sched_ule.c revision 165627
1109864Sjeff/*-
2165620Sjeff * Copyright (c) 2002-2006, 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
27116182Sobrien#include <sys/cdefs.h>
28116182Sobrien__FBSDID("$FreeBSD: head/sys/kern/sched_ule.c 165627 2006-12-29 12:55:32Z jeff $");
29116182Sobrien
30147565Speter#include "opt_hwpmc_hooks.h"
31147565Speter#include "opt_sched.h"
32134649Sscottl
33109864Sjeff#include <sys/param.h>
34109864Sjeff#include <sys/systm.h>
35131929Smarcel#include <sys/kdb.h>
36109864Sjeff#include <sys/kernel.h>
37109864Sjeff#include <sys/ktr.h>
38109864Sjeff#include <sys/lock.h>
39109864Sjeff#include <sys/mutex.h>
40109864Sjeff#include <sys/proc.h>
41112966Sjeff#include <sys/resource.h>
42122038Sjeff#include <sys/resourcevar.h>
43109864Sjeff#include <sys/sched.h>
44109864Sjeff#include <sys/smp.h>
45109864Sjeff#include <sys/sx.h>
46109864Sjeff#include <sys/sysctl.h>
47109864Sjeff#include <sys/sysproto.h>
48139453Sjhb#include <sys/turnstile.h>
49161599Sdavidxu#include <sys/umtx.h>
50109864Sjeff#include <sys/vmmeter.h>
51109864Sjeff#ifdef KTRACE
52109864Sjeff#include <sys/uio.h>
53109864Sjeff#include <sys/ktrace.h>
54109864Sjeff#endif
55109864Sjeff
56145256Sjkoshy#ifdef HWPMC_HOOKS
57145256Sjkoshy#include <sys/pmckern.h>
58145256Sjkoshy#endif
59145256Sjkoshy
60109864Sjeff#include <machine/cpu.h>
61121790Sjeff#include <machine/smp.h>
62109864Sjeff
63109864Sjeff/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
64109864Sjeff/* XXX This is bogus compatability crap for ps */
65109864Sjeffstatic fixpt_t  ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
66109864SjeffSYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
67109864Sjeff
68109864Sjeffstatic void sched_setup(void *dummy);
69109864SjeffSYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL)
70109864Sjeff
71153533Sdavidxustatic void sched_initticks(void *dummy);
72153533SdavidxuSYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, NULL)
73153533Sdavidxu
74132589Sscottlstatic SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
75113357Sjeff
76132589SscottlSYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ule", 0,
77132589Sscottl    "Scheduler name");
78130881Sscottl
79113357Sjeffstatic int slice_min = 1;
80113357SjeffSYSCTL_INT(_kern_sched, OID_AUTO, slice_min, CTLFLAG_RW, &slice_min, 0, "");
81113357Sjeff
82116365Sjeffstatic int slice_max = 10;
83113357SjeffSYSCTL_INT(_kern_sched, OID_AUTO, slice_max, CTLFLAG_RW, &slice_max, 0, "");
84113357Sjeff
85111857Sjeffint realstathz;
86153533Sdavidxuint tickincr = 1 << 10;
87111857Sjeff
88109864Sjeff/*
89164936Sjulian * Thread scheduler specific section.
90146954Sjeff */
91164936Sjulianstruct td_sched {
92164936Sjulian	TAILQ_ENTRY(td_sched) ts_procq;	/* (j/z) Run queue. */
93164936Sjulian	int		ts_flags;	/* (j) TSF_* flags. */
94164936Sjulian	struct thread	*ts_thread;	/* (*) Active associated thread. */
95164936Sjulian	fixpt_t		ts_pctcpu;	/* (j) %cpu during p_swtime. */
96164936Sjulian	u_char		ts_rqindex;	/* (j) Run queue index. */
97134791Sjulian	enum {
98164936Sjulian		TSS_THREAD = 0x0,	/* slaved to thread state */
99164936Sjulian		TSS_ONRUNQ
100164936Sjulian	} ts_state;			/* (j) thread sched specific status. */
101164936Sjulian	int		ts_slptime;
102164936Sjulian	int		ts_slice;
103164936Sjulian	struct runq	*ts_runq;
104164936Sjulian	u_char		ts_cpu;		/* CPU that we have affinity for. */
105134791Sjulian	/* The following variables are only used for pctcpu calculation */
106164936Sjulian	int		ts_ltick;	/* Last tick that we were running on */
107164936Sjulian	int		ts_ftick;	/* First tick that we were running on */
108164936Sjulian	int		ts_ticks;	/* Tick count */
109134791Sjulian
110163709Sjb	/* originally from kg_sched */
111163709Sjb	int	skg_slptime;		/* Number of ticks we vol. slept */
112163709Sjb	int	skg_runtime;		/* Number of ticks we were running */
113134791Sjulian};
114164936Sjulian#define	ts_assign		ts_procq.tqe_next
115164936Sjulian/* flags kept in ts_flags */
116164936Sjulian#define	TSF_ASSIGNED	0x0001		/* Thread is being migrated. */
117164936Sjulian#define	TSF_BOUND	0x0002		/* Thread can not migrate. */
118164936Sjulian#define	TSF_XFERABLE	0x0004		/* Thread was added as transferable. */
119164936Sjulian#define	TSF_HOLD	0x0008		/* Thread is temporarily bound. */
120164936Sjulian#define	TSF_REMOVED	0x0010		/* Thread was removed while ASSIGNED */
121164936Sjulian#define	TSF_INTERNAL	0x0020		/* Thread added due to migration. */
122164936Sjulian#define	TSF_PREEMPTED	0x0040		/* Thread was preempted */
123165620Sjeff#define	TSF_DIDRUN	0x2000		/* Thread actually ran. */
124165620Sjeff#define	TSF_EXIT	0x4000		/* Thread is being killed. */
125121790Sjeff
126164936Sjulianstatic struct td_sched td_sched0;
127109864Sjeff
128109864Sjeff/*
129116642Sjeff * The priority is primarily determined by the interactivity score.  Thus, we
130165620Sjeff * give lower(better) priorities to threads that use less CPU.  The nice
131116642Sjeff * value is then directly added to this to allow nice to have some effect
132116642Sjeff * on latency.
133111857Sjeff *
134111857Sjeff * PRI_RANGE:	Total priority range for timeshare threads.
135116642Sjeff * PRI_NRESV:	Number of nice values.
136111857Sjeff * PRI_BASE:	The start of the dynamic range.
137109864Sjeff */
138111857Sjeff#define	SCHED_PRI_RANGE		(PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
139121869Sjeff#define	SCHED_PRI_NRESV		((PRIO_MAX - PRIO_MIN) + 1)
140121869Sjeff#define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
141116642Sjeff#define	SCHED_PRI_BASE		(PRI_MIN_TIMESHARE)
142113357Sjeff#define	SCHED_PRI_INTERACT(score)					\
143116642Sjeff    ((score) * SCHED_PRI_RANGE / SCHED_INTERACT_MAX)
144109864Sjeff
145109864Sjeff/*
146111857Sjeff * These determine the interactivity of a process.
147109864Sjeff *
148110645Sjeff * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
149110645Sjeff *		before throttling back.
150121868Sjeff * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
151116365Sjeff * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
152111857Sjeff * INTERACT_THRESH:	Threshhold for placement on the current runq.
153109864Sjeff */
154121126Sjeff#define	SCHED_SLP_RUN_MAX	((hz * 5) << 10)
155121868Sjeff#define	SCHED_SLP_RUN_FORK	((hz / 2) << 10)
156116365Sjeff#define	SCHED_INTERACT_MAX	(100)
157116365Sjeff#define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
158121126Sjeff#define	SCHED_INTERACT_THRESH	(30)
159111857Sjeff
160109864Sjeff/*
161109864Sjeff * These parameters and macros determine the size of the time slice that is
162109864Sjeff * granted to each thread.
163109864Sjeff *
164109864Sjeff * SLICE_MIN:	Minimum time slice granted, in units of ticks.
165109864Sjeff * SLICE_MAX:	Maximum time slice granted.
166109864Sjeff * SLICE_RANGE:	Range of available time slices scaled by hz.
167112966Sjeff * SLICE_SCALE:	The number slices granted per val in the range of [0, max].
168112966Sjeff * SLICE_NICE:  Determine the amount of slice granted to a scaled nice.
169121871Sjeff * SLICE_NTHRESH:	The nice cutoff point for slice assignment.
170109864Sjeff */
171113357Sjeff#define	SCHED_SLICE_MIN			(slice_min)
172113357Sjeff#define	SCHED_SLICE_MAX			(slice_max)
173125299Sjeff#define	SCHED_SLICE_INTERACTIVE		(slice_max)
174121871Sjeff#define	SCHED_SLICE_NTHRESH	(SCHED_PRI_NHALF - 1)
175111857Sjeff#define	SCHED_SLICE_RANGE		(SCHED_SLICE_MAX - SCHED_SLICE_MIN + 1)
176109864Sjeff#define	SCHED_SLICE_SCALE(val, max)	(((val) * SCHED_SLICE_RANGE) / (max))
177112966Sjeff#define	SCHED_SLICE_NICE(nice)						\
178121871Sjeff    (SCHED_SLICE_MAX - SCHED_SLICE_SCALE((nice), SCHED_SLICE_NTHRESH))
179109864Sjeff
180109864Sjeff/*
181134791Sjulian * This macro determines whether or not the thread belongs on the current or
182109864Sjeff * next run queue.
183109864Sjeff */
184163709Sjb#define	SCHED_INTERACTIVE(td)						\
185163709Sjb    (sched_interact_score(td) < SCHED_INTERACT_THRESH)
186164936Sjulian#define	SCHED_CURR(td, ts)						\
187164936Sjulian    ((ts->ts_thread->td_flags & TDF_BORROWING) ||			\
188164936Sjulian     (ts->ts_flags & TSF_PREEMPTED) || SCHED_INTERACTIVE(td))
189109864Sjeff
190109864Sjeff/*
191109864Sjeff * Cpu percentage computation macros and defines.
192109864Sjeff *
193109864Sjeff * SCHED_CPU_TIME:	Number of seconds to average the cpu usage across.
194109864Sjeff * SCHED_CPU_TICKS:	Number of hz ticks to average the cpu usage across.
195109864Sjeff */
196109864Sjeff
197112971Sjeff#define	SCHED_CPU_TIME	10
198109864Sjeff#define	SCHED_CPU_TICKS	(hz * SCHED_CPU_TIME)
199109864Sjeff
200109864Sjeff/*
201164936Sjulian * tdq - per processor runqs and statistics.
202109864Sjeff */
203164936Sjulianstruct tdq {
204165620Sjeff	struct runq	tdq_idle;		/* Queue of IDLE threads. */
205165620Sjeff	struct runq	tdq_timeshare[2];	/* Run queues for !IDLE. */
206165620Sjeff	struct runq	*tdq_next;		/* Next timeshare queue. */
207165620Sjeff	struct runq	*tdq_curr;		/* Current queue. */
208165620Sjeff	int		tdq_load_timeshare;	/* Load for timeshare. */
209165620Sjeff	int		tdq_load;		/* Aggregate load. */
210165620Sjeff	short		tdq_nice[SCHED_PRI_NRESV]; /* threadss in each nice bin. */
211165620Sjeff	short		tdq_nicemin;		/* Least nice. */
212110267Sjeff#ifdef SMP
213165620Sjeff	int		tdq_transferable;
214165620Sjeff	LIST_ENTRY(tdq)	tdq_siblings;		/* Next in tdq group. */
215165620Sjeff	struct tdq_group *tdq_group;		/* Our processor group. */
216165620Sjeff	volatile struct td_sched *tdq_assigned;	/* assigned by another CPU. */
217125289Sjeff#else
218165620Sjeff	int		tdq_sysload;		/* For loadavg, !ITHD load. */
219110267Sjeff#endif
220109864Sjeff};
221109864Sjeff
222123433Sjeff#ifdef SMP
223109864Sjeff/*
224164936Sjulian * tdq groups are groups of processors which can cheaply share threads.  When
225123433Sjeff * one processor in the group goes idle it will check the runqs of the other
226123433Sjeff * processors in its group prior to halting and waiting for an interrupt.
227123433Sjeff * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA.
228123433Sjeff * In a numa environment we'd want an idle bitmap per group and a two tiered
229123433Sjeff * load balancer.
230123433Sjeff */
231164936Sjulianstruct tdq_group {
232165620Sjeff	int	tdg_cpus;		/* Count of CPUs in this tdq group. */
233165620Sjeff	cpumask_t tdg_cpumask;		/* Mask of cpus in this group. */
234165620Sjeff	cpumask_t tdg_idlemask;		/* Idle cpus in this group. */
235165620Sjeff	cpumask_t tdg_mask;		/* Bit mask for first cpu. */
236165620Sjeff	int	tdg_load;		/* Total load of this group. */
237165620Sjeff	int	tdg_transferable;	/* Transferable load of this group. */
238165620Sjeff	LIST_HEAD(, tdq) tdg_members;	/* Linked list of all members. */
239123433Sjeff};
240123433Sjeff#endif
241123433Sjeff
242123433Sjeff/*
243165620Sjeff * One thread queue per processor.
