xref: /freebsd/sys/kern/sched_ule.c (revision a0f1535205b5b23672cf0bed68f3e1119a06d44a)
1 /*-
2  * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice unmodified, this list of conditions, and the following
10  *    disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25  */
26 
27 /*
28  * This file implements the ULE scheduler.  ULE supports independent CPU
29  * run queues and fine grain locking.  It has superior interactive
30  * performance under load even on uni-processor systems.
31  *
32  * etymology:
33  *   ULE is the last three letters in schedule.  It owes its name to a
34  * generic user created for a scheduling system by Paul Mikesell at
35  * Isilon Systems and a general lack of creativity on the part of the author.
36  */
37 
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
40 
41 #include "opt_hwpmc_hooks.h"
42 #include "opt_kdtrace.h"
43 #include "opt_sched.h"
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/kdb.h>
48 #include <sys/kernel.h>
49 #include <sys/ktr.h>
50 #include <sys/lock.h>
51 #include <sys/mutex.h>
52 #include <sys/proc.h>
53 #include <sys/resource.h>
54 #include <sys/resourcevar.h>
55 #include <sys/sched.h>
56 #include <sys/smp.h>
57 #include <sys/sx.h>
58 #include <sys/sysctl.h>
59 #include <sys/sysproto.h>
60 #include <sys/turnstile.h>
61 #include <sys/umtx.h>
62 #include <sys/vmmeter.h>
63 #include <sys/cpuset.h>
64 #include <sys/sbuf.h>
65 #ifdef KTRACE
66 #include <sys/uio.h>
67 #include <sys/ktrace.h>
68 #endif
69 
70 #ifdef HWPMC_HOOKS
71 #include <sys/pmckern.h>
72 #endif
73 
74 #ifdef KDTRACE_HOOKS
75 #include <sys/dtrace_bsd.h>
76 int				dtrace_vtime_active;
77 dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
78 #endif
79 
80 #include <machine/cpu.h>
81 #include <machine/smp.h>
82 
83 #if defined(__sparc64__) || defined(__mips__)
84 #error "This architecture is not currently compatible with ULE"
85 #endif
86 
87 #define	KTR_ULE	0
88 
89 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
90 #define	TDQ_NAME_LEN	(sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
91 #define	TDQ_LOADNAME_LEN	(PCPU_NAME_LEN + sizeof(" load"))
92 
93 /*
94  * Thread scheduler specific section.  All fields are protected
95  * by the thread lock.
96  */
97 struct td_sched {
98 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
99 	short		ts_flags;	/* TSF_* flags. */
100 	u_char		ts_cpu;		/* CPU that we have affinity for. */
101 	int		ts_rltick;	/* Real last tick, for affinity. */
102 	int		ts_slice;	/* Ticks of slice remaining. */
103 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
104 	u_int		ts_runtime;	/* Number of ticks we were running */
105 	int		ts_ltick;	/* Last tick that we were running on */
106 	int		ts_ftick;	/* First tick that we were running on */
107 	int		ts_ticks;	/* Tick count */
108 #ifdef KTR
109 	char		ts_name[TS_NAME_LEN];
110 #endif
111 };
112 /* flags kept in ts_flags */
113 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
114 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
115 
116 static struct td_sched td_sched0;
117 
118 #define	THREAD_CAN_MIGRATE(td)	((td)->td_pinned == 0)
119 #define	THREAD_CAN_SCHED(td, cpu)	\
120     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
121 
122 /*
123  * Cpu percentage computation macros and defines.
124  *
125  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
126  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
127  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
128  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
129  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
130  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
131  */
132 #define	SCHED_TICK_SECS		10
133 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
134 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
135 #define	SCHED_TICK_SHIFT	10
136 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
137 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
138 
139 /*
140  * These macros determine priorities for non-interactive threads.  They are
141  * assigned a priority based on their recent cpu utilization as expressed
142  * by the ratio of ticks to the tick total.  NHALF priorities at the start
143  * and end of the MIN to MAX timeshare range are only reachable with negative
144  * or positive nice respectively.
145  *
146  * PRI_RANGE:	Priority range for utilization dependent priorities.
147  * PRI_NRESV:	Number of nice values.
148  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
149  * PRI_NICE:	Determines the part of the priority inherited from nice.
150  */
151 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
152 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
153 #define	SCHED_PRI_MIN		(PRI_MIN_TIMESHARE + SCHED_PRI_NHALF)
154 #define	SCHED_PRI_MAX		(PRI_MAX_TIMESHARE - SCHED_PRI_NHALF)
155 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN)
156 #define	SCHED_PRI_TICKS(ts)						\
157     (SCHED_TICK_HZ((ts)) /						\
158     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
159 #define	SCHED_PRI_NICE(nice)	(nice)
160 
161 /*
162  * These determine the interactivity of a process.  Interactivity differs from
163  * cpu utilization in that it expresses the voluntary time slept vs time ran
164  * while cpu utilization includes all time not running.  This more accurately
165  * models the intent of the thread.
166  *
167  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
168  *		before throttling back.
169  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
170  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
171  * INTERACT_THRESH:	Threshhold for placement on the current runq.
172  */
173 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
174 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
175 #define	SCHED_INTERACT_MAX	(100)
176 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
177 #define	SCHED_INTERACT_THRESH	(30)
178 
179 /*
180  * tickincr:		Converts a stathz tick into a hz domain scaled by
181  *			the shift factor.  Without the shift the error rate
182  *			due to rounding would be unacceptably high.
183  * realstathz:		stathz is sometimes 0 and run off of hz.
184  * sched_slice:		Runtime of each thread before rescheduling.
185  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
186  */
187 static int sched_interact = SCHED_INTERACT_THRESH;
188 static int realstathz;
189 static int tickincr;
190 static int sched_slice = 1;
191 #ifdef PREEMPTION
192 #ifdef FULL_PREEMPTION
193 static int preempt_thresh = PRI_MAX_IDLE;
194 #else
195 static int preempt_thresh = PRI_MIN_KERN;
196 #endif
197 #else
198 static int preempt_thresh = 0;
199 #endif
200 static int static_boost = PRI_MIN_TIMESHARE;
201 static int sched_idlespins = 10000;
202 static int sched_idlespinthresh = 4;
203 
204 /*
205  * tdq - per processor runqs and statistics.  All fields are protected by the
206  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
207  * locking in sched_pickcpu();
208  */
209 struct tdq {
210 	/* Ordered to improve efficiency of cpu_search() and switch(). */
211 	struct mtx	tdq_lock;		/* run queue lock. */
212 	struct cpu_group *tdq_cg;		/* Pointer to cpu topology. */
213 	volatile int	tdq_load;		/* Aggregate load. */
214 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
215 	int		tdq_transferable;	/* Transferable thread count. */
216 	short		tdq_switchcnt;		/* Switches this tick. */
217 	short		tdq_oldswitchcnt;	/* Switches last tick. */
218 	u_char		tdq_lowpri;		/* Lowest priority thread. */
219 	u_char		tdq_ipipending;		/* IPI pending. */
220 	u_char		tdq_idx;		/* Current insert index. */
221 	u_char		tdq_ridx;		/* Current removal index. */
222 	struct runq	tdq_realtime;		/* real-time run queue. */
223 	struct runq	tdq_timeshare;		/* timeshare run queue. */
224 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
225 	char		tdq_name[TDQ_NAME_LEN];
226 #ifdef KTR
227 	char		tdq_loadname[TDQ_LOADNAME_LEN];
228 #endif
229 } __aligned(64);
230 
231 /* Idle thread states and config. */
232 #define	TDQ_RUNNING	1
233 #define	TDQ_IDLE	2
234 
235 #ifdef SMP
236 struct cpu_group *cpu_top;		/* CPU topology */
237 
238 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
239 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
240 
241 /*
242  * Run-time tunables.
243  */
244 static int rebalance = 1;
245 static int balance_interval = 128;	/* Default set in sched_initticks(). */
246 static int affinity;
247 static int steal_htt = 1;
248 static int steal_idle = 1;
249 static int steal_thresh = 2;
250 
251 /*
252  * One thread queue per processor.
253  */
254 static struct tdq	tdq_cpu[MAXCPU];
255 static struct tdq	*balance_tdq;
256 static int balance_ticks;
257 
258 #define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
259 #define	TDQ_CPU(x)	(&tdq_cpu[(x)])
260 #define	TDQ_ID(x)	((int)((x) - tdq_cpu))
261 #else	/* !SMP */
262 static struct tdq	tdq_cpu;
263 
264 #define	TDQ_ID(x)	(0)
265 #define	TDQ_SELF()	(&tdq_cpu)
266 #define	TDQ_CPU(x)	(&tdq_cpu)
267 #endif
268 
269 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
270 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
271 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
272 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
273 #define	TDQ_LOCKPTR(t)		(&(t)->tdq_lock)
274 
275 static void sched_priority(struct thread *);
276 static void sched_thread_priority(struct thread *, u_char);
277 static int sched_interact_score(struct thread *);
278 static void sched_interact_update(struct thread *);
279 static void sched_interact_fork(struct thread *);
280 static void sched_pctcpu_update(struct td_sched *);
281 
282 /* Operations on per processor queues */
283 static struct thread *tdq_choose(struct tdq *);
284 static void tdq_setup(struct tdq *);
285 static void tdq_load_add(struct tdq *, struct thread *);
286 static void tdq_load_rem(struct tdq *, struct thread *);
287 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
288 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
289 static inline int sched_shouldpreempt(int, int, int);
290 void tdq_print(int cpu);
291 static void runq_print(struct runq *rq);
292 static void tdq_add(struct tdq *, struct thread *, int);
293 #ifdef SMP
294 static int tdq_move(struct tdq *, struct tdq *);
295 static int tdq_idled(struct tdq *);
296 static void tdq_notify(struct tdq *, struct thread *);
297 static struct thread *tdq_steal(struct tdq *, int);
298 static struct thread *runq_steal(struct runq *, int);
299 static int sched_pickcpu(struct thread *, int);
300 static void sched_balance(void);
301 static int sched_balance_pair(struct tdq *, struct tdq *);
302 static inline struct tdq *sched_setcpu(struct thread *, int, int);
303 static inline struct mtx *thread_block_switch(struct thread *);
304 static inline void thread_unblock_switch(struct thread *, struct mtx *);
305 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
306 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
307 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
308     struct cpu_group *cg, int indent);
309 #endif
310 
311 static void sched_setup(void *dummy);
312 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
313 
314 static void sched_initticks(void *dummy);
315 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
316     NULL);
317 
318 /*
319  * Print the threads waiting on a run-queue.
