xref: /freebsd/sys/kern/sched_ule.c (revision 8b238f4126d32df3e70056bc32536b7248ebffa0)
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_owepreempt;		/* Remote preemption 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_owepreempt)
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 
1100 	/*
1101 	 * The run queues have been updated, so any switch on the remote CPU
1102 	 * will satisfy the preemption request.
1103 	 */
1104 	tdq->tdq_owepreempt = 1;
1105 	ipi_cpu(cpu, IPI_PREEMPT);
1106 }
1107 
1108 /*
1109  * Steals load from a timeshare queue.  Honors the rotating queue head
1110  * index.
1111  */
1112 static struct thread *
1113 runq_steal_from(struct runq *rq, int cpu, u_char start)
1114 {
1115 	struct rqbits *rqb;
1116 	struct rqhead *rqh;
1117 	struct thread *td, *first;
1118 	int bit;
1119 	int i;
1120 
1121 	rqb = &rq->rq_status;
1122 	bit = start & (RQB_BPW -1);
1123 	first = NULL;
1124 again:
1125 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1126 		if (rqb->rqb_bits[i] == 0)
1127 			continue;
1128 		if (bit == 0)
1129 			bit = RQB_FFS(rqb->rqb_bits[i]);
1130 		for (; bit < RQB_BPW; bit++) {
1131 			if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1132 				continue;
1133 			rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1134 			TAILQ_FOREACH(td, rqh, td_runq) {
1135 				if (first && THREAD_CAN_MIGRATE(td) &&
1136 				    THREAD_CAN_SCHED(td, cpu))
1137 					return (td);
1138 				first = td;
1139 			}
1140 		}
1141 	}
1142 	if (start != 0) {
1143 		start = 0;
1144 		goto again;
1145 	}
1146 
1147 	if (first && THREAD_CAN_MIGRATE(first) &&
1148 	    THREAD_CAN_SCHED(first, cpu))
1149 		return (first);
1150 	return (NULL);
1151 }
1152 
1153 /*
1154  * Steals load from a standard linear queue.
1155  */
1156 static struct thread *
1157 runq_steal(struct runq *rq, int cpu)
1158 {
1159 	struct rqhead *rqh;
1160 	struct rqbits *rqb;
1161 	struct thread *td;
1162 	int word;
1163 	int bit;
1164 
1165 	rqb = &rq->rq_status;
1166 	for (word = 0; word < RQB_LEN; word++) {
1167 		if (rqb->rqb_bits[word] == 0)
1168 			continue;
1169 		for (bit = 0; bit < RQB_BPW; bit++) {
1170 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1171 				continue;
1172 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1173 			TAILQ_FOREACH(td, rqh, td_runq)
1174 				if (THREAD_CAN_MIGRATE(td) &&
1175 				    THREAD_CAN_SCHED(td, cpu))
1176 					return (td);
1177 		}
1178 	}
1179 	return (NULL);
1180 }
1181 
1182 /*
1183  * Attempt to steal a thread in priority order from a thread queue.
1184  */
1185 static struct thread *
1186 tdq_steal(struct tdq *tdq, int cpu)
1187 {
1188 	struct thread *td;
1189 
1190 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1191 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1192 		return (td);
1193 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1194 	    cpu, tdq->tdq_ridx)) != NULL)
1195 		return (td);
1196 	return (runq_steal(&tdq->tdq_idle, cpu));
1197 }
1198 
1199 /*
1200  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1201  * current lock and returns with the assigned queue locked.
1202  */
1203 static inline struct tdq *
1204 sched_setcpu(struct thread *td, int cpu, int flags)
1205 {
1206 
1207 	struct tdq *tdq;
1208 
1209 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1210 	tdq = TDQ_CPU(cpu);
1211 	td_get_sched(td)->ts_cpu = cpu;
1212 	/*
1213 	 * If the lock matches just return the queue.
1214 	 */
1215 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1216 		return (tdq);
1217 #ifdef notyet
1218 	/*
1219 	 * If the thread isn't running its lockptr is a
1220 	 * turnstile or a sleepqueue.  We can just lock_set without
1221 	 * blocking.
1222 	 */
1223 	if (TD_CAN_RUN(td)) {
1224 		TDQ_LOCK(tdq);
1225 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1226 		return (tdq);
1227 	}
1228 #endif
1229 	/*
1230 	 * The hard case, migration, we need to block the thread first to
1231 	 * prevent order reversals with other cpus locks.
1232 	 */
1233 	spinlock_enter();
1234 	thread_lock_block(td);
1235 	TDQ_LOCK(tdq);
1236 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1237 	spinlock_exit();
1238 	return (tdq);
1239 }
1240 
1241 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1242 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1243 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1244 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1245 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1246 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1247 
1248 static int
1249 sched_pickcpu(struct thread *td, int flags)
1250 {
1251 	struct cpu_group *cg, *ccg;
1252 	struct td_sched *ts;
1253 	struct tdq *tdq;
1254 	cpuset_t mask;
1255 	int cpu, pri, self, intr;
1256 
1257 	self = PCPU_GET(cpuid);
1258 	ts = td_get_sched(td);
1259 	KASSERT(!CPU_ABSENT(ts->ts_cpu), ("sched_pickcpu: Start scheduler on "
1260 	    "absent CPU %d for thread %s.", ts->ts_cpu, td->td_name));
1261 	if (smp_started == 0)
1262 		return (self);
1263 	/*
1264 	 * Don't migrate a running thread from sched_switch().
1265 	 */
1266 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1267 		return (ts->ts_cpu);
1268 	/*
1269 	 * Prefer to run interrupt threads on the processors that generate
1270 	 * the interrupt.
1271 	 */
1272 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1273 	    curthread->td_intr_nesting_level) {
1274 		tdq = TDQ_SELF();
1275 		if (tdq->tdq_lowpri >= PRI_MIN_IDLE) {
1276 			SCHED_STAT_INC(pickcpu_idle_affinity);
1277 			return (self);
1278 		}
1279 		ts->ts_cpu = self;
1280 		intr = 1;
1281 		cg = tdq->tdq_cg;
1282 		goto llc;
1283 	} else {
1284 		intr = 0;
1285 		tdq = TDQ_CPU(ts->ts_cpu);
1286 		cg = tdq->tdq_cg;
1287 	}
1288 	/*
1289 	 * If the thread can run on the last cpu and the affinity has not
1290 	 * expired and it is idle, run it there.
1291 	 */
1292 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1293 	    tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1294 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1295 		if (cg->cg_flags & CG_FLAG_THREAD) {
1296 			/* Check all SMT threads for being idle. */
1297 			for (cpu = CPU_FFS(&cg->cg_mask) - 1; ; cpu++) {
1298 				if (CPU_ISSET(cpu, &cg->cg_mask) &&
1299 				    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1300 					break;
1301 				if (cpu >= mp_maxid) {
1302 					SCHED_STAT_INC(pickcpu_idle_affinity);
1303 					return (ts->ts_cpu);
1304 				}
1305 			}
1306 		} else {
1307 			SCHED_STAT_INC(pickcpu_idle_affinity);
1308 			return (ts->ts_cpu);
1309 		}
1310 	}
1311 llc:
1312 	/*
1313 	 * Search for the last level cache CPU group in the tree.
1314 	 * Skip SMT, identical groups and caches with expired affinity.
1315 	 * Interrupt threads affinity is explicit and never expires.
1316 	 */
1317 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1318 		if (cg->cg_flags & CG_FLAG_THREAD)
1319 			continue;
1320 		if (cg->cg_children == 1 || cg->cg_count == 1)
1321 			continue;
1322 		if (cg->cg_level == CG_SHARE_NONE ||
1323 		    (!intr && !SCHED_AFFINITY(ts, cg->cg_level)))
1324 			continue;
1325 		ccg = cg;
1326 	}
1327 	/* Found LLC shared by all CPUs, so do a global search. */
1328 	if (ccg == cpu_top)
1329 		ccg = NULL;
1330 	cpu = -1;
1331 	mask = td->td_cpuset->cs_mask;
1332 	pri = td->td_priority;
1333 	/*
1334 	 * Try hard to keep interrupts within found LLC.  Search the LLC for
1335 	 * the least loaded CPU we can run now.  For NUMA systems it should
1336 	 * be within target domain, and it also reduces scheduling overhead.