244109864Sjeff */
245110028Sjeff#ifdef SMP
246164936Sjulianstatic cpumask_t tdq_idle;
247165620Sjeffstatic int tdg_maxid;
248164936Sjulianstatic struct tdq	tdq_cpu[MAXCPU];
249164936Sjulianstatic struct tdq_group tdq_groups[MAXCPU];
250129982Sjeffstatic int bal_tick;
251129982Sjeffstatic int gbal_tick;
252139334Sjeffstatic int balance_groups;
253129982Sjeff
254164936Sjulian#define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
255164936Sjulian#define	TDQ_CPU(x)	(&tdq_cpu[(x)])
256164936Sjulian#define	TDQ_ID(x)	((x) - tdq_cpu)
257164936Sjulian#define	TDQ_GROUP(x)	(&tdq_groups[(x)])
258123433Sjeff#else	/* !SMP */
259164936Sjulianstatic struct tdq	tdq_cpu;
260129982Sjeff
261164936Sjulian#define	TDQ_SELF()	(&tdq_cpu)
262164936Sjulian#define	TDQ_CPU(x)	(&tdq_cpu)
263110028Sjeff#endif
264109864Sjeff
265164936Sjulianstatic struct td_sched *sched_choose(void);		/* XXX Should be thread * */
266164936Sjulianstatic void sched_slice(struct td_sched *);
267163709Sjbstatic void sched_priority(struct thread *);
268146954Sjeffstatic void sched_thread_priority(struct thread *, u_char);
269163709Sjbstatic int sched_interact_score(struct thread *);
270163709Sjbstatic void sched_interact_update(struct thread *);
271163709Sjbstatic void sched_interact_fork(struct thread *);
272164936Sjulianstatic void sched_pctcpu_update(struct td_sched *);
273109864Sjeff
274110267Sjeff/* Operations on per processor queues */
275164936Sjulianstatic struct td_sched * tdq_choose(struct tdq *);
276164936Sjulianstatic void tdq_setup(struct tdq *);
277164936Sjulianstatic void tdq_load_add(struct tdq *, struct td_sched *);
278164936Sjulianstatic void tdq_load_rem(struct tdq *, struct td_sched *);
279164936Sjulianstatic __inline void tdq_runq_add(struct tdq *, struct td_sched *, int);
280164936Sjulianstatic __inline void tdq_runq_rem(struct tdq *, struct td_sched *);
281164936Sjulianstatic void tdq_nice_add(struct tdq *, int);
282164936Sjulianstatic void tdq_nice_rem(struct tdq *, int);
283164936Sjulianvoid tdq_print(int cpu);
284110267Sjeff#ifdef SMP
285164936Sjulianstatic int tdq_transfer(struct tdq *, struct td_sched *, int);
286164936Sjulianstatic struct td_sched *runq_steal(struct runq *);
287129982Sjeffstatic void sched_balance(void);
288129982Sjeffstatic void sched_balance_groups(void);
289164936Sjulianstatic void sched_balance_group(struct tdq_group *);
290164936Sjulianstatic void sched_balance_pair(struct tdq *, struct tdq *);
291164936Sjulianstatic void tdq_move(struct tdq *, int);
292164936Sjulianstatic int tdq_idled(struct tdq *);
293164936Sjulianstatic void tdq_notify(struct td_sched *, int);
294164936Sjulianstatic void tdq_assign(struct tdq *);
295164936Sjulianstatic struct td_sched *tdq_steal(struct tdq *, int);
296164936Sjulian#define	THREAD_CAN_MIGRATE(ts)						\
297164936Sjulian    ((ts)->ts_thread->td_pinned == 0 && ((ts)->ts_flags & TSF_BOUND) == 0)
298121790Sjeff#endif
299110028Sjeff
300113357Sjeffvoid
301164936Sjuliantdq_print(int cpu)
302110267Sjeff{
303164936Sjulian	struct tdq *tdq;
304113357Sjeff	int i;
305112994Sjeff
306164936Sjulian	tdq = TDQ_CPU(cpu);
307112994Sjeff
308164936Sjulian	printf("tdq:\n");
309165620Sjeff	printf("\tload:           %d\n", tdq->tdq_load);
310165620Sjeff	printf("\tload TIMESHARE: %d\n", tdq->tdq_load_timeshare);
311121896Sjeff#ifdef SMP
312165620Sjeff	printf("\tload transferable: %d\n", tdq->tdq_transferable);
313121896Sjeff#endif
314165620Sjeff	printf("\tnicemin:\t%d\n", tdq->tdq_nicemin);
315113357Sjeff	printf("\tnice counts:\n");
316121869Sjeff	for (i = 0; i < SCHED_PRI_NRESV; i++)
317165620Sjeff		if (tdq->tdq_nice[i])
318113357Sjeff			printf("\t\t%d = %d\n",
319165620Sjeff			    i - SCHED_PRI_NHALF, tdq->tdq_nice[i]);
320113357Sjeff}
321112994Sjeff
322122744Sjeffstatic __inline void
323164936Sjuliantdq_runq_add(struct tdq *tdq, struct td_sched *ts, int flags)
324122744Sjeff{
325122744Sjeff#ifdef SMP
326164936Sjulian	if (THREAD_CAN_MIGRATE(ts)) {
327165620Sjeff		tdq->tdq_transferable++;
328165620Sjeff		tdq->tdq_group->tdg_transferable++;
329164936Sjulian		ts->ts_flags |= TSF_XFERABLE;
330123433Sjeff	}
331122744Sjeff#endif
332164936Sjulian	if (ts->ts_flags & TSF_PREEMPTED)
333148856Sdavidxu		flags |= SRQ_PREEMPTED;
334164936Sjulian	runq_add(ts->ts_runq, ts, flags);
335122744Sjeff}
336122744Sjeff
337122744Sjeffstatic __inline void
338164936Sjuliantdq_runq_rem(struct tdq *tdq, struct td_sched *ts)
339122744Sjeff{
340122744Sjeff#ifdef SMP
341164936Sjulian	if (ts->ts_flags & TSF_XFERABLE) {
342165620Sjeff		tdq->tdq_transferable--;
343165620Sjeff		tdq->tdq_group->tdg_transferable--;
344164936Sjulian		ts->ts_flags &= ~TSF_XFERABLE;
345123433Sjeff	}
346122744Sjeff#endif
347164936Sjulian	runq_remove(ts->ts_runq, ts);
348122744Sjeff}
349122744Sjeff
350113357Sjeffstatic void
351164936Sjuliantdq_load_add(struct tdq *tdq, struct td_sched *ts)
352113357Sjeff{
353121896Sjeff	int class;
354115998Sjeff	mtx_assert(&sched_lock, MA_OWNED);
355164936Sjulian	class = PRI_BASE(ts->ts_thread->td_pri_class);
356121896Sjeff	if (class == PRI_TIMESHARE)
357165620Sjeff		tdq->tdq_load_timeshare++;
358165620Sjeff	tdq->tdq_load++;
359165620Sjeff	CTR1(KTR_SCHED, "load: %d", tdq->tdq_load);
360164936Sjulian	if (class != PRI_ITHD && (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0)
361123487Sjeff#ifdef SMP
362165620Sjeff		tdq->tdq_group->tdg_load++;
363125289Sjeff#else
364165620Sjeff		tdq->tdq_sysload++;
365123487Sjeff#endif
366164936Sjulian	if (ts->ts_thread->td_pri_class == PRI_TIMESHARE)
367164936Sjulian		tdq_nice_add(tdq, ts->ts_thread->td_proc->p_nice);
368110267Sjeff}
369113357Sjeff
370112994Sjeffstatic void
371164936Sjuliantdq_load_rem(struct tdq *tdq, struct td_sched *ts)
372110267Sjeff{
373121896Sjeff	int class;
374115998Sjeff	mtx_assert(&sched_lock, MA_OWNED);
375164936Sjulian	class = PRI_BASE(ts->ts_thread->td_pri_class);
376121896Sjeff	if (class == PRI_TIMESHARE)
377165620Sjeff		tdq->tdq_load_timeshare--;
378164936Sjulian	if (class != PRI_ITHD  && (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0)
379123487Sjeff#ifdef SMP
380165620Sjeff		tdq->tdq_group->tdg_load--;
381125289Sjeff#else
382165620Sjeff		tdq->tdq_sysload--;
383123487Sjeff#endif
384165620Sjeff	tdq->tdq_load--;
385165620Sjeff	CTR1(KTR_SCHED, "load: %d", tdq->tdq_load);
386164936Sjulian	ts->ts_runq = NULL;
387164936Sjulian	if (ts->ts_thread->td_pri_class == PRI_TIMESHARE)
388164936Sjulian		tdq_nice_rem(tdq, ts->ts_thread->td_proc->p_nice);
389110267Sjeff}
390110267Sjeff
391113357Sjeffstatic void
392164936Sjuliantdq_nice_add(struct tdq *tdq, int nice)
393110267Sjeff{
394115998Sjeff	mtx_assert(&sched_lock, MA_OWNED);
395113357Sjeff	/* Normalize to zero. */
396165620Sjeff	tdq->tdq_nice[nice + SCHED_PRI_NHALF]++;
397165620Sjeff	if (nice < tdq->tdq_nicemin || tdq->tdq_load_timeshare == 1)
398165620Sjeff		tdq->tdq_nicemin = nice;
399110267Sjeff}
400110267Sjeff
401113357Sjeffstatic void
402164936Sjuliantdq_nice_rem(struct tdq *tdq, int nice)
403110267Sjeff{
404113357Sjeff	int n;
405113357Sjeff
406115998Sjeff	mtx_assert(&sched_lock, MA_OWNED);
407113357Sjeff	/* Normalize to zero. */
408113357Sjeff	n = nice + SCHED_PRI_NHALF;
409165620Sjeff	tdq->tdq_nice[n]--;
410165620Sjeff	KASSERT(tdq->tdq_nice[n] >= 0, ("Negative nice count."));
411113357Sjeff
412113357Sjeff	/*
413113357Sjeff	 * If this wasn't the smallest nice value or there are more in
414113357Sjeff	 * this bucket we can just return.  Otherwise we have to recalculate
415113357Sjeff	 * the smallest nice.
416113357Sjeff	 */
417165620Sjeff	if (nice != tdq->tdq_nicemin ||
418165620Sjeff	    tdq->tdq_nice[n] != 0 ||
419165620Sjeff	    tdq->tdq_load_timeshare == 0)
420113357Sjeff		return;
421113357Sjeff
422121869Sjeff	for (; n < SCHED_PRI_NRESV; n++)
423165620Sjeff		if (tdq->tdq_nice[n]) {
424165620Sjeff			tdq->tdq_nicemin = n - SCHED_PRI_NHALF;
425113357Sjeff			return;
426113357Sjeff		}
427110267Sjeff}
428110267Sjeff
429113357Sjeff#ifdef SMP
430116069Sjeff/*
431122744Sjeff * sched_balance is a simple CPU load balancing algorithm.  It operates by
432116069Sjeff * finding the least loaded and most loaded cpu and equalizing their load
433116069Sjeff * by migrating some processes.
434116069Sjeff *
435116069Sjeff * Dealing only with two CPUs at a time has two advantages.  Firstly, most
436116069Sjeff * installations will only have 2 cpus.  Secondly, load balancing too much at
437116069Sjeff * once can have an unpleasant effect on the system.  The scheduler rarely has
438116069Sjeff * enough information to make perfect decisions.  So this algorithm chooses
439116069Sjeff * algorithm simplicity and more gradual effects on load in larger systems.
440116069Sjeff *
441116069Sjeff * It could be improved by considering the priorities and slices assigned to
442116069Sjeff * each task prior to balancing them.  There are many pathological cases with
443116069Sjeff * any approach and so the semi random algorithm below may work as well as any.
444116069Sjeff *
445116069Sjeff */
446121790Sjeffstatic void
447129982Sjeffsched_balance(void)
448116069Sjeff{
449164936Sjulian	struct tdq_group *high;
450164936Sjulian	struct tdq_group *low;
451165620Sjeff	struct tdq_group *tdg;
452123487Sjeff	int cnt;
453123487Sjeff	int i;
454123487Sjeff
455139334Sjeff	bal_tick = ticks + (random() % (hz * 2));
456123487Sjeff	if (smp_started == 0)
457139334Sjeff		return;
458123487Sjeff	low = high = NULL;
459165620Sjeff	i = random() % (tdg_maxid + 1);
460165620Sjeff	for (cnt = 0; cnt <= tdg_maxid; cnt++) {
461165620Sjeff		tdg = TDQ_GROUP(i);
462123487Sjeff		/*
463123487Sjeff		 * Find the CPU with the highest load that has some
464123487Sjeff		 * threads to transfer.
