sched_ule.c revision 216313
1/*-
2 * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 *    notice unmodified, this list of conditions, and the following
10 *    disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 *    notice, this list of conditions and the following disclaimer in the
13 *    documentation and/or other materials provided with the distribution.
14 *
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25 */
26
27/*
28 * This file implements the ULE scheduler.  ULE supports independent CPU
29 * run queues and fine grain locking.  It has superior interactive
30 * performance under load even on uni-processor systems.
31 *
32 * etymology:
33 *   ULE is the last three letters in schedule.  It owes its name to a
34 * generic user created for a scheduling system by Paul Mikesell at
35 * Isilon Systems and a general lack of creativity on the part of the author.
36 */
37
38#include <sys/cdefs.h>
39__FBSDID("$FreeBSD: head/sys/kern/sched_ule.c 216313 2010-12-09 02:42:02Z davidxu $");
40
41#include "opt_hwpmc_hooks.h"
42#include "opt_kdtrace.h"
43#include "opt_sched.h"
44
45#include <sys/param.h>
46#include <sys/systm.h>
47#include <sys/kdb.h>
48#include <sys/kernel.h>
49#include <sys/ktr.h>
50#include <sys/lock.h>
51#include <sys/mutex.h>
52#include <sys/proc.h>
53#include <sys/resource.h>
54#include <sys/resourcevar.h>
55#include <sys/sched.h>
56#include <sys/smp.h>
57#include <sys/sx.h>
58#include <sys/sysctl.h>
59#include <sys/sysproto.h>
60#include <sys/turnstile.h>
61#include <sys/umtx.h>
62#include <sys/vmmeter.h>
63#include <sys/cpuset.h>
64#include <sys/sbuf.h>
65
66#ifdef HWPMC_HOOKS
67#include <sys/pmckern.h>
68#endif
69
70#ifdef KDTRACE_HOOKS
71#include <sys/dtrace_bsd.h>
72int				dtrace_vtime_active;
73dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
74#endif
75
76#include <machine/cpu.h>
77#include <machine/smp.h>
78
79#if defined(__sparc64__)
80#error "This architecture is not currently compatible with ULE"
81#endif
82
83#define	KTR_ULE	0
84
85#define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
86#define	TDQ_NAME_LEN	(sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
87#define	TDQ_LOADNAME_LEN	(PCPU_NAME_LEN + sizeof(" load"))
88
89/*
90 * Thread scheduler specific section.  All fields are protected
91 * by the thread lock.
92 */
93struct td_sched {
94	struct runq	*ts_runq;	/* Run-queue we're queued on. */
95	short		ts_flags;	/* TSF_* flags. */
96	u_char		ts_cpu;		/* CPU that we have affinity for. */
97	int		ts_rltick;	/* Real last tick, for affinity. */
98	int		ts_slice;	/* Ticks of slice remaining. */
99	u_int		ts_slptime;	/* Number of ticks we vol. slept */
100	u_int		ts_runtime;	/* Number of ticks we were running */
101	int		ts_ltick;	/* Last tick that we were running on */
102	int		ts_incrtick;	/* Last tick that we incremented on */
103	int		ts_ftick;	/* First tick that we were running on */
104	int		ts_ticks;	/* Tick count */
105#ifdef KTR
106	char		ts_name[TS_NAME_LEN];
107#endif
108};
109/* flags kept in ts_flags */
110#define	TSF_BOUND	0x0001		/* Thread can not migrate. */
111#define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
112
113static struct td_sched td_sched0;
114
115#define	THREAD_CAN_MIGRATE(td)	((td)->td_pinned == 0)
116#define	THREAD_CAN_SCHED(td, cpu)	\
117    CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
118
119/*
120 * Cpu percentage computation macros and defines.
121 *
122 * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
123 * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
124 * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
125 * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
126 * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
127 * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
128 */
129#define	SCHED_TICK_SECS		10
130#define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
131#define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
132#define	SCHED_TICK_SHIFT	10
133#define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
134#define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
135
136/*
137 * These macros determine priorities for non-interactive threads.  They are
138 * assigned a priority based on their recent cpu utilization as expressed
139 * by the ratio of ticks to the tick total.  NHALF priorities at the start
140 * and end of the MIN to MAX timeshare range are only reachable with negative
141 * or positive nice respectively.
142 *
143 * PRI_RANGE:	Priority range for utilization dependent priorities.
144 * PRI_NRESV:	Number of nice values.
145 * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
146 * PRI_NICE:	Determines the part of the priority inherited from nice.
147 */
148#define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
149#define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
150#define	SCHED_PRI_MIN		(PRI_MIN_TIMESHARE + SCHED_PRI_NHALF)
151#define	SCHED_PRI_MAX		(PRI_MAX_TIMESHARE - SCHED_PRI_NHALF)
152#define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN)
153#define	SCHED_PRI_TICKS(ts)						\
154    (SCHED_TICK_HZ((ts)) /						\
155    (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
156#define	SCHED_PRI_NICE(nice)	(nice)
157
158/*
159 * These determine the interactivity of a process.  Interactivity differs from
160 * cpu utilization in that it expresses the voluntary time slept vs time ran
161 * while cpu utilization includes all time not running.  This more accurately
162 * models the intent of the thread.
163 *
164 * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
165 *		before throttling back.
166 * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
167 * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
168 * INTERACT_THRESH:	Threshold for placement on the current runq.
169 */
170#define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
171#define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
172#define	SCHED_INTERACT_MAX	(100)
173#define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
174#define	SCHED_INTERACT_THRESH	(30)
175
176/*
177 * tickincr:		Converts a stathz tick into a hz domain scaled by
178 *			the shift factor.  Without the shift the error rate
179 *			due to rounding would be unacceptably high.
180 * realstathz:		stathz is sometimes 0 and run off of hz.
181 * sched_slice:		Runtime of each thread before rescheduling.
182 * preempt_thresh:	Priority threshold for preemption and remote IPIs.
183 */
184static int sched_interact = SCHED_INTERACT_THRESH;
185static int realstathz;
186static int tickincr;
187static int sched_slice = 1;
188#ifdef PREEMPTION
189#ifdef FULL_PREEMPTION
190static int preempt_thresh = PRI_MAX_IDLE;
191#else
192static int preempt_thresh = PRI_MIN_KERN;
193#endif
194#else
195static int preempt_thresh = 0;
196#endif
197static int static_boost = PRI_MIN_TIMESHARE;
198static int sched_idlespins = 10000;
199static int sched_idlespinthresh = 16;
200
201/*
202 * tdq - per processor runqs and statistics.  All fields are protected by the
203 * tdq_lock.  The load and lowpri may be accessed without to avoid excess
204 * locking in sched_pickcpu();
205 */
206struct tdq {
207	/* Ordered to improve efficiency of cpu_search() and switch(). */
208	struct mtx	tdq_lock;		/* run queue lock. */
209	struct cpu_group *tdq_cg;		/* Pointer to cpu topology. */
210	volatile int	tdq_load;		/* Aggregate load. */
211	volatile int	tdq_cpu_idle;		/* cpu_idle() is active. */
212	int		tdq_sysload;		/* For loadavg, !ITHD load. */
213	int		tdq_transferable;	/* Transferable thread count. */
214	short		tdq_switchcnt;		/* Switches this tick. */
215	short		tdq_oldswitchcnt;	/* Switches last tick. */
216	u_char		tdq_lowpri;		/* Lowest priority thread. */
217	u_char		tdq_ipipending;		/* IPI pending. */
218	u_char		tdq_idx;		/* Current insert index. */
219	u_char		tdq_ridx;		/* Current removal index. */
220	struct runq	tdq_realtime;		/* real-time run queue. */
221	struct runq	tdq_timeshare;		/* timeshare run queue. */
222	struct runq	tdq_idle;		/* Queue of IDLE threads. */
223	char		tdq_name[TDQ_NAME_LEN];
224#ifdef KTR
225	char		tdq_loadname[TDQ_LOADNAME_LEN];
226#endif
227} __aligned(64);
228
229/* Idle thread states and config. */
230#define	TDQ_RUNNING	1
231#define	TDQ_IDLE	2
232
233#ifdef SMP
234struct cpu_group *cpu_top;		/* CPU topology */
235
236#define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
237#define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
238
239/*
240 * Run-time tunables.
241 */
242static int rebalance = 1;
243static int balance_interval = 128;	/* Default set in sched_initticks(). */
244static int affinity;
245static int steal_htt = 1;
246static int steal_idle = 1;
247static int steal_thresh = 2;
248
249/*
250 * One thread queue per processor.
251 */
252static struct tdq	tdq_cpu[MAXCPU];
253static struct tdq	*balance_tdq;
254static int balance_ticks;
255
256#define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
257#define	TDQ_CPU(x)	(&tdq_cpu[(x)])
258#define	TDQ_ID(x)	((int)((x) - tdq_cpu))
259#else	/* !SMP */
260static struct tdq	tdq_cpu;
261
262#define	TDQ_ID(x)	(0)
263#define	TDQ_SELF()	(&tdq_cpu)
264#define	TDQ_CPU(x)	(&tdq_cpu)
265#endif
266
267#define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
268#define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
269#define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
270#define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
271#define	TDQ_LOCKPTR(t)		(&(t)->tdq_lock)
272
273static void sched_priority(struct thread *);
274static void sched_thread_priority(struct thread *, u_char);
275static int sched_interact_score(struct thread *);
276static void sched_interact_update(struct thread *);
277static void sched_interact_fork(struct thread *);
278static void sched_pctcpu_update(struct td_sched *);
279
280/* Operations on per processor queues */
281static struct thread *tdq_choose(struct tdq *);
282static void tdq_setup(struct tdq *);
283static void tdq_load_add(struct tdq *, struct thread *);
284static void tdq_load_rem(struct tdq *, struct thread *);
285static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
286static __inline void tdq_runq_rem(struct tdq *, struct thread *);
287static inline int sched_shouldpreempt(int, int, int);
288void tdq_print(int cpu);
289static void runq_print(struct runq *rq);
290static void tdq_add(struct tdq *, struct thread *, int);
291#ifdef SMP
292static int tdq_move(struct tdq *, struct tdq *);
293static int tdq_idled(struct tdq *);
294static void tdq_notify(struct tdq *, struct thread *);
295static struct thread *tdq_steal(struct tdq *, int);
296static struct thread *runq_steal(struct runq *, int);
297static int sched_pickcpu(struct thread *, int);
298static void sched_balance(void);
299static int sched_balance_pair(struct tdq *, struct tdq *);
300static inline struct tdq *sched_setcpu(struct thread *, int, int);
301static inline void thread_unblock_switch(struct thread *, struct mtx *);
302static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
303static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
304static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
305    struct cpu_group *cg, int indent);
306#endif
307
308static void sched_setup(void *dummy);
309SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
310
311static void sched_initticks(void *dummy);
312SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
313    NULL);
314
315/*
316 * Print the threads waiting on a run-queue.
