xref: /freebsd/sys/kern/sched_ule.c (revision e7437ae907c89bf85a99c5cbb7ddd194a1ff1354)
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/umtxvar.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_PICKCPU	TDF_SCHED0	/* Thread should pick new CPU. */
200 #define	TDF_SLICEEND	TDF_SCHED2	/* Thread time slice is over. */
201 
202 /*
203  * tickincr:		Converts a stathz tick into a hz domain scaled by
204  *			the shift factor.  Without the shift the error rate
205  *			due to rounding would be unacceptably high.
206  * realstathz:		stathz is sometimes 0 and run off of hz.
207  * sched_slice:		Runtime of each thread before rescheduling.
208  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
209  */
210 static u_int __read_mostly sched_interact = SCHED_INTERACT_THRESH;
211 static int __read_mostly tickincr = 8 << SCHED_TICK_SHIFT;
212 static int __read_mostly realstathz = 127;	/* reset during boot. */
213 static int __read_mostly sched_slice = 10;	/* reset during boot. */
214 static int __read_mostly sched_slice_min = 1;	/* reset during boot. */
215 #ifdef PREEMPTION
216 #ifdef FULL_PREEMPTION
217 static int __read_mostly preempt_thresh = PRI_MAX_IDLE;
218 #else
219 static int __read_mostly preempt_thresh = PRI_MIN_KERN;
220 #endif
221 #else
222 static int __read_mostly preempt_thresh = 0;
223 #endif
224 static int __read_mostly static_boost = PRI_MIN_BATCH;
225 static int __read_mostly sched_idlespins = 10000;
226 static int __read_mostly sched_idlespinthresh = -1;
227 
228 /*
229  * tdq - per processor runqs and statistics.  A mutex synchronizes access to
230  * most fields.  Some fields are loaded or modified without the mutex.
231  *
232  * Locking protocols:
233  * (c)  constant after initialization
234  * (f)  flag, set with the tdq lock held, cleared on local CPU
235  * (l)  all accesses are CPU-local
236  * (ls) stores are performed by the local CPU, loads may be lockless
237  * (t)  all accesses are protected by the tdq mutex
238  * (ts) stores are serialized by the tdq mutex, loads may be lockless
239  */
240 struct tdq {
241 	/*
242 	 * Ordered to improve efficiency of cpu_search() and switch().
243 	 * tdq_lock is padded to avoid false sharing with tdq_load and
244 	 * tdq_cpu_idle.
245 	 */
246 	struct mtx_padalign tdq_lock;	/* run queue lock. */
247 	struct cpu_group *tdq_cg;	/* (c) Pointer to cpu topology. */
248 	struct thread	*tdq_curthread;	/* (t) Current executing thread. */
249 	int		tdq_load;	/* (ts) Aggregate load. */
250 	int		tdq_sysload;	/* (ts) For loadavg, !ITHD load. */
251 	int		tdq_cpu_idle;	/* (ls) cpu_idle() is active. */
252 	int		tdq_transferable; /* (ts) Transferable thread count. */
253 	short		tdq_switchcnt;	/* (l) Switches this tick. */
254 	short		tdq_oldswitchcnt; /* (l) Switches last tick. */
255 	u_char		tdq_lowpri;	/* (ts) Lowest priority thread. */
256 	u_char		tdq_owepreempt;	/* (f) Remote preemption pending. */
257 	u_char		tdq_idx;	/* (t) Current insert index. */
258 	u_char		tdq_ridx;	/* (t) Current removal index. */
259 	int		tdq_id;		/* (c) cpuid. */
260 	struct runq	tdq_realtime;	/* (t) real-time run queue. */
261 	struct runq	tdq_timeshare;	/* (t) timeshare run queue. */
262 	struct runq	tdq_idle;	/* (t) Queue of IDLE threads. */
263 	char		tdq_name[TDQ_NAME_LEN];
264 #ifdef KTR
265 	char		tdq_loadname[TDQ_LOADNAME_LEN];
266 #endif
267 };
268 
269 /* Idle thread states and config. */
270 #define	TDQ_RUNNING	1
271 #define	TDQ_IDLE	2
272 
273 /* Lockless accessors. */
274 #define	TDQ_LOAD(tdq)		atomic_load_int(&(tdq)->tdq_load)
275 #define	TDQ_TRANSFERABLE(tdq)	atomic_load_int(&(tdq)->tdq_transferable)
276 #define	TDQ_SWITCHCNT(tdq)	(atomic_load_short(&(tdq)->tdq_switchcnt) + \
277 				 atomic_load_short(&(tdq)->tdq_oldswitchcnt))
278 #define	TDQ_SWITCHCNT_INC(tdq)	(atomic_store_short(&(tdq)->tdq_switchcnt, \
279 				 atomic_load_short(&(tdq)->tdq_switchcnt) + 1))
280 
281 #ifdef SMP
282 struct cpu_group __read_mostly *cpu_top;		/* CPU topology */
283 
284 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
285 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
286 
287 /*
288  * Run-time tunables.
289  */
290 static int rebalance = 1;
291 static int balance_interval = 128;	/* Default set in sched_initticks(). */
292 static int __read_mostly affinity;
293 static int __read_mostly steal_idle = 1;
294 static int __read_mostly steal_thresh = 2;
295 static int __read_mostly always_steal = 0;
296 static int __read_mostly trysteal_limit = 2;
297 
298 /*
299  * One thread queue per processor.
300  */
301 static struct tdq __read_mostly *balance_tdq;
302 static int balance_ticks;
303 DPCPU_DEFINE_STATIC(struct tdq, tdq);
304 DPCPU_DEFINE_STATIC(uint32_t, randomval);
305 
306 #define	TDQ_SELF()	((struct tdq *)PCPU_GET(sched))
307 #define	TDQ_CPU(x)	(DPCPU_ID_PTR((x), tdq))
308 #define	TDQ_ID(x)	((x)->tdq_id)
309 #else	/* !SMP */
310 static struct tdq	tdq_cpu;
311 
312 #define	TDQ_ID(x)	(0)
313 #define	TDQ_SELF()	(&tdq_cpu)
314 #define	TDQ_CPU(x)	(&tdq_cpu)
315 #endif
316 
317 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
318 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
319 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
320 #define	TDQ_TRYLOCK(t)		mtx_trylock_spin(TDQ_LOCKPTR((t)))
321 #define	TDQ_TRYLOCK_FLAGS(t, f)	mtx_trylock_spin_flags(TDQ_LOCKPTR((t)), (f))
322 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
323 #define	TDQ_LOCKPTR(t)		((struct mtx *)(&(t)->tdq_lock))
324 
325 static void sched_setpreempt(int);
326 static void sched_priority(struct thread *);
327 static void sched_thread_priority(struct thread *, u_char);
328 static int sched_interact_score(struct thread *);
329 static void sched_interact_update(struct thread *);
330 static void sched_interact_fork(struct thread *);
331 static void sched_pctcpu_update(struct td_sched *, int);
332 
333 /* Operations on per processor queues */
334 static struct thread *tdq_choose(struct tdq *);
335 static void tdq_setup(struct tdq *, int i);
336 static void tdq_load_add(struct tdq *, struct thread *);
337 static void tdq_load_rem(struct tdq *, struct thread *);
338 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
339 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
340 static inline int sched_shouldpreempt(int, int, int);
341 static void tdq_print(int cpu);
342 static void runq_print(struct runq *rq);
343 static int tdq_add(struct tdq *, struct thread *, int);
344 #ifdef SMP
345 static int tdq_move(struct tdq *, struct tdq *);
346 static int tdq_idled(struct tdq *);
347 static void tdq_notify(struct tdq *, int lowpri);
348 static struct thread *tdq_steal(struct tdq *, int);
349 static struct thread *runq_steal(struct runq *, int);
350 static int sched_pickcpu(struct thread *, int);
351 static void sched_balance(void);
352 static bool sched_balance_pair(struct tdq *, struct tdq *);
353 static inline struct tdq *sched_setcpu(struct thread *, int, int);
354 static inline void thread_unblock_switch(struct thread *, struct mtx *);
355 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
356 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
357     struct cpu_group *cg, int indent);
358 #endif
359 
360 static void sched_setup(void *dummy);
361 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
362 
363 static void sched_initticks(void *dummy);
364 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
365     NULL);
366 
367 SDT_PROVIDER_DEFINE(sched);
368 
369 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
370     "struct proc *", "uint8_t");
371 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
372     "struct proc *", "void *");
373 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
374     "struct proc *", "void *", "int");
375 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
376     "struct proc *", "uint8_t", "struct thread *");
377 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
378 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
379     "struct proc *");
380 SDT_PROBE_DEFINE(sched, , , on__cpu);
381 SDT_PROBE_DEFINE(sched, , , remain__cpu);
382 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
383     "struct proc *");
384 
385 /*
386  * Print the threads waiting on a run-queue.
387  */
388 static void
389 runq_print(struct runq *rq)
390 {
391 	struct rqhead *rqh;
392 	struct thread *td;
393 	int pri;
394 	int j;
395 	int i;
396 
397 	for (i = 0; i < RQB_LEN; i++) {
398 		printf("\t\trunq bits %d 0x%zx\n",
399 		    i, rq->rq_status.rqb_bits[i]);
400 		for (j = 0; j < RQB_BPW; j++)
401 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
402 				pri = j + (i << RQB_L2BPW);
403 				rqh = &rq->rq_queues[pri];
404 				TAILQ_FOREACH(td, rqh, td_runq) {
405 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
406 					    td, td->td_name, td->td_priority,
407 					    td->td_rqindex, pri);
408 				}
409 			}
410 	}
411 }
412 
413 /*
414  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
415  */
416 static void __unused
417 tdq_print(int cpu)
418 {
419 	struct tdq *tdq;
420 
421 	tdq = TDQ_CPU(cpu);
422 
423 	printf("tdq %d:\n", TDQ_ID(tdq));
424 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
425 	printf("\tLock name:      %s\n", tdq->tdq_name);
426 	printf("\tload:           %d\n", tdq->tdq_load);
427 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
428 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
429 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
430 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
431 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
432 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
433 	printf("\trealtime runq:\n");
434 	runq_print(&tdq->tdq_realtime);
435 	printf("\ttimeshare runq:\n");
436 	runq_print(&tdq->tdq_timeshare);
437 	printf("\tidle runq:\n");
438 	runq_print(&tdq->tdq_idle);
439 }
440 
441 static inline int
442 sched_shouldpreempt(int pri, int cpri, int remote)
443 {
444 	/*
445 	 * If the new priority is not better than the current priority there is
446 	 * nothing to do.
447 	 */
448 	if (pri >= cpri)
449 		return (0);
450 	/*
451 	 * Always preempt idle.
452 	 */
453 	if (cpri >= PRI_MIN_IDLE)
454 		return (1);
455 	/*
456 	 * If preemption is disabled don't preempt others.
457 	 */
458 	if (preempt_thresh == 0)
459 		return (0);
460 	/*
461 	 * Preempt if we exceed the threshold.
462 	 */
463 	if (pri <= preempt_thresh)
464 		return (1);
465 	/*
466 	 * If we're interactive or better and there is non-interactive
467 	 * or worse running preempt only remote processors.
468 	 */
469 	if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
470 		return (1);
471 	return (0);
472 }
473 
474 /*
475  * Add a thread to the actual run-queue.  Keeps transferable counts up to
476  * date with what is actually on the run-queue.  Selects the correct
477  * queue position for timeshare threads.
478  */
479 static __inline void
480 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
481 {
482 	struct td_sched *ts;
483 	u_char pri;
484 
485 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
486 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
487 
488 	pri = td->td_priority;
489 	ts = td_get_sched(td);
490 	TD_SET_RUNQ(td);
491 	if (THREAD_CAN_MIGRATE(td)) {
492 		tdq->tdq_transferable++;
493 		ts->ts_flags |= TSF_XFERABLE;
494 	}
495 	if (pri < PRI_MIN_BATCH) {
496 		ts->ts_runq = &tdq->tdq_realtime;
497 	} else if (pri <= PRI_MAX_BATCH) {
498 		ts->ts_runq = &tdq->tdq_timeshare;
499 		KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
500 			("Invalid priority %d on timeshare runq", pri));
501 		/*
502 		 * This queue contains only priorities between MIN and MAX
503 		 * batch.  Use the whole queue to represent these values.
504 		 */
505 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
506 			pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
507 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
508 			/*
509 			 * This effectively shortens the queue by one so we
510 			 * can have a one slot difference between idx and
511 			 * ridx while we wait for threads to drain.
512 			 */
513 			if (tdq->tdq_ridx != tdq->tdq_idx &&
514 			    pri == tdq->tdq_ridx)
515 				pri = (unsigned char)(pri - 1) % RQ_NQS;
516 		} else
517 			pri = tdq->tdq_ridx;
518 		runq_add_pri(ts->ts_runq, td, pri, flags);
519 		return;
520 	} else
521 		ts->ts_runq = &tdq->tdq_idle;
522 	runq_add(ts->ts_runq, td, flags);
523 }
524 
525 /*
526  * Remove a thread from a run-queue.  This typically happens when a thread
527  * is selected to run.  Running threads are not on the queue and the
528  * transferable count does not reflect them.
