sched_ule.c revision 171715
11573Srgrimes/*- 214287Spst * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org> 31573Srgrimes * All rights reserved. 41573Srgrimes * 51573Srgrimes * Redistribution and use in source and binary forms, with or without 61573Srgrimes * modification, are permitted provided that the following conditions 71573Srgrimes * are met: 81573Srgrimes * 1. Redistributions of source code must retain the above copyright 91573Srgrimes * notice unmodified, this list of conditions, and the following 101573Srgrimes * disclaimer. 111573Srgrimes * 2. Redistributions in binary form must reproduce the above copyright 121573Srgrimes * notice, this list of conditions and the following disclaimer in the 131573Srgrimes * documentation and/or other materials provided with the distribution. 141573Srgrimes * 151573Srgrimes * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 161573Srgrimes * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 171573Srgrimes * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 181573Srgrimes * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 191573Srgrimes * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 201573Srgrimes * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 211573Srgrimes * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 221573Srgrimes * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 231573Srgrimes * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 241573Srgrimes * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 251573Srgrimes */ 261573Srgrimes 271573Srgrimes/* 281573Srgrimes * This file implements the ULE scheduler. ULE supports independent CPU 291573Srgrimes * run queues and fine grain locking. It has superior interactive 301573Srgrimes * performance under load even on uni-processor systems. 311573Srgrimes * 321573Srgrimes * etymology: 331573Srgrimes * ULE is the last three letters in schedule. It owes it's name to a 3414287Spst * generic user created for a scheduling system by Paul Mikesell at 351573Srgrimes * Isilon Systems and a general lack of creativity on the part of the author. 3692889Sobrien */ 3792889Sobrien 381573Srgrimes#include <sys/cdefs.h> 391573Srgrimes__FBSDID("$FreeBSD: head/sys/kern/sched_ule.c 171715 2007-08-04 01:21:28Z jeff $"); 401573Srgrimes 411573Srgrimes#include "opt_hwpmc_hooks.h" 421573Srgrimes#include "opt_sched.h" 431573Srgrimes 441573Srgrimes#include <sys/param.h> 451573Srgrimes#include <sys/systm.h> 461573Srgrimes#include <sys/kdb.h> 471573Srgrimes#include <sys/kernel.h> 481573Srgrimes#include <sys/ktr.h> 491573Srgrimes#include <sys/lock.h> 501573Srgrimes#include <sys/mutex.h> 511573Srgrimes#include <sys/proc.h> 521573Srgrimes#include <sys/resource.h> 531573Srgrimes#include <sys/resourcevar.h> 541573Srgrimes#include <sys/sched.h> 5571579Sdeischen#include <sys/smp.h> 56190485Sdelphij#include <sys/sx.h> 571573Srgrimes#include <sys/sysctl.h> 581573Srgrimes#include <sys/sysproto.h> 591573Srgrimes#include <sys/turnstile.h> 601573Srgrimes#include <sys/umtx.h> 611573Srgrimes#include <sys/vmmeter.h> 621573Srgrimes#ifdef KTRACE 631573Srgrimes#include <sys/uio.h> 641573Srgrimes#include <sys/ktrace.h> 651573Srgrimes#endif 661573Srgrimes 671573Srgrimes#ifdef HWPMC_HOOKS 6871579Sdeischen#include <sys/pmckern.h> 691573Srgrimes#endif 701573Srgrimes 711573Srgrimes#include <machine/cpu.h> 721573Srgrimes#include <machine/smp.h> 731573Srgrimes 741573Srgrimes#ifndef PREEMPTION 75189291Sdelphij#error "SCHED_ULE requires options PREEMPTION" 76189291Sdelphij#endif 77189291Sdelphij 78189291Sdelphij#define KTR_ULE 0 79189291Sdelphij 80189291Sdelphij/* 81189291Sdelphij * Thread scheduler specific section. All fields are protected 821573Srgrimes * by the thread lock. 831573Srgrimes */ 8414287Spststruct td_sched { 8514287Spst TAILQ_ENTRY(td_sched) ts_procq; /* Run queue. */ 8614287Spst struct thread *ts_thread; /* Active associated thread. */ 871573Srgrimes struct runq *ts_runq; /* Run-queue we're queued on. */ 881573Srgrimes short ts_flags; /* TSF_* flags. */ 891573Srgrimes u_char ts_rqindex; /* Run queue index. */ 901573Srgrimes u_char ts_cpu; /* CPU that we have affinity for. */ 911573Srgrimes int ts_slptick; /* Tick when we went to sleep. */ 921573Srgrimes int ts_slice; /* Ticks of slice remaining. */ 931573Srgrimes u_int ts_slptime; /* Number of ticks we vol. slept */ 941573Srgrimes u_int ts_runtime; /* Number of ticks we were running */ 95189291Sdelphij /* The following variables are only used for pctcpu calculation */ 961573Srgrimes int ts_ltick; /* Last tick that we were running on */ 9792889Sobrien int ts_ftick; /* First tick that we were running on */ 981573Srgrimes int ts_ticks; /* Tick count */ 9914287Spst#ifdef SMP 1001573Srgrimes int ts_rltick; /* Real last tick, for affinity. */ 1011573Srgrimes#endif 1021573Srgrimes}; 1031573Srgrimes/* flags kept in ts_flags */ 1041573Srgrimes#define TSF_BOUND 0x0001 /* Thread can not migrate. */ 1051573Srgrimes#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */ 1061573Srgrimes 1071573Srgrimesstatic struct td_sched td_sched0; 1081573Srgrimes 1091573Srgrimes/* 1101573Srgrimes * Cpu percentage computation macros and defines. 1111573Srgrimes * 1121573Srgrimes * SCHED_TICK_SECS: Number of seconds to average the cpu usage across. 1131573Srgrimes * SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across. 1141573Srgrimes * SCHED_TICK_MAX: Maximum number of ticks before scaling back. 11514287Spst * SCHED_TICK_SHIFT: Shift factor to avoid rounding away results. 1161573Srgrimes * SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count. 1171573Srgrimes * SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks. 1181573Srgrimes */ 1191573Srgrimes#define SCHED_TICK_SECS 10 1201573Srgrimes#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS) 1211573Srgrimes#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz) 1221573Srgrimes#define SCHED_TICK_SHIFT 10 1231573Srgrimes#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT) 124189291Sdelphij#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz)) 125189291Sdelphij 1261573Srgrimes/* 127190489Sdelphij * These macros determine priorities for non-interactive threads. They are 12892889Sobrien * assigned a priority based on their recent cpu utilization as expressed 1291573Srgrimes * by the ratio of ticks to the tick total. NHALF priorities at the start 13014287Spst * and end of the MIN to MAX timeshare range are only reachable with negative 1311573Srgrimes * or positive nice respectively. 1321573Srgrimes * 1331573Srgrimes * PRI_RANGE: Priority range for utilization dependent priorities. 1341573Srgrimes * PRI_NRESV: Number of nice values. 1351573Srgrimes * PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total. 1361573Srgrimes * PRI_NICE: Determines the part of the priority inherited from nice. 1371573Srgrimes */ 1381573Srgrimes#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN) 1391573Srgrimes#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2) 1401573Srgrimes#define SCHED_PRI_MIN (PRI_MIN_TIMESHARE + SCHED_PRI_NHALF) 1411573Srgrimes#define SCHED_PRI_MAX (PRI_MAX_TIMESHARE - SCHED_PRI_NHALF) 1421573Srgrimes#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN) 14392889Sobrien#define SCHED_PRI_TICKS(ts) \ 14492889Sobrien (SCHED_TICK_HZ((ts)) / \ 14592889Sobrien (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE)) 1461573Srgrimes#define SCHED_PRI_NICE(nice) (nice) 1471573Srgrimes 1481573Srgrimes/* 1491573Srgrimes * These determine the interactivity of a process. Interactivity differs from 1501573Srgrimes * cpu utilization in that it expresses the voluntary time slept vs time ran 1511573Srgrimes * while cpu utilization includes all time not running. This more accurately 1521573Srgrimes * models the intent of the thread. 1531573Srgrimes * 1541573Srgrimes * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate 1551573Srgrimes * before throttling back. 1561573Srgrimes * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time. 1571573Srgrimes * INTERACT_MAX: Maximum interactivity value. Smaller is better. 158190491Sdelphij * INTERACT_THRESH: Threshhold for placement on the current runq. 159190491Sdelphij */ 160190491Sdelphij#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT) 161190491Sdelphij#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT) 162190491Sdelphij#define SCHED_INTERACT_MAX (100) 163190491Sdelphij#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2) 164190491Sdelphij#define SCHED_INTERACT_THRESH (30) 165190491Sdelphij 1661573Srgrimes/* 1671573Srgrimes * tickincr: Converts a stathz tick into a hz domain scaled by 1681573Srgrimes * the shift factor. Without the shift the error rate 16914287Spst * due to rounding would be unacceptably high. 1701573Srgrimes * realstathz: stathz is sometimes 0 and run off of hz. 1711573Srgrimes * sched_slice: Runtime of each thread before rescheduling. 1721573Srgrimes * preempt_thresh: Priority threshold for preemption and remote IPIs. 1731573Srgrimes */ 1741573Srgrimesstatic int sched_interact = SCHED_INTERACT_THRESH; 1751573Srgrimesstatic int realstathz; 1761573Srgrimesstatic int tickincr; 1771573Srgrimesstatic int sched_slice; 1781573Srgrimesstatic int preempt_thresh = PRI_MIN_KERN; 1791573Srgrimes 1801573Srgrimes/* 181189291Sdelphij * tdq - per processor runqs and statistics. All fields are protected by the 182189291Sdelphij * tdq_lock. The load and lowpri may be accessed without to avoid excess 1831573Srgrimes * locking in sched_pickcpu(); 18492889Sobrien */ 18592889Sobrienstruct tdq { 18692889Sobrien struct mtx *tdq_lock; /* Pointer to group lock. */ 1871573Srgrimes struct runq tdq_realtime; /* real-time run queue. */ 1881573Srgrimes struct runq tdq_timeshare; /* timeshare run queue. */ 18914287Spst struct runq tdq_idle; /* Queue of IDLE threads. */ 1901573Srgrimes int tdq_load; /* Aggregate load. */ 1911573Srgrimes u_char tdq_idx; /* Current insert index. */ 19214287Spst u_char tdq_ridx; /* Current removal index. */ 19314287Spst#ifdef SMP 1941573Srgrimes u_char tdq_lowpri; /* Lowest priority thread. */ 1951573Srgrimes int tdq_transferable; /* Transferable thread count. */ 1961573Srgrimes LIST_ENTRY(tdq) tdq_siblings; /* Next in tdq group. */ 1971573Srgrimes struct tdq_group *tdq_group; /* Our processor group. */ 1981573Srgrimes#else 1991573Srgrimes int tdq_sysload; /* For loadavg, !ITHD load. */ 2001573Srgrimes#endif 2011573Srgrimes} __aligned(64); 2021573Srgrimes 2031573Srgrimes 20414287Spst#ifdef SMP 2051573Srgrimes/* 2061573Srgrimes * tdq groups are groups of processors which can cheaply share threads. When 2071573Srgrimes * one processor in the group goes idle it will check the runqs of the other 2081573Srgrimes * processors in its group prior to halting and waiting for an interrupt. 2091573Srgrimes * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA. 2101573Srgrimes * In a numa environment we'd want an idle bitmap per group and a two tiered 2111573Srgrimes * load balancer. 