xref: /freebsd/sys/kern/sched_ule.c (revision 2e3f49888ec8851bafb22011533217487764fdb0)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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 #include "opt_hwpmc_hooks.h"
42 #include "opt_sched.h"
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/kdb.h>
47 #include <sys/kernel.h>
48 #include <sys/ktr.h>
49 #include <sys/limits.h>
50 #include <sys/lock.h>
51 #include <sys/mutex.h>
52 #include <sys/proc.h>
53 #include <sys/resource.h>
54 #include <sys/resourcevar.h>
55 #include <sys/sched.h>
56 #include <sys/sdt.h>
57 #include <sys/smp.h>
58 #include <sys/sx.h>
59 #include <sys/sysctl.h>
60 #include <sys/sysproto.h>
61 #include <sys/turnstile.h>
62 #include <sys/umtxvar.h>
63 #include <sys/vmmeter.h>
64 #include <sys/cpuset.h>
65 #include <sys/sbuf.h>
66 
67 #ifdef HWPMC_HOOKS
68 #include <sys/pmckern.h>
69 #endif
70 
71 #ifdef KDTRACE_HOOKS
72 #include <sys/dtrace_bsd.h>
73 int __read_mostly		dtrace_vtime_active;
74 dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
75 #endif
76 
77 #include <machine/cpu.h>
78 #include <machine/smp.h>
79 
80 #define	KTR_ULE	0
81 
82 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
83 #define	TDQ_NAME_LEN	(sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
84 #define	TDQ_LOADNAME_LEN	(sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
85 
86 /*
87  * Thread scheduler specific section.  All fields are protected
88  * by the thread lock.
89  */
90 struct td_sched {
91 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
92 	short		ts_flags;	/* TSF_* flags. */
93 	int		ts_cpu;		/* CPU that we have affinity for. */
94 	int		ts_rltick;	/* Real last tick, for affinity. */
95 	int		ts_slice;	/* Ticks of slice remaining. */
96 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
97 	u_int		ts_runtime;	/* Number of ticks we were running */
98 	int		ts_ltick;	/* Last tick that we were running on */
99 	int		ts_ftick;	/* First tick that we were running on */
100 	int		ts_ticks;	/* Tick count */
101 #ifdef KTR
102 	char		ts_name[TS_NAME_LEN];
103 #endif
104 };
105 /* flags kept in ts_flags */
106 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
107 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
108 
109 #define	THREAD_CAN_MIGRATE(td)	((td)->td_pinned == 0)
110 #define	THREAD_CAN_SCHED(td, cpu)	\
111     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
112 
113 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <=
114     sizeof(struct thread0_storage),
115     "increase struct thread0_storage.t0st_sched size");
116 
117 /*
118  * Priority ranges used for interactive and non-interactive timeshare
119  * threads.  The timeshare priorities are split up into four ranges.
120  * The first range handles interactive threads.  The last three ranges
121  * (NHALF, x, and NHALF) handle non-interactive threads with the outer
122  * ranges supporting nice values.
123  */
124 #define	PRI_TIMESHARE_RANGE	(PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
125 #define	PRI_INTERACT_RANGE	((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
126 #define	PRI_BATCH_RANGE		(PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
127 
128 #define	PRI_MIN_INTERACT	PRI_MIN_TIMESHARE
129 #define	PRI_MAX_INTERACT	(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
130 #define	PRI_MIN_BATCH		(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
131 #define	PRI_MAX_BATCH		PRI_MAX_TIMESHARE
132 
133 /*
134  * Cpu percentage computation macros and defines.
135  *
136  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
137  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
138  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
139  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
140  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
141  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
142  */
143 #define	SCHED_TICK_SECS		10
144 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
145 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
146 #define	SCHED_TICK_SHIFT	10
147 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
148 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
149 
150 /*
151  * These macros determine priorities for non-interactive threads.  They are
152  * assigned a priority based on their recent cpu utilization as expressed
153  * by the ratio of ticks to the tick total.  NHALF priorities at the start
154  * and end of the MIN to MAX timeshare range are only reachable with negative
155  * or positive nice respectively.
156  *
157  * PRI_RANGE:	Priority range for utilization dependent priorities.
158  * PRI_NRESV:	Number of nice values.
159  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
160  * PRI_NICE:	Determines the part of the priority inherited from nice.
161  */
162 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
163 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
164 #define	SCHED_PRI_MIN		(PRI_MIN_BATCH + SCHED_PRI_NHALF)
165 #define	SCHED_PRI_MAX		(PRI_MAX_BATCH - SCHED_PRI_NHALF)
166 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
167 #define	SCHED_PRI_TICKS(ts)						\
168     (SCHED_TICK_HZ((ts)) /						\
169     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
170 #define	SCHED_PRI_NICE(nice)	(nice)
171 
172 /*
173  * These determine the interactivity of a process.  Interactivity differs from
174  * cpu utilization in that it expresses the voluntary time slept vs time ran
175  * while cpu utilization includes all time not running.  This more accurately
176  * models the intent of the thread.
177  *
178  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
179  *		before throttling back.
180  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
181  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
182  * INTERACT_THRESH:	Threshold for placement on the current runq.
183  */
184 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
185 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
186 #define	SCHED_INTERACT_MAX	(100)
187 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
188 #define	SCHED_INTERACT_THRESH	(30)
189 
190 /*
191  * These parameters determine the slice behavior for batch work.
192  */
193 #define	SCHED_SLICE_DEFAULT_DIVISOR	10	/* ~94 ms, 12 stathz ticks. */
194 #define	SCHED_SLICE_MIN_DIVISOR		6	/* DEFAULT/MIN = ~16 ms. */
195 
196 /* Flags kept in td_flags. */
197 #define	TDF_PICKCPU	TDF_SCHED0	/* Thread should pick new CPU. */
198 #define	TDF_SLICEEND	TDF_SCHED2	/* Thread time slice is over. */
199 
200 /*
201  * tickincr:		Converts a stathz tick into a hz domain scaled by
202  *			the shift factor.  Without the shift the error rate
203  *			due to rounding would be unacceptably high.
204  * realstathz:		stathz is sometimes 0 and run off of hz.
205  * sched_slice:		Runtime of each thread before rescheduling.
206  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
207  */
208 static u_int __read_mostly sched_interact = SCHED_INTERACT_THRESH;
209 static int __read_mostly tickincr = 8 << SCHED_TICK_SHIFT;
210 static int __read_mostly realstathz = 127;	/* reset during boot. */
211 static int __read_mostly sched_slice = 10;	/* reset during boot. */
212 static int __read_mostly sched_slice_min = 1;	/* reset during boot. */
213 #ifdef PREEMPTION
214 #ifdef FULL_PREEMPTION
215 static int __read_mostly preempt_thresh = PRI_MAX_IDLE;
216 #else
217 static int __read_mostly preempt_thresh = PRI_MIN_KERN;
218 #endif
219 #else
220 static int __read_mostly preempt_thresh = 0;
221 #endif
222 static int __read_mostly static_boost = PRI_MIN_BATCH;
223 static int __read_mostly sched_idlespins = 10000;
224 static int __read_mostly sched_idlespinthresh = -1;
225 
226 /*
227  * tdq - per processor runqs and statistics.  A mutex synchronizes access to
228  * most fields.  Some fields are loaded or modified without the mutex.
229  *
230  * Locking protocols:
231  * (c)  constant after initialization
232  * (f)  flag, set with the tdq lock held, cleared on local CPU
233  * (l)  all accesses are CPU-local
234  * (ls) stores are performed by the local CPU, loads may be lockless
235  * (t)  all accesses are protected by the tdq mutex
236  * (ts) stores are serialized by the tdq mutex, loads may be lockless
237  */
238 struct tdq {
239 	/*
240 	 * Ordered to improve efficiency of cpu_search() and switch().
241 	 * tdq_lock is padded to avoid false sharing with tdq_load and
242 	 * tdq_cpu_idle.
