xref: /freebsd/sys/kern/sched_ule.c (revision 1a0fda2b547365c9453523592a445dfe21266d4b)
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 void thread_unblock_switch(struct thread *, struct mtx *);
305 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
306 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
307 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
308     struct cpu_group *cg, int indent);
309 #endif
310 
311 static void sched_setup(void *dummy);
312 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
313 
314 static void sched_initticks(void *dummy);
315 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
316     NULL);
317 
318 /*
319  * Print the threads waiting on a run-queue.
320  */
321 static void
322 runq_print(struct runq *rq)
323 {
324 	struct rqhead *rqh;
325 	struct thread *td;
326 	int pri;
327 	int j;
328 	int i;
329 
330 	for (i = 0; i < RQB_LEN; i++) {
331 		printf("\t\trunq bits %d 0x%zx\n",
332 		    i, rq->rq_status.rqb_bits[i]);
333 		for (j = 0; j < RQB_BPW; j++)
334 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
335 				pri = j + (i << RQB_L2BPW);
336 				rqh = &rq->rq_queues[pri];
337 				TAILQ_FOREACH(td, rqh, td_runq) {
338 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
339 					    td, td->td_name, td->td_priority,
340 					    td->td_rqindex, pri);
341 				}
342 			}
343 	}
344 }
345 
346 /*
347  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
348  */
349 void
350 tdq_print(int cpu)
351 {
352 	struct tdq *tdq;
353 
354 	tdq = TDQ_CPU(cpu);
355 
356 	printf("tdq %d:\n", TDQ_ID(tdq));
357 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
358 	printf("\tLock name:      %s\n", tdq->tdq_name);
359 	printf("\tload:           %d\n", tdq->tdq_load);
360 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
361 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
362 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
363 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
364 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
365 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
366 	printf("\trealtime runq:\n");
367 	runq_print(&tdq->tdq_realtime);
368 	printf("\ttimeshare runq:\n");
369 	runq_print(&tdq->tdq_timeshare);
370 	printf("\tidle runq:\n");
371 	runq_print(&tdq->tdq_idle);
372 }
373 
374 static inline int
375 sched_shouldpreempt(int pri, int cpri, int remote)
376 {
377 	/*
378 	 * If the new priority is not better than the current priority there is
379 	 * nothing to do.
380 	 */
381 	if (pri >= cpri)
382 		return (0);
383 	/*
384 	 * Always preempt idle.
385 	 */
386 	if (cpri >= PRI_MIN_IDLE)
387 		return (1);
388 	/*
389 	 * If preemption is disabled don't preempt others.
390 	 */
391 	if (preempt_thresh == 0)
392 		return (0);
393 	/*
394 	 * Preempt if we exceed the threshold.
395 	 */
396 	if (pri <= preempt_thresh)
397 		return (1);
398 	/*
399 	 * If we're realtime or better and there is timeshare or worse running
400 	 * preempt only remote processors.
401 	 */
402 	if (remote && pri <= PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME)
403 		return (1);
404 	return (0);
405 }
406 
407 #define	TS_RQ_PPQ	(((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS)
408 /*
409  * Add a thread to the actual run-queue.  Keeps transferable counts up to
410  * date with what is actually on the run-queue.  Selects the correct
411  * queue position for timeshare threads.
412  */
413 static __inline void
414 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
415 {
416 	struct td_sched *ts;
417 	u_char pri;
418 
419 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
420 	THREAD_LOCK_ASSERT(td, MA_OWNED);
421 
422 	pri = td->td_priority;
423 	ts = td->td_sched;
424 	TD_SET_RUNQ(td);
425 	if (THREAD_CAN_MIGRATE(td)) {
426 		tdq->tdq_transferable++;
427 		ts->ts_flags |= TSF_XFERABLE;
428 	}
429 	if (pri <= PRI_MAX_REALTIME) {
430 		ts->ts_runq = &tdq->tdq_realtime;
431 	} else if (pri <= PRI_MAX_TIMESHARE) {
432 		ts->ts_runq = &tdq->tdq_timeshare;
433 		KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE,
434 			("Invalid priority %d on timeshare runq", pri));
435 		/*
436 		 * This queue contains only priorities between MIN and MAX
437 		 * realtime.  Use the whole queue to represent these values.
438 		 */
439 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
440 			pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ;
441 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
442 			/*
443 			 * This effectively shortens the queue by one so we
444 			 * can have a one slot difference between idx and
445 			 * ridx while we wait for threads to drain.
446 			 */
447 			if (tdq->tdq_ridx != tdq->tdq_idx &&
448 			    pri == tdq->tdq_ridx)
449 				pri = (unsigned char)(pri - 1) % RQ_NQS;
450 		} else
451 			pri = tdq->tdq_ridx;
452 		runq_add_pri(ts->ts_runq, td, pri, flags);
453 		return;
454 	} else
455 		ts->ts_runq = &tdq->tdq_idle;
456 	runq_add(ts->ts_runq, td, flags);
457 }
458 
459 /*
460  * Remove a thread from a run-queue.  This typically happens when a thread
461  * is selected to run.  Running threads are not on the queue and the
462  * transferable count does not reflect them.
463  */
464 static __inline void
465 tdq_runq_rem(struct tdq *tdq, struct thread *td)
466 {
467 	struct td_sched *ts;
468 
469 	ts = td->td_sched;
470 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
471 	KASSERT(ts->ts_runq != NULL,
472 	    ("tdq_runq_remove: thread %p null ts_runq", td));
473 	if (ts->ts_flags & TSF_XFERABLE) {
474 		tdq->tdq_transferable--;
475 		ts->ts_flags &= ~TSF_XFERABLE;
476 	}
477 	if (ts->ts_runq == &tdq->tdq_timeshare) {
478 		if (tdq->tdq_idx != tdq->tdq_ridx)
479 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
480 		else
481 			runq_remove_idx(ts->ts_runq, td, NULL);
482 	} else
483 		runq_remove(ts->ts_runq, td);
484 }
485 
486 /*
487  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
488  * for this thread to the referenced thread queue.
489  */
490 static void
491 tdq_load_add(struct tdq *tdq, struct thread *td)
492 {
493 
494 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
495 	THREAD_LOCK_ASSERT(td, MA_OWNED);
496 
497 	tdq->tdq_load++;
498 	if ((td->td_flags & TDF_NOLOAD) == 0)
499 		tdq->tdq_sysload++;
500 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
501 }
502 
503 /*
504  * Remove the load from a thread that is transitioning to a sleep state or
505  * exiting.
506  */
507 static void
508 tdq_load_rem(struct tdq *tdq, struct thread *td)
509 {
510 
511 	THREAD_LOCK_ASSERT(td, MA_OWNED);
512 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
513 	KASSERT(tdq->tdq_load != 0,
514 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
515 
516 	tdq->tdq_load--;
517 	if ((td->td_flags & TDF_NOLOAD) == 0)
518 		tdq->tdq_sysload--;
519 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
520 }
521 
522 /*
523  * Set lowpri to its exact value by searching the run-queue and
524  * evaluating curthread.  curthread may be passed as an optimization.
525  */
526 static void
527 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
528 {
529 	struct thread *td;
530 
531 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
532 	if (ctd == NULL)
533 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
534 	td = tdq_choose(tdq);
535 	if (td == NULL || td->td_priority > ctd->td_priority)
536 		tdq->tdq_lowpri = ctd->td_priority;
537 	else
538 		tdq->tdq_lowpri = td->td_priority;
539 }
540 
541 #ifdef SMP
542 struct cpu_search {
543 	cpuset_t cs_mask;
544 	u_int	cs_load;
545 	u_int	cs_cpu;
546 	int	cs_limit;	/* Min priority for low min load for high. */
547 };
548 
549 #define	CPU_SEARCH_LOWEST	0x1
550 #define	CPU_SEARCH_HIGHEST	0x2
551 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
552 
553 #define	CPUSET_FOREACH(cpu, mask)				\
554 	for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++)		\
555 		if ((mask) & 1 << (cpu))
556 
557 static __inline int cpu_search(struct cpu_group *cg, struct cpu_search *low,
558     struct cpu_search *high, const int match);
559 int cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low);
560 int cpu_search_highest(struct cpu_group *cg, struct cpu_search *high);
561 int cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
562     struct cpu_search *high);
563 
564 /*
565  * This routine compares according to the match argument and should be
566  * reduced in actual instantiations via constant propagation and dead code
567  * elimination.
568  */
569 static __inline int
570 cpu_compare(int cpu, struct cpu_search *low, struct cpu_search *high,
571     const int match)
572 {
573 	struct tdq *tdq;
574 
575 	tdq = TDQ_CPU(cpu);
576 	if (match & CPU_SEARCH_LOWEST)
577 		if (CPU_ISSET(cpu, &low->cs_mask) &&
578 		    tdq->tdq_load < low->cs_load &&
579 		    tdq->tdq_lowpri > low->cs_limit) {
580 			low->cs_cpu = cpu;
581 			low->cs_load = tdq->tdq_load;
582 		}
583 	if (match & CPU_SEARCH_HIGHEST)
584 		if (CPU_ISSET(cpu, &high->cs_mask) &&
585 		    tdq->tdq_load >= high->cs_limit &&
586 		    tdq->tdq_load > high->cs_load &&
587 		    tdq->tdq_transferable) {
588 			high->cs_cpu = cpu;
589 			high->cs_load = tdq->tdq_load;
590 		}
591 	return (tdq->tdq_load);
592 }
593 
594 /*
595  * Search the tree of cpu_groups for the lowest or highest loaded cpu
596  * according to the match argument.  This routine actually compares the
597  * load on all paths through the tree and finds the least loaded cpu on
598  * the least loaded path, which may differ from the least loaded cpu in
599  * the system.  This balances work among caches and busses.