2121573Srgrimes */ 2131573Srgrimesstruct tdq_group { 2141573Srgrimes struct mtx tdg_lock; /* Protects all fields below. */ 2151573Srgrimes int tdg_cpus; /* Count of CPUs in this tdq group. */ 2161573Srgrimes cpumask_t tdg_cpumask; /* Mask of cpus in this group. */ 2171573Srgrimes cpumask_t tdg_idlemask; /* Idle cpus in this group. */ 2181573Srgrimes cpumask_t tdg_mask; /* Bit mask for first cpu. */ 2191573Srgrimes int tdg_load; /* Total load of this group. */ 2201573Srgrimes int tdg_transferable; /* Transferable load of this group. */ 2211573Srgrimes LIST_HEAD(, tdq) tdg_members; /* Linked list of all members. */ 2221573Srgrimes char tdg_name[16]; /* lock name. */ 2231573Srgrimes} __aligned(64); 2241573Srgrimes 2251573Srgrimes#define SCHED_AFFINITY_DEFAULT (max(1, hz / 300)) 2261573Srgrimes#define SCHED_AFFINITY(ts) ((ts)->ts_rltick > ticks - affinity) 2271573Srgrimes 2281573Srgrimes/* 2291573Srgrimes * Run-time tunables. 2301573Srgrimes */ 2311573Srgrimesstatic int rebalance = 1; 2321573Srgrimesstatic int balance_secs = 1; 2331573Srgrimesstatic int pick_pri = 1; 2341573Srgrimesstatic int affinity; 2351573Srgrimesstatic int tryself = 1; 2361573Srgrimesstatic int steal_htt = 0; 2371573Srgrimesstatic int steal_idle = 1; 2381573Srgrimesstatic int steal_thresh = 2; 2391573Srgrimesstatic int topology = 0; 2401573Srgrimes 2411573Srgrimes/* 2421573Srgrimes * One thread queue per processor. 2431573Srgrimes */ 2441573Srgrimesstatic volatile cpumask_t tdq_idle; 2451573Srgrimesstatic int tdg_maxid; 24614287Spststatic struct tdq tdq_cpu[MAXCPU]; 2471573Srgrimesstatic struct tdq_group tdq_groups[MAXCPU]; 2481573Srgrimesstatic struct callout balco; 2491573Srgrimesstatic struct callout gbalco; 2501573Srgrimes 25114287Spst#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)]) 25214287Spst#define TDQ_CPU(x) (&tdq_cpu[(x)]) 2531573Srgrimes#define TDQ_ID(x) ((int)((x) - tdq_cpu)) 2541573Srgrimes#define TDQ_GROUP(x) (&tdq_groups[(x)]) 2551573Srgrimes#define TDG_ID(x) ((int)((x) - tdq_groups)) 2561573Srgrimes#else /* !SMP */ 2571573Srgrimesstatic struct tdq tdq_cpu; 2581573Srgrimesstatic struct mtx tdq_lock; 2591573Srgrimes 2601573Srgrimes#define TDQ_ID(x) (0) 2611573Srgrimes#define TDQ_SELF() (&tdq_cpu) 2621573Srgrimes#define TDQ_CPU(x) (&tdq_cpu) 2631573Srgrimes#endif 2641573Srgrimes 2651573Srgrimes#define TDQ_LOCK_ASSERT(t, type) mtx_assert(TDQ_LOCKPTR((t)), (type)) 2661573Srgrimes#define TDQ_LOCK(t) mtx_lock_spin(TDQ_LOCKPTR((t))) 2671573Srgrimes#define TDQ_LOCK_FLAGS(t, f) mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f)) 2681573Srgrimes#define TDQ_UNLOCK(t) mtx_unlock_spin(TDQ_LOCKPTR((t))) 2691573Srgrimes#define TDQ_LOCKPTR(t) ((t)->tdq_lock) 2701573Srgrimes 2711573Srgrimesstatic void sched_priority(struct thread *); 2721573Srgrimesstatic void sched_thread_priority(struct thread *, u_char); 2731573Srgrimesstatic int sched_interact_score(struct thread *); 2741573Srgrimesstatic void sched_interact_update(struct thread *); 2751573Srgrimesstatic void sched_interact_fork(struct thread *); 276189291Sdelphijstatic void sched_pctcpu_update(struct td_sched *); 277189291Sdelphij 278189291Sdelphij/* Operations on per processor queues */ 279189291Sdelphijstatic struct td_sched * tdq_choose(struct tdq *); 280189291Sdelphijstatic void tdq_setup(struct tdq *); 281189291Sdelphijstatic void tdq_load_add(struct tdq *, struct td_sched *); 2821573Srgrimesstatic void tdq_load_rem(struct tdq *, struct td_sched *); 283189291Sdelphijstatic __inline void tdq_runq_add(struct tdq *, struct td_sched *, int); 284189291Sdelphijstatic __inline void tdq_runq_rem(struct tdq *, struct td_sched *); 285189291Sdelphijvoid tdq_print(int cpu); 286189291Sdelphijstatic void runq_print(struct runq *rq); 2871573Srgrimesstatic void tdq_add(struct tdq *, struct thread *, int); 2881573Srgrimes#ifdef SMP 2891573Srgrimesstatic void tdq_move(struct tdq *, struct tdq *); 2901573Srgrimesstatic int tdq_idled(struct tdq *); 29114287Spststatic void tdq_notify(struct td_sched *); 2921573Srgrimesstatic struct td_sched *tdq_steal(struct tdq *, int); 2931573Srgrimesstatic struct td_sched *runq_steal(struct runq *); 2941573Srgrimesstatic int sched_pickcpu(struct td_sched *, int); 29514287Spststatic void sched_balance(void *); 29614287Spststatic void sched_balance_groups(void *); 29714287Spststatic void sched_balance_group(struct tdq_group *); 2981573Srgrimesstatic void sched_balance_pair(struct tdq *, struct tdq *); 29914287Spststatic inline struct tdq *sched_setcpu(struct td_sched *, int, int); 3001573Srgrimesstatic inline struct mtx *thread_block_switch(struct thread *); 3011573Srgrimesstatic inline void thread_unblock_switch(struct thread *, struct mtx *); 3021573Srgrimesstatic struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int); 3031573Srgrimes 3041573Srgrimes#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0) 3051573Srgrimes#endif 3061573Srgrimes 3071573Srgrimesstatic void sched_setup(void *dummy); 3081573SrgrimesSYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL) 3091573Srgrimes 31014287Spststatic void sched_initticks(void *dummy); 3111573SrgrimesSYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, NULL) 3121573Srgrimes 3131573Srgrimes/* 31414287Spst * Print the threads waiting on a run-queue. 3151573Srgrimes */ 3161573Srgrimesstatic void 3171573Srgrimesrunq_print(struct runq *rq) 3181573Srgrimes{ 31914287Spst struct rqhead *rqh; 3201573Srgrimes struct td_sched *ts; 3211573Srgrimes int pri; 3221573Srgrimes int j; 3231573Srgrimes int i; 3241573Srgrimes 3251573Srgrimes for (i = 0; i < RQB_LEN; i++) { 3261573Srgrimes printf("\t\trunq bits %d 0x%zx\n", 3271573Srgrimes i, rq->rq_status.rqb_bits[i]); 3281573Srgrimes for (j = 0; j < RQB_BPW; j++) 32914287Spst if (rq->rq_status.rqb_bits[i] & (1ul << j)) { 3301573Srgrimes pri = j + (i << RQB_L2BPW); 3311573Srgrimes rqh = &rq->rq_queues[pri]; 3321573Srgrimes TAILQ_FOREACH(ts, rqh, ts_procq) { 3331573Srgrimes printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n", 3341573Srgrimes ts->ts_thread, ts->ts_thread->td_proc->p_comm, ts->ts_thread->td_priority, ts->ts_rqindex, pri); 3351573Srgrimes } 33614287Spst } 3371573Srgrimes } 3381573Srgrimes} 3391573Srgrimes 3401573Srgrimes/* 3411573Srgrimes * Print the status of a per-cpu thread queue. Should be a ddb show cmd. 3421573Srgrimes */ 3431573Srgrimesvoid 3441573Srgrimestdq_print(int cpu) 3451573Srgrimes{ 3461573Srgrimes struct tdq *tdq; 3471573Srgrimes 3481573Srgrimes tdq = TDQ_CPU(cpu); 3491573Srgrimes 3501573Srgrimes printf("tdq %d:\n", TDQ_ID(tdq)); 3511573Srgrimes printf("\tlockptr %p\n", TDQ_LOCKPTR(tdq)); 3521573Srgrimes printf("\tload: %d\n", tdq->tdq_load); 3531573Srgrimes printf("\ttimeshare idx: %d\n", tdq->tdq_idx); 3541573Srgrimes printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx); 3551573Srgrimes printf("\trealtime runq:\n"); 3561573Srgrimes runq_print(&tdq->tdq_realtime); 3571573Srgrimes printf("\ttimeshare runq:\n"); 3581573Srgrimes runq_print(&tdq->tdq_timeshare); 3591573Srgrimes printf("\tidle runq:\n"); 3601573Srgrimes runq_print(&tdq->tdq_idle); 3611573Srgrimes#ifdef SMP 3621573Srgrimes printf("\tload transferable: %d\n", tdq->tdq_transferable); 3631573Srgrimes printf("\tlowest priority: %d\n", tdq->tdq_lowpri); 36414287Spst printf("\tgroup: %d\n", TDG_ID(tdq->tdq_group)); 3651573Srgrimes printf("\tLock name: %s\n", tdq->tdq_group->tdg_name); 3661573Srgrimes#endif 3671573Srgrimes} 3681573Srgrimes 3691573Srgrimes#define TS_RQ_PPQ (((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS) 3701573Srgrimes/* 3711573Srgrimes * Add a thread to the actual run-queue. Keeps transferable counts up to 3721573Srgrimes * date with what is actually on the run-queue. Selects the correct 3731573Srgrimes * queue position for timeshare threads. 3741573Srgrimes */ 3751573Srgrimesstatic __inline void 3761573Srgrimestdq_runq_add(struct tdq *tdq, struct td_sched *ts, int flags) 37714287Spst{ 3781573Srgrimes TDQ_LOCK_ASSERT(tdq, MA_OWNED); 3791573Srgrimes THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 3801573Srgrimes#ifdef SMP 3811573Srgrimes if (THREAD_CAN_MIGRATE(ts->ts_thread)) { 3821573Srgrimes tdq->tdq_transferable++; 3831573Srgrimes tdq->tdq_group->tdg_transferable++; 3841573Srgrimes ts->ts_flags |= TSF_XFERABLE; 3851573Srgrimes } 3861573Srgrimes#endif 3871573Srgrimes if (ts->ts_runq == &tdq->tdq_timeshare) { 3881573Srgrimes u_char pri; 3891573Srgrimes 3901573Srgrimes pri = ts->ts_thread->td_priority; 3911573Srgrimes KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE, 3921573Srgrimes ("Invalid priority %d on timeshare runq", pri)); 3931573Srgrimes /* 3941573Srgrimes * This queue contains only priorities between MIN and MAX 3951573Srgrimes * realtime. Use the whole queue to represent these values. 396189291Sdelphij */ 397189291Sdelphij if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) { 3981573Srgrimes pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ; 39992889Sobrien pri = (pri + tdq->tdq_idx) % RQ_NQS; 4001573Srgrimes /* 4011573Srgrimes * This effectively shortens the queue by one so we 40214287Spst * can have a one slot difference between idx and 4031573Srgrimes * ridx while we wait for threads to drain. 4041573Srgrimes */ 4051573Srgrimes if (tdq->tdq_ridx != tdq->tdq_idx && 4061573Srgrimes pri == tdq->tdq_ridx) 4071573Srgrimes pri = (unsigned char)(pri - 1) % RQ_NQS; 4081573Srgrimes } else 4091573Srgrimes pri = tdq->tdq_ridx; 4101573Srgrimes runq_add_pri(ts->ts_runq, ts, pri, flags); 4111573Srgrimes } else 4121573Srgrimes runq_add(ts->ts_runq, ts, flags); 4131573Srgrimes} 41414287Spst 415190490Sdelphij/* 416190490Sdelphij * Remove a thread from a run-queue. This typically happens when a thread 417190490Sdelphij * is selected to run. Running threads are not on the queue and the 418190490Sdelphij * transferable count does not reflect them. 4191573Srgrimes */ 420190490Sdelphijstatic __inline void 421190490Sdelphijtdq_runq_rem(struct tdq *tdq, struct td_sched *ts) 4221573Srgrimes{ 423190490Sdelphij TDQ_LOCK_ASSERT(tdq, MA_OWNED); 4241573Srgrimes KASSERT(ts->ts_runq != NULL, 4251573Srgrimes ("tdq_runq_remove: thread %p null ts_runq", ts->ts_thread)); 4261573Srgrimes#ifdef SMP 4271573Srgrimes if (ts->ts_flags & TSF_XFERABLE) { 42814287Spst tdq->tdq_transferable--; 4291573Srgrimes tdq->tdq_group->tdg_transferable--; 430190490Sdelphij ts->ts_flags &= ~TSF_XFERABLE; 4311573Srgrimes } 4321573Srgrimes#endif 4331573Srgrimes if (ts->ts_runq == &tdq->tdq_timeshare) { 4341573Srgrimes if (tdq->tdq_idx != tdq->tdq_ridx) 4351573Srgrimes runq_remove_idx(ts->ts_runq, ts, &tdq->tdq_ridx); 4361573Srgrimes else 4371573Srgrimes runq_remove_idx(ts->ts_runq, ts, NULL); 4381573Srgrimes /* 43914287Spst * For timeshare threads we update the priority here so 4401573Srgrimes * the priority reflects the time we've been sleeping. 