xref: /freebsd/sys/kern/sched_ule.c (revision f0adf7f5cdd241db2f2c817683191a6ef64a4e95)
1 /*-
2  * Copyright (c) 2002-2003, 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 <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/kdb.h>
33 #include <sys/kernel.h>
34 #include <sys/ktr.h>
35 #include <sys/lock.h>
36 #include <sys/mutex.h>
37 #include <sys/proc.h>
38 #include <sys/resource.h>
39 #include <sys/resourcevar.h>
40 #include <sys/sched.h>
41 #include <sys/smp.h>
42 #include <sys/sx.h>
43 #include <sys/sysctl.h>
44 #include <sys/sysproto.h>
45 #include <sys/vmmeter.h>
46 #ifdef KTRACE
47 #include <sys/uio.h>
48 #include <sys/ktrace.h>
49 #endif
50 
51 #include <machine/cpu.h>
52 #include <machine/smp.h>
53 
54 #define KTR_ULE         KTR_NFS
55 
56 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
57 /* XXX This is bogus compatability crap for ps */
58 static fixpt_t  ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
59 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
60 
61 static void sched_setup(void *dummy);
62 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL)
63 
64 static SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "SCHED");
65 
66 #define ULE_NAME	"ule"
67 #define ULE_NAME_LEN	3
68 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, ULE_NAME, ULE_NAME_LEN,
69 	      "System is using the ULE scheduler");
70 
71 static int slice_min = 1;
72 SYSCTL_INT(_kern_sched, OID_AUTO, slice_min, CTLFLAG_RW, &slice_min, 0, "");
73 
74 static int slice_max = 10;
75 SYSCTL_INT(_kern_sched, OID_AUTO, slice_max, CTLFLAG_RW, &slice_max, 0, "");
76 
77 int realstathz;
78 int tickincr = 1;
79 
80 /*
81  * These datastructures are allocated within their parent datastructure but
82  * are scheduler specific.
83  */
84 
85 struct ke_sched {
86 	int		ske_slice;
87 	struct runq	*ske_runq;
88 	/* The following variables are only used for pctcpu calculation */
89 	int		ske_ltick;	/* Last tick that we were running on */
90 	int		ske_ftick;	/* First tick that we were running on */
91 	int		ske_ticks;	/* Tick count */
92 	/* CPU that we have affinity for. */
93 	u_char		ske_cpu;
94 };
95 #define	ke_slice	ke_sched->ske_slice
96 #define	ke_runq		ke_sched->ske_runq
97 #define	ke_ltick	ke_sched->ske_ltick
98 #define	ke_ftick	ke_sched->ske_ftick
99 #define	ke_ticks	ke_sched->ske_ticks
100 #define	ke_cpu		ke_sched->ske_cpu
101 #define	ke_assign	ke_procq.tqe_next
102 
103 #define	KEF_ASSIGNED	KEF_SCHED0	/* KSE is being migrated. */
104 #define	KEF_BOUND	KEF_SCHED1	/* KSE can not migrate. */
105 
106 struct kg_sched {
107 	int	skg_slptime;		/* Number of ticks we vol. slept */
108 	int	skg_runtime;		/* Number of ticks we were running */
109 };
110 #define	kg_slptime	kg_sched->skg_slptime
111 #define	kg_runtime	kg_sched->skg_runtime
112 
113 struct td_sched {
114 	int	std_slptime;
115 };
116 #define	td_slptime	td_sched->std_slptime
117 
118 struct td_sched td_sched;
119 struct ke_sched ke_sched;
120 struct kg_sched kg_sched;
121 
122 struct ke_sched *kse0_sched = &ke_sched;
123 struct kg_sched *ksegrp0_sched = &kg_sched;
124 struct p_sched *proc0_sched = NULL;
125 struct td_sched *thread0_sched = &td_sched;
126 
127 /*
128  * The priority is primarily determined by the interactivity score.  Thus, we
129  * give lower(better) priorities to kse groups that use less CPU.  The nice
130  * value is then directly added to this to allow nice to have some effect
131  * on latency.
132  *
133  * PRI_RANGE:	Total priority range for timeshare threads.
134  * PRI_NRESV:	Number of nice values.
135  * PRI_BASE:	The start of the dynamic range.
136  */
137 #define	SCHED_PRI_RANGE		(PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
138 #define	SCHED_PRI_NRESV		((PRIO_MAX - PRIO_MIN) + 1)
139 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
140 #define	SCHED_PRI_BASE		(PRI_MIN_TIMESHARE)
141 #define	SCHED_PRI_INTERACT(score)					\
142     ((score) * SCHED_PRI_RANGE / SCHED_INTERACT_MAX)
143 
144 /*
145  * These determine the interactivity of a process.
146  *
147  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
148  *		before throttling back.
149  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
150  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
151  * INTERACT_THRESH:	Threshhold for placement on the current runq.
152  */
153 #define	SCHED_SLP_RUN_MAX	((hz * 5) << 10)
154 #define	SCHED_SLP_RUN_FORK	((hz / 2) << 10)
155 #define	SCHED_INTERACT_MAX	(100)
156 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
157 #define	SCHED_INTERACT_THRESH	(30)
158 
159 /*
160  * These parameters and macros determine the size of the time slice that is
161  * granted to each thread.
162  *
163  * SLICE_MIN:	Minimum time slice granted, in units of ticks.
164  * SLICE_MAX:	Maximum time slice granted.
165  * SLICE_RANGE:	Range of available time slices scaled by hz.
166  * SLICE_SCALE:	The number slices granted per val in the range of [0, max].
167  * SLICE_NICE:  Determine the amount of slice granted to a scaled nice.
168  * SLICE_NTHRESH:	The nice cutoff point for slice assignment.
169  */
170 #define	SCHED_SLICE_MIN			(slice_min)
171 #define	SCHED_SLICE_MAX			(slice_max)
172 #define	SCHED_SLICE_INTERACTIVE		(slice_max)
173 #define	SCHED_SLICE_NTHRESH	(SCHED_PRI_NHALF - 1)
174 #define	SCHED_SLICE_RANGE		(SCHED_SLICE_MAX - SCHED_SLICE_MIN + 1)
175 #define	SCHED_SLICE_SCALE(val, max)	(((val) * SCHED_SLICE_RANGE) / (max))
176 #define	SCHED_SLICE_NICE(nice)						\
177     (SCHED_SLICE_MAX - SCHED_SLICE_SCALE((nice), SCHED_SLICE_NTHRESH))
178 
179 /*
180  * This macro determines whether or not the kse belongs on the current or
181  * next run queue.
182  */
183 #define	SCHED_INTERACTIVE(kg)						\
184     (sched_interact_score(kg) < SCHED_INTERACT_THRESH)
185 #define	SCHED_CURR(kg, ke)						\
186     (ke->ke_thread->td_priority < kg->kg_user_pri ||			\
187     SCHED_INTERACTIVE(kg))
188 
189 /*
190  * Cpu percentage computation macros and defines.
191  *
192  * SCHED_CPU_TIME:	Number of seconds to average the cpu usage across.
193  * SCHED_CPU_TICKS:	Number of hz ticks to average the cpu usage across.
194  */
195 
196 #define	SCHED_CPU_TIME	10
197 #define	SCHED_CPU_TICKS	(hz * SCHED_CPU_TIME)
198 
199 /*
200  * kseq - per processor runqs and statistics.
201  */
202 struct kseq {
203 	struct runq	ksq_idle;		/* Queue of IDLE threads. */
204 	struct runq	ksq_timeshare[2];	/* Run queues for !IDLE. */
205 	struct runq	*ksq_next;		/* Next timeshare queue. */
206 	struct runq	*ksq_curr;		/* Current queue. */
207 	int		ksq_load_timeshare;	/* Load for timeshare. */
208 	int		ksq_load;		/* Aggregate load. */
209 	short		ksq_nice[SCHED_PRI_NRESV]; /* KSEs in each nice bin. */
210 	short		ksq_nicemin;		/* Least nice. */
211 #ifdef SMP
212 	int			ksq_transferable;
213 	LIST_ENTRY(kseq)	ksq_siblings;	/* Next in kseq group. */
214 	struct kseq_group	*ksq_group;	/* Our processor group. */
215 	volatile struct kse	*ksq_assigned;	/* assigned by another CPU. */
216 #else
217 	int		ksq_sysload;		/* For loadavg, !ITHD load. */
218 #endif
219 };
220 
221 #ifdef SMP
222 /*
223  * kseq groups are groups of processors which can cheaply share threads.  When
224  * one processor in the group goes idle it will check the runqs of the other
225  * processors in its group prior to halting and waiting for an interrupt.
