xref: /freebsd/sys/kern/sched_ule.c (revision 10b59a9b4add0320d52c15ce057dd697261e7dfc)
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(__powerpc__) && defined(E500)
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	(sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + 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  * Priority ranges used for interactive and non-interactive timeshare
121  * threads.  The timeshare priorities are split up into four ranges.
122  * The first range handles interactive threads.  The last three ranges
123  * (NHALF, x, and NHALF) handle non-interactive threads with the outer
124  * ranges supporting nice values.
125  */
126 #define	PRI_TIMESHARE_RANGE	(PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
127 #define	PRI_INTERACT_RANGE	((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
128 
129 #define	PRI_MIN_INTERACT	PRI_MIN_TIMESHARE
130 #define	PRI_MAX_INTERACT	(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
131 #define	PRI_MIN_BATCH		(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
132 #define	PRI_MAX_BATCH		PRI_MAX_TIMESHARE
133 
134 /*
135  * Cpu percentage computation macros and defines.
136  *
137  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
138  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
139  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
140  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
141  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
142  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
143  */
144 #define	SCHED_TICK_SECS		10
145 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
146 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
147 #define	SCHED_TICK_SHIFT	10
148 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
149 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
150 
151 /*
152  * These macros determine priorities for non-interactive threads.  They are
153  * assigned a priority based on their recent cpu utilization as expressed
154  * by the ratio of ticks to the tick total.  NHALF priorities at the start
155  * and end of the MIN to MAX timeshare range are only reachable with negative
156  * or positive nice respectively.
157  *
158  * PRI_RANGE:	Priority range for utilization dependent priorities.
159  * PRI_NRESV:	Number of nice values.
160  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
161  * PRI_NICE:	Determines the part of the priority inherited from nice.
162  */
163 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
164 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
165 #define	SCHED_PRI_MIN		(PRI_MIN_BATCH + SCHED_PRI_NHALF)
166 #define	SCHED_PRI_MAX		(PRI_MAX_BATCH - SCHED_PRI_NHALF)
167 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
168 #define	SCHED_PRI_TICKS(ts)						\
169     (SCHED_TICK_HZ((ts)) /						\
170     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
171 #define	SCHED_PRI_NICE(nice)	(nice)
172 
173 /*
174  * These determine the interactivity of a process.  Interactivity differs from
175  * cpu utilization in that it expresses the voluntary time slept vs time ran
176  * while cpu utilization includes all time not running.  This more accurately
177  * models the intent of the thread.
178  *
179  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
180  *		before throttling back.
181  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
182  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
183  * INTERACT_THRESH:	Threshold for placement on the current runq.
184  */
185 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
186 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
187 #define	SCHED_INTERACT_MAX	(100)
188 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
189 #define	SCHED_INTERACT_THRESH	(30)
190 
191 /*
192  * tickincr:		Converts a stathz tick into a hz domain scaled by
193  *			the shift factor.  Without the shift the error rate
194  *			due to rounding would be unacceptably high.
195  * realstathz:		stathz is sometimes 0 and run off of hz.
196  * sched_slice:		Runtime of each thread before rescheduling.
197  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
198  */
199 static int sched_interact = SCHED_INTERACT_THRESH;
200 static int realstathz;
201 static int tickincr;
202 static int sched_slice = 1;
203 #ifdef PREEMPTION
204 #ifdef FULL_PREEMPTION
205 static int preempt_thresh = PRI_MAX_IDLE;
206 #else
207 static int preempt_thresh = PRI_MIN_KERN;
208 #endif
209 #else
210 static int preempt_thresh = 0;
211 #endif
212 static int static_boost = PRI_MIN_BATCH;
213 static int sched_idlespins = 10000;
214 static int sched_idlespinthresh = 16;
215 
216 /*
217  * tdq - per processor runqs and statistics.  All fields are protected by the
218  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
219  * locking in sched_pickcpu();
220  */
221 struct tdq {
222 	/* Ordered to improve efficiency of cpu_search() and switch(). */
223 	struct mtx	tdq_lock;		/* run queue lock. */
224 	struct cpu_group *tdq_cg;		/* Pointer to cpu topology. */
225 	volatile int	tdq_load;		/* Aggregate load. */
226 	volatile int	tdq_cpu_idle;		/* cpu_idle() is active. */
227 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
228 	int		tdq_transferable;	/* Transferable thread count. */
229 	short		tdq_switchcnt;		/* Switches this tick. */
230 	short		tdq_oldswitchcnt;	/* Switches last tick. */
231 	u_char		tdq_lowpri;		/* Lowest priority thread. */
232 	u_char		tdq_ipipending;		/* IPI pending. */
233 	u_char		tdq_idx;		/* Current insert index. */
234 	u_char		tdq_ridx;		/* Current removal index. */
235 	struct runq	tdq_realtime;		/* real-time run queue. */
236 	struct runq	tdq_timeshare;		/* timeshare run queue. */
237 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
238 	char		tdq_name[TDQ_NAME_LEN];
239 #ifdef KTR
240 	char		tdq_loadname[TDQ_LOADNAME_LEN];
241 #endif
242 } __aligned(64);
243 
244 /* Idle thread states and config. */
245 #define	TDQ_RUNNING	1
246 #define	TDQ_IDLE	2
247 
248 #ifdef SMP
249 struct cpu_group *cpu_top;		/* CPU topology */
250 
251 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
252 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
253 
254 /*
255  * Run-time tunables.
256  */
257 static int rebalance = 1;
258 static int balance_interval = 128;	/* Default set in sched_initticks(). */
259 static int affinity;
260 static int steal_htt = 1;
261 static int steal_idle = 1;
262 static int steal_thresh = 2;
263 
264 /*
265  * One thread queue per processor.
266  */
267 static struct tdq	tdq_cpu[MAXCPU];
268 static struct tdq	*balance_tdq;
269 static int balance_ticks;
270 
271 #define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
272 #define	TDQ_CPU(x)	(&tdq_cpu[(x)])
273 #define	TDQ_ID(x)	((int)((x) - tdq_cpu))
274 #else	/* !SMP */
275 static struct tdq	tdq_cpu;
276 
277 #define	TDQ_ID(x)	(0)
278 #define	TDQ_SELF()	(&tdq_cpu)
279 #define	TDQ_CPU(x)	(&tdq_cpu)
280 #endif
281 
282 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
283 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
284 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
285 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
286 #define	TDQ_LOCKPTR(t)		(&(t)->tdq_lock)
287 
288 static void sched_priority(struct thread *);
289 static void sched_thread_priority(struct thread *, u_char);
290 static int sched_interact_score(struct thread *);
291 static void sched_interact_update(struct thread *);
292 static void sched_interact_fork(struct thread *);
293 static void sched_pctcpu_update(struct td_sched *);
294 
295 /* Operations on per processor queues */
296 static struct thread *tdq_choose(struct tdq *);
297 static void tdq_setup(struct tdq *);
298 static void tdq_load_add(struct tdq *, struct thread *);
299 static void tdq_load_rem(struct tdq *, struct thread *);
300 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
301 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
302 static inline int sched_shouldpreempt(int, int, int);
303 void tdq_print(int cpu);
304 static void runq_print(struct runq *rq);
305 static void tdq_add(struct tdq *, struct thread *, int);
306 #ifdef SMP
307 static int tdq_move(struct tdq *, struct tdq *);
308 static int tdq_idled(struct tdq *);
309 static void tdq_notify(struct tdq *, struct thread *);
310 static struct thread *tdq_steal(struct tdq *, int);
311 static struct thread *runq_steal(struct runq *, int);
312 static int sched_pickcpu(struct thread *, int);
313 static void sched_balance(void);
314 static int sched_balance_pair(struct tdq *, struct tdq *);
315 static inline struct tdq *sched_setcpu(struct thread *, int, int);
316 static inline void thread_unblock_switch(struct thread *, struct mtx *);
317 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
318 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
319 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
320     struct cpu_group *cg, int indent);
321 #endif
322 
323 static void sched_setup(void *dummy);
324 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
325 
326 static void sched_initticks(void *dummy);
327 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
328     NULL);
329 
330 /*
331  * Print the threads waiting on a run-queue.
332  */
333 static void
334 runq_print(struct runq *rq)
335 {
336 	struct rqhead *rqh;
337 	struct thread *td;
338 	int pri;
339 	int j;
340 	int i;
341 
342 	for (i = 0; i < RQB_LEN; i++) {
343 		printf("\t\trunq bits %d 0x%zx\n",
344 		    i, rq->rq_status.rqb_bits[i]);
345 		for (j = 0; j < RQB_BPW; j++)
346 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
347 				pri = j + (i << RQB_L2BPW);
348 				rqh = &rq->rq_queues[pri];
349 				TAILQ_FOREACH(td, rqh, td_runq) {
350 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
351 					    td, td->td_name, td->td_priority,
352 					    td->td_rqindex, pri);
353 				}
354 			}
355 	}
356 }
357 
358 /*
359  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
360  */
361 void
362 tdq_print(int cpu)
363 {
364 	struct tdq *tdq;
365 
366 	tdq = TDQ_CPU(cpu);
367 
368 	printf("tdq %d:\n", TDQ_ID(tdq));
369 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
370 	printf("\tLock name:      %s\n", tdq->tdq_name);
371 	printf("\tload:           %d\n", tdq->tdq_load);
372 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
373 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
374 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
375 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
376 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
377 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
378 	printf("\trealtime runq:\n");
379 	runq_print(&tdq->tdq_realtime);
380 	printf("\ttimeshare runq:\n");
381 	runq_print(&tdq->tdq_timeshare);
382 	printf("\tidle runq:\n");
383 	runq_print(&tdq->tdq_idle);
384 }
385 
386 static inline int
387 sched_shouldpreempt(int pri, int cpri, int remote)
388 {
389 	/*
390 	 * If the new priority is not better than the current priority there is
391 	 * nothing to do.