1337 	 */
1338 	if (ccg != NULL && intr) {
1339 		cpu = sched_lowest(ccg, mask, pri, INT_MAX, ts->ts_cpu);
1340 		if (cpu >= 0)
1341 			SCHED_STAT_INC(pickcpu_intrbind);
1342 	} else
1343 	/* Search the LLC for the least loaded idle CPU we can run now. */
1344 	if (ccg != NULL) {
1345 		cpu = sched_lowest(ccg, mask, max(pri, PRI_MAX_TIMESHARE),
1346 		    INT_MAX, ts->ts_cpu);
1347 		if (cpu >= 0)
1348 			SCHED_STAT_INC(pickcpu_affinity);
1349 	}
1350 	/* Search globally for the least loaded CPU we can run now. */
1351 	if (cpu < 0) {
1352 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1353 		if (cpu >= 0)
1354 			SCHED_STAT_INC(pickcpu_lowest);
1355 	}
1356 	/* Search globally for the least loaded CPU. */
1357 	if (cpu < 0) {
1358 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1359 		if (cpu >= 0)
1360 			SCHED_STAT_INC(pickcpu_lowest);
1361 	}
1362 	KASSERT(cpu >= 0, ("sched_pickcpu: Failed to find a cpu."));
1363 	KASSERT(!CPU_ABSENT(cpu), ("sched_pickcpu: Picked absent CPU %d.", cpu));
1364 	/*
1365 	 * Compare the lowest loaded cpu to current cpu.
1366 	 */
1367 	tdq = TDQ_CPU(cpu);
1368 	if (THREAD_CAN_SCHED(td, self) && TDQ_SELF()->tdq_lowpri > pri &&
1369 	    tdq->tdq_lowpri < PRI_MIN_IDLE &&
1370 	    TDQ_SELF()->tdq_load <= tdq->tdq_load + 1) {
1371 		SCHED_STAT_INC(pickcpu_local);
1372 		cpu = self;
1373 	}
1374 	if (cpu != ts->ts_cpu)
1375 		SCHED_STAT_INC(pickcpu_migration);
1376 	return (cpu);
1377 }
1378 #endif
1379 
1380 /*
1381  * Pick the highest priority task we have and return it.
1382  */
1383 static struct thread *
1384 tdq_choose(struct tdq *tdq)
1385 {
1386 	struct thread *td;
1387 
1388 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1389 	td = runq_choose(&tdq->tdq_realtime);
1390 	if (td != NULL)
1391 		return (td);
1392 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1393 	if (td != NULL) {
1394 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1395 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1396 		    td->td_priority));
1397 		return (td);
1398 	}
1399 	td = runq_choose(&tdq->tdq_idle);
1400 	if (td != NULL) {
1401 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1402 		    ("tdq_choose: Invalid priority on idle queue %d",
1403 		    td->td_priority));
1404 		return (td);
1405 	}
1406 
1407 	return (NULL);
1408 }
1409 
1410 /*
1411  * Initialize a thread queue.
1412  */
1413 static void
1414 tdq_setup(struct tdq *tdq, int id)
1415 {
1416 
1417 	if (bootverbose)
1418 		printf("ULE: setup cpu %d\n", id);
1419 	runq_init(&tdq->tdq_realtime);
1420 	runq_init(&tdq->tdq_timeshare);
1421 	runq_init(&tdq->tdq_idle);
1422 	tdq->tdq_id = id;
1423 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1424 	    "sched lock %d", (int)TDQ_ID(tdq));
1425 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1426 	    MTX_SPIN | MTX_RECURSE);
1427 #ifdef KTR
1428 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1429 	    "CPU %d load", (int)TDQ_ID(tdq));
1430 #endif
1431 }
1432 
1433 #ifdef SMP
1434 static void
1435 sched_setup_smp(void)
1436 {
1437 	struct tdq *tdq;
1438 	int i;
1439 
1440 	cpu_top = smp_topo();
1441 	CPU_FOREACH(i) {
1442 		tdq = DPCPU_ID_PTR(i, tdq);
1443 		tdq_setup(tdq, i);
1444 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1445 		if (tdq->tdq_cg == NULL)
1446 			panic("Can't find cpu group for %d\n", i);
1447 	}
1448 	PCPU_SET(sched, DPCPU_PTR(tdq));
1449 	balance_tdq = TDQ_SELF();
1450 }
1451 #endif
1452 
1453 /*
1454  * Setup the thread queues and initialize the topology based on MD
1455  * information.
1456  */
1457 static void
1458 sched_setup(void *dummy)
1459 {
1460 	struct tdq *tdq;
1461 
1462 #ifdef SMP
1463 	sched_setup_smp();
1464 #else
1465 	tdq_setup(TDQ_SELF(), 0);
1466 #endif
1467 	tdq = TDQ_SELF();
1468 
1469 	/* Add thread0's load since it's running. */
1470 	TDQ_LOCK(tdq);
1471 	thread0.td_lock = TDQ_LOCKPTR(tdq);
1472 	tdq_load_add(tdq, &thread0);
1473 	tdq->tdq_lowpri = thread0.td_priority;
1474 	TDQ_UNLOCK(tdq);
1475 }
1476 
1477 /*
1478  * This routine determines time constants after stathz and hz are setup.
1479  */
1480 /* ARGSUSED */
1481 static void
1482 sched_initticks(void *dummy)
1483 {
1484 	int incr;
1485 
1486 	realstathz = stathz ? stathz : hz;
1487 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1488 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1489 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1490 	    realstathz);
1491 
1492 	/*
1493 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1494 	 * hz not being evenly divisible by stathz on all platforms.
1495 	 */
1496 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1497 	/*
1498 	 * This does not work for values of stathz that are more than
1499 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1500 	 */
1501 	if (incr == 0)
1502 		incr = 1;
1503 	tickincr = incr;
1504 #ifdef SMP
1505 	/*
1506 	 * Set the default balance interval now that we know
1507 	 * what realstathz is.
1508 	 */
1509 	balance_interval = realstathz;
1510 	balance_ticks = balance_interval;
1511 	affinity = SCHED_AFFINITY_DEFAULT;
1512 #endif
1513 	if (sched_idlespinthresh < 0)
1514 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1515 }
1516 
1517 
1518 /*
1519  * This is the core of the interactivity algorithm.  Determines a score based
1520  * on past behavior.  It is the ratio of sleep time to run time scaled to
1521  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1522  * differs from the cpu usage because it does not account for time spent
1523  * waiting on a run-queue.  Would be prettier if we had floating point.
1524  *
1525  * When a thread's sleep time is greater than its run time the
1526  * calculation is:
1527  *
1528  *                           scaling factor
1529  * interactivity score =  ---------------------
1530  *                        sleep time / run time
1531  *
1532  *
1533  * When a thread's run time is greater than its sleep time the
1534  * calculation is:
1535  *
1536  *                           scaling factor
1537  * interactivity score =  ---------------------    + scaling factor
1538  *                        run time / sleep time
1539  */
1540 static int
1541 sched_interact_score(struct thread *td)
1542 {
1543 	struct td_sched *ts;
1544 	int div;
1545 
1546 	ts = td_get_sched(td);
1547 	/*
1548 	 * The score is only needed if this is likely to be an interactive
1549 	 * task.  Don't go through the expense of computing it if there's
1550 	 * no chance.
1551 	 */
1552 	if (sched_interact <= SCHED_INTERACT_HALF &&
1553 		ts->ts_runtime >= ts->ts_slptime)
1554 			return (SCHED_INTERACT_HALF);
1555 
1556 	if (ts->ts_runtime > ts->ts_slptime) {
1557 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1558 		return (SCHED_INTERACT_HALF +
1559 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1560 	}
1561 	if (ts->ts_slptime > ts->ts_runtime) {
1562 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1563 		return (ts->ts_runtime / div);
1564 	}
1565 	/* runtime == slptime */
1566 	if (ts->ts_runtime)
1567 		return (SCHED_INTERACT_HALF);
1568 
1569 	/*
1570 	 * This can happen if slptime and runtime are 0.
1571 	 */
1572 	return (0);
1573 
1574 }
1575 
1576 /*
1577  * Scale the scheduling priority according to the "interactivity" of this
1578  * process.