226  * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA.
227  * In a numa environment we'd want an idle bitmap per group and a two tiered
228  * load balancer.
229  */
230 struct kseq_group {
231 	int	ksg_cpus;		/* Count of CPUs in this kseq group. */
232 	cpumask_t ksg_cpumask;		/* Mask of cpus in this group. */
233 	cpumask_t ksg_idlemask;		/* Idle cpus in this group. */
234 	cpumask_t ksg_mask;		/* Bit mask for first cpu. */
235 	int	ksg_load;		/* Total load of this group. */
236 	int	ksg_transferable;	/* Transferable load of this group. */
237 	LIST_HEAD(, kseq)	ksg_members; /* Linked list of all members. */
238 };
239 #endif
240 
241 /*
242  * One kse queue per processor.
243  */
244 #ifdef SMP
245 static cpumask_t kseq_idle;
246 static int ksg_maxid;
247 static struct kseq	kseq_cpu[MAXCPU];
248 static struct kseq_group kseq_groups[MAXCPU];
249 static int bal_tick;
250 static int gbal_tick;
251 
252 #define	KSEQ_SELF()	(&kseq_cpu[PCPU_GET(cpuid)])
253 #define	KSEQ_CPU(x)	(&kseq_cpu[(x)])
254 #define	KSEQ_ID(x)	((x) - kseq_cpu)
255 #define	KSEQ_GROUP(x)	(&kseq_groups[(x)])
256 #else	/* !SMP */
257 static struct kseq	kseq_cpu;
258 
259 #define	KSEQ_SELF()	(&kseq_cpu)
260 #define	KSEQ_CPU(x)	(&kseq_cpu)
261 #endif
262 
263 static void sched_add_internal(struct thread *td, int preemptive);
264 static void sched_slice(struct kse *ke);
265 static void sched_priority(struct ksegrp *kg);
266 static int sched_interact_score(struct ksegrp *kg);
267 static void sched_interact_update(struct ksegrp *kg);
268 static void sched_interact_fork(struct ksegrp *kg);
269 static void sched_pctcpu_update(struct kse *ke);
270 
271 /* Operations on per processor queues */
272 static struct kse * kseq_choose(struct kseq *kseq);
273 static void kseq_setup(struct kseq *kseq);
274 static void kseq_load_add(struct kseq *kseq, struct kse *ke);
275 static void kseq_load_rem(struct kseq *kseq, struct kse *ke);
276 static __inline void kseq_runq_add(struct kseq *kseq, struct kse *ke);
277 static __inline void kseq_runq_rem(struct kseq *kseq, struct kse *ke);
278 static void kseq_nice_add(struct kseq *kseq, int nice);
279 static void kseq_nice_rem(struct kseq *kseq, int nice);
280 void kseq_print(int cpu);
281 #ifdef SMP
282 static int kseq_transfer(struct kseq *ksq, struct kse *ke, int class);
283 static struct kse *runq_steal(struct runq *rq);
284 static void sched_balance(void);
285 static void sched_balance_groups(void);
286 static void sched_balance_group(struct kseq_group *ksg);
287 static void sched_balance_pair(struct kseq *high, struct kseq *low);
288 static void kseq_move(struct kseq *from, int cpu);
289 static int kseq_idled(struct kseq *kseq);
290 static void kseq_notify(struct kse *ke, int cpu);
291 static void kseq_assign(struct kseq *);
292 static struct kse *kseq_steal(struct kseq *kseq, int stealidle);
293 /*
294  * On P4 Xeons the round-robin interrupt delivery is broken.  As a result of
295  * this, we can't pin interrupts to the cpu that they were delivered to,
296  * otherwise all ithreads only run on CPU 0.
297  */
298 #ifdef __i386__
299 #define	KSE_CAN_MIGRATE(ke, class)					\
300     ((ke)->ke_thread->td_pinned == 0 && ((ke)->ke_flags & KEF_BOUND) == 0)
301 #else /* !__i386__ */
302 #define	KSE_CAN_MIGRATE(ke, class)					\
303     ((class) != PRI_ITHD && (ke)->ke_thread->td_pinned == 0 &&		\
304     ((ke)->ke_flags & KEF_BOUND) == 0)
305 #endif /* !__i386__ */
306 #endif
307 
308 void
309 kseq_print(int cpu)
310 {
311 	struct kseq *kseq;
312 	int i;
313 
314 	kseq = KSEQ_CPU(cpu);
315 
316 	printf("kseq:\n");
317 	printf("\tload:           %d\n", kseq->ksq_load);
318 	printf("\tload TIMESHARE: %d\n", kseq->ksq_load_timeshare);
319 #ifdef SMP
320 	printf("\tload transferable: %d\n", kseq->ksq_transferable);
321 #endif
322 	printf("\tnicemin:\t%d\n", kseq->ksq_nicemin);
323 	printf("\tnice counts:\n");
324 	for (i = 0; i < SCHED_PRI_NRESV; i++)
325 		if (kseq->ksq_nice[i])
326 			printf("\t\t%d = %d\n",
327 			    i - SCHED_PRI_NHALF, kseq->ksq_nice[i]);
328 }
329 
330 static __inline void
331 kseq_runq_add(struct kseq *kseq, struct kse *ke)
332 {
333 #ifdef SMP
334 	if (KSE_CAN_MIGRATE(ke, PRI_BASE(ke->ke_ksegrp->kg_pri_class))) {
335 		kseq->ksq_transferable++;
336 		kseq->ksq_group->ksg_transferable++;
337 	}
338 #endif
339 	runq_add(ke->ke_runq, ke);
340 }
341 
342 static __inline void
343 kseq_runq_rem(struct kseq *kseq, struct kse *ke)
344 {
345 #ifdef SMP
346 	if (KSE_CAN_MIGRATE(ke, PRI_BASE(ke->ke_ksegrp->kg_pri_class))) {
347 		kseq->ksq_transferable--;
348 		kseq->ksq_group->ksg_transferable--;
349 	}
350 #endif
351 	runq_remove(ke->ke_runq, ke);
352 }
353 
354 static void
355 kseq_load_add(struct kseq *kseq, struct kse *ke)
356 {
357 	int class;
358 	mtx_assert(&sched_lock, MA_OWNED);
359 	class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
360 	if (class == PRI_TIMESHARE)
361 		kseq->ksq_load_timeshare++;
362 	kseq->ksq_load++;
363 	if (class != PRI_ITHD && (ke->ke_proc->p_flag & P_NOLOAD) == 0)
364 #ifdef SMP
365 		kseq->ksq_group->ksg_load++;
366 #else
367 		kseq->ksq_sysload++;
368 #endif
369 	if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE)
370 		CTR6(KTR_ULE,
371 		    "Add kse %p to %p (slice: %d, pri: %d, nice: %d(%d))",
372 		    ke, ke->ke_runq, ke->ke_slice, ke->ke_thread->td_priority,
373 		    ke->ke_proc->p_nice, kseq->ksq_nicemin);
374 	if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE)
375 		kseq_nice_add(kseq, ke->ke_proc->p_nice);
376 }
377 
378 static void
379 kseq_load_rem(struct kseq *kseq, struct kse *ke)
380 {
381 	int class;
382 	mtx_assert(&sched_lock, MA_OWNED);
383 	class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
384 	if (class == PRI_TIMESHARE)
385 		kseq->ksq_load_timeshare--;
386 	if (class != PRI_ITHD  && (ke->ke_proc->p_flag & P_NOLOAD) == 0)
387 #ifdef SMP
388 		kseq->ksq_group->ksg_load--;
389 #else
390 		kseq->ksq_sysload--;
391 #endif
392 	kseq->ksq_load--;
393 	ke->ke_runq = NULL;
394 	if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE)
395 		kseq_nice_rem(kseq, ke->ke_proc->p_nice);
396 }
397 
398 static void
399 kseq_nice_add(struct kseq *kseq, int nice)
400 {
401 	mtx_assert(&sched_lock, MA_OWNED);
402 	/* Normalize to zero. */
403 	kseq->ksq_nice[nice + SCHED_PRI_NHALF]++;
404 	if (nice < kseq->ksq_nicemin || kseq->ksq_load_timeshare == 1)
405 		kseq->ksq_nicemin = nice;
406 }
407 
408 static void
409 kseq_nice_rem(struct kseq *kseq, int nice)
410 {
411 	int n;
412 
413 	mtx_assert(&sched_lock, MA_OWNED);
414 	/* Normalize to zero. */
415 	n = nice + SCHED_PRI_NHALF;
416 	kseq->ksq_nice[n]--;
417 	KASSERT(kseq->ksq_nice[n] >= 0, ("Negative nice count."));
418 
419 	/*
420 	 * If this wasn't the smallest nice value or there are more in
421 	 * this bucket we can just return.  Otherwise we have to recalculate
422 	 * the smallest nice.
