xref: /freebsd/sys/kern/sched_ule.c (revision 282a3889ebf826db9839be296ff1dd903f6d6d6e)
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 it's 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 #ifdef KTRACE
63 #include <sys/uio.h>
64 #include <sys/ktrace.h>
65 #endif
66 
67 #ifdef HWPMC_HOOKS
68 #include <sys/pmckern.h>
69 #endif
70 
71 #include <machine/cpu.h>
72 #include <machine/smp.h>
73 
74 #ifndef PREEMPTION
75 #error	"SCHED_ULE requires options PREEMPTION"
76 #endif
77 
78 #define	KTR_ULE	0
79 
80 /*
81  * Thread scheduler specific section.  All fields are protected
82  * by the thread lock.
83  */
84 struct td_sched {
85 	TAILQ_ENTRY(td_sched) ts_procq;	/* Run queue. */
86 	struct thread	*ts_thread;	/* Active associated thread. */
87 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
88 	short		ts_flags;	/* TSF_* flags. */
89 	u_char		ts_rqindex;	/* Run queue index. */
90 	u_char		ts_cpu;		/* CPU that we have affinity for. */
91 	int		ts_slptick;	/* Tick when we went to sleep. */
92 	int		ts_slice;	/* Ticks of slice remaining. */
93 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
94 	u_int		ts_runtime;	/* Number of ticks we were running */
95 	/* The following variables are only used for pctcpu calculation */
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 #ifdef SMP
100 	int		ts_rltick;	/* Real last tick, for affinity. */
101 #endif
102 };
103 /* flags kept in ts_flags */
104 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
105 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
106 
107 static struct td_sched td_sched0;
108 
109 /*
110  * Cpu percentage computation macros and defines.
111  *
112  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
113  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
114  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
115  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
116  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
117  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
118  */
119 #define	SCHED_TICK_SECS		10
120 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
121 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
122 #define	SCHED_TICK_SHIFT	10
123 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
124 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
125 
126 /*
127  * These macros determine priorities for non-interactive threads.  They are
128  * assigned a priority based on their recent cpu utilization as expressed
129  * by the ratio of ticks to the tick total.  NHALF priorities at the start
130  * and end of the MIN to MAX timeshare range are only reachable with negative
131  * or positive nice respectively.
132  *
133  * PRI_RANGE:	Priority range for utilization dependent priorities.
134  * PRI_NRESV:	Number of nice values.
135  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
136  * PRI_NICE:	Determines the part of the priority inherited from nice.
137  */
138 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
139 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
140 #define	SCHED_PRI_MIN		(PRI_MIN_TIMESHARE + SCHED_PRI_NHALF)
141 #define	SCHED_PRI_MAX		(PRI_MAX_TIMESHARE - SCHED_PRI_NHALF)
142 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN)
143 #define	SCHED_PRI_TICKS(ts)						\
144     (SCHED_TICK_HZ((ts)) /						\
145     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
146 #define	SCHED_PRI_NICE(nice)	(nice)
147 
148 /*
149  * These determine the interactivity of a process.  Interactivity differs from
150  * cpu utilization in that it expresses the voluntary time slept vs time ran
151  * while cpu utilization includes all time not running.  This more accurately
152  * models the intent of the thread.
153  *
154  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
155  *		before throttling back.
156  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
157  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
158  * INTERACT_THRESH:	Threshhold for placement on the current runq.
159  */
160 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
161 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
162 #define	SCHED_INTERACT_MAX	(100)
163 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
164 #define	SCHED_INTERACT_THRESH	(30)
165 
166 /*
167  * tickincr:		Converts a stathz tick into a hz domain scaled by
168  *			the shift factor.  Without the shift the error rate
169  *			due to rounding would be unacceptably high.
170  * realstathz:		stathz is sometimes 0 and run off of hz.
171  * sched_slice:		Runtime of each thread before rescheduling.
172  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
173  */
174 static int sched_interact = SCHED_INTERACT_THRESH;
175 static int realstathz;
176 static int tickincr;
177 static int sched_slice;
178 static int preempt_thresh = PRI_MIN_KERN;
179 
180 /*
181  * tdq - per processor runqs and statistics.  All fields are protected by the
182  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
183  * locking in sched_pickcpu();
184  */
185 struct tdq {
186 	struct mtx	tdq_lock;		/* Protects all fields below. */
187 	struct runq	tdq_realtime;		/* real-time run queue. */
188 	struct runq	tdq_timeshare;		/* timeshare run queue. */
189 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
190 	int		tdq_load;		/* Aggregate load. */
191 	u_char		tdq_idx;		/* Current insert index. */
192 	u_char		tdq_ridx;		/* Current removal index. */
193 #ifdef SMP
194 	u_char		tdq_lowpri;		/* Lowest priority thread. */
195 	int		tdq_transferable;	/* Transferable thread count. */
196 	LIST_ENTRY(tdq)	tdq_siblings;		/* Next in tdq group. */
197 	struct tdq_group *tdq_group;		/* Our processor group. */
198 #else
199 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
200 #endif
201 	char		tdq_name[16];		/* lock name. */
202 } __aligned(64);
203 
204 
205 #ifdef SMP
206 /*
207  * tdq groups are groups of processors which can cheaply share threads.  When
208  * one processor in the group goes idle it will check the runqs of the other
209  * processors in its group prior to halting and waiting for an interrupt.
210  * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA.
211  * In a numa environment we'd want an idle bitmap per group and a two tiered
212  * load balancer.
213  */
214 struct tdq_group {
215 	int	tdg_cpus;		/* Count of CPUs in this tdq group. */
216 	cpumask_t tdg_cpumask;		/* Mask of cpus in this group. */
217 	cpumask_t tdg_idlemask;		/* Idle cpus in this group. */
218 	cpumask_t tdg_mask;		/* Bit mask for first cpu. */
219 	int	tdg_load;		/* Total load of this group. */
220 	int	tdg_transferable;	/* Transferable load of this group. */
221 	LIST_HEAD(, tdq) tdg_members;	/* Linked list of all members. */
222 } __aligned(64);
223 
224 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 300))
225 #define	SCHED_AFFINITY(ts)	((ts)->ts_rltick > ticks - affinity)
226 
227 /*
228  * Run-time tunables.
229  */
230 static int rebalance = 1;
231 static int balance_secs = 1;
232 static int pick_pri = 1;
233 static int affinity;
234 static int tryself = 1;
235 static int steal_htt = 0;
236 static int steal_idle = 1;
237 static int steal_thresh = 2;
238 static int topology = 0;
239 
240 /*
241  * One thread queue per processor.
242  */
243 static volatile cpumask_t tdq_idle;
244 static int tdg_maxid;
245 static struct tdq	tdq_cpu[MAXCPU];
246 static struct tdq_group tdq_groups[MAXCPU];
247 static struct callout balco;
248 static struct callout gbalco;
249 
250 #define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
251 #define	TDQ_CPU(x)	(&tdq_cpu[(x)])
252 #define	TDQ_ID(x)	((x) - tdq_cpu)
253 #define	TDQ_GROUP(x)	(&tdq_groups[(x)])
254 #else	/* !SMP */
255 static struct tdq	tdq_cpu;
256 
257 #define	TDQ_ID(x)	(0)
258 #define	TDQ_SELF()	(&tdq_cpu)
259 #define	TDQ_CPU(x)	(&tdq_cpu)
260 #endif
261 
262 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
263 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
264 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
265 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
266 #define	TDQ_LOCKPTR(t)		(&(t)->tdq_lock)
267 
268 static void sched_priority(struct thread *);
269 static void sched_thread_priority(struct thread *, u_char);
270 static int sched_interact_score(struct thread *);
271 static void sched_interact_update(struct thread *);
272 static void sched_interact_fork(struct thread *);
273 static void sched_pctcpu_update(struct td_sched *);
274 
275 /* Operations on per processor queues */
276 static struct td_sched * tdq_choose(struct tdq *);
277 static void tdq_setup(struct tdq *);
278 static void tdq_load_add(struct tdq *, struct td_sched *);
279 static void tdq_load_rem(struct tdq *, struct td_sched *);
280 static __inline void tdq_runq_add(struct tdq *, struct td_sched *, int);
281 static __inline void tdq_runq_rem(struct tdq *, struct td_sched *);
282 void tdq_print(int cpu);
283 static void runq_print(struct runq *rq);
284 static void tdq_add(struct tdq *, struct thread *, int);
285 #ifdef SMP
286 static void tdq_move(struct tdq *, struct tdq *);
287 static int tdq_idled(struct tdq *);
288 static void tdq_notify(struct td_sched *);
289 static struct td_sched *tdq_steal(struct tdq *, int);
290 static struct td_sched *runq_steal(struct runq *);
291 static int sched_pickcpu(struct td_sched *, int);
292 static void sched_balance(void *);
293 static void sched_balance_groups(void *);
294 static void sched_balance_group(struct tdq_group *);
295 static void sched_balance_pair(struct tdq *, struct tdq *);
296 static inline struct tdq *sched_setcpu(struct td_sched *, int, int);
297 static inline struct mtx *thread_block_switch(struct thread *);
298 static inline void thread_unblock_switch(struct thread *, struct mtx *);
299 
300 #define	THREAD_CAN_MIGRATE(td)	 ((td)->td_pinned == 0)
301 #endif
302 
303 static void sched_setup(void *dummy);
304 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL)
305 
306 static void sched_initticks(void *dummy);
307 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, NULL)
308 
309 /*
310  * Print the threads waiting on a run-queue.
311  */
312 static void
313 runq_print(struct runq *rq)
314 {
315 	struct rqhead *rqh;
316 	struct td_sched *ts;
317 	int pri;
318 	int j;
319 	int i;
320 
321 	for (i = 0; i < RQB_LEN; i++) {
322 		printf("\t\trunq bits %d 0x%zx\n",
323 		    i, rq->rq_status.rqb_bits[i]);
324 		for (j = 0; j < RQB_BPW; j++)
325 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
326 				pri = j + (i << RQB_L2BPW);
327 				rqh = &rq->rq_queues[pri];
328 				TAILQ_FOREACH(ts, rqh, ts_procq) {
329 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
330 					    ts->ts_thread, ts->ts_thread->td_proc->p_comm, ts->ts_thread->td_priority, ts->ts_rqindex, pri);
331 				}
332 			}
333 	}
334 }
335 
336 /*
337  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
338  */
339 void
340 tdq_print(int cpu)
341 {
342 	struct tdq *tdq;
343 
344 	tdq = TDQ_CPU(cpu);
345 
346 	printf("tdq:\n");
347 	printf("\tlockptr         %p\n", TDQ_LOCKPTR(tdq));
348 	printf("\tlock name       %s\n", tdq->tdq_name);
349 	printf("\tload:           %d\n", tdq->tdq_load);
350 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
351 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
352 	printf("\trealtime runq:\n");
353 	runq_print(&tdq->tdq_realtime);
354 	printf("\ttimeshare runq:\n");
355 	runq_print(&tdq->tdq_timeshare);
356 	printf("\tidle runq:\n");
357 	runq_print(&tdq->tdq_idle);
358 #ifdef SMP
359 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
360 	printf("\tlowest priority: %d\n", tdq->tdq_lowpri);
361 #endif
362 }
363 
364 #define	TS_RQ_PPQ	(((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS)
365 /*
366  * Add a thread to the actual run-queue.  Keeps transferable counts up to
367  * date with what is actually on the run-queue.  Selects the correct
368  * queue position for timeshare threads.