600  *
601  * This inline is instantiated in three forms below using constants for the
602  * match argument.  It is reduced to the minimum set for each case.  It is
603  * also recursive to the depth of the tree.
604  */
605 static __inline int
606 cpu_search(struct cpu_group *cg, struct cpu_search *low,
607     struct cpu_search *high, const int match)
608 {
609 	int total;
610 
611 	total = 0;
612 	if (cg->cg_children) {
613 		struct cpu_search lgroup;
614 		struct cpu_search hgroup;
615 		struct cpu_group *child;
616 		u_int lload;
617 		int hload;
618 		int load;
619 		int i;
620 
621 		lload = -1;
622 		hload = -1;
623 		for (i = 0; i < cg->cg_children; i++) {
624 			child = &cg->cg_child[i];
625 			if (match & CPU_SEARCH_LOWEST) {
626 				lgroup = *low;
627 				lgroup.cs_load = -1;
628 			}
629 			if (match & CPU_SEARCH_HIGHEST) {
630 				hgroup = *high;
631 				lgroup.cs_load = 0;
632 			}
633 			switch (match) {
634 			case CPU_SEARCH_LOWEST:
635 				load = cpu_search_lowest(child, &lgroup);
636 				break;
637 			case CPU_SEARCH_HIGHEST:
638 				load = cpu_search_highest(child, &hgroup);
639 				break;
640 			case CPU_SEARCH_BOTH:
641 				load = cpu_search_both(child, &lgroup, &hgroup);
642 				break;
643 			}
644 			total += load;
645 			if (match & CPU_SEARCH_LOWEST)
646 				if (load < lload || low->cs_cpu == -1) {
647 					*low = lgroup;
648 					lload = load;
649 				}
650 			if (match & CPU_SEARCH_HIGHEST)
651 				if (load > hload || high->cs_cpu == -1) {
652 					hload = load;
653 					*high = hgroup;
654 				}
655 		}
656 	} else {
657 		int cpu;
658 
659 		CPUSET_FOREACH(cpu, cg->cg_mask)
660 			total += cpu_compare(cpu, low, high, match);
661 	}
662 	return (total);
663 }
664 
665 /*
666  * cpu_search instantiations must pass constants to maintain the inline
667  * optimization.
668  */
669 int
670 cpu_search_lowest(struct cpu_group *cg, struct cpu_search *low)
671 {
672 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
673 }
674 
675 int
676 cpu_search_highest(struct cpu_group *cg, struct cpu_search *high)
677 {
678 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
679 }
680 
681 int
682 cpu_search_both(struct cpu_group *cg, struct cpu_search *low,
683     struct cpu_search *high)
684 {
685 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
686 }
687 
688 /*
689  * Find the cpu with the least load via the least loaded path that has a
690  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
691  * acceptable.
692  */
693 static inline int
694 sched_lowest(struct cpu_group *cg, cpuset_t mask, int pri)
695 {
696 	struct cpu_search low;
697 
698 	low.cs_cpu = -1;
699 	low.cs_load = -1;
700 	low.cs_mask = mask;
701 	low.cs_limit = pri;
702 	cpu_search_lowest(cg, &low);
703 	return low.cs_cpu;
704 }
705 
706 /*
707  * Find the cpu with the highest load via the highest loaded path.
708  */
709 static inline int
710 sched_highest(struct cpu_group *cg, cpuset_t mask, int minload)
711 {
712 	struct cpu_search high;
713 
714 	high.cs_cpu = -1;
715 	high.cs_load = 0;
716 	high.cs_mask = mask;
717 	high.cs_limit = minload;
718 	cpu_search_highest(cg, &high);
719 	return high.cs_cpu;
720 }
721 
722 /*
723  * Simultaneously find the highest and lowest loaded cpu reachable via
724  * cg.
725  */
726 static inline void
727 sched_both(struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu)
728 {
729 	struct cpu_search high;
730 	struct cpu_search low;
731 
732 	low.cs_cpu = -1;
733 	low.cs_limit = -1;
734 	low.cs_load = -1;
735 	low.cs_mask = mask;
736 	high.cs_load = 0;
737 	high.cs_cpu = -1;
738 	high.cs_limit = -1;
739 	high.cs_mask = mask;
740 	cpu_search_both(cg, &low, &high);
741 	*lowcpu = low.cs_cpu;
742 	*highcpu = high.cs_cpu;
743 	return;
744 }
745 
746 static void
747 sched_balance_group(struct cpu_group *cg)
748 {
749 	cpuset_t mask;
750 	int high;
751 	int low;
752 	int i;
753 
754 	CPU_FILL(&mask);
755 	for (;;) {
756 		sched_both(cg, mask, &low, &high);
757 		if (low == high || low == -1 || high == -1)
758 			break;
759 		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low)))
760 			break;
761 		/*
762 		 * If we failed to move any threads determine which cpu
763 		 * to kick out of the set and try again.
764 	 	 */
765 		if (TDQ_CPU(high)->tdq_transferable == 0)
766 			CPU_CLR(high, &mask);
767 		else
768 			CPU_CLR(low, &mask);
769 	}
770 
771 	for (i = 0; i < cg->cg_children; i++)
772 		sched_balance_group(&cg->cg_child[i]);
773 }
774 
775 static void
776 sched_balance(void)
777 {
778 	struct tdq *tdq;
779 
780 	/*
781 	 * Select a random time between .5 * balance_interval and
782 	 * 1.5 * balance_interval.
783 	 */
784 	balance_ticks = max(balance_interval / 2, 1);
785 	balance_ticks += random() % balance_interval;
786 	if (smp_started == 0 || rebalance == 0)
787 		return;
788 	tdq = TDQ_SELF();
789 	TDQ_UNLOCK(tdq);
790 	sched_balance_group(cpu_top);
791 	TDQ_LOCK(tdq);
792 }
793 
794 /*
795  * Lock two thread queues using their address to maintain lock order.
796  */
797 static void
798 tdq_lock_pair(struct tdq *one, struct tdq *two)
799 {
800 	if (one < two) {
801 		TDQ_LOCK(one);
802 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
803 	} else {
804 		TDQ_LOCK(two);
805 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
806 	}
807 }
808 
809 /*
810  * Unlock two thread queues.  Order is not important here.
811  */
812 static void
813 tdq_unlock_pair(struct tdq *one, struct tdq *two)
814 {
815 	TDQ_UNLOCK(one);
816 	TDQ_UNLOCK(two);
817 }
818 
819 /*
820  * Transfer load between two imbalanced thread queues.
821  */
822 static int
823 sched_balance_pair(struct tdq *high, struct tdq *low)
824 {
825 	int transferable;
826 	int high_load;
827 	int low_load;
828 	int moved;
829 	int move;
830 	int diff;
831 	int i;
832 
833 	tdq_lock_pair(high, low);
834 	transferable = high->tdq_transferable;
835 	high_load = high->tdq_load;
836 	low_load = low->tdq_load;
837 	moved = 0;
838 	/*
839 	 * Determine what the imbalance is and then adjust that to how many
840 	 * threads we actually have to give up (transferable).
841 	 */
842 	if (transferable != 0) {
843 		diff = high_load - low_load;
844 		move = diff / 2;
845 		if (diff & 0x1)
846 			move++;
847 		move = min(move, transferable);
848 		for (i = 0; i < move; i++)
849 			moved += tdq_move(high, low);
850 		/*
851 		 * IPI the target cpu to force it to reschedule with the new
852 		 * workload.
853 		 */
854 		ipi_selected(1 << TDQ_ID(low), IPI_PREEMPT);
855 	}
856 	tdq_unlock_pair(high, low);
857 	return (moved);
858 }
859 
860 /*
861  * Move a thread from one thread queue to another.
862  */
863 static int
864 tdq_move(struct tdq *from, struct tdq *to)
865 {
866 	struct td_sched *ts;
867 	struct thread *td;
868 	struct tdq *tdq;
869 	int cpu;
870 
871 	TDQ_LOCK_ASSERT(from, MA_OWNED);
872 	TDQ_LOCK_ASSERT(to, MA_OWNED);
873 
874 	tdq = from;
875 	cpu = TDQ_ID(to);
876 	td = tdq_steal(tdq, cpu);
877 	if (td == NULL)
878 		return (0);
879 	ts = td->td_sched;
880 	/*
881 	 * Although the run queue is locked the thread may be blocked.  Lock
882 	 * it to clear this and acquire the run-queue lock.
883 	 */
884 	thread_lock(td);
885 	/* Drop recursive lock on from acquired via thread_lock(). */
886 	TDQ_UNLOCK(from);
887 	sched_rem(td);
888 	ts->ts_cpu = cpu;
889 	td->td_lock = TDQ_LOCKPTR(to);
890 	tdq_add(to, td, SRQ_YIELDING);
891 	return (1);
892 }
893 
894 /*
895  * This tdq has idled.  Try to steal a thread from another cpu and switch
896  * to it.