317 */
318static void
319runq_print(struct runq *rq)
320{
321	struct rqhead *rqh;
322	struct thread *td;
323	int pri;
324	int j;
325	int i;
326
327	for (i = 0; i < RQB_LEN; i++) {
328		printf("\t\trunq bits %d 0x%zx\n",
329		    i, rq->rq_status.rqb_bits[i]);
330		for (j = 0; j < RQB_BPW; j++)
331			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
332				pri = j + (i << RQB_L2BPW);
333				rqh = &rq->rq_queues[pri];
334				TAILQ_FOREACH(td, rqh, td_runq) {
335					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
336					    td, td->td_name, td->td_priority,
337					    td->td_rqindex, pri);
338				}
339			}
340	}
341}
342
343/*
344 * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
345 */
346void
347tdq_print(int cpu)
348{
349	struct tdq *tdq;
350
351	tdq = TDQ_CPU(cpu);
352
353	printf("tdq %d:\n", TDQ_ID(tdq));
354	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
355	printf("\tLock name:      %s\n", tdq->tdq_name);
356	printf("\tload:           %d\n", tdq->tdq_load);
357	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
358	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
359	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
360	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
361	printf("\tload transferable: %d\n", tdq->tdq_transferable);
362	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
363	printf("\trealtime runq:\n");
364	runq_print(&tdq->tdq_realtime);
365	printf("\ttimeshare runq:\n");
366	runq_print(&tdq->tdq_timeshare);
367	printf("\tidle runq:\n");
368	runq_print(&tdq->tdq_idle);
369}
370
371static inline int
372sched_shouldpreempt(int pri, int cpri, int remote)
373{
374	/*
375	 * If the new priority is not better than the current priority there is
376	 * nothing to do.
377	 */
378	if (pri >= cpri)
379		return (0);
380	/*
381	 * Always preempt idle.
382	 */
383	if (cpri >= PRI_MIN_IDLE)
384		return (1);
385	/*
386	 * If preemption is disabled don't preempt others.
387	 */
388	if (preempt_thresh == 0)
389		return (0);
390	/*
391	 * Preempt if we exceed the threshold.
392	 */
393	if (pri <= preempt_thresh)
394		return (1);
395	/*
396	 * If we're realtime or better and there is timeshare or worse running
397	 * preempt only remote processors.
398	 */
399	if (remote && pri <= PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME)
400		return (1);
401	return (0);
402}
403
404#define	TS_RQ_PPQ	(((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS)
405/*
406 * Add a thread to the actual run-queue.  Keeps transferable counts up to
407 * date with what is actually on the run-queue.  Selects the correct
408 * queue position for timeshare threads.
409 */
410static __inline void
411tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
412{
413	struct td_sched *ts;
414	u_char pri;
415
416	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
417	THREAD_LOCK_ASSERT(td, MA_OWNED);
418
419	pri = td->td_priority;
420	ts = td->td_sched;
421	TD_SET_RUNQ(td);
422	if (THREAD_CAN_MIGRATE(td)) {
423		tdq->tdq_transferable++;
424		ts->ts_flags |= TSF_XFERABLE;
425	}
426	if (pri <= PRI_MAX_REALTIME) {
427		ts->ts_runq = &tdq->tdq_realtime;
428	} else if (pri <= PRI_MAX_TIMESHARE) {
429		ts->ts_runq = &tdq->tdq_timeshare;
430		KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE,
431			("Invalid priority %d on timeshare runq", pri));
432		/*
433		 * This queue contains only priorities between MIN and MAX
434		 * realtime.  Use the whole queue to represent these values.
435		 */
436		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
437			pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ;
438			pri = (pri + tdq->tdq_idx) % RQ_NQS;
439			/*
440			 * This effectively shortens the queue by one so we
441			 * can have a one slot difference between idx and
442			 * ridx while we wait for threads to drain.
443			 */
444			if (tdq->tdq_ridx != tdq->tdq_idx &&
445			    pri == tdq->tdq_ridx)
446				pri = (unsigned char)(pri - 1) % RQ_NQS;
447		} else
448			pri = tdq->tdq_ridx;
449		runq_add_pri(ts->ts_runq, td, pri, flags);
450		return;
451	} else
452		ts->ts_runq = &tdq->tdq_idle;
453	runq_add(ts->ts_runq, td, flags);
454}
455
456/*
457 * Remove a thread from a run-queue.  This typically happens when a thread
458 * is selected to run.  Running threads are not on the queue and the
459 * transferable count does not reflect them.
460 */
461static __inline void
462tdq_runq_rem(struct tdq *tdq, struct thread *td)
463{
464	struct td_sched *ts;
465
466	ts = td->td_sched;
467	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
468	KASSERT(ts->ts_runq != NULL,
469	    ("tdq_runq_remove: thread %p null ts_runq", td));
470	if (ts->ts_flags & TSF_XFERABLE) {
471		tdq->tdq_transferable--;
472		ts->ts_flags &= ~TSF_XFERABLE;
473	}
474	if (ts->ts_runq == &tdq->tdq_timeshare) {
475		if (tdq->tdq_idx != tdq->tdq_ridx)
476			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
477		else
478			runq_remove_idx(ts->ts_runq, td, NULL);
479	} else
480		runq_remove(ts->ts_runq, td);
481}
482
483/*
484 * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
485 * for this thread to the referenced thread queue.
486 */
487static void
488tdq_load_add(struct tdq *tdq, struct thread *td)
489{
490
491	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
492	THREAD_LOCK_ASSERT(td, MA_OWNED);
493
494	tdq->tdq_load++;
495	if ((td->td_flags & TDF_NOLOAD) == 0)
496		tdq->tdq_sysload++;
497	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
498}
499
500/*
501 * Remove the load from a thread that is transitioning to a sleep state or
502 * exiting.
503 */
504static void
505tdq_load_rem(struct tdq *tdq, struct thread *td)
506{
507
508	THREAD_LOCK_ASSERT(td, MA_OWNED);
509	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
510	KASSERT(tdq->tdq_load != 0,
511	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
512
513	tdq->tdq_load--;
514	if ((td->td_flags & TDF_NOLOAD) == 0)
515		tdq->tdq_sysload--;
516	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
517}
518
519/*
520 * Set lowpri to its exact value by searching the run-queue and
521 * evaluating curthread.  curthread may be passed as an optimization.
522 */
523static void
524tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
525{
526	struct thread *td;
527
528	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
529	if (ctd == NULL)
530		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
531	td = tdq_choose(tdq);
532	if (td == NULL || td->td_priority > ctd->td_priority)
533		tdq->tdq_lowpri = ctd->td_priority;
534	else
535		tdq->tdq_lowpri = td->td_priority;
536}
537
538#ifdef SMP
539struct cpu_search {
540	cpuset_t cs_mask;
541	u_int	cs_load;
542	u_int	cs_cpu;
543	int	cs_limit;	/* Min priority for low min load for high. */
544};
545
546#define	CPU_SEARCH_LOWEST	0x1
547#define	CPU_SEARCH_HIGHEST	0x2
548#define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
549
550#define	CPUSET_FOREACH(cpu, mask)				\
551	for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++)		\
552		if ((mask) & 1 << (cpu))
553
554static __inline int cpu_search(struct cpu_group *cg, struct cpu_search *low,
555    struct cpu_search *high, const int match);
556int cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low);
557int cpu_search_highest(struct cpu_group *cg, struct cpu_search *high);
558int cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
559    struct cpu_search *high);
560
561/*
562 * This routine compares according to the match argument and should be
563 * reduced in actual instantiations via constant propagation and dead code
564 * elimination.
565 */
566static __inline int
567cpu_compare(int cpu, struct cpu_search *low, struct cpu_search *high,
568    const int match)
569{
570	struct tdq *tdq;
571
572	tdq = TDQ_CPU(cpu);
573	if (match & CPU_SEARCH_LOWEST)
574		if (CPU_ISSET(cpu, &low->cs_mask) &&
575		    tdq->tdq_load < low->cs_load &&
576		    tdq->tdq_lowpri > low->cs_limit) {
577			low->cs_cpu = cpu;
578			low->cs_load = tdq->tdq_load;
579		}
580	if (match & CPU_SEARCH_HIGHEST)
581		if (CPU_ISSET(cpu, &high->cs_mask) &&
582		    tdq->tdq_load >= high->cs_limit &&
583		    tdq->tdq_load > high->cs_load &&
584		    tdq->tdq_transferable) {
585			high->cs_cpu = cpu;
586			high->cs_load = tdq->tdq_load;
587		}
588	return (tdq->tdq_load);
589}
590
591/*
592 * Search the tree of cpu_groups for the lowest or highest loaded cpu
593 * according to the match argument.  This routine actually compares the
594 * load on all paths through the tree and finds the least loaded cpu on
595 * the least loaded path, which may differ from the least loaded cpu in
596 * the system.  This balances work among caches and busses.
597 *
598 * This inline is instantiated in three forms below using constants for the
599 * match argument.  It is reduced to the minimum set for each case.  It is
600 * also recursive to the depth of the tree.
601 */
602static __inline int
603cpu_search(struct cpu_group *cg, struct cpu_search *low,
604    struct cpu_search *high, const int match)
605{
606	int total;
607
608	total = 0;
609	if (cg->cg_children) {
610		struct cpu_search lgroup;
611		struct cpu_search hgroup;
612		struct cpu_group *child;
613		u_int lload;
614		int hload;
615		int load;
616		int i;
617
618		lload = -1;
619		hload = -1;
620		for (i = 0; i < cg->cg_children; i++) {
621			child = &cg->cg_child[i];
622			if (match & CPU_SEARCH_LOWEST) {
623				lgroup = *low;
624				lgroup.cs_load = -1;
625			}
626			if (match & CPU_SEARCH_HIGHEST) {
627				hgroup = *high;
628				lgroup.cs_load = 0;
629			}
630			switch (match) {
631			case CPU_SEARCH_LOWEST:
632				load = cpu_search_lowest(child, &lgroup);
633				break;
634			case CPU_SEARCH_HIGHEST:
635				load = cpu_search_highest(child, &hgroup);
636				break;
637			case CPU_SEARCH_BOTH:
638				load = cpu_search_both(child, &lgroup, &hgroup);
639				break;
640			}
641			total += load;
642			if (match & CPU_SEARCH_LOWEST)
643				if (load < lload || low->cs_cpu == -1) {
644					*low = lgroup;
645					lload = load;
646				}
647			if (match & CPU_SEARCH_HIGHEST)
648				if (load > hload || high->cs_cpu == -1) {
649					hload = load;
650					*high = hgroup;
651				}
652		}
653	} else {
654		int cpu;
655
656		CPUSET_FOREACH(cpu, cg->cg_mask)
657			total += cpu_compare(cpu, low, high, match);
658	}
659	return (total);
660}
661
662/*
663 * cpu_search instantiations must pass constants to maintain the inline
664 * optimization.
