xref: /freebsd/sys/kern/sched_ule.c (revision 6829dae12bb055451fa467da4589c43bd03b1e64)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice unmodified, this list of conditions, and the following
12  *    disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 /*
30  * This file implements the ULE scheduler.  ULE supports independent CPU
31  * run queues and fine grain locking.  It has superior interactive
32  * performance under load even on uni-processor systems.
33  *
34  * etymology:
35  *   ULE is the last three letters in schedule.  It owes its name to a
36  * generic user created for a scheduling system by Paul Mikesell at
37  * Isilon Systems and a general lack of creativity on the part of the author.
38  */
39 
40 #include <sys/cdefs.h>
41 __FBSDID("$FreeBSD$");
42 
43 #include "opt_hwpmc_hooks.h"
44 #include "opt_sched.h"
45 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kdb.h>
49 #include <sys/kernel.h>
50 #include <sys/ktr.h>
51 #include <sys/limits.h>
52 #include <sys/lock.h>
53 #include <sys/mutex.h>
54 #include <sys/proc.h>
55 #include <sys/resource.h>
56 #include <sys/resourcevar.h>
57 #include <sys/sched.h>
58 #include <sys/sdt.h>
59 #include <sys/smp.h>
60 #include <sys/sx.h>
61 #include <sys/sysctl.h>
62 #include <sys/sysproto.h>
63 #include <sys/turnstile.h>
64 #include <sys/umtx.h>
65 #include <sys/vmmeter.h>
66 #include <sys/cpuset.h>
67 #include <sys/sbuf.h>
68 
69 #ifdef HWPMC_HOOKS
70 #include <sys/pmckern.h>
71 #endif
72 
73 #ifdef KDTRACE_HOOKS
74 #include <sys/dtrace_bsd.h>
75 int				dtrace_vtime_active;
76 dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
77 #endif
78 
79 #include <machine/cpu.h>
80 #include <machine/smp.h>
81 
82 #define	KTR_ULE	0
83 
84 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
85 #define	TDQ_NAME_LEN	(sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
86 #define	TDQ_LOADNAME_LEN	(sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
87 
88 /*
89  * Thread scheduler specific section.  All fields are protected
90  * by the thread lock.
91  */
92 struct td_sched {
93 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
94 	short		ts_flags;	/* TSF_* flags. */
95 	int		ts_cpu;		/* CPU that we have affinity for. */
96 	int		ts_rltick;	/* Real last tick, for affinity. */
97 	int		ts_slice;	/* Ticks of slice remaining. */
98 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
99 	u_int		ts_runtime;	/* Number of ticks we were running */
100 	int		ts_ltick;	/* Last tick that we were running on */
101 	int		ts_ftick;	/* First tick that we were running on */
102 	int		ts_ticks;	/* Tick count */
103 #ifdef KTR
104 	char		ts_name[TS_NAME_LEN];
105 #endif
106 };
107 /* flags kept in ts_flags */
108 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
109 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
110 
111 #define	THREAD_CAN_MIGRATE(td)	((td)->td_pinned == 0)
112 #define	THREAD_CAN_SCHED(td, cpu)	\
113     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
114 
115 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <=
116     sizeof(struct thread0_storage),
117     "increase struct thread0_storage.t0st_sched size");
118 
119 /*
120  * Priority ranges used for interactive and non-interactive timeshare
121  * threads.  The timeshare priorities are split up into four ranges.
122  * The first range handles interactive threads.  The last three ranges
123  * (NHALF, x, and NHALF) handle non-interactive threads with the outer
124  * ranges supporting nice values.
125  */
126 #define	PRI_TIMESHARE_RANGE	(PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
127 #define	PRI_INTERACT_RANGE	((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
128 #define	PRI_BATCH_RANGE		(PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
129 
130 #define	PRI_MIN_INTERACT	PRI_MIN_TIMESHARE
131 #define	PRI_MAX_INTERACT	(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
132 #define	PRI_MIN_BATCH		(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
133 #define	PRI_MAX_BATCH		PRI_MAX_TIMESHARE
134 
135 /*
136  * Cpu percentage computation macros and defines.
137  *
138  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
139  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
140  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
141  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
142  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
143  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
144  */
145 #define	SCHED_TICK_SECS		10
146 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
147 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
148 #define	SCHED_TICK_SHIFT	10
149 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
150 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
151 
152 /*
153  * These macros determine priorities for non-interactive threads.  They are
154  * assigned a priority based on their recent cpu utilization as expressed
155  * by the ratio of ticks to the tick total.  NHALF priorities at the start
156  * and end of the MIN to MAX timeshare range are only reachable with negative
157  * or positive nice respectively.
158  *
159  * PRI_RANGE:	Priority range for utilization dependent priorities.
160  * PRI_NRESV:	Number of nice values.
161  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
162  * PRI_NICE:	Determines the part of the priority inherited from nice.
163  */
164 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
165 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
166 #define	SCHED_PRI_MIN		(PRI_MIN_BATCH + SCHED_PRI_NHALF)
167 #define	SCHED_PRI_MAX		(PRI_MAX_BATCH - SCHED_PRI_NHALF)
168 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
169 #define	SCHED_PRI_TICKS(ts)						\
170     (SCHED_TICK_HZ((ts)) /						\
171     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
172 #define	SCHED_PRI_NICE(nice)	(nice)
173 
174 /*
175  * These determine the interactivity of a process.  Interactivity differs from
176  * cpu utilization in that it expresses the voluntary time slept vs time ran
177  * while cpu utilization includes all time not running.  This more accurately
178  * models the intent of the thread.
179  *
180  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
181  *		before throttling back.
182  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
183  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
184  * INTERACT_THRESH:	Threshold for placement on the current runq.
185  */
186 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
187 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
188 #define	SCHED_INTERACT_MAX	(100)
189 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
190 #define	SCHED_INTERACT_THRESH	(30)
191 
192 /*
193  * These parameters determine the slice behavior for batch work.
194  */
195 #define	SCHED_SLICE_DEFAULT_DIVISOR	10	/* ~94 ms, 12 stathz ticks. */
196 #define	SCHED_SLICE_MIN_DIVISOR		6	/* DEFAULT/MIN = ~16 ms. */
197 
198 /* Flags kept in td_flags. */
199 #define	TDF_SLICEEND	TDF_SCHED2	/* Thread time slice is over. */
200 
201 /*
202  * tickincr:		Converts a stathz tick into a hz domain scaled by
203  *			the shift factor.  Without the shift the error rate
204  *			due to rounding would be unacceptably high.
205  * realstathz:		stathz is sometimes 0 and run off of hz.
206  * sched_slice:		Runtime of each thread before rescheduling.
207  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
208  */
209 static int sched_interact = SCHED_INTERACT_THRESH;
210 static int tickincr = 8 << SCHED_TICK_SHIFT;
211 static int realstathz = 127;	/* reset during boot. */
212 static int sched_slice = 10;	/* reset during boot. */
213 static int sched_slice_min = 1;	/* reset during boot. */
214 #ifdef PREEMPTION
215 #ifdef FULL_PREEMPTION
216 static int preempt_thresh = PRI_MAX_IDLE;
217 #else
218 static int preempt_thresh = PRI_MIN_KERN;
219 #endif
220 #else
221 static int preempt_thresh = 0;
222 #endif
223 static int static_boost = PRI_MIN_BATCH;
224 static int sched_idlespins = 10000;
225 static int sched_idlespinthresh = -1;
226 
227 /*
228  * tdq - per processor runqs and statistics.  All fields are protected by the
229  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
230  * locking in sched_pickcpu();
231  */
232 struct tdq {
233 	/*
234 	 * Ordered to improve efficiency of cpu_search() and switch().
235 	 * tdq_lock is padded to avoid false sharing with tdq_load and
236 	 * tdq_cpu_idle.
237 	 */
238 	struct mtx_padalign tdq_lock;		/* run queue lock. */
239 	struct cpu_group *tdq_cg;		/* Pointer to cpu topology. */
240 	volatile int	tdq_load;		/* Aggregate load. */
241 	volatile int	tdq_cpu_idle;		/* cpu_idle() is active. */
242 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
243 	volatile int	tdq_transferable;	/* Transferable thread count. */
244 	volatile short	tdq_switchcnt;		/* Switches this tick. */
245 	volatile short	tdq_oldswitchcnt;	/* Switches last tick. */
246 	u_char		tdq_lowpri;		/* Lowest priority thread. */
247 	u_char		tdq_ipipending;		/* IPI pending. */
248 	u_char		tdq_idx;		/* Current insert index. */
249 	u_char		tdq_ridx;		/* Current removal index. */
250 	struct runq	tdq_realtime;		/* real-time run queue. */
251 	struct runq	tdq_timeshare;		/* timeshare run queue. */
252 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
253 	char		tdq_name[TDQ_NAME_LEN];
254 #ifdef KTR
255 	char		tdq_loadname[TDQ_LOADNAME_LEN];
256 #endif
257 } __aligned(64);
258 
259 /* Idle thread states and config. */
260 #define	TDQ_RUNNING	1
261 #define	TDQ_IDLE	2
262 
263 #ifdef SMP
264 struct cpu_group *cpu_top;		/* CPU topology */
265 
266 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
267 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
268 
269 /*
270  * Run-time tunables.
271  */
272 static int rebalance = 1;
273 static int balance_interval = 128;	/* Default set in sched_initticks(). */
274 static int affinity;
275 static int steal_idle = 1;
276 static int steal_thresh = 2;
277 static int always_steal = 0;
278 static int trysteal_limit = 2;
279 
280 /*
281  * One thread queue per processor.
282  */
283 static struct tdq	tdq_cpu[MAXCPU];
284 static struct tdq	*balance_tdq;
285 static int balance_ticks;
286 DPCPU_DEFINE_STATIC(uint32_t, randomval);
287 
288 #define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
289 #define	TDQ_CPU(x)	(&tdq_cpu[(x)])
290 #define	TDQ_ID(x)	((int)((x) - tdq_cpu))
291 #else	/* !SMP */
292 static struct tdq	tdq_cpu;
293 
294 #define	TDQ_ID(x)	(0)
295 #define	TDQ_SELF()	(&tdq_cpu)
296 #define	TDQ_CPU(x)	(&tdq_cpu)
297 #endif
298 
299 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
300 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
301 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
302 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
303 #define	TDQ_LOCKPTR(t)		((struct mtx *)(&(t)->tdq_lock))
304 
305 static void sched_priority(struct thread *);
306 static void sched_thread_priority(struct thread *, u_char);
307 static int sched_interact_score(struct thread *);
308 static void sched_interact_update(struct thread *);
309 static void sched_interact_fork(struct thread *);
310 static void sched_pctcpu_update(struct td_sched *, int);
311 
312 /* Operations on per processor queues */
313 static struct thread *tdq_choose(struct tdq *);
314 static void tdq_setup(struct tdq *);
315 static void tdq_load_add(struct tdq *, struct thread *);
316 static void tdq_load_rem(struct tdq *, struct thread *);
317 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
318 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
319 static inline int sched_shouldpreempt(int, int, int);
320 void tdq_print(int cpu);
321 static void runq_print(struct runq *rq);
322 static void tdq_add(struct tdq *, struct thread *, int);
323 #ifdef SMP
324 static struct thread *tdq_move(struct tdq *, struct tdq *);
325 static int tdq_idled(struct tdq *);
326 static void tdq_notify(struct tdq *, struct thread *);
327 static struct thread *tdq_steal(struct tdq *, int);
328 static struct thread *runq_steal(struct runq *, int);
329 static int sched_pickcpu(struct thread *, int);
330 static void sched_balance(void);
331 static int sched_balance_pair(struct tdq *, struct tdq *);
332 static inline struct tdq *sched_setcpu(struct thread *, int, int);
333 static inline void thread_unblock_switch(struct thread *, struct mtx *);
334 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
335 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
336 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
337     struct cpu_group *cg, int indent);
338 #endif
339 
340 static void sched_setup(void *dummy);
341 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
342 
343 static void sched_initticks(void *dummy);
344 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
345     NULL);
346 
347 SDT_PROVIDER_DEFINE(sched);
348 
349 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
350     "struct proc *", "uint8_t");
351 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
352     "struct proc *", "void *");
353 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
354     "struct proc *", "void *", "int");
355 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
356     "struct proc *", "uint8_t", "struct thread *");
357 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
358 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
359     "struct proc *");
360 SDT_PROBE_DEFINE(sched, , , on__cpu);
361 SDT_PROBE_DEFINE(sched, , , remain__cpu);
362 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
363     "struct proc *");
364 
365 /*
366  * Print the threads waiting on a run-queue.