897  */
898 static int
899 tdq_idled(struct tdq *tdq)
900 {
901 	struct cpu_group *cg;
902 	struct tdq *steal;
903 	cpuset_t mask;
904 	int thresh;
905 	int cpu;
906 
907 	if (smp_started == 0 || steal_idle == 0)
908 		return (1);
909 	CPU_FILL(&mask);
910 	CPU_CLR(PCPU_GET(cpuid), &mask);
911 	/* We don't want to be preempted while we're iterating. */
912 	spinlock_enter();
913 	for (cg = tdq->tdq_cg; cg != NULL; ) {
914 		if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
915 			thresh = steal_thresh;
916 		else
917 			thresh = 1;
918 		cpu = sched_highest(cg, mask, thresh);
919 		if (cpu == -1) {
920 			cg = cg->cg_parent;
921 			continue;
922 		}
923 		steal = TDQ_CPU(cpu);
924 		CPU_CLR(cpu, &mask);
925 		tdq_lock_pair(tdq, steal);
926 		if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
927 			tdq_unlock_pair(tdq, steal);
928 			continue;
929 		}
930 		/*
931 		 * If a thread was added while interrupts were disabled don't
932 		 * steal one here.  If we fail to acquire one due to affinity
933 		 * restrictions loop again with this cpu removed from the
934 		 * set.
935 		 */
936 		if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
937 			tdq_unlock_pair(tdq, steal);
938 			continue;
939 		}
940 		spinlock_exit();
941 		TDQ_UNLOCK(steal);
942 		mi_switch(SW_VOL | SWT_IDLE, NULL);
943 		thread_unlock(curthread);
944 
945 		return (0);
946 	}
947 	spinlock_exit();
948 	return (1);
949 }
950 
951 /*
952  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
953  */
954 static void
955 tdq_notify(struct tdq *tdq, struct thread *td)
956 {
957 	struct thread *ctd;
958 	int pri;
959 	int cpu;
960 
961 	if (tdq->tdq_ipipending)
962 		return;
963 	cpu = td->td_sched->ts_cpu;
964 	pri = td->td_priority;
965 	ctd = pcpu_find(cpu)->pc_curthread;
966 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
967 		return;
968 	if (TD_IS_IDLETHREAD(ctd)) {
969 		/*
970 		 * If the MD code has an idle wakeup routine try that before
971 		 * falling back to IPI.
972 		 */
973 		if (cpu_idle_wakeup(cpu))
974 			return;
975 	}
976 	tdq->tdq_ipipending = 1;
977 	ipi_selected(1 << cpu, IPI_PREEMPT);
978 }
979 
980 /*
981  * Steals load from a timeshare queue.  Honors the rotating queue head
982  * index.
983  */
984 static struct thread *
985 runq_steal_from(struct runq *rq, int cpu, u_char start)
986 {
987 	struct rqbits *rqb;
988 	struct rqhead *rqh;
989 	struct thread *td;
990 	int first;
991 	int bit;
992 	int pri;
993 	int i;
994 
995 	rqb = &rq->rq_status;
996 	bit = start & (RQB_BPW -1);
997 	pri = 0;
998 	first = 0;
999 again:
1000 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1001 		if (rqb->rqb_bits[i] == 0)
1002 			continue;
1003 		if (bit != 0) {
1004 			for (pri = bit; pri < RQB_BPW; pri++)
1005 				if (rqb->rqb_bits[i] & (1ul << pri))
1006 					break;
1007 			if (pri >= RQB_BPW)
1008 				continue;
1009 		} else
1010 			pri = RQB_FFS(rqb->rqb_bits[i]);
1011 		pri += (i << RQB_L2BPW);
1012 		rqh = &rq->rq_queues[pri];
1013 		TAILQ_FOREACH(td, rqh, td_runq) {
1014 			if (first && THREAD_CAN_MIGRATE(td) &&
1015 			    THREAD_CAN_SCHED(td, cpu))
1016 				return (td);
1017 			first = 1;
1018 		}
1019 	}
1020 	if (start != 0) {
1021 		start = 0;
1022 		goto again;
1023 	}
1024 
1025 	return (NULL);
1026 }
1027 
1028 /*
1029  * Steals load from a standard linear queue.
1030  */
1031 static struct thread *
1032 runq_steal(struct runq *rq, int cpu)
1033 {
1034 	struct rqhead *rqh;
1035 	struct rqbits *rqb;
1036 	struct thread *td;
1037 	int word;
1038 	int bit;
1039 
1040 	rqb = &rq->rq_status;
1041 	for (word = 0; word < RQB_LEN; word++) {
1042 		if (rqb->rqb_bits[word] == 0)
1043 			continue;
1044 		for (bit = 0; bit < RQB_BPW; bit++) {
1045 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1046 				continue;
1047 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1048 			TAILQ_FOREACH(td, rqh, td_runq)
1049 				if (THREAD_CAN_MIGRATE(td) &&
1050 				    THREAD_CAN_SCHED(td, cpu))
1051 					return (td);
1052 		}
1053 	}
1054 	return (NULL);
1055 }
1056 
1057 /*
1058  * Attempt to steal a thread in priority order from a thread queue.
1059  */
1060 static struct thread *
1061 tdq_steal(struct tdq *tdq, int cpu)
1062 {
1063 	struct thread *td;
1064 
1065 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1066 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1067 		return (td);
1068 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1069 	    cpu, tdq->tdq_ridx)) != NULL)
1070 		return (td);
1071 	return (runq_steal(&tdq->tdq_idle, cpu));
1072 }
1073 
1074 /*
1075  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1076  * current lock and returns with the assigned queue locked.
1077  */
1078 static inline struct tdq *
1079 sched_setcpu(struct thread *td, int cpu, int flags)
1080 {
1081 
1082 	struct tdq *tdq;
1083 
1084 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1085 	tdq = TDQ_CPU(cpu);
1086 	td->td_sched->ts_cpu = cpu;
1087 	/*
1088 	 * If the lock matches just return the queue.
1089 	 */
1090 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1091 		return (tdq);
1092 #ifdef notyet
1093 	/*
1094 	 * If the thread isn't running its lockptr is a
1095 	 * turnstile or a sleepqueue.  We can just lock_set without
1096 	 * blocking.
1097 	 */
1098 	if (TD_CAN_RUN(td)) {
1099 		TDQ_LOCK(tdq);
1100 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1101 		return (tdq);
1102 	}
1103 #endif
1104 	/*
1105 	 * The hard case, migration, we need to block the thread first to
1106 	 * prevent order reversals with other cpus locks.
1107 	 */
1108 	spinlock_enter();
1109 	thread_lock_block(td);
1110 	TDQ_LOCK(tdq);
1111 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1112 	spinlock_exit();
1113 	return (tdq);
1114 }
1115 
1116 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1117 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1118 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1119 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1120 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1121 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1122 
1123 static int
1124 sched_pickcpu(struct thread *td, int flags)
1125 {
1126 	struct cpu_group *cg;
1127 	struct td_sched *ts;
1128 	struct tdq *tdq;
1129 	cpuset_t mask;
1130 	int self;
1131 	int pri;
1132 	int cpu;
1133 
1134 	self = PCPU_GET(cpuid);
1135 	ts = td->td_sched;
1136 	if (smp_started == 0)
1137 		return (self);
1138 	/*
1139 	 * Don't migrate a running thread from sched_switch().
1140 	 */
1141 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1142 		return (ts->ts_cpu);
1143 	/*
1144 	 * Prefer to run interrupt threads on the processors that generate
1145 	 * the interrupt.
1146 	 */
1147 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1148 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1149 		SCHED_STAT_INC(pickcpu_intrbind);
1150 		ts->ts_cpu = self;
1151 	}
1152 	/*
1153 	 * If the thread can run on the last cpu and the affinity has not
1154 	 * expired or it is idle run it there.
1155 	 */
1156 	pri = td->td_priority;
1157 	tdq = TDQ_CPU(ts->ts_cpu);
1158 	if (THREAD_CAN_SCHED(td, ts->ts_cpu)) {
1159 		if (tdq->tdq_lowpri > PRI_MIN_IDLE) {
1160 			SCHED_STAT_INC(pickcpu_idle_affinity);
1161 			return (ts->ts_cpu);
1162 		}
1163 		if (SCHED_AFFINITY(ts, CG_SHARE_L2) && tdq->tdq_lowpri > pri) {
1164 			SCHED_STAT_INC(pickcpu_affinity);
1165 			return (ts->ts_cpu);
1166 		}
1167 	}
1168 	/*
1169 	 * Search for the highest level in the tree that still has affinity.
1170 	 */
1171 	cg = NULL;
1172 	for (cg = tdq->tdq_cg; cg != NULL; cg = cg->cg_parent)
1173 		if (SCHED_AFFINITY(ts, cg->cg_level))
1174 			break;
1175 	cpu = -1;
1176 	mask = td->td_cpuset->cs_mask;
1177 	if (cg)
1178 		cpu = sched_lowest(cg, mask, pri);
1179 	if (cpu == -1)
1180 		cpu = sched_lowest(cpu_top, mask, -1);
1181 	/*
1182 	 * Compare the lowest loaded cpu to current cpu.
1183 	 */
1184 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1185 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE) {
1186 		SCHED_STAT_INC(pickcpu_local);
1187 		cpu = self;
1188 	} else
1189 		SCHED_STAT_INC(pickcpu_lowest);
1190 	if (cpu != ts->ts_cpu)
1191 		SCHED_STAT_INC(pickcpu_migration);
1192 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1193 	return (cpu);
1194 }
1195 #endif
1196 
1197 /*
1198  * Pick the highest priority task we have and return it.