665 */
666int
667cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low)
668{
669	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
670}
671
672int
673cpu_search_highest(struct cpu_group *cg, struct cpu_search *high)
674{
675	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
676}
677
678int
679cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
680    struct cpu_search *high)
681{
682	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
683}
684
685/*
686 * Find the cpu with the least load via the least loaded path that has a
687 * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
688 * acceptable.
689 */
690static inline int
691sched_lowest(struct cpu_group *cg, cpuset_t mask, int pri)
692{
693	struct cpu_search low;
694
695	low.cs_cpu = -1;
696	low.cs_load = -1;
697	low.cs_mask = mask;
698	low.cs_limit = pri;
699	cpu_search_lowest(cg, &low);
700	return low.cs_cpu;
701}
702
703/*
704 * Find the cpu with the highest load via the highest loaded path.
705 */
706static inline int
707sched_highest(struct cpu_group *cg, cpuset_t mask, int minload)
708{
709	struct cpu_search high;
710
711	high.cs_cpu = -1;
712	high.cs_load = 0;
713	high.cs_mask = mask;
714	high.cs_limit = minload;
715	cpu_search_highest(cg, &high);
716	return high.cs_cpu;
717}
718
719/*
720 * Simultaneously find the highest and lowest loaded cpu reachable via
721 * cg.
722 */
723static inline void
724sched_both(struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu)
725{
726	struct cpu_search high;
727	struct cpu_search low;
728
729	low.cs_cpu = -1;
730	low.cs_limit = -1;
731	low.cs_load = -1;
732	low.cs_mask = mask;
733	high.cs_load = 0;
734	high.cs_cpu = -1;
735	high.cs_limit = -1;
736	high.cs_mask = mask;
737	cpu_search_both(cg, &low, &high);
738	*lowcpu = low.cs_cpu;
739	*highcpu = high.cs_cpu;
740	return;
741}
742
743static void
744sched_balance_group(struct cpu_group *cg)
745{
746	cpuset_t mask;
747	int high;
748	int low;
749	int i;
750
751	CPU_FILL(&mask);
752	for (;;) {
753		sched_both(cg, mask, &low, &high);
754		if (low == high || low == -1 || high == -1)
755			break;
756		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low)))
757			break;
758		/*
759		 * If we failed to move any threads determine which cpu
760		 * to kick out of the set and try again.
761	 	 */
762		if (TDQ_CPU(high)->tdq_transferable == 0)
763			CPU_CLR(high, &mask);
764		else
765			CPU_CLR(low, &mask);
766	}
767
768	for (i = 0; i < cg->cg_children; i++)
769		sched_balance_group(&cg->cg_child[i]);
770}
771
772static void
773sched_balance(void)
774{
775	struct tdq *tdq;
776
777	/*
778	 * Select a random time between .5 * balance_interval and
779	 * 1.5 * balance_interval.
780	 */
781	balance_ticks = max(balance_interval / 2, 1);
782	balance_ticks += random() % balance_interval;
783	if (smp_started == 0 || rebalance == 0)
784		return;
785	tdq = TDQ_SELF();
786	TDQ_UNLOCK(tdq);
787	sched_balance_group(cpu_top);
788	TDQ_LOCK(tdq);
789}
790
791/*
792 * Lock two thread queues using their address to maintain lock order.
793 */
794static void
795tdq_lock_pair(struct tdq *one, struct tdq *two)
796{
797	if (one < two) {
798		TDQ_LOCK(one);
799		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
800	} else {
801		TDQ_LOCK(two);
802		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
803	}
804}
805
806/*
807 * Unlock two thread queues.  Order is not important here.
808 */
809static void
810tdq_unlock_pair(struct tdq *one, struct tdq *two)
811{
812	TDQ_UNLOCK(one);
813	TDQ_UNLOCK(two);
814}
815
816/*
817 * Transfer load between two imbalanced thread queues.
818 */
819static int
820sched_balance_pair(struct tdq *high, struct tdq *low)
821{
822	int transferable;
823	int high_load;
824	int low_load;
825	int moved;
826	int move;
827	int diff;
828	int i;
829
830	tdq_lock_pair(high, low);
831	transferable = high->tdq_transferable;
832	high_load = high->tdq_load;
833	low_load = low->tdq_load;
834	moved = 0;
835	/*
836	 * Determine what the imbalance is and then adjust that to how many
837	 * threads we actually have to give up (transferable).
838	 */
839	if (transferable != 0) {
840		diff = high_load - low_load;
841		move = diff / 2;
842		if (diff & 0x1)
843			move++;
844		move = min(move, transferable);
845		for (i = 0; i < move; i++)
846			moved += tdq_move(high, low);
847		/*
848		 * IPI the target cpu to force it to reschedule with the new
849		 * workload.
850		 */
851		ipi_cpu(TDQ_ID(low), IPI_PREEMPT);
852	}
853	tdq_unlock_pair(high, low);
854	return (moved);
855}
856
857/*
858 * Move a thread from one thread queue to another.
859 */
860static int
861tdq_move(struct tdq *from, struct tdq *to)
862{
863	struct td_sched *ts;
864	struct thread *td;
865	struct tdq *tdq;
866	int cpu;
867
868	TDQ_LOCK_ASSERT(from, MA_OWNED);
869	TDQ_LOCK_ASSERT(to, MA_OWNED);
870
871	tdq = from;
872	cpu = TDQ_ID(to);
873	td = tdq_steal(tdq, cpu);
874	if (td == NULL)
875		return (0);
876	ts = td->td_sched;
877	/*
878	 * Although the run queue is locked the thread may be blocked.  Lock
879	 * it to clear this and acquire the run-queue lock.
880	 */
881	thread_lock(td);
882	/* Drop recursive lock on from acquired via thread_lock(). */
883	TDQ_UNLOCK(from);
884	sched_rem(td);
885	ts->ts_cpu = cpu;
886	td->td_lock = TDQ_LOCKPTR(to);
887	tdq_add(to, td, SRQ_YIELDING);
888	return (1);
889}
890
891/*
892 * This tdq has idled.  Try to steal a thread from another cpu and switch
893 * to it.
894 */
895static int
896tdq_idled(struct tdq *tdq)
897{
898	struct cpu_group *cg;
899	struct tdq *steal;
900	cpuset_t mask;
901	int thresh;
902	int cpu;
903
904	if (smp_started == 0 || steal_idle == 0)
905		return (1);
906	CPU_FILL(&mask);
907	CPU_CLR(PCPU_GET(cpuid), &mask);
908	/* We don't want to be preempted while we're iterating. */
909	spinlock_enter();
910	for (cg = tdq->tdq_cg; cg != NULL; ) {
911		if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
912			thresh = steal_thresh;
913		else
914			thresh = 1;
915		cpu = sched_highest(cg, mask, thresh);
916		if (cpu == -1) {
917			cg = cg->cg_parent;
918			continue;
919		}
920		steal = TDQ_CPU(cpu);
921		CPU_CLR(cpu, &mask);
922		tdq_lock_pair(tdq, steal);
923		if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
924			tdq_unlock_pair(tdq, steal);
925			continue;
926		}
927		/*
928		 * If a thread was added while interrupts were disabled don't
929		 * steal one here.  If we fail to acquire one due to affinity
930		 * restrictions loop again with this cpu removed from the
931		 * set.
932		 */
933		if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
934			tdq_unlock_pair(tdq, steal);
935			continue;
936		}
937		spinlock_exit();
938		TDQ_UNLOCK(steal);
939		mi_switch(SW_VOL | SWT_IDLE, NULL);
940		thread_unlock(curthread);
941
942		return (0);
943	}
944	spinlock_exit();
945	return (1);
946}
947
948/*
949 * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
950 */
951static void
952tdq_notify(struct tdq *tdq, struct thread *td)
953{
954	struct thread *ctd;
955	int pri;
956	int cpu;
957
958	if (tdq->tdq_ipipending)
959		return;
960	cpu = td->td_sched->ts_cpu;
961	pri = td->td_priority;
962	ctd = pcpu_find(cpu)->pc_curthread;
963	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
964		return;
965	if (TD_IS_IDLETHREAD(ctd)) {
966		/*
967		 * If the MD code has an idle wakeup routine try that before
968		 * falling back to IPI.
969		 */
970		if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
971			return;
972	}
973	tdq->tdq_ipipending = 1;
974	ipi_cpu(cpu, IPI_PREEMPT);
975}
976
977/*
978 * Steals load from a timeshare queue.  Honors the rotating queue head
979 * index.
980 */
981static struct thread *
982runq_steal_from(struct runq *rq, int cpu, u_char start)
983{
984	struct rqbits *rqb;
985	struct rqhead *rqh;
986	struct thread *td;
987	int first;
988	int bit;
989	int pri;
990	int i;
991
992	rqb = &rq->rq_status;
993	bit = start & (RQB_BPW -1);
994	pri = 0;
995	first = 0;
996again:
997	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
998		if (rqb->rqb_bits[i] == 0)
999			continue;
1000		if (bit != 0) {
1001			for (pri = bit; pri < RQB_BPW; pri++)
1002				if (rqb->rqb_bits[i] & (1ul << pri))
1003					break;
1004			if (pri >= RQB_BPW)
1005				continue;
1006		} else
1007			pri = RQB_FFS(rqb->rqb_bits[i]);
1008		pri += (i << RQB_L2BPW);
1009		rqh = &rq->rq_queues[pri];
1010		TAILQ_FOREACH(td, rqh, td_runq) {
1011			if (first && THREAD_CAN_MIGRATE(td) &&
1012			    THREAD_CAN_SCHED(td, cpu))
1013				return (td);
1014			first = 1;
1015		}
1016	}
1017	if (start != 0) {
1018		start = 0;
1019		goto again;
1020	}
1021
1022	return (NULL);
1023}
1024
1025/*
1026 * Steals load from a standard linear queue.