4411573Srgrimes */ 4421573Srgrimes ts->ts_ltick = ticks; 4431573Srgrimes sched_pctcpu_update(ts); 4441573Srgrimes sched_priority(ts->ts_thread); 4451573Srgrimes } else 4461573Srgrimes runq_remove(ts->ts_runq, ts); 447190490Sdelphij} 4481573Srgrimes 4491573Srgrimes/* 4501573Srgrimes * Load is maintained for all threads RUNNING and ON_RUNQ. Add the load 4511573Srgrimes * for this thread to the referenced thread queue. 4521573Srgrimes */ 4531573Srgrimesstatic void 4541573Srgrimestdq_load_add(struct tdq *tdq, struct td_sched *ts) 4551573Srgrimes{ 4561573Srgrimes int class; 4571573Srgrimes 4581573Srgrimes TDQ_LOCK_ASSERT(tdq, MA_OWNED); 4591573Srgrimes THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 4601573Srgrimes class = PRI_BASE(ts->ts_thread->td_pri_class); 4611573Srgrimes tdq->tdq_load++; 4621573Srgrimes CTR2(KTR_SCHED, "cpu %d load: %d", TDQ_ID(tdq), tdq->tdq_load); 4631573Srgrimes if (class != PRI_ITHD && 4641573Srgrimes (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 4651573Srgrimes#ifdef SMP 466189291Sdelphij tdq->tdq_group->tdg_load++; 467189291Sdelphij#else 4681573Srgrimes tdq->tdq_sysload++; 469190489Sdelphij#endif 4701573Srgrimes} 4711573Srgrimes 4721573Srgrimes/* 47314287Spst * Remove the load from a thread that is transitioning to a sleep state or 4741573Srgrimes * exiting. 4751573Srgrimes */ 4761573Srgrimesstatic void 4771573Srgrimestdq_load_rem(struct tdq *tdq, struct td_sched *ts) 4781573Srgrimes{ 4791573Srgrimes int class; 4801573Srgrimes 4811573Srgrimes THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 4821573Srgrimes TDQ_LOCK_ASSERT(tdq, MA_OWNED); 4831573Srgrimes class = PRI_BASE(ts->ts_thread->td_pri_class); 4841573Srgrimes if (class != PRI_ITHD && 4851573Srgrimes (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0) 4861573Srgrimes#ifdef SMP 4871573Srgrimes tdq->tdq_group->tdg_load--; 4881573Srgrimes#else 4891573Srgrimes tdq->tdq_sysload--; 4901573Srgrimes#endif 4911573Srgrimes KASSERT(tdq->tdq_load != 0, 4921573Srgrimes ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq))); 4931573Srgrimes tdq->tdq_load--; 4941573Srgrimes CTR1(KTR_SCHED, "load: %d", tdq->tdq_load); 4951573Srgrimes ts->ts_runq = NULL; 4961573Srgrimes} 4971573Srgrimes 4981573Srgrimes#ifdef SMP 4991573Srgrimes/* 5001573Srgrimes * sched_balance is a simple CPU load balancing algorithm. It operates by 5011573Srgrimes * finding the least loaded and most loaded cpu and equalizing their load 5021573Srgrimes * by migrating some processes. 5031573Srgrimes * 5041573Srgrimes * Dealing only with two CPUs at a time has two advantages. Firstly, most 5051573Srgrimes * installations will only have 2 cpus. Secondly, load balancing too much at 5061573Srgrimes * once can have an unpleasant effect on the system. The scheduler rarely has 5071573Srgrimes * enough information to make perfect decisions. So this algorithm chooses 5081573Srgrimes * simplicity and more gradual effects on load in larger systems. 5091573Srgrimes * 5101573Srgrimes */ 5111573Srgrimesstatic void 5121573Srgrimessched_balance(void *arg) 5131573Srgrimes{ 5141573Srgrimes struct tdq_group *high; 5151573Srgrimes struct tdq_group *low; 516189291Sdelphij struct tdq_group *tdg; 517189291Sdelphij int cnt; 518189291Sdelphij int i; 5191573Srgrimes 520190489Sdelphij callout_reset(&balco, max(hz / 2, random() % (hz * balance_secs)), 52114287Spst sched_balance, NULL); 5221573Srgrimes if (smp_started == 0 || rebalance == 0) 5231573Srgrimes return; 5241573Srgrimes low = high = NULL; 5251573Srgrimes i = random() % (tdg_maxid + 1); 5261573Srgrimes for (cnt = 0; cnt <= tdg_maxid; cnt++) { 5271573Srgrimes tdg = TDQ_GROUP(i); 5281573Srgrimes /* 5291573Srgrimes * Find the CPU with the highest load that has some 5301573Srgrimes * threads to transfer. 5311573Srgrimes */ 5321573Srgrimes if ((high == NULL || tdg->tdg_load > high->tdg_load) 5331573Srgrimes && tdg->tdg_transferable) 534190486Sdelphij high = tdg; 5351573Srgrimes if (low == NULL || tdg->tdg_load < low->tdg_load) 53614287Spst low = tdg; 5371573Srgrimes if (++i > tdg_maxid) 5381573Srgrimes i = 0; 5391573Srgrimes } 5401573Srgrimes if (low != NULL && high != NULL && high != low) 5411573Srgrimes sched_balance_pair(LIST_FIRST(&high->tdg_members), 5421573Srgrimes LIST_FIRST(&low->tdg_members)); 5431573Srgrimes} 5441573Srgrimes 5451573Srgrimes/* 5461573Srgrimes * Balance load between CPUs in a group. Will only migrate within the group. 5471573Srgrimes */ 54892889Sobrienstatic void 5491573Srgrimessched_balance_groups(void *arg) 5501573Srgrimes{ 5511573Srgrimes int i; 5521573Srgrimes 55314287Spst callout_reset(&gbalco, max(hz / 2, random() % (hz * balance_secs)), 5541573Srgrimes sched_balance_groups, NULL); 5551573Srgrimes if (smp_started == 0 || rebalance == 0) 5561573Srgrimes return; 5571573Srgrimes for (i = 0; i <= tdg_maxid; i++) 5581573Srgrimes sched_balance_group(TDQ_GROUP(i)); 5591573Srgrimes} 5601573Srgrimes 5611573Srgrimes/* 5621573Srgrimes * Finds the greatest imbalance between two tdqs in a group. 5631573Srgrimes */ 5641573Srgrimesstatic void 5651573Srgrimessched_balance_group(struct tdq_group *tdg) 5661573Srgrimes{ 5671573Srgrimes struct tdq *tdq; 5681573Srgrimes struct tdq *high; 5691573Srgrimes struct tdq *low; 5701573Srgrimes int load; 571189291Sdelphij 572189291Sdelphij if (tdg->tdg_transferable == 0) 5731573Srgrimes return; 574190489Sdelphij low = NULL; 5751573Srgrimes high = NULL; 5761573Srgrimes LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) { 5771573Srgrimes load = tdq->tdq_load; 5781573Srgrimes if (high == NULL || load > high->tdq_load) 5791573Srgrimes high = tdq; 5801573Srgrimes if (low == NULL || load < low->tdq_load) 5811573Srgrimes low = tdq; 582190489Sdelphij } 5831573Srgrimes if (high != NULL && low != NULL && high != low) 5841573Srgrimes sched_balance_pair(high, low); 5851573Srgrimes} 5861573Srgrimes 58714287Spst/* 5881573Srgrimes * Lock two thread queues using their address to maintain lock order. 58914287Spst */ 5901573Srgrimesstatic void 59114287Spsttdq_lock_pair(struct tdq *one, struct tdq *two) 5921573Srgrimes{ 5931573Srgrimes if (one < two) { 5941573Srgrimes TDQ_LOCK(one); 5951573Srgrimes TDQ_LOCK_FLAGS(two, MTX_DUPOK); 5961573Srgrimes } else { 5971573Srgrimes TDQ_LOCK(two); 598190486Sdelphij TDQ_LOCK_FLAGS(one, MTX_DUPOK); 5991573Srgrimes } 6001573Srgrimes} 6011573Srgrimes 6021573Srgrimes/* 6031573Srgrimes * Transfer load between two imbalanced thread queues. 6041573Srgrimes */ 6051573Srgrimesstatic void 6061573Srgrimessched_balance_pair(struct tdq *high, struct tdq *low) 6071573Srgrimes{ 6081573Srgrimes int transferable; 6091573Srgrimes int high_load; 6101573Srgrimes int low_load; 6111573Srgrimes int move; 6121573Srgrimes int diff; 613189291Sdelphij int i; 614189291Sdelphij 6151573Srgrimes tdq_lock_pair(high, low); 61614287Spst /* 6171573Srgrimes * If we're transfering within a group we have to use this specific 6181573Srgrimes * tdq's transferable count, otherwise we can steal from other members 61914287Spst * of the group. 6201573Srgrimes */ 6211573Srgrimes if (high->tdq_group == low->tdq_group) { 6221573Srgrimes transferable = high->tdq_transferable; 6231573Srgrimes high_load = high->tdq_load; 6241573Srgrimes low_load = low->tdq_load; 6251573Srgrimes } else { 6261573Srgrimes transferable = high->tdq_group->tdg_transferable; 6271573Srgrimes high_load = high->tdq_group->tdg_load; 6281573Srgrimes low_load = low->tdq_group->tdg_load; 62914287Spst } 6301573Srgrimes /* 6311573Srgrimes * Determine what the imbalance is and then adjust that to how many 6321573Srgrimes * threads we actually have to give up (transferable). 6331573Srgrimes */ 63414287Spst if (transferable != 0) { 635189291Sdelphij diff = high_load - low_load; 6361573Srgrimes move = diff / 2; 63792889Sobrien if (diff & 0x1) 6381573Srgrimes move++; 6391573Srgrimes move = min(move, transferable); 6401573Srgrimes for (i = 0; i < move; i++) 6411573Srgrimes tdq_move(high, low); 6421573Srgrimes } 6431573Srgrimes TDQ_UNLOCK(high); 6441573Srgrimes TDQ_UNLOCK(low); 6451573Srgrimes return; 6461573Srgrimes} 6471573Srgrimes 64814287Spst/* 649189291Sdelphij * Move a thread from one thread queue to another. 6501573Srgrimes */ 65192889Sobrienstatic void 65292889Sobrientdq_move(struct tdq *from, struct tdq *to) 65314287Spst{ 6541573Srgrimes struct td_sched *ts; 6551573Srgrimes struct thread *td; 6561573Srgrimes struct tdq *tdq; 6571573Srgrimes int cpu; 6581573Srgrimes 6591573Srgrimes tdq = from; 6601573Srgrimes cpu = TDQ_ID(to); 6611573Srgrimes ts = tdq_steal(tdq, 1); 6621573Srgrimes if (ts == NULL) { 6631573Srgrimes struct tdq_group *tdg; 6641573Srgrimes 6651573Srgrimes tdg = tdq->tdq_group; 6661573Srgrimes LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) { 66714287Spst if (tdq == from || tdq->tdq_transferable == 0) 6681573Srgrimes continue; 66914287Spst ts = tdq_steal(tdq, 1); 6701573Srgrimes break; 6711573Srgrimes } 6721573Srgrimes if (ts == NULL) 6731573Srgrimes return; 6748870Srgrimes } 6751573Srgrimes if (tdq == to) 6761573Srgrimes return; 6771573Srgrimes td = ts->ts_thread; 6781573Srgrimes /* 6791573Srgrimes * Although the run queue is locked the thread may be blocked. Lock 6801573Srgrimes * it to clear this. 6811573Srgrimes */ 6821573Srgrimes thread_lock(td); 6831573Srgrimes /* Drop recursive lock on from. */ 6841573Srgrimes TDQ_UNLOCK(from); 6851573Srgrimes sched_rem(td); 6861573Srgrimes ts->ts_cpu = cpu; 6871573Srgrimes td->td_lock = TDQ_LOCKPTR(to); 6881573Srgrimes tdq_add(to, td, SRQ_YIELDING); 6891573Srgrimes tdq_notify(ts); 6901573Srgrimes} 6911573Srgrimes 6921573Srgrimes/* 6931573Srgrimes * This tdq has idled. Try to steal a thread from another cpu and switch 6941573Srgrimes * to it. 6951573Srgrimes */ 6961573Srgrimesstatic int 6971573Srgrimestdq_idled(struct tdq *tdq) 69856698Sjasone{ 699190487Sdelphij struct tdq_group *tdg; 70014287Spst struct tdq *steal; 7011573Srgrimes struct td_sched *ts; 7021573Srgrimes struct thread *td; 7031573Srgrimes int highload; 7041573Srgrimes int highcpu; 7051573Srgrimes int load; 7061573Srgrimes int cpu; 7071573Srgrimes 7081573Srgrimes /* We don't want to be preempted while we're iterating over tdqs */ 7091573Srgrimes spinlock_enter(); 7101573Srgrimes tdg = tdq->tdq_group; 7111573Srgrimes /* 71256698Sjasone * If we're in a cpu group, try and steal threads from another cpu in 713190487Sdelphij * the group before idling. 