1199  */
1200 static struct thread *
1201 tdq_choose(struct tdq *tdq)
1202 {
1203 	struct thread *td;
1204 
1205 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1206 	td = runq_choose(&tdq->tdq_realtime);
1207 	if (td != NULL)
1208 		return (td);
1209 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1210 	if (td != NULL) {
1211 		KASSERT(td->td_priority >= PRI_MIN_TIMESHARE,
1212 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1213 		    td->td_priority));
1214 		return (td);
1215 	}
1216 	td = runq_choose(&tdq->tdq_idle);
1217 	if (td != NULL) {
1218 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1219 		    ("tdq_choose: Invalid priority on idle queue %d",
1220 		    td->td_priority));
1221 		return (td);
1222 	}
1223 
1224 	return (NULL);
1225 }
1226 
1227 /*
1228  * Initialize a thread queue.
1229  */
1230 static void
1231 tdq_setup(struct tdq *tdq)
1232 {
1233 
1234 	if (bootverbose)
1235 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1236 	runq_init(&tdq->tdq_realtime);
1237 	runq_init(&tdq->tdq_timeshare);
1238 	runq_init(&tdq->tdq_idle);
1239 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1240 	    "sched lock %d", (int)TDQ_ID(tdq));
1241 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1242 	    MTX_SPIN | MTX_RECURSE);
1243 #ifdef KTR
1244 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1245 	    "CPU %d load", (int)TDQ_ID(tdq));
1246 #endif
1247 }
1248 
1249 #ifdef SMP
1250 static void
1251 sched_setup_smp(void)
1252 {
1253 	struct tdq *tdq;
1254 	int i;
1255 
1256 	cpu_top = smp_topo();
1257 	for (i = 0; i < MAXCPU; i++) {
1258 		if (CPU_ABSENT(i))
1259 			continue;
1260 		tdq = TDQ_CPU(i);
1261 		tdq_setup(tdq);
1262 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1263 		if (tdq->tdq_cg == NULL)
1264 			panic("Can't find cpu group for %d\n", i);
1265 	}
1266 	balance_tdq = TDQ_SELF();
1267 	sched_balance();
1268 }
1269 #endif
1270 
1271 /*
1272  * Setup the thread queues and initialize the topology based on MD
1273  * information.
1274  */
1275 static void
1276 sched_setup(void *dummy)
1277 {
1278 	struct tdq *tdq;
1279 
1280 	tdq = TDQ_SELF();
1281 #ifdef SMP
1282 	sched_setup_smp();
1283 #else
1284 	tdq_setup(tdq);
1285 #endif
1286 	/*
1287 	 * To avoid divide-by-zero, we set realstathz a dummy value
1288 	 * in case which sched_clock() called before sched_initticks().
1289 	 */
1290 	realstathz = hz;
1291 	sched_slice = (realstathz/10);	/* ~100ms */
1292 	tickincr = 1 << SCHED_TICK_SHIFT;
1293 
1294 	/* Add thread0's load since it's running. */
1295 	TDQ_LOCK(tdq);
1296 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1297 	tdq_load_add(tdq, &thread0);
1298 	tdq->tdq_lowpri = thread0.td_priority;
1299 	TDQ_UNLOCK(tdq);
1300 }
1301 
1302 /*
1303  * This routine determines the tickincr after stathz and hz are setup.
1304  */
1305 /* ARGSUSED */
1306 static void
1307 sched_initticks(void *dummy)
1308 {
1309 	int incr;
1310 
1311 	realstathz = stathz ? stathz : hz;
1312 	sched_slice = (realstathz/10);	/* ~100ms */
1313 
1314 	/*
1315 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1316 	 * hz not being evenly divisible by stathz on all platforms.
1317 	 */
1318 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1319 	/*
1320 	 * This does not work for values of stathz that are more than
1321 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1322 	 */
1323 	if (incr == 0)
1324 		incr = 1;
1325 	tickincr = incr;
1326 #ifdef SMP
1327 	/*
1328 	 * Set the default balance interval now that we know
1329 	 * what realstathz is.
1330 	 */
1331 	balance_interval = realstathz;
1332 	/*
1333 	 * Set steal thresh to roughly log2(mp_ncpu) but no greater than 4.
1334 	 * This prevents excess thrashing on large machines and excess idle
1335 	 * on smaller machines.
1336 	 */
1337 	steal_thresh = min(fls(mp_ncpus) - 1, 3);
1338 	affinity = SCHED_AFFINITY_DEFAULT;
1339 #endif
1340 }
1341 
1342 
1343 /*
1344  * This is the core of the interactivity algorithm.  Determines a score based
1345  * on past behavior.  It is the ratio of sleep time to run time scaled to
1346  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1347  * differs from the cpu usage because it does not account for time spent
1348  * waiting on a run-queue.  Would be prettier if we had floating point.
1349  */
1350 static int
1351 sched_interact_score(struct thread *td)
1352 {
1353 	struct td_sched *ts;
1354 	int div;
1355 
1356 	ts = td->td_sched;
1357 	/*
1358 	 * The score is only needed if this is likely to be an interactive
1359 	 * task.  Don't go through the expense of computing it if there's
1360 	 * no chance.
1361 	 */
1362 	if (sched_interact <= SCHED_INTERACT_HALF &&
1363 		ts->ts_runtime >= ts->ts_slptime)
1364 			return (SCHED_INTERACT_HALF);
1365 
1366 	if (ts->ts_runtime > ts->ts_slptime) {
1367 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1368 		return (SCHED_INTERACT_HALF +
1369 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1370 	}
1371 	if (ts->ts_slptime > ts->ts_runtime) {
1372 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1373 		return (ts->ts_runtime / div);
1374 	}
1375 	/* runtime == slptime */
1376 	if (ts->ts_runtime)
1377 		return (SCHED_INTERACT_HALF);
1378 
1379 	/*
1380 	 * This can happen if slptime and runtime are 0.
1381 	 */
1382 	return (0);
1383 
1384 }
1385 
1386 /*
1387  * Scale the scheduling priority according to the "interactivity" of this
1388  * process.
1389  */
1390 static void
1391 sched_priority(struct thread *td)
1392 {
1393 	int score;
1394 	int pri;
1395 
1396 	if (td->td_pri_class != PRI_TIMESHARE)
1397 		return;
1398 	/*
1399 	 * If the score is interactive we place the thread in the realtime
1400 	 * queue with a priority that is less than kernel and interrupt
1401 	 * priorities.  These threads are not subject to nice restrictions.
1402 	 *
1403 	 * Scores greater than this are placed on the normal timeshare queue
1404 	 * where the priority is partially decided by the most recent cpu
1405 	 * utilization and the rest is decided by nice value.
1406 	 *
1407 	 * The nice value of the process has a linear effect on the calculated
1408 	 * score.  Negative nice values make it easier for a thread to be
1409 	 * considered interactive.
1410 	 */
1411 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1412 	if (score < sched_interact) {
1413 		pri = PRI_MIN_REALTIME;
1414 		pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact)
1415 		    * score;
1416 		KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME,
1417 		    ("sched_priority: invalid interactive priority %d score %d",
1418 		    pri, score));
1419 	} else {
1420 		pri = SCHED_PRI_MIN;
1421 		if (td->td_sched->ts_ticks)
1422 			pri += SCHED_PRI_TICKS(td->td_sched);
1423 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1424 		KASSERT(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE,
1425 		    ("sched_priority: invalid priority %d: nice %d, "
1426 		    "ticks %d ftick %d ltick %d tick pri %d",
1427 		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1428 		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1429 		    SCHED_PRI_TICKS(td->td_sched)));
1430 	}
1431 	sched_user_prio(td, pri);
1432 
1433 	return;
1434 }
1435 
1436 /*
1437  * This routine enforces a maximum limit on the amount of scheduling history
1438  * kept.  It is called after either the slptime or runtime is adjusted.  This
1439  * function is ugly due to integer math.
1440  */
1441 static void
1442 sched_interact_update(struct thread *td)
1443 {
1444 	struct td_sched *ts;
1445 	u_int sum;
1446 
1447 	ts = td->td_sched;
1448 	sum = ts->ts_runtime + ts->ts_slptime;
1449 	if (sum < SCHED_SLP_RUN_MAX)
1450 		return;
1451 	/*
1452 	 * This only happens from two places:
1453 	 * 1) We have added an unusual amount of run time from fork_exit.
1454 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1455 	 */
1456 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1457 		if (ts->ts_runtime > ts->ts_slptime) {
1458 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1459 			ts->ts_slptime = 1;
1460 		} else {
1461 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1462 			ts->ts_runtime = 1;
1463 		}
1464 		return;
1465 	}
1466 	/*
1467 	 * If we have exceeded by more than 1/5th then the algorithm below
1468 	 * will not bring us back into range.  Dividing by two here forces
1469 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1470 	 */
1471 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1472 		ts->ts_runtime /= 2;
1473 		ts->ts_slptime /= 2;
1474 		return;
1475 	}
1476 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1477 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1478 }
1479 
1480 /*
1481  * Scale back the interactivity history when a child thread is created.  The
1482  * history is inherited from the parent but the thread may behave totally
1483  * differently.  For example, a shell spawning a compiler process.  We want
1484  * to learn that the compiler is behaving badly very quickly.
