xref: /freebsd/sys/kern/sched_ule.c (revision b2db760808f74bb53c232900091c9da801ebbfcc)
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 
66 #ifdef HWPMC_HOOKS
67 #include <sys/pmckern.h>
68 #endif
69 
70 #ifdef KDTRACE_HOOKS
71 #include <sys/dtrace_bsd.h>
72 int				dtrace_vtime_active;
73 dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
74 #endif
75 
76 #include <machine/cpu.h>
77 #include <machine/smp.h>
78 
79 #if defined(__sparc64__)
80 #error "This architecture is not currently compatible with ULE"
81 #endif
82 
83 #define	KTR_ULE	0
84 
85 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
86 #define	TDQ_NAME_LEN	(sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
87 #define	TDQ_LOADNAME_LEN	(PCPU_NAME_LEN + sizeof(" load"))
88 
89 /*
90  * Thread scheduler specific section.  All fields are protected
91  * by the thread lock.
92  */
93 struct td_sched {
94 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
95 	short		ts_flags;	/* TSF_* flags. */
96 	u_char		ts_cpu;		/* CPU that we have affinity for. */
97 	int		ts_rltick;	/* Real last tick, for affinity. */
98 	int		ts_slice;	/* Ticks of slice remaining. */
99 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
100 	u_int		ts_runtime;	/* Number of ticks we were running */
101 	int		ts_ltick;	/* Last tick that we were running on */
102 	int		ts_incrtick;	/* Last tick that we incremented on */
103 	int		ts_ftick;	/* First tick that we were running on */
104 	int		ts_ticks;	/* Tick count */
105 #ifdef KTR
106 	char		ts_name[TS_NAME_LEN];
107 #endif
108 };
109 /* flags kept in ts_flags */
110 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
111 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
112 
113 static struct td_sched td_sched0;
114 
115 #define	THREAD_CAN_MIGRATE(td)	((td)->td_pinned == 0)
116 #define	THREAD_CAN_SCHED(td, cpu)	\
117     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
118 
119 /*
120  * Cpu percentage computation macros and defines.
121  *
122  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
123  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
124  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
125  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
126  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
127  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
128  */
129 #define	SCHED_TICK_SECS		10
130 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
131 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
132 #define	SCHED_TICK_SHIFT	10
133 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
134 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
135 
136 /*
137  * These macros determine priorities for non-interactive threads.  They are
138  * assigned a priority based on their recent cpu utilization as expressed
139  * by the ratio of ticks to the tick total.  NHALF priorities at the start
140  * and end of the MIN to MAX timeshare range are only reachable with negative
141  * or positive nice respectively.
142  *
143  * PRI_RANGE:	Priority range for utilization dependent priorities.
144  * PRI_NRESV:	Number of nice values.
145  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
146  * PRI_NICE:	Determines the part of the priority inherited from nice.
147  */
148 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
149 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
150 #define	SCHED_PRI_MIN		(PRI_MIN_TIMESHARE + SCHED_PRI_NHALF)
151 #define	SCHED_PRI_MAX		(PRI_MAX_TIMESHARE - SCHED_PRI_NHALF)
152 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN)
153 #define	SCHED_PRI_TICKS(ts)						\
154     (SCHED_TICK_HZ((ts)) /						\
155     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
156 #define	SCHED_PRI_NICE(nice)	(nice)
157 
158 /*
159  * These determine the interactivity of a process.  Interactivity differs from
160  * cpu utilization in that it expresses the voluntary time slept vs time ran
161  * while cpu utilization includes all time not running.  This more accurately
162  * models the intent of the thread.
163  *
164  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
165  *		before throttling back.
166  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
167  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
168  * INTERACT_THRESH:	Threshhold for placement on the current runq.
169  */
170 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
171 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
172 #define	SCHED_INTERACT_MAX	(100)
173 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
174 #define	SCHED_INTERACT_THRESH	(30)
175 
176 /*
177  * tickincr:		Converts a stathz tick into a hz domain scaled by
178  *			the shift factor.  Without the shift the error rate
179  *			due to rounding would be unacceptably high.
180  * realstathz:		stathz is sometimes 0 and run off of hz.
181  * sched_slice:		Runtime of each thread before rescheduling.
182  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
183  */
184 static int sched_interact = SCHED_INTERACT_THRESH;
185 static int realstathz;
186 static int tickincr;
187 static int sched_slice = 1;
188 #ifdef PREEMPTION
189 #ifdef FULL_PREEMPTION
190 static int preempt_thresh = PRI_MAX_IDLE;
191 #else
192 static int preempt_thresh = PRI_MIN_KERN;
193 #endif
194 #else
195 static int preempt_thresh = 0;
196 #endif
197 static int static_boost = PRI_MIN_TIMESHARE;
198 static int sched_idlespins = 10000;
199 static int sched_idlespinthresh = 4;
200 
201 /*
202  * tdq - per processor runqs and statistics.  All fields are protected by the
203  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
204  * locking in sched_pickcpu();
205  */
206 struct tdq {
207 	/* Ordered to improve efficiency of cpu_search() and switch(). */
208 	struct mtx	tdq_lock;		/* run queue lock. */
209 	struct cpu_group *tdq_cg;		/* Pointer to cpu topology. */
210 	volatile int	tdq_load;		/* Aggregate load. */
211 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
212 	int		tdq_transferable;	/* Transferable thread count. */
213 	short		tdq_switchcnt;		/* Switches this tick. */
214 	short		tdq_oldswitchcnt;	/* Switches last tick. */
215 	u_char		tdq_lowpri;		/* Lowest priority thread. */
216 	u_char		tdq_ipipending;		/* IPI pending. */
217 	u_char		tdq_idx;		/* Current insert index. */
218 	u_char		tdq_ridx;		/* Current removal index. */
219 	struct runq	tdq_realtime;		/* real-time run queue. */
220 	struct runq	tdq_timeshare;		/* timeshare run queue. */
221 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
222 	char		tdq_name[TDQ_NAME_LEN];
223 #ifdef KTR
224 	char		tdq_loadname[TDQ_LOADNAME_LEN];
225 #endif
226 } __aligned(64);
227 
228 /* Idle thread states and config. */
229 #define	TDQ_RUNNING	1
230 #define	TDQ_IDLE	2
231 
232 #ifdef SMP
233 struct cpu_group *cpu_top;		/* CPU topology */
234 
235 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
236 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
237 
238 /*
239  * Run-time tunables.
240  */
241 static int rebalance = 1;
242 static int balance_interval = 128;	/* Default set in sched_initticks(). */
243 static int affinity;
244 static int steal_htt = 1;
245 static int steal_idle = 1;
246 static int steal_thresh = 2;
247 
248 /*
249  * One thread queue per processor.
250  */
251 static struct tdq	tdq_cpu[MAXCPU];
252 static struct tdq	*balance_tdq;
253 static int balance_ticks;
254 
255 #define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
256 #define	TDQ_CPU(x)	(&tdq_cpu[(x)])
257 #define	TDQ_ID(x)	((int)((x) - tdq_cpu))
258 #else	/* !SMP */
259 static struct tdq	tdq_cpu;
260 
261 #define	TDQ_ID(x)	(0)
262 #define	TDQ_SELF()	(&tdq_cpu)
263 #define	TDQ_CPU(x)	(&tdq_cpu)
264 #endif
265 
266 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
267 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
268 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
269 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
270 #define	TDQ_LOCKPTR(t)		(&(t)->tdq_lock)
271 
272 static void sched_priority(struct thread *);
273 static void sched_thread_priority(struct thread *, u_char);
274 static int sched_interact_score(struct thread *);
275 static void sched_interact_update(struct thread *);
276 static void sched_interact_fork(struct thread *);
277 static void sched_pctcpu_update(struct td_sched *);
278 
279 /* Operations on per processor queues */
280 static struct thread *tdq_choose(struct tdq *);
281 static void tdq_setup(struct tdq *);
282 static void tdq_load_add(struct tdq *, struct thread *);
283 static void tdq_load_rem(struct tdq *, struct thread *);
284 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
285 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
286 static inline int sched_shouldpreempt(int, int, int);
287 void tdq_print(int cpu);
288 static void runq_print(struct runq *rq);
289 static void tdq_add(struct tdq *, struct thread *, int);
290 #ifdef SMP
291 static int tdq_move(struct tdq *, struct tdq *);
292 static int tdq_idled(struct tdq *);
293 static void tdq_notify(struct tdq *, struct thread *);
294 static struct thread *tdq_steal(struct tdq *, int);
295 static struct thread *runq_steal(struct runq *, int);
296 static int sched_pickcpu(struct thread *, int);
297 static void sched_balance(void);
298 static int sched_balance_pair(struct tdq *, struct tdq *);
299 static inline struct tdq *sched_setcpu(struct thread *, int, int);
300 static inline void thread_unblock_switch(struct thread *, struct mtx *);
301 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
302 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
303 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
304     struct cpu_group *cg, int indent);
305 #endif
306 
307 static void sched_setup(void *dummy);
308 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
309 
310 static void sched_initticks(void *dummy);
311 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
312     NULL);
313 
314 /*
315  * Print the threads waiting on a run-queue.
316  */
317 static void
318 runq_print(struct runq *rq)
319 {
320 	struct rqhead *rqh;
321 	struct thread *td;
322 	int pri;
323 	int j;
324 	int i;
325 
326 	for (i = 0; i < RQB_LEN; i++) {
327 		printf("\t\trunq bits %d 0x%zx\n",
328 		    i, rq->rq_status.rqb_bits[i]);
329 		for (j = 0; j < RQB_BPW; j++)
330 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
331 				pri = j + (i << RQB_L2BPW);
332 				rqh = &rq->rq_queues[pri];
333 				TAILQ_FOREACH(td, rqh, td_runq) {
334 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
335 					    td, td->td_name, td->td_priority,
336 					    td->td_rqindex, pri);
337 				}
338 			}
339 	}
340 }
341 
342 /*
343  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
344  */
345 void
346 tdq_print(int cpu)
347 {
348 	struct tdq *tdq;
349 
350 	tdq = TDQ_CPU(cpu);
351 
352 	printf("tdq %d:\n", TDQ_ID(tdq));
353 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
354 	printf("\tLock name:      %s\n", tdq->tdq_name);
355 	printf("\tload:           %d\n", tdq->tdq_load);
356 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
357 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
358 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
359 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
360 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
361 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
362 	printf("\trealtime runq:\n");
363 	runq_print(&tdq->tdq_realtime);
364 	printf("\ttimeshare runq:\n");
365 	runq_print(&tdq->tdq_timeshare);
366 	printf("\tidle runq:\n");
367 	runq_print(&tdq->tdq_idle);
368 }
369 
370 static inline int
371 sched_shouldpreempt(int pri, int cpri, int remote)
372 {
373 	/*
374 	 * If the new priority is not better than the current priority there is
375 	 * nothing to do.