367  */
368 static void
369 runq_print(struct runq *rq)
370 {
371 	struct rqhead *rqh;
372 	struct thread *td;
373 	int pri;
374 	int j;
375 	int i;
376 
377 	for (i = 0; i < RQB_LEN; i++) {
378 		printf("\t\trunq bits %d 0x%zx\n",
379 		    i, rq->rq_status.rqb_bits[i]);
380 		for (j = 0; j < RQB_BPW; j++)
381 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
382 				pri = j + (i << RQB_L2BPW);
383 				rqh = &rq->rq_queues[pri];
384 				TAILQ_FOREACH(td, rqh, td_runq) {
385 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
386 					    td, td->td_name, td->td_priority,
387 					    td->td_rqindex, pri);
388 				}
389 			}
390 	}
391 }
392 
393 /*
394  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
395  */
396 void
397 tdq_print(int cpu)
398 {
399 	struct tdq *tdq;
400 
401 	tdq = TDQ_CPU(cpu);
402 
403 	printf("tdq %d:\n", TDQ_ID(tdq));
404 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
405 	printf("\tLock name:      %s\n", tdq->tdq_name);
406 	printf("\tload:           %d\n", tdq->tdq_load);
407 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
408 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
409 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
410 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
411 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
412 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
413 	printf("\trealtime runq:\n");
414 	runq_print(&tdq->tdq_realtime);
415 	printf("\ttimeshare runq:\n");
416 	runq_print(&tdq->tdq_timeshare);
417 	printf("\tidle runq:\n");
418 	runq_print(&tdq->tdq_idle);
419 }
420 
421 static inline int
422 sched_shouldpreempt(int pri, int cpri, int remote)
423 {
424 	/*
425 	 * If the new priority is not better than the current priority there is
426 	 * nothing to do.
427 	 */
428 	if (pri >= cpri)
429 		return (0);
430 	/*
431 	 * Always preempt idle.
432 	 */
433 	if (cpri >= PRI_MIN_IDLE)
434 		return (1);
435 	/*
436 	 * If preemption is disabled don't preempt others.
437 	 */
438 	if (preempt_thresh == 0)
439 		return (0);
440 	/*
441 	 * Preempt if we exceed the threshold.
442 	 */
443 	if (pri <= preempt_thresh)
444 		return (1);
445 	/*
446 	 * If we're interactive or better and there is non-interactive
447 	 * or worse running preempt only remote processors.
448 	 */
449 	if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
450 		return (1);
451 	return (0);
452 }
453 
454 /*
455  * Add a thread to the actual run-queue.  Keeps transferable counts up to
456  * date with what is actually on the run-queue.  Selects the correct
457  * queue position for timeshare threads.
458  */
459 static __inline void
460 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
461 {
462 	struct td_sched *ts;
463 	u_char pri;
464 
465 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
466 	THREAD_LOCK_ASSERT(td, MA_OWNED);
467 
468 	pri = td->td_priority;
469 	ts = td_get_sched(td);
470 	TD_SET_RUNQ(td);
471 	if (THREAD_CAN_MIGRATE(td)) {
472 		tdq->tdq_transferable++;
473 		ts->ts_flags |= TSF_XFERABLE;
474 	}
475 	if (pri < PRI_MIN_BATCH) {
476 		ts->ts_runq = &tdq->tdq_realtime;
477 	} else if (pri <= PRI_MAX_BATCH) {
478 		ts->ts_runq = &tdq->tdq_timeshare;
479 		KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
480 			("Invalid priority %d on timeshare runq", pri));
481 		/*
482 		 * This queue contains only priorities between MIN and MAX
483 		 * realtime.  Use the whole queue to represent these values.
484 		 */
485 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
486 			pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
487 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
488 			/*
489 			 * This effectively shortens the queue by one so we
490 			 * can have a one slot difference between idx and
491 			 * ridx while we wait for threads to drain.
492 			 */
493 			if (tdq->tdq_ridx != tdq->tdq_idx &&
494 			    pri == tdq->tdq_ridx)
495 				pri = (unsigned char)(pri - 1) % RQ_NQS;
496 		} else
497 			pri = tdq->tdq_ridx;
498 		runq_add_pri(ts->ts_runq, td, pri, flags);
499 		return;
500 	} else
501 		ts->ts_runq = &tdq->tdq_idle;
502 	runq_add(ts->ts_runq, td, flags);
503 }
504 
505 /*
506  * Remove a thread from a run-queue.  This typically happens when a thread
507  * is selected to run.  Running threads are not on the queue and the
508  * transferable count does not reflect them.
509  */
510 static __inline void
511 tdq_runq_rem(struct tdq *tdq, struct thread *td)
512 {
513 	struct td_sched *ts;
514 
515 	ts = td_get_sched(td);
516 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
517 	KASSERT(ts->ts_runq != NULL,
518 	    ("tdq_runq_remove: thread %p null ts_runq", td));
519 	if (ts->ts_flags & TSF_XFERABLE) {
520 		tdq->tdq_transferable--;
521 		ts->ts_flags &= ~TSF_XFERABLE;
522 	}
523 	if (ts->ts_runq == &tdq->tdq_timeshare) {
524 		if (tdq->tdq_idx != tdq->tdq_ridx)
525 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
526 		else
527 			runq_remove_idx(ts->ts_runq, td, NULL);
528 	} else
529 		runq_remove(ts->ts_runq, td);
530 }
531 
532 /*
533  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
534  * for this thread to the referenced thread queue.
535  */
536 static void
537 tdq_load_add(struct tdq *tdq, struct thread *td)
538 {
539 
540 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
541 	THREAD_LOCK_ASSERT(td, MA_OWNED);
542 
543 	tdq->tdq_load++;
544 	if ((td->td_flags & TDF_NOLOAD) == 0)
545 		tdq->tdq_sysload++;
546 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
547 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
548 }
549 
550 /*
551  * Remove the load from a thread that is transitioning to a sleep state or
552  * exiting.
553  */
554 static void
555 tdq_load_rem(struct tdq *tdq, struct thread *td)
556 {
557 
558 	THREAD_LOCK_ASSERT(td, MA_OWNED);
559 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
560 	KASSERT(tdq->tdq_load != 0,
561 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
562 
563 	tdq->tdq_load--;
564 	if ((td->td_flags & TDF_NOLOAD) == 0)
565 		tdq->tdq_sysload--;
566 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
567 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
568 }
569 
570 /*
571  * Bound timeshare latency by decreasing slice size as load increases.  We
572  * consider the maximum latency as the sum of the threads waiting to run
573  * aside from curthread and target no more than sched_slice latency but
574  * no less than sched_slice_min runtime.
575  */
576 static inline int
577 tdq_slice(struct tdq *tdq)
578 {
579 	int load;
580 
581 	/*
582 	 * It is safe to use sys_load here because this is called from
583 	 * contexts where timeshare threads are running and so there
584 	 * cannot be higher priority load in the system.
585 	 */
586 	load = tdq->tdq_sysload - 1;
587 	if (load >= SCHED_SLICE_MIN_DIVISOR)
588 		return (sched_slice_min);
589 	if (load <= 1)
590 		return (sched_slice);
591 	return (sched_slice / load);
592 }
593 
594 /*
595  * Set lowpri to its exact value by searching the run-queue and
596  * evaluating curthread.  curthread may be passed as an optimization.
597  */
598 static void
599 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
600 {
601 	struct thread *td;
602 
603 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
604 	if (ctd == NULL)
605 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
606 	td = tdq_choose(tdq);
607 	if (td == NULL || td->td_priority > ctd->td_priority)
608 		tdq->tdq_lowpri = ctd->td_priority;
609 	else
610 		tdq->tdq_lowpri = td->td_priority;
611 }
612 
613 #ifdef SMP
614 /*
615  * We need some randomness. Implement a classic Linear Congruential
616  * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
617  * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
618  * of the random state (in the low bits of our answer) to keep
619  * the maximum randomness.
620  */
621 static uint32_t
622 sched_random(void)
623 {
624 	uint32_t *rndptr;
625 
626 	rndptr = DPCPU_PTR(randomval);
627 	*rndptr = *rndptr * 69069 + 5;
628 
629 	return (*rndptr >> 16);
630 }
631 
632 struct cpu_search {
633 	cpuset_t cs_mask;
634 	u_int	cs_prefer;
635 	int	cs_pri;		/* Min priority for low. */
636 	int	cs_limit;	/* Max load for low, min load for high. */
637 	int	cs_cpu;
638 	int	cs_load;
639 };
640 
641 #define	CPU_SEARCH_LOWEST	0x1
642 #define	CPU_SEARCH_HIGHEST	0x2
643 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
644 
645 #define	CPUSET_FOREACH(cpu, mask)				\
646 	for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++)		\
647 		if (CPU_ISSET(cpu, &mask))
648 
649 static __always_inline int cpu_search(const struct cpu_group *cg,
650     struct cpu_search *low, struct cpu_search *high, const int match);
651 int __noinline cpu_search_lowest(const struct cpu_group *cg,
652     struct cpu_search *low);
653 int __noinline cpu_search_highest(const struct cpu_group *cg,
654     struct cpu_search *high);
655 int __noinline cpu_search_both(const struct cpu_group *cg,
656     struct cpu_search *low, struct cpu_search *high);
657 
658 /*
659  * Search the tree of cpu_groups for the lowest or highest loaded cpu
660  * according to the match argument.  This routine actually compares the
661  * load on all paths through the tree and finds the least loaded cpu on
662  * the least loaded path, which may differ from the least loaded cpu in
663  * the system.  This balances work among caches and buses.
664  *
665  * This inline is instantiated in three forms below using constants for the
666  * match argument.  It is reduced to the minimum set for each case.  It is
667  * also recursive to the depth of the tree.
668  */
669 static __always_inline int
670 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
671     struct cpu_search *high, const int match)
672 {
673 	struct cpu_search lgroup;
674 	struct cpu_search hgroup;
675 	cpuset_t cpumask;
676 	struct cpu_group *child;
677 	struct tdq *tdq;
678 	int cpu, i, hload, lload, load, total, rnd;
679 
680 	total = 0;
681 	cpumask = cg->cg_mask;
682 	if (match & CPU_SEARCH_LOWEST) {
683 		lload = INT_MAX;
684 		lgroup = *low;
685 	}
686 	if (match & CPU_SEARCH_HIGHEST) {
687 		hload = INT_MIN;
688 		hgroup = *high;
689 	}
690 
691 	/* Iterate through the child CPU groups and then remaining CPUs. */
692 	for (i = cg->cg_children, cpu = mp_maxid; ; ) {
693 		if (i == 0) {
694 #ifdef HAVE_INLINE_FFSL
695 			cpu = CPU_FFS(&cpumask) - 1;
696 #else
697 			while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
698 				cpu--;
699 #endif
700 			if (cpu < 0)
701 				break;
702 			child = NULL;
703 		} else
704 			child = &cg->cg_child[i - 1];
705 
706 		if (match & CPU_SEARCH_LOWEST)
707 			lgroup.cs_cpu = -1;
708 		if (match & CPU_SEARCH_HIGHEST)
709 			hgroup.cs_cpu = -1;
710 		if (child) {			/* Handle child CPU group. */
711 			CPU_NAND(&cpumask, &child->cg_mask);
712 			switch (match) {
713 			case CPU_SEARCH_LOWEST:
714 				load = cpu_search_lowest(child, &lgroup);
715 				break;
716 			case CPU_SEARCH_HIGHEST:
717 				load = cpu_search_highest(child, &hgroup);
718 				break;
719 			case CPU_SEARCH_BOTH:
720 				load = cpu_search_both(child, &lgroup, &hgroup);
721 				break;
722 			}
723 		} else {			/* Handle child CPU. */
724 			CPU_CLR(cpu, &cpumask);
725 			tdq = TDQ_CPU(cpu);
726 			load = tdq->tdq_load * 256;
727 			rnd = sched_random() % 32;
728 			if (match & CPU_SEARCH_LOWEST) {
729 				if (cpu == low->cs_prefer)
730 					load -= 64;
731 				/* If that CPU is allowed and get data. */
732 				if (tdq->tdq_lowpri > lgroup.cs_pri &&
733 				    tdq->tdq_load <= lgroup.cs_limit &&
734 				    CPU_ISSET(cpu, &lgroup.cs_mask)) {
735 					lgroup.cs_cpu = cpu;
736 					lgroup.cs_load = load - rnd;
737 				}
738 			}
739 			if (match & CPU_SEARCH_HIGHEST)
740 				if (tdq->tdq_load >= hgroup.cs_limit &&
741 				    tdq->tdq_transferable &&
742 				    CPU_ISSET(cpu, &hgroup.cs_mask)) {
743 					hgroup.cs_cpu = cpu;
744 					hgroup.cs_load = load - rnd;
745 				}
746 		}
747 		total += load;
748 
749 		/* We have info about child item. Compare it. */
750 		if (match & CPU_SEARCH_LOWEST) {
751 			if (lgroup.cs_cpu >= 0 &&
752 			    (load < lload ||
753 			     (load == lload && lgroup.cs_load < low->cs_load))) {
754 				lload = load;
755 				low->cs_cpu = lgroup.cs_cpu;
756 				low->cs_load = lgroup.cs_load;
757 			}
758 		}
759 		if (match & CPU_SEARCH_HIGHEST)
760 			if (hgroup.cs_cpu >= 0 &&
761 			    (load > hload ||
762 			     (load == hload && hgroup.cs_load > high->cs_load))) {
763 				hload = load;
764 				high->cs_cpu = hgroup.cs_cpu;
765 				high->cs_load = hgroup.cs_load;
766 			}
767 		if (child) {
768 			i--;
769 			if (i == 0 && CPU_EMPTY(&cpumask))
770 				break;
771 		}
772 #ifndef HAVE_INLINE_FFSL
773 		else
774 			cpu--;
775 #endif
776 	}
777 	return (total);
778 }
779 
780 /*
781  * cpu_search instantiations must pass constants to maintain the inline
782  * optimization.
