xref: /freebsd/sys/kern/sched_ule.c (revision 4da0d332f46ff20329bbac1952a56fa912f5f514)
1 /*-
2  * Copyright (c) 2002-2005, 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 #define kse td_sched
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kdb.h>
38 #include <sys/kernel.h>
39 #include <sys/ktr.h>
40 #include <sys/lock.h>
41 #include <sys/mutex.h>
42 #include <sys/proc.h>
43 #include <sys/resource.h>
44 #include <sys/resourcevar.h>
45 #include <sys/sched.h>
46 #include <sys/smp.h>
47 #include <sys/sx.h>
48 #include <sys/sysctl.h>
49 #include <sys/sysproto.h>
50 #include <sys/turnstile.h>
51 #include <sys/vmmeter.h>
52 #ifdef KTRACE
53 #include <sys/uio.h>
54 #include <sys/ktrace.h>
55 #endif
56 
57 #ifdef HWPMC_HOOKS
58 #include <sys/pmckern.h>
59 #endif
60 
61 #include <machine/cpu.h>
62 #include <machine/smp.h>
63 
64 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
65 /* XXX This is bogus compatability crap for ps */
66 static fixpt_t  ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
67 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
68 
69 static void sched_setup(void *dummy);
70 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL)
71 
72 static SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
73 
74 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ule", 0,
75     "Scheduler name");
76 
77 static int slice_min = 1;
78 SYSCTL_INT(_kern_sched, OID_AUTO, slice_min, CTLFLAG_RW, &slice_min, 0, "");
79 
80 static int slice_max = 10;
81 SYSCTL_INT(_kern_sched, OID_AUTO, slice_max, CTLFLAG_RW, &slice_max, 0, "");
82 
83 int realstathz;
84 int tickincr = 1;
85 
86 /*
87  * The following datastructures are allocated within their parent structure
88  * but are scheduler specific.
89  */
90 /*
91  * The schedulable entity that can be given a context to run.  A process may
92  * have several of these.
93  */
94 struct kse {
95 	TAILQ_ENTRY(kse) ke_procq;	/* (j/z) Run queue. */
96 	int		ke_flags;	/* (j) KEF_* flags. */
97 	struct thread	*ke_thread;	/* (*) Active associated thread. */
98 	fixpt_t		ke_pctcpu;	/* (j) %cpu during p_swtime. */
99 	char		ke_rqindex;	/* (j) Run queue index. */
100 	enum {
101 		KES_THREAD = 0x0,	/* slaved to thread state */
102 		KES_ONRUNQ
103 	} ke_state;			/* (j) thread sched specific status. */
104 	int		ke_slptime;
105 	int		ke_slice;
106 	struct runq	*ke_runq;
107 	u_char		ke_cpu;		/* CPU that we have affinity for. */
108 	/* The following variables are only used for pctcpu calculation */
109 	int		ke_ltick;	/* Last tick that we were running on */
110 	int		ke_ftick;	/* First tick that we were running on */
111 	int		ke_ticks;	/* Tick count */
112 
113 };
114 #define	td_kse			td_sched
115 #define	td_slptime		td_kse->ke_slptime
116 #define ke_proc			ke_thread->td_proc
117 #define ke_ksegrp		ke_thread->td_ksegrp
118 #define	ke_assign		ke_procq.tqe_next
119 /* flags kept in ke_flags */
120 #define	KEF_ASSIGNED	0x0001		/* Thread is being migrated. */
121 #define	KEF_BOUND	0x0002		/* Thread can not migrate. */
122 #define	KEF_XFERABLE	0x0004		/* Thread was added as transferable. */
123 #define	KEF_HOLD	0x0008		/* Thread is temporarily bound. */
124 #define	KEF_REMOVED	0x0010		/* Thread was removed while ASSIGNED */
125 #define	KEF_INTERNAL	0x0020		/* Thread added due to migration. */
126 #define	KEF_DIDRUN	0x02000		/* Thread actually ran. */
127 #define	KEF_EXIT	0x04000		/* Thread is being killed. */
128 
129 struct kg_sched {
130 	struct thread	*skg_last_assigned; /* (j) Last thread assigned to */
131 					   /* the system scheduler */
132 	int	skg_slptime;		/* Number of ticks we vol. slept */
133 	int	skg_runtime;		/* Number of ticks we were running */
134 	int	skg_avail_opennings;	/* (j) Num unfilled slots in group.*/
135 	int	skg_concurrency;	/* (j) Num threads requested in group.*/
136 };
137 #define kg_last_assigned	kg_sched->skg_last_assigned
138 #define kg_avail_opennings	kg_sched->skg_avail_opennings
139 #define kg_concurrency		kg_sched->skg_concurrency
140 #define kg_runtime		kg_sched->skg_runtime
141 #define kg_slptime		kg_sched->skg_slptime
142 
143 #define SLOT_RELEASE(kg)	(kg)->kg_avail_opennings++
144 #define	SLOT_USE(kg)		(kg)->kg_avail_opennings--
145 
146 static struct kse kse0;
147 static struct kg_sched kg_sched0;
148 
149 /*
150  * The priority is primarily determined by the interactivity score.  Thus, we
151  * give lower(better) priorities to kse groups that use less CPU.  The nice
152  * value is then directly added to this to allow nice to have some effect
153  * on latency.
154  *
155  * PRI_RANGE:	Total priority range for timeshare threads.
156  * PRI_NRESV:	Number of nice values.
157  * PRI_BASE:	The start of the dynamic range.
158  */
159 #define	SCHED_PRI_RANGE		(PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
160 #define	SCHED_PRI_NRESV		((PRIO_MAX - PRIO_MIN) + 1)
161 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
162 #define	SCHED_PRI_BASE		(PRI_MIN_TIMESHARE)
163 #define	SCHED_PRI_INTERACT(score)					\
164     ((score) * SCHED_PRI_RANGE / SCHED_INTERACT_MAX)
165 
166 /*
167  * These determine the interactivity of a process.
168  *
169  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
170  *		before throttling back.
171  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
172  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
173  * INTERACT_THRESH:	Threshhold for placement on the current runq.
174  */
175 #define	SCHED_SLP_RUN_MAX	((hz * 5) << 10)
176 #define	SCHED_SLP_RUN_FORK	((hz / 2) << 10)
177 #define	SCHED_INTERACT_MAX	(100)
178 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
179 #define	SCHED_INTERACT_THRESH	(30)
180 
181 /*
182  * These parameters and macros determine the size of the time slice that is
183  * granted to each thread.
184  *
185  * SLICE_MIN:	Minimum time slice granted, in units of ticks.
186  * SLICE_MAX:	Maximum time slice granted.
187  * SLICE_RANGE:	Range of available time slices scaled by hz.
188  * SLICE_SCALE:	The number slices granted per val in the range of [0, max].
189  * SLICE_NICE:  Determine the amount of slice granted to a scaled nice.
190  * SLICE_NTHRESH:	The nice cutoff point for slice assignment.
191  */
192 #define	SCHED_SLICE_MIN			(slice_min)
193 #define	SCHED_SLICE_MAX			(slice_max)
194 #define	SCHED_SLICE_INTERACTIVE		(slice_max)
195 #define	SCHED_SLICE_NTHRESH	(SCHED_PRI_NHALF - 1)
196 #define	SCHED_SLICE_RANGE		(SCHED_SLICE_MAX - SCHED_SLICE_MIN + 1)
197 #define	SCHED_SLICE_SCALE(val, max)	(((val) * SCHED_SLICE_RANGE) / (max))
198 #define	SCHED_SLICE_NICE(nice)						\
199     (SCHED_SLICE_MAX - SCHED_SLICE_SCALE((nice), SCHED_SLICE_NTHRESH))
200 
201 /*
202  * This macro determines whether or not the thread belongs on the current or
203  * next run queue.
204  */
205 #define	SCHED_INTERACTIVE(kg)						\
206     (sched_interact_score(kg) < SCHED_INTERACT_THRESH)
207 #define	SCHED_CURR(kg, ke)						\
208     ((ke->ke_thread->td_flags & TDF_BORROWING) || SCHED_INTERACTIVE(kg))
209 
210 /*
211  * Cpu percentage computation macros and defines.
212  *
213  * SCHED_CPU_TIME:	Number of seconds to average the cpu usage across.
214  * SCHED_CPU_TICKS:	Number of hz ticks to average the cpu usage across.
215  */
216 
217 #define	SCHED_CPU_TIME	10
218 #define	SCHED_CPU_TICKS	(hz * SCHED_CPU_TIME)
219 
220 /*
221  * kseq - per processor runqs and statistics.
222  */
223 struct kseq {
224 	struct runq	ksq_idle;		/* Queue of IDLE threads. */
225 	struct runq	ksq_timeshare[2];	/* Run queues for !IDLE. */
226 	struct runq	*ksq_next;		/* Next timeshare queue. */
227 	struct runq	*ksq_curr;		/* Current queue. */
228 	int		ksq_load_timeshare;	/* Load for timeshare. */
229 	int		ksq_load;		/* Aggregate load. */
230 	short		ksq_nice[SCHED_PRI_NRESV]; /* KSEs in each nice bin. */
231 	short		ksq_nicemin;		/* Least nice. */
232 #ifdef SMP
233 	int			ksq_transferable;
234 	LIST_ENTRY(kseq)	ksq_siblings;	/* Next in kseq group. */
235 	struct kseq_group	*ksq_group;	/* Our processor group. */
236 	volatile struct kse	*ksq_assigned;	/* assigned by another CPU. */
237 #else
238 	int		ksq_sysload;		/* For loadavg, !ITHD load. */
239 #endif
240 };
241 
242 #ifdef SMP
243 /*
244  * kseq groups are groups of processors which can cheaply share threads.  When
245  * one processor in the group goes idle it will check the runqs of the other
246  * processors in its group prior to halting and waiting for an interrupt.
