xref: /freebsd/sys/kern/sched_ule.c (revision 8b16c208e6714d3a0d028016e84c5f20243558d3)
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("$FreeBSD$");
40 
41 #include "opt_hwpmc_hooks.h"
42 #include "opt_sched.h"
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/kdb.h>
47 #include <sys/kernel.h>
48 #include <sys/ktr.h>
49 #include <sys/lock.h>
50 #include <sys/mutex.h>
51 #include <sys/proc.h>
52 #include <sys/resource.h>
53 #include <sys/resourcevar.h>
54 #include <sys/sched.h>
55 #include <sys/smp.h>
56 #include <sys/sx.h>
57 #include <sys/sysctl.h>
58 #include <sys/sysproto.h>
59 #include <sys/turnstile.h>
60 #include <sys/umtx.h>
61 #include <sys/vmmeter.h>
62 #include <sys/cpuset.h>
63 #ifdef KTRACE
64 #include <sys/uio.h>
65 #include <sys/ktrace.h>
66 #endif
67 
68 #ifdef HWPMC_HOOKS
69 #include <sys/pmckern.h>
70 #endif
71 
72 #include <machine/cpu.h>
73 #include <machine/smp.h>
74 
75 #if !defined(__i386__) && !defined(__amd64__) && !defined(__powerpc__) && !defined(__arm__)
76 #error "This architecture is not currently compatible with ULE"
77 #endif
78 
79 #define	KTR_ULE	0
80 
81 /*
82  * Thread scheduler specific section.  All fields are protected
83  * by the thread lock.
84  */
85 struct td_sched {
86 	TAILQ_ENTRY(td_sched) ts_procq;	/* Run queue. */
87 	struct thread	*ts_thread;	/* Active associated thread. */
88 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
89 	short		ts_flags;	/* TSF_* flags. */
90 	u_char		ts_rqindex;	/* Run queue index. */
91 	u_char		ts_cpu;		/* CPU that we have affinity for. */
92 	int		ts_rltick;	/* Real last tick, for affinity. */
93 	int		ts_slice;	/* Ticks of slice remaining. */
94 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
95 	u_int		ts_runtime;	/* Number of ticks we were running */
96 	int		ts_ltick;	/* Last tick that we were running on */
97 	int		ts_ftick;	/* First tick that we were running on */
98 	int		ts_ticks;	/* Tick count */
99 };
100 /* flags kept in ts_flags */
101 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
102 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
103 
104 static struct td_sched td_sched0;
105 
106 #define	THREAD_CAN_MIGRATE(td)	((td)->td_pinned == 0)
107 #define	THREAD_CAN_SCHED(td, cpu)	\
108     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
109 
110 /*
111  * Cpu percentage computation macros and defines.
112  *
113  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
114  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
115  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
116  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
117  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
118  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
119  */
120 #define	SCHED_TICK_SECS		10
121 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
122 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
123 #define	SCHED_TICK_SHIFT	10
124 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
125 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
126 
127 /*
128  * These macros determine priorities for non-interactive threads.  They are
129  * assigned a priority based on their recent cpu utilization as expressed
130  * by the ratio of ticks to the tick total.  NHALF priorities at the start
131  * and end of the MIN to MAX timeshare range are only reachable with negative
132  * or positive nice respectively.
133  *
134  * PRI_RANGE:	Priority range for utilization dependent priorities.
135  * PRI_NRESV:	Number of nice values.
136  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
137  * PRI_NICE:	Determines the part of the priority inherited from nice.
138  */
139 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
140 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
141 #define	SCHED_PRI_MIN		(PRI_MIN_TIMESHARE + SCHED_PRI_NHALF)
142 #define	SCHED_PRI_MAX		(PRI_MAX_TIMESHARE - SCHED_PRI_NHALF)
143 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN)
144 #define	SCHED_PRI_TICKS(ts)						\
145     (SCHED_TICK_HZ((ts)) /						\
146     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
147 #define	SCHED_PRI_NICE(nice)	(nice)
148 
149 /*
150  * These determine the interactivity of a process.  Interactivity differs from
151  * cpu utilization in that it expresses the voluntary time slept vs time ran
152  * while cpu utilization includes all time not running.  This more accurately
153  * models the intent of the thread.
154  *
155  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
156  *		before throttling back.
157  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
158  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
159  * INTERACT_THRESH:	Threshhold for placement on the current runq.
160  */
161 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
162 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
163 #define	SCHED_INTERACT_MAX	(100)
164 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
165 #define	SCHED_INTERACT_THRESH	(30)
166 
167 /*
168  * tickincr:		Converts a stathz tick into a hz domain scaled by
169  *			the shift factor.  Without the shift the error rate
170  *			due to rounding would be unacceptably high.
171  * realstathz:		stathz is sometimes 0 and run off of hz.
172  * sched_slice:		Runtime of each thread before rescheduling.
173  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
174  */
175 static int sched_interact = SCHED_INTERACT_THRESH;
176 static int realstathz;
177 static int tickincr;
178 static int sched_slice = 1;
179 #ifdef PREEMPTION
180 #ifdef FULL_PREEMPTION
181 static int preempt_thresh = PRI_MAX_IDLE;
182 #else
183 static int preempt_thresh = PRI_MIN_KERN;
184 #endif
185 #else
186 static int preempt_thresh = 0;
187 #endif
188 static int static_boost = 1;
189 
190 /*
191  * tdq - per processor runqs and statistics.  All fields are protected by the
192  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
193  * locking in sched_pickcpu();
194  */
195 struct tdq {
196 	/* Ordered to improve efficiency of cpu_search() and switch(). */
197 	struct mtx	tdq_lock;		/* run queue lock. */
198 	struct cpu_group *tdq_cg;		/* Pointer to cpu topology. */
199 	int		tdq_load;		/* Aggregate load. */
200 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
201 	int		tdq_transferable;	/* Transferable thread count. */
202 	u_char		tdq_lowpri;		/* Lowest priority thread. */
203 	u_char		tdq_ipipending;		/* IPI pending. */
204 	u_char		tdq_idx;		/* Current insert index. */
205 	u_char		tdq_ridx;		/* Current removal index. */
206 	struct runq	tdq_realtime;		/* real-time run queue. */
207 	struct runq	tdq_timeshare;		/* timeshare run queue. */
208 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
209 	char		tdq_name[sizeof("sched lock") + 6];
210 } __aligned(64);
211 
212 
213 #ifdef SMP
214 struct cpu_group *cpu_top;
215 
216 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
217 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
218 
219 /*
220  * Run-time tunables.
221  */
222 static int rebalance = 1;
223 static int balance_interval = 128;	/* Default set in sched_initticks(). */
224 static int affinity;
225 static int steal_htt = 1;
226 static int steal_idle = 1;
227 static int steal_thresh = 2;
228 
229 /*
230  * One thread queue per processor.
231  */
232 static struct tdq	tdq_cpu[MAXCPU];
233 static struct tdq	*balance_tdq;
234 static int balance_ticks;
235 
236 #define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
237 #define	TDQ_CPU(x)	(&tdq_cpu[(x)])
238 #define	TDQ_ID(x)	((int)((x) - tdq_cpu))
239 #else	/* !SMP */
240 static struct tdq	tdq_cpu;
241 
242 #define	TDQ_ID(x)	(0)
243 #define	TDQ_SELF()	(&tdq_cpu)
244 #define	TDQ_CPU(x)	(&tdq_cpu)
245 #endif
246 
247 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
248 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
249 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
250 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
251 #define	TDQ_LOCKPTR(t)		(&(t)->tdq_lock)
252 
253 static void sched_priority(struct thread *);
254 static void sched_thread_priority(struct thread *, u_char);
255 static int sched_interact_score(struct thread *);
256 static void sched_interact_update(struct thread *);
257 static void sched_interact_fork(struct thread *);
258 static void sched_pctcpu_update(struct td_sched *);
259 
260 /* Operations on per processor queues */
261 static struct td_sched * tdq_choose(struct tdq *);
262 static void tdq_setup(struct tdq *);
263 static void tdq_load_add(struct tdq *, struct td_sched *);
264 static void tdq_load_rem(struct tdq *, struct td_sched *);
265 static __inline void tdq_runq_add(struct tdq *, struct td_sched *, int);
266 static __inline void tdq_runq_rem(struct tdq *, struct td_sched *);
267 static inline int sched_shouldpreempt(int, int, int);
268 void tdq_print(int cpu);
269 static void runq_print(struct runq *rq);
270 static void tdq_add(struct tdq *, struct thread *, int);
271 #ifdef SMP
272 static int tdq_move(struct tdq *, struct tdq *);
273 static int tdq_idled(struct tdq *);
274 static void tdq_notify(struct tdq *, struct td_sched *);
275 static struct td_sched *tdq_steal(struct tdq *, int);
276 static struct td_sched *runq_steal(struct runq *, int);
277 static int sched_pickcpu(struct td_sched *, int);
278 static void sched_balance(void);
279 static int sched_balance_pair(struct tdq *, struct tdq *);
280 static inline struct tdq *sched_setcpu(struct td_sched *, int, int);
281 static inline struct mtx *thread_block_switch(struct thread *);
282 static inline void thread_unblock_switch(struct thread *, struct mtx *);
283 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
284 #endif
285 
286 static void sched_setup(void *dummy);
287 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
288 
289 static void sched_initticks(void *dummy);
290 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
291     NULL);
292 
293 /*
294  * Print the threads waiting on a run-queue.
