xref: /freebsd/sys/kern/sched_ule.c (revision 2f3d68cd094257653be865d20bbdf0bf839dfa57)
1 /*-
2  * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice unmodified, this list of conditions, and the following
10  *    disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25  */
26 
27 /*
28  * This file implements the ULE scheduler.  ULE supports independent CPU
29  * run queues and fine grain locking.  It has superior interactive
30  * performance under load even on uni-processor systems.
31  *
32  * etymology:
33  *   ULE is the last three letters in schedule.  It owes its name to a
34  * generic user created for a scheduling system by Paul Mikesell at
35  * Isilon Systems and a general lack of creativity on the part of the author.
36  */
37 
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
40 
41 #include "opt_hwpmc_hooks.h"
42 #include "opt_kdtrace.h"
43 #include "opt_sched.h"
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/kdb.h>
48 #include <sys/kernel.h>
49 #include <sys/ktr.h>
50 #include <sys/lock.h>
51 #include <sys/mutex.h>
52 #include <sys/proc.h>
53 #include <sys/resource.h>
54 #include <sys/resourcevar.h>
55 #include <sys/sched.h>
56 #include <sys/smp.h>
57 #include <sys/sx.h>
58 #include <sys/sysctl.h>
59 #include <sys/sysproto.h>
60 #include <sys/turnstile.h>
61 #include <sys/umtx.h>
62 #include <sys/vmmeter.h>
63 #include <sys/cpuset.h>
64 #include <sys/sbuf.h>
65 #ifdef KTRACE
66 #include <sys/uio.h>
67 #include <sys/ktrace.h>
68 #endif
69 
70 #ifdef HWPMC_HOOKS
71 #include <sys/pmckern.h>
72 #endif
73 
74 #ifdef KDTRACE_HOOKS
75 #include <sys/dtrace_bsd.h>
76 int				dtrace_vtime_active;
77 dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
78 #endif
79 
80 #include <machine/cpu.h>
81 #include <machine/smp.h>
82 
83 #if defined(__sparc64__) || defined(__mips__)
84 #error "This architecture is not currently compatible with ULE"
85 #endif
86 
87 #define	KTR_ULE	0
88 
89 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
90 #define	TDQ_NAME_LEN	(sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
91 #define	TDQ_LOADNAME_LEN	(PCPU_NAME_LEN + sizeof(" load"))
92 
93 /*
94  * Thread scheduler specific section.  All fields are protected
95  * by the thread lock.
96  */
97 struct td_sched {
98 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
99 	short		ts_flags;	/* TSF_* flags. */
100 	u_char		ts_cpu;		/* CPU that we have affinity for. */
101 	int		ts_rltick;	/* Real last tick, for affinity. */
102 	int		ts_slice;	/* Ticks of slice remaining. */
103 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
104 	u_int		ts_runtime;	/* Number of ticks we were running */
105 	int		ts_ltick;	/* Last tick that we were running on */
106 	int		ts_incrtick;	/* Last tick that we incremented on */
107 	int		ts_ftick;	/* First tick that we were running on */
108 	int		ts_ticks;	/* Tick count */
109 #ifdef KTR
110 	char		ts_name[TS_NAME_LEN];
111 #endif
112 };
113 /* flags kept in ts_flags */
114 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
115 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
116 
117 static struct td_sched td_sched0;
118 
119 #define	THREAD_CAN_MIGRATE(td)	((td)->td_pinned == 0)
120 #define	THREAD_CAN_SCHED(td, cpu)	\
121     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
122 
123 /*
124  * Cpu percentage computation macros and defines.
125  *
126  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
127  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
128  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
129  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
130  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
131  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
132  */
133 #define	SCHED_TICK_SECS		10
134 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
135 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
136 #define	SCHED_TICK_SHIFT	10
137 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
138 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
139 
140 /*
141  * These macros determine priorities for non-interactive threads.  They are
142  * assigned a priority based on their recent cpu utilization as expressed
143  * by the ratio of ticks to the tick total.  NHALF priorities at the start
144  * and end of the MIN to MAX timeshare range are only reachable with negative
145  * or positive nice respectively.
146  *
147  * PRI_RANGE:	Priority range for utilization dependent priorities.
148  * PRI_NRESV:	Number of nice values.
149  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
150  * PRI_NICE:	Determines the part of the priority inherited from nice.
151  */
152 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
153 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
154 #define	SCHED_PRI_MIN		(PRI_MIN_TIMESHARE + SCHED_PRI_NHALF)
155 #define	SCHED_PRI_MAX		(PRI_MAX_TIMESHARE - SCHED_PRI_NHALF)
156 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN)
157 #define	SCHED_PRI_TICKS(ts)						\
158     (SCHED_TICK_HZ((ts)) /						\
159     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
160 #define	SCHED_PRI_NICE(nice)	(nice)
161 
162 /*
163  * These determine the interactivity of a process.  Interactivity differs from
164  * cpu utilization in that it expresses the voluntary time slept vs time ran
165  * while cpu utilization includes all time not running.  This more accurately
166  * models the intent of the thread.
167  *
168  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
169  *		before throttling back.
170  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
171  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
172  * INTERACT_THRESH:	Threshhold for placement on the current runq.
173  */
174 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
175 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
176 #define	SCHED_INTERACT_MAX	(100)
177 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
178 #define	SCHED_INTERACT_THRESH	(30)
179 
180 /*
181  * tickincr:		Converts a stathz tick into a hz domain scaled by
182  *			the shift factor.  Without the shift the error rate
183  *			due to rounding would be unacceptably high.
184  * realstathz:		stathz is sometimes 0 and run off of hz.
185  * sched_slice:		Runtime of each thread before rescheduling.
186  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
187  */
188 static int sched_interact = SCHED_INTERACT_THRESH;
189 static int realstathz;
190 static int tickincr;
191 static int sched_slice = 1;
192 #ifdef PREEMPTION
193 #ifdef FULL_PREEMPTION
194 static int preempt_thresh = PRI_MAX_IDLE;
195 #else
196 static int preempt_thresh = PRI_MIN_KERN;
197 #endif
198 #else
199 static int preempt_thresh = 0;
200 #endif
201 static int static_boost = PRI_MIN_TIMESHARE;
202 static int sched_idlespins = 10000;
203 static int sched_idlespinthresh = 4;
204 
205 /*
206  * tdq - per processor runqs and statistics.  All fields are protected by the
207  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
208  * locking in sched_pickcpu();
209  */
210 struct tdq {
211 	/* Ordered to improve efficiency of cpu_search() and switch(). */
212 	struct mtx	tdq_lock;		/* run queue lock. */
213 	struct cpu_group *tdq_cg;		/* Pointer to cpu topology. */
214 	volatile int	tdq_load;		/* Aggregate load. */
215 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
216 	int		tdq_transferable;	/* Transferable thread count. */
217 	short		tdq_switchcnt;		/* Switches this tick. */
218 	short		tdq_oldswitchcnt;	/* Switches last tick. */
219 	u_char		tdq_lowpri;		/* Lowest priority thread. */
220 	u_char		tdq_ipipending;		/* IPI pending. */
221 	u_char		tdq_idx;		/* Current insert index. */
222 	u_char		tdq_ridx;		/* Current removal index. */
223 	struct runq	tdq_realtime;		/* real-time run queue. */
224 	struct runq	tdq_timeshare;		/* timeshare run queue. */
225 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
226 	char		tdq_name[TDQ_NAME_LEN];
227 #ifdef KTR
228 	char		tdq_loadname[TDQ_LOADNAME_LEN];
229 #endif
230 } __aligned(64);
231 
232 /* Idle thread states and config. */
233 #define	TDQ_RUNNING	1
234 #define	TDQ_IDLE	2
235 
236 #ifdef SMP
237 struct cpu_group *cpu_top;		/* CPU topology */
238 
239 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
240 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
241 
242 /*
243  * Run-time tunables.
244  */
245 static int rebalance = 1;
246 static int balance_interval = 128;	/* Default set in sched_initticks(). */
247 static int affinity;
248 static int steal_htt = 1;
249 static int steal_idle = 1;
250 static int steal_thresh = 2;
251 
252 /*
253  * One thread queue per processor.
254  */
255 static struct tdq	tdq_cpu[MAXCPU];
256 static struct tdq	*balance_tdq;
257 static int balance_ticks;
258 
259 #define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
260 #define	TDQ_CPU(x)	(&tdq_cpu[(x)])
261 #define	TDQ_ID(x)	((int)((x) - tdq_cpu))
262 #else	/* !SMP */
263 static struct tdq	tdq_cpu;
264 
265 #define	TDQ_ID(x)	(0)
266 #define	TDQ_SELF()	(&tdq_cpu)
267 #define	TDQ_CPU(x)	(&tdq_cpu)
268 #endif
269 
270 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
271 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
272 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
273 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
274 #define	TDQ_LOCKPTR(t)		(&(t)->tdq_lock)
275 
276 static void sched_priority(struct thread *);
277 static void sched_thread_priority(struct thread *, u_char);
278 static int sched_interact_score(struct thread *);
279 static void sched_interact_update(struct thread *);
280 static void sched_interact_fork(struct thread *);
281 static void sched_pctcpu_update(struct td_sched *);
282 
283 /* Operations on per processor queues */
284 static struct thread *tdq_choose(struct tdq *);
285 static void tdq_setup(struct tdq *);
286 static void tdq_load_add(struct tdq *, struct thread *);
287 static void tdq_load_rem(struct tdq *, struct thread *);
288 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
289 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
290 static inline int sched_shouldpreempt(int, int, int);
291 void tdq_print(int cpu);
292 static void runq_print(struct runq *rq);
293 static void tdq_add(struct tdq *, struct thread *, int);
294 #ifdef SMP
295 static int tdq_move(struct tdq *, struct tdq *);
296 static int tdq_idled(struct tdq *);
297 static void tdq_notify(struct tdq *, struct thread *);
298 static struct thread *tdq_steal(struct tdq *, int);
299 static struct thread *runq_steal(struct runq *, int);
300 static int sched_pickcpu(struct thread *, int);
301 static void sched_balance(void);
302 static int sched_balance_pair(struct tdq *, struct tdq *);
303 static inline struct tdq *sched_setcpu(struct thread *, int, int);
304 static inline struct mtx *thread_block_switch(struct thread *);
305 static inline void thread_unblock_switch(struct thread *, struct mtx *);
306 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
307 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
308 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
309     struct cpu_group *cg, int indent);
310 #endif
311 
312 static void sched_setup(void *dummy);
313 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
314 
315 static void sched_initticks(void *dummy);
316 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
317     NULL);
318 
319 /*
320  * Print the threads waiting on a run-queue.