392 	 */
393 	if (pri >= cpri)
394 		return (0);
395 	/*
396 	 * Always preempt idle.
397 	 */
398 	if (cpri >= PRI_MIN_IDLE)
399 		return (1);
400 	/*
401 	 * If preemption is disabled don't preempt others.
402 	 */
403 	if (preempt_thresh == 0)
404 		return (0);
405 	/*
406 	 * Preempt if we exceed the threshold.
407 	 */
408 	if (pri <= preempt_thresh)
409 		return (1);
410 	/*
411 	 * If we're interactive or better and there is non-interactive
412 	 * or worse running preempt only remote processors.
413 	 */
414 	if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
415 		return (1);
416 	return (0);
417 }
418 
419 #define	TS_RQ_PPQ	(((PRI_MAX_BATCH - PRI_MIN_BATCH) + 1) / RQ_NQS)
420 /*
421  * Add a thread to the actual run-queue.  Keeps transferable counts up to
422  * date with what is actually on the run-queue.  Selects the correct
423  * queue position for timeshare threads.
424  */
425 static __inline void
426 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
427 {
428 	struct td_sched *ts;
429 	u_char pri;
430 
431 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
432 	THREAD_LOCK_ASSERT(td, MA_OWNED);
433 
434 	pri = td->td_priority;
435 	ts = td->td_sched;
436 	TD_SET_RUNQ(td);
437 	if (THREAD_CAN_MIGRATE(td)) {
438 		tdq->tdq_transferable++;
439 		ts->ts_flags |= TSF_XFERABLE;
440 	}
441 	if (pri < PRI_MIN_BATCH) {
442 		ts->ts_runq = &tdq->tdq_realtime;
443 	} else if (pri <= PRI_MAX_BATCH) {
444 		ts->ts_runq = &tdq->tdq_timeshare;
445 		KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
446 			("Invalid priority %d on timeshare runq", pri));
447 		/*
448 		 * This queue contains only priorities between MIN and MAX
449 		 * realtime.  Use the whole queue to represent these values.
450 		 */
451 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
452 			pri = (pri - PRI_MIN_BATCH) / TS_RQ_PPQ;
453 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
454 			/*
455 			 * This effectively shortens the queue by one so we
456 			 * can have a one slot difference between idx and
457 			 * ridx while we wait for threads to drain.
458 			 */
459 			if (tdq->tdq_ridx != tdq->tdq_idx &&
460 			    pri == tdq->tdq_ridx)
461 				pri = (unsigned char)(pri - 1) % RQ_NQS;
462 		} else
463 			pri = tdq->tdq_ridx;
464 		runq_add_pri(ts->ts_runq, td, pri, flags);
465 		return;
466 	} else
467 		ts->ts_runq = &tdq->tdq_idle;
468 	runq_add(ts->ts_runq, td, flags);
469 }
470 
471 /*
472  * Remove a thread from a run-queue.  This typically happens when a thread
473  * is selected to run.  Running threads are not on the queue and the
474  * transferable count does not reflect them.
475  */
476 static __inline void
477 tdq_runq_rem(struct tdq *tdq, struct thread *td)
478 {
479 	struct td_sched *ts;
480 
481 	ts = td->td_sched;
482 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
483 	KASSERT(ts->ts_runq != NULL,
484 	    ("tdq_runq_remove: thread %p null ts_runq", td));
485 	if (ts->ts_flags & TSF_XFERABLE) {
486 		tdq->tdq_transferable--;
487 		ts->ts_flags &= ~TSF_XFERABLE;
488 	}
489 	if (ts->ts_runq == &tdq->tdq_timeshare) {
490 		if (tdq->tdq_idx != tdq->tdq_ridx)
491 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
492 		else
493 			runq_remove_idx(ts->ts_runq, td, NULL);
494 	} else
495 		runq_remove(ts->ts_runq, td);
496 }
497 
498 /*
499  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
500  * for this thread to the referenced thread queue.
501  */
502 static void
503 tdq_load_add(struct tdq *tdq, struct thread *td)
504 {
505 
506 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
507 	THREAD_LOCK_ASSERT(td, MA_OWNED);
508 
509 	tdq->tdq_load++;
510 	if ((td->td_flags & TDF_NOLOAD) == 0)
511 		tdq->tdq_sysload++;
512 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
513 }
514 
515 /*
516  * Remove the load from a thread that is transitioning to a sleep state or
517  * exiting.
518  */
519 static void
520 tdq_load_rem(struct tdq *tdq, struct thread *td)
521 {
522 
523 	THREAD_LOCK_ASSERT(td, MA_OWNED);
524 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
525 	KASSERT(tdq->tdq_load != 0,
526 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
527 
528 	tdq->tdq_load--;
529 	if ((td->td_flags & TDF_NOLOAD) == 0)
530 		tdq->tdq_sysload--;
531 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
532 }
533 
534 /*
535  * Set lowpri to its exact value by searching the run-queue and
536  * evaluating curthread.  curthread may be passed as an optimization.
537  */
538 static void
539 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
540 {
541 	struct thread *td;
542 
543 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
544 	if (ctd == NULL)
545 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
546 	td = tdq_choose(tdq);
547 	if (td == NULL || td->td_priority > ctd->td_priority)
548 		tdq->tdq_lowpri = ctd->td_priority;
549 	else
550 		tdq->tdq_lowpri = td->td_priority;
551 }
552 
553 #ifdef SMP
554 struct cpu_search {
555 	cpuset_t cs_mask;
556 	u_int	cs_load;
557 	u_int	cs_cpu;
558 	int	cs_limit;	/* Min priority for low min load for high. */
559 };
560 
561 #define	CPU_SEARCH_LOWEST	0x1
562 #define	CPU_SEARCH_HIGHEST	0x2
563 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
564 
565 #define	CPUSET_FOREACH(cpu, mask)				\
566 	for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++)		\
567 		if (CPU_ISSET(cpu, &mask))
568 
569 static __inline int cpu_search(struct cpu_group *cg, struct cpu_search *low,
570     struct cpu_search *high, const int match);
571 int cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low);
572 int cpu_search_highest(struct cpu_group *cg, struct cpu_search *high);
573 int cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
574     struct cpu_search *high);
575 
576 /*
577  * This routine compares according to the match argument and should be
578  * reduced in actual instantiations via constant propagation and dead code
579  * elimination.
580  */
581 static __inline int
582 cpu_compare(int cpu, struct cpu_search *low, struct cpu_search *high,
583     const int match)
584 {
585 	struct tdq *tdq;
586 
587 	tdq = TDQ_CPU(cpu);
588 	if (match & CPU_SEARCH_LOWEST)
589 		if (CPU_ISSET(cpu, &low->cs_mask) &&
590 		    tdq->tdq_load < low->cs_load &&
591 		    tdq->tdq_lowpri > low->cs_limit) {
592 			low->cs_cpu = cpu;
593 			low->cs_load = tdq->tdq_load;
594 		}
595 	if (match & CPU_SEARCH_HIGHEST)
596 		if (CPU_ISSET(cpu, &high->cs_mask) &&
597 		    tdq->tdq_load >= high->cs_limit &&
598 		    tdq->tdq_load > high->cs_load &&
599 		    tdq->tdq_transferable) {
600 			high->cs_cpu = cpu;
601 			high->cs_load = tdq->tdq_load;
602 		}
603 	return (tdq->tdq_load);
604 }
605 
606 /*
607  * Search the tree of cpu_groups for the lowest or highest loaded cpu
608  * according to the match argument.  This routine actually compares the
609  * load on all paths through the tree and finds the least loaded cpu on
610  * the least loaded path, which may differ from the least loaded cpu in
611  * the system.  This balances work among caches and busses.
612  *
613  * This inline is instantiated in three forms below using constants for the
614  * match argument.  It is reduced to the minimum set for each case.  It is
615  * also recursive to the depth of the tree.
616  */
617 static __inline int
618 cpu_search(struct cpu_group *cg, struct cpu_search *low,
619     struct cpu_search *high, const int match)
620 {
621 	int total;
622 
623 	total = 0;
624 	if (cg->cg_children) {
625 		struct cpu_search lgroup;
626 		struct cpu_search hgroup;
627 		struct cpu_group *child;
628 		u_int lload;
629 		int hload;
630 		int load;
631 		int i;
632 
633 		lload = -1;
634 		hload = -1;
635 		for (i = 0; i < cg->cg_children; i++) {
636 			child = &cg->cg_child[i];
637 			if (match & CPU_SEARCH_LOWEST) {
638 				lgroup = *low;
639 				lgroup.cs_load = -1;
640 			}
641 			if (match & CPU_SEARCH_HIGHEST) {
642 				hgroup = *high;
643 				lgroup.cs_load = 0;
644 			}
645 			switch (match) {
646 			case CPU_SEARCH_LOWEST:
647 				load = cpu_search_lowest(child, &lgroup);
648 				break;
649 			case CPU_SEARCH_HIGHEST:
650 				load = cpu_search_highest(child, &hgroup);
651 				break;
652 			case CPU_SEARCH_BOTH:
653 				load = cpu_search_both(child, &lgroup, &hgroup);
654 				break;
655 			}
656 			total += load;
657 			if (match & CPU_SEARCH_LOWEST)
658 				if (load < lload || low->cs_cpu == -1) {
659 					*low = lgroup;
660 					lload = load;
661 				}
662 			if (match & CPU_SEARCH_HIGHEST)
663 				if (load > hload || high->cs_cpu == -1) {
664 					hload = load;
665 					*high = hgroup;
666 				}
667 		}
668 	} else {
669 		int cpu;
670 
671 		CPUSET_FOREACH(cpu, cg->cg_mask)
672 			total += cpu_compare(cpu, low, high, match);
673 	}
674 	return (total);
675 }
676 
677 /*
678  * cpu_search instantiations must pass constants to maintain the inline
679  * optimization.