1579  */
1580 static void
1581 sched_priority(struct thread *td)
1582 {
1583 	int score;
1584 	int pri;
1585 
1586 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1587 		return;
1588 	/*
1589 	 * If the score is interactive we place the thread in the realtime
1590 	 * queue with a priority that is less than kernel and interrupt
1591 	 * priorities.  These threads are not subject to nice restrictions.
1592 	 *
1593 	 * Scores greater than this are placed on the normal timeshare queue
1594 	 * where the priority is partially decided by the most recent cpu
1595 	 * utilization and the rest is decided by nice value.
1596 	 *
1597 	 * The nice value of the process has a linear effect on the calculated
1598 	 * score.  Negative nice values make it easier for a thread to be
1599 	 * considered interactive.
1600 	 */
1601 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1602 	if (score < sched_interact) {
1603 		pri = PRI_MIN_INTERACT;
1604 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1605 		    sched_interact) * score;
1606 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1607 		    ("sched_priority: invalid interactive priority %d score %d",
1608 		    pri, score));
1609 	} else {
1610 		pri = SCHED_PRI_MIN;
1611 		if (td_get_sched(td)->ts_ticks)
1612 			pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1613 			    SCHED_PRI_RANGE - 1);
1614 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1615 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1616 		    ("sched_priority: invalid priority %d: nice %d, "
1617 		    "ticks %d ftick %d ltick %d tick pri %d",
1618 		    pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1619 		    td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1620 		    SCHED_PRI_TICKS(td_get_sched(td))));
1621 	}
1622 	sched_user_prio(td, pri);
1623 
1624 	return;
1625 }
1626 
1627 /*
1628  * This routine enforces a maximum limit on the amount of scheduling history
1629  * kept.  It is called after either the slptime or runtime is adjusted.  This
1630  * function is ugly due to integer math.
1631  */
1632 static void
1633 sched_interact_update(struct thread *td)
1634 {
1635 	struct td_sched *ts;
1636 	u_int sum;
1637 
1638 	ts = td_get_sched(td);
1639 	sum = ts->ts_runtime + ts->ts_slptime;
1640 	if (sum < SCHED_SLP_RUN_MAX)
1641 		return;
1642 	/*
1643 	 * This only happens from two places:
1644 	 * 1) We have added an unusual amount of run time from fork_exit.
1645 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1646 	 */
1647 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1648 		if (ts->ts_runtime > ts->ts_slptime) {
1649 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1650 			ts->ts_slptime = 1;
1651 		} else {
1652 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1653 			ts->ts_runtime = 1;
1654 		}
1655 		return;
1656 	}
1657 	/*
1658 	 * If we have exceeded by more than 1/5th then the algorithm below
1659 	 * will not bring us back into range.  Dividing by two here forces
1660 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1661 	 */
1662 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1663 		ts->ts_runtime /= 2;
1664 		ts->ts_slptime /= 2;
1665 		return;
1666 	}
1667 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1668 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1669 }
1670 
1671 /*
1672  * Scale back the interactivity history when a child thread is created.  The
1673  * history is inherited from the parent but the thread may behave totally
1674  * differently.  For example, a shell spawning a compiler process.  We want
1675  * to learn that the compiler is behaving badly very quickly.
1676  */
1677 static void
1678 sched_interact_fork(struct thread *td)
1679 {
1680 	struct td_sched *ts;
1681 	int ratio;
1682 	int sum;
1683 
1684 	ts = td_get_sched(td);
1685 	sum = ts->ts_runtime + ts->ts_slptime;
1686 	if (sum > SCHED_SLP_RUN_FORK) {
1687 		ratio = sum / SCHED_SLP_RUN_FORK;
1688 		ts->ts_runtime /= ratio;
1689 		ts->ts_slptime /= ratio;
1690 	}
1691 }
1692 
1693 /*
1694  * Called from proc0_init() to setup the scheduler fields.
1695  */
1696 void
1697 schedinit(void)
1698 {
1699 	struct td_sched *ts0;
1700 
1701 	/*
1702 	 * Set up the scheduler specific parts of thread0.
1703 	 */
1704 	ts0 = td_get_sched(&thread0);
1705 	ts0->ts_ltick = ticks;
1706 	ts0->ts_ftick = ticks;
1707 	ts0->ts_slice = 0;
1708 	ts0->ts_cpu = curcpu;	/* set valid CPU number */
1709 }
1710 
1711 /*
1712  * This is only somewhat accurate since given many processes of the same
1713  * priority they will switch when their slices run out, which will be
1714  * at most sched_slice stathz ticks.
1715  */
1716 int
1717 sched_rr_interval(void)
1718 {
1719 
1720 	/* Convert sched_slice from stathz to hz. */
1721 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1722 }
1723 
1724 /*
1725  * Update the percent cpu tracking information when it is requested or
1726  * the total history exceeds the maximum.  We keep a sliding history of
1727  * tick counts that slowly decays.  This is less precise than the 4BSD
1728  * mechanism since it happens with less regular and frequent events.
1729  */
1730 static void
1731 sched_pctcpu_update(struct td_sched *ts, int run)
1732 {
1733 	int t = ticks;
1734 
1735 	/*
1736 	 * The signed difference may be negative if the thread hasn't run for
1737 	 * over half of the ticks rollover period.
1738 	 */
1739 	if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) {
1740 		ts->ts_ticks = 0;
1741 		ts->ts_ftick = t - SCHED_TICK_TARG;
1742 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1743 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1744 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1745 		ts->ts_ftick = t - SCHED_TICK_TARG;
1746 	}
1747 	if (run)
1748 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1749 	ts->ts_ltick = t;
1750 }
1751 
1752 /*
1753  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1754  * if necessary.  This is the back-end for several priority related
1755  * functions.
1756  */
1757 static void
1758 sched_thread_priority(struct thread *td, u_char prio)
1759 {
1760 	struct td_sched *ts;
1761 	struct tdq *tdq;
1762 	int oldpri;
1763 
1764 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1765 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1766 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1767 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1768 	if (td != curthread && prio < td->td_priority) {
1769 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1770 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1771 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1772 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1773 		    curthread);
1774 	}
1775 	ts = td_get_sched(td);
1776 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1777 	if (td->td_priority == prio)
1778 		return;
1779 	/*
1780 	 * If the priority has been elevated due to priority
1781 	 * propagation, we may have to move ourselves to a new
1782 	 * queue.  This could be optimized to not re-add in some
1783 	 * cases.
1784 	 */
1785 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1786 		sched_rem(td);
1787 		td->td_priority = prio;
1788 		sched_add(td, SRQ_BORROWING);
1789 		return;
1790 	}
1791 	/*
1792 	 * If the thread is currently running we may have to adjust the lowpri
1793 	 * information so other cpus are aware of our current priority.
1794 	 */
1795 	if (TD_IS_RUNNING(td)) {
1796 		tdq = TDQ_CPU(ts->ts_cpu);
1797 		oldpri = td->td_priority;
1798 		td->td_priority = prio;
1799 		if (prio < tdq->tdq_lowpri)
1800 			tdq->tdq_lowpri = prio;
1801 		else if (tdq->tdq_lowpri == oldpri)
1802 			tdq_setlowpri(tdq, td);
1803 		return;
1804 	}
1805 	td->td_priority = prio;
1806 }
1807 
1808 /*
1809  * Update a thread's priority when it is lent another thread's
1810  * priority.
1811  */
1812 void
1813 sched_lend_prio(struct thread *td, u_char prio)
1814 {
1815 
1816 	td->td_flags |= TDF_BORROWING;
1817 	sched_thread_priority(td, prio);
1818 }
1819 
1820 /*
1821  * Restore a thread's priority when priority propagation is
1822  * over.  The prio argument is the minimum priority the thread
1823  * needs to have to satisfy other possible priority lending
1824  * requests.  If the thread's regular priority is less
1825  * important than prio, the thread will keep a priority boost
1826  * of prio.
1827  */
1828 void
1829 sched_unlend_prio(struct thread *td, u_char prio)
1830 {
1831 	u_char base_pri;
1832 
1833 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1834 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1835 		base_pri = td->td_user_pri;
1836 	else
1837 		base_pri = td->td_base_pri;
1838 	if (prio >= base_pri) {
1839 		td->td_flags &= ~TDF_BORROWING;
1840 		sched_thread_priority(td, base_pri);
1841 	} else
1842 		sched_lend_prio(td, prio);
1843 }
1844 
1845 /*
1846  * Standard entry for setting the priority to an absolute value.