465123487Sjeff		 */
466165620Sjeff		if ((high == NULL || tdg->tdg_load > high->tdg_load)
467165620Sjeff		    && tdg->tdg_transferable)
468165620Sjeff			high = tdg;
469165620Sjeff		if (low == NULL || tdg->tdg_load < low->tdg_load)
470165620Sjeff			low = tdg;
471165620Sjeff		if (++i > tdg_maxid)
472123487Sjeff			i = 0;
473123487Sjeff	}
474123487Sjeff	if (low != NULL && high != NULL && high != low)
475165620Sjeff		sched_balance_pair(LIST_FIRST(&high->tdg_members),
476165620Sjeff		    LIST_FIRST(&low->tdg_members));
477123487Sjeff}
478123487Sjeff
479123487Sjeffstatic void
480129982Sjeffsched_balance_groups(void)
481123487Sjeff{
482123487Sjeff	int i;
483123487Sjeff
484139334Sjeff	gbal_tick = ticks + (random() % (hz * 2));
485129982Sjeff	mtx_assert(&sched_lock, MA_OWNED);
486123487Sjeff	if (smp_started)
487165620Sjeff		for (i = 0; i <= tdg_maxid; i++)
488164936Sjulian			sched_balance_group(TDQ_GROUP(i));
489123487Sjeff}
490123487Sjeff
491123487Sjeffstatic void
492165620Sjeffsched_balance_group(struct tdq_group *tdg)
493123487Sjeff{
494164936Sjulian	struct tdq *tdq;
495164936Sjulian	struct tdq *high;
496164936Sjulian	struct tdq *low;
497123487Sjeff	int load;
498123487Sjeff
499165620Sjeff	if (tdg->tdg_transferable == 0)
500123487Sjeff		return;
501123487Sjeff	low = NULL;
502123487Sjeff	high = NULL;
503165620Sjeff	LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) {
504165620Sjeff		load = tdq->tdq_load;
505165620Sjeff		if (high == NULL || load > high->tdq_load)
506164936Sjulian			high = tdq;
507165620Sjeff		if (low == NULL || load < low->tdq_load)
508164936Sjulian			low = tdq;
509123487Sjeff	}
510123487Sjeff	if (high != NULL && low != NULL && high != low)
511123487Sjeff		sched_balance_pair(high, low);
512123487Sjeff}
513123487Sjeff
514123487Sjeffstatic void
515164936Sjuliansched_balance_pair(struct tdq *high, struct tdq *low)
516123487Sjeff{
517123433Sjeff	int transferable;
518116069Sjeff	int high_load;
519116069Sjeff	int low_load;
520116069Sjeff	int move;
521116069Sjeff	int diff;
522116069Sjeff	int i;
523116069Sjeff
524116069Sjeff	/*
525123433Sjeff	 * If we're transfering within a group we have to use this specific
526164936Sjulian	 * tdq's transferable count, otherwise we can steal from other members
527123433Sjeff	 * of the group.
528123433Sjeff	 */
529165620Sjeff	if (high->tdq_group == low->tdq_group) {
530165620Sjeff		transferable = high->tdq_transferable;
531165620Sjeff		high_load = high->tdq_load;
532165620Sjeff		low_load = low->tdq_load;
533123487Sjeff	} else {
534165620Sjeff		transferable = high->tdq_group->tdg_transferable;
535165620Sjeff		high_load = high->tdq_group->tdg_load;
536165620Sjeff		low_load = low->tdq_group->tdg_load;
537123487Sjeff	}
538123433Sjeff	if (transferable == 0)
539123487Sjeff		return;
540123433Sjeff	/*
541122744Sjeff	 * Determine what the imbalance is and then adjust that to how many
542165620Sjeff	 * threads we actually have to give up (transferable).
543122744Sjeff	 */
544123487Sjeff	diff = high_load - low_load;
545116069Sjeff	move = diff / 2;
546116069Sjeff	if (diff & 0x1)
547116069Sjeff		move++;
548123433Sjeff	move = min(move, transferable);
549116069Sjeff	for (i = 0; i < move; i++)
550164936Sjulian		tdq_move(high, TDQ_ID(low));
551116069Sjeff	return;
552116069Sjeff}
553116069Sjeff
554121790Sjeffstatic void
555164936Sjuliantdq_move(struct tdq *from, int cpu)
556116069Sjeff{
557164936Sjulian	struct tdq *tdq;
558164936Sjulian	struct tdq *to;
559164936Sjulian	struct td_sched *ts;
560116069Sjeff
561164936Sjulian	tdq = from;
562164936Sjulian	to = TDQ_CPU(cpu);
563164936Sjulian	ts = tdq_steal(tdq, 1);
564164936Sjulian	if (ts == NULL) {
565165620Sjeff		struct tdq_group *tdg;
566123433Sjeff
567165620Sjeff		tdg = tdq->tdq_group;
568165620Sjeff		LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) {
569165620Sjeff			if (tdq == from || tdq->tdq_transferable == 0)
570123433Sjeff				continue;
571164936Sjulian			ts = tdq_steal(tdq, 1);
572123433Sjeff			break;
573123433Sjeff		}
574164936Sjulian		if (ts == NULL)
575164936Sjulian			panic("tdq_move: No threads available with a "
576123433Sjeff			    "transferable count of %d\n",
577165620Sjeff			    tdg->tdg_transferable);
578123433Sjeff	}
579164936Sjulian	if (tdq == to)
580123433Sjeff		return;
581164936Sjulian	ts->ts_state = TSS_THREAD;
582164936Sjulian	tdq_runq_rem(tdq, ts);
583164936Sjulian	tdq_load_rem(tdq, ts);
584164936Sjulian	tdq_notify(ts, cpu);
585116069Sjeff}
586110267Sjeff
587123433Sjeffstatic int
588164936Sjuliantdq_idled(struct tdq *tdq)
589121790Sjeff{
590165620Sjeff	struct tdq_group *tdg;
591164936Sjulian	struct tdq *steal;
592164936Sjulian	struct td_sched *ts;
593123433Sjeff
594165620Sjeff	tdg = tdq->tdq_group;
595123433Sjeff	/*
596165620Sjeff	 * If we're in a cpu group, try and steal threads from another cpu in
597123433Sjeff	 * the group before idling.
598123433Sjeff	 */
599165620Sjeff	if (tdg->tdg_cpus > 1 && tdg->tdg_transferable) {
600165620Sjeff		LIST_FOREACH(steal, &tdg->tdg_members, tdq_siblings) {
601165620Sjeff			if (steal == tdq || steal->tdq_transferable == 0)
602123433Sjeff				continue;
603164936Sjulian			ts = tdq_steal(steal, 0);
604164936Sjulian			if (ts == NULL)
605123433Sjeff				continue;
606164936Sjulian			ts->ts_state = TSS_THREAD;
607164936Sjulian			tdq_runq_rem(steal, ts);
608164936Sjulian			tdq_load_rem(steal, ts);
609164936Sjulian			ts->ts_cpu = PCPU_GET(cpuid);
610164936Sjulian			ts->ts_flags |= TSF_INTERNAL | TSF_HOLD;
611164936Sjulian			sched_add(ts->ts_thread, SRQ_YIELDING);
612123433Sjeff			return (0);
613123433Sjeff		}
614123433Sjeff	}
615123433Sjeff	/*
616123433Sjeff	 * We only set the idled bit when all of the cpus in the group are
617164936Sjulian	 * idle.  Otherwise we could get into a situation where a thread bounces
618123433Sjeff	 * back and forth between two idle cores on seperate physical CPUs.
619123433Sjeff	 */
620165620Sjeff	tdg->tdg_idlemask |= PCPU_GET(cpumask);
621165620Sjeff	if (tdg->tdg_idlemask != tdg->tdg_cpumask)
622123433Sjeff		return (1);
623165620Sjeff	atomic_set_int(&tdq_idle, tdg->tdg_mask);
624123433Sjeff	return (1);
625121790Sjeff}
626121790Sjeff
627121790Sjeffstatic void
628164936Sjuliantdq_assign(struct tdq *tdq)
629121790Sjeff{
630164936Sjulian	struct td_sched *nts;
631164936Sjulian	struct td_sched *ts;
632121790Sjeff
633121790Sjeff	do {
634165620Sjeff		*(volatile struct td_sched **)&ts = tdq->tdq_assigned;
635165620Sjeff	} while(!atomic_cmpset_ptr((volatile uintptr_t *)&tdq->tdq_assigned,
636164936Sjulian		(uintptr_t)ts, (uintptr_t)NULL));
637164936Sjulian	for (; ts != NULL; ts = nts) {
638164936Sjulian		nts = ts->ts_assign;
639165620Sjeff		tdq->tdq_group->tdg_load--;
640165620Sjeff		tdq->tdq_load--;
641164936Sjulian		ts->ts_flags &= ~TSF_ASSIGNED;
642164936Sjulian		if (ts->ts_flags & TSF_REMOVED) {
643164936Sjulian			ts->ts_flags &= ~TSF_REMOVED;
644148603Sdavidxu			continue;
645148603Sdavidxu		}
646164936Sjulian		ts->ts_flags |= TSF_INTERNAL | TSF_HOLD;
647164936Sjulian		sched_add(ts->ts_thread, SRQ_YIELDING);
648121790Sjeff	}
649121790Sjeff}
650121790Sjeff
651121790Sjeffstatic void
652164936Sjuliantdq_notify(struct td_sched *ts, int cpu)
653121790Sjeff{
654164936Sjulian	struct tdq *tdq;
655121790Sjeff	struct thread *td;
656121790Sjeff	struct pcpu *pcpu;
657139334Sjeff	int class;
658133427Sjeff	int prio;
659121790Sjeff
660164936Sjulian	tdq = TDQ_CPU(cpu);
661139334Sjeff	/* XXX */
662164936Sjulian	class = PRI_BASE(ts->ts_thread->td_pri_class);
663139334Sjeff	if ((class == PRI_TIMESHARE || class == PRI_REALTIME) &&
664165620Sjeff	    (tdq_idle & tdq->tdq_group->tdg_mask))
665165620Sjeff		atomic_clear_int(&tdq_idle, tdq->tdq_group->tdg_mask);
666165620Sjeff	tdq->tdq_group->tdg_load++;
667165620Sjeff	tdq->tdq_load++;
668164936Sjulian	ts->ts_cpu = cpu;
669164936Sjulian	ts->ts_flags |= TSF_ASSIGNED;
670164936Sjulian	prio = ts->ts_thread->td_priority;
671121790Sjeff
672121790Sjeff	/*
673164936Sjulian	 * Place a thread on another cpu's queue and force a resched.
674121790Sjeff	 */
675121790Sjeff	do {
676165620Sjeff		*(volatile struct td_sched **)&ts->ts_assign = tdq->tdq_assigned;
677165620Sjeff	} while(!atomic_cmpset_ptr((volatile uintptr_t *)&tdq->tdq_assigned,
678164936Sjulian		(uintptr_t)ts->ts_assign, (uintptr_t)ts));
679133427Sjeff	/*
680133427Sjeff	 * Without sched_lock we could lose a race where we set NEEDRESCHED
681133427Sjeff	 * on a thread that is switched out before the IPI is delivered.  This
682133427Sjeff	 * would lead us to miss the resched.  This will be a problem once
683133427Sjeff	 * sched_lock is pushed down.
684133427Sjeff	 */
685121790Sjeff	pcpu = pcpu_find(cpu);
686121790Sjeff	td = pcpu->pc_curthread;
687164936Sjulian	if (ts->ts_thread->td_priority < td->td_priority ||
688121790Sjeff	    td == pcpu->pc_idlethread) {
689121790Sjeff		td->td_flags |= TDF_NEEDRESCHED;
690121790Sjeff		ipi_selected(1 << cpu, IPI_AST);
691121790Sjeff	}
692121790Sjeff}
693121790Sjeff
694164936Sjulianstatic struct td_sched *
695121790Sjeffrunq_steal(struct runq *rq)
696121790Sjeff{
697121790Sjeff	struct rqhead *rqh;
698121790Sjeff	struct rqbits *rqb;
699164936Sjulian	struct td_sched *ts;
700121790Sjeff	int word;
701121790Sjeff	int bit;
702121790Sjeff
703121790Sjeff	mtx_assert(&sched_lock, MA_OWNED);
704121790Sjeff	rqb = &rq->rq_status;
705121790Sjeff	for (word = 0; word < RQB_LEN; word++) {
706121790Sjeff		if (rqb->rqb_bits[word] == 0)
707121790Sjeff			continue;
708121790Sjeff		for (bit = 0; bit < RQB_BPW; bit++) {
709123231Speter			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
710121790Sjeff				continue;
711121790Sjeff			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
712164936Sjulian			TAILQ_FOREACH(ts, rqh, ts_procq) {
713164936Sjulian				if (THREAD_CAN_MIGRATE(ts))
714164936Sjulian					return (ts);
715121790Sjeff			}
716121790Sjeff		}
717121790Sjeff	}
718121790Sjeff	return (NULL);
719121790Sjeff}
720121790Sjeff
721164936Sjulianstatic struct td_sched *
722164936Sjuliantdq_steal(struct tdq *tdq, int stealidle)
723121790Sjeff{
724164936Sjulian	struct td_sched *ts;
725121790Sjeff
726123433Sjeff	/*
727123433Sjeff	 * Steal from next first to try to get a non-interactive task that
728123433Sjeff	 * may not have run for a while.