247  * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA.
248  * In a numa environment we'd want an idle bitmap per group and a two tiered
249  * load balancer.
250  */
251 struct kseq_group {
252 	int	ksg_cpus;		/* Count of CPUs in this kseq group. */
253 	cpumask_t ksg_cpumask;		/* Mask of cpus in this group. */
254 	cpumask_t ksg_idlemask;		/* Idle cpus in this group. */
255 	cpumask_t ksg_mask;		/* Bit mask for first cpu. */
256 	int	ksg_load;		/* Total load of this group. */
257 	int	ksg_transferable;	/* Transferable load of this group. */
258 	LIST_HEAD(, kseq)	ksg_members; /* Linked list of all members. */
259 };
260 #endif
261 
262 /*
263  * One kse queue per processor.
264  */
265 #ifdef SMP
266 static cpumask_t kseq_idle;
267 static int ksg_maxid;
268 static struct kseq	kseq_cpu[MAXCPU];
269 static struct kseq_group kseq_groups[MAXCPU];
270 static int bal_tick;
271 static int gbal_tick;
272 static int balance_groups;
273 
274 #define	KSEQ_SELF()	(&kseq_cpu[PCPU_GET(cpuid)])
275 #define	KSEQ_CPU(x)	(&kseq_cpu[(x)])
276 #define	KSEQ_ID(x)	((x) - kseq_cpu)
277 #define	KSEQ_GROUP(x)	(&kseq_groups[(x)])
278 #else	/* !SMP */
279 static struct kseq	kseq_cpu;
280 
281 #define	KSEQ_SELF()	(&kseq_cpu)
282 #define	KSEQ_CPU(x)	(&kseq_cpu)
283 #endif
284 
285 static void slot_fill(struct ksegrp *);
286 static struct kse *sched_choose(void);		/* XXX Should be thread * */
287 static void sched_slice(struct kse *);
288 static void sched_priority(struct ksegrp *);
289 static void sched_thread_priority(struct thread *, u_char);
290 static int sched_interact_score(struct ksegrp *);
291 static void sched_interact_update(struct ksegrp *);
292 static void sched_interact_fork(struct ksegrp *);
293 static void sched_pctcpu_update(struct kse *);
294 
295 /* Operations on per processor queues */
296 static struct kse * kseq_choose(struct kseq *);
297 static void kseq_setup(struct kseq *);
298 static void kseq_load_add(struct kseq *, struct kse *);
299 static void kseq_load_rem(struct kseq *, struct kse *);
300 static __inline void kseq_runq_add(struct kseq *, struct kse *, int);
301 static __inline void kseq_runq_rem(struct kseq *, struct kse *);
302 static void kseq_nice_add(struct kseq *, int);
303 static void kseq_nice_rem(struct kseq *, int);
304 void kseq_print(int cpu);
305 #ifdef SMP
306 static int kseq_transfer(struct kseq *, struct kse *, int);
307 static struct kse *runq_steal(struct runq *);
308 static void sched_balance(void);
309 static void sched_balance_groups(void);
310 static void sched_balance_group(struct kseq_group *);
311 static void sched_balance_pair(struct kseq *, struct kseq *);
312 static void kseq_move(struct kseq *, int);
313 static int kseq_idled(struct kseq *);
314 static void kseq_notify(struct kse *, int);
315 static void kseq_assign(struct kseq *);
316 static struct kse *kseq_steal(struct kseq *, int);
317 #define	KSE_CAN_MIGRATE(ke)						\
318     ((ke)->ke_thread->td_pinned == 0 && ((ke)->ke_flags & KEF_BOUND) == 0)
319 #endif
320 
321 void
322 kseq_print(int cpu)
323 {
324 	struct kseq *kseq;
325 	int i;
326 
327 	kseq = KSEQ_CPU(cpu);
328 
329 	printf("kseq:\n");
330 	printf("\tload:           %d\n", kseq->ksq_load);
331 	printf("\tload TIMESHARE: %d\n", kseq->ksq_load_timeshare);
332 #ifdef SMP
333 	printf("\tload transferable: %d\n", kseq->ksq_transferable);
334 #endif
335 	printf("\tnicemin:\t%d\n", kseq->ksq_nicemin);
336 	printf("\tnice counts:\n");
337 	for (i = 0; i < SCHED_PRI_NRESV; i++)
338 		if (kseq->ksq_nice[i])
339 			printf("\t\t%d = %d\n",
340 			    i - SCHED_PRI_NHALF, kseq->ksq_nice[i]);
341 }
342 
343 static __inline void
344 kseq_runq_add(struct kseq *kseq, struct kse *ke, int flags)
345 {
346 #ifdef SMP
347 	if (KSE_CAN_MIGRATE(ke)) {
348 		kseq->ksq_transferable++;
349 		kseq->ksq_group->ksg_transferable++;
350 		ke->ke_flags |= KEF_XFERABLE;
351 	}
352 #endif
353 	runq_add(ke->ke_runq, ke, flags);
354 }
355 
356 static __inline void
357 kseq_runq_rem(struct kseq *kseq, struct kse *ke)
358 {
359 #ifdef SMP
360 	if (ke->ke_flags & KEF_XFERABLE) {
361 		kseq->ksq_transferable--;
362 		kseq->ksq_group->ksg_transferable--;
363 		ke->ke_flags &= ~KEF_XFERABLE;
364 	}
365 #endif
366 	runq_remove(ke->ke_runq, ke);
367 }
368 
369 static void
370 kseq_load_add(struct kseq *kseq, struct kse *ke)
371 {
372 	int class;
373 	mtx_assert(&sched_lock, MA_OWNED);
374 	class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
375 	if (class == PRI_TIMESHARE)
376 		kseq->ksq_load_timeshare++;
377 	kseq->ksq_load++;
378 	CTR1(KTR_SCHED, "load: %d", kseq->ksq_load);
379 	if (class != PRI_ITHD && (ke->ke_proc->p_flag & P_NOLOAD) == 0)
380 #ifdef SMP
381 		kseq->ksq_group->ksg_load++;
382 #else
383 		kseq->ksq_sysload++;
384 #endif
385 	if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE)
386 		kseq_nice_add(kseq, ke->ke_proc->p_nice);
387 }
388 
389 static void
390 kseq_load_rem(struct kseq *kseq, struct kse *ke)
391 {
392 	int class;
393 	mtx_assert(&sched_lock, MA_OWNED);
394 	class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
395 	if (class == PRI_TIMESHARE)
396 		kseq->ksq_load_timeshare--;
397 	if (class != PRI_ITHD  && (ke->ke_proc->p_flag & P_NOLOAD) == 0)
398 #ifdef SMP
399 		kseq->ksq_group->ksg_load--;
400 #else
401 		kseq->ksq_sysload--;
402 #endif
403 	kseq->ksq_load--;
404 	CTR1(KTR_SCHED, "load: %d", kseq->ksq_load);
405 	ke->ke_runq = NULL;
406 	if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE)
407 		kseq_nice_rem(kseq, ke->ke_proc->p_nice);
408 }
409 
410 static void
411 kseq_nice_add(struct kseq *kseq, int nice)
412 {
413 	mtx_assert(&sched_lock, MA_OWNED);
414 	/* Normalize to zero. */
415 	kseq->ksq_nice[nice + SCHED_PRI_NHALF]++;
416 	if (nice < kseq->ksq_nicemin || kseq->ksq_load_timeshare == 1)
417 		kseq->ksq_nicemin = nice;
418 }
419 
420 static void
421 kseq_nice_rem(struct kseq *kseq, int nice)
422 {
423 	int n;
424 
425 	mtx_assert(&sched_lock, MA_OWNED);
426 	/* Normalize to zero. */
427 	n = nice + SCHED_PRI_NHALF;
428 	kseq->ksq_nice[n]--;
429 	KASSERT(kseq->ksq_nice[n] >= 0, ("Negative nice count."));
430 
431 	/*
432 	 * If this wasn't the smallest nice value or there are more in
433 	 * this bucket we can just return.  Otherwise we have to recalculate
434 	 * the smallest nice.
435 	 */
436 	if (nice != kseq->ksq_nicemin ||
437 	    kseq->ksq_nice[n] != 0 ||
438 	    kseq->ksq_load_timeshare == 0)
439 		return;
440 
441 	for (; n < SCHED_PRI_NRESV; n++)
442 		if (kseq->ksq_nice[n]) {
443 			kseq->ksq_nicemin = n - SCHED_PRI_NHALF;
444 			return;
445 		}
446 }
447 
448 #ifdef SMP
449 /*
450  * sched_balance is a simple CPU load balancing algorithm.  It operates by
451  * finding the least loaded and most loaded cpu and equalizing their load
452  * by migrating some processes.
453  *
454  * Dealing only with two CPUs at a time has two advantages.  Firstly, most
455  * installations will only have 2 cpus.  Secondly, load balancing too much at
456  * once can have an unpleasant effect on the system.  The scheduler rarely has
457  * enough information to make perfect decisions.  So this algorithm chooses
458  * algorithm simplicity and more gradual effects on load in larger systems.
459  *
460  * It could be improved by considering the priorities and slices assigned to
461  * each task prior to balancing them.  There are many pathological cases with
462  * any approach and so the semi random algorithm below may work as well as any.