321  */
322 static void
323 runq_print(struct runq *rq)
324 {
325 	struct rqhead *rqh;
326 	struct thread *td;
327 	int pri;
328 	int j;
329 	int i;
330 
331 	for (i = 0; i < RQB_LEN; i++) {
332 		printf("\t\trunq bits %d 0x%zx\n",
333 		    i, rq->rq_status.rqb_bits[i]);
334 		for (j = 0; j < RQB_BPW; j++)
335 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
336 				pri = j + (i << RQB_L2BPW);
337 				rqh = &rq->rq_queues[pri];
338 				TAILQ_FOREACH(td, rqh, td_runq) {
339 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
340 					    td, td->td_name, td->td_priority,
341 					    td->td_rqindex, pri);
342 				}
343 			}
344 	}
345 }
346 
347 /*
348  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
349  */
350 void
351 tdq_print(int cpu)
352 {
353 	struct tdq *tdq;
354 
355 	tdq = TDQ_CPU(cpu);
356 
357 	printf("tdq %d:\n", TDQ_ID(tdq));
358 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
359 	printf("\tLock name:      %s\n", tdq->tdq_name);
360 	printf("\tload:           %d\n", tdq->tdq_load);
361 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
362 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
363 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
364 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
365 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
366 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
367 	printf("\trealtime runq:\n");
368 	runq_print(&tdq->tdq_realtime);
369 	printf("\ttimeshare runq:\n");
370 	runq_print(&tdq->tdq_timeshare);
371 	printf("\tidle runq:\n");
372 	runq_print(&tdq->tdq_idle);
373 }
374 
375 static inline int
376 sched_shouldpreempt(int pri, int cpri, int remote)
377 {
378 	/*
379 	 * If the new priority is not better than the current priority there is
380 	 * nothing to do.
381 	 */
382 	if (pri >= cpri)
383 		return (0);
384 	/*
385 	 * Always preempt idle.
386 	 */
387 	if (cpri >= PRI_MIN_IDLE)
388 		return (1);
389 	/*
390 	 * If preemption is disabled don't preempt others.
391 	 */
392 	if (preempt_thresh == 0)
393 		return (0);
394 	/*
395 	 * Preempt if we exceed the threshold.
396 	 */
397 	if (pri <= preempt_thresh)
398 		return (1);
399 	/*
400 	 * If we're realtime or better and there is timeshare or worse running
401 	 * preempt only remote processors.
402 	 */
403 	if (remote && pri <= PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME)
404 		return (1);
405 	return (0);
406 }
407 
408 #define	TS_RQ_PPQ	(((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS)
409 /*
410  * Add a thread to the actual run-queue.  Keeps transferable counts up to
411  * date with what is actually on the run-queue.  Selects the correct
412  * queue position for timeshare threads.
413  */
414 static __inline void
415 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
416 {
417 	struct td_sched *ts;
418 	u_char pri;
419 
420 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
421 	THREAD_LOCK_ASSERT(td, MA_OWNED);
422 
423 	pri = td->td_priority;
424 	ts = td->td_sched;
425 	TD_SET_RUNQ(td);
426 	if (THREAD_CAN_MIGRATE(td)) {
427 		tdq->tdq_transferable++;
428 		ts->ts_flags |= TSF_XFERABLE;
429 	}
430 	if (pri <= PRI_MAX_REALTIME) {
431 		ts->ts_runq = &tdq->tdq_realtime;
432 	} else if (pri <= PRI_MAX_TIMESHARE) {
433 		ts->ts_runq = &tdq->tdq_timeshare;
434 		KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE,
435 			("Invalid priority %d on timeshare runq", pri));
436 		/*
437 		 * This queue contains only priorities between MIN and MAX
438 		 * realtime.  Use the whole queue to represent these values.
439 		 */
440 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
441 			pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ;
442 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
443 			/*
444 			 * This effectively shortens the queue by one so we
445 			 * can have a one slot difference between idx and
446 			 * ridx while we wait for threads to drain.
447 			 */
448 			if (tdq->tdq_ridx != tdq->tdq_idx &&
449 			    pri == tdq->tdq_ridx)
450 				pri = (unsigned char)(pri - 1) % RQ_NQS;
451 		} else
452 			pri = tdq->tdq_ridx;
453 		runq_add_pri(ts->ts_runq, td, pri, flags);
454 		return;
455 	} else
456 		ts->ts_runq = &tdq->tdq_idle;
457 	runq_add(ts->ts_runq, td, flags);
458 }
459 
460 /*
461  * Remove a thread from a run-queue.  This typically happens when a thread
462  * is selected to run.  Running threads are not on the queue and the
463  * transferable count does not reflect them.
464  */
465 static __inline void
466 tdq_runq_rem(struct tdq *tdq, struct thread *td)
467 {
468 	struct td_sched *ts;
469 
470 	ts = td->td_sched;
471 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
472 	KASSERT(ts->ts_runq != NULL,
473 	    ("tdq_runq_remove: thread %p null ts_runq", td));
474 	if (ts->ts_flags & TSF_XFERABLE) {
475 		tdq->tdq_transferable--;
476 		ts->ts_flags &= ~TSF_XFERABLE;
477 	}
478 	if (ts->ts_runq == &tdq->tdq_timeshare) {
479 		if (tdq->tdq_idx != tdq->tdq_ridx)
480 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
481 		else
482 			runq_remove_idx(ts->ts_runq, td, NULL);
483 	} else
484 		runq_remove(ts->ts_runq, td);
485 }
486 
487 /*
488  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
489  * for this thread to the referenced thread queue.
490  */
491 static void
492 tdq_load_add(struct tdq *tdq, struct thread *td)
493 {
494 
495 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
496 	THREAD_LOCK_ASSERT(td, MA_OWNED);
497 
498 	tdq->tdq_load++;
499 	if ((td->td_proc->p_flag & P_NOLOAD) == 0)
500 		tdq->tdq_sysload++;
501 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
502 }
503 
504 /*
505  * Remove the load from a thread that is transitioning to a sleep state or
506  * exiting.
507  */
508 static void
509 tdq_load_rem(struct tdq *tdq, struct thread *td)
510 {
511 
512 	THREAD_LOCK_ASSERT(td, MA_OWNED);
513 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
514 	KASSERT(tdq->tdq_load != 0,
515 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
516 
517 	tdq->tdq_load--;
518 	if ((td->td_proc->p_flag & P_NOLOAD) == 0)
519 		tdq->tdq_sysload--;
520 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
521 }
522 
523 /*
524  * Set lowpri to its exact value by searching the run-queue and
525  * evaluating curthread.  curthread may be passed as an optimization.
526  */
527 static void
528 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
529 {
530 	struct thread *td;
531 
532 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
533 	if (ctd == NULL)
534 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
535 	td = tdq_choose(tdq);
536 	if (td == NULL || td->td_priority > ctd->td_priority)
537 		tdq->tdq_lowpri = ctd->td_priority;
538 	else
539 		tdq->tdq_lowpri = td->td_priority;
540 }
541 
542 #ifdef SMP
543 struct cpu_search {
544 	cpuset_t cs_mask;
545 	u_int	cs_load;
546 	u_int	cs_cpu;
547 	int	cs_limit;	/* Min priority for low min load for high. */
548 };
549 
550 #define	CPU_SEARCH_LOWEST	0x1
551 #define	CPU_SEARCH_HIGHEST	0x2
552 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
553 
554 #define	CPUSET_FOREACH(cpu, mask)				\
555 	for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++)		\
556 		if ((mask) & 1 << (cpu))
557 
558 static __inline int cpu_search(struct cpu_group *cg, struct cpu_search *low,
559     struct cpu_search *high, const int match);
560 int cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low);
561 int cpu_search_highest(struct cpu_group *cg, struct cpu_search *high);
562 int cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
563     struct cpu_search *high);
564 
565 /*
566  * This routine compares according to the match argument and should be
567  * reduced in actual instantiations via constant propagation and dead code
568  * elimination.
569  */
570 static __inline int
571 cpu_compare(int cpu, struct cpu_search *low, struct cpu_search *high,
572     const int match)
573 {
574 	struct tdq *tdq;
575 
576 	tdq = TDQ_CPU(cpu);
577 	if (match & CPU_SEARCH_LOWEST)
578 		if (CPU_ISSET(cpu, &low->cs_mask) &&
579 		    tdq->tdq_load < low->cs_load &&
580 		    tdq->tdq_lowpri > low->cs_limit) {
581 			low->cs_cpu = cpu;
582 			low->cs_load = tdq->tdq_load;
583 		}
584 	if (match & CPU_SEARCH_HIGHEST)
585 		if (CPU_ISSET(cpu, &high->cs_mask) &&
586 		    tdq->tdq_load >= high->cs_limit &&
587 		    tdq->tdq_load > high->cs_load &&
588 		    tdq->tdq_transferable) {
589 			high->cs_cpu = cpu;
590 			high->cs_load = tdq->tdq_load;
591 		}
592 	return (tdq->tdq_load);
593 }
594 
595 /*
596  * Search the tree of cpu_groups for the lowest or highest loaded cpu
597  * according to the match argument.  This routine actually compares the
598  * load on all paths through the tree and finds the least loaded cpu on
599  * the least loaded path, which may differ from the least loaded cpu in
600  * the system.  This balances work among caches and busses.
601  *
602  * This inline is instantiated in three forms below using constants for the
603  * match argument.  It is reduced to the minimum set for each case.  It is
604  * also recursive to the depth of the tree.
605  */
606 static __inline int
607 cpu_search(struct cpu_group *cg, struct cpu_search *low,
608     struct cpu_search *high, const int match)
609 {
610 	int total;
611 
612 	total = 0;
613 	if (cg->cg_children) {
614 		struct cpu_search lgroup;
615 		struct cpu_search hgroup;
616 		struct cpu_group *child;
617 		u_int lload;
618 		int hload;
619 		int load;
620 		int i;
621 
622 		lload = -1;
623 		hload = -1;
624 		for (i = 0; i < cg->cg_children; i++) {
625 			child = &cg->cg_child[i];
626 			if (match & CPU_SEARCH_LOWEST) {
627 				lgroup = *low;
628 				lgroup.cs_load = -1;
629 			}
630 			if (match & CPU_SEARCH_HIGHEST) {
631 				hgroup = *high;
632 				lgroup.cs_load = 0;
633 			}
634 			switch (match) {
635 			case CPU_SEARCH_LOWEST:
636 				load = cpu_search_lowest(child, &lgroup);
637 				break;
638 			case CPU_SEARCH_HIGHEST:
639 				load = cpu_search_highest(child, &hgroup);
640 				break;
641 			case CPU_SEARCH_BOTH:
642 				load = cpu_search_both(child, &lgroup, &hgroup);
643 				break;
644 			}
645 			total += load;
646 			if (match & CPU_SEARCH_LOWEST)
647 				if (load < lload || low->cs_cpu == -1) {
648 					*low = lgroup;
649 					lload = load;
650 				}
651 			if (match & CPU_SEARCH_HIGHEST)
652 				if (load > hload || high->cs_cpu == -1) {
653 					hload = load;
654 					*high = hgroup;
655 				}
656 		}
657 	} else {
658 		int cpu;
659 
660 		CPUSET_FOREACH(cpu, cg->cg_mask)
661 			total += cpu_compare(cpu, low, high, match);
662 	}
663 	return (total);
664 }
665 
666 /*
667  * cpu_search instantiations must pass constants to maintain the inline
668  * optimization.