1027 */
1028static struct thread *
1029runq_steal(struct runq *rq, int cpu)
1030{
1031	struct rqhead *rqh;
1032	struct rqbits *rqb;
1033	struct thread *td;
1034	int word;
1035	int bit;
1036
1037	rqb = &rq->rq_status;
1038	for (word = 0; word < RQB_LEN; word++) {
1039		if (rqb->rqb_bits[word] == 0)
1040			continue;
1041		for (bit = 0; bit < RQB_BPW; bit++) {
1042			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1043				continue;
1044			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1045			TAILQ_FOREACH(td, rqh, td_runq)
1046				if (THREAD_CAN_MIGRATE(td) &&
1047				    THREAD_CAN_SCHED(td, cpu))
1048					return (td);
1049		}
1050	}
1051	return (NULL);
1052}
1053
1054/*
1055 * Attempt to steal a thread in priority order from a thread queue.
1056 */
1057static struct thread *
1058tdq_steal(struct tdq *tdq, int cpu)
1059{
1060	struct thread *td;
1061
1062	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1063	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1064		return (td);
1065	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1066	    cpu, tdq->tdq_ridx)) != NULL)
1067		return (td);
1068	return (runq_steal(&tdq->tdq_idle, cpu));
1069}
1070
1071/*
1072 * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1073 * current lock and returns with the assigned queue locked.
1074 */
1075static inline struct tdq *
1076sched_setcpu(struct thread *td, int cpu, int flags)
1077{
1078
1079	struct tdq *tdq;
1080
1081	THREAD_LOCK_ASSERT(td, MA_OWNED);
1082	tdq = TDQ_CPU(cpu);
1083	td->td_sched->ts_cpu = cpu;
1084	/*
1085	 * If the lock matches just return the queue.
1086	 */
1087	if (td->td_lock == TDQ_LOCKPTR(tdq))
1088		return (tdq);
1089#ifdef notyet
1090	/*
1091	 * If the thread isn't running its lockptr is a
1092	 * turnstile or a sleepqueue.  We can just lock_set without
1093	 * blocking.
1094	 */
1095	if (TD_CAN_RUN(td)) {
1096		TDQ_LOCK(tdq);
1097		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1098		return (tdq);
1099	}
1100#endif
1101	/*
1102	 * The hard case, migration, we need to block the thread first to
1103	 * prevent order reversals with other cpus locks.
1104	 */
1105	spinlock_enter();
1106	thread_lock_block(td);
1107	TDQ_LOCK(tdq);
1108	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1109	spinlock_exit();
1110	return (tdq);
1111}
1112
1113SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1114SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1115SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1116SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1117SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1118SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1119
1120static int
1121sched_pickcpu(struct thread *td, int flags)
1122{
1123	struct cpu_group *cg;
1124	struct td_sched *ts;
1125	struct tdq *tdq;
1126	cpuset_t mask;
1127	int self;
1128	int pri;
1129	int cpu;
1130
1131	self = PCPU_GET(cpuid);
1132	ts = td->td_sched;
1133	if (smp_started == 0)
1134		return (self);
1135	/*
1136	 * Don't migrate a running thread from sched_switch().
1137	 */
1138	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1139		return (ts->ts_cpu);
1140	/*
1141	 * Prefer to run interrupt threads on the processors that generate
1142	 * the interrupt.
1143	 */
1144	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1145	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1146		SCHED_STAT_INC(pickcpu_intrbind);
1147		ts->ts_cpu = self;
1148	}
1149	/*
1150	 * If the thread can run on the last cpu and the affinity has not
1151	 * expired or it is idle run it there.
1152	 */
1153	pri = td->td_priority;
1154	tdq = TDQ_CPU(ts->ts_cpu);
1155	if (THREAD_CAN_SCHED(td, ts->ts_cpu)) {
1156		if (tdq->tdq_lowpri > PRI_MIN_IDLE) {
1157			SCHED_STAT_INC(pickcpu_idle_affinity);
1158			return (ts->ts_cpu);
1159		}
1160		if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri) {
1161			SCHED_STAT_INC(pickcpu_affinity);
1162			return (ts->ts_cpu);
1163		}
1164	}
1165	/*
1166	 * Search for the highest level in the tree that still has affinity.
1167	 */
1168	cg = NULL;
1169	for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent)
1170		if (SCHED_AFFINITY(ts, cg->cg_level))
1171			break;
1172	cpu = -1;
1173	mask = td->td_cpuset->cs_mask;
1174	if (cg)
1175		cpu = sched_lowest(cg, mask, pri);
1176	if (cpu == -1)
1177		cpu = sched_lowest(cpu_top, mask, -1);
1178	/*
1179	 * Compare the lowest loaded cpu to current cpu.
1180	 */
1181	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1182	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) {
1183		SCHED_STAT_INC(pickcpu_local);
1184		cpu = self;
1185	} else
1186		SCHED_STAT_INC(pickcpu_lowest);
1187	if (cpu != ts->ts_cpu)
1188		SCHED_STAT_INC(pickcpu_migration);
1189	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1190	return (cpu);
1191}
1192#endif
1193
1194/*
1195 * Pick the highest priority task we have and return it.
1196 */
1197static struct thread *
1198tdq_choose(struct tdq *tdq)
1199{
1200	struct thread *td;
1201
1202	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1203	td = runq_choose(&tdq->tdq_realtime);
1204	if (td != NULL)
1205		return (td);
1206	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1207	if (td != NULL) {
1208		KASSERT(td->td_priority >= PRI_MIN_TIMESHARE,
1209		    ("tdq_choose: Invalid priority on timeshare queue %d",
1210		    td->td_priority));
1211		return (td);
1212	}
1213	td = runq_choose(&tdq->tdq_idle);
1214	if (td != NULL) {
1215		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1216		    ("tdq_choose: Invalid priority on idle queue %d",
1217		    td->td_priority));
1218		return (td);
1219	}
1220
1221	return (NULL);
1222}
1223
1224/*
1225 * Initialize a thread queue.
1226 */
1227static void
1228tdq_setup(struct tdq *tdq)
1229{
1230
1231	if (bootverbose)
1232		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1233	runq_init(&tdq->tdq_realtime);
1234	runq_init(&tdq->tdq_timeshare);
1235	runq_init(&tdq->tdq_idle);
1236	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1237	    "sched lock %d", (int)TDQ_ID(tdq));
1238	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1239	    MTX_SPIN | MTX_RECURSE);
1240#ifdef KTR
1241	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1242	    "CPU %d load", (int)TDQ_ID(tdq));
1243#endif
1244}
1245
1246#ifdef SMP
1247static void
1248sched_setup_smp(void)
1249{
1250	struct tdq *tdq;
1251	int i;
1252
1253	cpu_top = smp_topo();
1254	CPU_FOREACH(i) {
1255		tdq = TDQ_CPU(i);
1256		tdq_setup(tdq);
1257		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1258		if (tdq->tdq_cg == NULL)
1259			panic("Can't find cpu group for %d\n", i);
1260	}
1261	balance_tdq = TDQ_SELF();
1262	sched_balance();
1263}
1264#endif
1265
1266/*
1267 * Setup the thread queues and initialize the topology based on MD
1268 * information.
1269 */
1270static void
1271sched_setup(void *dummy)
1272{
1273	struct tdq *tdq;
1274
1275	tdq = TDQ_SELF();
1276#ifdef SMP
1277	sched_setup_smp();
1278#else
1279	tdq_setup(tdq);
1280#endif
1281	/*
1282	 * To avoid divide-by-zero, we set realstathz a dummy value
1283	 * in case which sched_clock() called before sched_initticks().
1284	 */
1285	realstathz = hz;
1286	sched_slice = (realstathz/10);	/* ~100ms */
1287	tickincr = 1 << SCHED_TICK_SHIFT;
1288
1289	/* Add thread0's load since it's running. */
1290	TDQ_LOCK(tdq);
1291	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1292	tdq_load_add(tdq, &thread0);
1293	tdq->tdq_lowpri = thread0.td_priority;
1294	TDQ_UNLOCK(tdq);
1295}
1296
1297/*
1298 * This routine determines the tickincr after stathz and hz are setup.
1299 */
1300/* ARGSUSED */
1301static void
1302sched_initticks(void *dummy)
1303{
1304	int incr;
1305
1306	realstathz = stathz ? stathz : hz;
1307	sched_slice = (realstathz/10);	/* ~100ms */
1308
1309	/*
1310	 * tickincr is shifted out by 10 to avoid rounding errors due to
1311	 * hz not being evenly divisible by stathz on all platforms.
1312	 */
1313	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1314	/*
1315	 * This does not work for values of stathz that are more than
1316	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1317	 */
1318	if (incr == 0)
1319		incr = 1;
1320	tickincr = incr;
1321#ifdef SMP
1322	/*
1323	 * Set the default balance interval now that we know
1324	 * what realstathz is.
1325	 */
1326	balance_interval = realstathz;
1327	/*
1328	 * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4.
1329	 * This prevents excess thrashing on large machines and excess idle
1330	 * on smaller machines.
1331	 */
1332	steal_thresh = min(fls(mp_ncpus) - 1, 3);
1333	affinity = SCHED_AFFINITY_DEFAULT;
1334#endif
1335}
1336
1337
1338/*
1339 * This is the core of the interactivity algorithm.  Determines a score based
1340 * on past behavior.  It is the ratio of sleep time to run time scaled to
1341 * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1342 * differs from the cpu usage because it does not account for time spent
1343 * waiting on a run-queue.  Would be prettier if we had floating point.
1344 */
1345static int
1346sched_interact_score(struct thread *td)
1347{
1348	struct td_sched *ts;
1349	int div;
1350
1351	ts = td->td_sched;
1352	/*
1353	 * The score is only needed if this is likely to be an interactive
1354	 * task.  Don't go through the expense of computing it if there's
1355	 * no chance.
1356	 */
1357	if (sched_interact <= SCHED_INTERACT_HALF &&
1358		ts->ts_runtime >= ts->ts_slptime)
1359			return (SCHED_INTERACT_HALF);
1360
1361	if (ts->ts_runtime > ts->ts_slptime) {
1362		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1363		return (SCHED_INTERACT_HALF +
1364		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1365	}
1366	if (ts->ts_slptime > ts->ts_runtime) {
1367		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1368		return (ts->ts_runtime / div);
1369	}
1370	/* runtime == slptime */
1371	if (ts->ts_runtime)
1372		return (SCHED_INTERACT_HALF);
1373
1374	/*
1375	 * This can happen if slptime and runtime are 0.
1376	 */
1377	return (0);
1378
1379}
1380
1381/*
1382 * Scale the scheduling priority according to the "interactivity" of this
1383 * process.