243 	 */
244 	struct mtx_padalign tdq_lock;	/* run queue lock. */
245 	struct cpu_group *tdq_cg;	/* (c) Pointer to cpu topology. */
246 	struct thread	*tdq_curthread;	/* (t) Current executing thread. */
247 	int		tdq_load;	/* (ts) Aggregate load. */
248 	int		tdq_sysload;	/* (ts) For loadavg, !ITHD load. */
249 	int		tdq_cpu_idle;	/* (ls) cpu_idle() is active. */
250 	int		tdq_transferable; /* (ts) Transferable thread count. */
251 	short		tdq_switchcnt;	/* (l) Switches this tick. */
252 	short		tdq_oldswitchcnt; /* (l) Switches last tick. */
253 	u_char		tdq_lowpri;	/* (ts) Lowest priority thread. */
254 	u_char		tdq_owepreempt;	/* (f) Remote preemption pending. */
255 	u_char		tdq_idx;	/* (t) Current insert index. */
256 	u_char		tdq_ridx;	/* (t) Current removal index. */
257 	int		tdq_id;		/* (c) cpuid. */
258 	struct runq	tdq_realtime;	/* (t) real-time run queue. */
259 	struct runq	tdq_timeshare;	/* (t) timeshare run queue. */
260 	struct runq	tdq_idle;	/* (t) Queue of IDLE threads. */
261 	char		tdq_name[TDQ_NAME_LEN];
262 #ifdef KTR
263 	char		tdq_loadname[TDQ_LOADNAME_LEN];
264 #endif
265 };
266 
267 /* Idle thread states and config. */
268 #define	TDQ_RUNNING	1
269 #define	TDQ_IDLE	2
270 
271 /* Lockless accessors. */
272 #define	TDQ_LOAD(tdq)		atomic_load_int(&(tdq)->tdq_load)
273 #define	TDQ_TRANSFERABLE(tdq)	atomic_load_int(&(tdq)->tdq_transferable)
274 #define	TDQ_SWITCHCNT(tdq)	(atomic_load_short(&(tdq)->tdq_switchcnt) + \
275 				 atomic_load_short(&(tdq)->tdq_oldswitchcnt))
276 #define	TDQ_SWITCHCNT_INC(tdq)	(atomic_store_short(&(tdq)->tdq_switchcnt, \
277 				 atomic_load_short(&(tdq)->tdq_switchcnt) + 1))
278 
279 #ifdef SMP
280 struct cpu_group __read_mostly *cpu_top;		/* CPU topology */
281 
282 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
283 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
284 
285 /*
286  * Run-time tunables.
287  */
288 static int rebalance = 1;
289 static int balance_interval = 128;	/* Default set in sched_initticks(). */
290 static int __read_mostly affinity;
291 static int __read_mostly steal_idle = 1;
292 static int __read_mostly steal_thresh = 2;
293 static int __read_mostly always_steal = 0;
294 static int __read_mostly trysteal_limit = 2;
295 
296 /*
297  * One thread queue per processor.
298  */
299 static struct tdq __read_mostly *balance_tdq;
300 static int balance_ticks;
301 DPCPU_DEFINE_STATIC(struct tdq, tdq);
302 DPCPU_DEFINE_STATIC(uint32_t, randomval);
303 
304 #define	TDQ_SELF()	((struct tdq *)PCPU_GET(sched))
305 #define	TDQ_CPU(x)	(DPCPU_ID_PTR((x), tdq))
306 #define	TDQ_ID(x)	((x)->tdq_id)
307 #else	/* !SMP */
308 static struct tdq	tdq_cpu;
309 
310 #define	TDQ_ID(x)	(0)
311 #define	TDQ_SELF()	(&tdq_cpu)
312 #define	TDQ_CPU(x)	(&tdq_cpu)
313 #endif
314 
315 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
316 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
317 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
318 #define	TDQ_TRYLOCK(t)		mtx_trylock_spin(TDQ_LOCKPTR((t)))
319 #define	TDQ_TRYLOCK_FLAGS(t, f)	mtx_trylock_spin_flags(TDQ_LOCKPTR((t)), (f))
320 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
321 #define	TDQ_LOCKPTR(t)		((struct mtx *)(&(t)->tdq_lock))
322 
323 static void sched_setpreempt(int);
324 static void sched_priority(struct thread *);
325 static void sched_thread_priority(struct thread *, u_char);
326 static int sched_interact_score(struct thread *);
327 static void sched_interact_update(struct thread *);
328 static void sched_interact_fork(struct thread *);
329 static void sched_pctcpu_update(struct td_sched *, int);
330 
331 /* Operations on per processor queues */
332 static struct thread *tdq_choose(struct tdq *);
333 static void tdq_setup(struct tdq *, int i);
334 static void tdq_load_add(struct tdq *, struct thread *);
335 static void tdq_load_rem(struct tdq *, struct thread *);
336 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
337 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
338 static inline int sched_shouldpreempt(int, int, int);
339 static void tdq_print(int cpu);
340 static void runq_print(struct runq *rq);
341 static int tdq_add(struct tdq *, struct thread *, int);
342 #ifdef SMP
343 static int tdq_move(struct tdq *, struct tdq *);
344 static int tdq_idled(struct tdq *);
345 static void tdq_notify(struct tdq *, int lowpri);
346 static struct thread *tdq_steal(struct tdq *, int);
347 static struct thread *runq_steal(struct runq *, int);
348 static int sched_pickcpu(struct thread *, int);
349 static void sched_balance(void);
350 static bool sched_balance_pair(struct tdq *, struct tdq *);
351 static inline struct tdq *sched_setcpu(struct thread *, int, int);
352 static inline void thread_unblock_switch(struct thread *, struct mtx *);
353 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
354 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
355     struct cpu_group *cg, int indent);
356 #endif
357 
358 static void sched_setup(void *dummy);
359 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
360 
361 static void sched_initticks(void *dummy);
362 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
363     NULL);
364 
365 SDT_PROVIDER_DEFINE(sched);
366 
367 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
368     "struct proc *", "uint8_t");
369 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
370     "struct proc *", "void *");
371 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
372     "struct proc *", "void *", "int");
373 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
374     "struct proc *", "uint8_t", "struct thread *");
375 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
376 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
377     "struct proc *");
378 SDT_PROBE_DEFINE(sched, , , on__cpu);
379 SDT_PROBE_DEFINE(sched, , , remain__cpu);
380 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
381     "struct proc *");
382 
383 /*
384  * Print the threads waiting on a run-queue.
385  */
386 static void
387 runq_print(struct runq *rq)
388 {
389 	struct rqhead *rqh;
390 	struct thread *td;
391 	int pri;
392 	int j;
393 	int i;
394 
395 	for (i = 0; i < RQB_LEN; i++) {
396 		printf("\t\trunq bits %d 0x%zx\n",
397 		    i, rq->rq_status.rqb_bits[i]);
398 		for (j = 0; j < RQB_BPW; j++)
399 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
400 				pri = j + (i << RQB_L2BPW);
401 				rqh = &rq->rq_queues[pri];
402 				TAILQ_FOREACH(td, rqh, td_runq) {
403 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
404 					    td, td->td_name, td->td_priority,
405 					    td->td_rqindex, pri);
406 				}
407 			}
408 	}
409 }
410 
411 /*
412  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
413  */
414 static void __unused
415 tdq_print(int cpu)
416 {
417 	struct tdq *tdq;
418 
419 	tdq = TDQ_CPU(cpu);
420 
421 	printf("tdq %d:\n", TDQ_ID(tdq));
422 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
423 	printf("\tLock name:      %s\n", tdq->tdq_name);
424 	printf("\tload:           %d\n", tdq->tdq_load);
425 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
426 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
427 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
428 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
429 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
430 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
431 	printf("\trealtime runq:\n");
432 	runq_print(&tdq->tdq_realtime);
433 	printf("\ttimeshare runq:\n");
434 	runq_print(&tdq->tdq_timeshare);
435 	printf("\tidle runq:\n");
436 	runq_print(&tdq->tdq_idle);
437 }
438 
439 static inline int
440 sched_shouldpreempt(int pri, int cpri, int remote)
441 {
442 	/*
443 	 * If the new priority is not better than the current priority there is
444 	 * nothing to do.
445 	 */
446 	if (pri >= cpri)
447 		return (0);
448 	/*
449 	 * Always preempt idle.
450 	 */
451 	if (cpri >= PRI_MIN_IDLE)
452 		return (1);
453 	/*
454 	 * If preemption is disabled don't preempt others.
455 	 */
456 	if (preempt_thresh == 0)
457 		return (0);
458 	/*
459 	 * Preempt if we exceed the threshold.
460 	 */
461 	if (pri <= preempt_thresh)
462 		return (1);
463 	/*
464 	 * If we're interactive or better and there is non-interactive
465 	 * or worse running preempt only remote processors.
466 	 */
467 	if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
468 		return (1);
469 	return (0);
470 }
471 
472 /*
473  * Add a thread to the actual run-queue.  Keeps transferable counts up to
474  * date with what is actually on the run-queue.  Selects the correct
475  * queue position for timeshare threads.
476  */
477 static __inline void
478 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
479 {
480 	struct td_sched *ts;
481 	u_char pri;
482 
483 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
484 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
485 
486 	pri = td->td_priority;
487 	ts = td_get_sched(td);
488 	TD_SET_RUNQ(td);
489 	if (THREAD_CAN_MIGRATE(td)) {
490 		tdq->tdq_transferable++;
491 		ts->ts_flags |= TSF_XFERABLE;
492 	}
493 	if (pri < PRI_MIN_BATCH) {
494 		ts->ts_runq = &tdq->tdq_realtime;
495 	} else if (pri <= PRI_MAX_BATCH) {
496 		ts->ts_runq = &tdq->tdq_timeshare;
497 		KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
498 			("Invalid priority %d on timeshare runq", pri));
499 		/*
500 		 * This queue contains only priorities between MIN and MAX
501 		 * batch.  Use the whole queue to represent these values.
502 		 */
503 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
504 			pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
505 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
506 			/*
507 			 * This effectively shortens the queue by one so we
508 			 * can have a one slot difference between idx and
509 			 * ridx while we wait for threads to drain.
