#endif 82166190Sjeff 83171482Sjeff#define KTR_ULE 0 84166137Sjeff 85187679Sjeff#define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX))) 86187679Sjeff#define TDQ_NAME_LEN (sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU))) 87224221Sattilio#define TDQ_LOADNAME_LEN (sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load")) 88187357Sjeff 89166137Sjeff/* 90171482Sjeff * Thread scheduler specific section. All fields are protected 91171482Sjeff * by the thread lock. 92146954Sjeff */ 93164936Sjulianstruct td_sched { 94171482Sjeff struct runq *ts_runq; /* Run-queue we're queued on. */ 95171482Sjeff short ts_flags; /* TSF_* flags. */ 96164936Sjulian u_char ts_cpu; /* CPU that we have affinity for. */ 97177009Sjeff int ts_rltick; /* Real last tick, for affinity. */ 98171482Sjeff int ts_slice; /* Ticks of slice remaining. */ 99171482Sjeff u_int ts_slptime; /* Number of ticks we vol. slept */ 100171482Sjeff u_int ts_runtime; /* Number of ticks we were running */ 101164936Sjulian int ts_ltick; /* Last tick that we were running on */ 102199764Sivoras int ts_incrtick; /* Last tick that we incremented 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 192109864Sjeff/* 193165762Sjeff * tickincr: Converts a stathz tick into a hz domain scaled by 194165762Sjeff * the shift factor. Without the shift the error rate 195165762Sjeff * due to rounding would be unacceptably high. 196165762Sjeff * realstathz: stathz is sometimes 0 and run off of hz. 197165762Sjeff * sched_slice: Runtime of each thread before rescheduling. 198171482Sjeff * preempt_thresh: Priority threshold for preemption and remote IPIs. 199109864Sjeff */ 200165762Sjeffstatic int sched_interact = SCHED_INTERACT_THRESH; 201165762Sjeffstatic int realstathz; 202165762Sjeffstatic int tickincr; 203177009Sjeffstatic int sched_slice = 1; 204172345Sjeff#ifdef PREEMPTION 205172345Sjeff#ifdef FULL_PREEMPTION 206172345Sjeffstatic int preempt_thresh = PRI_MAX_IDLE; 207172345Sjeff#else 208171482Sjeffstatic int preempt_thresh = PRI_MIN_KERN; 209172345Sjeff#endif 210172345Sjeff#else 211172345Sjeffstatic int preempt_thresh = 0; 212172345Sjeff#endif 213217351Sjhbstatic int static_boost = PRI_MIN_BATCH; 214178277Sjeffstatic int sched_idlespins = 10000; 215212541Smavstatic int sched_idlespinthresh = 16; 216109864Sjeff 217109864Sjeff/* 218171482Sjeff * tdq - per processor runqs and statistics. All fields are protected by the 219171482Sjeff * tdq_lock. The load and lowpri may be accessed without to avoid excess 220171482Sjeff * locking in sched_pickcpu(); 221109864Sjeff */ 222164936Sjulianstruct tdq { 223177009Sjeff /* Ordered to improve efficiency of cpu_search() and switch(). */ 224177009Sjeff struct mtx tdq_lock; /* run queue lock. */ 225176735Sjeff struct cpu_group *tdq_cg; /* Pointer to cpu topology. */ 226178277Sjeff volatile int tdq_load; /* Aggregate load. */ 227212416Smav volatile int tdq_cpu_idle; /* cpu_idle() is active. */ 228176735Sjeff int tdq_sysload; /* For loadavg, !ITHD load. */ 229177009Sjeff int tdq_transferable; /* Transferable thread count. */ 230178277Sjeff short tdq_switchcnt; /* Switches this tick. */ 231178277Sjeff short tdq_oldswitchcnt; /* Switches last tick. */ 232177009Sjeff u_char tdq_lowpri; /* Lowest priority thread. */ 233177009Sjeff u_char tdq_ipipending; /* IPI pending. */ 234166557Sjeff u_char tdq_idx; /* Current insert index. */ 235166557Sjeff u_char tdq_ridx; /* Current removal index. */ 236177009Sjeff struct runq tdq_realtime; /* real-time run queue. */ 237177009Sjeff struct runq tdq_timeshare; /* timeshare run queue. */ 238177009Sjeff struct runq tdq_idle; /* Queue of IDLE threads. */ 239187357Sjeff char tdq_name[TDQ_NAME_LEN]; 240187357Sjeff#ifdef KTR 241187357Sjeff char tdq_loadname[TDQ_LOADNAME_LEN]; 242187357Sjeff#endif 243171482Sjeff} __aligned(64); 244109864Sjeff 245178277Sjeff/* Idle thread states and config. */ 246178277Sjeff#define TDQ_RUNNING 1 247178277Sjeff#define TDQ_IDLE 2 248166108Sjeff 249123433Sjeff#ifdef SMP 250184439Sivorasstruct cpu_group *cpu_top; /* CPU topology */ 251123433Sjeff 252176735Sjeff#define SCHED_AFFINITY_DEFAULT (max(1, hz / 1000)) 253176735Sjeff#define SCHED_AFFINITY(ts, t) ((ts)->ts_rltick > ticks - ((t) * affinity)) 254166108Sjeff 255123433Sjeff/* 256166108Sjeff * Run-time tunables. 257166108Sjeff */ 258171506Sjeffstatic int rebalance = 1; 259172409Sjeffstatic int balance_interval = 128; /* Default set in sched_initticks(). */ 260166108Sjeffstatic int affinity; 261171506Sjeffstatic int steal_idle = 1; 262171506Sjeffstatic int steal_thresh = 2; 263166108Sjeff 264166108Sjeff/* 265165620Sjeff * One thread queue per processor. 266109864Sjeff */ 267164936Sjulianstatic struct tdq tdq_cpu[MAXCPU]; 268172409Sjeffstatic struct tdq *balance_tdq; 269172409Sjeffstatic int balance_ticks; 270232207Smavstatic DPCPU_DEFINE(uint32_t, randomval); 271129982Sjeff 272164936Sjulian#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 273164936Sjulian#define TDQ_CPU(x) (&tdq_cpu[(x)]) 274171713Sjeff#define TDQ_ID(x) ((int)((x) - tdq_cpu)) 275123433Sjeff#else /* !SMP */ 276164936Sjulianstatic struct tdq tdq_cpu; 277129982Sjeff 278170315Sjeff#define TDQ_ID(x) (0) 279164936Sjulian#define TDQ_SELF() (&tdq_cpu) 280164936Sjulian#define TDQ_CPU(x) (&tdq_cpu) 281110028Sjeff#endif 282109864Sjeff 283171482Sjeff#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 284171482Sjeff#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 285171482Sjeff#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 286171482Sjeff#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 287176735Sjeff#define TDQ_LOCKPTR(t) (&(t)->tdq_lock) 288171482Sjeff 289163709Sjbstatic void sched_priority(struct thread *); 290146954Sjeffstatic void sched_thread_priority(struct thread *, u_char); 291163709Sjbstatic int sched_interact_score(struct thread *); 292163709Sjbstatic void sched_interact_update(struct thread *); 293163709Sjbstatic void sched_interact_fork(struct thread *); 294164936Sjulianstatic void sched_pctcpu_update(struct td_sched *); 295109864Sjeff 296110267Sjeff/* Operations on per processor queues */ 297177435Sjeffstatic struct thread *tdq_choose(struct tdq *); 298164936Sjulianstatic void tdq_setup(struct tdq *); 299177435Sjeffstatic void tdq_load_add(struct tdq *, struct thread *); 300177435Sjeffstatic void tdq_load_rem(struct tdq *, struct thread *); 301177435Sjeffstatic __inline void tdq_runq_add(struct tdq *, struct thread *, int); 302177435Sjeffstatic __inline void tdq_runq_rem(struct tdq *, struct thread *); 303177005Sjeffstatic inline int sched_shouldpreempt(int, int, int); 304164936Sjulianvoid tdq_print(int cpu); 305165762Sjeffstatic void runq_print(struct runq *rq); 306171482Sjeffstatic void tdq_add(struct tdq *, struct thread *, int); 307110267Sjeff#ifdef SMP 308176735Sjeffstatic int tdq_move(struct tdq *, struct tdq *); 309171482Sjeffstatic int tdq_idled(struct tdq *); 310177435Sjeffstatic void tdq_notify(struct tdq *, struct thread *); 311177435Sjeffstatic struct thread *tdq_steal(struct tdq *, int); 312177435Sjeffstatic struct thread *runq_steal(struct runq *, int); 313177435Sjeffstatic int sched_pickcpu(struct thread *, int); 314172409Sjeffstatic void sched_balance(void); 315176735Sjeffstatic int sched_balance_pair(struct tdq *, struct tdq *); 316177435Sjeffstatic inline struct tdq *sched_setcpu(struct thread *, int, int); 317171482Sjeffstatic inline void thread_unblock_switch(struct thread *, struct mtx *); 318171713Sjeffstatic struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 319184439Sivorasstatic int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS); 320184439Sivorasstatic int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, 321184439Sivoras struct cpu_group *cg, int indent); 322121790Sjeff#endif 323110028Sjeff 324165762Sjeffstatic void sched_setup(void *dummy); 325177253SrwatsonSYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); 326165762Sjeff 327165762Sjeffstatic void sched_initticks(void *dummy); 328177253SrwatsonSYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, 329177253Srwatson NULL); 330165762Sjeff 331171482Sjeff/* 332171482Sjeff * Print the threads waiting on a run-queue. 333171482Sjeff */ 334165762Sjeffstatic void 335165762Sjeffrunq_print(struct runq *rq) 336165762Sjeff{ 337165762Sjeff struct rqhead *rqh; 338177435Sjeff struct thread *td; 339165762Sjeff int pri; 340165762Sjeff int j; 341165762Sjeff int i; 342165762Sjeff 343165762Sjeff for (i = 0; i < RQB_LEN; i++) { 344165762Sjeff printf("\t\trunq bits %d 0x%zx\n", 345165762Sjeff i, rq->rq_status.rqb_bits[i]); 346165762Sjeff for (j = 0; j < RQB_BPW; j++) 347165762Sjeff if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 348165762Sjeff pri = j + (i << RQB_L2BPW); 349165762Sjeff rqh = &rq->rq_queues[pri]; 350177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) { 351165762Sjeff printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 352177435Sjeff td, td->td_name, td->td_priority, 353177435Sjeff td->td_rqindex, pri); 354165762Sjeff } 355165762Sjeff } 356165762Sjeff } 357165762Sjeff} 358165762Sjeff 359171482Sjeff/* 360171482Sjeff * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 361171482Sjeff */ 362113357Sjeffvoid 363164936Sjuliantdq_print(int cpu) 364110267Sjeff{ 365164936Sjulian struct tdq *tdq; 366112994Sjeff 367164936Sjulian tdq = TDQ_CPU(cpu); 368112994Sjeff 369171713Sjeff printf("tdq %d:\n", TDQ_ID(tdq)); 370176735Sjeff printf("\tlock %p\n", TDQ_LOCKPTR(tdq)); 371176735Sjeff printf("\tLock name: %s\n", tdq->tdq_name); 372165620Sjeff printf("\tload: %d\n", tdq->tdq_load); 373178277Sjeff printf("\tswitch cnt: %d\n", tdq->tdq_switchcnt); 374178277Sjeff printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt); 375171482Sjeff printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 376165766Sjeff printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 377178277Sjeff printf("\tload transferable: %d\n", tdq->tdq_transferable); 378178277Sjeff printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 379165762Sjeff printf("\trealtime runq:\n"); 380165762Sjeff runq_print(&tdq->tdq_realtime); 381165762Sjeff printf("\ttimeshare runq:\n"); 382165762Sjeff runq_print(&tdq->tdq_timeshare); 383165762Sjeff printf("\tidle runq:\n"); 384165762Sjeff runq_print(&tdq->tdq_idle); 385113357Sjeff} 386112994Sjeff 387177005Sjeffstatic inline int 388177005Sjeffsched_shouldpreempt(int pri, int cpri, int remote) 389177005Sjeff{ 390177005Sjeff /* 391177005Sjeff * If the new priority is not better than the current priority there is 392177005Sjeff * nothing to do. 393177005Sjeff */ 394177005Sjeff if (pri >= cpri) 395177005Sjeff return (0); 396177005Sjeff /* 397177005Sjeff * Always preempt idle. 398177005Sjeff */ 399177005Sjeff if (cpri >= PRI_MIN_IDLE) 400177005Sjeff return (1); 401177005Sjeff /* 402177005Sjeff * If preemption is disabled don't preempt others. 403177005Sjeff */ 404177005Sjeff if (preempt_thresh == 0) 405177005Sjeff return (0); 406177005Sjeff /* 407177005Sjeff * Preempt if we exceed the threshold. 408177005Sjeff */ 409177005Sjeff if (pri <= preempt_thresh) 410177005Sjeff return (1); 411177005Sjeff /* 412217351Sjhb * If we're interactive or better and there is non-interactive 413217351Sjhb * or worse running preempt only remote processors. 414177005Sjeff */ 415217351Sjhb if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT) 416177005Sjeff return (1); 417177005Sjeff return (0); 418177005Sjeff} 419177005Sjeff 420171482Sjeff/* 421171482Sjeff * Add a thread to the actual run-queue. Keeps transferable counts up to 422171482Sjeff * date with what is actually on the run-queue. Selects the correct 423171482Sjeff * queue position for timeshare threads. 424171482Sjeff */ 425122744Sjeffstatic __inline void 426177435Sjefftdq_runq_add(struct tdq *tdq, struct thread *td, int flags) 427122744Sjeff{ 428177435Sjeff struct td_sched *ts; 429177042Sjeff u_char pri; 430177042Sjeff 431171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 432177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 433177009Sjeff 434177435Sjeff pri = td->td_priority; 435177435Sjeff ts = td->td_sched; 436177435Sjeff TD_SET_RUNQ(td); 437177435Sjeff if (THREAD_CAN_MIGRATE(td)) { 438165620Sjeff tdq->tdq_transferable++; 439164936Sjulian ts->ts_flags |= TSF_XFERABLE; 440123433Sjeff } 441217351Sjhb if (pri < PRI_MIN_BATCH) { 442177042Sjeff ts->ts_runq = &tdq->tdq_realtime; 443217351Sjhb } else if (pri <= PRI_MAX_BATCH) { 444177042Sjeff ts->ts_runq = &tdq->tdq_timeshare; 445217351Sjhb KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH, 446165762Sjeff ("Invalid priority %d on timeshare runq", pri)); 447165762Sjeff /* 448165762Sjeff * This queue contains only priorities between MIN and MAX 449165762Sjeff * realtime. Use the whole queue to represent these values. 450165762Sjeff */ 451171713Sjeff if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 452228718Savg pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE; 453165762Sjeff pri = (pri + tdq->tdq_idx) % RQ_NQS; 454165766Sjeff /* 455165766Sjeff * This effectively shortens the queue by one so we 456165766Sjeff * can have a one slot difference between idx and 457165766Sjeff * ridx while we wait for threads to drain. 