529  */
530 static __inline void
531 tdq_runq_rem(struct tdq *tdq, struct thread *td)
532 {
533 	struct td_sched *ts;
534 
535 	ts = td_get_sched(td);
536 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
537 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
538 	KASSERT(ts->ts_runq != NULL,
539 	    ("tdq_runq_remove: thread %p null ts_runq", td));
540 	if (ts->ts_flags & TSF_XFERABLE) {
541 		tdq->tdq_transferable--;
542 		ts->ts_flags &= ~TSF_XFERABLE;
543 	}
544 	if (ts->ts_runq == &tdq->tdq_timeshare) {
545 		if (tdq->tdq_idx != tdq->tdq_ridx)
546 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
547 		else
548 			runq_remove_idx(ts->ts_runq, td, NULL);
549 	} else
550 		runq_remove(ts->ts_runq, td);
551 }
552 
553 /*
554  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
555  * for this thread to the referenced thread queue.
556  */
557 static void
558 tdq_load_add(struct tdq *tdq, struct thread *td)
559 {
560 
561 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
562 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
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  * Remove the load from a thread that is transitioning to a sleep state or
573  * exiting.
574  */
575 static void
576 tdq_load_rem(struct tdq *tdq, struct thread *td)
577 {
578 
579 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
580 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
581 	KASSERT(tdq->tdq_load != 0,
582 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
583 
584 	tdq->tdq_load--;
585 	if ((td->td_flags & TDF_NOLOAD) == 0)
586 		tdq->tdq_sysload--;
587 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
588 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
589 }
590 
591 /*
592  * Bound timeshare latency by decreasing slice size as load increases.  We
593  * consider the maximum latency as the sum of the threads waiting to run
594  * aside from curthread and target no more than sched_slice latency but
595  * no less than sched_slice_min runtime.
596  */
597 static inline int
598 tdq_slice(struct tdq *tdq)
599 {
600 	int load;
601 
602 	/*
603 	 * It is safe to use sys_load here because this is called from
604 	 * contexts where timeshare threads are running and so there
605 	 * cannot be higher priority load in the system.
606 	 */
607 	load = tdq->tdq_sysload - 1;
608 	if (load >= SCHED_SLICE_MIN_DIVISOR)
609 		return (sched_slice_min);
610 	if (load <= 1)
611 		return (sched_slice);
612 	return (sched_slice / load);
613 }
614 
615 /*
616  * Set lowpri to its exact value by searching the run-queue and
617  * evaluating curthread.  curthread may be passed as an optimization.
618  */
619 static void
620 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
621 {
622 	struct thread *td;
623 
624 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
625 	if (ctd == NULL)
626 		ctd = tdq->tdq_curthread;
627 	td = tdq_choose(tdq);
628 	if (td == NULL || td->td_priority > ctd->td_priority)
629 		tdq->tdq_lowpri = ctd->td_priority;
630 	else
631 		tdq->tdq_lowpri = td->td_priority;
632 }
633 
634 #ifdef SMP
635 /*
636  * We need some randomness. Implement a classic Linear Congruential
637  * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
638  * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
639  * of the random state (in the low bits of our answer) to keep
640  * the maximum randomness.
641  */
642 static uint32_t
643 sched_random(void)
644 {
645 	uint32_t *rndptr;
646 
647 	rndptr = DPCPU_PTR(randomval);
648 	*rndptr = *rndptr * 69069 + 5;
649 
650 	return (*rndptr >> 16);
651 }
652 
653 struct cpu_search {
654 	cpuset_t *cs_mask;	/* The mask of allowed CPUs to choose from. */
655 	int	cs_prefer;	/* Prefer this CPU and groups including it. */
656 	int	cs_running;	/* The thread is now running at cs_prefer. */
657 	int	cs_pri;		/* Min priority for low. */
658 	int	cs_load;	/* Max load for low, min load for high. */
659 	int	cs_trans;	/* Min transferable load for high. */
660 };
661 
662 struct cpu_search_res {
663 	int	csr_cpu;	/* The best CPU found. */
664 	int	csr_load;	/* The load of cs_cpu. */
665 };
666 
667 /*
668  * Search the tree of cpu_groups for the lowest or highest loaded CPU.
669  * These routines actually compare the load on all paths through the tree
670  * and find the least loaded cpu on the least loaded path, which may differ
671  * from the least loaded cpu in the system.  This balances work among caches
672  * and buses.
673  */
674 static int
675 cpu_search_lowest(const struct cpu_group *cg, const struct cpu_search *s,
676     struct cpu_search_res *r)
677 {
678 	struct cpu_search_res lr;
679 	struct tdq *tdq;
680 	int c, bload, l, load, p, total;
681 
682 	total = 0;
683 	bload = INT_MAX;
684 	r->csr_cpu = -1;
685 
686 	/* Loop through children CPU groups if there are any. */
687 	if (cg->cg_children > 0) {
688 		for (c = cg->cg_children - 1; c >= 0; c--) {
689 			load = cpu_search_lowest(&cg->cg_child[c], s, &lr);
690 			total += load;
691 
692 			/*
693 			 * When balancing do not prefer SMT groups with load >1.
694 			 * It allows round-robin between SMT groups with equal
695 			 * load within parent group for more fair scheduling.
696 			 */
697 			if (__predict_false(s->cs_running) &&
698 			    (cg->cg_child[c].cg_flags & CG_FLAG_THREAD) &&
699 			    load >= 128 && (load & 128) != 0)
700 				load += 128;
701 
702 			if (lr.csr_cpu >= 0 && (load < bload ||
703 			    (load == bload && lr.csr_load < r->csr_load))) {
704 				bload = load;
705 				r->csr_cpu = lr.csr_cpu;
706 				r->csr_load = lr.csr_load;
707 			}
708 		}
709 		return (total);
710 	}
711 
712 	/* Loop through children CPUs otherwise. */
713 	for (c = cg->cg_last; c >= cg->cg_first; c--) {
714 		if (!CPU_ISSET(c, &cg->cg_mask))
715 			continue;
716 		tdq = TDQ_CPU(c);
717 		l = TDQ_LOAD(tdq);
718 		if (c == s->cs_prefer) {
719 			if (__predict_false(s->cs_running))
720 				l--;
721 			p = 128;
722 		} else
723 			p = 0;
724 		load = l * 256;
725 		total += load - p;
726 
727 		/*
728 		 * Check this CPU is acceptable.
729 		 * If the threads is already on the CPU, don't look on the TDQ
730 		 * priority, since it can be the priority of the thread itself.
731 		 */
732 		if (l > s->cs_load ||
733 		    (atomic_load_char(&tdq->tdq_lowpri) <= s->cs_pri &&
734 		     (!s->cs_running || c != s->cs_prefer)) ||
735 		    !CPU_ISSET(c, s->cs_mask))
736 			continue;
737 
738 		/*
739 		 * When balancing do not prefer CPUs with load > 1.
740 		 * It allows round-robin between CPUs with equal load
741 		 * within the CPU group for more fair scheduling.
742 		 */
743 		if (__predict_false(s->cs_running) && l > 0)
744 			p = 0;
745 
746 		load -= sched_random() % 128;
747 		if (bload > load - p) {
748 			bload = load - p;
749 			r->csr_cpu = c;
750 			r->csr_load = load;
751 		}
752 	}
753 	return (total);
754 }
755 
756 static int
757 cpu_search_highest(const struct cpu_group *cg, const struct cpu_search *s,
758     struct cpu_search_res *r)
759 {
760 	struct cpu_search_res lr;
761 	struct tdq *tdq;
762 	int c, bload, l, load, total;
763 
764 	total = 0;
765 	bload = INT_MIN;
766 	r->csr_cpu = -1;
767 
768 	/* Loop through children CPU groups if there are any. */
769 	if (cg->cg_children > 0) {
770 		for (c = cg->cg_children - 1; c >= 0; c--) {
771 			load = cpu_search_highest(&cg->cg_child[c], s, &lr);
772 			total += load;
773 			if (lr.csr_cpu >= 0 && (load > bload ||
774 			    (load == bload && lr.csr_load > r->csr_load))) {
775 				bload = load;
776 				r->csr_cpu = lr.csr_cpu;
777 				r->csr_load = lr.csr_load;
778 			}
779 		}
780 		return (total);
781 	}
782 
783 	/* Loop through children CPUs otherwise. */
784 	for (c = cg->cg_last; c >= cg->cg_first; c--) {
785 		if (!CPU_ISSET(c, &cg->cg_mask))
786 			continue;
787 		tdq = TDQ_CPU(c);
788 		l = TDQ_LOAD(tdq);
789 		load = l * 256;
790 		total += load;
791 
792 		/*
793 		 * Check this CPU is acceptable.
794 		 */
795 		if (l < s->cs_load || TDQ_TRANSFERABLE(tdq) < s->cs_trans ||
796 		    !CPU_ISSET(c, s->cs_mask))
797 			continue;
798 
799 		load -= sched_random() % 256;
800 		if (load > bload) {
801 			bload = load;
802 			r->csr_cpu = c;
803 		}
804 	}
805 	r->csr_load = bload;
806 	return (total);
807 }
808 
809 /*
810  * Find the cpu with the least load via the least loaded path that has a
811  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
812  * acceptable.
813  */
814 static inline int
815 sched_lowest(const struct cpu_group *cg, cpuset_t *mask, int pri, int maxload,
816     int prefer, int running)
817 {
818 	struct cpu_search s;
819 	struct cpu_search_res r;
820 
821 	s.cs_prefer = prefer;
822 	s.cs_running = running;
823 	s.cs_mask = mask;
824 	s.cs_pri = pri;
825 	s.cs_load = maxload;
826 	cpu_search_lowest(cg, &s, &r);
827 	return (r.csr_cpu);
828 }
829 
830 /*
831  * Find the cpu with the highest load via the highest loaded path.
832  */
833 static inline int
834 sched_highest(const struct cpu_group *cg, cpuset_t *mask, int minload,
835     int mintrans)
836 {
837 	struct cpu_search s;
838 	struct cpu_search_res r;
839 
840 	s.cs_mask = mask;
841 	s.cs_load = minload;
842 	s.cs_trans = mintrans;
843 	cpu_search_highest(cg, &s, &r);
844 	return (r.csr_cpu);
845 }
846 
847 static void
848 sched_balance_group(struct cpu_group *cg)
849 {
850 	struct tdq *tdq;
851 	struct thread *td;
852 	cpuset_t hmask, lmask;
853 	int high, low, anylow;
854 
855 	CPU_FILL(&hmask);
856 	for (;;) {
857 		high = sched_highest(cg, &hmask, 1, 0);
858 		/* Stop if there is no more CPU with transferrable threads. */
859 		if (high == -1)
860 			break;
861 		CPU_CLR(high, &hmask);
862 		CPU_COPY(&hmask, &lmask);
863 		/* Stop if there is no more CPU left for low. */
864 		if (CPU_EMPTY(&lmask))
865 			break;
866 		tdq = TDQ_CPU(high);
867 		if (TDQ_LOAD(tdq) == 1) {
868 			/*
869 			 * There is only one running thread.  We can't move
870 			 * it from here, so tell it to pick new CPU by itself.
871 			 */
872 			TDQ_LOCK(tdq);
873 			td = tdq->tdq_curthread;
874 			if (td->td_lock == TDQ_LOCKPTR(tdq) &&
875 			    (td->td_flags & TDF_IDLETD) == 0 &&
876 			    THREAD_CAN_MIGRATE(td)) {
877 				td->td_flags |= TDF_NEEDRESCHED | TDF_PICKCPU;
878 				if (high != curcpu)
879 					ipi_cpu(high, IPI_AST);
880 			}
881 			TDQ_UNLOCK(tdq);
882 			break;
883 		}
884 		anylow = 1;
885 nextlow:
886 		if (TDQ_TRANSFERABLE(tdq) == 0)
887 			continue;
888 		low = sched_lowest(cg, &lmask, -1, TDQ_LOAD(tdq) - 1, high, 1);
889 		/* Stop if we looked well and found no less loaded CPU. */
890 		if (anylow && low == -1)
891 			break;
892 		/* Go to next high if we found no less loaded CPU. */
893 		if (low == -1)
894 			continue;
895 		/* Transfer thread from high to low. */
896 		if (sched_balance_pair(tdq, TDQ_CPU(low))) {
897 			/* CPU that got thread can no longer be a donor. */
898 			CPU_CLR(low, &hmask);
899 		} else {
900 			/*
901 			 * If failed, then there is no threads on high
902 			 * that can run on this low. Drop low from low
903 			 * mask and look for different one.
904 			 */
905 			CPU_CLR(low, &lmask);
906 			anylow = 0;
907 			goto nextlow;
908 		}
909 	}
910 }
911 
912 static void
913 sched_balance(void)
914 {
915 	struct tdq *tdq;
916 
917 	balance_ticks = max(balance_interval / 2, 1) +
918 	    (sched_random() % balance_interval);
919 	tdq = TDQ_SELF();
920 	TDQ_UNLOCK(tdq);
921 	sched_balance_group(cpu_top);
922 	TDQ_LOCK(tdq);
923 }
924 
925 /*
926  * Lock two thread queues using their address to maintain lock order.
927  */
928 static void
929 tdq_lock_pair(struct tdq *one, struct tdq *two)
930 {
931 	if (one < two) {
932 		TDQ_LOCK(one);
933 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
934 	} else {
935 		TDQ_LOCK(two);
936 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
937 	}
938 }
939 
940 /*
941  * Unlock two thread queues.  Order is not important here.
942  */
943 static void
944 tdq_unlock_pair(struct tdq *one, struct tdq *two)
945 {
946 	TDQ_UNLOCK(one);
947 	TDQ_UNLOCK(two);
948 }
949 
950 /*
951  * Transfer load between two imbalanced thread queues.  Returns true if a thread
952  * was moved between the queues, and false otherwise.
953  */
954 static bool
955 sched_balance_pair(struct tdq *high, struct tdq *low)
956 {
957 	int cpu, lowpri;
958 	bool ret;
959 
960 	ret = false;
961 	tdq_lock_pair(high, low);
962 
963 	/*
964 	 * Transfer a thread from high to low.