510 			 */
511 			if (tdq->tdq_ridx != tdq->tdq_idx &&
512 			    pri == tdq->tdq_ridx)
513 				pri = (unsigned char)(pri - 1) % RQ_NQS;
514 		} else
515 			pri = tdq->tdq_ridx;
516 		runq_add_pri(ts->ts_runq, td, pri, flags);
517 		return;
518 	} else
519 		ts->ts_runq = &tdq->tdq_idle;
520 	runq_add(ts->ts_runq, td, flags);
521 }
522 
523 /*
524  * Remove a thread from a run-queue.  This typically happens when a thread
525  * is selected to run.  Running threads are not on the queue and the
526  * transferable count does not reflect them.
527  */
528 static __inline void
529 tdq_runq_rem(struct tdq *tdq, struct thread *td)
530 {
531 	struct td_sched *ts;
532 
533 	ts = td_get_sched(td);
534 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
535 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
536 	KASSERT(ts->ts_runq != NULL,
537 	    ("tdq_runq_remove: thread %p null ts_runq", td));
538 	if (ts->ts_flags & TSF_XFERABLE) {
539 		tdq->tdq_transferable--;
540 		ts->ts_flags &= ~TSF_XFERABLE;
541 	}
542 	if (ts->ts_runq == &tdq->tdq_timeshare) {
543 		if (tdq->tdq_idx != tdq->tdq_ridx)
544 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
545 		else
546 			runq_remove_idx(ts->ts_runq, td, NULL);
547 	} else
548 		runq_remove(ts->ts_runq, td);
549 }
550 
551 /*
552  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
553  * for this thread to the referenced thread queue.
554  */
555 static void
556 tdq_load_add(struct tdq *tdq, struct thread *td)
557 {
558 
559 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
560 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
561 
562 	tdq->tdq_load++;
563 	if ((td->td_flags & TDF_NOLOAD) == 0)
564 		tdq->tdq_sysload++;
565 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
566 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
567 }
568 
569 /*
570  * Remove the load from a thread that is transitioning to a sleep state or
571  * exiting.
572  */
573 static void
574 tdq_load_rem(struct tdq *tdq, struct thread *td)
575 {
576 
577 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
578 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
579 	KASSERT(tdq->tdq_load != 0,
580 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
581 
582 	tdq->tdq_load--;
583 	if ((td->td_flags & TDF_NOLOAD) == 0)
584 		tdq->tdq_sysload--;
585 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
586 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
587 }
588 
589 /*
590  * Bound timeshare latency by decreasing slice size as load increases.  We
591  * consider the maximum latency as the sum of the threads waiting to run
592  * aside from curthread and target no more than sched_slice latency but
593  * no less than sched_slice_min runtime.
594  */
595 static inline int
596 tdq_slice(struct tdq *tdq)
597 {
598 	int load;
599 
600 	/*
601 	 * It is safe to use sys_load here because this is called from
602 	 * contexts where timeshare threads are running and so there
603 	 * cannot be higher priority load in the system.
604 	 */
605 	load = tdq->tdq_sysload - 1;
606 	if (load >= SCHED_SLICE_MIN_DIVISOR)
607 		return (sched_slice_min);
608 	if (load <= 1)
609 		return (sched_slice);
610 	return (sched_slice / load);
611 }
612 
613 /*
614  * Set lowpri to its exact value by searching the run-queue and
615  * evaluating curthread.  curthread may be passed as an optimization.
616  */
617 static void
618 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
619 {
620 	struct thread *td;
621 
622 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
623 	if (ctd == NULL)
624 		ctd = tdq->tdq_curthread;
625 	td = tdq_choose(tdq);
626 	if (td == NULL || td->td_priority > ctd->td_priority)
627 		tdq->tdq_lowpri = ctd->td_priority;
628 	else
629 		tdq->tdq_lowpri = td->td_priority;
630 }
631 
632 #ifdef SMP
633 /*
634  * We need some randomness. Implement a classic Linear Congruential
635  * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
636  * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
637  * of the random state (in the low bits of our answer) to keep
638  * the maximum randomness.
639  */
640 static uint32_t
641 sched_random(void)
642 {
643 	uint32_t *rndptr;
644 
645 	rndptr = DPCPU_PTR(randomval);
646 	*rndptr = *rndptr * 69069 + 5;
647 
648 	return (*rndptr >> 16);
649 }
650 
651 struct cpu_search {
652 	cpuset_t *cs_mask;	/* The mask of allowed CPUs to choose from. */
653 	int	cs_prefer;	/* Prefer this CPU and groups including it. */
654 	int	cs_running;	/* The thread is now running at cs_prefer. */
655 	int	cs_pri;		/* Min priority for low. */
656 	int	cs_load;	/* Max load for low, min load for high. */
657 	int	cs_trans;	/* Min transferable load for high. */
658 };
659 
660 struct cpu_search_res {
661 	int	csr_cpu;	/* The best CPU found. */
662 	int	csr_load;	/* The load of cs_cpu. */
663 };
664 
665 /*
666  * Search the tree of cpu_groups for the lowest or highest loaded CPU.
667  * These routines actually compare the load on all paths through the tree
668  * and find the least loaded cpu on the least loaded path, which may differ
669  * from the least loaded cpu in the system.  This balances work among caches
670  * and buses.
671  */
672 static int
673 cpu_search_lowest(const struct cpu_group *cg, const struct cpu_search *s,
674     struct cpu_search_res *r)
675 {
676 	struct cpu_search_res lr;
677 	struct tdq *tdq;
678 	int c, bload, l, load, p, total;
679 
680 	total = 0;
681 	bload = INT_MAX;
682 	r->csr_cpu = -1;
683 
684 	/* Loop through children CPU groups if there are any. */
685 	if (cg->cg_children > 0) {
686 		for (c = cg->cg_children - 1; c >= 0; c--) {
687 			load = cpu_search_lowest(&cg->cg_child[c], s, &lr);
688 			total += load;
689 
690 			/*
691 			 * When balancing do not prefer SMT groups with load >1.
692 			 * It allows round-robin between SMT groups with equal
693 			 * load within parent group for more fair scheduling.
694 			 */
695 			if (__predict_false(s->cs_running) &&
696 			    (cg->cg_child[c].cg_flags & CG_FLAG_THREAD) &&
697 			    load >= 128 && (load & 128) != 0)
698 				load += 128;
699 
700 			if (lr.csr_cpu >= 0 && (load < bload ||
701 			    (load == bload && lr.csr_load < r->csr_load))) {
702 				bload = load;
703 				r->csr_cpu = lr.csr_cpu;
704 				r->csr_load = lr.csr_load;
705 			}
706 		}
707 		return (total);
708 	}
709 
710 	/* Loop through children CPUs otherwise. */
711 	for (c = cg->cg_last; c >= cg->cg_first; c--) {
712 		if (!CPU_ISSET(c, &cg->cg_mask))
713 			continue;
714 		tdq = TDQ_CPU(c);
715 		l = TDQ_LOAD(tdq);
716 		if (c == s->cs_prefer) {
717 			if (__predict_false(s->cs_running))
718 				l--;
719 			p = 128;
720 		} else
721 			p = 0;
722 		load = l * 256;
723 		total += load - p;
724 
725 		/*
726 		 * Check this CPU is acceptable.
727 		 * If the threads is already on the CPU, don't look on the TDQ
728 		 * priority, since it can be the priority of the thread itself.
729 		 */
730 		if (l > s->cs_load ||
731 		    (atomic_load_char(&tdq->tdq_lowpri) <= s->cs_pri &&
732 		     (!s->cs_running || c != s->cs_prefer)) ||
733 		    !CPU_ISSET(c, s->cs_mask))
734 			continue;
735 
736 		/*
737 		 * When balancing do not prefer CPUs with load > 1.
738 		 * It allows round-robin between CPUs with equal load
739 		 * within the CPU group for more fair scheduling.
740 		 */
741 		if (__predict_false(s->cs_running) && l > 0)
742 			p = 0;
743 
744 		load -= sched_random() % 128;
745 		if (bload > load - p) {
746 			bload = load - p;
747 			r->csr_cpu = c;
748 			r->csr_load = load;
749 		}
750 	}
751 	return (total);
752 }
753 
754 static int
755 cpu_search_highest(const struct cpu_group *cg, const struct cpu_search *s,
756     struct cpu_search_res *r)
757 {
758 	struct cpu_search_res lr;
759 	struct tdq *tdq;
760 	int c, bload, l, load, total;
761 
762 	total = 0;
763 	bload = INT_MIN;
764 	r->csr_cpu = -1;
765 
766 	/* Loop through children CPU groups if there are any. */
767 	if (cg->cg_children > 0) {
768 		for (c = cg->cg_children - 1; c >= 0; c--) {
769 			load = cpu_search_highest(&cg->cg_child[c], s, &lr);
770 			total += load;
771 			if (lr.csr_cpu >= 0 && (load > bload ||
772 			    (load == bload && lr.csr_load > r->csr_load))) {
773 				bload = load;
774 				r->csr_cpu = lr.csr_cpu;
775 				r->csr_load = lr.csr_load;
776 			}
777 		}
778 		return (total);
779 	}
780 
781 	/* Loop through children CPUs otherwise. */
782 	for (c = cg->cg_last; c >= cg->cg_first; c--) {
783 		if (!CPU_ISSET(c, &cg->cg_mask))
784 			continue;
785 		tdq = TDQ_CPU(c);
786 		l = TDQ_LOAD(tdq);
787 		load = l * 256;
788 		total += load;
789 
790 		/*
791 		 * Check this CPU is acceptable.