458165766Sjeff */ 459165766Sjeff if (tdq->tdq_ridx != tdq->tdq_idx && 460165766Sjeff pri == tdq->tdq_ridx) 461167664Sjeff pri = (unsigned char)(pri - 1) % RQ_NQS; 462165762Sjeff } else 463165766Sjeff pri = tdq->tdq_ridx; 464177435Sjeff runq_add_pri(ts->ts_runq, td, pri, flags); 465177042Sjeff return; 466165762Sjeff } else 467177009Sjeff ts->ts_runq = &tdq->tdq_idle; 468177435Sjeff runq_add(ts->ts_runq, td, flags); 469177009Sjeff} 470177009Sjeff 471171482Sjeff/* 472171482Sjeff * Remove a thread from a run-queue. This typically happens when a thread 473171482Sjeff * is selected to run. Running threads are not on the queue and the 474171482Sjeff * transferable count does not reflect them. 475171482Sjeff */ 476122744Sjeffstatic __inline void 477177435Sjefftdq_runq_rem(struct tdq *tdq, struct thread *td) 478122744Sjeff{ 479177435Sjeff struct td_sched *ts; 480177435Sjeff 481177435Sjeff ts = td->td_sched; 482171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 483171482Sjeff KASSERT(ts->ts_runq != NULL, 484177435Sjeff ("tdq_runq_remove: thread %p null ts_runq", td)); 485164936Sjulian if (ts->ts_flags & TSF_XFERABLE) { 486165620Sjeff tdq->tdq_transferable--; 487164936Sjulian ts->ts_flags &= ~TSF_XFERABLE; 488123433Sjeff } 489165766Sjeff if (ts->ts_runq == &tdq->tdq_timeshare) { 490165766Sjeff if (tdq->tdq_idx != tdq->tdq_ridx) 491177435Sjeff runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx); 492165766Sjeff else 493177435Sjeff runq_remove_idx(ts->ts_runq, td, NULL); 494165766Sjeff } else 495177435Sjeff runq_remove(ts->ts_runq, td); 496122744Sjeff} 497122744Sjeff 498171482Sjeff/* 499171482Sjeff * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 500171482Sjeff * for this thread to the referenced thread queue. 501171482Sjeff */ 502113357Sjeffstatic void 503177435Sjefftdq_load_add(struct tdq *tdq, struct thread *td) 504113357Sjeff{ 505171482Sjeff 506171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 507177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 508177902Sjeff 509165620Sjeff tdq->tdq_load++; 510198854Sattilio if ((td->td_flags & TDF_NOLOAD) == 0) 511177902Sjeff tdq->tdq_sysload++; 512187357Sjeff KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 513110267Sjeff} 514113357Sjeff 515171482Sjeff/* 516171482Sjeff * Remove the load from a thread that is transitioning to a sleep state or 517171482Sjeff * exiting. 518171482Sjeff */ 519112994Sjeffstatic void 520177435Sjefftdq_load_rem(struct tdq *tdq, struct thread *td) 521110267Sjeff{ 522171482Sjeff 523177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 524171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 525171482Sjeff KASSERT(tdq->tdq_load != 0, 526171713Sjeff ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 527177902Sjeff 528165620Sjeff tdq->tdq_load--; 529198854Sattilio if ((td->td_flags & TDF_NOLOAD) == 0) 530177902Sjeff tdq->tdq_sysload--; 531187357Sjeff KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load); 532110267Sjeff} 533110267Sjeff 534176735Sjeff/* 535176735Sjeff * Set lowpri to its exact value by searching the run-queue and 536176735Sjeff * evaluating curthread. curthread may be passed as an optimization. 537176735Sjeff */ 538176735Sjeffstatic void 539176735Sjefftdq_setlowpri(struct tdq *tdq, struct thread *ctd) 540176735Sjeff{ 541176735Sjeff struct thread *td; 542176735Sjeff 543176735Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 544176735Sjeff if (ctd == NULL) 545176735Sjeff ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread; 546177435Sjeff td = tdq_choose(tdq); 547177435Sjeff if (td == NULL || td->td_priority > ctd->td_priority) 548176735Sjeff tdq->tdq_lowpri = ctd->td_priority; 549176735Sjeff else 550176735Sjeff tdq->tdq_lowpri = td->td_priority; 551176735Sjeff} 552176735Sjeff 553113357Sjeff#ifdef SMP 554176735Sjeffstruct cpu_search { 555194779Sjeff cpuset_t cs_mask; 556232207Smav u_int cs_prefer; 557232207Smav int cs_pri; /* Min priority for low. */ 558232207Smav int cs_limit; /* Max load for low, min load for high. */ 559232207Smav int cs_cpu; 560232207Smav int cs_load; 561176735Sjeff}; 562176735Sjeff 563176735Sjeff#define CPU_SEARCH_LOWEST 0x1 564176735Sjeff#define CPU_SEARCH_HIGHEST 0x2 565176735Sjeff#define CPU_SEARCH_BOTH (CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST) 566176735Sjeff 567194779Sjeff#define CPUSET_FOREACH(cpu, mask) \ 568194779Sjeff for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++) \ 569222813Sattilio if (CPU_ISSET(cpu, &mask)) 570176735Sjeff 571232207Smavstatic __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low, 572176735Sjeff struct cpu_search *high, const int match); 573232207Smavint cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low); 574232207Smavint cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high); 575232207Smavint cpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 576176735Sjeff struct cpu_search *high); 577176735Sjeff 578116069Sjeff/* 579176735Sjeff * Search the tree of cpu_groups for the lowest or highest loaded cpu 580176735Sjeff * according to the match argument. This routine actually compares the 581176735Sjeff * load on all paths through the tree and finds the least loaded cpu on 582176735Sjeff * the least loaded path, which may differ from the least loaded cpu in 583176735Sjeff * the system. This balances work among caches and busses. 584116069Sjeff * 585176735Sjeff * This inline is instantiated in three forms below using constants for the 586176735Sjeff * match argument. It is reduced to the minimum set for each case. It is 587176735Sjeff * also recursive to the depth of the tree. 588116069Sjeff */ 589177169Sjhbstatic __inline int 590232207Smavcpu_search(const struct cpu_group *cg, struct cpu_search *low, 591176735Sjeff struct cpu_search *high, const int match) 592176735Sjeff{ 593232207Smav struct cpu_search lgroup; 594232207Smav struct cpu_search hgroup; 595232207Smav cpuset_t cpumask; 596232207Smav struct cpu_group *child; 597232207Smav struct tdq *tdq; 598232207Smav int cpu, i, hload, lload, load, total, rnd; 599176735Sjeff 600176735Sjeff total = 0; 601232207Smav cpumask = cg->cg_mask; 602232207Smav if (match & CPU_SEARCH_LOWEST) { 603232207Smav lload = INT_MAX; 604232207Smav low->cs_load = INT_MAX; 605232207Smav lgroup = *low; 606232207Smav } 607232207Smav if (match & CPU_SEARCH_HIGHEST) { 608232207Smav hload = -1; 609232207Smav high->cs_load = -1; 610232207Smav hgroup = *high; 611232207Smav } 612176735Sjeff 613232207Smav /* Iterate through the child CPU groups and then remaining CPUs. */ 614232207Smav for (i = 0, cpu = 0; i <= cg->cg_children; ) { 615232207Smav if (i >= cg->cg_children) { 616232207Smav while (cpu <= mp_maxid && !CPU_ISSET(cpu, &cpumask)) 617232207Smav cpu++; 618232207Smav if (cpu > mp_maxid) 619232207Smav break; 620232207Smav child = NULL; 621232207Smav } else 622176735Sjeff child = &cg->cg_child[i]; 623232207Smav 624232207Smav if (child) { /* Handle child CPU group. */ 625232207Smav CPU_NAND(&cpumask, &child->cg_mask); 626176735Sjeff switch (match) { 627176735Sjeff case CPU_SEARCH_LOWEST: 628176735Sjeff load = cpu_search_lowest(child, &lgroup); 629176735Sjeff break; 630176735Sjeff case CPU_SEARCH_HIGHEST: 631176735Sjeff load = cpu_search_highest(child, &hgroup); 632176735Sjeff break; 633176735Sjeff case CPU_SEARCH_BOTH: 634176735Sjeff load = cpu_search_both(child, &lgroup, &hgroup); 635176735Sjeff break; 636176735Sjeff } 637232207Smav } else { /* Handle child CPU. */ 638232207Smav tdq = TDQ_CPU(cpu); 639232207Smav load = tdq->tdq_load * 256; 640232207Smav rnd = DPCPU_SET(randomval, 641232207Smav DPCPU_GET(randomval) * 69069 + 5) >> 26; 642232207Smav if (match & CPU_SEARCH_LOWEST) { 643232207Smav if (cpu == low->cs_prefer) 644232207Smav load -= 64; 645232207Smav /* If that CPU is allowed and get data. */ 646232207Smav if (CPU_ISSET(cpu, &lgroup.cs_mask) && 647232207Smav tdq->tdq_lowpri > lgroup.cs_pri && 648232207Smav tdq->tdq_load <= lgroup.cs_limit) { 649232207Smav lgroup.cs_cpu = cpu; 650232207Smav lgroup.cs_load = load - rnd; 651176735Sjeff } 652232207Smav } 653232207Smav if (match & CPU_SEARCH_HIGHEST) 654232207Smav if (CPU_ISSET(cpu, &hgroup.cs_mask) && 655232207Smav tdq->tdq_load >= hgroup.cs_limit && 656232207Smav tdq->tdq_transferable) { 657232207Smav hgroup.cs_cpu = cpu; 658232207Smav hgroup.cs_load = load - rnd; 659176735Sjeff } 660176735Sjeff } 661232207Smav total += load; 662176735Sjeff 663232207Smav /* We have info about child item. Compare it. */ 664232207Smav if (match & CPU_SEARCH_LOWEST) { 665232454Smav if (lgroup.cs_load != INT_MAX && 666232454Smav (load < lload || 667232454Smav (load == lload && lgroup.cs_load < low->cs_load))) { 668232207Smav lload = load; 669232207Smav low->cs_cpu = lgroup.cs_cpu; 670232207Smav low->cs_load = lgroup.cs_load; 671232207Smav } 672232207Smav } 673232207Smav if (match & CPU_SEARCH_HIGHEST) 674232454Smav if (hgroup.cs_load != -1 && 675232454Smav (load > hload || 676232454Smav (load == hload && hgroup.cs_load > high->cs_load))) { 677232207Smav hload = load; 678232207Smav high->cs_cpu = hgroup.cs_cpu; 679232207Smav high->cs_load = hgroup.cs_load; 680232207Smav } 681232207Smav if (child) 682232207Smav i++; 683232207Smav else 684232207Smav cpu++; 685176735Sjeff } 686176735Sjeff return (total); 687176735Sjeff} 688176735Sjeff 689176735Sjeff/* 690176735Sjeff * cpu_search instantiations must pass constants to maintain the inline 691176735Sjeff * optimization. 692176735Sjeff */ 693176735Sjeffint 694232207Smavcpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low) 695176735Sjeff{ 696176735Sjeff return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST); 697176735Sjeff} 698176735Sjeff 699176735Sjeffint 700232207Smavcpu_search_highest(const struct cpu_group *cg, struct cpu_search *high) 701176735Sjeff{ 702176735Sjeff return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST); 703176735Sjeff} 704176735Sjeff 705176735Sjeffint 706232207Smavcpu_search_both(const struct cpu_group *cg, struct cpu_search *low, 707176735Sjeff struct cpu_search *high) 708176735Sjeff{ 709176735Sjeff return cpu_search(cg, low, high, CPU_SEARCH_BOTH); 710176735Sjeff} 711176735Sjeff 712176735Sjeff/* 713176735Sjeff * Find the cpu with the least load via the least loaded path that has a 714176735Sjeff * lowpri greater than pri pri. A pri of -1 indicates any priority is 715176735Sjeff * acceptable. 716176735Sjeff */ 717176735Sjeffstatic inline int 718232207Smavsched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload, 719232207Smav int prefer) 720176735Sjeff{ 721176735Sjeff struct cpu_search low; 722176735Sjeff 723176735Sjeff low.cs_cpu = -1; 724232207Smav low.cs_prefer = prefer; 725176735Sjeff low.cs_mask = mask; 726232207Smav low.cs_pri = pri; 727232207Smav low.cs_limit = maxload; 728176735Sjeff cpu_search_lowest(cg, &low); 729176735Sjeff return low.cs_cpu; 730176735Sjeff} 731176735Sjeff 732176735Sjeff/* 733176735Sjeff * Find the cpu with the highest load via the highest loaded path. 734176735Sjeff */ 735176735Sjeffstatic inline int 736232207Smavsched_highest(const struct cpu_group *cg, cpuset_t mask, int minload) 737176735Sjeff{ 738176735Sjeff struct cpu_search high; 739176735Sjeff 740176735Sjeff high.cs_cpu = -1; 741176735Sjeff high.cs_mask = mask; 742176735Sjeff high.cs_limit = minload; 743176735Sjeff cpu_search_highest(cg, &high); 744176735Sjeff return high.cs_cpu; 745176735Sjeff} 746176735Sjeff 747176735Sjeff/* 748176735Sjeff * Simultaneously find the highest and lowest loaded cpu reachable via 749176735Sjeff * cg. 750176735Sjeff */ 751232207Smavstatic inline void 752232207Smavsched_both(const struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu) 753176735Sjeff{ 754176735Sjeff struct cpu_search high; 755176735Sjeff struct cpu_search low; 756176735Sjeff 757176735Sjeff low.cs_cpu = -1; 758232207Smav low.cs_prefer = -1; 759232207Smav low.cs_pri = -1; 760232207Smav low.cs_limit = INT_MAX; 761176735Sjeff low.cs_mask = mask; 762176735Sjeff high.cs_cpu = -1; 763176735Sjeff high.cs_limit = -1; 764176735Sjeff high.cs_mask = mask; 765176735Sjeff cpu_search_both(cg, &low, &high); 766176735Sjeff *lowcpu = low.cs_cpu; 767176735Sjeff *highcpu = high.cs_cpu; 768176735Sjeff return; 769176735Sjeff} 770176735Sjeff 771121790Sjeffstatic void 772176735Sjeffsched_balance_group(struct cpu_group *cg) 773116069Sjeff{ 774232207Smav cpuset_t hmask, lmask; 775232207Smav int high, low, anylow; 776123487Sjeff 777232207Smav CPU_FILL(&hmask); 778176735Sjeff for (;;) { 779232207Smav high = sched_highest(cg, hmask, 1); 780232207Smav /* Stop if there is no more CPU with transferrable threads. */ 781232207Smav if (high == -1) 782176735Sjeff break; 783232207Smav CPU_CLR(high, &hmask); 784232207Smav CPU_COPY(&hmask, &lmask); 785232207Smav /* Stop if there is no more CPU left for low. */ 786232207Smav if (CPU_EMPTY(&lmask)) 787176735Sjeff break; 788232207Smav anylow = 1; 789232207Smavnextlow: 790232207Smav low = sched_lowest(cg, lmask, -1, 791232207Smav TDQ_CPU(high)->tdq_load - 1, high); 792232207Smav /* Stop if we looked well and found no less loaded CPU. */ 793232207Smav if (anylow && low == -1) 794232207Smav break; 795232207Smav /* Go to next high if we found no less loaded CPU. */ 796232207Smav if (low == -1) 797232207Smav continue; 798232207Smav /* Transfer thread from high to low. */ 799232207Smav if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) { 800232207Smav /* CPU that got thread can no longer be a donor. */ 801232207Smav CPU_CLR(low, &hmask); 802232207Smav } else { 803232207Smav /* 804232207Smav * If failed, then there is no threads on high 805232207Smav * that can run on this low. Drop low from low 806232207Smav * mask and look for different one. 807232207Smav */ 808232207Smav CPU_CLR(low, &lmask); 809232207Smav anylow = 0; 810232207Smav goto nextlow; 811232207Smav } 812123487Sjeff } 813123487Sjeff} 814123487Sjeff 815123487Sjeffstatic void 816201148Sedsched_balance(void) 817123487Sjeff{ 818172409Sjeff struct tdq *tdq; 819123487Sjeff 820172409Sjeff /* 821172409Sjeff * Select a random time between .5 * balance_interval and 822172409Sjeff * 1.5 * balance_interval. 