320  */
321 static void
322 runq_print(struct runq *rq)
323 {
324 	struct rqhead *rqh;
325 	struct thread *td;
326 	int pri;
327 	int j;
328 	int i;
329 
330 	for (i = 0; i < RQB_LEN; i++) {
331 		printf("\t\trunq bits %d 0x%zx\n",
332 		    i, rq->rq_status.rqb_bits[i]);
333 		for (j = 0; j < RQB_BPW; j++)
334 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
335 				pri = j + (i << RQB_L2BPW);
336 				rqh = &rq->rq_queues[pri];
337 				TAILQ_FOREACH(td, rqh, td_runq) {
338 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
339 					    td, td->td_name, td->td_priority,
340 					    td->td_rqindex, pri);
341 				}
342 			}
343 	}
344 }
345 
346 /*
347  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
348  */
349 void
350 tdq_print(int cpu)
351 {
352 	struct tdq *tdq;
353 
354 	tdq = TDQ_CPU(cpu);
355 
356 	printf("tdq %d:\n", TDQ_ID(tdq));
357 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
358 	printf("\tLock name:      %s\n", tdq->tdq_name);
359 	printf("\tload:           %d\n", tdq->tdq_load);
360 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
361 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
362 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
363 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
364 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
365 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
366 	printf("\trealtime runq:\n");
367 	runq_print(&tdq->tdq_realtime);
368 	printf("\ttimeshare runq:\n");
369 	runq_print(&tdq->tdq_timeshare);
370 	printf("\tidle runq:\n");
371 	runq_print(&tdq->tdq_idle);
372 }
373 
374 static inline int
375 sched_shouldpreempt(int pri, int cpri, int remote)
376 {
377 	/*
378 	 * If the new priority is not better than the current priority there is
379 	 * nothing to do.
380 	 */
381 	if (pri >= cpri)
382 		return (0);
383 	/*
384 	 * Always preempt idle.
385 	 */
386 	if (cpri >= PRI_MIN_IDLE)
387 		return (1);
388 	/*
389 	 * If preemption is disabled don't preempt others.
390 	 */
391 	if (preempt_thresh == 0)
392 		return (0);
393 	/*
394 	 * Preempt if we exceed the threshold.
395 	 */
396 	if (pri <= preempt_thresh)
397 		return (1);
398 	/*
399 	 * If we're realtime or better and there is timeshare or worse running
400 	 * preempt only remote processors.
401 	 */
402 	if (remote && pri <= PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME)
403 		return (1);
404 	return (0);
405 }
406 
407 #define	TS_RQ_PPQ	(((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS)
408 /*
409  * Add a thread to the actual run-queue.  Keeps transferable counts up to
410  * date with what is actually on the run-queue.  Selects the correct
411  * queue position for timeshare threads.
412  */
413 static __inline void
414 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
415 {
416 	struct td_sched *ts;
417 	u_char pri;
418 
419 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
420 	THREAD_LOCK_ASSERT(td, MA_OWNED);
421 
422 	pri = td->td_priority;
423 	ts = td->td_sched;
424 	TD_SET_RUNQ(td);
425 	if (THREAD_CAN_MIGRATE(td)) {
426 		tdq->tdq_transferable++;
427 		ts->ts_flags |= TSF_XFERABLE;
428 	}
429 	if (pri <= PRI_MAX_REALTIME) {
430 		ts->ts_runq = &tdq->tdq_realtime;
431 	} else if (pri <= PRI_MAX_TIMESHARE) {
432 		ts->ts_runq = &tdq->tdq_timeshare;
433 		KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE,
434 			("Invalid priority %d on timeshare runq", pri));
435 		/*
436 		 * This queue contains only priorities between MIN and MAX
437 		 * realtime.  Use the whole queue to represent these values.
438 		 */
439 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
440 			pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ;
441 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
442 			/*
443 			 * This effectively shortens the queue by one so we
444 			 * can have a one slot difference between idx and
445 			 * ridx while we wait for threads to drain.
446 			 */
447 			if (tdq->tdq_ridx != tdq->tdq_idx &&
448 			    pri == tdq->tdq_ridx)
449 				pri = (unsigned char)(pri - 1) % RQ_NQS;
450 		} else
451 			pri = tdq->tdq_ridx;
452 		runq_add_pri(ts->ts_runq, td, pri, flags);
453 		return;
454 	} else
455 		ts->ts_runq = &tdq->tdq_idle;
456 	runq_add(ts->ts_runq, td, flags);
457 }
458 
459 /*
460  * Remove a thread from a run-queue.  This typically happens when a thread
461  * is selected to run.  Running threads are not on the queue and the
462  * transferable count does not reflect them.
463  */
464 static __inline void
465 tdq_runq_rem(struct tdq *tdq, struct thread *td)
466 {
467 	struct td_sched *ts;
468 
469 	ts = td->td_sched;
470 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
471 	KASSERT(ts->ts_runq != NULL,
472 	    ("tdq_runq_remove: thread %p null ts_runq", td));
473 	if (ts->ts_flags & TSF_XFERABLE) {
474 		tdq->tdq_transferable--;
475 		ts->ts_flags &= ~TSF_XFERABLE;
476 	}
477 	if (ts->ts_runq == &tdq->tdq_timeshare) {
478 		if (tdq->tdq_idx != tdq->tdq_ridx)
479 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
480 		else
481 			runq_remove_idx(ts->ts_runq, td, NULL);
482 	} else
483 		runq_remove(ts->ts_runq, td);
484 }
485 
486 /*
487  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
488  * for this thread to the referenced thread queue.
489  */
490 static void
491 tdq_load_add(struct tdq *tdq, struct thread *td)
492 {
493 
494 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
495 	THREAD_LOCK_ASSERT(td, MA_OWNED);
496 
497 	tdq->tdq_load++;
498 	if ((td->td_proc->p_flag & P_NOLOAD) == 0)
499 		tdq->tdq_sysload++;
500 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
501 }
502 
503 /*
504  * Remove the load from a thread that is transitioning to a sleep state or
505  * exiting.
506  */
507 static void
508 tdq_load_rem(struct tdq *tdq, struct thread *td)
509 {
510 
511 	THREAD_LOCK_ASSERT(td, MA_OWNED);
512 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
513 	KASSERT(tdq->tdq_load != 0,
514 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
515 
516 	tdq->tdq_load--;
517 	if ((td->td_proc->p_flag & P_NOLOAD) == 0)
518 		tdq->tdq_sysload--;
519 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
520 }
521 
522 /*
523  * Set lowpri to its exact value by searching the run-queue and
524  * evaluating curthread.  curthread may be passed as an optimization.
525  */
526 static void
527 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
528 {
529 	struct thread *td;
530 
531 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
532 	if (ctd == NULL)
533 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
534 	td = tdq_choose(tdq);
535 	if (td == NULL || td->td_priority > ctd->td_priority)
536 		tdq->tdq_lowpri = ctd->td_priority;
537 	else
538 		tdq->tdq_lowpri = td->td_priority;
539 }
540 
541 #ifdef SMP
542 struct cpu_search {
543 	cpuset_t cs_mask;
544 	u_int	cs_load;
545 	u_int	cs_cpu;
546 	int	cs_limit;	/* Min priority for low min load for high. */
547 };
548 
549 #define	CPU_SEARCH_LOWEST	0x1
550 #define	CPU_SEARCH_HIGHEST	0x2
551 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
552 
553 #define	CPUSET_FOREACH(cpu, mask)				\
554 	for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++)		\
555 		if ((mask) & 1 << (cpu))
556 
557 static __inline int cpu_search(struct cpu_group *cg, struct cpu_search *low,
558     struct cpu_search *high, const int match);
559 int cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low);
560 int cpu_search_highest(struct cpu_group *cg, struct cpu_search *high);
561 int cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
562     struct cpu_search *high);
563 
564 /*
565  * This routine compares according to the match argument and should be
566  * reduced in actual instantiations via constant propagation and dead code
567  * elimination.
568  */
569 static __inline int
570 cpu_compare(int cpu, struct cpu_search *low, struct cpu_search *high,
571     const int match)
572 {
573 	struct tdq *tdq;
574 
575 	tdq = TDQ_CPU(cpu);
576 	if (match & CPU_SEARCH_LOWEST)
577 		if (CPU_ISSET(cpu, &low->cs_mask) &&
578 		    tdq->tdq_load < low->cs_load &&
579 		    tdq->tdq_lowpri > low->cs_limit) {
580 			low->cs_cpu = cpu;
581 			low->cs_load = tdq->tdq_load;
582 		}
583 	if (match & CPU_SEARCH_HIGHEST)
584 		if (CPU_ISSET(cpu, &high->cs_mask) &&
585 		    tdq->tdq_load >= high->cs_limit &&
586 		    tdq->tdq_load > high->cs_load &&
587 		    tdq->tdq_transferable) {
588 			high->cs_cpu = cpu;
589 			high->cs_load = tdq->tdq_load;
590 		}
591 	return (tdq->tdq_load);
592 }
593 
594 /*
595  * Search the tree of cpu_groups for the lowest or highest loaded cpu
596  * according to the match argument.  This routine actually compares the
597  * load on all paths through the tree and finds the least loaded cpu on
598  * the least loaded path, which may differ from the least loaded cpu in
599  * the system.  This balances work among caches and busses.
600  *
601  * This inline is instantiated in three forms below using constants for the
602  * match argument.  It is reduced to the minimum set for each case.  It is
603  * also recursive to the depth of the tree.
604  */
605 static __inline int
606 cpu_search(struct cpu_group *cg, struct cpu_search *low,
607     struct cpu_search *high, const int match)
608 {
609 	int total;
610 
611 	total = 0;
612 	if (cg->cg_children) {
613 		struct cpu_search lgroup;
614 		struct cpu_search hgroup;
615 		struct cpu_group *child;
616 		u_int lload;
617 		int hload;
618 		int load;
619 		int i;
620 
621 		lload = -1;
622 		hload = -1;
623 		for (i = 0; i < cg->cg_children; i++) {
624 			child = &cg->cg_child[i];
625 			if (match & CPU_SEARCH_LOWEST) {
626 				lgroup = *low;
627 				lgroup.cs_load = -1;
628 			}
629 			if (match & CPU_SEARCH_HIGHEST) {
630 				hgroup = *high;
631 				lgroup.cs_load = 0;
632 			}
633 			switch (match) {
634 			case CPU_SEARCH_LOWEST:
635 				load = cpu_search_lowest(child, &lgroup);
636 				break;
637 			case CPU_SEARCH_HIGHEST:
638 				load = cpu_search_highest(child, &hgroup);
639 				break;
640 			case CPU_SEARCH_BOTH:
641 				load = cpu_search_both(child, &lgroup, &hgroup);
642 				break;
643 			}
644 			total += load;
645 			if (match & CPU_SEARCH_LOWEST)
646 				if (load < lload || low->cs_cpu == -1) {
647 					*low = lgroup;
648 					lload = load;
649 				}
650 			if (match & CPU_SEARCH_HIGHEST)
651 				if (load > hload || high->cs_cpu == -1) {
652 					hload = load;
653 					*high = hgroup;
654 				}
655 		}
656 	} else {
657 		int cpu;
658 
659 		CPUSET_FOREACH(cpu, cg->cg_mask)
660 			total += cpu_compare(cpu, low, high, match);
661 	}
662 	return (total);
663 }
664 
665 /*
666  * cpu_search instantiations must pass constants to maintain the inline
667  * optimization.