783  */
784 int
785 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
786 {
787 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
788 }
789 
790 int
791 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
792 {
793 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
794 }
795 
796 int
797 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
798     struct cpu_search *high)
799 {
800 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
801 }
802 
803 /*
804  * Find the cpu with the least load via the least loaded path that has a
805  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
806  * acceptable.
807  */
808 static inline int
809 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
810     int prefer)
811 {
812 	struct cpu_search low;
813 
814 	low.cs_cpu = -1;
815 	low.cs_prefer = prefer;
816 	low.cs_mask = mask;
817 	low.cs_pri = pri;
818 	low.cs_limit = maxload;
819 	cpu_search_lowest(cg, &low);
820 	return low.cs_cpu;
821 }
822 
823 /*
824  * Find the cpu with the highest load via the highest loaded path.
825  */
826 static inline int
827 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
828 {
829 	struct cpu_search high;
830 
831 	high.cs_cpu = -1;
832 	high.cs_mask = mask;
833 	high.cs_limit = minload;
834 	cpu_search_highest(cg, &high);
835 	return high.cs_cpu;
836 }
837 
838 static void
839 sched_balance_group(struct cpu_group *cg)
840 {
841 	cpuset_t hmask, lmask;
842 	int high, low, anylow;
843 
844 	CPU_FILL(&hmask);
845 	for (;;) {
846 		high = sched_highest(cg, hmask, 2);
847 		/* Stop if there is no more CPU with transferrable threads. */
848 		if (high == -1)
849 			break;
850 		CPU_CLR(high, &hmask);
851 		CPU_COPY(&hmask, &lmask);
852 		/* Stop if there is no more CPU left for low. */
853 		if (CPU_EMPTY(&lmask))
854 			break;
855 		anylow = 1;
856 nextlow:
857 		low = sched_lowest(cg, lmask, -1,
858 		    TDQ_CPU(high)->tdq_load - 1, high);
859 		/* Stop if we looked well and found no less loaded CPU. */
860 		if (anylow && low == -1)
861 			break;
862 		/* Go to next high if we found no less loaded CPU. */
863 		if (low == -1)
864 			continue;
865 		/* Transfer thread from high to low. */
866 		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
867 			/* CPU that got thread can no longer be a donor. */
868 			CPU_CLR(low, &hmask);
869 		} else {
870 			/*
871 			 * If failed, then there is no threads on high
872 			 * that can run on this low. Drop low from low
873 			 * mask and look for different one.
874 			 */
875 			CPU_CLR(low, &lmask);
876 			anylow = 0;
877 			goto nextlow;
878 		}
879 	}
880 }
881 
882 static void
883 sched_balance(void)
884 {
885 	struct tdq *tdq;
886 
887 	balance_ticks = max(balance_interval / 2, 1) +
888 	    (sched_random() % balance_interval);
889 	tdq = TDQ_SELF();
890 	TDQ_UNLOCK(tdq);
891 	sched_balance_group(cpu_top);
892 	TDQ_LOCK(tdq);
893 }
894 
895 /*
896  * Lock two thread queues using their address to maintain lock order.
897  */
898 static void
899 tdq_lock_pair(struct tdq *one, struct tdq *two)
900 {
901 	if (one < two) {
902 		TDQ_LOCK(one);
903 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
904 	} else {
905 		TDQ_LOCK(two);
906 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
907 	}
908 }
909 
910 /*
911  * Unlock two thread queues.  Order is not important here.
912  */
913 static void
914 tdq_unlock_pair(struct tdq *one, struct tdq *two)
915 {
916 	TDQ_UNLOCK(one);
917 	TDQ_UNLOCK(two);
918 }
919 
920 /*
921  * Transfer load between two imbalanced thread queues.
922  */
923 static int
924 sched_balance_pair(struct tdq *high, struct tdq *low)
925 {
926 	struct thread *td;
927 	int cpu;
928 
929 	tdq_lock_pair(high, low);
930 	td = NULL;
931 	/*
932 	 * Transfer a thread from high to low.
933 	 */
934 	if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
935 	    (td = tdq_move(high, low)) != NULL) {
936 		/*
937 		 * In case the target isn't the current cpu notify it of the
938 		 * new load, possibly sending an IPI to force it to reschedule.
939 		 */
940 		cpu = TDQ_ID(low);
941 		if (cpu != PCPU_GET(cpuid))
942 			tdq_notify(low, td);
943 	}
944 	tdq_unlock_pair(high, low);
945 	return (td != NULL);
946 }
947 
948 /*
949  * Move a thread from one thread queue to another.
950  */
951 static struct thread *
952 tdq_move(struct tdq *from, struct tdq *to)
953 {
954 	struct td_sched *ts;
955 	struct thread *td;
956 	struct tdq *tdq;
957 	int cpu;
958 
959 	TDQ_LOCK_ASSERT(from, MA_OWNED);
960 	TDQ_LOCK_ASSERT(to, MA_OWNED);
961 
962 	tdq = from;
963 	cpu = TDQ_ID(to);
964 	td = tdq_steal(tdq, cpu);
965 	if (td == NULL)
966 		return (NULL);
967 	ts = td_get_sched(td);
968 	/*
969 	 * Although the run queue is locked the thread may be blocked.  Lock
970 	 * it to clear this and acquire the run-queue lock.
971 	 */
972 	thread_lock(td);
973 	/* Drop recursive lock on from acquired via thread_lock(). */
974 	TDQ_UNLOCK(from);
975 	sched_rem(td);
976 	ts->ts_cpu = cpu;
977 	td->td_lock = TDQ_LOCKPTR(to);
978 	tdq_add(to, td, SRQ_YIELDING);
979 	return (td);
980 }
981 
982 /*
983  * This tdq has idled.  Try to steal a thread from another cpu and switch
984  * to it.
985  */
986 static int
987 tdq_idled(struct tdq *tdq)
988 {
989 	struct cpu_group *cg;
990 	struct tdq *steal;
991 	cpuset_t mask;
992 	int cpu, switchcnt;
993 
994 	if (smp_started == 0 || steal_idle == 0 || tdq->tdq_cg == NULL)
995 		return (1);
996 	CPU_FILL(&mask);
997 	CPU_CLR(PCPU_GET(cpuid), &mask);
998     restart:
999 	switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
1000 	for (cg = tdq->tdq_cg; ; ) {
1001 		cpu = sched_highest(cg, mask, steal_thresh);
1002 		/*
1003 		 * We were assigned a thread but not preempted.  Returning
1004 		 * 0 here will cause our caller to switch to it.
1005 		 */
1006 		if (tdq->tdq_load)
1007 			return (0);
1008 		if (cpu == -1) {
1009 			cg = cg->cg_parent;
1010 			if (cg == NULL)
1011 				return (1);
1012 			continue;
1013 		}
1014 		steal = TDQ_CPU(cpu);
1015 		/*
1016 		 * The data returned by sched_highest() is stale and
1017 		 * the chosen CPU no longer has an eligible thread.
1018 		 *
1019 		 * Testing this ahead of tdq_lock_pair() only catches
1020 		 * this situation about 20% of the time on an 8 core
1021 		 * 16 thread Ryzen 7, but it still helps performance.
1022 		 */
1023 		if (steal->tdq_load < steal_thresh ||
1024 		    steal->tdq_transferable == 0)
1025 			goto restart;
1026 		tdq_lock_pair(tdq, steal);
1027 		/*
1028 		 * We were assigned a thread while waiting for the locks.
1029 		 * Switch to it now instead of stealing a thread.
1030 		 */
1031 		if (tdq->tdq_load)
1032 			break;
1033 		/*
1034 		 * The data returned by sched_highest() is stale and
1035 		 * the chosen CPU no longer has an eligible thread, or
1036 		 * we were preempted and the CPU loading info may be out
1037 		 * of date.  The latter is rare.  In either case restart
1038 		 * the search.
1039 		 */
1040 		if (steal->tdq_load < steal_thresh ||
1041 		    steal->tdq_transferable == 0 ||
1042 		    switchcnt != tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt) {
1043 			tdq_unlock_pair(tdq, steal);
1044 			goto restart;
1045 		}
1046 		/*
1047 		 * Steal the thread and switch to it.
1048 		 */
1049 		if (tdq_move(steal, tdq) != NULL)
1050 			break;
1051 		/*
1052 		 * We failed to acquire a thread even though it looked
1053 		 * like one was available.  This could be due to affinity
1054 		 * restrictions or for other reasons.  Loop again after
1055 		 * removing this CPU from the set.  The restart logic
1056 		 * above does not restore this CPU to the set due to the
1057 		 * likelyhood of failing here again.
1058 		 */
1059 		CPU_CLR(cpu, &mask);
1060 		tdq_unlock_pair(tdq, steal);
1061 	}
1062 	TDQ_UNLOCK(steal);
1063 	mi_switch(SW_VOL | SWT_IDLE, NULL);
1064 	thread_unlock(curthread);
1065 	return (0);
1066 }
1067 
1068 /*
1069  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
1070  */
1071 static void
1072 tdq_notify(struct tdq *tdq, struct thread *td)
1073 {
1074 	struct thread *ctd;
1075 	int pri;
1076 	int cpu;
1077 
1078 	if (tdq->tdq_ipipending)
1079 		return;
1080 	cpu = td_get_sched(td)->ts_cpu;
1081 	pri = td->td_priority;
1082 	ctd = pcpu_find(cpu)->pc_curthread;
1083 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1084 		return;
1085 
1086 	/*
1087 	 * Make sure that our caller's earlier update to tdq_load is
1088 	 * globally visible before we read tdq_cpu_idle.  Idle thread
1089 	 * accesses both of them without locks, and the order is important.
1090 	 */
1091 	atomic_thread_fence_seq_cst();
1092 
1093 	if (TD_IS_IDLETHREAD(ctd)) {
1094 		/*
1095 		 * If the MD code has an idle wakeup routine try that before
1096 		 * falling back to IPI.
1097 		 */
1098 		if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1099 			return;
1100 	}
1101 	tdq->tdq_ipipending = 1;
1102 	ipi_cpu(cpu, IPI_PREEMPT);
1103 }
1104 
1105 /*
1106  * Steals load from a timeshare queue.  Honors the rotating queue head
1107  * index.
