xref: /freebsd/sys/kern/sched_ule.c (revision 588ff6c0cc9aaf10ba19080d9f8acbd8be36abf3)
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 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 #include "opt_hwpmc_hooks.h"
31 #include "opt_sched.h"
32 
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/kdb.h>
36 #include <sys/kernel.h>
37 #include <sys/ktr.h>
38 #include <sys/lock.h>
39 #include <sys/mutex.h>
40 #include <sys/proc.h>
41 #include <sys/resource.h>
42 #include <sys/resourcevar.h>
43 #include <sys/sched.h>
44 #include <sys/smp.h>
45 #include <sys/sx.h>
46 #include <sys/sysctl.h>
47 #include <sys/sysproto.h>
48 #include <sys/turnstile.h>
49 #include <sys/umtx.h>
50 #include <sys/vmmeter.h>
51 #ifdef KTRACE
52 #include <sys/uio.h>
53 #include <sys/ktrace.h>
54 #endif
55 
56 #ifdef HWPMC_HOOKS
57 #include <sys/pmckern.h>
58 #endif
59 
60 #include <machine/cpu.h>
61 #include <machine/smp.h>
62 
63 #ifndef PREEMPTION
64 #error	"SCHED_ULE requires options PREEMPTION"
65 #endif
66 
67 /*
68  * TODO:
69  *	Pick idle from affinity group or self group first.
70  *	Implement pick_score.
71  */
72 
73 #define	KTR_ULE	0x0		/* Enable for pickpri debugging. */
74 
75 /*
76  * Thread scheduler specific section.
77  */
78 struct td_sched {
79 	TAILQ_ENTRY(td_sched) ts_procq;	/* (j/z) Run queue. */
80 	int		ts_flags;	/* (j) TSF_* flags. */
81 	struct thread	*ts_thread;	/* (*) Active associated thread. */
82 	u_char		ts_rqindex;	/* (j) Run queue index. */
83 	int		ts_slptime;
84 	int		ts_slice;
85 	struct runq	*ts_runq;
86 	u_char		ts_cpu;		/* CPU that we have affinity for. */
87 	/* The following variables are only used for pctcpu calculation */
88 	int		ts_ltick;	/* Last tick that we were running on */
89 	int		ts_ftick;	/* First tick that we were running on */
90 	int		ts_ticks;	/* Tick count */
91 #ifdef SMP
92 	int		ts_rltick;	/* Real last tick, for affinity. */
93 #endif
94 
95 	/* originally from kg_sched */
96 	u_int	skg_slptime;		/* Number of ticks we vol. slept */
97 	u_int	skg_runtime;		/* Number of ticks we were running */
98 };
99 /* flags kept in ts_flags */
100 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
101 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
102 
103 static struct td_sched td_sched0;
104 
105 /*
106  * Cpu percentage computation macros and defines.
107  *
108  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
109  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
110  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
111  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
112  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
113  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
114  */
115 #define	SCHED_TICK_SECS		10
116 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
117 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
118 #define	SCHED_TICK_SHIFT	10
119 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
120 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
121 
122 /*
123  * These macros determine priorities for non-interactive threads.  They are
124  * assigned a priority based on their recent cpu utilization as expressed
125  * by the ratio of ticks to the tick total.  NHALF priorities at the start
126  * and end of the MIN to MAX timeshare range are only reachable with negative
127  * or positive nice respectively.
128  *
129  * PRI_RANGE:	Priority range for utilization dependent priorities.
130  * PRI_NRESV:	Number of nice values.
131  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
132  * PRI_NICE:	Determines the part of the priority inherited from nice.
133  */
134 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
135 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
136 #define	SCHED_PRI_MIN		(PRI_MIN_TIMESHARE + SCHED_PRI_NHALF)
137 #define	SCHED_PRI_MAX		(PRI_MAX_TIMESHARE - SCHED_PRI_NHALF)
138 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
139 #define	SCHED_PRI_TICKS(ts)						\
140     (SCHED_TICK_HZ((ts)) /						\
141     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
142 #define	SCHED_PRI_NICE(nice)	(nice)
143 
144 /*
145  * These determine the interactivity of a process.  Interactivity differs from
146  * cpu utilization in that it expresses the voluntary time slept vs time ran
147  * while cpu utilization includes all time not running.  This more accurately
148  * models the intent of the thread.
149  *
150  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
151  *		before throttling back.
152  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
153  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
154  * INTERACT_THRESH:	Threshhold for placement on the current runq.
155  */
156 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
157 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
158 #define	SCHED_INTERACT_MAX	(100)
159 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
160 #define	SCHED_INTERACT_THRESH	(30)
161 
162 /*
163  * tickincr:		Converts a stathz tick into a hz domain scaled by
164  *			the shift factor.  Without the shift the error rate
165  *			due to rounding would be unacceptably high.
166  * realstathz:		stathz is sometimes 0 and run off of hz.
167  * sched_slice:		Runtime of each thread before rescheduling.
168  */
169 static int sched_interact = SCHED_INTERACT_THRESH;
170 static int realstathz;
171 static int tickincr;
172 static int sched_slice;
173 
174 /*
175  * tdq - per processor runqs and statistics.
176  */
177 struct tdq {
178 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
179 	struct runq	tdq_timeshare;		/* timeshare run queue. */
180 	struct runq	tdq_realtime;		/* real-time run queue. */
181 	u_char		tdq_idx;		/* Current insert index. */
182 	u_char		tdq_ridx;		/* Current removal index. */
183 	short		tdq_flags;		/* Thread queue flags */
184 	int		tdq_load;		/* Aggregate load. */
185 #ifdef SMP
186 	int		tdq_transferable;
187 	LIST_ENTRY(tdq)	tdq_siblings;		/* Next in tdq group. */
188 	struct tdq_group *tdq_group;		/* Our processor group. */
189 #else
190 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
191 #endif
192 };
193 
194 #define	TDQF_BUSY	0x0001			/* Queue is marked as busy */
195 
196 #ifdef SMP
197 /*
198  * tdq groups are groups of processors which can cheaply share threads.  When
199  * one processor in the group goes idle it will check the runqs of the other
200  * processors in its group prior to halting and waiting for an interrupt.
201  * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA.
202  * In a numa environment we'd want an idle bitmap per group and a two tiered
203  * load balancer.
204  */
205 struct tdq_group {
206 	int	tdg_cpus;		/* Count of CPUs in this tdq group. */
207 	cpumask_t tdg_cpumask;		/* Mask of cpus in this group. */
208 	cpumask_t tdg_idlemask;		/* Idle cpus in this group. */
209 	cpumask_t tdg_mask;		/* Bit mask for first cpu. */
210 	int	tdg_load;		/* Total load of this group. */
211 	int	tdg_transferable;	/* Transferable load of this group. */
212 	LIST_HEAD(, tdq) tdg_members;	/* Linked list of all members. */
213 };
214 
215 #define	SCHED_AFFINITY_DEFAULT	(hz / 100)
216 #define	SCHED_AFFINITY(ts)	((ts)->ts_rltick > ticks - affinity)
217 
218 /*
219  * Run-time tunables.
220  */
221 static int rebalance = 0;
222 static int pick_pri = 1;
223 static int affinity;
224 static int tryself = 1;
225 static int tryselfidle = 1;
226 static int ipi_ast = 0;
227 static int ipi_preempt = 1;
228 static int ipi_thresh = PRI_MIN_KERN;
229 static int steal_htt = 1;
230 static int steal_busy = 1;
231 static int busy_thresh = 4;
232 
233 /*
234  * One thread queue per processor.