680  */
681 int
682 cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low)
683 {
684 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
685 }
686 
687 int
688 cpu_search_highest(struct cpu_group *cg, struct cpu_search *high)
689 {
690 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
691 }
692 
693 int
694 cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
695     struct cpu_search *high)
696 {
697 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
698 }
699 
700 /*
701  * Find the cpu with the least load via the least loaded path that has a
702  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
703  * acceptable.
704  */
705 static inline int
706 sched_lowest(struct cpu_group *cg, cpuset_t mask, int pri)
707 {
708 	struct cpu_search low;
709 
710 	low.cs_cpu = -1;
711 	low.cs_load = -1;
712 	low.cs_mask = mask;
713 	low.cs_limit = pri;
714 	cpu_search_lowest(cg, &low);
715 	return low.cs_cpu;
716 }
717 
718 /*
719  * Find the cpu with the highest load via the highest loaded path.
720  */
721 static inline int
722 sched_highest(struct cpu_group *cg, cpuset_t mask, int minload)
723 {
724 	struct cpu_search high;
725 
726 	high.cs_cpu = -1;
727 	high.cs_load = 0;
728 	high.cs_mask = mask;
729 	high.cs_limit = minload;
730 	cpu_search_highest(cg, &high);
731 	return high.cs_cpu;
732 }
733 
734 /*
735  * Simultaneously find the highest and lowest loaded cpu reachable via
736  * cg.
737  */
738 static inline void
739 sched_both(struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu)
740 {
741 	struct cpu_search high;
742 	struct cpu_search low;
743 
744 	low.cs_cpu = -1;
745 	low.cs_limit = -1;
746 	low.cs_load = -1;
747 	low.cs_mask = mask;
748 	high.cs_load = 0;
749 	high.cs_cpu = -1;
750 	high.cs_limit = -1;
751 	high.cs_mask = mask;
752 	cpu_search_both(cg, &low, &high);
753 	*lowcpu = low.cs_cpu;
754 	*highcpu = high.cs_cpu;
755 	return;
756 }
757 
758 static void
759 sched_balance_group(struct cpu_group *cg)
760 {
761 	cpuset_t mask;
762 	int high;
763 	int low;
764 	int i;
765 
766 	CPU_FILL(&mask);
767 	for (;;) {
768 		sched_both(cg, mask, &low, &high);
769 		if (low == high || low == -1 || high == -1)
770 			break;
771 		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low)))
772 			break;
773 		/*
774 		 * If we failed to move any threads determine which cpu
775 		 * to kick out of the set and try again.
776 	 	 */
777 		if (TDQ_CPU(high)->tdq_transferable == 0)
778 			CPU_CLR(high, &mask);
779 		else
780 			CPU_CLR(low, &mask);
781 	}
782 
783 	for (i = 0; i < cg->cg_children; i++)
784 		sched_balance_group(&cg->cg_child[i]);
785 }
786 
787 static void
788 sched_balance(void)
789 {
790 	struct tdq *tdq;
791 
792 	/*
793 	 * Select a random time between .5 * balance_interval and
794 	 * 1.5 * balance_interval.
795 	 */
796 	balance_ticks = max(balance_interval / 2, 1);
797 	balance_ticks += random() % balance_interval;
798 	if (smp_started == 0 || rebalance == 0)
799 		return;
800 	tdq = TDQ_SELF();
801 	TDQ_UNLOCK(tdq);
802 	sched_balance_group(cpu_top);
803 	TDQ_LOCK(tdq);
804 }
805 
806 /*
807  * Lock two thread queues using their address to maintain lock order.
808  */
809 static void
810 tdq_lock_pair(struct tdq *one, struct tdq *two)
811 {
812 	if (one < two) {
813 		TDQ_LOCK(one);
814 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
815 	} else {
816 		TDQ_LOCK(two);
817 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
818 	}
819 }
820 
821 /*
822  * Unlock two thread queues.  Order is not important here.
823  */
824 static void
825 tdq_unlock_pair(struct tdq *one, struct tdq *two)
826 {
827 	TDQ_UNLOCK(one);
828 	TDQ_UNLOCK(two);
829 }
830 
831 /*
832  * Transfer load between two imbalanced thread queues.
833  */
834 static int
835 sched_balance_pair(struct tdq *high, struct tdq *low)
836 {
837 	int transferable;
838 	int high_load;
839 	int low_load;
840 	int moved;
841 	int move;
842 	int cpu;
843 	int diff;
844 	int i;
845 
846 	tdq_lock_pair(high, low);
847 	transferable = high->tdq_transferable;
848 	high_load = high->tdq_load;
849 	low_load = low->tdq_load;
850 	moved = 0;
851 	/*
852 	 * Determine what the imbalance is and then adjust that to how many
853 	 * threads we actually have to give up (transferable).
854 	 */
855 	if (transferable != 0) {
856 		diff = high_load - low_load;
857 		move = diff / 2;
858 		if (diff & 0x1)
859 			move++;
860 		move = min(move, transferable);
861 		for (i = 0; i < move; i++)
862 			moved += tdq_move(high, low);
863 		/*
864 		 * In case the target isn't the current cpu IPI it to force a
865 		 * reschedule with the new workload.
866 		 */
867 		cpu = TDQ_ID(low);
868 		sched_pin();
869 		if (cpu != PCPU_GET(cpuid))
870 			ipi_cpu(cpu, IPI_PREEMPT);
871 		sched_unpin();
872 	}
873 	tdq_unlock_pair(high, low);
874 	return (moved);
875 }
876 
877 /*
878  * Move a thread from one thread queue to another.
879  */
880 static int
881 tdq_move(struct tdq *from, struct tdq *to)
882 {
883 	struct td_sched *ts;
884 	struct thread *td;
885 	struct tdq *tdq;
886 	int cpu;
887 
888 	TDQ_LOCK_ASSERT(from, MA_OWNED);
889 	TDQ_LOCK_ASSERT(to, MA_OWNED);
890 
891 	tdq = from;
892 	cpu = TDQ_ID(to);
893 	td = tdq_steal(tdq, cpu);
894 	if (td == NULL)
895 		return (0);
896 	ts = td->td_sched;
897 	/*
898 	 * Although the run queue is locked the thread may be blocked.  Lock
899 	 * it to clear this and acquire the run-queue lock.
900 	 */
901 	thread_lock(td);
902 	/* Drop recursive lock on from acquired via thread_lock(). */
903 	TDQ_UNLOCK(from);
904 	sched_rem(td);
905 	ts->ts_cpu = cpu;
906 	td->td_lock = TDQ_LOCKPTR(to);
907 	tdq_add(to, td, SRQ_YIELDING);
908 	return (1);
909 }
910 
911 /*
912  * This tdq has idled.  Try to steal a thread from another cpu and switch
913  * to it.
914  */
915 static int
916 tdq_idled(struct tdq *tdq)
917 {
918 	struct cpu_group *cg;
919 	struct tdq *steal;
920 	cpuset_t mask;
921 	int thresh;
922 	int cpu;
923 
924 	if (smp_started == 0 || steal_idle == 0)
925 		return (1);
926 	CPU_FILL(&mask);
927 	CPU_CLR(PCPU_GET(cpuid), &mask);
928 	/* We don't want to be preempted while we're iterating. */
929 	spinlock_enter();
930 	for (cg = tdq->tdq_cg; cg != NULL; ) {
931 		if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
932 			thresh = steal_thresh;
933 		else
934 			thresh = 1;
935 		cpu = sched_highest(cg, mask, thresh);
936 		if (cpu == -1) {
937 			cg = cg->cg_parent;
938 			continue;
939 		}
940 		steal = TDQ_CPU(cpu);
941 		CPU_CLR(cpu, &mask);
942 		tdq_lock_pair(tdq, steal);
943 		if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
944 			tdq_unlock_pair(tdq, steal);
945 			continue;
946 		}
947 		/*
948 		 * If a thread was added while interrupts were disabled don't
949 		 * steal one here.  If we fail to acquire one due to affinity
950 		 * restrictions loop again with this cpu removed from the
951 		 * set.
952 		 */
953 		if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
954 			tdq_unlock_pair(tdq, steal);
955 			continue;
956 		}
957 		spinlock_exit();
958 		TDQ_UNLOCK(steal);
959 		mi_switch(SW_VOL | SWT_IDLE, NULL);
960 		thread_unlock(curthread);
961 
962 		return (0);
963 	}
964 	spinlock_exit();
965 	return (1);
966 }
967 
968 /*
969  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
970  */
971 static void
972 tdq_notify(struct tdq *tdq, struct thread *td)
973 {
974 	struct thread *ctd;
975 	int pri;
976 	int cpu;
977 
978 	if (tdq->tdq_ipipending)
979 		return;
980 	cpu = td->td_sched->ts_cpu;
981 	pri = td->td_priority;
982 	ctd = pcpu_find(cpu)->pc_curthread;
983 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
984 		return;
985 	if (TD_IS_IDLETHREAD(ctd)) {
986 		/*
987 		 * If the MD code has an idle wakeup routine try that before
988 		 * falling back to IPI.
989 		 */
990 		if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
991 			return;
992 	}
993 	tdq->tdq_ipipending = 1;
994 	ipi_cpu(cpu, IPI_PREEMPT);
995 }
996 
997 /*
998  * Steals load from a timeshare queue.  Honors the rotating queue head
999  * index.
