xref: /freebsd/sys/kern/sched_ule.c (revision 61322a0a8affb0c41dcf441ea331f041e15879ff)
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 __read_mostly		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 __read_mostly sched_interact = SCHED_INTERACT_THRESH;
210 static int __read_mostly tickincr = 8 << SCHED_TICK_SHIFT;
211 static int __read_mostly realstathz = 127;	/* reset during boot. */
212 static int __read_mostly sched_slice = 10;	/* reset during boot. */
213 static int __read_mostly sched_slice_min = 1;	/* reset during boot. */
214 #ifdef PREEMPTION
215 #ifdef FULL_PREEMPTION
216 static int __read_mostly preempt_thresh = PRI_MAX_IDLE;
217 #else
218 static int __read_mostly preempt_thresh = PRI_MIN_KERN;
219 #endif
220 #else
221 static int __read_mostly preempt_thresh = 0;
222 #endif
223 static int __read_mostly static_boost = PRI_MIN_BATCH;
224 static int __read_mostly sched_idlespins = 10000;
225 static int __read_mostly 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 	int		tdq_id;			/* cpuid. */
251 	struct runq	tdq_realtime;		/* real-time run queue. */
252 	struct runq	tdq_timeshare;		/* timeshare run queue. */
253 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
254 	char		tdq_name[TDQ_NAME_LEN];
255 #ifdef KTR
256 	char		tdq_loadname[TDQ_LOADNAME_LEN];
257 #endif
258 } __aligned(64);
259 
260 /* Idle thread states and config. */
261 #define	TDQ_RUNNING	1
262 #define	TDQ_IDLE	2
263 
264 #ifdef SMP
265 struct cpu_group __read_mostly *cpu_top;		/* CPU topology */
266 
267 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
268 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
269 
270 /*
271  * Run-time tunables.
272  */
273 static int rebalance = 1;
274 static int balance_interval = 128;	/* Default set in sched_initticks(). */
275 static int __read_mostly affinity;
276 static int __read_mostly steal_idle = 1;
277 static int __read_mostly steal_thresh = 2;
278 static int __read_mostly always_steal = 0;
279 static int __read_mostly trysteal_limit = 2;
280 
281 /*
282  * One thread queue per processor.
283  */
284 static struct tdq __read_mostly *balance_tdq;
285 static int balance_ticks;
286 DPCPU_DEFINE_STATIC(struct tdq, tdq);
287 DPCPU_DEFINE_STATIC(uint32_t, randomval);
288 
289 #define	TDQ_SELF()	((struct tdq *)PCPU_GET(sched))
290 #define	TDQ_CPU(x)	(DPCPU_ID_PTR((x), tdq))
291 #define	TDQ_ID(x)	((x)->tdq_id)
292 #else	/* !SMP */
293 static struct tdq	tdq_cpu;
294 
295 #define	TDQ_ID(x)	(0)
296 #define	TDQ_SELF()	(&tdq_cpu)
297 #define	TDQ_CPU(x)	(&tdq_cpu)
298 #endif
299 
300 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
301 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
302 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
303 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
304 #define	TDQ_LOCKPTR(t)		((struct mtx *)(&(t)->tdq_lock))
305 
306 static void sched_priority(struct thread *);
307 static void sched_thread_priority(struct thread *, u_char);
308 static int sched_interact_score(struct thread *);
309 static void sched_interact_update(struct thread *);
310 static void sched_interact_fork(struct thread *);
311 static void sched_pctcpu_update(struct td_sched *, int);
312 
313 /* Operations on per processor queues */
314 static struct thread *tdq_choose(struct tdq *);
315 static void tdq_setup(struct tdq *, int i);
316 static void tdq_load_add(struct tdq *, struct thread *);
317 static void tdq_load_rem(struct tdq *, struct thread *);
318 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
319 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
320 static inline int sched_shouldpreempt(int, int, int);
321 void tdq_print(int cpu);
322 static void runq_print(struct runq *rq);
323 static void tdq_add(struct tdq *, struct thread *, int);
324 #ifdef SMP
325 static struct thread *tdq_move(struct tdq *, struct tdq *);
326 static int tdq_idled(struct tdq *);
327 static void tdq_notify(struct tdq *, struct thread *);
328 static struct thread *tdq_steal(struct tdq *, int);
329 static struct thread *runq_steal(struct runq *, int);
330 static int sched_pickcpu(struct thread *, int);
331 static void sched_balance(void);
332 static int sched_balance_pair(struct tdq *, struct tdq *);
333 static inline struct tdq *sched_setcpu(struct thread *, int, int);
334 static inline void thread_unblock_switch(struct thread *, struct mtx *);
335 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
336 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
337 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
338     struct cpu_group *cg, int indent);
339 #endif
340 
341 static void sched_setup(void *dummy);
342 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
343 
344 static void sched_initticks(void *dummy);
345 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
346     NULL);
347 
348 SDT_PROVIDER_DEFINE(sched);
349 
350 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
351     "struct proc *", "uint8_t");
352 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
353     "struct proc *", "void *");
354 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
355     "struct proc *", "void *", "int");
356 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
357     "struct proc *", "uint8_t", "struct thread *");
358 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
359 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
360     "struct proc *");
361 SDT_PROBE_DEFINE(sched, , , on__cpu);
362 SDT_PROBE_DEFINE(sched, , , remain__cpu);
363 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
364     "struct proc *");
365 
366 /*
367  * Print the threads waiting on a run-queue.
368  */
369 static void
370 runq_print(struct runq *rq)
371 {
372 	struct rqhead *rqh;
373 	struct thread *td;
374 	int pri;
375 	int j;
376 	int i;
377 
378 	for (i = 0; i < RQB_LEN; i++) {
379 		printf("\t\trunq bits %d 0x%zx\n",
380 		    i, rq->rq_status.rqb_bits[i]);
381 		for (j = 0; j < RQB_BPW; j++)
382 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
383 				pri = j + (i << RQB_L2BPW);
384 				rqh = &rq->rq_queues[pri];
385 				TAILQ_FOREACH(td, rqh, td_runq) {
386 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
387 					    td, td->td_name, td->td_priority,
388 					    td->td_rqindex, pri);
389 				}
390 			}
391 	}
392 }
393 
394 /*
395  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
396  */
397 void
398 tdq_print(int cpu)
399 {
400 	struct tdq *tdq;
401 
402 	tdq = TDQ_CPU(cpu);
403 
404 	printf("tdq %d:\n", TDQ_ID(tdq));
405 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
406 	printf("\tLock name:      %s\n", tdq->tdq_name);
407 	printf("\tload:           %d\n", tdq->tdq_load);
408 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
409 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
410 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
411 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
412 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
413 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
414 	printf("\trealtime runq:\n");
415 	runq_print(&tdq->tdq_realtime);
416 	printf("\ttimeshare runq:\n");
417 	runq_print(&tdq->tdq_timeshare);
418 	printf("\tidle runq:\n");
419 	runq_print(&tdq->tdq_idle);
420 }
421 
422 static inline int
423 sched_shouldpreempt(int pri, int cpri, int remote)
424 {
425 	/*
426 	 * If the new priority is not better than the current priority there is
427 	 * nothing to do.
428 	 */
429 	if (pri >= cpri)
430 		return (0);
431 	/*
432 	 * Always preempt idle.
433 	 */
434 	if (cpri >= PRI_MIN_IDLE)
435 		return (1);
436 	/*
437 	 * If preemption is disabled don't preempt others.
438 	 */
439 	if (preempt_thresh == 0)
440 		return (0);
441 	/*
442 	 * Preempt if we exceed the threshold.
443 	 */
444 	if (pri <= preempt_thresh)
445 		return (1);
446 	/*
447 	 * If we're interactive or better and there is non-interactive
448 	 * or worse running preempt only remote processors.
449 	 */
450 	if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
451 		return (1);
452 	return (0);
453 }
454 
455 /*
456  * Add a thread to the actual run-queue.  Keeps transferable counts up to
457  * date with what is actually on the run-queue.  Selects the correct
458  * queue position for timeshare threads.
459  */
460 static __inline void
461 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
462 {
463 	struct td_sched *ts;
464 	u_char pri;
465 
466 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
467 	THREAD_LOCK_ASSERT(td, MA_OWNED);
468 
469 	pri = td->td_priority;
470 	ts = td_get_sched(td);
471 	TD_SET_RUNQ(td);
472 	if (THREAD_CAN_MIGRATE(td)) {
473 		tdq->tdq_transferable++;
474 		ts->ts_flags |= TSF_XFERABLE;
475 	}
476 	if (pri < PRI_MIN_BATCH) {
477 		ts->ts_runq = &tdq->tdq_realtime;
478 	} else if (pri <= PRI_MAX_BATCH) {
479 		ts->ts_runq = &tdq->tdq_timeshare;
480 		KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
481 			("Invalid priority %d on timeshare runq", pri));
482 		/*
483 		 * This queue contains only priorities between MIN and MAX
484 		 * realtime.  Use the whole queue to represent these values.
485 		 */
486 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
487 			pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
488 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
489 			/*
490 			 * This effectively shortens the queue by one so we
491 			 * can have a one slot difference between idx and
492 			 * ridx while we wait for threads to drain.
493 			 */
494 			if (tdq->tdq_ridx != tdq->tdq_idx &&
495 			    pri == tdq->tdq_ridx)
496 				pri = (unsigned char)(pri - 1) % RQ_NQS;
497 		} else
498 			pri = tdq->tdq_ridx;
499 		runq_add_pri(ts->ts_runq, td, pri, flags);
500 		return;
501 	} else
502 		ts->ts_runq = &tdq->tdq_idle;
503 	runq_add(ts->ts_runq, td, flags);
504 }
505 
506 /*
507  * Remove a thread from a run-queue.  This typically happens when a thread
508  * is selected to run.  Running threads are not on the queue and the
509  * transferable count does not reflect them.
510  */
511 static __inline void
512 tdq_runq_rem(struct tdq *tdq, struct thread *td)
513 {
514 	struct td_sched *ts;
515 
516 	ts = td_get_sched(td);
517 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
518 	KASSERT(ts->ts_runq != NULL,
519 	    ("tdq_runq_remove: thread %p null ts_runq", td));
520 	if (ts->ts_flags & TSF_XFERABLE) {
521 		tdq->tdq_transferable--;
522 		ts->ts_flags &= ~TSF_XFERABLE;
523 	}
524 	if (ts->ts_runq == &tdq->tdq_timeshare) {
525 		if (tdq->tdq_idx != tdq->tdq_ridx)
526 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
527 		else
528 			runq_remove_idx(ts->ts_runq, td, NULL);
529 	} else
530 		runq_remove(ts->ts_runq, td);
531 }
532 
533 /*
534  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
535  * for this thread to the referenced thread queue.
536  */
537 static void
538 tdq_load_add(struct tdq *tdq, struct thread *td)
539 {
540 
541 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
542 	THREAD_LOCK_ASSERT(td, MA_OWNED);
543 
544 	tdq->tdq_load++;
545 	if ((td->td_flags & TDF_NOLOAD) == 0)
546 		tdq->tdq_sysload++;
547 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
548 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
549 }
550 
551 /*
552  * Remove the load from a thread that is transitioning to a sleep state or
553  * exiting.
554  */
555 static void
556 tdq_load_rem(struct tdq *tdq, struct thread *td)
557 {
558 
559 	THREAD_LOCK_ASSERT(td, MA_OWNED);
560 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
561 	KASSERT(tdq->tdq_load != 0,
562 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
563 
564 	tdq->tdq_load--;
565 	if ((td->td_flags & TDF_NOLOAD) == 0)
566 		tdq->tdq_sysload--;
567 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
568 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
569 }
570 
571 /*
572  * Bound timeshare latency by decreasing slice size as load increases.  We
573  * consider the maximum latency as the sum of the threads waiting to run
574  * aside from curthread and target no more than sched_slice latency but
575  * no less than sched_slice_min runtime.