235  */
236 static volatile cpumask_t tdq_idle;
237 static volatile cpumask_t tdq_busy;
238 static int tdg_maxid;
239 static struct tdq	tdq_cpu[MAXCPU];
240 static struct tdq_group tdq_groups[MAXCPU];
241 static int bal_tick;
242 static int gbal_tick;
243 static int balance_groups;
244 
245 #define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
246 #define	TDQ_CPU(x)	(&tdq_cpu[(x)])
247 #define	TDQ_ID(x)	((x) - tdq_cpu)
248 #define	TDQ_GROUP(x)	(&tdq_groups[(x)])
249 #else	/* !SMP */
250 static struct tdq	tdq_cpu;
251 
252 #define	TDQ_SELF()	(&tdq_cpu)
253 #define	TDQ_CPU(x)	(&tdq_cpu)
254 #endif
255 
256 static void sched_priority(struct thread *);
257 static void sched_thread_priority(struct thread *, u_char);
258 static int sched_interact_score(struct thread *);
259 static void sched_interact_update(struct thread *);
260 static void sched_interact_fork(struct thread *);
261 static void sched_pctcpu_update(struct td_sched *);
262 static inline void sched_pin_td(struct thread *td);
263 static inline void sched_unpin_td(struct thread *td);
264 
265 /* Operations on per processor queues */
266 static struct td_sched * tdq_choose(struct tdq *);
267 static void tdq_setup(struct tdq *);
268 static void tdq_load_add(struct tdq *, struct td_sched *);
269 static void tdq_load_rem(struct tdq *, struct td_sched *);
270 static __inline void tdq_runq_add(struct tdq *, struct td_sched *, int);
271 static __inline void tdq_runq_rem(struct tdq *, struct td_sched *);
272 void tdq_print(int cpu);
273 static void runq_print(struct runq *rq);
274 #ifdef SMP
275 static int tdq_pickidle(struct tdq *, struct td_sched *);
276 static int tdq_pickpri(struct tdq *, struct td_sched *, int);
277 static struct td_sched *runq_steal(struct runq *);
278 static void sched_balance(void);
279 static void sched_balance_groups(void);
280 static void sched_balance_group(struct tdq_group *);
281 static void sched_balance_pair(struct tdq *, struct tdq *);
282 static void sched_smp_tick(struct thread *);
283 static void tdq_move(struct tdq *, int);
284 static int tdq_idled(struct tdq *);
285 static void tdq_notify(struct td_sched *);
286 static struct td_sched *tdq_steal(struct tdq *, int);
287 
288 #define	THREAD_CAN_MIGRATE(td)	 ((td)->td_pinned == 0)
289 #endif
290 
291 static void sched_setup(void *dummy);
292 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL)
293 
294 static void sched_initticks(void *dummy);
295 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, NULL)
296 
297 static inline void
298 sched_pin_td(struct thread *td)
299 {
300 	td->td_pinned++;
301 }
302 
303 static inline void
304 sched_unpin_td(struct thread *td)
305 {
306 	td->td_pinned--;
307 }
308 
309 static void
310 runq_print(struct runq *rq)
311 {
312 	struct rqhead *rqh;
313 	struct td_sched *ts;
314 	int pri;
315 	int j;
316 	int i;
317 
318 	for (i = 0; i < RQB_LEN; i++) {
319 		printf("\t\trunq bits %d 0x%zx\n",
320 		    i, rq->rq_status.rqb_bits[i]);
321 		for (j = 0; j < RQB_BPW; j++)
322 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
323 				pri = j + (i << RQB_L2BPW);
324 				rqh = &rq->rq_queues[pri];
325 				TAILQ_FOREACH(ts, rqh, ts_procq) {
326 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
327 					    ts->ts_thread, ts->ts_thread->td_proc->p_comm, ts->ts_thread->td_priority, ts->ts_rqindex, pri);
328 				}
329 			}
330 	}
331 }
332 
333 void
334 tdq_print(int cpu)
335 {
336 	struct tdq *tdq;
337 
338 	tdq = TDQ_CPU(cpu);
339 
340 	printf("tdq:\n");
341 	printf("\tload:           %d\n", tdq->tdq_load);
342 	printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
343 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
344 	printf("\trealtime runq:\n");
345 	runq_print(&tdq->tdq_realtime);
346 	printf("\ttimeshare runq:\n");
347 	runq_print(&tdq->tdq_timeshare);
348 	printf("\tidle runq:\n");
349 	runq_print(&tdq->tdq_idle);
350 #ifdef SMP
351 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
352 #endif
353 }
354 
355 static __inline void
356 tdq_runq_add(struct tdq *tdq, struct td_sched *ts, int flags)
357 {
358 #ifdef SMP
359 	if (THREAD_CAN_MIGRATE(ts->ts_thread)) {
360 		tdq->tdq_transferable++;
361 		tdq->tdq_group->tdg_transferable++;
362 		ts->ts_flags |= TSF_XFERABLE;
363 		if (tdq->tdq_transferable >= busy_thresh &&
364 		    (tdq->tdq_flags & TDQF_BUSY) == 0) {
365 			tdq->tdq_flags |= TDQF_BUSY;
366 			atomic_set_int(&tdq_busy, 1 << TDQ_ID(tdq));
367 		}
368 	}
369 #endif
370 	if (ts->ts_runq == &tdq->tdq_timeshare) {
371 		u_char pri;
372 
373 		pri = ts->ts_thread->td_priority;
374 		KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE,
375 			("Invalid priority %d on timeshare runq", pri));
376 		/*
377 		 * This queue contains only priorities between MIN and MAX
378 		 * realtime.  Use the whole queue to represent these values.
379 		 */
380 #define	TS_RQ_PPQ	(((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS)
381 		if ((flags & SRQ_BORROWING) == 0) {
382 			pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ;
383 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
384 			/*
385 			 * This effectively shortens the queue by one so we
386 			 * can have a one slot difference between idx and
387 			 * ridx while we wait for threads to drain.
388 			 */
389 			if (tdq->tdq_ridx != tdq->tdq_idx &&
390 			    pri == tdq->tdq_ridx)
391 				pri = (pri - 1) % RQ_NQS;
392 		} else
393 			pri = tdq->tdq_ridx;
394 		runq_add_pri(ts->ts_runq, ts, pri, flags);
395 	} else
396 		runq_add(ts->ts_runq, ts, flags);
397 }
398 
399 static __inline void
400 tdq_runq_rem(struct tdq *tdq, struct td_sched *ts)
401 {
402 #ifdef SMP
403 	if (ts->ts_flags & TSF_XFERABLE) {
404 		tdq->tdq_transferable--;
405 		tdq->tdq_group->tdg_transferable--;
406 		ts->ts_flags &= ~TSF_XFERABLE;
407 		if (tdq->tdq_transferable < busy_thresh &&
408 		    (tdq->tdq_flags & TDQF_BUSY)) {
409 			atomic_clear_int(&tdq_busy, 1 << TDQ_ID(tdq));
410 			tdq->tdq_flags &= ~TDQF_BUSY;
411 		}
412 	}
413 #endif
414 	if (ts->ts_runq == &tdq->tdq_timeshare) {
415 		if (tdq->tdq_idx != tdq->tdq_ridx)
416 			runq_remove_idx(ts->ts_runq, ts, &tdq->tdq_ridx);
417 		else
418 			runq_remove_idx(ts->ts_runq, ts, NULL);
419 		/*
420 		 * For timeshare threads we update the priority here so
421 		 * the priority reflects the time we've been sleeping.
422 		 */
423 		ts->ts_ltick = ticks;
424 		sched_pctcpu_update(ts);
425 		sched_priority(ts->ts_thread);
426 	} else
427 		runq_remove(ts->ts_runq, ts);
428 }
429 
430 static void
431 tdq_load_add(struct tdq *tdq, struct td_sched *ts)
432 {
433 	int class;
434 	mtx_assert(&sched_lock, MA_OWNED);
435 	class = PRI_BASE(ts->ts_thread->td_pri_class);
436 	tdq->tdq_load++;
437 	CTR1(KTR_SCHED, "load: %d", tdq->tdq_load);
438 	if (class != PRI_ITHD &&
439 	    (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0)
440 #ifdef SMP
441 		tdq->tdq_group->tdg_load++;
442 #else
443 		tdq->tdq_sysload++;
444 #endif
445 }
446 
447 static void
448 tdq_load_rem(struct tdq *tdq, struct td_sched *ts)
449 {
450 	int class;
451 	mtx_assert(&sched_lock, MA_OWNED);
452 	class = PRI_BASE(ts->ts_thread->td_pri_class);
453 	if (class != PRI_ITHD &&
454 	    (ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0)
455 #ifdef SMP
456 		tdq->tdq_group->tdg_load--;
457 #else
458 		tdq->tdq_sysload--;
459 #endif
460 	tdq->tdq_load--;
461 	CTR1(KTR_SCHED, "load: %d", tdq->tdq_load);
462 	ts->ts_runq = NULL;
463 }
464 
465 #ifdef SMP
466 static void
467 sched_smp_tick(struct thread *td)
468 {
469 	struct tdq *tdq;
470 
471 	tdq = TDQ_SELF();
472 	if (rebalance) {
473 		if (ticks >= bal_tick)
474 			sched_balance();
475 		if (ticks >= gbal_tick && balance_groups)
476 			sched_balance_groups();
477 	}
478 	td->td_sched->ts_rltick = ticks;
479 }
480 
481 /*
482  * sched_balance is a simple CPU load balancing algorithm.  It operates by
483  * finding the least loaded and most loaded cpu and equalizing their load
484  * by migrating some processes.
485  *
486  * Dealing only with two CPUs at a time has two advantages.  Firstly, most
487  * installations will only have 2 cpus.  Secondly, load balancing too much at
488  * once can have an unpleasant effect on the system.  The scheduler rarely has
489  * enough information to make perfect decisions.  So this algorithm chooses
490  * algorithm simplicity and more gradual effects on load in larger systems.
491  *
492  * It could be improved by considering the priorities and slices assigned to
493  * each task prior to balancing them.  There are many pathological cases with
494  * any approach and so the semi random algorithm below may work as well as any.
495  *
496  */
497 static void
498 sched_balance(void)
499 {
500 	struct tdq_group *high;
501 	struct tdq_group *low;
502 	struct tdq_group *tdg;
503 	int cnt;
504 	int i;
505 
506 	bal_tick = ticks + (random() % (hz * 2));
507 	if (smp_started == 0)
508 		return;
509 	low = high = NULL;
510 	i = random() % (tdg_maxid + 1);
511 	for (cnt = 0; cnt <= tdg_maxid; cnt++) {
512 		tdg = TDQ_GROUP(i);
513 		/*
514 		 * Find the CPU with the highest load that has some
515 		 * threads to transfer.