965 	 */
966 	if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load) {
967 		lowpri = tdq_move(high, low);
968 		if (lowpri != -1) {
969 			/*
970 			 * In case the target isn't the current CPU notify it of
971 			 * the new load, possibly sending an IPI to force it to
972 			 * reschedule.  Otherwise maybe schedule a preemption.
973 			 */
974 			cpu = TDQ_ID(low);
975 			if (cpu != PCPU_GET(cpuid))
976 				tdq_notify(low, lowpri);
977 			else
978 				sched_setpreempt(low->tdq_lowpri);
979 			ret = true;
980 		}
981 	}
982 	tdq_unlock_pair(high, low);
983 	return (ret);
984 }
985 
986 /*
987  * Move a thread from one thread queue to another.  Returns -1 if the source
988  * queue was empty, else returns the maximum priority of all threads in
989  * the destination queue prior to the addition of the new thread.  In the latter
990  * case, this priority can be used to determine whether an IPI needs to be
991  * delivered.
992  */
993 static int
994 tdq_move(struct tdq *from, struct tdq *to)
995 {
996 	struct thread *td;
997 	int cpu;
998 
999 	TDQ_LOCK_ASSERT(from, MA_OWNED);
1000 	TDQ_LOCK_ASSERT(to, MA_OWNED);
1001 
1002 	cpu = TDQ_ID(to);
1003 	td = tdq_steal(from, cpu);
1004 	if (td == NULL)
1005 		return (-1);
1006 
1007 	/*
1008 	 * Although the run queue is locked the thread may be
1009 	 * blocked.  We can not set the lock until it is unblocked.
1010 	 */
1011 	thread_lock_block_wait(td);
1012 	sched_rem(td);
1013 	THREAD_LOCKPTR_ASSERT(td, TDQ_LOCKPTR(from));
1014 	td->td_lock = TDQ_LOCKPTR(to);
1015 	td_get_sched(td)->ts_cpu = cpu;
1016 	return (tdq_add(to, td, SRQ_YIELDING));
1017 }
1018 
1019 /*
1020  * This tdq has idled.  Try to steal a thread from another cpu and switch
1021  * to it.
1022  */
1023 static int
1024 tdq_idled(struct tdq *tdq)
1025 {
1026 	struct cpu_group *cg, *parent;
1027 	struct tdq *steal;
1028 	cpuset_t mask;
1029 	int cpu, switchcnt, goup;
1030 
1031 	if (smp_started == 0 || steal_idle == 0 || tdq->tdq_cg == NULL)
1032 		return (1);
1033 	CPU_FILL(&mask);
1034 	CPU_CLR(PCPU_GET(cpuid), &mask);
1035 restart:
1036 	switchcnt = TDQ_SWITCHCNT(tdq);
1037 	for (cg = tdq->tdq_cg, goup = 0; ; ) {
1038 		cpu = sched_highest(cg, &mask, steal_thresh, 1);
1039 		/*
1040 		 * We were assigned a thread but not preempted.  Returning
1041 		 * 0 here will cause our caller to switch to it.
1042 		 */
1043 		if (TDQ_LOAD(tdq))
1044 			return (0);
1045 
1046 		/*
1047 		 * We found no CPU to steal from in this group.  Escalate to
1048 		 * the parent and repeat.  But if parent has only two children
1049 		 * groups we can avoid searching this group again by searching
1050 		 * the other one specifically and then escalating two levels.
1051 		 */
1052 		if (cpu == -1) {
1053 			if (goup) {
1054 				cg = cg->cg_parent;
1055 				goup = 0;
1056 			}
1057 			parent = cg->cg_parent;
1058 			if (parent == NULL)
1059 				return (1);
1060 			if (parent->cg_children == 2) {
1061 				if (cg == &parent->cg_child[0])
1062 					cg = &parent->cg_child[1];
1063 				else
1064 					cg = &parent->cg_child[0];
1065 				goup = 1;
1066 			} else
1067 				cg = parent;
1068 			continue;
1069 		}
1070 		steal = TDQ_CPU(cpu);
1071 		/*
1072 		 * The data returned by sched_highest() is stale and
1073 		 * the chosen CPU no longer has an eligible thread.
1074 		 *
1075 		 * Testing this ahead of tdq_lock_pair() only catches
1076 		 * this situation about 20% of the time on an 8 core
1077 		 * 16 thread Ryzen 7, but it still helps performance.
1078 		 */
1079 		if (TDQ_LOAD(steal) < steal_thresh ||
1080 		    TDQ_TRANSFERABLE(steal) == 0)
1081 			goto restart;
1082 		/*
1083 		 * Try to lock both queues. If we are assigned a thread while
1084 		 * waited for the lock, switch to it now instead of stealing.
1085 		 * If we can't get the lock, then somebody likely got there
1086 		 * first so continue searching.
1087 		 */
1088 		TDQ_LOCK(tdq);
1089 		if (tdq->tdq_load > 0) {
1090 			mi_switch(SW_VOL | SWT_IDLE);
1091 			return (0);
1092 		}
1093 		if (TDQ_TRYLOCK_FLAGS(steal, MTX_DUPOK) == 0) {
1094 			TDQ_UNLOCK(tdq);
1095 			CPU_CLR(cpu, &mask);
1096 			continue;
1097 		}
1098 		/*
1099 		 * The data returned by sched_highest() is stale and
1100 		 * the chosen CPU no longer has an eligible thread, or
1101 		 * we were preempted and the CPU loading info may be out
1102 		 * of date.  The latter is rare.  In either case restart
1103 		 * the search.
1104 		 */
1105 		if (TDQ_LOAD(steal) < steal_thresh ||
1106 		    TDQ_TRANSFERABLE(steal) == 0 ||
1107 		    switchcnt != TDQ_SWITCHCNT(tdq)) {
1108 			tdq_unlock_pair(tdq, steal);
1109 			goto restart;
1110 		}
1111 		/*
1112 		 * Steal the thread and switch to it.
1113 		 */
1114 		if (tdq_move(steal, tdq) != -1)
1115 			break;
1116 		/*
1117 		 * We failed to acquire a thread even though it looked
1118 		 * like one was available.  This could be due to affinity
1119 		 * restrictions or for other reasons.  Loop again after
1120 		 * removing this CPU from the set.  The restart logic
1121 		 * above does not restore this CPU to the set due to the
1122 		 * likelyhood of failing here again.
1123 		 */
1124 		CPU_CLR(cpu, &mask);
1125 		tdq_unlock_pair(tdq, steal);
1126 	}
1127 	TDQ_UNLOCK(steal);
1128 	mi_switch(SW_VOL | SWT_IDLE);
1129 	return (0);
1130 }
1131 
1132 /*
1133  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
1134  *
1135  * "lowpri" is the minimum scheduling priority among all threads on
1136  * the queue prior to the addition of the new thread.
1137  */
1138 static void
1139 tdq_notify(struct tdq *tdq, int lowpri)
1140 {
1141 	int cpu;
1142 
1143 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1144 	KASSERT(tdq->tdq_lowpri <= lowpri,
1145 	    ("tdq_notify: lowpri %d > tdq_lowpri %d", lowpri, tdq->tdq_lowpri));
1146 
1147 	if (tdq->tdq_owepreempt)
1148 		return;
1149 
1150 	/*
1151 	 * Check to see if the newly added thread should preempt the one
1152 	 * currently running.
1153 	 */
1154 	if (!sched_shouldpreempt(tdq->tdq_lowpri, lowpri, 1))
1155 		return;
1156 
1157 	/*
1158 	 * Make sure that our caller's earlier update to tdq_load is
1159 	 * globally visible before we read tdq_cpu_idle.  Idle thread
1160 	 * accesses both of them without locks, and the order is important.
1161 	 */
1162 	atomic_thread_fence_seq_cst();
1163 
1164 	/*
1165 	 * Try to figure out if we can signal the idle thread instead of sending
1166 	 * an IPI.  This check is racy; at worst, we will deliever an IPI
1167 	 * unnecessarily.
1168 	 */
1169 	cpu = TDQ_ID(tdq);
1170 	if (TD_IS_IDLETHREAD(tdq->tdq_curthread) &&
1171 	    (atomic_load_int(&tdq->tdq_cpu_idle) == 0 || cpu_idle_wakeup(cpu)))
1172 		return;
1173 
1174 	/*
1175 	 * The run queues have been updated, so any switch on the remote CPU
1176 	 * will satisfy the preemption request.
1177 	 */
1178 	tdq->tdq_owepreempt = 1;
1179 	ipi_cpu(cpu, IPI_PREEMPT);
1180 }
1181 
1182 /*
1183  * Steals load from a timeshare queue.  Honors the rotating queue head
1184  * index.
1185  */
1186 static struct thread *
1187 runq_steal_from(struct runq *rq, int cpu, u_char start)
1188 {
1189 	struct rqbits *rqb;
1190 	struct rqhead *rqh;
1191 	struct thread *td, *first;
1192 	int bit;
1193 	int i;
1194 
1195 	rqb = &rq->rq_status;
1196 	bit = start & (RQB_BPW -1);
1197 	first = NULL;
1198 again:
1199 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1200 		if (rqb->rqb_bits[i] == 0)
1201 			continue;
1202 		if (bit == 0)
1203 			bit = RQB_FFS(rqb->rqb_bits[i]);
1204 		for (; bit < RQB_BPW; bit++) {
1205 			if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1206 				continue;
1207 			rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1208 			TAILQ_FOREACH(td, rqh, td_runq) {
1209 				if (first) {
1210 					if (THREAD_CAN_MIGRATE(td) &&
1211 					    THREAD_CAN_SCHED(td, cpu))
1212 						return (td);
1213 				} else
1214 					first = td;
1215 			}
1216 		}
1217 	}
1218 	if (start != 0) {
1219 		start = 0;
1220 		goto again;
1221 	}
1222 
1223 	if (first && THREAD_CAN_MIGRATE(first) &&
1224 	    THREAD_CAN_SCHED(first, cpu))
1225 		return (first);
1226 	return (NULL);
1227 }
1228 
1229 /*
1230  * Steals load from a standard linear queue.
1231  */
1232 static struct thread *
1233 runq_steal(struct runq *rq, int cpu)
1234 {
1235 	struct rqhead *rqh;
1236 	struct rqbits *rqb;
1237 	struct thread *td;
1238 	int word;
1239 	int bit;
1240 
1241 	rqb = &rq->rq_status;
1242 	for (word = 0; word < RQB_LEN; word++) {
1243 		if (rqb->rqb_bits[word] == 0)
1244 			continue;
1245 		for (bit = 0; bit < RQB_BPW; bit++) {
1246 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1247 				continue;
1248 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1249 			TAILQ_FOREACH(td, rqh, td_runq)
1250 				if (THREAD_CAN_MIGRATE(td) &&
1251 				    THREAD_CAN_SCHED(td, cpu))
1252 					return (td);
1253 		}
1254 	}
1255 	return (NULL);
1256 }
1257 
1258 /*
1259  * Attempt to steal a thread in priority order from a thread queue.
1260  */
1261 static struct thread *
1262 tdq_steal(struct tdq *tdq, int cpu)
1263 {
1264 	struct thread *td;
1265 
1266 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1267 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1268 		return (td);
1269 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1270 	    cpu, tdq->tdq_ridx)) != NULL)
1271 		return (td);
1272 	return (runq_steal(&tdq->tdq_idle, cpu));
1273 }
1274 
1275 /*
1276  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1277  * current lock and returns with the assigned queue locked.
1278  */
1279 static inline struct tdq *
1280 sched_setcpu(struct thread *td, int cpu, int flags)
1281 {
1282 
1283 	struct tdq *tdq;
1284 	struct mtx *mtx;
1285 
1286 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1287 	tdq = TDQ_CPU(cpu);
1288 	td_get_sched(td)->ts_cpu = cpu;
1289 	/*
1290 	 * If the lock matches just return the queue.
1291 	 */
1292 	if (td->td_lock == TDQ_LOCKPTR(tdq)) {
1293 		KASSERT((flags & SRQ_HOLD) == 0,
1294 		    ("sched_setcpu: Invalid lock for SRQ_HOLD"));
1295 		return (tdq);
1296 	}
1297 
1298 	/*
1299 	 * The hard case, migration, we need to block the thread first to
1300 	 * prevent order reversals with other cpus locks.
1301 	 */
1302 	spinlock_enter();
1303 	mtx = thread_lock_block(td);
1304 	if ((flags & SRQ_HOLD) == 0)
1305 		mtx_unlock_spin(mtx);
1306 	TDQ_LOCK(tdq);
1307 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1308 	spinlock_exit();
1309 	return (tdq);
1310 }
1311 
1312 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1313 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1314 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1315 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1316 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1317 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1318 
1319 static int
1320 sched_pickcpu(struct thread *td, int flags)
1321 {
1322 	struct cpu_group *cg, *ccg;
1323 	struct td_sched *ts;
1324 	struct tdq *tdq;
1325 	cpuset_t *mask;
1326 	int cpu, pri, r, self, intr;
1327 
1328 	self = PCPU_GET(cpuid);
1329 	ts = td_get_sched(td);
1330 	KASSERT(!CPU_ABSENT(ts->ts_cpu), ("sched_pickcpu: Start scheduler on "
1331 	    "absent CPU %d for thread %s.", ts->ts_cpu, td->td_name));
1332 	if (smp_started == 0)
1333 		return (self);
1334 	/*
1335 	 * Don't migrate a running thread from sched_switch().
1336 	 */
1337 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1338 		return (ts->ts_cpu);
1339 	/*
1340 	 * Prefer to run interrupt threads on the processors that generate
1341 	 * the interrupt.