71414287Spst */ 7151573Srgrimes if (steal_htt && tdg->tdg_cpus > 1 && tdg->tdg_transferable) { 7161573Srgrimes LIST_FOREACH(steal, &tdg->tdg_members, tdq_siblings) { 7171573Srgrimes if (steal == tdq || steal->tdq_transferable == 0) 7181573Srgrimes continue; 7191573Srgrimes TDQ_LOCK(steal); 7201573Srgrimes ts = tdq_steal(steal, 0); 7211573Srgrimes if (ts) 7221573Srgrimes goto steal; 7231573Srgrimes TDQ_UNLOCK(steal); 7241573Srgrimes } 7251573Srgrimes } 7261573Srgrimes for (;;) { 72714287Spst if (steal_idle == 0) 72814287Spst break; 72914287Spst highcpu = 0; 7301573Srgrimes highload = 0; 7311573Srgrimes for (cpu = 0; cpu <= mp_maxid; cpu++) { 7321573Srgrimes if (CPU_ABSENT(cpu)) 7331573Srgrimes continue; 7341573Srgrimes steal = TDQ_CPU(cpu); 7351573Srgrimes load = TDQ_CPU(cpu)->tdq_transferable; 7361573Srgrimes if (load < highload) 73756698Sjasone continue; 7381573Srgrimes highload = load; 739190487Sdelphij highcpu = cpu; 74014287Spst } 7411573Srgrimes if (highload < steal_thresh) 7421573Srgrimes break; 7431573Srgrimes steal = TDQ_CPU(highcpu); 7441573Srgrimes TDQ_LOCK(steal); 7451573Srgrimes if (steal->tdq_transferable >= steal_thresh && 7461573Srgrimes (ts = tdq_steal(steal, 1)) != NULL) 7471573Srgrimes goto steal; 7481573Srgrimes TDQ_UNLOCK(steal); 7491573Srgrimes break; 7501573Srgrimes } 7511573Srgrimes spinlock_exit(); 7521573Srgrimes return (1); 7531573Srgrimessteal: 7541573Srgrimes td = ts->ts_thread; 7551573Srgrimes thread_lock(td); 7561573Srgrimes spinlock_exit(); 7571573Srgrimes MPASS(td->td_lock == TDQ_LOCKPTR(steal)); 7581573Srgrimes TDQ_UNLOCK(steal); 7591573Srgrimes sched_rem(td); 7601573Srgrimes sched_setcpu(ts, PCPU_GET(cpuid), SRQ_YIELDING); 7611573Srgrimes tdq_add(tdq, td, SRQ_YIELDING); 7621573Srgrimes MPASS(td->td_lock == curthread->td_lock); 7631573Srgrimes mi_switch(SW_VOL, NULL); 7641573Srgrimes thread_unlock(curthread); 7651573Srgrimes 7661573Srgrimes return (0); 7671573Srgrimes} 7681573Srgrimes 7691573Srgrimes/* 7701573Srgrimes * Notify a remote cpu of new work. Sends an IPI if criteria are met. 7711573Srgrimes */ 7721573Srgrimesstatic void 7731573Srgrimestdq_notify(struct td_sched *ts) 7741573Srgrimes{ 7751573Srgrimes struct thread *ctd; 7761573Srgrimes struct pcpu *pcpu; 7771573Srgrimes int cpri; 7781573Srgrimes int pri; 7791573Srgrimes int cpu; 7801573Srgrimes 7811573Srgrimes cpu = ts->ts_cpu; 7821573Srgrimes pri = ts->ts_thread->td_priority; 783190487Sdelphij pcpu = pcpu_find(cpu); 784190487Sdelphij ctd = pcpu->pc_curthread; 785190487Sdelphij cpri = ctd->td_priority; 78614287Spst 787190487Sdelphij /* 7881573Srgrimes * If our priority is not better than the current priority there is 7891573Srgrimes * nothing to do. 7901573Srgrimes */ 7911573Srgrimes if (pri > cpri) 7921573Srgrimes return; 7931573Srgrimes /* 7941573Srgrimes * Always IPI idle. 7951573Srgrimes */ 7961573Srgrimes if (cpri > PRI_MIN_IDLE) 7971573Srgrimes goto sendipi; 7981573Srgrimes /* 7991573Srgrimes * If we're realtime or better and there is timeshare or worse running 8001573Srgrimes * send an IPI. 801189291Sdelphij */ 802189291Sdelphij if (pri < PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME) 8031573Srgrimes goto sendipi; 80492889Sobrien /* 80514287Spst * Otherwise only IPI if we exceed the threshold. 8061573Srgrimes */ 80714287Spst if (pri > preempt_thresh) 8081573Srgrimes return; 8091573Srgrimessendipi: 8101573Srgrimes ctd->td_flags |= TDF_NEEDRESCHED; 8111573Srgrimes ipi_selected(1 << cpu, IPI_PREEMPT); 8121573Srgrimes} 81314287Spst 8141573Srgrimes/* 8151573Srgrimes * Steals load from a timeshare queue. Honors the rotating queue head 8161573Srgrimes * index. 8171573Srgrimes */ 8181573Srgrimesstatic struct td_sched * 8191573Srgrimesrunq_steal_from(struct runq *rq, u_char start) 8201573Srgrimes{ 8211573Srgrimes struct td_sched *ts; 8221573Srgrimes struct rqbits *rqb; 8231573Srgrimes struct rqhead *rqh; 8241573Srgrimes int first; 8251573Srgrimes int bit; 8261573Srgrimes int pri; 8271573Srgrimes int i; 8281573Srgrimes 8291573Srgrimes rqb = &rq->rq_status; 8301573Srgrimes bit = start & (RQB_BPW -1); 8311573Srgrimes pri = 0; 8321573Srgrimes first = 0; 8331573Srgrimesagain: 8341573Srgrimes for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) { 8351573Srgrimes if (rqb->rqb_bits[i] == 0) 8361573Srgrimes continue; 8371573Srgrimes if (bit != 0) { 8381573Srgrimes for (pri = bit; pri < RQB_BPW; pri++) 8391573Srgrimes if (rqb->rqb_bits[i] & (1ul << pri)) 8401573Srgrimes break; 8411573Srgrimes if (pri >= RQB_BPW) 8421573Srgrimes continue; 8431573Srgrimes } else 8441573Srgrimes pri = RQB_FFS(rqb->rqb_bits[i]); 8451573Srgrimes pri += (i << RQB_L2BPW); 846189291Sdelphij rqh = &rq->rq_queues[pri]; 8471573Srgrimes TAILQ_FOREACH(ts, rqh, ts_procq) { 8481573Srgrimes if (first && THREAD_CAN_MIGRATE(ts->ts_thread)) 849190485Sdelphij return (ts); 850190485Sdelphij first = 1; 851190485Sdelphij } 8521573Srgrimes } 853190485Sdelphij if (start != 0) { 854190485Sdelphij start = 0; 855190485Sdelphij goto again; 856190485Sdelphij } 857190500Sdelphij 858190485Sdelphij return (NULL); 859190485Sdelphij} 860190485Sdelphij 861190485Sdelphij/* 8621573Srgrimes * Steals load from a standard linear queue. 8631573Srgrimes */ 86471579Sdeischenstatic struct td_sched * 865190485Sdelphijrunq_steal(struct runq *rq) 866190485Sdelphij{ 86756698Sjasone struct rqhead *rqh; 8681573Srgrimes struct rqbits *rqb; 86971579Sdeischen struct td_sched *ts; 8701573Srgrimes int word; 8711573Srgrimes int bit; 8721573Srgrimes 8731573Srgrimes rqb = &rq->rq_status; 8741573Srgrimes for (word = 0; word < RQB_LEN; word++) { 8751573Srgrimes if (rqb->rqb_bits[word] == 0) 8761573Srgrimes continue; 8771573Srgrimes for (bit = 0; bit < RQB_BPW; bit++) { 878189291Sdelphij if ((rqb->rqb_bits[word] & (1ul << bit)) == 0) 8791573Srgrimes continue; 88092889Sobrien rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)]; 88114287Spst TAILQ_FOREACH(ts, rqh, ts_procq) 8821573Srgrimes if (THREAD_CAN_MIGRATE(ts->ts_thread)) 8831573Srgrimes return (ts); 8841573Srgrimes } 8851573Srgrimes } 8861573Srgrimes return (NULL); 8871573Srgrimes} 8881573Srgrimes 8891573Srgrimes/* 8901573Srgrimes * Attempt to steal a thread in priority order from a thread queue. 8911573Srgrimes */ 8921573Srgrimesstatic struct td_sched * 8931573Srgrimestdq_steal(struct tdq *tdq, int stealidle) 8941573Srgrimes{ 8951573Srgrimes struct td_sched *ts; 8961573Srgrimes 8971573Srgrimes TDQ_LOCK_ASSERT(tdq, MA_OWNED); 8981573Srgrimes if ((ts = runq_steal(&tdq->tdq_realtime)) != NULL) 8991573Srgrimes return (ts); 9001573Srgrimes if ((ts = runq_steal_from(&tdq->tdq_timeshare, tdq->tdq_ridx)) != NULL) 9011573Srgrimes return (ts); 90214287Spst if (stealidle) 903189291Sdelphij return (runq_steal(&tdq->tdq_idle)); 9041573Srgrimes return (NULL); 9051573Srgrimes} 9061573Srgrimes 90714287Spst/* 9081573Srgrimes * Sets the thread lock and ts_cpu to match the requested cpu. Unlocks the 9091573Srgrimes * current lock and returns with the assigned queue locked. If this is 9101573Srgrimes * via sched_switch() we leave the thread in a blocked state as an 9111573Srgrimes * optimization. 9121573Srgrimes */ 9131573Srgrimesstatic inline struct tdq * 9141573Srgrimessched_setcpu(struct td_sched *ts, int cpu, int flags) 9151573Srgrimes{ 9161573Srgrimes struct thread *td; 9171573Srgrimes struct tdq *tdq; 9181573Srgrimes 919189291Sdelphij THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED); 9201573Srgrimes 9211573Srgrimes tdq = TDQ_CPU(cpu); 9221573Srgrimes td = ts->ts_thread; 9231573Srgrimes ts->ts_cpu = cpu; 9241573Srgrimes 9251573Srgrimes /* If the lock matches just return the queue. */ 9261573Srgrimes if (td->td_lock == TDQ_LOCKPTR(tdq)) 9271573Srgrimes return (tdq); 9281573Srgrimes#ifdef notyet 9291573Srgrimes /* 9301573Srgrimes * If the thread isn't running it's lockptr is a 9311573Srgrimes * turnstile or a sleepqueue. We can just lock_set without 9321573Srgrimes * blocking. 9331573Srgrimes */ 9341573Srgrimes if (TD_CAN_RUN(td)) { 9351573Srgrimes TDQ_LOCK(tdq); 9361573Srgrimes thread_lock_set(td, TDQ_LOCKPTR(tdq)); 937 return (tdq); 938 } 939#endif 940 /* 941 * The hard case, migration, we need to block the thread first to 942 * prevent order reversals with other cpus locks. 943 */ 944 thread_lock_block(td); 945 TDQ_LOCK(tdq); 946 thread_lock_unblock(td, TDQ_LOCKPTR(tdq)); 947 return (tdq); 948} 949 950/* 951 * Find the thread queue running the lowest priority thread. 952 */ 953static int 954tdq_lowestpri(void) 955{ 956 struct tdq *tdq; 957 int lowpri; 958 int lowcpu; 959 int lowload; 960 int load; 961 int cpu; 962 int pri; 963 964 lowload = 0; 965 lowpri = lowcpu = 0; 966 for (cpu = 0; cpu <= mp_maxid; cpu++) { 967 if (CPU_ABSENT(cpu)) 968 continue; 969 tdq = TDQ_CPU(cpu); 970 pri = tdq->tdq_lowpri; 971 load = TDQ_CPU(cpu)->tdq_load; 972 CTR4(KTR_ULE, 973 "cpu %d pri %d lowcpu %d lowpri %d", 974 cpu, pri, lowcpu, lowpri); 975 if (pri < lowpri) 976 continue; 977 if (lowpri && lowpri == pri && load > lowload) 978 continue; 979 lowpri = pri; 980 lowcpu = cpu; 981 lowload = load; 982 } 983 984 return (lowcpu); 985} 986 987/* 988 * Find the thread queue with the least load. 989 */ 990static int 991tdq_lowestload(void) 992{ 993 struct tdq *tdq; 994 int lowload; 995 int lowpri; 996 int lowcpu; 997 int load; 998 int cpu; 999 int pri; 1000 1001 lowcpu = 0; 1002 lowload = TDQ_CPU(0)->tdq_load; 1003 lowpri = TDQ_CPU(0)->tdq_lowpri; 1004 for (cpu = 1; cpu <= mp_maxid; cpu++) { 1005 if (CPU_ABSENT(cpu)) 1006 continue; 1007 tdq = TDQ_CPU(cpu); 1008 load = tdq->tdq_load; 1009 pri = tdq->tdq_lowpri; 1010 CTR4(KTR_ULE, "cpu %d load %d lowcpu %d lowload %d", 1011 cpu, load, lowcpu, lowload); 1012 if (load > lowload) 1013 continue; 1014 if (load == lowload && pri < lowpri) 1015 continue; 1016 lowcpu = cpu; 1017 lowload = load; 1018 lowpri = pri; 1019 } 1020 1021 return (lowcpu); 1022} 1023 1024/* 1025 * Pick the destination cpu for sched_add(). Respects affinity and makes 1026 * a determination based on load or priority of available processors. 1027 */ 1028static int 1029sched_pickcpu(struct td_sched *ts, int flags) 1030{ 1031 struct tdq *tdq; 1032 int self; 1033 int pri; 1034 int cpu; 1035 1036 cpu = self = PCPU_GET(cpuid); 1037 if (smp_started == 0) 1038 return (self); 1039 /* 1040 * Don't migrate a running thread from sched_switch(). 1041 */ 1042 if (flags & SRQ_OURSELF) { 1043 CTR1(KTR_ULE, "YIELDING %d", 1044 curthread->td_priority); 1045 return (self); 1046 } 1047 pri = ts->ts_thread->td_priority; 1048 cpu = ts->ts_cpu; 1049 /* 1050 * Regardless of affinity, if the last cpu is idle send it there. 1051 */ 1052 tdq = TDQ_CPU(cpu); 1053 if (tdq->tdq_lowpri > PRI_MIN_IDLE) { 1054 CTR5(KTR_ULE, 1055 "ts_cpu %d idle, ltick %d ticks %d pri %d curthread %d", 1056 ts->ts_cpu, ts->ts_rltick, ticks, pri, 1057 tdq->tdq_lowpri); 1058 return (ts->ts_cpu); 1059 } 1060 /* 1061 * If we have affinity, try to place it on the cpu we last ran on. 1062 */ 1063 if (SCHED_AFFINITY(ts) && tdq->tdq_lowpri > pri) { 1064 CTR5(KTR_ULE, 1065 "affinity for %d, ltick %d ticks %d pri %d curthread %d", 1066 ts->ts_cpu, ts->ts_rltick, ticks, pri, 1067 tdq->tdq_lowpri); 1068 return (ts->ts_cpu); 1069 } 1070 /* 1071 * Look for an idle group. 