463  *
464  */
465 static void
466 sched_balance(void)
467 {
468 	struct kseq_group *high;
469 	struct kseq_group *low;
470 	struct kseq_group *ksg;
471 	int cnt;
472 	int i;
473 
474 	bal_tick = ticks + (random() % (hz * 2));
475 	if (smp_started == 0)
476 		return;
477 	low = high = NULL;
478 	i = random() % (ksg_maxid + 1);
479 	for (cnt = 0; cnt <= ksg_maxid; cnt++) {
480 		ksg = KSEQ_GROUP(i);
481 		/*
482 		 * Find the CPU with the highest load that has some
483 		 * threads to transfer.
484 		 */
485 		if ((high == NULL || ksg->ksg_load > high->ksg_load)
486 		    && ksg->ksg_transferable)
487 			high = ksg;
488 		if (low == NULL || ksg->ksg_load < low->ksg_load)
489 			low = ksg;
490 		if (++i > ksg_maxid)
491 			i = 0;
492 	}
493 	if (low != NULL && high != NULL && high != low)
494 		sched_balance_pair(LIST_FIRST(&high->ksg_members),
495 		    LIST_FIRST(&low->ksg_members));
496 }
497 
498 static void
499 sched_balance_groups(void)
500 {
501 	int i;
502 
503 	gbal_tick = ticks + (random() % (hz * 2));
504 	mtx_assert(&sched_lock, MA_OWNED);
505 	if (smp_started)
506 		for (i = 0; i <= ksg_maxid; i++)
507 			sched_balance_group(KSEQ_GROUP(i));
508 }
509 
510 static void
511 sched_balance_group(struct kseq_group *ksg)
512 {
513 	struct kseq *kseq;
514 	struct kseq *high;
515 	struct kseq *low;
516 	int load;
517 
518 	if (ksg->ksg_transferable == 0)
519 		return;
520 	low = NULL;
521 	high = NULL;
522 	LIST_FOREACH(kseq, &ksg->ksg_members, ksq_siblings) {
523 		load = kseq->ksq_load;
524 		if (high == NULL || load > high->ksq_load)
525 			high = kseq;
526 		if (low == NULL || load < low->ksq_load)
527 			low = kseq;
528 	}
529 	if (high != NULL && low != NULL && high != low)
530 		sched_balance_pair(high, low);
531 }
532 
533 static void
534 sched_balance_pair(struct kseq *high, struct kseq *low)
535 {
536 	int transferable;
537 	int high_load;
538 	int low_load;
539 	int move;
540 	int diff;
541 	int i;
542 
543 	/*
544 	 * If we're transfering within a group we have to use this specific
545 	 * kseq's transferable count, otherwise we can steal from other members
546 	 * of the group.
547 	 */
548 	if (high->ksq_group == low->ksq_group) {
549 		transferable = high->ksq_transferable;
550 		high_load = high->ksq_load;
551 		low_load = low->ksq_load;
552 	} else {
553 		transferable = high->ksq_group->ksg_transferable;
554 		high_load = high->ksq_group->ksg_load;
555 		low_load = low->ksq_group->ksg_load;
556 	}
557 	if (transferable == 0)
558 		return;
559 	/*
560 	 * Determine what the imbalance is and then adjust that to how many
561 	 * kses we actually have to give up (transferable).
562 	 */
563 	diff = high_load - low_load;
564 	move = diff / 2;
565 	if (diff & 0x1)
566 		move++;
567 	move = min(move, transferable);
568 	for (i = 0; i < move; i++)
569 		kseq_move(high, KSEQ_ID(low));
570 	return;
571 }
572 
573 static void
574 kseq_move(struct kseq *from, int cpu)
575 {
576 	struct kseq *kseq;
577 	struct kseq *to;
578 	struct kse *ke;
579 
580 	kseq = from;
581 	to = KSEQ_CPU(cpu);
582 	ke = kseq_steal(kseq, 1);
583 	if (ke == NULL) {
584 		struct kseq_group *ksg;
585 
586 		ksg = kseq->ksq_group;
587 		LIST_FOREACH(kseq, &ksg->ksg_members, ksq_siblings) {
588 			if (kseq == from || kseq->ksq_transferable == 0)
589 				continue;
590 			ke = kseq_steal(kseq, 1);
591 			break;
592 		}
593 		if (ke == NULL)
594 			panic("kseq_move: No KSEs available with a "
595 			    "transferable count of %d\n",
596 			    ksg->ksg_transferable);
597 	}
598 	if (kseq == to)
599 		return;
600 	ke->ke_state = KES_THREAD;
601 	kseq_runq_rem(kseq, ke);
602 	kseq_load_rem(kseq, ke);
603 	kseq_notify(ke, cpu);
604 }
605 
606 static int
607 kseq_idled(struct kseq *kseq)
608 {
609 	struct kseq_group *ksg;
610 	struct kseq *steal;
611 	struct kse *ke;
612 
613 	ksg = kseq->ksq_group;
614 	/*
615 	 * If we're in a cpu group, try and steal kses from another cpu in
616 	 * the group before idling.
617 	 */
618 	if (ksg->ksg_cpus > 1 && ksg->ksg_transferable) {
619 		LIST_FOREACH(steal, &ksg->ksg_members, ksq_siblings) {
620 			if (steal == kseq || steal->ksq_transferable == 0)
621 				continue;
622 			ke = kseq_steal(steal, 0);
623 			if (ke == NULL)
624 				continue;
625 			ke->ke_state = KES_THREAD;
626 			kseq_runq_rem(steal, ke);
627 			kseq_load_rem(steal, ke);
628 			ke->ke_cpu = PCPU_GET(cpuid);
629 			ke->ke_flags |= KEF_INTERNAL | KEF_HOLD;
630 			sched_add(ke->ke_thread, SRQ_YIELDING);
631 			return (0);
632 		}
633 	}
634 	/*
635 	 * We only set the idled bit when all of the cpus in the group are
636 	 * idle.  Otherwise we could get into a situation where a KSE bounces
637 	 * back and forth between two idle cores on seperate physical CPUs.
638 	 */
639 	ksg->ksg_idlemask |= PCPU_GET(cpumask);
640 	if (ksg->ksg_idlemask != ksg->ksg_cpumask)
641 		return (1);
642 	atomic_set_int(&kseq_idle, ksg->ksg_mask);
643 	return (1);
644 }
645 
646 static void
647 kseq_assign(struct kseq *kseq)
648 {
649 	struct kse *nke;
650 	struct kse *ke;
651 
652 	do {
653 		*(volatile struct kse **)&ke = kseq->ksq_assigned;
654 	} while(!atomic_cmpset_ptr(&kseq->ksq_assigned, ke, NULL));
655 	for (; ke != NULL; ke = nke) {
656 		nke = ke->ke_assign;
657 		kseq->ksq_group->ksg_load--;
658 		kseq->ksq_load--;
659 		ke->ke_flags &= ~KEF_ASSIGNED;
660 		ke->ke_flags |= KEF_INTERNAL | KEF_HOLD;
661 		sched_add(ke->ke_thread, SRQ_YIELDING);
662 	}
663 }
664 
665 static void
666 kseq_notify(struct kse *ke, int cpu)
667 {
668 	struct kseq *kseq;
669 	struct thread *td;
670 	struct pcpu *pcpu;
671 	int class;
672 	int prio;
673 
674 	kseq = KSEQ_CPU(cpu);
675 	/* XXX */
676 	class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
677 	if ((class == PRI_TIMESHARE || class == PRI_REALTIME) &&
678 	    (kseq_idle & kseq->ksq_group->ksg_mask))
679 		atomic_clear_int(&kseq_idle, kseq->ksq_group->ksg_mask);
680 	kseq->ksq_group->ksg_load++;
681 	kseq->ksq_load++;
682 	ke->ke_cpu = cpu;
683 	ke->ke_flags |= KEF_ASSIGNED;
684 	prio = ke->ke_thread->td_priority;
685 
686 	/*
687 	 * Place a KSE on another cpu's queue and force a resched.
688 	 */
689 	do {
690 		*(volatile struct kse **)&ke->ke_assign = kseq->ksq_assigned;
691 	} while(!atomic_cmpset_ptr(&kseq->ksq_assigned, ke->ke_assign, ke));
692 	/*
693 	 * Without sched_lock we could lose a race where we set NEEDRESCHED
694 	 * on a thread that is switched out before the IPI is delivered.  This
695 	 * would lead us to miss the resched.  This will be a problem once
696 	 * sched_lock is pushed down.
697 	 */
698 	pcpu = pcpu_find(cpu);
699 	td = pcpu->pc_curthread;
700 	if (ke->ke_thread->td_priority < td->td_priority ||
701 	    td == pcpu->pc_idlethread) {
702 		td->td_flags |= TDF_NEEDRESCHED;
703 		ipi_selected(1 << cpu, IPI_AST);
704 	}
705 }
706 
707 static struct kse *
708 runq_steal(struct runq *rq)
709 {
710 	struct rqhead *rqh;
711 	struct rqbits *rqb;
712 	struct kse *ke;
713 	int word;
714 	int bit;
715 
716 	mtx_assert(&sched_lock, MA_OWNED);
717 	rqb = &rq->rq_status;
718 	for (word = 0; word < RQB_LEN; word++) {
719 		if (rqb->rqb_bits[word] == 0)
720 			continue;
721 		for (bit = 0; bit < RQB_BPW; bit++) {
722 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
723 				continue;
724 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
725 			TAILQ_FOREACH(ke, rqh, ke_procq) {
726 				if (KSE_CAN_MIGRATE(ke))
727 					return (ke);
728 			}
729 		}
730 	}
731 	return (NULL);
732 }
733 
734 static struct kse *
735 kseq_steal(struct kseq *kseq, int stealidle)
736 {
737 	struct kse *ke;
738 
739 	/*
740 	 * Steal from next first to try to get a non-interactive task that
741 	 * may not have run for a while.