376 	 */
377 	if (pri >= cpri)
378 		return (0);
379 	/*
380 	 * Always preempt idle.
381 	 */
382 	if (cpri >= PRI_MIN_IDLE)
383 		return (1);
384 	/*
385 	 * If preemption is disabled don't preempt others.
386 	 */
387 	if (preempt_thresh == 0)
388 		return (0);
389 	/*
390 	 * Preempt if we exceed the threshold.
391 	 */
392 	if (pri <= preempt_thresh)
393 		return (1);
394 	/*
395 	 * If we're realtime or better and there is timeshare or worse running
396 	 * preempt only remote processors.
397 	 */
398 	if (remote && pri <= PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME)
399 		return (1);
400 	return (0);
401 }
402 
403 #define	TS_RQ_PPQ	(((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS)
404 /*
405  * Add a thread to the actual run-queue.  Keeps transferable counts up to
406  * date with what is actually on the run-queue.  Selects the correct
407  * queue position for timeshare threads.
408  */
409 static __inline void
410 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
411 {
412 	struct td_sched *ts;
413 	u_char pri;
414 
415 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
416 	THREAD_LOCK_ASSERT(td, MA_OWNED);
417 
418 	pri = td->td_priority;
419 	ts = td->td_sched;
420 	TD_SET_RUNQ(td);
421 	if (THREAD_CAN_MIGRATE(td)) {
422 		tdq->tdq_transferable++;
423 		ts->ts_flags |= TSF_XFERABLE;
424 	}
425 	if (pri <= PRI_MAX_REALTIME) {
426 		ts->ts_runq = &tdq->tdq_realtime;
427 	} else if (pri <= PRI_MAX_TIMESHARE) {
428 		ts->ts_runq = &tdq->tdq_timeshare;
429 		KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE,
430 			("Invalid priority %d on timeshare runq", pri));
431 		/*
432 		 * This queue contains only priorities between MIN and MAX
433 		 * realtime.  Use the whole queue to represent these values.
434 		 */
435 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
436 			pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ;
437 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
438 			/*
439 			 * This effectively shortens the queue by one so we
440 			 * can have a one slot difference between idx and
441 			 * ridx while we wait for threads to drain.
442 			 */
443 			if (tdq->tdq_ridx != tdq->tdq_idx &&
444 			    pri == tdq->tdq_ridx)
445 				pri = (unsigned char)(pri - 1) % RQ_NQS;
446 		} else
447 			pri = tdq->tdq_ridx;
448 		runq_add_pri(ts->ts_runq, td, pri, flags);
449 		return;
450 	} else
451 		ts->ts_runq = &tdq->tdq_idle;
452 	runq_add(ts->ts_runq, td, flags);
453 }
454 
455 /*
456  * Remove a thread from a run-queue.  This typically happens when a thread
457  * is selected to run.  Running threads are not on the queue and the
458  * transferable count does not reflect them.
459  */
460 static __inline void
461 tdq_runq_rem(struct tdq *tdq, struct thread *td)
462 {
463 	struct td_sched *ts;
464 
465 	ts = td->td_sched;
466 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
467 	KASSERT(ts->ts_runq != NULL,
468 	    ("tdq_runq_remove: thread %p null ts_runq", td));
469 	if (ts->ts_flags & TSF_XFERABLE) {
470 		tdq->tdq_transferable--;
471 		ts->ts_flags &= ~TSF_XFERABLE;
472 	}
473 	if (ts->ts_runq == &tdq->tdq_timeshare) {
474 		if (tdq->tdq_idx != tdq->tdq_ridx)
475 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
476 		else
477 			runq_remove_idx(ts->ts_runq, td, NULL);
478 	} else
479 		runq_remove(ts->ts_runq, td);
480 }
481 
482 /*
483  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
484  * for this thread to the referenced thread queue.
485  */
486 static void
487 tdq_load_add(struct tdq *tdq, struct thread *td)
488 {
489 
490 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
491 	THREAD_LOCK_ASSERT(td, MA_OWNED);
492 
493 	tdq->tdq_load++;
494 	if ((td->td_flags & TDF_NOLOAD) == 0)
495 		tdq->tdq_sysload++;
496 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
497 }
498 
499 /*
500  * Remove the load from a thread that is transitioning to a sleep state or
501  * exiting.
502  */
503 static void
504 tdq_load_rem(struct tdq *tdq, struct thread *td)
505 {
506 
507 	THREAD_LOCK_ASSERT(td, MA_OWNED);
508 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
509 	KASSERT(tdq->tdq_load != 0,
510 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
511 
512 	tdq->tdq_load--;
513 	if ((td->td_flags & TDF_NOLOAD) == 0)
514 		tdq->tdq_sysload--;
515 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
516 }
517 
518 /*
519  * Set lowpri to its exact value by searching the run-queue and
520  * evaluating curthread.  curthread may be passed as an optimization.
521  */
522 static void
523 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
524 {
525 	struct thread *td;
526 
527 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
528 	if (ctd == NULL)
529 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
530 	td = tdq_choose(tdq);
531 	if (td == NULL || td->td_priority > ctd->td_priority)
532 		tdq->tdq_lowpri = ctd->td_priority;
533 	else
534 		tdq->tdq_lowpri = td->td_priority;
535 }
536 
537 #ifdef SMP
538 struct cpu_search {
539 	cpuset_t cs_mask;
540 	u_int	cs_load;
541 	u_int	cs_cpu;
542 	int	cs_limit;	/* Min priority for low min load for high. */
543 };
544 
545 #define	CPU_SEARCH_LOWEST	0x1
546 #define	CPU_SEARCH_HIGHEST	0x2
547 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
548 
549 #define	CPUSET_FOREACH(cpu, mask)				\
550 	for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++)		\
551 		if ((mask) & 1 << (cpu))
552 
553 static __inline int cpu_search(struct cpu_group *cg, struct cpu_search *low,
554     struct cpu_search *high, const int match);
555 int cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low);
556 int cpu_search_highest(struct cpu_group *cg, struct cpu_search *high);
557 int cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
558     struct cpu_search *high);
559 
560 /*
561  * This routine compares according to the match argument and should be
562  * reduced in actual instantiations via constant propagation and dead code
563  * elimination.
564  */
565 static __inline int
566 cpu_compare(int cpu, struct cpu_search *low, struct cpu_search *high,
567     const int match)
568 {
569 	struct tdq *tdq;
570 
571 	tdq = TDQ_CPU(cpu);
572 	if (match & CPU_SEARCH_LOWEST)
573 		if (CPU_ISSET(cpu, &low->cs_mask) &&
574 		    tdq->tdq_load < low->cs_load &&
575 		    tdq->tdq_lowpri > low->cs_limit) {
576 			low->cs_cpu = cpu;
577 			low->cs_load = tdq->tdq_load;
578 		}
579 	if (match & CPU_SEARCH_HIGHEST)
580 		if (CPU_ISSET(cpu, &high->cs_mask) &&
581 		    tdq->tdq_load >= high->cs_limit &&
582 		    tdq->tdq_load > high->cs_load &&
583 		    tdq->tdq_transferable) {
584 			high->cs_cpu = cpu;
585 			high->cs_load = tdq->tdq_load;
586 		}
587 	return (tdq->tdq_load);
588 }
589 
590 /*
591  * Search the tree of cpu_groups for the lowest or highest loaded cpu
592  * according to the match argument.  This routine actually compares the
593  * load on all paths through the tree and finds the least loaded cpu on
594  * the least loaded path, which may differ from the least loaded cpu in
595  * the system.  This balances work among caches and busses.
596  *
597  * This inline is instantiated in three forms below using constants for the
598  * match argument.  It is reduced to the minimum set for each case.  It is
599  * also recursive to the depth of the tree.
600  */
601 static __inline int
602 cpu_search(struct cpu_group *cg, struct cpu_search *low,
603     struct cpu_search *high, const int match)
604 {
605 	int total;
606 
607 	total = 0;
608 	if (cg->cg_children) {
609 		struct cpu_search lgroup;
610 		struct cpu_search hgroup;
611 		struct cpu_group *child;
612 		u_int lload;
613 		int hload;
614 		int load;
615 		int i;
616 
617 		lload = -1;
618 		hload = -1;
619 		for (i = 0; i < cg->cg_children; i++) {
620 			child = &cg->cg_child[i];
621 			if (match & CPU_SEARCH_LOWEST) {
622 				lgroup = *low;
623 				lgroup.cs_load = -1;
624 			}
625 			if (match & CPU_SEARCH_HIGHEST) {
626 				hgroup = *high;
627 				lgroup.cs_load = 0;
628 			}
629 			switch (match) {
630 			case CPU_SEARCH_LOWEST:
631 				load = cpu_search_lowest(child, &lgroup);
632 				break;
633 			case CPU_SEARCH_HIGHEST:
634 				load = cpu_search_highest(child, &hgroup);
635 				break;
636 			case CPU_SEARCH_BOTH:
637 				load = cpu_search_both(child, &lgroup, &hgroup);
638 				break;
639 			}
640 			total += load;
641 			if (match & CPU_SEARCH_LOWEST)
642 				if (load < lload || low->cs_cpu == -1) {
643 					*low = lgroup;
644 					lload = load;
645 				}
646 			if (match & CPU_SEARCH_HIGHEST)
647 				if (load > hload || high->cs_cpu == -1) {
648 					hload = load;
649 					*high = hgroup;
650 				}
651 		}
652 	} else {
653 		int cpu;
654 
655 		CPUSET_FOREACH(cpu, cg->cg_mask)
656 			total += cpu_compare(cpu, low, high, match);
657 	}
658 	return (total);
659 }
660 
661 /*
662  * cpu_search instantiations must pass constants to maintain the inline
663  * optimization.
664  */
665 int
666 cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low)
667 {
668 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
669 }
670 
671 int
672 cpu_search_highest(struct cpu_group *cg, struct cpu_search *high)
673 {
674 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
675 }
676 
677 int
678 cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
679     struct cpu_search *high)
680 {
681 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
682 }
683 
684 /*
685  * Find the cpu with the least load via the least loaded path that has a
686  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
687  * acceptable.
688  */
689 static inline int
690 sched_lowest(struct cpu_group *cg, cpuset_t mask, int pri)
691 {
692 	struct cpu_search low;
693 
694 	low.cs_cpu = -1;
695 	low.cs_load = -1;
696 	low.cs_mask = mask;
697 	low.cs_limit = pri;
698 	cpu_search_lowest(cg, &low);
699 	return low.cs_cpu;
700 }
701 
702 /*
703  * Find the cpu with the highest load via the highest loaded path.