295  */
296 static void
297 runq_print(struct runq *rq)
298 {
299 	struct rqhead *rqh;
300 	struct td_sched *ts;
301 	int pri;
302 	int j;
303 	int i;
304 
305 	for (i = 0; i < RQB_LEN; i++) {
306 		printf("\t\trunq bits %d 0x%zx\n",
307 		    i, rq->rq_status.rqb_bits[i]);
308 		for (j = 0; j < RQB_BPW; j++)
309 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
310 				pri = j + (i << RQB_L2BPW);
311 				rqh = &rq->rq_queues[pri];
312 				TAILQ_FOREACH(ts, rqh, ts_procq) {
313 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
314 					    ts->ts_thread, ts->ts_thread->td_name, ts->ts_thread->td_priority, ts->ts_rqindex, pri);
315 				}
316 			}
317 	}
318 }
319 
320 /*
321  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
322  */
323 void
324 tdq_print(int cpu)
325 {
326 	struct tdq *tdq;
327 
328 	tdq = TDQ_CPU(cpu);
329 
330 	printf("tdq %d:\n", TDQ_ID(tdq));
331 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
332 	printf("\tLock name:      %s\n", tdq->tdq_name);
333 	printf("\tload:           %d\n", tdq->tdq_load);
334 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
335 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
336 	printf("\trealtime runq:\n");
337 	runq_print(&tdq->tdq_realtime);
338 	printf("\ttimeshare runq:\n");
339 	runq_print(&tdq->tdq_timeshare);
340 	printf("\tidle runq:\n");
341 	runq_print(&tdq->tdq_idle);
342 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
343 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
344 }
345 
346 static inline int
347 sched_shouldpreempt(int pri, int cpri, int remote)
348 {
349 	/*
350 	 * If the new priority is not better than the current priority there is
351 	 * nothing to do.
352 	 */
353 	if (pri >= cpri)
354 		return (0);
355 	/*
356 	 * Always preempt idle.
357 	 */
358 	if (cpri >= PRI_MIN_IDLE)
359 		return (1);
360 	/*
361 	 * If preemption is disabled don't preempt others.
362 	 */
363 	if (preempt_thresh == 0)
364 		return (0);
365 	/*
366 	 * Preempt if we exceed the threshold.
367 	 */
368 	if (pri <= preempt_thresh)
369 		return (1);
370 	/*
371 	 * If we're realtime or better and there is timeshare or worse running
372 	 * preempt only remote processors.
373 	 */
374 	if (remote && pri <= PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME)
375 		return (1);
376 	return (0);
377 }
378 
379 #define	TS_RQ_PPQ	(((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS)
380 /*
381  * Add a thread to the actual run-queue.  Keeps transferable counts up to
382  * date with what is actually on the run-queue.  Selects the correct
383  * queue position for timeshare threads.
384  */
385 static __inline void
386 tdq_runq_add(struct tdq *tdq, struct td_sched *ts, int flags)
387 {
388 	u_char pri;
389 
390 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
391 	THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED);
392 
393 	TD_SET_RUNQ(ts->ts_thread);
394 	if (THREAD_CAN_MIGRATE(ts->ts_thread)) {
395 		tdq->tdq_transferable++;
396 		ts->ts_flags |= TSF_XFERABLE;
397 	}
398 	pri = ts->ts_thread->td_priority;
399 	if (pri <= PRI_MAX_REALTIME) {
400 		ts->ts_runq = &tdq->tdq_realtime;
401 	} else if (pri <= PRI_MAX_TIMESHARE) {
402 		ts->ts_runq = &tdq->tdq_timeshare;
403 		KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE,
404 			("Invalid priority %d on timeshare runq", pri));
405 		/*
406 		 * This queue contains only priorities between MIN and MAX
407 		 * realtime.  Use the whole queue to represent these values.
408 		 */
409 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
410 			pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ;
411 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
412 			/*
413 			 * This effectively shortens the queue by one so we
414 			 * can have a one slot difference between idx and
415 			 * ridx while we wait for threads to drain.
416 			 */
417 			if (tdq->tdq_ridx != tdq->tdq_idx &&
418 			    pri == tdq->tdq_ridx)
419 				pri = (unsigned char)(pri - 1) % RQ_NQS;
420 		} else
421 			pri = tdq->tdq_ridx;
422 		runq_add_pri(ts->ts_runq, ts, pri, flags);
423 		return;
424 	} else
425 		ts->ts_runq = &tdq->tdq_idle;
426 	runq_add(ts->ts_runq, ts, flags);
427 }
428 
429 /*
430  * Remove a thread from a run-queue.  This typically happens when a thread
431  * is selected to run.  Running threads are not on the queue and the
432  * transferable count does not reflect them.
433  */
434 static __inline void
435 tdq_runq_rem(struct tdq *tdq, struct td_sched *ts)
436 {
437 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
438 	KASSERT(ts->ts_runq != NULL,
439 	    ("tdq_runq_remove: thread %p null ts_runq", ts->ts_thread));
440 	if (ts->ts_flags & TSF_XFERABLE) {
441 		tdq->tdq_transferable--;
442 		ts->ts_flags &= ~TSF_XFERABLE;
443 	}
444 	if (ts->ts_runq == &tdq->tdq_timeshare) {
445 		if (tdq->tdq_idx != tdq->tdq_ridx)
446 			runq_remove_idx(ts->ts_runq, ts, &tdq->tdq_ridx);
447 		else
448 			runq_remove_idx(ts->ts_runq, ts, NULL);
449 	} else
450 		runq_remove(ts->ts_runq, ts);
451 }
452 
453 /*
454  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
455  * for this thread to the referenced thread queue.
456  */
457 static void
458 tdq_load_add(struct tdq *tdq, struct td_sched *ts)
459 {
460 	int class;
461 
462 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
463 	THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED);
464 	class = PRI_BASE(ts->ts_thread->td_pri_class);
465 	tdq->tdq_load++;
466 	CTR2(KTR_SCHED, "cpu %d load: %d", TDQ_ID(tdq), tdq->tdq_load);
467 	if (class != PRI_ITHD &&
468 	    (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0)
469 		tdq->tdq_sysload++;
470 }
471 
472 /*
473  * Remove the load from a thread that is transitioning to a sleep state or
474  * exiting.
475  */
476 static void
477 tdq_load_rem(struct tdq *tdq, struct td_sched *ts)
478 {
479 	int class;
480 
481 	THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED);
482 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
483 	class = PRI_BASE(ts->ts_thread->td_pri_class);
484 	if (class != PRI_ITHD &&
485 	    (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0)
486 		tdq->tdq_sysload--;
487 	KASSERT(tdq->tdq_load != 0,
488 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
489 	tdq->tdq_load--;
490 	CTR1(KTR_SCHED, "load: %d", tdq->tdq_load);
491 }
492 
493 /*
494  * Set lowpri to its exact value by searching the run-queue and
495  * evaluating curthread.  curthread may be passed as an optimization.
496  */
497 static void
498 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
499 {
500 	struct td_sched *ts;
501 	struct thread *td;
502 
503 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
504 	if (ctd == NULL)
505 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
506 	ts = tdq_choose(tdq);
507 	if (ts)
508 		td = ts->ts_thread;
509 	if (ts == NULL || td->td_priority > ctd->td_priority)
510 		tdq->tdq_lowpri = ctd->td_priority;
511 	else
512 		tdq->tdq_lowpri = td->td_priority;
513 }
514 
515 #ifdef SMP
516 struct cpu_search {
517 	cpumask_t cs_mask;	/* Mask of valid cpus. */
518 	u_int	cs_load;
519 	u_int	cs_cpu;
520 	int	cs_limit;	/* Min priority for low min load for high. */
521 };
522 
523 #define	CPU_SEARCH_LOWEST	0x1
524 #define	CPU_SEARCH_HIGHEST	0x2
525 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
526 
527 #define	CPUMASK_FOREACH(cpu, mask)				\
528 	for ((cpu) = 0; (cpu) < sizeof((mask)) * 8; (cpu)++)	\
529 		if ((mask) & 1 << (cpu))
530 
531 static __inline int cpu_search(struct cpu_group *cg, struct cpu_search *low,
532     struct cpu_search *high, const int match);
533 int cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low);
534 int cpu_search_highest(struct cpu_group *cg, struct cpu_search *high);
535 int cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
536     struct cpu_search *high);
537 
538 /*
539  * This routine compares according to the match argument and should be
540  * reduced in actual instantiations via constant propagation and dead code
541  * elimination.
542  */
543 static __inline int
544 cpu_compare(int cpu, struct cpu_search *low, struct cpu_search *high,
545     const int match)
546 {
547 	struct tdq *tdq;
548 
549 	tdq = TDQ_CPU(cpu);
550 	if (match & CPU_SEARCH_LOWEST)
551 		if (low->cs_mask & (1 << cpu) &&
552 		    tdq->tdq_load < low->cs_load &&
553 		    tdq->tdq_lowpri > low->cs_limit) {
554 			low->cs_cpu = cpu;
555 			low->cs_load = tdq->tdq_load;
556 		}
557 	if (match & CPU_SEARCH_HIGHEST)
558 		if (high->cs_mask & (1 << cpu) &&
559 		    tdq->tdq_load >= high->cs_limit &&
560 		    tdq->tdq_load > high->cs_load &&
561 		    tdq->tdq_transferable) {
562 			high->cs_cpu = cpu;
563 			high->cs_load = tdq->tdq_load;
564 		}
565 	return (tdq->tdq_load);
566 }
567 
568 /*
569  * Search the tree of cpu_groups for the lowest or highest loaded cpu
570  * according to the match argument.  This routine actually compares the
571  * load on all paths through the tree and finds the least loaded cpu on
572  * the least loaded path, which may differ from the least loaded cpu in
573  * the system.  This balances work among caches and busses.
574  *
575  * This inline is instantiated in three forms below using constants for the
576  * match argument.  It is reduced to the minimum set for each case.  It is
577  * also recursive to the depth of the tree.
578  */
579 static __inline int
580 cpu_search(struct cpu_group *cg, struct cpu_search *low,
581     struct cpu_search *high, const int match)
582 {
583 	int total;
584 
585 	total = 0;
586 	if (cg->cg_children) {
587 		struct cpu_search lgroup;
588 		struct cpu_search hgroup;
589 		struct cpu_group *child;
590 		u_int lload;
591 		int hload;
592 		int load;
593 		int i;
594 
595 		lload = -1;
596 		hload = -1;
597 		for (i = 0; i < cg->cg_children; i++) {
598 			child = &cg->cg_child[i];
599 			if (match & CPU_SEARCH_LOWEST) {
600 				lgroup = *low;
601 				lgroup.cs_load = -1;
602 			}
603 			if (match & CPU_SEARCH_HIGHEST) {
604 				hgroup = *high;
605 				lgroup.cs_load = 0;
606 			}
607 			switch (match) {
608 			case CPU_SEARCH_LOWEST:
609 				load = cpu_search_lowest(child, &lgroup);
610 				break;
611 			case CPU_SEARCH_HIGHEST:
612 				load = cpu_search_highest(child, &hgroup);
613 				break;
614 			case CPU_SEARCH_BOTH:
615 				load = cpu_search_both(child, &lgroup, &hgroup);
616 				break;
617 			}
618 			total += load;
619 			if (match & CPU_SEARCH_LOWEST)
620 				if (load < lload || low->cs_cpu == -1) {
621 					*low = lgroup;
622 					lload = load;
623 				}
624 			if (match & CPU_SEARCH_HIGHEST)
625 				if (load > hload || high->cs_cpu == -1) {
626 					hload = load;
627 					*high = hgroup;
628 				}
629 		}
630 	} else {
631 		int cpu;
632 
633 		CPUMASK_FOREACH(cpu, cg->cg_mask)
634 			total += cpu_compare(cpu, low, high, match);
635 	}
636 	return (total);
637 }
638 
639 /*
640  * cpu_search instantiations must pass constants to maintain the inline
641  * optimization.