742 	 */
743 	if ((ke = runq_steal(kseq->ksq_next)) != NULL)
744 		return (ke);
745 	if ((ke = runq_steal(kseq->ksq_curr)) != NULL)
746 		return (ke);
747 	if (stealidle)
748 		return (runq_steal(&kseq->ksq_idle));
749 	return (NULL);
750 }
751 
752 int
753 kseq_transfer(struct kseq *kseq, struct kse *ke, int class)
754 {
755 	struct kseq_group *nksg;
756 	struct kseq_group *ksg;
757 	struct kseq *old;
758 	int cpu;
759 	int idx;
760 
761 	if (smp_started == 0)
762 		return (0);
763 	cpu = 0;
764 	/*
765 	 * If our load exceeds a certain threshold we should attempt to
766 	 * reassign this thread.  The first candidate is the cpu that
767 	 * originally ran the thread.  If it is idle, assign it there,
768 	 * otherwise, pick an idle cpu.
769 	 *
770 	 * The threshold at which we start to reassign kses has a large impact
771 	 * on the overall performance of the system.  Tuned too high and
772 	 * some CPUs may idle.  Too low and there will be excess migration
773 	 * and context switches.
774 	 */
775 	old = KSEQ_CPU(ke->ke_cpu);
776 	nksg = old->ksq_group;
777 	ksg = kseq->ksq_group;
778 	if (kseq_idle) {
779 		if (kseq_idle & nksg->ksg_mask) {
780 			cpu = ffs(nksg->ksg_idlemask);
781 			if (cpu) {
782 				CTR2(KTR_SCHED,
783 				    "kseq_transfer: %p found old cpu %X "
784 				    "in idlemask.", ke, cpu);
785 				goto migrate;
786 			}
787 		}
788 		/*
789 		 * Multiple cpus could find this bit simultaneously
790 		 * but the race shouldn't be terrible.
791 		 */
792 		cpu = ffs(kseq_idle);
793 		if (cpu) {
794 			CTR2(KTR_SCHED, "kseq_transfer: %p found %X "
795 			    "in idlemask.", ke, cpu);
796 			goto migrate;
797 		}
798 	}
799 	idx = 0;
800 #if 0
801 	if (old->ksq_load < kseq->ksq_load) {
802 		cpu = ke->ke_cpu + 1;
803 		CTR2(KTR_SCHED, "kseq_transfer: %p old cpu %X "
804 		    "load less than ours.", ke, cpu);
805 		goto migrate;
806 	}
807 	/*
808 	 * No new CPU was found, look for one with less load.
809 	 */
810 	for (idx = 0; idx <= ksg_maxid; idx++) {
811 		nksg = KSEQ_GROUP(idx);
812 		if (nksg->ksg_load /*+ (nksg->ksg_cpus  * 2)*/ < ksg->ksg_load) {
813 			cpu = ffs(nksg->ksg_cpumask);
814 			CTR2(KTR_SCHED, "kseq_transfer: %p cpu %X load less "
815 			    "than ours.", ke, cpu);
816 			goto migrate;
817 		}
818 	}
819 #endif
820 	/*
821 	 * If another cpu in this group has idled, assign a thread over
822 	 * to them after checking to see if there are idled groups.
823 	 */
824 	if (ksg->ksg_idlemask) {
825 		cpu = ffs(ksg->ksg_idlemask);
826 		if (cpu) {
827 			CTR2(KTR_SCHED, "kseq_transfer: %p cpu %X idle in "
828 			    "group.", ke, cpu);
829 			goto migrate;
830 		}
831 	}
832 	return (0);
833 migrate:
834 	/*
835 	 * Now that we've found an idle CPU, migrate the thread.
836 	 */
837 	cpu--;
838 	ke->ke_runq = NULL;
839 	kseq_notify(ke, cpu);
840 
841 	return (1);
842 }
843 
844 #endif	/* SMP */
845 
846 /*
847  * Pick the highest priority task we have and return it.
848  */
849 
850 static struct kse *
851 kseq_choose(struct kseq *kseq)
852 {
853 	struct runq *swap;
854 	struct kse *ke;
855 	int nice;
856 
857 	mtx_assert(&sched_lock, MA_OWNED);
858 	swap = NULL;
859 
860 	for (;;) {
861 		ke = runq_choose(kseq->ksq_curr);
862 		if (ke == NULL) {
863 			/*
864 			 * We already swapped once and didn't get anywhere.
865 			 */
866 			if (swap)
867 				break;
868 			swap = kseq->ksq_curr;
869 			kseq->ksq_curr = kseq->ksq_next;
870 			kseq->ksq_next = swap;
871 			continue;
872 		}
873 		/*
874 		 * If we encounter a slice of 0 the kse is in a
875 		 * TIMESHARE kse group and its nice was too far out
876 		 * of the range that receives slices.
877 		 */
878 		nice = ke->ke_proc->p_nice + (0 - kseq->ksq_nicemin);
879 		if (ke->ke_slice == 0 || (nice > SCHED_SLICE_NTHRESH &&
880 		    ke->ke_proc->p_nice != 0)) {
881 			runq_remove(ke->ke_runq, ke);
882 			sched_slice(ke);
883 			ke->ke_runq = kseq->ksq_next;
884 			runq_add(ke->ke_runq, ke, 0);
885 			continue;
886 		}
887 		return (ke);
888 	}
889 
890 	return (runq_choose(&kseq->ksq_idle));
891 }
892 
893 static void
894 kseq_setup(struct kseq *kseq)
895 {
896 	runq_init(&kseq->ksq_timeshare[0]);
897 	runq_init(&kseq->ksq_timeshare[1]);
898 	runq_init(&kseq->ksq_idle);
899 	kseq->ksq_curr = &kseq->ksq_timeshare[0];
900 	kseq->ksq_next = &kseq->ksq_timeshare[1];
901 	kseq->ksq_load = 0;
902 	kseq->ksq_load_timeshare = 0;
903 }
904 
905 static void
906 sched_setup(void *dummy)
907 {
908 #ifdef SMP
909 	int i;
910 #endif
911 
912 	slice_min = (hz/100);	/* 10ms */
913 	slice_max = (hz/7);	/* ~140ms */
914 
915 #ifdef SMP
916 	balance_groups = 0;
917 	/*
918 	 * Initialize the kseqs.
919 	 */
920 	for (i = 0; i < MAXCPU; i++) {
921 		struct kseq *ksq;
922 
923 		ksq = &kseq_cpu[i];
924 		ksq->ksq_assigned = NULL;
925 		kseq_setup(&kseq_cpu[i]);
926 	}
927 	if (smp_topology == NULL) {
928 		struct kseq_group *ksg;
929 		struct kseq *ksq;
930 		int cpus;
931 
932 		for (cpus = 0, i = 0; i < MAXCPU; i++) {
933 			if (CPU_ABSENT(i))
934 				continue;
935 			ksq = &kseq_cpu[cpus];
936 			ksg = &kseq_groups[cpus];
937 			/*
938 			 * Setup a kseq group with one member.
939 			 */
940 			ksq->ksq_transferable = 0;
941 			ksq->ksq_group = ksg;
942 			ksg->ksg_cpus = 1;
943 			ksg->ksg_idlemask = 0;
944 			ksg->ksg_cpumask = ksg->ksg_mask = 1 << i;
945 			ksg->ksg_load = 0;
946 			ksg->ksg_transferable = 0;
947 			LIST_INIT(&ksg->ksg_members);
948 			LIST_INSERT_HEAD(&ksg->ksg_members, ksq, ksq_siblings);
949 			cpus++;
950 		}
951 		ksg_maxid = cpus - 1;
952 	} else {
953 		struct kseq_group *ksg;
954 		struct cpu_group *cg;
955 		int j;
956 
957 		for (i = 0; i < smp_topology->ct_count; i++) {
958 			cg = &smp_topology->ct_group[i];
959 			ksg = &kseq_groups[i];
960 			/*
961 			 * Initialize the group.
962 			 */
963 			ksg->ksg_idlemask = 0;
964 			ksg->ksg_load = 0;
965 			ksg->ksg_transferable = 0;
966 			ksg->ksg_cpus = cg->cg_count;
967 			ksg->ksg_cpumask = cg->cg_mask;
968 			LIST_INIT(&ksg->ksg_members);
969 			/*
970 			 * Find all of the group members and add them.
971 			 */
972 			for (j = 0; j < MAXCPU; j++) {
973 				if ((cg->cg_mask & (1 << j)) != 0) {
974 					if (ksg->ksg_mask == 0)
975 						ksg->ksg_mask = 1 << j;
976 					kseq_cpu[j].ksq_transferable = 0;
977 					kseq_cpu[j].ksq_group = ksg;
978 					LIST_INSERT_HEAD(&ksg->ksg_members,
979 					    &kseq_cpu[j], ksq_siblings);
980 				}
981 			}
982 			if (ksg->ksg_cpus > 1)
983 				balance_groups = 1;
984 		}
985 		ksg_maxid = smp_topology->ct_count - 1;
986 	}
987 	/*
988 	 * Stagger the group and global load balancer so they do not
989 	 * interfere with each other.