423 	 */
424 	if (nice != kseq->ksq_nicemin ||
425 	    kseq->ksq_nice[n] != 0 ||
426 	    kseq->ksq_load_timeshare == 0)
427 		return;
428 
429 	for (; n < SCHED_PRI_NRESV; n++)
430 		if (kseq->ksq_nice[n]) {
431 			kseq->ksq_nicemin = n - SCHED_PRI_NHALF;
432 			return;
433 		}
434 }
435 
436 #ifdef SMP
437 /*
438  * sched_balance is a simple CPU load balancing algorithm.  It operates by
439  * finding the least loaded and most loaded cpu and equalizing their load
440  * by migrating some processes.
441  *
442  * Dealing only with two CPUs at a time has two advantages.  Firstly, most
443  * installations will only have 2 cpus.  Secondly, load balancing too much at
444  * once can have an unpleasant effect on the system.  The scheduler rarely has
445  * enough information to make perfect decisions.  So this algorithm chooses
446  * algorithm simplicity and more gradual effects on load in larger systems.
447  *
448  * It could be improved by considering the priorities and slices assigned to
449  * each task prior to balancing them.  There are many pathological cases with
450  * any approach and so the semi random algorithm below may work as well as any.
451  *
452  */
453 static void
454 sched_balance(void)
455 {
456 	struct kseq_group *high;
457 	struct kseq_group *low;
458 	struct kseq_group *ksg;
459 	int cnt;
460 	int i;
461 
462 	if (smp_started == 0)
463 		goto out;
464 	low = high = NULL;
465 	i = random() % (ksg_maxid + 1);
466 	for (cnt = 0; cnt <= ksg_maxid; cnt++) {
467 		ksg = KSEQ_GROUP(i);
468 		/*
469 		 * Find the CPU with the highest load that has some
470 		 * threads to transfer.
471 		 */
472 		if ((high == NULL || ksg->ksg_load > high->ksg_load)
473 		    && ksg->ksg_transferable)
474 			high = ksg;
475 		if (low == NULL || ksg->ksg_load < low->ksg_load)
476 			low = ksg;
477 		if (++i > ksg_maxid)
478 			i = 0;
479 	}
480 	if (low != NULL && high != NULL && high != low)
481 		sched_balance_pair(LIST_FIRST(&high->ksg_members),
482 		    LIST_FIRST(&low->ksg_members));
483 out:
484 	bal_tick = ticks + (random() % (hz * 2));
485 }
486 
487 static void
488 sched_balance_groups(void)
489 {
490 	int i;
491 
492 	mtx_assert(&sched_lock, MA_OWNED);
493 	if (smp_started)
494 		for (i = 0; i <= ksg_maxid; i++)
495 			sched_balance_group(KSEQ_GROUP(i));
496 	gbal_tick = ticks + (random() % (hz * 2));
497 }
498 
499 static void
500 sched_balance_group(struct kseq_group *ksg)
501 {
502 	struct kseq *kseq;
503 	struct kseq *high;
504 	struct kseq *low;
505 	int load;
506 
507 	if (ksg->ksg_transferable == 0)
508 		return;
509 	low = NULL;
510 	high = NULL;
511 	LIST_FOREACH(kseq, &ksg->ksg_members, ksq_siblings) {
512 		load = kseq->ksq_load;
513 		if (high == NULL || load > high->ksq_load)
514 			high = kseq;
515 		if (low == NULL || load < low->ksq_load)
516 			low = kseq;
517 	}
518 	if (high != NULL && low != NULL && high != low)
519 		sched_balance_pair(high, low);
520 }
521 
522 static void
523 sched_balance_pair(struct kseq *high, struct kseq *low)
524 {
525 	int transferable;
526 	int high_load;
527 	int low_load;
528 	int move;
529 	int diff;
530 	int i;
531 
532 	/*
533 	 * If we're transfering within a group we have to use this specific
534 	 * kseq's transferable count, otherwise we can steal from other members
535 	 * of the group.
536 	 */
537 	if (high->ksq_group == low->ksq_group) {
538 		transferable = high->ksq_transferable;
539 		high_load = high->ksq_load;
540 		low_load = low->ksq_load;
541 	} else {
542 		transferable = high->ksq_group->ksg_transferable;
543 		high_load = high->ksq_group->ksg_load;
544 		low_load = low->ksq_group->ksg_load;
545 	}
546 	if (transferable == 0)
547 		return;
548 	/*
549 	 * Determine what the imbalance is and then adjust that to how many
550 	 * kses we actually have to give up (transferable).
551 	 */
552 	diff = high_load - low_load;
553 	move = diff / 2;
554 	if (diff & 0x1)
555 		move++;
556 	move = min(move, transferable);
557 	for (i = 0; i < move; i++)
558 		kseq_move(high, KSEQ_ID(low));
559 	return;
560 }
561 
562 static void
563 kseq_move(struct kseq *from, int cpu)
564 {
565 	struct kseq *kseq;
566 	struct kseq *to;
567 	struct kse *ke;
568 
569 	kseq = from;
570 	to = KSEQ_CPU(cpu);
571 	ke = kseq_steal(kseq, 1);
572 	if (ke == NULL) {
573 		struct kseq_group *ksg;
574 
575 		ksg = kseq->ksq_group;
576 		LIST_FOREACH(kseq, &ksg->ksg_members, ksq_siblings) {
577 			if (kseq == from || kseq->ksq_transferable == 0)
578 				continue;
579 			ke = kseq_steal(kseq, 1);
580 			break;
581 		}
582 		if (ke == NULL)
583 			panic("kseq_move: No KSEs available with a "
584 			    "transferable count of %d\n",
585 			    ksg->ksg_transferable);
586 	}
587 	if (kseq == to)
588 		return;
589 	ke->ke_state = KES_THREAD;
590 	kseq_runq_rem(kseq, ke);
591 	kseq_load_rem(kseq, ke);
592 	kseq_notify(ke, cpu);
593 }
594 
595 static int
596 kseq_idled(struct kseq *kseq)
597 {
598 	struct kseq_group *ksg;
599 	struct kseq *steal;
600 	struct kse *ke;
601 
602 	ksg = kseq->ksq_group;
603 	/*
604 	 * If we're in a cpu group, try and steal kses from another cpu in
605 	 * the group before idling.
606 	 */
607 	if (ksg->ksg_cpus > 1 && ksg->ksg_transferable) {
608 		LIST_FOREACH(steal, &ksg->ksg_members, ksq_siblings) {
609 			if (steal == kseq || steal->ksq_transferable == 0)
610 				continue;
611 			ke = kseq_steal(steal, 0);
612 			if (ke == NULL)
613 				continue;
614 			ke->ke_state = KES_THREAD;
615 			kseq_runq_rem(steal, ke);
616 			kseq_load_rem(steal, ke);
617 			ke->ke_cpu = PCPU_GET(cpuid);
618 			sched_add_internal(ke->ke_thread, 0);
619 			return (0);
620 		}
621 	}
622 	/*
623 	 * We only set the idled bit when all of the cpus in the group are
624 	 * idle.  Otherwise we could get into a situation where a KSE bounces
625 	 * back and forth between two idle cores on seperate physical CPUs.
626 	 */
627 	ksg->ksg_idlemask |= PCPU_GET(cpumask);
628 	if (ksg->ksg_idlemask != ksg->ksg_cpumask)
629 		return (1);
630 	atomic_set_int(&kseq_idle, ksg->ksg_mask);
631 	return (1);
632 }
633 
634 static void
635 kseq_assign(struct kseq *kseq)
636 {
637 	struct kse *nke;
638 	struct kse *ke;
639 
640 	do {
641 		(volatile struct kse *)ke = kseq->ksq_assigned;
642 	} while(!atomic_cmpset_ptr(&kseq->ksq_assigned, ke, NULL));
643 	for (; ke != NULL; ke = nke) {
644 		nke = ke->ke_assign;
645 		ke->ke_flags &= ~KEF_ASSIGNED;
646 		sched_add_internal(ke->ke_thread, 0);
647 	}
648 }
649 
650 static void
651 kseq_notify(struct kse *ke, int cpu)
652 {
653 	struct kseq *kseq;
654 	struct thread *td;
655 	struct pcpu *pcpu;
656 
657 	ke->ke_cpu = cpu;
658 	ke->ke_flags |= KEF_ASSIGNED;
659 
660 	kseq = KSEQ_CPU(cpu);
661 
662 	/*
663 	 * Place a KSE on another cpu's queue and force a resched.