642  */
643 int
644 cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low)
645 {
646 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
647 }
648 
649 int
650 cpu_search_highest(struct cpu_group *cg, struct cpu_search *high)
651 {
652 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
653 }
654 
655 int
656 cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
657     struct cpu_search *high)
658 {
659 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
660 }
661 
662 /*
663  * Find the cpu with the least load via the least loaded path that has a
664  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
665  * acceptable.
666  */
667 static inline int
668 sched_lowest(struct cpu_group *cg, cpumask_t mask, int pri)
669 {
670 	struct cpu_search low;
671 
672 	low.cs_cpu = -1;
673 	low.cs_load = -1;
674 	low.cs_mask = mask;
675 	low.cs_limit = pri;
676 	cpu_search_lowest(cg, &low);
677 	return low.cs_cpu;
678 }
679 
680 /*
681  * Find the cpu with the highest load via the highest loaded path.
682  */
683 static inline int
684 sched_highest(struct cpu_group *cg, cpumask_t mask, int minload)
685 {
686 	struct cpu_search high;
687 
688 	high.cs_cpu = -1;
689 	high.cs_load = 0;
690 	high.cs_mask = mask;
691 	high.cs_limit = minload;
692 	cpu_search_highest(cg, &high);
693 	return high.cs_cpu;
694 }
695 
696 /*
697  * Simultaneously find the highest and lowest loaded cpu reachable via
698  * cg.
699  */
700 static inline void
701 sched_both(struct cpu_group *cg, cpumask_t mask, int *lowcpu, int *highcpu)
702 {
703 	struct cpu_search high;
704 	struct cpu_search low;
705 
706 	low.cs_cpu = -1;
707 	low.cs_limit = -1;
708 	low.cs_load = -1;
709 	low.cs_mask = mask;
710 	high.cs_load = 0;
711 	high.cs_cpu = -1;
712 	high.cs_limit = -1;
713 	high.cs_mask = mask;
714 	cpu_search_both(cg, &low, &high);
715 	*lowcpu = low.cs_cpu;
716 	*highcpu = high.cs_cpu;
717 	return;
718 }
719 
720 static void
721 sched_balance_group(struct cpu_group *cg)
722 {
723 	cpumask_t mask;
724 	int high;
725 	int low;
726 	int i;
727 
728 	mask = -1;
729 	for (;;) {
730 		sched_both(cg, mask, &low, &high);
731 		if (low == high || low == -1 || high == -1)
732 			break;
733 		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low)))
734 			break;
735 		/*
736 		 * If we failed to move any threads determine which cpu
737 		 * to kick out of the set and try again.
738 	 	 */
739 		if (TDQ_CPU(high)->tdq_transferable == 0)
740 			mask &= ~(1 << high);
741 		else
742 			mask &= ~(1 << low);
743 	}
744 
745 	for (i = 0; i < cg->cg_children; i++)
746 		sched_balance_group(&cg->cg_child[i]);
747 }
748 
749 static void
750 sched_balance()
751 {
752 	struct tdq *tdq;
753 
754 	/*
755 	 * Select a random time between .5 * balance_interval and
756 	 * 1.5 * balance_interval.
757 	 */
758 	balance_ticks = max(balance_interval / 2, 1);
759 	balance_ticks += random() % balance_interval;
760 	if (smp_started == 0 || rebalance == 0)
761 		return;
762 	tdq = TDQ_SELF();
763 	TDQ_UNLOCK(tdq);
764 	sched_balance_group(cpu_top);
765 	TDQ_LOCK(tdq);
766 }
767 
768 /*
769  * Lock two thread queues using their address to maintain lock order.
770  */
771 static void
772 tdq_lock_pair(struct tdq *one, struct tdq *two)
773 {
774 	if (one < two) {
775 		TDQ_LOCK(one);
776 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
777 	} else {
778 		TDQ_LOCK(two);
779 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
780 	}
781 }
782 
783 /*
784  * Unlock two thread queues.  Order is not important here.
785  */
786 static void
787 tdq_unlock_pair(struct tdq *one, struct tdq *two)
788 {
789 	TDQ_UNLOCK(one);
790 	TDQ_UNLOCK(two);
791 }
792 
793 /*
794  * Transfer load between two imbalanced thread queues.
795  */
796 static int
797 sched_balance_pair(struct tdq *high, struct tdq *low)
798 {
799 	int transferable;
800 	int high_load;
801 	int low_load;
802 	int moved;
803 	int move;
804 	int diff;
805 	int i;
806 
807 	tdq_lock_pair(high, low);
808 	transferable = high->tdq_transferable;
809 	high_load = high->tdq_load;
810 	low_load = low->tdq_load;
811 	moved = 0;
812 	/*
813 	 * Determine what the imbalance is and then adjust that to how many
814 	 * threads we actually have to give up (transferable).
815 	 */
816 	if (transferable != 0) {
817 		diff = high_load - low_load;
818 		move = diff / 2;
819 		if (diff & 0x1)
820 			move++;
821 		move = min(move, transferable);
822 		for (i = 0; i < move; i++)
823 			moved += tdq_move(high, low);
824 		/*
825 		 * IPI the target cpu to force it to reschedule with the new
826 		 * workload.
827 		 */
828 		ipi_selected(1 << TDQ_ID(low), IPI_PREEMPT);
829 	}
830 	tdq_unlock_pair(high, low);
831 	return (moved);
832 }
833 
834 /*
835  * Move a thread from one thread queue to another.
836  */
837 static int
838 tdq_move(struct tdq *from, struct tdq *to)
839 {
840 	struct td_sched *ts;
841 	struct thread *td;
842 	struct tdq *tdq;
843 	int cpu;
844 
845 	TDQ_LOCK_ASSERT(from, MA_OWNED);
846 	TDQ_LOCK_ASSERT(to, MA_OWNED);
847 
848 	tdq = from;
849 	cpu = TDQ_ID(to);
850 	ts = tdq_steal(tdq, cpu);
851 	if (ts == NULL)
852 		return (0);
853 	td = ts->ts_thread;
854 	/*
855 	 * Although the run queue is locked the thread may be blocked.  Lock
856 	 * it to clear this and acquire the run-queue lock.
857 	 */
858 	thread_lock(td);
859 	/* Drop recursive lock on from acquired via thread_lock(). */
860 	TDQ_UNLOCK(from);
861 	sched_rem(td);
862 	ts->ts_cpu = cpu;
863 	td->td_lock = TDQ_LOCKPTR(to);
864 	tdq_add(to, td, SRQ_YIELDING);
865 	return (1);
866 }
867 
868 /*
869  * This tdq has idled.  Try to steal a thread from another cpu and switch
870  * to it.
871  */
872 static int
873 tdq_idled(struct tdq *tdq)
874 {
875 	struct cpu_group *cg;
876 	struct tdq *steal;
877 	cpumask_t mask;
878 	int thresh;
879 	int cpu;
880 
881 	if (smp_started == 0 || steal_idle == 0)
882 		return (1);
883 	mask = -1;
884 	mask &= ~PCPU_GET(cpumask);
885 	/* We don't want to be preempted while we're iterating. */
886 	spinlock_enter();
887 	for (cg = tdq->tdq_cg; cg != NULL; ) {
888 		if ((cg->cg_flags & (CG_FLAG_HTT | CG_FLAG_THREAD)) == 0)
889 			thresh = steal_thresh;
890 		else
891 			thresh = 1;
892 		cpu = sched_highest(cg, mask, thresh);
893 		if (cpu == -1) {
894 			cg = cg->cg_parent;
895 			continue;
896 		}
897 		steal = TDQ_CPU(cpu);
898 		mask &= ~(1 << cpu);
899 		tdq_lock_pair(tdq, steal);
900 		if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
901 			tdq_unlock_pair(tdq, steal);
902 			continue;
903 		}
904 		/*
905 		 * If a thread was added while interrupts were disabled don't
906 		 * steal one here.  If we fail to acquire one due to affinity
907 		 * restrictions loop again with this cpu removed from the
908 		 * set.
909 		 */
910 		if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
911 			tdq_unlock_pair(tdq, steal);
912 			continue;
913 		}
914 		spinlock_exit();
915 		TDQ_UNLOCK(steal);
916 		mi_switch(SW_VOL, NULL);
917 		thread_unlock(curthread);
918 
919 		return (0);
920 	}
921 	spinlock_exit();
922 	return (1);
923 }
924 
925 /*
926  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
927  */
928 static void
929 tdq_notify(struct tdq *tdq, struct td_sched *ts)
930 {
931 	int cpri;
932 	int pri;
933 	int cpu;
934 
935 	if (tdq->tdq_ipipending)
936 		return;
937 	cpu = ts->ts_cpu;
938 	pri = ts->ts_thread->td_priority;
939 	cpri = pcpu_find(cpu)->pc_curthread->td_priority;
940 	if (!sched_shouldpreempt(pri, cpri, 1))
941 		return;
942 	tdq->tdq_ipipending = 1;
943 	ipi_selected(1 << cpu, IPI_PREEMPT);
944 }
945 
946 /*
947  * Steals load from a timeshare queue.  Honors the rotating queue head
948  * index.