990 	 */
991 	bal_tick = ticks + hz;
992 	if (balance_groups)
993 		gbal_tick = ticks + (hz / 2);
994 #else
995 	kseq_setup(KSEQ_SELF());
996 #endif
997 	mtx_lock_spin(&sched_lock);
998 	kseq_load_add(KSEQ_SELF(), &kse0);
999 	mtx_unlock_spin(&sched_lock);
1000 }
1001 
1002 /*
1003  * Scale the scheduling priority according to the "interactivity" of this
1004  * process.
1005  */
1006 static void
1007 sched_priority(struct ksegrp *kg)
1008 {
1009 	int pri;
1010 
1011 	if (kg->kg_pri_class != PRI_TIMESHARE)
1012 		return;
1013 
1014 	pri = SCHED_PRI_INTERACT(sched_interact_score(kg));
1015 	pri += SCHED_PRI_BASE;
1016 	pri += kg->kg_proc->p_nice;
1017 
1018 	if (pri > PRI_MAX_TIMESHARE)
1019 		pri = PRI_MAX_TIMESHARE;
1020 	else if (pri < PRI_MIN_TIMESHARE)
1021 		pri = PRI_MIN_TIMESHARE;
1022 
1023 	kg->kg_user_pri = pri;
1024 
1025 	return;
1026 }
1027 
1028 /*
1029  * Calculate a time slice based on the properties of the kseg and the runq
1030  * that we're on.  This is only for PRI_TIMESHARE ksegrps.
1031  */
1032 static void
1033 sched_slice(struct kse *ke)
1034 {
1035 	struct kseq *kseq;
1036 	struct ksegrp *kg;
1037 
1038 	kg = ke->ke_ksegrp;
1039 	kseq = KSEQ_CPU(ke->ke_cpu);
1040 
1041 	if (ke->ke_thread->td_flags & TDF_BORROWING) {
1042 		ke->ke_slice = SCHED_SLICE_MIN;
1043 		return;
1044 	}
1045 
1046 	/*
1047 	 * Rationale:
1048 	 * KSEs in interactive ksegs get a minimal slice so that we
1049 	 * quickly notice if it abuses its advantage.
1050 	 *
1051 	 * KSEs in non-interactive ksegs are assigned a slice that is
1052 	 * based on the ksegs nice value relative to the least nice kseg
1053 	 * on the run queue for this cpu.
1054 	 *
1055 	 * If the KSE is less nice than all others it gets the maximum
1056 	 * slice and other KSEs will adjust their slice relative to
1057 	 * this when they first expire.
1058 	 *
1059 	 * There is 20 point window that starts relative to the least
1060 	 * nice kse on the run queue.  Slice size is determined by
1061 	 * the kse distance from the last nice ksegrp.
1062 	 *
1063 	 * If the kse is outside of the window it will get no slice
1064 	 * and will be reevaluated each time it is selected on the
1065 	 * run queue.  The exception to this is nice 0 ksegs when
1066 	 * a nice -20 is running.  They are always granted a minimum
1067 	 * slice.
1068 	 */
1069 	if (!SCHED_INTERACTIVE(kg)) {
1070 		int nice;
1071 
1072 		nice = kg->kg_proc->p_nice + (0 - kseq->ksq_nicemin);
1073 		if (kseq->ksq_load_timeshare == 0 ||
1074 		    kg->kg_proc->p_nice < kseq->ksq_nicemin)
1075 			ke->ke_slice = SCHED_SLICE_MAX;
1076 		else if (nice <= SCHED_SLICE_NTHRESH)
1077 			ke->ke_slice = SCHED_SLICE_NICE(nice);
1078 		else if (kg->kg_proc->p_nice == 0)
1079 			ke->ke_slice = SCHED_SLICE_MIN;
1080 		else
1081 			ke->ke_slice = 0;
1082 	} else
1083 		ke->ke_slice = SCHED_SLICE_INTERACTIVE;
1084 
1085 	return;
1086 }
1087 
1088 /*
1089  * This routine enforces a maximum limit on the amount of scheduling history
1090  * kept.  It is called after either the slptime or runtime is adjusted.
1091  * This routine will not operate correctly when slp or run times have been
1092  * adjusted to more than double their maximum.
1093  */
1094 static void
1095 sched_interact_update(struct ksegrp *kg)
1096 {
1097 	int sum;
1098 
1099 	sum = kg->kg_runtime + kg->kg_slptime;
1100 	if (sum < SCHED_SLP_RUN_MAX)
1101 		return;
1102 	/*
1103 	 * If we have exceeded by more than 1/5th then the algorithm below
1104 	 * will not bring us back into range.  Dividing by two here forces
1105 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1106 	 */
1107 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1108 		kg->kg_runtime /= 2;
1109 		kg->kg_slptime /= 2;
1110 		return;
1111 	}
1112 	kg->kg_runtime = (kg->kg_runtime / 5) * 4;
1113 	kg->kg_slptime = (kg->kg_slptime / 5) * 4;
1114 }
1115 
1116 static void
1117 sched_interact_fork(struct ksegrp *kg)
1118 {
1119 	int ratio;
1120 	int sum;
1121 
1122 	sum = kg->kg_runtime + kg->kg_slptime;
1123 	if (sum > SCHED_SLP_RUN_FORK) {
1124 		ratio = sum / SCHED_SLP_RUN_FORK;
1125 		kg->kg_runtime /= ratio;
1126 		kg->kg_slptime /= ratio;
1127 	}
1128 }
1129 
1130 static int
1131 sched_interact_score(struct ksegrp *kg)
1132 {
1133 	int div;
1134 
1135 	if (kg->kg_runtime > kg->kg_slptime) {
1136 		div = max(1, kg->kg_runtime / SCHED_INTERACT_HALF);
1137 		return (SCHED_INTERACT_HALF +
1138 		    (SCHED_INTERACT_HALF - (kg->kg_slptime / div)));
1139 	} if (kg->kg_slptime > kg->kg_runtime) {
1140 		div = max(1, kg->kg_slptime / SCHED_INTERACT_HALF);
1141 		return (kg->kg_runtime / div);
1142 	}
1143 
1144 	/*
1145 	 * This can happen if slptime and runtime are 0.
1146 	 */
1147 	return (0);
1148 
1149 }
1150 
1151 /*
1152  * Very early in the boot some setup of scheduler-specific
1153  * parts of proc0 and of soem scheduler resources needs to be done.
1154  * Called from:
1155  *  proc0_init()
1156  */
1157 void
1158 schedinit(void)
1159 {
1160 	/*
1161 	 * Set up the scheduler specific parts of proc0.
1162 	 */
1163 	proc0.p_sched = NULL; /* XXX */
1164 	ksegrp0.kg_sched = &kg_sched0;
1165 	thread0.td_sched = &kse0;
1166 	kse0.ke_thread = &thread0;
1167 	kse0.ke_state = KES_THREAD;
1168 	kg_sched0.skg_concurrency = 1;
1169 	kg_sched0.skg_avail_opennings = 0; /* we are already running */
1170 }
1171 
1172 /*
1173  * This is only somewhat accurate since given many processes of the same
1174  * priority they will switch when their slices run out, which will be
1175  * at most SCHED_SLICE_MAX.
1176  */
1177 int
1178 sched_rr_interval(void)
1179 {
1180 	return (SCHED_SLICE_MAX);
1181 }
1182 
1183 static void
1184 sched_pctcpu_update(struct kse *ke)
1185 {
1186 	/*
1187 	 * Adjust counters and watermark for pctcpu calc.
1188 	 */
1189 	if (ke->ke_ltick > ticks - SCHED_CPU_TICKS) {
1190 		/*
1191 		 * Shift the tick count out so that the divide doesn't
1192 		 * round away our results.
1193 		 */
1194 		ke->ke_ticks <<= 10;
1195 		ke->ke_ticks = (ke->ke_ticks / (ticks - ke->ke_ftick)) *
1196 			    SCHED_CPU_TICKS;
1197 		ke->ke_ticks >>= 10;
1198 	} else
1199 		ke->ke_ticks = 0;
1200 	ke->ke_ltick = ticks;
1201 	ke->ke_ftick = ke->ke_ltick - SCHED_CPU_TICKS;
1202 }
1203 
1204 void
1205 sched_thread_priority(struct thread *td, u_char prio)
1206 {
1207 	struct kse *ke;
1208 
1209 	CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)",
1210 	    td, td->td_proc->p_comm, td->td_priority, prio, curthread,
1211 	    curthread->td_proc->p_comm);
1212 	ke = td->td_kse;
1213 	mtx_assert(&sched_lock, MA_OWNED);
1214 	if (td->td_priority == prio)
1215 		return;
1216 	if (TD_ON_RUNQ(td)) {
1217 		/*
1218 		 * If the priority has been elevated due to priority
1219 		 * propagation, we may have to move ourselves to a new
1220 		 * queue.  We still call adjustrunqueue below in case kse
1221 		 * needs to fix things up.
1222 		 */
1223 		if (prio < td->td_priority && ke->ke_runq != NULL &&
1224 		    (ke->ke_flags & KEF_ASSIGNED) == 0 &&
1225 		    ke->ke_runq != KSEQ_CPU(ke->ke_cpu)->ksq_curr) {
1226 			runq_remove(ke->ke_runq, ke);
1227 			ke->ke_runq = KSEQ_CPU(ke->ke_cpu)->ksq_curr;
1228 			runq_add(ke->ke_runq, ke, 0);
1229 		}
1230 		/*
1231 		 * Hold this kse on this cpu so that sched_prio() doesn't
1232 		 * cause excessive migration.  We only want migration to
1233 		 * happen as the result of a wakeup.