1847  */
1848 void
1849 sched_prio(struct thread *td, u_char prio)
1850 {
1851 	u_char oldprio;
1852 
1853 	/* First, update the base priority. */
1854 	td->td_base_pri = prio;
1855 
1856 	/*
1857 	 * If the thread is borrowing another thread's priority, don't
1858 	 * ever lower the priority.
1859 	 */
1860 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1861 		return;
1862 
1863 	/* Change the real priority. */
1864 	oldprio = td->td_priority;
1865 	sched_thread_priority(td, prio);
1866 
1867 	/*
1868 	 * If the thread is on a turnstile, then let the turnstile update
1869 	 * its state.
1870 	 */
1871 	if (TD_ON_LOCK(td) && oldprio != prio)
1872 		turnstile_adjust(td, oldprio);
1873 }
1874 
1875 /*
1876  * Set the base user priority, does not effect current running priority.
1877  */
1878 void
1879 sched_user_prio(struct thread *td, u_char prio)
1880 {
1881 
1882 	td->td_base_user_pri = prio;
1883 	if (td->td_lend_user_pri <= prio)
1884 		return;
1885 	td->td_user_pri = prio;
1886 }
1887 
1888 void
1889 sched_lend_user_prio(struct thread *td, u_char prio)
1890 {
1891 
1892 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1893 	td->td_lend_user_pri = prio;
1894 	td->td_user_pri = min(prio, td->td_base_user_pri);
1895 	if (td->td_priority > td->td_user_pri)
1896 		sched_prio(td, td->td_user_pri);
1897 	else if (td->td_priority != td->td_user_pri)
1898 		td->td_flags |= TDF_NEEDRESCHED;
1899 }
1900 
1901 /*
1902  * Like the above but first check if there is anything to do.
1903  */
1904 void
1905 sched_lend_user_prio_cond(struct thread *td, u_char prio)
1906 {
1907 
1908 	if (td->td_lend_user_pri != prio)
1909 		goto lend;
1910 	if (td->td_user_pri != min(prio, td->td_base_user_pri))
1911 		goto lend;
1912 	if (td->td_priority >= td->td_user_pri)
1913 		goto lend;
1914 	return;
1915 
1916 lend:
1917 	thread_lock(td);
1918 	sched_lend_user_prio(td, prio);
1919 	thread_unlock(td);
1920 }
1921 
1922 #ifdef SMP
1923 /*
1924  * This tdq is about to idle.  Try to steal a thread from another CPU before
1925  * choosing the idle thread.
1926  */
1927 static void
1928 tdq_trysteal(struct tdq *tdq)
1929 {
1930 	struct cpu_group *cg;
1931 	struct tdq *steal;
1932 	cpuset_t mask;
1933 	int cpu, i;
1934 
1935 	if (smp_started == 0 || trysteal_limit == 0 || tdq->tdq_cg == NULL)
1936 		return;
1937 	CPU_FILL(&mask);
1938 	CPU_CLR(PCPU_GET(cpuid), &mask);
1939 	/* We don't want to be preempted while we're iterating. */
1940 	spinlock_enter();
1941 	TDQ_UNLOCK(tdq);
1942 	for (i = 1, cg = tdq->tdq_cg; ; ) {
1943 		cpu = sched_highest(cg, mask, steal_thresh);
1944 		/*
1945 		 * If a thread was added while interrupts were disabled don't
1946 		 * steal one here.
1947 		 */
1948 		if (tdq->tdq_load > 0) {
1949 			TDQ_LOCK(tdq);
1950 			break;
1951 		}
1952 		if (cpu == -1) {
1953 			i++;
1954 			cg = cg->cg_parent;
1955 			if (cg == NULL || i > trysteal_limit) {
1956 				TDQ_LOCK(tdq);
1957 				break;
1958 			}
1959 			continue;
1960 		}
1961 		steal = TDQ_CPU(cpu);
1962 		/*
1963 		 * The data returned by sched_highest() is stale and
1964                  * the chosen CPU no longer has an eligible thread.
1965 		 */
1966 		if (steal->tdq_load < steal_thresh ||
1967 		    steal->tdq_transferable == 0)
1968 			continue;
1969 		tdq_lock_pair(tdq, steal);
1970 		/*
1971 		 * If we get to this point, unconditonally exit the loop
1972 		 * to bound the time spent in the critcal section.
1973 		 *
1974 		 * If a thread was added while interrupts were disabled don't
1975 		 * steal one here.
1976 		 */
1977 		if (tdq->tdq_load > 0) {
1978 			TDQ_UNLOCK(steal);
1979 			break;
1980 		}
1981 		/*
1982 		 * The data returned by sched_highest() is stale and
1983                  * the chosen CPU no longer has an eligible thread.
1984 		 */
1985 		if (steal->tdq_load < steal_thresh ||
1986 		    steal->tdq_transferable == 0) {
1987 			TDQ_UNLOCK(steal);
1988 			break;
1989 		}
1990 		/*
1991 		 * If we fail to acquire one due to affinity restrictions,
1992 		 * bail out and let the idle thread to a more complete search
1993 		 * outside of a critical section.
1994 		 */
1995 		if (tdq_move(steal, tdq) == NULL) {
1996 			TDQ_UNLOCK(steal);
1997 			break;
1998 		}
1999 		TDQ_UNLOCK(steal);
2000 		break;
2001 	}
2002 	spinlock_exit();
2003 }
2004 #endif
2005 
2006 /*
2007  * Handle migration from sched_switch().  This happens only for
2008  * cpu binding.
2009  */
2010 static struct mtx *
2011 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
2012 {
2013 	struct tdq *tdn;
2014 
2015 	KASSERT(!CPU_ABSENT(td_get_sched(td)->ts_cpu), ("sched_switch_migrate: "
2016 	    "thread %s queued on absent CPU %d.", td->td_name,
2017 	    td_get_sched(td)->ts_cpu));
2018 	tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
2019 #ifdef SMP
2020 	tdq_load_rem(tdq, td);
2021 	/*
2022 	 * Do the lock dance required to avoid LOR.  We grab an extra
2023 	 * spinlock nesting to prevent preemption while we're
2024 	 * not holding either run-queue lock.
2025 	 */
2026 	spinlock_enter();
2027 	thread_lock_block(td);	/* This releases the lock on tdq. */
2028 
2029 	/*
2030 	 * Acquire both run-queue locks before placing the thread on the new
2031 	 * run-queue to avoid deadlocks created by placing a thread with a
2032 	 * blocked lock on the run-queue of a remote processor.  The deadlock
2033 	 * occurs when a third processor attempts to lock the two queues in
2034 	 * question while the target processor is spinning with its own
2035 	 * run-queue lock held while waiting for the blocked lock to clear.
2036 	 */
2037 	tdq_lock_pair(tdn, tdq);
2038 	tdq_add(tdn, td, flags);
2039 	tdq_notify(tdn, td);
2040 	TDQ_UNLOCK(tdn);
2041 	spinlock_exit();
2042 #endif
2043 	return (TDQ_LOCKPTR(tdn));
2044 }
2045 
2046 /*
2047  * Variadic version of thread_lock_unblock() that does not assume td_lock
2048  * is blocked.
2049  */
2050 static inline void
2051 thread_unblock_switch(struct thread *td, struct mtx *mtx)
2052 {
2053 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
2054 	    (uintptr_t)mtx);
2055 }
2056 
2057 /*
2058  * Switch threads.  This function has to handle threads coming in while
2059  * blocked for some reason, running, or idle.  It also must deal with
2060  * migrating a thread from one queue to another as running threads may
2061  * be assigned elsewhere via binding.