664 	 */
665 	do {
666 		(volatile struct kse *)ke->ke_assign = kseq->ksq_assigned;
667 	} while(!atomic_cmpset_ptr(&kseq->ksq_assigned, ke->ke_assign, ke));
668 	pcpu = pcpu_find(cpu);
669 	td = pcpu->pc_curthread;
670 	if (ke->ke_thread->td_priority < td->td_priority ||
671 	    td == pcpu->pc_idlethread) {
672 		td->td_flags |= TDF_NEEDRESCHED;
673 		ipi_selected(1 << cpu, IPI_AST);
674 	}
675 }
676 
677 static struct kse *
678 runq_steal(struct runq *rq)
679 {
680 	struct rqhead *rqh;
681 	struct rqbits *rqb;
682 	struct kse *ke;
683 	int word;
684 	int bit;
685 
686 	mtx_assert(&sched_lock, MA_OWNED);
687 	rqb = &rq->rq_status;
688 	for (word = 0; word < RQB_LEN; word++) {
689 		if (rqb->rqb_bits[word] == 0)
690 			continue;
691 		for (bit = 0; bit < RQB_BPW; bit++) {
692 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
693 				continue;
694 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
695 			TAILQ_FOREACH(ke, rqh, ke_procq) {
696 				if (KSE_CAN_MIGRATE(ke,
697 				    PRI_BASE(ke->ke_ksegrp->kg_pri_class)))
698 					return (ke);
699 			}
700 		}
701 	}
702 	return (NULL);
703 }
704 
705 static struct kse *
706 kseq_steal(struct kseq *kseq, int stealidle)
707 {
708 	struct kse *ke;
709 
710 	/*
711 	 * Steal from next first to try to get a non-interactive task that
712 	 * may not have run for a while.
713 	 */
714 	if ((ke = runq_steal(kseq->ksq_next)) != NULL)
715 		return (ke);
716 	if ((ke = runq_steal(kseq->ksq_curr)) != NULL)
717 		return (ke);
718 	if (stealidle)
719 		return (runq_steal(&kseq->ksq_idle));
720 	return (NULL);
721 }
722 
723 int
724 kseq_transfer(struct kseq *kseq, struct kse *ke, int class)
725 {
726 	struct kseq_group *ksg;
727 	int cpu;
728 
729 	if (smp_started == 0)
730 		return (0);
731 	cpu = 0;
732 	ksg = kseq->ksq_group;
733 
734 	/*
735 	 * If there are any idle groups, give them our extra load.  The
736 	 * threshold at which we start to reassign kses has a large impact
737 	 * on the overall performance of the system.  Tuned too high and
738 	 * some CPUs may idle.  Too low and there will be excess migration
739 	 * and context switches.
740 	 */
741 	if (ksg->ksg_load > (ksg->ksg_cpus * 2) && kseq_idle) {
742 		/*
743 		 * Multiple cpus could find this bit simultaneously
744 		 * but the race shouldn't be terrible.
745 		 */
746 		cpu = ffs(kseq_idle);
747 		if (cpu)
748 			atomic_clear_int(&kseq_idle, 1 << (cpu - 1));
749 	}
750 	/*
751 	 * If another cpu in this group has idled, assign a thread over
752 	 * to them after checking to see if there are idled groups.
753 	 */
754 	if (cpu == 0 && kseq->ksq_load > 1 && ksg->ksg_idlemask) {
755 		cpu = ffs(ksg->ksg_idlemask);
756 		if (cpu)
757 			ksg->ksg_idlemask &= ~(1 << (cpu - 1));
758 	}
759 	/*
760 	 * Now that we've found an idle CPU, migrate the thread.
761 	 */
762 	if (cpu) {
763 		cpu--;
764 		ke->ke_runq = NULL;
765 		kseq_notify(ke, cpu);
766 		return (1);
767 	}
768 	return (0);
769 }
770 
771 #endif	/* SMP */
772 
773 /*
774  * Pick the highest priority task we have and return it.
775  */
776 
777 static struct kse *
778 kseq_choose(struct kseq *kseq)
779 {
780 	struct kse *ke;
781 	struct runq *swap;
782 
783 	mtx_assert(&sched_lock, MA_OWNED);
784 	swap = NULL;
785 
786 	for (;;) {
787 		ke = runq_choose(kseq->ksq_curr);
788 		if (ke == NULL) {
789 			/*
790 			 * We already swapped once and didn't get anywhere.
791 			 */
792 			if (swap)
793 				break;
794 			swap = kseq->ksq_curr;
795 			kseq->ksq_curr = kseq->ksq_next;
796 			kseq->ksq_next = swap;
797 			continue;
798 		}
799 		/*
800 		 * If we encounter a slice of 0 the kse is in a
801 		 * TIMESHARE kse group and its nice was too far out
802 		 * of the range that receives slices.
803 		 */
804 		if (ke->ke_slice == 0) {
805 			runq_remove(ke->ke_runq, ke);
806 			sched_slice(ke);
807 			ke->ke_runq = kseq->ksq_next;
808 			runq_add(ke->ke_runq, ke);
809 			continue;
810 		}
811 		return (ke);
812 	}
813 
814 	return (runq_choose(&kseq->ksq_idle));
815 }
816 
817 static void
818 kseq_setup(struct kseq *kseq)
819 {
820 	runq_init(&kseq->ksq_timeshare[0]);
821 	runq_init(&kseq->ksq_timeshare[1]);
822 	runq_init(&kseq->ksq_idle);
823 	kseq->ksq_curr = &kseq->ksq_timeshare[0];
824 	kseq->ksq_next = &kseq->ksq_timeshare[1];
825 	kseq->ksq_load = 0;
826 	kseq->ksq_load_timeshare = 0;
827 }
828 
829 static void
830 sched_setup(void *dummy)
831 {
832 #ifdef SMP
833 	int balance_groups;
834 	int i;
835 #endif
836 
837 	slice_min = (hz/100);	/* 10ms */
838 	slice_max = (hz/7);	/* ~140ms */
839 
840 #ifdef SMP
841 	balance_groups = 0;
842 	/*
843 	 * Initialize the kseqs.
844 	 */
845 	for (i = 0; i < MAXCPU; i++) {
846 		struct kseq *ksq;
847 
848 		ksq = &kseq_cpu[i];
849 		ksq->ksq_assigned = NULL;
850 		kseq_setup(&kseq_cpu[i]);
851 	}
852 	if (smp_topology == NULL) {
853 		struct kseq_group *ksg;
854 		struct kseq *ksq;
855 
856 		for (i = 0; i < MAXCPU; i++) {
857 			ksq = &kseq_cpu[i];
858 			ksg = &kseq_groups[i];
859 			/*
860 			 * Setup a kseq group with one member.
861 			 */
862 			ksq->ksq_transferable = 0;
863 			ksq->ksq_group = ksg;
864 			ksg->ksg_cpus = 1;
865 			ksg->ksg_idlemask = 0;
866 			ksg->ksg_cpumask = ksg->ksg_mask = 1 << i;
867 			ksg->ksg_load = 0;
868 			ksg->ksg_transferable = 0;
869 			LIST_INIT(&ksg->ksg_members);
870 			LIST_INSERT_HEAD(&ksg->ksg_members, ksq, ksq_siblings);
871 		}
872 	} else {
873 		struct kseq_group *ksg;
874 		struct cpu_group *cg;
875 		int j;
876 
877 		for (i = 0; i < smp_topology->ct_count; i++) {
878 			cg = &smp_topology->ct_group[i];
879 			ksg = &kseq_groups[i];
880 			/*
881 			 * Initialize the group.
882 			 */
883 			ksg->ksg_idlemask = 0;
884 			ksg->ksg_load = 0;
885 			ksg->ksg_transferable = 0;
886 			ksg->ksg_cpus = cg->cg_count;
887 			ksg->ksg_cpumask = cg->cg_mask;
888 			LIST_INIT(&ksg->ksg_members);
889 			/*
890 			 * Find all of the group members and add them.