792 		 */
793 		if (l < s->cs_load || TDQ_TRANSFERABLE(tdq) < s->cs_trans ||
794 		    !CPU_ISSET(c, s->cs_mask))
795 			continue;
796 
797 		load -= sched_random() % 256;
798 		if (load > bload) {
799 			bload = load;
800 			r->csr_cpu = c;
801 		}
802 	}
803 	r->csr_load = bload;
804 	return (total);
805 }
806 
807 /*
808  * Find the cpu with the least load via the least loaded path that has a
809  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
810  * acceptable.
811  */
812 static inline int
813 sched_lowest(const struct cpu_group *cg, cpuset_t *mask, int pri, int maxload,
814     int prefer, int running)
815 {
816 	struct cpu_search s;
817 	struct cpu_search_res r;
818 
819 	s.cs_prefer = prefer;
820 	s.cs_running = running;
821 	s.cs_mask = mask;
822 	s.cs_pri = pri;
823 	s.cs_load = maxload;
824 	cpu_search_lowest(cg, &s, &r);
825 	return (r.csr_cpu);
826 }
827 
828 /*
829  * Find the cpu with the highest load via the highest loaded path.
830  */
831 static inline int
832 sched_highest(const struct cpu_group *cg, cpuset_t *mask, int minload,
833     int mintrans)
834 {
835 	struct cpu_search s;
836 	struct cpu_search_res r;
837 
838 	s.cs_mask = mask;
839 	s.cs_load = minload;
840 	s.cs_trans = mintrans;
841 	cpu_search_highest(cg, &s, &r);
842 	return (r.csr_cpu);
843 }
844 
845 static void
846 sched_balance_group(struct cpu_group *cg)
847 {
848 	struct tdq *tdq;
849 	struct thread *td;
850 	cpuset_t hmask, lmask;
851 	int high, low, anylow;
852 
853 	CPU_FILL(&hmask);
854 	for (;;) {
855 		high = sched_highest(cg, &hmask, 1, 0);
856 		/* Stop if there is no more CPU with transferrable threads. */
857 		if (high == -1)
858 			break;
859 		CPU_CLR(high, &hmask);
860 		CPU_COPY(&hmask, &lmask);
861 		/* Stop if there is no more CPU left for low. */
862 		if (CPU_EMPTY(&lmask))
863 			break;
864 		tdq = TDQ_CPU(high);
865 		if (TDQ_LOAD(tdq) == 1) {
866 			/*
867 			 * There is only one running thread.  We can't move
868 			 * it from here, so tell it to pick new CPU by itself.
869 			 */
870 			TDQ_LOCK(tdq);
871 			td = tdq->tdq_curthread;
872 			if (td->td_lock == TDQ_LOCKPTR(tdq) &&
873 			    (td->td_flags & TDF_IDLETD) == 0 &&
874 			    THREAD_CAN_MIGRATE(td)) {
875 				td->td_flags |= TDF_PICKCPU;
876 				ast_sched_locked(td, TDA_SCHED);
877 				if (high != curcpu)
878 					ipi_cpu(high, IPI_AST);
879 			}
880 			TDQ_UNLOCK(tdq);
881 			break;
882 		}
883 		anylow = 1;
884 nextlow:
885 		if (TDQ_TRANSFERABLE(tdq) == 0)
886 			continue;
887 		low = sched_lowest(cg, &lmask, -1, TDQ_LOAD(tdq) - 1, high, 1);
888 		/* Stop if we looked well and found no less loaded CPU. */
889 		if (anylow && low == -1)
890 			break;
891 		/* Go to next high if we found no less loaded CPU. */
892 		if (low == -1)
893 			continue;
894 		/* Transfer thread from high to low. */
895 		if (sched_balance_pair(tdq, TDQ_CPU(low))) {
896 			/* CPU that got thread can no longer be a donor. */
897 			CPU_CLR(low, &hmask);
898 		} else {
899 			/*
900 			 * If failed, then there is no threads on high
901 			 * that can run on this low. Drop low from low
902 			 * mask and look for different one.
903 			 */
904 			CPU_CLR(low, &lmask);
905 			anylow = 0;
906 			goto nextlow;
907 		}
908 	}
909 }
910 
911 static void
912 sched_balance(void)
913 {
914 	struct tdq *tdq;
915 
916 	balance_ticks = max(balance_interval / 2, 1) +
917 	    (sched_random() % balance_interval);
918 	tdq = TDQ_SELF();
919 	TDQ_UNLOCK(tdq);
920 	sched_balance_group(cpu_top);
921 	TDQ_LOCK(tdq);
922 }
923 
924 /*
925  * Lock two thread queues using their address to maintain lock order.
926  */
927 static void
928 tdq_lock_pair(struct tdq *one, struct tdq *two)
929 {
930 	if (one < two) {
931 		TDQ_LOCK(one);
932 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
933 	} else {
934 		TDQ_LOCK(two);
935 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
936 	}
937 }
938 
939 /*
940  * Unlock two thread queues.  Order is not important here.
941  */
942 static void
943 tdq_unlock_pair(struct tdq *one, struct tdq *two)
944 {
945 	TDQ_UNLOCK(one);
946 	TDQ_UNLOCK(two);
947 }
948 
949 /*
950  * Transfer load between two imbalanced thread queues.  Returns true if a thread
951  * was moved between the queues, and false otherwise.
952  */
953 static bool
954 sched_balance_pair(struct tdq *high, struct tdq *low)
955 {
956 	int cpu, lowpri;
957 	bool ret;
958 
959 	ret = false;
960 	tdq_lock_pair(high, low);
961 
962 	/*
963 	 * Transfer a thread from high to low.
964 	 */
965 	if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load) {
966 		lowpri = tdq_move(high, low);
967 		if (lowpri != -1) {
968 			/*
969 			 * In case the target isn't the current CPU notify it of
970 			 * the new load, possibly sending an IPI to force it to
971 			 * reschedule.  Otherwise maybe schedule a preemption.
972 			 */
973 			cpu = TDQ_ID(low);
974 			if (cpu != PCPU_GET(cpuid))
975 				tdq_notify(low, lowpri);
976 			else
977 				sched_setpreempt(low->tdq_lowpri);
978 			ret = true;
979 		}
980 	}
981 	tdq_unlock_pair(high, low);
982 	return (ret);
983 }
984 
985 /*
986  * Move a thread from one thread queue to another.  Returns -1 if the source
987  * queue was empty, else returns the maximum priority of all threads in
988  * the destination queue prior to the addition of the new thread.  In the latter
989  * case, this priority can be used to determine whether an IPI needs to be
990  * delivered.
991  */
992 static int
993 tdq_move(struct tdq *from, struct tdq *to)
994 {
995 	struct thread *td;
996 	int cpu;
997 
998 	TDQ_LOCK_ASSERT(from, MA_OWNED);
999 	TDQ_LOCK_ASSERT(to, MA_OWNED);
1000 
1001 	cpu = TDQ_ID(to);
1002 	td = tdq_steal(from, cpu);
1003 	if (td == NULL)
1004 		return (-1);
1005 
1006 	/*
1007 	 * Although the run queue is locked the thread may be
1008 	 * blocked.  We can not set the lock until it is unblocked.
1009 	 */
1010 	thread_lock_block_wait(td);
1011 	sched_rem(td);
1012 	THREAD_LOCKPTR_ASSERT(td, TDQ_LOCKPTR(from));
1013 	td->td_lock = TDQ_LOCKPTR(to);
1014 	td_get_sched(td)->ts_cpu = cpu;
1015 	return (tdq_add(to, td, SRQ_YIELDING));
1016 }
1017 
1018 /*
1019  * This tdq has idled.  Try to steal a thread from another cpu and switch
1020  * to it.
1021  */
1022 static int
1023 tdq_idled(struct tdq *tdq)
1024 {
1025 	struct cpu_group *cg, *parent;
1026 	struct tdq *steal;
1027 	cpuset_t mask;
1028 	int cpu, switchcnt, goup;
1029 
1030 	if (smp_started == 0 || steal_idle == 0 || tdq->tdq_cg == NULL)
1031 		return (1);
1032 	CPU_FILL(&mask);
1033 	CPU_CLR(PCPU_GET(cpuid), &mask);
1034 restart:
1035 	switchcnt = TDQ_SWITCHCNT(tdq);
1036 	for (cg = tdq->tdq_cg, goup = 0; ; ) {
1037 		cpu = sched_highest(cg, &mask, steal_thresh, 1);
1038 		/*
1039 		 * We were assigned a thread but not preempted.  Returning
1040 		 * 0 here will cause our caller to switch to it.