669  */
670 int
671 cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low)
672 {
673 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
674 }
675 
676 int
677 cpu_search_highest(struct cpu_group *cg, struct cpu_search *high)
678 {
679 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
680 }
681 
682 int
683 cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
684     struct cpu_search *high)
685 {
686 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
687 }
688 
689 /*
690  * Find the cpu with the least load via the least loaded path that has a
691  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
692  * acceptable.
693  */
694 static inline int
695 sched_lowest(struct cpu_group *cg, cpuset_t mask, int pri)
696 {
697 	struct cpu_search low;
698 
699 	low.cs_cpu = -1;
700 	low.cs_load = -1;
701 	low.cs_mask = mask;
702 	low.cs_limit = pri;
703 	cpu_search_lowest(cg, &low);
704 	return low.cs_cpu;
705 }
706 
707 /*
708  * Find the cpu with the highest load via the highest loaded path.
709  */
710 static inline int
711 sched_highest(struct cpu_group *cg, cpuset_t mask, int minload)
712 {
713 	struct cpu_search high;
714 
715 	high.cs_cpu = -1;
716 	high.cs_load = 0;
717 	high.cs_mask = mask;
718 	high.cs_limit = minload;
719 	cpu_search_highest(cg, &high);
720 	return high.cs_cpu;
721 }
722 
723 /*
724  * Simultaneously find the highest and lowest loaded cpu reachable via
725  * cg.
726  */
727 static inline void
728 sched_both(struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu)
729 {
730 	struct cpu_search high;
731 	struct cpu_search low;
732 
733 	low.cs_cpu = -1;
734 	low.cs_limit = -1;
735 	low.cs_load = -1;
736 	low.cs_mask = mask;
737 	high.cs_load = 0;
738 	high.cs_cpu = -1;
739 	high.cs_limit = -1;
740 	high.cs_mask = mask;
741 	cpu_search_both(cg, &low, &high);
742 	*lowcpu = low.cs_cpu;
743 	*highcpu = high.cs_cpu;
744 	return;
745 }
746 
747 static void
748 sched_balance_group(struct cpu_group *cg)
749 {
750 	cpuset_t mask;
751 	int high;
752 	int low;
753 	int i;
754 
755 	CPU_FILL(&mask);
756 	for (;;) {
757 		sched_both(cg, mask, &low, &high);
758 		if (low == high || low == -1 || high == -1)
759 			break;
760 		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low)))
761 			break;
762 		/*
763 		 * If we failed to move any threads determine which cpu
764 		 * to kick out of the set and try again.
765 	 	 */
766 		if (TDQ_CPU(high)->tdq_transferable == 0)
767 			CPU_CLR(high, &mask);
768 		else
769 			CPU_CLR(low, &mask);
770 	}
771 
772 	for (i = 0; i < cg->cg_children; i++)
773 		sched_balance_group(&cg->cg_child[i]);
774 }
775 
776 static void
777 sched_balance()
778 {
779 	struct tdq *tdq;
780 
781 	/*
782 	 * Select a random time between .5 * balance_interval and
783 	 * 1.5 * balance_interval.
784 	 */
785 	balance_ticks = max(balance_interval / 2, 1);
786 	balance_ticks += random() % balance_interval;
787 	if (smp_started == 0 || rebalance == 0)
788 		return;
789 	tdq = TDQ_SELF();
790 	TDQ_UNLOCK(tdq);
791 	sched_balance_group(cpu_top);
792 	TDQ_LOCK(tdq);
793 }
794 
795 /*
796  * Lock two thread queues using their address to maintain lock order.
797  */
798 static void
799 tdq_lock_pair(struct tdq *one, struct tdq *two)
800 {
801 	if (one < two) {
802 		TDQ_LOCK(one);
803 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
804 	} else {
805 		TDQ_LOCK(two);
806 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
807 	}
808 }
809 
810 /*
811  * Unlock two thread queues.  Order is not important here.
812  */
813 static void
814 tdq_unlock_pair(struct tdq *one, struct tdq *two)
815 {
816 	TDQ_UNLOCK(one);
817 	TDQ_UNLOCK(two);
818 }
819 
820 /*
821  * Transfer load between two imbalanced thread queues.
822  */
823 static int
824 sched_balance_pair(struct tdq *high, struct tdq *low)
825 {
826 	int transferable;
827 	int high_load;
828 	int low_load;
829 	int moved;
830 	int move;
831 	int diff;
832 	int i;
833 
834 	tdq_lock_pair(high, low);
835 	transferable = high->tdq_transferable;
836 	high_load = high->tdq_load;
837 	low_load = low->tdq_load;
838 	moved = 0;
839 	/*
840 	 * Determine what the imbalance is and then adjust that to how many
841 	 * threads we actually have to give up (transferable).
842 	 */
843 	if (transferable != 0) {
844 		diff = high_load - low_load;
845 		move = diff / 2;
846 		if (diff & 0x1)
847 			move++;
848 		move = min(move, transferable);
849 		for (i = 0; i < move; i++)
850 			moved += tdq_move(high, low);
851 		/*
852 		 * IPI the target cpu to force it to reschedule with the new
853 		 * workload.
854 		 */
855 		ipi_selected(1 << TDQ_ID(low), IPI_PREEMPT);
856 	}
857 	tdq_unlock_pair(high, low);
858 	return (moved);
859 }
860 
861 /*
862  * Move a thread from one thread queue to another.
863  */
864 static int
865 tdq_move(struct tdq *from, struct tdq *to)
866 {
867 	struct td_sched *ts;
868 	struct thread *td;
869 	struct tdq *tdq;
870 	int cpu;
871 
872 	TDQ_LOCK_ASSERT(from, MA_OWNED);
873 	TDQ_LOCK_ASSERT(to, MA_OWNED);
874 
875 	tdq = from;
876 	cpu = TDQ_ID(to);
877 	td = tdq_steal(tdq, cpu);
878 	if (td == NULL)
879 		return (0);
880 	ts = td->td_sched;
881 	/*
882 	 * Although the run queue is locked the thread may be blocked.  Lock
883 	 * it to clear this and acquire the run-queue lock.
884 	 */
885 	thread_lock(td);
886 	/* Drop recursive lock on from acquired via thread_lock(). */
887 	TDQ_UNLOCK(from);
888 	sched_rem(td);
889 	ts->ts_cpu = cpu;
890 	td->td_lock = TDQ_LOCKPTR(to);
891 	tdq_add(to, td, SRQ_YIELDING);
892 	return (1);
893 }
894 
895 /*
896  * This tdq has idled.  Try to steal a thread from another cpu and switch
897  * to it.
898  */
899 static int
900 tdq_idled(struct tdq *tdq)
901 {
902 	struct cpu_group *cg;
903 	struct tdq *steal;
904 	cpuset_t mask;
905 	int thresh;
906 	int cpu;
907 
908 	if (smp_started == 0 || steal_idle == 0)
909 		return (1);
910 	CPU_FILL(&mask);
911 	CPU_CLR(PCPU_GET(cpuid), &mask);
912 	/* We don't want to be preempted while we're iterating. */
913 	spinlock_enter();
914 	for (cg = tdq->tdq_cg; cg != NULL; ) {
915 		if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
916 			thresh = steal_thresh;
917 		else
918 			thresh = 1;
919 		cpu = sched_highest(cg, mask, thresh);
920 		if (cpu == -1) {
921 			cg = cg->cg_parent;
922 			continue;
923 		}
924 		steal = TDQ_CPU(cpu);
925 		CPU_CLR(cpu, &mask);
926 		tdq_lock_pair(tdq, steal);
927 		if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
928 			tdq_unlock_pair(tdq, steal);
929 			continue;
930 		}
931 		/*
932 		 * If a thread was added while interrupts were disabled don't
933 		 * steal one here.  If we fail to acquire one due to affinity
934 		 * restrictions loop again with this cpu removed from the
935 		 * set.
936 		 */
937 		if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
938 			tdq_unlock_pair(tdq, steal);
939 			continue;
940 		}
941 		spinlock_exit();
942 		TDQ_UNLOCK(steal);
943 		mi_switch(SW_VOL | SWT_IDLE, NULL);
944 		thread_unlock(curthread);
945 
946 		return (0);
947 	}
948 	spinlock_exit();
949 	return (1);
950 }
951 
952 /*
953  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
954  */
955 static void
956 tdq_notify(struct tdq *tdq, struct thread *td)
957 {
958 	struct thread *ctd;
959 	int pri;
960 	int cpu;
961 
962 	if (tdq->tdq_ipipending)
963 		return;
964 	cpu = td->td_sched->ts_cpu;
965 	pri = td->td_priority;
966 	ctd = pcpu_find(cpu)->pc_curthread;
967 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
968 		return;
969 	if (TD_IS_IDLETHREAD(ctd)) {
970 		/*
971 		 * If the MD code has an idle wakeup routine try that before
972 		 * falling back to IPI.
973 		 */
974 		if (cpu_idle_wakeup(cpu))
975 			return;
976 	}
977 	tdq->tdq_ipipending = 1;
978 	ipi_selected(1 << cpu, IPI_PREEMPT);
979 }
980 
981 /*
982  * Steals load from a timeshare queue.  Honors the rotating queue head
983  * index.
984  */
985 static struct thread *
986 runq_steal_from(struct runq *rq, int cpu, u_char start)
987 {
988 	struct rqbits *rqb;
989 	struct rqhead *rqh;
990 	struct thread *td;
991 	int first;
992 	int bit;
993 	int pri;
994 	int i;
995 
996 	rqb = &rq->rq_status;
997 	bit = start & (RQB_BPW -1);
998 	pri = 0;
999 	first = 0;
1000 again:
1001 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1002 		if (rqb->rqb_bits[i] == 0)
1003 			continue;
1004 		if (bit != 0) {
1005 			for (pri = bit; pri < RQB_BPW; pri++)
1006 				if (rqb->rqb_bits[i] & (1ul << pri))
1007 					break;
1008 			if (pri >= RQB_BPW)
1009 				continue;
1010 		} else
1011 			pri = RQB_FFS(rqb->rqb_bits[i]);
1012 		pri += (i << RQB_L2BPW);
1013 		rqh = &rq->rq_queues[pri];
1014 		TAILQ_FOREACH(td, rqh, td_runq) {
1015 			if (first && THREAD_CAN_MIGRATE(td) &&
1016 			    THREAD_CAN_SCHED(td, cpu))
1017 				return (td);
1018 			first = 1;
1019 		}
1020 	}
1021 	if (start != 0) {
1022 		start = 0;
1023 		goto again;
1024 	}
1025 
1026 	return (NULL);
1027 }
1028 
1029 /*
1030  * Steals load from a standard linear queue.