891 			 */
892 			for (j = 0; j < MAXCPU; j++) {
893 				if ((cg->cg_mask & (1 << j)) != 0) {
894 					if (ksg->ksg_mask == 0)
895 						ksg->ksg_mask = 1 << j;
896 					kseq_cpu[j].ksq_transferable = 0;
897 					kseq_cpu[j].ksq_group = ksg;
898 					LIST_INSERT_HEAD(&ksg->ksg_members,
899 					    &kseq_cpu[j], ksq_siblings);
900 				}
901 			}
902 			if (ksg->ksg_cpus > 1)
903 				balance_groups = 1;
904 		}
905 		ksg_maxid = smp_topology->ct_count - 1;
906 	}
907 	/*
908 	 * Stagger the group and global load balancer so they do not
909 	 * interfere with each other.
910 	 */
911 	bal_tick = ticks + hz;
912 	if (balance_groups)
913 		gbal_tick = ticks + (hz / 2);
914 #else
915 	kseq_setup(KSEQ_SELF());
916 #endif
917 	mtx_lock_spin(&sched_lock);
918 	kseq_load_add(KSEQ_SELF(), &kse0);
919 	mtx_unlock_spin(&sched_lock);
920 }
921 
922 /*
923  * Scale the scheduling priority according to the "interactivity" of this
924  * process.
925  */
926 static void
927 sched_priority(struct ksegrp *kg)
928 {
929 	int pri;
930 
931 	if (kg->kg_pri_class != PRI_TIMESHARE)
932 		return;
933 
934 	pri = SCHED_PRI_INTERACT(sched_interact_score(kg));
935 	pri += SCHED_PRI_BASE;
936 	pri += kg->kg_proc->p_nice;
937 
938 	if (pri > PRI_MAX_TIMESHARE)
939 		pri = PRI_MAX_TIMESHARE;
940 	else if (pri < PRI_MIN_TIMESHARE)
941 		pri = PRI_MIN_TIMESHARE;
942 
943 	kg->kg_user_pri = pri;
944 
945 	return;
946 }
947 
948 /*
949  * Calculate a time slice based on the properties of the kseg and the runq
950  * that we're on.  This is only for PRI_TIMESHARE ksegrps.
951  */
952 static void
953 sched_slice(struct kse *ke)
954 {
955 	struct kseq *kseq;
956 	struct ksegrp *kg;
957 
958 	kg = ke->ke_ksegrp;
959 	kseq = KSEQ_CPU(ke->ke_cpu);
960 
961 	/*
962 	 * Rationale:
963 	 * KSEs in interactive ksegs get the minimum slice so that we
964 	 * quickly notice if it abuses its advantage.
965 	 *
966 	 * KSEs in non-interactive ksegs are assigned a slice that is
967 	 * based on the ksegs nice value relative to the least nice kseg
968 	 * on the run queue for this cpu.
969 	 *
970 	 * If the KSE is less nice than all others it gets the maximum
971 	 * slice and other KSEs will adjust their slice relative to
972 	 * this when they first expire.
973 	 *
974 	 * There is 20 point window that starts relative to the least
975 	 * nice kse on the run queue.  Slice size is determined by
976 	 * the kse distance from the last nice ksegrp.
977 	 *
978 	 * If the kse is outside of the window it will get no slice
979 	 * and will be reevaluated each time it is selected on the
980 	 * run queue.  The exception to this is nice 0 ksegs when
981 	 * a nice -20 is running.  They are always granted a minimum
982 	 * slice.
983 	 */
984 	if (!SCHED_INTERACTIVE(kg)) {
985 		int nice;
986 
987 		nice = kg->kg_proc->p_nice + (0 - kseq->ksq_nicemin);
988 		if (kseq->ksq_load_timeshare == 0 ||
989 		    kg->kg_proc->p_nice < kseq->ksq_nicemin)
990 			ke->ke_slice = SCHED_SLICE_MAX;
991 		else if (nice <= SCHED_SLICE_NTHRESH)
992 			ke->ke_slice = SCHED_SLICE_NICE(nice);
993 		else if (kg->kg_proc->p_nice == 0)
994 			ke->ke_slice = SCHED_SLICE_MIN;
995 		else
996 			ke->ke_slice = 0;
997 	} else
998 		ke->ke_slice = SCHED_SLICE_INTERACTIVE;
999 
1000 	CTR6(KTR_ULE,
1001 	    "Sliced %p(%d) (nice: %d, nicemin: %d, load: %d, interactive: %d)",
1002 	    ke, ke->ke_slice, kg->kg_proc->p_nice, kseq->ksq_nicemin,
1003 	    kseq->ksq_load_timeshare, SCHED_INTERACTIVE(kg));
1004 
1005 	return;
1006 }
1007 
1008 /*
1009  * This routine enforces a maximum limit on the amount of scheduling history
1010  * kept.  It is called after either the slptime or runtime is adjusted.
1011  * This routine will not operate correctly when slp or run times have been
1012  * adjusted to more than double their maximum.
1013  */
1014 static void
1015 sched_interact_update(struct ksegrp *kg)
1016 {
1017 	int sum;
1018 
1019 	sum = kg->kg_runtime + kg->kg_slptime;
1020 	if (sum < SCHED_SLP_RUN_MAX)
1021 		return;
1022 	/*
1023 	 * If we have exceeded by more than 1/5th then the algorithm below
1024 	 * will not bring us back into range.  Dividing by two here forces
1025 	 * us into the range of [3/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1026 	 */
1027 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1028 		kg->kg_runtime /= 2;
1029 		kg->kg_slptime /= 2;
1030 		return;
1031 	}
1032 	kg->kg_runtime = (kg->kg_runtime / 5) * 4;
1033 	kg->kg_slptime = (kg->kg_slptime / 5) * 4;
1034 }
1035 
1036 static void
1037 sched_interact_fork(struct ksegrp *kg)
1038 {
1039 	int ratio;
1040 	int sum;
1041 
1042 	sum = kg->kg_runtime + kg->kg_slptime;
1043 	if (sum > SCHED_SLP_RUN_FORK) {
1044 		ratio = sum / SCHED_SLP_RUN_FORK;
1045 		kg->kg_runtime /= ratio;
1046 		kg->kg_slptime /= ratio;
1047 	}
1048 }
1049 
1050 static int
1051 sched_interact_score(struct ksegrp *kg)
1052 {
1053 	int div;
1054 
1055 	if (kg->kg_runtime > kg->kg_slptime) {
1056 		div = max(1, kg->kg_runtime / SCHED_INTERACT_HALF);
1057 		return (SCHED_INTERACT_HALF +
1058 		    (SCHED_INTERACT_HALF - (kg->kg_slptime / div)));
1059 	} if (kg->kg_slptime > kg->kg_runtime) {
1060 		div = max(1, kg->kg_slptime / SCHED_INTERACT_HALF);
1061 		return (kg->kg_runtime / div);
1062 	}
1063 
1064 	/*
1065 	 * This can happen if slptime and runtime are 0.
1066 	 */
1067 	return (0);
1068 
1069 }
1070 
1071 /*
1072  * This is only somewhat accurate since given many processes of the same
1073  * priority they will switch when their slices run out, which will be
1074  * at most SCHED_SLICE_MAX.
1075  */
1076 int
1077 sched_rr_interval(void)
1078 {
1079 	return (SCHED_SLICE_MAX);
1080 }
1081 
1082 static void
1083 sched_pctcpu_update(struct kse *ke)
1084 {
1085 	/*
1086 	 * Adjust counters and watermark for pctcpu calc.
1087 	 */
1088 	if (ke->ke_ltick > ticks - SCHED_CPU_TICKS) {
1089 		/*
1090 		 * Shift the tick count out so that the divide doesn't
1091 		 * round away our results.
1092 		 */
1093 		ke->ke_ticks <<= 10;
1094 		ke->ke_ticks = (ke->ke_ticks / (ticks - ke->ke_ftick)) *
1095 			    SCHED_CPU_TICKS;
1096 		ke->ke_ticks >>= 10;
1097 	} else
1098 		ke->ke_ticks = 0;
1099 	ke->ke_ltick = ticks;
1100 	ke->ke_ftick = ke->ke_ltick - SCHED_CPU_TICKS;
1101 }
1102 
1103 void
1104 sched_prio(struct thread *td, u_char prio)
1105 {
1106 	struct kse *ke;
1107 
1108 	ke = td->td_kse;
1109 	mtx_assert(&sched_lock, MA_OWNED);
1110 	if (TD_ON_RUNQ(td)) {
1111 		/*
1112 		 * If the priority has been elevated due to priority
1113 		 * propagation, we may have to move ourselves to a new
1114 		 * queue.  We still call adjustrunqueue below in case kse
1115 		 * needs to fix things up.