576  */
577 static inline int
578 tdq_slice(struct tdq *tdq)
579 {
580 	int load;
581 
582 	/*
583 	 * It is safe to use sys_load here because this is called from
584 	 * contexts where timeshare threads are running and so there
585 	 * cannot be higher priority load in the system.
586 	 */
587 	load = tdq->tdq_sysload - 1;
588 	if (load >= SCHED_SLICE_MIN_DIVISOR)
589 		return (sched_slice_min);
590 	if (load <= 1)
591 		return (sched_slice);
592 	return (sched_slice / load);
593 }
594 
595 /*
596  * Set lowpri to its exact value by searching the run-queue and
597  * evaluating curthread.  curthread may be passed as an optimization.
598  */
599 static void
600 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
601 {
602 	struct thread *td;
603 
604 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
605 	if (ctd == NULL)
606 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
607 	td = tdq_choose(tdq);
608 	if (td == NULL || td->td_priority > ctd->td_priority)
609 		tdq->tdq_lowpri = ctd->td_priority;
610 	else
611 		tdq->tdq_lowpri = td->td_priority;
612 }
613 
614 #ifdef SMP
615 /*
616  * We need some randomness. Implement a classic Linear Congruential
617  * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
618  * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
619  * of the random state (in the low bits of our answer) to keep
620  * the maximum randomness.
621  */
622 static uint32_t
623 sched_random(void)
624 {
625 	uint32_t *rndptr;
626 
627 	rndptr = DPCPU_PTR(randomval);
628 	*rndptr = *rndptr * 69069 + 5;
629 
630 	return (*rndptr >> 16);
631 }
632 
633 struct cpu_search {
634 	cpuset_t cs_mask;
635 	u_int	cs_prefer;
636 	int	cs_pri;		/* Min priority for low. */
637 	int	cs_limit;	/* Max load for low, min load for high. */
638 	int	cs_cpu;
639 	int	cs_load;
640 };
641 
642 #define	CPU_SEARCH_LOWEST	0x1
643 #define	CPU_SEARCH_HIGHEST	0x2
644 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
645 
646 static __always_inline int cpu_search(const struct cpu_group *cg,
647     struct cpu_search *low, struct cpu_search *high, const int match);
648 int __noinline cpu_search_lowest(const struct cpu_group *cg,
649     struct cpu_search *low);
650 int __noinline cpu_search_highest(const struct cpu_group *cg,
651     struct cpu_search *high);
652 int __noinline cpu_search_both(const struct cpu_group *cg,
653     struct cpu_search *low, struct cpu_search *high);
654 
655 /*
656  * Search the tree of cpu_groups for the lowest or highest loaded cpu
657  * according to the match argument.  This routine actually compares the
658  * load on all paths through the tree and finds the least loaded cpu on
659  * the least loaded path, which may differ from the least loaded cpu in
660  * the system.  This balances work among caches and buses.
661  *
662  * This inline is instantiated in three forms below using constants for the
663  * match argument.  It is reduced to the minimum set for each case.  It is
664  * also recursive to the depth of the tree.
665  */
666 static __always_inline int
667 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
668     struct cpu_search *high, const int match)
669 {
670 	struct cpu_search lgroup;
671 	struct cpu_search hgroup;
672 	cpuset_t cpumask;
673 	struct cpu_group *child;
674 	struct tdq *tdq;
675 	int cpu, i, hload, lload, load, total, rnd;
676 
677 	total = 0;
678 	cpumask = cg->cg_mask;
679 	if (match & CPU_SEARCH_LOWEST) {
680 		lload = INT_MAX;
681 		lgroup = *low;
682 	}
683 	if (match & CPU_SEARCH_HIGHEST) {
684 		hload = INT_MIN;
685 		hgroup = *high;
686 	}
687 
688 	/* Iterate through the child CPU groups and then remaining CPUs. */
689 	for (i = cg->cg_children, cpu = mp_maxid; ; ) {
690 		if (i == 0) {
691 #ifdef HAVE_INLINE_FFSL
692 			cpu = CPU_FFS(&cpumask) - 1;
693 #else
694 			while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
695 				cpu--;
696 #endif
697 			if (cpu < 0)
698 				break;
699 			child = NULL;
700 		} else
701 			child = &cg->cg_child[i - 1];
702 
703 		if (match & CPU_SEARCH_LOWEST)
704 			lgroup.cs_cpu = -1;
705 		if (match & CPU_SEARCH_HIGHEST)
706 			hgroup.cs_cpu = -1;
707 		if (child) {			/* Handle child CPU group. */
708 			CPU_NAND(&cpumask, &child->cg_mask);
709 			switch (match) {
710 			case CPU_SEARCH_LOWEST:
711 				load = cpu_search_lowest(child, &lgroup);
712 				break;
713 			case CPU_SEARCH_HIGHEST:
714 				load = cpu_search_highest(child, &hgroup);
715 				break;
716 			case CPU_SEARCH_BOTH:
717 				load = cpu_search_both(child, &lgroup, &hgroup);
718 				break;
719 			}
720 		} else {			/* Handle child CPU. */
721 			CPU_CLR(cpu, &cpumask);
722 			tdq = TDQ_CPU(cpu);
723 			load = tdq->tdq_load * 256;
724 			rnd = sched_random() % 32;
725 			if (match & CPU_SEARCH_LOWEST) {
726 				if (cpu == low->cs_prefer)
727 					load -= 64;
728 				/* If that CPU is allowed and get data. */
729 				if (tdq->tdq_lowpri > lgroup.cs_pri &&
730 				    tdq->tdq_load <= lgroup.cs_limit &&
731 				    CPU_ISSET(cpu, &lgroup.cs_mask)) {
732 					lgroup.cs_cpu = cpu;
733 					lgroup.cs_load = load - rnd;
734 				}
735 			}
736 			if (match & CPU_SEARCH_HIGHEST)
737 				if (tdq->tdq_load >= hgroup.cs_limit &&
738 				    tdq->tdq_transferable &&
739 				    CPU_ISSET(cpu, &hgroup.cs_mask)) {
740 					hgroup.cs_cpu = cpu;
741 					hgroup.cs_load = load - rnd;
742 				}
743 		}
744 		total += load;
745 
746 		/* We have info about child item. Compare it. */
747 		if (match & CPU_SEARCH_LOWEST) {
748 			if (lgroup.cs_cpu >= 0 &&
749 			    (load < lload ||
750 			     (load == lload && lgroup.cs_load < low->cs_load))) {
751 				lload = load;
752 				low->cs_cpu = lgroup.cs_cpu;
753 				low->cs_load = lgroup.cs_load;
754 			}
755 		}
756 		if (match & CPU_SEARCH_HIGHEST)
757 			if (hgroup.cs_cpu >= 0 &&
758 			    (load > hload ||
759 			     (load == hload && hgroup.cs_load > high->cs_load))) {
760 				hload = load;
761 				high->cs_cpu = hgroup.cs_cpu;
762 				high->cs_load = hgroup.cs_load;
763 			}
764 		if (child) {
765 			i--;
766 			if (i == 0 && CPU_EMPTY(&cpumask))
767 				break;
768 		}
769 #ifndef HAVE_INLINE_FFSL
770 		else
771 			cpu--;
772 #endif
773 	}
774 	return (total);
775 }
776 
777 /*
778  * cpu_search instantiations must pass constants to maintain the inline
779  * optimization.
780  */
781 int
782 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
783 {
784 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
785 }
786 
787 int
788 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
789 {
790 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
791 }
792 
793 int
794 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
795     struct cpu_search *high)
796 {
797 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
798 }
799 
800 /*
801  * Find the cpu with the least load via the least loaded path that has a
802  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
803  * acceptable.
804  */
805 static inline int
806 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
807     int prefer)
808 {
809 	struct cpu_search low;
810 
811 	low.cs_cpu = -1;
812 	low.cs_prefer = prefer;
813 	low.cs_mask = mask;
814 	low.cs_pri = pri;
815 	low.cs_limit = maxload;
816 	cpu_search_lowest(cg, &low);
817 	return low.cs_cpu;
818 }
819 
820 /*
821  * Find the cpu with the highest load via the highest loaded path.
822  */
823 static inline int
824 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
825 {
826 	struct cpu_search high;
827 
828 	high.cs_cpu = -1;
829 	high.cs_mask = mask;
830 	high.cs_limit = minload;
831 	cpu_search_highest(cg, &high);
832 	return high.cs_cpu;
833 }
834 
835 static void
836 sched_balance_group(struct cpu_group *cg)
837 {
838 	struct tdq *tdq;
839 	cpuset_t hmask, lmask;
840 	int high, low, anylow;
841 
842 	CPU_FILL(&hmask);
843 	for (;;) {
844 		high = sched_highest(cg, hmask, 2);
845 		/* Stop if there is no more CPU with transferrable threads. */
846 		if (high == -1)
847 			break;
848 		CPU_CLR(high, &hmask);
849 		CPU_COPY(&hmask, &lmask);
850 		/* Stop if there is no more CPU left for low. */
851 		if (CPU_EMPTY(&lmask))
852 			break;
853 		anylow = 1;
854 		tdq = TDQ_CPU(high);
855 nextlow:
856 		low = sched_lowest(cg, lmask, -1, tdq->tdq_load - 1, high);
857 		/* Stop if we looked well and found no less loaded CPU. */
858 		if (anylow && low == -1)
859 			break;
860 		/* Go to next high if we found no less loaded CPU. */
861 		if (low == -1)
862 			continue;
863 		/* Transfer thread from high to low. */
864 		if (sched_balance_pair(tdq, TDQ_CPU(low))) {
865 			/* CPU that got thread can no longer be a donor. */
866 			CPU_CLR(low, &hmask);
867 		} else {
868 			/*
869 			 * If failed, then there is no threads on high
870 			 * that can run on this low. Drop low from low
871 			 * mask and look for different one.
872 			 */
873 			CPU_CLR(low, &lmask);
874 			anylow = 0;
875 			goto nextlow;
876 		}
877 	}
878 }
879 
880 static void
881 sched_balance(void)
882 {
883 	struct tdq *tdq;
884 
885 	balance_ticks = max(balance_interval / 2, 1) +
886 	    (sched_random() % balance_interval);
887 	tdq = TDQ_SELF();
888 	TDQ_UNLOCK(tdq);
889 	sched_balance_group(cpu_top);
890 	TDQ_LOCK(tdq);
891 }
892 
893 /*
894  * Lock two thread queues using their address to maintain lock order.
895  */
896 static void
897 tdq_lock_pair(struct tdq *one, struct tdq *two)
898 {
899 	if (one < two) {
900 		TDQ_LOCK(one);
901 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
902 	} else {
903 		TDQ_LOCK(two);
904 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
905 	}
906 }
907 
908 /*
909  * Unlock two thread queues.  Order is not important here.
910  */
911 static void
912 tdq_unlock_pair(struct tdq *one, struct tdq *two)
913 {
914 	TDQ_UNLOCK(one);
915 	TDQ_UNLOCK(two);
916 }
917 
918 /*
919  * Transfer load between two imbalanced thread queues.
920  */
921 static int
922 sched_balance_pair(struct tdq *high, struct tdq *low)
923 {
924 	struct thread *td;
925 	int cpu;
926 
927 	tdq_lock_pair(high, low);
928 	td = NULL;
929 	/*
930 	 * Transfer a thread from high to low.
931 	 */
932 	if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
933 	    (td = tdq_move(high, low)) != NULL) {
934 		/*
935 		 * In case the target isn't the current cpu notify it of the
936 		 * new load, possibly sending an IPI to force it to reschedule.