516 		 */
517 		if ((high == NULL || tdg->tdg_load > high->tdg_load)
518 		    && tdg->tdg_transferable)
519 			high = tdg;
520 		if (low == NULL || tdg->tdg_load < low->tdg_load)
521 			low = tdg;
522 		if (++i > tdg_maxid)
523 			i = 0;
524 	}
525 	if (low != NULL && high != NULL && high != low)
526 		sched_balance_pair(LIST_FIRST(&high->tdg_members),
527 		    LIST_FIRST(&low->tdg_members));
528 }
529 
530 static void
531 sched_balance_groups(void)
532 {
533 	int i;
534 
535 	gbal_tick = ticks + (random() % (hz * 2));
536 	mtx_assert(&sched_lock, MA_OWNED);
537 	if (smp_started)
538 		for (i = 0; i <= tdg_maxid; i++)
539 			sched_balance_group(TDQ_GROUP(i));
540 }
541 
542 static void
543 sched_balance_group(struct tdq_group *tdg)
544 {
545 	struct tdq *tdq;
546 	struct tdq *high;
547 	struct tdq *low;
548 	int load;
549 
550 	if (tdg->tdg_transferable == 0)
551 		return;
552 	low = NULL;
553 	high = NULL;
554 	LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) {
555 		load = tdq->tdq_load;
556 		if (high == NULL || load > high->tdq_load)
557 			high = tdq;
558 		if (low == NULL || load < low->tdq_load)
559 			low = tdq;
560 	}
561 	if (high != NULL && low != NULL && high != low)
562 		sched_balance_pair(high, low);
563 }
564 
565 static void
566 sched_balance_pair(struct tdq *high, struct tdq *low)
567 {
568 	int transferable;
569 	int high_load;
570 	int low_load;
571 	int move;
572 	int diff;
573 	int i;
574 
575 	/*
576 	 * If we're transfering within a group we have to use this specific
577 	 * tdq's transferable count, otherwise we can steal from other members
578 	 * of the group.
579 	 */
580 	if (high->tdq_group == low->tdq_group) {
581 		transferable = high->tdq_transferable;
582 		high_load = high->tdq_load;
583 		low_load = low->tdq_load;
584 	} else {
585 		transferable = high->tdq_group->tdg_transferable;
586 		high_load = high->tdq_group->tdg_load;
587 		low_load = low->tdq_group->tdg_load;
588 	}
589 	if (transferable == 0)
590 		return;
591 	/*
592 	 * Determine what the imbalance is and then adjust that to how many
593 	 * threads we actually have to give up (transferable).
594 	 */
595 	diff = high_load - low_load;
596 	move = diff / 2;
597 	if (diff & 0x1)
598 		move++;
599 	move = min(move, transferable);
600 	for (i = 0; i < move; i++)
601 		tdq_move(high, TDQ_ID(low));
602 	return;
603 }
604 
605 static void
606 tdq_move(struct tdq *from, int cpu)
607 {
608 	struct tdq *tdq;
609 	struct tdq *to;
610 	struct td_sched *ts;
611 
612 	tdq = from;
613 	to = TDQ_CPU(cpu);
614 	ts = tdq_steal(tdq, 1);
615 	if (ts == NULL) {
616 		struct tdq_group *tdg;
617 
618 		tdg = tdq->tdq_group;
619 		LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) {
620 			if (tdq == from || tdq->tdq_transferable == 0)
621 				continue;
622 			ts = tdq_steal(tdq, 1);
623 			break;
624 		}
625 		if (ts == NULL)
626 			panic("tdq_move: No threads available with a "
627 			    "transferable count of %d\n",
628 			    tdg->tdg_transferable);
629 	}
630 	if (tdq == to)
631 		return;
632 	sched_rem(ts->ts_thread);
633 	ts->ts_cpu = cpu;
634 	sched_pin_td(ts->ts_thread);
635 	sched_add(ts->ts_thread, SRQ_YIELDING);
636 	sched_unpin_td(ts->ts_thread);
637 }
638 
639 static int
640 tdq_idled(struct tdq *tdq)
641 {
642 	struct tdq_group *tdg;
643 	struct tdq *steal;
644 	struct td_sched *ts;
645 
646 	tdg = tdq->tdq_group;
647 	/*
648 	 * If we're in a cpu group, try and steal threads from another cpu in
649 	 * the group before idling.
650 	 */
651 	if (steal_htt && tdg->tdg_cpus > 1 && tdg->tdg_transferable) {
652 		LIST_FOREACH(steal, &tdg->tdg_members, tdq_siblings) {
653 			if (steal == tdq || steal->tdq_transferable == 0)
654 				continue;
655 			ts = tdq_steal(steal, 0);
656 			if (ts)
657 				goto steal;
658 		}
659 	}
660 	if (steal_busy) {
661 		while (tdq_busy) {
662 			int cpu;
663 
664 			cpu = ffs(tdq_busy);
665 			if (cpu == 0)
666 				break;
667 			cpu--;
668 			steal = TDQ_CPU(cpu);
669 			if (steal->tdq_transferable == 0)
670 				continue;
671 			ts = tdq_steal(steal, 1);
672 			if (ts == NULL)
673 				continue;
674 			CTR5(KTR_ULE,
675 			    "tdq_idled: stealing td %p(%s) pri %d from %d busy 0x%X",
676 			    ts->ts_thread, ts->ts_thread->td_proc->p_comm,
677 			    ts->ts_thread->td_priority, cpu, tdq_busy);
678 			goto steal;
679 		}
680 	}
681 	/*
682 	 * We only set the idled bit when all of the cpus in the group are
683 	 * idle.  Otherwise we could get into a situation where a thread bounces
684 	 * back and forth between two idle cores on seperate physical CPUs.
685 	 */
686 	tdg->tdg_idlemask |= PCPU_GET(cpumask);
687 	if (tdg->tdg_idlemask == tdg->tdg_cpumask)
688 		atomic_set_int(&tdq_idle, tdg->tdg_mask);
689 	return (1);
690 steal:
691 	sched_rem(ts->ts_thread);
692 	ts->ts_cpu = PCPU_GET(cpuid);
693 	sched_pin_td(ts->ts_thread);
694 	sched_add(ts->ts_thread, SRQ_YIELDING);
695 	sched_unpin_td(ts->ts_thread);
696 
697 	return (0);
698 }
699 
700 static void
701 tdq_notify(struct td_sched *ts)
702 {
703 	struct thread *ctd;
704 	struct pcpu *pcpu;
705 	int cpri;
706 	int pri;
707 	int cpu;
708 
709 	cpu = ts->ts_cpu;
710 	pri = ts->ts_thread->td_priority;
711 	pcpu = pcpu_find(cpu);
712 	ctd = pcpu->pc_curthread;
713 	cpri = ctd->td_priority;
714 
715 	/*
716 	 * If our priority is not better than the current priority there is
717 	 * nothing to do.
718 	 */
719 	if (pri > cpri)
720 		return;
721 	/*
722 	 * Always IPI idle.
723 	 */
724 	if (cpri > PRI_MIN_IDLE)
725 		goto sendipi;
726 	/*
727 	 * If we're realtime or better and there is timeshare or worse running
728 	 * send an IPI.
729 	 */
730 	if (pri < PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME)
731 		goto sendipi;
732 	/*
733 	 * Otherwise only IPI if we exceed the threshold.
734 	 */
735 	if (pri > ipi_thresh)
736 		return;
737 sendipi:
738 	ctd->td_flags |= TDF_NEEDRESCHED;
739 	if (cpri < PRI_MIN_IDLE) {
740 		if (ipi_ast)
741 			ipi_selected(1 << cpu, IPI_AST);
742 		else if (ipi_preempt)
743 			ipi_selected(1 << cpu, IPI_PREEMPT);
744 	} else
745 		ipi_selected(1 << cpu, IPI_PREEMPT);
746 }
747 
748 static struct td_sched *
749 runq_steal(struct runq *rq)
750 {
751 	struct rqhead *rqh;
752 	struct rqbits *rqb;
753 	struct td_sched *ts;
754 	int word;
755 	int bit;
756 
757 	mtx_assert(&sched_lock, MA_OWNED);
758 	rqb = &rq->rq_status;
759 	for (word = 0; word < RQB_LEN; word++) {
760 		if (rqb->rqb_bits[word] == 0)
761 			continue;
762 		for (bit = 0; bit < RQB_BPW; bit++) {
763 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
764 				continue;
765 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
766 			TAILQ_FOREACH(ts, rqh, ts_procq) {
767 				if (THREAD_CAN_MIGRATE(ts->ts_thread))
768 					return (ts);
769 			}
770 		}
771 	}
772 	return (NULL);
773 }
774 
775 static struct td_sched *
776 tdq_steal(struct tdq *tdq, int stealidle)
777 {
778 	struct td_sched *ts;
779 
780 	/*
781 	 * Steal from next first to try to get a non-interactive task that
782 	 * may not have run for a while.
783 	 * XXX Need to effect steal order for timeshare threads.
784 	 */
785 	if ((ts = runq_steal(&tdq->tdq_realtime)) != NULL)
786 		return (ts);
787 	if ((ts = runq_steal(&tdq->tdq_timeshare)) != NULL)
788 		return (ts);
789 	if (stealidle)
790 		return (runq_steal(&tdq->tdq_idle));
791 	return (NULL);
792 }
793 
794 int
795 tdq_pickidle(struct tdq *tdq, struct td_sched *ts)
796 {
797 	struct tdq_group *tdg;
798 	int self;
799 	int cpu;
800 
801 	self = PCPU_GET(cpuid);
802 	if (smp_started == 0)
803 		return (self);
804 	/*
805 	 * If the current CPU has idled, just run it here.