1000  */
1001 static struct thread *
1002 runq_steal_from(struct runq *rq, int cpu, u_char start)
1003 {
1004 	struct rqbits *rqb;
1005 	struct rqhead *rqh;
1006 	struct thread *td;
1007 	int first;
1008 	int bit;
1009 	int pri;
1010 	int i;
1011 
1012 	rqb = &rq->rq_status;
1013 	bit = start & (RQB_BPW -1);
1014 	pri = 0;
1015 	first = 0;
1016 again:
1017 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1018 		if (rqb->rqb_bits[i] == 0)
1019 			continue;
1020 		if (bit != 0) {
1021 			for (pri = bit; pri < RQB_BPW; pri++)
1022 				if (rqb->rqb_bits[i] & (1ul << pri))
1023 					break;
1024 			if (pri >= RQB_BPW)
1025 				continue;
1026 		} else
1027 			pri = RQB_FFS(rqb->rqb_bits[i]);
1028 		pri += (i << RQB_L2BPW);
1029 		rqh = &rq->rq_queues[pri];
1030 		TAILQ_FOREACH(td, rqh, td_runq) {
1031 			if (first && THREAD_CAN_MIGRATE(td) &&
1032 			    THREAD_CAN_SCHED(td, cpu))
1033 				return (td);
1034 			first = 1;
1035 		}
1036 	}
1037 	if (start != 0) {
1038 		start = 0;
1039 		goto again;
1040 	}
1041 
1042 	return (NULL);
1043 }
1044 
1045 /*
1046  * Steals load from a standard linear queue.
1047  */
1048 static struct thread *
1049 runq_steal(struct runq *rq, int cpu)
1050 {
1051 	struct rqhead *rqh;
1052 	struct rqbits *rqb;
1053 	struct thread *td;
1054 	int word;
1055 	int bit;
1056 
1057 	rqb = &rq->rq_status;
1058 	for (word = 0; word < RQB_LEN; word++) {
1059 		if (rqb->rqb_bits[word] == 0)
1060 			continue;
1061 		for (bit = 0; bit < RQB_BPW; bit++) {
1062 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1063 				continue;
1064 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1065 			TAILQ_FOREACH(td, rqh, td_runq)
1066 				if (THREAD_CAN_MIGRATE(td) &&
1067 				    THREAD_CAN_SCHED(td, cpu))
1068 					return (td);
1069 		}
1070 	}
1071 	return (NULL);
1072 }
1073 
1074 /*
1075  * Attempt to steal a thread in priority order from a thread queue.
1076  */
1077 static struct thread *
1078 tdq_steal(struct tdq *tdq, int cpu)
1079 {
1080 	struct thread *td;
1081 
1082 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1083 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1084 		return (td);
1085 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1086 	    cpu, tdq->tdq_ridx)) != NULL)
1087 		return (td);
1088 	return (runq_steal(&tdq->tdq_idle, cpu));
1089 }
1090 
1091 /*
1092  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1093  * current lock and returns with the assigned queue locked.
1094  */
1095 static inline struct tdq *
1096 sched_setcpu(struct thread *td, int cpu, int flags)
1097 {
1098 
1099 	struct tdq *tdq;
1100 
1101 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1102 	tdq = TDQ_CPU(cpu);
1103 	td->td_sched->ts_cpu = cpu;
1104 	/*
1105 	 * If the lock matches just return the queue.
1106 	 */
1107 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1108 		return (tdq);
1109 #ifdef notyet
1110 	/*
1111 	 * If the thread isn't running its lockptr is a
1112 	 * turnstile or a sleepqueue.  We can just lock_set without
1113 	 * blocking.
1114 	 */
1115 	if (TD_CAN_RUN(td)) {
1116 		TDQ_LOCK(tdq);
1117 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1118 		return (tdq);
1119 	}
1120 #endif
1121 	/*
1122 	 * The hard case, migration, we need to block the thread first to
1123 	 * prevent order reversals with other cpus locks.
1124 	 */
1125 	spinlock_enter();
1126 	thread_lock_block(td);
1127 	TDQ_LOCK(tdq);
1128 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1129 	spinlock_exit();
1130 	return (tdq);
1131 }
1132 
1133 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1134 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1135 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1136 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1137 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1138 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1139 
1140 static int
1141 sched_pickcpu(struct thread *td, int flags)
1142 {
1143 	struct cpu_group *cg;
1144 	struct td_sched *ts;
1145 	struct tdq *tdq;
1146 	cpuset_t mask;
1147 	int self;
1148 	int pri;
1149 	int cpu;
1150 
1151 	self = PCPU_GET(cpuid);
1152 	ts = td->td_sched;
1153 	if (smp_started == 0)
1154 		return (self);
1155 	/*
1156 	 * Don't migrate a running thread from sched_switch().
1157 	 */
1158 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1159 		return (ts->ts_cpu);
1160 	/*
1161 	 * Prefer to run interrupt threads on the processors that generate
1162 	 * the interrupt.
1163 	 */
1164 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1165 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1166 		SCHED_STAT_INC(pickcpu_intrbind);
1167 		ts->ts_cpu = self;
1168 	}
1169 	/*
1170 	 * If the thread can run on the last cpu and the affinity has not
1171 	 * expired or it is idle run it there.
1172 	 */
1173 	pri = td->td_priority;
1174 	tdq = TDQ_CPU(ts->ts_cpu);
1175 	if (THREAD_CAN_SCHED(td, ts->ts_cpu)) {
1176 		if (tdq->tdq_lowpri > PRI_MIN_IDLE) {
1177 			SCHED_STAT_INC(pickcpu_idle_affinity);
1178 			return (ts->ts_cpu);
1179 		}
1180 		if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri) {
1181 			SCHED_STAT_INC(pickcpu_affinity);
1182 			return (ts->ts_cpu);
1183 		}
1184 	}
1185 	/*
1186 	 * Search for the highest level in the tree that still has affinity.
1187 	 */
1188 	cg = NULL;
1189 	for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent)
1190 		if (SCHED_AFFINITY(ts, cg->cg_level))
1191 			break;
1192 	cpu = -1;
1193 	mask = td->td_cpuset->cs_mask;
1194 	if (cg)
1195 		cpu = sched_lowest(cg, mask, pri);
1196 	if (cpu == -1)
1197 		cpu = sched_lowest(cpu_top, mask, -1);
1198 	/*
1199 	 * Compare the lowest loaded cpu to current cpu.
1200 	 */
1201 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1202 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) {
1203 		SCHED_STAT_INC(pickcpu_local);
1204 		cpu = self;
1205 	} else
1206 		SCHED_STAT_INC(pickcpu_lowest);
1207 	if (cpu != ts->ts_cpu)
1208 		SCHED_STAT_INC(pickcpu_migration);
1209 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1210 	return (cpu);
1211 }
1212 #endif
1213 
1214 /*
1215  * Pick the highest priority task we have and return it.
1216  */
1217 static struct thread *
1218 tdq_choose(struct tdq *tdq)
1219 {
1220 	struct thread *td;
1221 
1222 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1223 	td = runq_choose(&tdq->tdq_realtime);
1224 	if (td != NULL)
1225 		return (td);
1226 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1227 	if (td != NULL) {
1228 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1229 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1230 		    td->td_priority));
1231 		return (td);
1232 	}
1233 	td = runq_choose(&tdq->tdq_idle);
1234 	if (td != NULL) {
1235 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1236 		    ("tdq_choose: Invalid priority on idle queue %d",
1237 		    td->td_priority));
1238 		return (td);
1239 	}
1240 
1241 	return (NULL);
1242 }
1243 
1244 /*
1245  * Initialize a thread queue.
1246  */
1247 static void
1248 tdq_setup(struct tdq *tdq)
1249 {
1250 
1251 	if (bootverbose)
1252 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1253 	runq_init(&tdq->tdq_realtime);
1254 	runq_init(&tdq->tdq_timeshare);
1255 	runq_init(&tdq->tdq_idle);
1256 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1257 	    "sched lock %d", (int)TDQ_ID(tdq));
1258 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1259 	    MTX_SPIN | MTX_RECURSE);
1260 #ifdef KTR
1261 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1262 	    "CPU %d load", (int)TDQ_ID(tdq));
1263 #endif
1264 }
1265 
1266 #ifdef SMP
1267 static void
1268 sched_setup_smp(void)
1269 {
1270 	struct tdq *tdq;
1271 	int i;
1272 
1273 	cpu_top = smp_topo();
1274 	CPU_FOREACH(i) {
1275 		tdq = TDQ_CPU(i);
1276 		tdq_setup(tdq);
1277 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1278 		if (tdq->tdq_cg == NULL)
1279 			panic("Can't find cpu group for %d\n", i);
1280 	}
1281 	balance_tdq = TDQ_SELF();
1282 	sched_balance();
1283 }
1284 #endif
1285 
1286 /*
1287  * Setup the thread queues and initialize the topology based on MD
1288  * information.
1289  */
1290 static void
1291 sched_setup(void *dummy)
1292 {
1293 	struct tdq *tdq;
1294 
1295 	tdq = TDQ_SELF();
1296 #ifdef SMP
1297 	sched_setup_smp();
1298 #else
1299 	tdq_setup(tdq);
1300 #endif
1301 	/*
1302 	 * To avoid divide-by-zero, we set realstathz a dummy value
1303 	 * in case which sched_clock() called before sched_initticks().
1304 	 */
1305 	realstathz = hz;
1306 	sched_slice = (realstathz/10);	/* ~100ms */
1307 	tickincr = 1 << SCHED_TICK_SHIFT;
1308 
1309 	/* Add thread0's load since it's running. */
1310 	TDQ_LOCK(tdq);
1311 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1312 	tdq_load_add(tdq, &thread0);
1313 	tdq->tdq_lowpri = thread0.td_priority;
1314 	TDQ_UNLOCK(tdq);
1315 }
1316 
1317 /*
1318  * This routine determines the tickincr after stathz and hz are setup.