1485  */
1486 static void
1487 sched_interact_fork(struct thread *td)
1488 {
1489 	int ratio;
1490 	int sum;
1491 
1492 	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1493 	if (sum > SCHED_SLP_RUN_FORK) {
1494 		ratio = sum / SCHED_SLP_RUN_FORK;
1495 		td->td_sched->ts_runtime /= ratio;
1496 		td->td_sched->ts_slptime /= ratio;
1497 	}
1498 }
1499 
1500 /*
1501  * Called from proc0_init() to setup the scheduler fields.
1502  */
1503 void
1504 schedinit(void)
1505 {
1506 
1507 	/*
1508 	 * Set up the scheduler specific parts of proc0.
1509 	 */
1510 	proc0.p_sched = NULL; /* XXX */
1511 	thread0.td_sched = &td_sched0;
1512 	td_sched0.ts_ltick = ticks;
1513 	td_sched0.ts_ftick = ticks;
1514 	td_sched0.ts_slice = sched_slice;
1515 }
1516 
1517 /*
1518  * This is only somewhat accurate since given many processes of the same
1519  * priority they will switch when their slices run out, which will be
1520  * at most sched_slice stathz ticks.
1521  */
1522 int
1523 sched_rr_interval(void)
1524 {
1525 
1526 	/* Convert sched_slice to hz */
1527 	return (hz/(realstathz/sched_slice));
1528 }
1529 
1530 /*
1531  * Update the percent cpu tracking information when it is requested or
1532  * the total history exceeds the maximum.  We keep a sliding history of
1533  * tick counts that slowly decays.  This is less precise than the 4BSD
1534  * mechanism since it happens with less regular and frequent events.
1535  */
1536 static void
1537 sched_pctcpu_update(struct td_sched *ts)
1538 {
1539 
1540 	if (ts->ts_ticks == 0)
1541 		return;
1542 	if (ticks - (hz / 10) < ts->ts_ltick &&
1543 	    SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX)
1544 		return;
1545 	/*
1546 	 * Adjust counters and watermark for pctcpu calc.
1547 	 */
1548 	if (ts->ts_ltick > ticks - SCHED_TICK_TARG)
1549 		ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) *
1550 			    SCHED_TICK_TARG;
1551 	else
1552 		ts->ts_ticks = 0;
1553 	ts->ts_ltick = ticks;
1554 	ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG;
1555 }
1556 
1557 /*
1558  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1559  * if necessary.  This is the back-end for several priority related
1560  * functions.
1561  */
1562 static void
1563 sched_thread_priority(struct thread *td, u_char prio)
1564 {
1565 	struct td_sched *ts;
1566 	struct tdq *tdq;
1567 	int oldpri;
1568 
1569 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1570 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1571 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1572 	if (td != curthread && prio > td->td_priority) {
1573 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1574 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1575 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1576 	}
1577 	ts = td->td_sched;
1578 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1579 	if (td->td_priority == prio)
1580 		return;
1581 	/*
1582 	 * If the priority has been elevated due to priority
1583 	 * propagation, we may have to move ourselves to a new
1584 	 * queue.  This could be optimized to not re-add in some
1585 	 * cases.
1586 	 */
1587 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1588 		sched_rem(td);
1589 		td->td_priority = prio;
1590 		sched_add(td, SRQ_BORROWING);
1591 		return;
1592 	}
1593 	/*
1594 	 * If the thread is currently running we may have to adjust the lowpri
1595 	 * information so other cpus are aware of our current priority.
1596 	 */
1597 	if (TD_IS_RUNNING(td)) {
1598 		tdq = TDQ_CPU(ts->ts_cpu);
1599 		oldpri = td->td_priority;
1600 		td->td_priority = prio;
1601 		if (prio < tdq->tdq_lowpri)
1602 			tdq->tdq_lowpri = prio;
1603 		else if (tdq->tdq_lowpri == oldpri)
1604 			tdq_setlowpri(tdq, td);
1605 		return;
1606 	}
1607 	td->td_priority = prio;
1608 }
1609 
1610 /*
1611  * Update a thread's priority when it is lent another thread's
1612  * priority.
1613  */
1614 void
1615 sched_lend_prio(struct thread *td, u_char prio)
1616 {
1617 
1618 	td->td_flags |= TDF_BORROWING;
1619 	sched_thread_priority(td, prio);
1620 }
1621 
1622 /*
1623  * Restore a thread's priority when priority propagation is
1624  * over.  The prio argument is the minimum priority the thread
1625  * needs to have to satisfy other possible priority lending
1626  * requests.  If the thread's regular priority is less
1627  * important than prio, the thread will keep a priority boost
1628  * of prio.
1629  */
1630 void
1631 sched_unlend_prio(struct thread *td, u_char prio)
1632 {
1633 	u_char base_pri;
1634 
1635 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1636 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1637 		base_pri = td->td_user_pri;
1638 	else
1639 		base_pri = td->td_base_pri;
1640 	if (prio >= base_pri) {
1641 		td->td_flags &= ~TDF_BORROWING;
1642 		sched_thread_priority(td, base_pri);
1643 	} else
1644 		sched_lend_prio(td, prio);
1645 }
1646 
1647 /*
1648  * Standard entry for setting the priority to an absolute value.
1649  */
1650 void
1651 sched_prio(struct thread *td, u_char prio)
1652 {
1653 	u_char oldprio;
1654 
1655 	/* First, update the base priority. */
1656 	td->td_base_pri = prio;
1657 
1658 	/*
1659 	 * If the thread is borrowing another thread's priority, don't
1660 	 * ever lower the priority.
1661 	 */
1662 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1663 		return;
1664 
1665 	/* Change the real priority. */
1666 	oldprio = td->td_priority;
1667 	sched_thread_priority(td, prio);
1668 
1669 	/*
1670 	 * If the thread is on a turnstile, then let the turnstile update
1671 	 * its state.
1672 	 */
1673 	if (TD_ON_LOCK(td) && oldprio != prio)
1674 		turnstile_adjust(td, oldprio);
1675 }
1676 
1677 /*
1678  * Set the base user priority, does not effect current running priority.
1679  */
1680 void
1681 sched_user_prio(struct thread *td, u_char prio)
1682 {
1683 	u_char oldprio;
1684 
1685 	td->td_base_user_pri = prio;
1686 	if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
1687                 return;
1688 	oldprio = td->td_user_pri;
1689 	td->td_user_pri = prio;
1690 }
1691 
1692 void
1693 sched_lend_user_prio(struct thread *td, u_char prio)
1694 {
1695 	u_char oldprio;
1696 
1697 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1698 	td->td_flags |= TDF_UBORROWING;
1699 	oldprio = td->td_user_pri;
1700 	td->td_user_pri = prio;
1701 }
1702 
1703 void
1704 sched_unlend_user_prio(struct thread *td, u_char prio)
1705 {
1706 	u_char base_pri;
1707 
1708 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1709 	base_pri = td->td_base_user_pri;
1710 	if (prio >= base_pri) {
1711 		td->td_flags &= ~TDF_UBORROWING;
1712 		sched_user_prio(td, base_pri);
1713 	} else {
1714 		sched_lend_user_prio(td, prio);
1715 	}
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 	td->td_flags &= ~TDF_NEEDRESCHED;
1792 	td->td_owepreempt = 0;
1793 	tdq->tdq_switchcnt++;
1794 	/*
1795 	 * The lock pointer in an idle thread should never change.  Reset it
1796 	 * to CAN_RUN as well.
1797 	 */
1798 	if (TD_IS_IDLETHREAD(td)) {
1799 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1800 		TD_SET_CAN_RUN(td);
1801 	} else if (TD_IS_RUNNING(td)) {
1802 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1803 		srqflag = (flags & SW_PREEMPT) ?
1804 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1805 		    SRQ_OURSELF|SRQ_YIELDING;
1806 		if (ts->ts_cpu == cpuid)
1807 			tdq_runq_add(tdq, td, srqflag);
1808 		else
1809 			mtx = sched_switch_migrate(tdq, td, srqflag);
1810 	} else {
1811 		/* This thread must be going to sleep. */
1812 		TDQ_LOCK(tdq);
1813 		mtx = thread_lock_block(td);
1814 		tdq_load_rem(tdq, td);
1815 	}
1816 	/*
1817 	 * We enter here with the thread blocked and assigned to the
1818 	 * appropriate cpu run-queue or sleep-queue and with the current
1819 	 * thread-queue locked.
1820 	 */
1821 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1822 	newtd = choosethread();
1823 	/*
1824 	 * Call the MD code to switch contexts if necessary.
1825 	 */
1826 	if (td != newtd) {
1827 #ifdef	HWPMC_HOOKS
1828 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1829 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1830 #endif
1831 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1832 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1833 
1834 #ifdef KDTRACE_HOOKS
1835 		/*
1836 		 * If DTrace has set the active vtime enum to anything
1837 		 * other than INACTIVE (0), then it should have set the
1838 		 * function to call.
1839 		 */
1840 		if (dtrace_vtime_active)
1841 			(*dtrace_vtime_switch_func)(newtd);
1842 #endif
1843 
1844 		cpu_switch(td, newtd, mtx);
1845 		/*
1846 		 * We may return from cpu_switch on a different cpu.  However,
1847 		 * we always return with td_lock pointing to the current cpu's
1848 		 * run queue lock.