806 	 */
807 	if ((tdq->tdq_group->tdg_idlemask & PCPU_GET(cpumask)) != 0)
808 		return (self);
809 	/*
810 	 * Try the last group we ran on.
811 	 */
812 	tdg = TDQ_CPU(ts->ts_cpu)->tdq_group;
813 	cpu = ffs(tdg->tdg_idlemask);
814 	if (cpu)
815 		return (cpu - 1);
816 	/*
817 	 * Search for an idle group.
818 	 */
819 	cpu = ffs(tdq_idle);
820 	if (cpu)
821 		return (cpu - 1);
822 	/*
823 	 * XXX If there are no idle groups, check for an idle core.
824 	 */
825 	/*
826 	 * No idle CPUs?
827 	 */
828 	return (self);
829 }
830 
831 static int
832 tdq_pickpri(struct tdq *tdq, struct td_sched *ts, int flags)
833 {
834 	struct pcpu *pcpu;
835 	int lowpri;
836 	int lowcpu;
837 	int lowload;
838 	int load;
839 	int self;
840 	int pri;
841 	int cpu;
842 
843 	self = PCPU_GET(cpuid);
844 	if (smp_started == 0)
845 		return (self);
846 
847 	pri = ts->ts_thread->td_priority;
848 	/*
849 	 * Regardless of affinity, if the last cpu is idle send it there.
850 	 */
851 	pcpu = pcpu_find(ts->ts_cpu);
852 	if (pcpu->pc_curthread->td_priority > PRI_MIN_IDLE) {
853 		CTR5(KTR_ULE,
854 		    "ts_cpu %d idle, ltick %d ticks %d pri %d curthread %d",
855 		    ts->ts_cpu, ts->ts_rltick, ticks, pri,
856 		    pcpu->pc_curthread->td_priority);
857 		return (ts->ts_cpu);
858 	}
859 	/*
860 	 * If we have affinity, try to place it on the cpu we last ran on.
861 	 */
862 	if (SCHED_AFFINITY(ts) && pcpu->pc_curthread->td_priority > pri) {
863 		CTR5(KTR_ULE,
864 		    "affinity for %d, ltick %d ticks %d pri %d curthread %d",
865 		    ts->ts_cpu, ts->ts_rltick, ticks, pri,
866 		    pcpu->pc_curthread->td_priority);
867 		return (ts->ts_cpu);
868 	}
869 	/*
870 	 * Try ourself first; If we're running something lower priority this
871 	 * may have some locality with the waking thread and execute faster
872 	 * here.
873 	 */
874 	if (tryself) {
875 		/*
876 		 * If we're being awoken by an interrupt thread or the waker
877 		 * is going right to sleep run here as well.
878 		 */
879 		if ((TDQ_SELF()->tdq_load == 1) && (flags & SRQ_YIELDING ||
880 		    curthread->td_pri_class == PRI_ITHD)) {
881 			CTR2(KTR_ULE, "tryself load %d flags %d",
882 			    TDQ_SELF()->tdq_load, flags);
883 			return (self);
884 		}
885 	}
886 	/*
887 	 * Look for an idle group.
888 	 */
889 	CTR1(KTR_ULE, "tdq_idle %X", tdq_idle);
890 	cpu = ffs(tdq_idle);
891 	if (cpu)
892 		return (cpu - 1);
893 	if (tryselfidle && pri < curthread->td_priority) {
894 		CTR1(KTR_ULE, "tryself %d",
895 		    curthread->td_priority);
896 		return (self);
897 	}
898 	/*
899  	 * Now search for the cpu running the lowest priority thread with
900 	 * the least load.
901 	 */
902 	lowload = 0;
903 	lowpri = lowcpu = 0;
904 	for (cpu = 0; cpu <= mp_maxid; cpu++) {
905 		if (CPU_ABSENT(cpu))
906 			continue;
907 		pcpu = pcpu_find(cpu);
908 		pri = pcpu->pc_curthread->td_priority;
909 		CTR4(KTR_ULE,
910 		    "cpu %d pri %d lowcpu %d lowpri %d",
911 		    cpu, pri, lowcpu, lowpri);
912 		if (pri < lowpri)
913 			continue;
914 		load = TDQ_CPU(cpu)->tdq_load;
915 		if (lowpri && lowpri == pri && load > lowload)
916 			continue;
917 		lowpri = pri;
918 		lowcpu = cpu;
919 		lowload = load;
920 	}
921 
922 	return (lowcpu);
923 }
924 
925 #endif	/* SMP */
926 
927 /*
928  * Pick the highest priority task we have and return it.
929  */
930 
931 static struct td_sched *
932 tdq_choose(struct tdq *tdq)
933 {
934 	struct td_sched *ts;
935 
936 	mtx_assert(&sched_lock, MA_OWNED);
937 
938 	ts = runq_choose(&tdq->tdq_realtime);
939 	if (ts != NULL) {
940 		KASSERT(ts->ts_thread->td_priority <= PRI_MAX_REALTIME,
941 		    ("tdq_choose: Invalid priority on realtime queue %d",
942 		    ts->ts_thread->td_priority));
943 		return (ts);
944 	}
945 	ts = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
946 	if (ts != NULL) {
947 		KASSERT(ts->ts_thread->td_priority <= PRI_MAX_TIMESHARE &&
948 		    ts->ts_thread->td_priority >= PRI_MIN_TIMESHARE,
949 		    ("tdq_choose: Invalid priority on timeshare queue %d",
950 		    ts->ts_thread->td_priority));
951 		return (ts);
952 	}
953 
954 	ts = runq_choose(&tdq->tdq_idle);
955 	if (ts != NULL) {
956 		KASSERT(ts->ts_thread->td_priority >= PRI_MIN_IDLE,
957 		    ("tdq_choose: Invalid priority on idle queue %d",
958 		    ts->ts_thread->td_priority));
959 		return (ts);
960 	}
961 
962 	return (NULL);
963 }
964 
965 static void
966 tdq_setup(struct tdq *tdq)
967 {
968 	runq_init(&tdq->tdq_realtime);
969 	runq_init(&tdq->tdq_timeshare);
970 	runq_init(&tdq->tdq_idle);
971 	tdq->tdq_load = 0;
972 }
973 
974 static void
975 sched_setup(void *dummy)
976 {
977 #ifdef SMP
978 	int i;
979 #endif
980 
981 	/*
982 	 * To avoid divide-by-zero, we set realstathz a dummy value
983 	 * in case which sched_clock() called before sched_initticks().
984 	 */
985 	realstathz = hz;
986 	sched_slice = (realstathz/10);	/* ~100ms */
987 	tickincr = 1 << SCHED_TICK_SHIFT;
988 
989 #ifdef SMP
990 	balance_groups = 0;
991 	/*
992 	 * Initialize the tdqs.
993 	 */
994 	for (i = 0; i < MAXCPU; i++) {
995 		struct tdq *tdq;
996 
997 		tdq = &tdq_cpu[i];
998 		tdq_setup(&tdq_cpu[i]);
999 	}
1000 	if (smp_topology == NULL) {
1001 		struct tdq_group *tdg;
1002 		struct tdq *tdq;
1003 		int cpus;
1004 
1005 		for (cpus = 0, i = 0; i < MAXCPU; i++) {
1006 			if (CPU_ABSENT(i))
1007 				continue;
1008 			tdq = &tdq_cpu[i];
1009 			tdg = &tdq_groups[cpus];
1010 			/*
1011 			 * Setup a tdq group with one member.
1012 			 */
1013 			tdq->tdq_transferable = 0;
1014 			tdq->tdq_group = tdg;
1015 			tdg->tdg_cpus = 1;
1016 			tdg->tdg_idlemask = 0;
1017 			tdg->tdg_cpumask = tdg->tdg_mask = 1 << i;
1018 			tdg->tdg_load = 0;
1019 			tdg->tdg_transferable = 0;
1020 			LIST_INIT(&tdg->tdg_members);
1021 			LIST_INSERT_HEAD(&tdg->tdg_members, tdq, tdq_siblings);
1022 			cpus++;
1023 		}
1024 		tdg_maxid = cpus - 1;
1025 	} else {
1026 		struct tdq_group *tdg;
1027 		struct cpu_group *cg;
1028 		int j;
1029 
1030 		for (i = 0; i < smp_topology->ct_count; i++) {
1031 			cg = &smp_topology->ct_group[i];
1032 			tdg = &tdq_groups[i];
1033 			/*
1034 			 * Initialize the group.
1035 			 */
1036 			tdg->tdg_idlemask = 0;
1037 			tdg->tdg_load = 0;
1038 			tdg->tdg_transferable = 0;
1039 			tdg->tdg_cpus = cg->cg_count;
1040 			tdg->tdg_cpumask = cg->cg_mask;
1041 			LIST_INIT(&tdg->tdg_members);
1042 			/*
1043 			 * Find all of the group members and add them.