949  */
950 static struct td_sched *
951 runq_steal_from(struct runq *rq, int cpu, u_char start)
952 {
953 	struct td_sched *ts;
954 	struct rqbits *rqb;
955 	struct rqhead *rqh;
956 	int first;
957 	int bit;
958 	int pri;
959 	int i;
960 
961 	rqb = &rq->rq_status;
962 	bit = start & (RQB_BPW -1);
963 	pri = 0;
964 	first = 0;
965 again:
966 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
967 		if (rqb->rqb_bits[i] == 0)
968 			continue;
969 		if (bit != 0) {
970 			for (pri = bit; pri < RQB_BPW; pri++)
971 				if (rqb->rqb_bits[i] & (1ul << pri))
972 					break;
973 			if (pri >= RQB_BPW)
974 				continue;
975 		} else
976 			pri = RQB_FFS(rqb->rqb_bits[i]);
977 		pri += (i << RQB_L2BPW);
978 		rqh = &rq->rq_queues[pri];
979 		TAILQ_FOREACH(ts, rqh, ts_procq) {
980 			if (first && THREAD_CAN_MIGRATE(ts->ts_thread) &&
981 			    THREAD_CAN_SCHED(ts->ts_thread, cpu))
982 				return (ts);
983 			first = 1;
984 		}
985 	}
986 	if (start != 0) {
987 		start = 0;
988 		goto again;
989 	}
990 
991 	return (NULL);
992 }
993 
994 /*
995  * Steals load from a standard linear queue.
996  */
997 static struct td_sched *
998 runq_steal(struct runq *rq, int cpu)
999 {
1000 	struct rqhead *rqh;
1001 	struct rqbits *rqb;
1002 	struct td_sched *ts;
1003 	int word;
1004 	int bit;
1005 
1006 	rqb = &rq->rq_status;
1007 	for (word = 0; word < RQB_LEN; word++) {
1008 		if (rqb->rqb_bits[word] == 0)
1009 			continue;
1010 		for (bit = 0; bit < RQB_BPW; bit++) {
1011 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1012 				continue;
1013 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1014 			TAILQ_FOREACH(ts, rqh, ts_procq)
1015 				if (THREAD_CAN_MIGRATE(ts->ts_thread) &&
1016 				    THREAD_CAN_SCHED(ts->ts_thread, cpu))
1017 					return (ts);
1018 		}
1019 	}
1020 	return (NULL);
1021 }
1022 
1023 /*
1024  * Attempt to steal a thread in priority order from a thread queue.
1025  */
1026 static struct td_sched *
1027 tdq_steal(struct tdq *tdq, int cpu)
1028 {
1029 	struct td_sched *ts;
1030 
1031 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1032 	if ((ts = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1033 		return (ts);
1034 	if ((ts = runq_steal_from(&tdq->tdq_timeshare, cpu, tdq->tdq_ridx))
1035 	    != NULL)
1036 		return (ts);
1037 	return (runq_steal(&tdq->tdq_idle, cpu));
1038 }
1039 
1040 /*
1041  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1042  * current lock and returns with the assigned queue locked.
1043  */
1044 static inline struct tdq *
1045 sched_setcpu(struct td_sched *ts, int cpu, int flags)
1046 {
1047 	struct thread *td;
1048 	struct tdq *tdq;
1049 
1050 	THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED);
1051 
1052 	tdq = TDQ_CPU(cpu);
1053 	td = ts->ts_thread;
1054 	ts->ts_cpu = cpu;
1055 
1056 	/* If the lock matches just return the queue. */
1057 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1058 		return (tdq);
1059 #ifdef notyet
1060 	/*
1061 	 * If the thread isn't running its lockptr is a
1062 	 * turnstile or a sleepqueue.  We can just lock_set without
1063 	 * blocking.
1064 	 */
1065 	if (TD_CAN_RUN(td)) {
1066 		TDQ_LOCK(tdq);
1067 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1068 		return (tdq);
1069 	}
1070 #endif
1071 	/*
1072 	 * The hard case, migration, we need to block the thread first to
1073 	 * prevent order reversals with other cpus locks.
1074 	 */
1075 	thread_lock_block(td);
1076 	TDQ_LOCK(tdq);
1077 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1078 	return (tdq);
1079 }
1080 
1081 static int
1082 sched_pickcpu(struct td_sched *ts, int flags)
1083 {
1084 	struct cpu_group *cg;
1085 	struct thread *td;
1086 	struct tdq *tdq;
1087 	cpumask_t mask;
1088 	int self;
1089 	int pri;
1090 	int cpu;
1091 
1092 	self = PCPU_GET(cpuid);
1093 	td = ts->ts_thread;
1094 	if (smp_started == 0)
1095 		return (self);
1096 	/*
1097 	 * Don't migrate a running thread from sched_switch().
1098 	 */
1099 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1100 		return (ts->ts_cpu);
1101 	/*
1102 	 * Prefer to run interrupt threads on the processors that generate
1103 	 * the interrupt.
1104 	 */
1105 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1106 	    curthread->td_intr_nesting_level)
1107 		ts->ts_cpu = self;
1108 	/*
1109 	 * If the thread can run on the last cpu and the affinity has not
1110 	 * expired or it is idle run it there.
1111 	 */
1112 	pri = td->td_priority;
1113 	tdq = TDQ_CPU(ts->ts_cpu);
1114 	if (THREAD_CAN_SCHED(td, ts->ts_cpu)) {
1115 		if (tdq->tdq_lowpri > PRI_MIN_IDLE)
1116 			return (ts->ts_cpu);
1117 		if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri)
1118 			return (ts->ts_cpu);
1119 	}
1120 	/*
1121 	 * Search for the highest level in the tree that still has affinity.
1122 	 */
1123 	cg = NULL;
1124 	for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent)
1125 		if (SCHED_AFFINITY(ts, cg->cg_level))
1126 			break;
1127 	cpu = -1;
1128 	mask = td->td_cpuset->cs_mask.__bits[0];
1129 	if (cg)
1130 		cpu = sched_lowest(cg, mask, pri);
1131 	if (cpu == -1)
1132 		cpu = sched_lowest(cpu_top, mask, -1);
1133 	/*
1134 	 * Compare the lowest loaded cpu to current cpu.
1135 	 */
1136 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1137 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1138 		cpu = self;
1139 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1140 	return (cpu);
1141 }
1142 #endif
1143 
1144 /*
1145  * Pick the highest priority task we have and return it.
1146  */
1147 static struct td_sched *
1148 tdq_choose(struct tdq *tdq)
1149 {
1150 	struct td_sched *ts;
1151 
1152 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1153 	ts = runq_choose(&tdq->tdq_realtime);
1154 	if (ts != NULL)
1155 		return (ts);
1156 	ts = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1157 	if (ts != NULL) {
1158 		KASSERT(ts->ts_thread->td_priority >= PRI_MIN_TIMESHARE,
1159 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1160 		    ts->ts_thread->td_priority));
1161 		return (ts);
1162 	}
1163 
1164 	ts = runq_choose(&tdq->tdq_idle);
1165 	if (ts != NULL) {
1166 		KASSERT(ts->ts_thread->td_priority >= PRI_MIN_IDLE,
1167 		    ("tdq_choose: Invalid priority on idle queue %d",
1168 		    ts->ts_thread->td_priority));
1169 		return (ts);
1170 	}
1171 
1172 	return (NULL);
1173 }
1174 
1175 /*
1176  * Initialize a thread queue.
1177  */
1178 static void
1179 tdq_setup(struct tdq *tdq)
1180 {
1181 
1182 	if (bootverbose)
1183 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1184 	runq_init(&tdq->tdq_realtime);
1185 	runq_init(&tdq->tdq_timeshare);
1186 	runq_init(&tdq->tdq_idle);
1187 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1188 	    "sched lock %d", (int)TDQ_ID(tdq));
1189 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1190 	    MTX_SPIN | MTX_RECURSE);
1191 }
1192 
1193 #ifdef SMP
1194 static void
1195 sched_setup_smp(void)
1196 {
1197 	struct tdq *tdq;
1198 	int i;
1199 
1200 	cpu_top = smp_topo();
1201 	for (i = 0; i < MAXCPU; i++) {
1202 		if (CPU_ABSENT(i))
1203 			continue;
1204 		tdq = TDQ_CPU(i);
1205 		tdq_setup(tdq);
1206 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1207 		if (tdq->tdq_cg == NULL)
1208 			panic("Can't find cpu group for %d\n", i);
1209 	}
1210 	balance_tdq = TDQ_SELF();
1211 	sched_balance();
1212 }
1213 #endif
1214 
1215 /*
1216  * Setup the thread queues and initialize the topology based on MD
1217  * information.
1218  */
1219 static void
1220 sched_setup(void *dummy)
1221 {
1222 	struct tdq *tdq;
1223 
1224 	tdq = TDQ_SELF();
1225 #ifdef SMP
1226 	sched_setup_smp();
1227 #else
1228 	tdq_setup(tdq);
1229 #endif
1230 	/*
1231 	 * To avoid divide-by-zero, we set realstathz a dummy value
1232 	 * in case which sched_clock() called before sched_initticks().
1233 	 */
1234 	realstathz = hz;
1235 	sched_slice = (realstathz/10);	/* ~100ms */
1236 	tickincr = 1 << SCHED_TICK_SHIFT;
1237 
1238 	/* Add thread0's load since it's running. */
1239 	TDQ_LOCK(tdq);
1240 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1241 	tdq_load_add(tdq, &td_sched0);
1242 	tdq->tdq_lowpri = thread0.td_priority;
1243 	TDQ_UNLOCK(tdq);
1244 }
1245 
1246 /*
1247  * This routine determines the tickincr after stathz and hz are setup.
1248  */
1249 /* ARGSUSED */
1250 static void
1251 sched_initticks(void *dummy)
1252 {
1253 	int incr;
1254 
1255 	realstathz = stathz ? stathz : hz;
1256 	sched_slice = (realstathz/10);	/* ~100ms */
1257 
1258 	/*
1259 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1260 	 * hz not being evenly divisible by stathz on all platforms.
1261 	 */
1262 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1263 	/*
1264 	 * This does not work for values of stathz that are more than
1265 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1266 	 */
1267 	if (incr == 0)
1268 		incr = 1;
1269 	tickincr = incr;
1270 #ifdef SMP
1271 	/*
1272 	 * Set the default balance interval now that we know
1273 	 * what realstathz is.