2062  */
2063 void
2064 sched_switch(struct thread *td, struct thread *newtd, int flags)
2065 {
2066 	struct tdq *tdq;
2067 	struct td_sched *ts;
2068 	struct mtx *mtx;
2069 	int srqflag;
2070 	int cpuid, preempted;
2071 
2072 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2073 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
2074 
2075 	cpuid = PCPU_GET(cpuid);
2076 	tdq = TDQ_SELF();
2077 	ts = td_get_sched(td);
2078 	mtx = td->td_lock;
2079 	sched_pctcpu_update(ts, 1);
2080 	ts->ts_rltick = ticks;
2081 	td->td_lastcpu = td->td_oncpu;
2082 	td->td_oncpu = NOCPU;
2083 	preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
2084 	    (flags & SW_PREEMPT) != 0;
2085 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
2086 	td->td_owepreempt = 0;
2087 	tdq->tdq_owepreempt = 0;
2088 	if (!TD_IS_IDLETHREAD(td))
2089 		tdq->tdq_switchcnt++;
2090 
2091 	/*
2092 	 * The lock pointer in an idle thread should never change.  Reset it
2093 	 * to CAN_RUN as well.
2094 	 */
2095 	if (TD_IS_IDLETHREAD(td)) {
2096 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2097 		TD_SET_CAN_RUN(td);
2098 	} else if (TD_IS_RUNNING(td)) {
2099 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2100 		srqflag = preempted ?
2101 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
2102 		    SRQ_OURSELF|SRQ_YIELDING;
2103 #ifdef SMP
2104 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
2105 			ts->ts_cpu = sched_pickcpu(td, 0);
2106 #endif
2107 		if (ts->ts_cpu == cpuid)
2108 			tdq_runq_add(tdq, td, srqflag);
2109 		else {
2110 			KASSERT(THREAD_CAN_MIGRATE(td) ||
2111 			    (ts->ts_flags & TSF_BOUND) != 0,
2112 			    ("Thread %p shouldn't migrate", td));
2113 			mtx = sched_switch_migrate(tdq, td, srqflag);
2114 		}
2115 	} else {
2116 		/* This thread must be going to sleep. */
2117 		TDQ_LOCK(tdq);
2118 		mtx = thread_lock_block(td);
2119 		tdq_load_rem(tdq, td);
2120 #ifdef SMP
2121 		if (tdq->tdq_load == 0)
2122 			tdq_trysteal(tdq);
2123 #endif
2124 	}
2125 
2126 #if (KTR_COMPILE & KTR_SCHED) != 0
2127 	if (TD_IS_IDLETHREAD(td))
2128 		KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
2129 		    "prio:%d", td->td_priority);
2130 	else
2131 		KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
2132 		    "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
2133 		    "lockname:\"%s\"", td->td_lockname);
2134 #endif
2135 
2136 	/*
2137 	 * We enter here with the thread blocked and assigned to the
2138 	 * appropriate cpu run-queue or sleep-queue and with the current
2139 	 * thread-queue locked.
2140 	 */
2141 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2142 	newtd = choosethread();
2143 	/*
2144 	 * Call the MD code to switch contexts if necessary.
2145 	 */
2146 	if (td != newtd) {
2147 #ifdef	HWPMC_HOOKS
2148 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2149 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
2150 #endif
2151 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
2152 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2153 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2154 		sched_pctcpu_update(td_get_sched(newtd), 0);
2155 
2156 #ifdef KDTRACE_HOOKS
2157 		/*
2158 		 * If DTrace has set the active vtime enum to anything
2159 		 * other than INACTIVE (0), then it should have set the
2160 		 * function to call.
2161 		 */
2162 		if (dtrace_vtime_active)
2163 			(*dtrace_vtime_switch_func)(newtd);
2164 #endif
2165 
2166 		cpu_switch(td, newtd, mtx);
2167 		/*
2168 		 * We may return from cpu_switch on a different cpu.  However,
2169 		 * we always return with td_lock pointing to the current cpu's
2170 		 * run queue lock.
2171 		 */
2172 		cpuid = PCPU_GET(cpuid);
2173 		tdq = TDQ_SELF();
2174 		lock_profile_obtain_lock_success(
2175 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2176 
2177 		SDT_PROBE0(sched, , , on__cpu);
2178 #ifdef	HWPMC_HOOKS
2179 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2180 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
2181 #endif
2182 	} else {
2183 		thread_unblock_switch(td, mtx);
2184 		SDT_PROBE0(sched, , , remain__cpu);
2185 	}
2186 
2187 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2188 	    "prio:%d", td->td_priority);
2189 
2190 	/*
2191 	 * Assert that all went well and return.
2192 	 */
2193 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
2194 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2195 	td->td_oncpu = cpuid;
2196 }
2197 
2198 /*
2199  * Adjust thread priorities as a result of a nice request.
2200  */
2201 void
2202 sched_nice(struct proc *p, int nice)
2203 {
2204 	struct thread *td;
2205 
2206 	PROC_LOCK_ASSERT(p, MA_OWNED);
2207 
2208 	p->p_nice = nice;
2209 	FOREACH_THREAD_IN_PROC(p, td) {
2210 		thread_lock(td);
2211 		sched_priority(td);
2212 		sched_prio(td, td->td_base_user_pri);
2213 		thread_unlock(td);
2214 	}
2215 }
2216 
2217 /*
2218  * Record the sleep time for the interactivity scorer.
2219  */
2220 void
2221 sched_sleep(struct thread *td, int prio)
2222 {
2223 
2224 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2225 
2226 	td->td_slptick = ticks;
2227 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2228 		td->td_flags |= TDF_CANSWAP;
2229 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2230 		return;
2231 	if (static_boost == 1 && prio)
2232 		sched_prio(td, prio);
2233 	else if (static_boost && td->td_priority > static_boost)
2234 		sched_prio(td, static_boost);
2235 }
2236 
2237 /*
2238  * Schedule a thread to resume execution and record how long it voluntarily
2239  * slept.  We also update the pctcpu, interactivity, and priority.
2240  */
2241 void
2242 sched_wakeup(struct thread *td)
2243 {
2244 	struct td_sched *ts;
2245 	int slptick;
2246 
2247 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2248 	ts = td_get_sched(td);
2249 	td->td_flags &= ~TDF_CANSWAP;
2250 	/*
2251 	 * If we slept for more than a tick update our interactivity and
2252 	 * priority.
2253 	 */
2254 	slptick = td->td_slptick;
2255 	td->td_slptick = 0;
2256 	if (slptick && slptick != ticks) {
2257 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2258 		sched_interact_update(td);
2259 		sched_pctcpu_update(ts, 0);
2260 	}
2261 	/*
2262 	 * Reset the slice value since we slept and advanced the round-robin.
2263 	 */
2264 	ts->ts_slice = 0;
2265 	sched_add(td, SRQ_BORING);
2266 }
2267 
2268 /*
2269  * Penalize the parent for creating a new child and initialize the child's
2270  * priority.
2271  */
2272 void
2273 sched_fork(struct thread *td, struct thread *child)
2274 {
2275 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2276 	sched_pctcpu_update(td_get_sched(td), 1);
2277 	sched_fork_thread(td, child);
2278 	/*
2279 	 * Penalize the parent and child for forking.
2280 	 */
2281 	sched_interact_fork(child);
2282 	sched_priority(child);
2283 	td_get_sched(td)->ts_runtime += tickincr;
2284 	sched_interact_update(td);
2285 	sched_priority(td);
2286 }
2287 
2288 /*
2289  * Fork a new thread, may be within the same process.
2290  */
2291 void
2292 sched_fork_thread(struct thread *td, struct thread *child)
2293 {
2294 	struct td_sched *ts;
2295 	struct td_sched *ts2;
2296 	struct tdq *tdq;
2297 
2298 	tdq = TDQ_SELF();
2299 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2300 	/*
2301 	 * Initialize child.
2302 	 */
2303 	ts = td_get_sched(td);
2304 	ts2 = td_get_sched(child);
2305 	child->td_oncpu = NOCPU;
2306 	child->td_lastcpu = NOCPU;
2307 	child->td_lock = TDQ_LOCKPTR(tdq);
2308 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2309 	child->td_domain.dr_policy = td->td_cpuset->cs_domain;
2310 	ts2->ts_cpu = ts->ts_cpu;
2311 	ts2->ts_flags = 0;
2312 	/*
2313 	 * Grab our parents cpu estimation information.
2314 	 */
2315 	ts2->ts_ticks = ts->ts_ticks;
2316 	ts2->ts_ltick = ts->ts_ltick;
2317 	ts2->ts_ftick = ts->ts_ftick;
2318 	/*
2319 	 * Do not inherit any borrowed priority from the parent.
2320 	 */
2321 	child->td_priority = child->td_base_pri;
2322 	/*
2323 	 * And update interactivity score.