729123433Sjeff	 */
730165620Sjeff	if ((ts = runq_steal(tdq->tdq_next)) != NULL)
731164936Sjulian		return (ts);
732165620Sjeff	if ((ts = runq_steal(tdq->tdq_curr)) != NULL)
733164936Sjulian		return (ts);
734123433Sjeff	if (stealidle)
735165620Sjeff		return (runq_steal(&tdq->tdq_idle));
736123433Sjeff	return (NULL);
737121790Sjeff}
738123433Sjeff
739123433Sjeffint
740164936Sjuliantdq_transfer(struct tdq *tdq, struct td_sched *ts, int class)
741123433Sjeff{
742165620Sjeff	struct tdq_group *ntdg;
743165620Sjeff	struct tdq_group *tdg;
744164936Sjulian	struct tdq *old;
745123433Sjeff	int cpu;
746139334Sjeff	int idx;
747123433Sjeff
748123685Sjeff	if (smp_started == 0)
749123685Sjeff		return (0);
750123433Sjeff	cpu = 0;
751123433Sjeff	/*
752133427Sjeff	 * If our load exceeds a certain threshold we should attempt to
753133427Sjeff	 * reassign this thread.  The first candidate is the cpu that
754133427Sjeff	 * originally ran the thread.  If it is idle, assign it there,
755133427Sjeff	 * otherwise, pick an idle cpu.
756133427Sjeff	 *
757165620Sjeff	 * The threshold at which we start to reassign has a large impact
758123685Sjeff	 * on the overall performance of the system.  Tuned too high and
759123685Sjeff	 * some CPUs may idle.  Too low and there will be excess migration
760128055Scognet	 * and context switches.
761123685Sjeff	 */
762164936Sjulian	old = TDQ_CPU(ts->ts_cpu);
763165620Sjeff	ntdg = old->tdq_group;
764165620Sjeff	tdg = tdq->tdq_group;
765164936Sjulian	if (tdq_idle) {
766165620Sjeff		if (tdq_idle & ntdg->tdg_mask) {
767165620Sjeff			cpu = ffs(ntdg->tdg_idlemask);
768139334Sjeff			if (cpu) {
769139334Sjeff				CTR2(KTR_SCHED,
770164936Sjulian				    "tdq_transfer: %p found old cpu %X "
771164936Sjulian				    "in idlemask.", ts, cpu);
772133427Sjeff				goto migrate;
773139334Sjeff			}
774133427Sjeff		}
775123433Sjeff		/*
776123433Sjeff		 * Multiple cpus could find this bit simultaneously
777123433Sjeff		 * but the race shouldn't be terrible.
778123433Sjeff		 */
779164936Sjulian		cpu = ffs(tdq_idle);
780139334Sjeff		if (cpu) {
781164936Sjulian			CTR2(KTR_SCHED, "tdq_transfer: %p found %X "
782164936Sjulian			    "in idlemask.", ts, cpu);
783133427Sjeff			goto migrate;
784139334Sjeff		}
785123433Sjeff	}
786139334Sjeff	idx = 0;
787139334Sjeff#if 0
788165620Sjeff	if (old->tdq_load < tdq->tdq_load) {
789164936Sjulian		cpu = ts->ts_cpu + 1;
790164936Sjulian		CTR2(KTR_SCHED, "tdq_transfer: %p old cpu %X "
791164936Sjulian		    "load less than ours.", ts, cpu);
792139334Sjeff		goto migrate;
793139334Sjeff	}
794123433Sjeff	/*
795139334Sjeff	 * No new CPU was found, look for one with less load.
796139334Sjeff	 */
797165620Sjeff	for (idx = 0; idx <= tdg_maxid; idx++) {
798165620Sjeff		ntdg = TDQ_GROUP(idx);
799165620Sjeff		if (ntdg->tdg_load /*+ (ntdg->tdg_cpus  * 2)*/ < tdg->tdg_load) {
800165620Sjeff			cpu = ffs(ntdg->tdg_cpumask);
801164936Sjulian			CTR2(KTR_SCHED, "tdq_transfer: %p cpu %X load less "
802164936Sjulian			    "than ours.", ts, cpu);
803139334Sjeff			goto migrate;
804139334Sjeff		}
805139334Sjeff	}
806139334Sjeff#endif
807139334Sjeff	/*
808123433Sjeff	 * If another cpu in this group has idled, assign a thread over
809123433Sjeff	 * to them after checking to see if there are idled groups.
810123433Sjeff	 */
811165620Sjeff	if (tdg->tdg_idlemask) {
812165620Sjeff		cpu = ffs(tdg->tdg_idlemask);
813139334Sjeff		if (cpu) {
814164936Sjulian			CTR2(KTR_SCHED, "tdq_transfer: %p cpu %X idle in "
815164936Sjulian			    "group.", ts, cpu);
816133427Sjeff			goto migrate;
817139334Sjeff		}
818123433Sjeff	}
819133427Sjeff	return (0);
820133427Sjeffmigrate:
821133427Sjeff	/*
822123433Sjeff	 * Now that we've found an idle CPU, migrate the thread.
823123433Sjeff	 */
824133427Sjeff	cpu--;
825164936Sjulian	ts->ts_runq = NULL;
826164936Sjulian	tdq_notify(ts, cpu);
827133427Sjeff
828133427Sjeff	return (1);
829123433Sjeff}
830123433Sjeff
831121790Sjeff#endif	/* SMP */
832121790Sjeff
833117326Sjeff/*
834121790Sjeff * Pick the highest priority task we have and return it.
835117326Sjeff */
836117326Sjeff
837164936Sjulianstatic struct td_sched *
838164936Sjuliantdq_choose(struct tdq *tdq)
839110267Sjeff{
840137067Sjeff	struct runq *swap;
841164936Sjulian	struct td_sched *ts;
842137067Sjeff	int nice;
843110267Sjeff
844115998Sjeff	mtx_assert(&sched_lock, MA_OWNED);
845113357Sjeff	swap = NULL;
846112994Sjeff
847113357Sjeff	for (;;) {
848165620Sjeff		ts = runq_choose(tdq->tdq_curr);
849164936Sjulian		if (ts == NULL) {
850113357Sjeff			/*
851131473Sjhb			 * We already swapped once and didn't get anywhere.
852113357Sjeff			 */
853113357Sjeff			if (swap)
854113357Sjeff				break;
855165620Sjeff			swap = tdq->tdq_curr;
856165620Sjeff			tdq->tdq_curr = tdq->tdq_next;
857165620Sjeff			tdq->tdq_next = swap;
858113357Sjeff			continue;
859113357Sjeff		}
860113357Sjeff		/*
861164936Sjulian		 * If we encounter a slice of 0 the td_sched is in a
862164936Sjulian		 * TIMESHARE td_sched group and its nice was too far out
863113357Sjeff		 * of the range that receives slices.
864113357Sjeff		 */
865165620Sjeff		nice = ts->ts_thread->td_proc->p_nice + (0 - tdq->tdq_nicemin);
866150442Sdavidxu#if 0
867164936Sjulian		if (ts->ts_slice == 0 || (nice > SCHED_SLICE_NTHRESH &&
868164936Sjulian		    ts->ts_thread->td_proc->p_nice != 0)) {
869164936Sjulian			runq_remove(ts->ts_runq, ts);
870164936Sjulian			sched_slice(ts);
871165620Sjeff			ts->ts_runq = tdq->tdq_next;
872164936Sjulian			runq_add(ts->ts_runq, ts, 0);
873113357Sjeff			continue;
874113357Sjeff		}
875150442Sdavidxu#endif
876164936Sjulian		return (ts);
877110267Sjeff	}
878110267Sjeff
879165620Sjeff	return (runq_choose(&tdq->tdq_idle));
880110267Sjeff}
881110267Sjeff
882109864Sjeffstatic void
883164936Sjuliantdq_setup(struct tdq *tdq)
884110028Sjeff{
885165620Sjeff	runq_init(&tdq->tdq_timeshare[0]);
886165620Sjeff	runq_init(&tdq->tdq_timeshare[1]);
887165620Sjeff	runq_init(&tdq->tdq_idle);
888165620Sjeff	tdq->tdq_curr = &tdq->tdq_timeshare[0];
889165620Sjeff	tdq->tdq_next = &tdq->tdq_timeshare[1];
890165620Sjeff	tdq->tdq_load = 0;
891165620Sjeff	tdq->tdq_load_timeshare = 0;
892110028Sjeff}
893110028Sjeff
894110028Sjeffstatic void
895109864Sjeffsched_setup(void *dummy)
896109864Sjeff{
897117313Sjeff#ifdef SMP
898109864Sjeff	int i;
899117313Sjeff#endif
900109864Sjeff
901153533Sdavidxu	/*
902153533Sdavidxu	 * To avoid divide-by-zero, we set realstathz a dummy value
903153533Sdavidxu	 * in case which sched_clock() called before sched_initticks().
904153533Sdavidxu	 */
905153533Sdavidxu	realstathz = hz;
906116946Sjeff	slice_min = (hz/100);	/* 10ms */
907116946Sjeff	slice_max = (hz/7);	/* ~140ms */
908111857Sjeff
909117237Sjeff#ifdef SMP
910123487Sjeff	balance_groups = 0;
911123433Sjeff	/*
912164936Sjulian	 * Initialize the tdqs.
913123433Sjeff	 */
914123433Sjeff	for (i = 0; i < MAXCPU; i++) {
915165627Sjeff		struct tdq *tdq;
916123433Sjeff
917165627Sjeff		tdq = &tdq_cpu[i];
918165627Sjeff		tdq->tdq_assigned = NULL;
919164936Sjulian		tdq_setup(&tdq_cpu[i]);
920123433Sjeff	}
921117237Sjeff	if (smp_topology == NULL) {
922165620Sjeff		struct tdq_group *tdg;
923165627Sjeff		struct tdq *tdq;
924139334Sjeff		int cpus;
925123433Sjeff
926139334Sjeff		for (cpus = 0, i = 0; i < MAXCPU; i++) {
927139334Sjeff			if (CPU_ABSENT(i))
928139334Sjeff				continue;
929165627Sjeff			tdq = &tdq_cpu[i];
930165620Sjeff			tdg = &tdq_groups[cpus];
931123433Sjeff			/*
932164936Sjulian			 * Setup a tdq group with one member.
933123433Sjeff			 */
934165627Sjeff			tdq->tdq_transferable = 0;
935165627Sjeff			tdq->tdq_group = tdg;
936165620Sjeff			tdg->tdg_cpus = 1;
937165620Sjeff			tdg->tdg_idlemask = 0;
938165620Sjeff			tdg->tdg_cpumask = tdg->tdg_mask = 1 << i;
939165620Sjeff			tdg->tdg_load = 0;
940165620Sjeff			tdg->tdg_transferable = 0;
941165620Sjeff			LIST_INIT(&tdg->tdg_members);
942165627Sjeff			LIST_INSERT_HEAD(&tdg->tdg_members, tdq, tdq_siblings);
943139334Sjeff			cpus++;
944117237Sjeff		}
945165620Sjeff		tdg_maxid = cpus - 1;
946117237Sjeff	} else {
947165620Sjeff		struct tdq_group *tdg;
948123433Sjeff		struct cpu_group *cg;
949117237Sjeff		int j;
950113357Sjeff
951117237Sjeff		for (i = 0; i < smp_topology->ct_count; i++) {
952117237Sjeff			cg = &smp_topology->ct_group[i];
953165620Sjeff			tdg = &tdq_groups[i];
954123433Sjeff			/*
955123433Sjeff			 * Initialize the group.
956123433Sjeff			 */
957165620Sjeff			tdg->tdg_idlemask = 0;
958165620Sjeff			tdg->tdg_load = 0;
959165620Sjeff			tdg->tdg_transferable = 0;
960165620Sjeff			tdg->tdg_cpus = cg->cg_count;
961165620Sjeff			tdg->tdg_cpumask = cg->cg_mask;
962165620Sjeff			LIST_INIT(&tdg->tdg_members);
963123433Sjeff			/*
964123433Sjeff			 * Find all of the group members and add them.
965123433Sjeff			 */
966123433Sjeff			for (j = 0; j < MAXCPU; j++) {
967123433Sjeff				if ((cg->cg_mask & (1 << j)) != 0) {
968165620Sjeff					if (tdg->tdg_mask == 0)
969165620Sjeff						tdg->tdg_mask = 1 << j;
970165620Sjeff					tdq_cpu[j].tdq_transferable = 0;
971165620Sjeff					tdq_cpu[j].tdq_group = tdg;
972165620Sjeff					LIST_INSERT_HEAD(&tdg->tdg_members,
973165620Sjeff					    &tdq_cpu[j], tdq_siblings);
974123433Sjeff				}
975123433Sjeff			}
976165620Sjeff			if (tdg->tdg_cpus > 1)
977123487Sjeff				balance_groups = 1;
978117237Sjeff		}
979165620Sjeff		tdg_maxid = smp_topology->ct_count - 1;
980117237Sjeff	}
981123487Sjeff	/*
982123487Sjeff	 * Stagger the group and global load balancer so they do not
983123487Sjeff	 * interfere with each other.