1274 	 */
1275 	balance_interval = realstathz;
1276 	/*
1277 	 * Set steal thresh to log2(mp_ncpu) but no greater than 4.  This
1278 	 * prevents excess thrashing on large machines and excess idle on
1279 	 * smaller machines.
1280 	 */
1281 	steal_thresh = min(ffs(mp_ncpus) - 1, 3);
1282 	affinity = SCHED_AFFINITY_DEFAULT;
1283 #endif
1284 }
1285 
1286 
1287 /*
1288  * This is the core of the interactivity algorithm.  Determines a score based
1289  * on past behavior.  It is the ratio of sleep time to run time scaled to
1290  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1291  * differs from the cpu usage because it does not account for time spent
1292  * waiting on a run-queue.  Would be prettier if we had floating point.
1293  */
1294 static int
1295 sched_interact_score(struct thread *td)
1296 {
1297 	struct td_sched *ts;
1298 	int div;
1299 
1300 	ts = td->td_sched;
1301 	/*
1302 	 * The score is only needed if this is likely to be an interactive
1303 	 * task.  Don't go through the expense of computing it if there's
1304 	 * no chance.
1305 	 */
1306 	if (sched_interact <= SCHED_INTERACT_HALF &&
1307 		ts->ts_runtime >= ts->ts_slptime)
1308 			return (SCHED_INTERACT_HALF);
1309 
1310 	if (ts->ts_runtime > ts->ts_slptime) {
1311 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1312 		return (SCHED_INTERACT_HALF +
1313 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1314 	}
1315 	if (ts->ts_slptime > ts->ts_runtime) {
1316 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1317 		return (ts->ts_runtime / div);
1318 	}
1319 	/* runtime == slptime */
1320 	if (ts->ts_runtime)
1321 		return (SCHED_INTERACT_HALF);
1322 
1323 	/*
1324 	 * This can happen if slptime and runtime are 0.
1325 	 */
1326 	return (0);
1327 
1328 }
1329 
1330 /*
1331  * Scale the scheduling priority according to the "interactivity" of this
1332  * process.
1333  */
1334 static void
1335 sched_priority(struct thread *td)
1336 {
1337 	int score;
1338 	int pri;
1339 
1340 	if (td->td_pri_class != PRI_TIMESHARE)
1341 		return;
1342 	/*
1343 	 * If the score is interactive we place the thread in the realtime
1344 	 * queue with a priority that is less than kernel and interrupt
1345 	 * priorities.  These threads are not subject to nice restrictions.
1346 	 *
1347 	 * Scores greater than this are placed on the normal timeshare queue
1348 	 * where the priority is partially decided by the most recent cpu
1349 	 * utilization and the rest is decided by nice value.
1350 	 *
1351 	 * The nice value of the process has a linear effect on the calculated
1352 	 * score.  Negative nice values make it easier for a thread to be
1353 	 * considered interactive.
1354 	 */
1355 	score = imax(0, sched_interact_score(td) - td->td_proc->p_nice);
1356 	if (score < sched_interact) {
1357 		pri = PRI_MIN_REALTIME;
1358 		pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact)
1359 		    * score;
1360 		KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME,
1361 		    ("sched_priority: invalid interactive priority %d score %d",
1362 		    pri, score));
1363 	} else {
1364 		pri = SCHED_PRI_MIN;
1365 		if (td->td_sched->ts_ticks)
1366 			pri += SCHED_PRI_TICKS(td->td_sched);
1367 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1368 		KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE,
1369 		    ("sched_priority: invalid priority %d: nice %d, "
1370 		    "ticks %d ftick %d ltick %d tick pri %d",
1371 		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1372 		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1373 		    SCHED_PRI_TICKS(td->td_sched)));
1374 	}
1375 	sched_user_prio(td, pri);
1376 
1377 	return;
1378 }
1379 
1380 /*
1381  * This routine enforces a maximum limit on the amount of scheduling history
1382  * kept.  It is called after either the slptime or runtime is adjusted.  This
1383  * function is ugly due to integer math.
1384  */
1385 static void
1386 sched_interact_update(struct thread *td)
1387 {
1388 	struct td_sched *ts;
1389 	u_int sum;
1390 
1391 	ts = td->td_sched;
1392 	sum = ts->ts_runtime + ts->ts_slptime;
1393 	if (sum < SCHED_SLP_RUN_MAX)
1394 		return;
1395 	/*
1396 	 * This only happens from two places:
1397 	 * 1) We have added an unusual amount of run time from fork_exit.
1398 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1399 	 */
1400 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1401 		if (ts->ts_runtime > ts->ts_slptime) {
1402 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1403 			ts->ts_slptime = 1;
1404 		} else {
1405 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1406 			ts->ts_runtime = 1;
1407 		}
1408 		return;
1409 	}
1410 	/*
1411 	 * If we have exceeded by more than 1/5th then the algorithm below
1412 	 * will not bring us back into range.  Dividing by two here forces
1413 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1414 	 */
1415 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1416 		ts->ts_runtime /= 2;
1417 		ts->ts_slptime /= 2;
1418 		return;
1419 	}
1420 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1421 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1422 }
1423 
1424 /*
1425  * Scale back the interactivity history when a child thread is created.  The
1426  * history is inherited from the parent but the thread may behave totally
1427  * differently.  For example, a shell spawning a compiler process.  We want
1428  * to learn that the compiler is behaving badly very quickly.
1429  */
1430 static void
1431 sched_interact_fork(struct thread *td)
1432 {
1433 	int ratio;
1434 	int sum;
1435 
1436 	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1437 	if (sum > SCHED_SLP_RUN_FORK) {
1438 		ratio = sum / SCHED_SLP_RUN_FORK;
1439 		td->td_sched->ts_runtime /= ratio;
1440 		td->td_sched->ts_slptime /= ratio;
1441 	}
1442 }
1443 
1444 /*
1445  * Called from proc0_init() to setup the scheduler fields.
1446  */
1447 void
1448 schedinit(void)
1449 {
1450 
1451 	/*
1452 	 * Set up the scheduler specific parts of proc0.
1453 	 */
1454 	proc0.p_sched = NULL; /* XXX */
1455 	thread0.td_sched = &td_sched0;
1456 	td_sched0.ts_ltick = ticks;
1457 	td_sched0.ts_ftick = ticks;
1458 	td_sched0.ts_thread = &thread0;
1459 	td_sched0.ts_slice = sched_slice;
1460 }
1461 
1462 /*
1463  * This is only somewhat accurate since given many processes of the same
1464  * priority they will switch when their slices run out, which will be
1465  * at most sched_slice stathz ticks.
1466  */
1467 int
1468 sched_rr_interval(void)
1469 {
1470 
1471 	/* Convert sched_slice to hz */
1472 	return (hz/(realstathz/sched_slice));
1473 }
1474 
1475 /*
1476  * Update the percent cpu tracking information when it is requested or
1477  * the total history exceeds the maximum.  We keep a sliding history of
1478  * tick counts that slowly decays.  This is less precise than the 4BSD
1479  * mechanism since it happens with less regular and frequent events.
1480  */
1481 static void
1482 sched_pctcpu_update(struct td_sched *ts)
1483 {
1484 
1485 	if (ts->ts_ticks == 0)
1486 		return;
1487 	if (ticks - (hz / 10) < ts->ts_ltick &&
1488 	    SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX)
1489 		return;
1490 	/*
1491 	 * Adjust counters and watermark for pctcpu calc.
1492 	 */
1493 	if (ts->ts_ltick > ticks - SCHED_TICK_TARG)
1494 		ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) *
1495 			    SCHED_TICK_TARG;
1496 	else
1497 		ts->ts_ticks = 0;
1498 	ts->ts_ltick = ticks;
1499 	ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG;
1500 }
1501 
1502 /*
1503  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1504  * if necessary.  This is the back-end for several priority related
1505  * functions.
1506  */
1507 static void
1508 sched_thread_priority(struct thread *td, u_char prio)
1509 {
1510 	struct td_sched *ts;
1511 	struct tdq *tdq;
1512 	int oldpri;
1513 
1514 	CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)",
1515 	    td, td->td_name, td->td_priority, prio, curthread,
1516 	    curthread->td_name);
1517 	ts = td->td_sched;
1518 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1519 	if (td->td_priority == prio)
1520 		return;
1521 	/*
1522 	 * If the priority has been elevated due to priority
1523 	 * propagation, we may have to move ourselves to a new
1524 	 * queue.  This could be optimized to not re-add in some
1525 	 * cases.
1526 	 */
1527 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1528 		sched_rem(td);
1529 		td->td_priority = prio;
1530 		sched_add(td, SRQ_BORROWING);
1531 		return;
1532 	}
1533 	/*
1534 	 * If the thread is currently running we may have to adjust the lowpri
1535 	 * information so other cpus are aware of our current priority.
1536 	 */
1537 	if (TD_IS_RUNNING(td)) {
1538 		tdq = TDQ_CPU(ts->ts_cpu);
1539 		oldpri = td->td_priority;
1540 		td->td_priority = prio;
1541 		if (prio < tdq->tdq_lowpri)
1542 			tdq->tdq_lowpri = prio;
1543 		else if (tdq->tdq_lowpri == oldpri)
1544 			tdq_setlowpri(tdq, td);
1545 		return;
1546 	}
1547 	td->td_priority = prio;
1548 }
1549 
1550 /*
1551  * Update a thread's priority when it is lent another thread's
1552  * priority.
1553  */
1554 void
1555 sched_lend_prio(struct thread *td, u_char prio)
1556 {
1557 
1558 	td->td_flags |= TDF_BORROWING;
1559 	sched_thread_priority(td, prio);
1560 }
1561 
1562 /*
1563  * Restore a thread's priority when priority propagation is
1564  * over.  The prio argument is the minimum priority the thread
1565  * needs to have to satisfy other possible priority lending
1566  * requests.  If the thread's regular priority is less
1567  * important than prio, the thread will keep a priority boost
1568  * of prio.