1319  */
1320 /* ARGSUSED */
1321 static void
1322 sched_initticks(void *dummy)
1323 {
1324 	int incr;
1325 
1326 	realstathz = stathz ? stathz : hz;
1327 	sched_slice = (realstathz/10);	/* ~100ms */
1328 
1329 	/*
1330 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1331 	 * hz not being evenly divisible by stathz on all platforms.
1332 	 */
1333 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1334 	/*
1335 	 * This does not work for values of stathz that are more than
1336 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1337 	 */
1338 	if (incr == 0)
1339 		incr = 1;
1340 	tickincr = incr;
1341 #ifdef SMP
1342 	/*
1343 	 * Set the default balance interval now that we know
1344 	 * what realstathz is.
1345 	 */
1346 	balance_interval = realstathz;
1347 	/*
1348 	 * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4.
1349 	 * This prevents excess thrashing on large machines and excess idle
1350 	 * on smaller machines.
1351 	 */
1352 	steal_thresh = min(fls(mp_ncpus) - 1, 3);
1353 	affinity = SCHED_AFFINITY_DEFAULT;
1354 #endif
1355 }
1356 
1357 
1358 /*
1359  * This is the core of the interactivity algorithm.  Determines a score based
1360  * on past behavior.  It is the ratio of sleep time to run time scaled to
1361  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1362  * differs from the cpu usage because it does not account for time spent
1363  * waiting on a run-queue.  Would be prettier if we had floating point.
1364  */
1365 static int
1366 sched_interact_score(struct thread *td)
1367 {
1368 	struct td_sched *ts;
1369 	int div;
1370 
1371 	ts = td->td_sched;
1372 	/*
1373 	 * The score is only needed if this is likely to be an interactive
1374 	 * task.  Don't go through the expense of computing it if there's
1375 	 * no chance.
1376 	 */
1377 	if (sched_interact <= SCHED_INTERACT_HALF &&
1378 		ts->ts_runtime >= ts->ts_slptime)
1379 			return (SCHED_INTERACT_HALF);
1380 
1381 	if (ts->ts_runtime > ts->ts_slptime) {
1382 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1383 		return (SCHED_INTERACT_HALF +
1384 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1385 	}
1386 	if (ts->ts_slptime > ts->ts_runtime) {
1387 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1388 		return (ts->ts_runtime / div);
1389 	}
1390 	/* runtime == slptime */
1391 	if (ts->ts_runtime)
1392 		return (SCHED_INTERACT_HALF);
1393 
1394 	/*
1395 	 * This can happen if slptime and runtime are 0.
1396 	 */
1397 	return (0);
1398 
1399 }
1400 
1401 /*
1402  * Scale the scheduling priority according to the "interactivity" of this
1403  * process.
1404  */
1405 static void
1406 sched_priority(struct thread *td)
1407 {
1408 	int score;
1409 	int pri;
1410 
1411 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1412 		return;
1413 	/*
1414 	 * If the score is interactive we place the thread in the realtime
1415 	 * queue with a priority that is less than kernel and interrupt
1416 	 * priorities.  These threads are not subject to nice restrictions.
1417 	 *
1418 	 * Scores greater than this are placed on the normal timeshare queue
1419 	 * where the priority is partially decided by the most recent cpu
1420 	 * utilization and the rest is decided by nice value.
1421 	 *
1422 	 * The nice value of the process has a linear effect on the calculated
1423 	 * score.  Negative nice values make it easier for a thread to be
1424 	 * considered interactive.
1425 	 */
1426 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1427 	if (score < sched_interact) {
1428 		pri = PRI_MIN_INTERACT;
1429 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1430 		    sched_interact) * score;
1431 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1432 		    ("sched_priority: invalid interactive priority %d score %d",
1433 		    pri, score));
1434 	} else {
1435 		pri = SCHED_PRI_MIN;
1436 		if (td->td_sched->ts_ticks)
1437 			pri += SCHED_PRI_TICKS(td->td_sched);
1438 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1439 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1440 		    ("sched_priority: invalid priority %d: nice %d, "
1441 		    "ticks %d ftick %d ltick %d tick pri %d",
1442 		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1443 		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1444 		    SCHED_PRI_TICKS(td->td_sched)));
1445 	}
1446 	sched_user_prio(td, pri);
1447 
1448 	return;
1449 }
1450 
1451 /*
1452  * This routine enforces a maximum limit on the amount of scheduling history
1453  * kept.  It is called after either the slptime or runtime is adjusted.  This
1454  * function is ugly due to integer math.
1455  */
1456 static void
1457 sched_interact_update(struct thread *td)
1458 {
1459 	struct td_sched *ts;
1460 	u_int sum;
1461 
1462 	ts = td->td_sched;
1463 	sum = ts->ts_runtime + ts->ts_slptime;
1464 	if (sum < SCHED_SLP_RUN_MAX)
1465 		return;
1466 	/*
1467 	 * This only happens from two places:
1468 	 * 1) We have added an unusual amount of run time from fork_exit.
1469 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1470 	 */
1471 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1472 		if (ts->ts_runtime > ts->ts_slptime) {
1473 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1474 			ts->ts_slptime = 1;
1475 		} else {
1476 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1477 			ts->ts_runtime = 1;
1478 		}
1479 		return;
1480 	}
1481 	/*
1482 	 * If we have exceeded by more than 1/5th then the algorithm below
1483 	 * will not bring us back into range.  Dividing by two here forces
1484 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1485 	 */
1486 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1487 		ts->ts_runtime /= 2;
1488 		ts->ts_slptime /= 2;
1489 		return;
1490 	}
1491 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1492 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1493 }
1494 
1495 /*
1496  * Scale back the interactivity history when a child thread is created.  The
1497  * history is inherited from the parent but the thread may behave totally
1498  * differently.  For example, a shell spawning a compiler process.  We want
1499  * to learn that the compiler is behaving badly very quickly.
1500  */
1501 static void
1502 sched_interact_fork(struct thread *td)
1503 {
1504 	int ratio;
1505 	int sum;
1506 
1507 	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1508 	if (sum > SCHED_SLP_RUN_FORK) {
1509 		ratio = sum / SCHED_SLP_RUN_FORK;
1510 		td->td_sched->ts_runtime /= ratio;
1511 		td->td_sched->ts_slptime /= ratio;
1512 	}
1513 }
1514 
1515 /*
1516  * Called from proc0_init() to setup the scheduler fields.
1517  */
1518 void
1519 schedinit(void)
1520 {
1521 
1522 	/*
1523 	 * Set up the scheduler specific parts of proc0.
1524 	 */
1525 	proc0.p_sched = NULL; /* XXX */
1526 	thread0.td_sched = &td_sched0;
1527 	td_sched0.ts_ltick = ticks;
1528 	td_sched0.ts_ftick = ticks;
1529 	td_sched0.ts_slice = sched_slice;
1530 }
1531 
1532 /*
1533  * This is only somewhat accurate since given many processes of the same
1534  * priority they will switch when their slices run out, which will be
1535  * at most sched_slice stathz ticks.
1536  */
1537 int
1538 sched_rr_interval(void)
1539 {
1540 
1541 	/* Convert sched_slice to hz */
1542 	return (hz/(realstathz/sched_slice));
1543 }
1544 
1545 /*
1546  * Update the percent cpu tracking information when it is requested or
1547  * the total history exceeds the maximum.  We keep a sliding history of
1548  * tick counts that slowly decays.  This is less precise than the 4BSD
1549  * mechanism since it happens with less regular and frequent events.
1550  */
1551 static void
1552 sched_pctcpu_update(struct td_sched *ts)
1553 {
1554 
1555 	if (ts->ts_ticks == 0)
1556 		return;
1557 	if (ticks - (hz / 10) < ts->ts_ltick &&
1558 	    SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX)
1559 		return;
1560 	/*
1561 	 * Adjust counters and watermark for pctcpu calc.
1562 	 */
1563 	if (ts->ts_ltick > ticks - SCHED_TICK_TARG)
1564 		ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) *
1565 			    SCHED_TICK_TARG;
1566 	else
1567 		ts->ts_ticks = 0;
1568 	ts->ts_ltick = ticks;
1569 	ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG;
1570 }
1571 
1572 /*
1573  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1574  * if necessary.  This is the back-end for several priority related
1575  * functions.
1576  */
1577 static void
1578 sched_thread_priority(struct thread *td, u_char prio)
1579 {
1580 	struct td_sched *ts;
1581 	struct tdq *tdq;
1582 	int oldpri;
1583 
1584 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1585 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1586 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1587 	if (td != curthread && prio > td->td_priority) {
1588 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1589 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1590 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1591 	}
1592 	ts = td->td_sched;
1593 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1594 	if (td->td_priority == prio)
1595 		return;
1596 	/*
1597 	 * If the priority has been elevated due to priority
1598 	 * propagation, we may have to move ourselves to a new
1599 	 * queue.  This could be optimized to not re-add in some
1600 	 * cases.
1601 	 */
1602 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1603 		sched_rem(td);
1604 		td->td_priority = prio;
1605 		sched_add(td, SRQ_BORROWING);
1606 		return;
1607 	}
1608 	/*
1609 	 * If the thread is currently running we may have to adjust the lowpri
1610 	 * information so other cpus are aware of our current priority.
1611 	 */
1612 	if (TD_IS_RUNNING(td)) {
1613 		tdq = TDQ_CPU(ts->ts_cpu);
1614 		oldpri = td->td_priority;
1615 		td->td_priority = prio;
1616 		if (prio < tdq->tdq_lowpri)
1617 			tdq->tdq_lowpri = prio;
1618 		else if (tdq->tdq_lowpri == oldpri)
1619 			tdq_setlowpri(tdq, td);
1620 		return;
1621 	}
1622 	td->td_priority = prio;
1623 }
1624 
1625 /*
1626  * Update a thread's priority when it is lent another thread's
1627  * priority.