1044 			 */
1045 			for (j = 0; j < MAXCPU; j++) {
1046 				if ((cg->cg_mask & (1 << j)) != 0) {
1047 					if (tdg->tdg_mask == 0)
1048 						tdg->tdg_mask = 1 << j;
1049 					tdq_cpu[j].tdq_transferable = 0;
1050 					tdq_cpu[j].tdq_group = tdg;
1051 					LIST_INSERT_HEAD(&tdg->tdg_members,
1052 					    &tdq_cpu[j], tdq_siblings);
1053 				}
1054 			}
1055 			if (tdg->tdg_cpus > 1)
1056 				balance_groups = 1;
1057 		}
1058 		tdg_maxid = smp_topology->ct_count - 1;
1059 	}
1060 	/*
1061 	 * Stagger the group and global load balancer so they do not
1062 	 * interfere with each other.
1063 	 */
1064 	bal_tick = ticks + hz;
1065 	if (balance_groups)
1066 		gbal_tick = ticks + (hz / 2);
1067 #else
1068 	tdq_setup(TDQ_SELF());
1069 #endif
1070 	mtx_lock_spin(&sched_lock);
1071 	tdq_load_add(TDQ_SELF(), &td_sched0);
1072 	mtx_unlock_spin(&sched_lock);
1073 }
1074 
1075 /* ARGSUSED */
1076 static void
1077 sched_initticks(void *dummy)
1078 {
1079 	mtx_lock_spin(&sched_lock);
1080 	realstathz = stathz ? stathz : hz;
1081 	sched_slice = (realstathz/10);	/* ~100ms */
1082 
1083 	/*
1084 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1085 	 * hz not being evenly divisible by stathz on all platforms.
1086 	 */
1087 	tickincr = (hz << SCHED_TICK_SHIFT) / realstathz;
1088 	/*
1089 	 * This does not work for values of stathz that are more than
1090 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1091 	 */
1092 	if (tickincr == 0)
1093 		tickincr = 1;
1094 #ifdef SMP
1095 	affinity = SCHED_AFFINITY_DEFAULT;
1096 #endif
1097 	mtx_unlock_spin(&sched_lock);
1098 }
1099 
1100 
1101 /*
1102  * Scale the scheduling priority according to the "interactivity" of this
1103  * process.
1104  */
1105 static void
1106 sched_priority(struct thread *td)
1107 {
1108 	int score;
1109 	int pri;
1110 
1111 	if (td->td_pri_class != PRI_TIMESHARE)
1112 		return;
1113 	/*
1114 	 * If the score is interactive we place the thread in the realtime
1115 	 * queue with a priority that is less than kernel and interrupt
1116 	 * priorities.  These threads are not subject to nice restrictions.
1117 	 *
1118 	 * Scores greater than this are placed on the normal realtime queue
1119 	 * where the priority is partially decided by the most recent cpu
1120 	 * utilization and the rest is decided by nice value.
1121 	 */
1122 	score = sched_interact_score(td);
1123 	if (score < sched_interact) {
1124 		pri = PRI_MIN_REALTIME;
1125 		pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact)
1126 		    * score;
1127 		KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME,
1128 		    ("sched_priority: invalid interactive priority %d score %d",
1129 		    pri, score));
1130 	} else {
1131 		pri = SCHED_PRI_MIN;
1132 		if (td->td_sched->ts_ticks)
1133 			pri += SCHED_PRI_TICKS(td->td_sched);
1134 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1135 		if (!(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE)) {
1136 			static int once = 1;
1137 			if (once) {
1138 				printf("sched_priority: invalid priority %d",
1139 				    pri);
1140 				printf("nice %d, ticks %d ftick %d ltick %d tick pri %d\n",
1141 				    td->td_proc->p_nice,
1142 				    td->td_sched->ts_ticks,
1143 				    td->td_sched->ts_ftick,
1144 				    td->td_sched->ts_ltick,
1145 				    SCHED_PRI_TICKS(td->td_sched));
1146 				once = 0;
1147 			}
1148 			pri = min(max(pri, PRI_MIN_TIMESHARE),
1149 			    PRI_MAX_TIMESHARE);
1150 		}
1151 	}
1152 	sched_user_prio(td, pri);
1153 
1154 	return;
1155 }
1156 
1157 /*
1158  * This routine enforces a maximum limit on the amount of scheduling history
1159  * kept.  It is called after either the slptime or runtime is adjusted.
1160  */
1161 static void
1162 sched_interact_update(struct thread *td)
1163 {
1164 	struct td_sched *ts;
1165 	u_int sum;
1166 
1167 	ts = td->td_sched;
1168 	sum = ts->skg_runtime + ts->skg_slptime;
1169 	if (sum < SCHED_SLP_RUN_MAX)
1170 		return;
1171 	/*
1172 	 * This only happens from two places:
1173 	 * 1) We have added an unusual amount of run time from fork_exit.
1174 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1175 	 */
1176 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1177 		if (ts->skg_runtime > ts->skg_slptime) {
1178 			ts->skg_runtime = SCHED_SLP_RUN_MAX;
1179 			ts->skg_slptime = 1;
1180 		} else {
1181 			ts->skg_slptime = SCHED_SLP_RUN_MAX;
1182 			ts->skg_runtime = 1;
1183 		}
1184 		return;
1185 	}
1186 	/*
1187 	 * If we have exceeded by more than 1/5th then the algorithm below
1188 	 * will not bring us back into range.  Dividing by two here forces
1189 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1190 	 */
1191 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1192 		ts->skg_runtime /= 2;
1193 		ts->skg_slptime /= 2;
1194 		return;
1195 	}
1196 	ts->skg_runtime = (ts->skg_runtime / 5) * 4;
1197 	ts->skg_slptime = (ts->skg_slptime / 5) * 4;
1198 }
1199 
1200 static void
1201 sched_interact_fork(struct thread *td)
1202 {
1203 	int ratio;
1204 	int sum;
1205 
1206 	sum = td->td_sched->skg_runtime + td->td_sched->skg_slptime;
1207 	if (sum > SCHED_SLP_RUN_FORK) {
1208 		ratio = sum / SCHED_SLP_RUN_FORK;
1209 		td->td_sched->skg_runtime /= ratio;
1210 		td->td_sched->skg_slptime /= ratio;
1211 	}
1212 }
1213 
1214 static int
1215 sched_interact_score(struct thread *td)
1216 {
1217 	int div;
1218 
1219 	if (td->td_sched->skg_runtime > td->td_sched->skg_slptime) {
1220 		div = max(1, td->td_sched->skg_runtime / SCHED_INTERACT_HALF);
1221 		return (SCHED_INTERACT_HALF +
1222 		    (SCHED_INTERACT_HALF - (td->td_sched->skg_slptime / div)));
1223 	} if (td->td_sched->skg_slptime > td->td_sched->skg_runtime) {
1224 		div = max(1, td->td_sched->skg_slptime / SCHED_INTERACT_HALF);
1225 		return (td->td_sched->skg_runtime / div);
1226 	}
1227 
1228 	/*
1229 	 * This can happen if slptime and runtime are 0.
1230 	 */
1231 	return (0);
1232 
1233 }
1234 
1235 /*
1236  * Called from proc0_init() to bootstrap the scheduler.
1237  */
1238 void
1239 schedinit(void)
1240 {
1241 
1242 	/*
1243 	 * Set up the scheduler specific parts of proc0.
1244 	 */
1245 	proc0.p_sched = NULL; /* XXX */
1246 	thread0.td_sched = &td_sched0;
1247 	td_sched0.ts_ltick = ticks;
1248 	td_sched0.ts_ftick = ticks;
1249 	td_sched0.ts_thread = &thread0;
1250 }
1251 
1252 /*
1253  * This is only somewhat accurate since given many processes of the same
1254  * priority they will switch when their slices run out, which will be
1255  * at most sched_slice stathz ticks.
1256  */
1257 int
1258 sched_rr_interval(void)
1259 {
1260 
1261 	/* Convert sched_slice to hz */
1262 	return (hz/(realstathz/sched_slice));
1263 }
1264 
1265 static void
1266 sched_pctcpu_update(struct td_sched *ts)
1267 {
1268 
1269 	if (ts->ts_ticks == 0)
1270 		return;
1271 	if (ticks - (hz / 10) < ts->ts_ltick &&
1272 	    SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX)
1273 		return;
1274 	/*
1275 	 * Adjust counters and watermark for pctcpu calc.
1276 	 */
1277 	if (ts->ts_ltick > ticks - SCHED_TICK_TARG)
1278 		ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) *
1279 			    SCHED_TICK_TARG;
1280 	else
1281 		ts->ts_ticks = 0;
1282 	ts->ts_ltick = ticks;
1283 	ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG;
1284 }
1285 
1286 static void
1287 sched_thread_priority(struct thread *td, u_char prio)
1288 {
1289 	struct td_sched *ts;
1290 
1291 	CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)",
1292 	    td, td->td_proc->p_comm, td->td_priority, prio, curthread,
1293 	    curthread->td_proc->p_comm);
1294 	ts = td->td_sched;
1295 	mtx_assert(&sched_lock, MA_OWNED);
1296 	if (td->td_priority == prio)
1297 		return;
1298 
1299 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1300 		/*
1301 		 * If the priority has been elevated due to priority
1302 		 * propagation, we may have to move ourselves to a new
1303 		 * queue.  This could be optimized to not re-add in some
1304 		 * cases.
1305 		 */
1306 		sched_rem(td);
1307 		td->td_priority = prio;
1308 		sched_add(td, SRQ_BORROWING);
1309 	} else
1310 		td->td_priority = prio;
1311 }
1312 
1313 /*
1314  * Update a thread's priority when it is lent another thread's
1315  * priority.