1569  */
1570 void
1571 sched_unlend_prio(struct thread *td, u_char prio)
1572 {
1573 	u_char base_pri;
1574 
1575 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1576 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1577 		base_pri = td->td_user_pri;
1578 	else
1579 		base_pri = td->td_base_pri;
1580 	if (prio >= base_pri) {
1581 		td->td_flags &= ~TDF_BORROWING;
1582 		sched_thread_priority(td, base_pri);
1583 	} else
1584 		sched_lend_prio(td, prio);
1585 }
1586 
1587 /*
1588  * Standard entry for setting the priority to an absolute value.
1589  */
1590 void
1591 sched_prio(struct thread *td, u_char prio)
1592 {
1593 	u_char oldprio;
1594 
1595 	/* First, update the base priority. */
1596 	td->td_base_pri = prio;
1597 
1598 	/*
1599 	 * If the thread is borrowing another thread's priority, don't
1600 	 * ever lower the priority.
1601 	 */
1602 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1603 		return;
1604 
1605 	/* Change the real priority. */
1606 	oldprio = td->td_priority;
1607 	sched_thread_priority(td, prio);
1608 
1609 	/*
1610 	 * If the thread is on a turnstile, then let the turnstile update
1611 	 * its state.
1612 	 */
1613 	if (TD_ON_LOCK(td) && oldprio != prio)
1614 		turnstile_adjust(td, oldprio);
1615 }
1616 
1617 /*
1618  * Set the base user priority, does not effect current running priority.
1619  */
1620 void
1621 sched_user_prio(struct thread *td, u_char prio)
1622 {
1623 	u_char oldprio;
1624 
1625 	td->td_base_user_pri = prio;
1626 	if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
1627                 return;
1628 	oldprio = td->td_user_pri;
1629 	td->td_user_pri = prio;
1630 }
1631 
1632 void
1633 sched_lend_user_prio(struct thread *td, u_char prio)
1634 {
1635 	u_char oldprio;
1636 
1637 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1638 	td->td_flags |= TDF_UBORROWING;
1639 	oldprio = td->td_user_pri;
1640 	td->td_user_pri = prio;
1641 }
1642 
1643 void
1644 sched_unlend_user_prio(struct thread *td, u_char prio)
1645 {
1646 	u_char base_pri;
1647 
1648 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1649 	base_pri = td->td_base_user_pri;
1650 	if (prio >= base_pri) {
1651 		td->td_flags &= ~TDF_UBORROWING;
1652 		sched_user_prio(td, base_pri);
1653 	} else {
1654 		sched_lend_user_prio(td, prio);
1655 	}
1656 }
1657 
1658 /*
1659  * Block a thread for switching.  Similar to thread_block() but does not
1660  * bump the spin count.
1661  */
1662 static inline struct mtx *
1663 thread_block_switch(struct thread *td)
1664 {
1665 	struct mtx *lock;
1666 
1667 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1668 	lock = td->td_lock;
1669 	td->td_lock = &blocked_lock;
1670 	mtx_unlock_spin(lock);
1671 
1672 	return (lock);
1673 }
1674 
1675 /*
1676  * Handle migration from sched_switch().  This happens only for
1677  * cpu binding.
1678  */
1679 static struct mtx *
1680 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1681 {
1682 	struct tdq *tdn;
1683 
1684 	tdn = TDQ_CPU(td->td_sched->ts_cpu);
1685 #ifdef SMP
1686 	tdq_load_rem(tdq, td->td_sched);
1687 	/*
1688 	 * Do the lock dance required to avoid LOR.  We grab an extra
1689 	 * spinlock nesting to prevent preemption while we're
1690 	 * not holding either run-queue lock.
1691 	 */
1692 	spinlock_enter();
1693 	thread_block_switch(td);	/* This releases the lock on tdq. */
1694 	TDQ_LOCK(tdn);
1695 	tdq_add(tdn, td, flags);
1696 	tdq_notify(tdn, td->td_sched);
1697 	/*
1698 	 * After we unlock tdn the new cpu still can't switch into this
1699 	 * thread until we've unblocked it in cpu_switch().  The lock
1700 	 * pointers may match in the case of HTT cores.  Don't unlock here
1701 	 * or we can deadlock when the other CPU runs the IPI handler.
1702 	 */
1703 	if (TDQ_LOCKPTR(tdn) != TDQ_LOCKPTR(tdq)) {
1704 		TDQ_UNLOCK(tdn);
1705 		TDQ_LOCK(tdq);
1706 	}
1707 	spinlock_exit();
1708 #endif
1709 	return (TDQ_LOCKPTR(tdn));
1710 }
1711 
1712 /*
1713  * Release a thread that was blocked with thread_block_switch().
1714  */
1715 static inline void
1716 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1717 {
1718 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1719 	    (uintptr_t)mtx);
1720 }
1721 
1722 /*
1723  * Switch threads.  This function has to handle threads coming in while
1724  * blocked for some reason, running, or idle.  It also must deal with
1725  * migrating a thread from one queue to another as running threads may
1726  * be assigned elsewhere via binding.
1727  */
1728 void
1729 sched_switch(struct thread *td, struct thread *newtd, int flags)
1730 {
1731 	struct tdq *tdq;
1732 	struct td_sched *ts;
1733 	struct mtx *mtx;
1734 	int srqflag;
1735 	int cpuid;
1736 
1737 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1738 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1739 
1740 	cpuid = PCPU_GET(cpuid);
1741 	tdq = TDQ_CPU(cpuid);
1742 	ts = td->td_sched;
1743 	mtx = td->td_lock;
1744 	ts->ts_rltick = ticks;
1745 	td->td_lastcpu = td->td_oncpu;
1746 	td->td_oncpu = NOCPU;
1747 	td->td_flags &= ~TDF_NEEDRESCHED;
1748 	td->td_owepreempt = 0;
1749 	/*
1750 	 * The lock pointer in an idle thread should never change.  Reset it
1751 	 * to CAN_RUN as well.
1752 	 */
1753 	if (TD_IS_IDLETHREAD(td)) {
1754 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1755 		TD_SET_CAN_RUN(td);
1756 	} else if (TD_IS_RUNNING(td)) {
1757 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1758 		srqflag = (flags & SW_PREEMPT) ?
1759 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1760 		    SRQ_OURSELF|SRQ_YIELDING;
1761 		if (ts->ts_cpu == cpuid)
1762 			tdq_runq_add(tdq, ts, srqflag);
1763 		else
1764 			mtx = sched_switch_migrate(tdq, td, srqflag);
1765 	} else {
1766 		/* This thread must be going to sleep. */
1767 		TDQ_LOCK(tdq);
1768 		mtx = thread_block_switch(td);
1769 		tdq_load_rem(tdq, ts);
1770 	}
1771 	/*
1772 	 * We enter here with the thread blocked and assigned to the
1773 	 * appropriate cpu run-queue or sleep-queue and with the current
1774 	 * thread-queue locked.
1775 	 */
1776 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1777 	newtd = choosethread();
1778 	/*
1779 	 * Call the MD code to switch contexts if necessary.
1780 	 */
1781 	if (td != newtd) {
1782 #ifdef	HWPMC_HOOKS
1783 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1784 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1785 #endif
1786 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1787 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1788 		cpu_switch(td, newtd, mtx);
1789 		/*
1790 		 * We may return from cpu_switch on a different cpu.  However,
1791 		 * we always return with td_lock pointing to the current cpu's
1792 		 * run queue lock.
1793 		 */
1794 		cpuid = PCPU_GET(cpuid);
1795 		tdq = TDQ_CPU(cpuid);
1796 		lock_profile_obtain_lock_success(
1797 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1798 #ifdef	HWPMC_HOOKS
1799 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1800 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1801 #endif
1802 	} else
1803 		thread_unblock_switch(td, mtx);
1804 	/*
1805 	 * Assert that all went well and return.
1806 	 */
1807 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1808 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1809 	td->td_oncpu = cpuid;
1810 }
1811 
1812 /*
1813  * Adjust thread priorities as a result of a nice request.
1814  */
1815 void
1816 sched_nice(struct proc *p, int nice)
1817 {
1818 	struct thread *td;
1819 
1820 	PROC_LOCK_ASSERT(p, MA_OWNED);
1821 
1822 	p->p_nice = nice;
1823 	FOREACH_THREAD_IN_PROC(p, td) {
1824 		thread_lock(td);
1825 		sched_priority(td);
1826 		sched_prio(td, td->td_base_user_pri);
1827 		thread_unlock(td);
1828 	}
1829 }
1830 
1831 /*
1832  * Record the sleep time for the interactivity scorer.
1833  */
1834 void
1835 sched_sleep(struct thread *td, int prio)
1836 {
1837 
1838 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1839 
1840 	td->td_slptick = ticks;
1841 	if (TD_IS_SUSPENDED(td) || prio <= PSOCK)
1842 		td->td_flags |= TDF_CANSWAP;
1843 	if (static_boost && prio)
1844 		sched_prio(td, prio);
1845 }
1846 
1847 /*
1848  * Schedule a thread to resume execution and record how long it voluntarily
1849  * slept.  We also update the pctcpu, interactivity, and priority.
1850  */
1851 void
1852 sched_wakeup(struct thread *td)
1853 {
1854 	struct td_sched *ts;
1855 	int slptick;
1856 
1857 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1858 	ts = td->td_sched;
1859 	td->td_flags &= ~TDF_CANSWAP;
1860 	/*
1861 	 * If we slept for more than a tick update our interactivity and
1862 	 * priority.
1863 	 */
1864 	slptick = td->td_slptick;
1865 	td->td_slptick = 0;
1866 	if (slptick && slptick != ticks) {
1867 		u_int hzticks;
1868 
1869 		hzticks = (ticks - slptick) << SCHED_TICK_SHIFT;
1870 		ts->ts_slptime += hzticks;
1871 		sched_interact_update(td);
1872 		sched_pctcpu_update(ts);
1873 	}
1874 	/* Reset the slice value after we sleep. */
1875 	ts->ts_slice = sched_slice;
1876 	sched_add(td, SRQ_BORING);
1877 }
1878 
1879 /*
1880  * Penalize the parent for creating a new child and initialize the child's
1881  * priority.
1882  */
1883 void
1884 sched_fork(struct thread *td, struct thread *child)
1885 {
1886 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1887 	sched_fork_thread(td, child);
1888 	/*
1889 	 * Penalize the parent and child for forking.