704  */
705 static inline int
706 sched_highest(struct cpu_group *cg, cpuset_t mask, int minload)
707 {
708 	struct cpu_search high;
709 
710 	high.cs_cpu = -1;
711 	high.cs_load = 0;
712 	high.cs_mask = mask;
713 	high.cs_limit = minload;
714 	cpu_search_highest(cg, &high);
715 	return high.cs_cpu;
716 }
717 
718 /*
719  * Simultaneously find the highest and lowest loaded cpu reachable via
720  * cg.
721  */
722 static inline void
723 sched_both(struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu)
724 {
725 	struct cpu_search high;
726 	struct cpu_search low;
727 
728 	low.cs_cpu = -1;
729 	low.cs_limit = -1;
730 	low.cs_load = -1;
731 	low.cs_mask = mask;
732 	high.cs_load = 0;
733 	high.cs_cpu = -1;
734 	high.cs_limit = -1;
735 	high.cs_mask = mask;
736 	cpu_search_both(cg, &low, &high);
737 	*lowcpu = low.cs_cpu;
738 	*highcpu = high.cs_cpu;
739 	return;
740 }
741 
742 static void
743 sched_balance_group(struct cpu_group *cg)
744 {
745 	cpuset_t mask;
746 	int high;
747 	int low;
748 	int i;
749 
750 	CPU_FILL(&mask);
751 	for (;;) {
752 		sched_both(cg, mask, &low, &high);
753 		if (low == high || low == -1 || high == -1)
754 			break;
755 		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low)))
756 			break;
757 		/*
758 		 * If we failed to move any threads determine which cpu
759 		 * to kick out of the set and try again.
760 	 	 */
761 		if (TDQ_CPU(high)->tdq_transferable == 0)
762 			CPU_CLR(high, &mask);
763 		else
764 			CPU_CLR(low, &mask);
765 	}
766 
767 	for (i = 0; i < cg->cg_children; i++)
768 		sched_balance_group(&cg->cg_child[i]);
769 }
770 
771 static void
772 sched_balance(void)
773 {
774 	struct tdq *tdq;
775 
776 	/*
777 	 * Select a random time between .5 * balance_interval and
778 	 * 1.5 * balance_interval.
779 	 */
780 	balance_ticks = max(balance_interval / 2, 1);
781 	balance_ticks += random() % balance_interval;
782 	if (smp_started == 0 || rebalance == 0)
783 		return;
784 	tdq = TDQ_SELF();
785 	TDQ_UNLOCK(tdq);
786 	sched_balance_group(cpu_top);
787 	TDQ_LOCK(tdq);
788 }
789 
790 /*
791  * Lock two thread queues using their address to maintain lock order.
792  */
793 static void
794 tdq_lock_pair(struct tdq *one, struct tdq *two)
795 {
796 	if (one < two) {
797 		TDQ_LOCK(one);
798 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
799 	} else {
800 		TDQ_LOCK(two);
801 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
802 	}
803 }
804 
805 /*
806  * Unlock two thread queues.  Order is not important here.
807  */
808 static void
809 tdq_unlock_pair(struct tdq *one, struct tdq *two)
810 {
811 	TDQ_UNLOCK(one);
812 	TDQ_UNLOCK(two);
813 }
814 
815 /*
816  * Transfer load between two imbalanced thread queues.
817  */
818 static int
819 sched_balance_pair(struct tdq *high, struct tdq *low)
820 {
821 	int transferable;
822 	int high_load;
823 	int low_load;
824 	int moved;
825 	int move;
826 	int diff;
827 	int i;
828 
829 	tdq_lock_pair(high, low);
830 	transferable = high->tdq_transferable;
831 	high_load = high->tdq_load;
832 	low_load = low->tdq_load;
833 	moved = 0;
834 	/*
835 	 * Determine what the imbalance is and then adjust that to how many
836 	 * threads we actually have to give up (transferable).
837 	 */
838 	if (transferable != 0) {
839 		diff = high_load - low_load;
840 		move = diff / 2;
841 		if (diff & 0x1)
842 			move++;
843 		move = min(move, transferable);
844 		for (i = 0; i < move; i++)
845 			moved += tdq_move(high, low);
846 		/*
847 		 * IPI the target cpu to force it to reschedule with the new
848 		 * workload.
849 		 */
850 		ipi_cpu(TDQ_ID(low), IPI_PREEMPT);
851 	}
852 	tdq_unlock_pair(high, low);
853 	return (moved);
854 }
855 
856 /*
857  * Move a thread from one thread queue to another.
858  */
859 static int
860 tdq_move(struct tdq *from, struct tdq *to)
861 {
862 	struct td_sched *ts;
863 	struct thread *td;
864 	struct tdq *tdq;
865 	int cpu;
866 
867 	TDQ_LOCK_ASSERT(from, MA_OWNED);
868 	TDQ_LOCK_ASSERT(to, MA_OWNED);
869 
870 	tdq = from;
871 	cpu = TDQ_ID(to);
872 	td = tdq_steal(tdq, cpu);
873 	if (td == NULL)
874 		return (0);
875 	ts = td->td_sched;
876 	/*
877 	 * Although the run queue is locked the thread may be blocked.  Lock
878 	 * it to clear this and acquire the run-queue lock.
879 	 */
880 	thread_lock(td);
881 	/* Drop recursive lock on from acquired via thread_lock(). */
882 	TDQ_UNLOCK(from);
883 	sched_rem(td);
884 	ts->ts_cpu = cpu;
885 	td->td_lock = TDQ_LOCKPTR(to);
886 	tdq_add(to, td, SRQ_YIELDING);
887 	return (1);
888 }
889 
890 /*
891  * This tdq has idled.  Try to steal a thread from another cpu and switch
892  * to it.
893  */
894 static int
895 tdq_idled(struct tdq *tdq)
896 {
897 	struct cpu_group *cg;
898 	struct tdq *steal;
899 	cpuset_t mask;
900 	int thresh;
901 	int cpu;
902 
903 	if (smp_started == 0 || steal_idle == 0)
904 		return (1);
905 	CPU_FILL(&mask);
906 	CPU_CLR(PCPU_GET(cpuid), &mask);
907 	/* We don't want to be preempted while we're iterating. */
908 	spinlock_enter();
909 	for (cg = tdq->tdq_cg; cg != NULL; ) {
910 		if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
911 			thresh = steal_thresh;
912 		else
913 			thresh = 1;
914 		cpu = sched_highest(cg, mask, thresh);
915 		if (cpu == -1) {
916 			cg = cg->cg_parent;
917 			continue;
918 		}
919 		steal = TDQ_CPU(cpu);
920 		CPU_CLR(cpu, &mask);
921 		tdq_lock_pair(tdq, steal);
922 		if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
923 			tdq_unlock_pair(tdq, steal);
924 			continue;
925 		}
926 		/*
927 		 * If a thread was added while interrupts were disabled don't
928 		 * steal one here.  If we fail to acquire one due to affinity
929 		 * restrictions loop again with this cpu removed from the
930 		 * set.
931 		 */
932 		if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
933 			tdq_unlock_pair(tdq, steal);
934 			continue;
935 		}
936 		spinlock_exit();
937 		TDQ_UNLOCK(steal);
938 		mi_switch(SW_VOL | SWT_IDLE, NULL);
939 		thread_unlock(curthread);
940 
941 		return (0);
942 	}
943 	spinlock_exit();
944 	return (1);
945 }
946 
947 /*
948  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
949  */
950 static void
951 tdq_notify(struct tdq *tdq, struct thread *td)
952 {
953 	struct thread *ctd;
954 	int pri;
955 	int cpu;
956 
957 	if (tdq->tdq_ipipending)
958 		return;
959 	cpu = td->td_sched->ts_cpu;
960 	pri = td->td_priority;
961 	ctd = pcpu_find(cpu)->pc_curthread;
962 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
963 		return;
964 	if (TD_IS_IDLETHREAD(ctd)) {
965 		/*
966 		 * If the MD code has an idle wakeup routine try that before
967 		 * falling back to IPI.
968 		 */
969 		if (cpu_idle_wakeup(cpu))
970 			return;
971 	}
972 	tdq->tdq_ipipending = 1;
973 	ipi_cpu(cpu, IPI_PREEMPT);
974 }
975 
976 /*
977  * Steals load from a timeshare queue.  Honors the rotating queue head
978  * index.
979  */
980 static struct thread *
981 runq_steal_from(struct runq *rq, int cpu, u_char start)
982 {
983 	struct rqbits *rqb;
984 	struct rqhead *rqh;
985 	struct thread *td;
986 	int first;
987 	int bit;
988 	int pri;
989 	int i;
990 
991 	rqb = &rq->rq_status;
992 	bit = start & (RQB_BPW -1);
993 	pri = 0;
994 	first = 0;
995 again:
996 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
997 		if (rqb->rqb_bits[i] == 0)
998 			continue;
999 		if (bit != 0) {
1000 			for (pri = bit; pri < RQB_BPW; pri++)
1001 				if (rqb->rqb_bits[i] & (1ul << pri))
1002 					break;
1003 			if (pri >= RQB_BPW)
1004 				continue;
1005 		} else
1006 			pri = RQB_FFS(rqb->rqb_bits[i]);
1007 		pri += (i << RQB_L2BPW);
1008 		rqh = &rq->rq_queues[pri];
1009 		TAILQ_FOREACH(td, rqh, td_runq) {
1010 			if (first && THREAD_CAN_MIGRATE(td) &&
1011 			    THREAD_CAN_SCHED(td, cpu))
1012 				return (td);
1013 			first = 1;
1014 		}
1015 	}
1016 	if (start != 0) {
1017 		start = 0;
1018 		goto again;
1019 	}
1020 
1021 	return (NULL);
1022 }
1023 
1024 /*
1025  * Steals load from a standard linear queue.
1026  */
1027 static struct thread *
1028 runq_steal(struct runq *rq, int cpu)
1029 {
1030 	struct rqhead *rqh;
1031 	struct rqbits *rqb;
1032 	struct thread *td;
1033 	int word;
1034 	int bit;
1035 
1036 	rqb = &rq->rq_status;
1037 	for (word = 0; word < RQB_LEN; word++) {
1038 		if (rqb->rqb_bits[word] == 0)
1039 			continue;
1040 		for (bit = 0; bit < RQB_BPW; bit++) {
1041 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1042 				continue;
1043 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1044 			TAILQ_FOREACH(td, rqh, td_runq)
1045 				if (THREAD_CAN_MIGRATE(td) &&
1046 				    THREAD_CAN_SCHED(td, cpu))
1047 					return (td);
1048 		}
1049 	}
1050 	return (NULL);
1051 }
1052 
1053 /*
1054  * Attempt to steal a thread in priority order from a thread queue.