1342 	 */
1343 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1344 	    curthread->td_intr_nesting_level) {
1345 		tdq = TDQ_SELF();
1346 		if (tdq->tdq_lowpri >= PRI_MIN_IDLE) {
1347 			SCHED_STAT_INC(pickcpu_idle_affinity);
1348 			return (self);
1349 		}
1350 		ts->ts_cpu = self;
1351 		intr = 1;
1352 		cg = tdq->tdq_cg;
1353 		goto llc;
1354 	} else {
1355 		intr = 0;
1356 		tdq = TDQ_CPU(ts->ts_cpu);
1357 		cg = tdq->tdq_cg;
1358 	}
1359 	/*
1360 	 * If the thread can run on the last cpu and the affinity has not
1361 	 * expired and it is idle, run it there.
1362 	 */
1363 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1364 	    atomic_load_char(&tdq->tdq_lowpri) >= PRI_MIN_IDLE &&
1365 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1366 		if (cg->cg_flags & CG_FLAG_THREAD) {
1367 			/* Check all SMT threads for being idle. */
1368 			for (cpu = cg->cg_first; cpu <= cg->cg_last; cpu++) {
1369 				pri =
1370 				    atomic_load_char(&TDQ_CPU(cpu)->tdq_lowpri);
1371 				if (CPU_ISSET(cpu, &cg->cg_mask) &&
1372 				    pri < PRI_MIN_IDLE)
1373 					break;
1374 			}
1375 			if (cpu > cg->cg_last) {
1376 				SCHED_STAT_INC(pickcpu_idle_affinity);
1377 				return (ts->ts_cpu);
1378 			}
1379 		} else {
1380 			SCHED_STAT_INC(pickcpu_idle_affinity);
1381 			return (ts->ts_cpu);
1382 		}
1383 	}
1384 llc:
1385 	/*
1386 	 * Search for the last level cache CPU group in the tree.
1387 	 * Skip SMT, identical groups and caches with expired affinity.
1388 	 * Interrupt threads affinity is explicit and never expires.
1389 	 */
1390 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1391 		if (cg->cg_flags & CG_FLAG_THREAD)
1392 			continue;
1393 		if (cg->cg_children == 1 || cg->cg_count == 1)
1394 			continue;
1395 		if (cg->cg_level == CG_SHARE_NONE ||
1396 		    (!intr && !SCHED_AFFINITY(ts, cg->cg_level)))
1397 			continue;
1398 		ccg = cg;
1399 	}
1400 	/* Found LLC shared by all CPUs, so do a global search. */
1401 	if (ccg == cpu_top)
1402 		ccg = NULL;
1403 	cpu = -1;
1404 	mask = &td->td_cpuset->cs_mask;
1405 	pri = td->td_priority;
1406 	r = TD_IS_RUNNING(td);
1407 	/*
1408 	 * Try hard to keep interrupts within found LLC.  Search the LLC for
1409 	 * the least loaded CPU we can run now.  For NUMA systems it should
1410 	 * be within target domain, and it also reduces scheduling overhead.
1411 	 */
1412 	if (ccg != NULL && intr) {
1413 		cpu = sched_lowest(ccg, mask, pri, INT_MAX, ts->ts_cpu, r);
1414 		if (cpu >= 0)
1415 			SCHED_STAT_INC(pickcpu_intrbind);
1416 	} else
1417 	/* Search the LLC for the least loaded idle CPU we can run now. */
1418 	if (ccg != NULL) {
1419 		cpu = sched_lowest(ccg, mask, max(pri, PRI_MAX_TIMESHARE),
1420 		    INT_MAX, ts->ts_cpu, r);
1421 		if (cpu >= 0)
1422 			SCHED_STAT_INC(pickcpu_affinity);
1423 	}
1424 	/* Search globally for the least loaded CPU we can run now. */
1425 	if (cpu < 0) {
1426 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu, r);
1427 		if (cpu >= 0)
1428 			SCHED_STAT_INC(pickcpu_lowest);
1429 	}
1430 	/* Search globally for the least loaded CPU. */
1431 	if (cpu < 0) {
1432 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu, r);
1433 		if (cpu >= 0)
1434 			SCHED_STAT_INC(pickcpu_lowest);
1435 	}
1436 	KASSERT(cpu >= 0, ("sched_pickcpu: Failed to find a cpu."));
1437 	KASSERT(!CPU_ABSENT(cpu), ("sched_pickcpu: Picked absent CPU %d.", cpu));
1438 	/*
1439 	 * Compare the lowest loaded cpu to current cpu.
1440 	 */
1441 	tdq = TDQ_CPU(cpu);
1442 	if (THREAD_CAN_SCHED(td, self) && TDQ_SELF()->tdq_lowpri > pri &&
1443 	    atomic_load_char(&tdq->tdq_lowpri) < PRI_MIN_IDLE &&
1444 	    TDQ_LOAD(TDQ_SELF()) <= TDQ_LOAD(tdq) + 1) {
1445 		SCHED_STAT_INC(pickcpu_local);
1446 		cpu = self;
1447 	}
1448 	if (cpu != ts->ts_cpu)
1449 		SCHED_STAT_INC(pickcpu_migration);
1450 	return (cpu);
1451 }
1452 #endif
1453 
1454 /*
1455  * Pick the highest priority task we have and return it.
1456  */
1457 static struct thread *
1458 tdq_choose(struct tdq *tdq)
1459 {
1460 	struct thread *td;
1461 
1462 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1463 	td = runq_choose(&tdq->tdq_realtime);
1464 	if (td != NULL)
1465 		return (td);
1466 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1467 	if (td != NULL) {
1468 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1469 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1470 		    td->td_priority));
1471 		return (td);
1472 	}
1473 	td = runq_choose(&tdq->tdq_idle);
1474 	if (td != NULL) {
1475 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1476 		    ("tdq_choose: Invalid priority on idle queue %d",
1477 		    td->td_priority));
1478 		return (td);
1479 	}
1480 
1481 	return (NULL);
1482 }
1483 
1484 /*
1485  * Initialize a thread queue.
1486  */
1487 static void
1488 tdq_setup(struct tdq *tdq, int id)
1489 {
1490 
1491 	if (bootverbose)
1492 		printf("ULE: setup cpu %d\n", id);
1493 	runq_init(&tdq->tdq_realtime);
1494 	runq_init(&tdq->tdq_timeshare);
1495 	runq_init(&tdq->tdq_idle);
1496 	tdq->tdq_id = id;
1497 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1498 	    "sched lock %d", (int)TDQ_ID(tdq));
1499 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", MTX_SPIN);
1500 #ifdef KTR
1501 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1502 	    "CPU %d load", (int)TDQ_ID(tdq));
1503 #endif
1504 }
1505 
1506 #ifdef SMP
1507 static void
1508 sched_setup_smp(void)
1509 {
1510 	struct tdq *tdq;
1511 	int i;
1512 
1513 	cpu_top = smp_topo();
1514 	CPU_FOREACH(i) {
1515 		tdq = DPCPU_ID_PTR(i, tdq);
1516 		tdq_setup(tdq, i);
1517 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1518 		if (tdq->tdq_cg == NULL)
1519 			panic("Can't find cpu group for %d\n", i);
1520 		DPCPU_ID_SET(i, randomval, i * 69069 + 5);
1521 	}
1522 	PCPU_SET(sched, DPCPU_PTR(tdq));
1523 	balance_tdq = TDQ_SELF();
1524 }
1525 #endif
1526 
1527 /*
1528  * Setup the thread queues and initialize the topology based on MD
1529  * information.
1530  */
1531 static void
1532 sched_setup(void *dummy)
1533 {
1534 	struct tdq *tdq;
1535 
1536 #ifdef SMP
1537 	sched_setup_smp();
1538 #else
1539 	tdq_setup(TDQ_SELF(), 0);
1540 #endif
1541 	tdq = TDQ_SELF();
1542 
1543 	/* Add thread0's load since it's running. */
1544 	TDQ_LOCK(tdq);
1545 	thread0.td_lock = TDQ_LOCKPTR(tdq);
1546 	tdq_load_add(tdq, &thread0);
1547 	tdq->tdq_curthread = &thread0;
1548 	tdq->tdq_lowpri = thread0.td_priority;
1549 	TDQ_UNLOCK(tdq);
1550 }
1551 
1552 /*
1553  * This routine determines time constants after stathz and hz are setup.
1554  */
1555 /* ARGSUSED */
1556 static void
1557 sched_initticks(void *dummy)
1558 {
1559 	int incr;
1560 
1561 	realstathz = stathz ? stathz : hz;
1562 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1563 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1564 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1565 	    realstathz);
1566 
1567 	/*
1568 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1569 	 * hz not being evenly divisible by stathz on all platforms.
1570 	 */
1571 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1572 	/*
1573 	 * This does not work for values of stathz that are more than
1574 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1575 	 */
1576 	if (incr == 0)
1577 		incr = 1;
1578 	tickincr = incr;
1579 #ifdef SMP
1580 	/*
1581 	 * Set the default balance interval now that we know
1582 	 * what realstathz is.
1583 	 */
1584 	balance_interval = realstathz;
1585 	balance_ticks = balance_interval;
1586 	affinity = SCHED_AFFINITY_DEFAULT;
1587 #endif
1588 	if (sched_idlespinthresh < 0)
1589 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1590 }
1591 
1592 /*
1593  * This is the core of the interactivity algorithm.  Determines a score based
1594  * on past behavior.  It is the ratio of sleep time to run time scaled to
1595  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1596  * differs from the cpu usage because it does not account for time spent
1597  * waiting on a run-queue.  Would be prettier if we had floating point.
1598  *
1599  * When a thread's sleep time is greater than its run time the
1600  * calculation is:
1601  *
1602  *                           scaling factor
1603  * interactivity score =  ---------------------
1604  *                        sleep time / run time
1605  *
1606  *
1607  * When a thread's run time is greater than its sleep time the
1608  * calculation is:
1609  *
1610  *                                                 scaling factor
1611  * interactivity score = 2 * scaling factor  -  ---------------------
1612  *                                              run time / sleep time
1613  */
1614 static int
1615 sched_interact_score(struct thread *td)
1616 {
1617 	struct td_sched *ts;
1618 	int div;
1619 
1620 	ts = td_get_sched(td);
1621 	/*
1622 	 * The score is only needed if this is likely to be an interactive
1623 	 * task.  Don't go through the expense of computing it if there's
1624 	 * no chance.
1625 	 */
1626 	if (sched_interact <= SCHED_INTERACT_HALF &&
1627 		ts->ts_runtime >= ts->ts_slptime)
1628 			return (SCHED_INTERACT_HALF);
1629 
1630 	if (ts->ts_runtime > ts->ts_slptime) {
1631 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1632 		return (SCHED_INTERACT_HALF +
1633 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1634 	}
1635 	if (ts->ts_slptime > ts->ts_runtime) {
1636 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1637 		return (ts->ts_runtime / div);
1638 	}
1639 	/* runtime == slptime */
1640 	if (ts->ts_runtime)
1641 		return (SCHED_INTERACT_HALF);
1642 
1643 	/*
1644 	 * This can happen if slptime and runtime are 0.
1645 	 */
1646 	return (0);
1647 
1648 }
1649 
1650 /*
1651  * Scale the scheduling priority according to the "interactivity" of this
1652  * process.
1653  */
1654 static void
1655 sched_priority(struct thread *td)
1656 {
1657 	u_int pri, score;
1658 
1659 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1660 		return;
1661 	/*
1662 	 * If the score is interactive we place the thread in the realtime
1663 	 * queue with a priority that is less than kernel and interrupt
1664 	 * priorities.  These threads are not subject to nice restrictions.
1665 	 *
1666 	 * Scores greater than this are placed on the normal timeshare queue
1667 	 * where the priority is partially decided by the most recent cpu
1668 	 * utilization and the rest is decided by nice value.
1669 	 *
1670 	 * The nice value of the process has a linear effect on the calculated
1671 	 * score.  Negative nice values make it easier for a thread to be
1672 	 * considered interactive.
1673 	 */
1674 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1675 	if (score < sched_interact) {
1676 		pri = PRI_MIN_INTERACT;
1677 		pri += (PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) * score /
1678 		    sched_interact;
1679 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1680 		    ("sched_priority: invalid interactive priority %u score %u",
1681 		    pri, score));
1682 	} else {
1683 		pri = SCHED_PRI_MIN;
1684 		if (td_get_sched(td)->ts_ticks)
1685 			pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1686 			    SCHED_PRI_RANGE - 1);
1687 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1688 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1689 		    ("sched_priority: invalid priority %u: nice %d, "
1690 		    "ticks %d ftick %d ltick %d tick pri %d",
1691 		    pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1692 		    td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1693 		    SCHED_PRI_TICKS(td_get_sched(td))));
1694 	}
1695 	sched_user_prio(td, pri);
1696 
1697 	return;
1698 }
1699 
1700 /*
1701  * This routine enforces a maximum limit on the amount of scheduling history
1702  * kept.  It is called after either the slptime or runtime is adjusted.  This
1703  * function is ugly due to integer math.
1704  */
1705 static void
1706 sched_interact_update(struct thread *td)
1707 {
1708 	struct td_sched *ts;
1709 	u_int sum;
1710 
1711 	ts = td_get_sched(td);
1712 	sum = ts->ts_runtime + ts->ts_slptime;
1713 	if (sum < SCHED_SLP_RUN_MAX)
1714 		return;
1715 	/*
1716 	 * This only happens from two places:
1717 	 * 1) We have added an unusual amount of run time from fork_exit.