1384 */
1385static void
1386sched_priority(struct thread *td)
1387{
1388	int score;
1389	int pri;
1390
1391	if (td->td_pri_class != PRI_TIMESHARE)
1392		return;
1393	/*
1394	 * If the score is interactive we place the thread in the realtime
1395	 * queue with a priority that is less than kernel and interrupt
1396	 * priorities.  These threads are not subject to nice restrictions.
1397	 *
1398	 * Scores greater than this are placed on the normal timeshare queue
1399	 * where the priority is partially decided by the most recent cpu
1400	 * utilization and the rest is decided by nice value.
1401	 *
1402	 * The nice value of the process has a linear effect on the calculated
1403	 * score.  Negative nice values make it easier for a thread to be
1404	 * considered interactive.
1405	 */
1406	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1407	if (score < sched_interact) {
1408		pri = PRI_MIN_REALTIME;
1409		pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact)
1410		    * score;
1411		KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME,
1412		    ("sched_priority: invalid interactive priority %d score %d",
1413		    pri, score));
1414	} else {
1415		pri = SCHED_PRI_MIN;
1416		if (td->td_sched->ts_ticks)
1417			pri += SCHED_PRI_TICKS(td->td_sched);
1418		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1419		KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE,
1420		    ("sched_priority: invalid priority %d: nice %d, "
1421		    "ticks %d ftick %d ltick %d tick pri %d",
1422		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1423		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1424		    SCHED_PRI_TICKS(td->td_sched)));
1425	}
1426	sched_user_prio(td, pri);
1427
1428	return;
1429}
1430
1431/*
1432 * This routine enforces a maximum limit on the amount of scheduling history
1433 * kept.  It is called after either the slptime or runtime is adjusted.  This
1434 * function is ugly due to integer math.
1435 */
1436static void
1437sched_interact_update(struct thread *td)
1438{
1439	struct td_sched *ts;
1440	u_int sum;
1441
1442	ts = td->td_sched;
1443	sum = ts->ts_runtime + ts->ts_slptime;
1444	if (sum < SCHED_SLP_RUN_MAX)
1445		return;
1446	/*
1447	 * This only happens from two places:
1448	 * 1) We have added an unusual amount of run time from fork_exit.
1449	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1450	 */
1451	if (sum > SCHED_SLP_RUN_MAX * 2) {
1452		if (ts->ts_runtime > ts->ts_slptime) {
1453			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1454			ts->ts_slptime = 1;
1455		} else {
1456			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1457			ts->ts_runtime = 1;
1458		}
1459		return;
1460	}
1461	/*
1462	 * If we have exceeded by more than 1/5th then the algorithm below
1463	 * will not bring us back into range.  Dividing by two here forces
1464	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1465	 */
1466	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1467		ts->ts_runtime /= 2;
1468		ts->ts_slptime /= 2;
1469		return;
1470	}
1471	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1472	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1473}
1474
1475/*
1476 * Scale back the interactivity history when a child thread is created.  The
1477 * history is inherited from the parent but the thread may behave totally
1478 * differently.  For example, a shell spawning a compiler process.  We want
1479 * to learn that the compiler is behaving badly very quickly.
1480 */
1481static void
1482sched_interact_fork(struct thread *td)
1483{
1484	int ratio;
1485	int sum;
1486
1487	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1488	if (sum > SCHED_SLP_RUN_FORK) {
1489		ratio = sum / SCHED_SLP_RUN_FORK;
1490		td->td_sched->ts_runtime /= ratio;
1491		td->td_sched->ts_slptime /= ratio;
1492	}
1493}
1494
1495/*
1496 * Called from proc0_init() to setup the scheduler fields.
1497 */
1498void
1499schedinit(void)
1500{
1501
1502	/*
1503	 * Set up the scheduler specific parts of proc0.
1504	 */
1505	proc0.p_sched = NULL; /* XXX */
1506	thread0.td_sched = &td_sched0;
1507	td_sched0.ts_ltick = ticks;
1508	td_sched0.ts_ftick = ticks;
1509	td_sched0.ts_slice = sched_slice;
1510}
1511
1512/*
1513 * This is only somewhat accurate since given many processes of the same
1514 * priority they will switch when their slices run out, which will be
1515 * at most sched_slice stathz ticks.
1516 */
1517int
1518sched_rr_interval(void)
1519{
1520
1521	/* Convert sched_slice to hz */
1522	return (hz/(realstathz/sched_slice));
1523}
1524
1525/*
1526 * Update the percent cpu tracking information when it is requested or
1527 * the total history exceeds the maximum.  We keep a sliding history of
1528 * tick counts that slowly decays.  This is less precise than the 4BSD
1529 * mechanism since it happens with less regular and frequent events.
1530 */
1531static void
1532sched_pctcpu_update(struct td_sched *ts)
1533{
1534
1535	if (ts->ts_ticks == 0)
1536		return;
1537	if (ticks - (hz / 10) < ts->ts_ltick &&
1538	    SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX)
1539		return;
1540	/*
1541	 * Adjust counters and watermark for pctcpu calc.
1542	 */
1543	if (ts->ts_ltick > ticks - SCHED_TICK_TARG)
1544		ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) *
1545			    SCHED_TICK_TARG;
1546	else
1547		ts->ts_ticks = 0;
1548	ts->ts_ltick = ticks;
1549	ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG;
1550}
1551
1552/*
1553 * Adjust the priority of a thread.  Move it to the appropriate run-queue
1554 * if necessary.  This is the back-end for several priority related
1555 * functions.
1556 */
1557static void
1558sched_thread_priority(struct thread *td, u_char prio)
1559{
1560	struct td_sched *ts;
1561	struct tdq *tdq;
1562	int oldpri;
1563
1564	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1565	    "prio:%d", td->td_priority, "new prio:%d", prio,
1566	    KTR_ATTR_LINKED, sched_tdname(curthread));
1567	if (td != curthread && prio > td->td_priority) {
1568		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1569		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1570		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1571	}
1572	ts = td->td_sched;
1573	THREAD_LOCK_ASSERT(td, MA_OWNED);
1574	if (td->td_priority == prio)
1575		return;
1576	/*
1577	 * If the priority has been elevated due to priority
1578	 * propagation, we may have to move ourselves to a new
1579	 * queue.  This could be optimized to not re-add in some
1580	 * cases.
1581	 */
1582	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1583		sched_rem(td);
1584		td->td_priority = prio;
1585		sched_add(td, SRQ_BORROWING);
1586		return;
1587	}
1588	/*
1589	 * If the thread is currently running we may have to adjust the lowpri
1590	 * information so other cpus are aware of our current priority.
1591	 */
1592	if (TD_IS_RUNNING(td)) {
1593		tdq = TDQ_CPU(ts->ts_cpu);
1594		oldpri = td->td_priority;
1595		td->td_priority = prio;
1596		if (prio < tdq->tdq_lowpri)
1597			tdq->tdq_lowpri = prio;
1598		else if (tdq->tdq_lowpri == oldpri)
1599			tdq_setlowpri(tdq, td);
1600		return;
1601	}
1602	td->td_priority = prio;
1603}
1604
1605/*
1606 * Update a thread's priority when it is lent another thread's
1607 * priority.
1608 */
1609void
1610sched_lend_prio(struct thread *td, u_char prio)
1611{
1612
1613	td->td_flags |= TDF_BORROWING;
1614	sched_thread_priority(td, prio);
1615}
1616
1617/*
1618 * Restore a thread's priority when priority propagation is
1619 * over.  The prio argument is the minimum priority the thread
1620 * needs to have to satisfy other possible priority lending
1621 * requests.  If the thread's regular priority is less
1622 * important than prio, the thread will keep a priority boost
1623 * of prio.
1624 */
1625void
1626sched_unlend_prio(struct thread *td, u_char prio)
1627{
1628	u_char base_pri;
1629
1630	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1631	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1632		base_pri = td->td_user_pri;
1633	else
1634		base_pri = td->td_base_pri;
1635	if (prio >= base_pri) {
1636		td->td_flags &= ~TDF_BORROWING;
1637		sched_thread_priority(td, base_pri);
1638	} else
1639		sched_lend_prio(td, prio);
1640}
1641
1642/*
1643 * Standard entry for setting the priority to an absolute value.
1644 */
1645void
1646sched_prio(struct thread *td, u_char prio)
1647{
1648	u_char oldprio;
1649
1650	/* First, update the base priority. */
1651	td->td_base_pri = prio;
1652
1653	/*
1654	 * If the thread is borrowing another thread's priority, don't
1655	 * ever lower the priority.
1656	 */
1657	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1658		return;
1659
1660	/* Change the real priority. */
1661	oldprio = td->td_priority;
1662	sched_thread_priority(td, prio);
1663
1664	/*
1665	 * If the thread is on a turnstile, then let the turnstile update
1666	 * its state.
1667	 */
1668	if (TD_ON_LOCK(td) && oldprio != prio)
1669		turnstile_adjust(td, oldprio);
1670}
1671
1672/*
1673 * Set the base user priority, does not effect current running priority.
1674 */
1675void
1676sched_user_prio(struct thread *td, u_char prio)
1677{
1678
1679	td->td_base_user_pri = prio;
1680	if (td->td_lend_user_pri <= prio)
1681		return;
1682	td->td_user_pri = prio;
1683}
1684
1685void
1686sched_lend_user_prio(struct thread *td, u_char prio)
1687{
1688
1689	THREAD_LOCK_ASSERT(td, MA_OWNED);
1690	if (prio < td->td_lend_user_pri)
1691		td->td_lend_user_pri = prio;
1692	if (prio < td->td_user_pri)
1693		td->td_user_pri = prio;
1694}
1695
1696void
1697sched_unlend_user_prio(struct thread *td, u_char prio)
1698{
1699	u_char base_pri;
1700
1701	THREAD_LOCK_ASSERT(td, MA_OWNED);
1702	base_pri = td->td_base_user_pri;
1703	td->td_lend_user_pri = prio;
1704	if (prio > base_pri)
1705		td->td_user_pri = base_pri;
1706	else
1707		td->td_user_pri = prio;
1708}
1709
1710/*
1711 * Handle migration from sched_switch().  This happens only for
1712 * cpu binding.
1713 */
1714static struct mtx *
1715sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1716{
1717	struct tdq *tdn;
1718
1719	tdn = TDQ_CPU(td->td_sched->ts_cpu);
1720#ifdef SMP
1721	tdq_load_rem(tdq, td);
1722	/*
1723	 * Do the lock dance required to avoid LOR.  We grab an extra
1724	 * spinlock nesting to prevent preemption while we're
1725	 * not holding either run-queue lock.