1055  */
1056 static struct thread *
1057 tdq_steal(struct tdq *tdq, int cpu)
1058 {
1059 	struct thread *td;
1060 
1061 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1062 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1063 		return (td);
1064 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1065 	    cpu, tdq->tdq_ridx)) != NULL)
1066 		return (td);
1067 	return (runq_steal(&tdq->tdq_idle, cpu));
1068 }
1069 
1070 /*
1071  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1072  * current lock and returns with the assigned queue locked.
1073  */
1074 static inline struct tdq *
1075 sched_setcpu(struct thread *td, int cpu, int flags)
1076 {
1077 
1078 	struct tdq *tdq;
1079 
1080 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1081 	tdq = TDQ_CPU(cpu);
1082 	td->td_sched->ts_cpu = cpu;
1083 	/*
1084 	 * If the lock matches just return the queue.
1085 	 */
1086 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1087 		return (tdq);
1088 #ifdef notyet
1089 	/*
1090 	 * If the thread isn't running its lockptr is a
1091 	 * turnstile or a sleepqueue.  We can just lock_set without
1092 	 * blocking.
1093 	 */
1094 	if (TD_CAN_RUN(td)) {
1095 		TDQ_LOCK(tdq);
1096 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1097 		return (tdq);
1098 	}
1099 #endif
1100 	/*
1101 	 * The hard case, migration, we need to block the thread first to
1102 	 * prevent order reversals with other cpus locks.
1103 	 */
1104 	spinlock_enter();
1105 	thread_lock_block(td);
1106 	TDQ_LOCK(tdq);
1107 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1108 	spinlock_exit();
1109 	return (tdq);
1110 }
1111 
1112 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1113 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1114 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1115 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1116 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1117 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1118 
1119 static int
1120 sched_pickcpu(struct thread *td, int flags)
1121 {
1122 	struct cpu_group *cg;
1123 	struct td_sched *ts;
1124 	struct tdq *tdq;
1125 	cpuset_t mask;
1126 	int self;
1127 	int pri;
1128 	int cpu;
1129 
1130 	self = PCPU_GET(cpuid);
1131 	ts = td->td_sched;
1132 	if (smp_started == 0)
1133 		return (self);
1134 	/*
1135 	 * Don't migrate a running thread from sched_switch().
1136 	 */
1137 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1138 		return (ts->ts_cpu);
1139 	/*
1140 	 * Prefer to run interrupt threads on the processors that generate
1141 	 * the interrupt.
1142 	 */
1143 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1144 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1145 		SCHED_STAT_INC(pickcpu_intrbind);
1146 		ts->ts_cpu = self;
1147 	}
1148 	/*
1149 	 * If the thread can run on the last cpu and the affinity has not
1150 	 * expired or it is idle run it there.
1151 	 */
1152 	pri = td->td_priority;
1153 	tdq = TDQ_CPU(ts->ts_cpu);
1154 	if (THREAD_CAN_SCHED(td, ts->ts_cpu)) {
1155 		if (tdq->tdq_lowpri > PRI_MIN_IDLE) {
1156 			SCHED_STAT_INC(pickcpu_idle_affinity);
1157 			return (ts->ts_cpu);
1158 		}
1159 		if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri) {
1160 			SCHED_STAT_INC(pickcpu_affinity);
1161 			return (ts->ts_cpu);
1162 		}
1163 	}
1164 	/*
1165 	 * Search for the highest level in the tree that still has affinity.
1166 	 */
1167 	cg = NULL;
1168 	for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent)
1169 		if (SCHED_AFFINITY(ts, cg->cg_level))
1170 			break;
1171 	cpu = -1;
1172 	mask = td->td_cpuset->cs_mask;
1173 	if (cg)
1174 		cpu = sched_lowest(cg, mask, pri);
1175 	if (cpu == -1)
1176 		cpu = sched_lowest(cpu_top, mask, -1);
1177 	/*
1178 	 * Compare the lowest loaded cpu to current cpu.
1179 	 */
1180 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1181 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) {
1182 		SCHED_STAT_INC(pickcpu_local);
1183 		cpu = self;
1184 	} else
1185 		SCHED_STAT_INC(pickcpu_lowest);
1186 	if (cpu != ts->ts_cpu)
1187 		SCHED_STAT_INC(pickcpu_migration);
1188 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1189 	return (cpu);
1190 }
1191 #endif
1192 
1193 /*
1194  * Pick the highest priority task we have and return it.
1195  */
1196 static struct thread *
1197 tdq_choose(struct tdq *tdq)
1198 {
1199 	struct thread *td;
1200 
1201 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1202 	td = runq_choose(&tdq->tdq_realtime);
1203 	if (td != NULL)
1204 		return (td);
1205 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1206 	if (td != NULL) {
1207 		KASSERT(td->td_priority >= PRI_MIN_TIMESHARE,
1208 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1209 		    td->td_priority));
1210 		return (td);
1211 	}
1212 	td = runq_choose(&tdq->tdq_idle);
1213 	if (td != NULL) {
1214 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1215 		    ("tdq_choose: Invalid priority on idle queue %d",
1216 		    td->td_priority));
1217 		return (td);
1218 	}
1219 
1220 	return (NULL);
1221 }
1222 
1223 /*
1224  * Initialize a thread queue.
1225  */
1226 static void
1227 tdq_setup(struct tdq *tdq)
1228 {
1229 
1230 	if (bootverbose)
1231 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1232 	runq_init(&tdq->tdq_realtime);
1233 	runq_init(&tdq->tdq_timeshare);
1234 	runq_init(&tdq->tdq_idle);
1235 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1236 	    "sched lock %d", (int)TDQ_ID(tdq));
1237 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1238 	    MTX_SPIN | MTX_RECURSE);
1239 #ifdef KTR
1240 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1241 	    "CPU %d load", (int)TDQ_ID(tdq));
1242 #endif
1243 }
1244 
1245 #ifdef SMP
1246 static void
1247 sched_setup_smp(void)
1248 {
1249 	struct tdq *tdq;
1250 	int i;
1251 
1252 	cpu_top = smp_topo();
1253 	CPU_FOREACH(i) {
1254 		tdq = TDQ_CPU(i);
1255 		tdq_setup(tdq);
1256 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1257 		if (tdq->tdq_cg == NULL)
1258 			panic("Can't find cpu group for %d\n", i);
1259 	}
1260 	balance_tdq = TDQ_SELF();
1261 	sched_balance();
1262 }
1263 #endif
1264 
1265 /*
1266  * Setup the thread queues and initialize the topology based on MD
1267  * information.
1268  */
1269 static void
1270 sched_setup(void *dummy)
1271 {
1272 	struct tdq *tdq;
1273 
1274 	tdq = TDQ_SELF();
1275 #ifdef SMP
1276 	sched_setup_smp();
1277 #else
1278 	tdq_setup(tdq);
1279 #endif
1280 	/*
1281 	 * To avoid divide-by-zero, we set realstathz a dummy value
1282 	 * in case which sched_clock() called before sched_initticks().
1283 	 */
1284 	realstathz = hz;
1285 	sched_slice = (realstathz/10);	/* ~100ms */
1286 	tickincr = 1 << SCHED_TICK_SHIFT;
1287 
1288 	/* Add thread0's load since it's running. */
1289 	TDQ_LOCK(tdq);
1290 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1291 	tdq_load_add(tdq, &thread0);
1292 	tdq->tdq_lowpri = thread0.td_priority;
1293 	TDQ_UNLOCK(tdq);
1294 }
1295 
1296 /*
1297  * This routine determines the tickincr after stathz and hz are setup.
1298  */
1299 /* ARGSUSED */
1300 static void
1301 sched_initticks(void *dummy)
1302 {
1303 	int incr;
1304 
1305 	realstathz = stathz ? stathz : hz;
1306 	sched_slice = (realstathz/10);	/* ~100ms */
1307 
1308 	/*
1309 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1310 	 * hz not being evenly divisible by stathz on all platforms.
1311 	 */
1312 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1313 	/*
1314 	 * This does not work for values of stathz that are more than
1315 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1316 	 */
1317 	if (incr == 0)
1318 		incr = 1;
1319 	tickincr = incr;
1320 #ifdef SMP
1321 	/*
1322 	 * Set the default balance interval now that we know
1323 	 * what realstathz is.
1324 	 */
1325 	balance_interval = realstathz;
1326 	/*
1327 	 * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4.
1328 	 * This prevents excess thrashing on large machines and excess idle
1329 	 * on smaller machines.
1330 	 */
1331 	steal_thresh = min(fls(mp_ncpus) - 1, 3);
1332 	affinity = SCHED_AFFINITY_DEFAULT;
1333 #endif
1334 }
1335 
1336 
1337 /*
1338  * This is the core of the interactivity algorithm.  Determines a score based
1339  * on past behavior.  It is the ratio of sleep time to run time scaled to
1340  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1341  * differs from the cpu usage because it does not account for time spent
1342  * waiting on a run-queue.  Would be prettier if we had floating point.
1343  */
1344 static int
1345 sched_interact_score(struct thread *td)
1346 {
1347 	struct td_sched *ts;
1348 	int div;
1349 
1350 	ts = td->td_sched;
1351 	/*
1352 	 * The score is only needed if this is likely to be an interactive
1353 	 * task.  Don't go through the expense of computing it if there's
1354 	 * no chance.
1355 	 */
1356 	if (sched_interact <= SCHED_INTERACT_HALF &&
1357 		ts->ts_runtime >= ts->ts_slptime)
1358 			return (SCHED_INTERACT_HALF);
1359 
1360 	if (ts->ts_runtime > ts->ts_slptime) {
1361 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1362 		return (SCHED_INTERACT_HALF +
1363 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1364 	}
1365 	if (ts->ts_slptime > ts->ts_runtime) {
1366 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1367 		return (ts->ts_runtime / div);
1368 	}
1369 	/* runtime == slptime */
1370 	if (ts->ts_runtime)
1371 		return (SCHED_INTERACT_HALF);
1372 
1373 	/*
1374 	 * This can happen if slptime and runtime are 0.
1375 	 */
1376 	return (0);
1377 
1378 }
1379 
1380 /*
1381  * Scale the scheduling priority according to the "interactivity" of this
1382  * process.
1383  */
1384 static void
1385 sched_priority(struct thread *td)
1386 {
1387 	int score;
1388 	int pri;
1389 
1390 	if (td->td_pri_class != PRI_TIMESHARE)
1391 		return;
1392 	/*
1393 	 * If the score is interactive we place the thread in the realtime
1394 	 * queue with a priority that is less than kernel and interrupt
1395 	 * priorities.  These threads are not subject to nice restrictions.