937 		 */
938 		cpu = TDQ_ID(low);
939 		if (cpu != PCPU_GET(cpuid))
940 			tdq_notify(low, td);
941 	}
942 	tdq_unlock_pair(high, low);
943 	return (td != NULL);
944 }
945 
946 /*
947  * Move a thread from one thread queue to another.
948  */
949 static struct thread *
950 tdq_move(struct tdq *from, struct tdq *to)
951 {
952 	struct td_sched *ts;
953 	struct thread *td;
954 	struct tdq *tdq;
955 	int cpu;
956 
957 	TDQ_LOCK_ASSERT(from, MA_OWNED);
958 	TDQ_LOCK_ASSERT(to, MA_OWNED);
959 
960 	tdq = from;
961 	cpu = TDQ_ID(to);
962 	td = tdq_steal(tdq, cpu);
963 	if (td == NULL)
964 		return (NULL);
965 	ts = td_get_sched(td);
966 	/*
967 	 * Although the run queue is locked the thread may be blocked.  Lock
968 	 * it to clear this and acquire the run-queue lock.
969 	 */
970 	thread_lock(td);
971 	/* Drop recursive lock on from acquired via thread_lock(). */
972 	TDQ_UNLOCK(from);
973 	sched_rem(td);
974 	ts->ts_cpu = cpu;
975 	td->td_lock = TDQ_LOCKPTR(to);
976 	tdq_add(to, td, SRQ_YIELDING);
977 	return (td);
978 }
979 
980 /*
981  * This tdq has idled.  Try to steal a thread from another cpu and switch
982  * to it.
983  */
984 static int
985 tdq_idled(struct tdq *tdq)
986 {
987 	struct cpu_group *cg;
988 	struct tdq *steal;
989 	cpuset_t mask;
990 	int cpu, switchcnt;
991 
992 	if (smp_started == 0 || steal_idle == 0 || tdq->tdq_cg == NULL)
993 		return (1);
994 	CPU_FILL(&mask);
995 	CPU_CLR(PCPU_GET(cpuid), &mask);
996     restart:
997 	switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
998 	for (cg = tdq->tdq_cg; ; ) {
999 		cpu = sched_highest(cg, mask, steal_thresh);
1000 		/*
1001 		 * We were assigned a thread but not preempted.  Returning
1002 		 * 0 here will cause our caller to switch to it.
1003 		 */
1004 		if (tdq->tdq_load)
1005 			return (0);
1006 		if (cpu == -1) {
1007 			cg = cg->cg_parent;
1008 			if (cg == NULL)
1009 				return (1);
1010 			continue;
1011 		}
1012 		steal = TDQ_CPU(cpu);
1013 		/*
1014 		 * The data returned by sched_highest() is stale and
1015 		 * the chosen CPU no longer has an eligible thread.
1016 		 *
1017 		 * Testing this ahead of tdq_lock_pair() only catches
1018 		 * this situation about 20% of the time on an 8 core
1019 		 * 16 thread Ryzen 7, but it still helps performance.
1020 		 */
1021 		if (steal->tdq_load < steal_thresh ||
1022 		    steal->tdq_transferable == 0)
1023 			goto restart;
1024 		tdq_lock_pair(tdq, steal);
1025 		/*
1026 		 * We were assigned a thread while waiting for the locks.
1027 		 * Switch to it now instead of stealing a thread.
1028 		 */
1029 		if (tdq->tdq_load)
1030 			break;
1031 		/*
1032 		 * The data returned by sched_highest() is stale and
1033 		 * the chosen CPU no longer has an eligible thread, or
1034 		 * we were preempted and the CPU loading info may be out
1035 		 * of date.  The latter is rare.  In either case restart
1036 		 * the search.
1037 		 */
1038 		if (steal->tdq_load < steal_thresh ||
1039 		    steal->tdq_transferable == 0 ||
1040 		    switchcnt != tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt) {
1041 			tdq_unlock_pair(tdq, steal);
1042 			goto restart;
1043 		}
1044 		/*
1045 		 * Steal the thread and switch to it.
1046 		 */
1047 		if (tdq_move(steal, tdq) != NULL)
1048 			break;
1049 		/*
1050 		 * We failed to acquire a thread even though it looked
1051 		 * like one was available.  This could be due to affinity
1052 		 * restrictions or for other reasons.  Loop again after
1053 		 * removing this CPU from the set.  The restart logic
1054 		 * above does not restore this CPU to the set due to the
1055 		 * likelyhood of failing here again.
1056 		 */
1057 		CPU_CLR(cpu, &mask);
1058 		tdq_unlock_pair(tdq, steal);
1059 	}
1060 	TDQ_UNLOCK(steal);
1061 	mi_switch(SW_VOL | SWT_IDLE, NULL);
1062 	thread_unlock(curthread);
1063 	return (0);
1064 }
1065 
1066 /*
1067  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
1068  */
1069 static void
1070 tdq_notify(struct tdq *tdq, struct thread *td)
1071 {
1072 	struct thread *ctd;
1073 	int pri;
1074 	int cpu;
1075 
1076 	if (tdq->tdq_ipipending)
1077 		return;
1078 	cpu = td_get_sched(td)->ts_cpu;
1079 	pri = td->td_priority;
1080 	ctd = pcpu_find(cpu)->pc_curthread;
1081 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1082 		return;
1083 
1084 	/*
1085 	 * Make sure that our caller's earlier update to tdq_load is
1086 	 * globally visible before we read tdq_cpu_idle.  Idle thread
1087 	 * accesses both of them without locks, and the order is important.
1088 	 */
1089 	atomic_thread_fence_seq_cst();
1090 
1091 	if (TD_IS_IDLETHREAD(ctd)) {
1092 		/*
1093 		 * If the MD code has an idle wakeup routine try that before
1094 		 * falling back to IPI.
1095 		 */
1096 		if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1097 			return;
1098 	}
1099 	tdq->tdq_ipipending = 1;
1100 	ipi_cpu(cpu, IPI_PREEMPT);
1101 }
1102 
1103 /*
1104  * Steals load from a timeshare queue.  Honors the rotating queue head
1105  * index.
1106  */
1107 static struct thread *
1108 runq_steal_from(struct runq *rq, int cpu, u_char start)
1109 {
1110 	struct rqbits *rqb;
1111 	struct rqhead *rqh;
1112 	struct thread *td, *first;
1113 	int bit;
1114 	int i;
1115 
1116 	rqb = &rq->rq_status;
1117 	bit = start & (RQB_BPW -1);
1118 	first = NULL;
1119 again:
1120 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1121 		if (rqb->rqb_bits[i] == 0)
1122 			continue;
1123 		if (bit == 0)
1124 			bit = RQB_FFS(rqb->rqb_bits[i]);
1125 		for (; bit < RQB_BPW; bit++) {
1126 			if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1127 				continue;
1128 			rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1129 			TAILQ_FOREACH(td, rqh, td_runq) {
1130 				if (first && THREAD_CAN_MIGRATE(td) &&
1131 				    THREAD_CAN_SCHED(td, cpu))
1132 					return (td);
1133 				first = td;
1134 			}
1135 		}
1136 	}
1137 	if (start != 0) {
1138 		start = 0;
1139 		goto again;
1140 	}
1141 
1142 	if (first && THREAD_CAN_MIGRATE(first) &&
1143 	    THREAD_CAN_SCHED(first, cpu))
1144 		return (first);
1145 	return (NULL);
1146 }
1147 
1148 /*
1149  * Steals load from a standard linear queue.
1150  */
1151 static struct thread *
1152 runq_steal(struct runq *rq, int cpu)
1153 {
1154 	struct rqhead *rqh;
1155 	struct rqbits *rqb;
1156 	struct thread *td;
1157 	int word;
1158 	int bit;
1159 
1160 	rqb = &rq->rq_status;
1161 	for (word = 0; word < RQB_LEN; word++) {
1162 		if (rqb->rqb_bits[word] == 0)
1163 			continue;
1164 		for (bit = 0; bit < RQB_BPW; bit++) {
1165 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1166 				continue;
1167 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1168 			TAILQ_FOREACH(td, rqh, td_runq)
1169 				if (THREAD_CAN_MIGRATE(td) &&
1170 				    THREAD_CAN_SCHED(td, cpu))
1171 					return (td);
1172 		}
1173 	}
1174 	return (NULL);
1175 }
1176 
1177 /*
1178  * Attempt to steal a thread in priority order from a thread queue.
1179  */
1180 static struct thread *
1181 tdq_steal(struct tdq *tdq, int cpu)
1182 {
1183 	struct thread *td;
1184 
1185 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1186 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1187 		return (td);
1188 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1189 	    cpu, tdq->tdq_ridx)) != NULL)
1190 		return (td);
1191 	return (runq_steal(&tdq->tdq_idle, cpu));
1192 }
1193 
1194 /*
1195  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1196  * current lock and returns with the assigned queue locked.
1197  */
1198 static inline struct tdq *
1199 sched_setcpu(struct thread *td, int cpu, int flags)
1200 {
1201 
1202 	struct tdq *tdq;
1203 
1204 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1205 	tdq = TDQ_CPU(cpu);
1206 	td_get_sched(td)->ts_cpu = cpu;
1207 	/*
1208 	 * If the lock matches just return the queue.
1209 	 */
1210 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1211 		return (tdq);
1212 #ifdef notyet
1213 	/*
1214 	 * If the thread isn't running its lockptr is a
1215 	 * turnstile or a sleepqueue.  We can just lock_set without
1216 	 * blocking.
1217 	 */
1218 	if (TD_CAN_RUN(td)) {
1219 		TDQ_LOCK(tdq);
1220 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1221 		return (tdq);
1222 	}
1223 #endif
1224 	/*
1225 	 * The hard case, migration, we need to block the thread first to
1226 	 * prevent order reversals with other cpus locks.
1227 	 */
1228 	spinlock_enter();
1229 	thread_lock_block(td);
1230 	TDQ_LOCK(tdq);
1231 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1232 	spinlock_exit();
1233 	return (tdq);
1234 }
1235 
1236 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1237 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1238 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1239 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1240 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1241 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1242 
1243 static int
1244 sched_pickcpu(struct thread *td, int flags)
1245 {
1246 	struct cpu_group *cg, *ccg;
1247 	struct td_sched *ts;
1248 	struct tdq *tdq;
1249 	cpuset_t mask;
1250 	int cpu, pri, self, intr;
1251 
1252 	self = PCPU_GET(cpuid);
1253 	ts = td_get_sched(td);
1254 	KASSERT(!CPU_ABSENT(ts->ts_cpu), ("sched_pickcpu: Start scheduler on "
1255 	    "absent CPU %d for thread %s.", ts->ts_cpu, td->td_name));
1256 	if (smp_started == 0)
1257 		return (self);
1258 	/*
1259 	 * Don't migrate a running thread from sched_switch().
1260 	 */
1261 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1262 		return (ts->ts_cpu);
1263 	/*
1264 	 * Prefer to run interrupt threads on the processors that generate
1265 	 * the interrupt.
1266 	 */
1267 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1268 	    curthread->td_intr_nesting_level) {
1269 		tdq = TDQ_SELF();
1270 		if (tdq->tdq_lowpri >= PRI_MIN_IDLE) {
1271 			SCHED_STAT_INC(pickcpu_idle_affinity);
1272 			return (self);
1273 		}
1274 		ts->ts_cpu = self;
1275 		intr = 1;
1276 		cg = tdq->tdq_cg;
1277 		goto llc;
1278 	} else {
1279 		intr = 0;
1280 		tdq = TDQ_CPU(ts->ts_cpu);
1281 		cg = tdq->tdq_cg;
1282 	}
1283 	/*
1284 	 * If the thread can run on the last cpu and the affinity has not
1285 	 * expired and it is idle, run it there.