1108  */
1109 static struct thread *
1110 runq_steal_from(struct runq *rq, int cpu, u_char start)
1111 {
1112 	struct rqbits *rqb;
1113 	struct rqhead *rqh;
1114 	struct thread *td, *first;
1115 	int bit;
1116 	int i;
1117 
1118 	rqb = &rq->rq_status;
1119 	bit = start & (RQB_BPW -1);
1120 	first = NULL;
1121 again:
1122 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1123 		if (rqb->rqb_bits[i] == 0)
1124 			continue;
1125 		if (bit == 0)
1126 			bit = RQB_FFS(rqb->rqb_bits[i]);
1127 		for (; bit < RQB_BPW; bit++) {
1128 			if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1129 				continue;
1130 			rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1131 			TAILQ_FOREACH(td, rqh, td_runq) {
1132 				if (first && THREAD_CAN_MIGRATE(td) &&
1133 				    THREAD_CAN_SCHED(td, cpu))
1134 					return (td);
1135 				first = td;
1136 			}
1137 		}
1138 	}
1139 	if (start != 0) {
1140 		start = 0;
1141 		goto again;
1142 	}
1143 
1144 	if (first && THREAD_CAN_MIGRATE(first) &&
1145 	    THREAD_CAN_SCHED(first, cpu))
1146 		return (first);
1147 	return (NULL);
1148 }
1149 
1150 /*
1151  * Steals load from a standard linear queue.
1152  */
1153 static struct thread *
1154 runq_steal(struct runq *rq, int cpu)
1155 {
1156 	struct rqhead *rqh;
1157 	struct rqbits *rqb;
1158 	struct thread *td;
1159 	int word;
1160 	int bit;
1161 
1162 	rqb = &rq->rq_status;
1163 	for (word = 0; word < RQB_LEN; word++) {
1164 		if (rqb->rqb_bits[word] == 0)
1165 			continue;
1166 		for (bit = 0; bit < RQB_BPW; bit++) {
1167 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1168 				continue;
1169 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1170 			TAILQ_FOREACH(td, rqh, td_runq)
1171 				if (THREAD_CAN_MIGRATE(td) &&
1172 				    THREAD_CAN_SCHED(td, cpu))
1173 					return (td);
1174 		}
1175 	}
1176 	return (NULL);
1177 }
1178 
1179 /*
1180  * Attempt to steal a thread in priority order from a thread queue.
1181  */
1182 static struct thread *
1183 tdq_steal(struct tdq *tdq, int cpu)
1184 {
1185 	struct thread *td;
1186 
1187 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1188 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1189 		return (td);
1190 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1191 	    cpu, tdq->tdq_ridx)) != NULL)
1192 		return (td);
1193 	return (runq_steal(&tdq->tdq_idle, cpu));
1194 }
1195 
1196 /*
1197  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1198  * current lock and returns with the assigned queue locked.
1199  */
1200 static inline struct tdq *
1201 sched_setcpu(struct thread *td, int cpu, int flags)
1202 {
1203 
1204 	struct tdq *tdq;
1205 
1206 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1207 	tdq = TDQ_CPU(cpu);
1208 	td_get_sched(td)->ts_cpu = cpu;
1209 	/*
1210 	 * If the lock matches just return the queue.
1211 	 */
1212 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1213 		return (tdq);
1214 #ifdef notyet
1215 	/*
1216 	 * If the thread isn't running its lockptr is a
1217 	 * turnstile or a sleepqueue.  We can just lock_set without
1218 	 * blocking.
1219 	 */
1220 	if (TD_CAN_RUN(td)) {
1221 		TDQ_LOCK(tdq);
1222 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1223 		return (tdq);
1224 	}
1225 #endif
1226 	/*
1227 	 * The hard case, migration, we need to block the thread first to
1228 	 * prevent order reversals with other cpus locks.
1229 	 */
1230 	spinlock_enter();
1231 	thread_lock_block(td);
1232 	TDQ_LOCK(tdq);
1233 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1234 	spinlock_exit();
1235 	return (tdq);
1236 }
1237 
1238 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1239 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1240 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1241 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1242 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1243 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1244 
1245 static int
1246 sched_pickcpu(struct thread *td, int flags)
1247 {
1248 	struct cpu_group *cg, *ccg;
1249 	struct td_sched *ts;
1250 	struct tdq *tdq;
1251 	cpuset_t mask;
1252 	int cpu, pri, self;
1253 
1254 	self = PCPU_GET(cpuid);
1255 	ts = td_get_sched(td);
1256 	KASSERT(!CPU_ABSENT(ts->ts_cpu), ("sched_pickcpu: Start scheduler on "
1257 	    "absent CPU %d for thread %s.", ts->ts_cpu, td->td_name));
1258 	if (smp_started == 0)
1259 		return (self);
1260 	/*
1261 	 * Don't migrate a running thread from sched_switch().
1262 	 */
1263 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1264 		return (ts->ts_cpu);
1265 	/*
1266 	 * Prefer to run interrupt threads on the processors that generate
1267 	 * the interrupt.
1268 	 */
1269 	pri = td->td_priority;
1270 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1271 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1272 		SCHED_STAT_INC(pickcpu_intrbind);
1273 		ts->ts_cpu = self;
1274 		if (TDQ_CPU(self)->tdq_lowpri > pri) {
1275 			SCHED_STAT_INC(pickcpu_affinity);
1276 			return (ts->ts_cpu);
1277 		}
1278 	}
1279 	/*
1280 	 * If the thread can run on the last cpu and the affinity has not
1281 	 * expired and it is idle, run it there.
1282 	 */
1283 	tdq = TDQ_CPU(ts->ts_cpu);
1284 	cg = tdq->tdq_cg;
1285 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1286 	    tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1287 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1288 		if (cg->cg_flags & CG_FLAG_THREAD) {
1289 			CPUSET_FOREACH(cpu, cg->cg_mask) {
1290 				if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1291 					break;
1292 			}
1293 		} else
1294 			cpu = INT_MAX;
1295 		if (cpu > mp_maxid) {
1296 			SCHED_STAT_INC(pickcpu_idle_affinity);
1297 			return (ts->ts_cpu);
1298 		}
1299 	}
1300 	/*
1301 	 * Search for the last level cache CPU group in the tree.
1302 	 * Skip caches with expired affinity time and SMT groups.
1303 	 * Affinity to higher level caches will be handled less aggressively.
1304 	 */
1305 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1306 		if (cg->cg_flags & CG_FLAG_THREAD)
1307 			continue;
1308 		if (!SCHED_AFFINITY(ts, cg->cg_level))
1309 			continue;
1310 		ccg = cg;
1311 	}
1312 	if (ccg != NULL)
1313 		cg = ccg;
1314 	cpu = -1;
1315 	/* Search the group for the less loaded idle CPU we can run now. */
1316 	mask = td->td_cpuset->cs_mask;
1317 	if (cg != NULL && cg != cpu_top &&
1318 	    CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1319 		cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1320 		    INT_MAX, ts->ts_cpu);
1321 	/* Search globally for the less loaded CPU we can run now. */
1322 	if (cpu == -1)
1323 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1324 	/* Search globally for the less loaded CPU. */
1325 	if (cpu == -1)
1326 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1327 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1328 	KASSERT(!CPU_ABSENT(cpu), ("sched_pickcpu: Picked absent CPU %d.", cpu));
1329 	/*
1330 	 * Compare the lowest loaded cpu to current cpu.
1331 	 */
1332 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1333 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1334 	    TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1335 		SCHED_STAT_INC(pickcpu_local);
1336 		cpu = self;
1337 	} else
1338 		SCHED_STAT_INC(pickcpu_lowest);
1339 	if (cpu != ts->ts_cpu)
1340 		SCHED_STAT_INC(pickcpu_migration);
1341 	return (cpu);
1342 }
1343 #endif
1344 
1345 /*
1346  * Pick the highest priority task we have and return it.
1347  */
1348 static struct thread *
1349 tdq_choose(struct tdq *tdq)
1350 {
1351 	struct thread *td;
1352 
1353 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1354 	td = runq_choose(&tdq->tdq_realtime);
1355 	if (td != NULL)
1356 		return (td);
1357 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1358 	if (td != NULL) {
1359 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1360 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1361 		    td->td_priority));
1362 		return (td);
1363 	}
1364 	td = runq_choose(&tdq->tdq_idle);
1365 	if (td != NULL) {
1366 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1367 		    ("tdq_choose: Invalid priority on idle queue %d",
1368 		    td->td_priority));
1369 		return (td);
1370 	}
1371 
1372 	return (NULL);
1373 }
1374 
1375 /*
1376  * Initialize a thread queue.
1377  */
1378 static void
1379 tdq_setup(struct tdq *tdq)
1380 {
1381 
1382 	if (bootverbose)
1383 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1384 	runq_init(&tdq->tdq_realtime);
1385 	runq_init(&tdq->tdq_timeshare);
1386 	runq_init(&tdq->tdq_idle);
1387 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1388 	    "sched lock %d", (int)TDQ_ID(tdq));
1389 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1390 	    MTX_SPIN | MTX_RECURSE);
1391 #ifdef KTR
1392 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1393 	    "CPU %d load", (int)TDQ_ID(tdq));
1394 #endif
1395 }
1396 
1397 #ifdef SMP
1398 static void
1399 sched_setup_smp(void)
1400 {
1401 	struct tdq *tdq;
1402 	int i;
1403 
1404 	cpu_top = smp_topo();
1405 	CPU_FOREACH(i) {
1406 		tdq = TDQ_CPU(i);
1407 		tdq_setup(tdq);
1408 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1409 		if (tdq->tdq_cg == NULL)
1410 			panic("Can't find cpu group for %d\n", i);
1411 	}
1412 	balance_tdq = TDQ_SELF();
1413 }
1414 #endif
1415 
1416 /*
1417  * Setup the thread queues and initialize the topology based on MD
1418  * information.
1419  */
1420 static void
1421 sched_setup(void *dummy)
1422 {
1423 	struct tdq *tdq;
1424 
1425 	tdq = TDQ_SELF();
1426 #ifdef SMP
1427 	sched_setup_smp();
1428 #else
1429 	tdq_setup(tdq);
1430 #endif
1431 
1432 	/* Add thread0's load since it's running. */
1433 	TDQ_LOCK(tdq);
1434 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1435 	tdq_load_add(tdq, &thread0);
1436 	tdq->tdq_lowpri = thread0.td_priority;
1437 	TDQ_UNLOCK(tdq);
1438 }
1439 
1440 /*
1441  * This routine determines time constants after stathz and hz are setup.
1442  */
1443 /* ARGSUSED */
1444 static void
1445 sched_initticks(void *dummy)
1446 {
1447 	int incr;
1448 
1449 	realstathz = stathz ? stathz : hz;
1450 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1451 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1452 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1453 	    realstathz);
1454 
1455 	/*
1456 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1457 	 * hz not being evenly divisible by stathz on all platforms.
1458 	 */
1459 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1460 	/*
1461 	 * This does not work for values of stathz that are more than
1462 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1463 	 */
1464 	if (incr == 0)
1465 		incr = 1;
1466 	tickincr = incr;
1467 #ifdef SMP
1468 	/*
1469 	 * Set the default balance interval now that we know
1470 	 * what realstathz is.
1471 	 */
1472 	balance_interval = realstathz;
1473 	balance_ticks = balance_interval;
1474 	affinity = SCHED_AFFINITY_DEFAULT;
1475 #endif
1476 	if (sched_idlespinthresh < 0)
1477 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1478 }
1479 
1480 
1481 /*
1482  * This is the core of the interactivity algorithm.  Determines a score based
1483  * on past behavior.  It is the ratio of sleep time to run time scaled to
1484  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1485  * differs from the cpu usage because it does not account for time spent
1486  * waiting on a run-queue.  Would be prettier if we had floating point.
1487  *
1488  * When a thread's sleep time is greater than its run time the
1489  * calculation is:
1490  *
1491  *                           scaling factor
1492  * interactivity score =  ---------------------
1493  *                        sleep time / run time
1494  *
1495  *
1496  * When a thread's run time is greater than its sleep time the
1497  * calculation is:
1498  *
1499  *                           scaling factor
1500  * interactivity score =  ---------------------    + scaling factor
1501  *                        run time / sleep time
1502  */
1503 static int
1504 sched_interact_score(struct thread *td)
1505 {
1506 	struct td_sched *ts;
1507 	int div;
1508 
1509 	ts = td_get_sched(td);
1510 	/*
1511 	 * The score is only needed if this is likely to be an interactive
1512 	 * task.  Don't go through the expense of computing it if there's
1513 	 * no chance.