1396 	 *
1397 	 * Scores greater than this are placed on the normal timeshare queue
1398 	 * where the priority is partially decided by the most recent cpu
1399 	 * utilization and the rest is decided by nice value.
1400 	 *
1401 	 * The nice value of the process has a linear effect on the calculated
1402 	 * score.  Negative nice values make it easier for a thread to be
1403 	 * considered interactive.
1404 	 */
1405 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1406 	if (score < sched_interact) {
1407 		pri = PRI_MIN_REALTIME;
1408 		pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact)
1409 		    * score;
1410 		KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME,
1411 		    ("sched_priority: invalid interactive priority %d score %d",
1412 		    pri, score));
1413 	} else {
1414 		pri = SCHED_PRI_MIN;
1415 		if (td->td_sched->ts_ticks)
1416 			pri += SCHED_PRI_TICKS(td->td_sched);
1417 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1418 		KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE,
1419 		    ("sched_priority: invalid priority %d: nice %d, "
1420 		    "ticks %d ftick %d ltick %d tick pri %d",
1421 		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1422 		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1423 		    SCHED_PRI_TICKS(td->td_sched)));
1424 	}
1425 	sched_user_prio(td, pri);
1426 
1427 	return;
1428 }
1429 
1430 /*
1431  * This routine enforces a maximum limit on the amount of scheduling history
1432  * kept.  It is called after either the slptime or runtime is adjusted.  This
1433  * function is ugly due to integer math.
1434  */
1435 static void
1436 sched_interact_update(struct thread *td)
1437 {
1438 	struct td_sched *ts;
1439 	u_int sum;
1440 
1441 	ts = td->td_sched;
1442 	sum = ts->ts_runtime + ts->ts_slptime;
1443 	if (sum < SCHED_SLP_RUN_MAX)
1444 		return;
1445 	/*
1446 	 * This only happens from two places:
1447 	 * 1) We have added an unusual amount of run time from fork_exit.
1448 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1449 	 */
1450 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1451 		if (ts->ts_runtime > ts->ts_slptime) {
1452 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1453 			ts->ts_slptime = 1;
1454 		} else {
1455 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1456 			ts->ts_runtime = 1;
1457 		}
1458 		return;
1459 	}
1460 	/*
1461 	 * If we have exceeded by more than 1/5th then the algorithm below
1462 	 * will not bring us back into range.  Dividing by two here forces
1463 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1464 	 */
1465 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1466 		ts->ts_runtime /= 2;
1467 		ts->ts_slptime /= 2;
1468 		return;
1469 	}
1470 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1471 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1472 }
1473 
1474 /*
1475  * Scale back the interactivity history when a child thread is created.  The
1476  * history is inherited from the parent but the thread may behave totally
1477  * differently.  For example, a shell spawning a compiler process.  We want
1478  * to learn that the compiler is behaving badly very quickly.
1479  */
1480 static void
1481 sched_interact_fork(struct thread *td)
1482 {
1483 	int ratio;
1484 	int sum;
1485 
1486 	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1487 	if (sum > SCHED_SLP_RUN_FORK) {
1488 		ratio = sum / SCHED_SLP_RUN_FORK;
1489 		td->td_sched->ts_runtime /= ratio;
1490 		td->td_sched->ts_slptime /= ratio;
1491 	}
1492 }
1493 
1494 /*
1495  * Called from proc0_init() to setup the scheduler fields.
1496  */
1497 void
1498 schedinit(void)
1499 {
1500 
1501 	/*
1502 	 * Set up the scheduler specific parts of proc0.
1503 	 */
1504 	proc0.p_sched = NULL; /* XXX */
1505 	thread0.td_sched = &td_sched0;
1506 	td_sched0.ts_ltick = ticks;
1507 	td_sched0.ts_ftick = ticks;
1508 	td_sched0.ts_slice = sched_slice;
1509 }
1510 
1511 /*
1512  * This is only somewhat accurate since given many processes of the same
1513  * priority they will switch when their slices run out, which will be
1514  * at most sched_slice stathz ticks.
1515  */
1516 int
1517 sched_rr_interval(void)
1518 {
1519 
1520 	/* Convert sched_slice to hz */
1521 	return (hz/(realstathz/sched_slice));
1522 }
1523 
1524 /*
1525  * Update the percent cpu tracking information when it is requested or
1526  * the total history exceeds the maximum.  We keep a sliding history of
1527  * tick counts that slowly decays.  This is less precise than the 4BSD
1528  * mechanism since it happens with less regular and frequent events.
1529  */
1530 static void
1531 sched_pctcpu_update(struct td_sched *ts)
1532 {
1533 
1534 	if (ts->ts_ticks == 0)
1535 		return;
1536 	if (ticks - (hz / 10) < ts->ts_ltick &&
1537 	    SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX)
1538 		return;
1539 	/*
1540 	 * Adjust counters and watermark for pctcpu calc.
1541 	 */
1542 	if (ts->ts_ltick > ticks - SCHED_TICK_TARG)
1543 		ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) *
1544 			    SCHED_TICK_TARG;
1545 	else
1546 		ts->ts_ticks = 0;
1547 	ts->ts_ltick = ticks;
1548 	ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG;
1549 }
1550 
1551 /*
1552  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1553  * if necessary.  This is the back-end for several priority related
1554  * functions.
1555  */
1556 static void
1557 sched_thread_priority(struct thread *td, u_char prio)
1558 {
1559 	struct td_sched *ts;
1560 	struct tdq *tdq;
1561 	int oldpri;
1562 
1563 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1564 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1565 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1566 	if (td != curthread && prio > td->td_priority) {
1567 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1568 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1569 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1570 	}
1571 	ts = td->td_sched;
1572 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1573 	if (td->td_priority == prio)
1574 		return;
1575 	/*
1576 	 * If the priority has been elevated due to priority
1577 	 * propagation, we may have to move ourselves to a new
1578 	 * queue.  This could be optimized to not re-add in some
1579 	 * cases.
1580 	 */
1581 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1582 		sched_rem(td);
1583 		td->td_priority = prio;
1584 		sched_add(td, SRQ_BORROWING);
1585 		return;
1586 	}
1587 	/*
1588 	 * If the thread is currently running we may have to adjust the lowpri
1589 	 * information so other cpus are aware of our current priority.
1590 	 */
1591 	if (TD_IS_RUNNING(td)) {
1592 		tdq = TDQ_CPU(ts->ts_cpu);
1593 		oldpri = td->td_priority;
1594 		td->td_priority = prio;
1595 		if (prio < tdq->tdq_lowpri)
1596 			tdq->tdq_lowpri = prio;
1597 		else if (tdq->tdq_lowpri == oldpri)
1598 			tdq_setlowpri(tdq, td);
1599 		return;
1600 	}
1601 	td->td_priority = prio;
1602 }
1603 
1604 /*
1605  * Update a thread's priority when it is lent another thread's
1606  * priority.
1607  */
1608 void
1609 sched_lend_prio(struct thread *td, u_char prio)
1610 {
1611 
1612 	td->td_flags |= TDF_BORROWING;
1613 	sched_thread_priority(td, prio);
1614 }
1615 
1616 /*
1617  * Restore a thread's priority when priority propagation is
1618  * over.  The prio argument is the minimum priority the thread
1619  * needs to have to satisfy other possible priority lending
1620  * requests.  If the thread's regular priority is less
1621  * important than prio, the thread will keep a priority boost
1622  * of prio.
1623  */
1624 void
1625 sched_unlend_prio(struct thread *td, u_char prio)
1626 {
1627 	u_char base_pri;
1628 
1629 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1630 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1631 		base_pri = td->td_user_pri;
1632 	else
1633 		base_pri = td->td_base_pri;
1634 	if (prio >= base_pri) {
1635 		td->td_flags &= ~TDF_BORROWING;
1636 		sched_thread_priority(td, base_pri);
1637 	} else
1638 		sched_lend_prio(td, prio);
1639 }
1640 
1641 /*
1642  * Standard entry for setting the priority to an absolute value.
1643  */
1644 void
1645 sched_prio(struct thread *td, u_char prio)
1646 {
1647 	u_char oldprio;
1648 
1649 	/* First, update the base priority. */
1650 	td->td_base_pri = prio;
1651 
1652 	/*
1653 	 * If the thread is borrowing another thread's priority, don't
1654 	 * ever lower the priority.
1655 	 */
1656 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1657 		return;
1658 
1659 	/* Change the real priority. */
1660 	oldprio = td->td_priority;
1661 	sched_thread_priority(td, prio);
1662 
1663 	/*
1664 	 * If the thread is on a turnstile, then let the turnstile update
1665 	 * its state.
1666 	 */
1667 	if (TD_ON_LOCK(td) && oldprio != prio)
1668 		turnstile_adjust(td, oldprio);
1669 }
1670 
1671 /*
1672  * Set the base user priority, does not effect current running priority.
1673  */
1674 void
1675 sched_user_prio(struct thread *td, u_char prio)
1676 {
1677 	u_char oldprio;
1678 
1679 	td->td_base_user_pri = prio;
1680 	if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
1681                 return;
1682 	oldprio = td->td_user_pri;
1683 	td->td_user_pri = prio;
1684 }
1685 
1686 void
1687 sched_lend_user_prio(struct thread *td, u_char prio)
1688 {
1689 	u_char oldprio;
1690 
1691 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1692 	td->td_flags |= TDF_UBORROWING;
1693 	oldprio = td->td_user_pri;
1694 	td->td_user_pri = prio;
1695 }
1696 
1697 void
1698 sched_unlend_user_prio(struct thread *td, u_char prio)
1699 {
1700 	u_char base_pri;
1701 
1702 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1703 	base_pri = td->td_base_user_pri;
1704 	if (prio >= base_pri) {
1705 		td->td_flags &= ~TDF_UBORROWING;
1706 		sched_user_prio(td, base_pri);
1707 	} else {
1708 		sched_lend_user_prio(td, prio);
1709 	}
1710 }
1711 
1712 /*
1713  * Handle migration from sched_switch().  This happens only for
1714  * cpu binding.
1715  */
1716 static struct mtx *
1717 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1718 {
1719 	struct tdq *tdn;
1720 
1721 	tdn = TDQ_CPU(td->td_sched->ts_cpu);
1722 #ifdef SMP
1723 	tdq_load_rem(tdq, td);
1724 	/*
1725 	 * Do the lock dance required to avoid LOR.  We grab an extra
1726 	 * spinlock nesting to prevent preemption while we're
1727 	 * not holding either run-queue lock.