823172409Sjeff */ 824176735Sjeff balance_ticks = max(balance_interval / 2, 1); 825176735Sjeff balance_ticks += random() % balance_interval; 826171482Sjeff if (smp_started == 0 || rebalance == 0) 827171482Sjeff return; 828172409Sjeff tdq = TDQ_SELF(); 829172409Sjeff TDQ_UNLOCK(tdq); 830176735Sjeff sched_balance_group(cpu_top); 831172409Sjeff TDQ_LOCK(tdq); 832123487Sjeff} 833123487Sjeff 834171482Sjeff/* 835171482Sjeff * Lock two thread queues using their address to maintain lock order. 836171482Sjeff */ 837123487Sjeffstatic void 838171482Sjefftdq_lock_pair(struct tdq *one, struct tdq *two) 839171482Sjeff{ 840171482Sjeff if (one < two) { 841171482Sjeff TDQ_LOCK(one); 842171482Sjeff TDQ_LOCK_FLAGS(two, MTX_DUPOK); 843171482Sjeff } else { 844171482Sjeff TDQ_LOCK(two); 845171482Sjeff TDQ_LOCK_FLAGS(one, MTX_DUPOK); 846171482Sjeff } 847171482Sjeff} 848171482Sjeff 849171482Sjeff/* 850172409Sjeff * Unlock two thread queues. Order is not important here. 851172409Sjeff */ 852172409Sjeffstatic void 853172409Sjefftdq_unlock_pair(struct tdq *one, struct tdq *two) 854172409Sjeff{ 855172409Sjeff TDQ_UNLOCK(one); 856172409Sjeff TDQ_UNLOCK(two); 857172409Sjeff} 858172409Sjeff 859172409Sjeff/* 860171482Sjeff * Transfer load between two imbalanced thread queues. 861171482Sjeff */ 862176735Sjeffstatic int 863164936Sjuliansched_balance_pair(struct tdq *high, struct tdq *low) 864123487Sjeff{ 865176735Sjeff int moved; 866226057Smarius int cpu; 867116069Sjeff 868171482Sjeff tdq_lock_pair(high, low); 869176735Sjeff moved = 0; 870116069Sjeff /* 871122744Sjeff * Determine what the imbalance is and then adjust that to how many 872165620Sjeff * threads we actually have to give up (transferable). 873122744Sjeff */ 874232207Smav if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load && 875232207Smav (moved = tdq_move(high, low)) > 0) { 876172293Sjeff /* 877226057Smarius * In case the target isn't the current cpu IPI it to force a 878226057Smarius * reschedule with the new workload. 879172293Sjeff */ 880226057Smarius cpu = TDQ_ID(low); 881226057Smarius sched_pin(); 882226057Smarius if (cpu != PCPU_GET(cpuid)) 883226057Smarius ipi_cpu(cpu, IPI_PREEMPT); 884226057Smarius sched_unpin(); 885171482Sjeff } 886172409Sjeff tdq_unlock_pair(high, low); 887176735Sjeff return (moved); 888116069Sjeff} 889116069Sjeff 890171482Sjeff/* 891171482Sjeff * Move a thread from one thread queue to another. 892171482Sjeff */ 893176735Sjeffstatic int 894171482Sjefftdq_move(struct tdq *from, struct tdq *to) 895116069Sjeff{ 896171482Sjeff struct td_sched *ts; 897171482Sjeff struct thread *td; 898164936Sjulian struct tdq *tdq; 899171482Sjeff int cpu; 900116069Sjeff 901172409Sjeff TDQ_LOCK_ASSERT(from, MA_OWNED); 902172409Sjeff TDQ_LOCK_ASSERT(to, MA_OWNED); 903172409Sjeff 904164936Sjulian tdq = from; 905171482Sjeff cpu = TDQ_ID(to); 906177435Sjeff td = tdq_steal(tdq, cpu); 907177435Sjeff if (td == NULL) 908176735Sjeff return (0); 909177435Sjeff ts = td->td_sched; 910171482Sjeff /* 911171482Sjeff * Although the run queue is locked the thread may be blocked. Lock 912172409Sjeff * it to clear this and acquire the run-queue lock. 913171482Sjeff */ 914171482Sjeff thread_lock(td); 915172409Sjeff /* Drop recursive lock on from acquired via thread_lock(). */ 916171482Sjeff TDQ_UNLOCK(from); 917171482Sjeff sched_rem(td); 918166108Sjeff ts->ts_cpu = cpu; 919171482Sjeff td->td_lock = TDQ_LOCKPTR(to); 920171482Sjeff tdq_add(to, td, SRQ_YIELDING); 921176735Sjeff return (1); 922116069Sjeff} 923110267Sjeff 924171482Sjeff/* 925171482Sjeff * This tdq has idled. Try to steal a thread from another cpu and switch 926171482Sjeff * to it. 927171482Sjeff */ 928123433Sjeffstatic int 929164936Sjuliantdq_idled(struct tdq *tdq) 930121790Sjeff{ 931176735Sjeff struct cpu_group *cg; 932164936Sjulian struct tdq *steal; 933194779Sjeff cpuset_t mask; 934176735Sjeff int thresh; 935171482Sjeff int cpu; 936123433Sjeff 937172484Sjeff if (smp_started == 0 || steal_idle == 0) 938172484Sjeff return (1); 939194779Sjeff CPU_FILL(&mask); 940194779Sjeff CPU_CLR(PCPU_GET(cpuid), &mask); 941176735Sjeff /* We don't want to be preempted while we're iterating. */ 942171482Sjeff spinlock_enter(); 943176735Sjeff for (cg = tdq->tdq_cg; cg != NULL; ) { 944191643Sjeff if ((cg->cg_flags & CG_FLAG_THREAD) == 0) 945176735Sjeff thresh = steal_thresh; 946176735Sjeff else 947176735Sjeff thresh = 1; 948176735Sjeff cpu = sched_highest(cg, mask, thresh); 949176735Sjeff if (cpu == -1) { 950176735Sjeff cg = cg->cg_parent; 951176735Sjeff continue; 952166108Sjeff } 953176735Sjeff steal = TDQ_CPU(cpu); 954194779Sjeff CPU_CLR(cpu, &mask); 955176735Sjeff tdq_lock_pair(tdq, steal); 956176735Sjeff if (steal->tdq_load < thresh || steal->tdq_transferable == 0) { 957176735Sjeff tdq_unlock_pair(tdq, steal); 958176735Sjeff continue; 959171482Sjeff } 960176735Sjeff /* 961176735Sjeff * If a thread was added while interrupts were disabled don't 962176735Sjeff * steal one here. If we fail to acquire one due to affinity 963176735Sjeff * restrictions loop again with this cpu removed from the 964176735Sjeff * set. 965176735Sjeff */ 966176735Sjeff if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) { 967176735Sjeff tdq_unlock_pair(tdq, steal); 968176735Sjeff continue; 969176735Sjeff } 970176735Sjeff spinlock_exit(); 971176735Sjeff TDQ_UNLOCK(steal); 972178272Sjeff mi_switch(SW_VOL | SWT_IDLE, NULL); 973176735Sjeff thread_unlock(curthread); 974176735Sjeff 975176735Sjeff return (0); 976123433Sjeff } 977171482Sjeff spinlock_exit(); 978123433Sjeff return (1); 979121790Sjeff} 980121790Sjeff 981171482Sjeff/* 982171482Sjeff * Notify a remote cpu of new work. Sends an IPI if criteria are met. 983171482Sjeff */ 984121790Sjeffstatic void 985177435Sjefftdq_notify(struct tdq *tdq, struct thread *td) 986121790Sjeff{ 987185047Sjhb struct thread *ctd; 988166247Sjeff int pri; 989166108Sjeff int cpu; 990121790Sjeff 991177005Sjeff if (tdq->tdq_ipipending) 992177005Sjeff return; 993177435Sjeff cpu = td->td_sched->ts_cpu; 994177435Sjeff pri = td->td_priority; 995185047Sjhb ctd = pcpu_find(cpu)->pc_curthread; 996185047Sjhb if (!sched_shouldpreempt(pri, ctd->td_priority, 1)) 997166137Sjeff return; 998185047Sjhb if (TD_IS_IDLETHREAD(ctd)) { 999178277Sjeff /* 1000178471Sjeff * If the MD code has an idle wakeup routine try that before 1001178471Sjeff * falling back to IPI. 1002178471Sjeff */ 1003212416Smav if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu)) 1004178471Sjeff return; 1005178277Sjeff } 1006177005Sjeff tdq->tdq_ipipending = 1; 1007210939Sjhb ipi_cpu(cpu, IPI_PREEMPT); 1008121790Sjeff} 1009121790Sjeff 1010171482Sjeff/* 1011171482Sjeff * Steals load from a timeshare queue. Honors the rotating queue head 1012171482Sjeff * index. 1013171482Sjeff */ 1014177435Sjeffstatic struct thread * 1015176735Sjeffrunq_steal_from(struct runq *rq, int cpu, u_char start) 1016171482Sjeff{ 1017171482Sjeff struct rqbits *rqb; 1018171482Sjeff struct rqhead *rqh; 1019232207Smav struct thread *td, *first; 1020171482Sjeff int bit; 1021171482Sjeff int pri; 1022171482Sjeff int i; 1023171482Sjeff 1024171482Sjeff rqb = &rq->rq_status; 1025171482Sjeff bit = start & (RQB_BPW -1); 1026171482Sjeff pri = 0; 1027232207Smav first = NULL; 1028171482Sjeffagain: 1029171482Sjeff for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 1030171482Sjeff if (rqb->rqb_bits[i] == 0) 1031171482Sjeff continue; 1032171482Sjeff if (bit != 0) { 1033171482Sjeff for (pri = bit; pri < RQB_BPW; pri++) 1034171482Sjeff if (rqb->rqb_bits[i] & (1ul << pri)) 1035171482Sjeff break; 1036171482Sjeff if (pri >= RQB_BPW) 1037171482Sjeff continue; 1038171482Sjeff } else 1039171482Sjeff pri = RQB_FFS(rqb->rqb_bits[i]); 1040171482Sjeff pri += (i << RQB_L2BPW); 1041171482Sjeff rqh = &rq->rq_queues[pri]; 1042177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) { 1043177435Sjeff if (first && THREAD_CAN_MIGRATE(td) && 1044177435Sjeff THREAD_CAN_SCHED(td, cpu)) 1045177435Sjeff return (td); 1046232207Smav first = td; 1047171482Sjeff } 1048171482Sjeff } 1049171482Sjeff if (start != 0) { 1050171482Sjeff start = 0; 1051171482Sjeff goto again; 1052171482Sjeff } 1053171482Sjeff 1054232207Smav if (first && THREAD_CAN_MIGRATE(first) && 1055232207Smav THREAD_CAN_SCHED(first, cpu)) 1056232207Smav return (first); 1057171482Sjeff return (NULL); 1058171482Sjeff} 1059171482Sjeff 1060171482Sjeff/* 1061171482Sjeff * Steals load from a standard linear queue. 1062171482Sjeff */ 1063177435Sjeffstatic struct thread * 1064176735Sjeffrunq_steal(struct runq *rq, int cpu) 1065121790Sjeff{ 1066121790Sjeff struct rqhead *rqh; 1067121790Sjeff struct rqbits *rqb; 1068177435Sjeff struct thread *td; 1069121790Sjeff int word; 1070121790Sjeff int bit; 1071121790Sjeff 1072121790Sjeff rqb = &rq->rq_status; 1073121790Sjeff for (word = 0; word < RQB_LEN; word++) { 1074121790Sjeff if (rqb->rqb_bits[word] == 0) 1075121790Sjeff continue; 1076121790Sjeff for (bit = 0; bit < RQB_BPW; bit++) { 1077123231Speter if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 1078121790Sjeff continue; 1079121790Sjeff rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 1080177435Sjeff TAILQ_FOREACH(td, rqh, td_runq) 1081177435Sjeff if (THREAD_CAN_MIGRATE(td) && 1082177435Sjeff THREAD_CAN_SCHED(td, cpu)) 1083177435Sjeff return (td); 1084121790Sjeff } 1085121790Sjeff } 1086121790Sjeff return (NULL); 1087121790Sjeff} 1088121790Sjeff 1089171482Sjeff/* 1090171482Sjeff * Attempt to steal a thread in priority order from a thread queue. 1091171482Sjeff */ 1092177435Sjeffstatic struct thread * 1093176735Sjefftdq_steal(struct tdq *tdq, int cpu) 1094121790Sjeff{ 1095177435Sjeff struct thread *td; 1096121790Sjeff 1097171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1098177435Sjeff if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL) 1099177435Sjeff return (td); 1100177435Sjeff if ((td = runq_steal_from(&tdq->tdq_timeshare, 1101177435Sjeff cpu, tdq->tdq_ridx)) != NULL) 1102177435Sjeff return (td); 1103176735Sjeff return (runq_steal(&tdq->tdq_idle, cpu)); 1104121790Sjeff} 1105123433Sjeff 1106171482Sjeff/* 1107171482Sjeff * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 1108172409Sjeff * current lock and returns with the assigned queue locked. 1109171482Sjeff */ 1110171482Sjeffstatic inline struct tdq * 1111177435Sjeffsched_setcpu(struct thread *td, int cpu, int flags) 1112123433Sjeff{ 1113177435Sjeff 1114171482Sjeff struct tdq *tdq; 1115123433Sjeff 1116177435Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1117171482Sjeff tdq = TDQ_CPU(cpu); 1118177435Sjeff td->td_sched->ts_cpu = cpu; 1119177435Sjeff /* 1120177435Sjeff * If the lock matches just return the queue. 1121177435Sjeff */ 1122171482Sjeff if (td->td_lock == TDQ_LOCKPTR(tdq)) 1123171482Sjeff return (tdq); 1124171482Sjeff#ifdef notyet 1125123433Sjeff /* 1126172293Sjeff * If the thread isn't running its lockptr is a 1127171482Sjeff * turnstile or a sleepqueue. We can just lock_set without 1128171482Sjeff * blocking. 1129123685Sjeff */ 1130171482Sjeff if (TD_CAN_RUN(td)) { 1131171482Sjeff TDQ_LOCK(tdq); 1132171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 1133171482Sjeff return (tdq); 1134171482Sjeff } 1135171482Sjeff#endif 1136166108Sjeff /* 1137171482Sjeff * The hard case, migration, we need to block the thread first to 1138171482Sjeff * prevent order reversals with other cpus locks. 1139166108Sjeff */ 1140202889Sattilio spinlock_enter(); 1141171482Sjeff thread_lock_block(td); 1142171482Sjeff TDQ_LOCK(tdq); 1143171713Sjeff thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 1144202889Sattilio spinlock_exit(); 1145171482Sjeff return (tdq); 1146166108Sjeff} 1147166108Sjeff 1148178272SjeffSCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding"); 1149178272SjeffSCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity"); 1150178272SjeffSCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity"); 1151178272SjeffSCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load"); 1152178272SjeffSCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu"); 1153178272SjeffSCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration"); 1154178272Sjeff 1155166108Sjeffstatic int 1156177435Sjeffsched_pickcpu(struct thread *td, int flags) 1157171482Sjeff{ 1158232207Smav struct cpu_group *cg, *ccg; 1159177435Sjeff struct td_sched *ts; 1160171482Sjeff struct tdq *tdq; 1161194779Sjeff cpuset_t mask; 1162232207Smav int cpu, pri, self; 1163166108Sjeff 1164176735Sjeff self = PCPU_GET(cpuid); 1165177435Sjeff ts = td->td_sched; 1166166108Sjeff if (smp_started == 0) 1167166108Sjeff return (self); 1168171506Sjeff /* 1169171506Sjeff * Don't migrate a running thread from sched_switch(). 