984123487Sjeff	 */
985129982Sjeff	bal_tick = ticks + hz;
986123487Sjeff	if (balance_groups)
987129982Sjeff		gbal_tick = ticks + (hz / 2);
988117237Sjeff#else
989164936Sjulian	tdq_setup(TDQ_SELF());
990116069Sjeff#endif
991117237Sjeff	mtx_lock_spin(&sched_lock);
992164936Sjulian	tdq_load_add(TDQ_SELF(), &td_sched0);
993117237Sjeff	mtx_unlock_spin(&sched_lock);
994109864Sjeff}
995109864Sjeff
996153533Sdavidxu/* ARGSUSED */
997153533Sdavidxustatic void
998153533Sdavidxusched_initticks(void *dummy)
999153533Sdavidxu{
1000153533Sdavidxu	mtx_lock_spin(&sched_lock);
1001153533Sdavidxu	realstathz = stathz ? stathz : hz;
1002153533Sdavidxu	slice_min = (realstathz/100);	/* 10ms */
1003153533Sdavidxu	slice_max = (realstathz/7);	/* ~140ms */
1004153533Sdavidxu
1005153533Sdavidxu	tickincr = (hz << 10) / realstathz;
1006153533Sdavidxu	/*
1007153533Sdavidxu	 * XXX This does not work for values of stathz that are much
1008153533Sdavidxu	 * larger than hz.
1009153533Sdavidxu	 */
1010153533Sdavidxu	if (tickincr == 0)
1011153533Sdavidxu		tickincr = 1;
1012153533Sdavidxu	mtx_unlock_spin(&sched_lock);
1013153533Sdavidxu}
1014153533Sdavidxu
1015153533Sdavidxu
1016109864Sjeff/*
1017109864Sjeff * Scale the scheduling priority according to the "interactivity" of this
1018109864Sjeff * process.
1019109864Sjeff */
1020113357Sjeffstatic void
1021163709Sjbsched_priority(struct thread *td)
1022109864Sjeff{
1023109864Sjeff	int pri;
1024109864Sjeff
1025163709Sjb	if (td->td_pri_class != PRI_TIMESHARE)
1026113357Sjeff		return;
1027109864Sjeff
1028163709Sjb	pri = SCHED_PRI_INTERACT(sched_interact_score(td));
1029111857Sjeff	pri += SCHED_PRI_BASE;
1030163709Sjb	pri += td->td_proc->p_nice;
1031109864Sjeff
1032109864Sjeff	if (pri > PRI_MAX_TIMESHARE)
1033109864Sjeff		pri = PRI_MAX_TIMESHARE;
1034109864Sjeff	else if (pri < PRI_MIN_TIMESHARE)
1035109864Sjeff		pri = PRI_MIN_TIMESHARE;
1036109864Sjeff
1037163709Sjb	sched_user_prio(td, pri);
1038109864Sjeff
1039113357Sjeff	return;
1040109864Sjeff}
1041109864Sjeff
1042109864Sjeff/*
1043164936Sjulian * Calculate a time slice based on the properties of the process
1044164936Sjulian * and the runq that we're on.  This is only for PRI_TIMESHARE threads.
1045109864Sjeff */
1046112966Sjeffstatic void
1047164936Sjuliansched_slice(struct td_sched *ts)
1048109864Sjeff{
1049164936Sjulian	struct tdq *tdq;
1050163709Sjb	struct thread *td;
1051109864Sjeff
1052164936Sjulian	td = ts->ts_thread;
1053164936Sjulian	tdq = TDQ_CPU(ts->ts_cpu);
1054109864Sjeff
1055163709Sjb	if (td->td_flags & TDF_BORROWING) {
1056164936Sjulian		ts->ts_slice = SCHED_SLICE_MIN;
1057138842Sjeff		return;
1058138842Sjeff	}
1059138842Sjeff
1060112966Sjeff	/*
1061112966Sjeff	 * Rationale:
1062164936Sjulian	 * Threads in interactive procs get a minimal slice so that we
1063112966Sjeff	 * quickly notice if it abuses its advantage.
1064112966Sjeff	 *
1065164936Sjulian	 * Threads in non-interactive procs are assigned a slice that is
1066164936Sjulian	 * based on the procs nice value relative to the least nice procs
1067112966Sjeff	 * on the run queue for this cpu.
1068112966Sjeff	 *
1069164936Sjulian	 * If the thread is less nice than all others it gets the maximum
1070164936Sjulian	 * slice and other threads will adjust their slice relative to
1071112966Sjeff	 * this when they first expire.
1072112966Sjeff	 *
1073112966Sjeff	 * There is 20 point window that starts relative to the least
1074164936Sjulian	 * nice td_sched on the run queue.  Slice size is determined by
1075164936Sjulian	 * the td_sched distance from the last nice thread.
1076112966Sjeff	 *
1077164936Sjulian	 * If the td_sched is outside of the window it will get no slice
1078121871Sjeff	 * and will be reevaluated each time it is selected on the
1079164936Sjulian	 * run queue.  The exception to this is nice 0 procs when
1080121871Sjeff	 * a nice -20 is running.  They are always granted a minimum
1081121871Sjeff	 * slice.
1082112966Sjeff	 */
1083163709Sjb	if (!SCHED_INTERACTIVE(td)) {
1084112966Sjeff		int nice;
1085112966Sjeff
1086165620Sjeff		nice = td->td_proc->p_nice + (0 - tdq->tdq_nicemin);
1087165620Sjeff		if (tdq->tdq_load_timeshare == 0 ||
1088165620Sjeff		    td->td_proc->p_nice < tdq->tdq_nicemin)
1089164936Sjulian			ts->ts_slice = SCHED_SLICE_MAX;
1090121871Sjeff		else if (nice <= SCHED_SLICE_NTHRESH)
1091164936Sjulian			ts->ts_slice = SCHED_SLICE_NICE(nice);
1092163709Sjb		else if (td->td_proc->p_nice == 0)
1093164936Sjulian			ts->ts_slice = SCHED_SLICE_MIN;
1094112966Sjeff		else
1095164936Sjulian			ts->ts_slice = SCHED_SLICE_MIN; /* 0 */
1096112966Sjeff	} else
1097164936Sjulian		ts->ts_slice = SCHED_SLICE_INTERACTIVE;
1098112966Sjeff
1099112966Sjeff	return;
1100109864Sjeff}
1101109864Sjeff
1102121868Sjeff/*
1103121868Sjeff * This routine enforces a maximum limit on the amount of scheduling history
1104121868Sjeff * kept.  It is called after either the slptime or runtime is adjusted.
1105121868Sjeff * This routine will not operate correctly when slp or run times have been
1106121868Sjeff * adjusted to more than double their maximum.
1107121868Sjeff */
1108116463Sjeffstatic void
1109163709Sjbsched_interact_update(struct thread *td)
1110116463Sjeff{
1111121868Sjeff	int sum;
1112121605Sjeff
1113163709Sjb	sum = td->td_sched->skg_runtime + td->td_sched->skg_slptime;
1114121868Sjeff	if (sum < SCHED_SLP_RUN_MAX)
1115121868Sjeff		return;
1116121868Sjeff	/*
1117121868Sjeff	 * If we have exceeded by more than 1/5th then the algorithm below
1118121868Sjeff	 * will not bring us back into range.  Dividing by two here forces
1119133427Sjeff	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1120121868Sjeff	 */
1121127850Sjeff	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1122163709Sjb		td->td_sched->skg_runtime /= 2;
1123163709Sjb		td->td_sched->skg_slptime /= 2;
1124121868Sjeff		return;
1125116463Sjeff	}
1126163709Sjb	td->td_sched->skg_runtime = (td->td_sched->skg_runtime / 5) * 4;
1127163709Sjb	td->td_sched->skg_slptime = (td->td_sched->skg_slptime / 5) * 4;
1128116463Sjeff}
1129116463Sjeff
1130121868Sjeffstatic void
1131163709Sjbsched_interact_fork(struct thread *td)
1132121868Sjeff{
1133121868Sjeff	int ratio;
1134121868Sjeff	int sum;
1135121868Sjeff
1136163709Sjb	sum = td->td_sched->skg_runtime + td->td_sched->skg_slptime;
1137121868Sjeff	if (sum > SCHED_SLP_RUN_FORK) {
1138121868Sjeff		ratio = sum / SCHED_SLP_RUN_FORK;
1139163709Sjb		td->td_sched->skg_runtime /= ratio;
1140163709Sjb		td->td_sched->skg_slptime /= ratio;
1141121868Sjeff	}
1142121868Sjeff}
1143121868Sjeff
1144111857Sjeffstatic int
1145163709Sjbsched_interact_score(struct thread *td)
1146111857Sjeff{
1147116365Sjeff	int div;
1148111857Sjeff
1149163709Sjb	if (td->td_sched->skg_runtime > td->td_sched->skg_slptime) {
1150163709Sjb		div = max(1, td->td_sched->skg_runtime / SCHED_INTERACT_HALF);
1151116365Sjeff		return (SCHED_INTERACT_HALF +
1152163709Sjb		    (SCHED_INTERACT_HALF - (td->td_sched->skg_slptime / div)));
1153163709Sjb	} if (td->td_sched->skg_slptime > td->td_sched->skg_runtime) {
1154163709Sjb		div = max(1, td->td_sched->skg_slptime / SCHED_INTERACT_HALF);
1155163709Sjb		return (td->td_sched->skg_runtime / div);
1156111857Sjeff	}
1157111857Sjeff
1158116365Sjeff	/*
1159116365Sjeff	 * This can happen if slptime and runtime are 0.
1160116365Sjeff	 */
1161116365Sjeff	return (0);
1162111857Sjeff
1163111857Sjeff}
1164111857Sjeff
1165113357Sjeff/*
1166134791Sjulian * Very early in the boot some setup of scheduler-specific
1167134791Sjulian * parts of proc0 and of soem scheduler resources needs to be done.
1168134791Sjulian * Called from:
1169134791Sjulian *  proc0_init()
1170134791Sjulian */
1171134791Sjulianvoid
1172134791Sjulianschedinit(void)
1173134791Sjulian{
1174134791Sjulian	/*
1175134791Sjulian	 * Set up the scheduler specific parts of proc0.
1176134791Sjulian	 */
1177136167Sjulian	proc0.p_sched = NULL; /* XXX */
1178164936Sjulian	thread0.td_sched = &td_sched0;
1179164936Sjulian	td_sched0.ts_thread = &thread0;
1180164936Sjulian	td_sched0.ts_state = TSS_THREAD;
1181134791Sjulian}
1182134791Sjulian
1183134791Sjulian/*
1184113357Sjeff * This is only somewhat accurate since given many processes of the same
1185113357Sjeff * priority they will switch when their slices run out, which will be
1186113357Sjeff * at most SCHED_SLICE_MAX.
1187113357Sjeff */
1188109864Sjeffint
1189109864Sjeffsched_rr_interval(void)
1190109864Sjeff{
1191109864Sjeff	return (SCHED_SLICE_MAX);
1192109864Sjeff}
1193109864Sjeff
1194121790Sjeffstatic void
1195164936Sjuliansched_pctcpu_update(struct td_sched *ts)
1196109864Sjeff{
1197109864Sjeff	/*
1198109864Sjeff	 * Adjust counters and watermark for pctcpu calc.
1199116365Sjeff	 */
1200164936Sjulian	if (ts->ts_ltick > ticks - SCHED_CPU_TICKS) {
1201120272Sjeff		/*
1202120272Sjeff		 * Shift the tick count out so that the divide doesn't
1203120272Sjeff		 * round away our results.
1204120272Sjeff		 */
1205164936Sjulian		ts->ts_ticks <<= 10;
1206164936Sjulian		ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) *
1207120272Sjeff			    SCHED_CPU_TICKS;
1208164936Sjulian		ts->ts_ticks >>= 10;
1209120272Sjeff	} else
1210164936Sjulian		ts->ts_ticks = 0;
1211164936Sjulian	ts->ts_ltick = ticks;
1212164936Sjulian	ts->ts_ftick = ts->ts_ltick - SCHED_CPU_TICKS;
1213109864Sjeff}
1214109864Sjeff
1215109864Sjeffvoid
1216139453Sjhbsched_thread_priority(struct thread *td, u_char prio)
1217109864Sjeff{
1218164936Sjulian	struct td_sched *ts;
1219109864Sjeff
1220139316Sjeff	CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)",
1221139316Sjeff	    td, td->td_proc->p_comm, td->td_priority, prio, curthread,
1222139316Sjeff	    curthread->td_proc->p_comm);
1223164936Sjulian	ts = td->td_sched;
1224109864Sjeff	mtx_assert(&sched_lock, MA_OWNED);
1225139453Sjhb	if (td->td_priority == prio)
1226139453Sjhb		return;
1227109864Sjeff	if (TD_ON_RUNQ(td)) {
1228121605Sjeff		/*
1229121605Sjeff		 * If the priority has been elevated due to priority
1230121605Sjeff		 * propagation, we may have to move ourselves to a new
1231121605Sjeff		 * queue.  We still call adjustrunqueue below in case kse
1232121605Sjeff		 * needs to fix things up.
1233121605Sjeff		 */
1234164936Sjulian		if (prio < td->td_priority && ts->ts_runq != NULL &&
1235164936Sjulian		    (ts->ts_flags & TSF_ASSIGNED) == 0 &&
1236165620Sjeff		    ts->ts_runq != TDQ_CPU(ts->ts_cpu)->tdq_curr) {
1237164936Sjulian			runq_remove(ts->ts_runq, ts);
1238165620Sjeff			ts->ts_runq = TDQ_CPU(ts->ts_cpu)->tdq_curr;
1239164936Sjulian			runq_add(ts->ts_runq, ts, 0);
1240121605Sjeff		}
1241133555Sjeff		/*
1242164936Sjulian		 * Hold this td_sched on this cpu so that sched_prio() doesn't
1243133555Sjeff		 * cause excessive migration.  We only want migration to
1244133555Sjeff		 * happen as the result of a wakeup.