1072 */ 1073 CTR1(KTR_ULE, "tdq_idle %X", tdq_idle); 1074 cpu = ffs(tdq_idle); 1075 if (cpu) 1076 return (--cpu); 1077 /* 1078 * If there are no idle cores see if we can run the thread locally. This may 1079 * improve locality among sleepers and wakers when there is shared data. 1080 */ 1081 if (tryself && pri < curthread->td_priority) { 1082 CTR1(KTR_ULE, "tryself %d", 1083 curthread->td_priority); 1084 return (self); 1085 } 1086 /* 1087 * Now search for the cpu running the lowest priority thread with 1088 * the least load. 1089 */ 1090 if (pick_pri) 1091 cpu = tdq_lowestpri(); 1092 else 1093 cpu = tdq_lowestload(); 1094 return (cpu); 1095} 1096 1097#endif /* SMP */ 1098 1099/* 1100 * Pick the highest priority task we have and return it. 1101 */ 1102static struct td_sched * 1103tdq_choose(struct tdq *tdq) 1104{ 1105 struct td_sched *ts; 1106 1107 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 1108 ts = runq_choose(&tdq->tdq_realtime); 1109 if (ts != NULL) 1110 return (ts); 1111 ts = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx); 1112 if (ts != NULL) { 1113 KASSERT(ts->ts_thread->td_priority >= PRI_MIN_TIMESHARE, 1114 ("tdq_choose: Invalid priority on timeshare queue %d", 1115 ts->ts_thread->td_priority)); 1116 return (ts); 1117 } 1118 1119 ts = runq_choose(&tdq->tdq_idle); 1120 if (ts != NULL) { 1121 KASSERT(ts->ts_thread->td_priority >= PRI_MIN_IDLE, 1122 ("tdq_choose: Invalid priority on idle queue %d", 1123 ts->ts_thread->td_priority)); 1124 return (ts); 1125 } 1126 1127 return (NULL); 1128} 1129 1130/* 1131 * Initialize a thread queue. 1132 */ 1133static void 1134tdq_setup(struct tdq *tdq) 1135{ 1136 1137 if (bootverbose) 1138 printf("ULE: setup cpu %d\n", TDQ_ID(tdq)); 1139 runq_init(&tdq->tdq_realtime); 1140 runq_init(&tdq->tdq_timeshare); 1141 runq_init(&tdq->tdq_idle); 1142 tdq->tdq_load = 0; 1143} 1144 1145#ifdef SMP 1146static void 1147tdg_setup(struct tdq_group *tdg) 1148{ 1149 if (bootverbose) 1150 printf("ULE: setup cpu group %d\n", TDG_ID(tdg)); 1151 snprintf(tdg->tdg_name, sizeof(tdg->tdg_name), 1152 "sched lock %d", (int)TDG_ID(tdg)); 1153 mtx_init(&tdg->tdg_lock, tdg->tdg_name, "sched lock", 1154 MTX_SPIN | MTX_RECURSE); 1155 LIST_INIT(&tdg->tdg_members); 1156 tdg->tdg_load = 0; 1157 tdg->tdg_transferable = 0; 1158 tdg->tdg_cpus = 0; 1159 tdg->tdg_mask = 0; 1160 tdg->tdg_cpumask = 0; 1161 tdg->tdg_idlemask = 0; 1162} 1163 1164static void 1165tdg_add(struct tdq_group *tdg, struct tdq *tdq) 1166{ 1167 if (tdg->tdg_mask == 0) 1168 tdg->tdg_mask |= 1 << TDQ_ID(tdq); 1169 tdg->tdg_cpumask |= 1 << TDQ_ID(tdq); 1170 tdg->tdg_cpus++; 1171 tdq->tdq_group = tdg; 1172 tdq->tdq_lock = &tdg->tdg_lock; 1173 LIST_INSERT_HEAD(&tdg->tdg_members, tdq, tdq_siblings); 1174 if (bootverbose) 1175 printf("ULE: adding cpu %d to group %d: cpus %d mask 0x%X\n", 1176 TDQ_ID(tdq), TDG_ID(tdg), tdg->tdg_cpus, tdg->tdg_cpumask); 1177} 1178 1179static void 1180sched_setup_topology(void) 1181{ 1182 struct tdq_group *tdg; 1183 struct cpu_group *cg; 1184 int balance_groups; 1185 struct tdq *tdq; 1186 int i; 1187 int j; 1188 1189 topology = 1; 1190 balance_groups = 0; 1191 for (i = 0; i < smp_topology->ct_count; i++) { 1192 cg = &smp_topology->ct_group[i]; 1193 tdg = &tdq_groups[i]; 1194 /* 1195 * Initialize the group. 1196 */ 1197 tdg_setup(tdg); 1198 /* 1199 * Find all of the group members and add them. 1200 */ 1201 for (j = 0; j < MAXCPU; j++) { 1202 if ((cg->cg_mask & (1 << j)) != 0) { 1203 tdq = TDQ_CPU(j); 1204 tdq_setup(tdq); 1205 tdg_add(tdg, tdq); 1206 } 1207 } 1208 if (tdg->tdg_cpus > 1) 1209 balance_groups = 1; 1210 } 1211 tdg_maxid = smp_topology->ct_count - 1; 1212 if (balance_groups) 1213 sched_balance_groups(NULL); 1214} 1215 1216static void 1217sched_setup_smp(void) 1218{ 1219 struct tdq_group *tdg; 1220 struct tdq *tdq; 1221 int cpus; 1222 int i; 1223 1224 for (cpus = 0, i = 0; i < MAXCPU; i++) { 1225 if (CPU_ABSENT(i)) 1226 continue; 1227 tdq = &tdq_cpu[i]; 1228 tdg = &tdq_groups[i]; 1229 /* 1230 * Setup a tdq group with one member. 1231 */ 1232 tdg_setup(tdg); 1233 tdq_setup(tdq); 1234 tdg_add(tdg, tdq); 1235 cpus++; 1236 } 1237 tdg_maxid = cpus - 1; 1238} 1239 1240/* 1241 * Fake a topology with one group containing all CPUs. 1242 */ 1243static void 1244sched_fake_topo(void) 1245{ 1246#ifdef SCHED_FAKE_TOPOLOGY 1247 static struct cpu_top top; 1248 static struct cpu_group group; 1249 1250 top.ct_count = 1; 1251 top.ct_group = &group; 1252 group.cg_mask = all_cpus; 1253 group.cg_count = mp_ncpus; 1254 group.cg_children = 0; 1255 smp_topology = ⊤ 1256#endif 1257} 1258#endif 1259 1260/* 1261 * Setup the thread queues and initialize the topology based on MD 1262 * information. 1263 */ 1264static void 1265sched_setup(void *dummy) 1266{ 1267 struct tdq *tdq; 1268 1269 tdq = TDQ_SELF(); 1270#ifdef SMP 1271 /* 1272 * Initialize long-term cpu balancing algorithm. 1273 */ 1274 callout_init(&balco, CALLOUT_MPSAFE); 1275 callout_init(&gbalco, CALLOUT_MPSAFE); 1276 sched_fake_topo(); 1277 /* 1278 * Setup tdqs based on a topology configuration or vanilla SMP based 1279 * on mp_maxid. 1280 */ 1281 if (smp_topology == NULL) 1282 sched_setup_smp(); 1283 else 1284 sched_setup_topology(); 1285 sched_balance(NULL); 1286#else 1287 tdq_setup(tdq); 1288 mtx_init(&tdq_lock, "sched lock", "sched lock", MTX_SPIN | MTX_RECURSE); 1289 tdq->tdq_lock = &tdq_lock; 1290#endif 1291 /* 1292 * To avoid divide-by-zero, we set realstathz a dummy value 1293 * in case which sched_clock() called before sched_initticks(). 1294 */ 1295 realstathz = hz; 1296 sched_slice = (realstathz/10); /* ~100ms */ 1297 tickincr = 1 << SCHED_TICK_SHIFT; 1298 1299 /* Add thread0's load since it's running. */ 1300 TDQ_LOCK(tdq); 1301 thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1302 tdq_load_add(tdq, &td_sched0); 1303 TDQ_UNLOCK(tdq); 1304} 1305 1306/* 1307 * This routine determines the tickincr after stathz and hz are setup. 1308 */ 1309/* ARGSUSED */ 1310static void 1311sched_initticks(void *dummy) 1312{ 1313 int incr; 1314 1315 realstathz = stathz ? stathz : hz; 1316 sched_slice = (realstathz/10); /* ~100ms */ 1317 1318 /* 1319 * tickincr is shifted out by 10 to avoid rounding errors due to 1320 * hz not being evenly divisible by stathz on all platforms. 1321 */ 1322 incr = (hz << SCHED_TICK_SHIFT) / realstathz; 1323 /* 1324 * This does not work for values of stathz that are more than 1325 * 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen. 1326 */ 1327 if (incr == 0) 1328 incr = 1; 1329 tickincr = incr; 1330#ifdef SMP 1331 affinity = SCHED_AFFINITY_DEFAULT; 1332#endif 1333} 1334 1335 1336/* 1337 * This is the core of the interactivity algorithm. Determines a score based 1338 * on past behavior. It is the ratio of sleep time to run time scaled to 1339 * a [0, 100] integer. This is the voluntary sleep time of a process, which 1340 * differs from the cpu usage because it does not account for time spent 1341 * waiting on a run-queue. Would be prettier if we had floating point. 1342 */ 1343static int 1344sched_interact_score(struct thread *td) 1345{ 1346 struct td_sched *ts; 1347 int div; 1348 1349 ts = td->td_sched; 1350 /* 1351 * The score is only needed if this is likely to be an interactive 1352 * task. Don't go through the expense of computing it if there's 1353 * no chance. 1354 */ 1355 if (sched_interact <= SCHED_INTERACT_HALF && 1356 ts->ts_runtime >= ts->ts_slptime) 1357 return (SCHED_INTERACT_HALF); 1358 1359 if (ts->ts_runtime > ts->ts_slptime) { 1360 div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF); 1361 return (SCHED_INTERACT_HALF + 1362 (SCHED_INTERACT_HALF - (ts->ts_slptime / div))); 1363 } 1364 if (ts->ts_slptime > ts->ts_runtime) { 1365 div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF); 1366 return (ts->ts_runtime / div); 1367 } 1368 /* runtime == slptime */ 1369 if (ts->ts_runtime) 1370 return (SCHED_INTERACT_HALF); 1371 1372 /* 1373 * This can happen if slptime and runtime are 0. 1374 */ 1375 return (0); 1376 1377} 1378 1379/* 1380 * Scale the scheduling priority according to the "interactivity" of this 1381 * process. 1382 */ 1383static void 1384sched_priority(struct thread *td) 1385{ 1386 int score; 1387 int pri; 1388 1389 if (td->td_pri_class != PRI_TIMESHARE) 1390 return; 1391 /* 1392 * If the score is interactive we place the thread in the realtime 1393 * queue with a priority that is less than kernel and interrupt 1394 * priorities. These threads are not subject to nice restrictions. 1395 * 1396 * Scores greater than this are placed on the normal timeshare queue 1397 * where the priority is partially decided by the most recent cpu 1398 * utilization and the rest is decided by nice value. 1399 */ 1400 score = sched_interact_score(td); 1401 if (score < sched_interact) { 1402 pri = PRI_MIN_REALTIME; 1403 pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact) 1404 * score; 1405 KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME, 1406 ("sched_priority: invalid interactive priority %d score %d", 1407 pri, score)); 1408 } else { 1409 pri = SCHED_PRI_MIN; 1410 if (td->td_sched->ts_ticks) 1411 pri += SCHED_PRI_TICKS(td->td_sched); 1412 pri += SCHED_PRI_NICE(td->td_proc->p_nice); 1413 KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE, 1414 ("sched_priority: invalid priority %d: nice %d, " 1415 "ticks %d ftick %d ltick %d tick pri %d", 1416 pri, td->td_proc->p_nice, td->td_sched->ts_ticks, 1417 td->td_sched->ts_ftick, td->td_sched->ts_ltick, 1418 SCHED_PRI_TICKS(td->td_sched))); 1419 } 1420 sched_user_prio(td, pri); 1421 1422 return; 1423} 1424 1425/* 1426 * This routine enforces a maximum limit on the amount of scheduling history 1427 * kept. It is called after either the slptime or runtime is adjusted. This 1428 * function is ugly due to integer math. 1429 */ 1430static void 1431sched_interact_update(struct thread *td) 1432{ 1433 struct td_sched *ts; 1434 u_int sum; 1435 1436 ts = td->td_sched; 1437 sum = ts->ts_runtime + ts->ts_slptime; 1438 if (sum < SCHED_SLP_RUN_MAX) 1439 return; 1440 /* 1441 * This only happens from two places: 1442 * 1) We have added an unusual amount of run time from fork_exit. 1443 * 2) We have added an unusual amount of sleep time from sched_sleep(). 1444 */ 1445 if (sum > SCHED_SLP_RUN_MAX * 2) { 1446 if (ts->ts_runtime > ts->ts_slptime) { 1447 ts->ts_runtime = SCHED_SLP_RUN_MAX; 1448 ts->ts_slptime = 1; 1449 } else { 1450 ts->ts_slptime = SCHED_SLP_RUN_MAX; 1451 ts->ts_runtime = 1; 1452 } 1453 return; 1454 } 1455 /* 1456 * If we have exceeded by more than 1/5th then the algorithm below 1457 * will not bring us back into range. Dividing by two here forces 1458 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX] 1459 */ 1460 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) { 1461 ts->ts_runtime /= 2; 1462 ts->ts_slptime /= 2; 1463 return; 1464 } 1465 ts->ts_runtime = (ts->ts_runtime / 5) * 4; 1466 ts->ts_slptime = (ts->ts_slptime / 5) * 4; 1467} 1468 1469/* 1470 * Scale back the interactivity history when a child thread is created. The 1471 * history is inherited from the parent but the thread may behave totally 1472 * differently. For example, a shell spawning a compiler process. We want 1473 * to learn that the compiler is behaving badly very quickly. 