1041 		 */
1042 		if (TDQ_LOAD(tdq))
1043 			return (0);
1044 
1045 		/*
1046 		 * We found no CPU to steal from in this group.  Escalate to
1047 		 * the parent and repeat.  But if parent has only two children
1048 		 * groups we can avoid searching this group again by searching
1049 		 * the other one specifically and then escalating two levels.
1050 		 */
1051 		if (cpu == -1) {
1052 			if (goup) {
1053 				cg = cg->cg_parent;
1054 				goup = 0;
1055 			}
1056 			parent = cg->cg_parent;
1057 			if (parent == NULL)
1058 				return (1);
1059 			if (parent->cg_children == 2) {
1060 				if (cg == &parent->cg_child[0])
1061 					cg = &parent->cg_child[1];
1062 				else
1063 					cg = &parent->cg_child[0];
1064 				goup = 1;
1065 			} else
1066 				cg = parent;
1067 			continue;
1068 		}
1069 		steal = TDQ_CPU(cpu);
1070 		/*
1071 		 * The data returned by sched_highest() is stale and
1072 		 * the chosen CPU no longer has an eligible thread.
1073 		 *
1074 		 * Testing this ahead of tdq_lock_pair() only catches
1075 		 * this situation about 20% of the time on an 8 core
1076 		 * 16 thread Ryzen 7, but it still helps performance.
1077 		 */
1078 		if (TDQ_LOAD(steal) < steal_thresh ||
1079 		    TDQ_TRANSFERABLE(steal) == 0)
1080 			goto restart;
1081 		/*
1082 		 * Try to lock both queues. If we are assigned a thread while
1083 		 * waited for the lock, switch to it now instead of stealing.
1084 		 * If we can't get the lock, then somebody likely got there
1085 		 * first so continue searching.
1086 		 */
1087 		TDQ_LOCK(tdq);
1088 		if (tdq->tdq_load > 0) {
1089 			mi_switch(SW_VOL | SWT_IDLE);
1090 			return (0);
1091 		}
1092 		if (TDQ_TRYLOCK_FLAGS(steal, MTX_DUPOK) == 0) {
1093 			TDQ_UNLOCK(tdq);
1094 			CPU_CLR(cpu, &mask);
1095 			continue;
1096 		}
1097 		/*
1098 		 * The data returned by sched_highest() is stale and
1099 		 * the chosen CPU no longer has an eligible thread, or
1100 		 * we were preempted and the CPU loading info may be out
1101 		 * of date.  The latter is rare.  In either case restart
1102 		 * the search.
1103 		 */
1104 		if (TDQ_LOAD(steal) < steal_thresh ||
1105 		    TDQ_TRANSFERABLE(steal) == 0 ||
1106 		    switchcnt != TDQ_SWITCHCNT(tdq)) {
1107 			tdq_unlock_pair(tdq, steal);
1108 			goto restart;
1109 		}
1110 		/*
1111 		 * Steal the thread and switch to it.
1112 		 */
1113 		if (tdq_move(steal, tdq) != -1)
1114 			break;
1115 		/*
1116 		 * We failed to acquire a thread even though it looked
1117 		 * like one was available.  This could be due to affinity
1118 		 * restrictions or for other reasons.  Loop again after
1119 		 * removing this CPU from the set.  The restart logic
1120 		 * above does not restore this CPU to the set due to the
1121 		 * likelyhood of failing here again.
1122 		 */
1123 		CPU_CLR(cpu, &mask);
1124 		tdq_unlock_pair(tdq, steal);
1125 	}
1126 	TDQ_UNLOCK(steal);
1127 	mi_switch(SW_VOL | SWT_IDLE);
1128 	return (0);
1129 }
1130 
1131 /*
1132  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
1133  *
1134  * "lowpri" is the minimum scheduling priority among all threads on
1135  * the queue prior to the addition of the new thread.
1136  */
1137 static void
1138 tdq_notify(struct tdq *tdq, int lowpri)
1139 {
1140 	int cpu;
1141 
1142 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1143 	KASSERT(tdq->tdq_lowpri <= lowpri,
1144 	    ("tdq_notify: lowpri %d > tdq_lowpri %d", lowpri, tdq->tdq_lowpri));
1145 
1146 	if (tdq->tdq_owepreempt)
1147 		return;
1148 
1149 	/*
1150 	 * Check to see if the newly added thread should preempt the one
1151 	 * currently running.
1152 	 */
1153 	if (!sched_shouldpreempt(tdq->tdq_lowpri, lowpri, 1))
1154 		return;
1155 
1156 	/*
1157 	 * Make sure that our caller's earlier update to tdq_load is
1158 	 * globally visible before we read tdq_cpu_idle.  Idle thread
1159 	 * accesses both of them without locks, and the order is important.
1160 	 */
1161 	atomic_thread_fence_seq_cst();
1162 
1163 	/*
1164 	 * Try to figure out if we can signal the idle thread instead of sending
1165 	 * an IPI.  This check is racy; at worst, we will deliever an IPI
1166 	 * unnecessarily.
1167 	 */
1168 	cpu = TDQ_ID(tdq);
1169 	if (TD_IS_IDLETHREAD(tdq->tdq_curthread) &&
1170 	    (atomic_load_int(&tdq->tdq_cpu_idle) == 0 || cpu_idle_wakeup(cpu)))
1171 		return;
1172 
1173 	/*
1174 	 * The run queues have been updated, so any switch on the remote CPU
1175 	 * will satisfy the preemption request.
1176 	 */
1177 	tdq->tdq_owepreempt = 1;
1178 	ipi_cpu(cpu, IPI_PREEMPT);
1179 }
1180 
1181 /*
1182  * Steals load from a timeshare queue.  Honors the rotating queue head
1183  * index.
1184  */
1185 static struct thread *
1186 runq_steal_from(struct runq *rq, int cpu, u_char start)
1187 {
1188 	struct rqbits *rqb;
1189 	struct rqhead *rqh;
1190 	struct thread *td, *first;
1191 	int bit;
1192 	int i;
1193 
1194 	rqb = &rq->rq_status;
1195 	bit = start & (RQB_BPW -1);
1196 	first = NULL;
1197 again:
1198 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1199 		if (rqb->rqb_bits[i] == 0)
1200 			continue;
1201 		if (bit == 0)
1202 			bit = RQB_FFS(rqb->rqb_bits[i]);
1203 		for (; bit < RQB_BPW; bit++) {
1204 			if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1205 				continue;
1206 			rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1207 			TAILQ_FOREACH(td, rqh, td_runq) {
1208 				if (first) {
1209 					if (THREAD_CAN_MIGRATE(td) &&
1210 					    THREAD_CAN_SCHED(td, cpu))
1211 						return (td);
1212 				} else
1213 					first = td;
1214 			}
1215 		}
1216 	}
1217 	if (start != 0) {
1218 		start = 0;
1219 		goto again;
1220 	}
1221 
1222 	if (first && THREAD_CAN_MIGRATE(first) &&
1223 	    THREAD_CAN_SCHED(first, cpu))
1224 		return (first);
1225 	return (NULL);
1226 }
1227 
1228 /*
1229  * Steals load from a standard linear queue.
1230  */
1231 static struct thread *
1232 runq_steal(struct runq *rq, int cpu)
1233 {
1234 	struct rqhead *rqh;
1235 	struct rqbits *rqb;
1236 	struct thread *td;
1237 	int word;
1238 	int bit;
1239 
1240 	rqb = &rq->rq_status;
1241 	for (word = 0; word < RQB_LEN; word++) {
1242 		if (rqb->rqb_bits[word] == 0)
1243 			continue;
1244 		for (bit = 0; bit < RQB_BPW; bit++) {
1245 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1246 				continue;
1247 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1248 			TAILQ_FOREACH(td, rqh, td_runq)
1249 				if (THREAD_CAN_MIGRATE(td) &&
1250 				    THREAD_CAN_SCHED(td, cpu))
1251 					return (td);
1252 		}
1253 	}
1254 	return (NULL);
1255 }
1256 
1257 /*
1258  * Attempt to steal a thread in priority order from a thread queue.
1259  */
1260 static struct thread *
1261 tdq_steal(struct tdq *tdq, int cpu)
1262 {
1263 	struct thread *td;
1264 
1265 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1266 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1267 		return (td);
1268 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1269 	    cpu, tdq->tdq_ridx)) != NULL)
1270 		return (td);
1271 	return (runq_steal(&tdq->tdq_idle, cpu));
1272 }
1273 
1274 /*
1275  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1276  * current lock and returns with the assigned queue locked.
1277  */
1278 static inline struct tdq *
1279 sched_setcpu(struct thread *td, int cpu, int flags)
1280 {
1281 
1282 	struct tdq *tdq;
1283 	struct mtx *mtx;
1284 
1285 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1286 	tdq = TDQ_CPU(cpu);
1287 	td_get_sched(td)->ts_cpu = cpu;
1288 	/*
1289 	 * If the lock matches just return the queue.