1849 		 */
1850 		cpuid = PCPU_GET(cpuid);
1851 		tdq = TDQ_CPU(cpuid);
1852 		lock_profile_obtain_lock_success(
1853 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1854 #ifdef	HWPMC_HOOKS
1855 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1856 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1857 #endif
1858 	} else
1859 		thread_unblock_switch(td, mtx);
1860 	/*
1861 	 * Assert that all went well and return.
1862 	 */
1863 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1864 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1865 	td->td_oncpu = cpuid;
1866 }
1867 
1868 /*
1869  * Adjust thread priorities as a result of a nice request.
1870  */
1871 void
1872 sched_nice(struct proc *p, int nice)
1873 {
1874 	struct thread *td;
1875 
1876 	PROC_LOCK_ASSERT(p, MA_OWNED);
1877 
1878 	p->p_nice = nice;
1879 	FOREACH_THREAD_IN_PROC(p, td) {
1880 		thread_lock(td);
1881 		sched_priority(td);
1882 		sched_prio(td, td->td_base_user_pri);
1883 		thread_unlock(td);
1884 	}
1885 }
1886 
1887 /*
1888  * Record the sleep time for the interactivity scorer.
1889  */
1890 void
1891 sched_sleep(struct thread *td, int prio)
1892 {
1893 
1894 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1895 
1896 	td->td_slptick = ticks;
1897 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1898 		td->td_flags |= TDF_CANSWAP;
1899 	if (static_boost == 1 && prio)
1900 		sched_prio(td, prio);
1901 	else if (static_boost && td->td_priority > static_boost)
1902 		sched_prio(td, static_boost);
1903 }
1904 
1905 /*
1906  * Schedule a thread to resume execution and record how long it voluntarily
1907  * slept.  We also update the pctcpu, interactivity, and priority.
1908  */
1909 void
1910 sched_wakeup(struct thread *td)
1911 {
1912 	struct td_sched *ts;
1913 	int slptick;
1914 
1915 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1916 	ts = td->td_sched;
1917 	td->td_flags &= ~TDF_CANSWAP;
1918 	/*
1919 	 * If we slept for more than a tick update our interactivity and
1920 	 * priority.
1921 	 */
1922 	slptick = td->td_slptick;
1923 	td->td_slptick = 0;
1924 	if (slptick && slptick != ticks) {
1925 		u_int hzticks;
1926 
1927 		hzticks = (ticks - slptick) << SCHED_TICK_SHIFT;
1928 		ts->ts_slptime += hzticks;
1929 		sched_interact_update(td);
1930 		sched_pctcpu_update(ts);
1931 	}
1932 	/* Reset the slice value after we sleep. */
1933 	ts->ts_slice = sched_slice;
1934 	sched_add(td, SRQ_BORING);
1935 }
1936 
1937 /*
1938  * Penalize the parent for creating a new child and initialize the child's
1939  * priority.
1940  */
1941 void
1942 sched_fork(struct thread *td, struct thread *child)
1943 {
1944 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1945 	sched_fork_thread(td, child);
1946 	/*
1947 	 * Penalize the parent and child for forking.
1948 	 */
1949 	sched_interact_fork(child);
1950 	sched_priority(child);
1951 	td->td_sched->ts_runtime += tickincr;
1952 	sched_interact_update(td);
1953 	sched_priority(td);
1954 }
1955 
1956 /*
1957  * Fork a new thread, may be within the same process.
1958  */
1959 void
1960 sched_fork_thread(struct thread *td, struct thread *child)
1961 {
1962 	struct td_sched *ts;
1963 	struct td_sched *ts2;
1964 
1965 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1966 	/*
1967 	 * Initialize child.
1968 	 */
1969 	ts = td->td_sched;
1970 	ts2 = child->td_sched;
1971 	child->td_lock = TDQ_LOCKPTR(TDQ_SELF());
1972 	child->td_cpuset = cpuset_ref(td->td_cpuset);
1973 	ts2->ts_cpu = ts->ts_cpu;
1974 	ts2->ts_flags = 0;
1975 	/*
1976 	 * Grab our parents cpu estimation information and priority.
1977 	 */
1978 	ts2->ts_ticks = ts->ts_ticks;
1979 	ts2->ts_ltick = ts->ts_ltick;
1980 	ts2->ts_incrtick = ts->ts_incrtick;
1981 	ts2->ts_ftick = ts->ts_ftick;
1982 	child->td_user_pri = td->td_user_pri;
1983 	child->td_base_user_pri = td->td_base_user_pri;
1984 	/*
1985 	 * And update interactivity score.
1986 	 */
1987 	ts2->ts_slptime = ts->ts_slptime;
1988 	ts2->ts_runtime = ts->ts_runtime;
1989 	ts2->ts_slice = 1;	/* Attempt to quickly learn interactivity. */
1990 #ifdef KTR
1991 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
1992 #endif
1993 }
1994 
1995 /*
1996  * Adjust the priority class of a thread.
1997  */
1998 void
1999 sched_class(struct thread *td, int class)
2000 {
2001 
2002 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2003 	if (td->td_pri_class == class)
2004 		return;
2005 	td->td_pri_class = class;
2006 }
2007 
2008 /*
2009  * Return some of the child's priority and interactivity to the parent.
2010  */
2011 void
2012 sched_exit(struct proc *p, struct thread *child)
2013 {
2014 	struct thread *td;
2015 
2016 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2017 	    "prio:td", child->td_priority);
2018 	PROC_LOCK_ASSERT(p, MA_OWNED);
2019 	td = FIRST_THREAD_IN_PROC(p);
2020 	sched_exit_thread(td, child);
2021 }
2022 
2023 /*
2024  * Penalize another thread for the time spent on this one.  This helps to
2025  * worsen the priority and interactivity of processes which schedule batch
2026  * jobs such as make.  This has little effect on the make process itself but
2027  * causes new processes spawned by it to receive worse scores immediately.
2028  */
2029 void
2030 sched_exit_thread(struct thread *td, struct thread *child)
2031 {
2032 
2033 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2034 	    "prio:td", child->td_priority);
2035 	/*
2036 	 * Give the child's runtime to the parent without returning the
2037 	 * sleep time as a penalty to the parent.  This causes shells that
2038 	 * launch expensive things to mark their children as expensive.
2039 	 */
2040 	thread_lock(td);
2041 	td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2042 	sched_interact_update(td);
2043 	sched_priority(td);
2044 	thread_unlock(td);
2045 }
2046 
2047 void
2048 sched_preempt(struct thread *td)
2049 {
2050 	struct tdq *tdq;
2051 
2052 	thread_lock(td);
2053 	tdq = TDQ_SELF();
2054 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2055 	tdq->tdq_ipipending = 0;
2056 	if (td->td_priority > tdq->tdq_lowpri) {
2057 		int flags;
2058 
2059 		flags = SW_INVOL | SW_PREEMPT;
2060 		if (td->td_critnest > 1)
2061 			td->td_owepreempt = 1;
2062 		else if (TD_IS_IDLETHREAD(td))
2063 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2064 		else
2065 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2066 	}
2067 	thread_unlock(td);
2068 }
2069 
2070 /*
2071  * Fix priorities on return to user-space.  Priorities may be elevated due
2072  * to static priorities in msleep() or similar.
2073  */
2074 void
2075 sched_userret(struct thread *td)
2076 {
2077 	/*
2078 	 * XXX we cheat slightly on the locking here to avoid locking in
2079 	 * the usual case.  Setting td_priority here is essentially an
2080 	 * incomplete workaround for not setting it properly elsewhere.
2081 	 * Now that some interrupt handlers are threads, not setting it
2082 	 * properly elsewhere can clobber it in the window between setting
2083 	 * it here and returning to user mode, so don't waste time setting
2084 	 * it perfectly here.
2085 	 */
2086 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2087 	    ("thread with borrowed priority returning to userland"));
2088 	if (td->td_priority != td->td_user_pri) {
2089 		thread_lock(td);
2090 		td->td_priority = td->td_user_pri;
2091 		td->td_base_pri = td->td_user_pri;
2092 		tdq_setlowpri(TDQ_SELF(), td);
2093 		thread_unlock(td);
2094         }
2095 }
2096 
2097 /*
2098  * Handle a stathz tick.  This is really only relevant for timeshare
2099  * threads.
2100  */
2101 void
2102 sched_clock(struct thread *td)
2103 {
2104 	struct tdq *tdq;
2105 	struct td_sched *ts;
2106 
2107 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2108 	tdq = TDQ_SELF();
2109 #ifdef SMP
2110 	/*
2111 	 * We run the long term load balancer infrequently on the first cpu.
2112 	 */
2113 	if (balance_tdq == tdq) {
2114 		if (balance_ticks && --balance_ticks == 0)
2115 			sched_balance();
2116 	}
2117 #endif
2118 	/*
2119 	 * Save the old switch count so we have a record of the last ticks
2120 	 * activity.   Initialize the new switch count based on our load.
2121 	 * If there is some activity seed it to reflect that.
2122 	 */
2123 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2124 	tdq->tdq_switchcnt = tdq->tdq_load;
2125 	/*
2126 	 * Advance the insert index once for each tick to ensure that all
2127 	 * threads get a chance to run.