1170171506Sjeff */ 1171176735Sjeff if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td)) 1172176735Sjeff return (ts->ts_cpu); 1173166108Sjeff /* 1174176735Sjeff * Prefer to run interrupt threads on the processors that generate 1175176735Sjeff * the interrupt. 1176166108Sjeff */ 1177232207Smav pri = td->td_priority; 1178176735Sjeff if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) && 1179178272Sjeff curthread->td_intr_nesting_level && ts->ts_cpu != self) { 1180178272Sjeff SCHED_STAT_INC(pickcpu_intrbind); 1181176735Sjeff ts->ts_cpu = self; 1182232207Smav if (TDQ_CPU(self)->tdq_lowpri > pri) { 1183232207Smav SCHED_STAT_INC(pickcpu_affinity); 1184232207Smav return (ts->ts_cpu); 1185232207Smav } 1186178272Sjeff } 1187166108Sjeff /* 1188176735Sjeff * If the thread can run on the last cpu and the affinity has not 1189176735Sjeff * expired or it is idle run it there. 1190166108Sjeff */ 1191176735Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1192232207Smav cg = tdq->tdq_cg; 1193232207Smav if (THREAD_CAN_SCHED(td, ts->ts_cpu) && 1194232207Smav tdq->tdq_lowpri >= PRI_MIN_IDLE && 1195232207Smav SCHED_AFFINITY(ts, CG_SHARE_L2)) { 1196232207Smav if (cg->cg_flags & CG_FLAG_THREAD) { 1197232207Smav CPUSET_FOREACH(cpu, cg->cg_mask) { 1198232207Smav if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) 1199232207Smav break; 1200232207Smav } 1201232207Smav } else 1202232207Smav cpu = INT_MAX; 1203232207Smav if (cpu > mp_maxid) { 1204178272Sjeff SCHED_STAT_INC(pickcpu_idle_affinity); 1205176735Sjeff return (ts->ts_cpu); 1206178272Sjeff } 1207139334Sjeff } 1208123433Sjeff /* 1209232207Smav * Search for the last level cache CPU group in the tree. 1210232207Smav * Skip caches with expired affinity time and SMT groups. 1211232207Smav * Affinity to higher level caches will be handled less aggressively. 1212123433Sjeff */ 1213232207Smav for (ccg = NULL; cg != NULL; cg = cg->cg_parent) { 1214232207Smav if (cg->cg_flags & CG_FLAG_THREAD) 1215232207Smav continue; 1216232207Smav if (!SCHED_AFFINITY(ts, cg->cg_level)) 1217232207Smav continue; 1218232207Smav ccg = cg; 1219232207Smav } 1220232207Smav if (ccg != NULL) 1221232207Smav cg = ccg; 1222176735Sjeff cpu = -1; 1223232207Smav /* Search the group for the less loaded idle CPU we can run now. */ 1224194779Sjeff mask = td->td_cpuset->cs_mask; 1225232207Smav if (cg != NULL && cg != cpu_top && 1226232207Smav CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0) 1227232207Smav cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE), 1228232207Smav INT_MAX, ts->ts_cpu); 1229232207Smav /* Search globally for the less loaded CPU we can run now. */ 1230176735Sjeff if (cpu == -1) 1231232207Smav cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu); 1232232207Smav /* Search globally for the less loaded CPU. */ 1233232207Smav if (cpu == -1) 1234232207Smav cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu); 1235232454Smav KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu.")); 1236171506Sjeff /* 1237176735Sjeff * Compare the lowest loaded cpu to current cpu. 1238171506Sjeff */ 1239177005Sjeff if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri && 1240232207Smav TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE && 1241232207Smav TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) { 1242178272Sjeff SCHED_STAT_INC(pickcpu_local); 1243177005Sjeff cpu = self; 1244178272Sjeff } else 1245178272Sjeff SCHED_STAT_INC(pickcpu_lowest); 1246178272Sjeff if (cpu != ts->ts_cpu) 1247178272Sjeff SCHED_STAT_INC(pickcpu_migration); 1248171482Sjeff return (cpu); 1249123433Sjeff} 1250176735Sjeff#endif 1251123433Sjeff 1252117326Sjeff/* 1253121790Sjeff * Pick the highest priority task we have and return it. 1254117326Sjeff */ 1255177435Sjeffstatic struct thread * 1256164936Sjuliantdq_choose(struct tdq *tdq) 1257110267Sjeff{ 1258177435Sjeff struct thread *td; 1259110267Sjeff 1260171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1261177435Sjeff td = runq_choose(&tdq->tdq_realtime); 1262177435Sjeff if (td != NULL) 1263177435Sjeff return (td); 1264177435Sjeff td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1265177435Sjeff if (td != NULL) { 1266217351Sjhb KASSERT(td->td_priority >= PRI_MIN_BATCH, 1267165762Sjeff ("tdq_choose: Invalid priority on timeshare queue %d", 1268177435Sjeff td->td_priority)); 1269177435Sjeff return (td); 1270165762Sjeff } 1271177435Sjeff td = runq_choose(&tdq->tdq_idle); 1272177435Sjeff if (td != NULL) { 1273177435Sjeff KASSERT(td->td_priority >= PRI_MIN_IDLE, 1274165762Sjeff ("tdq_choose: Invalid priority on idle queue %d", 1275177435Sjeff td->td_priority)); 1276177435Sjeff return (td); 1277165762Sjeff } 1278165762Sjeff 1279165762Sjeff return (NULL); 1280110267Sjeff} 1281110267Sjeff 1282171482Sjeff/* 1283171482Sjeff * Initialize a thread queue. 1284171482Sjeff */ 1285109864Sjeffstatic void 1286164936Sjuliantdq_setup(struct tdq *tdq) 1287110028Sjeff{ 1288171482Sjeff 1289171713Sjeff if (bootverbose) 1290171713Sjeff printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1291165762Sjeff runq_init(&tdq->tdq_realtime); 1292165762Sjeff runq_init(&tdq->tdq_timeshare); 1293165620Sjeff runq_init(&tdq->tdq_idle); 1294176735Sjeff snprintf(tdq->tdq_name, sizeof(tdq->tdq_name), 1295176735Sjeff "sched lock %d", (int)TDQ_ID(tdq)); 1296176735Sjeff mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", 1297176735Sjeff MTX_SPIN | MTX_RECURSE); 1298187357Sjeff#ifdef KTR 1299187357Sjeff snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname), 1300187357Sjeff "CPU %d load", (int)TDQ_ID(tdq)); 1301187357Sjeff#endif 1302110028Sjeff} 1303110028Sjeff 1304171713Sjeff#ifdef SMP 1305110028Sjeffstatic void 1306171713Sjeffsched_setup_smp(void) 1307171713Sjeff{ 1308171713Sjeff struct tdq *tdq; 1309171713Sjeff int i; 1310171713Sjeff 1311176735Sjeff cpu_top = smp_topo(); 1312209059Sjhb CPU_FOREACH(i) { 1313176735Sjeff tdq = TDQ_CPU(i); 1314171713Sjeff tdq_setup(tdq); 1315176735Sjeff tdq->tdq_cg = smp_topo_find(cpu_top, i); 1316176735Sjeff if (tdq->tdq_cg == NULL) 1317176735Sjeff panic("Can't find cpu group for %d\n", i); 1318123433Sjeff } 1319176735Sjeff balance_tdq = TDQ_SELF(); 1320176735Sjeff sched_balance(); 1321171713Sjeff} 1322171713Sjeff#endif 1323171713Sjeff 1324171713Sjeff/* 1325171713Sjeff * Setup the thread queues and initialize the topology based on MD 1326171713Sjeff * information. 1327171713Sjeff */ 1328171713Sjeffstatic void 1329171713Sjeffsched_setup(void *dummy) 1330171713Sjeff{ 1331171713Sjeff struct tdq *tdq; 1332171713Sjeff 1333171713Sjeff tdq = TDQ_SELF(); 1334171713Sjeff#ifdef SMP 1335176734Sjeff sched_setup_smp(); 1336117237Sjeff#else 1337171713Sjeff tdq_setup(tdq); 1338116069Sjeff#endif 1339171482Sjeff /* 1340171482Sjeff * To avoid divide-by-zero, we set realstathz a dummy value 1341171482Sjeff * in case which sched_clock() called before sched_initticks(). 1342171482Sjeff */ 1343171482Sjeff realstathz = hz; 1344171482Sjeff sched_slice = (realstathz/10); /* ~100ms */ 1345171482Sjeff tickincr = 1 << SCHED_TICK_SHIFT; 1346171482Sjeff 1347171482Sjeff /* Add thread0's load since it's running. */ 1348171482Sjeff TDQ_LOCK(tdq); 1349171713Sjeff thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1350177435Sjeff tdq_load_add(tdq, &thread0); 1351176735Sjeff tdq->tdq_lowpri = thread0.td_priority; 1352171482Sjeff TDQ_UNLOCK(tdq); 1353109864Sjeff} 1354109864Sjeff 1355171482Sjeff/* 1356171482Sjeff * This routine determines the tickincr after stathz and hz are setup. 1357171482Sjeff */ 1358153533Sdavidxu/* ARGSUSED */ 1359153533Sdavidxustatic void 1360153533Sdavidxusched_initticks(void *dummy) 1361153533Sdavidxu{ 1362171482Sjeff int incr; 1363171482Sjeff 1364153533Sdavidxu realstathz = stathz ? stathz : hz; 1365166229Sjeff sched_slice = (realstathz/10); /* ~100ms */ 1366153533Sdavidxu 1367153533Sdavidxu /* 1368165762Sjeff * tickincr is shifted out by 10 to avoid rounding errors due to 1369165766Sjeff * hz not being evenly divisible by stathz on all platforms. 1370153533Sdavidxu */ 1371171482Sjeff incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1372165762Sjeff /* 1373165762Sjeff * This does not work for values of stathz that are more than 1374165762Sjeff * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1375165762Sjeff */ 1376171482Sjeff if (incr == 0) 1377171482Sjeff incr = 1; 1378171482Sjeff tickincr = incr; 1379166108Sjeff#ifdef SMP 1380171899Sjeff /* 1381172409Sjeff * Set the default balance interval now that we know 1382172409Sjeff * what realstathz is. 1383172409Sjeff */ 1384172409Sjeff balance_interval = realstathz; 1385172409Sjeff /* 1386189787Sjeff * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4. 1387189787Sjeff * This prevents excess thrashing on large machines and excess idle 1388189787Sjeff * on smaller machines. 1389171899Sjeff */ 1390189787Sjeff steal_thresh = min(fls(mp_ncpus) - 1, 3); 1391166108Sjeff affinity = SCHED_AFFINITY_DEFAULT; 1392166108Sjeff#endif 1393153533Sdavidxu} 1394153533Sdavidxu 1395153533Sdavidxu 1396109864Sjeff/* 1397171482Sjeff * This is the core of the interactivity algorithm. Determines a score based 1398171482Sjeff * on past behavior. It is the ratio of sleep time to run time scaled to 1399171482Sjeff * a [0, 100] integer. This is the voluntary sleep time of a process, which 1400171482Sjeff * differs from the cpu usage because it does not account for time spent 1401171482Sjeff * waiting on a run-queue. Would be prettier if we had floating point. 1402171482Sjeff */ 1403171482Sjeffstatic int 1404171482Sjeffsched_interact_score(struct thread *td) 1405171482Sjeff{ 1406171482Sjeff struct td_sched *ts; 1407171482Sjeff int div; 1408171482Sjeff 1409171482Sjeff ts = td->td_sched; 1410171482Sjeff /* 1411171482Sjeff * The score is only needed if this is likely to be an interactive 1412171482Sjeff * task. Don't go through the expense of computing it if there's 1413171482Sjeff * no chance. 1414171482Sjeff */ 1415171482Sjeff if (sched_interact <= SCHED_INTERACT_HALF && 1416171482Sjeff ts->ts_runtime >= ts->ts_slptime) 1417171482Sjeff return (SCHED_INTERACT_HALF); 1418171482Sjeff 1419171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1420171482Sjeff div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1421171482Sjeff return (SCHED_INTERACT_HALF + 1422171482Sjeff (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1423171482Sjeff } 1424171482Sjeff if (ts->ts_slptime > ts->ts_runtime) { 1425171482Sjeff div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1426171482Sjeff return (ts->ts_runtime / div); 1427171482Sjeff } 1428171482Sjeff /* runtime == slptime */ 1429171482Sjeff if (ts->ts_runtime) 1430171482Sjeff return (SCHED_INTERACT_HALF); 1431171482Sjeff 1432171482Sjeff /* 1433171482Sjeff * This can happen if slptime and runtime are 0. 1434171482Sjeff */ 1435171482Sjeff return (0); 1436171482Sjeff 1437171482Sjeff} 1438171482Sjeff 1439171482Sjeff/* 1440109864Sjeff * Scale the scheduling priority according to the "interactivity" of this 1441109864Sjeff * process. 1442109864Sjeff */ 1443113357Sjeffstatic void 1444163709Sjbsched_priority(struct thread *td) 1445109864Sjeff{ 1446165762Sjeff int score; 1447109864Sjeff int pri; 1448109864Sjeff 1449217291Sjhb if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1450113357Sjeff return; 1451112966Sjeff /* 1452165762Sjeff * If the score is interactive we place the thread in the realtime 1453165762Sjeff * queue with a priority that is less than kernel and interrupt 1454165762Sjeff * priorities. These threads are not subject to nice restrictions. 1455112966Sjeff * 1456171482Sjeff * Scores greater than this are placed on the normal timeshare queue 1457165762Sjeff * where the priority is partially decided by the most recent cpu 1458165762Sjeff * utilization and the rest is decided by nice value. 1459172293Sjeff * 1460172293Sjeff * The nice value of the process has a linear effect on the calculated 1461172293Sjeff * score. Negative nice values make it easier for a thread to be 1462172293Sjeff * considered interactive. 