1286 	 */
1287 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1288 	    tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1289 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1290 		if (cg->cg_flags & CG_FLAG_THREAD) {
1291 			/* Check all SMT threads for being idle. */
1292 			for (cpu = CPU_FFS(&cg->cg_mask) - 1; ; cpu++) {
1293 				if (CPU_ISSET(cpu, &cg->cg_mask) &&
1294 				    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1295 					break;
1296 				if (cpu >= mp_maxid) {
1297 					SCHED_STAT_INC(pickcpu_idle_affinity);
1298 					return (ts->ts_cpu);
1299 				}
1300 			}
1301 		} else {
1302 			SCHED_STAT_INC(pickcpu_idle_affinity);
1303 			return (ts->ts_cpu);
1304 		}
1305 	}
1306 llc:
1307 	/*
1308 	 * Search for the last level cache CPU group in the tree.
1309 	 * Skip SMT, identical groups and caches with expired affinity.
1310 	 * Interrupt threads affinity is explicit and never expires.
1311 	 */
1312 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1313 		if (cg->cg_flags & CG_FLAG_THREAD)
1314 			continue;
1315 		if (cg->cg_children == 1 || cg->cg_count == 1)
1316 			continue;
1317 		if (cg->cg_level == CG_SHARE_NONE ||
1318 		    (!intr && !SCHED_AFFINITY(ts, cg->cg_level)))
1319 			continue;
1320 		ccg = cg;
1321 	}
1322 	/* Found LLC shared by all CPUs, so do a global search. */
1323 	if (ccg == cpu_top)
1324 		ccg = NULL;
1325 	cpu = -1;
1326 	mask = td->td_cpuset->cs_mask;
1327 	pri = td->td_priority;
1328 	/*
1329 	 * Try hard to keep interrupts within found LLC.  Search the LLC for
1330 	 * the least loaded CPU we can run now.  For NUMA systems it should
1331 	 * be within target domain, and it also reduces scheduling overhead.
1332 	 */
1333 	if (ccg != NULL && intr) {
1334 		cpu = sched_lowest(ccg, mask, pri, INT_MAX, ts->ts_cpu);
1335 		if (cpu >= 0)
1336 			SCHED_STAT_INC(pickcpu_intrbind);
1337 	} else
1338 	/* Search the LLC for the least loaded idle CPU we can run now. */
1339 	if (ccg != NULL) {
1340 		cpu = sched_lowest(ccg, mask, max(pri, PRI_MAX_TIMESHARE),
1341 		    INT_MAX, ts->ts_cpu);
1342 		if (cpu >= 0)
1343 			SCHED_STAT_INC(pickcpu_affinity);
1344 	}
1345 	/* Search globally for the least loaded CPU we can run now. */
1346 	if (cpu < 0) {
1347 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1348 		if (cpu >= 0)
1349 			SCHED_STAT_INC(pickcpu_lowest);
1350 	}
1351 	/* Search globally for the least loaded CPU. */
1352 	if (cpu < 0) {
1353 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1354 		if (cpu >= 0)
1355 			SCHED_STAT_INC(pickcpu_lowest);
1356 	}
1357 	KASSERT(cpu >= 0, ("sched_pickcpu: Failed to find a cpu."));
1358 	KASSERT(!CPU_ABSENT(cpu), ("sched_pickcpu: Picked absent CPU %d.", cpu));
1359 	/*
1360 	 * Compare the lowest loaded cpu to current cpu.
1361 	 */
1362 	tdq = TDQ_CPU(cpu);
1363 	if (THREAD_CAN_SCHED(td, self) && TDQ_SELF()->tdq_lowpri > pri &&
1364 	    tdq->tdq_lowpri < PRI_MIN_IDLE &&
1365 	    TDQ_SELF()->tdq_load <= tdq->tdq_load + 1) {
1366 		SCHED_STAT_INC(pickcpu_local);
1367 		cpu = self;
1368 	}
1369 	if (cpu != ts->ts_cpu)
1370 		SCHED_STAT_INC(pickcpu_migration);
1371 	return (cpu);
1372 }
1373 #endif
1374 
1375 /*
1376  * Pick the highest priority task we have and return it.
1377  */
1378 static struct thread *
1379 tdq_choose(struct tdq *tdq)
1380 {
1381 	struct thread *td;
1382 
1383 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1384 	td = runq_choose(&tdq->tdq_realtime);
1385 	if (td != NULL)
1386 		return (td);
1387 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1388 	if (td != NULL) {
1389 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1390 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1391 		    td->td_priority));
1392 		return (td);
1393 	}
1394 	td = runq_choose(&tdq->tdq_idle);
1395 	if (td != NULL) {
1396 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1397 		    ("tdq_choose: Invalid priority on idle queue %d",
1398 		    td->td_priority));
1399 		return (td);
1400 	}
1401 
1402 	return (NULL);
1403 }
1404 
1405 /*
1406  * Initialize a thread queue.
1407  */
1408 static void
1409 tdq_setup(struct tdq *tdq, int id)
1410 {
1411 
1412 	if (bootverbose)
1413 		printf("ULE: setup cpu %d\n", id);
1414 	runq_init(&tdq->tdq_realtime);
1415 	runq_init(&tdq->tdq_timeshare);
1416 	runq_init(&tdq->tdq_idle);
1417 	tdq->tdq_id = id;
1418 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1419 	    "sched lock %d", (int)TDQ_ID(tdq));
1420 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1421 	    MTX_SPIN | MTX_RECURSE);
1422 #ifdef KTR
1423 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1424 	    "CPU %d load", (int)TDQ_ID(tdq));
1425 #endif
1426 }
1427 
1428 #ifdef SMP
1429 static void
1430 sched_setup_smp(void)
1431 {
1432 	struct tdq *tdq;
1433 	int i;
1434 
1435 	cpu_top = smp_topo();
1436 	CPU_FOREACH(i) {
1437 		tdq = DPCPU_ID_PTR(i, tdq);
1438 		tdq_setup(tdq, i);
1439 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1440 		if (tdq->tdq_cg == NULL)
1441 			panic("Can't find cpu group for %d\n", i);
1442 	}
1443 	PCPU_SET(sched, DPCPU_PTR(tdq));
1444 	balance_tdq = TDQ_SELF();
1445 }
1446 #endif
1447 
1448 /*
1449  * Setup the thread queues and initialize the topology based on MD
1450  * information.
1451  */
1452 static void
1453 sched_setup(void *dummy)
1454 {
1455 	struct tdq *tdq;
1456 
1457 #ifdef SMP
1458 	sched_setup_smp();
1459 #else
1460 	tdq_setup(TDQ_SELF(), 0);
1461 #endif
1462 	tdq = TDQ_SELF();
1463 
1464 	/* Add thread0's load since it's running. */
1465 	TDQ_LOCK(tdq);
1466 	thread0.td_lock = TDQ_LOCKPTR(tdq);
1467 	tdq_load_add(tdq, &thread0);
1468 	tdq->tdq_lowpri = thread0.td_priority;
1469 	TDQ_UNLOCK(tdq);
1470 }
1471 
1472 /*
1473  * This routine determines time constants after stathz and hz are setup.
1474  */
1475 /* ARGSUSED */
1476 static void
1477 sched_initticks(void *dummy)
1478 {
1479 	int incr;
1480 
1481 	realstathz = stathz ? stathz : hz;
1482 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1483 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1484 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1485 	    realstathz);
1486 
1487 	/*
1488 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1489 	 * hz not being evenly divisible by stathz on all platforms.
1490 	 */
1491 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1492 	/*
1493 	 * This does not work for values of stathz that are more than
1494 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1495 	 */
1496 	if (incr == 0)
1497 		incr = 1;
1498 	tickincr = incr;
1499 #ifdef SMP
1500 	/*
1501 	 * Set the default balance interval now that we know
1502 	 * what realstathz is.
1503 	 */
1504 	balance_interval = realstathz;
1505 	balance_ticks = balance_interval;
1506 	affinity = SCHED_AFFINITY_DEFAULT;
1507 #endif
1508 	if (sched_idlespinthresh < 0)
1509 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1510 }
1511 
1512 
1513 /*
1514  * This is the core of the interactivity algorithm.  Determines a score based
1515  * on past behavior.  It is the ratio of sleep time to run time scaled to
1516  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1517  * differs from the cpu usage because it does not account for time spent
1518  * waiting on a run-queue.  Would be prettier if we had floating point.
1519  *
1520  * When a thread's sleep time is greater than its run time the
1521  * calculation is:
1522  *
1523  *                           scaling factor
1524  * interactivity score =  ---------------------
1525  *                        sleep time / run time
1526  *
1527  *
1528  * When a thread's run time is greater than its sleep time the
1529  * calculation is:
1530  *
1531  *                           scaling factor
1532  * interactivity score =  ---------------------    + scaling factor
1533  *                        run time / sleep time
1534  */
1535 static int
1536 sched_interact_score(struct thread *td)
1537 {
1538 	struct td_sched *ts;
1539 	int div;
1540 
1541 	ts = td_get_sched(td);
1542 	/*
1543 	 * The score is only needed if this is likely to be an interactive
1544 	 * task.  Don't go through the expense of computing it if there's
1545 	 * no chance.
1546 	 */
1547 	if (sched_interact <= SCHED_INTERACT_HALF &&
1548 		ts->ts_runtime >= ts->ts_slptime)
1549 			return (SCHED_INTERACT_HALF);
1550 
1551 	if (ts->ts_runtime > ts->ts_slptime) {
1552 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1553 		return (SCHED_INTERACT_HALF +
1554 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1555 	}
1556 	if (ts->ts_slptime > ts->ts_runtime) {
1557 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1558 		return (ts->ts_runtime / div);
1559 	}
1560 	/* runtime == slptime */
1561 	if (ts->ts_runtime)
1562 		return (SCHED_INTERACT_HALF);
1563 
1564 	/*
1565 	 * This can happen if slptime and runtime are 0.
1566 	 */
1567 	return (0);
1568 
1569 }
1570 
1571 /*
1572  * Scale the scheduling priority according to the "interactivity" of this
1573  * process.
1574  */
1575 static void
1576 sched_priority(struct thread *td)
1577 {
1578 	int score;
1579 	int pri;
1580 
1581 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1582 		return;
1583 	/*
1584 	 * If the score is interactive we place the thread in the realtime
1585 	 * queue with a priority that is less than kernel and interrupt
1586 	 * priorities.  These threads are not subject to nice restrictions.
1587 	 *
1588 	 * Scores greater than this are placed on the normal timeshare queue
1589 	 * where the priority is partially decided by the most recent cpu
1590 	 * utilization and the rest is decided by nice value.
1591 	 *
1592 	 * The nice value of the process has a linear effect on the calculated
1593 	 * score.  Negative nice values make it easier for a thread to be
1594 	 * considered interactive.
1595 	 */
1596 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1597 	if (score < sched_interact) {
1598 		pri = PRI_MIN_INTERACT;
1599 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1600 		    sched_interact) * score;
1601 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1602 		    ("sched_priority: invalid interactive priority %d score %d",
1603 		    pri, score));
1604 	} else {
1605 		pri = SCHED_PRI_MIN;
1606 		if (td_get_sched(td)->ts_ticks)
1607 			pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1608 			    SCHED_PRI_RANGE - 1);
1609 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1610 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1611 		    ("sched_priority: invalid priority %d: nice %d, "
1612 		    "ticks %d ftick %d ltick %d tick pri %d",
1613 		    pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1614 		    td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1615 		    SCHED_PRI_TICKS(td_get_sched(td))));
1616 	}
1617 	sched_user_prio(td, pri);
1618 
1619 	return;
1620 }
1621 
1622 /*
1623  * This routine enforces a maximum limit on the amount of scheduling history
1624  * kept.  It is called after either the slptime or runtime is adjusted.  This
1625  * function is ugly due to integer math.