668  */
669 int
670 cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low)
671 {
672 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
673 }
674 
675 int
676 cpu_search_highest(struct cpu_group *cg, struct cpu_search *high)
677 {
678 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
679 }
680 
681 int
682 cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
683     struct cpu_search *high)
684 {
685 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
686 }
687 
688 /*
689  * Find the cpu with the least load via the least loaded path that has a
690  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
691  * acceptable.
692  */
693 static inline int
694 sched_lowest(struct cpu_group *cg, cpuset_t mask, int pri)
695 {
696 	struct cpu_search low;
697 
698 	low.cs_cpu = -1;
699 	low.cs_load = -1;
700 	low.cs_mask = mask;
701 	low.cs_limit = pri;
702 	cpu_search_lowest(cg, &low);
703 	return low.cs_cpu;
704 }
705 
706 /*
707  * Find the cpu with the highest load via the highest loaded path.
708  */
709 static inline int
710 sched_highest(struct cpu_group *cg, cpuset_t mask, int minload)
711 {
712 	struct cpu_search high;
713 
714 	high.cs_cpu = -1;
715 	high.cs_load = 0;
716 	high.cs_mask = mask;
717 	high.cs_limit = minload;
718 	cpu_search_highest(cg, &high);
719 	return high.cs_cpu;
720 }
721 
722 /*
723  * Simultaneously find the highest and lowest loaded cpu reachable via
724  * cg.
725  */
726 static inline void
727 sched_both(struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu)
728 {
729 	struct cpu_search high;
730 	struct cpu_search low;
731 
732 	low.cs_cpu = -1;
733 	low.cs_limit = -1;
734 	low.cs_load = -1;
735 	low.cs_mask = mask;
736 	high.cs_load = 0;
737 	high.cs_cpu = -1;
738 	high.cs_limit = -1;
739 	high.cs_mask = mask;
740 	cpu_search_both(cg, &low, &high);
741 	*lowcpu = low.cs_cpu;
742 	*highcpu = high.cs_cpu;
743 	return;
744 }
745 
746 static void
747 sched_balance_group(struct cpu_group *cg)
748 {
749 	cpuset_t mask;
750 	int high;
751 	int low;
752 	int i;
753 
754 	CPU_FILL(&mask);
755 	for (;;) {
756 		sched_both(cg, mask, &low, &high);
757 		if (low == high || low == -1 || high == -1)
758 			break;
759 		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low)))
760 			break;
761 		/*
762 		 * If we failed to move any threads determine which cpu
763 		 * to kick out of the set and try again.
764 	 	 */
765 		if (TDQ_CPU(high)->tdq_transferable == 0)
766 			CPU_CLR(high, &mask);
767 		else
768 			CPU_CLR(low, &mask);
769 	}
770 
771 	for (i = 0; i < cg->cg_children; i++)
772 		sched_balance_group(&cg->cg_child[i]);
773 }
774 
775 static void
776 sched_balance()
777 {
778 	struct tdq *tdq;
779 
780 	/*
781 	 * Select a random time between .5 * balance_interval and
782 	 * 1.5 * balance_interval.
783 	 */
784 	balance_ticks = max(balance_interval / 2, 1);
785 	balance_ticks += random() % balance_interval;
786 	if (smp_started == 0 || rebalance == 0)
787 		return;
788 	tdq = TDQ_SELF();
789 	TDQ_UNLOCK(tdq);
790 	sched_balance_group(cpu_top);
791 	TDQ_LOCK(tdq);
792 }
793 
794 /*
795  * Lock two thread queues using their address to maintain lock order.
796  */
797 static void
798 tdq_lock_pair(struct tdq *one, struct tdq *two)
799 {
800 	if (one < two) {
801 		TDQ_LOCK(one);
802 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
803 	} else {
804 		TDQ_LOCK(two);
805 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
806 	}
807 }
808 
809 /*
810  * Unlock two thread queues.  Order is not important here.
811  */
812 static void
813 tdq_unlock_pair(struct tdq *one, struct tdq *two)
814 {
815 	TDQ_UNLOCK(one);
816 	TDQ_UNLOCK(two);
817 }
818 
819 /*
820  * Transfer load between two imbalanced thread queues.
821  */
822 static int
823 sched_balance_pair(struct tdq *high, struct tdq *low)
824 {
825 	int transferable;
826 	int high_load;
827 	int low_load;
828 	int moved;
829 	int move;
830 	int diff;
831 	int i;
832 
833 	tdq_lock_pair(high, low);
834 	transferable = high->tdq_transferable;
835 	high_load = high->tdq_load;
836 	low_load = low->tdq_load;
837 	moved = 0;
838 	/*
839 	 * Determine what the imbalance is and then adjust that to how many
840 	 * threads we actually have to give up (transferable).
841 	 */
842 	if (transferable != 0) {
843 		diff = high_load - low_load;
844 		move = diff / 2;
845 		if (diff & 0x1)
846 			move++;
847 		move = min(move, transferable);
848 		for (i = 0; i < move; i++)
849 			moved += tdq_move(high, low);
850 		/*
851 		 * IPI the target cpu to force it to reschedule with the new
852 		 * workload.
853 		 */
854 		ipi_selected(1 << TDQ_ID(low), IPI_PREEMPT);
855 	}
856 	tdq_unlock_pair(high, low);
857 	return (moved);
858 }
859 
860 /*
861  * Move a thread from one thread queue to another.
862  */
863 static int
864 tdq_move(struct tdq *from, struct tdq *to)
865 {
866 	struct td_sched *ts;
867 	struct thread *td;
868 	struct tdq *tdq;
869 	int cpu;
870 
871 	TDQ_LOCK_ASSERT(from, MA_OWNED);
872 	TDQ_LOCK_ASSERT(to, MA_OWNED);
873 
874 	tdq = from;
875 	cpu = TDQ_ID(to);
876 	td = tdq_steal(tdq, cpu);
877 	if (td == NULL)
878 		return (0);
879 	ts = td->td_sched;
880 	/*
881 	 * Although the run queue is locked the thread may be blocked.  Lock
882 	 * it to clear this and acquire the run-queue lock.
883 	 */
884 	thread_lock(td);
885 	/* Drop recursive lock on from acquired via thread_lock(). */
886 	TDQ_UNLOCK(from);
887 	sched_rem(td);
888 	ts->ts_cpu = cpu;
889 	td->td_lock = TDQ_LOCKPTR(to);
890 	tdq_add(to, td, SRQ_YIELDING);
891 	return (1);
892 }
893 
894 /*
895  * This tdq has idled.  Try to steal a thread from another cpu and switch
896  * to it.
897  */
898 static int
899 tdq_idled(struct tdq *tdq)
900 {
901 	struct cpu_group *cg;
902 	struct tdq *steal;
903 	cpuset_t mask;
904 	int thresh;
905 	int cpu;
906 
907 	if (smp_started == 0 || steal_idle == 0)
908 		return (1);
909 	CPU_FILL(&mask);
910 	CPU_CLR(PCPU_GET(cpuid), &mask);
911 	/* We don't want to be preempted while we're iterating. */
912 	spinlock_enter();
913 	for (cg = tdq->tdq_cg; cg != NULL; ) {
914 		if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
915 			thresh = steal_thresh;
916 		else
917 			thresh = 1;
918 		cpu = sched_highest(cg, mask, thresh);
919 		if (cpu == -1) {
920 			cg = cg->cg_parent;
921 			continue;
922 		}
923 		steal = TDQ_CPU(cpu);
924 		CPU_CLR(cpu, &mask);
925 		tdq_lock_pair(tdq, steal);
926 		if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
927 			tdq_unlock_pair(tdq, steal);
928 			continue;
929 		}
930 		/*
931 		 * If a thread was added while interrupts were disabled don't
932 		 * steal one here.  If we fail to acquire one due to affinity
933 		 * restrictions loop again with this cpu removed from the
934 		 * set.
935 		 */
936 		if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
937 			tdq_unlock_pair(tdq, steal);
938 			continue;
939 		}
940 		spinlock_exit();
941 		TDQ_UNLOCK(steal);
942 		mi_switch(SW_VOL | SWT_IDLE, NULL);
943 		thread_unlock(curthread);
944 
945 		return (0);
946 	}
947 	spinlock_exit();
948 	return (1);
949 }
950 
951 /*
952  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
953  */
954 static void
955 tdq_notify(struct tdq *tdq, struct thread *td)
956 {
957 	struct thread *ctd;
958 	int pri;
959 	int cpu;
960 
961 	if (tdq->tdq_ipipending)
962 		return;
963 	cpu = td->td_sched->ts_cpu;
964 	pri = td->td_priority;
965 	ctd = pcpu_find(cpu)->pc_curthread;
966 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
967 		return;
968 	if (TD_IS_IDLETHREAD(ctd)) {
969 		/*
970 		 * If the MD code has an idle wakeup routine try that before
971 		 * falling back to IPI.
972 		 */
973 		if (cpu_idle_wakeup(cpu))
974 			return;
975 	}
976 	tdq->tdq_ipipending = 1;
977 	ipi_selected(1 << cpu, IPI_PREEMPT);
978 }
979 
980 /*
981  * Steals load from a timeshare queue.  Honors the rotating queue head
982  * index.
983  */
984 static struct thread *
985 runq_steal_from(struct runq *rq, int cpu, u_char start)
986 {
987 	struct rqbits *rqb;
988 	struct rqhead *rqh;
989 	struct thread *td;
990 	int first;
991 	int bit;
992 	int pri;
993 	int i;
994 
995 	rqb = &rq->rq_status;
996 	bit = start & (RQB_BPW -1);
997 	pri = 0;
998 	first = 0;
999 again:
1000 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1001 		if (rqb->rqb_bits[i] == 0)
1002 			continue;
1003 		if (bit != 0) {
1004 			for (pri = bit; pri < RQB_BPW; pri++)
1005 				if (rqb->rqb_bits[i] & (1ul << pri))
1006 					break;
1007 			if (pri >= RQB_BPW)
1008 				continue;
1009 		} else
1010 			pri = RQB_FFS(rqb->rqb_bits[i]);
1011 		pri += (i << RQB_L2BPW);
1012 		rqh = &rq->rq_queues[pri];
1013 		TAILQ_FOREACH(td, rqh, td_runq) {
1014 			if (first && THREAD_CAN_MIGRATE(td) &&
1015 			    THREAD_CAN_SCHED(td, cpu))
1016 				return (td);
1017 			first = 1;
1018 		}
1019 	}
1020 	if (start != 0) {
1021 		start = 0;
1022 		goto again;
1023 	}
1024 
1025 	return (NULL);
1026 }
1027 
1028 /*
1029  * Steals load from a standard linear queue.