1463112966Sjeff */ 1464198126Sjhb score = imax(0, sched_interact_score(td) + td->td_proc->p_nice); 1465165762Sjeff if (score < sched_interact) { 1466217351Sjhb pri = PRI_MIN_INTERACT; 1467217351Sjhb pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) / 1468217237Sjhb sched_interact) * score; 1469217351Sjhb KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT, 1470166208Sjeff ("sched_priority: invalid interactive priority %d score %d", 1471166208Sjeff pri, score)); 1472165762Sjeff } else { 1473165762Sjeff pri = SCHED_PRI_MIN; 1474165762Sjeff if (td->td_sched->ts_ticks) 1475228960Sjhb pri += min(SCHED_PRI_TICKS(td->td_sched), 1476228960Sjhb SCHED_PRI_RANGE); 1477165762Sjeff pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1478217351Sjhb KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH, 1479171482Sjeff ("sched_priority: invalid priority %d: nice %d, " 1480171482Sjeff "ticks %d ftick %d ltick %d tick pri %d", 1481171482Sjeff pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1482171482Sjeff td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1483171482Sjeff SCHED_PRI_TICKS(td->td_sched))); 1484165762Sjeff } 1485165762Sjeff sched_user_prio(td, pri); 1486112966Sjeff 1487112966Sjeff return; 1488109864Sjeff} 1489109864Sjeff 1490121868Sjeff/* 1491121868Sjeff * This routine enforces a maximum limit on the amount of scheduling history 1492171482Sjeff * kept. It is called after either the slptime or runtime is adjusted. This 1493171482Sjeff * function is ugly due to integer math. 1494121868Sjeff */ 1495116463Sjeffstatic void 1496163709Sjbsched_interact_update(struct thread *td) 1497116463Sjeff{ 1498165819Sjeff struct td_sched *ts; 1499166208Sjeff u_int sum; 1500121605Sjeff 1501165819Sjeff ts = td->td_sched; 1502171482Sjeff sum = ts->ts_runtime + ts->ts_slptime; 1503121868Sjeff if (sum < SCHED_SLP_RUN_MAX) 1504121868Sjeff return; 1505121868Sjeff /* 1506165819Sjeff * This only happens from two places: 1507165819Sjeff * 1) We have added an unusual amount of run time from fork_exit. 1508165819Sjeff * 2) We have added an unusual amount of sleep time from sched_sleep(). 1509165819Sjeff */ 1510165819Sjeff if (sum > SCHED_SLP_RUN_MAX * 2) { 1511171482Sjeff if (ts->ts_runtime > ts->ts_slptime) { 1512171482Sjeff ts->ts_runtime = SCHED_SLP_RUN_MAX; 1513171482Sjeff ts->ts_slptime = 1; 1514165819Sjeff } else { 1515171482Sjeff ts->ts_slptime = SCHED_SLP_RUN_MAX; 1516171482Sjeff ts->ts_runtime = 1; 1517165819Sjeff } 1518165819Sjeff return; 1519165819Sjeff } 1520165819Sjeff /* 1521121868Sjeff * If we have exceeded by more than 1/5th then the algorithm below 1522121868Sjeff * will not bring us back into range. Dividing by two here forces 1523133427Sjeff * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1524121868Sjeff */ 1525127850Sjeff if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1526171482Sjeff ts->ts_runtime /= 2; 1527171482Sjeff ts->ts_slptime /= 2; 1528121868Sjeff return; 1529116463Sjeff } 1530171482Sjeff ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1531171482Sjeff ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1532116463Sjeff} 1533116463Sjeff 1534171482Sjeff/* 1535171482Sjeff * Scale back the interactivity history when a child thread is created. The 1536171482Sjeff * history is inherited from the parent but the thread may behave totally 1537171482Sjeff * differently. For example, a shell spawning a compiler process. We want 1538171482Sjeff * to learn that the compiler is behaving badly very quickly. 1539171482Sjeff */ 1540121868Sjeffstatic void 1541163709Sjbsched_interact_fork(struct thread *td) 1542121868Sjeff{ 1543121868Sjeff int ratio; 1544121868Sjeff int sum; 1545121868Sjeff 1546171482Sjeff sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1547121868Sjeff if (sum > SCHED_SLP_RUN_FORK) { 1548121868Sjeff ratio = sum / SCHED_SLP_RUN_FORK; 1549171482Sjeff td->td_sched->ts_runtime /= ratio; 1550171482Sjeff td->td_sched->ts_slptime /= ratio; 1551121868Sjeff } 1552121868Sjeff} 1553121868Sjeff 1554113357Sjeff/* 1555171482Sjeff * Called from proc0_init() to setup the scheduler fields. 1556134791Sjulian */ 1557134791Sjulianvoid 1558134791Sjulianschedinit(void) 1559134791Sjulian{ 1560165762Sjeff 1561134791Sjulian /* 1562134791Sjulian * Set up the scheduler specific parts of proc0. 1563134791Sjulian */ 1564136167Sjulian proc0.p_sched = NULL; /* XXX */ 1565164936Sjulian thread0.td_sched = &td_sched0; 1566165762Sjeff td_sched0.ts_ltick = ticks; 1567165796Sjeff td_sched0.ts_ftick = ticks; 1568177009Sjeff td_sched0.ts_slice = sched_slice; 1569134791Sjulian} 1570134791Sjulian 1571134791Sjulian/* 1572113357Sjeff * This is only somewhat accurate since given many processes of the same 1573113357Sjeff * priority they will switch when their slices run out, which will be 1574165762Sjeff * at most sched_slice stathz ticks. 1575113357Sjeff */ 1576109864Sjeffint 1577109864Sjeffsched_rr_interval(void) 1578109864Sjeff{ 1579165762Sjeff 1580165762Sjeff /* Convert sched_slice to hz */ 1581165762Sjeff return (hz/(realstathz/sched_slice)); 1582109864Sjeff} 1583109864Sjeff 1584171482Sjeff/* 1585171482Sjeff * Update the percent cpu tracking information when it is requested or 1586171482Sjeff * the total history exceeds the maximum. We keep a sliding history of 1587171482Sjeff * tick counts that slowly decays. This is less precise than the 4BSD 1588171482Sjeff * mechanism since it happens with less regular and frequent events. 1589171482Sjeff */ 1590121790Sjeffstatic void 1591164936Sjuliansched_pctcpu_update(struct td_sched *ts) 1592109864Sjeff{ 1593165762Sjeff 1594165762Sjeff if (ts->ts_ticks == 0) 1595165762Sjeff return; 1596165796Sjeff if (ticks - (hz / 10) < ts->ts_ltick && 1597165796Sjeff SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1598165796Sjeff return; 1599109864Sjeff /* 1600109864Sjeff * Adjust counters and watermark for pctcpu calc. 1601116365Sjeff */ 1602165762Sjeff if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1603164936Sjulian ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1604165762Sjeff SCHED_TICK_TARG; 1605165762Sjeff else 1606164936Sjulian ts->ts_ticks = 0; 1607164936Sjulian ts->ts_ltick = ticks; 1608165762Sjeff ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1609109864Sjeff} 1610109864Sjeff 1611171482Sjeff/* 1612171482Sjeff * Adjust the priority of a thread. Move it to the appropriate run-queue 1613171482Sjeff * if necessary. This is the back-end for several priority related 1614171482Sjeff * functions. 1615171482Sjeff */ 1616165762Sjeffstatic void 1617139453Sjhbsched_thread_priority(struct thread *td, u_char prio) 1618109864Sjeff{ 1619164936Sjulian struct td_sched *ts; 1620177009Sjeff struct tdq *tdq; 1621177009Sjeff int oldpri; 1622109864Sjeff 1623187357Sjeff KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio", 1624187357Sjeff "prio:%d", td->td_priority, "new prio:%d", prio, 1625187357Sjeff KTR_ATTR_LINKED, sched_tdname(curthread)); 1626187357Sjeff if (td != curthread && prio > td->td_priority) { 1627187357Sjeff KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), 1628187357Sjeff "lend prio", "prio:%d", td->td_priority, "new prio:%d", 1629187357Sjeff prio, KTR_ATTR_LINKED, sched_tdname(td)); 1630187357Sjeff } 1631164936Sjulian ts = td->td_sched; 1632170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1633139453Sjhb if (td->td_priority == prio) 1634139453Sjhb return; 1635177376Sjeff /* 1636177376Sjeff * If the priority has been elevated due to priority 1637177376Sjeff * propagation, we may have to move ourselves to a new 1638177376Sjeff * queue. This could be optimized to not re-add in some 1639177376Sjeff * cases. 1640177376Sjeff */ 1641165766Sjeff if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1642165762Sjeff sched_rem(td); 1643165762Sjeff td->td_priority = prio; 1644171482Sjeff sched_add(td, SRQ_BORROWING); 1645177009Sjeff return; 1646177009Sjeff } 1647177376Sjeff /* 1648177376Sjeff * If the thread is currently running we may have to adjust the lowpri 1649177376Sjeff * information so other cpus are aware of our current priority. 1650177376Sjeff */ 1651177009Sjeff if (TD_IS_RUNNING(td)) { 1652177376Sjeff tdq = TDQ_CPU(ts->ts_cpu); 1653177376Sjeff oldpri = td->td_priority; 1654177376Sjeff td->td_priority = prio; 1655176735Sjeff if (prio < tdq->tdq_lowpri) 1656171482Sjeff tdq->tdq_lowpri = prio; 1657176735Sjeff else if (tdq->tdq_lowpri == oldpri) 1658176735Sjeff tdq_setlowpri(tdq, td); 1659177376Sjeff return; 1660177009Sjeff } 1661177376Sjeff td->td_priority = prio; 1662109864Sjeff} 1663109864Sjeff 1664139453Sjhb/* 1665139453Sjhb * Update a thread's priority when it is lent another thread's 1666139453Sjhb * priority. 1667139453Sjhb */ 1668109864Sjeffvoid 1669139453Sjhbsched_lend_prio(struct thread *td, u_char prio) 1670139453Sjhb{ 1671139453Sjhb 1672139453Sjhb td->td_flags |= TDF_BORROWING; 1673139453Sjhb sched_thread_priority(td, prio); 1674139453Sjhb} 1675139453Sjhb 1676139453Sjhb/* 1677139453Sjhb * Restore a thread's priority when priority propagation is 1678139453Sjhb * over. The prio argument is the minimum priority the thread 1679139453Sjhb * needs to have to satisfy other possible priority lending 1680139453Sjhb * requests. If the thread's regular priority is less 1681139453Sjhb * important than prio, the thread will keep a priority boost 1682139453Sjhb * of prio. 1683139453Sjhb */ 1684139453Sjhbvoid 1685139453Sjhbsched_unlend_prio(struct thread *td, u_char prio) 1686139453Sjhb{ 1687139453Sjhb u_char base_pri; 1688139453Sjhb 1689139453Sjhb if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1690139453Sjhb td->td_base_pri <= PRI_MAX_TIMESHARE) 1691163709Sjb base_pri = td->td_user_pri; 1692139453Sjhb else 1693139453Sjhb base_pri = td->td_base_pri; 1694139453Sjhb if (prio >= base_pri) { 1695139455Sjhb td->td_flags &= ~TDF_BORROWING; 1696139453Sjhb sched_thread_priority(td, base_pri); 1697139453Sjhb } else 1698139453Sjhb sched_lend_prio(td, prio); 1699139453Sjhb} 1700139453Sjhb 1701171482Sjeff/* 1702171482Sjeff * Standard entry for setting the priority to an absolute value. 1703171482Sjeff */ 1704139453Sjhbvoid 1705139453Sjhbsched_prio(struct thread *td, u_char prio) 1706139453Sjhb{ 1707139453Sjhb u_char oldprio; 1708139453Sjhb 1709139453Sjhb /* First, update the base priority. */ 1710139453Sjhb td->td_base_pri = prio; 1711139453Sjhb 1712139453Sjhb /* 1713139455Sjhb * If the thread is borrowing another thread's priority, don't 1714139453Sjhb * ever lower the priority. 1715139453Sjhb */ 1716139453Sjhb if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1717139453Sjhb return; 1718139453Sjhb 1719139453Sjhb /* Change the real priority. */ 1720139453Sjhb oldprio = td->td_priority; 1721139453Sjhb sched_thread_priority(td, prio); 1722139453Sjhb 1723139453Sjhb /* 1724139453Sjhb * If the thread is on a turnstile, then let the turnstile update 1725139453Sjhb * its state. 1726139453Sjhb */ 1727139453Sjhb if (TD_ON_LOCK(td) && oldprio != prio) 1728139453Sjhb turnstile_adjust(td, oldprio); 1729139453Sjhb} 1730139455Sjhb 1731171482Sjeff/* 1732171482Sjeff * Set the base user priority, does not effect current running priority. 1733171482Sjeff */ 1734139453Sjhbvoid 1735163709Sjbsched_user_prio(struct thread *td, u_char prio) 1736161599Sdavidxu{ 1737161599Sdavidxu 1738163709Sjb td->td_base_user_pri = prio; 1739216313Sdavidxu if (td->td_lend_user_pri <= prio) 1740216313Sdavidxu return; 1741163709Sjb td->td_user_pri = prio; 1742161599Sdavidxu} 1743161599Sdavidxu 1744161599Sdavidxuvoid 1745161599Sdavidxusched_lend_user_prio(struct thread *td, u_char prio) 1746161599Sdavidxu{ 1747161599Sdavidxu 1748174536Sdavidxu THREAD_LOCK_ASSERT(td, MA_OWNED); 1749216313Sdavidxu td->td_lend_user_pri = prio; 1750216791Sdavidxu td->td_user_pri = min(prio, td->td_base_user_pri); 1751216791Sdavidxu if (td->td_priority > td->td_user_pri) 1752216791Sdavidxu sched_prio(td, td->td_user_pri); 1753216791Sdavidxu else if (td->td_priority != td->td_user_pri) 1754216791Sdavidxu td->td_flags |= TDF_NEEDRESCHED; 1755161599Sdavidxu} 1756161599Sdavidxu 1757171482Sjeff/* 1758171713Sjeff * Handle migration from sched_switch(). This happens only for 1759171713Sjeff * cpu binding. 1760171713Sjeff */ 1761171713Sjeffstatic struct mtx * 1762171713Sjeffsched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1763171713Sjeff{ 1764171713Sjeff struct tdq *tdn; 1765171713Sjeff 1766171713Sjeff tdn = TDQ_CPU(td->td_sched->ts_cpu); 1767171713Sjeff#ifdef SMP 1768177435Sjeff tdq_load_rem(tdq, td); 1769171713Sjeff /* 1770171713Sjeff * Do the lock dance required to avoid LOR. We grab an extra 1771171713Sjeff * spinlock nesting to prevent preemption while we're 1772171713Sjeff * not holding either run-queue lock. 1773171713Sjeff */ 1774171713Sjeff spinlock_enter(); 1775202889Sattilio thread_lock_block(td); /* This releases the lock on tdq. */ 1776197223Sattilio 1777197223Sattilio /* 1778197223Sattilio * Acquire both run-queue locks before placing the thread on the new 1779197223Sattilio * run-queue to avoid deadlocks created by placing a thread with a 1780197223Sattilio * blocked lock on the run-queue of a remote processor. The deadlock 1781197223Sattilio * occurs when a third processor attempts to lock the two queues in 1782197223Sattilio * question while the target processor is spinning with its own 1783197223Sattilio * run-queue lock held while waiting for the blocked lock to clear. 1784197223Sattilio */ 1785197223Sattilio tdq_lock_pair(tdn, tdq); 1786171713Sjeff tdq_add(tdn, td, flags); 1787177435Sjeff tdq_notify(tdn, td); 1788197223Sattilio TDQ_UNLOCK(tdn); 1789171713Sjeff spinlock_exit(); 1790171713Sjeff#endif 1791171713Sjeff return (TDQ_LOCKPTR(tdn)); 1792171713Sjeff} 1793171713Sjeff 1794171713Sjeff/* 1795202889Sattilio * Variadic version of thread_lock_unblock() that does not assume td_lock 1796202889Sattilio * is blocked. 1797171482Sjeff */ 1798171482Sjeffstatic inline void 1799171482Sjeffthread_unblock_switch(struct thread *td, struct mtx *mtx) 1800171482Sjeff{ 1801171482Sjeff atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1802171482Sjeff (uintptr_t)mtx); 1803171482Sjeff} 1804171482Sjeff 1805171482Sjeff/* 1806171482Sjeff * Switch threads. This function has to handle threads coming in while 1807171482Sjeff * blocked for some reason, running, or idle. It also must deal with 1808171482Sjeff * migrating a thread from one queue to another as running threads may 1809171482Sjeff * be assigned elsewhere via binding. 