1031  */
1032 static struct thread *
1033 runq_steal(struct runq *rq, int cpu)
1034 {
1035 	struct rqhead *rqh;
1036 	struct rqbits *rqb;
1037 	struct thread *td;
1038 	int word;
1039 	int bit;
1040 
1041 	rqb = &rq->rq_status;
1042 	for (word = 0; word < RQB_LEN; word++) {
1043 		if (rqb->rqb_bits[word] == 0)
1044 			continue;
1045 		for (bit = 0; bit < RQB_BPW; bit++) {
1046 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1047 				continue;
1048 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1049 			TAILQ_FOREACH(td, rqh, td_runq)
1050 				if (THREAD_CAN_MIGRATE(td) &&
1051 				    THREAD_CAN_SCHED(td, cpu))
1052 					return (td);
1053 		}
1054 	}
1055 	return (NULL);
1056 }
1057 
1058 /*
1059  * Attempt to steal a thread in priority order from a thread queue.
1060  */
1061 static struct thread *
1062 tdq_steal(struct tdq *tdq, int cpu)
1063 {
1064 	struct thread *td;
1065 
1066 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1067 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1068 		return (td);
1069 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1070 	    cpu, tdq->tdq_ridx)) != NULL)
1071 		return (td);
1072 	return (runq_steal(&tdq->tdq_idle, cpu));
1073 }
1074 
1075 /*
1076  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1077  * current lock and returns with the assigned queue locked.
1078  */
1079 static inline struct tdq *
1080 sched_setcpu(struct thread *td, int cpu, int flags)
1081 {
1082 
1083 	struct tdq *tdq;
1084 
1085 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1086 	tdq = TDQ_CPU(cpu);
1087 	td->td_sched->ts_cpu = cpu;
1088 	/*
1089 	 * If the lock matches just return the queue.
1090 	 */
1091 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1092 		return (tdq);
1093 #ifdef notyet
1094 	/*
1095 	 * If the thread isn't running its lockptr is a
1096 	 * turnstile or a sleepqueue.  We can just lock_set without
1097 	 * blocking.
1098 	 */
1099 	if (TD_CAN_RUN(td)) {
1100 		TDQ_LOCK(tdq);
1101 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1102 		return (tdq);
1103 	}
1104 #endif
1105 	/*
1106 	 * The hard case, migration, we need to block the thread first to
1107 	 * prevent order reversals with other cpus locks.
1108 	 */
1109 	thread_lock_block(td);
1110 	TDQ_LOCK(tdq);
1111 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1112 	return (tdq);
1113 }
1114 
1115 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1116 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1117 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1118 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1119 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1120 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1121 
1122 static int
1123 sched_pickcpu(struct thread *td, int flags)
1124 {
1125 	struct cpu_group *cg;
1126 	struct td_sched *ts;
1127 	struct tdq *tdq;
1128 	cpuset_t mask;
1129 	int self;
1130 	int pri;
1131 	int cpu;
1132 
1133 	self = PCPU_GET(cpuid);
1134 	ts = td->td_sched;
1135 	if (smp_started == 0)
1136 		return (self);
1137 	/*
1138 	 * Don't migrate a running thread from sched_switch().
1139 	 */
1140 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1141 		return (ts->ts_cpu);
1142 	/*
1143 	 * Prefer to run interrupt threads on the processors that generate
1144 	 * the interrupt.
1145 	 */
1146 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1147 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1148 		SCHED_STAT_INC(pickcpu_intrbind);
1149 		ts->ts_cpu = self;
1150 	}
1151 	/*
1152 	 * If the thread can run on the last cpu and the affinity has not
1153 	 * expired or it is idle run it there.
1154 	 */
1155 	pri = td->td_priority;
1156 	tdq = TDQ_CPU(ts->ts_cpu);
1157 	if (THREAD_CAN_SCHED(td, ts->ts_cpu)) {
1158 		if (tdq->tdq_lowpri > PRI_MIN_IDLE) {
1159 			SCHED_STAT_INC(pickcpu_idle_affinity);
1160 			return (ts->ts_cpu);
1161 		}
1162 		if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri) {
1163 			SCHED_STAT_INC(pickcpu_affinity);
1164 			return (ts->ts_cpu);
1165 		}
1166 	}
1167 	/*
1168 	 * Search for the highest level in the tree that still has affinity.
1169 	 */
1170 	cg = NULL;
1171 	for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent)
1172 		if (SCHED_AFFINITY(ts, cg->cg_level))
1173 			break;
1174 	cpu = -1;
1175 	mask = td->td_cpuset->cs_mask;
1176 	if (cg)
1177 		cpu = sched_lowest(cg, mask, pri);
1178 	if (cpu == -1)
1179 		cpu = sched_lowest(cpu_top, mask, -1);
1180 	/*
1181 	 * Compare the lowest loaded cpu to current cpu.
1182 	 */
1183 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1184 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) {
1185 		SCHED_STAT_INC(pickcpu_local);
1186 		cpu = self;
1187 	} else
1188 		SCHED_STAT_INC(pickcpu_lowest);
1189 	if (cpu != ts->ts_cpu)
1190 		SCHED_STAT_INC(pickcpu_migration);
1191 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1192 	return (cpu);
1193 }
1194 #endif
1195 
1196 /*
1197  * Pick the highest priority task we have and return it.
1198  */
1199 static struct thread *
1200 tdq_choose(struct tdq *tdq)
1201 {
1202 	struct thread *td;
1203 
1204 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1205 	td = runq_choose(&tdq->tdq_realtime);
1206 	if (td != NULL)
1207 		return (td);
1208 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1209 	if (td != NULL) {
1210 		KASSERT(td->td_priority >= PRI_MIN_TIMESHARE,
1211 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1212 		    td->td_priority));
1213 		return (td);
1214 	}
1215 	td = runq_choose(&tdq->tdq_idle);
1216 	if (td != NULL) {
1217 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1218 		    ("tdq_choose: Invalid priority on idle queue %d",
1219 		    td->td_priority));
1220 		return (td);
1221 	}
1222 
1223 	return (NULL);
1224 }
1225 
1226 /*
1227  * Initialize a thread queue.
1228  */
1229 static void
1230 tdq_setup(struct tdq *tdq)
1231 {
1232 
1233 	if (bootverbose)
1234 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1235 	runq_init(&tdq->tdq_realtime);
1236 	runq_init(&tdq->tdq_timeshare);
1237 	runq_init(&tdq->tdq_idle);
1238 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1239 	    "sched lock %d", (int)TDQ_ID(tdq));
1240 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1241 	    MTX_SPIN | MTX_RECURSE);
1242 #ifdef KTR
1243 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1244 	    "CPU %d load", (int)TDQ_ID(tdq));
1245 #endif
1246 }
1247 
1248 #ifdef SMP
1249 static void
1250 sched_setup_smp(void)
1251 {
1252 	struct tdq *tdq;
1253 	int i;
1254 
1255 	cpu_top = smp_topo();
1256 	for (i = 0; i < MAXCPU; i++) {
1257 		if (CPU_ABSENT(i))
1258 			continue;
1259 		tdq = TDQ_CPU(i);
1260 		tdq_setup(tdq);
1261 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1262 		if (tdq->tdq_cg == NULL)
1263 			panic("Can't find cpu group for %d\n", i);
1264 	}
1265 	balance_tdq = TDQ_SELF();
1266 	sched_balance();
1267 }
1268 #endif
1269 
1270 /*
1271  * Setup the thread queues and initialize the topology based on MD
1272  * information.
1273  */
1274 static void
1275 sched_setup(void *dummy)
1276 {
1277 	struct tdq *tdq;
1278 
1279 	tdq = TDQ_SELF();
1280 #ifdef SMP
1281 	sched_setup_smp();
1282 #else
1283 	tdq_setup(tdq);
1284 #endif
1285 	/*
1286 	 * To avoid divide-by-zero, we set realstathz a dummy value
1287 	 * in case which sched_clock() called before sched_initticks().
1288 	 */
1289 	realstathz = hz;
1290 	sched_slice = (realstathz/10);	/* ~100ms */
1291 	tickincr = 1 << SCHED_TICK_SHIFT;
1292 
1293 	/* Add thread0's load since it's running. */
1294 	TDQ_LOCK(tdq);
1295 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1296 	tdq_load_add(tdq, &thread0);
1297 	tdq->tdq_lowpri = thread0.td_priority;
1298 	TDQ_UNLOCK(tdq);
1299 }
1300 
1301 /*
1302  * This routine determines the tickincr after stathz and hz are setup.
1303  */
1304 /* ARGSUSED */
1305 static void
1306 sched_initticks(void *dummy)
1307 {
1308 	int incr;
1309 
1310 	realstathz = stathz ? stathz : hz;
1311 	sched_slice = (realstathz/10);	/* ~100ms */
1312 
1313 	/*
1314 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1315 	 * hz not being evenly divisible by stathz on all platforms.
1316 	 */
1317 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1318 	/*
1319 	 * This does not work for values of stathz that are more than
1320 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1321 	 */
1322 	if (incr == 0)
1323 		incr = 1;
1324 	tickincr = incr;
1325 #ifdef SMP
1326 	/*
1327 	 * Set the default balance interval now that we know
1328 	 * what realstathz is.
1329 	 */
1330 	balance_interval = realstathz;
1331 	/*
1332 	 * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4.
1333 	 * This prevents excess thrashing on large machines and excess idle
1334 	 * on smaller machines.
1335 	 */
1336 	steal_thresh = min(fls(mp_ncpus) - 1, 3);
1337 	affinity = SCHED_AFFINITY_DEFAULT;
1338 #endif
1339 }
1340 
1341 
1342 /*
1343  * This is the core of the interactivity algorithm.  Determines a score based
1344  * on past behavior.  It is the ratio of sleep time to run time scaled to
1345  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1346  * differs from the cpu usage because it does not account for time spent
1347  * waiting on a run-queue.  Would be prettier if we had floating point.
1348  */
1349 static int
1350 sched_interact_score(struct thread *td)
1351 {
1352 	struct td_sched *ts;
1353 	int div;
1354 
1355 	ts = td->td_sched;
1356 	/*
1357 	 * The score is only needed if this is likely to be an interactive
1358 	 * task.  Don't go through the expense of computing it if there's
1359 	 * no chance.
1360 	 */
1361 	if (sched_interact <= SCHED_INTERACT_HALF &&
1362 		ts->ts_runtime >= ts->ts_slptime)
1363 			return (SCHED_INTERACT_HALF);
1364 
1365 	if (ts->ts_runtime > ts->ts_slptime) {
1366 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1367 		return (SCHED_INTERACT_HALF +
1368 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1369 	}
1370 	if (ts->ts_slptime > ts->ts_runtime) {
1371 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1372 		return (ts->ts_runtime / div);
1373 	}
1374 	/* runtime == slptime */
1375 	if (ts->ts_runtime)
1376 		return (SCHED_INTERACT_HALF);
1377 
1378 	/*
1379 	 * This can happen if slptime and runtime are 0.