1514 	 */
1515 	if (sched_interact <= SCHED_INTERACT_HALF &&
1516 		ts->ts_runtime >= ts->ts_slptime)
1517 			return (SCHED_INTERACT_HALF);
1518 
1519 	if (ts->ts_runtime > ts->ts_slptime) {
1520 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1521 		return (SCHED_INTERACT_HALF +
1522 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1523 	}
1524 	if (ts->ts_slptime > ts->ts_runtime) {
1525 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1526 		return (ts->ts_runtime / div);
1527 	}
1528 	/* runtime == slptime */
1529 	if (ts->ts_runtime)
1530 		return (SCHED_INTERACT_HALF);
1531 
1532 	/*
1533 	 * This can happen if slptime and runtime are 0.
1534 	 */
1535 	return (0);
1536 
1537 }
1538 
1539 /*
1540  * Scale the scheduling priority according to the "interactivity" of this
1541  * process.
1542  */
1543 static void
1544 sched_priority(struct thread *td)
1545 {
1546 	int score;
1547 	int pri;
1548 
1549 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1550 		return;
1551 	/*
1552 	 * If the score is interactive we place the thread in the realtime
1553 	 * queue with a priority that is less than kernel and interrupt
1554 	 * priorities.  These threads are not subject to nice restrictions.
1555 	 *
1556 	 * Scores greater than this are placed on the normal timeshare queue
1557 	 * where the priority is partially decided by the most recent cpu
1558 	 * utilization and the rest is decided by nice value.
1559 	 *
1560 	 * The nice value of the process has a linear effect on the calculated
1561 	 * score.  Negative nice values make it easier for a thread to be
1562 	 * considered interactive.
1563 	 */
1564 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1565 	if (score < sched_interact) {
1566 		pri = PRI_MIN_INTERACT;
1567 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1568 		    sched_interact) * score;
1569 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1570 		    ("sched_priority: invalid interactive priority %d score %d",
1571 		    pri, score));
1572 	} else {
1573 		pri = SCHED_PRI_MIN;
1574 		if (td_get_sched(td)->ts_ticks)
1575 			pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1576 			    SCHED_PRI_RANGE - 1);
1577 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1578 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1579 		    ("sched_priority: invalid priority %d: nice %d, "
1580 		    "ticks %d ftick %d ltick %d tick pri %d",
1581 		    pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1582 		    td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1583 		    SCHED_PRI_TICKS(td_get_sched(td))));
1584 	}
1585 	sched_user_prio(td, pri);
1586 
1587 	return;
1588 }
1589 
1590 /*
1591  * This routine enforces a maximum limit on the amount of scheduling history
1592  * kept.  It is called after either the slptime or runtime is adjusted.  This
1593  * function is ugly due to integer math.
1594  */
1595 static void
1596 sched_interact_update(struct thread *td)
1597 {
1598 	struct td_sched *ts;
1599 	u_int sum;
1600 
1601 	ts = td_get_sched(td);
1602 	sum = ts->ts_runtime + ts->ts_slptime;
1603 	if (sum < SCHED_SLP_RUN_MAX)
1604 		return;
1605 	/*
1606 	 * This only happens from two places:
1607 	 * 1) We have added an unusual amount of run time from fork_exit.
1608 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1609 	 */
1610 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1611 		if (ts->ts_runtime > ts->ts_slptime) {
1612 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1613 			ts->ts_slptime = 1;
1614 		} else {
1615 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1616 			ts->ts_runtime = 1;
1617 		}
1618 		return;
1619 	}
1620 	/*
1621 	 * If we have exceeded by more than 1/5th then the algorithm below
1622 	 * will not bring us back into range.  Dividing by two here forces
1623 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1624 	 */
1625 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1626 		ts->ts_runtime /= 2;
1627 		ts->ts_slptime /= 2;
1628 		return;
1629 	}
1630 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1631 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1632 }
1633 
1634 /*
1635  * Scale back the interactivity history when a child thread is created.  The
1636  * history is inherited from the parent but the thread may behave totally
1637  * differently.  For example, a shell spawning a compiler process.  We want
1638  * to learn that the compiler is behaving badly very quickly.
1639  */
1640 static void
1641 sched_interact_fork(struct thread *td)
1642 {
1643 	struct td_sched *ts;
1644 	int ratio;
1645 	int sum;
1646 
1647 	ts = td_get_sched(td);
1648 	sum = ts->ts_runtime + ts->ts_slptime;
1649 	if (sum > SCHED_SLP_RUN_FORK) {
1650 		ratio = sum / SCHED_SLP_RUN_FORK;
1651 		ts->ts_runtime /= ratio;
1652 		ts->ts_slptime /= ratio;
1653 	}
1654 }
1655 
1656 /*
1657  * Called from proc0_init() to setup the scheduler fields.
1658  */
1659 void
1660 schedinit(void)
1661 {
1662 	struct td_sched *ts0;
1663 
1664 	/*
1665 	 * Set up the scheduler specific parts of thread0.
1666 	 */
1667 	ts0 = td_get_sched(&thread0);
1668 	ts0->ts_ltick = ticks;
1669 	ts0->ts_ftick = ticks;
1670 	ts0->ts_slice = 0;
1671 	ts0->ts_cpu = curcpu;	/* set valid CPU number */
1672 }
1673 
1674 /*
1675  * This is only somewhat accurate since given many processes of the same
1676  * priority they will switch when their slices run out, which will be
1677  * at most sched_slice stathz ticks.
1678  */
1679 int
1680 sched_rr_interval(void)
1681 {
1682 
1683 	/* Convert sched_slice from stathz to hz. */
1684 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1685 }
1686 
1687 /*
1688  * Update the percent cpu tracking information when it is requested or
1689  * the total history exceeds the maximum.  We keep a sliding history of
1690  * tick counts that slowly decays.  This is less precise than the 4BSD
1691  * mechanism since it happens with less regular and frequent events.
1692  */
1693 static void
1694 sched_pctcpu_update(struct td_sched *ts, int run)
1695 {
1696 	int t = ticks;
1697 
1698 	/*
1699 	 * The signed difference may be negative if the thread hasn't run for
1700 	 * over half of the ticks rollover period.
1701 	 */
1702 	if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) {
1703 		ts->ts_ticks = 0;
1704 		ts->ts_ftick = t - SCHED_TICK_TARG;
1705 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1706 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1707 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1708 		ts->ts_ftick = t - SCHED_TICK_TARG;
1709 	}
1710 	if (run)
1711 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1712 	ts->ts_ltick = t;
1713 }
1714 
1715 /*
1716  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1717  * if necessary.  This is the back-end for several priority related
1718  * functions.
1719  */
1720 static void
1721 sched_thread_priority(struct thread *td, u_char prio)
1722 {
1723 	struct td_sched *ts;
1724 	struct tdq *tdq;
1725 	int oldpri;
1726 
1727 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1728 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1729 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1730 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1731 	if (td != curthread && prio < td->td_priority) {
1732 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1733 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1734 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1735 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1736 		    curthread);
1737 	}
1738 	ts = td_get_sched(td);
1739 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1740 	if (td->td_priority == prio)
1741 		return;
1742 	/*
1743 	 * If the priority has been elevated due to priority
1744 	 * propagation, we may have to move ourselves to a new
1745 	 * queue.  This could be optimized to not re-add in some
1746 	 * cases.
1747 	 */
1748 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1749 		sched_rem(td);
1750 		td->td_priority = prio;
1751 		sched_add(td, SRQ_BORROWING);
1752 		return;
1753 	}
1754 	/*
1755 	 * If the thread is currently running we may have to adjust the lowpri
1756 	 * information so other cpus are aware of our current priority.
1757 	 */
1758 	if (TD_IS_RUNNING(td)) {
1759 		tdq = TDQ_CPU(ts->ts_cpu);
1760 		oldpri = td->td_priority;
1761 		td->td_priority = prio;
1762 		if (prio < tdq->tdq_lowpri)
1763 			tdq->tdq_lowpri = prio;
1764 		else if (tdq->tdq_lowpri == oldpri)
1765 			tdq_setlowpri(tdq, td);
1766 		return;
1767 	}
1768 	td->td_priority = prio;
1769 }
1770 
1771 /*
1772  * Update a thread's priority when it is lent another thread's
1773  * priority.
1774  */
1775 void
1776 sched_lend_prio(struct thread *td, u_char prio)
1777 {
1778 
1779 	td->td_flags |= TDF_BORROWING;
1780 	sched_thread_priority(td, prio);
1781 }
1782 
1783 /*
1784  * Restore a thread's priority when priority propagation is
1785  * over.  The prio argument is the minimum priority the thread
1786  * needs to have to satisfy other possible priority lending
1787  * requests.  If the thread's regular priority is less
1788  * important than prio, the thread will keep a priority boost
1789  * of prio.
1790  */
1791 void
1792 sched_unlend_prio(struct thread *td, u_char prio)
1793 {
1794 	u_char base_pri;
1795 
1796 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1797 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1798 		base_pri = td->td_user_pri;
1799 	else
1800 		base_pri = td->td_base_pri;
1801 	if (prio >= base_pri) {
1802 		td->td_flags &= ~TDF_BORROWING;
1803 		sched_thread_priority(td, base_pri);
1804 	} else
1805 		sched_lend_prio(td, prio);
1806 }
1807 
1808 /*
1809  * Standard entry for setting the priority to an absolute value.
1810  */
1811 void
1812 sched_prio(struct thread *td, u_char prio)
1813 {
1814 	u_char oldprio;
1815 
1816 	/* First, update the base priority. */
1817 	td->td_base_pri = prio;
1818 
1819 	/*
1820 	 * If the thread is borrowing another thread's priority, don't
1821 	 * ever lower the priority.
1822 	 */
1823 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1824 		return;
1825 
1826 	/* Change the real priority. */
1827 	oldprio = td->td_priority;
1828 	sched_thread_priority(td, prio);
1829 
1830 	/*
1831 	 * If the thread is on a turnstile, then let the turnstile update
1832 	 * its state.
1833 	 */
1834 	if (TD_ON_LOCK(td) && oldprio != prio)
1835 		turnstile_adjust(td, oldprio);
1836 }
1837 
1838 /*
1839  * Set the base user priority, does not effect current running priority.
1840  */
1841 void
1842 sched_user_prio(struct thread *td, u_char prio)
1843 {
1844 
1845 	td->td_base_user_pri = prio;
1846 	if (td->td_lend_user_pri <= prio)
1847 		return;
1848 	td->td_user_pri = prio;
1849 }
1850 
1851 void
1852 sched_lend_user_prio(struct thread *td, u_char prio)
1853 {
1854 
1855 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1856 	td->td_lend_user_pri = prio;
1857 	td->td_user_pri = min(prio, td->td_base_user_pri);
1858 	if (td->td_priority > td->td_user_pri)
1859 		sched_prio(td, td->td_user_pri);
1860 	else if (td->td_priority != td->td_user_pri)
1861 		td->td_flags |= TDF_NEEDRESCHED;
1862 }
1863 
1864 #ifdef SMP
1865 /*
1866  * This tdq is about to idle.  Try to steal a thread from another CPU before
1867  * choosing the idle thread.
1868  */
1869 static void
1870 tdq_trysteal(struct tdq *tdq)
1871 {
1872 	struct cpu_group *cg;
1873 	struct tdq *steal;
1874 	cpuset_t mask;
1875 	int cpu, i;
1876 
1877 	if (smp_started == 0 || trysteal_limit == 0 || tdq->tdq_cg == NULL)
1878 		return;
1879 	CPU_FILL(&mask);
1880 	CPU_CLR(PCPU_GET(cpuid), &mask);
1881 	/* We don't want to be preempted while we're iterating. */
1882 	spinlock_enter();
1883 	TDQ_UNLOCK(tdq);
1884 	for (i = 1, cg = tdq->tdq_cg; ; ) {
1885 		cpu = sched_highest(cg, mask, steal_thresh);
1886 		/*
1887 		 * If a thread was added while interrupts were disabled don't
1888 		 * steal one here.