1245133555Sjeff		 */
1246164936Sjulian		ts->ts_flags |= TSF_HOLD;
1247119488Sdavidxu		adjustrunqueue(td, prio);
1248164936Sjulian		ts->ts_flags &= ~TSF_HOLD;
1249121605Sjeff	} else
1250119488Sdavidxu		td->td_priority = prio;
1251109864Sjeff}
1252109864Sjeff
1253139453Sjhb/*
1254139453Sjhb * Update a thread's priority when it is lent another thread's
1255139453Sjhb * priority.
1256139453Sjhb */
1257109864Sjeffvoid
1258139453Sjhbsched_lend_prio(struct thread *td, u_char prio)
1259139453Sjhb{
1260139453Sjhb
1261139453Sjhb	td->td_flags |= TDF_BORROWING;
1262139453Sjhb	sched_thread_priority(td, prio);
1263139453Sjhb}
1264139453Sjhb
1265139453Sjhb/*
1266139453Sjhb * Restore a thread's priority when priority propagation is
1267139453Sjhb * over.  The prio argument is the minimum priority the thread
1268139453Sjhb * needs to have to satisfy other possible priority lending
1269139453Sjhb * requests.  If the thread's regular priority is less
1270139453Sjhb * important than prio, the thread will keep a priority boost
1271139453Sjhb * of prio.
1272139453Sjhb */
1273139453Sjhbvoid
1274139453Sjhbsched_unlend_prio(struct thread *td, u_char prio)
1275139453Sjhb{
1276139453Sjhb	u_char base_pri;
1277139453Sjhb
1278139453Sjhb	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1279139453Sjhb	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1280163709Sjb		base_pri = td->td_user_pri;
1281139453Sjhb	else
1282139453Sjhb		base_pri = td->td_base_pri;
1283139453Sjhb	if (prio >= base_pri) {
1284139455Sjhb		td->td_flags &= ~TDF_BORROWING;
1285139453Sjhb		sched_thread_priority(td, base_pri);
1286139453Sjhb	} else
1287139453Sjhb		sched_lend_prio(td, prio);
1288139453Sjhb}
1289139453Sjhb
1290139453Sjhbvoid
1291139453Sjhbsched_prio(struct thread *td, u_char prio)
1292139453Sjhb{
1293139453Sjhb	u_char oldprio;
1294139453Sjhb
1295139453Sjhb	/* First, update the base priority. */
1296139453Sjhb	td->td_base_pri = prio;
1297139453Sjhb
1298139453Sjhb	/*
1299139455Sjhb	 * If the thread is borrowing another thread's priority, don't
1300139453Sjhb	 * ever lower the priority.
1301139453Sjhb	 */
1302139453Sjhb	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1303139453Sjhb		return;
1304139453Sjhb
1305139453Sjhb	/* Change the real priority. */
1306139453Sjhb	oldprio = td->td_priority;
1307139453Sjhb	sched_thread_priority(td, prio);
1308139453Sjhb
1309139453Sjhb	/*
1310139453Sjhb	 * If the thread is on a turnstile, then let the turnstile update
1311139453Sjhb	 * its state.
1312139453Sjhb	 */
1313139453Sjhb	if (TD_ON_LOCK(td) && oldprio != prio)
1314139453Sjhb		turnstile_adjust(td, oldprio);
1315139453Sjhb}
1316139455Sjhb
1317139453Sjhbvoid
1318163709Sjbsched_user_prio(struct thread *td, u_char prio)
1319161599Sdavidxu{
1320161599Sdavidxu	u_char oldprio;
1321161599Sdavidxu
1322163709Sjb	td->td_base_user_pri = prio;
1323164939Sjulian	if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
1324164939Sjulian                return;
1325163709Sjb	oldprio = td->td_user_pri;
1326163709Sjb	td->td_user_pri = prio;
1327163709Sjb
1328161599Sdavidxu	if (TD_ON_UPILOCK(td) && oldprio != prio)
1329161599Sdavidxu		umtx_pi_adjust(td, oldprio);
1330161599Sdavidxu}
1331161599Sdavidxu
1332161599Sdavidxuvoid
1333161599Sdavidxusched_lend_user_prio(struct thread *td, u_char prio)
1334161599Sdavidxu{
1335161599Sdavidxu	u_char oldprio;
1336161599Sdavidxu
1337161599Sdavidxu	td->td_flags |= TDF_UBORROWING;
1338161599Sdavidxu
1339164091Smaxim	oldprio = td->td_user_pri;
1340163709Sjb	td->td_user_pri = prio;
1341161599Sdavidxu
1342161599Sdavidxu	if (TD_ON_UPILOCK(td) && oldprio != prio)
1343161599Sdavidxu		umtx_pi_adjust(td, oldprio);
1344161599Sdavidxu}
1345161599Sdavidxu
1346161599Sdavidxuvoid
1347161599Sdavidxusched_unlend_user_prio(struct thread *td, u_char prio)
1348161599Sdavidxu{
1349161599Sdavidxu	u_char base_pri;
1350161599Sdavidxu
1351163709Sjb	base_pri = td->td_base_user_pri;
1352161599Sdavidxu	if (prio >= base_pri) {
1353161599Sdavidxu		td->td_flags &= ~TDF_UBORROWING;
1354163709Sjb		sched_user_prio(td, base_pri);
1355161599Sdavidxu	} else
1356161599Sdavidxu		sched_lend_user_prio(td, prio);
1357161599Sdavidxu}
1358161599Sdavidxu
1359161599Sdavidxuvoid
1360135051Sjuliansched_switch(struct thread *td, struct thread *newtd, int flags)
1361109864Sjeff{
1362165627Sjeff	struct tdq *tdq;
1363164936Sjulian	struct td_sched *ts;
1364109864Sjeff
1365109864Sjeff	mtx_assert(&sched_lock, MA_OWNED);
1366109864Sjeff
1367164936Sjulian	ts = td->td_sched;
1368165627Sjeff	tdq = TDQ_SELF();
1369109864Sjeff
1370133555Sjeff	td->td_lastcpu = td->td_oncpu;
1371113339Sjulian	td->td_oncpu = NOCPU;
1372132266Sjhb	td->td_flags &= ~TDF_NEEDRESCHED;
1373144777Sups	td->td_owepreempt = 0;
1374109864Sjeff
1375123434Sjeff	/*
1376164936Sjulian	 * If the thread has been assigned it may be in the process of switching
1377123434Sjeff	 * to the new cpu.  This is the case in sched_bind().
1378123434Sjeff	 */
1379139334Sjeff	if (td == PCPU_GET(idlethread)) {
1380139334Sjeff		TD_SET_CAN_RUN(td);
1381164936Sjulian	} else if ((ts->ts_flags & TSF_ASSIGNED) == 0) {
1382139334Sjeff		/* We are ending our run so make our slot available again */
1383165627Sjeff		tdq_load_rem(tdq, ts);
1384139334Sjeff		if (TD_IS_RUNNING(td)) {
1385139334Sjeff			/*
1386139334Sjeff			 * Don't allow the thread to migrate
1387139334Sjeff			 * from a preemption.
1388139334Sjeff			 */
1389164936Sjulian			ts->ts_flags |= TSF_HOLD;
1390139334Sjeff			setrunqueue(td, (flags & SW_PREEMPT) ?
1391139334Sjeff			    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1392139334Sjeff			    SRQ_OURSELF|SRQ_YIELDING);
1393164936Sjulian			ts->ts_flags &= ~TSF_HOLD;
1394163709Sjb		}
1395121146Sjeff	}
1396136167Sjulian	if (newtd != NULL) {
1397136170Sjulian		/*
1398147068Sjeff		 * If we bring in a thread account for it as if it had been
1399147068Sjeff		 * added to the run queue and then chosen.
1400136170Sjulian		 */
1401164936Sjulian		newtd->td_sched->ts_flags |= TSF_DIDRUN;
1402165627Sjeff		newtd->td_sched->ts_runq = tdq->tdq_curr;
1403136173Sjulian		TD_SET_RUNNING(newtd);
1404164936Sjulian		tdq_load_add(TDQ_SELF(), newtd->td_sched);
1405136167Sjulian	} else
1406131473Sjhb		newtd = choosethread();
1407145256Sjkoshy	if (td != newtd) {
1408145256Sjkoshy#ifdef	HWPMC_HOOKS
1409145256Sjkoshy		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1410145256Sjkoshy			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1411145256Sjkoshy#endif
1412163709Sjb
1413121128Sjeff		cpu_switch(td, newtd);
1414145256Sjkoshy#ifdef	HWPMC_HOOKS
1415145256Sjkoshy		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1416145256Sjkoshy			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1417145256Sjkoshy#endif
1418145256Sjkoshy	}
1419145256Sjkoshy
1420121128Sjeff	sched_lock.mtx_lock = (uintptr_t)td;
1421109864Sjeff
1422113339Sjulian	td->td_oncpu = PCPU_GET(cpuid);
1423109864Sjeff}
1424109864Sjeff
1425109864Sjeffvoid
1426130551Sjuliansched_nice(struct proc *p, int nice)
1427109864Sjeff{
1428164936Sjulian	struct td_sched *ts;
1429109864Sjeff	struct thread *td;
1430164936Sjulian	struct tdq *tdq;
1431109864Sjeff
1432130551Sjulian	PROC_LOCK_ASSERT(p, MA_OWNED);
1433113873Sjhb	mtx_assert(&sched_lock, MA_OWNED);
1434113357Sjeff	/*
1435164936Sjulian	 * We need to adjust the nice counts for running threads.
1436113357Sjeff	 */
1437163709Sjb	FOREACH_THREAD_IN_PROC(p, td) {
1438163709Sjb		if (td->td_pri_class == PRI_TIMESHARE) {
1439164936Sjulian			ts = td->td_sched;
1440164936Sjulian			if (ts->ts_runq == NULL)
1441163709Sjb				continue;
1442164936Sjulian			tdq = TDQ_CPU(ts->ts_cpu);
1443164936Sjulian			tdq_nice_rem(tdq, p->p_nice);
1444164936Sjulian			tdq_nice_add(tdq, nice);
1445113357Sjeff		}
1446130551Sjulian	}
1447130551Sjulian	p->p_nice = nice;
1448163709Sjb	FOREACH_THREAD_IN_PROC(p, td) {
1449163709Sjb		sched_priority(td);
1450163709Sjb		td->td_flags |= TDF_NEEDRESCHED;
1451130551Sjulian	}
1452109864Sjeff}
1453109864Sjeff
1454109864Sjeffvoid
1455126326Sjhbsched_sleep(struct thread *td)
1456109864Sjeff{
1457109864Sjeff	mtx_assert(&sched_lock, MA_OWNED);
1458109864Sjeff
1459164936Sjulian	td->td_sched->ts_slptime = ticks;
1460109864Sjeff}
1461109864Sjeff
1462109864Sjeffvoid
1463109864Sjeffsched_wakeup(struct thread *td)
1464109864Sjeff{
1465109864Sjeff	mtx_assert(&sched_lock, MA_OWNED);
1466109864Sjeff
1467109864Sjeff	/*
1468164936Sjulian	 * Let the procs know how long we slept for.  This is because process
1469164936Sjulian	 * interactivity behavior is modeled in the procs.
1470109864Sjeff	 */
1471164936Sjulian	if (td->td_sched->ts_slptime) {
1472113357Sjeff		int hzticks;
1473109864Sjeff
1474164936Sjulian		hzticks = (ticks - td->td_sched->ts_slptime) << 10;
1475121868Sjeff		if (hzticks >= SCHED_SLP_RUN_MAX) {
1476163709Sjb			td->td_sched->skg_slptime = SCHED_SLP_RUN_MAX;
1477163709Sjb			td->td_sched->skg_runtime = 1;
1478121868Sjeff		} else {
1479163709Sjb			td->td_sched->skg_slptime += hzticks;
1480163709Sjb			sched_interact_update(td);
1481121868Sjeff		}
1482163709Sjb		sched_priority(td);
1483164936Sjulian		sched_slice(td->td_sched);
1484164936Sjulian		td->td_sched->ts_slptime = 0;
1485109864Sjeff	}
1486134586Sjulian	setrunqueue(td, SRQ_BORING);
1487109864Sjeff}
1488109864Sjeff
1489109864Sjeff/*
1490109864Sjeff * Penalize the parent for creating a new child and initialize the child's
1491109864Sjeff * priority.