369  */
370 static __inline void
371 tdq_runq_add(struct tdq *tdq, struct td_sched *ts, int flags)
372 {
373 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
374 	THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED);
375 #ifdef SMP
376 	if (THREAD_CAN_MIGRATE(ts->ts_thread)) {
377 		tdq->tdq_transferable++;
378 		tdq->tdq_group->tdg_transferable++;
379 		ts->ts_flags |= TSF_XFERABLE;
380 	}
381 #endif
382 	if (ts->ts_runq == &tdq->tdq_timeshare) {
383 		u_char pri;
384 
385 		pri = ts->ts_thread->td_priority;
386 		KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE,
387 			("Invalid priority %d on timeshare runq", pri));
388 		/*
389 		 * This queue contains only priorities between MIN and MAX
390 		 * realtime.  Use the whole queue to represent these values.
391 		 */
392 		if ((flags & SRQ_BORROWING) == 0) {
393 			pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ;
394 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
395 			/*
396 			 * This effectively shortens the queue by one so we
397 			 * can have a one slot difference between idx and
398 			 * ridx while we wait for threads to drain.
399 			 */
400 			if (tdq->tdq_ridx != tdq->tdq_idx &&
401 			    pri == tdq->tdq_ridx)
402 				pri = (unsigned char)(pri - 1) % RQ_NQS;
403 		} else
404 			pri = tdq->tdq_ridx;
405 		runq_add_pri(ts->ts_runq, ts, pri, flags);
406 	} else
407 		runq_add(ts->ts_runq, ts, flags);
408 }
409 
410 /*
411  * Remove a thread from a run-queue.  This typically happens when a thread
412  * is selected to run.  Running threads are not on the queue and the
413  * transferable count does not reflect them.
414  */
415 static __inline void
416 tdq_runq_rem(struct tdq *tdq, struct td_sched *ts)
417 {
418 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
419 	KASSERT(ts->ts_runq != NULL,
420 	    ("tdq_runq_remove: thread %p null ts_runq", ts->ts_thread));
421 #ifdef SMP
422 	if (ts->ts_flags & TSF_XFERABLE) {
423 		tdq->tdq_transferable--;
424 		tdq->tdq_group->tdg_transferable--;
425 		ts->ts_flags &= ~TSF_XFERABLE;
426 	}
427 #endif
428 	if (ts->ts_runq == &tdq->tdq_timeshare) {
429 		if (tdq->tdq_idx != tdq->tdq_ridx)
430 			runq_remove_idx(ts->ts_runq, ts, &tdq->tdq_ridx);
431 		else
432 			runq_remove_idx(ts->ts_runq, ts, NULL);
433 		/*
434 		 * For timeshare threads we update the priority here so
435 		 * the priority reflects the time we've been sleeping.
436 		 */
437 		ts->ts_ltick = ticks;
438 		sched_pctcpu_update(ts);
439 		sched_priority(ts->ts_thread);
440 	} else
441 		runq_remove(ts->ts_runq, ts);
442 }
443 
444 /*
445  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
446  * for this thread to the referenced thread queue.
447  */
448 static void
449 tdq_load_add(struct tdq *tdq, struct td_sched *ts)
450 {
451 	int class;
452 
453 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
454 	THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED);
455 	class = PRI_BASE(ts->ts_thread->td_pri_class);
456 	tdq->tdq_load++;
457 	CTR2(KTR_SCHED, "cpu %jd load: %d", TDQ_ID(tdq), tdq->tdq_load);
458 	if (class != PRI_ITHD &&
459 	    (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0)
460 #ifdef SMP
461 		tdq->tdq_group->tdg_load++;
462 #else
463 		tdq->tdq_sysload++;
464 #endif
465 }
466 
467 /*
468  * Remove the load from a thread that is transitioning to a sleep state or
469  * exiting.
470  */
471 static void
472 tdq_load_rem(struct tdq *tdq, struct td_sched *ts)
473 {
474 	int class;
475 
476 	THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED);
477 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
478 	class = PRI_BASE(ts->ts_thread->td_pri_class);
479 	if (class != PRI_ITHD &&
480 	    (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0)
481 #ifdef SMP
482 		tdq->tdq_group->tdg_load--;
483 #else
484 		tdq->tdq_sysload--;
485 #endif
486 	KASSERT(tdq->tdq_load != 0,
487 	    ("tdq_load_rem: Removing with 0 load on queue %d", (int)TDQ_ID(tdq)));
488 	tdq->tdq_load--;
489 	CTR1(KTR_SCHED, "load: %d", tdq->tdq_load);
490 	ts->ts_runq = NULL;
491 }
492 
493 #ifdef SMP
494 /*
495  * sched_balance is a simple CPU load balancing algorithm.  It operates by
496  * finding the least loaded and most loaded cpu and equalizing their load
497  * by migrating some processes.
498  *
499  * Dealing only with two CPUs at a time has two advantages.  Firstly, most
500  * installations will only have 2 cpus.  Secondly, load balancing too much at
501  * once can have an unpleasant effect on the system.  The scheduler rarely has
502  * enough information to make perfect decisions.  So this algorithm chooses
503  * simplicity and more gradual effects on load in larger systems.
504  *
505  */
506 static void
507 sched_balance(void *arg)
508 {
509 	struct tdq_group *high;
510 	struct tdq_group *low;
511 	struct tdq_group *tdg;
512 	int cnt;
513 	int i;
514 
515 	callout_reset(&balco, max(hz / 2, random() % (hz * balance_secs)),
516 	    sched_balance, NULL);
517 	if (smp_started == 0 || rebalance == 0)
518 		return;
519 	low = high = NULL;
520 	i = random() % (tdg_maxid + 1);
521 	for (cnt = 0; cnt <= tdg_maxid; cnt++) {
522 		tdg = TDQ_GROUP(i);
523 		/*
524 		 * Find the CPU with the highest load that has some
525 		 * threads to transfer.
526 		 */
527 		if ((high == NULL || tdg->tdg_load > high->tdg_load)
528 		    && tdg->tdg_transferable)
529 			high = tdg;
530 		if (low == NULL || tdg->tdg_load < low->tdg_load)
531 			low = tdg;
532 		if (++i > tdg_maxid)
533 			i = 0;
534 	}
535 	if (low != NULL && high != NULL && high != low)
536 		sched_balance_pair(LIST_FIRST(&high->tdg_members),
537 		    LIST_FIRST(&low->tdg_members));
538 }
539 
540 /*
541  * Balance load between CPUs in a group.  Will only migrate within the group.
542  */
543 static void
544 sched_balance_groups(void *arg)
545 {
546 	int i;
547 
548 	callout_reset(&gbalco, max(hz / 2, random() % (hz * balance_secs)),
549 	    sched_balance_groups, NULL);
550 	if (smp_started == 0 || rebalance == 0)
551 		return;
552 	for (i = 0; i <= tdg_maxid; i++)
553 		sched_balance_group(TDQ_GROUP(i));
554 }
555 
556 /*
557  * Finds the greatest imbalance between two tdqs in a group.
558  */
559 static void
560 sched_balance_group(struct tdq_group *tdg)
561 {
562 	struct tdq *tdq;
563 	struct tdq *high;
564 	struct tdq *low;
565 	int load;
566 
567 	if (tdg->tdg_transferable == 0)
568 		return;
569 	low = NULL;
570 	high = NULL;
571 	LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) {
572 		load = tdq->tdq_load;
573 		if (high == NULL || load > high->tdq_load)
574 			high = tdq;
575 		if (low == NULL || load < low->tdq_load)
576 			low = tdq;
577 	}
578 	if (high != NULL && low != NULL && high != low)
579 		sched_balance_pair(high, low);
580 }
581 
582 /*
583  * Lock two thread queues using their address to maintain lock order.
584  */
585 static void
586 tdq_lock_pair(struct tdq *one, struct tdq *two)
587 {
588 	if (one < two) {
589 		TDQ_LOCK(one);
590 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
591 	} else {
592 		TDQ_LOCK(two);
593 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
594 	}
595 }
596 
597 /*
598  * Transfer load between two imbalanced thread queues.
599  */
600 static void
601 sched_balance_pair(struct tdq *high, struct tdq *low)
602 {
603 	int transferable;
604 	int high_load;
605 	int low_load;
606 	int move;
607 	int diff;
608 	int i;
609 
610 	tdq_lock_pair(high, low);
611 	/*
612 	 * If we're transfering within a group we have to use this specific
613 	 * tdq's transferable count, otherwise we can steal from other members
614 	 * of the group.
615 	 */
616 	if (high->tdq_group == low->tdq_group) {
617 		transferable = high->tdq_transferable;
618 		high_load = high->tdq_load;
619 		low_load = low->tdq_load;
620 	} else {
621 		transferable = high->tdq_group->tdg_transferable;
622 		high_load = high->tdq_group->tdg_load;
623 		low_load = low->tdq_group->tdg_load;
624 	}
625 	/*
626 	 * Determine what the imbalance is and then adjust that to how many
627 	 * threads we actually have to give up (transferable).
628 	 */
629 	if (transferable != 0) {
630 		diff = high_load - low_load;
631 		move = diff / 2;
632 		if (diff & 0x1)
633 			move++;
634 		move = min(move, transferable);
635 		for (i = 0; i < move; i++)
636 			tdq_move(high, low);
637 	}
638 	TDQ_UNLOCK(high);
639 	TDQ_UNLOCK(low);
640 	return;
641 }
642 
643 /*
644  * Move a thread from one thread queue to another.