1889 		 */
1890 		if (tdq->tdq_load > 0) {
1891 			TDQ_LOCK(tdq);
1892 			break;
1893 		}
1894 		if (cpu == -1) {
1895 			i++;
1896 			cg = cg->cg_parent;
1897 			if (cg == NULL || i > trysteal_limit) {
1898 				TDQ_LOCK(tdq);
1899 				break;
1900 			}
1901 			continue;
1902 		}
1903 		steal = TDQ_CPU(cpu);
1904 		/*
1905 		 * The data returned by sched_highest() is stale and
1906                  * the chosen CPU no longer has an eligible thread.
1907 		 */
1908 		if (steal->tdq_load < steal_thresh ||
1909 		    steal->tdq_transferable == 0)
1910 			continue;
1911 		tdq_lock_pair(tdq, steal);
1912 		/*
1913 		 * If we get to this point, unconditonally exit the loop
1914 		 * to bound the time spent in the critcal section.
1915 		 *
1916 		 * If a thread was added while interrupts were disabled don't
1917 		 * steal one here.
1918 		 */
1919 		if (tdq->tdq_load > 0) {
1920 			TDQ_UNLOCK(steal);
1921 			break;
1922 		}
1923 		/*
1924 		 * The data returned by sched_highest() is stale and
1925                  * the chosen CPU no longer has an eligible thread.
1926 		 */
1927 		if (steal->tdq_load < steal_thresh ||
1928 		    steal->tdq_transferable == 0) {
1929 			TDQ_UNLOCK(steal);
1930 			break;
1931 		}
1932 		/*
1933 		 * If we fail to acquire one due to affinity restrictions,
1934 		 * bail out and let the idle thread to a more complete search
1935 		 * outside of a critical section.
1936 		 */
1937 		if (tdq_move(steal, tdq) == NULL) {
1938 			TDQ_UNLOCK(steal);
1939 			break;
1940 		}
1941 		TDQ_UNLOCK(steal);
1942 		break;
1943 	}
1944 	spinlock_exit();
1945 }
1946 #endif
1947 
1948 /*
1949  * Handle migration from sched_switch().  This happens only for
1950  * cpu binding.
1951  */
1952 static struct mtx *
1953 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1954 {
1955 	struct tdq *tdn;
1956 
1957 	KASSERT(!CPU_ABSENT(td_get_sched(td)->ts_cpu), ("sched_switch_migrate: "
1958 	    "thread %s queued on absent CPU %d.", td->td_name,
1959 	    td_get_sched(td)->ts_cpu));
1960 	tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
1961 #ifdef SMP
1962 	tdq_load_rem(tdq, td);
1963 	/*
1964 	 * Do the lock dance required to avoid LOR.  We grab an extra
1965 	 * spinlock nesting to prevent preemption while we're
1966 	 * not holding either run-queue lock.
1967 	 */
1968 	spinlock_enter();
1969 	thread_lock_block(td);	/* This releases the lock on tdq. */
1970 
1971 	/*
1972 	 * Acquire both run-queue locks before placing the thread on the new
1973 	 * run-queue to avoid deadlocks created by placing a thread with a
1974 	 * blocked lock on the run-queue of a remote processor.  The deadlock
1975 	 * occurs when a third processor attempts to lock the two queues in
1976 	 * question while the target processor is spinning with its own
1977 	 * run-queue lock held while waiting for the blocked lock to clear.
1978 	 */
1979 	tdq_lock_pair(tdn, tdq);
1980 	tdq_add(tdn, td, flags);
1981 	tdq_notify(tdn, td);
1982 	TDQ_UNLOCK(tdn);
1983 	spinlock_exit();
1984 #endif
1985 	return (TDQ_LOCKPTR(tdn));
1986 }
1987 
1988 /*
1989  * Variadic version of thread_lock_unblock() that does not assume td_lock
1990  * is blocked.
1991  */
1992 static inline void
1993 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1994 {
1995 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1996 	    (uintptr_t)mtx);
1997 }
1998 
1999 /*
2000  * Switch threads.  This function has to handle threads coming in while
2001  * blocked for some reason, running, or idle.  It also must deal with
2002  * migrating a thread from one queue to another as running threads may
2003  * be assigned elsewhere via binding.
2004  */
2005 void
2006 sched_switch(struct thread *td, struct thread *newtd, int flags)
2007 {
2008 	struct tdq *tdq;
2009 	struct td_sched *ts;
2010 	struct mtx *mtx;
2011 	int srqflag;
2012 	int cpuid, preempted;
2013 
2014 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2015 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
2016 
2017 	cpuid = PCPU_GET(cpuid);
2018 	tdq = TDQ_CPU(cpuid);
2019 	ts = td_get_sched(td);
2020 	mtx = td->td_lock;
2021 	sched_pctcpu_update(ts, 1);
2022 	ts->ts_rltick = ticks;
2023 	td->td_lastcpu = td->td_oncpu;
2024 	td->td_oncpu = NOCPU;
2025 	preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
2026 	    (flags & SW_PREEMPT) != 0;
2027 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
2028 	td->td_owepreempt = 0;
2029 	if (!TD_IS_IDLETHREAD(td))
2030 		tdq->tdq_switchcnt++;
2031 	/*
2032 	 * The lock pointer in an idle thread should never change.  Reset it
2033 	 * to CAN_RUN as well.
2034 	 */
2035 	if (TD_IS_IDLETHREAD(td)) {
2036 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2037 		TD_SET_CAN_RUN(td);
2038 	} else if (TD_IS_RUNNING(td)) {
2039 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2040 		srqflag = preempted ?
2041 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
2042 		    SRQ_OURSELF|SRQ_YIELDING;
2043 #ifdef SMP
2044 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
2045 			ts->ts_cpu = sched_pickcpu(td, 0);
2046 #endif
2047 		if (ts->ts_cpu == cpuid)
2048 			tdq_runq_add(tdq, td, srqflag);
2049 		else {
2050 			KASSERT(THREAD_CAN_MIGRATE(td) ||
2051 			    (ts->ts_flags & TSF_BOUND) != 0,
2052 			    ("Thread %p shouldn't migrate", td));
2053 			mtx = sched_switch_migrate(tdq, td, srqflag);
2054 		}
2055 	} else {
2056 		/* This thread must be going to sleep. */
2057 		TDQ_LOCK(tdq);
2058 		mtx = thread_lock_block(td);
2059 		tdq_load_rem(tdq, td);
2060 #ifdef SMP
2061 		if (tdq->tdq_load == 0)
2062 			tdq_trysteal(tdq);
2063 #endif
2064 	}
2065 
2066 #if (KTR_COMPILE & KTR_SCHED) != 0
2067 	if (TD_IS_IDLETHREAD(td))
2068 		KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
2069 		    "prio:%d", td->td_priority);
2070 	else
2071 		KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
2072 		    "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
2073 		    "lockname:\"%s\"", td->td_lockname);
2074 #endif
2075 
2076 	/*
2077 	 * We enter here with the thread blocked and assigned to the
2078 	 * appropriate cpu run-queue or sleep-queue and with the current
2079 	 * thread-queue locked.
2080 	 */
2081 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2082 	newtd = choosethread();
2083 	/*
2084 	 * Call the MD code to switch contexts if necessary.
2085 	 */
2086 	if (td != newtd) {
2087 #ifdef	HWPMC_HOOKS
2088 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2089 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
2090 #endif
2091 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
2092 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2093 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2094 		sched_pctcpu_update(td_get_sched(newtd), 0);
2095 
2096 #ifdef KDTRACE_HOOKS
2097 		/*
2098 		 * If DTrace has set the active vtime enum to anything
2099 		 * other than INACTIVE (0), then it should have set the
2100 		 * function to call.
2101 		 */
2102 		if (dtrace_vtime_active)
2103 			(*dtrace_vtime_switch_func)(newtd);
2104 #endif
2105 
2106 		cpu_switch(td, newtd, mtx);
2107 		/*
2108 		 * We may return from cpu_switch on a different cpu.  However,
2109 		 * we always return with td_lock pointing to the current cpu's
2110 		 * run queue lock.
2111 		 */
2112 		cpuid = PCPU_GET(cpuid);
2113 		tdq = TDQ_CPU(cpuid);
2114 		lock_profile_obtain_lock_success(
2115 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2116 
2117 		SDT_PROBE0(sched, , , on__cpu);
2118 #ifdef	HWPMC_HOOKS
2119 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2120 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
2121 #endif
2122 	} else {
2123 		thread_unblock_switch(td, mtx);
2124 		SDT_PROBE0(sched, , , remain__cpu);
2125 	}
2126 
2127 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2128 	    "prio:%d", td->td_priority);
2129 
2130 	/*
2131 	 * Assert that all went well and return.
2132 	 */
2133 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
2134 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2135 	td->td_oncpu = cpuid;
2136 }
2137 
2138 /*
2139  * Adjust thread priorities as a result of a nice request.
2140  */
2141 void
2142 sched_nice(struct proc *p, int nice)
2143 {
2144 	struct thread *td;
2145 
2146 	PROC_LOCK_ASSERT(p, MA_OWNED);
2147 
2148 	p->p_nice = nice;
2149 	FOREACH_THREAD_IN_PROC(p, td) {
2150 		thread_lock(td);
2151 		sched_priority(td);
2152 		sched_prio(td, td->td_base_user_pri);
2153 		thread_unlock(td);
2154 	}
2155 }
2156 
2157 /*
2158  * Record the sleep time for the interactivity scorer.
2159  */
2160 void
2161 sched_sleep(struct thread *td, int prio)
2162 {
2163 
2164 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2165 
2166 	td->td_slptick = ticks;
2167 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2168 		td->td_flags |= TDF_CANSWAP;
2169 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2170 		return;
2171 	if (static_boost == 1 && prio)
2172 		sched_prio(td, prio);
2173 	else if (static_boost && td->td_priority > static_boost)
2174 		sched_prio(td, static_boost);
2175 }
2176 
2177 /*
2178  * Schedule a thread to resume execution and record how long it voluntarily
2179  * slept.  We also update the pctcpu, interactivity, and priority.
2180  */
2181 void
2182 sched_wakeup(struct thread *td)
2183 {
2184 	struct td_sched *ts;
2185 	int slptick;
2186 
2187 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2188 	ts = td_get_sched(td);
2189 	td->td_flags &= ~TDF_CANSWAP;
2190 	/*
2191 	 * If we slept for more than a tick update our interactivity and
2192 	 * priority.
2193 	 */
2194 	slptick = td->td_slptick;
2195 	td->td_slptick = 0;
2196 	if (slptick && slptick != ticks) {
2197 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2198 		sched_interact_update(td);
2199 		sched_pctcpu_update(ts, 0);
2200 	}
2201 	/*
2202 	 * Reset the slice value since we slept and advanced the round-robin.
2203 	 */
2204 	ts->ts_slice = 0;
2205 	sched_add(td, SRQ_BORING);
2206 }
2207 
2208 /*
2209  * Penalize the parent for creating a new child and initialize the child's
2210  * priority.
2211  */
2212 void
2213 sched_fork(struct thread *td, struct thread *child)
2214 {
2215 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2216 	sched_pctcpu_update(td_get_sched(td), 1);
2217 	sched_fork_thread(td, child);
2218 	/*
2219 	 * Penalize the parent and child for forking.
2220 	 */
2221 	sched_interact_fork(child);
2222 	sched_priority(child);
2223 	td_get_sched(td)->ts_runtime += tickincr;
2224 	sched_interact_update(td);
2225 	sched_priority(td);
2226 }
2227 
2228 /*
2229  * Fork a new thread, may be within the same process.
2230  */
2231 void
2232 sched_fork_thread(struct thread *td, struct thread *child)
2233 {
2234 	struct td_sched *ts;
2235 	struct td_sched *ts2;
2236 	struct tdq *tdq;
2237 
2238 	tdq = TDQ_SELF();
2239 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2240 	/*
2241 	 * Initialize child.
2242 	 */
2243 	ts = td_get_sched(td);
2244 	ts2 = td_get_sched(child);
2245 	child->td_oncpu = NOCPU;
2246 	child->td_lastcpu = NOCPU;
2247 	child->td_lock = TDQ_LOCKPTR(tdq);
2248 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2249 	child->td_domain.dr_policy = td->td_cpuset->cs_domain;
2250 	ts2->ts_cpu = ts->ts_cpu;
2251 	ts2->ts_flags = 0;
2252 	/*
2253 	 * Grab our parents cpu estimation information.