1626  */
1627 static void
1628 sched_interact_update(struct thread *td)
1629 {
1630 	struct td_sched *ts;
1631 	u_int sum;
1632 
1633 	ts = td_get_sched(td);
1634 	sum = ts->ts_runtime + ts->ts_slptime;
1635 	if (sum < SCHED_SLP_RUN_MAX)
1636 		return;
1637 	/*
1638 	 * This only happens from two places:
1639 	 * 1) We have added an unusual amount of run time from fork_exit.
1640 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1641 	 */
1642 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1643 		if (ts->ts_runtime > ts->ts_slptime) {
1644 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1645 			ts->ts_slptime = 1;
1646 		} else {
1647 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1648 			ts->ts_runtime = 1;
1649 		}
1650 		return;
1651 	}
1652 	/*
1653 	 * If we have exceeded by more than 1/5th then the algorithm below
1654 	 * will not bring us back into range.  Dividing by two here forces
1655 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1656 	 */
1657 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1658 		ts->ts_runtime /= 2;
1659 		ts->ts_slptime /= 2;
1660 		return;
1661 	}
1662 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1663 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1664 }
1665 
1666 /*
1667  * Scale back the interactivity history when a child thread is created.  The
1668  * history is inherited from the parent but the thread may behave totally
1669  * differently.  For example, a shell spawning a compiler process.  We want
1670  * to learn that the compiler is behaving badly very quickly.
1671  */
1672 static void
1673 sched_interact_fork(struct thread *td)
1674 {
1675 	struct td_sched *ts;
1676 	int ratio;
1677 	int sum;
1678 
1679 	ts = td_get_sched(td);
1680 	sum = ts->ts_runtime + ts->ts_slptime;
1681 	if (sum > SCHED_SLP_RUN_FORK) {
1682 		ratio = sum / SCHED_SLP_RUN_FORK;
1683 		ts->ts_runtime /= ratio;
1684 		ts->ts_slptime /= ratio;
1685 	}
1686 }
1687 
1688 /*
1689  * Called from proc0_init() to setup the scheduler fields.
1690  */
1691 void
1692 schedinit(void)
1693 {
1694 	struct td_sched *ts0;
1695 
1696 	/*
1697 	 * Set up the scheduler specific parts of thread0.
1698 	 */
1699 	ts0 = td_get_sched(&thread0);
1700 	ts0->ts_ltick = ticks;
1701 	ts0->ts_ftick = ticks;
1702 	ts0->ts_slice = 0;
1703 	ts0->ts_cpu = curcpu;	/* set valid CPU number */
1704 }
1705 
1706 /*
1707  * This is only somewhat accurate since given many processes of the same
1708  * priority they will switch when their slices run out, which will be
1709  * at most sched_slice stathz ticks.
1710  */
1711 int
1712 sched_rr_interval(void)
1713 {
1714 
1715 	/* Convert sched_slice from stathz to hz. */
1716 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1717 }
1718 
1719 /*
1720  * Update the percent cpu tracking information when it is requested or
1721  * the total history exceeds the maximum.  We keep a sliding history of
1722  * tick counts that slowly decays.  This is less precise than the 4BSD
1723  * mechanism since it happens with less regular and frequent events.
1724  */
1725 static void
1726 sched_pctcpu_update(struct td_sched *ts, int run)
1727 {
1728 	int t = ticks;
1729 
1730 	/*
1731 	 * The signed difference may be negative if the thread hasn't run for
1732 	 * over half of the ticks rollover period.
1733 	 */
1734 	if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) {
1735 		ts->ts_ticks = 0;
1736 		ts->ts_ftick = t - SCHED_TICK_TARG;
1737 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1738 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1739 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1740 		ts->ts_ftick = t - SCHED_TICK_TARG;
1741 	}
1742 	if (run)
1743 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1744 	ts->ts_ltick = t;
1745 }
1746 
1747 /*
1748  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1749  * if necessary.  This is the back-end for several priority related
1750  * functions.
1751  */
1752 static void
1753 sched_thread_priority(struct thread *td, u_char prio)
1754 {
1755 	struct td_sched *ts;
1756 	struct tdq *tdq;
1757 	int oldpri;
1758 
1759 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1760 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1761 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1762 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1763 	if (td != curthread && prio < td->td_priority) {
1764 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1765 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1766 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1767 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1768 		    curthread);
1769 	}
1770 	ts = td_get_sched(td);
1771 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1772 	if (td->td_priority == prio)
1773 		return;
1774 	/*
1775 	 * If the priority has been elevated due to priority
1776 	 * propagation, we may have to move ourselves to a new
1777 	 * queue.  This could be optimized to not re-add in some
1778 	 * cases.
1779 	 */
1780 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1781 		sched_rem(td);
1782 		td->td_priority = prio;
1783 		sched_add(td, SRQ_BORROWING);
1784 		return;
1785 	}
1786 	/*
1787 	 * If the thread is currently running we may have to adjust the lowpri
1788 	 * information so other cpus are aware of our current priority.
1789 	 */
1790 	if (TD_IS_RUNNING(td)) {
1791 		tdq = TDQ_CPU(ts->ts_cpu);
1792 		oldpri = td->td_priority;
1793 		td->td_priority = prio;
1794 		if (prio < tdq->tdq_lowpri)
1795 			tdq->tdq_lowpri = prio;
1796 		else if (tdq->tdq_lowpri == oldpri)
1797 			tdq_setlowpri(tdq, td);
1798 		return;
1799 	}
1800 	td->td_priority = prio;
1801 }
1802 
1803 /*
1804  * Update a thread's priority when it is lent another thread's
1805  * priority.
1806  */
1807 void
1808 sched_lend_prio(struct thread *td, u_char prio)
1809 {
1810 
1811 	td->td_flags |= TDF_BORROWING;
1812 	sched_thread_priority(td, prio);
1813 }
1814 
1815 /*
1816  * Restore a thread's priority when priority propagation is
1817  * over.  The prio argument is the minimum priority the thread
1818  * needs to have to satisfy other possible priority lending
1819  * requests.  If the thread's regular priority is less
1820  * important than prio, the thread will keep a priority boost
1821  * of prio.
1822  */
1823 void
1824 sched_unlend_prio(struct thread *td, u_char prio)
1825 {
1826 	u_char base_pri;
1827 
1828 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1829 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1830 		base_pri = td->td_user_pri;
1831 	else
1832 		base_pri = td->td_base_pri;
1833 	if (prio >= base_pri) {
1834 		td->td_flags &= ~TDF_BORROWING;
1835 		sched_thread_priority(td, base_pri);
1836 	} else
1837 		sched_lend_prio(td, prio);
1838 }
1839 
1840 /*
1841  * Standard entry for setting the priority to an absolute value.
1842  */
1843 void
1844 sched_prio(struct thread *td, u_char prio)
1845 {
1846 	u_char oldprio;
1847 
1848 	/* First, update the base priority. */
1849 	td->td_base_pri = prio;
1850 
1851 	/*
1852 	 * If the thread is borrowing another thread's priority, don't
1853 	 * ever lower the priority.
1854 	 */
1855 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1856 		return;
1857 
1858 	/* Change the real priority. */
1859 	oldprio = td->td_priority;
1860 	sched_thread_priority(td, prio);
1861 
1862 	/*
1863 	 * If the thread is on a turnstile, then let the turnstile update
1864 	 * its state.
1865 	 */
1866 	if (TD_ON_LOCK(td) && oldprio != prio)
1867 		turnstile_adjust(td, oldprio);
1868 }
1869 
1870 /*
1871  * Set the base user priority, does not effect current running priority.
1872  */
1873 void
1874 sched_user_prio(struct thread *td, u_char prio)
1875 {
1876 
1877 	td->td_base_user_pri = prio;
1878 	if (td->td_lend_user_pri <= prio)
1879 		return;
1880 	td->td_user_pri = prio;
1881 }
1882 
1883 void
1884 sched_lend_user_prio(struct thread *td, u_char prio)
1885 {
1886 
1887 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1888 	td->td_lend_user_pri = prio;
1889 	td->td_user_pri = min(prio, td->td_base_user_pri);
1890 	if (td->td_priority > td->td_user_pri)
1891 		sched_prio(td, td->td_user_pri);
1892 	else if (td->td_priority != td->td_user_pri)
1893 		td->td_flags |= TDF_NEEDRESCHED;
1894 }
1895 
1896 /*
1897  * Like the above but first check if there is anything to do.
1898  */
1899 void
1900 sched_lend_user_prio_cond(struct thread *td, u_char prio)
1901 {
1902 
1903 	if (td->td_lend_user_pri != prio)
1904 		goto lend;
1905 	if (td->td_user_pri != min(prio, td->td_base_user_pri))
1906 		goto lend;
1907 	if (td->td_priority >= td->td_user_pri)
1908 		goto lend;
1909 	return;
1910 
1911 lend:
1912 	thread_lock(td);
1913 	sched_lend_user_prio(td, prio);
1914 	thread_unlock(td);
1915 }
1916 
1917 #ifdef SMP
1918 /*
1919  * This tdq is about to idle.  Try to steal a thread from another CPU before
1920  * choosing the idle thread.
1921  */
1922 static void
1923 tdq_trysteal(struct tdq *tdq)
1924 {
1925 	struct cpu_group *cg;
1926 	struct tdq *steal;
1927 	cpuset_t mask;
1928 	int cpu, i;
1929 
1930 	if (smp_started == 0 || trysteal_limit == 0 || tdq->tdq_cg == NULL)
1931 		return;
1932 	CPU_FILL(&mask);
1933 	CPU_CLR(PCPU_GET(cpuid), &mask);
1934 	/* We don't want to be preempted while we're iterating. */
1935 	spinlock_enter();
1936 	TDQ_UNLOCK(tdq);
1937 	for (i = 1, cg = tdq->tdq_cg; ; ) {
1938 		cpu = sched_highest(cg, mask, steal_thresh);
1939 		/*
1940 		 * If a thread was added while interrupts were disabled don't
1941 		 * steal one here.
1942 		 */
1943 		if (tdq->tdq_load > 0) {
1944 			TDQ_LOCK(tdq);
1945 			break;
1946 		}
1947 		if (cpu == -1) {
1948 			i++;
1949 			cg = cg->cg_parent;
1950 			if (cg == NULL || i > trysteal_limit) {
1951 				TDQ_LOCK(tdq);
1952 				break;
1953 			}
1954 			continue;
1955 		}
1956 		steal = TDQ_CPU(cpu);
1957 		/*
1958 		 * The data returned by sched_highest() is stale and
1959                  * the chosen CPU no longer has an eligible thread.
1960 		 */
1961 		if (steal->tdq_load < steal_thresh ||
1962 		    steal->tdq_transferable == 0)
1963 			continue;
1964 		tdq_lock_pair(tdq, steal);
1965 		/*
1966 		 * If we get to this point, unconditonally exit the loop
1967 		 * to bound the time spent in the critcal section.
1968 		 *
1969 		 * If a thread was added while interrupts were disabled don't
1970 		 * steal one here.
1971 		 */
1972 		if (tdq->tdq_load > 0) {
1973 			TDQ_UNLOCK(steal);
1974 			break;
1975 		}
1976 		/*
1977 		 * The data returned by sched_highest() is stale and
1978                  * the chosen CPU no longer has an eligible thread.
1979 		 */
1980 		if (steal->tdq_load < steal_thresh ||
1981 		    steal->tdq_transferable == 0) {
1982 			TDQ_UNLOCK(steal);
1983 			break;
1984 		}
1985 		/*
1986 		 * If we fail to acquire one due to affinity restrictions,
1987 		 * bail out and let the idle thread to a more complete search
1988 		 * outside of a critical section.