2128 	 */
2129 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2130 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2131 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2132 			tdq->tdq_ridx = tdq->tdq_idx;
2133 	}
2134 	ts = td->td_sched;
2135 	if (td->td_pri_class & PRI_FIFO_BIT)
2136 		return;
2137 	if (td->td_pri_class == PRI_TIMESHARE) {
2138 		/*
2139 		 * We used a tick; charge it to the thread so
2140 		 * that we can compute our interactivity.
2141 		 */
2142 		td->td_sched->ts_runtime += tickincr;
2143 		sched_interact_update(td);
2144 		sched_priority(td);
2145 	}
2146 	/*
2147 	 * We used up one time slice.
2148 	 */
2149 	if (--ts->ts_slice > 0)
2150 		return;
2151 	/*
2152 	 * We're out of time, force a requeue at userret().
2153 	 */
2154 	ts->ts_slice = sched_slice;
2155 	td->td_flags |= TDF_NEEDRESCHED;
2156 }
2157 
2158 /*
2159  * Called once per hz tick.  Used for cpu utilization information.  This
2160  * is easier than trying to scale based on stathz.
2161  */
2162 void
2163 sched_tick(void)
2164 {
2165 	struct td_sched *ts;
2166 
2167 	ts = curthread->td_sched;
2168 	/*
2169 	 * Ticks is updated asynchronously on a single cpu.  Check here to
2170 	 * avoid incrementing ts_ticks multiple times in a single tick.
2171 	 */
2172 	if (ts->ts_incrtick == ticks)
2173 		return;
2174 	/* Adjust ticks for pctcpu */
2175 	ts->ts_ticks += 1 << SCHED_TICK_SHIFT;
2176 	ts->ts_ltick = ticks;
2177 	ts->ts_incrtick = ticks;
2178 	/*
2179 	 * Update if we've exceeded our desired tick threshhold by over one
2180 	 * second.
2181 	 */
2182 	if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick)
2183 		sched_pctcpu_update(ts);
2184 }
2185 
2186 /*
2187  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2188  * cooperative idle threads.
2189  */
2190 int
2191 sched_runnable(void)
2192 {
2193 	struct tdq *tdq;
2194 	int load;
2195 
2196 	load = 1;
2197 
2198 	tdq = TDQ_SELF();
2199 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2200 		if (tdq->tdq_load > 0)
2201 			goto out;
2202 	} else
2203 		if (tdq->tdq_load - 1 > 0)
2204 			goto out;
2205 	load = 0;
2206 out:
2207 	return (load);
2208 }
2209 
2210 /*
2211  * Choose the highest priority thread to run.  The thread is removed from
2212  * the run-queue while running however the load remains.  For SMP we set
2213  * the tdq in the global idle bitmask if it idles here.
2214  */
2215 struct thread *
2216 sched_choose(void)
2217 {
2218 	struct thread *td;
2219 	struct tdq *tdq;
2220 
2221 	tdq = TDQ_SELF();
2222 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2223 	td = tdq_choose(tdq);
2224 	if (td) {
2225 		td->td_sched->ts_ltick = ticks;
2226 		tdq_runq_rem(tdq, td);
2227 		tdq->tdq_lowpri = td->td_priority;
2228 		return (td);
2229 	}
2230 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2231 	return (PCPU_GET(idlethread));
2232 }
2233 
2234 /*
2235  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2236  * we always request it once we exit a critical section.
2237  */
2238 static inline void
2239 sched_setpreempt(struct thread *td)
2240 {
2241 	struct thread *ctd;
2242 	int cpri;
2243 	int pri;
2244 
2245 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2246 
2247 	ctd = curthread;
2248 	pri = td->td_priority;
2249 	cpri = ctd->td_priority;
2250 	if (pri < cpri)
2251 		ctd->td_flags |= TDF_NEEDRESCHED;
2252 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2253 		return;
2254 	if (!sched_shouldpreempt(pri, cpri, 0))
2255 		return;
2256 	ctd->td_owepreempt = 1;
2257 }
2258 
2259 /*
2260  * Add a thread to a thread queue.  Select the appropriate runq and add the
2261  * thread to it.  This is the internal function called when the tdq is
2262  * predetermined.
2263  */
2264 void
2265 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2266 {
2267 
2268 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2269 	KASSERT((td->td_inhibitors == 0),
2270 	    ("sched_add: trying to run inhibited thread"));
2271 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2272 	    ("sched_add: bad thread state"));
2273 	KASSERT(td->td_flags & TDF_INMEM,
2274 	    ("sched_add: thread swapped out"));
2275 
2276 	if (td->td_priority < tdq->tdq_lowpri)
2277 		tdq->tdq_lowpri = td->td_priority;
2278 	tdq_runq_add(tdq, td, flags);
2279 	tdq_load_add(tdq, td);
2280 }
2281 
2282 /*
2283  * Select the target thread queue and add a thread to it.  Request
2284  * preemption or IPI a remote processor if required.
2285  */
2286 void
2287 sched_add(struct thread *td, int flags)
2288 {
2289 	struct tdq *tdq;
2290 #ifdef SMP
2291 	int cpu;
2292 #endif
2293 
2294 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2295 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2296 	    sched_tdname(curthread));
2297 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2298 	    KTR_ATTR_LINKED, sched_tdname(td));
2299 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2300 	/*
2301 	 * Recalculate the priority before we select the target cpu or
2302 	 * run-queue.
2303 	 */
2304 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2305 		sched_priority(td);
2306 #ifdef SMP
2307 	/*
2308 	 * Pick the destination cpu and if it isn't ours transfer to the
2309 	 * target cpu.
2310 	 */
2311 	cpu = sched_pickcpu(td, flags);
2312 	tdq = sched_setcpu(td, cpu, flags);
2313 	tdq_add(tdq, td, flags);
2314 	if (cpu != PCPU_GET(cpuid)) {
2315 		tdq_notify(tdq, td);
2316 		return;
2317 	}
2318 #else
2319 	tdq = TDQ_SELF();
2320 	TDQ_LOCK(tdq);
2321 	/*
2322 	 * Now that the thread is moving to the run-queue, set the lock
2323 	 * to the scheduler's lock.
2324 	 */
2325 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2326 	tdq_add(tdq, td, flags);
2327 #endif
2328 	if (!(flags & SRQ_YIELDING))
2329 		sched_setpreempt(td);
2330 }
2331 
2332 /*
2333  * Remove a thread from a run-queue without running it.  This is used
2334  * when we're stealing a thread from a remote queue.  Otherwise all threads
2335  * exit by calling sched_exit_thread() and sched_throw() themselves.
2336  */
2337 void
2338 sched_rem(struct thread *td)
2339 {
2340 	struct tdq *tdq;
2341 
2342 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2343 	    "prio:%d", td->td_priority);
2344 	tdq = TDQ_CPU(td->td_sched->ts_cpu);
2345 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2346 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2347 	KASSERT(TD_ON_RUNQ(td),
2348 	    ("sched_rem: thread not on run queue"));
2349 	tdq_runq_rem(tdq, td);
2350 	tdq_load_rem(tdq, td);
2351 	TD_SET_CAN_RUN(td);
2352 	if (td->td_priority == tdq->tdq_lowpri)
2353 		tdq_setlowpri(tdq, NULL);
2354 }
2355 
2356 /*
2357  * Fetch cpu utilization information.  Updates on demand.
2358  */
2359 fixpt_t
2360 sched_pctcpu(struct thread *td)
2361 {
2362 	fixpt_t pctcpu;
2363 	struct td_sched *ts;
2364 
2365 	pctcpu = 0;
2366 	ts = td->td_sched;
2367 	if (ts == NULL)
2368 		return (0);
2369 
2370 	thread_lock(td);
2371 	if (ts->ts_ticks) {
2372 		int rtick;
2373 
2374 		sched_pctcpu_update(ts);
2375 		/* How many rtick per second ? */
2376 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2377 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2378 	}
2379 	thread_unlock(td);
2380 
2381 	return (pctcpu);
2382 }
2383 
2384 /*
2385  * Enforce affinity settings for a thread.  Called after adjustments to
2386  * cpumask.
2387  */
2388 void
2389 sched_affinity(struct thread *td)
2390 {
2391 #ifdef SMP
2392 	struct td_sched *ts;
2393 	int cpu;
2394 
2395 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2396 	ts = td->td_sched;
2397 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2398 		return;
2399 	if (TD_ON_RUNQ(td)) {
2400 		sched_rem(td);
2401 		sched_add(td, SRQ_BORING);
2402 		return;
2403 	}
2404 	if (!TD_IS_RUNNING(td))
2405 		return;
2406 	td->td_flags |= TDF_NEEDRESCHED;
2407 	if (!THREAD_CAN_MIGRATE(td))
2408 		return;
2409 	/*
2410 	 * Assign the new cpu and force a switch before returning to
2411 	 * userspace.  If the target thread is not running locally send
2412 	 * an ipi to force the issue.
2413 	 */
2414 	cpu = ts->ts_cpu;
2415 	ts->ts_cpu = sched_pickcpu(td, 0);
2416 	if (cpu != PCPU_GET(cpuid))
2417 		ipi_selected(1 << cpu, IPI_PREEMPT);
2418 #endif
2419 }
2420 
2421 /*
2422  * Bind a thread to a target cpu.