645  */
646 static void
647 tdq_move(struct tdq *from, struct tdq *to)
648 {
649 	struct td_sched *ts;
650 	struct thread *td;
651 	struct tdq *tdq;
652 	int cpu;
653 
654 	tdq = from;
655 	cpu = TDQ_ID(to);
656 	ts = tdq_steal(tdq, 1);
657 	if (ts == NULL) {
658 		struct tdq_group *tdg;
659 
660 		tdg = tdq->tdq_group;
661 		LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) {
662 			if (tdq == from || tdq->tdq_transferable == 0)
663 				continue;
664 			ts = tdq_steal(tdq, 1);
665 			break;
666 		}
667 		if (ts == NULL)
668 			return;
669 	}
670 	if (tdq == to)
671 		return;
672 	td = ts->ts_thread;
673 	/*
674 	 * Although the run queue is locked the thread may be blocked.  Lock
675 	 * it to clear this.
676 	 */
677 	thread_lock(td);
678 	/* Drop recursive lock on from. */
679 	TDQ_UNLOCK(from);
680 	sched_rem(td);
681 	ts->ts_cpu = cpu;
682 	td->td_lock = TDQ_LOCKPTR(to);
683 	tdq_add(to, td, SRQ_YIELDING);
684 	tdq_notify(ts);
685 }
686 
687 /*
688  * This tdq has idled.  Try to steal a thread from another cpu and switch
689  * to it.
690  */
691 static int
692 tdq_idled(struct tdq *tdq)
693 {
694 	struct tdq_group *tdg;
695 	struct tdq *steal;
696 	struct td_sched *ts;
697 	struct thread *td;
698 	int highload;
699 	int highcpu;
700 	int load;
701 	int cpu;
702 
703 	/* We don't want to be preempted while we're iterating over tdqs */
704 	spinlock_enter();
705 	tdg = tdq->tdq_group;
706 	/*
707 	 * If we're in a cpu group, try and steal threads from another cpu in
708 	 * the group before idling.
709 	 */
710 	if (steal_htt && tdg->tdg_cpus > 1 && tdg->tdg_transferable) {
711 		LIST_FOREACH(steal, &tdg->tdg_members, tdq_siblings) {
712 			if (steal == tdq || steal->tdq_transferable == 0)
713 				continue;
714 			TDQ_LOCK(steal);
715 			ts = tdq_steal(steal, 0);
716 			if (ts)
717 				goto steal;
718 			TDQ_UNLOCK(steal);
719 		}
720 	}
721 	for (;;) {
722 		if (steal_idle == 0)
723 			break;
724 		highcpu = 0;
725 		highload = 0;
726 		for (cpu = 0; cpu <= mp_maxid; cpu++) {
727 			if (CPU_ABSENT(cpu))
728 				continue;
729 			steal = TDQ_CPU(cpu);
730 			load = TDQ_CPU(cpu)->tdq_transferable;
731 			if (load < highload)
732 				continue;
733 			highload = load;
734 			highcpu = cpu;
735 		}
736 		if (highload < steal_thresh)
737 			break;
738 		steal = TDQ_CPU(highcpu);
739 		TDQ_LOCK(steal);
740 		if (steal->tdq_transferable >= steal_thresh &&
741 		    (ts = tdq_steal(steal, 1)) != NULL)
742 			goto steal;
743 		TDQ_UNLOCK(steal);
744 		break;
745 	}
746 	spinlock_exit();
747 	return (1);
748 steal:
749 	td = ts->ts_thread;
750 	thread_lock(td);
751 	spinlock_exit();
752 	MPASS(td->td_lock == TDQ_LOCKPTR(steal));
753 	TDQ_UNLOCK(steal);
754 	sched_rem(td);
755 	sched_setcpu(ts, PCPU_GET(cpuid), SRQ_YIELDING);
756 	tdq_add(tdq, td, SRQ_YIELDING);
757 	MPASS(td->td_lock == curthread->td_lock);
758 	mi_switch(SW_VOL, NULL);
759 	thread_unlock(curthread);
760 
761 	return (0);
762 }
763 
764 /*
765  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
766  */
767 static void
768 tdq_notify(struct td_sched *ts)
769 {
770 	struct thread *ctd;
771 	struct pcpu *pcpu;
772 	int cpri;
773 	int pri;
774 	int cpu;
775 
776 	cpu = ts->ts_cpu;
777 	pri = ts->ts_thread->td_priority;
778 	pcpu = pcpu_find(cpu);
779 	ctd = pcpu->pc_curthread;
780 	cpri = ctd->td_priority;
781 
782 	/*
783 	 * If our priority is not better than the current priority there is
784 	 * nothing to do.
785 	 */
786 	if (pri > cpri)
787 		return;
788 	/*
789 	 * Always IPI idle.
790 	 */
791 	if (cpri > PRI_MIN_IDLE)
792 		goto sendipi;
793 	/*
794 	 * If we're realtime or better and there is timeshare or worse running
795 	 * send an IPI.
796 	 */
797 	if (pri < PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME)
798 		goto sendipi;
799 	/*
800 	 * Otherwise only IPI if we exceed the threshold.
801 	 */
802 	if (pri > preempt_thresh)
803 		return;
804 sendipi:
805 	ctd->td_flags |= TDF_NEEDRESCHED;
806 	ipi_selected(1 << cpu, IPI_PREEMPT);
807 }
808 
809 /*
810  * Steals load from a timeshare queue.  Honors the rotating queue head
811  * index.
812  */
813 static struct td_sched *
814 runq_steal_from(struct runq *rq, u_char start)
815 {
816 	struct td_sched *ts;
817 	struct rqbits *rqb;
818 	struct rqhead *rqh;
819 	int first;
820 	int bit;
821 	int pri;
822 	int i;
823 
824 	rqb = &rq->rq_status;
825 	bit = start & (RQB_BPW -1);
826 	pri = 0;
827 	first = 0;
828 again:
829 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
830 		if (rqb->rqb_bits[i] == 0)
831 			continue;
832 		if (bit != 0) {
833 			for (pri = bit; pri < RQB_BPW; pri++)
834 				if (rqb->rqb_bits[i] & (1ul << pri))
835 					break;
836 			if (pri >= RQB_BPW)
837 				continue;
838 		} else
839 			pri = RQB_FFS(rqb->rqb_bits[i]);
840 		pri += (i << RQB_L2BPW);
841 		rqh = &rq->rq_queues[pri];
842 		TAILQ_FOREACH(ts, rqh, ts_procq) {
843 			if (first && THREAD_CAN_MIGRATE(ts->ts_thread))
844 				return (ts);
845 			first = 1;
846 		}
847 	}
848 	if (start != 0) {
849 		start = 0;
850 		goto again;
851 	}
852 
853 	return (NULL);
854 }
855 
856 /*
857  * Steals load from a standard linear queue.
858  */
859 static struct td_sched *
860 runq_steal(struct runq *rq)
861 {
862 	struct rqhead *rqh;
863 	struct rqbits *rqb;
864 	struct td_sched *ts;
865 	int word;
866 	int bit;
867 
868 	rqb = &rq->rq_status;
869 	for (word = 0; word < RQB_LEN; word++) {
870 		if (rqb->rqb_bits[word] == 0)
871 			continue;
872 		for (bit = 0; bit < RQB_BPW; bit++) {
873 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
874 				continue;
875 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
876 			TAILQ_FOREACH(ts, rqh, ts_procq)
877 				if (THREAD_CAN_MIGRATE(ts->ts_thread))
878 					return (ts);
879 		}
880 	}
881 	return (NULL);
882 }
883 
884 /*
885  * Attempt to steal a thread in priority order from a thread queue.
886  */
887 static struct td_sched *
888 tdq_steal(struct tdq *tdq, int stealidle)
889 {
890 	struct td_sched *ts;
891 
892 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
893 	if ((ts = runq_steal(&tdq->tdq_realtime)) != NULL)
894 		return (ts);
895 	if ((ts = runq_steal_from(&tdq->tdq_timeshare, tdq->tdq_ridx)) != NULL)
896 		return (ts);
897 	if (stealidle)
898 		return (runq_steal(&tdq->tdq_idle));
899 	return (NULL);
900 }
901 
902 /*
903  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
904  * current lock and returns with the assigned queue locked.  If this is
905  * via sched_switch() we leave the thread in a blocked state as an
906  * optimization.
907  */
908 static inline struct tdq *
909 sched_setcpu(struct td_sched *ts, int cpu, int flags)
910 {
911 	struct thread *td;
912 	struct tdq *tdq;
913 
914 	THREAD_LOCK_ASSERT(ts->ts_thread, MA_OWNED);
915 
916 	tdq = TDQ_CPU(cpu);
917 	td = ts->ts_thread;
918 	ts->ts_cpu = cpu;
919 	if (td->td_lock == TDQ_LOCKPTR(tdq))
920 		return (tdq);
921 #ifdef notyet
922 	/*
923 	 * If the thread isn't running it's lockptr is a
924 	 * turnstile or a sleepqueue.  We can just lock_set without
925 	 * blocking.
926 	 */
927 	if (TD_CAN_RUN(td)) {
928 		TDQ_LOCK(tdq);
929 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
930 		return (tdq);
931 	}
932 #endif
933 	/*
934 	 * The hard case, migration, we need to block the thread first to
935 	 * prevent order reversals with other cpus locks.
936 	 */
937 	thread_lock_block(td);
938 	TDQ_LOCK(tdq);
939 	/* Return to sched_switch() with the lock still blocked */
940 	if ((flags & SRQ_OURSELF) == 0)
941 		thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
942 	return (tdq);
943 }
944 
945 /*
946  * Find the thread queue running the lowest priority thread.
947  */
948 static int
949 tdq_lowestpri(void)
950 {
951 	struct tdq *tdq;
952 	int lowpri;
953 	int lowcpu;
954 	int lowload;
955 	int load;
956 	int cpu;
957 	int pri;
958 
959 	lowload = 0;
960 	lowpri = lowcpu = 0;
961 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
962 		if (CPU_ABSENT(cpu))
963 			continue;
964 		tdq = TDQ_CPU(cpu);
965 		pri = tdq->tdq_lowpri;
966 		load = TDQ_CPU(cpu)->tdq_load;
967 		CTR4(KTR_ULE,
968 		    "cpu %d pri %d lowcpu %d lowpri %d",
969 		    cpu, pri, lowcpu, lowpri);
970 		if (pri < lowpri)
971 			continue;
972 		if (lowpri && lowpri == pri && load > lowload)
973 			continue;
974 		lowpri = pri;
975 		lowcpu = cpu;
976 		lowload = load;
977 	}
978 
979 	return (lowcpu);
980 }
981 
982 /*
983  * Find the thread queue with the least load.