1810171482Sjeff */ 1811161599Sdavidxuvoid 1812135051Sjuliansched_switch(struct thread *td, struct thread *newtd, int flags) 1813109864Sjeff{ 1814165627Sjeff struct tdq *tdq; 1815164936Sjulian struct td_sched *ts; 1816171482Sjeff struct mtx *mtx; 1817171713Sjeff int srqflag; 1818171482Sjeff int cpuid; 1819109864Sjeff 1820170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1821177376Sjeff KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument")); 1822109864Sjeff 1823171482Sjeff cpuid = PCPU_GET(cpuid); 1824171482Sjeff tdq = TDQ_CPU(cpuid); 1825164936Sjulian ts = td->td_sched; 1826171713Sjeff mtx = td->td_lock; 1827171482Sjeff ts->ts_rltick = ticks; 1828133555Sjeff td->td_lastcpu = td->td_oncpu; 1829113339Sjulian td->td_oncpu = NOCPU; 1830220198Sfabient if (!(flags & SW_PREEMPT)) 1831220198Sfabient td->td_flags &= ~TDF_NEEDRESCHED; 1832144777Sups td->td_owepreempt = 0; 1833178277Sjeff tdq->tdq_switchcnt++; 1834123434Sjeff /* 1835171482Sjeff * The lock pointer in an idle thread should never change. Reset it 1836171482Sjeff * to CAN_RUN as well. 1837123434Sjeff */ 1838167327Sjulian if (TD_IS_IDLETHREAD(td)) { 1839171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1840139334Sjeff TD_SET_CAN_RUN(td); 1841170293Sjeff } else if (TD_IS_RUNNING(td)) { 1842171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1843171713Sjeff srqflag = (flags & SW_PREEMPT) ? 1844170293Sjeff SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1845171713Sjeff SRQ_OURSELF|SRQ_YIELDING; 1846212153Smdf#ifdef SMP 1847212115Smdf if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu)) 1848212115Smdf ts->ts_cpu = sched_pickcpu(td, 0); 1849212153Smdf#endif 1850171713Sjeff if (ts->ts_cpu == cpuid) 1851177435Sjeff tdq_runq_add(tdq, td, srqflag); 1852212115Smdf else { 1853212115Smdf KASSERT(THREAD_CAN_MIGRATE(td) || 1854212115Smdf (ts->ts_flags & TSF_BOUND) != 0, 1855212115Smdf ("Thread %p shouldn't migrate", td)); 1856171713Sjeff mtx = sched_switch_migrate(tdq, td, srqflag); 1857212115Smdf } 1858171482Sjeff } else { 1859171482Sjeff /* This thread must be going to sleep. */ 1860171482Sjeff TDQ_LOCK(tdq); 1861202889Sattilio mtx = thread_lock_block(td); 1862177435Sjeff tdq_load_rem(tdq, td); 1863171482Sjeff } 1864171482Sjeff /* 1865171482Sjeff * We enter here with the thread blocked and assigned to the 1866171482Sjeff * appropriate cpu run-queue or sleep-queue and with the current 1867171482Sjeff * thread-queue locked. 1868171482Sjeff */ 1869171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1870171482Sjeff newtd = choosethread(); 1871171482Sjeff /* 1872171482Sjeff * Call the MD code to switch contexts if necessary. 1873171482Sjeff */ 1874145256Sjkoshy if (td != newtd) { 1875145256Sjkoshy#ifdef HWPMC_HOOKS 1876145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1877145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1878145256Sjkoshy#endif 1879174629Sjeff lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 1880172411Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 1881179297Sjb 1882179297Sjb#ifdef KDTRACE_HOOKS 1883179297Sjb /* 1884179297Sjb * If DTrace has set the active vtime enum to anything 1885179297Sjb * other than INACTIVE (0), then it should have set the 1886179297Sjb * function to call. 1887179297Sjb */ 1888179297Sjb if (dtrace_vtime_active) 1889179297Sjb (*dtrace_vtime_switch_func)(newtd); 1890179297Sjb#endif 1891179297Sjb 1892171482Sjeff cpu_switch(td, newtd, mtx); 1893171482Sjeff /* 1894171482Sjeff * We may return from cpu_switch on a different cpu. However, 1895171482Sjeff * we always return with td_lock pointing to the current cpu's 1896171482Sjeff * run queue lock. 1897171482Sjeff */ 1898171482Sjeff cpuid = PCPU_GET(cpuid); 1899171482Sjeff tdq = TDQ_CPU(cpuid); 1900174629Sjeff lock_profile_obtain_lock_success( 1901174629Sjeff &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 1902145256Sjkoshy#ifdef HWPMC_HOOKS 1903145256Sjkoshy if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1904145256Sjkoshy PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1905145256Sjkoshy#endif 1906171482Sjeff } else 1907171482Sjeff thread_unblock_switch(td, mtx); 1908171482Sjeff /* 1909171482Sjeff * Assert that all went well and return. 1910171482Sjeff */ 1911171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1912171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1913171482Sjeff td->td_oncpu = cpuid; 1914109864Sjeff} 1915109864Sjeff 1916171482Sjeff/* 1917171482Sjeff * Adjust thread priorities as a result of a nice request. 1918171482Sjeff */ 1919109864Sjeffvoid 1920130551Sjuliansched_nice(struct proc *p, int nice) 1921109864Sjeff{ 1922109864Sjeff struct thread *td; 1923109864Sjeff 1924130551Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 1925165762Sjeff 1926130551Sjulian p->p_nice = nice; 1927163709Sjb FOREACH_THREAD_IN_PROC(p, td) { 1928170293Sjeff thread_lock(td); 1929163709Sjb sched_priority(td); 1930165762Sjeff sched_prio(td, td->td_base_user_pri); 1931170293Sjeff thread_unlock(td); 1932130551Sjulian } 1933109864Sjeff} 1934109864Sjeff 1935171482Sjeff/* 1936171482Sjeff * Record the sleep time for the interactivity scorer. 1937171482Sjeff */ 1938109864Sjeffvoid 1939177085Sjeffsched_sleep(struct thread *td, int prio) 1940109864Sjeff{ 1941165762Sjeff 1942170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1943109864Sjeff 1944172264Sjeff td->td_slptick = ticks; 1945201347Skib if (TD_IS_SUSPENDED(td) || prio >= PSOCK) 1946177085Sjeff td->td_flags |= TDF_CANSWAP; 1947217410Sjhb if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE) 1948217410Sjhb return; 1949177903Sjeff if (static_boost == 1 && prio) 1950177085Sjeff sched_prio(td, prio); 1951177903Sjeff else if (static_boost && td->td_priority > static_boost) 1952177903Sjeff sched_prio(td, static_boost); 1953109864Sjeff} 1954109864Sjeff 1955171482Sjeff/* 1956171482Sjeff * Schedule a thread to resume execution and record how long it voluntarily 1957171482Sjeff * slept. We also update the pctcpu, interactivity, and priority. 1958171482Sjeff */ 1959109864Sjeffvoid 1960109864Sjeffsched_wakeup(struct thread *td) 1961109864Sjeff{ 1962166229Sjeff struct td_sched *ts; 1963171482Sjeff int slptick; 1964165762Sjeff 1965170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1966166229Sjeff ts = td->td_sched; 1967177085Sjeff td->td_flags &= ~TDF_CANSWAP; 1968109864Sjeff /* 1969165762Sjeff * If we slept for more than a tick update our interactivity and 1970165762Sjeff * priority. 1971109864Sjeff */ 1972172264Sjeff slptick = td->td_slptick; 1973172264Sjeff td->td_slptick = 0; 1974171482Sjeff if (slptick && slptick != ticks) { 1975166208Sjeff u_int hzticks; 1976109864Sjeff 1977171482Sjeff hzticks = (ticks - slptick) << SCHED_TICK_SHIFT; 1978171482Sjeff ts->ts_slptime += hzticks; 1979165819Sjeff sched_interact_update(td); 1980166229Sjeff sched_pctcpu_update(ts); 1981109864Sjeff } 1982166229Sjeff /* Reset the slice value after we sleep. */ 1983166229Sjeff ts->ts_slice = sched_slice; 1984166190Sjeff sched_add(td, SRQ_BORING); 1985109864Sjeff} 1986109864Sjeff 1987109864Sjeff/* 1988109864Sjeff * Penalize the parent for creating a new child and initialize the child's 1989109864Sjeff * priority. 1990109864Sjeff */ 1991109864Sjeffvoid 1992163709Sjbsched_fork(struct thread *td, struct thread *child) 1993109864Sjeff{ 1994170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 1995164936Sjulian sched_fork_thread(td, child); 1996165762Sjeff /* 1997165762Sjeff * Penalize the parent and child for forking. 1998165762Sjeff */ 1999165762Sjeff sched_interact_fork(child); 2000165762Sjeff sched_priority(child); 2001171482Sjeff td->td_sched->ts_runtime += tickincr; 2002165762Sjeff sched_interact_update(td); 2003165762Sjeff sched_priority(td); 2004164936Sjulian} 2005109864Sjeff 2006171482Sjeff/* 2007171482Sjeff * Fork a new thread, may be within the same process. 2008171482Sjeff */ 2009164936Sjulianvoid 2010164936Sjuliansched_fork_thread(struct thread *td, struct thread *child) 2011164936Sjulian{ 2012164936Sjulian struct td_sched *ts; 2013164936Sjulian struct td_sched *ts2; 2014164936Sjulian 2015177426Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2016165762Sjeff /* 2017165762Sjeff * Initialize child. 2018165762Sjeff */ 2019177426Sjeff ts = td->td_sched; 2020177426Sjeff ts2 = child->td_sched; 2021171482Sjeff child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 2022176735Sjeff child->td_cpuset = cpuset_ref(td->td_cpuset); 2023164936Sjulian ts2->ts_cpu = ts->ts_cpu; 2024177426Sjeff ts2->ts_flags = 0; 2025165762Sjeff /* 2026217078Sjhb * Grab our parents cpu estimation information. 2027165762Sjeff */ 2028164936Sjulian ts2->ts_ticks = ts->ts_ticks; 2029164936Sjulian ts2->ts_ltick = ts->ts_ltick; 2030199764Sivoras ts2->ts_incrtick = ts->ts_incrtick; 2031164936Sjulian ts2->ts_ftick = ts->ts_ftick; 2032165762Sjeff /* 2033217078Sjhb * Do not inherit any borrowed priority from the parent. 2034217078Sjhb */ 2035217078Sjhb child->td_priority = child->td_base_pri; 2036217078Sjhb /* 2037165762Sjeff * And update interactivity score. 2038165762Sjeff */ 2039171482Sjeff ts2->ts_slptime = ts->ts_slptime; 2040171482Sjeff ts2->ts_runtime = ts->ts_runtime; 2041165762Sjeff ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 2042187357Sjeff#ifdef KTR 2043187357Sjeff bzero(ts2->ts_name, sizeof(ts2->ts_name)); 2044187357Sjeff#endif 2045113357Sjeff} 2046113357Sjeff 2047171482Sjeff/* 2048171482Sjeff * Adjust the priority class of a thread. 2049171482Sjeff */ 2050113357Sjeffvoid 2051163709Sjbsched_class(struct thread *td, int class) 2052113357Sjeff{ 2053113357Sjeff 2054170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2055163709Sjb if (td->td_pri_class == class) 2056113357Sjeff return; 2057163709Sjb td->td_pri_class = class; 2058109864Sjeff} 2059109864Sjeff 2060109864Sjeff/* 2061109864Sjeff * Return some of the child's priority and interactivity to the parent. 2062109864Sjeff */ 2063109864Sjeffvoid 2064164939Sjuliansched_exit(struct proc *p, struct thread *child) 2065109864Sjeff{ 2066165762Sjeff struct thread *td; 2067113372Sjeff 2068187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit", 2069225199Sdelphij "prio:%d", child->td_priority); 2070177368Sjeff PROC_LOCK_ASSERT(p, MA_OWNED); 2071165762Sjeff td = FIRST_THREAD_IN_PROC(p); 2072165762Sjeff sched_exit_thread(td, child); 2073113372Sjeff} 2074113372Sjeff 2075171482Sjeff/* 2076171482Sjeff * Penalize another thread for the time spent on this one. This helps to 2077171482Sjeff * worsen the priority and interactivity of processes which schedule batch 2078171482Sjeff * jobs such as make. This has little effect on the make process itself but 2079171482Sjeff * causes new processes spawned by it to receive worse scores immediately. 2080171482Sjeff */ 2081113372Sjeffvoid 2082164939Sjuliansched_exit_thread(struct thread *td, struct thread *child) 2083164936Sjulian{ 2084165762Sjeff 2085187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit", 2086225199Sdelphij "prio:%d", child->td_priority); 2087165762Sjeff /* 2088165762Sjeff * Give the child's runtime to the parent without returning the 2089165762Sjeff * sleep time as a penalty to the parent. This causes shells that 2090165762Sjeff * launch expensive things to mark their children as expensive. 2091165762Sjeff */ 2092170293Sjeff thread_lock(td); 2093171482Sjeff td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2094164939Sjulian sched_interact_update(td); 2095165762Sjeff sched_priority(td); 2096170293Sjeff thread_unlock(td); 2097164936Sjulian} 2098164936Sjulian 2099177005Sjeffvoid 2100177005Sjeffsched_preempt(struct thread *td) 2101177005Sjeff{ 2102177005Sjeff struct tdq *tdq; 2103177005Sjeff 2104177005Sjeff thread_lock(td); 2105177005Sjeff tdq = TDQ_SELF(); 2106177005Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2107177005Sjeff tdq->tdq_ipipending = 0; 2108177005Sjeff if (td->td_priority > tdq->tdq_lowpri) { 2109178272Sjeff int flags; 2110178272Sjeff 2111178272Sjeff flags = SW_INVOL | SW_PREEMPT; 2112177005Sjeff if (td->td_critnest > 1) 2113177005Sjeff td->td_owepreempt = 1; 2114178272Sjeff else if (TD_IS_IDLETHREAD(td)) 2115178272Sjeff mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL); 2116177005Sjeff else 2117178272Sjeff mi_switch(flags | SWT_REMOTEPREEMPT, NULL); 2118177005Sjeff } 2119177005Sjeff thread_unlock(td); 2120177005Sjeff} 2121177005Sjeff 2122171482Sjeff/* 2123171482Sjeff * Fix priorities on return to user-space. Priorities may be elevated due 2124171482Sjeff * to static priorities in msleep() or similar. 2125171482Sjeff */ 2126164936Sjulianvoid 2127164936Sjuliansched_userret(struct thread *td) 2128164936Sjulian{ 2129164936Sjulian /* 2130164936Sjulian * XXX we cheat slightly on the locking here to avoid locking in 2131164936Sjulian * the usual case. Setting td_priority here is essentially an 2132164936Sjulian * incomplete workaround for not setting it properly elsewhere. 