1474 */ 1475static void 1476sched_interact_fork(struct thread *td) 1477{ 1478 int ratio; 1479 int sum; 1480 1481 sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime; 1482 if (sum > SCHED_SLP_RUN_FORK) { 1483 ratio = sum / SCHED_SLP_RUN_FORK; 1484 td->td_sched->ts_runtime /= ratio; 1485 td->td_sched->ts_slptime /= ratio; 1486 } 1487} 1488 1489/* 1490 * Called from proc0_init() to setup the scheduler fields. 1491 */ 1492void 1493schedinit(void) 1494{ 1495 1496 /* 1497 * Set up the scheduler specific parts of proc0. 1498 */ 1499 proc0.p_sched = NULL; /* XXX */ 1500 thread0.td_sched = &td_sched0; 1501 td_sched0.ts_ltick = ticks; 1502 td_sched0.ts_ftick = ticks; 1503 td_sched0.ts_thread = &thread0; 1504} 1505 1506/* 1507 * This is only somewhat accurate since given many processes of the same 1508 * priority they will switch when their slices run out, which will be 1509 * at most sched_slice stathz ticks. 1510 */ 1511int 1512sched_rr_interval(void) 1513{ 1514 1515 /* Convert sched_slice to hz */ 1516 return (hz/(realstathz/sched_slice)); 1517} 1518 1519/* 1520 * Update the percent cpu tracking information when it is requested or 1521 * the total history exceeds the maximum. We keep a sliding history of 1522 * tick counts that slowly decays. This is less precise than the 4BSD 1523 * mechanism since it happens with less regular and frequent events. 1524 */ 1525static void 1526sched_pctcpu_update(struct td_sched *ts) 1527{ 1528 1529 if (ts->ts_ticks == 0) 1530 return; 1531 if (ticks - (hz / 10) < ts->ts_ltick && 1532 SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX) 1533 return; 1534 /* 1535 * Adjust counters and watermark for pctcpu calc. 1536 */ 1537 if (ts->ts_ltick > ticks - SCHED_TICK_TARG) 1538 ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) * 1539 SCHED_TICK_TARG; 1540 else 1541 ts->ts_ticks = 0; 1542 ts->ts_ltick = ticks; 1543 ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG; 1544} 1545 1546/* 1547 * Adjust the priority of a thread. Move it to the appropriate run-queue 1548 * if necessary. This is the back-end for several priority related 1549 * functions. 1550 */ 1551static void 1552sched_thread_priority(struct thread *td, u_char prio) 1553{ 1554 struct td_sched *ts; 1555 1556 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)", 1557 td, td->td_proc->p_comm, td->td_priority, prio, curthread, 1558 curthread->td_proc->p_comm); 1559 ts = td->td_sched; 1560 THREAD_LOCK_ASSERT(td, MA_OWNED); 1561 if (td->td_priority == prio) 1562 return; 1563 1564 if (TD_ON_RUNQ(td) && prio < td->td_priority) { 1565 /* 1566 * If the priority has been elevated due to priority 1567 * propagation, we may have to move ourselves to a new 1568 * queue. This could be optimized to not re-add in some 1569 * cases. 1570 */ 1571 sched_rem(td); 1572 td->td_priority = prio; 1573 sched_add(td, SRQ_BORROWING); 1574 } else { 1575#ifdef SMP 1576 struct tdq *tdq; 1577 1578 tdq = TDQ_CPU(ts->ts_cpu); 1579 if (prio < tdq->tdq_lowpri) 1580 tdq->tdq_lowpri = prio; 1581#endif 1582 td->td_priority = prio; 1583 } 1584} 1585 1586/* 1587 * Update a thread's priority when it is lent another thread's 1588 * priority. 1589 */ 1590void 1591sched_lend_prio(struct thread *td, u_char prio) 1592{ 1593 1594 td->td_flags |= TDF_BORROWING; 1595 sched_thread_priority(td, prio); 1596} 1597 1598/* 1599 * Restore a thread's priority when priority propagation is 1600 * over. The prio argument is the minimum priority the thread 1601 * needs to have to satisfy other possible priority lending 1602 * requests. If the thread's regular priority is less 1603 * important than prio, the thread will keep a priority boost 1604 * of prio. 1605 */ 1606void 1607sched_unlend_prio(struct thread *td, u_char prio) 1608{ 1609 u_char base_pri; 1610 1611 if (td->td_base_pri >= PRI_MIN_TIMESHARE && 1612 td->td_base_pri <= PRI_MAX_TIMESHARE) 1613 base_pri = td->td_user_pri; 1614 else 1615 base_pri = td->td_base_pri; 1616 if (prio >= base_pri) { 1617 td->td_flags &= ~TDF_BORROWING; 1618 sched_thread_priority(td, base_pri); 1619 } else 1620 sched_lend_prio(td, prio); 1621} 1622 1623/* 1624 * Standard entry for setting the priority to an absolute value. 1625 */ 1626void 1627sched_prio(struct thread *td, u_char prio) 1628{ 1629 u_char oldprio; 1630 1631 /* First, update the base priority. */ 1632 td->td_base_pri = prio; 1633 1634 /* 1635 * If the thread is borrowing another thread's priority, don't 1636 * ever lower the priority. 1637 */ 1638 if (td->td_flags & TDF_BORROWING && td->td_priority < prio) 1639 return; 1640 1641 /* Change the real priority. */ 1642 oldprio = td->td_priority; 1643 sched_thread_priority(td, prio); 1644 1645 /* 1646 * If the thread is on a turnstile, then let the turnstile update 1647 * its state. 1648 */ 1649 if (TD_ON_LOCK(td) && oldprio != prio) 1650 turnstile_adjust(td, oldprio); 1651} 1652 1653/* 1654 * Set the base user priority, does not effect current running priority. 1655 */ 1656void 1657sched_user_prio(struct thread *td, u_char prio) 1658{ 1659 u_char oldprio; 1660 1661 td->td_base_user_pri = prio; 1662 if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio) 1663 return; 1664 oldprio = td->td_user_pri; 1665 td->td_user_pri = prio; 1666 1667 if (TD_ON_UPILOCK(td) && oldprio != prio) 1668 umtx_pi_adjust(td, oldprio); 1669} 1670 1671void 1672sched_lend_user_prio(struct thread *td, u_char prio) 1673{ 1674 u_char oldprio; 1675 1676 td->td_flags |= TDF_UBORROWING; 1677 1678 oldprio = td->td_user_pri; 1679 td->td_user_pri = prio; 1680 1681 if (TD_ON_UPILOCK(td) && oldprio != prio) 1682 umtx_pi_adjust(td, oldprio); 1683} 1684 1685void 1686sched_unlend_user_prio(struct thread *td, u_char prio) 1687{ 1688 u_char base_pri; 1689 1690 base_pri = td->td_base_user_pri; 1691 if (prio >= base_pri) { 1692 td->td_flags &= ~TDF_UBORROWING; 1693 sched_user_prio(td, base_pri); 1694 } else 1695 sched_lend_user_prio(td, prio); 1696} 1697 1698/* 1699 * Add the thread passed as 'newtd' to the run queue before selecting 1700 * the next thread to run. This is only used for KSE. 1701 */ 1702static void 1703sched_switchin(struct tdq *tdq, struct thread *td) 1704{ 1705#ifdef SMP 1706 spinlock_enter(); 1707 TDQ_UNLOCK(tdq); 1708 thread_lock(td); 1709 spinlock_exit(); 1710 sched_setcpu(td->td_sched, TDQ_ID(tdq), SRQ_YIELDING); 1711#else 1712 td->td_lock = TDQ_LOCKPTR(tdq); 1713#endif 1714 tdq_add(tdq, td, SRQ_YIELDING); 1715 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1716} 1717 1718/* 1719 * Handle migration from sched_switch(). This happens only for 1720 * cpu binding. 1721 */ 1722static struct mtx * 1723sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags) 1724{ 1725 struct tdq *tdn; 1726 1727 tdn = TDQ_CPU(td->td_sched->ts_cpu); 1728#ifdef SMP 1729 /* 1730 * Do the lock dance required to avoid LOR. We grab an extra 1731 * spinlock nesting to prevent preemption while we're 1732 * not holding either run-queue lock. 1733 */ 1734 spinlock_enter(); 1735 thread_block_switch(td); /* This releases the lock on tdq. */ 1736 TDQ_LOCK(tdn); 1737 tdq_add(tdn, td, flags); 1738 tdq_notify(td->td_sched); 1739 /* 1740 * After we unlock tdn the new cpu still can't switch into this 1741 * thread until we've unblocked it in cpu_switch(). The lock 1742 * pointers may match in the case of HTT cores. Don't unlock here 1743 * or we can deadlock when the other CPU runs the IPI handler. 1744 */ 1745 if (TDQ_LOCKPTR(tdn) != TDQ_LOCKPTR(tdq)) { 1746 TDQ_UNLOCK(tdn); 1747 TDQ_LOCK(tdq); 1748 } 1749 spinlock_exit(); 1750#endif 1751 return (TDQ_LOCKPTR(tdn)); 1752} 1753 1754/* 1755 * Block a thread for switching. Similar to thread_block() but does not 1756 * bump the spin count. 1757 */ 1758static inline struct mtx * 1759thread_block_switch(struct thread *td) 1760{ 1761 struct mtx *lock; 1762 1763 THREAD_LOCK_ASSERT(td, MA_OWNED); 1764 lock = td->td_lock; 1765 td->td_lock = &blocked_lock; 1766 mtx_unlock_spin(lock); 1767 1768 return (lock); 1769} 1770 1771/* 1772 * Release a thread that was blocked with thread_block_switch(). 1773 */ 1774static inline void 1775thread_unblock_switch(struct thread *td, struct mtx *mtx) 1776{ 1777 atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock, 1778 (uintptr_t)mtx); 1779} 1780 1781/* 1782 * Switch threads. This function has to handle threads coming in while 1783 * blocked for some reason, running, or idle. It also must deal with 1784 * migrating a thread from one queue to another as running threads may 1785 * be assigned elsewhere via binding. 1786 */ 1787void 1788sched_switch(struct thread *td, struct thread *newtd, int flags) 1789{ 1790 struct tdq *tdq; 1791 struct td_sched *ts; 1792 struct mtx *mtx; 1793 int srqflag; 1794 int cpuid; 1795 1796 THREAD_LOCK_ASSERT(td, MA_OWNED); 1797 1798 cpuid = PCPU_GET(cpuid); 1799 tdq = TDQ_CPU(cpuid); 1800 ts = td->td_sched; 1801 mtx = td->td_lock; 1802#ifdef SMP 1803 ts->ts_rltick = ticks; 1804 if (newtd && newtd->td_priority < tdq->tdq_lowpri) 1805 tdq->tdq_lowpri = newtd->td_priority; 1806#endif 1807 td->td_lastcpu = td->td_oncpu; 1808 td->td_oncpu = NOCPU; 1809 td->td_flags &= ~TDF_NEEDRESCHED; 1810 td->td_owepreempt = 0; 1811 /* 1812 * The lock pointer in an idle thread should never change. Reset it 1813 * to CAN_RUN as well. 1814 */ 1815 if (TD_IS_IDLETHREAD(td)) { 1816 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1817 TD_SET_CAN_RUN(td); 1818 } else if (TD_IS_RUNNING(td)) { 1819 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1820 tdq_load_rem(tdq, ts); 1821 srqflag = (flags & SW_PREEMPT) ? 1822 SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : 1823 SRQ_OURSELF|SRQ_YIELDING; 1824 if (ts->ts_cpu == cpuid) 1825 tdq_add(tdq, td, srqflag); 1826 else 1827 mtx = sched_switch_migrate(tdq, td, srqflag); 1828 } else { 1829 /* This thread must be going to sleep. */ 1830 TDQ_LOCK(tdq); 1831 mtx = thread_block_switch(td); 1832 tdq_load_rem(tdq, ts); 1833 } 1834 /* 1835 * We enter here with the thread blocked and assigned to the 1836 * appropriate cpu run-queue or sleep-queue and with the current 1837 * thread-queue locked. 1838 */ 1839 TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED); 1840 /* 1841 * If KSE assigned a new thread just add it here and let choosethread 1842 * select the best one. 1843 */ 1844 if (newtd != NULL) 1845 sched_switchin(tdq, newtd); 1846 newtd = choosethread(); 1847 /* 1848 * Call the MD code to switch contexts if necessary. 