1728 	 */
1729 	spinlock_enter();
1730 	thread_lock_block(td);	/* This releases the lock on tdq. */
1731 
1732 	/*
1733 	 * Acquire both run-queue locks before placing the thread on the new
1734 	 * run-queue to avoid deadlocks created by placing a thread with a
1735 	 * blocked lock on the run-queue of a remote processor.  The deadlock
1736 	 * occurs when a third processor attempts to lock the two queues in
1737 	 * question while the target processor is spinning with its own
1738 	 * run-queue lock held while waiting for the blocked lock to clear.
1739 	 */
1740 	tdq_lock_pair(tdn, tdq);
1741 	tdq_add(tdn, td, flags);
1742 	tdq_notify(tdn, td);
1743 	TDQ_UNLOCK(tdn);
1744 	spinlock_exit();
1745 #endif
1746 	return (TDQ_LOCKPTR(tdn));
1747 }
1748 
1749 /*
1750  * Variadic version of thread_lock_unblock() that does not assume td_lock
1751  * is blocked.
1752  */
1753 static inline void
1754 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1755 {
1756 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1757 	    (uintptr_t)mtx);
1758 }
1759 
1760 /*
1761  * Switch threads.  This function has to handle threads coming in while
1762  * blocked for some reason, running, or idle.  It also must deal with
1763  * migrating a thread from one queue to another as running threads may
1764  * be assigned elsewhere via binding.
1765  */
1766 void
1767 sched_switch(struct thread *td, struct thread *newtd, int flags)
1768 {
1769 	struct tdq *tdq;
1770 	struct td_sched *ts;
1771 	struct mtx *mtx;
1772 	int srqflag;
1773 	int cpuid;
1774 
1775 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1776 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1777 
1778 	cpuid = PCPU_GET(cpuid);
1779 	tdq = TDQ_CPU(cpuid);
1780 	ts = td->td_sched;
1781 	mtx = td->td_lock;
1782 	ts->ts_rltick = ticks;
1783 	td->td_lastcpu = td->td_oncpu;
1784 	td->td_oncpu = NOCPU;
1785 	td->td_flags &= ~TDF_NEEDRESCHED;
1786 	td->td_owepreempt = 0;
1787 	tdq->tdq_switchcnt++;
1788 	/*
1789 	 * The lock pointer in an idle thread should never change.  Reset it
1790 	 * to CAN_RUN as well.
1791 	 */
1792 	if (TD_IS_IDLETHREAD(td)) {
1793 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1794 		TD_SET_CAN_RUN(td);
1795 	} else if (TD_IS_RUNNING(td)) {
1796 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1797 		srqflag = (flags & SW_PREEMPT) ?
1798 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1799 		    SRQ_OURSELF|SRQ_YIELDING;
1800 		if (ts->ts_cpu == cpuid)
1801 			tdq_runq_add(tdq, td, srqflag);
1802 		else
1803 			mtx = sched_switch_migrate(tdq, td, srqflag);
1804 	} else {
1805 		/* This thread must be going to sleep. */
1806 		TDQ_LOCK(tdq);
1807 		mtx = thread_lock_block(td);
1808 		tdq_load_rem(tdq, td);
1809 	}
1810 	/*
1811 	 * We enter here with the thread blocked and assigned to the
1812 	 * appropriate cpu run-queue or sleep-queue and with the current
1813 	 * thread-queue locked.
1814 	 */
1815 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1816 	newtd = choosethread();
1817 	/*
1818 	 * Call the MD code to switch contexts if necessary.
1819 	 */
1820 	if (td != newtd) {
1821 #ifdef	HWPMC_HOOKS
1822 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1823 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1824 #endif
1825 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1826 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1827 
1828 #ifdef KDTRACE_HOOKS
1829 		/*
1830 		 * If DTrace has set the active vtime enum to anything
1831 		 * other than INACTIVE (0), then it should have set the
1832 		 * function to call.
1833 		 */
1834 		if (dtrace_vtime_active)
1835 			(*dtrace_vtime_switch_func)(newtd);
1836 #endif
1837 
1838 		cpu_switch(td, newtd, mtx);
1839 		/*
1840 		 * We may return from cpu_switch on a different cpu.  However,
1841 		 * we always return with td_lock pointing to the current cpu's
1842 		 * run queue lock.
1843 		 */
1844 		cpuid = PCPU_GET(cpuid);
1845 		tdq = TDQ_CPU(cpuid);
1846 		lock_profile_obtain_lock_success(
1847 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1848 #ifdef	HWPMC_HOOKS
1849 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1850 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1851 #endif
1852 	} else
1853 		thread_unblock_switch(td, mtx);
1854 	/*
1855 	 * Assert that all went well and return.
1856 	 */
1857 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1858 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1859 	td->td_oncpu = cpuid;
1860 }
1861 
1862 /*
1863  * Adjust thread priorities as a result of a nice request.
1864  */
1865 void
1866 sched_nice(struct proc *p, int nice)
1867 {
1868 	struct thread *td;
1869 
1870 	PROC_LOCK_ASSERT(p, MA_OWNED);
1871 
1872 	p->p_nice = nice;
1873 	FOREACH_THREAD_IN_PROC(p, td) {
1874 		thread_lock(td);
1875 		sched_priority(td);
1876 		sched_prio(td, td->td_base_user_pri);
1877 		thread_unlock(td);
1878 	}
1879 }
1880 
1881 /*
1882  * Record the sleep time for the interactivity scorer.
1883  */
1884 void
1885 sched_sleep(struct thread *td, int prio)
1886 {
1887 
1888 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1889 
1890 	td->td_slptick = ticks;
1891 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1892 		td->td_flags |= TDF_CANSWAP;
1893 	if (static_boost == 1 && prio)
1894 		sched_prio(td, prio);
1895 	else if (static_boost && td->td_priority > static_boost)
1896 		sched_prio(td, static_boost);
1897 }
1898 
1899 /*
1900  * Schedule a thread to resume execution and record how long it voluntarily
1901  * slept.  We also update the pctcpu, interactivity, and priority.
1902  */
1903 void
1904 sched_wakeup(struct thread *td)
1905 {
1906 	struct td_sched *ts;
1907 	int slptick;
1908 
1909 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1910 	ts = td->td_sched;
1911 	td->td_flags &= ~TDF_CANSWAP;
1912 	/*
1913 	 * If we slept for more than a tick update our interactivity and
1914 	 * priority.
1915 	 */
1916 	slptick = td->td_slptick;
1917 	td->td_slptick = 0;
1918 	if (slptick && slptick != ticks) {
1919 		u_int hzticks;
1920 
1921 		hzticks = (ticks - slptick) << SCHED_TICK_SHIFT;
1922 		ts->ts_slptime += hzticks;
1923 		sched_interact_update(td);
1924 		sched_pctcpu_update(ts);
1925 	}
1926 	/* Reset the slice value after we sleep. */
1927 	ts->ts_slice = sched_slice;
1928 	sched_add(td, SRQ_BORING);
1929 }
1930 
1931 /*
1932  * Penalize the parent for creating a new child and initialize the child's
1933  * priority.
1934  */
1935 void
1936 sched_fork(struct thread *td, struct thread *child)
1937 {
1938 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1939 	sched_fork_thread(td, child);
1940 	/*
1941 	 * Penalize the parent and child for forking.
1942 	 */
1943 	sched_interact_fork(child);
1944 	sched_priority(child);
1945 	td->td_sched->ts_runtime += tickincr;
1946 	sched_interact_update(td);
1947 	sched_priority(td);
1948 }
1949 
1950 /*
1951  * Fork a new thread, may be within the same process.
1952  */
1953 void
1954 sched_fork_thread(struct thread *td, struct thread *child)
1955 {
1956 	struct td_sched *ts;
1957 	struct td_sched *ts2;
1958 
1959 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1960 	/*
1961 	 * Initialize child.
1962 	 */
1963 	ts = td->td_sched;
1964 	ts2 = child->td_sched;
1965 	child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
1966 	child->td_cpuset = cpuset_ref(td->td_cpuset);
1967 	ts2->ts_cpu = ts->ts_cpu;
1968 	ts2->ts_flags = 0;
1969 	/*
1970 	 * Grab our parents cpu estimation information and priority.
1971 	 */
1972 	ts2->ts_ticks = ts->ts_ticks;
1973 	ts2->ts_ltick = ts->ts_ltick;
1974 	ts2->ts_incrtick = ts->ts_incrtick;
1975 	ts2->ts_ftick = ts->ts_ftick;
1976 	child->td_user_pri = td->td_user_pri;
1977 	child->td_base_user_pri = td->td_base_user_pri;
1978 	/*
1979 	 * And update interactivity score.
1980 	 */
1981 	ts2->ts_slptime = ts->ts_slptime;
1982 	ts2->ts_runtime = ts->ts_runtime;
1983 	ts2->ts_slice = 1;	/* Attempt to quickly learn interactivity. */
1984 #ifdef KTR
1985 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
1986 #endif
1987 }
1988 
1989 /*
1990  * Adjust the priority class of a thread.
1991  */
1992 void
1993 sched_class(struct thread *td, int class)
1994 {
1995 
1996 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1997 	if (td->td_pri_class == class)
1998 		return;
1999 	td->td_pri_class = class;
2000 }
2001 
2002 /*
2003  * Return some of the child's priority and interactivity to the parent.
2004  */
2005 void
2006 sched_exit(struct proc *p, struct thread *child)
2007 {
2008 	struct thread *td;
2009 
2010 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2011 	    "prio:td", child->td_priority);
2012 	PROC_LOCK_ASSERT(p, MA_OWNED);
2013 	td = FIRST_THREAD_IN_PROC(p);
2014 	sched_exit_thread(td, child);
2015 }
2016 
2017 /*
2018  * Penalize another thread for the time spent on this one.  This helps to
2019  * worsen the priority and interactivity of processes which schedule batch
2020  * jobs such as make.  This has little effect on the make process itself but
2021  * causes new processes spawned by it to receive worse scores immediately.
2022  */
2023 void
2024 sched_exit_thread(struct thread *td, struct thread *child)
2025 {
2026 
2027 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2028 	    "prio:td", child->td_priority);
2029 	/*
2030 	 * Give the child's runtime to the parent without returning the
2031 	 * sleep time as a penalty to the parent.  This causes shells that
2032 	 * launch expensive things to mark their children as expensive.