2133164936Sjulian * Now that some interrupt handlers are threads, not setting it 2134164936Sjulian * properly elsewhere can clobber it in the window between setting 2135164936Sjulian * it here and returning to user mode, so don't waste time setting 2136164936Sjulian * it perfectly here. 2137164936Sjulian */ 2138164936Sjulian KASSERT((td->td_flags & TDF_BORROWING) == 0, 2139164936Sjulian ("thread with borrowed priority returning to userland")); 2140164936Sjulian if (td->td_priority != td->td_user_pri) { 2141170293Sjeff thread_lock(td); 2142164936Sjulian td->td_priority = td->td_user_pri; 2143164936Sjulian td->td_base_pri = td->td_user_pri; 2144177005Sjeff tdq_setlowpri(TDQ_SELF(), td); 2145170293Sjeff thread_unlock(td); 2146164936Sjulian } 2147164936Sjulian} 2148164936Sjulian 2149171482Sjeff/* 2150171482Sjeff * Handle a stathz tick. This is really only relevant for timeshare 2151171482Sjeff * threads. 2152171482Sjeff */ 2153164936Sjulianvoid 2154121127Sjeffsched_clock(struct thread *td) 2155109864Sjeff{ 2156164936Sjulian struct tdq *tdq; 2157164936Sjulian struct td_sched *ts; 2158109864Sjeff 2159171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2160164936Sjulian tdq = TDQ_SELF(); 2161172409Sjeff#ifdef SMP 2162133427Sjeff /* 2163172409Sjeff * We run the long term load balancer infrequently on the first cpu. 2164172409Sjeff */ 2165172409Sjeff if (balance_tdq == tdq) { 2166172409Sjeff if (balance_ticks && --balance_ticks == 0) 2167172409Sjeff sched_balance(); 2168172409Sjeff } 2169172409Sjeff#endif 2170172409Sjeff /* 2171178277Sjeff * Save the old switch count so we have a record of the last ticks 2172178277Sjeff * activity. Initialize the new switch count based on our load. 2173178277Sjeff * If there is some activity seed it to reflect that. 2174178277Sjeff */ 2175178277Sjeff tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt; 2176178471Sjeff tdq->tdq_switchcnt = tdq->tdq_load; 2177178277Sjeff /* 2178165766Sjeff * Advance the insert index once for each tick to ensure that all 2179165766Sjeff * threads get a chance to run. 2180133427Sjeff */ 2181165766Sjeff if (tdq->tdq_idx == tdq->tdq_ridx) { 2182165766Sjeff tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2183165766Sjeff if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2184165766Sjeff tdq->tdq_ridx = tdq->tdq_idx; 2185165766Sjeff } 2186165766Sjeff ts = td->td_sched; 2187175104Sjeff if (td->td_pri_class & PRI_FIFO_BIT) 2188113357Sjeff return; 2189217291Sjhb if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) { 2190175104Sjeff /* 2191175104Sjeff * We used a tick; charge it to the thread so 2192175104Sjeff * that we can compute our interactivity. 2193175104Sjeff */ 2194175104Sjeff td->td_sched->ts_runtime += tickincr; 2195175104Sjeff sched_interact_update(td); 2196177009Sjeff sched_priority(td); 2197175104Sjeff } 2198113357Sjeff /* 2199109864Sjeff * We used up one time slice. 2200109864Sjeff */ 2201164936Sjulian if (--ts->ts_slice > 0) 2202113357Sjeff return; 2203109864Sjeff /* 2204177009Sjeff * We're out of time, force a requeue at userret(). 2205109864Sjeff */ 2206177009Sjeff ts->ts_slice = sched_slice; 2207113357Sjeff td->td_flags |= TDF_NEEDRESCHED; 2208109864Sjeff} 2209109864Sjeff 2210171482Sjeff/* 2211171482Sjeff * Called once per hz tick. Used for cpu utilization information. This 2212171482Sjeff * is easier than trying to scale based on stathz. 2213171482Sjeff */ 2214171482Sjeffvoid 2215212541Smavsched_tick(int cnt) 2216171482Sjeff{ 2217171482Sjeff struct td_sched *ts; 2218171482Sjeff 2219171482Sjeff ts = curthread->td_sched; 2220180607Sjeff /* 2221180607Sjeff * Ticks is updated asynchronously on a single cpu. Check here to 2222180607Sjeff * avoid incrementing ts_ticks multiple times in a single tick. 2223180607Sjeff */ 2224199764Sivoras if (ts->ts_incrtick == ticks) 2225180607Sjeff return; 2226171482Sjeff /* Adjust ticks for pctcpu */ 2227212541Smav ts->ts_ticks += cnt << SCHED_TICK_SHIFT; 2228171482Sjeff ts->ts_ltick = ticks; 2229199764Sivoras ts->ts_incrtick = ticks; 2230171482Sjeff /* 2231215102Sattilio * Update if we've exceeded our desired tick threshold by over one 2232171482Sjeff * second. 2233171482Sjeff */ 2234171482Sjeff if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2235171482Sjeff sched_pctcpu_update(ts); 2236171482Sjeff} 2237171482Sjeff 2238171482Sjeff/* 2239171482Sjeff * Return whether the current CPU has runnable tasks. Used for in-kernel 2240171482Sjeff * cooperative idle threads. 2241171482Sjeff */ 2242109864Sjeffint 2243109864Sjeffsched_runnable(void) 2244109864Sjeff{ 2245164936Sjulian struct tdq *tdq; 2246115998Sjeff int load; 2247109864Sjeff 2248115998Sjeff load = 1; 2249115998Sjeff 2250164936Sjulian tdq = TDQ_SELF(); 2251121605Sjeff if ((curthread->td_flags & TDF_IDLETD) != 0) { 2252165620Sjeff if (tdq->tdq_load > 0) 2253121605Sjeff goto out; 2254121605Sjeff } else 2255165620Sjeff if (tdq->tdq_load - 1 > 0) 2256121605Sjeff goto out; 2257115998Sjeff load = 0; 2258115998Sjeffout: 2259115998Sjeff return (load); 2260109864Sjeff} 2261109864Sjeff 2262171482Sjeff/* 2263171482Sjeff * Choose the highest priority thread to run. The thread is removed from 2264171482Sjeff * the run-queue while running however the load remains. For SMP we set 2265171482Sjeff * the tdq in the global idle bitmask if it idles here. 2266171482Sjeff */ 2267166190Sjeffstruct thread * 2268109970Sjeffsched_choose(void) 2269109970Sjeff{ 2270177435Sjeff struct thread *td; 2271164936Sjulian struct tdq *tdq; 2272109970Sjeff 2273164936Sjulian tdq = TDQ_SELF(); 2274171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2275177435Sjeff td = tdq_choose(tdq); 2276177435Sjeff if (td) { 2277177435Sjeff td->td_sched->ts_ltick = ticks; 2278177435Sjeff tdq_runq_rem(tdq, td); 2279177903Sjeff tdq->tdq_lowpri = td->td_priority; 2280177435Sjeff return (td); 2281109864Sjeff } 2282177903Sjeff tdq->tdq_lowpri = PRI_MAX_IDLE; 2283176735Sjeff return (PCPU_GET(idlethread)); 2284109864Sjeff} 2285109864Sjeff 2286171482Sjeff/* 2287171482Sjeff * Set owepreempt if necessary. Preemption never happens directly in ULE, 2288171482Sjeff * we always request it once we exit a critical section. 2289171482Sjeff */ 2290171482Sjeffstatic inline void 2291171482Sjeffsched_setpreempt(struct thread *td) 2292166190Sjeff{ 2293166190Sjeff struct thread *ctd; 2294166190Sjeff int cpri; 2295166190Sjeff int pri; 2296166190Sjeff 2297177005Sjeff THREAD_LOCK_ASSERT(curthread, MA_OWNED); 2298177005Sjeff 2299166190Sjeff ctd = curthread; 2300166190Sjeff pri = td->td_priority; 2301166190Sjeff cpri = ctd->td_priority; 2302177005Sjeff if (pri < cpri) 2303177005Sjeff ctd->td_flags |= TDF_NEEDRESCHED; 2304166190Sjeff if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2305171482Sjeff return; 2306177005Sjeff if (!sched_shouldpreempt(pri, cpri, 0)) 2307171482Sjeff return; 2308171482Sjeff ctd->td_owepreempt = 1; 2309166190Sjeff} 2310166190Sjeff 2311171482Sjeff/* 2312177009Sjeff * Add a thread to a thread queue. Select the appropriate runq and add the 2313177009Sjeff * thread to it. This is the internal function called when the tdq is 2314177009Sjeff * predetermined. 2315171482Sjeff */ 2316109864Sjeffvoid 2317171482Sjefftdq_add(struct tdq *tdq, struct thread *td, int flags) 2318109864Sjeff{ 2319109864Sjeff 2320171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2321166190Sjeff KASSERT((td->td_inhibitors == 0), 2322166190Sjeff ("sched_add: trying to run inhibited thread")); 2323166190Sjeff KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2324166190Sjeff ("sched_add: bad thread state")); 2325172207Sjeff KASSERT(td->td_flags & TDF_INMEM, 2326172207Sjeff ("sched_add: thread swapped out")); 2327171482Sjeff 2328171482Sjeff if (td->td_priority < tdq->tdq_lowpri) 2329171482Sjeff tdq->tdq_lowpri = td->td_priority; 2330177435Sjeff tdq_runq_add(tdq, td, flags); 2331177435Sjeff tdq_load_add(tdq, td); 2332171482Sjeff} 2333171482Sjeff 2334171482Sjeff/* 2335171482Sjeff * Select the target thread queue and add a thread to it. Request 2336171482Sjeff * preemption or IPI a remote processor if required. 2337171482Sjeff */ 2338171482Sjeffvoid 2339171482Sjeffsched_add(struct thread *td, int flags) 2340171482Sjeff{ 2341171482Sjeff struct tdq *tdq; 2342171482Sjeff#ifdef SMP 2343171482Sjeff int cpu; 2344171482Sjeff#endif 2345187357Sjeff 2346187357Sjeff KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", 2347187357Sjeff "prio:%d", td->td_priority, KTR_ATTR_LINKED, 2348187357Sjeff sched_tdname(curthread)); 2349187357Sjeff KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", 2350187357Sjeff KTR_ATTR_LINKED, sched_tdname(td)); 2351171482Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2352166108Sjeff /* 2353171482Sjeff * Recalculate the priority before we select the target cpu or 2354171482Sjeff * run-queue. 2355166108Sjeff */ 2356171482Sjeff if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2357171482Sjeff sched_priority(td); 2358171482Sjeff#ifdef SMP 2359171482Sjeff /* 2360171482Sjeff * Pick the destination cpu and if it isn't ours transfer to the 2361171482Sjeff * target cpu. 2362171482Sjeff */ 2363177435Sjeff cpu = sched_pickcpu(td, flags); 2364177435Sjeff tdq = sched_setcpu(td, cpu, flags); 2365171482Sjeff tdq_add(tdq, td, flags); 2366177009Sjeff if (cpu != PCPU_GET(cpuid)) { 2367177435Sjeff tdq_notify(tdq, td); 2368166108Sjeff return; 2369166108Sjeff } 2370171482Sjeff#else 2371171482Sjeff tdq = TDQ_SELF(); 2372171482Sjeff TDQ_LOCK(tdq); 2373171482Sjeff /* 2374171482Sjeff * Now that the thread is moving to the run-queue, set the lock 2375171482Sjeff * to the scheduler's lock. 2376171482Sjeff */ 2377171482Sjeff thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2378171482Sjeff tdq_add(tdq, td, flags); 2379166108Sjeff#endif 2380171482Sjeff if (!(flags & SRQ_YIELDING)) 2381171482Sjeff sched_setpreempt(td); 2382109864Sjeff} 2383109864Sjeff 2384171482Sjeff/* 2385171482Sjeff * Remove a thread from a run-queue without running it. This is used 2386171482Sjeff * when we're stealing a thread from a remote queue. Otherwise all threads 2387171482Sjeff * exit by calling sched_exit_thread() and sched_throw() themselves. 2388171482Sjeff */ 2389109864Sjeffvoid 2390121127Sjeffsched_rem(struct thread *td) 2391109864Sjeff{ 2392164936Sjulian struct tdq *tdq; 2393113357Sjeff 2394187357Sjeff KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem", 2395187357Sjeff "prio:%d", td->td_priority); 2396177435Sjeff tdq = TDQ_CPU(td->td_sched->ts_cpu); 2397171482Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2398171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2399166190Sjeff KASSERT(TD_ON_RUNQ(td), 2400164936Sjulian ("sched_rem: thread not on run queue")); 2401177435Sjeff tdq_runq_rem(tdq, td); 2402177435Sjeff tdq_load_rem(tdq, td); 2403166190Sjeff TD_SET_CAN_RUN(td); 2404176735Sjeff if (td->td_priority == tdq->tdq_lowpri) 2405176735Sjeff tdq_setlowpri(tdq, NULL); 2406109864Sjeff} 2407109864Sjeff 2408171482Sjeff/* 2409171482Sjeff * Fetch cpu utilization information. Updates on demand. 2410171482Sjeff */ 2411109864Sjefffixpt_t 2412121127Sjeffsched_pctcpu(struct thread *td) 2413109864Sjeff{ 2414109864Sjeff fixpt_t pctcpu; 2415164936Sjulian struct td_sched *ts; 2416109864Sjeff 2417109864Sjeff pctcpu = 0; 2418164936Sjulian ts = td->td_sched; 2419164936Sjulian if (ts == NULL) 2420121290Sjeff return (0); 2421109864Sjeff 2422208787Sjhb THREAD_LOCK_ASSERT(td, MA_OWNED); 2423164936Sjulian if (ts->ts_ticks) { 2424109864Sjeff int rtick; 2425109864Sjeff 2426165796Sjeff sched_pctcpu_update(ts); 2427109864Sjeff /* How many rtick per second ? */ 2428165762Sjeff rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2429165762Sjeff pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2430109864Sjeff } 2431109864Sjeff 2432109864Sjeff return (pctcpu); 2433109864Sjeff} 2434109864Sjeff 2435176735Sjeff/* 2436176735Sjeff * Enforce affinity settings for a thread. Called after adjustments to 2437176735Sjeff * cpumask. 2438176735Sjeff */ 2439176729Sjeffvoid 2440176729Sjeffsched_affinity(struct thread *td) 2441176729Sjeff{ 2442176735Sjeff#ifdef SMP 2443176735Sjeff struct td_sched *ts; 2444176735Sjeff 2445176735Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2446176735Sjeff ts = td->td_sched; 2447176735Sjeff if (THREAD_CAN_SCHED(td, ts->ts_cpu)) 2448176735Sjeff return; 2449189787Sjeff if (TD_ON_RUNQ(td)) { 2450189787Sjeff sched_rem(td); 2451189787Sjeff sched_add(td, SRQ_BORING); 2452189787Sjeff return; 2453189787Sjeff } 2454176735Sjeff if (!TD_IS_RUNNING(td)) 2455176735Sjeff return; 2456176735Sjeff /* 2457212115Smdf * Force a switch before returning to userspace. If the 2458212115Smdf * target thread is not running locally send an ipi to force 2459212115Smdf * the issue. 