1316  */
1317 void
1318 sched_lend_prio(struct thread *td, u_char prio)
1319 {
1320 
1321 	td->td_flags |= TDF_BORROWING;
1322 	sched_thread_priority(td, prio);
1323 }
1324 
1325 /*
1326  * Restore a thread's priority when priority propagation is
1327  * over.  The prio argument is the minimum priority the thread
1328  * needs to have to satisfy other possible priority lending
1329  * requests.  If the thread's regular priority is less
1330  * important than prio, the thread will keep a priority boost
1331  * of prio.
1332  */
1333 void
1334 sched_unlend_prio(struct thread *td, u_char prio)
1335 {
1336 	u_char base_pri;
1337 
1338 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1339 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1340 		base_pri = td->td_user_pri;
1341 	else
1342 		base_pri = td->td_base_pri;
1343 	if (prio >= base_pri) {
1344 		td->td_flags &= ~TDF_BORROWING;
1345 		sched_thread_priority(td, base_pri);
1346 	} else
1347 		sched_lend_prio(td, prio);
1348 }
1349 
1350 void
1351 sched_prio(struct thread *td, u_char prio)
1352 {
1353 	u_char oldprio;
1354 
1355 	/* First, update the base priority. */
1356 	td->td_base_pri = prio;
1357 
1358 	/*
1359 	 * If the thread is borrowing another thread's priority, don't
1360 	 * ever lower the priority.
1361 	 */
1362 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1363 		return;
1364 
1365 	/* Change the real priority. */
1366 	oldprio = td->td_priority;
1367 	sched_thread_priority(td, prio);
1368 
1369 	/*
1370 	 * If the thread is on a turnstile, then let the turnstile update
1371 	 * its state.
1372 	 */
1373 	if (TD_ON_LOCK(td) && oldprio != prio)
1374 		turnstile_adjust(td, oldprio);
1375 }
1376 
1377 void
1378 sched_user_prio(struct thread *td, u_char prio)
1379 {
1380 	u_char oldprio;
1381 
1382 	td->td_base_user_pri = prio;
1383 	if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
1384                 return;
1385 	oldprio = td->td_user_pri;
1386 	td->td_user_pri = prio;
1387 
1388 	if (TD_ON_UPILOCK(td) && oldprio != prio)
1389 		umtx_pi_adjust(td, oldprio);
1390 }
1391 
1392 void
1393 sched_lend_user_prio(struct thread *td, u_char prio)
1394 {
1395 	u_char oldprio;
1396 
1397 	td->td_flags |= TDF_UBORROWING;
1398 
1399 	oldprio = td->td_user_pri;
1400 	td->td_user_pri = prio;
1401 
1402 	if (TD_ON_UPILOCK(td) && oldprio != prio)
1403 		umtx_pi_adjust(td, oldprio);
1404 }
1405 
1406 void
1407 sched_unlend_user_prio(struct thread *td, u_char prio)
1408 {
1409 	u_char base_pri;
1410 
1411 	base_pri = td->td_base_user_pri;
1412 	if (prio >= base_pri) {
1413 		td->td_flags &= ~TDF_UBORROWING;
1414 		sched_user_prio(td, base_pri);
1415 	} else
1416 		sched_lend_user_prio(td, prio);
1417 }
1418 
1419 void
1420 sched_switch(struct thread *td, struct thread *newtd, int flags)
1421 {
1422 	struct tdq *tdq;
1423 	struct td_sched *ts;
1424 	int preempt;
1425 
1426 	mtx_assert(&sched_lock, MA_OWNED);
1427 
1428 	preempt = flags & SW_PREEMPT;
1429 	tdq = TDQ_SELF();
1430 	ts = td->td_sched;
1431 	td->td_lastcpu = td->td_oncpu;
1432 	td->td_oncpu = NOCPU;
1433 	td->td_flags &= ~TDF_NEEDRESCHED;
1434 	td->td_owepreempt = 0;
1435 	/*
1436 	 * If the thread has been assigned it may be in the process of switching
1437 	 * to the new cpu.  This is the case in sched_bind().
1438 	 */
1439 	if (td == PCPU_GET(idlethread)) {
1440 		TD_SET_CAN_RUN(td);
1441 	} else {
1442 		tdq_load_rem(tdq, ts);
1443 		if (TD_IS_RUNNING(td)) {
1444 			/*
1445 			 * Don't allow the thread to migrate
1446 			 * from a preemption.
1447 			 */
1448 			if (preempt)
1449 				sched_pin_td(td);
1450 			sched_add(td, preempt ?
1451 			    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1452 			    SRQ_OURSELF|SRQ_YIELDING);
1453 			if (preempt)
1454 				sched_unpin_td(td);
1455 		}
1456 	}
1457 	if (newtd != NULL) {
1458 		/*
1459 		 * If we bring in a thread account for it as if it had been
1460 		 * added to the run queue and then chosen.
1461 		 */
1462 		TD_SET_RUNNING(newtd);
1463 		tdq_load_add(TDQ_SELF(), newtd->td_sched);
1464 	} else
1465 		newtd = choosethread();
1466 	if (td != newtd) {
1467 #ifdef	HWPMC_HOOKS
1468 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1469 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1470 #endif
1471 
1472 		cpu_switch(td, newtd);
1473 #ifdef	HWPMC_HOOKS
1474 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1475 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1476 #endif
1477 	}
1478 	sched_lock.mtx_lock = (uintptr_t)td;
1479 	td->td_oncpu = PCPU_GET(cpuid);
1480 }
1481 
1482 void
1483 sched_nice(struct proc *p, int nice)
1484 {
1485 	struct thread *td;
1486 
1487 	PROC_LOCK_ASSERT(p, MA_OWNED);
1488 	mtx_assert(&sched_lock, MA_OWNED);
1489 
1490 	p->p_nice = nice;
1491 	FOREACH_THREAD_IN_PROC(p, td) {
1492 		sched_priority(td);
1493 		sched_prio(td, td->td_base_user_pri);
1494 	}
1495 }
1496 
1497 void
1498 sched_sleep(struct thread *td)
1499 {
1500 
1501 	mtx_assert(&sched_lock, MA_OWNED);
1502 
1503 	td->td_sched->ts_slptime = ticks;
1504 }
1505 
1506 void
1507 sched_wakeup(struct thread *td)
1508 {
1509 	struct td_sched *ts;
1510 	int slptime;
1511 
1512 	mtx_assert(&sched_lock, MA_OWNED);
1513 	ts = td->td_sched;
1514 	/*
1515 	 * If we slept for more than a tick update our interactivity and
1516 	 * priority.
1517 	 */
1518 	slptime = ts->ts_slptime;
1519 	ts->ts_slptime = 0;
1520 	if (slptime && slptime != ticks) {
1521 		u_int hzticks;
1522 
1523 		hzticks = (ticks - slptime) << SCHED_TICK_SHIFT;
1524 		ts->skg_slptime += hzticks;
1525 		sched_interact_update(td);
1526 		sched_pctcpu_update(ts);
1527 		sched_priority(td);
1528 	}
1529 	/* Reset the slice value after we sleep. */
1530 	ts->ts_slice = sched_slice;
1531 	sched_add(td, SRQ_BORING);
1532 }
1533 
1534 /*
1535  * Penalize the parent for creating a new child and initialize the child's
1536  * priority.
1537  */
1538 void
1539 sched_fork(struct thread *td, struct thread *child)
1540 {
1541 	mtx_assert(&sched_lock, MA_OWNED);
1542 	sched_fork_thread(td, child);
1543 	/*
1544 	 * Penalize the parent and child for forking.
1545 	 */
1546 	sched_interact_fork(child);
1547 	sched_priority(child);
1548 	td->td_sched->skg_runtime += tickincr;
1549 	sched_interact_update(td);
1550 	sched_priority(td);
1551 }
1552 
1553 void
1554 sched_fork_thread(struct thread *td, struct thread *child)
1555 {
1556 	struct td_sched *ts;
1557 	struct td_sched *ts2;
1558 
1559 	/*
1560 	 * Initialize child.
1561 	 */
1562 	sched_newthread(child);
1563 	ts = td->td_sched;
1564 	ts2 = child->td_sched;
1565 	ts2->ts_cpu = ts->ts_cpu;
1566 	ts2->ts_runq = NULL;
1567 	/*
1568 	 * Grab our parents cpu estimation information and priority.
1569 	 */
1570 	ts2->ts_ticks = ts->ts_ticks;
1571 	ts2->ts_ltick = ts->ts_ltick;
1572 	ts2->ts_ftick = ts->ts_ftick;
1573 	child->td_user_pri = td->td_user_pri;
1574 	child->td_base_user_pri = td->td_base_user_pri;
1575 	/*
1576 	 * And update interactivity score.
1577 	 */
1578 	ts2->skg_slptime = ts->skg_slptime;
1579 	ts2->skg_runtime = ts->skg_runtime;
1580 	ts2->ts_slice = 1;	/* Attempt to quickly learn interactivity. */
1581 }
1582 
1583 void
1584 sched_class(struct thread *td, int class)
1585 {
1586 
1587 	mtx_assert(&sched_lock, MA_OWNED);
1588 	if (td->td_pri_class == class)
1589 		return;
1590 
1591 #ifdef SMP
1592 	/*
1593 	 * On SMP if we're on the RUNQ we must adjust the transferable
1594 	 * count because could be changing to or from an interrupt
1595 	 * class.