1726	 */
1727	spinlock_enter();
1728	thread_lock_block(td);	/* This releases the lock on tdq. */
1729
1730	/*
1731	 * Acquire both run-queue locks before placing the thread on the new
1732	 * run-queue to avoid deadlocks created by placing a thread with a
1733	 * blocked lock on the run-queue of a remote processor.  The deadlock
1734	 * occurs when a third processor attempts to lock the two queues in
1735	 * question while the target processor is spinning with its own
1736	 * run-queue lock held while waiting for the blocked lock to clear.
1737	 */
1738	tdq_lock_pair(tdn, tdq);
1739	tdq_add(tdn, td, flags);
1740	tdq_notify(tdn, td);
1741	TDQ_UNLOCK(tdn);
1742	spinlock_exit();
1743#endif
1744	return (TDQ_LOCKPTR(tdn));
1745}
1746
1747/*
1748 * Variadic version of thread_lock_unblock() that does not assume td_lock
1749 * is blocked.
1750 */
1751static inline void
1752thread_unblock_switch(struct thread *td, struct mtx *mtx)
1753{
1754	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1755	    (uintptr_t)mtx);
1756}
1757
1758/*
1759 * Switch threads.  This function has to handle threads coming in while
1760 * blocked for some reason, running, or idle.  It also must deal with
1761 * migrating a thread from one queue to another as running threads may
1762 * be assigned elsewhere via binding.
1763 */
1764void
1765sched_switch(struct thread *td, struct thread *newtd, int flags)
1766{
1767	struct tdq *tdq;
1768	struct td_sched *ts;
1769	struct mtx *mtx;
1770	int srqflag;
1771	int cpuid;
1772
1773	THREAD_LOCK_ASSERT(td, MA_OWNED);
1774	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1775
1776	cpuid = PCPU_GET(cpuid);
1777	tdq = TDQ_CPU(cpuid);
1778	ts = td->td_sched;
1779	mtx = td->td_lock;
1780	ts->ts_rltick = ticks;
1781	td->td_lastcpu = td->td_oncpu;
1782	td->td_oncpu = NOCPU;
1783	td->td_flags &= ~TDF_NEEDRESCHED;
1784	td->td_owepreempt = 0;
1785	tdq->tdq_switchcnt++;
1786	/*
1787	 * The lock pointer in an idle thread should never change.  Reset it
1788	 * to CAN_RUN as well.
1789	 */
1790	if (TD_IS_IDLETHREAD(td)) {
1791		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1792		TD_SET_CAN_RUN(td);
1793	} else if (TD_IS_RUNNING(td)) {
1794		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1795		srqflag = (flags & SW_PREEMPT) ?
1796		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1797		    SRQ_OURSELF|SRQ_YIELDING;
1798#ifdef SMP
1799		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1800			ts->ts_cpu = sched_pickcpu(td, 0);
1801#endif
1802		if (ts->ts_cpu == cpuid)
1803			tdq_runq_add(tdq, td, srqflag);
1804		else {
1805			KASSERT(THREAD_CAN_MIGRATE(td) ||
1806			    (ts->ts_flags & TSF_BOUND) != 0,
1807			    ("Thread %p shouldn't migrate", td));
1808			mtx = sched_switch_migrate(tdq, td, srqflag);
1809		}
1810	} else {
1811		/* This thread must be going to sleep. */
1812		TDQ_LOCK(tdq);
1813		mtx = thread_lock_block(td);
1814		tdq_load_rem(tdq, td);
1815	}
1816	/*
1817	 * We enter here with the thread blocked and assigned to the
1818	 * appropriate cpu run-queue or sleep-queue and with the current
1819	 * thread-queue locked.
1820	 */
1821	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1822	newtd = choosethread();
1823	/*
1824	 * Call the MD code to switch contexts if necessary.
1825	 */
1826	if (td != newtd) {
1827#ifdef	HWPMC_HOOKS
1828		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1829			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1830#endif
1831		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1832		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1833
1834#ifdef KDTRACE_HOOKS
1835		/*
1836		 * If DTrace has set the active vtime enum to anything
1837		 * other than INACTIVE (0), then it should have set the
1838		 * function to call.
1839		 */
1840		if (dtrace_vtime_active)
1841			(*dtrace_vtime_switch_func)(newtd);
1842#endif
1843
1844		cpu_switch(td, newtd, mtx);
1845		/*
1846		 * We may return from cpu_switch on a different cpu.  However,
1847		 * we always return with td_lock pointing to the current cpu's
1848		 * run queue lock.
1849		 */
1850		cpuid = PCPU_GET(cpuid);
1851		tdq = TDQ_CPU(cpuid);
1852		lock_profile_obtain_lock_success(
1853		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1854#ifdef	HWPMC_HOOKS
1855		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1856			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1857#endif
1858	} else
1859		thread_unblock_switch(td, mtx);
1860	/*
1861	 * Assert that all went well and return.
1862	 */
1863	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1864	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1865	td->td_oncpu = cpuid;
1866}
1867
1868/*
1869 * Adjust thread priorities as a result of a nice request.
1870 */
1871void
1872sched_nice(struct proc *p, int nice)
1873{
1874	struct thread *td;
1875
1876	PROC_LOCK_ASSERT(p, MA_OWNED);
1877
1878	p->p_nice = nice;
1879	FOREACH_THREAD_IN_PROC(p, td) {
1880		thread_lock(td);
1881		sched_priority(td);
1882		sched_prio(td, td->td_base_user_pri);
1883		thread_unlock(td);
1884	}
1885}
1886
1887/*
1888 * Record the sleep time for the interactivity scorer.
1889 */
1890void
1891sched_sleep(struct thread *td, int prio)
1892{
1893
1894	THREAD_LOCK_ASSERT(td, MA_OWNED);
1895
1896	td->td_slptick = ticks;
1897	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1898		td->td_flags |= TDF_CANSWAP;
1899	if (static_boost == 1 && prio)
1900		sched_prio(td, prio);
1901	else if (static_boost && td->td_priority > static_boost)
1902		sched_prio(td, static_boost);
1903}
1904
1905/*
1906 * Schedule a thread to resume execution and record how long it voluntarily
1907 * slept.  We also update the pctcpu, interactivity, and priority.
1908 */
1909void
1910sched_wakeup(struct thread *td)
1911{
1912	struct td_sched *ts;
1913	int slptick;
1914
1915	THREAD_LOCK_ASSERT(td, MA_OWNED);
1916	ts = td->td_sched;
1917	td->td_flags &= ~TDF_CANSWAP;
1918	/*
1919	 * If we slept for more than a tick update our interactivity and
1920	 * priority.
1921	 */
1922	slptick = td->td_slptick;
1923	td->td_slptick = 0;
1924	if (slptick && slptick != ticks) {
1925		u_int hzticks;
1926
1927		hzticks = (ticks - slptick) << SCHED_TICK_SHIFT;
1928		ts->ts_slptime += hzticks;
1929		sched_interact_update(td);
1930		sched_pctcpu_update(ts);
1931	}
1932	/* Reset the slice value after we sleep. */
1933	ts->ts_slice = sched_slice;
1934	sched_add(td, SRQ_BORING);
1935}
1936
1937/*
1938 * Penalize the parent for creating a new child and initialize the child's
1939 * priority.
1940 */
1941void
1942sched_fork(struct thread *td, struct thread *child)
1943{
1944	THREAD_LOCK_ASSERT(td, MA_OWNED);
1945	sched_fork_thread(td, child);
1946	/*
1947	 * Penalize the parent and child for forking.
1948	 */
1949	sched_interact_fork(child);
1950	sched_priority(child);
1951	td->td_sched->ts_runtime += tickincr;
1952	sched_interact_update(td);
1953	sched_priority(td);
1954}
1955
1956/*
1957 * Fork a new thread, may be within the same process.
1958 */
1959void
1960sched_fork_thread(struct thread *td, struct thread *child)
1961{
1962	struct td_sched *ts;
1963	struct td_sched *ts2;
1964
1965	THREAD_LOCK_ASSERT(td, MA_OWNED);
1966	/*
1967	 * Initialize child.
1968	 */
1969	ts = td->td_sched;
1970	ts2 = child->td_sched;
1971	child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
1972	child->td_cpuset = cpuset_ref(td->td_cpuset);
1973	ts2->ts_cpu = ts->ts_cpu;
1974	ts2->ts_flags = 0;
1975	/*
1976	 * Grab our parents cpu estimation information and priority.
1977	 */
1978	ts2->ts_ticks = ts->ts_ticks;
1979	ts2->ts_ltick = ts->ts_ltick;
1980	ts2->ts_incrtick = ts->ts_incrtick;
1981	ts2->ts_ftick = ts->ts_ftick;
1982	child->td_user_pri = td->td_user_pri;
1983	child->td_base_user_pri = td->td_base_user_pri;
1984	/*
1985	 * And update interactivity score.
1986	 */
1987	ts2->ts_slptime = ts->ts_slptime;
1988	ts2->ts_runtime = ts->ts_runtime;
1989	ts2->ts_slice = 1;	/* Attempt to quickly learn interactivity. */
1990#ifdef KTR
1991	bzero(ts2->ts_name, sizeof(ts2->ts_name));
1992#endif
1993}
1994
1995/*
1996 * Adjust the priority class of a thread.
1997 */
1998void
1999sched_class(struct thread *td, int class)
2000{
2001
2002	THREAD_LOCK_ASSERT(td, MA_OWNED);
2003	if (td->td_pri_class == class)
2004		return;
2005	td->td_pri_class = class;
2006}
2007
2008/*
2009 * Return some of the child's priority and interactivity to the parent.
2010 */
2011void
2012sched_exit(struct proc *p, struct thread *child)
2013{
2014	struct thread *td;
2015
2016	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2017	    "prio:td", child->td_priority);
2018	PROC_LOCK_ASSERT(p, MA_OWNED);
2019	td = FIRST_THREAD_IN_PROC(p);
2020	sched_exit_thread(td, child);
2021}
2022
2023/*
2024 * Penalize another thread for the time spent on this one.  This helps to
2025 * worsen the priority and interactivity of processes which schedule batch
2026 * jobs such as make.  This has little effect on the make process itself but
2027 * causes new processes spawned by it to receive worse scores immediately.
2028 */
2029void
2030sched_exit_thread(struct thread *td, struct thread *child)
2031{
2032
2033	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2034	    "prio:td", child->td_priority);
2035	/*
2036	 * Give the child's runtime to the parent without returning the
2037	 * sleep time as a penalty to the parent.  This causes shells that
2038	 * launch expensive things to mark their children as expensive.