2460176735Sjeff */ 2461212974Sjhb td->td_flags |= TDF_NEEDRESCHED; 2462212115Smdf if (td != curthread) 2463212115Smdf ipi_cpu(ts->ts_cpu, IPI_PREEMPT); 2464176735Sjeff#endif 2465176729Sjeff} 2466176729Sjeff 2467171482Sjeff/* 2468171482Sjeff * Bind a thread to a target cpu. 2469171482Sjeff */ 2470122038Sjeffvoid 2471122038Sjeffsched_bind(struct thread *td, int cpu) 2472122038Sjeff{ 2473164936Sjulian struct td_sched *ts; 2474122038Sjeff 2475171713Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2476208391Sjhb KASSERT(td == curthread, ("sched_bind: can only bind curthread")); 2477164936Sjulian ts = td->td_sched; 2478166137Sjeff if (ts->ts_flags & TSF_BOUND) 2479166152Sjeff sched_unbind(td); 2480212115Smdf KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td)); 2481164936Sjulian ts->ts_flags |= TSF_BOUND; 2482166137Sjeff sched_pin(); 2483123433Sjeff if (PCPU_GET(cpuid) == cpu) 2484122038Sjeff return; 2485166137Sjeff ts->ts_cpu = cpu; 2486122038Sjeff /* When we return from mi_switch we'll be on the correct cpu. */ 2487131527Sphk mi_switch(SW_VOL, NULL); 2488122038Sjeff} 2489122038Sjeff 2490171482Sjeff/* 2491171482Sjeff * Release a bound thread. 2492171482Sjeff */ 2493122038Sjeffvoid 2494122038Sjeffsched_unbind(struct thread *td) 2495122038Sjeff{ 2496165762Sjeff struct td_sched *ts; 2497165762Sjeff 2498170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2499208391Sjhb KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); 2500165762Sjeff ts = td->td_sched; 2501166137Sjeff if ((ts->ts_flags & TSF_BOUND) == 0) 2502166137Sjeff return; 2503165762Sjeff ts->ts_flags &= ~TSF_BOUND; 2504165762Sjeff sched_unpin(); 2505122038Sjeff} 2506122038Sjeff 2507109864Sjeffint 2508145256Sjkoshysched_is_bound(struct thread *td) 2509145256Sjkoshy{ 2510170293Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 2511164936Sjulian return (td->td_sched->ts_flags & TSF_BOUND); 2512145256Sjkoshy} 2513145256Sjkoshy 2514171482Sjeff/* 2515171482Sjeff * Basic yield call. 2516171482Sjeff */ 2517159630Sdavidxuvoid 2518159630Sdavidxusched_relinquish(struct thread *td) 2519159630Sdavidxu{ 2520170293Sjeff thread_lock(td); 2521178272Sjeff mi_switch(SW_VOL | SWT_RELINQUISH, NULL); 2522170293Sjeff thread_unlock(td); 2523159630Sdavidxu} 2524159630Sdavidxu 2525171482Sjeff/* 2526171482Sjeff * Return the total system load. 2527171482Sjeff */ 2528145256Sjkoshyint 2529125289Sjeffsched_load(void) 2530125289Sjeff{ 2531125289Sjeff#ifdef SMP 2532125289Sjeff int total; 2533125289Sjeff int i; 2534125289Sjeff 2535125289Sjeff total = 0; 2536209059Sjhb CPU_FOREACH(i) 2537176735Sjeff total += TDQ_CPU(i)->tdq_sysload; 2538125289Sjeff return (total); 2539125289Sjeff#else 2540165620Sjeff return (TDQ_SELF()->tdq_sysload); 2541125289Sjeff#endif 2542125289Sjeff} 2543125289Sjeff 2544125289Sjeffint 2545109864Sjeffsched_sizeof_proc(void) 2546109864Sjeff{ 2547109864Sjeff return (sizeof(struct proc)); 2548109864Sjeff} 2549109864Sjeff 2550109864Sjeffint 2551109864Sjeffsched_sizeof_thread(void) 2552109864Sjeff{ 2553109864Sjeff return (sizeof(struct thread) + sizeof(struct td_sched)); 2554109864Sjeff} 2555159570Sdavidxu 2556191676Sjeff#ifdef SMP 2557191676Sjeff#define TDQ_IDLESPIN(tdq) \ 2558191676Sjeff ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0) 2559191676Sjeff#else 2560191676Sjeff#define TDQ_IDLESPIN(tdq) 1 2561191676Sjeff#endif 2562191676Sjeff 2563166190Sjeff/* 2564166190Sjeff * The actual idle process. 2565166190Sjeff */ 2566166190Sjeffvoid 2567166190Sjeffsched_idletd(void *dummy) 2568166190Sjeff{ 2569166190Sjeff struct thread *td; 2570171482Sjeff struct tdq *tdq; 2571178277Sjeff int switchcnt; 2572178277Sjeff int i; 2573166190Sjeff 2574191643Sjeff mtx_assert(&Giant, MA_NOTOWNED); 2575166190Sjeff td = curthread; 2576171482Sjeff tdq = TDQ_SELF(); 2577171482Sjeff for (;;) { 2578171482Sjeff#ifdef SMP 2579178277Sjeff if (tdq_idled(tdq) == 0) 2580178277Sjeff continue; 2581171482Sjeff#endif 2582178277Sjeff switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2583178277Sjeff /* 2584178277Sjeff * If we're switching very frequently, spin while checking 2585178277Sjeff * for load rather than entering a low power state that 2586191643Sjeff * may require an IPI. However, don't do any busy 2587191643Sjeff * loops while on SMT machines as this simply steals 2588191643Sjeff * cycles from cores doing useful work. 2589178277Sjeff */ 2590191676Sjeff if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) { 2591178277Sjeff for (i = 0; i < sched_idlespins; i++) { 2592178277Sjeff if (tdq->tdq_load) 2593178277Sjeff break; 2594178277Sjeff cpu_spinwait(); 2595178277Sjeff } 2596178277Sjeff } 2597191643Sjeff switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt; 2598212416Smav if (tdq->tdq_load == 0) { 2599212416Smav tdq->tdq_cpu_idle = 1; 2600212416Smav if (tdq->tdq_load == 0) { 2601212541Smav cpu_idle(switchcnt > sched_idlespinthresh * 4); 2602212416Smav tdq->tdq_switchcnt++; 2603212416Smav } 2604212416Smav tdq->tdq_cpu_idle = 0; 2605212416Smav } 2606178277Sjeff if (tdq->tdq_load) { 2607178277Sjeff thread_lock(td); 2608178277Sjeff mi_switch(SW_VOL | SWT_IDLE, NULL); 2609178277Sjeff thread_unlock(td); 2610178277Sjeff } 2611171482Sjeff } 2612166190Sjeff} 2613166190Sjeff 2614170293Sjeff/* 2615170293Sjeff * A CPU is entering for the first time or a thread is exiting. 2616170293Sjeff */ 2617170293Sjeffvoid 2618170293Sjeffsched_throw(struct thread *td) 2619170293Sjeff{ 2620172411Sjeff struct thread *newtd; 2621171482Sjeff struct tdq *tdq; 2622171482Sjeff 2623171482Sjeff tdq = TDQ_SELF(); 2624170293Sjeff if (td == NULL) { 2625171482Sjeff /* Correct spinlock nesting and acquire the correct lock. */ 2626171482Sjeff TDQ_LOCK(tdq); 2627170293Sjeff spinlock_exit(); 2628229429Sjhb PCPU_SET(switchtime, cpu_ticks()); 2629229429Sjhb PCPU_SET(switchticks, ticks); 2630170293Sjeff } else { 2631171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2632177435Sjeff tdq_load_rem(tdq, td); 2633174629Sjeff lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object); 2634170293Sjeff } 2635170293Sjeff KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2636172411Sjeff newtd = choosethread(); 2637172411Sjeff TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd; 2638172411Sjeff cpu_throw(td, newtd); /* doesn't return */ 2639170293Sjeff} 2640170293Sjeff 2641171482Sjeff/* 2642171482Sjeff * This is called from fork_exit(). Just acquire the correct locks and 2643171482Sjeff * let fork do the rest of the work. 2644171482Sjeff */ 2645170293Sjeffvoid 2646170600Sjeffsched_fork_exit(struct thread *td) 2647170293Sjeff{ 2648171482Sjeff struct td_sched *ts; 2649171482Sjeff struct tdq *tdq; 2650171482Sjeff int cpuid; 2651170293Sjeff 2652170293Sjeff /* 2653170293Sjeff * Finish setting up thread glue so that it begins execution in a 2654171482Sjeff * non-nested critical section with the scheduler lock held. 2655170293Sjeff */ 2656171482Sjeff cpuid = PCPU_GET(cpuid); 2657171482Sjeff tdq = TDQ_CPU(cpuid); 2658171482Sjeff ts = td->td_sched; 2659171482Sjeff if (TD_IS_IDLETHREAD(td)) 2660171482Sjeff td->td_lock = TDQ_LOCKPTR(tdq); 2661171482Sjeff MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2662171482Sjeff td->td_oncpu = cpuid; 2663172411Sjeff TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 2664174629Sjeff lock_profile_obtain_lock_success( 2665174629Sjeff &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__); 2666170293Sjeff} 2667170293Sjeff 2668187357Sjeff/* 2669187357Sjeff * Create on first use to catch odd startup conditons. 2670187357Sjeff */ 2671187357Sjeffchar * 2672187357Sjeffsched_tdname(struct thread *td) 2673187357Sjeff{ 2674187357Sjeff#ifdef KTR 2675187357Sjeff struct td_sched *ts; 2676187357Sjeff 2677187357Sjeff ts = td->td_sched; 2678187357Sjeff if (ts->ts_name[0] == '\0') 2679187357Sjeff snprintf(ts->ts_name, sizeof(ts->ts_name), 2680187357Sjeff "%s tid %d", td->td_name, td->td_tid); 2681187357Sjeff return (ts->ts_name); 2682187357Sjeff#else 2683187357Sjeff return (td->td_name); 2684187357Sjeff#endif 2685187357Sjeff} 2686187357Sjeff 2687184439Sivoras#ifdef SMP 2688184439Sivoras 2689184439Sivoras/* 2690184439Sivoras * Build the CPU topology dump string. Is recursively called to collect 2691184439Sivoras * the topology tree. 2692184439Sivoras */ 2693184439Sivorasstatic int 2694184439Sivorassysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg, 2695184439Sivoras int indent) 2696184439Sivoras{ 2697222813Sattilio char cpusetbuf[CPUSETBUFSIZ]; 2698184439Sivoras int i, first; 2699184439Sivoras 2700184439Sivoras sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent, 2701212821Savg "", 1 + indent / 2, cg->cg_level); 2702222813Sattilio sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "", 2703222813Sattilio cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask)); 2704184439Sivoras first = TRUE; 2705184439Sivoras for (i = 0; i < MAXCPU; i++) { 2706222813Sattilio if (CPU_ISSET(i, &cg->cg_mask)) { 2707184439Sivoras if (!first) 2708184439Sivoras sbuf_printf(sb, ", "); 2709184439Sivoras else 2710184439Sivoras first = FALSE; 2711184439Sivoras sbuf_printf(sb, "%d", i); 2712184439Sivoras } 2713184439Sivoras } 2714184439Sivoras sbuf_printf(sb, "</cpu>\n"); 2715184439Sivoras 2716184439Sivoras if (cg->cg_flags != 0) { 2717210117Sivoras sbuf_printf(sb, "%*s <flags>", indent, ""); 2718184439Sivoras if ((cg->cg_flags & CG_FLAG_HTT) != 0) 2719208982Sivoras sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>"); 2720208983Sivoras if ((cg->cg_flags & CG_FLAG_THREAD) != 0) 2721208983Sivoras sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>"); 2722191643Sjeff if ((cg->cg_flags & CG_FLAG_SMT) != 0) 2723208983Sivoras sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>"); 2724210117Sivoras sbuf_printf(sb, "</flags>\n"); 2725184439Sivoras } 2726184439Sivoras 2727184439Sivoras if (cg->cg_children > 0) { 2728184439Sivoras sbuf_printf(sb, "%*s <children>\n", indent, ""); 2729184439Sivoras for (i = 0; i < cg->cg_children; i++) 2730184439Sivoras sysctl_kern_sched_topology_spec_internal(sb, 2731184439Sivoras &cg->cg_child[i], indent+2); 2732184439Sivoras sbuf_printf(sb, "%*s </children>\n", indent, ""); 2733184439Sivoras } 2734184439Sivoras sbuf_printf(sb, "%*s</group>\n", indent, ""); 2735184439Sivoras return (0); 2736184439Sivoras} 2737184439Sivoras 2738184439Sivoras/* 2739184439Sivoras * Sysctl handler for retrieving topology dump. It's a wrapper for 2740184439Sivoras * the recursive sysctl_kern_smp_topology_spec_internal(). 2741184439Sivoras */ 2742184439Sivorasstatic int 2743184439Sivorassysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS) 2744184439Sivoras{ 2745184439Sivoras struct sbuf *topo; 2746184439Sivoras int err; 2747184439Sivoras 2748184439Sivoras KASSERT(cpu_top != NULL, ("cpu_top isn't initialized")); 2749184439Sivoras 2750184570Sivoras topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND); 2751184439Sivoras if (topo == NULL) 2752184439Sivoras return (ENOMEM); 2753184439Sivoras 2754184439Sivoras sbuf_printf(topo, "<groups>\n"); 2755184439Sivoras err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1); 2756184439Sivoras sbuf_printf(topo, "</groups>\n"); 2757184439Sivoras 2758184439Sivoras if (err == 0) { 2759184439Sivoras sbuf_finish(topo); 2760184439Sivoras err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo)); 2761184439Sivoras } 2762184439Sivoras sbuf_delete(topo); 2763184439Sivoras return (err); 2764184439Sivoras} 2765214510Sdavidxu 2766184439Sivoras#endif 2767184439Sivoras 2768177435SjeffSYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler"); 2769171482SjeffSYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2770165762Sjeff "Scheduler name"); 2771171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2772171482Sjeff "Slice size for timeshare threads"); 2773171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2774171482Sjeff "Interactivity score threshold"); 2775171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2776171482Sjeff 0,"Min priority for preemption, lower priorities have greater precedence"); 2777177085SjeffSYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 2778177085Sjeff 0,"Controls whether static kernel priorities are assigned to sleeping threads."); 2779178277SjeffSYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 2780178277Sjeff 0,"Number of times idle will spin waiting for new work."); 2781178277SjeffSYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, &sched_idlespinthresh, 2782178277Sjeff 0,"Threshold before we will permit idle spinning."); 2783166108Sjeff#ifdef SMP 2784171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2785171482Sjeff "Number of hz ticks to keep thread affinity for"); 2786171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2787171482Sjeff "Enables the long-term load balancer"); 2788172409SjeffSYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW, 2789172409Sjeff &balance_interval, 0, 2790172409Sjeff "Average frequency in stathz ticks to run the long-term balancer"); 2791171482SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2792171482Sjeff "Attempts to steal work from other cores before idling"); 2793171506SjeffSYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2794171506Sjeff "Minimum load on remote cpu before we'll steal"); 2795184439Sivoras 2796184439Sivoras/* Retrieve SMP topology */ 2797184439SivorasSYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING | 2798184439Sivoras CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A", 2799184439Sivoras "XML dump of detected CPU topology"); 2800214510Sdavidxu 2801166108Sjeff#endif 2802165762Sjeff 2803172264Sjeff/* ps compat. All cpu percentages from ULE are weighted. */ 2804172293Sjeffstatic int ccpu = 0; 2805165762SjeffSYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2806