2254 	 */
2255 	ts2->ts_ticks = ts->ts_ticks;
2256 	ts2->ts_ltick = ts->ts_ltick;
2257 	ts2->ts_ftick = ts->ts_ftick;
2258 	/*
2259 	 * Do not inherit any borrowed priority from the parent.
2260 	 */
2261 	child->td_priority = child->td_base_pri;
2262 	/*
2263 	 * And update interactivity score.
2264 	 */
2265 	ts2->ts_slptime = ts->ts_slptime;
2266 	ts2->ts_runtime = ts->ts_runtime;
2267 	/* Attempt to quickly learn interactivity. */
2268 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2269 #ifdef KTR
2270 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2271 #endif
2272 }
2273 
2274 /*
2275  * Adjust the priority class of a thread.
2276  */
2277 void
2278 sched_class(struct thread *td, int class)
2279 {
2280 
2281 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2282 	if (td->td_pri_class == class)
2283 		return;
2284 	td->td_pri_class = class;
2285 }
2286 
2287 /*
2288  * Return some of the child's priority and interactivity to the parent.
2289  */
2290 void
2291 sched_exit(struct proc *p, struct thread *child)
2292 {
2293 	struct thread *td;
2294 
2295 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2296 	    "prio:%d", child->td_priority);
2297 	PROC_LOCK_ASSERT(p, MA_OWNED);
2298 	td = FIRST_THREAD_IN_PROC(p);
2299 	sched_exit_thread(td, child);
2300 }
2301 
2302 /*
2303  * Penalize another thread for the time spent on this one.  This helps to
2304  * worsen the priority and interactivity of processes which schedule batch
2305  * jobs such as make.  This has little effect on the make process itself but
2306  * causes new processes spawned by it to receive worse scores immediately.
2307  */
2308 void
2309 sched_exit_thread(struct thread *td, struct thread *child)
2310 {
2311 
2312 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2313 	    "prio:%d", child->td_priority);
2314 	/*
2315 	 * Give the child's runtime to the parent without returning the
2316 	 * sleep time as a penalty to the parent.  This causes shells that
2317 	 * launch expensive things to mark their children as expensive.
2318 	 */
2319 	thread_lock(td);
2320 	td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2321 	sched_interact_update(td);
2322 	sched_priority(td);
2323 	thread_unlock(td);
2324 }
2325 
2326 void
2327 sched_preempt(struct thread *td)
2328 {
2329 	struct tdq *tdq;
2330 
2331 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2332 
2333 	thread_lock(td);
2334 	tdq = TDQ_SELF();
2335 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2336 	tdq->tdq_ipipending = 0;
2337 	if (td->td_priority > tdq->tdq_lowpri) {
2338 		int flags;
2339 
2340 		flags = SW_INVOL | SW_PREEMPT;
2341 		if (td->td_critnest > 1)
2342 			td->td_owepreempt = 1;
2343 		else if (TD_IS_IDLETHREAD(td))
2344 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2345 		else
2346 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2347 	}
2348 	thread_unlock(td);
2349 }
2350 
2351 /*
2352  * Fix priorities on return to user-space.  Priorities may be elevated due
2353  * to static priorities in msleep() or similar.
2354  */
2355 void
2356 sched_userret_slowpath(struct thread *td)
2357 {
2358 
2359 	thread_lock(td);
2360 	td->td_priority = td->td_user_pri;
2361 	td->td_base_pri = td->td_user_pri;
2362 	tdq_setlowpri(TDQ_SELF(), td);
2363 	thread_unlock(td);
2364 }
2365 
2366 /*
2367  * Handle a stathz tick.  This is really only relevant for timeshare
2368  * threads.
2369  */
2370 void
2371 sched_clock(struct thread *td)
2372 {
2373 	struct tdq *tdq;
2374 	struct td_sched *ts;
2375 
2376 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2377 	tdq = TDQ_SELF();
2378 #ifdef SMP
2379 	/*
2380 	 * We run the long term load balancer infrequently on the first cpu.
2381 	 */
2382 	if (balance_tdq == tdq && smp_started != 0 && rebalance != 0) {
2383 		if (balance_ticks && --balance_ticks == 0)
2384 			sched_balance();
2385 	}
2386 #endif
2387 	/*
2388 	 * Save the old switch count so we have a record of the last ticks
2389 	 * activity.   Initialize the new switch count based on our load.
2390 	 * If there is some activity seed it to reflect that.
2391 	 */
2392 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2393 	tdq->tdq_switchcnt = tdq->tdq_load;
2394 	/*
2395 	 * Advance the insert index once for each tick to ensure that all
2396 	 * threads get a chance to run.
2397 	 */
2398 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2399 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2400 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2401 			tdq->tdq_ridx = tdq->tdq_idx;
2402 	}
2403 	ts = td_get_sched(td);
2404 	sched_pctcpu_update(ts, 1);
2405 	if (td->td_pri_class & PRI_FIFO_BIT)
2406 		return;
2407 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2408 		/*
2409 		 * We used a tick; charge it to the thread so
2410 		 * that we can compute our interactivity.
2411 		 */
2412 		td_get_sched(td)->ts_runtime += tickincr;
2413 		sched_interact_update(td);
2414 		sched_priority(td);
2415 	}
2416 
2417 	/*
2418 	 * Force a context switch if the current thread has used up a full
2419 	 * time slice (default is 100ms).
2420 	 */
2421 	if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2422 		ts->ts_slice = 0;
2423 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2424 	}
2425 }
2426 
2427 u_int
2428 sched_estcpu(struct thread *td __unused)
2429 {
2430 
2431 	return (0);
2432 }
2433 
2434 /*
2435  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2436  * cooperative idle threads.
2437  */
2438 int
2439 sched_runnable(void)
2440 {
2441 	struct tdq *tdq;
2442 	int load;
2443 
2444 	load = 1;
2445 
2446 	tdq = TDQ_SELF();
2447 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2448 		if (tdq->tdq_load > 0)
2449 			goto out;
2450 	} else
2451 		if (tdq->tdq_load - 1 > 0)
2452 			goto out;
2453 	load = 0;
2454 out:
2455 	return (load);
2456 }
2457 
2458 /*
2459  * Choose the highest priority thread to run.  The thread is removed from
2460  * the run-queue while running however the load remains.  For SMP we set
2461  * the tdq in the global idle bitmask if it idles here.
2462  */
2463 struct thread *
2464 sched_choose(void)
2465 {
2466 	struct thread *td;
2467 	struct tdq *tdq;
2468 
2469 	tdq = TDQ_SELF();
2470 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2471 	td = tdq_choose(tdq);
2472 	if (td) {
2473 		tdq_runq_rem(tdq, td);
2474 		tdq->tdq_lowpri = td->td_priority;
2475 		return (td);
2476 	}
2477 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2478 	return (PCPU_GET(idlethread));
2479 }
2480 
2481 /*
2482  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2483  * we always request it once we exit a critical section.
2484  */
2485 static inline void
2486 sched_setpreempt(struct thread *td)
2487 {
2488 	struct thread *ctd;
2489 	int cpri;
2490 	int pri;
2491 
2492 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2493 
2494 	ctd = curthread;
2495 	pri = td->td_priority;
2496 	cpri = ctd->td_priority;
2497 	if (pri < cpri)
2498 		ctd->td_flags |= TDF_NEEDRESCHED;
2499 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2500 		return;
2501 	if (!sched_shouldpreempt(pri, cpri, 0))
2502 		return;
2503 	ctd->td_owepreempt = 1;
2504 }
2505 
2506 /*
2507  * Add a thread to a thread queue.  Select the appropriate runq and add the
2508  * thread to it.  This is the internal function called when the tdq is
2509  * predetermined.
2510  */
2511 void
2512 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2513 {
2514 
2515 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2516 	KASSERT((td->td_inhibitors == 0),
2517 	    ("sched_add: trying to run inhibited thread"));
2518 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2519 	    ("sched_add: bad thread state"));
2520 	KASSERT(td->td_flags & TDF_INMEM,
2521 	    ("sched_add: thread swapped out"));
2522 
2523 	if (td->td_priority < tdq->tdq_lowpri)
2524 		tdq->tdq_lowpri = td->td_priority;
2525 	tdq_runq_add(tdq, td, flags);
2526 	tdq_load_add(tdq, td);
2527 }
2528 
2529 /*
2530  * Select the target thread queue and add a thread to it.  Request
2531  * preemption or IPI a remote processor if required.
2532  */
2533 void
2534 sched_add(struct thread *td, int flags)
2535 {
2536 	struct tdq *tdq;
2537 #ifdef SMP
2538 	int cpu;
2539 #endif
2540 
2541 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2542 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2543 	    sched_tdname(curthread));
2544 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2545 	    KTR_ATTR_LINKED, sched_tdname(td));
2546 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2547 	    flags & SRQ_PREEMPTED);
2548 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2549 	/*
2550 	 * Recalculate the priority before we select the target cpu or
2551 	 * run-queue.
2552 	 */
2553 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2554 		sched_priority(td);
2555 #ifdef SMP
2556 	/*
2557 	 * Pick the destination cpu and if it isn't ours transfer to the
2558 	 * target cpu.
2559 	 */
2560 	cpu = sched_pickcpu(td, flags);
2561 	tdq = sched_setcpu(td, cpu, flags);
2562 	tdq_add(tdq, td, flags);
2563 	if (cpu != PCPU_GET(cpuid)) {
2564 		tdq_notify(tdq, td);
2565 		return;
2566 	}
2567 #else
2568 	tdq = TDQ_SELF();
2569 	TDQ_LOCK(tdq);
2570 	/*
2571 	 * Now that the thread is moving to the run-queue, set the lock
2572 	 * to the scheduler's lock.
2573 	 */
2574 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2575 	tdq_add(tdq, td, flags);
2576 #endif
2577 	if (!(flags & SRQ_YIELDING))
2578 		sched_setpreempt(td);
2579 }
2580 
2581 /*
2582  * Remove a thread from a run-queue without running it.  This is used
2583  * when we're stealing a thread from a remote queue.  Otherwise all threads
2584  * exit by calling sched_exit_thread() and sched_throw() themselves.
2585  */
2586 void
2587 sched_rem(struct thread *td)
2588 {
2589 	struct tdq *tdq;
2590 
2591 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2592 	    "prio:%d", td->td_priority);
2593 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2594 	tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2595 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2596 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2597 	KASSERT(TD_ON_RUNQ(td),
2598 	    ("sched_rem: thread not on run queue"));
2599 	tdq_runq_rem(tdq, td);
2600 	tdq_load_rem(tdq, td);
2601 	TD_SET_CAN_RUN(td);
2602 	if (td->td_priority == tdq->tdq_lowpri)
2603 		tdq_setlowpri(tdq, NULL);
2604 }
2605 
2606 /*
2607  * Fetch cpu utilization information.  Updates on demand.
2608  */
2609 fixpt_t
2610 sched_pctcpu(struct thread *td)
2611 {
2612 	fixpt_t pctcpu;
2613 	struct td_sched *ts;
2614 
2615 	pctcpu = 0;
2616 	ts = td_get_sched(td);
2617 
2618 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2619 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2620 	if (ts->ts_ticks) {
2621 		int rtick;
2622 
2623 		/* How many rtick per second ? */
2624 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2625 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2626 	}
2627 
2628 	return (pctcpu);
2629 }
2630 
2631 /*
2632  * Enforce affinity settings for a thread.  Called after adjustments to
2633  * cpumask.
2634  */
2635 void
2636 sched_affinity(struct thread *td)
2637 {
2638 #ifdef SMP
2639 	struct td_sched *ts;
2640 
2641 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2642 	ts = td_get_sched(td);
2643 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2644 		return;
2645 	if (TD_ON_RUNQ(td)) {
2646 		sched_rem(td);
2647 		sched_add(td, SRQ_BORING);
2648 		return;
2649 	}
2650 	if (!TD_IS_RUNNING(td))
2651 		return;
2652 	/*
2653 	 * Force a switch before returning to userspace.  If the
2654 	 * target thread is not running locally send an ipi to force
2655 	 * the issue.