1596 	 */
1597 	if (TD_ON_RUNQ(td)) {
1598 		struct tdq *tdq;
1599 
1600 		tdq = TDQ_CPU(td->td_sched->ts_cpu);
1601 		if (THREAD_CAN_MIGRATE(td)) {
1602 			tdq->tdq_transferable--;
1603 			tdq->tdq_group->tdg_transferable--;
1604 		}
1605 		td->td_pri_class = class;
1606 		if (THREAD_CAN_MIGRATE(td)) {
1607 			tdq->tdq_transferable++;
1608 			tdq->tdq_group->tdg_transferable++;
1609 		}
1610 	}
1611 #endif
1612 	td->td_pri_class = class;
1613 }
1614 
1615 /*
1616  * Return some of the child's priority and interactivity to the parent.
1617  */
1618 void
1619 sched_exit(struct proc *p, struct thread *child)
1620 {
1621 	struct thread *td;
1622 
1623 	CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d",
1624 	    child, child->td_proc->p_comm, child->td_priority);
1625 
1626 	td = FIRST_THREAD_IN_PROC(p);
1627 	sched_exit_thread(td, child);
1628 }
1629 
1630 void
1631 sched_exit_thread(struct thread *td, struct thread *child)
1632 {
1633 
1634 	CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d",
1635 	    child, child->td_proc->p_comm, child->td_priority);
1636 
1637 	tdq_load_rem(TDQ_CPU(child->td_sched->ts_cpu), child->td_sched);
1638 #ifdef KSE
1639 	/*
1640 	 * KSE forks and exits so often that this penalty causes short-lived
1641 	 * threads to always be non-interactive.  This causes mozilla to
1642 	 * crawl under load.
1643 	 */
1644 	if ((td->td_pflags & TDP_SA) && td->td_proc == child->td_proc)
1645 		return;
1646 #endif
1647 	/*
1648 	 * Give the child's runtime to the parent without returning the
1649 	 * sleep time as a penalty to the parent.  This causes shells that
1650 	 * launch expensive things to mark their children as expensive.
1651 	 */
1652 	td->td_sched->skg_runtime += child->td_sched->skg_runtime;
1653 	sched_interact_update(td);
1654 	sched_priority(td);
1655 }
1656 
1657 void
1658 sched_userret(struct thread *td)
1659 {
1660 	/*
1661 	 * XXX we cheat slightly on the locking here to avoid locking in
1662 	 * the usual case.  Setting td_priority here is essentially an
1663 	 * incomplete workaround for not setting it properly elsewhere.
1664 	 * Now that some interrupt handlers are threads, not setting it
1665 	 * properly elsewhere can clobber it in the window between setting
1666 	 * it here and returning to user mode, so don't waste time setting
1667 	 * it perfectly here.
1668 	 */
1669 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
1670 	    ("thread with borrowed priority returning to userland"));
1671 	if (td->td_priority != td->td_user_pri) {
1672 		mtx_lock_spin(&sched_lock);
1673 		td->td_priority = td->td_user_pri;
1674 		td->td_base_pri = td->td_user_pri;
1675 		mtx_unlock_spin(&sched_lock);
1676         }
1677 }
1678 
1679 void
1680 sched_clock(struct thread *td)
1681 {
1682 	struct tdq *tdq;
1683 	struct td_sched *ts;
1684 
1685 	mtx_assert(&sched_lock, MA_OWNED);
1686 #ifdef SMP
1687 	sched_smp_tick(td);
1688 #endif
1689 	tdq = TDQ_SELF();
1690 	/*
1691 	 * Advance the insert index once for each tick to ensure that all
1692 	 * threads get a chance to run.
1693 	 */
1694 	if (tdq->tdq_idx == tdq->tdq_ridx) {
1695 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
1696 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
1697 			tdq->tdq_ridx = tdq->tdq_idx;
1698 	}
1699 	ts = td->td_sched;
1700 	/*
1701 	 * We only do slicing code for TIMESHARE threads.
1702 	 */
1703 	if (td->td_pri_class != PRI_TIMESHARE)
1704 		return;
1705 	/*
1706 	 * We used a tick; charge it to the thread so that we can compute our
1707 	 * interactivity.
1708 	 */
1709 	td->td_sched->skg_runtime += tickincr;
1710 	sched_interact_update(td);
1711 	/*
1712 	 * We used up one time slice.
1713 	 */
1714 	if (--ts->ts_slice > 0)
1715 		return;
1716 	/*
1717 	 * We're out of time, recompute priorities and requeue.
1718 	 */
1719 	sched_priority(td);
1720 	td->td_flags |= TDF_NEEDRESCHED;
1721 }
1722 
1723 int
1724 sched_runnable(void)
1725 {
1726 	struct tdq *tdq;
1727 	int load;
1728 
1729 	load = 1;
1730 
1731 	tdq = TDQ_SELF();
1732 #ifdef SMP
1733 	if (tdq_busy)
1734 		goto out;
1735 #endif
1736 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
1737 		if (tdq->tdq_load > 0)
1738 			goto out;
1739 	} else
1740 		if (tdq->tdq_load - 1 > 0)
1741 			goto out;
1742 	load = 0;
1743 out:
1744 	return (load);
1745 }
1746 
1747 struct thread *
1748 sched_choose(void)
1749 {
1750 	struct tdq *tdq;
1751 	struct td_sched *ts;
1752 
1753 	mtx_assert(&sched_lock, MA_OWNED);
1754 	tdq = TDQ_SELF();
1755 #ifdef SMP
1756 restart:
1757 #endif
1758 	ts = tdq_choose(tdq);
1759 	if (ts) {
1760 #ifdef SMP
1761 		if (ts->ts_thread->td_priority > PRI_MIN_IDLE)
1762 			if (tdq_idled(tdq) == 0)
1763 				goto restart;
1764 #endif
1765 		tdq_runq_rem(tdq, ts);
1766 		return (ts->ts_thread);
1767 	}
1768 #ifdef SMP
1769 	if (tdq_idled(tdq) == 0)
1770 		goto restart;
1771 #endif
1772 	return (PCPU_GET(idlethread));
1773 }
1774 
1775 static int
1776 sched_preempt(struct thread *td)
1777 {
1778 	struct thread *ctd;
1779 	int cpri;
1780 	int pri;
1781 
1782 	ctd = curthread;
1783 	pri = td->td_priority;
1784 	cpri = ctd->td_priority;
1785 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
1786 		return (0);
1787 	/*
1788 	 * Always preempt IDLE threads.  Otherwise only if the preempting
1789 	 * thread is an ithread.
1790 	 */
1791 	if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
1792 		return (0);
1793 	if (ctd->td_critnest > 1) {
1794 		CTR1(KTR_PROC, "sched_preempt: in critical section %d",
1795 		    ctd->td_critnest);
1796 		ctd->td_owepreempt = 1;
1797 		return (0);
1798 	}
1799 	/*
1800 	 * Thread is runnable but not yet put on system run queue.
1801 	 */
1802 	MPASS(TD_ON_RUNQ(td));
1803 	TD_SET_RUNNING(td);
1804 	CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td,
1805 	    td->td_proc->p_pid, td->td_proc->p_comm);
1806 	mi_switch(SW_INVOL|SW_PREEMPT, td);
1807 	return (1);
1808 }
1809 
1810 void
1811 sched_add(struct thread *td, int flags)
1812 {
1813 	struct tdq *tdq;
1814 	struct td_sched *ts;
1815 	int preemptive;
1816 	int class;
1817 #ifdef SMP
1818 	int cpuid;
1819 	int cpumask;
1820 #endif
1821 	ts = td->td_sched;
1822 
1823 	mtx_assert(&sched_lock, MA_OWNED);
1824 	CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1825 	    td, td->td_proc->p_comm, td->td_priority, curthread,
1826 	    curthread->td_proc->p_comm);
1827 	KASSERT((td->td_inhibitors == 0),
1828 	    ("sched_add: trying to run inhibited thread"));
1829 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
1830 	    ("sched_add: bad thread state"));
1831 	KASSERT(td->td_proc->p_sflag & PS_INMEM,
1832 	    ("sched_add: process swapped out"));
1833 	KASSERT(ts->ts_runq == NULL,
1834 	    ("sched_add: thread %p is still assigned to a run queue", td));
1835         TD_SET_RUNQ(td);
1836 	tdq = TDQ_SELF();
1837 	class = PRI_BASE(td->td_pri_class);
1838 	preemptive = !(flags & SRQ_YIELDING);
1839 	/*
1840 	 * Recalculate the priority before we select the target cpu or
1841 	 * run-queue.
1842 	 */
1843 	if (class == PRI_TIMESHARE)
1844 		sched_priority(td);
1845 	if (ts->ts_slice == 0)
1846 		ts->ts_slice = sched_slice;
1847 #ifdef SMP
1848 	cpuid = PCPU_GET(cpuid);
1849 	/*
1850 	 * Pick the destination cpu and if it isn't ours transfer to the
1851 	 * target cpu.