1030  */
1031 static struct thread *
1032 runq_steal(struct runq *rq, int cpu)
1033 {
1034 	struct rqhead *rqh;
1035 	struct rqbits *rqb;
1036 	struct thread *td;
1037 	int word;
1038 	int bit;
1039 
1040 	rqb = &rq->rq_status;
1041 	for (word = 0; word < RQB_LEN; word++) {
1042 		if (rqb->rqb_bits[word] == 0)
1043 			continue;
1044 		for (bit = 0; bit < RQB_BPW; bit++) {
1045 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1046 				continue;
1047 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1048 			TAILQ_FOREACH(td, rqh, td_runq)
1049 				if (THREAD_CAN_MIGRATE(td) &&
1050 				    THREAD_CAN_SCHED(td, cpu))
1051 					return (td);
1052 		}
1053 	}
1054 	return (NULL);
1055 }
1056 
1057 /*
1058  * Attempt to steal a thread in priority order from a thread queue.
1059  */
1060 static struct thread *
1061 tdq_steal(struct tdq *tdq, int cpu)
1062 {
1063 	struct thread *td;
1064 
1065 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1066 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1067 		return (td);
1068 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1069 	    cpu, tdq->tdq_ridx)) != NULL)
1070 		return (td);
1071 	return (runq_steal(&tdq->tdq_idle, cpu));
1072 }
1073 
1074 /*
1075  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1076  * current lock and returns with the assigned queue locked.
1077  */
1078 static inline struct tdq *
1079 sched_setcpu(struct thread *td, int cpu, int flags)
1080 {
1081 
1082 	struct tdq *tdq;
1083 
1084 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1085 	tdq = TDQ_CPU(cpu);
1086 	td->td_sched->ts_cpu = cpu;
1087 	/*
1088 	 * If the lock matches just return the queue.
1089 	 */
1090 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1091 		return (tdq);
1092 #ifdef notyet
1093 	/*
1094 	 * If the thread isn't running its lockptr is a
1095 	 * turnstile or a sleepqueue.  We can just lock_set without
1096 	 * blocking.
1097 	 */
1098 	if (TD_CAN_RUN(td)) {
1099 		TDQ_LOCK(tdq);
1100 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1101 		return (tdq);
1102 	}
1103 #endif
1104 	/*
1105 	 * The hard case, migration, we need to block the thread first to
1106 	 * prevent order reversals with other cpus locks.
1107 	 */
1108 	thread_lock_block(td);
1109 	TDQ_LOCK(tdq);
1110 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1111 	return (tdq);
1112 }
1113 
1114 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1115 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1116 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1117 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1118 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1119 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1120 
1121 static int
1122 sched_pickcpu(struct thread *td, int flags)
1123 {
1124 	struct cpu_group *cg;
1125 	struct td_sched *ts;
1126 	struct tdq *tdq;
1127 	cpuset_t mask;
1128 	int self;
1129 	int pri;
1130 	int cpu;
1131 
1132 	self = PCPU_GET(cpuid);
1133 	ts = td->td_sched;
1134 	if (smp_started == 0)
1135 		return (self);
1136 	/*
1137 	 * Don't migrate a running thread from sched_switch().
1138 	 */
1139 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1140 		return (ts->ts_cpu);
1141 	/*
1142 	 * Prefer to run interrupt threads on the processors that generate
1143 	 * the interrupt.
1144 	 */
1145 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1146 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1147 		SCHED_STAT_INC(pickcpu_intrbind);
1148 		ts->ts_cpu = self;
1149 	}
1150 	/*
1151 	 * If the thread can run on the last cpu and the affinity has not
1152 	 * expired or it is idle run it there.
1153 	 */
1154 	pri = td->td_priority;
1155 	tdq = TDQ_CPU(ts->ts_cpu);
1156 	if (THREAD_CAN_SCHED(td, ts->ts_cpu)) {
1157 		if (tdq->tdq_lowpri > PRI_MIN_IDLE) {
1158 			SCHED_STAT_INC(pickcpu_idle_affinity);
1159 			return (ts->ts_cpu);
1160 		}
1161 		if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri) {
1162 			SCHED_STAT_INC(pickcpu_affinity);
1163 			return (ts->ts_cpu);
1164 		}
1165 	}
1166 	/*
1167 	 * Search for the highest level in the tree that still has affinity.
1168 	 */
1169 	cg = NULL;
1170 	for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent)
1171 		if (SCHED_AFFINITY(ts, cg->cg_level))
1172 			break;
1173 	cpu = -1;
1174 	mask = td->td_cpuset->cs_mask;
1175 	if (cg)
1176 		cpu = sched_lowest(cg, mask, pri);
1177 	if (cpu == -1)
1178 		cpu = sched_lowest(cpu_top, mask, -1);
1179 	/*
1180 	 * Compare the lowest loaded cpu to current cpu.
1181 	 */
1182 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1183 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) {
1184 		SCHED_STAT_INC(pickcpu_local);
1185 		cpu = self;
1186 	} else
1187 		SCHED_STAT_INC(pickcpu_lowest);
1188 	if (cpu != ts->ts_cpu)
1189 		SCHED_STAT_INC(pickcpu_migration);
1190 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1191 	return (cpu);
1192 }
1193 #endif
1194 
1195 /*
1196  * Pick the highest priority task we have and return it.
1197  */
1198 static struct thread *
1199 tdq_choose(struct tdq *tdq)
1200 {
1201 	struct thread *td;
1202 
1203 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1204 	td = runq_choose(&tdq->tdq_realtime);
1205 	if (td != NULL)
1206 		return (td);
1207 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1208 	if (td != NULL) {
1209 		KASSERT(td->td_priority >= PRI_MIN_TIMESHARE,
1210 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1211 		    td->td_priority));
1212 		return (td);
1213 	}
1214 	td = runq_choose(&tdq->tdq_idle);
1215 	if (td != NULL) {
1216 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1217 		    ("tdq_choose: Invalid priority on idle queue %d",
1218 		    td->td_priority));
1219 		return (td);
1220 	}
1221 
1222 	return (NULL);
1223 }
1224 
1225 /*
1226  * Initialize a thread queue.
1227  */
1228 static void
1229 tdq_setup(struct tdq *tdq)
1230 {
1231 
1232 	if (bootverbose)
1233 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1234 	runq_init(&tdq->tdq_realtime);
1235 	runq_init(&tdq->tdq_timeshare);
1236 	runq_init(&tdq->tdq_idle);
1237 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1238 	    "sched lock %d", (int)TDQ_ID(tdq));
1239 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1240 	    MTX_SPIN | MTX_RECURSE);
1241 #ifdef KTR
1242 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1243 	    "CPU %d load", (int)TDQ_ID(tdq));
1244 #endif
1245 }
1246 
1247 #ifdef SMP
1248 static void
1249 sched_setup_smp(void)
1250 {
1251 	struct tdq *tdq;
1252 	int i;
1253 
1254 	cpu_top = smp_topo();
1255 	for (i = 0; i < MAXCPU; i++) {
1256 		if (CPU_ABSENT(i))
1257 			continue;
1258 		tdq = TDQ_CPU(i);
1259 		tdq_setup(tdq);
1260 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1261 		if (tdq->tdq_cg == NULL)
1262 			panic("Can't find cpu group for %d\n", i);
1263 	}
1264 	balance_tdq = TDQ_SELF();
1265 	sched_balance();
1266 }
1267 #endif
1268 
1269 /*
1270  * Setup the thread queues and initialize the topology based on MD
1271  * information.
1272  */
1273 static void
1274 sched_setup(void *dummy)
1275 {
1276 	struct tdq *tdq;
1277 
1278 	tdq = TDQ_SELF();
1279 #ifdef SMP
1280 	sched_setup_smp();
1281 #else
1282 	tdq_setup(tdq);
1283 #endif
1284 	/*
1285 	 * To avoid divide-by-zero, we set realstathz a dummy value
1286 	 * in case which sched_clock() called before sched_initticks().
1287 	 */
1288 	realstathz = hz;
1289 	sched_slice = (realstathz/10);	/* ~100ms */
1290 	tickincr = 1 << SCHED_TICK_SHIFT;
1291 
1292 	/* Add thread0's load since it's running. */
1293 	TDQ_LOCK(tdq);
1294 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1295 	tdq_load_add(tdq, &thread0);
1296 	tdq->tdq_lowpri = thread0.td_priority;
1297 	TDQ_UNLOCK(tdq);
1298 }
1299 
1300 /*
1301  * This routine determines the tickincr after stathz and hz are setup.
1302  */
1303 /* ARGSUSED */
1304 static void
1305 sched_initticks(void *dummy)
1306 {
1307 	int incr;
1308 
1309 	realstathz = stathz ? stathz : hz;
1310 	sched_slice = (realstathz/10);	/* ~100ms */
1311 
1312 	/*
1313 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1314 	 * hz not being evenly divisible by stathz on all platforms.
1315 	 */
1316 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1317 	/*
1318 	 * This does not work for values of stathz that are more than
1319 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1320 	 */
1321 	if (incr == 0)
1322 		incr = 1;
1323 	tickincr = incr;
1324 #ifdef SMP
1325 	/*
1326 	 * Set the default balance interval now that we know
1327 	 * what realstathz is.
1328 	 */
1329 	balance_interval = realstathz;
1330 	/*
1331 	 * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4.
1332 	 * This prevents excess thrashing on large machines and excess idle
1333 	 * on smaller machines.
1334 	 */
1335 	steal_thresh = min(fls(mp_ncpus) - 1, 3);
1336 	affinity = SCHED_AFFINITY_DEFAULT;
1337 #endif
1338 }
1339 
1340 
1341 /*
1342  * This is the core of the interactivity algorithm.  Determines a score based
1343  * on past behavior.  It is the ratio of sleep time to run time scaled to
1344  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1345  * differs from the cpu usage because it does not account for time spent
1346  * waiting on a run-queue.  Would be prettier if we had floating point.
1347  */
1348 static int
1349 sched_interact_score(struct thread *td)
1350 {
1351 	struct td_sched *ts;
1352 	int div;
1353 
1354 	ts = td->td_sched;
1355 	/*
1356 	 * The score is only needed if this is likely to be an interactive
1357 	 * task.  Don't go through the expense of computing it if there's
1358 	 * no chance.
1359 	 */
1360 	if (sched_interact <= SCHED_INTERACT_HALF &&
1361 		ts->ts_runtime >= ts->ts_slptime)
1362 			return (SCHED_INTERACT_HALF);
1363 
1364 	if (ts->ts_runtime > ts->ts_slptime) {
1365 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1366 		return (SCHED_INTERACT_HALF +
1367 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1368 	}
1369 	if (ts->ts_slptime > ts->ts_runtime) {
1370 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1371 		return (ts->ts_runtime / div);
1372 	}
1373 	/* runtime == slptime */
1374 	if (ts->ts_runtime)
1375 		return (SCHED_INTERACT_HALF);
1376 
1377 	/*
1378 	 * This can happen if slptime and runtime are 0.