1628  */
1629 void
1630 sched_lend_prio(struct thread *td, u_char prio)
1631 {
1632 
1633 	td->td_flags |= TDF_BORROWING;
1634 	sched_thread_priority(td, prio);
1635 }
1636 
1637 /*
1638  * Restore a thread's priority when priority propagation is
1639  * over.  The prio argument is the minimum priority the thread
1640  * needs to have to satisfy other possible priority lending
1641  * requests.  If the thread's regular priority is less
1642  * important than prio, the thread will keep a priority boost
1643  * of prio.
1644  */
1645 void
1646 sched_unlend_prio(struct thread *td, u_char prio)
1647 {
1648 	u_char base_pri;
1649 
1650 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1651 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1652 		base_pri = td->td_user_pri;
1653 	else
1654 		base_pri = td->td_base_pri;
1655 	if (prio >= base_pri) {
1656 		td->td_flags &= ~TDF_BORROWING;
1657 		sched_thread_priority(td, base_pri);
1658 	} else
1659 		sched_lend_prio(td, prio);
1660 }
1661 
1662 /*
1663  * Standard entry for setting the priority to an absolute value.
1664  */
1665 void
1666 sched_prio(struct thread *td, u_char prio)
1667 {
1668 	u_char oldprio;
1669 
1670 	/* First, update the base priority. */
1671 	td->td_base_pri = prio;
1672 
1673 	/*
1674 	 * If the thread is borrowing another thread's priority, don't
1675 	 * ever lower the priority.
1676 	 */
1677 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1678 		return;
1679 
1680 	/* Change the real priority. */
1681 	oldprio = td->td_priority;
1682 	sched_thread_priority(td, prio);
1683 
1684 	/*
1685 	 * If the thread is on a turnstile, then let the turnstile update
1686 	 * its state.
1687 	 */
1688 	if (TD_ON_LOCK(td) && oldprio != prio)
1689 		turnstile_adjust(td, oldprio);
1690 }
1691 
1692 /*
1693  * Set the base user priority, does not effect current running priority.
1694  */
1695 void
1696 sched_user_prio(struct thread *td, u_char prio)
1697 {
1698 
1699 	td->td_base_user_pri = prio;
1700 	if (td->td_lend_user_pri <= prio)
1701 		return;
1702 	td->td_user_pri = prio;
1703 }
1704 
1705 void
1706 sched_lend_user_prio(struct thread *td, u_char prio)
1707 {
1708 
1709 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1710 	td->td_lend_user_pri = prio;
1711 	td->td_user_pri = min(prio, td->td_base_user_pri);
1712 	if (td->td_priority > td->td_user_pri)
1713 		sched_prio(td, td->td_user_pri);
1714 	else if (td->td_priority != td->td_user_pri)
1715 		td->td_flags |= TDF_NEEDRESCHED;
1716 }
1717 
1718 /*
1719  * Handle migration from sched_switch().  This happens only for
1720  * cpu binding.
1721  */
1722 static struct mtx *
1723 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1724 {
1725 	struct tdq *tdn;
1726 
1727 	tdn = TDQ_CPU(td->td_sched->ts_cpu);
1728 #ifdef SMP
1729 	tdq_load_rem(tdq, td);
1730 	/*
1731 	 * Do the lock dance required to avoid LOR.  We grab an extra
1732 	 * spinlock nesting to prevent preemption while we're
1733 	 * not holding either run-queue lock.
1734 	 */
1735 	spinlock_enter();
1736 	thread_lock_block(td);	/* This releases the lock on tdq. */
1737 
1738 	/*
1739 	 * Acquire both run-queue locks before placing the thread on the new
1740 	 * run-queue to avoid deadlocks created by placing a thread with a
1741 	 * blocked lock on the run-queue of a remote processor.  The deadlock
1742 	 * occurs when a third processor attempts to lock the two queues in
1743 	 * question while the target processor is spinning with its own
1744 	 * run-queue lock held while waiting for the blocked lock to clear.
1745 	 */
1746 	tdq_lock_pair(tdn, tdq);
1747 	tdq_add(tdn, td, flags);
1748 	tdq_notify(tdn, td);
1749 	TDQ_UNLOCK(tdn);
1750 	spinlock_exit();
1751 #endif
1752 	return (TDQ_LOCKPTR(tdn));
1753 }
1754 
1755 /*
1756  * Variadic version of thread_lock_unblock() that does not assume td_lock
1757  * is blocked.
1758  */
1759 static inline void
1760 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1761 {
1762 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1763 	    (uintptr_t)mtx);
1764 }
1765 
1766 /*
1767  * Switch threads.  This function has to handle threads coming in while
1768  * blocked for some reason, running, or idle.  It also must deal with
1769  * migrating a thread from one queue to another as running threads may
1770  * be assigned elsewhere via binding.
1771  */
1772 void
1773 sched_switch(struct thread *td, struct thread *newtd, int flags)
1774 {
1775 	struct tdq *tdq;
1776 	struct td_sched *ts;
1777 	struct mtx *mtx;
1778 	int srqflag;
1779 	int cpuid;
1780 
1781 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1782 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1783 
1784 	cpuid = PCPU_GET(cpuid);
1785 	tdq = TDQ_CPU(cpuid);
1786 	ts = td->td_sched;
1787 	mtx = td->td_lock;
1788 	ts->ts_rltick = ticks;
1789 	td->td_lastcpu = td->td_oncpu;
1790 	td->td_oncpu = NOCPU;
1791 	if (!(flags & SW_PREEMPT))
1792 		td->td_flags &= ~TDF_NEEDRESCHED;
1793 	td->td_owepreempt = 0;
1794 	tdq->tdq_switchcnt++;
1795 	/*
1796 	 * The lock pointer in an idle thread should never change.  Reset it
1797 	 * to CAN_RUN as well.
1798 	 */
1799 	if (TD_IS_IDLETHREAD(td)) {
1800 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1801 		TD_SET_CAN_RUN(td);
1802 	} else if (TD_IS_RUNNING(td)) {
1803 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1804 		srqflag = (flags & SW_PREEMPT) ?
1805 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1806 		    SRQ_OURSELF|SRQ_YIELDING;
1807 #ifdef SMP
1808 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1809 			ts->ts_cpu = sched_pickcpu(td, 0);
1810 #endif
1811 		if (ts->ts_cpu == cpuid)
1812 			tdq_runq_add(tdq, td, srqflag);
1813 		else {
1814 			KASSERT(THREAD_CAN_MIGRATE(td) ||
1815 			    (ts->ts_flags & TSF_BOUND) != 0,
1816 			    ("Thread %p shouldn't migrate", td));
1817 			mtx = sched_switch_migrate(tdq, td, srqflag);
1818 		}
1819 	} else {
1820 		/* This thread must be going to sleep. */
1821 		TDQ_LOCK(tdq);
1822 		mtx = thread_lock_block(td);
1823 		tdq_load_rem(tdq, td);
1824 	}
1825 	/*
1826 	 * We enter here with the thread blocked and assigned to the
1827 	 * appropriate cpu run-queue or sleep-queue and with the current
1828 	 * thread-queue locked.
1829 	 */
1830 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1831 	newtd = choosethread();
1832 	/*
1833 	 * Call the MD code to switch contexts if necessary.
1834 	 */
1835 	if (td != newtd) {
1836 #ifdef	HWPMC_HOOKS
1837 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1838 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1839 #endif
1840 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1841 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1842 
1843 #ifdef KDTRACE_HOOKS
1844 		/*
1845 		 * If DTrace has set the active vtime enum to anything
1846 		 * other than INACTIVE (0), then it should have set the
1847 		 * function to call.
1848 		 */
1849 		if (dtrace_vtime_active)
1850 			(*dtrace_vtime_switch_func)(newtd);
1851 #endif
1852 
1853 		cpu_switch(td, newtd, mtx);
1854 		/*
1855 		 * We may return from cpu_switch on a different cpu.  However,
1856 		 * we always return with td_lock pointing to the current cpu's
1857 		 * run queue lock.
1858 		 */
1859 		cpuid = PCPU_GET(cpuid);
1860 		tdq = TDQ_CPU(cpuid);
1861 		lock_profile_obtain_lock_success(
1862 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1863 #ifdef	HWPMC_HOOKS
1864 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1865 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1866 #endif
1867 	} else
1868 		thread_unblock_switch(td, mtx);
1869 	/*
1870 	 * Assert that all went well and return.
1871 	 */
1872 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1873 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1874 	td->td_oncpu = cpuid;
1875 }
1876 
1877 /*
1878  * Adjust thread priorities as a result of a nice request.
1879  */
1880 void
1881 sched_nice(struct proc *p, int nice)
1882 {
1883 	struct thread *td;
1884 
1885 	PROC_LOCK_ASSERT(p, MA_OWNED);
1886 
1887 	p->p_nice = nice;
1888 	FOREACH_THREAD_IN_PROC(p, td) {
1889 		thread_lock(td);
1890 		sched_priority(td);
1891 		sched_prio(td, td->td_base_user_pri);
1892 		thread_unlock(td);
1893 	}
1894 }
1895 
1896 /*
1897  * Record the sleep time for the interactivity scorer.
1898  */
1899 void
1900 sched_sleep(struct thread *td, int prio)
1901 {
1902 
1903 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1904 
1905 	td->td_slptick = ticks;
1906 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1907 		td->td_flags |= TDF_CANSWAP;
1908 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1909 		return;
1910 	if (static_boost == 1 && prio)
1911 		sched_prio(td, prio);
1912 	else if (static_boost && td->td_priority > static_boost)
1913 		sched_prio(td, static_boost);
1914 }
1915 
1916 /*
1917  * Schedule a thread to resume execution and record how long it voluntarily
1918  * slept.  We also update the pctcpu, interactivity, and priority.