1890 	 */
1891 	sched_interact_fork(child);
1892 	sched_priority(child);
1893 	td->td_sched->ts_runtime += tickincr;
1894 	sched_interact_update(td);
1895 	sched_priority(td);
1896 }
1897 
1898 /*
1899  * Fork a new thread, may be within the same process.
1900  */
1901 void
1902 sched_fork_thread(struct thread *td, struct thread *child)
1903 {
1904 	struct td_sched *ts;
1905 	struct td_sched *ts2;
1906 
1907 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1908 	/*
1909 	 * Initialize child.
1910 	 */
1911 	ts = td->td_sched;
1912 	ts2 = child->td_sched;
1913 	child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
1914 	child->td_cpuset = cpuset_ref(td->td_cpuset);
1915 	ts2->ts_thread = child;
1916 	ts2->ts_cpu = ts->ts_cpu;
1917 	ts2->ts_flags = 0;
1918 	/*
1919 	 * Grab our parents cpu estimation information and priority.
1920 	 */
1921 	ts2->ts_ticks = ts->ts_ticks;
1922 	ts2->ts_ltick = ts->ts_ltick;
1923 	ts2->ts_ftick = ts->ts_ftick;
1924 	child->td_user_pri = td->td_user_pri;
1925 	child->td_base_user_pri = td->td_base_user_pri;
1926 	/*
1927 	 * And update interactivity score.
1928 	 */
1929 	ts2->ts_slptime = ts->ts_slptime;
1930 	ts2->ts_runtime = ts->ts_runtime;
1931 	ts2->ts_slice = 1;	/* Attempt to quickly learn interactivity. */
1932 }
1933 
1934 /*
1935  * Adjust the priority class of a thread.
1936  */
1937 void
1938 sched_class(struct thread *td, int class)
1939 {
1940 
1941 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1942 	if (td->td_pri_class == class)
1943 		return;
1944 	td->td_pri_class = class;
1945 }
1946 
1947 /*
1948  * Return some of the child's priority and interactivity to the parent.
1949  */
1950 void
1951 sched_exit(struct proc *p, struct thread *child)
1952 {
1953 	struct thread *td;
1954 
1955 	CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d",
1956 	    child, child->td_name, child->td_priority);
1957 
1958 	PROC_LOCK_ASSERT(p, MA_OWNED);
1959 	td = FIRST_THREAD_IN_PROC(p);
1960 	sched_exit_thread(td, child);
1961 }
1962 
1963 /*
1964  * Penalize another thread for the time spent on this one.  This helps to
1965  * worsen the priority and interactivity of processes which schedule batch
1966  * jobs such as make.  This has little effect on the make process itself but
1967  * causes new processes spawned by it to receive worse scores immediately.
1968  */
1969 void
1970 sched_exit_thread(struct thread *td, struct thread *child)
1971 {
1972 
1973 	CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d",
1974 	    child, child->td_name, child->td_priority);
1975 
1976 	/*
1977 	 * Give the child's runtime to the parent without returning the
1978 	 * sleep time as a penalty to the parent.  This causes shells that
1979 	 * launch expensive things to mark their children as expensive.
1980 	 */
1981 	thread_lock(td);
1982 	td->td_sched->ts_runtime += child->td_sched->ts_runtime;
1983 	sched_interact_update(td);
1984 	sched_priority(td);
1985 	thread_unlock(td);
1986 }
1987 
1988 void
1989 sched_preempt(struct thread *td)
1990 {
1991 	struct tdq *tdq;
1992 
1993 	thread_lock(td);
1994 	tdq = TDQ_SELF();
1995 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1996 	tdq->tdq_ipipending = 0;
1997 	if (td->td_priority > tdq->tdq_lowpri) {
1998 		if (td->td_critnest > 1)
1999 			td->td_owepreempt = 1;
2000 		else
2001 			mi_switch(SW_INVOL | SW_PREEMPT, NULL);
2002 	}
2003 	thread_unlock(td);
2004 }
2005 
2006 /*
2007  * Fix priorities on return to user-space.  Priorities may be elevated due
2008  * to static priorities in msleep() or similar.
2009  */
2010 void
2011 sched_userret(struct thread *td)
2012 {
2013 	/*
2014 	 * XXX we cheat slightly on the locking here to avoid locking in
2015 	 * the usual case.  Setting td_priority here is essentially an
2016 	 * incomplete workaround for not setting it properly elsewhere.
2017 	 * Now that some interrupt handlers are threads, not setting it
2018 	 * properly elsewhere can clobber it in the window between setting
2019 	 * it here and returning to user mode, so don't waste time setting
2020 	 * it perfectly here.
2021 	 */
2022 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2023 	    ("thread with borrowed priority returning to userland"));
2024 	if (td->td_priority != td->td_user_pri) {
2025 		thread_lock(td);
2026 		td->td_priority = td->td_user_pri;
2027 		td->td_base_pri = td->td_user_pri;
2028 		tdq_setlowpri(TDQ_SELF(), td);
2029 		thread_unlock(td);
2030         }
2031 }
2032 
2033 /*
2034  * Handle a stathz tick.  This is really only relevant for timeshare
2035  * threads.
2036  */
2037 void
2038 sched_clock(struct thread *td)
2039 {
2040 	struct tdq *tdq;
2041 	struct td_sched *ts;
2042 
2043 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2044 	tdq = TDQ_SELF();
2045 #ifdef SMP
2046 	/*
2047 	 * We run the long term load balancer infrequently on the first cpu.
2048 	 */
2049 	if (balance_tdq == tdq) {
2050 		if (balance_ticks && --balance_ticks == 0)
2051 			sched_balance();
2052 	}
2053 #endif
2054 	/*
2055 	 * Advance the insert index once for each tick to ensure that all
2056 	 * threads get a chance to run.
2057 	 */
2058 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2059 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2060 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2061 			tdq->tdq_ridx = tdq->tdq_idx;
2062 	}
2063 	ts = td->td_sched;
2064 	if (td->td_pri_class & PRI_FIFO_BIT)
2065 		return;
2066 	if (td->td_pri_class == PRI_TIMESHARE) {
2067 		/*
2068 		 * We used a tick; charge it to the thread so
2069 		 * that we can compute our interactivity.
2070 		 */
2071 		td->td_sched->ts_runtime += tickincr;
2072 		sched_interact_update(td);
2073 		sched_priority(td);
2074 	}
2075 	/*
2076 	 * We used up one time slice.
2077 	 */
2078 	if (--ts->ts_slice > 0)
2079 		return;
2080 	/*
2081 	 * We're out of time, force a requeue at userret().
2082 	 */
2083 	ts->ts_slice = sched_slice;
2084 	td->td_flags |= TDF_NEEDRESCHED;
2085 }
2086 
2087 /*
2088  * Called once per hz tick.  Used for cpu utilization information.  This
2089  * is easier than trying to scale based on stathz.
2090  */
2091 void
2092 sched_tick(void)
2093 {
2094 	struct td_sched *ts;
2095 
2096 	ts = curthread->td_sched;
2097 	/* Adjust ticks for pctcpu */
2098 	ts->ts_ticks += 1 << SCHED_TICK_SHIFT;
2099 	ts->ts_ltick = ticks;
2100 	/*
2101 	 * Update if we've exceeded our desired tick threshhold by over one
2102 	 * second.
2103 	 */
2104 	if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick)
2105 		sched_pctcpu_update(ts);
2106 }
2107 
2108 /*
2109  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2110  * cooperative idle threads.
2111  */
2112 int
2113 sched_runnable(void)
2114 {
2115 	struct tdq *tdq;
2116 	int load;
2117 
2118 	load = 1;
2119 
2120 	tdq = TDQ_SELF();
2121 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2122 		if (tdq->tdq_load > 0)
2123 			goto out;
2124 	} else
2125 		if (tdq->tdq_load - 1 > 0)
2126 			goto out;
2127 	load = 0;
2128 out:
2129 	return (load);
2130 }
2131 
2132 /*
2133  * Choose the highest priority thread to run.  The thread is removed from
2134  * the run-queue while running however the load remains.  For SMP we set
2135  * the tdq in the global idle bitmask if it idles here.
2136  */
2137 struct thread *
2138 sched_choose(void)
2139 {
2140 	struct td_sched *ts;
2141 	struct tdq *tdq;
2142 
2143 	tdq = TDQ_SELF();
2144 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2145 	ts = tdq_choose(tdq);
2146 	if (ts) {
2147 		ts->ts_ltick = ticks;
2148 		tdq_runq_rem(tdq, ts);
2149 		return (ts->ts_thread);
2150 	}
2151 	return (PCPU_GET(idlethread));
2152 }
2153 
2154 /*
2155  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2156  * we always request it once we exit a critical section.
2157  */
2158 static inline void
2159 sched_setpreempt(struct thread *td)
2160 {
2161 	struct thread *ctd;
2162 	int cpri;
2163 	int pri;
2164 
2165 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2166 
2167 	ctd = curthread;
2168 	pri = td->td_priority;
2169 	cpri = ctd->td_priority;
2170 	if (pri < cpri)
2171 		ctd->td_flags |= TDF_NEEDRESCHED;
2172 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2173 		return;
2174 	if (!sched_shouldpreempt(pri, cpri, 0))
2175 		return;
2176 	ctd->td_owepreempt = 1;
2177 }
2178 
2179 /*
2180  * Add a thread to a thread queue.  Select the appropriate runq and add the
2181  * thread to it.  This is the internal function called when the tdq is
2182  * predetermined.
2183  */
2184 void
2185 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2186 {
2187 	struct td_sched *ts;
2188 
2189 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2190 	KASSERT((td->td_inhibitors == 0),
2191 	    ("sched_add: trying to run inhibited thread"));
2192 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2193 	    ("sched_add: bad thread state"));
2194 	KASSERT(td->td_flags & TDF_INMEM,
2195 	    ("sched_add: thread swapped out"));
2196 
2197 	ts = td->td_sched;
2198 	if (td->td_priority < tdq->tdq_lowpri)
2199 		tdq->tdq_lowpri = td->td_priority;
2200 	tdq_runq_add(tdq, ts, flags);
2201 	tdq_load_add(tdq, ts);
2202 }
2203 
2204 /*
2205  * Select the target thread queue and add a thread to it.  Request
2206  * preemption or IPI a remote processor if required.