1290 	 */
1291 	if (td->td_lock == TDQ_LOCKPTR(tdq)) {
1292 		KASSERT((flags & SRQ_HOLD) == 0,
1293 		    ("sched_setcpu: Invalid lock for SRQ_HOLD"));
1294 		return (tdq);
1295 	}
1296 
1297 	/*
1298 	 * The hard case, migration, we need to block the thread first to
1299 	 * prevent order reversals with other cpus locks.
1300 	 */
1301 	spinlock_enter();
1302 	mtx = thread_lock_block(td);
1303 	if ((flags & SRQ_HOLD) == 0)
1304 		mtx_unlock_spin(mtx);
1305 	TDQ_LOCK(tdq);
1306 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1307 	spinlock_exit();
1308 	return (tdq);
1309 }
1310 
1311 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1312 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1313 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1314 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1315 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1316 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1317 
1318 static int
1319 sched_pickcpu(struct thread *td, int flags)
1320 {
1321 	struct cpu_group *cg, *ccg;
1322 	struct td_sched *ts;
1323 	struct tdq *tdq;
1324 	cpuset_t *mask;
1325 	int cpu, pri, r, self, intr;
1326 
1327 	self = PCPU_GET(cpuid);
1328 	ts = td_get_sched(td);
1329 	KASSERT(!CPU_ABSENT(ts->ts_cpu), ("sched_pickcpu: Start scheduler on "
1330 	    "absent CPU %d for thread %s.", ts->ts_cpu, td->td_name));
1331 	if (smp_started == 0)
1332 		return (self);
1333 	/*
1334 	 * Don't migrate a running thread from sched_switch().
1335 	 */
1336 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1337 		return (ts->ts_cpu);
1338 	/*
1339 	 * Prefer to run interrupt threads on the processors that generate
1340 	 * the interrupt.
1341 	 */
1342 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1343 	    curthread->td_intr_nesting_level) {
1344 		tdq = TDQ_SELF();
1345 		if (tdq->tdq_lowpri >= PRI_MIN_IDLE) {
1346 			SCHED_STAT_INC(pickcpu_idle_affinity);
1347 			return (self);
1348 		}
1349 		ts->ts_cpu = self;
1350 		intr = 1;
1351 		cg = tdq->tdq_cg;
1352 		goto llc;
1353 	} else {
1354 		intr = 0;
1355 		tdq = TDQ_CPU(ts->ts_cpu);
1356 		cg = tdq->tdq_cg;
1357 	}
1358 	/*
1359 	 * If the thread can run on the last cpu and the affinity has not
1360 	 * expired and it is idle, run it there.
1361 	 */
1362 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1363 	    atomic_load_char(&tdq->tdq_lowpri) >= PRI_MIN_IDLE &&
1364 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1365 		if (cg->cg_flags & CG_FLAG_THREAD) {
1366 			/* Check all SMT threads for being idle. */
1367 			for (cpu = cg->cg_first; cpu <= cg->cg_last; cpu++) {
1368 				pri =
1369 				    atomic_load_char(&TDQ_CPU(cpu)->tdq_lowpri);
1370 				if (CPU_ISSET(cpu, &cg->cg_mask) &&
1371 				    pri < PRI_MIN_IDLE)
1372 					break;
1373 			}
1374 			if (cpu > cg->cg_last) {
1375 				SCHED_STAT_INC(pickcpu_idle_affinity);
1376 				return (ts->ts_cpu);
1377 			}
1378 		} else {
1379 			SCHED_STAT_INC(pickcpu_idle_affinity);
1380 			return (ts->ts_cpu);
1381 		}
1382 	}
1383 llc:
1384 	/*
1385 	 * Search for the last level cache CPU group in the tree.
1386 	 * Skip SMT, identical groups and caches with expired affinity.
1387 	 * Interrupt threads affinity is explicit and never expires.
1388 	 */
1389 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1390 		if (cg->cg_flags & CG_FLAG_THREAD)
1391 			continue;
1392 		if (cg->cg_children == 1 || cg->cg_count == 1)
1393 			continue;
1394 		if (cg->cg_level == CG_SHARE_NONE ||
1395 		    (!intr && !SCHED_AFFINITY(ts, cg->cg_level)))
1396 			continue;
1397 		ccg = cg;
1398 	}
1399 	/* Found LLC shared by all CPUs, so do a global search. */
1400 	if (ccg == cpu_top)
1401 		ccg = NULL;
1402 	cpu = -1;
1403 	mask = &td->td_cpuset->cs_mask;
1404 	pri = td->td_priority;
1405 	r = TD_IS_RUNNING(td);
1406 	/*
1407 	 * Try hard to keep interrupts within found LLC.  Search the LLC for
1408 	 * the least loaded CPU we can run now.  For NUMA systems it should
1409 	 * be within target domain, and it also reduces scheduling overhead.
1410 	 */
1411 	if (ccg != NULL && intr) {
1412 		cpu = sched_lowest(ccg, mask, pri, INT_MAX, ts->ts_cpu, r);
1413 		if (cpu >= 0)
1414 			SCHED_STAT_INC(pickcpu_intrbind);
1415 	} else
1416 	/* Search the LLC for the least loaded idle CPU we can run now. */
1417 	if (ccg != NULL) {
1418 		cpu = sched_lowest(ccg, mask, max(pri, PRI_MAX_TIMESHARE),
1419 		    INT_MAX, ts->ts_cpu, r);
1420 		if (cpu >= 0)
1421 			SCHED_STAT_INC(pickcpu_affinity);
1422 	}
1423 	/* Search globally for the least loaded CPU we can run now. */
1424 	if (cpu < 0) {
1425 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu, r);
1426 		if (cpu >= 0)
1427 			SCHED_STAT_INC(pickcpu_lowest);
1428 	}
1429 	/* Search globally for the least loaded CPU. */
1430 	if (cpu < 0) {
1431 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu, r);
1432 		if (cpu >= 0)
1433 			SCHED_STAT_INC(pickcpu_lowest);
1434 	}
1435 	KASSERT(cpu >= 0, ("sched_pickcpu: Failed to find a cpu."));
1436 	KASSERT(!CPU_ABSENT(cpu), ("sched_pickcpu: Picked absent CPU %d.", cpu));
1437 	/*
1438 	 * Compare the lowest loaded cpu to current cpu.
1439 	 */
1440 	tdq = TDQ_CPU(cpu);
1441 	if (THREAD_CAN_SCHED(td, self) && TDQ_SELF()->tdq_lowpri > pri &&
1442 	    atomic_load_char(&tdq->tdq_lowpri) < PRI_MIN_IDLE &&
1443 	    TDQ_LOAD(TDQ_SELF()) <= TDQ_LOAD(tdq) + 1) {
1444 		SCHED_STAT_INC(pickcpu_local);
1445 		cpu = self;
1446 	}
1447 	if (cpu != ts->ts_cpu)
1448 		SCHED_STAT_INC(pickcpu_migration);
1449 	return (cpu);
1450 }
1451 #endif
1452 
1453 /*
1454  * Pick the highest priority task we have and return it.
1455  */
1456 static struct thread *
1457 tdq_choose(struct tdq *tdq)
1458 {
1459 	struct thread *td;
1460 
1461 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1462 	td = runq_choose(&tdq->tdq_realtime);
1463 	if (td != NULL)
1464 		return (td);
1465 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1466 	if (td != NULL) {
1467 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1468 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1469 		    td->td_priority));
1470 		return (td);
1471 	}
1472 	td = runq_choose(&tdq->tdq_idle);
1473 	if (td != NULL) {
1474 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1475 		    ("tdq_choose: Invalid priority on idle queue %d",
1476 		    td->td_priority));
1477 		return (td);
1478 	}
1479 
1480 	return (NULL);
1481 }
1482 
1483 /*
1484  * Initialize a thread queue.
1485  */
1486 static void
1487 tdq_setup(struct tdq *tdq, int id)
1488 {
1489 
1490 	if (bootverbose)
1491 		printf("ULE: setup cpu %d\n", id);
1492 	runq_init(&tdq->tdq_realtime);
1493 	runq_init(&tdq->tdq_timeshare);
1494 	runq_init(&tdq->tdq_idle);
1495 	tdq->tdq_id = id;
1496 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1497 	    "sched lock %d", (int)TDQ_ID(tdq));
1498 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", MTX_SPIN);
1499 #ifdef KTR
1500 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1501 	    "CPU %d load", (int)TDQ_ID(tdq));
1502 #endif
1503 }
1504 
1505 #ifdef SMP
1506 static void
1507 sched_setup_smp(void)
1508 {
1509 	struct tdq *tdq;
1510 	int i;
1511 
1512 	cpu_top = smp_topo();
1513 	CPU_FOREACH(i) {
1514 		tdq = DPCPU_ID_PTR(i, tdq);
1515 		tdq_setup(tdq, i);
1516 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1517 		if (tdq->tdq_cg == NULL)
1518 			panic("Can't find cpu group for %d\n", i);
1519 		DPCPU_ID_SET(i, randomval, i * 69069 + 5);
1520 	}
1521 	PCPU_SET(sched, DPCPU_PTR(tdq));
1522 	balance_tdq = TDQ_SELF();
1523 }
1524 #endif
1525 
1526 /*
1527  * Setup the thread queues and initialize the topology based on MD
1528  * information.