1234 		 */
1235 		ke->ke_flags |= KEF_HOLD;
1236 		adjustrunqueue(td, prio);
1237 		ke->ke_flags &= ~KEF_HOLD;
1238 	} else
1239 		td->td_priority = prio;
1240 }
1241 
1242 /*
1243  * Update a thread's priority when it is lent another thread's
1244  * priority.
1245  */
1246 void
1247 sched_lend_prio(struct thread *td, u_char prio)
1248 {
1249 
1250 	td->td_flags |= TDF_BORROWING;
1251 	sched_thread_priority(td, prio);
1252 }
1253 
1254 /*
1255  * Restore a thread's priority when priority propagation is
1256  * over.  The prio argument is the minimum priority the thread
1257  * needs to have to satisfy other possible priority lending
1258  * requests.  If the thread's regular priority is less
1259  * important than prio, the thread will keep a priority boost
1260  * of prio.
1261  */
1262 void
1263 sched_unlend_prio(struct thread *td, u_char prio)
1264 {
1265 	u_char base_pri;
1266 
1267 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1268 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1269 		base_pri = td->td_ksegrp->kg_user_pri;
1270 	else
1271 		base_pri = td->td_base_pri;
1272 	if (prio >= base_pri) {
1273 		td->td_flags &= ~TDF_BORROWING;
1274 		sched_thread_priority(td, base_pri);
1275 	} else
1276 		sched_lend_prio(td, prio);
1277 }
1278 
1279 void
1280 sched_prio(struct thread *td, u_char prio)
1281 {
1282 	u_char oldprio;
1283 
1284 	/* First, update the base priority. */
1285 	td->td_base_pri = prio;
1286 
1287 	/*
1288 	 * If the thread is borrowing another thread's priority, don't
1289 	 * ever lower the priority.
1290 	 */
1291 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1292 		return;
1293 
1294 	/* Change the real priority. */
1295 	oldprio = td->td_priority;
1296 	sched_thread_priority(td, prio);
1297 
1298 	/*
1299 	 * If the thread is on a turnstile, then let the turnstile update
1300 	 * its state.
1301 	 */
1302 	if (TD_ON_LOCK(td) && oldprio != prio)
1303 		turnstile_adjust(td, oldprio);
1304 }
1305 
1306 void
1307 sched_switch(struct thread *td, struct thread *newtd, int flags)
1308 {
1309 	struct kseq *ksq;
1310 	struct kse *ke;
1311 
1312 	mtx_assert(&sched_lock, MA_OWNED);
1313 
1314 	ke = td->td_kse;
1315 	ksq = KSEQ_SELF();
1316 
1317 	td->td_lastcpu = td->td_oncpu;
1318 	td->td_oncpu = NOCPU;
1319 	td->td_flags &= ~TDF_NEEDRESCHED;
1320 	td->td_owepreempt = 0;
1321 
1322 	/*
1323 	 * If the KSE has been assigned it may be in the process of switching
1324 	 * to the new cpu.  This is the case in sched_bind().
1325 	 */
1326 	if (td == PCPU_GET(idlethread)) {
1327 		TD_SET_CAN_RUN(td);
1328 	} else if ((ke->ke_flags & KEF_ASSIGNED) == 0) {
1329 		/* We are ending our run so make our slot available again */
1330 		SLOT_RELEASE(td->td_ksegrp);
1331 		kseq_load_rem(ksq, ke);
1332 		if (TD_IS_RUNNING(td)) {
1333 			/*
1334 			 * Don't allow the thread to migrate
1335 			 * from a preemption.
1336 			 */
1337 			ke->ke_flags |= KEF_HOLD;
1338 			setrunqueue(td, (flags & SW_PREEMPT) ?
1339 			    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1340 			    SRQ_OURSELF|SRQ_YIELDING);
1341 			ke->ke_flags &= ~KEF_HOLD;
1342 		} else if ((td->td_proc->p_flag & P_HADTHREADS) &&
1343 		    (newtd == NULL || newtd->td_ksegrp != td->td_ksegrp))
1344 			/*
1345 			 * We will not be on the run queue.
1346 			 * So we must be sleeping or similar.
1347 			 * Don't use the slot if we will need it
1348 			 * for newtd.
1349 			 */
1350 			slot_fill(td->td_ksegrp);
1351 	}
1352 	if (newtd != NULL) {
1353 		/*
1354 		 * If we bring in a thread account for it as if it had been
1355 		 * added to the run queue and then chosen.
1356 		 */
1357 		newtd->td_kse->ke_flags |= KEF_DIDRUN;
1358 		newtd->td_kse->ke_runq = ksq->ksq_curr;
1359 		TD_SET_RUNNING(newtd);
1360 		kseq_load_add(KSEQ_SELF(), newtd->td_kse);
1361 		/*
1362 		 * XXX When we preempt, we've already consumed a slot because
1363 		 * we got here through sched_add().  However, newtd can come
1364 		 * from thread_switchout() which can't SLOT_USE() because
1365 		 * the SLOT code is scheduler dependent.  We must use the
1366 		 * slot here otherwise.
1367 		 */
1368 		if ((flags & SW_PREEMPT) == 0)
1369 			SLOT_USE(newtd->td_ksegrp);
1370 	} else
1371 		newtd = choosethread();
1372 	if (td != newtd) {
1373 #ifdef	HWPMC_HOOKS
1374 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1375 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1376 #endif
1377 		cpu_switch(td, newtd);
1378 #ifdef	HWPMC_HOOKS
1379 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1380 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1381 #endif
1382 	}
1383 
1384 	sched_lock.mtx_lock = (uintptr_t)td;
1385 
1386 	td->td_oncpu = PCPU_GET(cpuid);
1387 }
1388 
1389 void
1390 sched_nice(struct proc *p, int nice)
1391 {
1392 	struct ksegrp *kg;
1393 	struct kse *ke;
1394 	struct thread *td;
1395 	struct kseq *kseq;
1396 
1397 	PROC_LOCK_ASSERT(p, MA_OWNED);
1398 	mtx_assert(&sched_lock, MA_OWNED);
1399 	/*
1400 	 * We need to adjust the nice counts for running KSEs.
1401 	 */
1402 	FOREACH_KSEGRP_IN_PROC(p, kg) {
1403 		if (kg->kg_pri_class == PRI_TIMESHARE) {
1404 			FOREACH_THREAD_IN_GROUP(kg, td) {
1405 				ke = td->td_kse;
1406 				if (ke->ke_runq == NULL)
1407 					continue;
1408 				kseq = KSEQ_CPU(ke->ke_cpu);
1409 				kseq_nice_rem(kseq, p->p_nice);
1410 				kseq_nice_add(kseq, nice);
1411 			}
1412 		}
1413 	}
1414 	p->p_nice = nice;
1415 	FOREACH_KSEGRP_IN_PROC(p, kg) {
1416 		sched_priority(kg);
1417 		FOREACH_THREAD_IN_GROUP(kg, td)
1418 			td->td_flags |= TDF_NEEDRESCHED;
1419 	}
1420 }
1421 
1422 void
1423 sched_sleep(struct thread *td)
1424 {
1425 	mtx_assert(&sched_lock, MA_OWNED);
1426 
1427 	td->td_slptime = ticks;
1428 }
1429 
1430 void
1431 sched_wakeup(struct thread *td)
1432 {
1433 	mtx_assert(&sched_lock, MA_OWNED);
1434 
1435 	/*
1436 	 * Let the kseg know how long we slept for.  This is because process
1437 	 * interactivity behavior is modeled in the kseg.
1438 	 */
1439 	if (td->td_slptime) {
1440 		struct ksegrp *kg;
1441 		int hzticks;
1442 
1443 		kg = td->td_ksegrp;
1444 		hzticks = (ticks - td->td_slptime) << 10;
1445 		if (hzticks >= SCHED_SLP_RUN_MAX) {
1446 			kg->kg_slptime = SCHED_SLP_RUN_MAX;
1447 			kg->kg_runtime = 1;
1448 		} else {
1449 			kg->kg_slptime += hzticks;
1450 			sched_interact_update(kg);
1451 		}
1452 		sched_priority(kg);
1453 		sched_slice(td->td_kse);
1454 		td->td_slptime = 0;
1455 	}
1456 	setrunqueue(td, SRQ_BORING);
1457 }
1458 
1459 /*
1460  * Penalize the parent for creating a new child and initialize the child's
1461  * priority.