1492109864Sjeff */
1493109864Sjeffvoid
1494163709Sjbsched_fork(struct thread *td, struct thread *child)
1495109864Sjeff{
1496109864Sjeff	mtx_assert(&sched_lock, MA_OWNED);
1497164936Sjulian	sched_fork_thread(td, child);
1498164936Sjulian}
1499109864Sjeff
1500164936Sjulianvoid
1501164936Sjuliansched_fork_thread(struct thread *td, struct thread *child)
1502164936Sjulian{
1503164936Sjulian	struct td_sched *ts;
1504164936Sjulian	struct td_sched *ts2;
1505164936Sjulian
1506163709Sjb	child->td_sched->skg_slptime = td->td_sched->skg_slptime;
1507163709Sjb	child->td_sched->skg_runtime = td->td_sched->skg_runtime;
1508163709Sjb	child->td_user_pri = td->td_user_pri;
1509164091Smaxim	child->td_base_user_pri = td->td_base_user_pri;
1510121868Sjeff	sched_interact_fork(child);
1511163709Sjb	td->td_sched->skg_runtime += tickincr;
1512163709Sjb	sched_interact_update(td);
1513109864Sjeff
1514163709Sjb	sched_newthread(child);
1515134791Sjulian
1516164936Sjulian	ts = td->td_sched;
1517164936Sjulian	ts2 = child->td_sched;
1518164936Sjulian	ts2->ts_slice = 1;	/* Attempt to quickly learn interactivity. */
1519164936Sjulian	ts2->ts_cpu = ts->ts_cpu;
1520164936Sjulian	ts2->ts_runq = NULL;
1521134791Sjulian
1522134791Sjulian	/* Grab our parents cpu estimation information. */
1523164936Sjulian	ts2->ts_ticks = ts->ts_ticks;
1524164936Sjulian	ts2->ts_ltick = ts->ts_ltick;
1525164936Sjulian	ts2->ts_ftick = ts->ts_ftick;
1526113357Sjeff}
1527113357Sjeff
1528113357Sjeffvoid
1529163709Sjbsched_class(struct thread *td, int class)
1530113357Sjeff{
1531164936Sjulian	struct tdq *tdq;
1532164936Sjulian	struct td_sched *ts;
1533121896Sjeff	int nclass;
1534121896Sjeff	int oclass;
1535113357Sjeff
1536113923Sjhb	mtx_assert(&sched_lock, MA_OWNED);
1537163709Sjb	if (td->td_pri_class == class)
1538113357Sjeff		return;
1539113357Sjeff
1540121896Sjeff	nclass = PRI_BASE(class);
1541163709Sjb	oclass = PRI_BASE(td->td_pri_class);
1542164936Sjulian	ts = td->td_sched;
1543164936Sjulian	if (!((ts->ts_state != TSS_ONRUNQ &&
1544164936Sjulian	    ts->ts_state != TSS_THREAD) || ts->ts_runq == NULL)) {
1545164936Sjulian		tdq = TDQ_CPU(ts->ts_cpu);
1546113357Sjeff
1547121896Sjeff#ifdef SMP
1548164936Sjulian		/*
1549164936Sjulian		 * On SMP if we're on the RUNQ we must adjust the transferable
1550164936Sjulian		 * count because could be changing to or from an interrupt
1551164936Sjulian		 * class.
1552164936Sjulian		 */
1553164936Sjulian		if (ts->ts_state == TSS_ONRUNQ) {
1554164936Sjulian			if (THREAD_CAN_MIGRATE(ts)) {
1555165620Sjeff				tdq->tdq_transferable--;
1556165620Sjeff				tdq->tdq_group->tdg_transferable--;
1557164936Sjulian			}
1558164936Sjulian			if (THREAD_CAN_MIGRATE(ts)) {
1559165620Sjeff				tdq->tdq_transferable++;
1560165620Sjeff				tdq->tdq_group->tdg_transferable++;
1561164936Sjulian			}
1562122744Sjeff		}
1563164936Sjulian#endif
1564164936Sjulian		if (oclass == PRI_TIMESHARE) {
1565165620Sjeff			tdq->tdq_load_timeshare--;
1566164936Sjulian			tdq_nice_rem(tdq, td->td_proc->p_nice);
1567122744Sjeff		}
1568164936Sjulian		if (nclass == PRI_TIMESHARE) {
1569165620Sjeff			tdq->tdq_load_timeshare++;
1570164936Sjulian			tdq_nice_add(tdq, td->td_proc->p_nice);
1571164936Sjulian		}
1572109970Sjeff	}
1573109970Sjeff
1574163709Sjb	td->td_pri_class = class;
1575109864Sjeff}
1576109864Sjeff
1577109864Sjeff/*
1578109864Sjeff * Return some of the child's priority and interactivity to the parent.
1579109864Sjeff */
1580109864Sjeffvoid
1581164939Sjuliansched_exit(struct proc *p, struct thread *child)
1582109864Sjeff{
1583164939Sjulian
1584163709Sjb	CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d",
1585164939Sjulian	    child, child->td_proc->p_comm, child->td_priority);
1586113372Sjeff
1587164939Sjulian	sched_exit_thread(FIRST_THREAD_IN_PROC(p), child);
1588113372Sjeff}
1589113372Sjeff
1590113372Sjeffvoid
1591164939Sjuliansched_exit_thread(struct thread *td, struct thread *child)
1592164936Sjulian{
1593164939Sjulian	CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d",
1594164939Sjulian	    child, childproc->p_comm, child->td_priority);
1595164939Sjulian
1596164939Sjulian	td->td_sched->skg_runtime += child->td_sched->skg_runtime;
1597164939Sjulian	sched_interact_update(td);
1598164939Sjulian	tdq_load_rem(TDQ_CPU(child->td_sched->ts_cpu), child->td_sched);
1599164936Sjulian}
1600164936Sjulian
1601164936Sjulianvoid
1602164936Sjuliansched_userret(struct thread *td)
1603164936Sjulian{
1604164936Sjulian	/*
1605164936Sjulian	 * XXX we cheat slightly on the locking here to avoid locking in
1606164936Sjulian	 * the usual case.  Setting td_priority here is essentially an
1607164936Sjulian	 * incomplete workaround for not setting it properly elsewhere.
1608164936Sjulian	 * Now that some interrupt handlers are threads, not setting it
1609164936Sjulian	 * properly elsewhere can clobber it in the window between setting
1610164936Sjulian	 * it here and returning to user mode, so don't waste time setting
1611164936Sjulian	 * it perfectly here.
1612164936Sjulian	 */
1613164936Sjulian	KASSERT((td->td_flags & TDF_BORROWING) == 0,
1614164936Sjulian	    ("thread with borrowed priority returning to userland"));
1615164936Sjulian	if (td->td_priority != td->td_user_pri) {
1616164936Sjulian		mtx_lock_spin(&sched_lock);
1617164936Sjulian		td->td_priority = td->td_user_pri;
1618164936Sjulian		td->td_base_pri = td->td_user_pri;
1619164936Sjulian		mtx_unlock_spin(&sched_lock);
1620164936Sjulian        }
1621164936Sjulian}
1622164936Sjulian
1623164936Sjulianvoid
1624121127Sjeffsched_clock(struct thread *td)
1625109864Sjeff{
1626164936Sjulian	struct tdq *tdq;
1627164936Sjulian	struct td_sched *ts;
1628109864Sjeff
1629129982Sjeff	mtx_assert(&sched_lock, MA_OWNED);
1630164936Sjulian	tdq = TDQ_SELF();
1631129982Sjeff#ifdef SMP
1632139334Sjeff	if (ticks >= bal_tick)
1633129982Sjeff		sched_balance();
1634139334Sjeff	if (ticks >= gbal_tick && balance_groups)
1635129982Sjeff		sched_balance_groups();
1636133427Sjeff	/*
1637133427Sjeff	 * We could have been assigned a non real-time thread without an
1638133427Sjeff	 * IPI.
1639133427Sjeff	 */
1640165620Sjeff	if (tdq->tdq_assigned)
1641164936Sjulian		tdq_assign(tdq);	/* Potentially sets NEEDRESCHED */
1642129982Sjeff#endif
1643164936Sjulian	ts = td->td_sched;
1644109864Sjeff
1645110028Sjeff	/* Adjust ticks for pctcpu */
1646164936Sjulian	ts->ts_ticks++;
1647164936Sjulian	ts->ts_ltick = ticks;
1648112994Sjeff
1649109971Sjeff	/* Go up to one second beyond our max and then trim back down */
1650164936Sjulian	if (ts->ts_ftick + SCHED_CPU_TICKS + hz < ts->ts_ltick)
1651164936Sjulian		sched_pctcpu_update(ts);
1652109971Sjeff
1653114496Sjulian	if (td->td_flags & TDF_IDLETD)
1654109864Sjeff		return;
1655110028Sjeff	/*
1656163709Sjb	 * We only do slicing code for TIMESHARE threads.
1657113357Sjeff	 */
1658163709Sjb	if (td->td_pri_class != PRI_TIMESHARE)
1659113357Sjeff		return;
1660113357Sjeff	/*
1661163709Sjb	 * We used a tick charge it to the thread so that we can compute our
1662113357Sjeff	 * interactivity.
1663109864Sjeff	 */
1664163709Sjb	td->td_sched->skg_runtime += tickincr;
1665163709Sjb	sched_interact_update(td);
1666110645Sjeff
1667109864Sjeff	/*
1668109864Sjeff	 * We used up one time slice.
1669109864Sjeff	 */
1670164936Sjulian	if (--ts->ts_slice > 0)
1671113357Sjeff		return;
1672109864Sjeff	/*
1673113357Sjeff	 * We're out of time, recompute priorities and requeue.
1674109864Sjeff	 */
1675164936Sjulian	tdq_load_rem(tdq, ts);
1676163709Sjb	sched_priority(td);
1677164936Sjulian	sched_slice(ts);
1678164936Sjulian	if (SCHED_CURR(td, ts))
1679165620Sjeff		ts->ts_runq = tdq->tdq_curr;
1680113357Sjeff	else
1681165620Sjeff		ts->ts_runq = tdq->tdq_next;
1682164936Sjulian	tdq_load_add(tdq, ts);
1683113357Sjeff	td->td_flags |= TDF_NEEDRESCHED;
1684109864Sjeff}
1685109864Sjeff
1686109864Sjeffint
1687109864Sjeffsched_runnable(void)
1688109864Sjeff{
1689164936Sjulian	struct tdq *tdq;
1690115998Sjeff	int load;
1691109864Sjeff
1692115998Sjeff	load = 1;
1693115998Sjeff
1694164936Sjulian	tdq = TDQ_SELF();
1695121790Sjeff#ifdef SMP
1696165620Sjeff	if (tdq->tdq_assigned) {
1697122094Sjeff		mtx_lock_spin(&sched_lock);
1698164936Sjulian		tdq_assign(tdq);
1699122094Sjeff		mtx_unlock_spin(&sched_lock);
1700122094Sjeff	}
1701121790Sjeff#endif
1702121605Sjeff	if ((curthread->td_flags & TDF_IDLETD) != 0) {
1703165620Sjeff		if (tdq->tdq_load > 0)
1704121605Sjeff			goto out;
1705121605Sjeff	} else
1706165620Sjeff		if (tdq->tdq_load - 1 > 0)
1707121605Sjeff			goto out;
1708115998Sjeff	load = 0;
1709115998Sjeffout:
1710115998Sjeff	return (load);
1711109864Sjeff}
1712109864Sjeff
1713164936Sjulianstruct td_sched *
1714109970Sjeffsched_choose(void)
1715109970Sjeff{
1716164936Sjulian	struct tdq *tdq;
1717164936Sjulian	struct td_sched *ts;
1718109970Sjeff
1719115998Sjeff	mtx_assert(&sched_lock, MA_OWNED);
1720164936Sjulian	tdq = TDQ_SELF();
1721113357Sjeff#ifdef SMP
1722123433Sjeffrestart:
1723165620Sjeff	if (tdq->tdq_assigned)
1724164936Sjulian		tdq_assign(tdq);
1725113357Sjeff#endif
1726164936Sjulian	ts = tdq_choose(tdq);
1727164936Sjulian	if (ts) {
1728121790Sjeff#ifdef SMP
1729164936Sjulian		if (ts->ts_thread->td_pri_class == PRI_IDLE)
1730164936Sjulian			if (tdq_idled(tdq) == 0)
1731123433Sjeff				goto restart;
1732121790Sjeff#endif
1733164936Sjulian		tdq_runq_rem(tdq, ts);
1734164936Sjulian		ts->ts_state = TSS_THREAD;
1735164936Sjulian		ts->ts_flags &= ~TSF_PREEMPTED;
1736164936Sjulian		return (ts);
1737109864Sjeff	}
1738109970Sjeff#ifdef SMP
1739164936Sjulian	if (tdq_idled(tdq) == 0)
1740123433Sjeff		goto restart;
1741109970Sjeff#endif
1742113357Sjeff	return (NULL);
1743109864Sjeff}
1744109864Sjeff
1745109864Sjeffvoid
1746134586Sjuliansched_add(struct thread *td, int flags)
1747109864Sjeff{
1748164936Sjulian	struct tdq *tdq;
1749164936Sjulian	struct td_sched *ts;
1750139334Sjeff	int preemptive;
1751133427Sjeff	int canmigrate;
1752121790Sjeff	int class;
1753109864Sjeff
1754139316Sjeff	CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1755139316Sjeff	    td, td->td_proc->p_comm, td->td_priority, curthread,
1756139316Sjeff	    curthread->td_proc->p_comm);
1757121790Sjeff	mtx_assert(&sched_lock, MA_OWNED);
1758164936Sjulian	ts = td->td_sched;
1759139334Sjeff	canmigrate = 1;
1760139334Sjeff	preemptive = !(flags & SRQ_YIELDING);
1761163709Sjb	class = PRI_BASE(td->td_pri_class);
1762164936Sjulian	tdq = TDQ_SELF();
1763164936Sjulian	ts->ts_flags &= ~TSF_INTERNAL;
1764139334Sjeff#ifdef SMP
1765164936Sjulian	if (ts->ts_flags & TSF_ASSIGNED) {
1766164936Sjulian		if (ts->ts_flags & TSF_REMOVED)
1767164936Sjulian			ts->ts_flags &= ~TSF_REMOVED;
1768121790Sjeff		return;
1769138802Sjeff	}
1770164936Sjulian	canmigrate = THREAD_CAN_MIGRATE(ts);
1771149278Sdavidxu	/*
1772149278Sdavidxu	 * Don't migrate running threads here.  Force the long term balancer
1773149278Sdavidxu	 * to do it.