1529  */
1530 static void
1531 sched_setup(void *dummy)
1532 {
1533 	struct tdq *tdq;
1534 
1535 #ifdef SMP
1536 	sched_setup_smp();
1537 #else
1538 	tdq_setup(TDQ_SELF(), 0);
1539 #endif
1540 	tdq = TDQ_SELF();
1541 
1542 	/* Add thread0's load since it's running. */
1543 	TDQ_LOCK(tdq);
1544 	thread0.td_lock = TDQ_LOCKPTR(tdq);
1545 	tdq_load_add(tdq, &thread0);
1546 	tdq->tdq_curthread = &thread0;
1547 	tdq->tdq_lowpri = thread0.td_priority;
1548 	TDQ_UNLOCK(tdq);
1549 }
1550 
1551 /*
1552  * This routine determines time constants after stathz and hz are setup.
1553  */
1554 /* ARGSUSED */
1555 static void
1556 sched_initticks(void *dummy)
1557 {
1558 	int incr;
1559 
1560 	realstathz = stathz ? stathz : hz;
1561 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1562 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1563 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1564 	    realstathz);
1565 
1566 	/*
1567 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1568 	 * hz not being evenly divisible by stathz on all platforms.
1569 	 */
1570 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1571 	/*
1572 	 * This does not work for values of stathz that are more than
1573 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1574 	 */
1575 	if (incr == 0)
1576 		incr = 1;
1577 	tickincr = incr;
1578 #ifdef SMP
1579 	/*
1580 	 * Set the default balance interval now that we know
1581 	 * what realstathz is.
1582 	 */
1583 	balance_interval = realstathz;
1584 	balance_ticks = balance_interval;
1585 	affinity = SCHED_AFFINITY_DEFAULT;
1586 #endif
1587 	if (sched_idlespinthresh < 0)
1588 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1589 }
1590 
1591 /*
1592  * This is the core of the interactivity algorithm.  Determines a score based
1593  * on past behavior.  It is the ratio of sleep time to run time scaled to
1594  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1595  * differs from the cpu usage because it does not account for time spent
1596  * waiting on a run-queue.  Would be prettier if we had floating point.
1597  *
1598  * When a thread's sleep time is greater than its run time the
1599  * calculation is:
1600  *
1601  *                           scaling factor
1602  * interactivity score =  ---------------------
1603  *                        sleep time / run time
1604  *
1605  *
1606  * When a thread's run time is greater than its sleep time the
1607  * calculation is:
1608  *
1609  *                                                 scaling factor
1610  * interactivity score = 2 * scaling factor  -  ---------------------
1611  *                                              run time / sleep time
1612  */
1613 static int
1614 sched_interact_score(struct thread *td)
1615 {
1616 	struct td_sched *ts;
1617 	int div;
1618 
1619 	ts = td_get_sched(td);
1620 	/*
1621 	 * The score is only needed if this is likely to be an interactive
1622 	 * task.  Don't go through the expense of computing it if there's
1623 	 * no chance.
1624 	 */
1625 	if (sched_interact <= SCHED_INTERACT_HALF &&
1626 		ts->ts_runtime >= ts->ts_slptime)
1627 			return (SCHED_INTERACT_HALF);
1628 
1629 	if (ts->ts_runtime > ts->ts_slptime) {
1630 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1631 		return (SCHED_INTERACT_HALF +
1632 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1633 	}
1634 	if (ts->ts_slptime > ts->ts_runtime) {
1635 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1636 		return (ts->ts_runtime / div);
1637 	}
1638 	/* runtime == slptime */
1639 	if (ts->ts_runtime)
1640 		return (SCHED_INTERACT_HALF);
1641 
1642 	/*
1643 	 * This can happen if slptime and runtime are 0.
1644 	 */
1645 	return (0);
1646 
1647 }
1648 
1649 /*
1650  * Scale the scheduling priority according to the "interactivity" of this
1651  * process.
1652  */
1653 static void
1654 sched_priority(struct thread *td)
1655 {
1656 	u_int pri, score;
1657 
1658 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1659 		return;
1660 	/*
1661 	 * If the score is interactive we place the thread in the realtime
1662 	 * queue with a priority that is less than kernel and interrupt
1663 	 * priorities.  These threads are not subject to nice restrictions.
1664 	 *
1665 	 * Scores greater than this are placed on the normal timeshare queue
1666 	 * where the priority is partially decided by the most recent cpu
1667 	 * utilization and the rest is decided by nice value.
1668 	 *
1669 	 * The nice value of the process has a linear effect on the calculated
1670 	 * score.  Negative nice values make it easier for a thread to be
1671 	 * considered interactive.
1672 	 */
1673 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1674 	if (score < sched_interact) {
1675 		pri = PRI_MIN_INTERACT;
1676 		pri += (PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) * score /
1677 		    sched_interact;
1678 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1679 		    ("sched_priority: invalid interactive priority %u score %u",
1680 		    pri, score));
1681 	} else {
1682 		pri = SCHED_PRI_MIN;
1683 		if (td_get_sched(td)->ts_ticks)
1684 			pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1685 			    SCHED_PRI_RANGE - 1);
1686 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1687 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1688 		    ("sched_priority: invalid priority %u: nice %d, "
1689 		    "ticks %d ftick %d ltick %d tick pri %d",
1690 		    pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1691 		    td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1692 		    SCHED_PRI_TICKS(td_get_sched(td))));
1693 	}
1694 	sched_user_prio(td, pri);
1695 
1696 	return;
1697 }
1698 
1699 /*
1700  * This routine enforces a maximum limit on the amount of scheduling history
1701  * kept.  It is called after either the slptime or runtime is adjusted.  This
1702  * function is ugly due to integer math.
1703  */
1704 static void
1705 sched_interact_update(struct thread *td)
1706 {
1707 	struct td_sched *ts;
1708 	u_int sum;
1709 
1710 	ts = td_get_sched(td);
1711 	sum = ts->ts_runtime + ts->ts_slptime;
1712 	if (sum < SCHED_SLP_RUN_MAX)
1713 		return;
1714 	/*
1715 	 * This only happens from two places:
1716 	 * 1) We have added an unusual amount of run time from fork_exit.
1717 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1718 	 */
1719 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1720 		if (ts->ts_runtime > ts->ts_slptime) {
1721 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1722 			ts->ts_slptime = 1;
1723 		} else {
1724 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1725 			ts->ts_runtime = 1;
1726 		}
1727 		return;
1728 	}
1729 	/*
1730 	 * If we have exceeded by more than 1/5th then the algorithm below
1731 	 * will not bring us back into range.  Dividing by two here forces
1732 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1733 	 */
1734 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1735 		ts->ts_runtime /= 2;
1736 		ts->ts_slptime /= 2;
1737 		return;
1738 	}
1739 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1740 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1741 }
1742 
1743 /*
1744  * Scale back the interactivity history when a child thread is created.  The
1745  * history is inherited from the parent but the thread may behave totally
1746  * differently.  For example, a shell spawning a compiler process.  We want
1747  * to learn that the compiler is behaving badly very quickly.
1748  */
1749 static void
1750 sched_interact_fork(struct thread *td)
1751 {
1752 	struct td_sched *ts;
1753 	int ratio;
1754 	int sum;
1755 
1756 	ts = td_get_sched(td);
1757 	sum = ts->ts_runtime + ts->ts_slptime;
1758 	if (sum > SCHED_SLP_RUN_FORK) {
1759 		ratio = sum / SCHED_SLP_RUN_FORK;
1760 		ts->ts_runtime /= ratio;
1761 		ts->ts_slptime /= ratio;
1762 	}
1763 }
1764 
1765 /*
1766  * Called from proc0_init() to setup the scheduler fields.
1767  */
1768 void
1769 schedinit(void)
1770 {
1771 	struct td_sched *ts0;
1772 
1773 	/*
1774 	 * Set up the scheduler specific parts of thread0.
1775 	 */
1776 	ts0 = td_get_sched(&thread0);
1777 	ts0->ts_ltick = ticks;
1778 	ts0->ts_ftick = ticks;
1779 	ts0->ts_slice = 0;
1780 	ts0->ts_cpu = curcpu;	/* set valid CPU number */
1781 }
1782 
1783 /*
1784  * schedinit_ap() is needed prior to calling sched_throw(NULL) to ensure that
1785  * the pcpu requirements are met for any calls in the period between curthread
1786  * initialization and sched_throw().  One can safely add threads to the queue
1787  * before sched_throw(), for instance, as long as the thread lock is setup
1788  * correctly.