2039	 */
2040	thread_lock(td);
2041	td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2042	sched_interact_update(td);
2043	sched_priority(td);
2044	thread_unlock(td);
2045}
2046
2047void
2048sched_preempt(struct thread *td)
2049{
2050	struct tdq *tdq;
2051
2052	thread_lock(td);
2053	tdq = TDQ_SELF();
2054	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2055	tdq->tdq_ipipending = 0;
2056	if (td->td_priority > tdq->tdq_lowpri) {
2057		int flags;
2058
2059		flags = SW_INVOL | SW_PREEMPT;
2060		if (td->td_critnest > 1)
2061			td->td_owepreempt = 1;
2062		else if (TD_IS_IDLETHREAD(td))
2063			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2064		else
2065			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2066	}
2067	thread_unlock(td);
2068}
2069
2070/*
2071 * Fix priorities on return to user-space.  Priorities may be elevated due
2072 * to static priorities in msleep() or similar.
2073 */
2074void
2075sched_userret(struct thread *td)
2076{
2077	/*
2078	 * XXX we cheat slightly on the locking here to avoid locking in
2079	 * the usual case.  Setting td_priority here is essentially an
2080	 * incomplete workaround for not setting it properly elsewhere.
2081	 * Now that some interrupt handlers are threads, not setting it
2082	 * properly elsewhere can clobber it in the window between setting
2083	 * it here and returning to user mode, so don't waste time setting
2084	 * it perfectly here.
2085	 */
2086	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2087	    ("thread with borrowed priority returning to userland"));
2088	if (td->td_priority != td->td_user_pri) {
2089		thread_lock(td);
2090		td->td_priority = td->td_user_pri;
2091		td->td_base_pri = td->td_user_pri;
2092		tdq_setlowpri(TDQ_SELF(), td);
2093		thread_unlock(td);
2094        }
2095}
2096
2097/*
2098 * Handle a stathz tick.  This is really only relevant for timeshare
2099 * threads.
2100 */
2101void
2102sched_clock(struct thread *td)
2103{
2104	struct tdq *tdq;
2105	struct td_sched *ts;
2106
2107	THREAD_LOCK_ASSERT(td, MA_OWNED);
2108	tdq = TDQ_SELF();
2109#ifdef SMP
2110	/*
2111	 * We run the long term load balancer infrequently on the first cpu.
2112	 */
2113	if (balance_tdq == tdq) {
2114		if (balance_ticks && --balance_ticks == 0)
2115			sched_balance();
2116	}
2117#endif
2118	/*
2119	 * Save the old switch count so we have a record of the last ticks
2120	 * activity.   Initialize the new switch count based on our load.
2121	 * If there is some activity seed it to reflect that.
2122	 */
2123	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2124	tdq->tdq_switchcnt = tdq->tdq_load;
2125	/*
2126	 * Advance the insert index once for each tick to ensure that all
2127	 * threads get a chance to run.
2128	 */
2129	if (tdq->tdq_idx == tdq->tdq_ridx) {
2130		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2131		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2132			tdq->tdq_ridx = tdq->tdq_idx;
2133	}
2134	ts = td->td_sched;
2135	if (td->td_pri_class & PRI_FIFO_BIT)
2136		return;
2137	if (td->td_pri_class == PRI_TIMESHARE) {
2138		/*
2139		 * We used a tick; charge it to the thread so
2140		 * that we can compute our interactivity.
2141		 */
2142		td->td_sched->ts_runtime += tickincr;
2143		sched_interact_update(td);
2144		sched_priority(td);
2145	}
2146	/*
2147	 * We used up one time slice.
2148	 */
2149	if (--ts->ts_slice > 0)
2150		return;
2151	/*
2152	 * We're out of time, force a requeue at userret().
2153	 */
2154	ts->ts_slice = sched_slice;
2155	td->td_flags |= TDF_NEEDRESCHED;
2156}
2157
2158/*
2159 * Called once per hz tick.  Used for cpu utilization information.  This
2160 * is easier than trying to scale based on stathz.
2161 */
2162void
2163sched_tick(int cnt)
2164{
2165	struct td_sched *ts;
2166
2167	ts = curthread->td_sched;
2168	/*
2169	 * Ticks is updated asynchronously on a single cpu.  Check here to
2170	 * avoid incrementing ts_ticks multiple times in a single tick.
2171	 */
2172	if (ts->ts_incrtick == ticks)
2173		return;
2174	/* Adjust ticks for pctcpu */
2175	ts->ts_ticks += cnt << SCHED_TICK_SHIFT;
2176	ts->ts_ltick = ticks;
2177	ts->ts_incrtick = ticks;
2178	/*
2179	 * Update if we've exceeded our desired tick threshold by over one
2180	 * second.
2181	 */
2182	if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick)
2183		sched_pctcpu_update(ts);
2184}
2185
2186/*
2187 * Return whether the current CPU has runnable tasks.  Used for in-kernel
2188 * cooperative idle threads.
2189 */
2190int
2191sched_runnable(void)
2192{
2193	struct tdq *tdq;
2194	int load;
2195
2196	load = 1;
2197
2198	tdq = TDQ_SELF();
2199	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2200		if (tdq->tdq_load > 0)
2201			goto out;
2202	} else
2203		if (tdq->tdq_load - 1 > 0)
2204			goto out;
2205	load = 0;
2206out:
2207	return (load);
2208}
2209
2210/*
2211 * Choose the highest priority thread to run.  The thread is removed from
2212 * the run-queue while running however the load remains.  For SMP we set
2213 * the tdq in the global idle bitmask if it idles here.
2214 */
2215struct thread *
2216sched_choose(void)
2217{
2218	struct thread *td;
2219	struct tdq *tdq;
2220
2221	tdq = TDQ_SELF();
2222	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2223	td = tdq_choose(tdq);
2224	if (td) {
2225		td->td_sched->ts_ltick = ticks;
2226		tdq_runq_rem(tdq, td);
2227		tdq->tdq_lowpri = td->td_priority;
2228		return (td);
2229	}
2230	tdq->tdq_lowpri = PRI_MAX_IDLE;
2231	return (PCPU_GET(idlethread));
2232}
2233
2234/*
2235 * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2236 * we always request it once we exit a critical section.
2237 */
2238static inline void
2239sched_setpreempt(struct thread *td)
2240{
2241	struct thread *ctd;
2242	int cpri;
2243	int pri;
2244
2245	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2246
2247	ctd = curthread;
2248	pri = td->td_priority;
2249	cpri = ctd->td_priority;
2250	if (pri < cpri)
2251		ctd->td_flags |= TDF_NEEDRESCHED;
2252	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2253		return;
2254	if (!sched_shouldpreempt(pri, cpri, 0))
2255		return;
2256	ctd->td_owepreempt = 1;
2257}
2258
2259/*
2260 * Add a thread to a thread queue.  Select the appropriate runq and add the
2261 * thread to it.  This is the internal function called when the tdq is
2262 * predetermined.
2263 */
2264void
2265tdq_add(struct tdq *tdq, struct thread *td, int flags)
2266{
2267
2268	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2269	KASSERT((td->td_inhibitors == 0),
2270	    ("sched_add: trying to run inhibited thread"));
2271	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2272	    ("sched_add: bad thread state"));
2273	KASSERT(td->td_flags & TDF_INMEM,
2274	    ("sched_add: thread swapped out"));
2275
2276	if (td->td_priority < tdq->tdq_lowpri)
2277		tdq->tdq_lowpri = td->td_priority;
2278	tdq_runq_add(tdq, td, flags);
2279	tdq_load_add(tdq, td);
2280}
2281
2282/*
2283 * Select the target thread queue and add a thread to it.  Request
2284 * preemption or IPI a remote processor if required.
2285 */
2286void
2287sched_add(struct thread *td, int flags)
2288{
2289	struct tdq *tdq;
2290#ifdef SMP
2291	int cpu;
2292#endif
2293
2294	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2295	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2296	    sched_tdname(curthread));
2297	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2298	    KTR_ATTR_LINKED, sched_tdname(td));
2299	THREAD_LOCK_ASSERT(td, MA_OWNED);
2300	/*
2301	 * Recalculate the priority before we select the target cpu or
2302	 * run-queue.
2303	 */
2304	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2305		sched_priority(td);
2306#ifdef SMP
2307	/*
2308	 * Pick the destination cpu and if it isn't ours transfer to the
2309	 * target cpu.
2310	 */
2311	cpu = sched_pickcpu(td, flags);
2312	tdq = sched_setcpu(td, cpu, flags);
2313	tdq_add(tdq, td, flags);
2314	if (cpu != PCPU_GET(cpuid)) {
2315		tdq_notify(tdq, td);
2316		return;
2317	}
2318#else
2319	tdq = TDQ_SELF();
2320	TDQ_LOCK(tdq);
2321	/*
2322	 * Now that the thread is moving to the run-queue, set the lock
2323	 * to the scheduler's lock.
2324	 */
2325	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2326	tdq_add(tdq, td, flags);
2327#endif
2328	if (!(flags & SRQ_YIELDING))
2329		sched_setpreempt(td);
2330}
2331
2332/*
2333 * Remove a thread from a run-queue without running it.  This is used
2334 * when we're stealing a thread from a remote queue.  Otherwise all threads
2335 * exit by calling sched_exit_thread() and sched_throw() themselves.
2336 */
2337void
2338sched_rem(struct thread *td)
2339{
2340	struct tdq *tdq;
2341
2342	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2343	    "prio:%d", td->td_priority);
2344	tdq = TDQ_CPU(td->td_sched->ts_cpu);
2345	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2346	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2347	KASSERT(TD_ON_RUNQ(td),
2348	    ("sched_rem: thread not on run queue"));
2349	tdq_runq_rem(tdq, td);
2350	tdq_load_rem(tdq, td);
2351	TD_SET_CAN_RUN(td);
2352	if (td->td_priority == tdq->tdq_lowpri)
2353		tdq_setlowpri(tdq, NULL);
2354}
2355
2356/*
2357 * Fetch cpu utilization information.  Updates on demand.
2358 */
2359fixpt_t
2360sched_pctcpu(struct thread *td)
2361{
2362	fixpt_t pctcpu;
2363	struct td_sched *ts;
2364
2365	pctcpu = 0;
2366	ts = td->td_sched;
2367	if (ts == NULL)
2368		return (0);
2369
2370	THREAD_LOCK_ASSERT(td, MA_OWNED);
2371	if (ts->ts_ticks) {
2372		int rtick;
2373
2374		sched_pctcpu_update(ts);
2375		/* How many rtick per second ? */
2376		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2377		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2378	}
2379
2380	return (pctcpu);
2381}
2382
2383/*
2384 * Enforce affinity settings for a thread.  Called after adjustments to
2385 * cpumask.