1849 */ 1850 if (td != newtd) { 1851#ifdef HWPMC_HOOKS 1852 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1853 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); 1854#endif 1855 cpu_switch(td, newtd, mtx); 1856 /* 1857 * We may return from cpu_switch on a different cpu. However, 1858 * we always return with td_lock pointing to the current cpu's 1859 * run queue lock. 1860 */ 1861 cpuid = PCPU_GET(cpuid); 1862 tdq = TDQ_CPU(cpuid); 1863 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)td; 1864#ifdef HWPMC_HOOKS 1865 if (PMC_PROC_IS_USING_PMCS(td->td_proc)) 1866 PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); 1867#endif 1868 } else 1869 thread_unblock_switch(td, mtx); 1870 /* 1871 * Assert that all went well and return. 1872 */ 1873#ifdef SMP 1874 /* We should always get here with the lowest priority td possible */ 1875 tdq->tdq_lowpri = td->td_priority; 1876#endif 1877 TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED); 1878 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 1879 td->td_oncpu = cpuid; 1880} 1881 1882/* 1883 * Adjust thread priorities as a result of a nice request. 1884 */ 1885void 1886sched_nice(struct proc *p, int nice) 1887{ 1888 struct thread *td; 1889 1890 PROC_LOCK_ASSERT(p, MA_OWNED); 1891 PROC_SLOCK_ASSERT(p, MA_OWNED); 1892 1893 p->p_nice = nice; 1894 FOREACH_THREAD_IN_PROC(p, td) { 1895 thread_lock(td); 1896 sched_priority(td); 1897 sched_prio(td, td->td_base_user_pri); 1898 thread_unlock(td); 1899 } 1900} 1901 1902/* 1903 * Record the sleep time for the interactivity scorer. 1904 */ 1905void 1906sched_sleep(struct thread *td) 1907{ 1908 1909 THREAD_LOCK_ASSERT(td, MA_OWNED); 1910 1911 td->td_sched->ts_slptick = ticks; 1912} 1913 1914/* 1915 * Schedule a thread to resume execution and record how long it voluntarily 1916 * slept. We also update the pctcpu, interactivity, and priority. 1917 */ 1918void 1919sched_wakeup(struct thread *td) 1920{ 1921 struct td_sched *ts; 1922 int slptick; 1923 1924 THREAD_LOCK_ASSERT(td, MA_OWNED); 1925 ts = td->td_sched; 1926 /* 1927 * If we slept for more than a tick update our interactivity and 1928 * priority. 1929 */ 1930 slptick = ts->ts_slptick; 1931 ts->ts_slptick = 0; 1932 if (slptick && slptick != ticks) { 1933 u_int hzticks; 1934 1935 hzticks = (ticks - slptick) << SCHED_TICK_SHIFT; 1936 ts->ts_slptime += hzticks; 1937 sched_interact_update(td); 1938 sched_pctcpu_update(ts); 1939 sched_priority(td); 1940 } 1941 /* Reset the slice value after we sleep. */ 1942 ts->ts_slice = sched_slice; 1943 sched_add(td, SRQ_BORING); 1944} 1945 1946/* 1947 * Penalize the parent for creating a new child and initialize the child's 1948 * priority. 1949 */ 1950void 1951sched_fork(struct thread *td, struct thread *child) 1952{ 1953 THREAD_LOCK_ASSERT(td, MA_OWNED); 1954 sched_fork_thread(td, child); 1955 /* 1956 * Penalize the parent and child for forking. 1957 */ 1958 sched_interact_fork(child); 1959 sched_priority(child); 1960 td->td_sched->ts_runtime += tickincr; 1961 sched_interact_update(td); 1962 sched_priority(td); 1963} 1964 1965/* 1966 * Fork a new thread, may be within the same process. 1967 */ 1968void 1969sched_fork_thread(struct thread *td, struct thread *child) 1970{ 1971 struct td_sched *ts; 1972 struct td_sched *ts2; 1973 1974 /* 1975 * Initialize child. 1976 */ 1977 THREAD_LOCK_ASSERT(td, MA_OWNED); 1978 sched_newthread(child); 1979 child->td_lock = TDQ_LOCKPTR(TDQ_SELF()); 1980 ts = td->td_sched; 1981 ts2 = child->td_sched; 1982 ts2->ts_cpu = ts->ts_cpu; 1983 ts2->ts_runq = NULL; 1984 /* 1985 * Grab our parents cpu estimation information and priority. 1986 */ 1987 ts2->ts_ticks = ts->ts_ticks; 1988 ts2->ts_ltick = ts->ts_ltick; 1989 ts2->ts_ftick = ts->ts_ftick; 1990 child->td_user_pri = td->td_user_pri; 1991 child->td_base_user_pri = td->td_base_user_pri; 1992 /* 1993 * And update interactivity score. 1994 */ 1995 ts2->ts_slptime = ts->ts_slptime; 1996 ts2->ts_runtime = ts->ts_runtime; 1997 ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */ 1998} 1999 2000/* 2001 * Adjust the priority class of a thread. 2002 */ 2003void 2004sched_class(struct thread *td, int class) 2005{ 2006 2007 THREAD_LOCK_ASSERT(td, MA_OWNED); 2008 if (td->td_pri_class == class) 2009 return; 2010 2011#ifdef SMP 2012 /* 2013 * On SMP if we're on the RUNQ we must adjust the transferable 2014 * count because could be changing to or from an interrupt 2015 * class. 2016 */ 2017 if (TD_ON_RUNQ(td)) { 2018 struct tdq *tdq; 2019 2020 tdq = TDQ_CPU(td->td_sched->ts_cpu); 2021 if (THREAD_CAN_MIGRATE(td)) { 2022 tdq->tdq_transferable--; 2023 tdq->tdq_group->tdg_transferable--; 2024 } 2025 td->td_pri_class = class; 2026 if (THREAD_CAN_MIGRATE(td)) { 2027 tdq->tdq_transferable++; 2028 tdq->tdq_group->tdg_transferable++; 2029 } 2030 } 2031#endif 2032 td->td_pri_class = class; 2033} 2034 2035/* 2036 * Return some of the child's priority and interactivity to the parent. 2037 */ 2038void 2039sched_exit(struct proc *p, struct thread *child) 2040{ 2041 struct thread *td; 2042 2043 CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d", 2044 child, child->td_proc->p_comm, child->td_priority); 2045 2046 PROC_SLOCK_ASSERT(p, MA_OWNED); 2047 td = FIRST_THREAD_IN_PROC(p); 2048 sched_exit_thread(td, child); 2049} 2050 2051/* 2052 * Penalize another thread for the time spent on this one. This helps to 2053 * worsen the priority and interactivity of processes which schedule batch 2054 * jobs such as make. This has little effect on the make process itself but 2055 * causes new processes spawned by it to receive worse scores immediately. 2056 */ 2057void 2058sched_exit_thread(struct thread *td, struct thread *child) 2059{ 2060 2061 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d", 2062 child, child->td_proc->p_comm, child->td_priority); 2063 2064#ifdef KSE 2065 /* 2066 * KSE forks and exits so often that this penalty causes short-lived 2067 * threads to always be non-interactive. This causes mozilla to 2068 * crawl under load. 2069 */ 2070 if ((td->td_pflags & TDP_SA) && td->td_proc == child->td_proc) 2071 return; 2072#endif 2073 /* 2074 * Give the child's runtime to the parent without returning the 2075 * sleep time as a penalty to the parent. This causes shells that 2076 * launch expensive things to mark their children as expensive. 2077 */ 2078 thread_lock(td); 2079 td->td_sched->ts_runtime += child->td_sched->ts_runtime; 2080 sched_interact_update(td); 2081 sched_priority(td); 2082 thread_unlock(td); 2083} 2084 2085/* 2086 * Fix priorities on return to user-space. Priorities may be elevated due 2087 * to static priorities in msleep() or similar. 2088 */ 2089void 2090sched_userret(struct thread *td) 2091{ 2092 /* 2093 * XXX we cheat slightly on the locking here to avoid locking in 2094 * the usual case. Setting td_priority here is essentially an 2095 * incomplete workaround for not setting it properly elsewhere. 2096 * Now that some interrupt handlers are threads, not setting it 2097 * properly elsewhere can clobber it in the window between setting 2098 * it here and returning to user mode, so don't waste time setting 2099 * it perfectly here. 2100 */ 2101 KASSERT((td->td_flags & TDF_BORROWING) == 0, 2102 ("thread with borrowed priority returning to userland")); 2103 if (td->td_priority != td->td_user_pri) { 2104 thread_lock(td); 2105 td->td_priority = td->td_user_pri; 2106 td->td_base_pri = td->td_user_pri; 2107 thread_unlock(td); 2108 } 2109} 2110 2111/* 2112 * Handle a stathz tick. This is really only relevant for timeshare 2113 * threads. 2114 */ 2115void 2116sched_clock(struct thread *td) 2117{ 2118 struct tdq *tdq; 2119 struct td_sched *ts; 2120 2121 THREAD_LOCK_ASSERT(td, MA_OWNED); 2122 tdq = TDQ_SELF(); 2123 /* 2124 * Advance the insert index once for each tick to ensure that all 2125 * threads get a chance to run. 2126 */ 2127 if (tdq->tdq_idx == tdq->tdq_ridx) { 2128 tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS; 2129 if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx])) 2130 tdq->tdq_ridx = tdq->tdq_idx; 2131 } 2132 ts = td->td_sched; 2133 /* 2134 * We only do slicing code for TIMESHARE threads. 2135 */ 2136 if (td->td_pri_class != PRI_TIMESHARE) 2137 return; 2138 /* 2139 * We used a tick; charge it to the thread so that we can compute our 2140 * interactivity. 2141 */ 2142 td->td_sched->ts_runtime += tickincr; 2143 sched_interact_update(td); 2144 /* 2145 * We used up one time slice. 2146 */ 2147 if (--ts->ts_slice > 0) 2148 return; 2149 /* 2150 * We're out of time, recompute priorities and requeue. 2151 */ 2152 sched_priority(td); 2153 td->td_flags |= TDF_NEEDRESCHED; 2154} 2155 2156/* 2157 * Called once per hz tick. Used for cpu utilization information. This 2158 * is easier than trying to scale based on stathz. 2159 */ 2160void 2161sched_tick(void) 2162{ 2163 struct td_sched *ts; 2164 2165 ts = curthread->td_sched; 2166 /* Adjust ticks for pctcpu */ 2167 ts->ts_ticks += 1 << SCHED_TICK_SHIFT; 2168 ts->ts_ltick = ticks; 2169 /* 2170 * Update if we've exceeded our desired tick threshhold by over one 2171 * second. 2172 */ 2173 if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick) 2174 sched_pctcpu_update(ts); 2175} 2176 2177/* 2178 * Return whether the current CPU has runnable tasks. Used for in-kernel 2179 * cooperative idle threads. 2180 */ 2181int 2182sched_runnable(void) 2183{ 2184 struct tdq *tdq; 2185 int load; 2186 2187 load = 1; 2188 2189 tdq = TDQ_SELF(); 2190 if ((curthread->td_flags & TDF_IDLETD) != 0) { 2191 if (tdq->tdq_load > 0) 2192 goto out; 2193 } else 2194 if (tdq->tdq_load - 1 > 0) 2195 goto out; 2196 load = 0; 2197out: 2198 return (load); 2199} 2200 2201/* 2202 * Choose the highest priority thread to run. The thread is removed from 2203 * the run-queue while running however the load remains. For SMP we set 2204 * the tdq in the global idle bitmask if it idles here. 2205 */ 2206struct thread * 2207sched_choose(void) 2208{ 2209#ifdef SMP 2210 struct tdq_group *tdg; 2211#endif 2212 struct td_sched *ts; 2213 struct tdq *tdq; 2214 2215 tdq = TDQ_SELF(); 2216 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2217 ts = tdq_choose(tdq); 2218 if (ts) { 2219 tdq_runq_rem(tdq, ts); 2220 return (ts->ts_thread); 2221 } 2222#ifdef SMP 2223 /* 2224 * We only set the idled bit when all of the cpus in the group are 2225 * idle. Otherwise we could get into a situation where a thread bounces 2226 * back and forth between two idle cores on seperate physical CPUs. 2227 */ 2228 tdg = tdq->tdq_group; 2229 tdg->tdg_idlemask |= PCPU_GET(cpumask); 2230 if (tdg->tdg_idlemask == tdg->tdg_cpumask) 2231 atomic_set_int(&tdq_idle, tdg->tdg_mask); 2232 tdq->tdq_lowpri = PRI_MAX_IDLE; 2233#endif 2234 return (PCPU_GET(idlethread)); 2235} 2236 2237/* 2238 * Set owepreempt if necessary. Preemption never happens directly in ULE, 2239 * we always request it once we exit a critical section. 2240 */ 2241static inline void 2242sched_setpreempt(struct thread *td) 2243{ 2244 struct thread *ctd; 2245 int cpri; 2246 int pri; 2247 2248 ctd = curthread; 2249 pri = td->td_priority; 2250 cpri = ctd->td_priority; 2251 if (td->td_priority < ctd->td_priority) 2252 curthread->td_flags |= TDF_NEEDRESCHED; 2253 if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd)) 2254 return; 2255 /* 2256 * Always preempt IDLE threads. Otherwise only if the preempting 2257 * thread is an ithread. 