1116 		 */
1117 		if (prio < td->td_priority && ke &&
1118 		    (ke->ke_flags & KEF_ASSIGNED) == 0 &&
1119 		    ke->ke_runq != KSEQ_CPU(ke->ke_cpu)->ksq_curr) {
1120 			runq_remove(ke->ke_runq, ke);
1121 			ke->ke_runq = KSEQ_CPU(ke->ke_cpu)->ksq_curr;
1122 			runq_add(ke->ke_runq, ke);
1123 		}
1124 		adjustrunqueue(td, prio);
1125 	} else
1126 		td->td_priority = prio;
1127 }
1128 
1129 void
1130 sched_switch(struct thread *td, struct thread *newtd)
1131 {
1132 	struct kse *ke;
1133 
1134 	mtx_assert(&sched_lock, MA_OWNED);
1135 
1136 	ke = td->td_kse;
1137 
1138 	td->td_last_kse = ke;
1139         td->td_lastcpu = td->td_oncpu;
1140 	td->td_oncpu = NOCPU;
1141 	td->td_flags &= ~TDF_NEEDRESCHED;
1142 	td->td_pflags &= ~TDP_OWEPREEMPT;
1143 
1144 	/*
1145 	 * If the KSE has been assigned it may be in the process of switching
1146 	 * to the new cpu.  This is the case in sched_bind().
1147 	 */
1148 	if ((ke->ke_flags & KEF_ASSIGNED) == 0) {
1149 		if (td == PCPU_GET(idlethread))
1150 			TD_SET_CAN_RUN(td);
1151 		else if (TD_IS_RUNNING(td)) {
1152 			kseq_load_rem(KSEQ_CPU(ke->ke_cpu), ke);
1153 			setrunqueue(td);
1154 		} else {
1155 			if (ke->ke_runq) {
1156 				kseq_load_rem(KSEQ_CPU(ke->ke_cpu), ke);
1157 			} else if ((td->td_flags & TDF_IDLETD) == 0)
1158 				kdb_backtrace();
1159 			/*
1160 			 * We will not be on the run queue. So we must be
1161 			 * sleeping or similar.
1162 			 */
1163 			if (td->td_proc->p_flag & P_SA)
1164 				kse_reassign(ke);
1165 		}
1166 	}
1167 	if (newtd == NULL)
1168 		newtd = choosethread();
1169 	else
1170 		kseq_load_add(KSEQ_SELF(), newtd->td_kse);
1171 	if (td != newtd)
1172 		cpu_switch(td, newtd);
1173 	sched_lock.mtx_lock = (uintptr_t)td;
1174 
1175 	td->td_oncpu = PCPU_GET(cpuid);
1176 }
1177 
1178 void
1179 sched_nice(struct proc *p, int nice)
1180 {
1181 	struct ksegrp *kg;
1182 	struct kse *ke;
1183 	struct thread *td;
1184 	struct kseq *kseq;
1185 
1186 	PROC_LOCK_ASSERT(p, MA_OWNED);
1187 	mtx_assert(&sched_lock, MA_OWNED);
1188 	/*
1189 	 * We need to adjust the nice counts for running KSEs.
1190 	 */
1191 	FOREACH_KSEGRP_IN_PROC(p, kg) {
1192 		if (kg->kg_pri_class == PRI_TIMESHARE) {
1193 			FOREACH_KSE_IN_GROUP(kg, ke) {
1194 				if (ke->ke_runq == NULL)
1195 					continue;
1196 				kseq = KSEQ_CPU(ke->ke_cpu);
1197 				kseq_nice_rem(kseq, p->p_nice);
1198 				kseq_nice_add(kseq, nice);
1199 			}
1200 		}
1201 	}
1202 	p->p_nice = nice;
1203 	FOREACH_KSEGRP_IN_PROC(p, kg) {
1204 		sched_priority(kg);
1205 		FOREACH_THREAD_IN_GROUP(kg, td)
1206 			td->td_flags |= TDF_NEEDRESCHED;
1207 	}
1208 }
1209 
1210 void
1211 sched_sleep(struct thread *td)
1212 {
1213 	mtx_assert(&sched_lock, MA_OWNED);
1214 
1215 	td->td_slptime = ticks;
1216 	td->td_base_pri = td->td_priority;
1217 
1218 	CTR2(KTR_ULE, "sleep kse %p (tick: %d)",
1219 	    td->td_kse, td->td_slptime);
1220 }
1221 
1222 void
1223 sched_wakeup(struct thread *td)
1224 {
1225 	mtx_assert(&sched_lock, MA_OWNED);
1226 
1227 	/*
1228 	 * Let the kseg know how long we slept for.  This is because process
1229 	 * interactivity behavior is modeled in the kseg.
1230 	 */
1231 	if (td->td_slptime) {
1232 		struct ksegrp *kg;
1233 		int hzticks;
1234 
1235 		kg = td->td_ksegrp;
1236 		hzticks = (ticks - td->td_slptime) << 10;
1237 		if (hzticks >= SCHED_SLP_RUN_MAX) {
1238 			kg->kg_slptime = SCHED_SLP_RUN_MAX;
1239 			kg->kg_runtime = 1;
1240 		} else {
1241 			kg->kg_slptime += hzticks;
1242 			sched_interact_update(kg);
1243 		}
1244 		sched_priority(kg);
1245 		if (td->td_kse)
1246 			sched_slice(td->td_kse);
1247 		CTR2(KTR_ULE, "wakeup kse %p (%d ticks)",
1248 		    td->td_kse, hzticks);
1249 		td->td_slptime = 0;
1250 	}
1251 	setrunqueue(td);
1252 }
1253 
1254 /*
1255  * Penalize the parent for creating a new child and initialize the child's
1256  * priority.
1257  */
1258 void
1259 sched_fork(struct thread *td, struct proc *p1)
1260 {
1261 
1262 	mtx_assert(&sched_lock, MA_OWNED);
1263 
1264 	p1->p_nice = td->td_proc->p_nice;
1265 	sched_fork_ksegrp(td, FIRST_KSEGRP_IN_PROC(p1));
1266 	sched_fork_kse(td, FIRST_KSE_IN_PROC(p1));
1267 	sched_fork_thread(td, FIRST_THREAD_IN_PROC(p1));
1268 }
1269 
1270 void
1271 sched_fork_kse(struct thread *td, struct kse *child)
1272 {
1273 
1274 	struct kse *ke = td->td_kse;
1275 
1276 	child->ke_slice = 1;	/* Attempt to quickly learn interactivity. */
1277 	child->ke_cpu = ke->ke_cpu;
1278 	child->ke_runq = NULL;
1279 
1280 	/* Grab our parents cpu estimation information. */
1281 	child->ke_ticks = ke->ke_ticks;
1282 	child->ke_ltick = ke->ke_ltick;
1283 	child->ke_ftick = ke->ke_ftick;
1284 }
1285 
1286 void
1287 sched_fork_ksegrp(struct thread *td, struct ksegrp *child)
1288 {
1289 	struct ksegrp *kg = td->td_ksegrp;
1290 	PROC_LOCK_ASSERT(child->kg_proc, MA_OWNED);
1291 
1292 	child->kg_slptime = kg->kg_slptime;
1293 	child->kg_runtime = kg->kg_runtime;
1294 	child->kg_user_pri = kg->kg_user_pri;
1295 	sched_interact_fork(child);
1296 	kg->kg_runtime += tickincr << 10;
1297 	sched_interact_update(kg);
1298 
1299 	CTR6(KTR_ULE, "sched_fork_ksegrp: %d(%d, %d) - %d(%d, %d)",
1300 	    kg->kg_proc->p_pid, kg->kg_slptime, kg->kg_runtime,
1301 	    child->kg_proc->p_pid, child->kg_slptime, child->kg_runtime);
1302 }
1303 
1304 void
1305 sched_fork_thread(struct thread *td, struct thread *child)
1306 {
1307 }
1308 
1309 void
1310 sched_class(struct ksegrp *kg, int class)
1311 {
1312 	struct kseq *kseq;
1313 	struct kse *ke;
1314 	int nclass;
1315 	int oclass;
1316 
1317 	mtx_assert(&sched_lock, MA_OWNED);
1318 	if (kg->kg_pri_class == class)
1319 		return;
1320 
1321 	nclass = PRI_BASE(class);
1322 	oclass = PRI_BASE(kg->kg_pri_class);
1323 	FOREACH_KSE_IN_GROUP(kg, ke) {
1324 		if (ke->ke_state != KES_ONRUNQ &&
1325 		    ke->ke_state != KES_THREAD)
1326 			continue;
1327 		kseq = KSEQ_CPU(ke->ke_cpu);
1328 
1329 #ifdef SMP
1330 		/*
1331 		 * On SMP if we're on the RUNQ we must adjust the transferable
1332 		 * count because could be changing to or from an interrupt
1333 		 * class.