1852 	 */
1853 	if (THREAD_CAN_MIGRATE(td)) {
1854 		if (td->td_priority <= PRI_MAX_ITHD) {
1855 			CTR2(KTR_ULE, "ithd %d < %d",
1856 			    td->td_priority, PRI_MAX_ITHD);
1857 			ts->ts_cpu = cpuid;
1858 		}
1859 		if (pick_pri)
1860 			ts->ts_cpu = tdq_pickpri(tdq, ts, flags);
1861 		else
1862 			ts->ts_cpu = tdq_pickidle(tdq, ts);
1863 	} else
1864 		CTR1(KTR_ULE, "pinned %d", td->td_pinned);
1865 	if (ts->ts_cpu != cpuid)
1866 		preemptive = 0;
1867 	tdq = TDQ_CPU(ts->ts_cpu);
1868 	cpumask = 1 << ts->ts_cpu;
1869 	/*
1870 	 * If we had been idle, clear our bit in the group and potentially
1871 	 * the global bitmap.
1872 	 */
1873 	if ((class != PRI_IDLE && class != PRI_ITHD) &&
1874 	    (tdq->tdq_group->tdg_idlemask & cpumask) != 0) {
1875 		/*
1876 		 * Check to see if our group is unidling, and if so, remove it
1877 		 * from the global idle mask.
1878 		 */
1879 		if (tdq->tdq_group->tdg_idlemask ==
1880 		    tdq->tdq_group->tdg_cpumask)
1881 			atomic_clear_int(&tdq_idle, tdq->tdq_group->tdg_mask);
1882 		/*
1883 		 * Now remove ourselves from the group specific idle mask.
1884 		 */
1885 		tdq->tdq_group->tdg_idlemask &= ~cpumask;
1886 	}
1887 #endif
1888 	/*
1889 	 * Pick the run queue based on priority.
1890 	 */
1891 	if (td->td_priority <= PRI_MAX_REALTIME)
1892 		ts->ts_runq = &tdq->tdq_realtime;
1893 	else if (td->td_priority <= PRI_MAX_TIMESHARE)
1894 		ts->ts_runq = &tdq->tdq_timeshare;
1895 	else
1896 		ts->ts_runq = &tdq->tdq_idle;
1897 	if (preemptive && sched_preempt(td))
1898 		return;
1899 	tdq_runq_add(tdq, ts, flags);
1900 	tdq_load_add(tdq, ts);
1901 #ifdef SMP
1902 	if (ts->ts_cpu != cpuid) {
1903 		tdq_notify(ts);
1904 		return;
1905 	}
1906 #endif
1907 	if (td->td_priority < curthread->td_priority)
1908 		curthread->td_flags |= TDF_NEEDRESCHED;
1909 }
1910 
1911 void
1912 sched_rem(struct thread *td)
1913 {
1914 	struct tdq *tdq;
1915 	struct td_sched *ts;
1916 
1917 	CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)",
1918 	    td, td->td_proc->p_comm, td->td_priority, curthread,
1919 	    curthread->td_proc->p_comm);
1920 	mtx_assert(&sched_lock, MA_OWNED);
1921 	ts = td->td_sched;
1922 	KASSERT(TD_ON_RUNQ(td),
1923 	    ("sched_rem: thread not on run queue"));
1924 
1925 	tdq = TDQ_CPU(ts->ts_cpu);
1926 	tdq_runq_rem(tdq, ts);
1927 	tdq_load_rem(tdq, ts);
1928 	TD_SET_CAN_RUN(td);
1929 }
1930 
1931 fixpt_t
1932 sched_pctcpu(struct thread *td)
1933 {
1934 	fixpt_t pctcpu;
1935 	struct td_sched *ts;
1936 
1937 	pctcpu = 0;
1938 	ts = td->td_sched;
1939 	if (ts == NULL)
1940 		return (0);
1941 
1942 	mtx_lock_spin(&sched_lock);
1943 	if (ts->ts_ticks) {
1944 		int rtick;
1945 
1946 		sched_pctcpu_update(ts);
1947 		/* How many rtick per second ? */
1948 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
1949 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
1950 	}
1951 	td->td_proc->p_swtime = ts->ts_ltick - ts->ts_ftick;
1952 	mtx_unlock_spin(&sched_lock);
1953 
1954 	return (pctcpu);
1955 }
1956 
1957 void
1958 sched_bind(struct thread *td, int cpu)
1959 {
1960 	struct td_sched *ts;
1961 
1962 	mtx_assert(&sched_lock, MA_OWNED);
1963 	ts = td->td_sched;
1964 	if (ts->ts_flags & TSF_BOUND)
1965 		sched_unbind(td);
1966 	ts->ts_flags |= TSF_BOUND;
1967 #ifdef SMP
1968 	sched_pin();
1969 	if (PCPU_GET(cpuid) == cpu)
1970 		return;
1971 	ts->ts_cpu = cpu;
1972 	/* When we return from mi_switch we'll be on the correct cpu. */
1973 	mi_switch(SW_VOL, NULL);
1974 #endif
1975 }
1976 
1977 void
1978 sched_unbind(struct thread *td)
1979 {
1980 	struct td_sched *ts;
1981 
1982 	mtx_assert(&sched_lock, MA_OWNED);
1983 	ts = td->td_sched;
1984 	if ((ts->ts_flags & TSF_BOUND) == 0)
1985 		return;
1986 	ts->ts_flags &= ~TSF_BOUND;
1987 #ifdef SMP
1988 	sched_unpin();
1989 #endif
1990 }
1991 
1992 int
1993 sched_is_bound(struct thread *td)
1994 {
1995 	mtx_assert(&sched_lock, MA_OWNED);
1996 	return (td->td_sched->ts_flags & TSF_BOUND);
1997 }
1998 
1999 void
2000 sched_relinquish(struct thread *td)
2001 {
2002 	mtx_lock_spin(&sched_lock);
2003 	if (td->td_pri_class == PRI_TIMESHARE)
2004 		sched_prio(td, PRI_MAX_TIMESHARE);
2005 	mi_switch(SW_VOL, NULL);
2006 	mtx_unlock_spin(&sched_lock);
2007 }
2008 
2009 int
2010 sched_load(void)
2011 {
2012 #ifdef SMP
2013 	int total;
2014 	int i;
2015 
2016 	total = 0;
2017 	for (i = 0; i <= tdg_maxid; i++)
2018 		total += TDQ_GROUP(i)->tdg_load;
2019 	return (total);
2020 #else
2021 	return (TDQ_SELF()->tdq_sysload);
2022 #endif
2023 }
2024 
2025 int
2026 sched_sizeof_proc(void)
2027 {
2028 	return (sizeof(struct proc));
2029 }
2030 
2031 int
2032 sched_sizeof_thread(void)
2033 {
2034 	return (sizeof(struct thread) + sizeof(struct td_sched));
2035 }
2036 
2037 void
2038 sched_tick(void)
2039 {
2040 	struct td_sched *ts;
2041 
2042 	ts = curthread->td_sched;
2043 	/* Adjust ticks for pctcpu */
2044 	ts->ts_ticks += 1 << SCHED_TICK_SHIFT;
2045 	ts->ts_ltick = ticks;
2046 	/*
2047 	 * Update if we've exceeded our desired tick threshhold by over one
2048 	 * second.
2049 	 */
2050 	if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick)
2051 		sched_pctcpu_update(ts);
2052 }
2053 
2054 /*
2055  * The actual idle process.
2056  */
2057 void
2058 sched_idletd(void *dummy)
2059 {
2060 	struct proc *p;
2061 	struct thread *td;
2062 
2063 	td = curthread;
2064 	p = td->td_proc;
2065 	mtx_assert(&Giant, MA_NOTOWNED);
2066 	/* ULE Relies on preemption for idle interruption. */
2067 	for (;;)
2068 		cpu_idle();
2069 }
2070 
2071 static SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2072 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ule", 0,
2073     "Scheduler name");
2074 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, "");
2075 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, "");
2076 SYSCTL_INT(_kern_sched, OID_AUTO, tickincr, CTLFLAG_RD, &tickincr, 0, "");
2077 SYSCTL_INT(_kern_sched, OID_AUTO, realstathz, CTLFLAG_RD, &realstathz, 0, "");
2078 #ifdef SMP
2079 SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri, CTLFLAG_RW, &pick_pri, 0, "");
2080 SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri_affinity, CTLFLAG_RW,
2081     &affinity, 0, "");
2082 SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri_tryself, CTLFLAG_RW,
2083     &tryself, 0, "");
2084 SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri_tryselfidle, CTLFLAG_RW,
2085     &tryselfidle, 0, "");
2086 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, "");
2087 SYSCTL_INT(_kern_sched, OID_AUTO, ipi_preempt, CTLFLAG_RW, &ipi_preempt, 0, "");
2088 SYSCTL_INT(_kern_sched, OID_AUTO, ipi_ast, CTLFLAG_RW, &ipi_ast, 0, "");
2089 SYSCTL_INT(_kern_sched, OID_AUTO, ipi_thresh, CTLFLAG_RW, &ipi_thresh, 0, "");
2090 SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, "");
2091 SYSCTL_INT(_kern_sched, OID_AUTO, steal_busy, CTLFLAG_RW, &steal_busy, 0, "");
2092 SYSCTL_INT(_kern_sched, OID_AUTO, busy_thresh, CTLFLAG_RW, &busy_thresh, 0, "");
2093 #endif
2094 
2095 /* ps compat */
2096 static fixpt_t  ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
2097 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2098 
2099 
2100 #define KERN_SWITCH_INCLUDE 1
2101 #include "kern/kern_switch.c"
2102