2423  */
2424 void
2425 sched_bind(struct thread *td, int cpu)
2426 {
2427 	struct td_sched *ts;
2428 
2429 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2430 	ts = td->td_sched;
2431 	if (ts->ts_flags & TSF_BOUND)
2432 		sched_unbind(td);
2433 	ts->ts_flags |= TSF_BOUND;
2434 	sched_pin();
2435 	if (PCPU_GET(cpuid) == cpu)
2436 		return;
2437 	ts->ts_cpu = cpu;
2438 	/* When we return from mi_switch we'll be on the correct cpu. */
2439 	mi_switch(SW_VOL, NULL);
2440 }
2441 
2442 /*
2443  * Release a bound thread.
2444  */
2445 void
2446 sched_unbind(struct thread *td)
2447 {
2448 	struct td_sched *ts;
2449 
2450 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2451 	ts = td->td_sched;
2452 	if ((ts->ts_flags & TSF_BOUND) == 0)
2453 		return;
2454 	ts->ts_flags &= ~TSF_BOUND;
2455 	sched_unpin();
2456 }
2457 
2458 int
2459 sched_is_bound(struct thread *td)
2460 {
2461 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2462 	return (td->td_sched->ts_flags & TSF_BOUND);
2463 }
2464 
2465 /*
2466  * Basic yield call.
2467  */
2468 void
2469 sched_relinquish(struct thread *td)
2470 {
2471 	thread_lock(td);
2472 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2473 	thread_unlock(td);
2474 }
2475 
2476 /*
2477  * Return the total system load.
2478  */
2479 int
2480 sched_load(void)
2481 {
2482 #ifdef SMP
2483 	int total;
2484 	int i;
2485 
2486 	total = 0;
2487 	for (i = 0; i <= mp_maxid; i++)
2488 		total += TDQ_CPU(i)->tdq_sysload;
2489 	return (total);
2490 #else
2491 	return (TDQ_SELF()->tdq_sysload);
2492 #endif
2493 }
2494 
2495 int
2496 sched_sizeof_proc(void)
2497 {
2498 	return (sizeof(struct proc));
2499 }
2500 
2501 int
2502 sched_sizeof_thread(void)
2503 {
2504 	return (sizeof(struct thread) + sizeof(struct td_sched));
2505 }
2506 
2507 #ifdef SMP
2508 #define	TDQ_IDLESPIN(tdq)						\
2509     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2510 #else
2511 #define	TDQ_IDLESPIN(tdq)	1
2512 #endif
2513 
2514 /*
2515  * The actual idle process.
2516  */
2517 void
2518 sched_idletd(void *dummy)
2519 {
2520 	struct thread *td;
2521 	struct tdq *tdq;
2522 	int switchcnt;
2523 	int i;
2524 
2525 	mtx_assert(&Giant, MA_NOTOWNED);
2526 	td = curthread;
2527 	tdq = TDQ_SELF();
2528 	for (;;) {
2529 #ifdef SMP
2530 		if (tdq_idled(tdq) == 0)
2531 			continue;
2532 #endif
2533 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2534 		/*
2535 		 * If we're switching very frequently, spin while checking
2536 		 * for load rather than entering a low power state that
2537 		 * may require an IPI.  However, don't do any busy
2538 		 * loops while on SMT machines as this simply steals
2539 		 * cycles from cores doing useful work.
2540 		 */
2541 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2542 			for (i = 0; i < sched_idlespins; i++) {
2543 				if (tdq->tdq_load)
2544 					break;
2545 				cpu_spinwait();
2546 			}
2547 		}
2548 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2549 		if (tdq->tdq_load == 0)
2550 			cpu_idle(switchcnt > 1);
2551 		if (tdq->tdq_load) {
2552 			thread_lock(td);
2553 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2554 			thread_unlock(td);
2555 		}
2556 	}
2557 }
2558 
2559 /*
2560  * A CPU is entering for the first time or a thread is exiting.
2561  */
2562 void
2563 sched_throw(struct thread *td)
2564 {
2565 	struct thread *newtd;
2566 	struct tdq *tdq;
2567 
2568 	tdq = TDQ_SELF();
2569 	if (td == NULL) {
2570 		/* Correct spinlock nesting and acquire the correct lock. */
2571 		TDQ_LOCK(tdq);
2572 		spinlock_exit();
2573 	} else {
2574 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2575 		tdq_load_rem(tdq, td);
2576 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2577 	}
2578 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2579 	newtd = choosethread();
2580 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2581 	PCPU_SET(switchtime, cpu_ticks());
2582 	PCPU_SET(switchticks, ticks);
2583 	cpu_throw(td, newtd);		/* doesn't return */
2584 }
2585 
2586 /*
2587  * This is called from fork_exit().  Just acquire the correct locks and
2588  * let fork do the rest of the work.
2589  */
2590 void
2591 sched_fork_exit(struct thread *td)
2592 {
2593 	struct td_sched *ts;
2594 	struct tdq *tdq;
2595 	int cpuid;
2596 
2597 	/*
2598 	 * Finish setting up thread glue so that it begins execution in a
2599 	 * non-nested critical section with the scheduler lock held.
2600 	 */
2601 	cpuid = PCPU_GET(cpuid);
2602 	tdq = TDQ_CPU(cpuid);
2603 	ts = td->td_sched;
2604 	if (TD_IS_IDLETHREAD(td))
2605 		td->td_lock = TDQ_LOCKPTR(tdq);
2606 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2607 	td->td_oncpu = cpuid;
2608 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2609 	lock_profile_obtain_lock_success(
2610 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2611 }
2612 
2613 /*
2614  * Create on first use to catch odd startup conditons.
2615  */
2616 char *
2617 sched_tdname(struct thread *td)
2618 {
2619 #ifdef KTR
2620 	struct td_sched *ts;
2621 
2622 	ts = td->td_sched;
2623 	if (ts->ts_name[0] == '\0')
2624 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2625 		    "%s tid %d", td->td_name, td->td_tid);
2626 	return (ts->ts_name);
2627 #else
2628 	return (td->td_name);
2629 #endif
2630 }
2631 
2632 #ifdef SMP
2633 
2634 /*
2635  * Build the CPU topology dump string. Is recursively called to collect
2636  * the topology tree.
2637  */
2638 static int
2639 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2640     int indent)
2641 {
2642 	int i, first;
2643 
2644 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2645 	    "", indent, cg->cg_level);
2646 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"0x%x\">", indent, "",
2647 	    cg->cg_count, cg->cg_mask);
2648 	first = TRUE;
2649 	for (i = 0; i < MAXCPU; i++) {
2650 		if ((cg->cg_mask & (1 << i)) != 0) {
2651 			if (!first)
2652 				sbuf_printf(sb, ", ");
2653 			else
2654 				first = FALSE;
2655 			sbuf_printf(sb, "%d", i);
2656 		}
2657 	}
2658 	sbuf_printf(sb, "</cpu>\n");
2659 
2660 	sbuf_printf(sb, "%*s <flags>", indent, "");
2661 	if (cg->cg_flags != 0) {
2662 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2663 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>\n");
2664 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2665 			sbuf_printf(sb, "<flag name=\"THREAD\">SMT group</flag>\n");
2666 	}
2667 	sbuf_printf(sb, "</flags>\n");
2668 
2669 	if (cg->cg_children > 0) {
2670 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2671 		for (i = 0; i < cg->cg_children; i++)
2672 			sysctl_kern_sched_topology_spec_internal(sb,
2673 			    &cg->cg_child[i], indent+2);
2674 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2675 	}
2676 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2677 	return (0);
2678 }
2679 
2680 /*
2681  * Sysctl handler for retrieving topology dump. It's a wrapper for
2682  * the recursive sysctl_kern_smp_topology_spec_internal().
2683  */
2684 static int
2685 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2686 {
2687 	struct sbuf *topo;
2688 	int err;
2689 
2690 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2691 
2692 	topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2693 	if (topo == NULL)
2694 		return (ENOMEM);
2695 
2696 	sbuf_printf(topo, "<groups>\n");
2697 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2698 	sbuf_printf(topo, "</groups>\n");
2699 
2700 	if (err == 0) {
2701 		sbuf_finish(topo);
2702 		err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2703 	}
2704 	sbuf_delete(topo);
2705 	return (err);
2706 }
2707 #endif
2708 
2709 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2710 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2711     "Scheduler name");
2712 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2713     "Slice size for timeshare threads");
2714 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2715      "Interactivity score threshold");
2716 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW, &preempt_thresh,
2717      0,"Min priority for preemption, lower priorities have greater precedence");
2718 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost,
2719      0,"Controls whether static kernel priorities are assigned to sleeping threads.");
2720 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins,
2721      0,"Number of times idle will spin waiting for new work.");
2722 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW, &sched_idlespinthresh,
2723      0,"Threshold before we will permit idle spinning.");
2724 #ifdef SMP
2725 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2726     "Number of hz ticks to keep thread affinity for");
2727 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2728     "Enables the long-term load balancer");
2729 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2730     &balance_interval, 0,
2731     "Average frequency in stathz ticks to run the long-term balancer");
2732 SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0,
2733     "Steals work from another hyper-threaded core on idle");
2734 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2735     "Attempts to steal work from other cores before idling");
2736 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2737     "Minimum load on remote cpu before we'll steal");
2738 
2739 /* Retrieve SMP topology */
2740 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2741     CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2742     "XML dump of detected CPU topology");
2743 #endif
2744 
2745 /* ps compat.  All cpu percentages from ULE are weighted. */
2746 static int ccpu = 0;
2747 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2748