2324 	 */
2325 	ts2->ts_slptime = ts->ts_slptime;
2326 	ts2->ts_runtime = ts->ts_runtime;
2327 	/* Attempt to quickly learn interactivity. */
2328 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2329 #ifdef KTR
2330 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2331 #endif
2332 }
2333 
2334 /*
2335  * Adjust the priority class of a thread.
2336  */
2337 void
2338 sched_class(struct thread *td, int class)
2339 {
2340 
2341 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2342 	if (td->td_pri_class == class)
2343 		return;
2344 	td->td_pri_class = class;
2345 }
2346 
2347 /*
2348  * Return some of the child's priority and interactivity to the parent.
2349  */
2350 void
2351 sched_exit(struct proc *p, struct thread *child)
2352 {
2353 	struct thread *td;
2354 
2355 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2356 	    "prio:%d", child->td_priority);
2357 	PROC_LOCK_ASSERT(p, MA_OWNED);
2358 	td = FIRST_THREAD_IN_PROC(p);
2359 	sched_exit_thread(td, child);
2360 }
2361 
2362 /*
2363  * Penalize another thread for the time spent on this one.  This helps to
2364  * worsen the priority and interactivity of processes which schedule batch
2365  * jobs such as make.  This has little effect on the make process itself but
2366  * causes new processes spawned by it to receive worse scores immediately.
2367  */
2368 void
2369 sched_exit_thread(struct thread *td, struct thread *child)
2370 {
2371 
2372 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2373 	    "prio:%d", child->td_priority);
2374 	/*
2375 	 * Give the child's runtime to the parent without returning the
2376 	 * sleep time as a penalty to the parent.  This causes shells that
2377 	 * launch expensive things to mark their children as expensive.
2378 	 */
2379 	thread_lock(td);
2380 	td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2381 	sched_interact_update(td);
2382 	sched_priority(td);
2383 	thread_unlock(td);
2384 }
2385 
2386 void
2387 sched_preempt(struct thread *td)
2388 {
2389 	struct tdq *tdq;
2390 
2391 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2392 
2393 	thread_lock(td);
2394 	tdq = TDQ_SELF();
2395 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2396 	if (td->td_priority > tdq->tdq_lowpri) {
2397 		int flags;
2398 
2399 		flags = SW_INVOL | SW_PREEMPT;
2400 		if (td->td_critnest > 1)
2401 			td->td_owepreempt = 1;
2402 		else if (TD_IS_IDLETHREAD(td))
2403 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2404 		else
2405 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2406 	} else {
2407 		tdq->tdq_owepreempt = 0;
2408 	}
2409 	thread_unlock(td);
2410 }
2411 
2412 /*
2413  * Fix priorities on return to user-space.  Priorities may be elevated due
2414  * to static priorities in msleep() or similar.
2415  */
2416 void
2417 sched_userret_slowpath(struct thread *td)
2418 {
2419 
2420 	thread_lock(td);
2421 	td->td_priority = td->td_user_pri;
2422 	td->td_base_pri = td->td_user_pri;
2423 	tdq_setlowpri(TDQ_SELF(), td);
2424 	thread_unlock(td);
2425 }
2426 
2427 /*
2428  * Handle a stathz tick.  This is really only relevant for timeshare
2429  * threads.
2430  */
2431 void
2432 sched_clock(struct thread *td, int cnt)
2433 {
2434 	struct tdq *tdq;
2435 	struct td_sched *ts;
2436 
2437 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2438 	tdq = TDQ_SELF();
2439 #ifdef SMP
2440 	/*
2441 	 * We run the long term load balancer infrequently on the first cpu.
2442 	 */
2443 	if (balance_tdq == tdq && smp_started != 0 && rebalance != 0 &&
2444 	    balance_ticks != 0) {
2445 		balance_ticks -= cnt;
2446 		if (balance_ticks <= 0)
2447 			sched_balance();
2448 	}
2449 #endif
2450 	/*
2451 	 * Save the old switch count so we have a record of the last ticks
2452 	 * activity.   Initialize the new switch count based on our load.
2453 	 * If there is some activity seed it to reflect that.
2454 	 */
2455 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2456 	tdq->tdq_switchcnt = tdq->tdq_load;
2457 	/*
2458 	 * Advance the insert index once for each tick to ensure that all
2459 	 * threads get a chance to run.
2460 	 */
2461 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2462 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2463 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2464 			tdq->tdq_ridx = tdq->tdq_idx;
2465 	}
2466 	ts = td_get_sched(td);
2467 	sched_pctcpu_update(ts, 1);
2468 	if ((td->td_pri_class & PRI_FIFO_BIT) || TD_IS_IDLETHREAD(td))
2469 		return;
2470 
2471 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2472 		/*
2473 		 * We used a tick; charge it to the thread so
2474 		 * that we can compute our interactivity.
2475 		 */
2476 		td_get_sched(td)->ts_runtime += tickincr * cnt;
2477 		sched_interact_update(td);
2478 		sched_priority(td);
2479 	}
2480 
2481 	/*
2482 	 * Force a context switch if the current thread has used up a full
2483 	 * time slice (default is 100ms).
2484 	 */
2485 	ts->ts_slice += cnt;
2486 	if (ts->ts_slice >= tdq_slice(tdq)) {
2487 		ts->ts_slice = 0;
2488 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2489 	}
2490 }
2491 
2492 u_int
2493 sched_estcpu(struct thread *td __unused)
2494 {
2495 
2496 	return (0);
2497 }
2498 
2499 /*
2500  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2501  * cooperative idle threads.
2502  */
2503 int
2504 sched_runnable(void)
2505 {
2506 	struct tdq *tdq;
2507 	int load;
2508 
2509 	load = 1;
2510 
2511 	tdq = TDQ_SELF();
2512 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2513 		if (tdq->tdq_load > 0)
2514 			goto out;
2515 	} else
2516 		if (tdq->tdq_load - 1 > 0)
2517 			goto out;
2518 	load = 0;
2519 out:
2520 	return (load);
2521 }
2522 
2523 /*
2524  * Choose the highest priority thread to run.  The thread is removed from
2525  * the run-queue while running however the load remains.  For SMP we set
2526  * the tdq in the global idle bitmask if it idles here.
2527  */
2528 struct thread *
2529 sched_choose(void)
2530 {
2531 	struct thread *td;
2532 	struct tdq *tdq;
2533 
2534 	tdq = TDQ_SELF();
2535 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2536 	td = tdq_choose(tdq);
2537 	if (td) {
2538 		tdq_runq_rem(tdq, td);
2539 		tdq->tdq_lowpri = td->td_priority;
2540 		return (td);
2541 	}
2542 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2543 	return (PCPU_GET(idlethread));
2544 }
2545 
2546 /*
2547  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2548  * we always request it once we exit a critical section.
2549  */
2550 static inline void
2551 sched_setpreempt(struct thread *td)
2552 {
2553 	struct thread *ctd;
2554 	int cpri;
2555 	int pri;
2556 
2557 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2558 
2559 	ctd = curthread;
2560 	pri = td->td_priority;
2561 	cpri = ctd->td_priority;
2562 	if (pri < cpri)
2563 		ctd->td_flags |= TDF_NEEDRESCHED;
2564 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2565 		return;
2566 	if (!sched_shouldpreempt(pri, cpri, 0))
2567 		return;
2568 	ctd->td_owepreempt = 1;
2569 }
2570 
2571 /*
2572  * Add a thread to a thread queue.  Select the appropriate runq and add the
2573  * thread to it.  This is the internal function called when the tdq is
2574  * predetermined.
2575  */
2576 void
2577 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2578 {
2579 
2580 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2581 	KASSERT((td->td_inhibitors == 0),
2582 	    ("sched_add: trying to run inhibited thread"));
2583 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2584 	    ("sched_add: bad thread state"));
2585 	KASSERT(td->td_flags & TDF_INMEM,
2586 	    ("sched_add: thread swapped out"));
2587 
2588 	if (td->td_priority < tdq->tdq_lowpri)
2589 		tdq->tdq_lowpri = td->td_priority;
2590 	tdq_runq_add(tdq, td, flags);
2591 	tdq_load_add(tdq, td);
2592 }
2593 
2594 /*
2595  * Select the target thread queue and add a thread to it.  Request
2596  * preemption or IPI a remote processor if required.