1379 	 */
1380 	return (0);
1381 
1382 }
1383 
1384 /*
1385  * Scale the scheduling priority according to the "interactivity" of this
1386  * process.
1387  */
1388 static void
1389 sched_priority(struct thread *td)
1390 {
1391 	int score;
1392 	int pri;
1393 
1394 	if (td->td_pri_class != PRI_TIMESHARE)
1395 		return;
1396 	/*
1397 	 * If the score is interactive we place the thread in the realtime
1398 	 * queue with a priority that is less than kernel and interrupt
1399 	 * priorities.  These threads are not subject to nice restrictions.
1400 	 *
1401 	 * Scores greater than this are placed on the normal timeshare queue
1402 	 * where the priority is partially decided by the most recent cpu
1403 	 * utilization and the rest is decided by nice value.
1404 	 *
1405 	 * The nice value of the process has a linear effect on the calculated
1406 	 * score.  Negative nice values make it easier for a thread to be
1407 	 * considered interactive.
1408 	 */
1409 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1410 	if (score < sched_interact) {
1411 		pri = PRI_MIN_REALTIME;
1412 		pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact)
1413 		    * score;
1414 		KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME,
1415 		    ("sched_priority: invalid interactive priority %d score %d",
1416 		    pri, score));
1417 	} else {
1418 		pri = SCHED_PRI_MIN;
1419 		if (td->td_sched->ts_ticks)
1420 			pri += SCHED_PRI_TICKS(td->td_sched);
1421 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1422 		KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE,
1423 		    ("sched_priority: invalid priority %d: nice %d, "
1424 		    "ticks %d ftick %d ltick %d tick pri %d",
1425 		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1426 		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1427 		    SCHED_PRI_TICKS(td->td_sched)));
1428 	}
1429 	sched_user_prio(td, pri);
1430 
1431 	return;
1432 }
1433 
1434 /*
1435  * This routine enforces a maximum limit on the amount of scheduling history
1436  * kept.  It is called after either the slptime or runtime is adjusted.  This
1437  * function is ugly due to integer math.
1438  */
1439 static void
1440 sched_interact_update(struct thread *td)
1441 {
1442 	struct td_sched *ts;
1443 	u_int sum;
1444 
1445 	ts = td->td_sched;
1446 	sum = ts->ts_runtime + ts->ts_slptime;
1447 	if (sum < SCHED_SLP_RUN_MAX)
1448 		return;
1449 	/*
1450 	 * This only happens from two places:
1451 	 * 1) We have added an unusual amount of run time from fork_exit.
1452 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1453 	 */
1454 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1455 		if (ts->ts_runtime > ts->ts_slptime) {
1456 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1457 			ts->ts_slptime = 1;
1458 		} else {
1459 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1460 			ts->ts_runtime = 1;
1461 		}
1462 		return;
1463 	}
1464 	/*
1465 	 * If we have exceeded by more than 1/5th then the algorithm below
1466 	 * will not bring us back into range.  Dividing by two here forces
1467 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1468 	 */
1469 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1470 		ts->ts_runtime /= 2;
1471 		ts->ts_slptime /= 2;
1472 		return;
1473 	}
1474 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1475 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1476 }
1477 
1478 /*
1479  * Scale back the interactivity history when a child thread is created.  The
1480  * history is inherited from the parent but the thread may behave totally
1481  * differently.  For example, a shell spawning a compiler process.  We want
1482  * to learn that the compiler is behaving badly very quickly.
1483  */
1484 static void
1485 sched_interact_fork(struct thread *td)
1486 {
1487 	int ratio;
1488 	int sum;
1489 
1490 	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1491 	if (sum > SCHED_SLP_RUN_FORK) {
1492 		ratio = sum / SCHED_SLP_RUN_FORK;
1493 		td->td_sched->ts_runtime /= ratio;
1494 		td->td_sched->ts_slptime /= ratio;
1495 	}
1496 }
1497 
1498 /*
1499  * Called from proc0_init() to setup the scheduler fields.
1500  */
1501 void
1502 schedinit(void)
1503 {
1504 
1505 	/*
1506 	 * Set up the scheduler specific parts of proc0.
1507 	 */
1508 	proc0.p_sched = NULL; /* XXX */
1509 	thread0.td_sched = &td_sched0;
1510 	td_sched0.ts_ltick = ticks;
1511 	td_sched0.ts_ftick = ticks;
1512 	td_sched0.ts_slice = sched_slice;
1513 }
1514 
1515 /*
1516  * This is only somewhat accurate since given many processes of the same
1517  * priority they will switch when their slices run out, which will be
1518  * at most sched_slice stathz ticks.
1519  */
1520 int
1521 sched_rr_interval(void)
1522 {
1523 
1524 	/* Convert sched_slice to hz */
1525 	return (hz/(realstathz/sched_slice));
1526 }
1527 
1528 /*
1529  * Update the percent cpu tracking information when it is requested or
1530  * the total history exceeds the maximum.  We keep a sliding history of
1531  * tick counts that slowly decays.  This is less precise than the 4BSD
1532  * mechanism since it happens with less regular and frequent events.
1533  */
1534 static void
1535 sched_pctcpu_update(struct td_sched *ts)
1536 {
1537 
1538 	if (ts->ts_ticks == 0)
1539 		return;
1540 	if (ticks - (hz / 10) < ts->ts_ltick &&
1541 	    SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX)
1542 		return;
1543 	/*
1544 	 * Adjust counters and watermark for pctcpu calc.
1545 	 */
1546 	if (ts->ts_ltick > ticks - SCHED_TICK_TARG)
1547 		ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) *
1548 			    SCHED_TICK_TARG;
1549 	else
1550 		ts->ts_ticks = 0;
1551 	ts->ts_ltick = ticks;
1552 	ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG;
1553 }
1554 
1555 /*
1556  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1557  * if necessary.  This is the back-end for several priority related
1558  * functions.
1559  */
1560 static void
1561 sched_thread_priority(struct thread *td, u_char prio)
1562 {
1563 	struct td_sched *ts;
1564 	struct tdq *tdq;
1565 	int oldpri;
1566 
1567 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1568 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1569 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1570 	if (td != curthread && prio > td->td_priority) {
1571 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1572 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1573 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1574 	}
1575 	ts = td->td_sched;
1576 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1577 	if (td->td_priority == prio)
1578 		return;
1579 	/*
1580 	 * If the priority has been elevated due to priority
1581 	 * propagation, we may have to move ourselves to a new
1582 	 * queue.  This could be optimized to not re-add in some
1583 	 * cases.
1584 	 */
1585 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1586 		sched_rem(td);
1587 		td->td_priority = prio;
1588 		sched_add(td, SRQ_BORROWING);
1589 		return;
1590 	}
1591 	/*
1592 	 * If the thread is currently running we may have to adjust the lowpri
1593 	 * information so other cpus are aware of our current priority.
1594 	 */
1595 	if (TD_IS_RUNNING(td)) {
1596 		tdq = TDQ_CPU(ts->ts_cpu);
1597 		oldpri = td->td_priority;
1598 		td->td_priority = prio;
1599 		if (prio < tdq->tdq_lowpri)
1600 			tdq->tdq_lowpri = prio;
1601 		else if (tdq->tdq_lowpri == oldpri)
1602 			tdq_setlowpri(tdq, td);
1603 		return;
1604 	}
1605 	td->td_priority = prio;
1606 }
1607 
1608 /*
1609  * Update a thread's priority when it is lent another thread's
1610  * priority.
1611  */
1612 void
1613 sched_lend_prio(struct thread *td, u_char prio)
1614 {
1615 
1616 	td->td_flags |= TDF_BORROWING;
1617 	sched_thread_priority(td, prio);
1618 }
1619 
1620 /*
1621  * Restore a thread's priority when priority propagation is
1622  * over.  The prio argument is the minimum priority the thread
1623  * needs to have to satisfy other possible priority lending
1624  * requests.  If the thread's regular priority is less
1625  * important than prio, the thread will keep a priority boost
1626  * of prio.
1627  */
1628 void
1629 sched_unlend_prio(struct thread *td, u_char prio)
1630 {
1631 	u_char base_pri;
1632 
1633 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1634 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1635 		base_pri = td->td_user_pri;
1636 	else
1637 		base_pri = td->td_base_pri;
1638 	if (prio >= base_pri) {
1639 		td->td_flags &= ~TDF_BORROWING;
1640 		sched_thread_priority(td, base_pri);
1641 	} else
1642 		sched_lend_prio(td, prio);
1643 }
1644 
1645 /*
1646  * Standard entry for setting the priority to an absolute value.
1647  */
1648 void
1649 sched_prio(struct thread *td, u_char prio)
1650 {
1651 	u_char oldprio;
1652 
1653 	/* First, update the base priority. */
1654 	td->td_base_pri = prio;
1655 
1656 	/*
1657 	 * If the thread is borrowing another thread's priority, don't
1658 	 * ever lower the priority.
1659 	 */
1660 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1661 		return;
1662 
1663 	/* Change the real priority. */
1664 	oldprio = td->td_priority;
1665 	sched_thread_priority(td, prio);
1666 
1667 	/*
1668 	 * If the thread is on a turnstile, then let the turnstile update
1669 	 * its state.
1670 	 */
1671 	if (TD_ON_LOCK(td) && oldprio != prio)
1672 		turnstile_adjust(td, oldprio);
1673 }
1674 
1675 /*
1676  * Set the base user priority, does not effect current running priority.
1677  */
1678 void
1679 sched_user_prio(struct thread *td, u_char prio)
1680 {
1681 	u_char oldprio;
1682 
1683 	td->td_base_user_pri = prio;
1684 	if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
1685                 return;
1686 	oldprio = td->td_user_pri;
1687 	td->td_user_pri = prio;
1688 }
1689 
1690 void
1691 sched_lend_user_prio(struct thread *td, u_char prio)
1692 {
1693 	u_char oldprio;
1694 
1695 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1696 	td->td_flags |= TDF_UBORROWING;
1697 	oldprio = td->td_user_pri;
1698 	td->td_user_pri = prio;
1699 }
1700 
1701 void
1702 sched_unlend_user_prio(struct thread *td, u_char prio)
1703 {
1704 	u_char base_pri;
1705 
1706 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1707 	base_pri = td->td_base_user_pri;
1708 	if (prio >= base_pri) {
1709 		td->td_flags &= ~TDF_UBORROWING;
1710 		sched_user_prio(td, base_pri);
1711 	} else {
1712 		sched_lend_user_prio(td, prio);
1713 	}
1714 }
1715 
1716 /*
1717  * Block a thread for switching.  Similar to thread_block() but does not
1718  * bump the spin count.