1380 	 */
1381 	return (0);
1382 
1383 }
1384 
1385 /*
1386  * Scale the scheduling priority according to the "interactivity" of this
1387  * process.
1388  */
1389 static void
1390 sched_priority(struct thread *td)
1391 {
1392 	int score;
1393 	int pri;
1394 
1395 	if (td->td_pri_class != PRI_TIMESHARE)
1396 		return;
1397 	/*
1398 	 * If the score is interactive we place the thread in the realtime
1399 	 * queue with a priority that is less than kernel and interrupt
1400 	 * priorities.  These threads are not subject to nice restrictions.
1401 	 *
1402 	 * Scores greater than this are placed on the normal timeshare queue
1403 	 * where the priority is partially decided by the most recent cpu
1404 	 * utilization and the rest is decided by nice value.
1405 	 *
1406 	 * The nice value of the process has a linear effect on the calculated
1407 	 * score.  Negative nice values make it easier for a thread to be
1408 	 * considered interactive.
1409 	 */
1410 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1411 	if (score < sched_interact) {
1412 		pri = PRI_MIN_REALTIME;
1413 		pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact)
1414 		    * score;
1415 		KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME,
1416 		    ("sched_priority: invalid interactive priority %d score %d",
1417 		    pri, score));
1418 	} else {
1419 		pri = SCHED_PRI_MIN;
1420 		if (td->td_sched->ts_ticks)
1421 			pri += SCHED_PRI_TICKS(td->td_sched);
1422 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1423 		KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE,
1424 		    ("sched_priority: invalid priority %d: nice %d, "
1425 		    "ticks %d ftick %d ltick %d tick pri %d",
1426 		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1427 		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1428 		    SCHED_PRI_TICKS(td->td_sched)));
1429 	}
1430 	sched_user_prio(td, pri);
1431 
1432 	return;
1433 }
1434 
1435 /*
1436  * This routine enforces a maximum limit on the amount of scheduling history
1437  * kept.  It is called after either the slptime or runtime is adjusted.  This
1438  * function is ugly due to integer math.
1439  */
1440 static void
1441 sched_interact_update(struct thread *td)
1442 {
1443 	struct td_sched *ts;
1444 	u_int sum;
1445 
1446 	ts = td->td_sched;
1447 	sum = ts->ts_runtime + ts->ts_slptime;
1448 	if (sum < SCHED_SLP_RUN_MAX)
1449 		return;
1450 	/*
1451 	 * This only happens from two places:
1452 	 * 1) We have added an unusual amount of run time from fork_exit.
1453 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1454 	 */
1455 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1456 		if (ts->ts_runtime > ts->ts_slptime) {
1457 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1458 			ts->ts_slptime = 1;
1459 		} else {
1460 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1461 			ts->ts_runtime = 1;
1462 		}
1463 		return;
1464 	}
1465 	/*
1466 	 * If we have exceeded by more than 1/5th then the algorithm below
1467 	 * will not bring us back into range.  Dividing by two here forces
1468 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1469 	 */
1470 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1471 		ts->ts_runtime /= 2;
1472 		ts->ts_slptime /= 2;
1473 		return;
1474 	}
1475 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1476 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1477 }
1478 
1479 /*
1480  * Scale back the interactivity history when a child thread is created.  The
1481  * history is inherited from the parent but the thread may behave totally
1482  * differently.  For example, a shell spawning a compiler process.  We want
1483  * to learn that the compiler is behaving badly very quickly.
1484  */
1485 static void
1486 sched_interact_fork(struct thread *td)
1487 {
1488 	int ratio;
1489 	int sum;
1490 
1491 	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1492 	if (sum > SCHED_SLP_RUN_FORK) {
1493 		ratio = sum / SCHED_SLP_RUN_FORK;
1494 		td->td_sched->ts_runtime /= ratio;
1495 		td->td_sched->ts_slptime /= ratio;
1496 	}
1497 }
1498 
1499 /*
1500  * Called from proc0_init() to setup the scheduler fields.
1501  */
1502 void
1503 schedinit(void)
1504 {
1505 
1506 	/*
1507 	 * Set up the scheduler specific parts of proc0.
1508 	 */
1509 	proc0.p_sched = NULL; /* XXX */
1510 	thread0.td_sched = &td_sched0;
1511 	td_sched0.ts_ltick = ticks;
1512 	td_sched0.ts_ftick = ticks;
1513 	td_sched0.ts_slice = sched_slice;
1514 }
1515 
1516 /*
1517  * This is only somewhat accurate since given many processes of the same
1518  * priority they will switch when their slices run out, which will be
1519  * at most sched_slice stathz ticks.
1520  */
1521 int
1522 sched_rr_interval(void)
1523 {
1524 
1525 	/* Convert sched_slice to hz */
1526 	return (hz/(realstathz/sched_slice));
1527 }
1528 
1529 /*
1530  * Update the percent cpu tracking information when it is requested or
1531  * the total history exceeds the maximum.  We keep a sliding history of
1532  * tick counts that slowly decays.  This is less precise than the 4BSD
1533  * mechanism since it happens with less regular and frequent events.
1534  */
1535 static void
1536 sched_pctcpu_update(struct td_sched *ts)
1537 {
1538 
1539 	if (ts->ts_ticks == 0)
1540 		return;
1541 	if (ticks - (hz / 10) < ts->ts_ltick &&
1542 	    SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX)
1543 		return;
1544 	/*
1545 	 * Adjust counters and watermark for pctcpu calc.
1546 	 */
1547 	if (ts->ts_ltick > ticks - SCHED_TICK_TARG)
1548 		ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) *
1549 			    SCHED_TICK_TARG;
1550 	else
1551 		ts->ts_ticks = 0;
1552 	ts->ts_ltick = ticks;
1553 	ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG;
1554 }
1555 
1556 /*
1557  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1558  * if necessary.  This is the back-end for several priority related
1559  * functions.
1560  */
1561 static void
1562 sched_thread_priority(struct thread *td, u_char prio)
1563 {
1564 	struct td_sched *ts;
1565 	struct tdq *tdq;
1566 	int oldpri;
1567 
1568 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1569 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1570 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1571 	if (td != curthread && prio > td->td_priority) {
1572 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1573 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1574 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1575 	}
1576 	ts = td->td_sched;
1577 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1578 	if (td->td_priority == prio)
1579 		return;
1580 	/*
1581 	 * If the priority has been elevated due to priority
1582 	 * propagation, we may have to move ourselves to a new
1583 	 * queue.  This could be optimized to not re-add in some
1584 	 * cases.
1585 	 */
1586 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1587 		sched_rem(td);
1588 		td->td_priority = prio;
1589 		sched_add(td, SRQ_BORROWING);
1590 		return;
1591 	}
1592 	/*
1593 	 * If the thread is currently running we may have to adjust the lowpri
1594 	 * information so other cpus are aware of our current priority.
1595 	 */
1596 	if (TD_IS_RUNNING(td)) {
1597 		tdq = TDQ_CPU(ts->ts_cpu);
1598 		oldpri = td->td_priority;
1599 		td->td_priority = prio;
1600 		if (prio < tdq->tdq_lowpri)
1601 			tdq->tdq_lowpri = prio;
1602 		else if (tdq->tdq_lowpri == oldpri)
1603 			tdq_setlowpri(tdq, td);
1604 		return;
1605 	}
1606 	td->td_priority = prio;
1607 }
1608 
1609 /*
1610  * Update a thread's priority when it is lent another thread's
1611  * priority.
1612  */
1613 void
1614 sched_lend_prio(struct thread *td, u_char prio)
1615 {
1616 
1617 	td->td_flags |= TDF_BORROWING;
1618 	sched_thread_priority(td, prio);
1619 }
1620 
1621 /*
1622  * Restore a thread's priority when priority propagation is
1623  * over.  The prio argument is the minimum priority the thread
1624  * needs to have to satisfy other possible priority lending
1625  * requests.  If the thread's regular priority is less
1626  * important than prio, the thread will keep a priority boost
1627  * of prio.
1628  */
1629 void
1630 sched_unlend_prio(struct thread *td, u_char prio)
1631 {
1632 	u_char base_pri;
1633 
1634 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1635 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1636 		base_pri = td->td_user_pri;
1637 	else
1638 		base_pri = td->td_base_pri;
1639 	if (prio >= base_pri) {
1640 		td->td_flags &= ~TDF_BORROWING;
1641 		sched_thread_priority(td, base_pri);
1642 	} else
1643 		sched_lend_prio(td, prio);
1644 }
1645 
1646 /*
1647  * Standard entry for setting the priority to an absolute value.
1648  */
1649 void
1650 sched_prio(struct thread *td, u_char prio)
1651 {
1652 	u_char oldprio;
1653 
1654 	/* First, update the base priority. */
1655 	td->td_base_pri = prio;
1656 
1657 	/*
1658 	 * If the thread is borrowing another thread's priority, don't
1659 	 * ever lower the priority.
1660 	 */
1661 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1662 		return;
1663 
1664 	/* Change the real priority. */
1665 	oldprio = td->td_priority;
1666 	sched_thread_priority(td, prio);
1667 
1668 	/*
1669 	 * If the thread is on a turnstile, then let the turnstile update
1670 	 * its state.
1671 	 */
1672 	if (TD_ON_LOCK(td) && oldprio != prio)
1673 		turnstile_adjust(td, oldprio);
1674 }
1675 
1676 /*
1677  * Set the base user priority, does not effect current running priority.
1678  */
1679 void
1680 sched_user_prio(struct thread *td, u_char prio)
1681 {
1682 	u_char oldprio;
1683 
1684 	td->td_base_user_pri = prio;
1685 	if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
1686                 return;
1687 	oldprio = td->td_user_pri;
1688 	td->td_user_pri = prio;
1689 }
1690 
1691 void
1692 sched_lend_user_prio(struct thread *td, u_char prio)
1693 {
1694 	u_char oldprio;
1695 
1696 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1697 	td->td_flags |= TDF_UBORROWING;
1698 	oldprio = td->td_user_pri;
1699 	td->td_user_pri = prio;
1700 }
1701 
1702 void
1703 sched_unlend_user_prio(struct thread *td, u_char prio)
1704 {
1705 	u_char base_pri;
1706 
1707 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1708 	base_pri = td->td_base_user_pri;
1709 	if (prio >= base_pri) {
1710 		td->td_flags &= ~TDF_UBORROWING;
1711 		sched_user_prio(td, base_pri);
1712 	} else {
1713 		sched_lend_user_prio(td, prio);
1714 	}
1715 }
1716 
1717 /*
1718  * Block a thread for switching.  Similar to thread_block() but does not
1719  * bump the spin count.