984  */
985 static int
986 tdq_lowestload(void)
987 {
988 	struct tdq *tdq;
989 	int lowload;
990 	int lowpri;
991 	int lowcpu;
992 	int load;
993 	int cpu;
994 	int pri;
995 
996 	lowcpu = 0;
997 	lowload = TDQ_CPU(0)->tdq_load;
998 	lowpri = TDQ_CPU(0)->tdq_lowpri;
999 	for (cpu = 1; cpu <= mp_maxid; cpu++) {
1000 		if (CPU_ABSENT(cpu))
1001 			continue;
1002 		tdq = TDQ_CPU(cpu);
1003 		load = tdq->tdq_load;
1004 		pri = tdq->tdq_lowpri;
1005 		CTR4(KTR_ULE, "cpu %d load %d lowcpu %d lowload %d",
1006 		    cpu, load, lowcpu, lowload);
1007 		if (load > lowload)
1008 			continue;
1009 		if (load == lowload && pri < lowpri)
1010 			continue;
1011 		lowcpu = cpu;
1012 		lowload = load;
1013 		lowpri = pri;
1014 	}
1015 
1016 	return (lowcpu);
1017 }
1018 
1019 /*
1020  * Pick the destination cpu for sched_add().  Respects affinity and makes
1021  * a determination based on load or priority of available processors.
1022  */
1023 static int
1024 sched_pickcpu(struct td_sched *ts, int flags)
1025 {
1026 	struct tdq *tdq;
1027 	int self;
1028 	int pri;
1029 	int cpu;
1030 
1031 	cpu = self = PCPU_GET(cpuid);
1032 	if (smp_started == 0)
1033 		return (self);
1034 	/*
1035 	 * Don't migrate a running thread from sched_switch().
1036 	 */
1037 	if (flags & SRQ_OURSELF) {
1038 		CTR1(KTR_ULE, "YIELDING %d",
1039 		    curthread->td_priority);
1040 		return (self);
1041 	}
1042 	pri = ts->ts_thread->td_priority;
1043 	cpu = ts->ts_cpu;
1044 	/*
1045 	 * Regardless of affinity, if the last cpu is idle send it there.
1046 	 */
1047 	tdq = TDQ_CPU(cpu);
1048 	if (tdq->tdq_lowpri > PRI_MIN_IDLE) {
1049 		CTR5(KTR_ULE,
1050 		    "ts_cpu %d idle, ltick %d ticks %d pri %d curthread %d",
1051 		    ts->ts_cpu, ts->ts_rltick, ticks, pri,
1052 		    tdq->tdq_lowpri);
1053 		return (ts->ts_cpu);
1054 	}
1055 	/*
1056 	 * If we have affinity, try to place it on the cpu we last ran on.
1057 	 */
1058 	if (SCHED_AFFINITY(ts) && tdq->tdq_lowpri > pri) {
1059 		CTR5(KTR_ULE,
1060 		    "affinity for %d, ltick %d ticks %d pri %d curthread %d",
1061 		    ts->ts_cpu, ts->ts_rltick, ticks, pri,
1062 		    tdq->tdq_lowpri);
1063 		return (ts->ts_cpu);
1064 	}
1065 	/*
1066 	 * Look for an idle group.
1067 	 */
1068 	CTR1(KTR_ULE, "tdq_idle %X", tdq_idle);
1069 	cpu = ffs(tdq_idle);
1070 	if (cpu)
1071 		return (--cpu);
1072 	/*
1073 	 * If there are no idle cores see if we can run the thread locally.  This may
1074 	 * improve locality among sleepers and wakers when there is shared data.
1075 	 */
1076 	if (tryself && pri < curthread->td_priority) {
1077 		CTR1(KTR_ULE, "tryself %d",
1078 		    curthread->td_priority);
1079 		return (self);
1080 	}
1081 	/*
1082  	 * Now search for the cpu running the lowest priority thread with
1083 	 * the least load.
1084 	 */
1085 	if (pick_pri)
1086 		cpu = tdq_lowestpri();
1087 	else
1088 		cpu = tdq_lowestload();
1089 	return (cpu);
1090 }
1091 
1092 #endif	/* SMP */
1093 
1094 /*
1095  * Pick the highest priority task we have and return it.
1096  */
1097 static struct td_sched *
1098 tdq_choose(struct tdq *tdq)
1099 {
1100 	struct td_sched *ts;
1101 
1102 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1103 	ts = runq_choose(&tdq->tdq_realtime);
1104 	if (ts != NULL)
1105 		return (ts);
1106 	ts = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1107 	if (ts != NULL) {
1108 		KASSERT(ts->ts_thread->td_priority >= PRI_MIN_TIMESHARE,
1109 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1110 		    ts->ts_thread->td_priority));
1111 		return (ts);
1112 	}
1113 
1114 	ts = runq_choose(&tdq->tdq_idle);
1115 	if (ts != NULL) {
1116 		KASSERT(ts->ts_thread->td_priority >= PRI_MIN_IDLE,
1117 		    ("tdq_choose: Invalid priority on idle queue %d",
1118 		    ts->ts_thread->td_priority));
1119 		return (ts);
1120 	}
1121 
1122 	return (NULL);
1123 }
1124 
1125 /*
1126  * Initialize a thread queue.
1127  */
1128 static void
1129 tdq_setup(struct tdq *tdq)
1130 {
1131 
1132 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1133 	    "sched lock %d", (int)TDQ_ID(tdq));
1134 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1135 	    MTX_SPIN | MTX_RECURSE);
1136 	runq_init(&tdq->tdq_realtime);
1137 	runq_init(&tdq->tdq_timeshare);
1138 	runq_init(&tdq->tdq_idle);
1139 	tdq->tdq_load = 0;
1140 }
1141 
1142 /*
1143  * Setup the thread queues and initialize the topology based on MD
1144  * information.
1145  */
1146 static void
1147 sched_setup(void *dummy)
1148 {
1149 	struct tdq *tdq;
1150 #ifdef SMP
1151 	int balance_groups;
1152 	int i;
1153 
1154 	balance_groups = 0;
1155 	/*
1156 	 * Initialize the tdqs.
1157 	 */
1158 	for (i = 0; i < MAXCPU; i++) {
1159 		tdq = &tdq_cpu[i];
1160 		tdq_setup(&tdq_cpu[i]);
1161 	}
1162 	if (smp_topology == NULL) {
1163 		struct tdq_group *tdg;
1164 		int cpus;
1165 
1166 		for (cpus = 0, i = 0; i < MAXCPU; i++) {
1167 			if (CPU_ABSENT(i))
1168 				continue;
1169 			tdq = &tdq_cpu[i];
1170 			tdg = &tdq_groups[cpus];
1171 			/*
1172 			 * Setup a tdq group with one member.
1173 			 */
1174 			tdq->tdq_transferable = 0;
1175 			tdq->tdq_group = tdg;
1176 			tdg->tdg_cpus = 1;
1177 			tdg->tdg_idlemask = 0;
1178 			tdg->tdg_cpumask = tdg->tdg_mask = 1 << i;
1179 			tdg->tdg_load = 0;
1180 			tdg->tdg_transferable = 0;
1181 			LIST_INIT(&tdg->tdg_members);
1182 			LIST_INSERT_HEAD(&tdg->tdg_members, tdq, tdq_siblings);
1183 			cpus++;
1184 		}
1185 		tdg_maxid = cpus - 1;
1186 	} else {
1187 		struct tdq_group *tdg;
1188 		struct cpu_group *cg;
1189 		int j;
1190 
1191 		topology = 1;
1192 		for (i = 0; i < smp_topology->ct_count; i++) {
1193 			cg = &smp_topology->ct_group[i];
1194 			tdg = &tdq_groups[i];
1195 			/*
1196 			 * Initialize the group.
1197 			 */
1198 			tdg->tdg_idlemask = 0;
1199 			tdg->tdg_load = 0;
1200 			tdg->tdg_transferable = 0;
1201 			tdg->tdg_cpus = cg->cg_count;
1202 			tdg->tdg_cpumask = cg->cg_mask;
1203 			LIST_INIT(&tdg->tdg_members);
1204 			/*
1205 			 * Find all of the group members and add them.
1206 			 */
1207 			for (j = 0; j < MAXCPU; j++) {
1208 				if ((cg->cg_mask & (1 << j)) != 0) {
1209 					if (tdg->tdg_mask == 0)
1210 						tdg->tdg_mask = 1 << j;
1211 					tdq_cpu[j].tdq_transferable = 0;
1212 					tdq_cpu[j].tdq_group = tdg;
1213 					LIST_INSERT_HEAD(&tdg->tdg_members,
1214 					    &tdq_cpu[j], tdq_siblings);
1215 				}
1216 			}
1217 			if (tdg->tdg_cpus > 1)
1218 				balance_groups = 1;
1219 		}
1220 		tdg_maxid = smp_topology->ct_count - 1;
1221 	}
1222 	/*
1223 	 * Initialize long-term cpu balancing algorithm.
1224 	 */
1225 	callout_init(&balco, CALLOUT_MPSAFE);
1226 	callout_init(&gbalco, CALLOUT_MPSAFE);
1227 	sched_balance(NULL);
1228 	if (balance_groups)
1229 		sched_balance_groups(NULL);
1230 
1231 #else
1232 	tdq_setup(TDQ_SELF());
1233 #endif
1234 	/*
1235 	 * To avoid divide-by-zero, we set realstathz a dummy value
1236 	 * in case which sched_clock() called before sched_initticks().
1237 	 */
1238 	realstathz = hz;
1239 	sched_slice = (realstathz/10);	/* ~100ms */
1240 	tickincr = 1 << SCHED_TICK_SHIFT;
1241 
1242 	/* Add thread0's load since it's running. */
1243 	tdq = TDQ_SELF();
1244 	TDQ_LOCK(tdq);
1245 	tdq_load_add(tdq, &td_sched0);
1246 	TDQ_UNLOCK(tdq);
1247 }
1248 
1249 /*
1250  * This routine determines the tickincr after stathz and hz are setup.
1251  */
1252 /* ARGSUSED */
1253 static void
1254 sched_initticks(void *dummy)
1255 {
1256 	int incr;
1257 
1258 	realstathz = stathz ? stathz : hz;
1259 	sched_slice = (realstathz/10);	/* ~100ms */
1260 
1261 	/*
1262 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1263 	 * hz not being evenly divisible by stathz on all platforms.
1264 	 */
1265 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1266 	/*
1267 	 * This does not work for values of stathz that are more than
1268 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1269 	 */
1270 	if (incr == 0)
1271 		incr = 1;
1272 	tickincr = incr;
1273 #ifdef SMP
1274 	affinity = SCHED_AFFINITY_DEFAULT;
1275 #endif
1276 }
1277 
1278 
1279 /*
1280  * This is the core of the interactivity algorithm.  Determines a score based
1281  * on past behavior.  It is the ratio of sleep time to run time scaled to
1282  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1283  * differs from the cpu usage because it does not account for time spent
1284  * waiting on a run-queue.  Would be prettier if we had floating point.