1719  */
1720 static inline struct mtx *
1721 thread_block_switch(struct thread *td)
1722 {
1723 	struct mtx *lock;
1724 
1725 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1726 	lock = td->td_lock;
1727 	td->td_lock = &blocked_lock;
1728 	mtx_unlock_spin(lock);
1729 
1730 	return (lock);
1731 }
1732 
1733 /*
1734  * Handle migration from sched_switch().  This happens only for
1735  * cpu binding.
1736  */
1737 static struct mtx *
1738 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1739 {
1740 	struct tdq *tdn;
1741 
1742 	tdn = TDQ_CPU(td->td_sched->ts_cpu);
1743 #ifdef SMP
1744 	tdq_load_rem(tdq, td);
1745 	/*
1746 	 * Do the lock dance required to avoid LOR.  We grab an extra
1747 	 * spinlock nesting to prevent preemption while we're
1748 	 * not holding either run-queue lock.
1749 	 */
1750 	spinlock_enter();
1751 	thread_block_switch(td);	/* This releases the lock on tdq. */
1752 
1753 	/*
1754 	 * Acquire both run-queue locks before placing the thread on the new
1755 	 * run-queue to avoid deadlocks created by placing a thread with a
1756 	 * blocked lock on the run-queue of a remote processor.  The deadlock
1757 	 * occurs when a third processor attempts to lock the two queues in
1758 	 * question while the target processor is spinning with its own
1759 	 * run-queue lock held while waiting for the blocked lock to clear.
1760 	 */
1761 	tdq_lock_pair(tdn, tdq);
1762 	tdq_add(tdn, td, flags);
1763 	tdq_notify(tdn, td);
1764 	TDQ_UNLOCK(tdn);
1765 	spinlock_exit();
1766 #endif
1767 	return (TDQ_LOCKPTR(tdn));
1768 }
1769 
1770 /*
1771  * Release a thread that was blocked with thread_block_switch().
1772  */
1773 static inline void
1774 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1775 {
1776 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1777 	    (uintptr_t)mtx);
1778 }
1779 
1780 /*
1781  * Switch threads.  This function has to handle threads coming in while
1782  * blocked for some reason, running, or idle.  It also must deal with
1783  * migrating a thread from one queue to another as running threads may
1784  * be assigned elsewhere via binding.
1785  */
1786 void
1787 sched_switch(struct thread *td, struct thread *newtd, int flags)
1788 {
1789 	struct tdq *tdq;
1790 	struct td_sched *ts;
1791 	struct mtx *mtx;
1792 	int srqflag;
1793 	int cpuid;
1794 
1795 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1796 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1797 
1798 	cpuid = PCPU_GET(cpuid);
1799 	tdq = TDQ_CPU(cpuid);
1800 	ts = td->td_sched;
1801 	mtx = td->td_lock;
1802 	ts->ts_rltick = ticks;
1803 	td->td_lastcpu = td->td_oncpu;
1804 	td->td_oncpu = NOCPU;
1805 	td->td_flags &= ~TDF_NEEDRESCHED;
1806 	td->td_owepreempt = 0;
1807 	tdq->tdq_switchcnt++;
1808 	/*
1809 	 * The lock pointer in an idle thread should never change.  Reset it
1810 	 * to CAN_RUN as well.
1811 	 */
1812 	if (TD_IS_IDLETHREAD(td)) {
1813 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1814 		TD_SET_CAN_RUN(td);
1815 	} else if (TD_IS_RUNNING(td)) {
1816 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1817 		srqflag = (flags & SW_PREEMPT) ?
1818 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1819 		    SRQ_OURSELF|SRQ_YIELDING;
1820 		if (ts->ts_cpu == cpuid)
1821 			tdq_runq_add(tdq, td, srqflag);
1822 		else
1823 			mtx = sched_switch_migrate(tdq, td, srqflag);
1824 	} else {
1825 		/* This thread must be going to sleep. */
1826 		TDQ_LOCK(tdq);
1827 		mtx = thread_block_switch(td);
1828 		tdq_load_rem(tdq, td);
1829 	}
1830 	/*
1831 	 * We enter here with the thread blocked and assigned to the
1832 	 * appropriate cpu run-queue or sleep-queue and with the current
1833 	 * thread-queue locked.
1834 	 */
1835 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1836 	newtd = choosethread();
1837 	/*
1838 	 * Call the MD code to switch contexts if necessary.
1839 	 */
1840 	if (td != newtd) {
1841 #ifdef	HWPMC_HOOKS
1842 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1843 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1844 #endif
1845 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1846 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1847 
1848 #ifdef KDTRACE_HOOKS
1849 		/*
1850 		 * If DTrace has set the active vtime enum to anything
1851 		 * other than INACTIVE (0), then it should have set the
1852 		 * function to call.
1853 		 */
1854 		if (dtrace_vtime_active)
1855 			(*dtrace_vtime_switch_func)(newtd);
1856 #endif
1857 
1858 		cpu_switch(td, newtd, mtx);
1859 		/*
1860 		 * We may return from cpu_switch on a different cpu.  However,
1861 		 * we always return with td_lock pointing to the current cpu's
1862 		 * run queue lock.
1863 		 */
1864 		cpuid = PCPU_GET(cpuid);
1865 		tdq = TDQ_CPU(cpuid);
1866 		lock_profile_obtain_lock_success(
1867 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1868 #ifdef	HWPMC_HOOKS
1869 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1870 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1871 #endif
1872 	} else
1873 		thread_unblock_switch(td, mtx);
1874 	/*
1875 	 * Assert that all went well and return.
1876 	 */
1877 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1878 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1879 	td->td_oncpu = cpuid;
1880 }
1881 
1882 /*
1883  * Adjust thread priorities as a result of a nice request.
1884  */
1885 void
1886 sched_nice(struct proc *p, int nice)
1887 {
1888 	struct thread *td;
1889 
1890 	PROC_LOCK_ASSERT(p, MA_OWNED);
1891 
1892 	p->p_nice = nice;
1893 	FOREACH_THREAD_IN_PROC(p, td) {
1894 		thread_lock(td);
1895 		sched_priority(td);
1896 		sched_prio(td, td->td_base_user_pri);
1897 		thread_unlock(td);
1898 	}
1899 }
1900 
1901 /*
1902  * Record the sleep time for the interactivity scorer.
1903  */
1904 void
1905 sched_sleep(struct thread *td, int prio)
1906 {
1907 
1908 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1909 
1910 	td->td_slptick = ticks;
1911 	if (TD_IS_SUSPENDED(td) || prio <= PSOCK)
1912 		td->td_flags |= TDF_CANSWAP;
1913 	if (static_boost == 1 && prio)
1914 		sched_prio(td, prio);
1915 	else if (static_boost && td->td_priority > static_boost)
1916 		sched_prio(td, static_boost);
1917 }
1918 
1919 /*
1920  * Schedule a thread to resume execution and record how long it voluntarily
1921  * slept.  We also update the pctcpu, interactivity, and priority.
1922  */
1923 void
1924 sched_wakeup(struct thread *td)
1925 {
1926 	struct td_sched *ts;
1927 	int slptick;
1928 
1929 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1930 	ts = td->td_sched;
1931 	td->td_flags &= ~TDF_CANSWAP;
1932 	/*
1933 	 * If we slept for more than a tick update our interactivity and
1934 	 * priority.
1935 	 */
1936 	slptick = td->td_slptick;
1937 	td->td_slptick = 0;
1938 	if (slptick && slptick != ticks) {
1939 		u_int hzticks;
1940 
1941 		hzticks = (ticks - slptick) << SCHED_TICK_SHIFT;
1942 		ts->ts_slptime += hzticks;
1943 		sched_interact_update(td);
1944 		sched_pctcpu_update(ts);
1945 	}
1946 	/* Reset the slice value after we sleep. */
1947 	ts->ts_slice = sched_slice;
1948 	sched_add(td, SRQ_BORING);
1949 }
1950 
1951 /*
1952  * Penalize the parent for creating a new child and initialize the child's
1953  * priority.
1954  */
1955 void
1956 sched_fork(struct thread *td, struct thread *child)
1957 {
1958 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1959 	sched_fork_thread(td, child);
1960 	/*
1961 	 * Penalize the parent and child for forking.
1962 	 */
1963 	sched_interact_fork(child);
1964 	sched_priority(child);
1965 	td->td_sched->ts_runtime += tickincr;
1966 	sched_interact_update(td);
1967 	sched_priority(td);
1968 }
1969 
1970 /*
1971  * Fork a new thread, may be within the same process.
1972  */
1973 void
1974 sched_fork_thread(struct thread *td, struct thread *child)
1975 {
1976 	struct td_sched *ts;
1977 	struct td_sched *ts2;
1978 
1979 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1980 	/*
1981 	 * Initialize child.
1982 	 */
1983 	ts = td->td_sched;
1984 	ts2 = child->td_sched;
1985 	child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
1986 	child->td_cpuset = cpuset_ref(td->td_cpuset);
1987 	ts2->ts_cpu = ts->ts_cpu;
1988 	ts2->ts_flags = 0;
1989 	/*
1990 	 * Grab our parents cpu estimation information and priority.
1991 	 */
1992 	ts2->ts_ticks = ts->ts_ticks;
1993 	ts2->ts_ltick = ts->ts_ltick;
1994 	ts2->ts_ftick = ts->ts_ftick;
1995 	child->td_user_pri = td->td_user_pri;
1996 	child->td_base_user_pri = td->td_base_user_pri;
1997 	/*
1998 	 * And update interactivity score.
1999 	 */
2000 	ts2->ts_slptime = ts->ts_slptime;
2001 	ts2->ts_runtime = ts->ts_runtime;
2002 	ts2->ts_slice = 1;	/* Attempt to quickly learn interactivity. */
2003 #ifdef KTR
2004 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2005 #endif
2006 }
2007 
2008 /*
2009  * Adjust the priority class of a thread.
2010  */
2011 void
2012 sched_class(struct thread *td, int class)
2013 {
2014 
2015 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2016 	if (td->td_pri_class == class)
2017 		return;
2018 	td->td_pri_class = class;
2019 }
2020 
2021 /*
2022  * Return some of the child's priority and interactivity to the parent.
2023  */
2024 void
2025 sched_exit(struct proc *p, struct thread *child)
2026 {
2027 	struct thread *td;
2028 
2029 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2030 	    "prio:td", child->td_priority);
2031 	PROC_LOCK_ASSERT(p, MA_OWNED);
2032 	td = FIRST_THREAD_IN_PROC(p);
2033 	sched_exit_thread(td, child);
2034 }
2035 
2036 /*
2037  * Penalize another thread for the time spent on this one.  This helps to
2038  * worsen the priority and interactivity of processes which schedule batch
2039  * jobs such as make.  This has little effect on the make process itself but
2040  * causes new processes spawned by it to receive worse scores immediately.
2041  */
2042 void
2043 sched_exit_thread(struct thread *td, struct thread *child)
2044 {
2045 
2046 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2047 	    "prio:td", child->td_priority);
2048 	/*
2049 	 * Give the child's runtime to the parent without returning the
2050 	 * sleep time as a penalty to the parent.  This causes shells that
2051 	 * launch expensive things to mark their children as expensive.