1720  */
1721 static inline struct mtx *
1722 thread_block_switch(struct thread *td)
1723 {
1724 	struct mtx *lock;
1725 
1726 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1727 	lock = td->td_lock;
1728 	td->td_lock = &blocked_lock;
1729 	mtx_unlock_spin(lock);
1730 
1731 	return (lock);
1732 }
1733 
1734 /*
1735  * Handle migration from sched_switch().  This happens only for
1736  * cpu binding.
1737  */
1738 static struct mtx *
1739 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1740 {
1741 	struct tdq *tdn;
1742 
1743 	tdn = TDQ_CPU(td->td_sched->ts_cpu);
1744 #ifdef SMP
1745 	tdq_load_rem(tdq, td);
1746 	/*
1747 	 * Do the lock dance required to avoid LOR.  We grab an extra
1748 	 * spinlock nesting to prevent preemption while we're
1749 	 * not holding either run-queue lock.
1750 	 */
1751 	spinlock_enter();
1752 	thread_block_switch(td);	/* This releases the lock on tdq. */
1753 
1754 	/*
1755 	 * Acquire both run-queue locks before placing the thread on the new
1756 	 * run-queue to avoid deadlocks created by placing a thread with a
1757 	 * blocked lock on the run-queue of a remote processor.  The deadlock
1758 	 * occurs when a third processor attempts to lock the two queues in
1759 	 * question while the target processor is spinning with its own
1760 	 * run-queue lock held while waiting for the blocked lock to clear.
1761 	 */
1762 	tdq_lock_pair(tdn, tdq);
1763 	tdq_add(tdn, td, flags);
1764 	tdq_notify(tdn, td);
1765 	TDQ_UNLOCK(tdn);
1766 	spinlock_exit();
1767 #endif
1768 	return (TDQ_LOCKPTR(tdn));
1769 }
1770 
1771 /*
1772  * Release a thread that was blocked with thread_block_switch().
1773  */
1774 static inline void
1775 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1776 {
1777 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1778 	    (uintptr_t)mtx);
1779 }
1780 
1781 /*
1782  * Switch threads.  This function has to handle threads coming in while
1783  * blocked for some reason, running, or idle.  It also must deal with
1784  * migrating a thread from one queue to another as running threads may
1785  * be assigned elsewhere via binding.
1786  */
1787 void
1788 sched_switch(struct thread *td, struct thread *newtd, int flags)
1789 {
1790 	struct tdq *tdq;
1791 	struct td_sched *ts;
1792 	struct mtx *mtx;
1793 	int srqflag;
1794 	int cpuid;
1795 
1796 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1797 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1798 
1799 	cpuid = PCPU_GET(cpuid);
1800 	tdq = TDQ_CPU(cpuid);
1801 	ts = td->td_sched;
1802 	mtx = td->td_lock;
1803 	ts->ts_rltick = ticks;
1804 	td->td_lastcpu = td->td_oncpu;
1805 	td->td_oncpu = NOCPU;
1806 	td->td_flags &= ~TDF_NEEDRESCHED;
1807 	td->td_owepreempt = 0;
1808 	tdq->tdq_switchcnt++;
1809 	/*
1810 	 * The lock pointer in an idle thread should never change.  Reset it
1811 	 * to CAN_RUN as well.
1812 	 */
1813 	if (TD_IS_IDLETHREAD(td)) {
1814 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1815 		TD_SET_CAN_RUN(td);
1816 	} else if (TD_IS_RUNNING(td)) {
1817 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1818 		srqflag = (flags & SW_PREEMPT) ?
1819 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1820 		    SRQ_OURSELF|SRQ_YIELDING;
1821 		if (ts->ts_cpu == cpuid)
1822 			tdq_runq_add(tdq, td, srqflag);
1823 		else
1824 			mtx = sched_switch_migrate(tdq, td, srqflag);
1825 	} else {
1826 		/* This thread must be going to sleep. */
1827 		TDQ_LOCK(tdq);
1828 		mtx = thread_block_switch(td);
1829 		tdq_load_rem(tdq, td);
1830 	}
1831 	/*
1832 	 * We enter here with the thread blocked and assigned to the
1833 	 * appropriate cpu run-queue or sleep-queue and with the current
1834 	 * thread-queue locked.
1835 	 */
1836 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1837 	newtd = choosethread();
1838 	/*
1839 	 * Call the MD code to switch contexts if necessary.
1840 	 */
1841 	if (td != newtd) {
1842 #ifdef	HWPMC_HOOKS
1843 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1844 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1845 #endif
1846 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1847 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1848 
1849 #ifdef KDTRACE_HOOKS
1850 		/*
1851 		 * If DTrace has set the active vtime enum to anything
1852 		 * other than INACTIVE (0), then it should have set the
1853 		 * function to call.
1854 		 */
1855 		if (dtrace_vtime_active)
1856 			(*dtrace_vtime_switch_func)(newtd);
1857 #endif
1858 
1859 		cpu_switch(td, newtd, mtx);
1860 		/*
1861 		 * We may return from cpu_switch on a different cpu.  However,
1862 		 * we always return with td_lock pointing to the current cpu's
1863 		 * run queue lock.
1864 		 */
1865 		cpuid = PCPU_GET(cpuid);
1866 		tdq = TDQ_CPU(cpuid);
1867 		lock_profile_obtain_lock_success(
1868 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1869 #ifdef	HWPMC_HOOKS
1870 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1871 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1872 #endif
1873 	} else
1874 		thread_unblock_switch(td, mtx);
1875 	/*
1876 	 * Assert that all went well and return.
1877 	 */
1878 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1879 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1880 	td->td_oncpu = cpuid;
1881 }
1882 
1883 /*
1884  * Adjust thread priorities as a result of a nice request.
1885  */
1886 void
1887 sched_nice(struct proc *p, int nice)
1888 {
1889 	struct thread *td;
1890 
1891 	PROC_LOCK_ASSERT(p, MA_OWNED);
1892 
1893 	p->p_nice = nice;
1894 	FOREACH_THREAD_IN_PROC(p, td) {
1895 		thread_lock(td);
1896 		sched_priority(td);
1897 		sched_prio(td, td->td_base_user_pri);
1898 		thread_unlock(td);
1899 	}
1900 }
1901 
1902 /*
1903  * Record the sleep time for the interactivity scorer.
1904  */
1905 void
1906 sched_sleep(struct thread *td, int prio)
1907 {
1908 
1909 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1910 
1911 	td->td_slptick = ticks;
1912 	if (TD_IS_SUSPENDED(td) || prio <= PSOCK)
1913 		td->td_flags |= TDF_CANSWAP;
1914 	if (static_boost == 1 && prio)
1915 		sched_prio(td, prio);
1916 	else if (static_boost && td->td_priority > static_boost)
1917 		sched_prio(td, static_boost);
1918 }
1919 
1920 /*
1921  * Schedule a thread to resume execution and record how long it voluntarily
1922  * slept.  We also update the pctcpu, interactivity, and priority.
1923  */
1924 void
1925 sched_wakeup(struct thread *td)
1926 {
1927 	struct td_sched *ts;
1928 	int slptick;
1929 
1930 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1931 	ts = td->td_sched;
1932 	td->td_flags &= ~TDF_CANSWAP;
1933 	/*
1934 	 * If we slept for more than a tick update our interactivity and
1935 	 * priority.
1936 	 */
1937 	slptick = td->td_slptick;
1938 	td->td_slptick = 0;
1939 	if (slptick && slptick != ticks) {
1940 		u_int hzticks;
1941 
1942 		hzticks = (ticks - slptick) << SCHED_TICK_SHIFT;
1943 		ts->ts_slptime += hzticks;
1944 		sched_interact_update(td);
1945 		sched_pctcpu_update(ts);
1946 	}
1947 	/* Reset the slice value after we sleep. */
1948 	ts->ts_slice = sched_slice;
1949 	sched_add(td, SRQ_BORING);
1950 }
1951 
1952 /*
1953  * Penalize the parent for creating a new child and initialize the child's
1954  * priority.
1955  */
1956 void
1957 sched_fork(struct thread *td, struct thread *child)
1958 {
1959 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1960 	sched_fork_thread(td, child);
1961 	/*
1962 	 * Penalize the parent and child for forking.
1963 	 */
1964 	sched_interact_fork(child);
1965 	sched_priority(child);
1966 	td->td_sched->ts_runtime += tickincr;
1967 	sched_interact_update(td);
1968 	sched_priority(td);
1969 }
1970 
1971 /*
1972  * Fork a new thread, may be within the same process.
1973  */
1974 void
1975 sched_fork_thread(struct thread *td, struct thread *child)
1976 {
1977 	struct td_sched *ts;
1978 	struct td_sched *ts2;
1979 
1980 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1981 	/*
1982 	 * Initialize child.
1983 	 */
1984 	ts = td->td_sched;
1985 	ts2 = child->td_sched;
1986 	child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
1987 	child->td_cpuset = cpuset_ref(td->td_cpuset);
1988 	ts2->ts_cpu = ts->ts_cpu;
1989 	ts2->ts_flags = 0;
1990 	/*
1991 	 * Grab our parents cpu estimation information and priority.
1992 	 */
1993 	ts2->ts_ticks = ts->ts_ticks;
1994 	ts2->ts_ltick = ts->ts_ltick;
1995 	ts2->ts_incrtick = ts->ts_incrtick;
1996 	ts2->ts_ftick = ts->ts_ftick;
1997 	child->td_user_pri = td->td_user_pri;
1998 	child->td_base_user_pri = td->td_base_user_pri;
1999 	/*
2000 	 * And update interactivity score.
2001 	 */
2002 	ts2->ts_slptime = ts->ts_slptime;
2003 	ts2->ts_runtime = ts->ts_runtime;
2004 	ts2->ts_slice = 1;	/* Attempt to quickly learn interactivity. */
2005 #ifdef KTR
2006 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2007 #endif
2008 }
2009 
2010 /*
2011  * Adjust the priority class of a thread.
2012  */
2013 void
2014 sched_class(struct thread *td, int class)
2015 {
2016 
2017 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2018 	if (td->td_pri_class == class)
2019 		return;
2020 	td->td_pri_class = class;
2021 }
2022 
2023 /*
2024  * Return some of the child's priority and interactivity to the parent.
2025  */
2026 void
2027 sched_exit(struct proc *p, struct thread *child)
2028 {
2029 	struct thread *td;
2030 
2031 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2032 	    "prio:td", child->td_priority);
2033 	PROC_LOCK_ASSERT(p, MA_OWNED);
2034 	td = FIRST_THREAD_IN_PROC(p);
2035 	sched_exit_thread(td, child);
2036 }
2037 
2038 /*
2039  * Penalize another thread for the time spent on this one.  This helps to
2040  * worsen the priority and interactivity of processes which schedule batch
2041  * jobs such as make.  This has little effect on the make process itself but
2042  * causes new processes spawned by it to receive worse scores immediately.
2043  */
2044 void
2045 sched_exit_thread(struct thread *td, struct thread *child)
2046 {
2047 
2048 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2049 	    "prio:td", child->td_priority);
2050 	/*
2051 	 * Give the child's runtime to the parent without returning the
2052 	 * sleep time as a penalty to the parent.  This causes shells that
2053 	 * launch expensive things to mark their children as expensive.