2386 */
2387void
2388sched_affinity(struct thread *td)
2389{
2390#ifdef SMP
2391	struct td_sched *ts;
2392
2393	THREAD_LOCK_ASSERT(td, MA_OWNED);
2394	ts = td->td_sched;
2395	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2396		return;
2397	if (TD_ON_RUNQ(td)) {
2398		sched_rem(td);
2399		sched_add(td, SRQ_BORING);
2400		return;
2401	}
2402	if (!TD_IS_RUNNING(td))
2403		return;
2404	/*
2405	 * Force a switch before returning to userspace.  If the
2406	 * target thread is not running locally send an ipi to force
2407	 * the issue.
2408	 */
2409	td->td_flags |= TDF_NEEDRESCHED;
2410	if (td != curthread)
2411		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2412#endif
2413}
2414
2415/*
2416 * Bind a thread to a target cpu.
2417 */
2418void
2419sched_bind(struct thread *td, int cpu)
2420{
2421	struct td_sched *ts;
2422
2423	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2424	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2425	ts = td->td_sched;
2426	if (ts->ts_flags & TSF_BOUND)
2427		sched_unbind(td);
2428	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2429	ts->ts_flags |= TSF_BOUND;
2430	sched_pin();
2431	if (PCPU_GET(cpuid) == cpu)
2432		return;
2433	ts->ts_cpu = cpu;
2434	/* When we return from mi_switch we'll be on the correct cpu. */
2435	mi_switch(SW_VOL, NULL);
2436}
2437
2438/*
2439 * Release a bound thread.
2440 */
2441void
2442sched_unbind(struct thread *td)
2443{
2444	struct td_sched *ts;
2445
2446	THREAD_LOCK_ASSERT(td, MA_OWNED);
2447	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2448	ts = td->td_sched;
2449	if ((ts->ts_flags & TSF_BOUND) == 0)
2450		return;
2451	ts->ts_flags &= ~TSF_BOUND;
2452	sched_unpin();
2453}
2454
2455int
2456sched_is_bound(struct thread *td)
2457{
2458	THREAD_LOCK_ASSERT(td, MA_OWNED);
2459	return (td->td_sched->ts_flags & TSF_BOUND);
2460}
2461
2462/*
2463 * Basic yield call.
2464 */
2465void
2466sched_relinquish(struct thread *td)
2467{
2468	thread_lock(td);
2469	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2470	thread_unlock(td);
2471}
2472
2473/*
2474 * Return the total system load.
2475 */
2476int
2477sched_load(void)
2478{
2479#ifdef SMP
2480	int total;
2481	int i;
2482
2483	total = 0;
2484	CPU_FOREACH(i)
2485		total += TDQ_CPU(i)->tdq_sysload;
2486	return (total);
2487#else
2488	return (TDQ_SELF()->tdq_sysload);
2489#endif
2490}
2491
2492int
2493sched_sizeof_proc(void)
2494{
2495	return (sizeof(struct proc));
2496}
2497
2498int
2499sched_sizeof_thread(void)
2500{
2501	return (sizeof(struct thread) + sizeof(struct td_sched));
2502}
2503
2504#ifdef SMP
2505#define	TDQ_IDLESPIN(tdq)						\
2506    ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2507#else
2508#define	TDQ_IDLESPIN(tdq)	1
2509#endif
2510
2511/*
2512 * The actual idle process.
2513 */
2514void
2515sched_idletd(void *dummy)
2516{
2517	struct thread *td;
2518	struct tdq *tdq;
2519	int switchcnt;
2520	int i;
2521
2522	mtx_assert(&Giant, MA_NOTOWNED);
2523	td = curthread;
2524	tdq = TDQ_SELF();
2525	for (;;) {
2526#ifdef SMP
2527		if (tdq_idled(tdq) == 0)
2528			continue;
2529#endif
2530		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2531		/*
2532		 * If we're switching very frequently, spin while checking
2533		 * for load rather than entering a low power state that
2534		 * may require an IPI.  However, don't do any busy
2535		 * loops while on SMT machines as this simply steals
2536		 * cycles from cores doing useful work.
2537		 */
2538		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2539			for (i = 0; i < sched_idlespins; i++) {
2540				if (tdq->tdq_load)
2541					break;
2542				cpu_spinwait();
2543			}
2544		}
2545		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2546		if (tdq->tdq_load == 0) {
2547			tdq->tdq_cpu_idle = 1;
2548			if (tdq->tdq_load == 0) {
2549				cpu_idle(switchcnt > sched_idlespinthresh * 4);
2550				tdq->tdq_switchcnt++;
2551			}
2552			tdq->tdq_cpu_idle = 0;
2553		}
2554		if (tdq->tdq_load) {
2555			thread_lock(td);
2556			mi_switch(SW_VOL | SWT_IDLE, NULL);
2557			thread_unlock(td);
2558		}
2559	}
2560}
2561
2562/*
2563 * A CPU is entering for the first time or a thread is exiting.
2564 */
2565void
2566sched_throw(struct thread *td)
2567{
2568	struct thread *newtd;
2569	struct tdq *tdq;
2570
2571	tdq = TDQ_SELF();
2572	if (td == NULL) {
2573		/* Correct spinlock nesting and acquire the correct lock. */
2574		TDQ_LOCK(tdq);
2575		spinlock_exit();
2576	} else {
2577		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2578		tdq_load_rem(tdq, td);
2579		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2580	}
2581	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2582	newtd = choosethread();
2583	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2584	PCPU_SET(switchtime, cpu_ticks());
2585	PCPU_SET(switchticks, ticks);
2586	cpu_throw(td, newtd);		/* doesn't return */
2587}
2588
2589/*
2590 * This is called from fork_exit().  Just acquire the correct locks and
2591 * let fork do the rest of the work.
2592 */
2593void
2594sched_fork_exit(struct thread *td)
2595{
2596	struct td_sched *ts;
2597	struct tdq *tdq;
2598	int cpuid;
2599
2600	/*
2601	 * Finish setting up thread glue so that it begins execution in a
2602	 * non-nested critical section with the scheduler lock held.
2603	 */
2604	cpuid = PCPU_GET(cpuid);
2605	tdq = TDQ_CPU(cpuid);
2606	ts = td->td_sched;
2607	if (TD_IS_IDLETHREAD(td))
2608		td->td_lock = TDQ_LOCKPTR(tdq);
2609	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2610	td->td_oncpu = cpuid;
2611	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2612	lock_profile_obtain_lock_success(
2613	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2614}
2615
2616/*
2617 * Create on first use to catch odd startup conditons.
2618 */
2619char *
2620sched_tdname(struct thread *td)
2621{
2622#ifdef KTR
2623	struct td_sched *ts;
2624
2625	ts = td->td_sched;
2626	if (ts->ts_name[0] == '\0')
2627		snprintf(ts->ts_name, sizeof(ts->ts_name),
2628		    "%s tid %d", td->td_name, td->td_tid);
2629	return (ts->ts_name);
2630#else
2631	return (td->td_name);
2632#endif
2633}
2634
2635#ifdef SMP
2636
2637/*
2638 * Build the CPU topology dump string. Is recursively called to collect
2639 * the topology tree.
2640 */
2641static int
2642sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2643    int indent)
2644{
2645	int i, first;
2646
2647	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2648	    "", 1 + indent / 2, cg->cg_level);
2649	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"0x%x\">", indent, "",
2650	    cg->cg_count, cg->cg_mask);
2651	first = TRUE;
2652	for (i = 0; i < MAXCPU; i++) {
2653		if ((cg->cg_mask & (1 << i)) != 0) {
2654			if (!first)
2655				sbuf_printf(sb, ", ");
2656			else
2657				first = FALSE;
2658			sbuf_printf(sb, "%d", i);
2659		}
2660	}
2661	sbuf_printf(sb, "</cpu>\n");
2662
2663	if (cg->cg_flags != 0) {
2664		sbuf_printf(sb, "%*s <flags>", indent, "");
2665		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2666			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2667		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2668			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2669		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2670			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2671		sbuf_printf(sb, "</flags>\n");
2672	}
2673
2674	if (cg->cg_children > 0) {
2675		sbuf_printf(sb, "%*s <children>\n", indent, "");
2676		for (i = 0; i < cg->cg_children; i++)
2677			sysctl_kern_sched_topology_spec_internal(sb,
2678			    &cg->cg_child[i], indent+2);
2679		sbuf_printf(sb, "%*s </children>\n", indent, "");
2680	}
2681	sbuf_printf(sb, "%*s</group>\n", indent, "");
2682	return (0);
2683}
2684
2685/*
2686 * Sysctl handler for retrieving topology dump. It's a wrapper for
2687 * the recursive sysctl_kern_smp_topology_spec_internal().
2688 */
2689static int
2690sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2691{
2692	struct sbuf *topo;
2693	int err;
2694
2695	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2696
2697	topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2698	if (topo == NULL)
2699		return (ENOMEM);
2700
2701	sbuf_printf(topo, "<groups>\n");
2702	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2703	sbuf_printf(topo, "</groups>\n");
2704
2705	if (err == 0) {
2706		sbuf_finish(topo);
2707		err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2708	}
2709	sbuf_delete(topo);
2710	return (err);
2711}
2712
2713#endif
2714
2715SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2716SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2717    "Scheduler name");
2718SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2719    "Slice size for timeshare threads");
2720SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2721     "Interactivity score threshold");
2722SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh,
2723     0,"Min priority for preemption, lower priorities have greater precedence");
2724SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost,
2725     0,"Controls whether static kernel priorities are assigned to sleeping threads.");
2726SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins,
2727     0,"Number of times idle will spin waiting for new work.");
2728SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, &sched_idlespinthresh,
2729     0,"Threshold before we will permit idle spinning.");
2730#ifdef SMP
2731SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2732    "Number of hz ticks to keep thread affinity for");
2733SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2734    "Enables the long-term load balancer");
2735SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2736    &balance_interval, 0,
2737    "Average frequency in stathz ticks to run the long-term balancer");
2738SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0,
2739    "Steals work from another hyper-threaded core on idle");
2740SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2741    "Attempts to steal work from other cores before idling");
2742SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2743    "Minimum load on remote cpu before we'll steal");
2744
2745/* Retrieve SMP topology */
2746SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2747    CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2748    "XML dump of detected CPU topology");
2749
2750#endif
2751
2752/* ps compat.  All cpu percentages from ULE are weighted. */
2753static int ccpu = 0;
2754SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2755