1989 		 */
1990 		if (tdq_move(steal, tdq) == NULL) {
1991 			TDQ_UNLOCK(steal);
1992 			break;
1993 		}
1994 		TDQ_UNLOCK(steal);
1995 		break;
1996 	}
1997 	spinlock_exit();
1998 }
1999 #endif
2000 
2001 /*
2002  * Handle migration from sched_switch().  This happens only for
2003  * cpu binding.
2004  */
2005 static struct mtx *
2006 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
2007 {
2008 	struct tdq *tdn;
2009 
2010 	KASSERT(!CPU_ABSENT(td_get_sched(td)->ts_cpu), ("sched_switch_migrate: "
2011 	    "thread %s queued on absent CPU %d.", td->td_name,
2012 	    td_get_sched(td)->ts_cpu));
2013 	tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
2014 #ifdef SMP
2015 	tdq_load_rem(tdq, td);
2016 	/*
2017 	 * Do the lock dance required to avoid LOR.  We grab an extra
2018 	 * spinlock nesting to prevent preemption while we're
2019 	 * not holding either run-queue lock.
2020 	 */
2021 	spinlock_enter();
2022 	thread_lock_block(td);	/* This releases the lock on tdq. */
2023 
2024 	/*
2025 	 * Acquire both run-queue locks before placing the thread on the new
2026 	 * run-queue to avoid deadlocks created by placing a thread with a
2027 	 * blocked lock on the run-queue of a remote processor.  The deadlock
2028 	 * occurs when a third processor attempts to lock the two queues in
2029 	 * question while the target processor is spinning with its own
2030 	 * run-queue lock held while waiting for the blocked lock to clear.
2031 	 */
2032 	tdq_lock_pair(tdn, tdq);
2033 	tdq_add(tdn, td, flags);
2034 	tdq_notify(tdn, td);
2035 	TDQ_UNLOCK(tdn);
2036 	spinlock_exit();
2037 #endif
2038 	return (TDQ_LOCKPTR(tdn));
2039 }
2040 
2041 /*
2042  * Variadic version of thread_lock_unblock() that does not assume td_lock
2043  * is blocked.
2044  */
2045 static inline void
2046 thread_unblock_switch(struct thread *td, struct mtx *mtx)
2047 {
2048 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
2049 	    (uintptr_t)mtx);
2050 }
2051 
2052 /*
2053  * Switch threads.  This function has to handle threads coming in while
2054  * blocked for some reason, running, or idle.  It also must deal with
2055  * migrating a thread from one queue to another as running threads may
2056  * be assigned elsewhere via binding.
2057  */
2058 void
2059 sched_switch(struct thread *td, struct thread *newtd, int flags)
2060 {
2061 	struct tdq *tdq;
2062 	struct td_sched *ts;
2063 	struct mtx *mtx;
2064 	int srqflag;
2065 	int cpuid, preempted;
2066 
2067 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2068 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
2069 
2070 	cpuid = PCPU_GET(cpuid);
2071 	tdq = TDQ_SELF();
2072 	ts = td_get_sched(td);
2073 	mtx = td->td_lock;
2074 	sched_pctcpu_update(ts, 1);
2075 	ts->ts_rltick = ticks;
2076 	td->td_lastcpu = td->td_oncpu;
2077 	td->td_oncpu = NOCPU;
2078 	preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
2079 	    (flags & SW_PREEMPT) != 0;
2080 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
2081 	td->td_owepreempt = 0;
2082 	if (!TD_IS_IDLETHREAD(td))
2083 		tdq->tdq_switchcnt++;
2084 	/*
2085 	 * The lock pointer in an idle thread should never change.  Reset it
2086 	 * to CAN_RUN as well.
2087 	 */
2088 	if (TD_IS_IDLETHREAD(td)) {
2089 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2090 		TD_SET_CAN_RUN(td);
2091 	} else if (TD_IS_RUNNING(td)) {
2092 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2093 		srqflag = preempted ?
2094 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
2095 		    SRQ_OURSELF|SRQ_YIELDING;
2096 #ifdef SMP
2097 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
2098 			ts->ts_cpu = sched_pickcpu(td, 0);
2099 #endif
2100 		if (ts->ts_cpu == cpuid)
2101 			tdq_runq_add(tdq, td, srqflag);
2102 		else {
2103 			KASSERT(THREAD_CAN_MIGRATE(td) ||
2104 			    (ts->ts_flags & TSF_BOUND) != 0,
2105 			    ("Thread %p shouldn't migrate", td));
2106 			mtx = sched_switch_migrate(tdq, td, srqflag);
2107 		}
2108 	} else {
2109 		/* This thread must be going to sleep. */
2110 		TDQ_LOCK(tdq);
2111 		mtx = thread_lock_block(td);
2112 		tdq_load_rem(tdq, td);
2113 #ifdef SMP
2114 		if (tdq->tdq_load == 0)
2115 			tdq_trysteal(tdq);
2116 #endif
2117 	}
2118 
2119 #if (KTR_COMPILE & KTR_SCHED) != 0
2120 	if (TD_IS_IDLETHREAD(td))
2121 		KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
2122 		    "prio:%d", td->td_priority);
2123 	else
2124 		KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
2125 		    "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
2126 		    "lockname:\"%s\"", td->td_lockname);
2127 #endif
2128 
2129 	/*
2130 	 * We enter here with the thread blocked and assigned to the
2131 	 * appropriate cpu run-queue or sleep-queue and with the current
2132 	 * thread-queue locked.
2133 	 */
2134 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2135 	newtd = choosethread();
2136 	/*
2137 	 * Call the MD code to switch contexts if necessary.
2138 	 */
2139 	if (td != newtd) {
2140 #ifdef	HWPMC_HOOKS
2141 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2142 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
2143 #endif
2144 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
2145 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2146 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2147 		sched_pctcpu_update(td_get_sched(newtd), 0);
2148 
2149 #ifdef KDTRACE_HOOKS
2150 		/*
2151 		 * If DTrace has set the active vtime enum to anything
2152 		 * other than INACTIVE (0), then it should have set the
2153 		 * function to call.
2154 		 */
2155 		if (dtrace_vtime_active)
2156 			(*dtrace_vtime_switch_func)(newtd);
2157 #endif
2158 
2159 		cpu_switch(td, newtd, mtx);
2160 		/*
2161 		 * We may return from cpu_switch on a different cpu.  However,
2162 		 * we always return with td_lock pointing to the current cpu's
2163 		 * run queue lock.
2164 		 */
2165 		cpuid = PCPU_GET(cpuid);
2166 		tdq = TDQ_SELF();
2167 		lock_profile_obtain_lock_success(
2168 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2169 
2170 		SDT_PROBE0(sched, , , on__cpu);
2171 #ifdef	HWPMC_HOOKS
2172 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2173 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
2174 #endif
2175 	} else {
2176 		thread_unblock_switch(td, mtx);
2177 		SDT_PROBE0(sched, , , remain__cpu);
2178 	}
2179 
2180 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2181 	    "prio:%d", td->td_priority);
2182 
2183 	/*
2184 	 * Assert that all went well and return.
2185 	 */
2186 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
2187 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2188 	td->td_oncpu = cpuid;
2189 }
2190 
2191 /*
2192  * Adjust thread priorities as a result of a nice request.
2193  */
2194 void
2195 sched_nice(struct proc *p, int nice)
2196 {
2197 	struct thread *td;
2198 
2199 	PROC_LOCK_ASSERT(p, MA_OWNED);
2200 
2201 	p->p_nice = nice;
2202 	FOREACH_THREAD_IN_PROC(p, td) {
2203 		thread_lock(td);
2204 		sched_priority(td);
2205 		sched_prio(td, td->td_base_user_pri);
2206 		thread_unlock(td);
2207 	}
2208 }
2209 
2210 /*
2211  * Record the sleep time for the interactivity scorer.
2212  */
2213 void
2214 sched_sleep(struct thread *td, int prio)
2215 {
2216 
2217 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2218 
2219 	td->td_slptick = ticks;
2220 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2221 		td->td_flags |= TDF_CANSWAP;
2222 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2223 		return;
2224 	if (static_boost == 1 && prio)
2225 		sched_prio(td, prio);
2226 	else if (static_boost && td->td_priority > static_boost)
2227 		sched_prio(td, static_boost);
2228 }
2229 
2230 /*
2231  * Schedule a thread to resume execution and record how long it voluntarily
2232  * slept.  We also update the pctcpu, interactivity, and priority.
2233  */
2234 void
2235 sched_wakeup(struct thread *td)
2236 {
2237 	struct td_sched *ts;
2238 	int slptick;
2239 
2240 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2241 	ts = td_get_sched(td);
2242 	td->td_flags &= ~TDF_CANSWAP;
2243 	/*
2244 	 * If we slept for more than a tick update our interactivity and
2245 	 * priority.
2246 	 */
2247 	slptick = td->td_slptick;
2248 	td->td_slptick = 0;
2249 	if (slptick && slptick != ticks) {
2250 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2251 		sched_interact_update(td);
2252 		sched_pctcpu_update(ts, 0);
2253 	}
2254 	/*
2255 	 * Reset the slice value since we slept and advanced the round-robin.
2256 	 */
2257 	ts->ts_slice = 0;
2258 	sched_add(td, SRQ_BORING);
2259 }
2260 
2261 /*
2262  * Penalize the parent for creating a new child and initialize the child's
2263  * priority.
2264  */
2265 void
2266 sched_fork(struct thread *td, struct thread *child)
2267 {
2268 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2269 	sched_pctcpu_update(td_get_sched(td), 1);
2270 	sched_fork_thread(td, child);
2271 	/*
2272 	 * Penalize the parent and child for forking.
2273 	 */
2274 	sched_interact_fork(child);
2275 	sched_priority(child);
2276 	td_get_sched(td)->ts_runtime += tickincr;
2277 	sched_interact_update(td);
2278 	sched_priority(td);
2279 }
2280 
2281 /*
2282  * Fork a new thread, may be within the same process.
2283  */
2284 void
2285 sched_fork_thread(struct thread *td, struct thread *child)
2286 {
2287 	struct td_sched *ts;
2288 	struct td_sched *ts2;
2289 	struct tdq *tdq;
2290 
2291 	tdq = TDQ_SELF();
2292 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2293 	/*
2294 	 * Initialize child.
2295 	 */
2296 	ts = td_get_sched(td);
2297 	ts2 = td_get_sched(child);
2298 	child->td_oncpu = NOCPU;
2299 	child->td_lastcpu = NOCPU;
2300 	child->td_lock = TDQ_LOCKPTR(tdq);
2301 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2302 	child->td_domain.dr_policy = td->td_cpuset->cs_domain;
2303 	ts2->ts_cpu = ts->ts_cpu;
2304 	ts2->ts_flags = 0;
2305 	/*
2306 	 * Grab our parents cpu estimation information.
2307 	 */
2308 	ts2->ts_ticks = ts->ts_ticks;
2309 	ts2->ts_ltick = ts->ts_ltick;
2310 	ts2->ts_ftick = ts->ts_ftick;
2311 	/*
2312 	 * Do not inherit any borrowed priority from the parent.
2313 	 */
2314 	child->td_priority = child->td_base_pri;
2315 	/*
2316 	 * And update interactivity score.
2317 	 */
2318 	ts2->ts_slptime = ts->ts_slptime;
2319 	ts2->ts_runtime = ts->ts_runtime;
2320 	/* Attempt to quickly learn interactivity. */
2321 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2322 #ifdef KTR
2323 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2324 #endif
2325 }
2326 
2327 /*
2328  * Adjust the priority class of a thread.
2329  */
2330 void
2331 sched_class(struct thread *td, int class)
2332 {
2333 
2334 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2335 	if (td->td_pri_class == class)
2336 		return;
2337 	td->td_pri_class = class;
2338 }
2339 
2340 /*
2341  * Return some of the child's priority and interactivity to the parent.