1718 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1719 	 */
1720 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1721 		if (ts->ts_runtime > ts->ts_slptime) {
1722 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1723 			ts->ts_slptime = 1;
1724 		} else {
1725 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1726 			ts->ts_runtime = 1;
1727 		}
1728 		return;
1729 	}
1730 	/*
1731 	 * If we have exceeded by more than 1/5th then the algorithm below
1732 	 * will not bring us back into range.  Dividing by two here forces
1733 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1734 	 */
1735 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1736 		ts->ts_runtime /= 2;
1737 		ts->ts_slptime /= 2;
1738 		return;
1739 	}
1740 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1741 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1742 }
1743 
1744 /*
1745  * Scale back the interactivity history when a child thread is created.  The
1746  * history is inherited from the parent but the thread may behave totally
1747  * differently.  For example, a shell spawning a compiler process.  We want
1748  * to learn that the compiler is behaving badly very quickly.
1749  */
1750 static void
1751 sched_interact_fork(struct thread *td)
1752 {
1753 	struct td_sched *ts;
1754 	int ratio;
1755 	int sum;
1756 
1757 	ts = td_get_sched(td);
1758 	sum = ts->ts_runtime + ts->ts_slptime;
1759 	if (sum > SCHED_SLP_RUN_FORK) {
1760 		ratio = sum / SCHED_SLP_RUN_FORK;
1761 		ts->ts_runtime /= ratio;
1762 		ts->ts_slptime /= ratio;
1763 	}
1764 }
1765 
1766 /*
1767  * Called from proc0_init() to setup the scheduler fields.
1768  */
1769 void
1770 schedinit(void)
1771 {
1772 	struct td_sched *ts0;
1773 
1774 	/*
1775 	 * Set up the scheduler specific parts of thread0.
1776 	 */
1777 	ts0 = td_get_sched(&thread0);
1778 	ts0->ts_ltick = ticks;
1779 	ts0->ts_ftick = ticks;
1780 	ts0->ts_slice = 0;
1781 	ts0->ts_cpu = curcpu;	/* set valid CPU number */
1782 }
1783 
1784 /*
1785  * schedinit_ap() is needed prior to calling sched_throw(NULL) to ensure that
1786  * the pcpu requirements are met for any calls in the period between curthread
1787  * initialization and sched_throw().  One can safely add threads to the queue
1788  * before sched_throw(), for instance, as long as the thread lock is setup
1789  * correctly.
1790  *
1791  * TDQ_SELF() relies on the below sched pcpu setting; it may be used only
1792  * after schedinit_ap().
1793  */
1794 void
1795 schedinit_ap(void)
1796 {
1797 
1798 #ifdef SMP
1799 	PCPU_SET(sched, DPCPU_PTR(tdq));
1800 #endif
1801 	PCPU_GET(idlethread)->td_lock = TDQ_LOCKPTR(TDQ_SELF());
1802 }
1803 
1804 /*
1805  * This is only somewhat accurate since given many processes of the same
1806  * priority they will switch when their slices run out, which will be
1807  * at most sched_slice stathz ticks.
1808  */
1809 int
1810 sched_rr_interval(void)
1811 {
1812 
1813 	/* Convert sched_slice from stathz to hz. */
1814 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1815 }
1816 
1817 /*
1818  * Update the percent cpu tracking information when it is requested or
1819  * the total history exceeds the maximum.  We keep a sliding history of
1820  * tick counts that slowly decays.  This is less precise than the 4BSD
1821  * mechanism since it happens with less regular and frequent events.
1822  */
1823 static void
1824 sched_pctcpu_update(struct td_sched *ts, int run)
1825 {
1826 	int t = ticks;
1827 
1828 	/*
1829 	 * The signed difference may be negative if the thread hasn't run for
1830 	 * over half of the ticks rollover period.
1831 	 */
1832 	if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) {
1833 		ts->ts_ticks = 0;
1834 		ts->ts_ftick = t - SCHED_TICK_TARG;
1835 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1836 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1837 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1838 		ts->ts_ftick = t - SCHED_TICK_TARG;
1839 	}
1840 	if (run)
1841 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1842 	ts->ts_ltick = t;
1843 }
1844 
1845 /*
1846  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1847  * if necessary.  This is the back-end for several priority related
1848  * functions.
1849  */
1850 static void
1851 sched_thread_priority(struct thread *td, u_char prio)
1852 {
1853 	struct tdq *tdq;
1854 	int oldpri;
1855 
1856 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1857 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1858 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1859 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1860 	if (td != curthread && prio < td->td_priority) {
1861 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1862 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1863 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1864 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1865 		    curthread);
1866 	}
1867 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1868 	if (td->td_priority == prio)
1869 		return;
1870 	/*
1871 	 * If the priority has been elevated due to priority
1872 	 * propagation, we may have to move ourselves to a new
1873 	 * queue.  This could be optimized to not re-add in some
1874 	 * cases.
1875 	 */
1876 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1877 		sched_rem(td);
1878 		td->td_priority = prio;
1879 		sched_add(td, SRQ_BORROWING | SRQ_HOLDTD);
1880 		return;
1881 	}
1882 	/*
1883 	 * If the thread is currently running we may have to adjust the lowpri
1884 	 * information so other cpus are aware of our current priority.
1885 	 */
1886 	if (TD_IS_RUNNING(td)) {
1887 		tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
1888 		oldpri = td->td_priority;
1889 		td->td_priority = prio;
1890 		if (prio < tdq->tdq_lowpri)
1891 			tdq->tdq_lowpri = prio;
1892 		else if (tdq->tdq_lowpri == oldpri)
1893 			tdq_setlowpri(tdq, td);
1894 		return;
1895 	}
1896 	td->td_priority = prio;
1897 }
1898 
1899 /*
1900  * Update a thread's priority when it is lent another thread's
1901  * priority.
1902  */
1903 void
1904 sched_lend_prio(struct thread *td, u_char prio)
1905 {
1906 
1907 	td->td_flags |= TDF_BORROWING;
1908 	sched_thread_priority(td, prio);
1909 }
1910 
1911 /*
1912  * Restore a thread's priority when priority propagation is
1913  * over.  The prio argument is the minimum priority the thread
1914  * needs to have to satisfy other possible priority lending
1915  * requests.  If the thread's regular priority is less
1916  * important than prio, the thread will keep a priority boost
1917  * of prio.
1918  */
1919 void
1920 sched_unlend_prio(struct thread *td, u_char prio)
1921 {
1922 	u_char base_pri;
1923 
1924 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1925 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1926 		base_pri = td->td_user_pri;
1927 	else
1928 		base_pri = td->td_base_pri;
1929 	if (prio >= base_pri) {
1930 		td->td_flags &= ~TDF_BORROWING;
1931 		sched_thread_priority(td, base_pri);
1932 	} else
1933 		sched_lend_prio(td, prio);
1934 }
1935 
1936 /*
1937  * Standard entry for setting the priority to an absolute value.
1938  */
1939 void
1940 sched_prio(struct thread *td, u_char prio)
1941 {
1942 	u_char oldprio;
1943 
1944 	/* First, update the base priority. */
1945 	td->td_base_pri = prio;
1946 
1947 	/*
1948 	 * If the thread is borrowing another thread's priority, don't
1949 	 * ever lower the priority.
1950 	 */
1951 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1952 		return;
1953 
1954 	/* Change the real priority. */
1955 	oldprio = td->td_priority;
1956 	sched_thread_priority(td, prio);
1957 
1958 	/*
1959 	 * If the thread is on a turnstile, then let the turnstile update
1960 	 * its state.
1961 	 */
1962 	if (TD_ON_LOCK(td) && oldprio != prio)
1963 		turnstile_adjust(td, oldprio);
1964 }
1965 
1966 /*
1967  * Set the base interrupt thread priority.
1968  */
1969 void
1970 sched_ithread_prio(struct thread *td, u_char prio)
1971 {
1972 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1973 	MPASS(td->td_pri_class == PRI_ITHD);
1974 	td->td_base_ithread_pri = prio;
1975 	sched_prio(td, prio);
1976 }
1977 
1978 /*
1979  * Set the base user priority, does not effect current running priority.
1980  */
1981 void
1982 sched_user_prio(struct thread *td, u_char prio)
1983 {
1984 
1985 	td->td_base_user_pri = prio;
1986 	if (td->td_lend_user_pri <= prio)
1987 		return;
1988 	td->td_user_pri = prio;
1989 }
1990 
1991 void
1992 sched_lend_user_prio(struct thread *td, u_char prio)
1993 {
1994 
1995 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1996 	td->td_lend_user_pri = prio;
1997 	td->td_user_pri = min(prio, td->td_base_user_pri);
1998 	if (td->td_priority > td->td_user_pri)
1999 		sched_prio(td, td->td_user_pri);
2000 	else if (td->td_priority != td->td_user_pri)
2001 		td->td_flags |= TDF_NEEDRESCHED;
2002 }
2003 
2004 /*
2005  * Like the above but first check if there is anything to do.
2006  */
2007 void
2008 sched_lend_user_prio_cond(struct thread *td, u_char prio)
2009 {
2010 
2011 	if (td->td_lend_user_pri != prio)
2012 		goto lend;
2013 	if (td->td_user_pri != min(prio, td->td_base_user_pri))
2014 		goto lend;
2015 	if (td->td_priority != td->td_user_pri)
2016 		goto lend;
2017 	return;
2018 
2019 lend:
2020 	thread_lock(td);
2021 	sched_lend_user_prio(td, prio);
2022 	thread_unlock(td);
2023 }
2024 
2025 #ifdef SMP
2026 /*
2027  * This tdq is about to idle.  Try to steal a thread from another CPU before
2028  * choosing the idle thread.
2029  */
2030 static void
2031 tdq_trysteal(struct tdq *tdq)
2032 {
2033 	struct cpu_group *cg, *parent;
2034 	struct tdq *steal;
2035 	cpuset_t mask;
2036 	int cpu, i, goup;
2037 
2038 	if (smp_started == 0 || steal_idle == 0 || trysteal_limit == 0 ||
2039 	    tdq->tdq_cg == NULL)
2040 		return;
2041 	CPU_FILL(&mask);
2042 	CPU_CLR(PCPU_GET(cpuid), &mask);
2043 	/* We don't want to be preempted while we're iterating. */
2044 	spinlock_enter();
2045 	TDQ_UNLOCK(tdq);
2046 	for (i = 1, cg = tdq->tdq_cg, goup = 0; ; ) {
2047 		cpu = sched_highest(cg, &mask, steal_thresh, 1);
2048 		/*
2049 		 * If a thread was added while interrupts were disabled don't
2050 		 * steal one here.
2051 		 */
2052 		if (TDQ_LOAD(tdq) > 0) {
2053 			TDQ_LOCK(tdq);
2054 			break;
2055 		}
2056 
2057 		/*
2058 		 * We found no CPU to steal from in this group.  Escalate to
2059 		 * the parent and repeat.  But if parent has only two children
2060 		 * groups we can avoid searching this group again by searching
2061 		 * the other one specifically and then escalating two levels.
2062 		 */
2063 		if (cpu == -1) {
2064 			if (goup) {
2065 				cg = cg->cg_parent;
2066 				goup = 0;
2067 			}
2068 			if (++i > trysteal_limit) {
2069 				TDQ_LOCK(tdq);
2070 				break;
2071 			}
2072 			parent = cg->cg_parent;
2073 			if (parent == NULL) {
2074 				TDQ_LOCK(tdq);
2075 				break;
2076 			}
2077 			if (parent->cg_children == 2) {
2078 				if (cg == &parent->cg_child[0])
2079 					cg = &parent->cg_child[1];
2080 				else
2081 					cg = &parent->cg_child[0];
2082 				goup = 1;
2083 			} else
2084 				cg = parent;
2085 			continue;
2086 		}
2087 		steal = TDQ_CPU(cpu);
2088 		/*
2089 		 * The data returned by sched_highest() is stale and
2090 		 * the chosen CPU no longer has an eligible thread.
2091 		 * At this point unconditionally exit the loop to bound
2092 		 * the time spent in the critcal section.
2093 		 */
2094 		if (TDQ_LOAD(steal) < steal_thresh ||
2095 		    TDQ_TRANSFERABLE(steal) == 0)
2096 			continue;
2097 		/*
2098 		 * Try to lock both queues. If we are assigned a thread while
2099 		 * waited for the lock, switch to it now instead of stealing.
2100 		 * If we can't get the lock, then somebody likely got there
2101 		 * first.
2102 		 */
2103 		TDQ_LOCK(tdq);
2104 		if (tdq->tdq_load > 0)
2105 			break;
2106 		if (TDQ_TRYLOCK_FLAGS(steal, MTX_DUPOK) == 0)
2107 			break;
2108 		/*
2109 		 * The data returned by sched_highest() is stale and
2110                  * the chosen CPU no longer has an eligible thread.
2111 		 */
2112 		if (TDQ_LOAD(steal) < steal_thresh ||
2113 		    TDQ_TRANSFERABLE(steal) == 0) {
2114 			TDQ_UNLOCK(steal);
2115 			break;
2116 		}
2117 		/*
2118 		 * If we fail to acquire one due to affinity restrictions,
2119 		 * bail out and let the idle thread to a more complete search
2120 		 * outside of a critical section.
2121 		 */
2122 		if (tdq_move(steal, tdq) == -1) {
2123 			TDQ_UNLOCK(steal);
2124 			break;
2125 		}
2126 		TDQ_UNLOCK(steal);
2127 		break;
2128 	}
2129 	spinlock_exit();
2130 }
2131 #endif
2132 
2133 /*
2134  * Handle migration from sched_switch().  This happens only for
2135  * cpu binding.