2258 */ 2259 if (pri > preempt_thresh && cpri < PRI_MIN_IDLE) 2260 return; 2261 ctd->td_owepreempt = 1; 2262 return; 2263} 2264 2265/* 2266 * Add a thread to a thread queue. Initializes priority, slice, runq, and 2267 * add it to the appropriate queue. This is the internal function called 2268 * when the tdq is predetermined. 2269 */ 2270void 2271tdq_add(struct tdq *tdq, struct thread *td, int flags) 2272{ 2273 struct td_sched *ts; 2274 int class; 2275#ifdef SMP 2276 int cpumask; 2277#endif 2278 2279 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2280 KASSERT((td->td_inhibitors == 0), 2281 ("sched_add: trying to run inhibited thread")); 2282 KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), 2283 ("sched_add: bad thread state")); 2284 KASSERT(td->td_proc->p_sflag & PS_INMEM, 2285 ("sched_add: process swapped out")); 2286 2287 ts = td->td_sched; 2288 class = PRI_BASE(td->td_pri_class); 2289 TD_SET_RUNQ(td); 2290 if (ts->ts_slice == 0) 2291 ts->ts_slice = sched_slice; 2292 /* 2293 * Pick the run queue based on priority. 2294 */ 2295 if (td->td_priority <= PRI_MAX_REALTIME) 2296 ts->ts_runq = &tdq->tdq_realtime; 2297 else if (td->td_priority <= PRI_MAX_TIMESHARE) 2298 ts->ts_runq = &tdq->tdq_timeshare; 2299 else 2300 ts->ts_runq = &tdq->tdq_idle; 2301#ifdef SMP 2302 cpumask = 1 << ts->ts_cpu; 2303 /* 2304 * If we had been idle, clear our bit in the group and potentially 2305 * the global bitmap. 2306 */ 2307 if ((class != PRI_IDLE && class != PRI_ITHD) && 2308 (tdq->tdq_group->tdg_idlemask & cpumask) != 0) { 2309 /* 2310 * Check to see if our group is unidling, and if so, remove it 2311 * from the global idle mask. 2312 */ 2313 if (tdq->tdq_group->tdg_idlemask == 2314 tdq->tdq_group->tdg_cpumask) 2315 atomic_clear_int(&tdq_idle, tdq->tdq_group->tdg_mask); 2316 /* 2317 * Now remove ourselves from the group specific idle mask. 2318 */ 2319 tdq->tdq_group->tdg_idlemask &= ~cpumask; 2320 } 2321 if (td->td_priority < tdq->tdq_lowpri) 2322 tdq->tdq_lowpri = td->td_priority; 2323#endif 2324 tdq_runq_add(tdq, ts, flags); 2325 tdq_load_add(tdq, ts); 2326} 2327 2328/* 2329 * Select the target thread queue and add a thread to it. Request 2330 * preemption or IPI a remote processor if required. 2331 */ 2332void 2333sched_add(struct thread *td, int flags) 2334{ 2335 struct td_sched *ts; 2336 struct tdq *tdq; 2337#ifdef SMP 2338 int cpuid; 2339 int cpu; 2340#endif 2341 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)", 2342 td, td->td_proc->p_comm, td->td_priority, curthread, 2343 curthread->td_proc->p_comm); 2344 THREAD_LOCK_ASSERT(td, MA_OWNED); 2345 ts = td->td_sched; 2346 /* 2347 * Recalculate the priority before we select the target cpu or 2348 * run-queue. 2349 */ 2350 if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) 2351 sched_priority(td); 2352#ifdef SMP 2353 cpuid = PCPU_GET(cpuid); 2354 /* 2355 * Pick the destination cpu and if it isn't ours transfer to the 2356 * target cpu. 2357 */ 2358 if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_MIGRATE(td)) 2359 cpu = cpuid; 2360 else if (!THREAD_CAN_MIGRATE(td)) 2361 cpu = ts->ts_cpu; 2362 else 2363 cpu = sched_pickcpu(ts, flags); 2364 tdq = sched_setcpu(ts, cpu, flags); 2365 tdq_add(tdq, td, flags); 2366 if (cpu != cpuid) { 2367 tdq_notify(ts); 2368 return; 2369 } 2370#else 2371 tdq = TDQ_SELF(); 2372 TDQ_LOCK(tdq); 2373 /* 2374 * Now that the thread is moving to the run-queue, set the lock 2375 * to the scheduler's lock. 2376 */ 2377 thread_lock_set(td, TDQ_LOCKPTR(tdq)); 2378 tdq_add(tdq, td, flags); 2379#endif 2380 if (!(flags & SRQ_YIELDING)) 2381 sched_setpreempt(td); 2382} 2383 2384/* 2385 * Remove a thread from a run-queue without running it. This is used 2386 * when we're stealing a thread from a remote queue. Otherwise all threads 2387 * exit by calling sched_exit_thread() and sched_throw() themselves. 2388 */ 2389void 2390sched_rem(struct thread *td) 2391{ 2392 struct tdq *tdq; 2393 struct td_sched *ts; 2394 2395 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)", 2396 td, td->td_proc->p_comm, td->td_priority, curthread, 2397 curthread->td_proc->p_comm); 2398 ts = td->td_sched; 2399 tdq = TDQ_CPU(ts->ts_cpu); 2400 TDQ_LOCK_ASSERT(tdq, MA_OWNED); 2401 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2402 KASSERT(TD_ON_RUNQ(td), 2403 ("sched_rem: thread not on run queue")); 2404 tdq_runq_rem(tdq, ts); 2405 tdq_load_rem(tdq, ts); 2406 TD_SET_CAN_RUN(td); 2407} 2408 2409/* 2410 * Fetch cpu utilization information. Updates on demand. 2411 */ 2412fixpt_t 2413sched_pctcpu(struct thread *td) 2414{ 2415 fixpt_t pctcpu; 2416 struct td_sched *ts; 2417 2418 pctcpu = 0; 2419 ts = td->td_sched; 2420 if (ts == NULL) 2421 return (0); 2422 2423 thread_lock(td); 2424 if (ts->ts_ticks) { 2425 int rtick; 2426 2427 sched_pctcpu_update(ts); 2428 /* How many rtick per second ? */ 2429 rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz); 2430 pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT; 2431 } 2432 td->td_proc->p_swtime = ts->ts_ltick - ts->ts_ftick; 2433 thread_unlock(td); 2434 2435 return (pctcpu); 2436} 2437 2438/* 2439 * Bind a thread to a target cpu. 2440 */ 2441void 2442sched_bind(struct thread *td, int cpu) 2443{ 2444 struct td_sched *ts; 2445 2446 THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); 2447 ts = td->td_sched; 2448 if (ts->ts_flags & TSF_BOUND) 2449 sched_unbind(td); 2450 ts->ts_flags |= TSF_BOUND; 2451#ifdef SMP 2452 sched_pin(); 2453 if (PCPU_GET(cpuid) == cpu) 2454 return; 2455 ts->ts_cpu = cpu; 2456 /* When we return from mi_switch we'll be on the correct cpu. */ 2457 mi_switch(SW_VOL, NULL); 2458#endif 2459} 2460 2461/* 2462 * Release a bound thread. 2463 */ 2464void 2465sched_unbind(struct thread *td) 2466{ 2467 struct td_sched *ts; 2468 2469 THREAD_LOCK_ASSERT(td, MA_OWNED); 2470 ts = td->td_sched; 2471 if ((ts->ts_flags & TSF_BOUND) == 0) 2472 return; 2473 ts->ts_flags &= ~TSF_BOUND; 2474#ifdef SMP 2475 sched_unpin(); 2476#endif 2477} 2478 2479int 2480sched_is_bound(struct thread *td) 2481{ 2482 THREAD_LOCK_ASSERT(td, MA_OWNED); 2483 return (td->td_sched->ts_flags & TSF_BOUND); 2484} 2485 2486/* 2487 * Basic yield call. 2488 */ 2489void 2490sched_relinquish(struct thread *td) 2491{ 2492 thread_lock(td); 2493 if (td->td_pri_class == PRI_TIMESHARE) 2494 sched_prio(td, PRI_MAX_TIMESHARE); 2495 SCHED_STAT_INC(switch_relinquish); 2496 mi_switch(SW_VOL, NULL); 2497 thread_unlock(td); 2498} 2499 2500/* 2501 * Return the total system load. 2502 */ 2503int 2504sched_load(void) 2505{ 2506#ifdef SMP 2507 int total; 2508 int i; 2509 2510 total = 0; 2511 for (i = 0; i <= tdg_maxid; i++) 2512 total += TDQ_GROUP(i)->tdg_load; 2513 return (total); 2514#else 2515 return (TDQ_SELF()->tdq_sysload); 2516#endif 2517} 2518 2519int 2520sched_sizeof_proc(void) 2521{ 2522 return (sizeof(struct proc)); 2523} 2524 2525int 2526sched_sizeof_thread(void) 2527{ 2528 return (sizeof(struct thread) + sizeof(struct td_sched)); 2529} 2530 2531/* 2532 * The actual idle process. 2533 */ 2534void 2535sched_idletd(void *dummy) 2536{ 2537 struct thread *td; 2538 struct tdq *tdq; 2539 2540 td = curthread; 2541 tdq = TDQ_SELF(); 2542 mtx_assert(&Giant, MA_NOTOWNED); 2543 /* ULE relies on preemption for idle interruption. */ 2544 for (;;) { 2545#ifdef SMP 2546 if (tdq_idled(tdq)) 2547 cpu_idle(); 2548#else 2549 cpu_idle(); 2550#endif 2551 } 2552} 2553 2554/* 2555 * A CPU is entering for the first time or a thread is exiting. 2556 */ 2557void 2558sched_throw(struct thread *td) 2559{ 2560 struct tdq *tdq; 2561 2562 tdq = TDQ_SELF(); 2563 if (td == NULL) { 2564 /* Correct spinlock nesting and acquire the correct lock. */ 2565 TDQ_LOCK(tdq); 2566 spinlock_exit(); 2567 } else { 2568 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2569 tdq_load_rem(tdq, td->td_sched); 2570 } 2571 KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); 2572 PCPU_SET(switchtime, cpu_ticks()); 2573 PCPU_SET(switchticks, ticks); 2574 cpu_throw(td, choosethread()); /* doesn't return */ 2575} 2576 2577/* 2578 * This is called from fork_exit(). Just acquire the correct locks and 2579 * let fork do the rest of the work. 2580 */ 2581void 2582sched_fork_exit(struct thread *td) 2583{ 2584 struct td_sched *ts; 2585 struct tdq *tdq; 2586 int cpuid; 2587 2588 /* 2589 * Finish setting up thread glue so that it begins execution in a 2590 * non-nested critical section with the scheduler lock held. 2591 */ 2592 cpuid = PCPU_GET(cpuid); 2593 tdq = TDQ_CPU(cpuid); 2594 ts = td->td_sched; 2595 if (TD_IS_IDLETHREAD(td)) 2596 td->td_lock = TDQ_LOCKPTR(tdq); 2597 MPASS(td->td_lock == TDQ_LOCKPTR(tdq)); 2598 td->td_oncpu = cpuid; 2599 TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)td; 2600 THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED); 2601} 2602 2603static SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, 2604 "Scheduler"); 2605SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0, 2606 "Scheduler name"); 2607SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, 2608 "Slice size for timeshare threads"); 2609SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, 2610 "Interactivity score threshold"); 2611SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh, 2612 0,"Min priority for preemption, lower priorities have greater precedence"); 2613#ifdef SMP 2614SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri, CTLFLAG_RW, &pick_pri, 0, 2615 "Pick the target cpu based on priority rather than load."); 2616SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0, 2617 "Number of hz ticks to keep thread affinity for"); 2618SYSCTL_INT(_kern_sched, OID_AUTO, tryself, CTLFLAG_RW, &tryself, 0, ""); 2619SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, 2620 "Enables the long-term load balancer"); 2621SYSCTL_INT(_kern_sched, OID_AUTO, balance_secs, CTLFLAG_RW, &balance_secs, 0, 2622 "Average frequence in seconds to run the long-term balancer"); 2623SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, 2624 "Steals work from another hyper-threaded core on idle"); 2625SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0, 2626 "Attempts to steal work from other cores before idling"); 2627SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0, 2628 "Minimum load on remote cpu before we'll steal"); 2629SYSCTL_INT(_kern_sched, OID_AUTO, topology, CTLFLAG_RD, &topology, 0, 2630 "True when a topology has been specified by the MD code."); 2631#endif 2632 2633/* ps compat */ 2634static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 2635SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 2636 2637 2638#define KERN_SWITCH_INCLUDE 1 2639#include "kern/kern_switch.c" 2640