2033 	 */
2034 	thread_lock(td);
2035 	td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2036 	sched_interact_update(td);
2037 	sched_priority(td);
2038 	thread_unlock(td);
2039 }
2040 
2041 void
2042 sched_preempt(struct thread *td)
2043 {
2044 	struct tdq *tdq;
2045 
2046 	thread_lock(td);
2047 	tdq = TDQ_SELF();
2048 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2049 	tdq->tdq_ipipending = 0;
2050 	if (td->td_priority > tdq->tdq_lowpri) {
2051 		int flags;
2052 
2053 		flags = SW_INVOL | SW_PREEMPT;
2054 		if (td->td_critnest > 1)
2055 			td->td_owepreempt = 1;
2056 		else if (TD_IS_IDLETHREAD(td))
2057 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2058 		else
2059 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2060 	}
2061 	thread_unlock(td);
2062 }
2063 
2064 /*
2065  * Fix priorities on return to user-space.  Priorities may be elevated due
2066  * to static priorities in msleep() or similar.
2067  */
2068 void
2069 sched_userret(struct thread *td)
2070 {
2071 	/*
2072 	 * XXX we cheat slightly on the locking here to avoid locking in
2073 	 * the usual case.  Setting td_priority here is essentially an
2074 	 * incomplete workaround for not setting it properly elsewhere.
2075 	 * Now that some interrupt handlers are threads, not setting it
2076 	 * properly elsewhere can clobber it in the window between setting
2077 	 * it here and returning to user mode, so don't waste time setting
2078 	 * it perfectly here.
2079 	 */
2080 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2081 	    ("thread with borrowed priority returning to userland"));
2082 	if (td->td_priority != td->td_user_pri) {
2083 		thread_lock(td);
2084 		td->td_priority = td->td_user_pri;
2085 		td->td_base_pri = td->td_user_pri;
2086 		tdq_setlowpri(TDQ_SELF(), td);
2087 		thread_unlock(td);
2088         }
2089 }
2090 
2091 /*
2092  * Handle a stathz tick.  This is really only relevant for timeshare
2093  * threads.
2094  */
2095 void
2096 sched_clock(struct thread *td)
2097 {
2098 	struct tdq *tdq;
2099 	struct td_sched *ts;
2100 
2101 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2102 	tdq = TDQ_SELF();
2103 #ifdef SMP
2104 	/*
2105 	 * We run the long term load balancer infrequently on the first cpu.
2106 	 */
2107 	if (balance_tdq == tdq) {
2108 		if (balance_ticks && --balance_ticks == 0)
2109 			sched_balance();
2110 	}
2111 #endif
2112 	/*
2113 	 * Save the old switch count so we have a record of the last ticks
2114 	 * activity.   Initialize the new switch count based on our load.
2115 	 * If there is some activity seed it to reflect that.
2116 	 */
2117 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2118 	tdq->tdq_switchcnt = tdq->tdq_load;
2119 	/*
2120 	 * Advance the insert index once for each tick to ensure that all
2121 	 * threads get a chance to run.
2122 	 */
2123 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2124 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2125 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2126 			tdq->tdq_ridx = tdq->tdq_idx;
2127 	}
2128 	ts = td->td_sched;
2129 	if (td->td_pri_class & PRI_FIFO_BIT)
2130 		return;
2131 	if (td->td_pri_class == PRI_TIMESHARE) {
2132 		/*
2133 		 * We used a tick; charge it to the thread so
2134 		 * that we can compute our interactivity.
2135 		 */
2136 		td->td_sched->ts_runtime += tickincr;
2137 		sched_interact_update(td);
2138 		sched_priority(td);
2139 	}
2140 	/*
2141 	 * We used up one time slice.
2142 	 */
2143 	if (--ts->ts_slice > 0)
2144 		return;
2145 	/*
2146 	 * We're out of time, force a requeue at userret().
2147 	 */
2148 	ts->ts_slice = sched_slice;
2149 	td->td_flags |= TDF_NEEDRESCHED;
2150 }
2151 
2152 /*
2153  * Called once per hz tick.  Used for cpu utilization information.  This
2154  * is easier than trying to scale based on stathz.
2155  */
2156 void
2157 sched_tick(void)
2158 {
2159 	struct td_sched *ts;
2160 
2161 	ts = curthread->td_sched;
2162 	/*
2163 	 * Ticks is updated asynchronously on a single cpu.  Check here to
2164 	 * avoid incrementing ts_ticks multiple times in a single tick.
2165 	 */
2166 	if (ts->ts_incrtick == ticks)
2167 		return;
2168 	/* Adjust ticks for pctcpu */
2169 	ts->ts_ticks += 1 << SCHED_TICK_SHIFT;
2170 	ts->ts_ltick = ticks;
2171 	ts->ts_incrtick = ticks;
2172 	/*
2173 	 * Update if we've exceeded our desired tick threshhold by over one
2174 	 * second.
2175 	 */
2176 	if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick)
2177 		sched_pctcpu_update(ts);
2178 }
2179 
2180 /*
2181  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2182  * cooperative idle threads.
2183  */
2184 int
2185 sched_runnable(void)
2186 {
2187 	struct tdq *tdq;
2188 	int load;
2189 
2190 	load = 1;
2191 
2192 	tdq = TDQ_SELF();
2193 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2194 		if (tdq->tdq_load > 0)
2195 			goto out;
2196 	} else
2197 		if (tdq->tdq_load - 1 > 0)
2198 			goto out;
2199 	load = 0;
2200 out:
2201 	return (load);
2202 }
2203 
2204 /*
2205  * Choose the highest priority thread to run.  The thread is removed from
2206  * the run-queue while running however the load remains.  For SMP we set
2207  * the tdq in the global idle bitmask if it idles here.
2208  */
2209 struct thread *
2210 sched_choose(void)
2211 {
2212 	struct thread *td;
2213 	struct tdq *tdq;
2214 
2215 	tdq = TDQ_SELF();
2216 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2217 	td = tdq_choose(tdq);
2218 	if (td) {
2219 		td->td_sched->ts_ltick = ticks;
2220 		tdq_runq_rem(tdq, td);
2221 		tdq->tdq_lowpri = td->td_priority;
2222 		return (td);
2223 	}
2224 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2225 	return (PCPU_GET(idlethread));
2226 }
2227 
2228 /*
2229  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2230  * we always request it once we exit a critical section.
2231  */
2232 static inline void
2233 sched_setpreempt(struct thread *td)
2234 {
2235 	struct thread *ctd;
2236 	int cpri;
2237 	int pri;
2238 
2239 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2240 
2241 	ctd = curthread;
2242 	pri = td->td_priority;
2243 	cpri = ctd->td_priority;
2244 	if (pri < cpri)
2245 		ctd->td_flags |= TDF_NEEDRESCHED;
2246 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2247 		return;
2248 	if (!sched_shouldpreempt(pri, cpri, 0))
2249 		return;
2250 	ctd->td_owepreempt = 1;
2251 }
2252 
2253 /*
2254  * Add a thread to a thread queue.  Select the appropriate runq and add the
2255  * thread to it.  This is the internal function called when the tdq is
2256  * predetermined.
2257  */
2258 void
2259 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2260 {
2261 
2262 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2263 	KASSERT((td->td_inhibitors == 0),
2264 	    ("sched_add: trying to run inhibited thread"));
2265 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2266 	    ("sched_add: bad thread state"));
2267 	KASSERT(td->td_flags & TDF_INMEM,
2268 	    ("sched_add: thread swapped out"));
2269 
2270 	if (td->td_priority < tdq->tdq_lowpri)
2271 		tdq->tdq_lowpri = td->td_priority;
2272 	tdq_runq_add(tdq, td, flags);
2273 	tdq_load_add(tdq, td);
2274 }
2275 
2276 /*
2277  * Select the target thread queue and add a thread to it.  Request
2278  * preemption or IPI a remote processor if required.
2279  */
2280 void
2281 sched_add(struct thread *td, int flags)
2282 {
2283 	struct tdq *tdq;
2284 #ifdef SMP
2285 	int cpu;
2286 #endif
2287 
2288 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2289 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2290 	    sched_tdname(curthread));
2291 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2292 	    KTR_ATTR_LINKED, sched_tdname(td));
2293 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2294 	/*
2295 	 * Recalculate the priority before we select the target cpu or
2296 	 * run-queue.
2297 	 */
2298 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2299 		sched_priority(td);
2300 #ifdef SMP
2301 	/*
2302 	 * Pick the destination cpu and if it isn't ours transfer to the
2303 	 * target cpu.
2304 	 */
2305 	cpu = sched_pickcpu(td, flags);
2306 	tdq = sched_setcpu(td, cpu, flags);
2307 	tdq_add(tdq, td, flags);
2308 	if (cpu != PCPU_GET(cpuid)) {
2309 		tdq_notify(tdq, td);
2310 		return;
2311 	}
2312 #else
2313 	tdq = TDQ_SELF();
2314 	TDQ_LOCK(tdq);
2315 	/*
2316 	 * Now that the thread is moving to the run-queue, set the lock
2317 	 * to the scheduler's lock.
2318 	 */
2319 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2320 	tdq_add(tdq, td, flags);
2321 #endif
2322 	if (!(flags & SRQ_YIELDING))
2323 		sched_setpreempt(td);
2324 }
2325 
2326 /*
2327  * Remove a thread from a run-queue without running it.  This is used
2328  * when we're stealing a thread from a remote queue.  Otherwise all threads
2329  * exit by calling sched_exit_thread() and sched_throw() themselves.
2330  */
2331 void
2332 sched_rem(struct thread *td)
2333 {
2334 	struct tdq *tdq;
2335 
2336 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2337 	    "prio:%d", td->td_priority);
2338 	tdq = TDQ_CPU(td->td_sched->ts_cpu);
2339 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2340 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2341 	KASSERT(TD_ON_RUNQ(td),
2342 	    ("sched_rem: thread not on run queue"));
2343 	tdq_runq_rem(tdq, td);
2344 	tdq_load_rem(tdq, td);
2345 	TD_SET_CAN_RUN(td);
2346 	if (td->td_priority == tdq->tdq_lowpri)
2347 		tdq_setlowpri(tdq, NULL);
2348 }
2349 
2350 /*
2351  * Fetch cpu utilization information.  Updates on demand.
2352  */
2353 fixpt_t
2354 sched_pctcpu(struct thread *td)
2355 {
2356 	fixpt_t pctcpu;
2357 	struct td_sched *ts;
2358 
2359 	pctcpu = 0;
2360 	ts = td->td_sched;
2361 	if (ts == NULL)
2362 		return (0);
2363 
2364 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2365 	if (ts->ts_ticks) {
2366 		int rtick;
2367 
2368 		sched_pctcpu_update(ts);
2369 		/* How many rtick per second ? */
2370 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2371 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2372 	}
2373 
2374 	return (pctcpu);
2375 }
2376 
2377 /*
2378  * Enforce affinity settings for a thread.  Called after adjustments to
2379  * cpumask.