1919  */
1920 void
1921 sched_wakeup(struct thread *td)
1922 {
1923 	struct td_sched *ts;
1924 	int slptick;
1925 
1926 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1927 	ts = td->td_sched;
1928 	td->td_flags &= ~TDF_CANSWAP;
1929 	/*
1930 	 * If we slept for more than a tick update our interactivity and
1931 	 * priority.
1932 	 */
1933 	slptick = td->td_slptick;
1934 	td->td_slptick = 0;
1935 	if (slptick && slptick != ticks) {
1936 		u_int hzticks;
1937 
1938 		hzticks = (ticks - slptick) << SCHED_TICK_SHIFT;
1939 		ts->ts_slptime += hzticks;
1940 		sched_interact_update(td);
1941 		sched_pctcpu_update(ts);
1942 	}
1943 	/* Reset the slice value after we sleep. */
1944 	ts->ts_slice = sched_slice;
1945 	sched_add(td, SRQ_BORING);
1946 }
1947 
1948 /*
1949  * Penalize the parent for creating a new child and initialize the child's
1950  * priority.
1951  */
1952 void
1953 sched_fork(struct thread *td, struct thread *child)
1954 {
1955 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1956 	sched_fork_thread(td, child);
1957 	/*
1958 	 * Penalize the parent and child for forking.
1959 	 */
1960 	sched_interact_fork(child);
1961 	sched_priority(child);
1962 	td->td_sched->ts_runtime += tickincr;
1963 	sched_interact_update(td);
1964 	sched_priority(td);
1965 }
1966 
1967 /*
1968  * Fork a new thread, may be within the same process.
1969  */
1970 void
1971 sched_fork_thread(struct thread *td, struct thread *child)
1972 {
1973 	struct td_sched *ts;
1974 	struct td_sched *ts2;
1975 
1976 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1977 	/*
1978 	 * Initialize child.
1979 	 */
1980 	ts = td->td_sched;
1981 	ts2 = child->td_sched;
1982 	child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
1983 	child->td_cpuset = cpuset_ref(td->td_cpuset);
1984 	ts2->ts_cpu = ts->ts_cpu;
1985 	ts2->ts_flags = 0;
1986 	/*
1987 	 * Grab our parents cpu estimation information.
1988 	 */
1989 	ts2->ts_ticks = ts->ts_ticks;
1990 	ts2->ts_ltick = ts->ts_ltick;
1991 	ts2->ts_incrtick = ts->ts_incrtick;
1992 	ts2->ts_ftick = ts->ts_ftick;
1993 	/*
1994 	 * Do not inherit any borrowed priority from the parent.
1995 	 */
1996 	child->td_priority = child->td_base_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:%d", 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:%d", 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 (PRI_BASE(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(int cnt)
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_incrtick == ticks)
2186 		return;
2187 	/* Adjust ticks for pctcpu */
2188 	ts->ts_ticks += cnt << SCHED_TICK_SHIFT;
2189 	ts->ts_ltick = ticks;
2190 	ts->ts_incrtick = ticks;
2191 	/*
2192 	 * Update if we've exceeded our desired tick threshold by over one
2193 	 * second.
2194 	 */
2195 	if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick)
2196 		sched_pctcpu_update(ts);
2197 }
2198 
2199 /*
2200  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2201  * cooperative idle threads.
2202  */
2203 int
2204 sched_runnable(void)
2205 {
2206 	struct tdq *tdq;
2207 	int load;
2208 
2209 	load = 1;
2210 
2211 	tdq = TDQ_SELF();
2212 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2213 		if (tdq->tdq_load > 0)
2214 			goto out;
2215 	} else
2216 		if (tdq->tdq_load - 1 > 0)
2217 			goto out;
2218 	load = 0;
2219 out:
2220 	return (load);
2221 }
2222 
2223 /*
2224  * Choose the highest priority thread to run.  The thread is removed from
2225  * the run-queue while running however the load remains.  For SMP we set
2226  * the tdq in the global idle bitmask if it idles here.
2227  */
2228 struct thread *
2229 sched_choose(void)
2230 {
2231 	struct thread *td;
2232 	struct tdq *tdq;
2233 
2234 	tdq = TDQ_SELF();
2235 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2236 	td = tdq_choose(tdq);
2237 	if (td) {
2238 		td->td_sched->ts_ltick = ticks;
2239 		tdq_runq_rem(tdq, td);
2240 		tdq->tdq_lowpri = td->td_priority;
2241 		return (td);
2242 	}
2243 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2244 	return (PCPU_GET(idlethread));
2245 }
2246 
2247 /*
2248  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2249  * we always request it once we exit a critical section.
2250  */
2251 static inline void
2252 sched_setpreempt(struct thread *td)
2253 {
2254 	struct thread *ctd;
2255 	int cpri;
2256 	int pri;
2257 
2258 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2259 
2260 	ctd = curthread;
2261 	pri = td->td_priority;
2262 	cpri = ctd->td_priority;
2263 	if (pri < cpri)
2264 		ctd->td_flags |= TDF_NEEDRESCHED;
2265 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2266 		return;
2267 	if (!sched_shouldpreempt(pri, cpri, 0))
2268 		return;
2269 	ctd->td_owepreempt = 1;
2270 }
2271 
2272 /*
2273  * Add a thread to a thread queue.  Select the appropriate runq and add the
2274  * thread to it.  This is the internal function called when the tdq is
2275  * predetermined.
2276  */
2277 void
2278 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2279 {
2280 
2281 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2282 	KASSERT((td->td_inhibitors == 0),
2283 	    ("sched_add: trying to run inhibited thread"));
2284 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2285 	    ("sched_add: bad thread state"));
2286 	KASSERT(td->td_flags & TDF_INMEM,
2287 	    ("sched_add: thread swapped out"));
2288 
2289 	if (td->td_priority < tdq->tdq_lowpri)
2290 		tdq->tdq_lowpri = td->td_priority;
2291 	tdq_runq_add(tdq, td, flags);
2292 	tdq_load_add(tdq, td);
2293 }
2294 
2295 /*
2296  * Select the target thread queue and add a thread to it.  Request
2297  * preemption or IPI a remote processor if required.
2298  */
2299 void
2300 sched_add(struct thread *td, int flags)
2301 {
2302 	struct tdq *tdq;
2303 #ifdef SMP
2304 	int cpu;
2305 #endif
2306 
2307 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2308 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2309 	    sched_tdname(curthread));
2310 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2311 	    KTR_ATTR_LINKED, sched_tdname(td));
2312 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2313 	/*
2314 	 * Recalculate the priority before we select the target cpu or
2315 	 * run-queue.
2316 	 */
2317 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2318 		sched_priority(td);
2319 #ifdef SMP
2320 	/*
2321 	 * Pick the destination cpu and if it isn't ours transfer to the
2322 	 * target cpu.
2323 	 */
2324 	cpu = sched_pickcpu(td, flags);
2325 	tdq = sched_setcpu(td, cpu, flags);
2326 	tdq_add(tdq, td, flags);
2327 	if (cpu != PCPU_GET(cpuid)) {
2328 		tdq_notify(tdq, td);
2329 		return;
2330 	}
2331 #else
2332 	tdq = TDQ_SELF();
2333 	TDQ_LOCK(tdq);
2334 	/*
2335 	 * Now that the thread is moving to the run-queue, set the lock
2336 	 * to the scheduler's lock.
2337 	 */
2338 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2339 	tdq_add(tdq, td, flags);
2340 #endif
2341 	if (!(flags & SRQ_YIELDING))
2342 		sched_setpreempt(td);
2343 }
2344 
2345 /*
2346  * Remove a thread from a run-queue without running it.  This is used
2347  * when we're stealing a thread from a remote queue.  Otherwise all threads
2348  * exit by calling sched_exit_thread() and sched_throw() themselves.
2349  */
2350 void
2351 sched_rem(struct thread *td)
2352 {
2353 	struct tdq *tdq;
2354 
2355 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2356 	    "prio:%d", td->td_priority);
2357 	tdq = TDQ_CPU(td->td_sched->ts_cpu);
2358 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2359 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2360 	KASSERT(TD_ON_RUNQ(td),
2361 	    ("sched_rem: thread not on run queue"));
2362 	tdq_runq_rem(tdq, td);
2363 	tdq_load_rem(tdq, td);
2364 	TD_SET_CAN_RUN(td);
2365 	if (td->td_priority == tdq->tdq_lowpri)
2366 		tdq_setlowpri(tdq, NULL);
2367 }
2368 
2369 /*
2370  * Fetch cpu utilization information.  Updates on demand.
2371  */
2372 fixpt_t
2373 sched_pctcpu(struct thread *td)
2374 {
2375 	fixpt_t pctcpu;
2376 	struct td_sched *ts;
2377 
2378 	pctcpu = 0;
2379 	ts = td->td_sched;
2380 	if (ts == NULL)
2381 		return (0);
2382 
2383 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2384 	if (ts->ts_ticks) {
2385 		int rtick;
2386 
2387 		sched_pctcpu_update(ts);
2388 		/* How many rtick per second ? */
2389 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2390 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2391 	}
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 
2406 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2407 	ts = td->td_sched;
2408 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2409 		return;
2410 	if (TD_ON_RUNQ(td)) {
2411 		sched_rem(td);
2412 		sched_add(td, SRQ_BORING);
2413 		return;
2414 	}
2415 	if (!TD_IS_RUNNING(td))
2416 		return;
2417 	/*
2418 	 * Force a switch before returning to userspace.  If the
2419 	 * target thread is not running locally send an ipi to force
2420 	 * the issue.
2421 	 */
2422 	td->td_flags |= TDF_NEEDRESCHED;
2423 	if (td != curthread)
2424 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2425 #endif
2426 }
2427 
2428 /*
2429  * Bind a thread to a target cpu.