1285  */
1286 static int
1287 sched_interact_score(struct thread *td)
1288 {
1289 	struct td_sched *ts;
1290 	int div;
1291 
1292 	ts = td->td_sched;
1293 	/*
1294 	 * The score is only needed if this is likely to be an interactive
1295 	 * task.  Don't go through the expense of computing it if there's
1296 	 * no chance.
1297 	 */
1298 	if (sched_interact <= SCHED_INTERACT_HALF &&
1299 		ts->ts_runtime >= ts->ts_slptime)
1300 			return (SCHED_INTERACT_HALF);
1301 
1302 	if (ts->ts_runtime > ts->ts_slptime) {
1303 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1304 		return (SCHED_INTERACT_HALF +
1305 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1306 	}
1307 	if (ts->ts_slptime > ts->ts_runtime) {
1308 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1309 		return (ts->ts_runtime / div);
1310 	}
1311 	/* runtime == slptime */
1312 	if (ts->ts_runtime)
1313 		return (SCHED_INTERACT_HALF);
1314 
1315 	/*
1316 	 * This can happen if slptime and runtime are 0.
1317 	 */
1318 	return (0);
1319 
1320 }
1321 
1322 /*
1323  * Scale the scheduling priority according to the "interactivity" of this
1324  * process.
1325  */
1326 static void
1327 sched_priority(struct thread *td)
1328 {
1329 	int score;
1330 	int pri;
1331 
1332 	if (td->td_pri_class != PRI_TIMESHARE)
1333 		return;
1334 	/*
1335 	 * If the score is interactive we place the thread in the realtime
1336 	 * queue with a priority that is less than kernel and interrupt
1337 	 * priorities.  These threads are not subject to nice restrictions.
1338 	 *
1339 	 * Scores greater than this are placed on the normal timeshare queue
1340 	 * where the priority is partially decided by the most recent cpu
1341 	 * utilization and the rest is decided by nice value.
1342 	 */
1343 	score = sched_interact_score(td);
1344 	if (score < sched_interact) {
1345 		pri = PRI_MIN_REALTIME;
1346 		pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact)
1347 		    * score;
1348 		KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME,
1349 		    ("sched_priority: invalid interactive priority %d score %d",
1350 		    pri, score));
1351 	} else {
1352 		pri = SCHED_PRI_MIN;
1353 		if (td->td_sched->ts_ticks)
1354 			pri += SCHED_PRI_TICKS(td->td_sched);
1355 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1356 		KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE,
1357 		    ("sched_priority: invalid priority %d: nice %d, "
1358 		    "ticks %d ftick %d ltick %d tick pri %d",
1359 		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1360 		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1361 		    SCHED_PRI_TICKS(td->td_sched)));
1362 	}
1363 	sched_user_prio(td, pri);
1364 
1365 	return;
1366 }
1367 
1368 /*
1369  * This routine enforces a maximum limit on the amount of scheduling history
1370  * kept.  It is called after either the slptime or runtime is adjusted.  This
1371  * function is ugly due to integer math.
1372  */
1373 static void
1374 sched_interact_update(struct thread *td)
1375 {
1376 	struct td_sched *ts;
1377 	u_int sum;
1378 
1379 	ts = td->td_sched;
1380 	sum = ts->ts_runtime + ts->ts_slptime;
1381 	if (sum < SCHED_SLP_RUN_MAX)
1382 		return;
1383 	/*
1384 	 * This only happens from two places:
1385 	 * 1) We have added an unusual amount of run time from fork_exit.
1386 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1387 	 */
1388 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1389 		if (ts->ts_runtime > ts->ts_slptime) {
1390 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1391 			ts->ts_slptime = 1;
1392 		} else {
1393 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1394 			ts->ts_runtime = 1;
1395 		}
1396 		return;
1397 	}
1398 	/*
1399 	 * If we have exceeded by more than 1/5th then the algorithm below
1400 	 * will not bring us back into range.  Dividing by two here forces
1401 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1402 	 */
1403 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1404 		ts->ts_runtime /= 2;
1405 		ts->ts_slptime /= 2;
1406 		return;
1407 	}
1408 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1409 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1410 }
1411 
1412 /*
1413  * Scale back the interactivity history when a child thread is created.  The
1414  * history is inherited from the parent but the thread may behave totally
1415  * differently.  For example, a shell spawning a compiler process.  We want
1416  * to learn that the compiler is behaving badly very quickly.
1417  */
1418 static void
1419 sched_interact_fork(struct thread *td)
1420 {
1421 	int ratio;
1422 	int sum;
1423 
1424 	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1425 	if (sum > SCHED_SLP_RUN_FORK) {
1426 		ratio = sum / SCHED_SLP_RUN_FORK;
1427 		td->td_sched->ts_runtime /= ratio;
1428 		td->td_sched->ts_slptime /= ratio;
1429 	}
1430 }
1431 
1432 /*
1433  * Called from proc0_init() to setup the scheduler fields.
1434  */
1435 void
1436 schedinit(void)
1437 {
1438 
1439 	/*
1440 	 * Set up the scheduler specific parts of proc0.
1441 	 */
1442 	proc0.p_sched = NULL; /* XXX */
1443 	thread0.td_sched = &td_sched0;
1444 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1445 	td_sched0.ts_ltick = ticks;
1446 	td_sched0.ts_ftick = ticks;
1447 	td_sched0.ts_thread = &thread0;
1448 }
1449 
1450 /*
1451  * This is only somewhat accurate since given many processes of the same
1452  * priority they will switch when their slices run out, which will be
1453  * at most sched_slice stathz ticks.
1454  */
1455 int
1456 sched_rr_interval(void)
1457 {
1458 
1459 	/* Convert sched_slice to hz */
1460 	return (hz/(realstathz/sched_slice));
1461 }
1462 
1463 /*
1464  * Update the percent cpu tracking information when it is requested or
1465  * the total history exceeds the maximum.  We keep a sliding history of
1466  * tick counts that slowly decays.  This is less precise than the 4BSD
1467  * mechanism since it happens with less regular and frequent events.
1468  */
1469 static void
1470 sched_pctcpu_update(struct td_sched *ts)
1471 {
1472 
1473 	if (ts->ts_ticks == 0)
1474 		return;
1475 	if (ticks - (hz / 10) < ts->ts_ltick &&
1476 	    SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX)
1477 		return;
1478 	/*
1479 	 * Adjust counters and watermark for pctcpu calc.
1480 	 */
1481 	if (ts->ts_ltick > ticks - SCHED_TICK_TARG)
1482 		ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) *
1483 			    SCHED_TICK_TARG;
1484 	else
1485 		ts->ts_ticks = 0;
1486 	ts->ts_ltick = ticks;
1487 	ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG;
1488 }
1489 
1490 /*
1491  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1492  * if necessary.  This is the back-end for several priority related
1493  * functions.
1494  */
1495 static void
1496 sched_thread_priority(struct thread *td, u_char prio)
1497 {
1498 	struct td_sched *ts;
1499 
1500 	CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)",
1501 	    td, td->td_proc->p_comm, td->td_priority, prio, curthread,
1502 	    curthread->td_proc->p_comm);
1503 	ts = td->td_sched;
1504 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1505 	if (td->td_priority == prio)
1506 		return;
1507 
1508 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1509 		/*
1510 		 * If the priority has been elevated due to priority
1511 		 * propagation, we may have to move ourselves to a new
1512 		 * queue.  This could be optimized to not re-add in some
1513 		 * cases.
1514 		 */
1515 		sched_rem(td);
1516 		td->td_priority = prio;
1517 		sched_add(td, SRQ_BORROWING);
1518 	} else {
1519 #ifdef SMP
1520 		struct tdq *tdq;
1521 
1522 		tdq = TDQ_CPU(ts->ts_cpu);
1523 		if (prio < tdq->tdq_lowpri)
1524 			tdq->tdq_lowpri = prio;
1525 #endif
1526 		td->td_priority = prio;
1527 	}
1528 }
1529 
1530 /*
1531  * Update a thread's priority when it is lent another thread's
1532  * priority.
1533  */
1534 void
1535 sched_lend_prio(struct thread *td, u_char prio)
1536 {
1537 
1538 	td->td_flags |= TDF_BORROWING;
1539 	sched_thread_priority(td, prio);
1540 }
1541 
1542 /*
1543  * Restore a thread's priority when priority propagation is
1544  * over.  The prio argument is the minimum priority the thread
1545  * needs to have to satisfy other possible priority lending
1546  * requests.  If the thread's regular priority is less
1547  * important than prio, the thread will keep a priority boost
1548  * of prio.
1549  */
1550 void
1551 sched_unlend_prio(struct thread *td, u_char prio)
1552 {
1553 	u_char base_pri;
1554 
1555 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1556 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1557 		base_pri = td->td_user_pri;
1558 	else
1559 		base_pri = td->td_base_pri;
1560 	if (prio >= base_pri) {
1561 		td->td_flags &= ~TDF_BORROWING;
1562 		sched_thread_priority(td, base_pri);
1563 	} else
1564 		sched_lend_prio(td, prio);
1565 }
1566 
1567 /*
1568  * Standard entry for setting the priority to an absolute value.
1569  */
1570 void
1571 sched_prio(struct thread *td, u_char prio)
1572 {
1573 	u_char oldprio;
1574 
1575 	/* First, update the base priority. */
1576 	td->td_base_pri = prio;
1577 
1578 	/*
1579 	 * If the thread is borrowing another thread's priority, don't
1580 	 * ever lower the priority.
1581 	 */
1582 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1583 		return;
1584 
1585 	/* Change the real priority. */
1586 	oldprio = td->td_priority;
1587 	sched_thread_priority(td, prio);
1588 
1589 	/*
1590 	 * If the thread is on a turnstile, then let the turnstile update
1591 	 * its state.
1592 	 */
1593 	if (TD_ON_LOCK(td) && oldprio != prio)
1594 		turnstile_adjust(td, oldprio);
1595 }
1596 
1597 /*
1598  * Set the base user priority, does not effect current running priority.