1774149278Sdavidxu	 */
1775164936Sjulian	if (ts->ts_flags & TSF_HOLD) {
1776164936Sjulian		ts->ts_flags &= ~TSF_HOLD;
1777149278Sdavidxu		canmigrate = 0;
1778149278Sdavidxu	}
1779139334Sjeff#endif
1780164936Sjulian	KASSERT(ts->ts_state != TSS_ONRUNQ,
1781164936Sjulian	    ("sched_add: thread %p (%s) already in run queue", td,
1782163709Sjb	    td->td_proc->p_comm));
1783163709Sjb	KASSERT(td->td_proc->p_sflag & PS_INMEM,
1784110267Sjeff	    ("sched_add: process swapped out"));
1785164936Sjulian	KASSERT(ts->ts_runq == NULL,
1786164936Sjulian	    ("sched_add: thread %p is still assigned to a run queue", td));
1787148856Sdavidxu	if (flags & SRQ_PREEMPTED)
1788164936Sjulian		ts->ts_flags |= TSF_PREEMPTED;
1789121790Sjeff	switch (class) {
1790112994Sjeff	case PRI_ITHD:
1791112994Sjeff	case PRI_REALTIME:
1792165620Sjeff		ts->ts_runq = tdq->tdq_curr;
1793164936Sjulian		ts->ts_slice = SCHED_SLICE_MAX;
1794139334Sjeff		if (canmigrate)
1795164936Sjulian			ts->ts_cpu = PCPU_GET(cpuid);
1796112994Sjeff		break;
1797112994Sjeff	case PRI_TIMESHARE:
1798164936Sjulian		if (SCHED_CURR(td, ts))
1799165620Sjeff			ts->ts_runq = tdq->tdq_curr;
1800113387Sjeff		else
1801165620Sjeff			ts->ts_runq = tdq->tdq_next;
1802113357Sjeff		break;
1803112994Sjeff	case PRI_IDLE:
1804113357Sjeff		/*
1805113357Sjeff		 * This is for priority prop.
1806113357Sjeff		 */
1807164936Sjulian		if (ts->ts_thread->td_priority < PRI_MIN_IDLE)
1808165620Sjeff			ts->ts_runq = tdq->tdq_curr;
1809113357Sjeff		else
1810165620Sjeff			ts->ts_runq = &tdq->tdq_idle;
1811164936Sjulian		ts->ts_slice = SCHED_SLICE_MIN;
1812112994Sjeff		break;
1813113357Sjeff	default:
1814121868Sjeff		panic("Unknown pri class.");
1815113357Sjeff		break;
1816112994Sjeff	}
1817121790Sjeff#ifdef SMP
1818133427Sjeff	/*
1819133427Sjeff	 * If this thread is pinned or bound, notify the target cpu.
1820133427Sjeff	 */
1821164936Sjulian	if (!canmigrate && ts->ts_cpu != PCPU_GET(cpuid) ) {
1822164936Sjulian		ts->ts_runq = NULL;
1823164936Sjulian		tdq_notify(ts, ts->ts_cpu);
1824123433Sjeff		return;
1825123433Sjeff	}
1826121790Sjeff	/*
1827123685Sjeff	 * If we had been idle, clear our bit in the group and potentially
1828123685Sjeff	 * the global bitmap.  If not, see if we should transfer this thread.
1829121790Sjeff	 */
1830123433Sjeff	if ((class == PRI_TIMESHARE || class == PRI_REALTIME) &&
1831165620Sjeff	    (tdq->tdq_group->tdg_idlemask & PCPU_GET(cpumask)) != 0) {
1832121790Sjeff		/*
1833123433Sjeff		 * Check to see if our group is unidling, and if so, remove it
1834123433Sjeff		 * from the global idle mask.
1835121790Sjeff		 */
1836165620Sjeff		if (tdq->tdq_group->tdg_idlemask ==
1837165620Sjeff		    tdq->tdq_group->tdg_cpumask)
1838165620Sjeff			atomic_clear_int(&tdq_idle, tdq->tdq_group->tdg_mask);
1839123433Sjeff		/*
1840123433Sjeff		 * Now remove ourselves from the group specific idle mask.
1841123433Sjeff		 */
1842165620Sjeff		tdq->tdq_group->tdg_idlemask &= ~PCPU_GET(cpumask);
1843165620Sjeff	} else if (canmigrate && tdq->tdq_load > 1 && class != PRI_ITHD)
1844164936Sjulian		if (tdq_transfer(tdq, ts, class))
1845123685Sjeff			return;
1846164936Sjulian	ts->ts_cpu = PCPU_GET(cpuid);
1847121790Sjeff#endif
1848133555Sjeff	if (td->td_priority < curthread->td_priority &&
1849165620Sjeff	    ts->ts_runq == tdq->tdq_curr)
1850133555Sjeff		curthread->td_flags |= TDF_NEEDRESCHED;
1851131839Sjhb	if (preemptive && maybe_preempt(td))
1852131481Sjhb		return;
1853164936Sjulian	ts->ts_state = TSS_ONRUNQ;
1854109864Sjeff
1855164936Sjulian	tdq_runq_add(tdq, ts, flags);
1856164936Sjulian	tdq_load_add(tdq, ts);
1857109864Sjeff}
1858109864Sjeff
1859109864Sjeffvoid
1860121127Sjeffsched_rem(struct thread *td)
1861109864Sjeff{
1862164936Sjulian	struct tdq *tdq;
1863164936Sjulian	struct td_sched *ts;
1864113357Sjeff
1865139316Sjeff	CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)",
1866139316Sjeff	    td, td->td_proc->p_comm, td->td_priority, curthread,
1867139316Sjeff	    curthread->td_proc->p_comm);
1868139334Sjeff	mtx_assert(&sched_lock, MA_OWNED);
1869164936Sjulian	ts = td->td_sched;
1870164936Sjulian	ts->ts_flags &= ~TSF_PREEMPTED;
1871164936Sjulian	if (ts->ts_flags & TSF_ASSIGNED) {
1872164936Sjulian		ts->ts_flags |= TSF_REMOVED;
1873121790Sjeff		return;
1874138802Sjeff	}
1875164936Sjulian	KASSERT((ts->ts_state == TSS_ONRUNQ),
1876164936Sjulian	    ("sched_rem: thread not on run queue"));
1877109864Sjeff
1878164936Sjulian	ts->ts_state = TSS_THREAD;
1879164936Sjulian	tdq = TDQ_CPU(ts->ts_cpu);
1880164936Sjulian	tdq_runq_rem(tdq, ts);
1881164936Sjulian	tdq_load_rem(tdq, ts);
1882109864Sjeff}
1883109864Sjeff
1884109864Sjefffixpt_t
1885121127Sjeffsched_pctcpu(struct thread *td)
1886109864Sjeff{
1887109864Sjeff	fixpt_t pctcpu;
1888164936Sjulian	struct td_sched *ts;
1889109864Sjeff
1890109864Sjeff	pctcpu = 0;
1891164936Sjulian	ts = td->td_sched;
1892164936Sjulian	if (ts == NULL)
1893121290Sjeff		return (0);
1894109864Sjeff
1895115998Sjeff	mtx_lock_spin(&sched_lock);
1896164936Sjulian	if (ts->ts_ticks) {
1897109864Sjeff		int rtick;
1898109864Sjeff
1899116365Sjeff		/*
1900116365Sjeff		 * Don't update more frequently than twice a second.  Allowing
1901116365Sjeff		 * this causes the cpu usage to decay away too quickly due to
1902116365Sjeff		 * rounding errors.
1903116365Sjeff		 */
1904164936Sjulian		if (ts->ts_ftick + SCHED_CPU_TICKS < ts->ts_ltick ||
1905164936Sjulian		    ts->ts_ltick < (ticks - (hz / 2)))
1906164936Sjulian			sched_pctcpu_update(ts);
1907109864Sjeff		/* How many rtick per second ? */
1908164936Sjulian		rtick = min(ts->ts_ticks / SCHED_CPU_TIME, SCHED_CPU_TICKS);
1909110226Sscottl		pctcpu = (FSCALE * ((FSCALE * rtick)/realstathz)) >> FSHIFT;
1910109864Sjeff	}
1911109864Sjeff
1912164936Sjulian	td->td_proc->p_swtime = ts->ts_ltick - ts->ts_ftick;
1913113865Sjhb	mtx_unlock_spin(&sched_lock);
1914109864Sjeff
1915109864Sjeff	return (pctcpu);
1916109864Sjeff}
1917109864Sjeff
1918122038Sjeffvoid
1919122038Sjeffsched_bind(struct thread *td, int cpu)
1920122038Sjeff{
1921164936Sjulian	struct td_sched *ts;
1922122038Sjeff
1923122038Sjeff	mtx_assert(&sched_lock, MA_OWNED);
1924164936Sjulian	ts = td->td_sched;
1925164936Sjulian	ts->ts_flags |= TSF_BOUND;
1926123433Sjeff#ifdef SMP
1927123433Sjeff	if (PCPU_GET(cpuid) == cpu)
1928122038Sjeff		return;
1929122038Sjeff	/* sched_rem without the runq_remove */
1930164936Sjulian	ts->ts_state = TSS_THREAD;
1931164936Sjulian	tdq_load_rem(TDQ_CPU(ts->ts_cpu), ts);
1932164936Sjulian	tdq_notify(ts, cpu);
1933122038Sjeff	/* When we return from mi_switch we'll be on the correct cpu. */
1934131527Sphk	mi_switch(SW_VOL, NULL);
1935122038Sjeff#endif
1936122038Sjeff}
1937122038Sjeff
1938122038Sjeffvoid
1939122038Sjeffsched_unbind(struct thread *td)
1940122038Sjeff{
1941122038Sjeff	mtx_assert(&sched_lock, MA_OWNED);
1942164936Sjulian	td->td_sched->ts_flags &= ~TSF_BOUND;
1943122038Sjeff}
1944122038Sjeff
1945109864Sjeffint
1946145256Sjkoshysched_is_bound(struct thread *td)
1947145256Sjkoshy{
1948145256Sjkoshy	mtx_assert(&sched_lock, MA_OWNED);
1949164936Sjulian	return (td->td_sched->ts_flags & TSF_BOUND);
1950145256Sjkoshy}
1951145256Sjkoshy
1952159630Sdavidxuvoid
1953159630Sdavidxusched_relinquish(struct thread *td)
1954159630Sdavidxu{
1955159630Sdavidxu	mtx_lock_spin(&sched_lock);
1956163709Sjb	if (td->td_pri_class == PRI_TIMESHARE)
1957159630Sdavidxu		sched_prio(td, PRI_MAX_TIMESHARE);
1958159630Sdavidxu	mi_switch(SW_VOL, NULL);
1959159630Sdavidxu	mtx_unlock_spin(&sched_lock);
1960159630Sdavidxu}
1961159630Sdavidxu
1962145256Sjkoshyint
1963125289Sjeffsched_load(void)
1964125289Sjeff{
1965125289Sjeff#ifdef SMP
1966125289Sjeff	int total;
1967125289Sjeff	int i;
1968125289Sjeff
1969125289Sjeff	total = 0;
1970165620Sjeff	for (i = 0; i <= tdg_maxid; i++)
1971165620Sjeff		total += TDQ_GROUP(i)->tdg_load;
1972125289Sjeff	return (total);
1973125289Sjeff#else
1974165620Sjeff	return (TDQ_SELF()->tdq_sysload);
1975125289Sjeff#endif
1976125289Sjeff}
1977125289Sjeff
1978125289Sjeffint
1979109864Sjeffsched_sizeof_proc(void)
1980109864Sjeff{
1981109864Sjeff	return (sizeof(struct proc));
1982109864Sjeff}
1983109864Sjeff
1984109864Sjeffint
1985109864Sjeffsched_sizeof_thread(void)
1986109864Sjeff{
1987109864Sjeff	return (sizeof(struct thread) + sizeof(struct td_sched));
1988109864Sjeff}
1989159570Sdavidxu
1990159570Sdavidxuvoid
1991159570Sdavidxusched_tick(void)
1992159570Sdavidxu{
1993159570Sdavidxu}
1994134791Sjulian#define KERN_SWITCH_INCLUDE 1
1995134791Sjulian#include "kern/kern_switch.c"
1996