2597  */
2598 void
2599 sched_add(struct thread *td, int flags)
2600 {
2601 	struct tdq *tdq;
2602 #ifdef SMP
2603 	int cpu;
2604 #endif
2605 
2606 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2607 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2608 	    sched_tdname(curthread));
2609 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2610 	    KTR_ATTR_LINKED, sched_tdname(td));
2611 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2612 	    flags & SRQ_PREEMPTED);
2613 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2614 	/*
2615 	 * Recalculate the priority before we select the target cpu or
2616 	 * run-queue.
2617 	 */
2618 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2619 		sched_priority(td);
2620 #ifdef SMP
2621 	/*
2622 	 * Pick the destination cpu and if it isn't ours transfer to the
2623 	 * target cpu.
2624 	 */
2625 	cpu = sched_pickcpu(td, flags);
2626 	tdq = sched_setcpu(td, cpu, flags);
2627 	tdq_add(tdq, td, flags);
2628 	if (cpu != PCPU_GET(cpuid)) {
2629 		tdq_notify(tdq, td);
2630 		return;
2631 	}
2632 #else
2633 	tdq = TDQ_SELF();
2634 	TDQ_LOCK(tdq);
2635 	/*
2636 	 * Now that the thread is moving to the run-queue, set the lock
2637 	 * to the scheduler's lock.
2638 	 */
2639 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2640 	tdq_add(tdq, td, flags);
2641 #endif
2642 	if (!(flags & SRQ_YIELDING))
2643 		sched_setpreempt(td);
2644 }
2645 
2646 /*
2647  * Remove a thread from a run-queue without running it.  This is used
2648  * when we're stealing a thread from a remote queue.  Otherwise all threads
2649  * exit by calling sched_exit_thread() and sched_throw() themselves.
2650  */
2651 void
2652 sched_rem(struct thread *td)
2653 {
2654 	struct tdq *tdq;
2655 
2656 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2657 	    "prio:%d", td->td_priority);
2658 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2659 	tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2660 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2661 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2662 	KASSERT(TD_ON_RUNQ(td),
2663 	    ("sched_rem: thread not on run queue"));
2664 	tdq_runq_rem(tdq, td);
2665 	tdq_load_rem(tdq, td);
2666 	TD_SET_CAN_RUN(td);
2667 	if (td->td_priority == tdq->tdq_lowpri)
2668 		tdq_setlowpri(tdq, NULL);
2669 }
2670 
2671 /*
2672  * Fetch cpu utilization information.  Updates on demand.
2673  */
2674 fixpt_t
2675 sched_pctcpu(struct thread *td)
2676 {
2677 	fixpt_t pctcpu;
2678 	struct td_sched *ts;
2679 
2680 	pctcpu = 0;
2681 	ts = td_get_sched(td);
2682 
2683 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2684 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2685 	if (ts->ts_ticks) {
2686 		int rtick;
2687 
2688 		/* How many rtick per second ? */
2689 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2690 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2691 	}
2692 
2693 	return (pctcpu);
2694 }
2695 
2696 /*
2697  * Enforce affinity settings for a thread.  Called after adjustments to
2698  * cpumask.
2699  */
2700 void
2701 sched_affinity(struct thread *td)
2702 {
2703 #ifdef SMP
2704 	struct td_sched *ts;
2705 
2706 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2707 	ts = td_get_sched(td);
2708 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2709 		return;
2710 	if (TD_ON_RUNQ(td)) {
2711 		sched_rem(td);
2712 		sched_add(td, SRQ_BORING);
2713 		return;
2714 	}
2715 	if (!TD_IS_RUNNING(td))
2716 		return;
2717 	/*
2718 	 * Force a switch before returning to userspace.  If the
2719 	 * target thread is not running locally send an ipi to force
2720 	 * the issue.
2721 	 */
2722 	td->td_flags |= TDF_NEEDRESCHED;
2723 	if (td != curthread)
2724 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2725 #endif
2726 }
2727 
2728 /*
2729  * Bind a thread to a target cpu.
2730  */
2731 void
2732 sched_bind(struct thread *td, int cpu)
2733 {
2734 	struct td_sched *ts;
2735 
2736 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2737 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2738 	ts = td_get_sched(td);
2739 	if (ts->ts_flags & TSF_BOUND)
2740 		sched_unbind(td);
2741 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2742 	ts->ts_flags |= TSF_BOUND;
2743 	sched_pin();
2744 	if (PCPU_GET(cpuid) == cpu)
2745 		return;
2746 	ts->ts_cpu = cpu;
2747 	/* When we return from mi_switch we'll be on the correct cpu. */
2748 	mi_switch(SW_VOL, NULL);
2749 }
2750 
2751 /*
2752  * Release a bound thread.
2753  */
2754 void
2755 sched_unbind(struct thread *td)
2756 {
2757 	struct td_sched *ts;
2758 
2759 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2760 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2761 	ts = td_get_sched(td);
2762 	if ((ts->ts_flags & TSF_BOUND) == 0)
2763 		return;
2764 	ts->ts_flags &= ~TSF_BOUND;
2765 	sched_unpin();
2766 }
2767 
2768 int
2769 sched_is_bound(struct thread *td)
2770 {
2771 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2772 	return (td_get_sched(td)->ts_flags & TSF_BOUND);
2773 }
2774 
2775 /*
2776  * Basic yield call.
2777  */
2778 void
2779 sched_relinquish(struct thread *td)
2780 {
2781 	thread_lock(td);
2782 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2783 	thread_unlock(td);
2784 }
2785 
2786 /*
2787  * Return the total system load.
2788  */
2789 int
2790 sched_load(void)
2791 {
2792 #ifdef SMP
2793 	int total;
2794 	int i;
2795 
2796 	total = 0;
2797 	CPU_FOREACH(i)
2798 		total += TDQ_CPU(i)->tdq_sysload;
2799 	return (total);
2800 #else
2801 	return (TDQ_SELF()->tdq_sysload);
2802 #endif
2803 }
2804 
2805 int
2806 sched_sizeof_proc(void)
2807 {
2808 	return (sizeof(struct proc));
2809 }
2810 
2811 int
2812 sched_sizeof_thread(void)
2813 {
2814 	return (sizeof(struct thread) + sizeof(struct td_sched));
2815 }
2816 
2817 #ifdef SMP
2818 #define	TDQ_IDLESPIN(tdq)						\
2819     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2820 #else
2821 #define	TDQ_IDLESPIN(tdq)	1
2822 #endif
2823 
2824 /*
2825  * The actual idle process.
2826  */
2827 void
2828 sched_idletd(void *dummy)
2829 {
2830 	struct thread *td;
2831 	struct tdq *tdq;
2832 	int oldswitchcnt, switchcnt;
2833 	int i;
2834 
2835 	mtx_assert(&Giant, MA_NOTOWNED);
2836 	td = curthread;
2837 	tdq = TDQ_SELF();
2838 	THREAD_NO_SLEEPING();
2839 	oldswitchcnt = -1;
2840 	for (;;) {
2841 		if (tdq->tdq_load) {
2842 			thread_lock(td);
2843 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2844 			thread_unlock(td);
2845 		}
2846 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2847 #ifdef SMP
2848 		if (always_steal || switchcnt != oldswitchcnt) {
2849 			oldswitchcnt = switchcnt;
2850 			if (tdq_idled(tdq) == 0)
2851 				continue;
2852 		}
2853 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2854 #else
2855 		oldswitchcnt = switchcnt;
2856 #endif
2857 		/*
2858 		 * If we're switching very frequently, spin while checking
2859 		 * for load rather than entering a low power state that
2860 		 * may require an IPI.  However, don't do any busy
2861 		 * loops while on SMT machines as this simply steals
2862 		 * cycles from cores doing useful work.
2863 		 */
2864 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2865 			for (i = 0; i < sched_idlespins; i++) {
2866 				if (tdq->tdq_load)
2867 					break;
2868 				cpu_spinwait();
2869 			}
2870 		}
2871 
2872 		/* If there was context switch during spin, restart it. */
2873 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2874 		if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2875 			continue;
2876 
2877 		/* Run main MD idle handler. */
2878 		tdq->tdq_cpu_idle = 1;
2879 		/*
2880 		 * Make sure that tdq_cpu_idle update is globally visible
2881 		 * before cpu_idle() read tdq_load.  The order is important
2882 		 * to avoid race with tdq_notify.