2052 	 */
2053 	thread_lock(td);
2054 	td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2055 	sched_interact_update(td);
2056 	sched_priority(td);
2057 	thread_unlock(td);
2058 }
2059 
2060 void
2061 sched_preempt(struct thread *td)
2062 {
2063 	struct tdq *tdq;
2064 
2065 	thread_lock(td);
2066 	tdq = TDQ_SELF();
2067 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2068 	tdq->tdq_ipipending = 0;
2069 	if (td->td_priority > tdq->tdq_lowpri) {
2070 		int flags;
2071 
2072 		flags = SW_INVOL | SW_PREEMPT;
2073 		if (td->td_critnest > 1)
2074 			td->td_owepreempt = 1;
2075 		else if (TD_IS_IDLETHREAD(td))
2076 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2077 		else
2078 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2079 	}
2080 	thread_unlock(td);
2081 }
2082 
2083 /*
2084  * Fix priorities on return to user-space.  Priorities may be elevated due
2085  * to static priorities in msleep() or similar.
2086  */
2087 void
2088 sched_userret(struct thread *td)
2089 {
2090 	/*
2091 	 * XXX we cheat slightly on the locking here to avoid locking in
2092 	 * the usual case.  Setting td_priority here is essentially an
2093 	 * incomplete workaround for not setting it properly elsewhere.
2094 	 * Now that some interrupt handlers are threads, not setting it
2095 	 * properly elsewhere can clobber it in the window between setting
2096 	 * it here and returning to user mode, so don't waste time setting
2097 	 * it perfectly here.
2098 	 */
2099 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2100 	    ("thread with borrowed priority returning to userland"));
2101 	if (td->td_priority != td->td_user_pri) {
2102 		thread_lock(td);
2103 		td->td_priority = td->td_user_pri;
2104 		td->td_base_pri = td->td_user_pri;
2105 		tdq_setlowpri(TDQ_SELF(), td);
2106 		thread_unlock(td);
2107         }
2108 }
2109 
2110 /*
2111  * Handle a stathz tick.  This is really only relevant for timeshare
2112  * threads.
2113  */
2114 void
2115 sched_clock(struct thread *td)
2116 {
2117 	struct tdq *tdq;
2118 	struct td_sched *ts;
2119 
2120 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2121 	tdq = TDQ_SELF();
2122 #ifdef SMP
2123 	/*
2124 	 * We run the long term load balancer infrequently on the first cpu.
2125 	 */
2126 	if (balance_tdq == tdq) {
2127 		if (balance_ticks && --balance_ticks == 0)
2128 			sched_balance();
2129 	}
2130 #endif
2131 	/*
2132 	 * Save the old switch count so we have a record of the last ticks
2133 	 * activity.   Initialize the new switch count based on our load.
2134 	 * If there is some activity seed it to reflect that.
2135 	 */
2136 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2137 	tdq->tdq_switchcnt = tdq->tdq_load;
2138 	/*
2139 	 * Advance the insert index once for each tick to ensure that all
2140 	 * threads get a chance to run.
2141 	 */
2142 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2143 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2144 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2145 			tdq->tdq_ridx = tdq->tdq_idx;
2146 	}
2147 	ts = td->td_sched;
2148 	if (td->td_pri_class & PRI_FIFO_BIT)
2149 		return;
2150 	if (td->td_pri_class == PRI_TIMESHARE) {
2151 		/*
2152 		 * We used a tick; charge it to the thread so
2153 		 * that we can compute our interactivity.
2154 		 */
2155 		td->td_sched->ts_runtime += tickincr;
2156 		sched_interact_update(td);
2157 		sched_priority(td);
2158 	}
2159 	/*
2160 	 * We used up one time slice.
2161 	 */
2162 	if (--ts->ts_slice > 0)
2163 		return;
2164 	/*
2165 	 * We're out of time, force a requeue at userret().
2166 	 */
2167 	ts->ts_slice = sched_slice;
2168 	td->td_flags |= TDF_NEEDRESCHED;
2169 }
2170 
2171 /*
2172  * Called once per hz tick.  Used for cpu utilization information.  This
2173  * is easier than trying to scale based on stathz.
2174  */
2175 void
2176 sched_tick(void)
2177 {
2178 	struct td_sched *ts;
2179 
2180 	ts = curthread->td_sched;
2181 	/*
2182 	 * Ticks is updated asynchronously on a single cpu.  Check here to
2183 	 * avoid incrementing ts_ticks multiple times in a single tick.
2184 	 */
2185 	if (ts->ts_ltick == ticks)
2186 		return;
2187 	/* Adjust ticks for pctcpu */
2188 	ts->ts_ticks += 1 << SCHED_TICK_SHIFT;
2189 	ts->ts_ltick = ticks;
2190 	/*
2191 	 * Update if we've exceeded our desired tick threshhold by over one
2192 	 * second.
2193 	 */
2194 	if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick)
2195 		sched_pctcpu_update(ts);
2196 }
2197 
2198 /*
2199  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2200  * cooperative idle threads.
2201  */
2202 int
2203 sched_runnable(void)
2204 {
2205 	struct tdq *tdq;
2206 	int load;
2207 
2208 	load = 1;
2209 
2210 	tdq = TDQ_SELF();
2211 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2212 		if (tdq->tdq_load > 0)
2213 			goto out;
2214 	} else
2215 		if (tdq->tdq_load - 1 > 0)
2216 			goto out;
2217 	load = 0;
2218 out:
2219 	return (load);
2220 }
2221 
2222 /*
2223  * Choose the highest priority thread to run.  The thread is removed from
2224  * the run-queue while running however the load remains.  For SMP we set
2225  * the tdq in the global idle bitmask if it idles here.
2226  */
2227 struct thread *
2228 sched_choose(void)
2229 {
2230 	struct thread *td;
2231 	struct tdq *tdq;
2232 
2233 	tdq = TDQ_SELF();
2234 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2235 	td = tdq_choose(tdq);
2236 	if (td) {
2237 		td->td_sched->ts_ltick = ticks;
2238 		tdq_runq_rem(tdq, td);
2239 		tdq->tdq_lowpri = td->td_priority;
2240 		return (td);
2241 	}
2242 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2243 	return (PCPU_GET(idlethread));
2244 }
2245 
2246 /*
2247  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2248  * we always request it once we exit a critical section.
2249  */
2250 static inline void
2251 sched_setpreempt(struct thread *td)
2252 {
2253 	struct thread *ctd;
2254 	int cpri;
2255 	int pri;
2256 
2257 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2258 
2259 	ctd = curthread;
2260 	pri = td->td_priority;
2261 	cpri = ctd->td_priority;
2262 	if (pri < cpri)
2263 		ctd->td_flags |= TDF_NEEDRESCHED;
2264 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2265 		return;
2266 	if (!sched_shouldpreempt(pri, cpri, 0))
2267 		return;
2268 	ctd->td_owepreempt = 1;
2269 }
2270 
2271 /*
2272  * Add a thread to a thread queue.  Select the appropriate runq and add the
2273  * thread to it.  This is the internal function called when the tdq is
2274  * predetermined.
2275  */
2276 void
2277 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2278 {
2279 
2280 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2281 	KASSERT((td->td_inhibitors == 0),
2282 	    ("sched_add: trying to run inhibited thread"));
2283 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2284 	    ("sched_add: bad thread state"));
2285 	KASSERT(td->td_flags & TDF_INMEM,
2286 	    ("sched_add: thread swapped out"));
2287 
2288 	if (td->td_priority < tdq->tdq_lowpri)
2289 		tdq->tdq_lowpri = td->td_priority;
2290 	tdq_runq_add(tdq, td, flags);
2291 	tdq_load_add(tdq, td);
2292 }
2293 
2294 /*
2295  * Select the target thread queue and add a thread to it.  Request
2296  * preemption or IPI a remote processor if required.
2297  */
2298 void
2299 sched_add(struct thread *td, int flags)
2300 {
2301 	struct tdq *tdq;
2302 #ifdef SMP
2303 	int cpu;
2304 #endif
2305 
2306 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2307 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2308 	    sched_tdname(curthread));
2309 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2310 	    KTR_ATTR_LINKED, sched_tdname(td));
2311 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2312 	/*
2313 	 * Recalculate the priority before we select the target cpu or
2314 	 * run-queue.
2315 	 */
2316 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2317 		sched_priority(td);
2318 #ifdef SMP
2319 	/*
2320 	 * Pick the destination cpu and if it isn't ours transfer to the
2321 	 * target cpu.
2322 	 */
2323 	cpu = sched_pickcpu(td, flags);
2324 	tdq = sched_setcpu(td, cpu, flags);
2325 	tdq_add(tdq, td, flags);
2326 	if (cpu != PCPU_GET(cpuid)) {
2327 		tdq_notify(tdq, td);
2328 		return;
2329 	}
2330 #else
2331 	tdq = TDQ_SELF();
2332 	TDQ_LOCK(tdq);
2333 	/*
2334 	 * Now that the thread is moving to the run-queue, set the lock
2335 	 * to the scheduler's lock.
2336 	 */
2337 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2338 	tdq_add(tdq, td, flags);
2339 #endif
2340 	if (!(flags & SRQ_YIELDING))
2341 		sched_setpreempt(td);
2342 }
2343 
2344 /*
2345  * Remove a thread from a run-queue without running it.  This is used
2346  * when we're stealing a thread from a remote queue.  Otherwise all threads
2347  * exit by calling sched_exit_thread() and sched_throw() themselves.
2348  */
2349 void
2350 sched_rem(struct thread *td)
2351 {
2352 	struct tdq *tdq;
2353 
2354 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2355 	    "prio:%d", td->td_priority);
2356 	tdq = TDQ_CPU(td->td_sched->ts_cpu);
2357 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2358 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2359 	KASSERT(TD_ON_RUNQ(td),
2360 	    ("sched_rem: thread not on run queue"));
2361 	tdq_runq_rem(tdq, td);
2362 	tdq_load_rem(tdq, td);
2363 	TD_SET_CAN_RUN(td);
2364 	if (td->td_priority == tdq->tdq_lowpri)
2365 		tdq_setlowpri(tdq, NULL);
2366 }
2367 
2368 /*
2369  * Fetch cpu utilization information.  Updates on demand.
2370  */
2371 fixpt_t
2372 sched_pctcpu(struct thread *td)
2373 {
2374 	fixpt_t pctcpu;
2375 	struct td_sched *ts;
2376 
2377 	pctcpu = 0;
2378 	ts = td->td_sched;
2379 	if (ts == NULL)
2380 		return (0);
2381 
2382 	thread_lock(td);
2383 	if (ts->ts_ticks) {
2384 		int rtick;
2385 
2386 		sched_pctcpu_update(ts);
2387 		/* How many rtick per second ? */
2388 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2389 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2390 	}
2391 	thread_unlock(td);
2392 
2393 	return (pctcpu);
2394 }
2395 
2396 /*
2397  * Enforce affinity settings for a thread.  Called after adjustments to
2398  * cpumask.