1599  */
1600 void
1601 sched_user_prio(struct thread *td, u_char prio)
1602 {
1603 	u_char oldprio;
1604 
1605 	td->td_base_user_pri = prio;
1606 	if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
1607                 return;
1608 	oldprio = td->td_user_pri;
1609 	td->td_user_pri = prio;
1610 
1611 	if (TD_ON_UPILOCK(td) && oldprio != prio)
1612 		umtx_pi_adjust(td, oldprio);
1613 }
1614 
1615 void
1616 sched_lend_user_prio(struct thread *td, u_char prio)
1617 {
1618 	u_char oldprio;
1619 
1620 	td->td_flags |= TDF_UBORROWING;
1621 
1622 	oldprio = td->td_user_pri;
1623 	td->td_user_pri = prio;
1624 
1625 	if (TD_ON_UPILOCK(td) && oldprio != prio)
1626 		umtx_pi_adjust(td, oldprio);
1627 }
1628 
1629 void
1630 sched_unlend_user_prio(struct thread *td, u_char prio)
1631 {
1632 	u_char base_pri;
1633 
1634 	base_pri = td->td_base_user_pri;
1635 	if (prio >= base_pri) {
1636 		td->td_flags &= ~TDF_UBORROWING;
1637 		sched_user_prio(td, base_pri);
1638 	} else
1639 		sched_lend_user_prio(td, prio);
1640 }
1641 
1642 /*
1643  * Add the thread passed as 'newtd' to the run queue before selecting
1644  * the next thread to run.  This is only used for KSE.
1645  */
1646 static void
1647 sched_switchin(struct tdq *tdq, struct thread *td)
1648 {
1649 #ifdef SMP
1650 	spinlock_enter();
1651 	TDQ_UNLOCK(tdq);
1652 	thread_lock(td);
1653 	spinlock_exit();
1654 	sched_setcpu(td->td_sched, TDQ_ID(tdq), SRQ_YIELDING);
1655 #else
1656 	td->td_lock = TDQ_LOCKPTR(tdq);
1657 #endif
1658 	tdq_add(tdq, td, SRQ_YIELDING);
1659 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1660 }
1661 
1662 /*
1663  * Block a thread for switching.  Similar to thread_block() but does not
1664  * bump the spin count.
1665  */
1666 static inline struct mtx *
1667 thread_block_switch(struct thread *td)
1668 {
1669 	struct mtx *lock;
1670 
1671 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1672 	lock = td->td_lock;
1673 	td->td_lock = &blocked_lock;
1674 	mtx_unlock_spin(lock);
1675 
1676 	return (lock);
1677 }
1678 
1679 /*
1680  * Release a thread that was blocked with thread_block_switch().
1681  */
1682 static inline void
1683 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1684 {
1685 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1686 	    (uintptr_t)mtx);
1687 }
1688 
1689 /*
1690  * Switch threads.  This function has to handle threads coming in while
1691  * blocked for some reason, running, or idle.  It also must deal with
1692  * migrating a thread from one queue to another as running threads may
1693  * be assigned elsewhere via binding.
1694  */
1695 void
1696 sched_switch(struct thread *td, struct thread *newtd, int flags)
1697 {
1698 	struct tdq *tdq;
1699 	struct td_sched *ts;
1700 	struct mtx *mtx;
1701 	int cpuid;
1702 
1703 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1704 
1705 	cpuid = PCPU_GET(cpuid);
1706 	tdq = TDQ_CPU(cpuid);
1707 	ts = td->td_sched;
1708 	mtx = TDQ_LOCKPTR(tdq);
1709 #ifdef SMP
1710 	ts->ts_rltick = ticks;
1711 	if (newtd && newtd->td_priority < tdq->tdq_lowpri)
1712 		tdq->tdq_lowpri = newtd->td_priority;
1713 #endif
1714 	td->td_lastcpu = td->td_oncpu;
1715 	td->td_oncpu = NOCPU;
1716 	td->td_flags &= ~TDF_NEEDRESCHED;
1717 	td->td_owepreempt = 0;
1718 	/*
1719 	 * The lock pointer in an idle thread should never change.  Reset it
1720 	 * to CAN_RUN as well.
1721 	 */
1722 	if (TD_IS_IDLETHREAD(td)) {
1723 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1724 		TD_SET_CAN_RUN(td);
1725 	} else if (TD_IS_RUNNING(td)) {
1726 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1727 		/* Remove our load so the selection algorithm is not biased. */
1728 		tdq_load_rem(tdq, ts);
1729 		sched_add(td, (flags & SW_PREEMPT) ?
1730 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1731 		    SRQ_OURSELF|SRQ_YIELDING);
1732 		/*
1733 		 * When migrating we return from sched_add with an extra
1734 		 * spinlock nesting, the tdq locked, and a blocked thread.
1735 		 * This is to optimize out an extra block/unblock cycle here.
1736 		 */
1737 		if (ts->ts_cpu != cpuid) {
1738 			mtx = TDQ_LOCKPTR(TDQ_CPU(ts->ts_cpu));
1739 			mtx_unlock_spin(mtx);
1740 			TDQ_LOCK(tdq);
1741 			spinlock_exit();
1742 		}
1743 	} else {
1744 		/* This thread must be going to sleep. */
1745 		TDQ_LOCK(tdq);
1746 		mtx = thread_block_switch(td);
1747 		tdq_load_rem(tdq, ts);
1748 	}
1749 	/*
1750 	 * We enter here with the thread blocked and assigned to the
1751 	 * appropriate cpu run-queue or sleep-queue and with the current
1752 	 * thread-queue locked.
1753 	 */
1754 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1755 	/*
1756 	 * If KSE assigned a new thread just add it here and let choosethread
1757 	 * select the best one.
1758 	 */
1759 	if (newtd != NULL)
1760 		sched_switchin(tdq, newtd);
1761 	newtd = choosethread();
1762 	/*
1763 	 * Call the MD code to switch contexts if necessary.
1764 	 */
1765 	if (td != newtd) {
1766 #ifdef	HWPMC_HOOKS
1767 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1768 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1769 #endif
1770 		cpu_switch(td, newtd, mtx);
1771 		/*
1772 		 * We may return from cpu_switch on a different cpu.  However,
1773 		 * we always return with td_lock pointing to the current cpu's
1774 		 * run queue lock.
1775 		 */
1776 		cpuid = PCPU_GET(cpuid);
1777 		tdq = TDQ_CPU(cpuid);
1778 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)td;
1779 #ifdef	HWPMC_HOOKS
1780 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1781 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1782 #endif
1783 	} else
1784 		thread_unblock_switch(td, mtx);
1785 	/*
1786 	 * Assert that all went well and return.
1787 	 */
1788 #ifdef SMP
1789 	/* We should always get here with the lowest priority td possible */
1790 	tdq->tdq_lowpri = td->td_priority;
1791 #endif
1792 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1793 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1794 	td->td_oncpu = cpuid;
1795 }
1796 
1797 /*
1798  * Adjust thread priorities as a result of a nice request.
1799  */
1800 void
1801 sched_nice(struct proc *p, int nice)
1802 {
1803 	struct thread *td;
1804 
1805 	PROC_LOCK_ASSERT(p, MA_OWNED);
1806 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1807 
1808 	p->p_nice = nice;
1809 	FOREACH_THREAD_IN_PROC(p, td) {
1810 		thread_lock(td);
1811 		sched_priority(td);
1812 		sched_prio(td, td->td_base_user_pri);
1813 		thread_unlock(td);
1814 	}
1815 }
1816 
1817 /*
1818  * Record the sleep time for the interactivity scorer.
1819  */
1820 void
1821 sched_sleep(struct thread *td)
1822 {
1823 
1824 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1825 
1826 	td->td_sched->ts_slptick = ticks;
1827 }
1828 
1829 /*
1830  * Schedule a thread to resume execution and record how long it voluntarily
1831  * slept.  We also update the pctcpu, interactivity, and priority.
1832  */
1833 void
1834 sched_wakeup(struct thread *td)
1835 {
1836 	struct td_sched *ts;
1837 	int slptick;
1838 
1839 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1840 	ts = td->td_sched;
1841 	/*
1842 	 * If we slept for more than a tick update our interactivity and
1843 	 * priority.
1844 	 */
1845 	slptick = ts->ts_slptick;
1846 	ts->ts_slptick = 0;
1847 	if (slptick && slptick != ticks) {
1848 		u_int hzticks;
1849 
1850 		hzticks = (ticks - slptick) << SCHED_TICK_SHIFT;
1851 		ts->ts_slptime += hzticks;
1852 		sched_interact_update(td);
1853 		sched_pctcpu_update(ts);
1854 		sched_priority(td);
1855 	}
1856 	/* Reset the slice value after we sleep. */
1857 	ts->ts_slice = sched_slice;
1858 	sched_add(td, SRQ_BORING);
1859 }
1860 
1861 /*
1862  * Penalize the parent for creating a new child and initialize the child's
1863  * priority.
1864  */
1865 void
1866 sched_fork(struct thread *td, struct thread *child)
1867 {
1868 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1869 	sched_fork_thread(td, child);
1870 	/*
1871 	 * Penalize the parent and child for forking.
1872 	 */
1873 	sched_interact_fork(child);
1874 	sched_priority(child);
1875 	td->td_sched->ts_runtime += tickincr;
1876 	sched_interact_update(td);
1877 	sched_priority(td);
1878 }
1879 
1880 /*
1881  * Fork a new thread, may be within the same process.
1882  */
1883 void
1884 sched_fork_thread(struct thread *td, struct thread *child)
1885 {
1886 	struct td_sched *ts;
1887 	struct td_sched *ts2;
1888 
1889 	/*
1890 	 * Initialize child.
1891 	 */
1892 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1893 	sched_newthread(child);
1894 	child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
1895 	ts = td->td_sched;
1896 	ts2 = child->td_sched;
1897 	ts2->ts_cpu = ts->ts_cpu;
1898 	ts2->ts_runq = NULL;
1899 	/*
1900 	 * Grab our parents cpu estimation information and priority.
1901 	 */
1902 	ts2->ts_ticks = ts->ts_ticks;
1903 	ts2->ts_ltick = ts->ts_ltick;
1904 	ts2->ts_ftick = ts->ts_ftick;
1905 	child->td_user_pri = td->td_user_pri;
1906 	child->td_base_user_pri = td->td_base_user_pri;
1907 	/*
1908 	 * And update interactivity score.
1909 	 */
1910 	ts2->ts_slptime = ts->ts_slptime;
1911 	ts2->ts_runtime = ts->ts_runtime;
1912 	ts2->ts_slice = 1;	/* Attempt to quickly learn interactivity. */
1913 }
1914 
1915 /*
1916  * Adjust the priority class of a thread.