2380  */
2381 void
2382 sched_affinity(struct thread *td)
2383 {
2384 #ifdef SMP
2385 	struct td_sched *ts;
2386 	int cpu;
2387 
2388 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2389 	ts = td->td_sched;
2390 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2391 		return;
2392 	if (TD_ON_RUNQ(td)) {
2393 		sched_rem(td);
2394 		sched_add(td, SRQ_BORING);
2395 		return;
2396 	}
2397 	if (!TD_IS_RUNNING(td))
2398 		return;
2399 	td->td_flags |= TDF_NEEDRESCHED;
2400 	if (!THREAD_CAN_MIGRATE(td))
2401 		return;
2402 	/*
2403 	 * Assign the new cpu and force a switch before returning to
2404 	 * userspace.  If the target thread is not running locally send
2405 	 * an ipi to force the issue.
2406 	 */
2407 	cpu = ts->ts_cpu;
2408 	ts->ts_cpu = sched_pickcpu(td, 0);
2409 	if (cpu != PCPU_GET(cpuid))
2410 		ipi_cpu(cpu, IPI_PREEMPT);
2411 #endif
2412 }
2413 
2414 /*
2415  * Bind a thread to a target cpu.
2416  */
2417 void
2418 sched_bind(struct thread *td, int cpu)
2419 {
2420 	struct td_sched *ts;
2421 
2422 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2423 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2424 	ts = td->td_sched;
2425 	if (ts->ts_flags & TSF_BOUND)
2426 		sched_unbind(td);
2427 	ts->ts_flags |= TSF_BOUND;
2428 	sched_pin();
2429 	if (PCPU_GET(cpuid) == cpu)
2430 		return;
2431 	ts->ts_cpu = cpu;
2432 	/* When we return from mi_switch we'll be on the correct cpu. */
2433 	mi_switch(SW_VOL, NULL);
2434 }
2435 
2436 /*
2437  * Release a bound thread.
2438  */
2439 void
2440 sched_unbind(struct thread *td)
2441 {
2442 	struct td_sched *ts;
2443 
2444 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2445 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2446 	ts = td->td_sched;
2447 	if ((ts->ts_flags & TSF_BOUND) == 0)
2448 		return;
2449 	ts->ts_flags &= ~TSF_BOUND;
2450 	sched_unpin();
2451 }
2452 
2453 int
2454 sched_is_bound(struct thread *td)
2455 {
2456 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2457 	return (td->td_sched->ts_flags & TSF_BOUND);
2458 }
2459 
2460 /*
2461  * Basic yield call.
2462  */
2463 void
2464 sched_relinquish(struct thread *td)
2465 {
2466 	thread_lock(td);
2467 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2468 	thread_unlock(td);
2469 }
2470 
2471 /*
2472  * Return the total system load.
2473  */
2474 int
2475 sched_load(void)
2476 {
2477 #ifdef SMP
2478 	int total;
2479 	int i;
2480 
2481 	total = 0;
2482 	CPU_FOREACH(i)
2483 		total += TDQ_CPU(i)->tdq_sysload;
2484 	return (total);
2485 #else
2486 	return (TDQ_SELF()->tdq_sysload);
2487 #endif
2488 }
2489 
2490 int
2491 sched_sizeof_proc(void)
2492 {
2493 	return (sizeof(struct proc));
2494 }
2495 
2496 int
2497 sched_sizeof_thread(void)
2498 {
2499 	return (sizeof(struct thread) + sizeof(struct td_sched));
2500 }
2501 
2502 #ifdef SMP
2503 #define	TDQ_IDLESPIN(tdq)						\
2504     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2505 #else
2506 #define	TDQ_IDLESPIN(tdq)	1
2507 #endif
2508 
2509 /*
2510  * The actual idle process.
2511  */
2512 void
2513 sched_idletd(void *dummy)
2514 {
2515 	struct thread *td;
2516 	struct tdq *tdq;
2517 	int switchcnt;
2518 	int i;
2519 
2520 	mtx_assert(&Giant, MA_NOTOWNED);
2521 	td = curthread;
2522 	tdq = TDQ_SELF();
2523 	for (;;) {
2524 #ifdef SMP
2525 		if (tdq_idled(tdq) == 0)
2526 			continue;
2527 #endif
2528 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2529 		/*
2530 		 * If we're switching very frequently, spin while checking
2531 		 * for load rather than entering a low power state that
2532 		 * may require an IPI.  However, don't do any busy
2533 		 * loops while on SMT machines as this simply steals
2534 		 * cycles from cores doing useful work.
2535 		 */
2536 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2537 			for (i = 0; i < sched_idlespins; i++) {
2538 				if (tdq->tdq_load)
2539 					break;
2540 				cpu_spinwait();
2541 			}
2542 		}
2543 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2544 		if (tdq->tdq_load == 0)
2545 			cpu_idle(switchcnt > 1);
2546 		if (tdq->tdq_load) {
2547 			thread_lock(td);
2548 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2549 			thread_unlock(td);
2550 		}
2551 	}
2552 }
2553 
2554 /*
2555  * A CPU is entering for the first time or a thread is exiting.
2556  */
2557 void
2558 sched_throw(struct thread *td)
2559 {
2560 	struct thread *newtd;
2561 	struct tdq *tdq;
2562 
2563 	tdq = TDQ_SELF();
2564 	if (td == NULL) {
2565 		/* Correct spinlock nesting and acquire the correct lock. */
2566 		TDQ_LOCK(tdq);
2567 		spinlock_exit();
2568 	} else {
2569 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2570 		tdq_load_rem(tdq, td);
2571 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2572 	}
2573 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2574 	newtd = choosethread();
2575 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2576 	PCPU_SET(switchtime, cpu_ticks());
2577 	PCPU_SET(switchticks, ticks);
2578 	cpu_throw(td, newtd);		/* doesn't return */
2579 }
2580 
2581 /*
2582  * This is called from fork_exit().  Just acquire the correct locks and
2583  * let fork do the rest of the work.
2584  */
2585 void
2586 sched_fork_exit(struct thread *td)
2587 {
2588 	struct td_sched *ts;
2589 	struct tdq *tdq;
2590 	int cpuid;
2591 
2592 	/*
2593 	 * Finish setting up thread glue so that it begins execution in a
2594 	 * non-nested critical section with the scheduler lock held.
2595 	 */
2596 	cpuid = PCPU_GET(cpuid);
2597 	tdq = TDQ_CPU(cpuid);
2598 	ts = td->td_sched;
2599 	if (TD_IS_IDLETHREAD(td))
2600 		td->td_lock = TDQ_LOCKPTR(tdq);
2601 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2602 	td->td_oncpu = cpuid;
2603 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2604 	lock_profile_obtain_lock_success(
2605 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2606 }
2607 
2608 /*
2609  * Create on first use to catch odd startup conditons.
2610  */
2611 char *
2612 sched_tdname(struct thread *td)
2613 {
2614 #ifdef KTR
2615 	struct td_sched *ts;
2616 
2617 	ts = td->td_sched;
2618 	if (ts->ts_name[0] == '\0')
2619 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2620 		    "%s tid %d", td->td_name, td->td_tid);
2621 	return (ts->ts_name);
2622 #else
2623 	return (td->td_name);
2624 #endif
2625 }
2626 
2627 #ifdef SMP
2628 
2629 /*
2630  * Build the CPU topology dump string. Is recursively called to collect
2631  * the topology tree.
2632  */
2633 static int
2634 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2635     int indent)
2636 {
2637 	int i, first;
2638 
2639 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2640 	    "", indent, cg->cg_level);
2641 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"0x%x\">", indent, "",
2642 	    cg->cg_count, cg->cg_mask);
2643 	first = TRUE;
2644 	for (i = 0; i < MAXCPU; i++) {
2645 		if ((cg->cg_mask & (1 << i)) != 0) {
2646 			if (!first)
2647 				sbuf_printf(sb, ", ");
2648 			else
2649 				first = FALSE;
2650 			sbuf_printf(sb, "%d", i);
2651 		}
2652 	}
2653 	sbuf_printf(sb, "</cpu>\n");
2654 
2655 	if (cg->cg_flags != 0) {
2656 		sbuf_printf(sb, "%*s <flags>", indent, "");
2657 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2658 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2659 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2660 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2661 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2662 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2663 		sbuf_printf(sb, "</flags>\n");
2664 	}
2665 
2666 	if (cg->cg_children > 0) {
2667 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2668 		for (i = 0; i < cg->cg_children; i++)
2669 			sysctl_kern_sched_topology_spec_internal(sb,
2670 			    &cg->cg_child[i], indent+2);
2671 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2672 	}
2673 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2674 	return (0);
2675 }
2676 
2677 /*
2678  * Sysctl handler for retrieving topology dump. It's a wrapper for
2679  * the recursive sysctl_kern_smp_topology_spec_internal().
2680  */
2681 static int
2682 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2683 {
2684 	struct sbuf *topo;
2685 	int err;
2686 
2687 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2688 
2689 	topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2690 	if (topo == NULL)
2691 		return (ENOMEM);
2692 
2693 	sbuf_printf(topo, "<groups>\n");
2694 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2695 	sbuf_printf(topo, "</groups>\n");
2696 
2697 	if (err == 0) {
2698 		sbuf_finish(topo);
2699 		err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2700 	}
2701 	sbuf_delete(topo);
2702 	return (err);
2703 }
2704 #endif
2705 
2706 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2707 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2708     "Scheduler name");
2709 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2710     "Slice size for timeshare threads");
2711 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2712      "Interactivity score threshold");
2713 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh,
2714      0,"Min priority for preemption, lower priorities have greater precedence");
2715 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost,
2716      0,"Controls whether static kernel priorities are assigned to sleeping threads.");
2717 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins,
2718      0,"Number of times idle will spin waiting for new work.");
2719 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, &sched_idlespinthresh,
2720      0,"Threshold before we will permit idle spinning.");
2721 #ifdef SMP
2722 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2723     "Number of hz ticks to keep thread affinity for");
2724 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2725     "Enables the long-term load balancer");
2726 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2727     &balance_interval, 0,
2728     "Average frequency in stathz ticks to run the long-term balancer");
2729 SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0,
2730     "Steals work from another hyper-threaded core on idle");
2731 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2732     "Attempts to steal work from other cores before idling");
2733 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2734     "Minimum load on remote cpu before we'll steal");
2735 
2736 /* Retrieve SMP topology */
2737 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2738     CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2739     "XML dump of detected CPU topology");
2740 #endif
2741 
2742 /* ps compat.  All cpu percentages from ULE are weighted. */
2743 static int ccpu = 0;
2744 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2745