2430  */
2431 void
2432 sched_bind(struct thread *td, int cpu)
2433 {
2434 	struct td_sched *ts;
2435 
2436 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2437 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2438 	ts = td->td_sched;
2439 	if (ts->ts_flags & TSF_BOUND)
2440 		sched_unbind(td);
2441 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2442 	ts->ts_flags |= TSF_BOUND;
2443 	sched_pin();
2444 	if (PCPU_GET(cpuid) == cpu)
2445 		return;
2446 	ts->ts_cpu = cpu;
2447 	/* When we return from mi_switch we'll be on the correct cpu. */
2448 	mi_switch(SW_VOL, NULL);
2449 }
2450 
2451 /*
2452  * Release a bound thread.
2453  */
2454 void
2455 sched_unbind(struct thread *td)
2456 {
2457 	struct td_sched *ts;
2458 
2459 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2460 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2461 	ts = td->td_sched;
2462 	if ((ts->ts_flags & TSF_BOUND) == 0)
2463 		return;
2464 	ts->ts_flags &= ~TSF_BOUND;
2465 	sched_unpin();
2466 }
2467 
2468 int
2469 sched_is_bound(struct thread *td)
2470 {
2471 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2472 	return (td->td_sched->ts_flags & TSF_BOUND);
2473 }
2474 
2475 /*
2476  * Basic yield call.
2477  */
2478 void
2479 sched_relinquish(struct thread *td)
2480 {
2481 	thread_lock(td);
2482 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2483 	thread_unlock(td);
2484 }
2485 
2486 /*
2487  * Return the total system load.
2488  */
2489 int
2490 sched_load(void)
2491 {
2492 #ifdef SMP
2493 	int total;
2494 	int i;
2495 
2496 	total = 0;
2497 	CPU_FOREACH(i)
2498 		total += TDQ_CPU(i)->tdq_sysload;
2499 	return (total);
2500 #else
2501 	return (TDQ_SELF()->tdq_sysload);
2502 #endif
2503 }
2504 
2505 int
2506 sched_sizeof_proc(void)
2507 {
2508 	return (sizeof(struct proc));
2509 }
2510 
2511 int
2512 sched_sizeof_thread(void)
2513 {
2514 	return (sizeof(struct thread) + sizeof(struct td_sched));
2515 }
2516 
2517 #ifdef SMP
2518 #define	TDQ_IDLESPIN(tdq)						\
2519     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2520 #else
2521 #define	TDQ_IDLESPIN(tdq)	1
2522 #endif
2523 
2524 /*
2525  * The actual idle process.
2526  */
2527 void
2528 sched_idletd(void *dummy)
2529 {
2530 	struct thread *td;
2531 	struct tdq *tdq;
2532 	int switchcnt;
2533 	int i;
2534 
2535 	mtx_assert(&Giant, MA_NOTOWNED);
2536 	td = curthread;
2537 	tdq = TDQ_SELF();
2538 	for (;;) {
2539 #ifdef SMP
2540 		if (tdq_idled(tdq) == 0)
2541 			continue;
2542 #endif
2543 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2544 		/*
2545 		 * If we're switching very frequently, spin while checking
2546 		 * for load rather than entering a low power state that
2547 		 * may require an IPI.  However, don't do any busy
2548 		 * loops while on SMT machines as this simply steals
2549 		 * cycles from cores doing useful work.
2550 		 */
2551 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2552 			for (i = 0; i < sched_idlespins; i++) {
2553 				if (tdq->tdq_load)
2554 					break;
2555 				cpu_spinwait();
2556 			}
2557 		}
2558 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2559 		if (tdq->tdq_load == 0) {
2560 			tdq->tdq_cpu_idle = 1;
2561 			if (tdq->tdq_load == 0) {
2562 				cpu_idle(switchcnt > sched_idlespinthresh * 4);
2563 				tdq->tdq_switchcnt++;
2564 			}
2565 			tdq->tdq_cpu_idle = 0;
2566 		}
2567 		if (tdq->tdq_load) {
2568 			thread_lock(td);
2569 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2570 			thread_unlock(td);
2571 		}
2572 	}
2573 }
2574 
2575 /*
2576  * A CPU is entering for the first time or a thread is exiting.
2577  */
2578 void
2579 sched_throw(struct thread *td)
2580 {
2581 	struct thread *newtd;
2582 	struct tdq *tdq;
2583 
2584 	tdq = TDQ_SELF();
2585 	if (td == NULL) {
2586 		/* Correct spinlock nesting and acquire the correct lock. */
2587 		TDQ_LOCK(tdq);
2588 		spinlock_exit();
2589 	} else {
2590 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2591 		tdq_load_rem(tdq, td);
2592 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2593 	}
2594 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2595 	newtd = choosethread();
2596 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2597 	PCPU_SET(switchtime, cpu_ticks());
2598 	PCPU_SET(switchticks, ticks);
2599 	cpu_throw(td, newtd);		/* doesn't return */
2600 }
2601 
2602 /*
2603  * This is called from fork_exit().  Just acquire the correct locks and
2604  * let fork do the rest of the work.
2605  */
2606 void
2607 sched_fork_exit(struct thread *td)
2608 {
2609 	struct td_sched *ts;
2610 	struct tdq *tdq;
2611 	int cpuid;
2612 
2613 	/*
2614 	 * Finish setting up thread glue so that it begins execution in a
2615 	 * non-nested critical section with the scheduler lock held.
2616 	 */
2617 	cpuid = PCPU_GET(cpuid);
2618 	tdq = TDQ_CPU(cpuid);
2619 	ts = td->td_sched;
2620 	if (TD_IS_IDLETHREAD(td))
2621 		td->td_lock = TDQ_LOCKPTR(tdq);
2622 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2623 	td->td_oncpu = cpuid;
2624 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2625 	lock_profile_obtain_lock_success(
2626 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2627 }
2628 
2629 /*
2630  * Create on first use to catch odd startup conditons.
2631  */
2632 char *
2633 sched_tdname(struct thread *td)
2634 {
2635 #ifdef KTR
2636 	struct td_sched *ts;
2637 
2638 	ts = td->td_sched;
2639 	if (ts->ts_name[0] == '\0')
2640 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2641 		    "%s tid %d", td->td_name, td->td_tid);
2642 	return (ts->ts_name);
2643 #else
2644 	return (td->td_name);
2645 #endif
2646 }
2647 
2648 #ifdef SMP
2649 
2650 /*
2651  * Build the CPU topology dump string. Is recursively called to collect
2652  * the topology tree.
2653  */
2654 static int
2655 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2656     int indent)
2657 {
2658 	char cpusetbuf[CPUSETBUFSIZ];
2659 	int i, first;
2660 
2661 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2662 	    "", 1 + indent / 2, cg->cg_level);
2663 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2664 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2665 	first = TRUE;
2666 	for (i = 0; i < MAXCPU; i++) {
2667 		if (CPU_ISSET(i, &cg->cg_mask)) {
2668 			if (!first)
2669 				sbuf_printf(sb, ", ");
2670 			else
2671 				first = FALSE;
2672 			sbuf_printf(sb, "%d", i);
2673 		}
2674 	}
2675 	sbuf_printf(sb, "</cpu>\n");
2676 
2677 	if (cg->cg_flags != 0) {
2678 		sbuf_printf(sb, "%*s <flags>", indent, "");
2679 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2680 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2681 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2682 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2683 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2684 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2685 		sbuf_printf(sb, "</flags>\n");
2686 	}
2687 
2688 	if (cg->cg_children > 0) {
2689 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2690 		for (i = 0; i < cg->cg_children; i++)
2691 			sysctl_kern_sched_topology_spec_internal(sb,
2692 			    &cg->cg_child[i], indent+2);
2693 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2694 	}
2695 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2696 	return (0);
2697 }
2698 
2699 /*
2700  * Sysctl handler for retrieving topology dump. It's a wrapper for
2701  * the recursive sysctl_kern_smp_topology_spec_internal().
2702  */
2703 static int
2704 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2705 {
2706 	struct sbuf *topo;
2707 	int err;
2708 
2709 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2710 
2711 	topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2712 	if (topo == NULL)
2713 		return (ENOMEM);
2714 
2715 	sbuf_printf(topo, "<groups>\n");
2716 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2717 	sbuf_printf(topo, "</groups>\n");
2718 
2719 	if (err == 0) {
2720 		sbuf_finish(topo);
2721 		err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2722 	}
2723 	sbuf_delete(topo);
2724 	return (err);
2725 }
2726 
2727 #endif
2728 
2729 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2730 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2731     "Scheduler name");
2732 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2733     "Slice size for timeshare threads");
2734 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2735      "Interactivity score threshold");
2736 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh,
2737      0,"Min priority for preemption, lower priorities have greater precedence");
2738 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost,
2739      0,"Controls whether static kernel priorities are assigned to sleeping threads.");
2740 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins,
2741      0,"Number of times idle will spin waiting for new work.");
2742 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, &sched_idlespinthresh,
2743      0,"Threshold before we will permit idle spinning.");
2744 #ifdef SMP
2745 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2746     "Number of hz ticks to keep thread affinity for");
2747 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2748     "Enables the long-term load balancer");
2749 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2750     &balance_interval, 0,
2751     "Average frequency in stathz ticks to run the long-term balancer");
2752 SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0,
2753     "Steals work from another hyper-threaded core on idle");
2754 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2755     "Attempts to steal work from other cores before idling");
2756 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2757     "Minimum load on remote cpu before we'll steal");
2758 
2759 /* Retrieve SMP topology */
2760 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2761     CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2762     "XML dump of detected CPU topology");
2763 
2764 #endif
2765 
2766 /* ps compat.  All cpu percentages from ULE are weighted. */
2767 static int ccpu = 0;
2768 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2769