2136  */
2137 static struct mtx *
2138 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
2139 {
2140 	struct tdq *tdn;
2141 #ifdef SMP
2142 	int lowpri;
2143 #endif
2144 
2145 	KASSERT(THREAD_CAN_MIGRATE(td) ||
2146 	    (td_get_sched(td)->ts_flags & TSF_BOUND) != 0,
2147 	    ("Thread %p shouldn't migrate", td));
2148 	KASSERT(!CPU_ABSENT(td_get_sched(td)->ts_cpu), ("sched_switch_migrate: "
2149 	    "thread %s queued on absent CPU %d.", td->td_name,
2150 	    td_get_sched(td)->ts_cpu));
2151 	tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
2152 #ifdef SMP
2153 	tdq_load_rem(tdq, td);
2154 	/*
2155 	 * Do the lock dance required to avoid LOR.  We have an
2156 	 * extra spinlock nesting from sched_switch() which will
2157 	 * prevent preemption while we're holding neither run-queue lock.
2158 	 */
2159 	TDQ_UNLOCK(tdq);
2160 	TDQ_LOCK(tdn);
2161 	lowpri = tdq_add(tdn, td, flags);
2162 	tdq_notify(tdn, lowpri);
2163 	TDQ_UNLOCK(tdn);
2164 	TDQ_LOCK(tdq);
2165 #endif
2166 	return (TDQ_LOCKPTR(tdn));
2167 }
2168 
2169 /*
2170  * thread_lock_unblock() that does not assume td_lock is blocked.
2171  */
2172 static inline void
2173 thread_unblock_switch(struct thread *td, struct mtx *mtx)
2174 {
2175 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
2176 	    (uintptr_t)mtx);
2177 }
2178 
2179 /*
2180  * Switch threads.  This function has to handle threads coming in while
2181  * blocked for some reason, running, or idle.  It also must deal with
2182  * migrating a thread from one queue to another as running threads may
2183  * be assigned elsewhere via binding.
2184  */
2185 void
2186 sched_switch(struct thread *td, int flags)
2187 {
2188 	struct thread *newtd;
2189 	struct tdq *tdq;
2190 	struct td_sched *ts;
2191 	struct mtx *mtx;
2192 	int srqflag;
2193 	int cpuid, preempted;
2194 #ifdef SMP
2195 	int pickcpu;
2196 #endif
2197 
2198 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2199 
2200 	cpuid = PCPU_GET(cpuid);
2201 	tdq = TDQ_SELF();
2202 	ts = td_get_sched(td);
2203 	sched_pctcpu_update(ts, 1);
2204 #ifdef SMP
2205 	pickcpu = (td->td_flags & TDF_PICKCPU) != 0;
2206 	if (pickcpu)
2207 		ts->ts_rltick = ticks - affinity * MAX_CACHE_LEVELS;
2208 	else
2209 		ts->ts_rltick = ticks;
2210 #endif
2211 	td->td_lastcpu = td->td_oncpu;
2212 	preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
2213 	    (flags & SW_PREEMPT) != 0;
2214 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_PICKCPU | TDF_SLICEEND);
2215 	td->td_owepreempt = 0;
2216 	atomic_store_char(&tdq->tdq_owepreempt, 0);
2217 	if (!TD_IS_IDLETHREAD(td))
2218 		TDQ_SWITCHCNT_INC(tdq);
2219 
2220 	/*
2221 	 * Always block the thread lock so we can drop the tdq lock early.
2222 	 */
2223 	mtx = thread_lock_block(td);
2224 	spinlock_enter();
2225 	if (TD_IS_IDLETHREAD(td)) {
2226 		MPASS(mtx == TDQ_LOCKPTR(tdq));
2227 		TD_SET_CAN_RUN(td);
2228 	} else if (TD_IS_RUNNING(td)) {
2229 		MPASS(mtx == TDQ_LOCKPTR(tdq));
2230 		srqflag = preempted ?
2231 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
2232 		    SRQ_OURSELF|SRQ_YIELDING;
2233 #ifdef SMP
2234 		if (THREAD_CAN_MIGRATE(td) && (!THREAD_CAN_SCHED(td, ts->ts_cpu)
2235 		    || pickcpu))
2236 			ts->ts_cpu = sched_pickcpu(td, 0);
2237 #endif
2238 		if (ts->ts_cpu == cpuid)
2239 			tdq_runq_add(tdq, td, srqflag);
2240 		else
2241 			mtx = sched_switch_migrate(tdq, td, srqflag);
2242 	} else {
2243 		/* This thread must be going to sleep. */
2244 		if (mtx != TDQ_LOCKPTR(tdq)) {
2245 			mtx_unlock_spin(mtx);
2246 			TDQ_LOCK(tdq);
2247 		}
2248 		tdq_load_rem(tdq, td);
2249 #ifdef SMP
2250 		if (tdq->tdq_load == 0)
2251 			tdq_trysteal(tdq);
2252 #endif
2253 	}
2254 
2255 #if (KTR_COMPILE & KTR_SCHED) != 0
2256 	if (TD_IS_IDLETHREAD(td))
2257 		KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
2258 		    "prio:%d", td->td_priority);
2259 	else
2260 		KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
2261 		    "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
2262 		    "lockname:\"%s\"", td->td_lockname);
2263 #endif
2264 
2265 	/*
2266 	 * We enter here with the thread blocked and assigned to the
2267 	 * appropriate cpu run-queue or sleep-queue and with the current
2268 	 * thread-queue locked.
2269 	 */
2270 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2271 	MPASS(td == tdq->tdq_curthread);
2272 	newtd = choosethread();
2273 	sched_pctcpu_update(td_get_sched(newtd), 0);
2274 	TDQ_UNLOCK(tdq);
2275 
2276 	/*
2277 	 * Call the MD code to switch contexts if necessary.
2278 	 */
2279 	if (td != newtd) {
2280 #ifdef	HWPMC_HOOKS
2281 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2282 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
2283 #endif
2284 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
2285 
2286 #ifdef KDTRACE_HOOKS
2287 		/*
2288 		 * If DTrace has set the active vtime enum to anything
2289 		 * other than INACTIVE (0), then it should have set the
2290 		 * function to call.
2291 		 */
2292 		if (dtrace_vtime_active)
2293 			(*dtrace_vtime_switch_func)(newtd);
2294 #endif
2295 		td->td_oncpu = NOCPU;
2296 		cpu_switch(td, newtd, mtx);
2297 		cpuid = td->td_oncpu = PCPU_GET(cpuid);
2298 
2299 		SDT_PROBE0(sched, , , on__cpu);
2300 #ifdef	HWPMC_HOOKS
2301 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2302 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
2303 #endif
2304 	} else {
2305 		thread_unblock_switch(td, mtx);
2306 		SDT_PROBE0(sched, , , remain__cpu);
2307 	}
2308 	KASSERT(curthread->td_md.md_spinlock_count == 1,
2309 	    ("invalid count %d", curthread->td_md.md_spinlock_count));
2310 
2311 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2312 	    "prio:%d", td->td_priority);
2313 }
2314 
2315 /*
2316  * Adjust thread priorities as a result of a nice request.
2317  */
2318 void
2319 sched_nice(struct proc *p, int nice)
2320 {
2321 	struct thread *td;
2322 
2323 	PROC_LOCK_ASSERT(p, MA_OWNED);
2324 
2325 	p->p_nice = nice;
2326 	FOREACH_THREAD_IN_PROC(p, td) {
2327 		thread_lock(td);
2328 		sched_priority(td);
2329 		sched_prio(td, td->td_base_user_pri);
2330 		thread_unlock(td);
2331 	}
2332 }
2333 
2334 /*
2335  * Record the sleep time for the interactivity scorer.
2336  */
2337 void
2338 sched_sleep(struct thread *td, int prio)
2339 {
2340 
2341 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2342 
2343 	td->td_slptick = ticks;
2344 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2345 		td->td_flags |= TDF_CANSWAP;
2346 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2347 		return;
2348 	if (static_boost == 1 && prio)
2349 		sched_prio(td, prio);
2350 	else if (static_boost && td->td_priority > static_boost)
2351 		sched_prio(td, static_boost);
2352 }
2353 
2354 /*
2355  * Schedule a thread to resume execution and record how long it voluntarily
2356  * slept.  We also update the pctcpu, interactivity, and priority.
2357  *
2358  * Requires the thread lock on entry, drops on exit.
2359  */
2360 void
2361 sched_wakeup(struct thread *td, int srqflags)
2362 {
2363 	struct td_sched *ts;
2364 	int slptick;
2365 
2366 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2367 	ts = td_get_sched(td);
2368 	td->td_flags &= ~TDF_CANSWAP;
2369 
2370 	/*
2371 	 * If we slept for more than a tick update our interactivity and
2372 	 * priority.
2373 	 */
2374 	slptick = td->td_slptick;
2375 	td->td_slptick = 0;
2376 	if (slptick && slptick != ticks) {
2377 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2378 		sched_interact_update(td);
2379 		sched_pctcpu_update(ts, 0);
2380 	}
2381 
2382 	/*
2383 	 * When resuming an idle ithread, restore its base ithread
2384 	 * priority.
2385 	 */
2386 	if (PRI_BASE(td->td_pri_class) == PRI_ITHD &&
2387 	    td->td_priority != td->td_base_ithread_pri)
2388 		sched_prio(td, td->td_base_ithread_pri);
2389 
2390 	/*
2391 	 * Reset the slice value since we slept and advanced the round-robin.
2392 	 */
2393 	ts->ts_slice = 0;
2394 	sched_add(td, SRQ_BORING | srqflags);
2395 }
2396 
2397 /*
2398  * Penalize the parent for creating a new child and initialize the child's
2399  * priority.
2400  */
2401 void
2402 sched_fork(struct thread *td, struct thread *child)
2403 {
2404 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2405 	sched_pctcpu_update(td_get_sched(td), 1);
2406 	sched_fork_thread(td, child);
2407 	/*
2408 	 * Penalize the parent and child for forking.
2409 	 */
2410 	sched_interact_fork(child);
2411 	sched_priority(child);
2412 	td_get_sched(td)->ts_runtime += tickincr;
2413 	sched_interact_update(td);
2414 	sched_priority(td);
2415 }
2416 
2417 /*
2418  * Fork a new thread, may be within the same process.
2419  */
2420 void
2421 sched_fork_thread(struct thread *td, struct thread *child)
2422 {
2423 	struct td_sched *ts;
2424 	struct td_sched *ts2;
2425 	struct tdq *tdq;
2426 
2427 	tdq = TDQ_SELF();
2428 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2429 	/*
2430 	 * Initialize child.
2431 	 */
2432 	ts = td_get_sched(td);
2433 	ts2 = td_get_sched(child);
2434 	child->td_oncpu = NOCPU;
2435 	child->td_lastcpu = NOCPU;
2436 	child->td_lock = TDQ_LOCKPTR(tdq);
2437 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2438 	child->td_domain.dr_policy = td->td_cpuset->cs_domain;
2439 	ts2->ts_cpu = ts->ts_cpu;
2440 	ts2->ts_flags = 0;
2441 	/*
2442 	 * Grab our parents cpu estimation information.
2443 	 */
2444 	ts2->ts_ticks = ts->ts_ticks;
2445 	ts2->ts_ltick = ts->ts_ltick;
2446 	ts2->ts_ftick = ts->ts_ftick;
2447 	/*
2448 	 * Do not inherit any borrowed priority from the parent.
2449 	 */
2450 	child->td_priority = child->td_base_pri;
2451 	/*
2452 	 * And update interactivity score.
2453 	 */
2454 	ts2->ts_slptime = ts->ts_slptime;
2455 	ts2->ts_runtime = ts->ts_runtime;
2456 	/* Attempt to quickly learn interactivity. */
2457 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2458 #ifdef KTR
2459 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2460 #endif
2461 }
2462 
2463 /*
2464  * Adjust the priority class of a thread.
2465  */
2466 void
2467 sched_class(struct thread *td, int class)
2468 {
2469 
2470 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2471 	if (td->td_pri_class == class)
2472 		return;
2473 	td->td_pri_class = class;
2474 }
2475 
2476 /*
2477  * Return some of the child's priority and interactivity to the parent.
2478  */
2479 void
2480 sched_exit(struct proc *p, struct thread *child)
2481 {
2482 	struct thread *td;
2483 
2484 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2485 	    "prio:%d", child->td_priority);
2486 	PROC_LOCK_ASSERT(p, MA_OWNED);
2487 	td = FIRST_THREAD_IN_PROC(p);
2488 	sched_exit_thread(td, child);
2489 }
2490 
2491 /*
2492  * Penalize another thread for the time spent on this one.  This helps to
2493  * worsen the priority and interactivity of processes which schedule batch
2494  * jobs such as make.  This has little effect on the make process itself but
2495  * causes new processes spawned by it to receive worse scores immediately.
2496  */
2497 void
2498 sched_exit_thread(struct thread *td, struct thread *child)
2499 {
2500 
2501 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2502 	    "prio:%d", child->td_priority);
2503 	/*
2504 	 * Give the child's runtime to the parent without returning the
2505 	 * sleep time as a penalty to the parent.  This causes shells that
2506 	 * launch expensive things to mark their children as expensive.