2342  */
2343 void
2344 sched_exit(struct proc *p, struct thread *child)
2345 {
2346 	struct thread *td;
2347 
2348 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2349 	    "prio:%d", child->td_priority);
2350 	PROC_LOCK_ASSERT(p, MA_OWNED);
2351 	td = FIRST_THREAD_IN_PROC(p);
2352 	sched_exit_thread(td, child);
2353 }
2354 
2355 /*
2356  * Penalize another thread for the time spent on this one.  This helps to
2357  * worsen the priority and interactivity of processes which schedule batch
2358  * jobs such as make.  This has little effect on the make process itself but
2359  * causes new processes spawned by it to receive worse scores immediately.
2360  */
2361 void
2362 sched_exit_thread(struct thread *td, struct thread *child)
2363 {
2364 
2365 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2366 	    "prio:%d", child->td_priority);
2367 	/*
2368 	 * Give the child's runtime to the parent without returning the
2369 	 * sleep time as a penalty to the parent.  This causes shells that
2370 	 * launch expensive things to mark their children as expensive.
2371 	 */
2372 	thread_lock(td);
2373 	td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2374 	sched_interact_update(td);
2375 	sched_priority(td);
2376 	thread_unlock(td);
2377 }
2378 
2379 void
2380 sched_preempt(struct thread *td)
2381 {
2382 	struct tdq *tdq;
2383 
2384 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2385 
2386 	thread_lock(td);
2387 	tdq = TDQ_SELF();
2388 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2389 	tdq->tdq_ipipending = 0;
2390 	if (td->td_priority > tdq->tdq_lowpri) {
2391 		int flags;
2392 
2393 		flags = SW_INVOL | SW_PREEMPT;
2394 		if (td->td_critnest > 1)
2395 			td->td_owepreempt = 1;
2396 		else if (TD_IS_IDLETHREAD(td))
2397 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2398 		else
2399 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2400 	}
2401 	thread_unlock(td);
2402 }
2403 
2404 /*
2405  * Fix priorities on return to user-space.  Priorities may be elevated due
2406  * to static priorities in msleep() or similar.
2407  */
2408 void
2409 sched_userret_slowpath(struct thread *td)
2410 {
2411 
2412 	thread_lock(td);
2413 	td->td_priority = td->td_user_pri;
2414 	td->td_base_pri = td->td_user_pri;
2415 	tdq_setlowpri(TDQ_SELF(), td);
2416 	thread_unlock(td);
2417 }
2418 
2419 /*
2420  * Handle a stathz tick.  This is really only relevant for timeshare
2421  * threads.
2422  */
2423 void
2424 sched_clock(struct thread *td)
2425 {
2426 	struct tdq *tdq;
2427 	struct td_sched *ts;
2428 
2429 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2430 	tdq = TDQ_SELF();
2431 #ifdef SMP
2432 	/*
2433 	 * We run the long term load balancer infrequently on the first cpu.
2434 	 */
2435 	if (balance_tdq == tdq && smp_started != 0 && rebalance != 0) {
2436 		if (balance_ticks && --balance_ticks == 0)
2437 			sched_balance();
2438 	}
2439 #endif
2440 	/*
2441 	 * Save the old switch count so we have a record of the last ticks
2442 	 * activity.   Initialize the new switch count based on our load.
2443 	 * If there is some activity seed it to reflect that.
2444 	 */
2445 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2446 	tdq->tdq_switchcnt = tdq->tdq_load;
2447 	/*
2448 	 * Advance the insert index once for each tick to ensure that all
2449 	 * threads get a chance to run.
2450 	 */
2451 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2452 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2453 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2454 			tdq->tdq_ridx = tdq->tdq_idx;
2455 	}
2456 	ts = td_get_sched(td);
2457 	sched_pctcpu_update(ts, 1);
2458 	if (td->td_pri_class & PRI_FIFO_BIT)
2459 		return;
2460 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2461 		/*
2462 		 * We used a tick; charge it to the thread so
2463 		 * that we can compute our interactivity.
2464 		 */
2465 		td_get_sched(td)->ts_runtime += tickincr;
2466 		sched_interact_update(td);
2467 		sched_priority(td);
2468 	}
2469 
2470 	/*
2471 	 * Force a context switch if the current thread has used up a full
2472 	 * time slice (default is 100ms).
2473 	 */
2474 	if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2475 		ts->ts_slice = 0;
2476 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2477 	}
2478 }
2479 
2480 u_int
2481 sched_estcpu(struct thread *td __unused)
2482 {
2483 
2484 	return (0);
2485 }
2486 
2487 /*
2488  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2489  * cooperative idle threads.
2490  */
2491 int
2492 sched_runnable(void)
2493 {
2494 	struct tdq *tdq;
2495 	int load;
2496 
2497 	load = 1;
2498 
2499 	tdq = TDQ_SELF();
2500 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2501 		if (tdq->tdq_load > 0)
2502 			goto out;
2503 	} else
2504 		if (tdq->tdq_load - 1 > 0)
2505 			goto out;
2506 	load = 0;
2507 out:
2508 	return (load);
2509 }
2510 
2511 /*
2512  * Choose the highest priority thread to run.  The thread is removed from
2513  * the run-queue while running however the load remains.  For SMP we set
2514  * the tdq in the global idle bitmask if it idles here.
2515  */
2516 struct thread *
2517 sched_choose(void)
2518 {
2519 	struct thread *td;
2520 	struct tdq *tdq;
2521 
2522 	tdq = TDQ_SELF();
2523 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2524 	td = tdq_choose(tdq);
2525 	if (td) {
2526 		tdq_runq_rem(tdq, td);
2527 		tdq->tdq_lowpri = td->td_priority;
2528 		return (td);
2529 	}
2530 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2531 	return (PCPU_GET(idlethread));
2532 }
2533 
2534 /*
2535  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2536  * we always request it once we exit a critical section.
2537  */
2538 static inline void
2539 sched_setpreempt(struct thread *td)
2540 {
2541 	struct thread *ctd;
2542 	int cpri;
2543 	int pri;
2544 
2545 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2546 
2547 	ctd = curthread;
2548 	pri = td->td_priority;
2549 	cpri = ctd->td_priority;
2550 	if (pri < cpri)
2551 		ctd->td_flags |= TDF_NEEDRESCHED;
2552 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2553 		return;
2554 	if (!sched_shouldpreempt(pri, cpri, 0))
2555 		return;
2556 	ctd->td_owepreempt = 1;
2557 }
2558 
2559 /*
2560  * Add a thread to a thread queue.  Select the appropriate runq and add the
2561  * thread to it.  This is the internal function called when the tdq is
2562  * predetermined.
2563  */
2564 void
2565 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2566 {
2567 
2568 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2569 	KASSERT((td->td_inhibitors == 0),
2570 	    ("sched_add: trying to run inhibited thread"));
2571 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2572 	    ("sched_add: bad thread state"));
2573 	KASSERT(td->td_flags & TDF_INMEM,
2574 	    ("sched_add: thread swapped out"));
2575 
2576 	if (td->td_priority < tdq->tdq_lowpri)
2577 		tdq->tdq_lowpri = td->td_priority;
2578 	tdq_runq_add(tdq, td, flags);
2579 	tdq_load_add(tdq, td);
2580 }
2581 
2582 /*
2583  * Select the target thread queue and add a thread to it.  Request
2584  * preemption or IPI a remote processor if required.
2585  */
2586 void
2587 sched_add(struct thread *td, int flags)
2588 {
2589 	struct tdq *tdq;
2590 #ifdef SMP
2591 	int cpu;
2592 #endif
2593 
2594 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2595 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2596 	    sched_tdname(curthread));
2597 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2598 	    KTR_ATTR_LINKED, sched_tdname(td));
2599 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2600 	    flags & SRQ_PREEMPTED);
2601 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2602 	/*
2603 	 * Recalculate the priority before we select the target cpu or
2604 	 * run-queue.
2605 	 */
2606 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2607 		sched_priority(td);
2608 #ifdef SMP
2609 	/*
2610 	 * Pick the destination cpu and if it isn't ours transfer to the
2611 	 * target cpu.
2612 	 */
2613 	cpu = sched_pickcpu(td, flags);
2614 	tdq = sched_setcpu(td, cpu, flags);
2615 	tdq_add(tdq, td, flags);
2616 	if (cpu != PCPU_GET(cpuid)) {
2617 		tdq_notify(tdq, td);
2618 		return;
2619 	}
2620 #else
2621 	tdq = TDQ_SELF();
2622 	TDQ_LOCK(tdq);
2623 	/*
2624 	 * Now that the thread is moving to the run-queue, set the lock
2625 	 * to the scheduler's lock.
2626 	 */
2627 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2628 	tdq_add(tdq, td, flags);
2629 #endif
2630 	if (!(flags & SRQ_YIELDING))
2631 		sched_setpreempt(td);
2632 }
2633 
2634 /*
2635  * Remove a thread from a run-queue without running it.  This is used
2636  * when we're stealing a thread from a remote queue.  Otherwise all threads
2637  * exit by calling sched_exit_thread() and sched_throw() themselves.
2638  */
2639 void
2640 sched_rem(struct thread *td)
2641 {
2642 	struct tdq *tdq;
2643 
2644 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2645 	    "prio:%d", td->td_priority);
2646 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2647 	tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2648 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2649 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2650 	KASSERT(TD_ON_RUNQ(td),
2651 	    ("sched_rem: thread not on run queue"));
2652 	tdq_runq_rem(tdq, td);
2653 	tdq_load_rem(tdq, td);
2654 	TD_SET_CAN_RUN(td);
2655 	if (td->td_priority == tdq->tdq_lowpri)
2656 		tdq_setlowpri(tdq, NULL);
2657 }
2658 
2659 /*
2660  * Fetch cpu utilization information.  Updates on demand.
2661  */
2662 fixpt_t
2663 sched_pctcpu(struct thread *td)
2664 {
2665 	fixpt_t pctcpu;
2666 	struct td_sched *ts;
2667 
2668 	pctcpu = 0;
2669 	ts = td_get_sched(td);
2670 
2671 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2672 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2673 	if (ts->ts_ticks) {
2674 		int rtick;
2675 
2676 		/* How many rtick per second ? */
2677 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2678 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2679 	}
2680 
2681 	return (pctcpu);
2682 }
2683 
2684 /*
2685  * Enforce affinity settings for a thread.  Called after adjustments to
2686  * cpumask.
2687  */
2688 void
2689 sched_affinity(struct thread *td)
2690 {
2691 #ifdef SMP
2692 	struct td_sched *ts;
2693 
2694 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2695 	ts = td_get_sched(td);
2696 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2697 		return;
2698 	if (TD_ON_RUNQ(td)) {
2699 		sched_rem(td);
2700 		sched_add(td, SRQ_BORING);
2701 		return;
2702 	}
2703 	if (!TD_IS_RUNNING(td))
2704 		return;
2705 	/*
2706 	 * Force a switch before returning to userspace.  If the
2707 	 * target thread is not running locally send an ipi to force
2708 	 * the issue.
2709 	 */
2710 	td->td_flags |= TDF_NEEDRESCHED;
2711 	if (td != curthread)
2712 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2713 #endif
2714 }
2715 
2716 /*
2717  * Bind a thread to a target cpu.
2718  */
2719 void
2720 sched_bind(struct thread *td, int cpu)
2721 {
2722 	struct td_sched *ts;
2723 
2724 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2725 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2726 	ts = td_get_sched(td);
2727 	if (ts->ts_flags & TSF_BOUND)
2728 		sched_unbind(td);
2729 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2730 	ts->ts_flags |= TSF_BOUND;
2731 	sched_pin();
2732 	if (PCPU_GET(cpuid) == cpu)
2733 		return;
2734 	ts->ts_cpu = cpu;
2735 	/* When we return from mi_switch we'll be on the correct cpu. */
2736 	mi_switch(SW_VOL, NULL);
2737 }
2738 
2739 /*
2740  * Release a bound thread.