2207  */
2208 void
2209 sched_add(struct thread *td, int flags)
2210 {
2211 	struct tdq *tdq;
2212 #ifdef SMP
2213 	struct td_sched *ts;
2214 	int cpu;
2215 #endif
2216 	CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
2217 	    td, td->td_name, td->td_priority, curthread,
2218 	    curthread->td_name);
2219 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2220 	/*
2221 	 * Recalculate the priority before we select the target cpu or
2222 	 * run-queue.
2223 	 */
2224 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2225 		sched_priority(td);
2226 #ifdef SMP
2227 	/*
2228 	 * Pick the destination cpu and if it isn't ours transfer to the
2229 	 * target cpu.
2230 	 */
2231 	ts = td->td_sched;
2232 	cpu = sched_pickcpu(ts, flags);
2233 	tdq = sched_setcpu(ts, cpu, flags);
2234 	tdq_add(tdq, td, flags);
2235 	if (cpu != PCPU_GET(cpuid)) {
2236 		tdq_notify(tdq, ts);
2237 		return;
2238 	}
2239 #else
2240 	tdq = TDQ_SELF();
2241 	TDQ_LOCK(tdq);
2242 	/*
2243 	 * Now that the thread is moving to the run-queue, set the lock
2244 	 * to the scheduler's lock.
2245 	 */
2246 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2247 	tdq_add(tdq, td, flags);
2248 #endif
2249 	if (!(flags & SRQ_YIELDING))
2250 		sched_setpreempt(td);
2251 }
2252 
2253 /*
2254  * Remove a thread from a run-queue without running it.  This is used
2255  * when we're stealing a thread from a remote queue.  Otherwise all threads
2256  * exit by calling sched_exit_thread() and sched_throw() themselves.
2257  */
2258 void
2259 sched_rem(struct thread *td)
2260 {
2261 	struct tdq *tdq;
2262 	struct td_sched *ts;
2263 
2264 	CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)",
2265 	    td, td->td_name, td->td_priority, curthread,
2266 	    curthread->td_name);
2267 	ts = td->td_sched;
2268 	tdq = TDQ_CPU(ts->ts_cpu);
2269 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2270 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2271 	KASSERT(TD_ON_RUNQ(td),
2272 	    ("sched_rem: thread not on run queue"));
2273 	tdq_runq_rem(tdq, ts);
2274 	tdq_load_rem(tdq, ts);
2275 	TD_SET_CAN_RUN(td);
2276 	if (td->td_priority == tdq->tdq_lowpri)
2277 		tdq_setlowpri(tdq, NULL);
2278 }
2279 
2280 /*
2281  * Fetch cpu utilization information.  Updates on demand.
2282  */
2283 fixpt_t
2284 sched_pctcpu(struct thread *td)
2285 {
2286 	fixpt_t pctcpu;
2287 	struct td_sched *ts;
2288 
2289 	pctcpu = 0;
2290 	ts = td->td_sched;
2291 	if (ts == NULL)
2292 		return (0);
2293 
2294 	thread_lock(td);
2295 	if (ts->ts_ticks) {
2296 		int rtick;
2297 
2298 		sched_pctcpu_update(ts);
2299 		/* How many rtick per second ? */
2300 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2301 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2302 	}
2303 	thread_unlock(td);
2304 
2305 	return (pctcpu);
2306 }
2307 
2308 /*
2309  * Enforce affinity settings for a thread.  Called after adjustments to
2310  * cpumask.
2311  */
2312 void
2313 sched_affinity(struct thread *td)
2314 {
2315 #ifdef SMP
2316 	struct td_sched *ts;
2317 	int cpu;
2318 
2319 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2320 	ts = td->td_sched;
2321 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2322 		return;
2323 	if (!TD_IS_RUNNING(td))
2324 		return;
2325 	td->td_flags |= TDF_NEEDRESCHED;
2326 	if (!THREAD_CAN_MIGRATE(td))
2327 		return;
2328 	/*
2329 	 * Assign the new cpu and force a switch before returning to
2330 	 * userspace.  If the target thread is not running locally send
2331 	 * an ipi to force the issue.
2332 	 */
2333 	cpu = ts->ts_cpu;
2334 	ts->ts_cpu = sched_pickcpu(ts, 0);
2335 	if (cpu != PCPU_GET(cpuid))
2336 		ipi_selected(1 << cpu, IPI_PREEMPT);
2337 #endif
2338 }
2339 
2340 /*
2341  * Bind a thread to a target cpu.
2342  */
2343 void
2344 sched_bind(struct thread *td, int cpu)
2345 {
2346 	struct td_sched *ts;
2347 
2348 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2349 	ts = td->td_sched;
2350 	if (ts->ts_flags & TSF_BOUND)
2351 		sched_unbind(td);
2352 	ts->ts_flags |= TSF_BOUND;
2353 	sched_pin();
2354 	if (PCPU_GET(cpuid) == cpu)
2355 		return;
2356 	ts->ts_cpu = cpu;
2357 	/* When we return from mi_switch we'll be on the correct cpu. */
2358 	mi_switch(SW_VOL, NULL);
2359 }
2360 
2361 /*
2362  * Release a bound thread.
2363  */
2364 void
2365 sched_unbind(struct thread *td)
2366 {
2367 	struct td_sched *ts;
2368 
2369 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2370 	ts = td->td_sched;
2371 	if ((ts->ts_flags & TSF_BOUND) == 0)
2372 		return;
2373 	ts->ts_flags &= ~TSF_BOUND;
2374 	sched_unpin();
2375 }
2376 
2377 int
2378 sched_is_bound(struct thread *td)
2379 {
2380 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2381 	return (td->td_sched->ts_flags & TSF_BOUND);
2382 }
2383 
2384 /*
2385  * Basic yield call.
2386  */
2387 void
2388 sched_relinquish(struct thread *td)
2389 {
2390 	thread_lock(td);
2391 	SCHED_STAT_INC(switch_relinquish);
2392 	mi_switch(SW_VOL, NULL);
2393 	thread_unlock(td);
2394 }
2395 
2396 /*
2397  * Return the total system load.
2398  */
2399 int
2400 sched_load(void)
2401 {
2402 #ifdef SMP
2403 	int total;
2404 	int i;
2405 
2406 	total = 0;
2407 	for (i = 0; i <= mp_maxid; i++)
2408 		total += TDQ_CPU(i)->tdq_sysload;
2409 	return (total);
2410 #else
2411 	return (TDQ_SELF()->tdq_sysload);
2412 #endif
2413 }
2414 
2415 int
2416 sched_sizeof_proc(void)
2417 {
2418 	return (sizeof(struct proc));
2419 }
2420 
2421 int
2422 sched_sizeof_thread(void)
2423 {
2424 	return (sizeof(struct thread) + sizeof(struct td_sched));
2425 }
2426 
2427 /*
2428  * The actual idle process.
2429  */
2430 void
2431 sched_idletd(void *dummy)
2432 {
2433 	struct thread *td;
2434 	struct tdq *tdq;
2435 
2436 	td = curthread;
2437 	tdq = TDQ_SELF();
2438 	mtx_assert(&Giant, MA_NOTOWNED);
2439 	/* ULE relies on preemption for idle interruption. */
2440 	for (;;) {
2441 #ifdef SMP
2442 		if (tdq_idled(tdq))
2443 			cpu_idle();
2444 #else
2445 		cpu_idle();
2446 #endif
2447 	}
2448 }
2449 
2450 /*
2451  * A CPU is entering for the first time or a thread is exiting.
2452  */
2453 void
2454 sched_throw(struct thread *td)
2455 {
2456 	struct thread *newtd;
2457 	struct tdq *tdq;
2458 
2459 	tdq = TDQ_SELF();
2460 	if (td == NULL) {
2461 		/* Correct spinlock nesting and acquire the correct lock. */
2462 		TDQ_LOCK(tdq);
2463 		spinlock_exit();
2464 	} else {
2465 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2466 		tdq_load_rem(tdq, td->td_sched);
2467 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2468 	}
2469 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2470 	newtd = choosethread();
2471 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2472 	PCPU_SET(switchtime, cpu_ticks());
2473 	PCPU_SET(switchticks, ticks);
2474 	cpu_throw(td, newtd);		/* doesn't return */
2475 }
2476 
2477 /*
2478  * This is called from fork_exit().  Just acquire the correct locks and
2479  * let fork do the rest of the work.
2480  */
2481 void
2482 sched_fork_exit(struct thread *td)
2483 {
2484 	struct td_sched *ts;
2485 	struct tdq *tdq;
2486 	int cpuid;
2487 
2488 	/*
2489 	 * Finish setting up thread glue so that it begins execution in a
2490 	 * non-nested critical section with the scheduler lock held.
2491 	 */
2492 	cpuid = PCPU_GET(cpuid);
2493 	tdq = TDQ_CPU(cpuid);
2494 	ts = td->td_sched;
2495 	if (TD_IS_IDLETHREAD(td))
2496 		td->td_lock = TDQ_LOCKPTR(tdq);
2497 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2498 	td->td_oncpu = cpuid;
2499 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2500 	lock_profile_obtain_lock_success(
2501 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2502 }
2503 
2504 static SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0,
2505     "Scheduler");
2506 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2507     "Scheduler name");
2508 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2509     "Slice size for timeshare threads");
2510 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2511      "Interactivity score threshold");
2512 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh,
2513      0,"Min priority for preemption, lower priorities have greater precedence");
2514 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost,
2515      0,"Controls whether static kernel priorities are assigned to sleeping threads.");
2516 #ifdef SMP
2517 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2518     "Number of hz ticks to keep thread affinity for");
2519 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2520     "Enables the long-term load balancer");
2521 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2522     &balance_interval, 0,
2523     "Average frequency in stathz ticks to run the long-term balancer");
2524 SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0,
2525     "Steals work from another hyper-threaded core on idle");
2526 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2527     "Attempts to steal work from other cores before idling");
2528 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2529     "Minimum load on remote cpu before we'll steal");
2530 #endif
2531 
2532 /* ps compat.  All cpu percentages from ULE are weighted. */
2533 static int ccpu = 0;
2534 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2535 
2536 
2537 #define KERN_SWITCH_INCLUDE 1
2538 #include "kern/kern_switch.c"
2539