2507 	 */
2508 	thread_lock(td);
2509 	td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2510 	sched_interact_update(td);
2511 	sched_priority(td);
2512 	thread_unlock(td);
2513 }
2514 
2515 void
2516 sched_preempt(struct thread *td)
2517 {
2518 	struct tdq *tdq;
2519 	int flags;
2520 
2521 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2522 
2523 	thread_lock(td);
2524 	tdq = TDQ_SELF();
2525 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2526 	if (td->td_priority > tdq->tdq_lowpri) {
2527 		if (td->td_critnest == 1) {
2528 			flags = SW_INVOL | SW_PREEMPT;
2529 			flags |= TD_IS_IDLETHREAD(td) ? SWT_REMOTEWAKEIDLE :
2530 			    SWT_REMOTEPREEMPT;
2531 			mi_switch(flags);
2532 			/* Switch dropped thread lock. */
2533 			return;
2534 		}
2535 		td->td_owepreempt = 1;
2536 	} else {
2537 		tdq->tdq_owepreempt = 0;
2538 	}
2539 	thread_unlock(td);
2540 }
2541 
2542 /*
2543  * Fix priorities on return to user-space.  Priorities may be elevated due
2544  * to static priorities in msleep() or similar.
2545  */
2546 void
2547 sched_userret_slowpath(struct thread *td)
2548 {
2549 
2550 	thread_lock(td);
2551 	td->td_priority = td->td_user_pri;
2552 	td->td_base_pri = td->td_user_pri;
2553 	tdq_setlowpri(TDQ_SELF(), td);
2554 	thread_unlock(td);
2555 }
2556 
2557 SCHED_STAT_DEFINE(ithread_demotions, "Interrupt thread priority demotions");
2558 SCHED_STAT_DEFINE(ithread_preemptions,
2559     "Interrupt thread preemptions due to time-sharing");
2560 
2561 /*
2562  * Return time slice for a given thread.  For ithreads this is
2563  * sched_slice.  For other threads it is tdq_slice(tdq).
2564  */
2565 static inline int
2566 td_slice(struct thread *td, struct tdq *tdq)
2567 {
2568 	if (PRI_BASE(td->td_pri_class) == PRI_ITHD)
2569 		return (sched_slice);
2570 	return (tdq_slice(tdq));
2571 }
2572 
2573 /*
2574  * Handle a stathz tick.  This is really only relevant for timeshare
2575  * and interrupt threads.
2576  */
2577 void
2578 sched_clock(struct thread *td, int cnt)
2579 {
2580 	struct tdq *tdq;
2581 	struct td_sched *ts;
2582 
2583 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2584 	tdq = TDQ_SELF();
2585 #ifdef SMP
2586 	/*
2587 	 * We run the long term load balancer infrequently on the first cpu.
2588 	 */
2589 	if (balance_tdq == tdq && smp_started != 0 && rebalance != 0 &&
2590 	    balance_ticks != 0) {
2591 		balance_ticks -= cnt;
2592 		if (balance_ticks <= 0)
2593 			sched_balance();
2594 	}
2595 #endif
2596 	/*
2597 	 * Save the old switch count so we have a record of the last ticks
2598 	 * activity.   Initialize the new switch count based on our load.
2599 	 * If there is some activity seed it to reflect that.
2600 	 */
2601 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2602 	tdq->tdq_switchcnt = tdq->tdq_load;
2603 
2604 	/*
2605 	 * Advance the insert index once for each tick to ensure that all
2606 	 * threads get a chance to run.
2607 	 */
2608 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2609 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2610 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2611 			tdq->tdq_ridx = tdq->tdq_idx;
2612 	}
2613 	ts = td_get_sched(td);
2614 	sched_pctcpu_update(ts, 1);
2615 	if ((td->td_pri_class & PRI_FIFO_BIT) || TD_IS_IDLETHREAD(td))
2616 		return;
2617 
2618 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2619 		/*
2620 		 * We used a tick; charge it to the thread so
2621 		 * that we can compute our interactivity.
2622 		 */
2623 		td_get_sched(td)->ts_runtime += tickincr * cnt;
2624 		sched_interact_update(td);
2625 		sched_priority(td);
2626 	}
2627 
2628 	/*
2629 	 * Force a context switch if the current thread has used up a full
2630 	 * time slice (default is 100ms).
2631 	 */
2632 	ts->ts_slice += cnt;
2633 	if (ts->ts_slice >= td_slice(td, tdq)) {
2634 		ts->ts_slice = 0;
2635 
2636 		/*
2637 		 * If an ithread uses a full quantum, demote its
2638 		 * priority and preempt it.
2639 		 */
2640 		if (PRI_BASE(td->td_pri_class) == PRI_ITHD) {
2641 			SCHED_STAT_INC(ithread_preemptions);
2642 			td->td_owepreempt = 1;
2643 			if (td->td_base_pri + RQ_PPQ < PRI_MAX_ITHD) {
2644 				SCHED_STAT_INC(ithread_demotions);
2645 				sched_prio(td, td->td_base_pri + RQ_PPQ);
2646 			}
2647 		} else
2648 			td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2649 	}
2650 }
2651 
2652 u_int
2653 sched_estcpu(struct thread *td __unused)
2654 {
2655 
2656 	return (0);
2657 }
2658 
2659 /*
2660  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2661  * cooperative idle threads.
2662  */
2663 int
2664 sched_runnable(void)
2665 {
2666 	struct tdq *tdq;
2667 	int load;
2668 
2669 	load = 1;
2670 
2671 	tdq = TDQ_SELF();
2672 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2673 		if (TDQ_LOAD(tdq) > 0)
2674 			goto out;
2675 	} else
2676 		if (TDQ_LOAD(tdq) - 1 > 0)
2677 			goto out;
2678 	load = 0;
2679 out:
2680 	return (load);
2681 }
2682 
2683 /*
2684  * Choose the highest priority thread to run.  The thread is removed from
2685  * the run-queue while running however the load remains.
2686  */
2687 struct thread *
2688 sched_choose(void)
2689 {
2690 	struct thread *td;
2691 	struct tdq *tdq;
2692 
2693 	tdq = TDQ_SELF();
2694 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2695 	td = tdq_choose(tdq);
2696 	if (td != NULL) {
2697 		tdq_runq_rem(tdq, td);
2698 		tdq->tdq_lowpri = td->td_priority;
2699 	} else {
2700 		tdq->tdq_lowpri = PRI_MAX_IDLE;
2701 		td = PCPU_GET(idlethread);
2702 	}
2703 	tdq->tdq_curthread = td;
2704 	return (td);
2705 }
2706 
2707 /*
2708  * Set owepreempt if the currently running thread has lower priority than "pri".
2709  * Preemption never happens directly in ULE, we always request it once we exit a
2710  * critical section.
2711  */
2712 static void
2713 sched_setpreempt(int pri)
2714 {
2715 	struct thread *ctd;
2716 	int cpri;
2717 
2718 	ctd = curthread;
2719 	THREAD_LOCK_ASSERT(ctd, MA_OWNED);
2720 
2721 	cpri = ctd->td_priority;
2722 	if (pri < cpri)
2723 		ctd->td_flags |= TDF_NEEDRESCHED;
2724 	if (KERNEL_PANICKED() || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2725 		return;
2726 	if (!sched_shouldpreempt(pri, cpri, 0))
2727 		return;
2728 	ctd->td_owepreempt = 1;
2729 }
2730 
2731 /*
2732  * Add a thread to a thread queue.  Select the appropriate runq and add the
2733  * thread to it.  This is the internal function called when the tdq is
2734  * predetermined.
2735  */
2736 static int
2737 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2738 {
2739 	int lowpri;
2740 
2741 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2742 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
2743 	KASSERT((td->td_inhibitors == 0),
2744 	    ("sched_add: trying to run inhibited thread"));
2745 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2746 	    ("sched_add: bad thread state"));
2747 	KASSERT(td->td_flags & TDF_INMEM,
2748 	    ("sched_add: thread swapped out"));
2749 
2750 	lowpri = tdq->tdq_lowpri;
2751 	if (td->td_priority < lowpri)
2752 		tdq->tdq_lowpri = td->td_priority;
2753 	tdq_runq_add(tdq, td, flags);
2754 	tdq_load_add(tdq, td);
2755 	return (lowpri);
2756 }
2757 
2758 /*
2759  * Select the target thread queue and add a thread to it.  Request
2760  * preemption or IPI a remote processor if required.
2761  *
2762  * Requires the thread lock on entry, drops on exit.
2763  */
2764 void
2765 sched_add(struct thread *td, int flags)
2766 {
2767 	struct tdq *tdq;
2768 #ifdef SMP
2769 	int cpu, lowpri;
2770 #endif
2771 
2772 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2773 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2774 	    sched_tdname(curthread));
2775 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2776 	    KTR_ATTR_LINKED, sched_tdname(td));
2777 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2778 	    flags & SRQ_PREEMPTED);
2779 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2780 	/*
2781 	 * Recalculate the priority before we select the target cpu or
2782 	 * run-queue.
2783 	 */
2784 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2785 		sched_priority(td);
2786 #ifdef SMP
2787 	/*
2788 	 * Pick the destination cpu and if it isn't ours transfer to the
2789 	 * target cpu.
2790 	 */
2791 	cpu = sched_pickcpu(td, flags);
2792 	tdq = sched_setcpu(td, cpu, flags);
2793 	lowpri = tdq_add(tdq, td, flags);
2794 	if (cpu != PCPU_GET(cpuid))
2795 		tdq_notify(tdq, lowpri);
2796 	else if (!(flags & SRQ_YIELDING))
2797 		sched_setpreempt(td->td_priority);
2798 #else
2799 	tdq = TDQ_SELF();
2800 	/*
2801 	 * Now that the thread is moving to the run-queue, set the lock
2802 	 * to the scheduler's lock.
2803 	 */
2804 	if (td->td_lock != TDQ_LOCKPTR(tdq)) {
2805 		TDQ_LOCK(tdq);
2806 		if ((flags & SRQ_HOLD) != 0)
2807 			td->td_lock = TDQ_LOCKPTR(tdq);
2808 		else
2809 			thread_lock_set(td, TDQ_LOCKPTR(tdq));
2810 	}
2811 	(void)tdq_add(tdq, td, flags);
2812 	if (!(flags & SRQ_YIELDING))
2813 		sched_setpreempt(td->td_priority);
2814 #endif
2815 	if (!(flags & SRQ_HOLDTD))
2816 		thread_unlock(td);
2817 }
2818 
2819 /*
2820  * Remove a thread from a run-queue without running it.  This is used
2821  * when we're stealing a thread from a remote queue.  Otherwise all threads
2822  * exit by calling sched_exit_thread() and sched_throw() themselves.
2823  */
2824 void
2825 sched_rem(struct thread *td)
2826 {
2827 	struct tdq *tdq;
2828 
2829 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2830 	    "prio:%d", td->td_priority);
2831 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2832 	tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2833 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2834 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2835 	KASSERT(TD_ON_RUNQ(td),
2836 	    ("sched_rem: thread not on run queue"));
2837 	tdq_runq_rem(tdq, td);
2838 	tdq_load_rem(tdq, td);
2839 	TD_SET_CAN_RUN(td);
2840 	if (td->td_priority == tdq->tdq_lowpri)
2841 		tdq_setlowpri(tdq, NULL);
2842 }
2843 
2844 /*
2845  * Fetch cpu utilization information.  Updates on demand.
2846  */
2847 fixpt_t
2848 sched_pctcpu(struct thread *td)
2849 {
2850 	fixpt_t pctcpu;
2851 	struct td_sched *ts;
2852 
2853 	pctcpu = 0;
2854 	ts = td_get_sched(td);
2855 
2856 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2857 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2858 	if (ts->ts_ticks) {
2859 		int rtick;
2860 
2861 		/* How many rtick per second ? */
2862 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2863 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2864 	}
2865 
2866 	return (pctcpu);
2867 }
2868 
2869 /*
2870  * Enforce affinity settings for a thread.  Called after adjustments to
2871  * cpumask.
2872  */
2873 void
2874 sched_affinity(struct thread *td)
2875 {
2876 #ifdef SMP
2877 	struct td_sched *ts;
2878 
2879 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2880 	ts = td_get_sched(td);
2881 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2882 		return;
2883 	if (TD_ON_RUNQ(td)) {
2884 		sched_rem(td);
2885 		sched_add(td, SRQ_BORING | SRQ_HOLDTD);
2886 		return;
2887 	}
2888 	if (!TD_IS_RUNNING(td))
2889 		return;
2890 	/*
2891 	 * Force a switch before returning to userspace.  If the
2892 	 * target thread is not running locally send an ipi to force
2893 	 * the issue.
2894 	 */
2895 	td->td_flags |= TDF_NEEDRESCHED;
2896 	if (td != curthread)
2897 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2898 #endif
2899 }
2900 
2901 /*
2902  * Bind a thread to a target cpu.
2903  */
2904 void
2905 sched_bind(struct thread *td, int cpu)
2906 {
2907 	struct td_sched *ts;
2908 
2909 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2910 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2911 	ts = td_get_sched(td);
2912 	if (ts->ts_flags & TSF_BOUND)
2913 		sched_unbind(td);
2914 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2915 	ts->ts_flags |= TSF_BOUND;
2916 	sched_pin();
2917 	if (PCPU_GET(cpuid) == cpu)
2918 		return;
2919 	ts->ts_cpu = cpu;
2920 	/* When we return from mi_switch we'll be on the correct cpu. */
2921 	mi_switch(SW_VOL);
2922 	thread_lock(td);
2923 }
2924 
2925 /*
2926  * Release a bound thread.
2927  */
2928 void
2929 sched_unbind(struct thread *td)
2930 {
2931 	struct td_sched *ts;
2932 
2933 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2934 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2935 	ts = td_get_sched(td);
2936 	if ((ts->ts_flags & TSF_BOUND) == 0)
2937 		return;
2938 	ts->ts_flags &= ~TSF_BOUND;
2939 	sched_unpin();
2940 }
2941 
2942 int
2943 sched_is_bound(struct thread *td)
2944 {
2945 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2946 	return (td_get_sched(td)->ts_flags & TSF_BOUND);
2947 }
2948 
2949 /*
2950  * Basic yield call.