2399  */
2400 void
2401 sched_affinity(struct thread *td)
2402 {
2403 #ifdef SMP
2404 	struct td_sched *ts;
2405 	int cpu;
2406 
2407 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2408 	ts = td->td_sched;
2409 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2410 		return;
2411 	if (TD_ON_RUNQ(td)) {
2412 		sched_rem(td);
2413 		sched_add(td, SRQ_BORING);
2414 		return;
2415 	}
2416 	if (!TD_IS_RUNNING(td))
2417 		return;
2418 	td->td_flags |= TDF_NEEDRESCHED;
2419 	if (!THREAD_CAN_MIGRATE(td))
2420 		return;
2421 	/*
2422 	 * Assign the new cpu and force a switch before returning to
2423 	 * userspace.  If the target thread is not running locally send
2424 	 * an ipi to force the issue.
2425 	 */
2426 	cpu = ts->ts_cpu;
2427 	ts->ts_cpu = sched_pickcpu(td, 0);
2428 	if (cpu != PCPU_GET(cpuid))
2429 		ipi_selected(1 << cpu, IPI_PREEMPT);
2430 #endif
2431 }
2432 
2433 /*
2434  * Bind a thread to a target cpu.
2435  */
2436 void
2437 sched_bind(struct thread *td, int cpu)
2438 {
2439 	struct td_sched *ts;
2440 
2441 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2442 	ts = td->td_sched;
2443 	if (ts->ts_flags & TSF_BOUND)
2444 		sched_unbind(td);
2445 	ts->ts_flags |= TSF_BOUND;
2446 	sched_pin();
2447 	if (PCPU_GET(cpuid) == cpu)
2448 		return;
2449 	ts->ts_cpu = cpu;
2450 	/* When we return from mi_switch we'll be on the correct cpu. */
2451 	mi_switch(SW_VOL, NULL);
2452 }
2453 
2454 /*
2455  * Release a bound thread.
2456  */
2457 void
2458 sched_unbind(struct thread *td)
2459 {
2460 	struct td_sched *ts;
2461 
2462 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2463 	ts = td->td_sched;
2464 	if ((ts->ts_flags & TSF_BOUND) == 0)
2465 		return;
2466 	ts->ts_flags &= ~TSF_BOUND;
2467 	sched_unpin();
2468 }
2469 
2470 int
2471 sched_is_bound(struct thread *td)
2472 {
2473 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2474 	return (td->td_sched->ts_flags & TSF_BOUND);
2475 }
2476 
2477 /*
2478  * Basic yield call.
2479  */
2480 void
2481 sched_relinquish(struct thread *td)
2482 {
2483 	thread_lock(td);
2484 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2485 	thread_unlock(td);
2486 }
2487 
2488 /*
2489  * Return the total system load.
2490  */
2491 int
2492 sched_load(void)
2493 {
2494 #ifdef SMP
2495 	int total;
2496 	int i;
2497 
2498 	total = 0;
2499 	for (i = 0; i <= mp_maxid; i++)
2500 		total += TDQ_CPU(i)->tdq_sysload;
2501 	return (total);
2502 #else
2503 	return (TDQ_SELF()->tdq_sysload);
2504 #endif
2505 }
2506 
2507 int
2508 sched_sizeof_proc(void)
2509 {
2510 	return (sizeof(struct proc));
2511 }
2512 
2513 int
2514 sched_sizeof_thread(void)
2515 {
2516 	return (sizeof(struct thread) + sizeof(struct td_sched));
2517 }
2518 
2519 #ifdef SMP
2520 #define	TDQ_IDLESPIN(tdq)						\
2521     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2522 #else
2523 #define	TDQ_IDLESPIN(tdq)	1
2524 #endif
2525 
2526 /*
2527  * The actual idle process.
2528  */
2529 void
2530 sched_idletd(void *dummy)
2531 {
2532 	struct thread *td;
2533 	struct tdq *tdq;
2534 	int switchcnt;
2535 	int i;
2536 
2537 	mtx_assert(&Giant, MA_NOTOWNED);
2538 	td = curthread;
2539 	tdq = TDQ_SELF();
2540 	for (;;) {
2541 #ifdef SMP
2542 		if (tdq_idled(tdq) == 0)
2543 			continue;
2544 #endif
2545 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2546 		/*
2547 		 * If we're switching very frequently, spin while checking
2548 		 * for load rather than entering a low power state that
2549 		 * may require an IPI.  However, don't do any busy
2550 		 * loops while on SMT machines as this simply steals
2551 		 * cycles from cores doing useful work.
2552 		 */
2553 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2554 			for (i = 0; i < sched_idlespins; i++) {
2555 				if (tdq->tdq_load)
2556 					break;
2557 				cpu_spinwait();
2558 			}
2559 		}
2560 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2561 		if (tdq->tdq_load == 0)
2562 			cpu_idle(switchcnt > 1);
2563 		if (tdq->tdq_load) {
2564 			thread_lock(td);
2565 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2566 			thread_unlock(td);
2567 		}
2568 	}
2569 }
2570 
2571 /*
2572  * A CPU is entering for the first time or a thread is exiting.
2573  */
2574 void
2575 sched_throw(struct thread *td)
2576 {
2577 	struct thread *newtd;
2578 	struct tdq *tdq;
2579 
2580 	tdq = TDQ_SELF();
2581 	if (td == NULL) {
2582 		/* Correct spinlock nesting and acquire the correct lock. */
2583 		TDQ_LOCK(tdq);
2584 		spinlock_exit();
2585 	} else {
2586 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2587 		tdq_load_rem(tdq, td);
2588 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2589 	}
2590 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2591 	newtd = choosethread();
2592 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2593 	PCPU_SET(switchtime, cpu_ticks());
2594 	PCPU_SET(switchticks, ticks);
2595 	cpu_throw(td, newtd);		/* doesn't return */
2596 }
2597 
2598 /*
2599  * This is called from fork_exit().  Just acquire the correct locks and
2600  * let fork do the rest of the work.
2601  */
2602 void
2603 sched_fork_exit(struct thread *td)
2604 {
2605 	struct td_sched *ts;
2606 	struct tdq *tdq;
2607 	int cpuid;
2608 
2609 	/*
2610 	 * Finish setting up thread glue so that it begins execution in a
2611 	 * non-nested critical section with the scheduler lock held.
2612 	 */
2613 	cpuid = PCPU_GET(cpuid);
2614 	tdq = TDQ_CPU(cpuid);
2615 	ts = td->td_sched;
2616 	if (TD_IS_IDLETHREAD(td))
2617 		td->td_lock = TDQ_LOCKPTR(tdq);
2618 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2619 	td->td_oncpu = cpuid;
2620 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2621 	lock_profile_obtain_lock_success(
2622 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2623 }
2624 
2625 /*
2626  * Create on first use to catch odd startup conditons.
2627  */
2628 char *
2629 sched_tdname(struct thread *td)
2630 {
2631 #ifdef KTR
2632 	struct td_sched *ts;
2633 
2634 	ts = td->td_sched;
2635 	if (ts->ts_name[0] == '\0')
2636 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2637 		    "%s tid %d", td->td_name, td->td_tid);
2638 	return (ts->ts_name);
2639 #else
2640 	return (td->td_name);
2641 #endif
2642 }
2643 
2644 #ifdef SMP
2645 
2646 /*
2647  * Build the CPU topology dump string. Is recursively called to collect
2648  * the topology tree.
2649  */
2650 static int
2651 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2652     int indent)
2653 {
2654 	int i, first;
2655 
2656 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2657 	    "", indent, cg->cg_level);
2658 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"0x%x\">", indent, "",
2659 	    cg->cg_count, cg->cg_mask);
2660 	first = TRUE;
2661 	for (i = 0; i < MAXCPU; i++) {
2662 		if ((cg->cg_mask & (1 << i)) != 0) {
2663 			if (!first)
2664 				sbuf_printf(sb, ", ");
2665 			else
2666 				first = FALSE;
2667 			sbuf_printf(sb, "%d", i);
2668 		}
2669 	}
2670 	sbuf_printf(sb, "</cpu>\n");
2671 
2672 	sbuf_printf(sb, "%*s <flags>", indent, "");
2673 	if (cg->cg_flags != 0) {
2674 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2675 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>\n");
2676 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2677 			sbuf_printf(sb, "<flag name=\"THREAD\">SMT group</flag>\n");
2678 	}
2679 	sbuf_printf(sb, "</flags>\n");
2680 
2681 	if (cg->cg_children > 0) {
2682 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2683 		for (i = 0; i < cg->cg_children; i++)
2684 			sysctl_kern_sched_topology_spec_internal(sb,
2685 			    &cg->cg_child[i], indent+2);
2686 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2687 	}
2688 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2689 	return (0);
2690 }
2691 
2692 /*
2693  * Sysctl handler for retrieving topology dump. It's a wrapper for
2694  * the recursive sysctl_kern_smp_topology_spec_internal().
2695  */
2696 static int
2697 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2698 {
2699 	struct sbuf *topo;
2700 	int err;
2701 
2702 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2703 
2704 	topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2705 	if (topo == NULL)
2706 		return (ENOMEM);
2707 
2708 	sbuf_printf(topo, "<groups>\n");
2709 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2710 	sbuf_printf(topo, "</groups>\n");
2711 
2712 	if (err == 0) {
2713 		sbuf_finish(topo);
2714 		err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2715 	}
2716 	sbuf_delete(topo);
2717 	return (err);
2718 }
2719 #endif
2720 
2721 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2722 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2723     "Scheduler name");
2724 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2725     "Slice size for timeshare threads");
2726 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2727      "Interactivity score threshold");
2728 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh,
2729      0,"Min priority for preemption, lower priorities have greater precedence");
2730 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost,
2731      0,"Controls whether static kernel priorities are assigned to sleeping threads.");
2732 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins,
2733      0,"Number of times idle will spin waiting for new work.");
2734 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, &sched_idlespinthresh,
2735      0,"Threshold before we will permit idle spinning.");
2736 #ifdef SMP
2737 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2738     "Number of hz ticks to keep thread affinity for");
2739 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2740     "Enables the long-term load balancer");
2741 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2742     &balance_interval, 0,
2743     "Average frequency in stathz ticks to run the long-term balancer");
2744 SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0,
2745     "Steals work from another hyper-threaded core on idle");
2746 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2747     "Attempts to steal work from other cores before idling");
2748 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2749     "Minimum load on remote cpu before we'll steal");
2750 
2751 /* Retrieve SMP topology */
2752 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2753     CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2754     "XML dump of detected CPU topology");
2755 #endif
2756 
2757 /* ps compat.  All cpu percentages from ULE are weighted. */
2758 static int ccpu = 0;
2759 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2760