1334 		 */
1335 		if (ke->ke_state == KES_ONRUNQ) {
1336 			if (KSE_CAN_MIGRATE(ke, oclass)) {
1337 				kseq->ksq_transferable--;
1338 				kseq->ksq_group->ksg_transferable--;
1339 			}
1340 			if (KSE_CAN_MIGRATE(ke, nclass)) {
1341 				kseq->ksq_transferable++;
1342 				kseq->ksq_group->ksg_transferable++;
1343 			}
1344 		}
1345 #endif
1346 		if (oclass == PRI_TIMESHARE) {
1347 			kseq->ksq_load_timeshare--;
1348 			kseq_nice_rem(kseq, kg->kg_proc->p_nice);
1349 		}
1350 		if (nclass == PRI_TIMESHARE) {
1351 			kseq->ksq_load_timeshare++;
1352 			kseq_nice_add(kseq, kg->kg_proc->p_nice);
1353 		}
1354 	}
1355 
1356 	kg->kg_pri_class = class;
1357 }
1358 
1359 /*
1360  * Return some of the child's priority and interactivity to the parent.
1361  */
1362 void
1363 sched_exit(struct proc *p, struct thread *td)
1364 {
1365 	mtx_assert(&sched_lock, MA_OWNED);
1366 	sched_exit_kse(FIRST_KSE_IN_PROC(p), td);
1367 	sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), td);
1368 }
1369 
1370 void
1371 sched_exit_kse(struct kse *ke, struct thread *td)
1372 {
1373 	kseq_load_rem(KSEQ_CPU(td->td_kse->ke_cpu), td->td_kse);
1374 }
1375 
1376 void
1377 sched_exit_ksegrp(struct ksegrp *kg, struct thread *td)
1378 {
1379 	/* kg->kg_slptime += td->td_ksegrp->kg_slptime; */
1380 	kg->kg_runtime += td->td_ksegrp->kg_runtime;
1381 	sched_interact_update(kg);
1382 }
1383 
1384 void
1385 sched_exit_thread(struct thread *td, struct thread *child)
1386 {
1387 }
1388 
1389 void
1390 sched_clock(struct thread *td)
1391 {
1392 	struct kseq *kseq;
1393 	struct ksegrp *kg;
1394 	struct kse *ke;
1395 
1396 	mtx_assert(&sched_lock, MA_OWNED);
1397 #ifdef SMP
1398 	if (ticks == bal_tick)
1399 		sched_balance();
1400 	if (ticks == gbal_tick)
1401 		sched_balance_groups();
1402 #endif
1403 	/*
1404 	 * sched_setup() apparently happens prior to stathz being set.  We
1405 	 * need to resolve the timers earlier in the boot so we can avoid
1406 	 * calculating this here.
1407 	 */
1408 	if (realstathz == 0) {
1409 		realstathz = stathz ? stathz : hz;
1410 		tickincr = hz / realstathz;
1411 		/*
1412 		 * XXX This does not work for values of stathz that are much
1413 		 * larger than hz.
1414 		 */
1415 		if (tickincr == 0)
1416 			tickincr = 1;
1417 	}
1418 
1419 	ke = td->td_kse;
1420 	kg = ke->ke_ksegrp;
1421 
1422 	/* Adjust ticks for pctcpu */
1423 	ke->ke_ticks++;
1424 	ke->ke_ltick = ticks;
1425 
1426 	/* Go up to one second beyond our max and then trim back down */
1427 	if (ke->ke_ftick + SCHED_CPU_TICKS + hz < ke->ke_ltick)
1428 		sched_pctcpu_update(ke);
1429 
1430 	if (td->td_flags & TDF_IDLETD)
1431 		return;
1432 
1433 	CTR4(KTR_ULE, "Tick kse %p (slice: %d, slptime: %d, runtime: %d)",
1434 	    ke, ke->ke_slice, kg->kg_slptime >> 10, kg->kg_runtime >> 10);
1435 	/*
1436 	 * We only do slicing code for TIMESHARE ksegrps.
1437 	 */
1438 	if (kg->kg_pri_class != PRI_TIMESHARE)
1439 		return;
1440 	/*
1441 	 * We used a tick charge it to the ksegrp so that we can compute our
1442 	 * interactivity.
1443 	 */
1444 	kg->kg_runtime += tickincr << 10;
1445 	sched_interact_update(kg);
1446 
1447 	/*
1448 	 * We used up one time slice.
1449 	 */
1450 	if (--ke->ke_slice > 0)
1451 		return;
1452 	/*
1453 	 * We're out of time, recompute priorities and requeue.
1454 	 */
1455 	kseq = KSEQ_SELF();
1456 	kseq_load_rem(kseq, ke);
1457 	sched_priority(kg);
1458 	sched_slice(ke);
1459 	if (SCHED_CURR(kg, ke))
1460 		ke->ke_runq = kseq->ksq_curr;
1461 	else
1462 		ke->ke_runq = kseq->ksq_next;
1463 	kseq_load_add(kseq, ke);
1464 	td->td_flags |= TDF_NEEDRESCHED;
1465 }
1466 
1467 int
1468 sched_runnable(void)
1469 {
1470 	struct kseq *kseq;
1471 	int load;
1472 
1473 	load = 1;
1474 
1475 	kseq = KSEQ_SELF();
1476 #ifdef SMP
1477 	if (kseq->ksq_assigned) {
1478 		mtx_lock_spin(&sched_lock);
1479 		kseq_assign(kseq);
1480 		mtx_unlock_spin(&sched_lock);
1481 	}
1482 #endif
1483 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
1484 		if (kseq->ksq_load > 0)
1485 			goto out;
1486 	} else
1487 		if (kseq->ksq_load - 1 > 0)
1488 			goto out;
1489 	load = 0;
1490 out:
1491 	return (load);
1492 }
1493 
1494 void
1495 sched_userret(struct thread *td)
1496 {
1497 	struct ksegrp *kg;
1498 
1499 	kg = td->td_ksegrp;
1500 
1501 	if (td->td_priority != kg->kg_user_pri) {
1502 		mtx_lock_spin(&sched_lock);
1503 		td->td_priority = kg->kg_user_pri;
1504 		mtx_unlock_spin(&sched_lock);
1505 	}
1506 }
1507 
1508 struct kse *
1509 sched_choose(void)
1510 {
1511 	struct kseq *kseq;
1512 	struct kse *ke;
1513 
1514 	mtx_assert(&sched_lock, MA_OWNED);
1515 	kseq = KSEQ_SELF();
1516 #ifdef SMP
1517 restart:
1518 	if (kseq->ksq_assigned)
1519 		kseq_assign(kseq);
1520 #endif
1521 	ke = kseq_choose(kseq);
1522 	if (ke) {
1523 #ifdef SMP
1524 		if (ke->ke_ksegrp->kg_pri_class == PRI_IDLE)
1525 			if (kseq_idled(kseq) == 0)
1526 				goto restart;
1527 #endif
1528 		kseq_runq_rem(kseq, ke);
1529 		ke->ke_state = KES_THREAD;
1530 
1531 		if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE) {
1532 			CTR4(KTR_ULE, "Run kse %p from %p (slice: %d, pri: %d)",
1533 			    ke, ke->ke_runq, ke->ke_slice,
1534 			    ke->ke_thread->td_priority);
1535 		}
1536 		return (ke);
1537 	}
1538 #ifdef SMP
1539 	if (kseq_idled(kseq) == 0)
1540 		goto restart;
1541 #endif
1542 	return (NULL);
1543 }
1544 
1545 void
1546 sched_add(struct thread *td)
1547 {
1548 
1549 	sched_add_internal(td, 1);
1550 }
1551 
1552 static void
1553 sched_add_internal(struct thread *td, int preemptive)
1554 {
1555 	struct kseq *kseq;
1556 	struct ksegrp *kg;
1557 	struct kse *ke;
1558 	int class;
1559 
1560 	mtx_assert(&sched_lock, MA_OWNED);
1561 	ke = td->td_kse;
1562 	kg = td->td_ksegrp;
1563 	if (ke->ke_flags & KEF_ASSIGNED)
1564 		return;
1565 	kseq = KSEQ_SELF();
1566 	KASSERT((ke->ke_thread != NULL),
1567 	    ("sched_add: No thread on KSE"));
1568 	KASSERT((ke->ke_thread->td_kse != NULL),
1569 	    ("sched_add: No KSE on thread"));
1570 	KASSERT(ke->ke_state != KES_ONRUNQ,
1571 	    ("sched_add: kse %p (%s) already in run queue", ke,
1572 	    ke->ke_proc->p_comm));
1573 	KASSERT(ke->ke_proc->p_sflag & PS_INMEM,
1574 	    ("sched_add: process swapped out"));
1575 	KASSERT(ke->ke_runq == NULL,
1576 	    ("sched_add: KSE %p is still assigned to a run queue", ke));
1577 
1578 	class = PRI_BASE(kg->kg_pri_class);
1579 	switch (class) {
1580 	case PRI_ITHD:
1581 	case PRI_REALTIME:
1582 		ke->ke_runq = kseq->ksq_curr;
1583 		ke->ke_slice = SCHED_SLICE_MAX;
1584 		ke->ke_cpu = PCPU_GET(cpuid);
1585 		break;
1586 	case PRI_TIMESHARE:
1587 		if (SCHED_CURR(kg, ke))
1588 			ke->ke_runq = kseq->ksq_curr;
1589 		else
1590 			ke->ke_runq = kseq->ksq_next;
1591 		break;
1592 	case PRI_IDLE:
1593 		/*
1594 		 * This is for priority prop.