2656 	 */
2657 	td->td_flags |= TDF_NEEDRESCHED;
2658 	if (td != curthread)
2659 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2660 #endif
2661 }
2662 
2663 /*
2664  * Bind a thread to a target cpu.
2665  */
2666 void
2667 sched_bind(struct thread *td, int cpu)
2668 {
2669 	struct td_sched *ts;
2670 
2671 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2672 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2673 	ts = td_get_sched(td);
2674 	if (ts->ts_flags & TSF_BOUND)
2675 		sched_unbind(td);
2676 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2677 	ts->ts_flags |= TSF_BOUND;
2678 	sched_pin();
2679 	if (PCPU_GET(cpuid) == cpu)
2680 		return;
2681 	ts->ts_cpu = cpu;
2682 	/* When we return from mi_switch we'll be on the correct cpu. */
2683 	mi_switch(SW_VOL, NULL);
2684 }
2685 
2686 /*
2687  * Release a bound thread.
2688  */
2689 void
2690 sched_unbind(struct thread *td)
2691 {
2692 	struct td_sched *ts;
2693 
2694 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2695 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2696 	ts = td_get_sched(td);
2697 	if ((ts->ts_flags & TSF_BOUND) == 0)
2698 		return;
2699 	ts->ts_flags &= ~TSF_BOUND;
2700 	sched_unpin();
2701 }
2702 
2703 int
2704 sched_is_bound(struct thread *td)
2705 {
2706 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2707 	return (td_get_sched(td)->ts_flags & TSF_BOUND);
2708 }
2709 
2710 /*
2711  * Basic yield call.
2712  */
2713 void
2714 sched_relinquish(struct thread *td)
2715 {
2716 	thread_lock(td);
2717 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2718 	thread_unlock(td);
2719 }
2720 
2721 /*
2722  * Return the total system load.
2723  */
2724 int
2725 sched_load(void)
2726 {
2727 #ifdef SMP
2728 	int total;
2729 	int i;
2730 
2731 	total = 0;
2732 	CPU_FOREACH(i)
2733 		total += TDQ_CPU(i)->tdq_sysload;
2734 	return (total);
2735 #else
2736 	return (TDQ_SELF()->tdq_sysload);
2737 #endif
2738 }
2739 
2740 int
2741 sched_sizeof_proc(void)
2742 {
2743 	return (sizeof(struct proc));
2744 }
2745 
2746 int
2747 sched_sizeof_thread(void)
2748 {
2749 	return (sizeof(struct thread) + sizeof(struct td_sched));
2750 }
2751 
2752 #ifdef SMP
2753 #define	TDQ_IDLESPIN(tdq)						\
2754     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2755 #else
2756 #define	TDQ_IDLESPIN(tdq)	1
2757 #endif
2758 
2759 /*
2760  * The actual idle process.
2761  */
2762 void
2763 sched_idletd(void *dummy)
2764 {
2765 	struct thread *td;
2766 	struct tdq *tdq;
2767 	int oldswitchcnt, switchcnt;
2768 	int i;
2769 
2770 	mtx_assert(&Giant, MA_NOTOWNED);
2771 	td = curthread;
2772 	tdq = TDQ_SELF();
2773 	THREAD_NO_SLEEPING();
2774 	oldswitchcnt = -1;
2775 	for (;;) {
2776 		if (tdq->tdq_load) {
2777 			thread_lock(td);
2778 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2779 			thread_unlock(td);
2780 		}
2781 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2782 #ifdef SMP
2783 		if (always_steal || switchcnt != oldswitchcnt) {
2784 			oldswitchcnt = switchcnt;
2785 			if (tdq_idled(tdq) == 0)
2786 				continue;
2787 		}
2788 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2789 #else
2790 		oldswitchcnt = switchcnt;
2791 #endif
2792 		/*
2793 		 * If we're switching very frequently, spin while checking
2794 		 * for load rather than entering a low power state that
2795 		 * may require an IPI.  However, don't do any busy
2796 		 * loops while on SMT machines as this simply steals
2797 		 * cycles from cores doing useful work.
2798 		 */
2799 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2800 			for (i = 0; i < sched_idlespins; i++) {
2801 				if (tdq->tdq_load)
2802 					break;
2803 				cpu_spinwait();
2804 			}
2805 		}
2806 
2807 		/* If there was context switch during spin, restart it. */
2808 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2809 		if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2810 			continue;
2811 
2812 		/* Run main MD idle handler. */
2813 		tdq->tdq_cpu_idle = 1;
2814 		/*
2815 		 * Make sure that tdq_cpu_idle update is globally visible
2816 		 * before cpu_idle() read tdq_load.  The order is important
2817 		 * to avoid race with tdq_notify.
2818 		 */
2819 		atomic_thread_fence_seq_cst();
2820 		/*
2821 		 * Checking for again after the fence picks up assigned
2822 		 * threads often enough to make it worthwhile to do so in
2823 		 * order to avoid calling cpu_idle().
2824 		 */
2825 		if (tdq->tdq_load != 0) {
2826 			tdq->tdq_cpu_idle = 0;
2827 			continue;
2828 		}
2829 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2830 		tdq->tdq_cpu_idle = 0;
2831 
2832 		/*
2833 		 * Account thread-less hardware interrupts and
2834 		 * other wakeup reasons equal to context switches.
2835 		 */
2836 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2837 		if (switchcnt != oldswitchcnt)
2838 			continue;
2839 		tdq->tdq_switchcnt++;
2840 		oldswitchcnt++;
2841 	}
2842 }
2843 
2844 /*
2845  * A CPU is entering for the first time or a thread is exiting.
2846  */
2847 void
2848 sched_throw(struct thread *td)
2849 {
2850 	struct thread *newtd;
2851 	struct tdq *tdq;
2852 
2853 	tdq = TDQ_SELF();
2854 	if (td == NULL) {
2855 		/* Correct spinlock nesting and acquire the correct lock. */
2856 		TDQ_LOCK(tdq);
2857 		spinlock_exit();
2858 		PCPU_SET(switchtime, cpu_ticks());
2859 		PCPU_SET(switchticks, ticks);
2860 	} else {
2861 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2862 		tdq_load_rem(tdq, td);
2863 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2864 		td->td_lastcpu = td->td_oncpu;
2865 		td->td_oncpu = NOCPU;
2866 	}
2867 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2868 	newtd = choosethread();
2869 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2870 	cpu_throw(td, newtd);		/* doesn't return */
2871 }
2872 
2873 /*
2874  * This is called from fork_exit().  Just acquire the correct locks and
2875  * let fork do the rest of the work.
2876  */
2877 void
2878 sched_fork_exit(struct thread *td)
2879 {
2880 	struct tdq *tdq;
2881 	int cpuid;
2882 
2883 	/*
2884 	 * Finish setting up thread glue so that it begins execution in a
2885 	 * non-nested critical section with the scheduler lock held.
2886 	 */
2887 	cpuid = PCPU_GET(cpuid);
2888 	tdq = TDQ_CPU(cpuid);
2889 	if (TD_IS_IDLETHREAD(td))
2890 		td->td_lock = TDQ_LOCKPTR(tdq);
2891 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2892 	td->td_oncpu = cpuid;
2893 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2894 	lock_profile_obtain_lock_success(
2895 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2896 
2897 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2898 	    "prio:%d", td->td_priority);
2899 	SDT_PROBE0(sched, , , on__cpu);
2900 }
2901 
2902 /*
2903  * Create on first use to catch odd startup conditons.
2904  */
2905 char *
2906 sched_tdname(struct thread *td)
2907 {
2908 #ifdef KTR
2909 	struct td_sched *ts;
2910 
2911 	ts = td_get_sched(td);
2912 	if (ts->ts_name[0] == '\0')
2913 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2914 		    "%s tid %d", td->td_name, td->td_tid);
2915 	return (ts->ts_name);
2916 #else
2917 	return (td->td_name);
2918 #endif
2919 }
2920 
2921 #ifdef KTR
2922 void
2923 sched_clear_tdname(struct thread *td)
2924 {
2925 	struct td_sched *ts;
2926 
2927 	ts = td_get_sched(td);
2928 	ts->ts_name[0] = '\0';
2929 }
2930 #endif
2931 
2932 #ifdef SMP
2933 
2934 /*
2935  * Build the CPU topology dump string. Is recursively called to collect
2936  * the topology tree.
2937  */
2938 static int
2939 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2940     int indent)
2941 {
2942 	char cpusetbuf[CPUSETBUFSIZ];
2943 	int i, first;
2944 
2945 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2946 	    "", 1 + indent / 2, cg->cg_level);
2947 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2948 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2949 	first = TRUE;
2950 	for (i = 0; i < MAXCPU; i++) {
2951 		if (CPU_ISSET(i, &cg->cg_mask)) {
2952 			if (!first)
2953 				sbuf_printf(sb, ", ");
2954 			else
2955 				first = FALSE;
2956 			sbuf_printf(sb, "%d", i);
2957 		}
2958 	}
2959 	sbuf_printf(sb, "</cpu>\n");
2960 
2961 	if (cg->cg_flags != 0) {
2962 		sbuf_printf(sb, "%*s <flags>", indent, "");
2963 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2964 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2965 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2966 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2967 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2968 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2969 		sbuf_printf(sb, "</flags>\n");
2970 	}
2971 
2972 	if (cg->cg_children > 0) {
2973 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2974 		for (i = 0; i < cg->cg_children; i++)
2975 			sysctl_kern_sched_topology_spec_internal(sb,
2976 			    &cg->cg_child[i], indent+2);
2977 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2978 	}
2979 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2980 	return (0);
2981 }
2982 
2983 /*
2984  * Sysctl handler for retrieving topology dump. It's a wrapper for
2985  * the recursive sysctl_kern_smp_topology_spec_internal().
2986  */
2987 static int
2988 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2989 {
2990 	struct sbuf *topo;
2991 	int err;
2992 
2993 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2994 
2995 	topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
2996 	if (topo == NULL)
2997 		return (ENOMEM);
2998 
2999 	sbuf_printf(topo, "<groups>\n");
3000 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
3001 	sbuf_printf(topo, "</groups>\n");
3002 
3003 	if (err == 0) {
3004 		err = sbuf_finish(topo);
3005 	}
3006 	sbuf_delete(topo);
3007 	return (err);
3008 }
3009 
3010 #endif
3011 
3012 static int
3013 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
3014 {
3015 	int error, new_val, period;
3016 
3017 	period = 1000000 / realstathz;
3018 	new_val = period * sched_slice;
3019 	error = sysctl_handle_int(oidp, &new_val, 0, req);
3020 	if (error != 0 || req->newptr == NULL)
3021 		return (error);
3022 	if (new_val <= 0)
3023 		return (EINVAL);
3024 	sched_slice = imax(1, (new_val + period / 2) / period);
3025 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
3026 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
3027 	    realstathz);
3028 	return (0);
3029 }
3030 
3031 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
3032 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
3033     "Scheduler name");
3034 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
3035     NULL, 0, sysctl_kern_quantum, "I",
3036     "Quantum for timeshare threads in microseconds");
3037 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
3038     "Quantum for timeshare threads in stathz ticks");
3039 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
3040     "Interactivity score threshold");
3041 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
3042     &preempt_thresh, 0,
3043     "Maximal (lowest) priority for preemption");
3044 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
3045     "Assign static kernel priorities to sleeping threads");
3046 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
3047     "Number of times idle thread will spin waiting for new work");
3048 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
3049     &sched_idlespinthresh, 0,
3050     "Threshold before we will permit idle thread spinning");
3051 #ifdef SMP
3052 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
3053     "Number of hz ticks to keep thread affinity for");
3054 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
3055     "Enables the long-term load balancer");
3056 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
3057     &balance_interval, 0,
3058     "Average period in stathz ticks to run the long-term balancer");
3059 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
3060     "Attempts to steal work from other cores before idling");
3061 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
3062     "Minimum load on remote CPU before we'll steal");
3063 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit,
3064     0, "Topological distance limit for stealing threads in sched_switch()");
3065 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0,
3066     "Always run the stealer from the idle thread");
3067 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
3068     CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
3069     "XML dump of detected CPU topology");
3070 #endif
3071 
3072 /* ps compat.  All cpu percentages from ULE are weighted. */
3073 static int ccpu = 0;
3074 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
3075