2054 	 */
2055 	thread_lock(td);
2056 	td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2057 	sched_interact_update(td);
2058 	sched_priority(td);
2059 	thread_unlock(td);
2060 }
2061 
2062 void
2063 sched_preempt(struct thread *td)
2064 {
2065 	struct tdq *tdq;
2066 
2067 	thread_lock(td);
2068 	tdq = TDQ_SELF();
2069 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2070 	tdq->tdq_ipipending = 0;
2071 	if (td->td_priority > tdq->tdq_lowpri) {
2072 		int flags;
2073 
2074 		flags = SW_INVOL | SW_PREEMPT;
2075 		if (td->td_critnest > 1)
2076 			td->td_owepreempt = 1;
2077 		else if (TD_IS_IDLETHREAD(td))
2078 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2079 		else
2080 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2081 	}
2082 	thread_unlock(td);
2083 }
2084 
2085 /*
2086  * Fix priorities on return to user-space.  Priorities may be elevated due
2087  * to static priorities in msleep() or similar.
2088  */
2089 void
2090 sched_userret(struct thread *td)
2091 {
2092 	/*
2093 	 * XXX we cheat slightly on the locking here to avoid locking in
2094 	 * the usual case.  Setting td_priority here is essentially an
2095 	 * incomplete workaround for not setting it properly elsewhere.
2096 	 * Now that some interrupt handlers are threads, not setting it
2097 	 * properly elsewhere can clobber it in the window between setting
2098 	 * it here and returning to user mode, so don't waste time setting
2099 	 * it perfectly here.
2100 	 */
2101 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2102 	    ("thread with borrowed priority returning to userland"));
2103 	if (td->td_priority != td->td_user_pri) {
2104 		thread_lock(td);
2105 		td->td_priority = td->td_user_pri;
2106 		td->td_base_pri = td->td_user_pri;
2107 		tdq_setlowpri(TDQ_SELF(), td);
2108 		thread_unlock(td);
2109         }
2110 }
2111 
2112 /*
2113  * Handle a stathz tick.  This is really only relevant for timeshare
2114  * threads.
2115  */
2116 void
2117 sched_clock(struct thread *td)
2118 {
2119 	struct tdq *tdq;
2120 	struct td_sched *ts;
2121 
2122 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2123 	tdq = TDQ_SELF();
2124 #ifdef SMP
2125 	/*
2126 	 * We run the long term load balancer infrequently on the first cpu.
2127 	 */
2128 	if (balance_tdq == tdq) {
2129 		if (balance_ticks && --balance_ticks == 0)
2130 			sched_balance();
2131 	}
2132 #endif
2133 	/*
2134 	 * Save the old switch count so we have a record of the last ticks
2135 	 * activity.   Initialize the new switch count based on our load.
2136 	 * If there is some activity seed it to reflect that.
2137 	 */
2138 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2139 	tdq->tdq_switchcnt = tdq->tdq_load;
2140 	/*
2141 	 * Advance the insert index once for each tick to ensure that all
2142 	 * threads get a chance to run.
2143 	 */
2144 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2145 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2146 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2147 			tdq->tdq_ridx = tdq->tdq_idx;
2148 	}
2149 	ts = td->td_sched;
2150 	if (td->td_pri_class & PRI_FIFO_BIT)
2151 		return;
2152 	if (td->td_pri_class == PRI_TIMESHARE) {
2153 		/*
2154 		 * We used a tick; charge it to the thread so
2155 		 * that we can compute our interactivity.
2156 		 */
2157 		td->td_sched->ts_runtime += tickincr;
2158 		sched_interact_update(td);
2159 		sched_priority(td);
2160 	}
2161 	/*
2162 	 * We used up one time slice.
2163 	 */
2164 	if (--ts->ts_slice > 0)
2165 		return;
2166 	/*
2167 	 * We're out of time, force a requeue at userret().
2168 	 */
2169 	ts->ts_slice = sched_slice;
2170 	td->td_flags |= TDF_NEEDRESCHED;
2171 }
2172 
2173 /*
2174  * Called once per hz tick.  Used for cpu utilization information.  This
2175  * is easier than trying to scale based on stathz.
2176  */
2177 void
2178 sched_tick(void)
2179 {
2180 	struct td_sched *ts;
2181 
2182 	ts = curthread->td_sched;
2183 	/*
2184 	 * Ticks is updated asynchronously on a single cpu.  Check here to
2185 	 * avoid incrementing ts_ticks multiple times in a single tick.
2186 	 */
2187 	if (ts->ts_incrtick == ticks)
2188 		return;
2189 	/* Adjust ticks for pctcpu */
2190 	ts->ts_ticks += 1 << SCHED_TICK_SHIFT;
2191 	ts->ts_ltick = ticks;
2192 	ts->ts_incrtick = ticks;
2193 	/*
2194 	 * Update if we've exceeded our desired tick threshhold by over one
2195 	 * second.
2196 	 */
2197 	if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick)
2198 		sched_pctcpu_update(ts);
2199 }
2200 
2201 /*
2202  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2203  * cooperative idle threads.
2204  */
2205 int
2206 sched_runnable(void)
2207 {
2208 	struct tdq *tdq;
2209 	int load;
2210 
2211 	load = 1;
2212 
2213 	tdq = TDQ_SELF();
2214 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2215 		if (tdq->tdq_load > 0)
2216 			goto out;
2217 	} else
2218 		if (tdq->tdq_load - 1 > 0)
2219 			goto out;
2220 	load = 0;
2221 out:
2222 	return (load);
2223 }
2224 
2225 /*
2226  * Choose the highest priority thread to run.  The thread is removed from
2227  * the run-queue while running however the load remains.  For SMP we set
2228  * the tdq in the global idle bitmask if it idles here.
2229  */
2230 struct thread *
2231 sched_choose(void)
2232 {
2233 	struct thread *td;
2234 	struct tdq *tdq;
2235 
2236 	tdq = TDQ_SELF();
2237 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2238 	td = tdq_choose(tdq);
2239 	if (td) {
2240 		td->td_sched->ts_ltick = ticks;
2241 		tdq_runq_rem(tdq, td);
2242 		tdq->tdq_lowpri = td->td_priority;
2243 		return (td);
2244 	}
2245 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2246 	return (PCPU_GET(idlethread));
2247 }
2248 
2249 /*
2250  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2251  * we always request it once we exit a critical section.
2252  */
2253 static inline void
2254 sched_setpreempt(struct thread *td)
2255 {
2256 	struct thread *ctd;
2257 	int cpri;
2258 	int pri;
2259 
2260 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2261 
2262 	ctd = curthread;
2263 	pri = td->td_priority;
2264 	cpri = ctd->td_priority;
2265 	if (pri < cpri)
2266 		ctd->td_flags |= TDF_NEEDRESCHED;
2267 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2268 		return;
2269 	if (!sched_shouldpreempt(pri, cpri, 0))
2270 		return;
2271 	ctd->td_owepreempt = 1;
2272 }
2273 
2274 /*
2275  * Add a thread to a thread queue.  Select the appropriate runq and add the
2276  * thread to it.  This is the internal function called when the tdq is
2277  * predetermined.
2278  */
2279 void
2280 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2281 {
2282 
2283 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2284 	KASSERT((td->td_inhibitors == 0),
2285 	    ("sched_add: trying to run inhibited thread"));
2286 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2287 	    ("sched_add: bad thread state"));
2288 	KASSERT(td->td_flags & TDF_INMEM,
2289 	    ("sched_add: thread swapped out"));
2290 
2291 	if (td->td_priority < tdq->tdq_lowpri)
2292 		tdq->tdq_lowpri = td->td_priority;
2293 	tdq_runq_add(tdq, td, flags);
2294 	tdq_load_add(tdq, td);
2295 }
2296 
2297 /*
2298  * Select the target thread queue and add a thread to it.  Request
2299  * preemption or IPI a remote processor if required.
2300  */
2301 void
2302 sched_add(struct thread *td, int flags)
2303 {
2304 	struct tdq *tdq;
2305 #ifdef SMP
2306 	int cpu;
2307 #endif
2308 
2309 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2310 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2311 	    sched_tdname(curthread));
2312 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2313 	    KTR_ATTR_LINKED, sched_tdname(td));
2314 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2315 	/*
2316 	 * Recalculate the priority before we select the target cpu or
2317 	 * run-queue.
2318 	 */
2319 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2320 		sched_priority(td);
2321 #ifdef SMP
2322 	/*
2323 	 * Pick the destination cpu and if it isn't ours transfer to the
2324 	 * target cpu.
2325 	 */
2326 	cpu = sched_pickcpu(td, flags);
2327 	tdq = sched_setcpu(td, cpu, flags);
2328 	tdq_add(tdq, td, flags);
2329 	if (cpu != PCPU_GET(cpuid)) {
2330 		tdq_notify(tdq, td);
2331 		return;
2332 	}
2333 #else
2334 	tdq = TDQ_SELF();
2335 	TDQ_LOCK(tdq);
2336 	/*
2337 	 * Now that the thread is moving to the run-queue, set the lock
2338 	 * to the scheduler's lock.
2339 	 */
2340 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2341 	tdq_add(tdq, td, flags);
2342 #endif
2343 	if (!(flags & SRQ_YIELDING))
2344 		sched_setpreempt(td);
2345 }
2346 
2347 /*
2348  * Remove a thread from a run-queue without running it.  This is used
2349  * when we're stealing a thread from a remote queue.  Otherwise all threads
2350  * exit by calling sched_exit_thread() and sched_throw() themselves.
2351  */
2352 void
2353 sched_rem(struct thread *td)
2354 {
2355 	struct tdq *tdq;
2356 
2357 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2358 	    "prio:%d", td->td_priority);
2359 	tdq = TDQ_CPU(td->td_sched->ts_cpu);
2360 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2361 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2362 	KASSERT(TD_ON_RUNQ(td),
2363 	    ("sched_rem: thread not on run queue"));
2364 	tdq_runq_rem(tdq, td);
2365 	tdq_load_rem(tdq, td);
2366 	TD_SET_CAN_RUN(td);
2367 	if (td->td_priority == tdq->tdq_lowpri)
2368 		tdq_setlowpri(tdq, NULL);
2369 }
2370 
2371 /*
2372  * Fetch cpu utilization information.  Updates on demand.
2373  */
2374 fixpt_t
2375 sched_pctcpu(struct thread *td)
2376 {
2377 	fixpt_t pctcpu;
2378 	struct td_sched *ts;
2379 
2380 	pctcpu = 0;
2381 	ts = td->td_sched;
2382 	if (ts == NULL)
2383 		return (0);
2384 
2385 	thread_lock(td);
2386 	if (ts->ts_ticks) {
2387 		int rtick;
2388 
2389 		sched_pctcpu_update(ts);
2390 		/* How many rtick per second ? */
2391 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2392 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2393 	}
2394 	thread_unlock(td);
2395 
2396 	return (pctcpu);
2397 }
2398 
2399 /*
2400  * Enforce affinity settings for a thread.  Called after adjustments to
2401  * cpumask.