1595 		 */
1596 		if (ke->ke_thread->td_priority < PRI_MIN_IDLE)
1597 			ke->ke_runq = kseq->ksq_curr;
1598 		else
1599 			ke->ke_runq = &kseq->ksq_idle;
1600 		ke->ke_slice = SCHED_SLICE_MIN;
1601 		break;
1602 	default:
1603 		panic("Unknown pri class.");
1604 		break;
1605 	}
1606 #ifdef SMP
1607 	if (ke->ke_cpu != PCPU_GET(cpuid)) {
1608 		ke->ke_runq = NULL;
1609 		kseq_notify(ke, ke->ke_cpu);
1610 		return;
1611 	}
1612 	/*
1613 	 * If we had been idle, clear our bit in the group and potentially
1614 	 * the global bitmap.  If not, see if we should transfer this thread.
1615 	 */
1616 	if ((class == PRI_TIMESHARE || class == PRI_REALTIME) &&
1617 	    (kseq->ksq_group->ksg_idlemask & PCPU_GET(cpumask)) != 0) {
1618 		/*
1619 		 * Check to see if our group is unidling, and if so, remove it
1620 		 * from the global idle mask.
1621 		 */
1622 		if (kseq->ksq_group->ksg_idlemask ==
1623 		    kseq->ksq_group->ksg_cpumask)
1624 			atomic_clear_int(&kseq_idle, kseq->ksq_group->ksg_mask);
1625 		/*
1626 		 * Now remove ourselves from the group specific idle mask.
1627 		 */
1628 		kseq->ksq_group->ksg_idlemask &= ~PCPU_GET(cpumask);
1629 	} else if (kseq->ksq_load > 1 && KSE_CAN_MIGRATE(ke, class))
1630 		if (kseq_transfer(kseq, ke, class))
1631 			return;
1632 #endif
1633         if (td->td_priority < curthread->td_priority)
1634                 curthread->td_flags |= TDF_NEEDRESCHED;
1635 
1636 #ifdef SMP
1637 	/*
1638 	 * Only try to preempt if the thread is unpinned or pinned to the
1639 	 * current CPU.
1640 	 */
1641 	if (KSE_CAN_MIGRATE(ke, class) || ke->ke_cpu == PCPU_GET(cpuid))
1642 #endif
1643 	if (preemptive && maybe_preempt(td))
1644 		return;
1645 	ke->ke_ksegrp->kg_runq_kses++;
1646 	ke->ke_state = KES_ONRUNQ;
1647 
1648 	kseq_runq_add(kseq, ke);
1649 	kseq_load_add(kseq, ke);
1650 }
1651 
1652 void
1653 sched_rem(struct thread *td)
1654 {
1655 	struct kseq *kseq;
1656 	struct kse *ke;
1657 
1658 	ke = td->td_kse;
1659 	/*
1660 	 * It is safe to just return here because sched_rem() is only ever
1661 	 * used in places where we're immediately going to add the
1662 	 * kse back on again.  In that case it'll be added with the correct
1663 	 * thread and priority when the caller drops the sched_lock.
1664 	 */
1665 	if (ke->ke_flags & KEF_ASSIGNED)
1666 		return;
1667 	mtx_assert(&sched_lock, MA_OWNED);
1668 	KASSERT((ke->ke_state == KES_ONRUNQ),
1669 	    ("sched_rem: KSE not on run queue"));
1670 
1671 	ke->ke_state = KES_THREAD;
1672 	ke->ke_ksegrp->kg_runq_kses--;
1673 	kseq = KSEQ_CPU(ke->ke_cpu);
1674 	kseq_runq_rem(kseq, ke);
1675 	kseq_load_rem(kseq, ke);
1676 }
1677 
1678 fixpt_t
1679 sched_pctcpu(struct thread *td)
1680 {
1681 	fixpt_t pctcpu;
1682 	struct kse *ke;
1683 
1684 	pctcpu = 0;
1685 	ke = td->td_kse;
1686 	if (ke == NULL)
1687 		return (0);
1688 
1689 	mtx_lock_spin(&sched_lock);
1690 	if (ke->ke_ticks) {
1691 		int rtick;
1692 
1693 		/*
1694 		 * Don't update more frequently than twice a second.  Allowing
1695 		 * this causes the cpu usage to decay away too quickly due to
1696 		 * rounding errors.
1697 		 */
1698 		if (ke->ke_ftick + SCHED_CPU_TICKS < ke->ke_ltick ||
1699 		    ke->ke_ltick < (ticks - (hz / 2)))
1700 			sched_pctcpu_update(ke);
1701 		/* How many rtick per second ? */
1702 		rtick = min(ke->ke_ticks / SCHED_CPU_TIME, SCHED_CPU_TICKS);
1703 		pctcpu = (FSCALE * ((FSCALE * rtick)/realstathz)) >> FSHIFT;
1704 	}
1705 
1706 	ke->ke_proc->p_swtime = ke->ke_ltick - ke->ke_ftick;
1707 	mtx_unlock_spin(&sched_lock);
1708 
1709 	return (pctcpu);
1710 }
1711 
1712 void
1713 sched_bind(struct thread *td, int cpu)
1714 {
1715 	struct kse *ke;
1716 
1717 	mtx_assert(&sched_lock, MA_OWNED);
1718 	ke = td->td_kse;
1719 	ke->ke_flags |= KEF_BOUND;
1720 #ifdef SMP
1721 	if (PCPU_GET(cpuid) == cpu)
1722 		return;
1723 	/* sched_rem without the runq_remove */
1724 	ke->ke_state = KES_THREAD;
1725 	ke->ke_ksegrp->kg_runq_kses--;
1726 	kseq_load_rem(KSEQ_CPU(ke->ke_cpu), ke);
1727 	kseq_notify(ke, cpu);
1728 	/* When we return from mi_switch we'll be on the correct cpu. */
1729 	mi_switch(SW_VOL, NULL);
1730 #endif
1731 }
1732 
1733 void
1734 sched_unbind(struct thread *td)
1735 {
1736 	mtx_assert(&sched_lock, MA_OWNED);
1737 	td->td_kse->ke_flags &= ~KEF_BOUND;
1738 }
1739 
1740 int
1741 sched_load(void)
1742 {
1743 #ifdef SMP
1744 	int total;
1745 	int i;
1746 
1747 	total = 0;
1748 	for (i = 0; i <= ksg_maxid; i++)
1749 		total += KSEQ_GROUP(i)->ksg_load;
1750 	return (total);
1751 #else
1752 	return (KSEQ_SELF()->ksq_sysload);
1753 #endif
1754 }
1755 
1756 int
1757 sched_sizeof_kse(void)
1758 {
1759 	return (sizeof(struct kse) + sizeof(struct ke_sched));
1760 }
1761 
1762 int
1763 sched_sizeof_ksegrp(void)
1764 {
1765 	return (sizeof(struct ksegrp) + sizeof(struct kg_sched));
1766 }
1767 
1768 int
1769 sched_sizeof_proc(void)
1770 {
1771 	return (sizeof(struct proc));
1772 }
1773 
1774 int
1775 sched_sizeof_thread(void)
1776 {
1777 	return (sizeof(struct thread) + sizeof(struct td_sched));
1778 }
1779