2951  */
2952 void
2953 sched_relinquish(struct thread *td)
2954 {
2955 	thread_lock(td);
2956 	mi_switch(SW_VOL | SWT_RELINQUISH);
2957 }
2958 
2959 /*
2960  * Return the total system load.
2961  */
2962 int
2963 sched_load(void)
2964 {
2965 #ifdef SMP
2966 	int total;
2967 	int i;
2968 
2969 	total = 0;
2970 	CPU_FOREACH(i)
2971 		total += atomic_load_int(&TDQ_CPU(i)->tdq_sysload);
2972 	return (total);
2973 #else
2974 	return (atomic_load_int(&TDQ_SELF()->tdq_sysload));
2975 #endif
2976 }
2977 
2978 int
2979 sched_sizeof_proc(void)
2980 {
2981 	return (sizeof(struct proc));
2982 }
2983 
2984 int
2985 sched_sizeof_thread(void)
2986 {
2987 	return (sizeof(struct thread) + sizeof(struct td_sched));
2988 }
2989 
2990 #ifdef SMP
2991 #define	TDQ_IDLESPIN(tdq)						\
2992     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2993 #else
2994 #define	TDQ_IDLESPIN(tdq)	1
2995 #endif
2996 
2997 /*
2998  * The actual idle process.
2999  */
3000 void
3001 sched_idletd(void *dummy)
3002 {
3003 	struct thread *td;
3004 	struct tdq *tdq;
3005 	int oldswitchcnt, switchcnt;
3006 	int i;
3007 
3008 	mtx_assert(&Giant, MA_NOTOWNED);
3009 	td = curthread;
3010 	tdq = TDQ_SELF();
3011 	THREAD_NO_SLEEPING();
3012 	oldswitchcnt = -1;
3013 	for (;;) {
3014 		if (TDQ_LOAD(tdq)) {
3015 			thread_lock(td);
3016 			mi_switch(SW_VOL | SWT_IDLE);
3017 		}
3018 		switchcnt = TDQ_SWITCHCNT(tdq);
3019 #ifdef SMP
3020 		if (always_steal || switchcnt != oldswitchcnt) {
3021 			oldswitchcnt = switchcnt;
3022 			if (tdq_idled(tdq) == 0)
3023 				continue;
3024 		}
3025 		switchcnt = TDQ_SWITCHCNT(tdq);
3026 #else
3027 		oldswitchcnt = switchcnt;
3028 #endif
3029 		/*
3030 		 * If we're switching very frequently, spin while checking
3031 		 * for load rather than entering a low power state that
3032 		 * may require an IPI.  However, don't do any busy
3033 		 * loops while on SMT machines as this simply steals
3034 		 * cycles from cores doing useful work.
3035 		 */
3036 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
3037 			for (i = 0; i < sched_idlespins; i++) {
3038 				if (TDQ_LOAD(tdq))
3039 					break;
3040 				cpu_spinwait();
3041 			}
3042 		}
3043 
3044 		/* If there was context switch during spin, restart it. */
3045 		switchcnt = TDQ_SWITCHCNT(tdq);
3046 		if (TDQ_LOAD(tdq) != 0 || switchcnt != oldswitchcnt)
3047 			continue;
3048 
3049 		/* Run main MD idle handler. */
3050 		atomic_store_int(&tdq->tdq_cpu_idle, 1);
3051 		/*
3052 		 * Make sure that the tdq_cpu_idle update is globally visible
3053 		 * before cpu_idle() reads tdq_load.  The order is important
3054 		 * to avoid races with tdq_notify().
3055 		 */
3056 		atomic_thread_fence_seq_cst();
3057 		/*
3058 		 * Checking for again after the fence picks up assigned
3059 		 * threads often enough to make it worthwhile to do so in
3060 		 * order to avoid calling cpu_idle().
3061 		 */
3062 		if (TDQ_LOAD(tdq) != 0) {
3063 			atomic_store_int(&tdq->tdq_cpu_idle, 0);
3064 			continue;
3065 		}
3066 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
3067 		atomic_store_int(&tdq->tdq_cpu_idle, 0);
3068 
3069 		/*
3070 		 * Account thread-less hardware interrupts and
3071 		 * other wakeup reasons equal to context switches.
3072 		 */
3073 		switchcnt = TDQ_SWITCHCNT(tdq);
3074 		if (switchcnt != oldswitchcnt)
3075 			continue;
3076 		TDQ_SWITCHCNT_INC(tdq);
3077 		oldswitchcnt++;
3078 	}
3079 }
3080 
3081 /*
3082  * sched_throw_grab() chooses a thread from the queue to switch to
3083  * next.  It returns with the tdq lock dropped in a spinlock section to
3084  * keep interrupts disabled until the CPU is running in a proper threaded
3085  * context.
3086  */
3087 static struct thread *
3088 sched_throw_grab(struct tdq *tdq)
3089 {
3090 	struct thread *newtd;
3091 
3092 	newtd = choosethread();
3093 	spinlock_enter();
3094 	TDQ_UNLOCK(tdq);
3095 	KASSERT(curthread->td_md.md_spinlock_count == 1,
3096 	    ("invalid count %d", curthread->td_md.md_spinlock_count));
3097 	return (newtd);
3098 }
3099 
3100 /*
3101  * A CPU is entering for the first time.
3102  */
3103 void
3104 sched_ap_entry(void)
3105 {
3106 	struct thread *newtd;
3107 	struct tdq *tdq;
3108 
3109 	tdq = TDQ_SELF();
3110 
3111 	/* This should have been setup in schedinit_ap(). */
3112 	THREAD_LOCKPTR_ASSERT(curthread, TDQ_LOCKPTR(tdq));
3113 
3114 	TDQ_LOCK(tdq);
3115 	/* Correct spinlock nesting. */
3116 	spinlock_exit();
3117 	PCPU_SET(switchtime, cpu_ticks());
3118 	PCPU_SET(switchticks, ticks);
3119 
3120 	newtd = sched_throw_grab(tdq);
3121 
3122 	/* doesn't return */
3123 	cpu_throw(NULL, newtd);
3124 }
3125 
3126 /*
3127  * A thread is exiting.
3128  */
3129 void
3130 sched_throw(struct thread *td)
3131 {
3132 	struct thread *newtd;
3133 	struct tdq *tdq;
3134 
3135 	tdq = TDQ_SELF();
3136 
3137 	MPASS(td != NULL);
3138 	THREAD_LOCK_ASSERT(td, MA_OWNED);
3139 	THREAD_LOCKPTR_ASSERT(td, TDQ_LOCKPTR(tdq));
3140 
3141 	tdq_load_rem(tdq, td);
3142 	td->td_lastcpu = td->td_oncpu;
3143 	td->td_oncpu = NOCPU;
3144 	thread_lock_block(td);
3145 
3146 	newtd = sched_throw_grab(tdq);
3147 
3148 	/* doesn't return */
3149 	cpu_switch(td, newtd, TDQ_LOCKPTR(tdq));
3150 }
3151 
3152 /*
3153  * This is called from fork_exit().  Just acquire the correct locks and
3154  * let fork do the rest of the work.
3155  */
3156 void
3157 sched_fork_exit(struct thread *td)
3158 {
3159 	struct tdq *tdq;
3160 	int cpuid;
3161 
3162 	/*
3163 	 * Finish setting up thread glue so that it begins execution in a
3164 	 * non-nested critical section with the scheduler lock held.
3165 	 */
3166 	KASSERT(curthread->td_md.md_spinlock_count == 1,
3167 	    ("invalid count %d", curthread->td_md.md_spinlock_count));
3168 	cpuid = PCPU_GET(cpuid);
3169 	tdq = TDQ_SELF();
3170 	TDQ_LOCK(tdq);
3171 	spinlock_exit();
3172 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
3173 	td->td_oncpu = cpuid;
3174 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
3175 	    "prio:%d", td->td_priority);
3176 	SDT_PROBE0(sched, , , on__cpu);
3177 }
3178 
3179 /*
3180  * Create on first use to catch odd startup conditions.
3181  */
3182 char *
3183 sched_tdname(struct thread *td)
3184 {
3185 #ifdef KTR
3186 	struct td_sched *ts;
3187 
3188 	ts = td_get_sched(td);
3189 	if (ts->ts_name[0] == '\0')
3190 		snprintf(ts->ts_name, sizeof(ts->ts_name),
3191 		    "%s tid %d", td->td_name, td->td_tid);
3192 	return (ts->ts_name);
3193 #else
3194 	return (td->td_name);
3195 #endif
3196 }
3197 
3198 #ifdef KTR
3199 void
3200 sched_clear_tdname(struct thread *td)
3201 {
3202 	struct td_sched *ts;
3203 
3204 	ts = td_get_sched(td);
3205 	ts->ts_name[0] = '\0';
3206 }
3207 #endif
3208 
3209 #ifdef SMP
3210 
3211 /*
3212  * Build the CPU topology dump string. Is recursively called to collect
3213  * the topology tree.
3214  */
3215 static int
3216 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
3217     int indent)
3218 {
3219 	char cpusetbuf[CPUSETBUFSIZ];
3220 	int i, first;
3221 
3222 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
3223 	    "", 1 + indent / 2, cg->cg_level);
3224 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
3225 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
3226 	first = TRUE;
3227 	for (i = cg->cg_first; i <= cg->cg_last; i++) {
3228 		if (CPU_ISSET(i, &cg->cg_mask)) {
3229 			if (!first)
3230 				sbuf_printf(sb, ", ");
3231 			else
3232 				first = FALSE;
3233 			sbuf_printf(sb, "%d", i);
3234 		}
3235 	}
3236 	sbuf_printf(sb, "</cpu>\n");
3237 
3238 	if (cg->cg_flags != 0) {
3239 		sbuf_printf(sb, "%*s <flags>", indent, "");
3240 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
3241 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
3242 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
3243 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
3244 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
3245 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
3246 		if ((cg->cg_flags & CG_FLAG_NODE) != 0)
3247 			sbuf_printf(sb, "<flag name=\"NODE\">NUMA node</flag>");
3248 		sbuf_printf(sb, "</flags>\n");
3249 	}
3250 
3251 	if (cg->cg_children > 0) {
3252 		sbuf_printf(sb, "%*s <children>\n", indent, "");
3253 		for (i = 0; i < cg->cg_children; i++)
3254 			sysctl_kern_sched_topology_spec_internal(sb,
3255 			    &cg->cg_child[i], indent+2);
3256 		sbuf_printf(sb, "%*s </children>\n", indent, "");
3257 	}
3258 	sbuf_printf(sb, "%*s</group>\n", indent, "");
3259 	return (0);
3260 }
3261 
3262 /*
3263  * Sysctl handler for retrieving topology dump. It's a wrapper for
3264  * the recursive sysctl_kern_smp_topology_spec_internal().
3265  */
3266 static int
3267 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
3268 {
3269 	struct sbuf *topo;
3270 	int err;
3271 
3272 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
3273 
3274 	topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
3275 	if (topo == NULL)
3276 		return (ENOMEM);
3277 
3278 	sbuf_printf(topo, "<groups>\n");
3279 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
3280 	sbuf_printf(topo, "</groups>\n");
3281 
3282 	if (err == 0) {
3283 		err = sbuf_finish(topo);
3284 	}
3285 	sbuf_delete(topo);
3286 	return (err);
3287 }
3288 
3289 #endif
3290 
3291 static int
3292 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
3293 {
3294 	int error, new_val, period;
3295 
3296 	period = 1000000 / realstathz;
3297 	new_val = period * sched_slice;
3298 	error = sysctl_handle_int(oidp, &new_val, 0, req);
3299 	if (error != 0 || req->newptr == NULL)
3300 		return (error);
3301 	if (new_val <= 0)
3302 		return (EINVAL);
3303 	sched_slice = imax(1, (new_val + period / 2) / period);
3304 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
3305 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
3306 	    realstathz);
3307 	return (0);
3308 }
3309 
3310 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
3311     "Scheduler");
3312 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
3313     "Scheduler name");
3314 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum,
3315     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
3316     sysctl_kern_quantum, "I",
3317     "Quantum for timeshare threads in microseconds");
3318 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
3319     "Quantum for timeshare threads in stathz ticks");
3320 SYSCTL_UINT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
3321     "Interactivity score threshold");
3322 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
3323     &preempt_thresh, 0,
3324     "Maximal (lowest) priority for preemption");
3325 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
3326     "Assign static kernel priorities to sleeping threads");
3327 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
3328     "Number of times idle thread will spin waiting for new work");
3329 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
3330     &sched_idlespinthresh, 0,
3331     "Threshold before we will permit idle thread spinning");
3332 #ifdef SMP
3333 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
3334     "Number of hz ticks to keep thread affinity for");
3335 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
3336     "Enables the long-term load balancer");
3337 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
3338     &balance_interval, 0,
3339     "Average period in stathz ticks to run the long-term balancer");
3340 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
3341     "Attempts to steal work from other cores before idling");
3342 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
3343     "Minimum load on remote CPU before we'll steal");
3344 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit,
3345     0, "Topological distance limit for stealing threads in sched_switch()");
3346 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0,
3347     "Always run the stealer from the idle thread");
3348 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
3349     CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
3350     "XML dump of detected CPU topology");
3351 #endif
3352 
3353 /* ps compat.  All cpu percentages from ULE are weighted. */
3354 static int ccpu = 0;
3355 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0,
3356     "Decay factor used for updating %CPU in 4BSD scheduler");
3357