2741  */
2742 void
2743 sched_unbind(struct thread *td)
2744 {
2745 	struct td_sched *ts;
2746 
2747 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2748 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2749 	ts = td_get_sched(td);
2750 	if ((ts->ts_flags & TSF_BOUND) == 0)
2751 		return;
2752 	ts->ts_flags &= ~TSF_BOUND;
2753 	sched_unpin();
2754 }
2755 
2756 int
2757 sched_is_bound(struct thread *td)
2758 {
2759 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2760 	return (td_get_sched(td)->ts_flags & TSF_BOUND);
2761 }
2762 
2763 /*
2764  * Basic yield call.
2765  */
2766 void
2767 sched_relinquish(struct thread *td)
2768 {
2769 	thread_lock(td);
2770 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2771 	thread_unlock(td);
2772 }
2773 
2774 /*
2775  * Return the total system load.
2776  */
2777 int
2778 sched_load(void)
2779 {
2780 #ifdef SMP
2781 	int total;
2782 	int i;
2783 
2784 	total = 0;
2785 	CPU_FOREACH(i)
2786 		total += TDQ_CPU(i)->tdq_sysload;
2787 	return (total);
2788 #else
2789 	return (TDQ_SELF()->tdq_sysload);
2790 #endif
2791 }
2792 
2793 int
2794 sched_sizeof_proc(void)
2795 {
2796 	return (sizeof(struct proc));
2797 }
2798 
2799 int
2800 sched_sizeof_thread(void)
2801 {
2802 	return (sizeof(struct thread) + sizeof(struct td_sched));
2803 }
2804 
2805 #ifdef SMP
2806 #define	TDQ_IDLESPIN(tdq)						\
2807     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2808 #else
2809 #define	TDQ_IDLESPIN(tdq)	1
2810 #endif
2811 
2812 /*
2813  * The actual idle process.
2814  */
2815 void
2816 sched_idletd(void *dummy)
2817 {
2818 	struct thread *td;
2819 	struct tdq *tdq;
2820 	int oldswitchcnt, switchcnt;
2821 	int i;
2822 
2823 	mtx_assert(&Giant, MA_NOTOWNED);
2824 	td = curthread;
2825 	tdq = TDQ_SELF();
2826 	THREAD_NO_SLEEPING();
2827 	oldswitchcnt = -1;
2828 	for (;;) {
2829 		if (tdq->tdq_load) {
2830 			thread_lock(td);
2831 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2832 			thread_unlock(td);
2833 		}
2834 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2835 #ifdef SMP
2836 		if (always_steal || switchcnt != oldswitchcnt) {
2837 			oldswitchcnt = switchcnt;
2838 			if (tdq_idled(tdq) == 0)
2839 				continue;
2840 		}
2841 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2842 #else
2843 		oldswitchcnt = switchcnt;
2844 #endif
2845 		/*
2846 		 * If we're switching very frequently, spin while checking
2847 		 * for load rather than entering a low power state that
2848 		 * may require an IPI.  However, don't do any busy
2849 		 * loops while on SMT machines as this simply steals
2850 		 * cycles from cores doing useful work.
2851 		 */
2852 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2853 			for (i = 0; i < sched_idlespins; i++) {
2854 				if (tdq->tdq_load)
2855 					break;
2856 				cpu_spinwait();
2857 			}
2858 		}
2859 
2860 		/* If there was context switch during spin, restart it. */
2861 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2862 		if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2863 			continue;
2864 
2865 		/* Run main MD idle handler. */
2866 		tdq->tdq_cpu_idle = 1;
2867 		/*
2868 		 * Make sure that tdq_cpu_idle update is globally visible
2869 		 * before cpu_idle() read tdq_load.  The order is important
2870 		 * to avoid race with tdq_notify.
2871 		 */
2872 		atomic_thread_fence_seq_cst();
2873 		/*
2874 		 * Checking for again after the fence picks up assigned
2875 		 * threads often enough to make it worthwhile to do so in
2876 		 * order to avoid calling cpu_idle().
2877 		 */
2878 		if (tdq->tdq_load != 0) {
2879 			tdq->tdq_cpu_idle = 0;
2880 			continue;
2881 		}
2882 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2883 		tdq->tdq_cpu_idle = 0;
2884 
2885 		/*
2886 		 * Account thread-less hardware interrupts and
2887 		 * other wakeup reasons equal to context switches.
2888 		 */
2889 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2890 		if (switchcnt != oldswitchcnt)
2891 			continue;
2892 		tdq->tdq_switchcnt++;
2893 		oldswitchcnt++;
2894 	}
2895 }
2896 
2897 /*
2898  * A CPU is entering for the first time or a thread is exiting.
2899  */
2900 void
2901 sched_throw(struct thread *td)
2902 {
2903 	struct thread *newtd;
2904 	struct tdq *tdq;
2905 
2906 	if (td == NULL) {
2907 #ifdef SMP
2908 		PCPU_SET(sched, DPCPU_PTR(tdq));
2909 #endif
2910 		/* Correct spinlock nesting and acquire the correct lock. */
2911 		tdq = TDQ_SELF();
2912 		TDQ_LOCK(tdq);
2913 		spinlock_exit();
2914 		PCPU_SET(switchtime, cpu_ticks());
2915 		PCPU_SET(switchticks, ticks);
2916 		PCPU_GET(idlethread)->td_lock = TDQ_LOCKPTR(tdq);
2917 	} else {
2918 		tdq = TDQ_SELF();
2919 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2920 		tdq_load_rem(tdq, td);
2921 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2922 		td->td_lastcpu = td->td_oncpu;
2923 		td->td_oncpu = NOCPU;
2924 	}
2925 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2926 	newtd = choosethread();
2927 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2928 	cpu_throw(td, newtd);		/* doesn't return */
2929 }
2930 
2931 /*
2932  * This is called from fork_exit().  Just acquire the correct locks and
2933  * let fork do the rest of the work.
2934  */
2935 void
2936 sched_fork_exit(struct thread *td)
2937 {
2938 	struct tdq *tdq;
2939 	int cpuid;
2940 
2941 	/*
2942 	 * Finish setting up thread glue so that it begins execution in a
2943 	 * non-nested critical section with the scheduler lock held.
2944 	 */
2945 	cpuid = PCPU_GET(cpuid);
2946 	tdq = TDQ_SELF();
2947 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2948 	td->td_oncpu = cpuid;
2949 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2950 	lock_profile_obtain_lock_success(
2951 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2952 
2953 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2954 	    "prio:%d", td->td_priority);
2955 	SDT_PROBE0(sched, , , on__cpu);
2956 }
2957 
2958 /*
2959  * Create on first use to catch odd startup conditons.
2960  */
2961 char *
2962 sched_tdname(struct thread *td)
2963 {
2964 #ifdef KTR
2965 	struct td_sched *ts;
2966 
2967 	ts = td_get_sched(td);
2968 	if (ts->ts_name[0] == '\0')
2969 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2970 		    "%s tid %d", td->td_name, td->td_tid);
2971 	return (ts->ts_name);
2972 #else
2973 	return (td->td_name);
2974 #endif
2975 }
2976 
2977 #ifdef KTR
2978 void
2979 sched_clear_tdname(struct thread *td)
2980 {
2981 	struct td_sched *ts;
2982 
2983 	ts = td_get_sched(td);
2984 	ts->ts_name[0] = '\0';
2985 }
2986 #endif
2987 
2988 #ifdef SMP
2989 
2990 /*
2991  * Build the CPU topology dump string. Is recursively called to collect
2992  * the topology tree.
2993  */
2994 static int
2995 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2996     int indent)
2997 {
2998 	char cpusetbuf[CPUSETBUFSIZ];
2999 	int i, first;
3000 
3001 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
3002 	    "", 1 + indent / 2, cg->cg_level);
3003 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
3004 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
3005 	first = TRUE;
3006 	for (i = 0; i < MAXCPU; i++) {
3007 		if (CPU_ISSET(i, &cg->cg_mask)) {
3008 			if (!first)
3009 				sbuf_printf(sb, ", ");
3010 			else
3011 				first = FALSE;
3012 			sbuf_printf(sb, "%d", i);
3013 		}
3014 	}
3015 	sbuf_printf(sb, "</cpu>\n");
3016 
3017 	if (cg->cg_flags != 0) {
3018 		sbuf_printf(sb, "%*s <flags>", indent, "");
3019 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
3020 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
3021 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
3022 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
3023 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
3024 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
3025 		sbuf_printf(sb, "</flags>\n");
3026 	}
3027 
3028 	if (cg->cg_children > 0) {
3029 		sbuf_printf(sb, "%*s <children>\n", indent, "");
3030 		for (i = 0; i < cg->cg_children; i++)
3031 			sysctl_kern_sched_topology_spec_internal(sb,
3032 			    &cg->cg_child[i], indent+2);
3033 		sbuf_printf(sb, "%*s </children>\n", indent, "");
3034 	}
3035 	sbuf_printf(sb, "%*s</group>\n", indent, "");
3036 	return (0);
3037 }
3038 
3039 /*
3040  * Sysctl handler for retrieving topology dump. It's a wrapper for
3041  * the recursive sysctl_kern_smp_topology_spec_internal().
3042  */
3043 static int
3044 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
3045 {
3046 	struct sbuf *topo;
3047 	int err;
3048 
3049 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
3050 
3051 	topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
3052 	if (topo == NULL)
3053 		return (ENOMEM);
3054 
3055 	sbuf_printf(topo, "<groups>\n");
3056 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
3057 	sbuf_printf(topo, "</groups>\n");
3058 
3059 	if (err == 0) {
3060 		err = sbuf_finish(topo);
3061 	}
3062 	sbuf_delete(topo);
3063 	return (err);
3064 }
3065 
3066 #endif
3067 
3068 static int
3069 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
3070 {
3071 	int error, new_val, period;
3072 
3073 	period = 1000000 / realstathz;
3074 	new_val = period * sched_slice;
3075 	error = sysctl_handle_int(oidp, &new_val, 0, req);
3076 	if (error != 0 || req->newptr == NULL)
3077 		return (error);
3078 	if (new_val <= 0)
3079 		return (EINVAL);
3080 	sched_slice = imax(1, (new_val + period / 2) / period);
3081 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
3082 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
3083 	    realstathz);
3084 	return (0);
3085 }
3086 
3087 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
3088 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
3089     "Scheduler name");
3090 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
3091     NULL, 0, sysctl_kern_quantum, "I",
3092     "Quantum for timeshare threads in microseconds");
3093 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
3094     "Quantum for timeshare threads in stathz ticks");
3095 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
3096     "Interactivity score threshold");
3097 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
3098     &preempt_thresh, 0,
3099     "Maximal (lowest) priority for preemption");
3100 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
3101     "Assign static kernel priorities to sleeping threads");
3102 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
3103     "Number of times idle thread will spin waiting for new work");
3104 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
3105     &sched_idlespinthresh, 0,
3106     "Threshold before we will permit idle thread spinning");
3107 #ifdef SMP
3108 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
3109     "Number of hz ticks to keep thread affinity for");
3110 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
3111     "Enables the long-term load balancer");
3112 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
3113     &balance_interval, 0,
3114     "Average period in stathz ticks to run the long-term balancer");
3115 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
3116     "Attempts to steal work from other cores before idling");
3117 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
3118     "Minimum load on remote CPU before we'll steal");
3119 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit,
3120     0, "Topological distance limit for stealing threads in sched_switch()");
3121 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0,
3122     "Always run the stealer from the idle thread");
3123 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
3124     CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
3125     "XML dump of detected CPU topology");
3126 #endif
3127 
3128 /* ps compat.  All cpu percentages from ULE are weighted. */
3129 static int ccpu = 0;
3130 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
3131