1917  */
1918 void
1919 sched_class(struct thread *td, int class)
1920 {
1921 
1922 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1923 	if (td->td_pri_class == class)
1924 		return;
1925 
1926 #ifdef SMP
1927 	/*
1928 	 * On SMP if we're on the RUNQ we must adjust the transferable
1929 	 * count because could be changing to or from an interrupt
1930 	 * class.
1931 	 */
1932 	if (TD_ON_RUNQ(td)) {
1933 		struct tdq *tdq;
1934 
1935 		tdq = TDQ_CPU(td->td_sched->ts_cpu);
1936 		if (THREAD_CAN_MIGRATE(td)) {
1937 			tdq->tdq_transferable--;
1938 			tdq->tdq_group->tdg_transferable--;
1939 		}
1940 		td->td_pri_class = class;
1941 		if (THREAD_CAN_MIGRATE(td)) {
1942 			tdq->tdq_transferable++;
1943 			tdq->tdq_group->tdg_transferable++;
1944 		}
1945 	}
1946 #endif
1947 	td->td_pri_class = class;
1948 }
1949 
1950 /*
1951  * Return some of the child's priority and interactivity to the parent.
1952  */
1953 void
1954 sched_exit(struct proc *p, struct thread *child)
1955 {
1956 	struct thread *td;
1957 
1958 	CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d",
1959 	    child, child->td_proc->p_comm, child->td_priority);
1960 
1961 	PROC_SLOCK_ASSERT(p, MA_OWNED);
1962 	td = FIRST_THREAD_IN_PROC(p);
1963 	sched_exit_thread(td, child);
1964 }
1965 
1966 /*
1967  * Penalize another thread for the time spent on this one.  This helps to
1968  * worsen the priority and interactivity of processes which schedule batch
1969  * jobs such as make.  This has little effect on the make process itself but
1970  * causes new processes spawned by it to receive worse scores immediately.
1971  */
1972 void
1973 sched_exit_thread(struct thread *td, struct thread *child)
1974 {
1975 
1976 	CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d",
1977 	    child, child->td_proc->p_comm, child->td_priority);
1978 
1979 #ifdef KSE
1980 	/*
1981 	 * KSE forks and exits so often that this penalty causes short-lived
1982 	 * threads to always be non-interactive.  This causes mozilla to
1983 	 * crawl under load.
1984 	 */
1985 	if ((td->td_pflags & TDP_SA) && td->td_proc == child->td_proc)
1986 		return;
1987 #endif
1988 	/*
1989 	 * Give the child's runtime to the parent without returning the
1990 	 * sleep time as a penalty to the parent.  This causes shells that
1991 	 * launch expensive things to mark their children as expensive.
1992 	 */
1993 	thread_lock(td);
1994 	td->td_sched->ts_runtime += child->td_sched->ts_runtime;
1995 	sched_interact_update(td);
1996 	sched_priority(td);
1997 	thread_unlock(td);
1998 }
1999 
2000 /*
2001  * Fix priorities on return to user-space.  Priorities may be elevated due
2002  * to static priorities in msleep() or similar.
2003  */
2004 void
2005 sched_userret(struct thread *td)
2006 {
2007 	/*
2008 	 * XXX we cheat slightly on the locking here to avoid locking in
2009 	 * the usual case.  Setting td_priority here is essentially an
2010 	 * incomplete workaround for not setting it properly elsewhere.
2011 	 * Now that some interrupt handlers are threads, not setting it
2012 	 * properly elsewhere can clobber it in the window between setting
2013 	 * it here and returning to user mode, so don't waste time setting
2014 	 * it perfectly here.
2015 	 */
2016 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2017 	    ("thread with borrowed priority returning to userland"));
2018 	if (td->td_priority != td->td_user_pri) {
2019 		thread_lock(td);
2020 		td->td_priority = td->td_user_pri;
2021 		td->td_base_pri = td->td_user_pri;
2022 		thread_unlock(td);
2023         }
2024 }
2025 
2026 /*
2027  * Handle a stathz tick.  This is really only relevant for timeshare
2028  * threads.
2029  */
2030 void
2031 sched_clock(struct thread *td)
2032 {
2033 	struct tdq *tdq;
2034 	struct td_sched *ts;
2035 
2036 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2037 	tdq = TDQ_SELF();
2038 	/*
2039 	 * Advance the insert index once for each tick to ensure that all
2040 	 * threads get a chance to run.
2041 	 */
2042 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2043 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2044 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2045 			tdq->tdq_ridx = tdq->tdq_idx;
2046 	}
2047 	ts = td->td_sched;
2048 	/*
2049 	 * We only do slicing code for TIMESHARE threads.
2050 	 */
2051 	if (td->td_pri_class != PRI_TIMESHARE)
2052 		return;
2053 	/*
2054 	 * We used a tick; charge it to the thread so that we can compute our
2055 	 * interactivity.
2056 	 */
2057 	td->td_sched->ts_runtime += tickincr;
2058 	sched_interact_update(td);
2059 	/*
2060 	 * We used up one time slice.
2061 	 */
2062 	if (--ts->ts_slice > 0)
2063 		return;
2064 	/*
2065 	 * We're out of time, recompute priorities and requeue.
2066 	 */
2067 	sched_priority(td);
2068 	td->td_flags |= TDF_NEEDRESCHED;
2069 }
2070 
2071 /*
2072  * Called once per hz tick.  Used for cpu utilization information.  This
2073  * is easier than trying to scale based on stathz.
2074  */
2075 void
2076 sched_tick(void)
2077 {
2078 	struct td_sched *ts;
2079 
2080 	ts = curthread->td_sched;
2081 	/* Adjust ticks for pctcpu */
2082 	ts->ts_ticks += 1 << SCHED_TICK_SHIFT;
2083 	ts->ts_ltick = ticks;
2084 	/*
2085 	 * Update if we've exceeded our desired tick threshhold by over one
2086 	 * second.
2087 	 */
2088 	if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick)
2089 		sched_pctcpu_update(ts);
2090 }
2091 
2092 /*
2093  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2094  * cooperative idle threads.
2095  */
2096 int
2097 sched_runnable(void)
2098 {
2099 	struct tdq *tdq;
2100 	int load;
2101 
2102 	load = 1;
2103 
2104 	tdq = TDQ_SELF();
2105 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2106 		if (tdq->tdq_load > 0)
2107 			goto out;
2108 	} else
2109 		if (tdq->tdq_load - 1 > 0)
2110 			goto out;
2111 	load = 0;
2112 out:
2113 	return (load);
2114 }
2115 
2116 /*
2117  * Choose the highest priority thread to run.  The thread is removed from
2118  * the run-queue while running however the load remains.  For SMP we set
2119  * the tdq in the global idle bitmask if it idles here.
2120  */
2121 struct thread *
2122 sched_choose(void)
2123 {
2124 #ifdef SMP
2125 	struct tdq_group *tdg;
2126 #endif
2127 	struct td_sched *ts;
2128 	struct tdq *tdq;
2129 
2130 	tdq = TDQ_SELF();
2131 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2132 	ts = tdq_choose(tdq);
2133 	if (ts) {
2134 		tdq_runq_rem(tdq, ts);
2135 		return (ts->ts_thread);
2136 	}
2137 #ifdef SMP
2138 	/*
2139 	 * We only set the idled bit when all of the cpus in the group are
2140 	 * idle.  Otherwise we could get into a situation where a thread bounces
2141 	 * back and forth between two idle cores on seperate physical CPUs.
2142 	 */
2143 	tdg = tdq->tdq_group;
2144 	tdg->tdg_idlemask |= PCPU_GET(cpumask);
2145 	if (tdg->tdg_idlemask == tdg->tdg_cpumask)
2146 		atomic_set_int(&tdq_idle, tdg->tdg_mask);
2147 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2148 #endif
2149 	return (PCPU_GET(idlethread));
2150 }
2151 
2152 /*
2153  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2154  * we always request it once we exit a critical section.
2155  */
2156 static inline void
2157 sched_setpreempt(struct thread *td)
2158 {
2159 	struct thread *ctd;
2160 	int cpri;
2161 	int pri;
2162 
2163 	ctd = curthread;
2164 	pri = td->td_priority;
2165 	cpri = ctd->td_priority;
2166 	if (td->td_priority < ctd->td_priority)
2167 		curthread->td_flags |= TDF_NEEDRESCHED;
2168 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2169 		return;
2170 	/*
2171 	 * Always preempt IDLE threads.  Otherwise only if the preempting
2172 	 * thread is an ithread.
2173 	 */
2174 	if (pri > preempt_thresh && cpri < PRI_MIN_IDLE)
2175 		return;
2176 	ctd->td_owepreempt = 1;
2177 	return;
2178 }
2179 
2180 /*
2181  * Add a thread to a thread queue.  Initializes priority, slice, runq, and
2182  * add it to the appropriate queue.  This is the internal function called
2183  * when the tdq is predetermined.
2184  */
2185 void
2186 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2187 {
2188 	struct td_sched *ts;
2189 	int class;
2190 #ifdef SMP
2191 	int cpumask;
2192 #endif
2193 
2194 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2195 	KASSERT((td->td_inhibitors == 0),
2196 	    ("sched_add: trying to run inhibited thread"));
2197 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2198 	    ("sched_add: bad thread state"));
2199 	KASSERT(td->td_proc->p_sflag & PS_INMEM,
2200 	    ("sched_add: process swapped out"));
2201 
2202 	ts = td->td_sched;
2203 	class = PRI_BASE(td->td_pri_class);
2204         TD_SET_RUNQ(td);
2205 	if (ts->ts_slice == 0)
2206 		ts->ts_slice = sched_slice;
2207 	/*
2208 	 * Pick the run queue based on priority.
2209 	 */
2210 	if (td->td_priority <= PRI_MAX_REALTIME)
2211 		ts->ts_runq = &tdq->tdq_realtime;
2212 	else if (td->td_priority <= PRI_MAX_TIMESHARE)
2213 		ts->ts_runq = &tdq->tdq_timeshare;
2214 	else
2215 		ts->ts_runq = &tdq->tdq_idle;
2216 #ifdef SMP
2217 	cpumask = 1 << ts->ts_cpu;
2218 	/*
2219 	 * If we had been idle, clear our bit in the group and potentially
2220 	 * the global bitmap.
2221 	 */
2222 	if ((class != PRI_IDLE && class != PRI_ITHD) &&
2223 	    (tdq->tdq_group->tdg_idlemask & cpumask) != 0) {
2224 		/*
2225 		 * Check to see if our group is unidling, and if so, remove it
2226 		 * from the global idle mask.
2227 		 */
2228 		if (tdq->tdq_group->tdg_idlemask ==
2229 		    tdq->tdq_group->tdg_cpumask)
2230 			atomic_clear_int(&tdq_idle, tdq->tdq_group->tdg_mask);
2231 		/*
2232 		 * Now remove ourselves from the group specific idle mask.