2402  */
2403 void
2404 sched_affinity(struct thread *td)
2405 {
2406 #ifdef SMP
2407 	struct td_sched *ts;
2408 	int cpu;
2409 
2410 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2411 	ts = td->td_sched;
2412 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2413 		return;
2414 	if (TD_ON_RUNQ(td)) {
2415 		sched_rem(td);
2416 		sched_add(td, SRQ_BORING);
2417 		return;
2418 	}
2419 	if (!TD_IS_RUNNING(td))
2420 		return;
2421 	td->td_flags |= TDF_NEEDRESCHED;
2422 	if (!THREAD_CAN_MIGRATE(td))
2423 		return;
2424 	/*
2425 	 * Assign the new cpu and force a switch before returning to
2426 	 * userspace.  If the target thread is not running locally send
2427 	 * an ipi to force the issue.
2428 	 */
2429 	cpu = ts->ts_cpu;
2430 	ts->ts_cpu = sched_pickcpu(td, 0);
2431 	if (cpu != PCPU_GET(cpuid))
2432 		ipi_selected(1 << cpu, IPI_PREEMPT);
2433 #endif
2434 }
2435 
2436 /*
2437  * Bind a thread to a target cpu.
2438  */
2439 void
2440 sched_bind(struct thread *td, int cpu)
2441 {
2442 	struct td_sched *ts;
2443 
2444 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2445 	ts = td->td_sched;
2446 	if (ts->ts_flags & TSF_BOUND)
2447 		sched_unbind(td);
2448 	ts->ts_flags |= TSF_BOUND;
2449 	sched_pin();
2450 	if (PCPU_GET(cpuid) == cpu)
2451 		return;
2452 	ts->ts_cpu = cpu;
2453 	/* When we return from mi_switch we'll be on the correct cpu. */
2454 	mi_switch(SW_VOL, NULL);
2455 }
2456 
2457 /*
2458  * Release a bound thread.
2459  */
2460 void
2461 sched_unbind(struct thread *td)
2462 {
2463 	struct td_sched *ts;
2464 
2465 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2466 	ts = td->td_sched;
2467 	if ((ts->ts_flags & TSF_BOUND) == 0)
2468 		return;
2469 	ts->ts_flags &= ~TSF_BOUND;
2470 	sched_unpin();
2471 }
2472 
2473 int
2474 sched_is_bound(struct thread *td)
2475 {
2476 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2477 	return (td->td_sched->ts_flags & TSF_BOUND);
2478 }
2479 
2480 /*
2481  * Basic yield call.
2482  */
2483 void
2484 sched_relinquish(struct thread *td)
2485 {
2486 	thread_lock(td);
2487 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2488 	thread_unlock(td);
2489 }
2490 
2491 /*
2492  * Return the total system load.
2493  */
2494 int
2495 sched_load(void)
2496 {
2497 #ifdef SMP
2498 	int total;
2499 	int i;
2500 
2501 	total = 0;
2502 	for (i = 0; i <= mp_maxid; i++)
2503 		total += TDQ_CPU(i)->tdq_sysload;
2504 	return (total);
2505 #else
2506 	return (TDQ_SELF()->tdq_sysload);
2507 #endif
2508 }
2509 
2510 int
2511 sched_sizeof_proc(void)
2512 {
2513 	return (sizeof(struct proc));
2514 }
2515 
2516 int
2517 sched_sizeof_thread(void)
2518 {
2519 	return (sizeof(struct thread) + sizeof(struct td_sched));
2520 }
2521 
2522 #ifdef SMP
2523 #define	TDQ_IDLESPIN(tdq)						\
2524     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2525 #else
2526 #define	TDQ_IDLESPIN(tdq)	1
2527 #endif
2528 
2529 /*
2530  * The actual idle process.
2531  */
2532 void
2533 sched_idletd(void *dummy)
2534 {
2535 	struct thread *td;
2536 	struct tdq *tdq;
2537 	int switchcnt;
2538 	int i;
2539 
2540 	mtx_assert(&Giant, MA_NOTOWNED);
2541 	td = curthread;
2542 	tdq = TDQ_SELF();
2543 	for (;;) {
2544 #ifdef SMP
2545 		if (tdq_idled(tdq) == 0)
2546 			continue;
2547 #endif
2548 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2549 		/*
2550 		 * If we're switching very frequently, spin while checking
2551 		 * for load rather than entering a low power state that
2552 		 * may require an IPI.  However, don't do any busy
2553 		 * loops while on SMT machines as this simply steals
2554 		 * cycles from cores doing useful work.
2555 		 */
2556 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2557 			for (i = 0; i < sched_idlespins; i++) {
2558 				if (tdq->tdq_load)
2559 					break;
2560 				cpu_spinwait();
2561 			}
2562 		}
2563 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2564 		if (tdq->tdq_load == 0)
2565 			cpu_idle(switchcnt > 1);
2566 		if (tdq->tdq_load) {
2567 			thread_lock(td);
2568 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2569 			thread_unlock(td);
2570 		}
2571 	}
2572 }
2573 
2574 /*
2575  * A CPU is entering for the first time or a thread is exiting.
2576  */
2577 void
2578 sched_throw(struct thread *td)
2579 {
2580 	struct thread *newtd;
2581 	struct tdq *tdq;
2582 
2583 	tdq = TDQ_SELF();
2584 	if (td == NULL) {
2585 		/* Correct spinlock nesting and acquire the correct lock. */
2586 		TDQ_LOCK(tdq);
2587 		spinlock_exit();
2588 	} else {
2589 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2590 		tdq_load_rem(tdq, td);
2591 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2592 	}
2593 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2594 	newtd = choosethread();
2595 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2596 	PCPU_SET(switchtime, cpu_ticks());
2597 	PCPU_SET(switchticks, ticks);
2598 	cpu_throw(td, newtd);		/* doesn't return */
2599 }
2600 
2601 /*
2602  * This is called from fork_exit().  Just acquire the correct locks and
2603  * let fork do the rest of the work.
2604  */
2605 void
2606 sched_fork_exit(struct thread *td)
2607 {
2608 	struct td_sched *ts;
2609 	struct tdq *tdq;
2610 	int cpuid;
2611 
2612 	/*
2613 	 * Finish setting up thread glue so that it begins execution in a
2614 	 * non-nested critical section with the scheduler lock held.
2615 	 */
2616 	cpuid = PCPU_GET(cpuid);
2617 	tdq = TDQ_CPU(cpuid);
2618 	ts = td->td_sched;
2619 	if (TD_IS_IDLETHREAD(td))
2620 		td->td_lock = TDQ_LOCKPTR(tdq);
2621 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2622 	td->td_oncpu = cpuid;
2623 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2624 	lock_profile_obtain_lock_success(
2625 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2626 }
2627 
2628 /*
2629  * Create on first use to catch odd startup conditons.
2630  */
2631 char *
2632 sched_tdname(struct thread *td)
2633 {
2634 #ifdef KTR
2635 	struct td_sched *ts;
2636 
2637 	ts = td->td_sched;
2638 	if (ts->ts_name[0] == '\0')
2639 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2640 		    "%s tid %d", td->td_name, td->td_tid);
2641 	return (ts->ts_name);
2642 #else
2643 	return (td->td_name);
2644 #endif
2645 }
2646 
2647 #ifdef SMP
2648 
2649 /*
2650  * Build the CPU topology dump string. Is recursively called to collect
2651  * the topology tree.
2652  */
2653 static int
2654 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2655     int indent)
2656 {
2657 	int i, first;
2658 
2659 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2660 	    "", indent, cg->cg_level);
2661 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"0x%x\">", indent, "",
2662 	    cg->cg_count, cg->cg_mask);
2663 	first = TRUE;
2664 	for (i = 0; i < MAXCPU; i++) {
2665 		if ((cg->cg_mask & (1 << i)) != 0) {
2666 			if (!first)
2667 				sbuf_printf(sb, ", ");
2668 			else
2669 				first = FALSE;
2670 			sbuf_printf(sb, "%d", i);
2671 		}
2672 	}
2673 	sbuf_printf(sb, "</cpu>\n");
2674 
2675 	sbuf_printf(sb, "%*s <flags>", indent, "");
2676 	if (cg->cg_flags != 0) {
2677 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2678 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>\n");
2679 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2680 			sbuf_printf(sb, "<flag name=\"THREAD\">SMT group</flag>\n");
2681 	}
2682 	sbuf_printf(sb, "</flags>\n");
2683 
2684 	if (cg->cg_children > 0) {
2685 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2686 		for (i = 0; i < cg->cg_children; i++)
2687 			sysctl_kern_sched_topology_spec_internal(sb,
2688 			    &cg->cg_child[i], indent+2);
2689 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2690 	}
2691 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2692 	return (0);
2693 }
2694 
2695 /*
2696  * Sysctl handler for retrieving topology dump. It's a wrapper for
2697  * the recursive sysctl_kern_smp_topology_spec_internal().
2698  */
2699 static int
2700 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2701 {
2702 	struct sbuf *topo;
2703 	int err;
2704 
2705 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2706 
2707 	topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2708 	if (topo == NULL)
2709 		return (ENOMEM);
2710 
2711 	sbuf_printf(topo, "<groups>\n");
2712 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2713 	sbuf_printf(topo, "</groups>\n");
2714 
2715 	if (err == 0) {
2716 		sbuf_finish(topo);
2717 		err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2718 	}
2719 	sbuf_delete(topo);
2720 	return (err);
2721 }
2722 #endif
2723 
2724 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2725 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2726     "Scheduler name");
2727 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2728     "Slice size for timeshare threads");
2729 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2730      "Interactivity score threshold");
2731 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh,
2732      0,"Min priority for preemption, lower priorities have greater precedence");
2733 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost,
2734      0,"Controls whether static kernel priorities are assigned to sleeping threads.");
2735 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins,
2736      0,"Number of times idle will spin waiting for new work.");
2737 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, &sched_idlespinthresh,
2738      0,"Threshold before we will permit idle spinning.");
2739 #ifdef SMP
2740 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2741     "Number of hz ticks to keep thread affinity for");
2742 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2743     "Enables the long-term load balancer");
2744 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2745     &balance_interval, 0,
2746     "Average frequency in stathz ticks to run the long-term balancer");
2747 SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0,
2748     "Steals work from another hyper-threaded core on idle");
2749 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2750     "Attempts to steal work from other cores before idling");
2751 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2752     "Minimum load on remote cpu before we'll steal");
2753 
2754 /* Retrieve SMP topology */
2755 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2756     CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2757     "XML dump of detected CPU topology");
2758 #endif
2759 
2760 /* ps compat.  All cpu percentages from ULE are weighted. */
2761 static int ccpu = 0;
2762 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2763