1462  */
1463 void
1464 sched_fork(struct thread *td, struct thread *childtd)
1465 {
1466 
1467 	mtx_assert(&sched_lock, MA_OWNED);
1468 
1469 	sched_fork_ksegrp(td, childtd->td_ksegrp);
1470 	sched_fork_thread(td, childtd);
1471 }
1472 
1473 void
1474 sched_fork_ksegrp(struct thread *td, struct ksegrp *child)
1475 {
1476 	struct ksegrp *kg = td->td_ksegrp;
1477 	mtx_assert(&sched_lock, MA_OWNED);
1478 
1479 	child->kg_slptime = kg->kg_slptime;
1480 	child->kg_runtime = kg->kg_runtime;
1481 	child->kg_user_pri = kg->kg_user_pri;
1482 	sched_interact_fork(child);
1483 	kg->kg_runtime += tickincr << 10;
1484 	sched_interact_update(kg);
1485 }
1486 
1487 void
1488 sched_fork_thread(struct thread *td, struct thread *child)
1489 {
1490 	struct kse *ke;
1491 	struct kse *ke2;
1492 
1493 	sched_newthread(child);
1494 	ke = td->td_kse;
1495 	ke2 = child->td_kse;
1496 	ke2->ke_slice = 1;	/* Attempt to quickly learn interactivity. */
1497 	ke2->ke_cpu = ke->ke_cpu;
1498 	ke2->ke_runq = NULL;
1499 
1500 	/* Grab our parents cpu estimation information. */
1501 	ke2->ke_ticks = ke->ke_ticks;
1502 	ke2->ke_ltick = ke->ke_ltick;
1503 	ke2->ke_ftick = ke->ke_ftick;
1504 }
1505 
1506 void
1507 sched_class(struct ksegrp *kg, int class)
1508 {
1509 	struct kseq *kseq;
1510 	struct kse *ke;
1511 	struct thread *td;
1512 	int nclass;
1513 	int oclass;
1514 
1515 	mtx_assert(&sched_lock, MA_OWNED);
1516 	if (kg->kg_pri_class == class)
1517 		return;
1518 
1519 	nclass = PRI_BASE(class);
1520 	oclass = PRI_BASE(kg->kg_pri_class);
1521 	FOREACH_THREAD_IN_GROUP(kg, td) {
1522 		ke = td->td_kse;
1523 		if ((ke->ke_state != KES_ONRUNQ &&
1524 		    ke->ke_state != KES_THREAD) || ke->ke_runq == NULL)
1525 			continue;
1526 		kseq = KSEQ_CPU(ke->ke_cpu);
1527 
1528 #ifdef SMP
1529 		/*
1530 		 * On SMP if we're on the RUNQ we must adjust the transferable
1531 		 * count because could be changing to or from an interrupt
1532 		 * class.
1533 		 */
1534 		if (ke->ke_state == KES_ONRUNQ) {
1535 			if (KSE_CAN_MIGRATE(ke)) {
1536 				kseq->ksq_transferable--;
1537 				kseq->ksq_group->ksg_transferable--;
1538 			}
1539 			if (KSE_CAN_MIGRATE(ke)) {
1540 				kseq->ksq_transferable++;
1541 				kseq->ksq_group->ksg_transferable++;
1542 			}
1543 		}
1544 #endif
1545 		if (oclass == PRI_TIMESHARE) {
1546 			kseq->ksq_load_timeshare--;
1547 			kseq_nice_rem(kseq, kg->kg_proc->p_nice);
1548 		}
1549 		if (nclass == PRI_TIMESHARE) {
1550 			kseq->ksq_load_timeshare++;
1551 			kseq_nice_add(kseq, kg->kg_proc->p_nice);
1552 		}
1553 	}
1554 
1555 	kg->kg_pri_class = class;
1556 }
1557 
1558 /*
1559  * Return some of the child's priority and interactivity to the parent.
1560  */
1561 void
1562 sched_exit(struct proc *p, struct thread *childtd)
1563 {
1564 	mtx_assert(&sched_lock, MA_OWNED);
1565 	sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), childtd);
1566 	sched_exit_thread(NULL, childtd);
1567 }
1568 
1569 void
1570 sched_exit_ksegrp(struct ksegrp *kg, struct thread *td)
1571 {
1572 	/* kg->kg_slptime += td->td_ksegrp->kg_slptime; */
1573 	kg->kg_runtime += td->td_ksegrp->kg_runtime;
1574 	sched_interact_update(kg);
1575 }
1576 
1577 void
1578 sched_exit_thread(struct thread *td, struct thread *childtd)
1579 {
1580 	CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d",
1581 	    childtd, childtd->td_proc->p_comm, childtd->td_priority);
1582 	kseq_load_rem(KSEQ_CPU(childtd->td_kse->ke_cpu), childtd->td_kse);
1583 }
1584 
1585 void
1586 sched_clock(struct thread *td)
1587 {
1588 	struct kseq *kseq;
1589 	struct ksegrp *kg;
1590 	struct kse *ke;
1591 
1592 	mtx_assert(&sched_lock, MA_OWNED);
1593 	kseq = KSEQ_SELF();
1594 #ifdef SMP
1595 	if (ticks >= bal_tick)
1596 		sched_balance();
1597 	if (ticks >= gbal_tick && balance_groups)
1598 		sched_balance_groups();
1599 	/*
1600 	 * We could have been assigned a non real-time thread without an
1601 	 * IPI.
1602 	 */
1603 	if (kseq->ksq_assigned)
1604 		kseq_assign(kseq);	/* Potentially sets NEEDRESCHED */
1605 #endif
1606 	/*
1607 	 * sched_setup() apparently happens prior to stathz being set.  We
1608 	 * need to resolve the timers earlier in the boot so we can avoid
1609 	 * calculating this here.
1610 	 */
1611 	if (realstathz == 0) {
1612 		realstathz = stathz ? stathz : hz;
1613 		tickincr = hz / realstathz;
1614 		/*
1615 		 * XXX This does not work for values of stathz that are much
1616 		 * larger than hz.
1617 		 */
1618 		if (tickincr == 0)
1619 			tickincr = 1;
1620 	}
1621 
1622 	ke = td->td_kse;
1623 	kg = ke->ke_ksegrp;
1624 
1625 	/* Adjust ticks for pctcpu */
1626 	ke->ke_ticks++;
1627 	ke->ke_ltick = ticks;
1628 
1629 	/* Go up to one second beyond our max and then trim back down */
1630 	if (ke->ke_ftick + SCHED_CPU_TICKS + hz < ke->ke_ltick)
1631 		sched_pctcpu_update(ke);
1632 
1633 	if (td->td_flags & TDF_IDLETD)
1634 		return;
1635 	/*
1636 	 * We only do slicing code for TIMESHARE ksegrps.
1637 	 */
1638 	if (kg->kg_pri_class != PRI_TIMESHARE)
1639 		return;
1640 	/*
1641 	 * We used a tick charge it to the ksegrp so that we can compute our
1642 	 * interactivity.
1643 	 */
1644 	kg->kg_runtime += tickincr << 10;
1645 	sched_interact_update(kg);
1646 
1647 	/*
1648 	 * We used up one time slice.
1649 	 */
1650 	if (--ke->ke_slice > 0)
1651 		return;
1652 	/*
1653 	 * We're out of time, recompute priorities and requeue.
1654 	 */
1655 	kseq_load_rem(kseq, ke);
1656 	sched_priority(kg);
1657 	sched_slice(ke);
1658 	if (SCHED_CURR(kg, ke))
1659 		ke->ke_runq = kseq->ksq_curr;
1660 	else
1661 		ke->ke_runq = kseq->ksq_next;
1662 	kseq_load_add(kseq, ke);
1663 	td->td_flags |= TDF_NEEDRESCHED;
1664 }
1665 
1666 int
1667 sched_runnable(void)
1668 {
1669 	struct kseq *kseq;
1670 	int load;
1671 
1672 	load = 1;
1673 
1674 	kseq = KSEQ_SELF();
1675 #ifdef SMP
1676 	if (kseq->ksq_assigned) {
1677 		mtx_lock_spin(&sched_lock);
1678 		kseq_assign(kseq);
1679 		mtx_unlock_spin(&sched_lock);
1680 	}
1681 #endif
1682 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
1683 		if (kseq->ksq_load > 0)
1684 			goto out;
1685 	} else
1686 		if (kseq->ksq_load - 1 > 0)
1687 			goto out;
1688 	load = 0;
1689 out:
1690 	return (load);
1691 }
1692 
1693 void
1694 sched_userret(struct thread *td)
1695 {
1696 	struct ksegrp *kg;
1697 
1698 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
1699 	    ("thread with borrowed priority returning to userland"));
1700 	kg = td->td_ksegrp;
1701 	if (td->td_priority != kg->kg_user_pri) {
1702 		mtx_lock_spin(&sched_lock);
1703 		td->td_priority = kg->kg_user_pri;
1704 		td->td_base_pri = kg->kg_user_pri;
1705 		mtx_unlock_spin(&sched_lock);
1706 	}
1707 }
1708 
1709 struct kse *
1710 sched_choose(void)
1711 {
1712 	struct kseq *kseq;
1713 	struct kse *ke;
1714 
1715 	mtx_assert(&sched_lock, MA_OWNED);
1716 	kseq = KSEQ_SELF();
1717 #ifdef SMP
1718 restart:
1719 	if (kseq->ksq_assigned)
1720 		kseq_assign(kseq);
1721 #endif
1722 	ke = kseq_choose(kseq);
1723 	if (ke) {
1724 #ifdef SMP
1725 		if (ke->ke_ksegrp->kg_pri_class == PRI_IDLE)
1726 			if (kseq_idled(kseq) == 0)
1727 				goto restart;
1728 #endif
1729 		kseq_runq_rem(kseq, ke);
1730 		ke->ke_state = KES_THREAD;
1731 		return (ke);
1732 	}
1733 #ifdef SMP
1734 	if (kseq_idled(kseq) == 0)
1735 		goto restart;
1736 #endif
1737 	return (NULL);
1738 }
1739 
1740 void
1741 sched_add(struct thread *td, int flags)
1742 {
1743 	struct kseq *kseq;
1744 	struct ksegrp *kg;
1745 	struct kse *ke;
1746 	int preemptive;
1747 	int canmigrate;
1748 	int class;
1749 
1750 	CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1751 	    td, td->td_proc->p_comm, td->td_priority, curthread,
1752 	    curthread->td_proc->p_comm);
1753 	mtx_assert(&sched_lock, MA_OWNED);
1754 	ke = td->td_kse;
1755 	kg = td->td_ksegrp;
1756 	canmigrate = 1;
1757 	preemptive = !(flags & SRQ_YIELDING);
1758 	class = PRI_BASE(kg->kg_pri_class);
1759 	kseq = KSEQ_SELF();
1760 	if ((ke->ke_flags & KEF_INTERNAL) == 0)
1761 		SLOT_USE(td->td_ksegrp);
1762 	ke->ke_flags &= ~KEF_INTERNAL;
1763 #ifdef SMP
1764 	if (ke->ke_flags & KEF_ASSIGNED) {
1765 		if (ke->ke_flags & KEF_REMOVED)
1766 			ke->ke_flags &= ~KEF_REMOVED;
1767 		return;
1768 	}
1769 	canmigrate = KSE_CAN_MIGRATE(ke);
1770 #endif
1771 	KASSERT(ke->ke_state != KES_ONRUNQ,
1772 	    ("sched_add: kse %p (%s) already in run queue", ke,
1773 	    ke->ke_proc->p_comm));
1774 	KASSERT(ke->ke_proc->p_sflag & PS_INMEM,
1775 	    ("sched_add: process swapped out"));
1776 	KASSERT(ke->ke_runq == NULL,
1777 	    ("sched_add: KSE %p is still assigned to a run queue", ke));
1778 	switch (class) {
1779 	case PRI_ITHD:
1780 	case PRI_REALTIME:
1781 		ke->ke_runq = kseq->ksq_curr;
1782 		ke->ke_slice = SCHED_SLICE_MAX;
1783 		if (canmigrate)
1784 			ke->ke_cpu = PCPU_GET(cpuid);
1785 		break;
1786 	case PRI_TIMESHARE:
1787 		if (SCHED_CURR(kg, ke))
1788 			ke->ke_runq = kseq->ksq_curr;
1789 		else
1790 			ke->ke_runq = kseq->ksq_next;
1791 		break;
1792 	case PRI_IDLE:
1793 		/*
1794 		 * This is for priority prop.