2233 		 */
2234 		tdq->tdq_group->tdg_idlemask &= ~cpumask;
2235 	}
2236 	if (td->td_priority < tdq->tdq_lowpri)
2237 		tdq->tdq_lowpri = td->td_priority;
2238 #endif
2239 	tdq_runq_add(tdq, ts, flags);
2240 	tdq_load_add(tdq, ts);
2241 }
2242 
2243 /*
2244  * Select the target thread queue and add a thread to it.  Request
2245  * preemption or IPI a remote processor if required.
2246  */
2247 void
2248 sched_add(struct thread *td, int flags)
2249 {
2250 	struct td_sched *ts;
2251 	struct tdq *tdq;
2252 #ifdef SMP
2253 	int cpuid;
2254 	int cpu;
2255 #endif
2256 	CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
2257 	    td, td->td_proc->p_comm, td->td_priority, curthread,
2258 	    curthread->td_proc->p_comm);
2259 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2260 	ts = td->td_sched;
2261 	/*
2262 	 * Recalculate the priority before we select the target cpu or
2263 	 * run-queue.
2264 	 */
2265 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2266 		sched_priority(td);
2267 #ifdef SMP
2268 	cpuid = PCPU_GET(cpuid);
2269 	/*
2270 	 * Pick the destination cpu and if it isn't ours transfer to the
2271 	 * target cpu.
2272 	 */
2273 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_MIGRATE(td))
2274 		cpu = cpuid;
2275 	else if (!THREAD_CAN_MIGRATE(td))
2276 		cpu = ts->ts_cpu;
2277 	else
2278 		cpu = sched_pickcpu(ts, flags);
2279 	tdq = sched_setcpu(ts, cpu, flags);
2280 	tdq_add(tdq, td, flags);
2281 	if (cpu != cpuid) {
2282 		tdq_notify(ts);
2283 		return;
2284 	}
2285 #else
2286 	tdq = TDQ_SELF();
2287 	TDQ_LOCK(tdq);
2288 	/*
2289 	 * Now that the thread is moving to the run-queue, set the lock
2290 	 * to the scheduler's lock.
2291 	 */
2292 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2293 	tdq_add(tdq, td, flags);
2294 #endif
2295 	if (!(flags & SRQ_YIELDING))
2296 		sched_setpreempt(td);
2297 }
2298 
2299 /*
2300  * Remove a thread from a run-queue without running it.  This is used
2301  * when we're stealing a thread from a remote queue.  Otherwise all threads
2302  * exit by calling sched_exit_thread() and sched_throw() themselves.
2303  */
2304 void
2305 sched_rem(struct thread *td)
2306 {
2307 	struct tdq *tdq;
2308 	struct td_sched *ts;
2309 
2310 	CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)",
2311 	    td, td->td_proc->p_comm, td->td_priority, curthread,
2312 	    curthread->td_proc->p_comm);
2313 	ts = td->td_sched;
2314 	tdq = TDQ_CPU(ts->ts_cpu);
2315 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2316 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2317 	KASSERT(TD_ON_RUNQ(td),
2318 	    ("sched_rem: thread not on run queue"));
2319 	tdq_runq_rem(tdq, ts);
2320 	tdq_load_rem(tdq, ts);
2321 	TD_SET_CAN_RUN(td);
2322 }
2323 
2324 /*
2325  * Fetch cpu utilization information.  Updates on demand.
2326  */
2327 fixpt_t
2328 sched_pctcpu(struct thread *td)
2329 {
2330 	fixpt_t pctcpu;
2331 	struct td_sched *ts;
2332 
2333 	pctcpu = 0;
2334 	ts = td->td_sched;
2335 	if (ts == NULL)
2336 		return (0);
2337 
2338 	thread_lock(td);
2339 	if (ts->ts_ticks) {
2340 		int rtick;
2341 
2342 		sched_pctcpu_update(ts);
2343 		/* How many rtick per second ? */
2344 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2345 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2346 	}
2347 	td->td_proc->p_swtime = ts->ts_ltick - ts->ts_ftick;
2348 	thread_unlock(td);
2349 
2350 	return (pctcpu);
2351 }
2352 
2353 /*
2354  * Bind a thread to a target cpu.
2355  */
2356 void
2357 sched_bind(struct thread *td, int cpu)
2358 {
2359 	struct td_sched *ts;
2360 
2361 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2362 	ts = td->td_sched;
2363 	if (ts->ts_flags & TSF_BOUND)
2364 		sched_unbind(td);
2365 	ts->ts_flags |= TSF_BOUND;
2366 #ifdef SMP
2367 	sched_pin();
2368 	if (PCPU_GET(cpuid) == cpu)
2369 		return;
2370 	ts->ts_cpu = cpu;
2371 	/* When we return from mi_switch we'll be on the correct cpu. */
2372 	mi_switch(SW_VOL, NULL);
2373 #endif
2374 }
2375 
2376 /*
2377  * Release a bound thread.
2378  */
2379 void
2380 sched_unbind(struct thread *td)
2381 {
2382 	struct td_sched *ts;
2383 
2384 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2385 	ts = td->td_sched;
2386 	if ((ts->ts_flags & TSF_BOUND) == 0)
2387 		return;
2388 	ts->ts_flags &= ~TSF_BOUND;
2389 #ifdef SMP
2390 	sched_unpin();
2391 #endif
2392 }
2393 
2394 int
2395 sched_is_bound(struct thread *td)
2396 {
2397 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2398 	return (td->td_sched->ts_flags & TSF_BOUND);
2399 }
2400 
2401 /*
2402  * Basic yield call.
2403  */
2404 void
2405 sched_relinquish(struct thread *td)
2406 {
2407 	thread_lock(td);
2408 	if (td->td_pri_class == PRI_TIMESHARE)
2409 		sched_prio(td, PRI_MAX_TIMESHARE);
2410 	SCHED_STAT_INC(switch_relinquish);
2411 	mi_switch(SW_VOL, NULL);
2412 	thread_unlock(td);
2413 }
2414 
2415 /*
2416  * Return the total system load.
2417  */
2418 int
2419 sched_load(void)
2420 {
2421 #ifdef SMP
2422 	int total;
2423 	int i;
2424 
2425 	total = 0;
2426 	for (i = 0; i <= tdg_maxid; i++)
2427 		total += TDQ_GROUP(i)->tdg_load;
2428 	return (total);
2429 #else
2430 	return (TDQ_SELF()->tdq_sysload);
2431 #endif
2432 }
2433 
2434 int
2435 sched_sizeof_proc(void)
2436 {
2437 	return (sizeof(struct proc));
2438 }
2439 
2440 int
2441 sched_sizeof_thread(void)
2442 {
2443 	return (sizeof(struct thread) + sizeof(struct td_sched));
2444 }
2445 
2446 /*
2447  * The actual idle process.
2448  */
2449 void
2450 sched_idletd(void *dummy)
2451 {
2452 	struct thread *td;
2453 	struct tdq *tdq;
2454 
2455 	td = curthread;
2456 	tdq = TDQ_SELF();
2457 	mtx_assert(&Giant, MA_NOTOWNED);
2458 	/* ULE relies on preemption for idle interruption. */
2459 	for (;;) {
2460 #ifdef SMP
2461 		if (tdq_idled(tdq))
2462 			cpu_idle();
2463 #else
2464 		cpu_idle();
2465 #endif
2466 	}
2467 }
2468 
2469 /*
2470  * A CPU is entering for the first time or a thread is exiting.
2471  */
2472 void
2473 sched_throw(struct thread *td)
2474 {
2475 	struct tdq *tdq;
2476 
2477 	tdq = TDQ_SELF();
2478 	if (td == NULL) {
2479 		/* Correct spinlock nesting and acquire the correct lock. */
2480 		TDQ_LOCK(tdq);
2481 		spinlock_exit();
2482 	} else {
2483 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2484 		tdq_load_rem(tdq, td->td_sched);
2485 	}
2486 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2487 	PCPU_SET(switchtime, cpu_ticks());
2488 	PCPU_SET(switchticks, ticks);
2489 	cpu_throw(td, choosethread());	/* doesn't return */
2490 }
2491 
2492 /*
2493  * This is called from fork_exit().  Just acquire the correct locks and
2494  * let fork do the rest of the work.
2495  */
2496 void
2497 sched_fork_exit(struct thread *td)
2498 {
2499 	struct td_sched *ts;
2500 	struct tdq *tdq;
2501 	int cpuid;
2502 
2503 	/*
2504 	 * Finish setting up thread glue so that it begins execution in a
2505 	 * non-nested critical section with the scheduler lock held.
2506 	 */
2507 	cpuid = PCPU_GET(cpuid);
2508 	tdq = TDQ_CPU(cpuid);
2509 	ts = td->td_sched;
2510 	if (TD_IS_IDLETHREAD(td))
2511 		td->td_lock = TDQ_LOCKPTR(tdq);
2512 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2513 	td->td_oncpu = cpuid;
2514 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)td;
2515 	THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
2516 }
2517 
2518 static SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0,
2519     "Scheduler");
2520 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2521     "Scheduler name");
2522 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2523     "Slice size for timeshare threads");
2524 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2525      "Interactivity score threshold");
2526 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh,
2527      0,"Min priority for preemption, lower priorities have greater precedence");
2528 #ifdef SMP
2529 SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri, CTLFLAG_RW, &pick_pri, 0,
2530     "Pick the target cpu based on priority rather than load.");
2531 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2532     "Number of hz ticks to keep thread affinity for");
2533 SYSCTL_INT(_kern_sched, OID_AUTO, tryself, CTLFLAG_RW, &tryself, 0, "");
2534 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2535     "Enables the long-term load balancer");
2536 SYSCTL_INT(_kern_sched, OID_AUTO, balance_secs, CTLFLAG_RW, &balance_secs, 0,
2537     "Average frequence in seconds to run the long-term balancer");
2538 SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0,
2539     "Steals work from another hyper-threaded core on idle");
2540 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2541     "Attempts to steal work from other cores before idling");
2542 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2543     "Minimum load on remote cpu before we'll steal");
2544 SYSCTL_INT(_kern_sched, OID_AUTO, topology, CTLFLAG_RD, &topology, 0,
2545     "True when a topology has been specified by the MD code.");
2546 #endif
2547 
2548 /* ps compat */
2549 static fixpt_t  ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
2550 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2551 
2552 
2553 #define KERN_SWITCH_INCLUDE 1
2554 #include "kern/kern_switch.c"
2555