2883 		 */
2884 		atomic_thread_fence_seq_cst();
2885 		/*
2886 		 * Checking for again after the fence picks up assigned
2887 		 * threads often enough to make it worthwhile to do so in
2888 		 * order to avoid calling cpu_idle().
2889 		 */
2890 		if (tdq->tdq_load != 0) {
2891 			tdq->tdq_cpu_idle = 0;
2892 			continue;
2893 		}
2894 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2895 		tdq->tdq_cpu_idle = 0;
2896 
2897 		/*
2898 		 * Account thread-less hardware interrupts and
2899 		 * other wakeup reasons equal to context switches.
2900 		 */
2901 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2902 		if (switchcnt != oldswitchcnt)
2903 			continue;
2904 		tdq->tdq_switchcnt++;
2905 		oldswitchcnt++;
2906 	}
2907 }
2908 
2909 /*
2910  * A CPU is entering for the first time or a thread is exiting.
2911  */
2912 void
2913 sched_throw(struct thread *td)
2914 {
2915 	struct thread *newtd;
2916 	struct tdq *tdq;
2917 
2918 	if (td == NULL) {
2919 #ifdef SMP
2920 		PCPU_SET(sched, DPCPU_PTR(tdq));
2921 #endif
2922 		/* Correct spinlock nesting and acquire the correct lock. */
2923 		tdq = TDQ_SELF();
2924 		TDQ_LOCK(tdq);
2925 		spinlock_exit();
2926 		PCPU_SET(switchtime, cpu_ticks());
2927 		PCPU_SET(switchticks, ticks);
2928 		PCPU_GET(idlethread)->td_lock = TDQ_LOCKPTR(tdq);
2929 	} else {
2930 		tdq = TDQ_SELF();
2931 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2932 		tdq_load_rem(tdq, td);
2933 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2934 		td->td_lastcpu = td->td_oncpu;
2935 		td->td_oncpu = NOCPU;
2936 	}
2937 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2938 	newtd = choosethread();
2939 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2940 	cpu_throw(td, newtd);		/* doesn't return */
2941 }
2942 
2943 /*
2944  * This is called from fork_exit().  Just acquire the correct locks and
2945  * let fork do the rest of the work.
2946  */
2947 void
2948 sched_fork_exit(struct thread *td)
2949 {
2950 	struct tdq *tdq;
2951 	int cpuid;
2952 
2953 	/*
2954 	 * Finish setting up thread glue so that it begins execution in a
2955 	 * non-nested critical section with the scheduler lock held.
2956 	 */
2957 	cpuid = PCPU_GET(cpuid);
2958 	tdq = TDQ_SELF();
2959 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2960 	td->td_oncpu = cpuid;
2961 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2962 	lock_profile_obtain_lock_success(
2963 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2964 
2965 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2966 	    "prio:%d", td->td_priority);
2967 	SDT_PROBE0(sched, , , on__cpu);
2968 }
2969 
2970 /*
2971  * Create on first use to catch odd startup conditons.
2972  */
2973 char *
2974 sched_tdname(struct thread *td)
2975 {
2976 #ifdef KTR
2977 	struct td_sched *ts;
2978 
2979 	ts = td_get_sched(td);
2980 	if (ts->ts_name[0] == '\0')
2981 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2982 		    "%s tid %d", td->td_name, td->td_tid);
2983 	return (ts->ts_name);
2984 #else
2985 	return (td->td_name);
2986 #endif
2987 }
2988 
2989 #ifdef KTR
2990 void
2991 sched_clear_tdname(struct thread *td)
2992 {
2993 	struct td_sched *ts;
2994 
2995 	ts = td_get_sched(td);
2996 	ts->ts_name[0] = '\0';
2997 }
2998 #endif
2999 
3000 #ifdef SMP
3001 
3002 /*
3003  * Build the CPU topology dump string. Is recursively called to collect
3004  * the topology tree.
3005  */
3006 static int
3007 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
3008     int indent)
3009 {
3010 	char cpusetbuf[CPUSETBUFSIZ];
3011 	int i, first;
3012 
3013 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
3014 	    "", 1 + indent / 2, cg->cg_level);
3015 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
3016 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
3017 	first = TRUE;
3018 	for (i = 0; i < MAXCPU; i++) {
3019 		if (CPU_ISSET(i, &cg->cg_mask)) {
3020 			if (!first)
3021 				sbuf_printf(sb, ", ");
3022 			else
3023 				first = FALSE;
3024 			sbuf_printf(sb, "%d", i);
3025 		}
3026 	}
3027 	sbuf_printf(sb, "</cpu>\n");
3028 
3029 	if (cg->cg_flags != 0) {
3030 		sbuf_printf(sb, "%*s <flags>", indent, "");
3031 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
3032 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
3033 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
3034 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
3035 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
3036 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
3037 		sbuf_printf(sb, "</flags>\n");
3038 	}
3039 
3040 	if (cg->cg_children > 0) {
3041 		sbuf_printf(sb, "%*s <children>\n", indent, "");
3042 		for (i = 0; i < cg->cg_children; i++)
3043 			sysctl_kern_sched_topology_spec_internal(sb,
3044 			    &cg->cg_child[i], indent+2);
3045 		sbuf_printf(sb, "%*s </children>\n", indent, "");
3046 	}
3047 	sbuf_printf(sb, "%*s</group>\n", indent, "");
3048 	return (0);
3049 }
3050 
3051 /*
3052  * Sysctl handler for retrieving topology dump. It's a wrapper for
3053  * the recursive sysctl_kern_smp_topology_spec_internal().
3054  */
3055 static int
3056 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
3057 {
3058 	struct sbuf *topo;
3059 	int err;
3060 
3061 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
3062 
3063 	topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
3064 	if (topo == NULL)
3065 		return (ENOMEM);
3066 
3067 	sbuf_printf(topo, "<groups>\n");
3068 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
3069 	sbuf_printf(topo, "</groups>\n");
3070 
3071 	if (err == 0) {
3072 		err = sbuf_finish(topo);
3073 	}
3074 	sbuf_delete(topo);
3075 	return (err);
3076 }
3077 
3078 #endif
3079 
3080 static int
3081 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
3082 {
3083 	int error, new_val, period;
3084 
3085 	period = 1000000 / realstathz;
3086 	new_val = period * sched_slice;
3087 	error = sysctl_handle_int(oidp, &new_val, 0, req);
3088 	if (error != 0 || req->newptr == NULL)
3089 		return (error);
3090 	if (new_val <= 0)
3091 		return (EINVAL);
3092 	sched_slice = imax(1, (new_val + period / 2) / period);
3093 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
3094 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
3095 	    realstathz);
3096 	return (0);
3097 }
3098 
3099 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
3100 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
3101     "Scheduler name");
3102 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
3103     NULL, 0, sysctl_kern_quantum, "I",
3104     "Quantum for timeshare threads in microseconds");
3105 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
3106     "Quantum for timeshare threads in stathz ticks");
3107 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
3108     "Interactivity score threshold");
3109 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
3110     &preempt_thresh, 0,
3111     "Maximal (lowest) priority for preemption");
3112 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
3113     "Assign static kernel priorities to sleeping threads");
3114 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
3115     "Number of times idle thread will spin waiting for new work");
3116 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
3117     &sched_idlespinthresh, 0,
3118     "Threshold before we will permit idle thread spinning");
3119 #ifdef SMP
3120 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
3121     "Number of hz ticks to keep thread affinity for");
3122 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
3123     "Enables the long-term load balancer");
3124 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
3125     &balance_interval, 0,
3126     "Average period in stathz ticks to run the long-term balancer");
3127 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
3128     "Attempts to steal work from other cores before idling");
3129 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
3130     "Minimum load on remote CPU before we'll steal");
3131 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit,
3132     0, "Topological distance limit for stealing threads in sched_switch()");
3133 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0,
3134     "Always run the stealer from the idle thread");
3135 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
3136     CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
3137     "XML dump of detected CPU topology");
3138 #endif
3139 
3140 /* ps compat.  All cpu percentages from ULE are weighted. */
3141 static int ccpu = 0;
3142 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
3143