1795 		 */
1796 		if (ke->ke_thread->td_priority < PRI_MIN_IDLE)
1797 			ke->ke_runq = kseq->ksq_curr;
1798 		else
1799 			ke->ke_runq = &kseq->ksq_idle;
1800 		ke->ke_slice = SCHED_SLICE_MIN;
1801 		break;
1802 	default:
1803 		panic("Unknown pri class.");
1804 		break;
1805 	}
1806 #ifdef SMP
1807 	/*
1808 	 * Don't migrate running threads here.  Force the long term balancer
1809 	 * to do it.
1810 	 */
1811 	if (ke->ke_flags & KEF_HOLD) {
1812 		ke->ke_flags &= ~KEF_HOLD;
1813 		canmigrate = 0;
1814 	}
1815 	/*
1816 	 * If this thread is pinned or bound, notify the target cpu.
1817 	 */
1818 	if (!canmigrate && ke->ke_cpu != PCPU_GET(cpuid) ) {
1819 		ke->ke_runq = NULL;
1820 		kseq_notify(ke, ke->ke_cpu);
1821 		return;
1822 	}
1823 	/*
1824 	 * If we had been idle, clear our bit in the group and potentially
1825 	 * the global bitmap.  If not, see if we should transfer this thread.
1826 	 */
1827 	if ((class == PRI_TIMESHARE || class == PRI_REALTIME) &&
1828 	    (kseq->ksq_group->ksg_idlemask & PCPU_GET(cpumask)) != 0) {
1829 		/*
1830 		 * Check to see if our group is unidling, and if so, remove it
1831 		 * from the global idle mask.
1832 		 */
1833 		if (kseq->ksq_group->ksg_idlemask ==
1834 		    kseq->ksq_group->ksg_cpumask)
1835 			atomic_clear_int(&kseq_idle, kseq->ksq_group->ksg_mask);
1836 		/*
1837 		 * Now remove ourselves from the group specific idle mask.
1838 		 */
1839 		kseq->ksq_group->ksg_idlemask &= ~PCPU_GET(cpumask);
1840 	} else if (canmigrate && kseq->ksq_load > 1 && class != PRI_ITHD)
1841 		if (kseq_transfer(kseq, ke, class))
1842 			return;
1843 	ke->ke_cpu = PCPU_GET(cpuid);
1844 #endif
1845 	if (td->td_priority < curthread->td_priority &&
1846 	    ke->ke_runq == kseq->ksq_curr)
1847 		curthread->td_flags |= TDF_NEEDRESCHED;
1848 	if (preemptive && maybe_preempt(td))
1849 		return;
1850 	ke->ke_state = KES_ONRUNQ;
1851 
1852 	kseq_runq_add(kseq, ke, flags);
1853 	kseq_load_add(kseq, ke);
1854 }
1855 
1856 void
1857 sched_rem(struct thread *td)
1858 {
1859 	struct kseq *kseq;
1860 	struct kse *ke;
1861 
1862 	CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)",
1863 	    td, td->td_proc->p_comm, td->td_priority, curthread,
1864 	    curthread->td_proc->p_comm);
1865 	mtx_assert(&sched_lock, MA_OWNED);
1866 	ke = td->td_kse;
1867 	SLOT_RELEASE(td->td_ksegrp);
1868 	if (ke->ke_flags & KEF_ASSIGNED) {
1869 		ke->ke_flags |= KEF_REMOVED;
1870 		return;
1871 	}
1872 	KASSERT((ke->ke_state == KES_ONRUNQ),
1873 	    ("sched_rem: KSE not on run queue"));
1874 
1875 	ke->ke_state = KES_THREAD;
1876 	kseq = KSEQ_CPU(ke->ke_cpu);
1877 	kseq_runq_rem(kseq, ke);
1878 	kseq_load_rem(kseq, ke);
1879 }
1880 
1881 fixpt_t
1882 sched_pctcpu(struct thread *td)
1883 {
1884 	fixpt_t pctcpu;
1885 	struct kse *ke;
1886 
1887 	pctcpu = 0;
1888 	ke = td->td_kse;
1889 	if (ke == NULL)
1890 		return (0);
1891 
1892 	mtx_lock_spin(&sched_lock);
1893 	if (ke->ke_ticks) {
1894 		int rtick;
1895 
1896 		/*
1897 		 * Don't update more frequently than twice a second.  Allowing
1898 		 * this causes the cpu usage to decay away too quickly due to
1899 		 * rounding errors.
1900 		 */
1901 		if (ke->ke_ftick + SCHED_CPU_TICKS < ke->ke_ltick ||
1902 		    ke->ke_ltick < (ticks - (hz / 2)))
1903 			sched_pctcpu_update(ke);
1904 		/* How many rtick per second ? */
1905 		rtick = min(ke->ke_ticks / SCHED_CPU_TIME, SCHED_CPU_TICKS);
1906 		pctcpu = (FSCALE * ((FSCALE * rtick)/realstathz)) >> FSHIFT;
1907 	}
1908 
1909 	ke->ke_proc->p_swtime = ke->ke_ltick - ke->ke_ftick;
1910 	mtx_unlock_spin(&sched_lock);
1911 
1912 	return (pctcpu);
1913 }
1914 
1915 void
1916 sched_bind(struct thread *td, int cpu)
1917 {
1918 	struct kse *ke;
1919 
1920 	mtx_assert(&sched_lock, MA_OWNED);
1921 	ke = td->td_kse;
1922 	ke->ke_flags |= KEF_BOUND;
1923 #ifdef SMP
1924 	if (PCPU_GET(cpuid) == cpu)
1925 		return;
1926 	/* sched_rem without the runq_remove */
1927 	ke->ke_state = KES_THREAD;
1928 	kseq_load_rem(KSEQ_CPU(ke->ke_cpu), ke);
1929 	kseq_notify(ke, cpu);
1930 	/* When we return from mi_switch we'll be on the correct cpu. */
1931 	mi_switch(SW_VOL, NULL);
1932 #endif
1933 }
1934 
1935 void
1936 sched_unbind(struct thread *td)
1937 {
1938 	mtx_assert(&sched_lock, MA_OWNED);
1939 	td->td_kse->ke_flags &= ~KEF_BOUND;
1940 }
1941 
1942 int
1943 sched_is_bound(struct thread *td)
1944 {
1945 	mtx_assert(&sched_lock, MA_OWNED);
1946 	return (td->td_kse->ke_flags & KEF_BOUND);
1947 }
1948 
1949 int
1950 sched_load(void)
1951 {
1952 #ifdef SMP
1953 	int total;
1954 	int i;
1955 
1956 	total = 0;
1957 	for (i = 0; i <= ksg_maxid; i++)
1958 		total += KSEQ_GROUP(i)->ksg_load;
1959 	return (total);
1960 #else
1961 	return (KSEQ_SELF()->ksq_sysload);
1962 #endif
1963 }
1964 
1965 int
1966 sched_sizeof_ksegrp(void)
1967 {
1968 	return (sizeof(struct ksegrp) + sizeof(struct kg_sched));
1969 }
1970 
1971 int
1972 sched_sizeof_proc(void)
1973 {
1974 	return (sizeof(struct proc));
1975 }
1976 
1977 int
1978 sched_sizeof_thread(void)
1979 {
1980 	return (sizeof(struct thread) + sizeof(struct td_sched));
1981 }
1982 #define KERN_SWITCH_INCLUDE 1
1983 #include "kern/kern_switch.c"
1984