1789  *
1790  * TDQ_SELF() relies on the below sched pcpu setting; it may be used only
1791  * after schedinit_ap().
1792  */
1793 void
1794 schedinit_ap(void)
1795 {
1796 
1797 #ifdef SMP
1798 	PCPU_SET(sched, DPCPU_PTR(tdq));
1799 #endif
1800 	PCPU_GET(idlethread)->td_lock = TDQ_LOCKPTR(TDQ_SELF());
1801 }
1802 
1803 /*
1804  * This is only somewhat accurate since given many processes of the same
1805  * priority they will switch when their slices run out, which will be
1806  * at most sched_slice stathz ticks.
1807  */
1808 int
1809 sched_rr_interval(void)
1810 {
1811 
1812 	/* Convert sched_slice from stathz to hz. */
1813 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1814 }
1815 
1816 /*
1817  * Update the percent cpu tracking information when it is requested or
1818  * the total history exceeds the maximum.  We keep a sliding history of
1819  * tick counts that slowly decays.  This is less precise than the 4BSD
1820  * mechanism since it happens with less regular and frequent events.
1821  */
1822 static void
1823 sched_pctcpu_update(struct td_sched *ts, int run)
1824 {
1825 	int t = ticks;
1826 
1827 	/*
1828 	 * The signed difference may be negative if the thread hasn't run for
1829 	 * over half of the ticks rollover period.
1830 	 */
1831 	if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) {
1832 		ts->ts_ticks = 0;
1833 		ts->ts_ftick = t - SCHED_TICK_TARG;
1834 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1835 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1836 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1837 		ts->ts_ftick = t - SCHED_TICK_TARG;
1838 	}
1839 	if (run)
1840 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1841 	ts->ts_ltick = t;
1842 }
1843 
1844 /*
1845  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1846  * if necessary.  This is the back-end for several priority related
1847  * functions.
1848  */
1849 static void
1850 sched_thread_priority(struct thread *td, u_char prio)
1851 {
1852 	struct tdq *tdq;
1853 	int oldpri;
1854 
1855 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1856 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1857 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1858 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1859 	if (td != curthread && prio < td->td_priority) {
1860 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1861 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1862 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1863 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1864 		    curthread);
1865 	}
1866 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1867 	if (td->td_priority == prio)
1868 		return;
1869 	/*
1870 	 * If the priority has been elevated due to priority
1871 	 * propagation, we may have to move ourselves to a new
1872 	 * queue.  This could be optimized to not re-add in some
1873 	 * cases.
1874 	 */
1875 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1876 		sched_rem(td);
1877 		td->td_priority = prio;
1878 		sched_add(td, SRQ_BORROWING | SRQ_HOLDTD);
1879 		return;
1880 	}
1881 	/*
1882 	 * If the thread is currently running we may have to adjust the lowpri
1883 	 * information so other cpus are aware of our current priority.
1884 	 */
1885 	if (TD_IS_RUNNING(td)) {
1886 		tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
1887 		oldpri = td->td_priority;
1888 		td->td_priority = prio;
1889 		if (prio < tdq->tdq_lowpri)
1890 			tdq->tdq_lowpri = prio;
1891 		else if (tdq->tdq_lowpri == oldpri)
1892 			tdq_setlowpri(tdq, td);
1893 		return;
1894 	}
1895 	td->td_priority = prio;
1896 }
1897 
1898 /*
1899  * Update a thread's priority when it is lent another thread's
1900  * priority.
1901  */
1902 void
1903 sched_lend_prio(struct thread *td, u_char prio)
1904 {
1905 
1906 	td->td_flags |= TDF_BORROWING;
1907 	sched_thread_priority(td, prio);
1908 }
1909 
1910 /*
1911  * Restore a thread's priority when priority propagation is
1912  * over.  The prio argument is the minimum priority the thread
1913  * needs to have to satisfy other possible priority lending
1914  * requests.  If the thread's regular priority is less
1915  * important than prio, the thread will keep a priority boost
1916  * of prio.
1917  */
1918 void
1919 sched_unlend_prio(struct thread *td, u_char prio)
1920 {
1921 	u_char base_pri;
1922 
1923 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1924 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1925 		base_pri = td->td_user_pri;
1926 	else
1927 		base_pri = td->td_base_pri;
1928 	if (prio >= base_pri) {
1929 		td->td_flags &= ~TDF_BORROWING;
1930 		sched_thread_priority(td, base_pri);
1931 	} else
1932 		sched_lend_prio(td, prio);
1933 }
1934 
1935 /*
1936  * Standard entry for setting the priority to an absolute value.
1937  */
1938 void
1939 sched_prio(struct thread *td, u_char prio)
1940 {
1941 	u_char oldprio;
1942 
1943 	/* First, update the base priority. */
1944 	td->td_base_pri = prio;
1945 
1946 	/*
1947 	 * If the thread is borrowing another thread's priority, don't
1948 	 * ever lower the priority.
1949 	 */
1950 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1951 		return;
1952 
1953 	/* Change the real priority. */
1954 	oldprio = td->td_priority;
1955 	sched_thread_priority(td, prio);
1956 
1957 	/*
1958 	 * If the thread is on a turnstile, then let the turnstile update
1959 	 * its state.
1960 	 */
1961 	if (TD_ON_LOCK(td) && oldprio != prio)
1962 		turnstile_adjust(td, oldprio);
1963 }
1964 
1965 /*
1966  * Set the base interrupt thread priority.
1967  */
1968 void
1969 sched_ithread_prio(struct thread *td, u_char prio)
1970 {
1971 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1972 	MPASS(td->td_pri_class == PRI_ITHD);
1973 	td->td_base_ithread_pri = prio;
1974 	sched_prio(td, prio);
1975 }
1976 
1977 /*
1978  * Set the base user priority, does not effect current running priority.
1979  */
1980 void
1981 sched_user_prio(struct thread *td, u_char prio)
1982 {
1983 
1984 	td->td_base_user_pri = prio;
1985 	if (td->td_lend_user_pri <= prio)
1986 		return;
1987 	td->td_user_pri = prio;
1988 }
1989 
1990 void
1991 sched_lend_user_prio(struct thread *td, u_char prio)
1992 {
1993 
1994 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1995 	td->td_lend_user_pri = prio;
1996 	td->td_user_pri = min(prio, td->td_base_user_pri);
1997 	if (td->td_priority > td->td_user_pri)
1998 		sched_prio(td, td->td_user_pri);
1999 	else if (td->td_priority != td->td_user_pri)
2000 		ast_sched_locked(td, TDA_SCHED);
2001 }
2002 
2003 /*
2004  * Like the above but first check if there is anything to do.
2005  */
2006 void
2007 sched_lend_user_prio_cond(struct thread *td, u_char prio)
2008 {
2009 
2010 	if (td->td_lend_user_pri != prio)
2011 		goto lend;
2012 	if (td->td_user_pri != min(prio, td->td_base_user_pri))
2013 		goto lend;
2014 	if (td->td_priority != td->td_user_pri)
2015 		goto lend;
2016 	return;
2017 
2018 lend:
2019 	thread_lock(td);
2020 	sched_lend_user_prio(td, prio);
2021 	thread_unlock(td);
2022 }
2023 
2024 #ifdef SMP
2025 /*
2026  * This tdq is about to idle.  Try to steal a thread from another CPU before
2027  * choosing the idle thread.
2028  */
2029 static void
2030 tdq_trysteal(struct tdq *tdq)
2031 {
2032 	struct cpu_group *cg, *parent;
2033 	struct tdq *steal;
2034 	cpuset_t mask;
2035 	int cpu, i, goup;
2036 
2037 	if (smp_started == 0 || steal_idle == 0 || trysteal_limit == 0 ||
2038 	    tdq->tdq_cg == NULL)
2039 		return;
2040 	CPU_FILL(&mask);
2041 	CPU_CLR(PCPU_GET(cpuid), &mask);
2042 	/* We don't want to be preempted while we're iterating. */
2043 	spinlock_enter();
2044 	TDQ_UNLOCK(tdq);
2045 	for (i = 1, cg = tdq->tdq_cg, goup = 0; ; ) {
2046 		cpu = sched_highest(cg, &mask, steal_thresh, 1);
2047 		/*
2048 		 * If a thread was added while interrupts were disabled don't
2049 		 * steal one here.
2050 		 */
2051 		if (TDQ_LOAD(tdq) > 0) {
2052 			TDQ_LOCK(tdq);
2053 			break;
2054 		}
2055 
2056 		/*
2057 		 * We found no CPU to steal from in this group.  Escalate to
2058 		 * the parent and repeat.  But if parent has only two children
2059 		 * groups we can avoid searching this group again by searching
2060 		 * the other one specifically and then escalating two levels.
2061 		 */
2062 		if (cpu == -1) {
2063 			if (goup) {
2064 				cg = cg->cg_parent;
2065 				goup = 0;
2066 			}
2067 			if (++i > trysteal_limit) {
2068 				TDQ_LOCK(tdq);
2069 				break;
2070 			}
2071 			parent = cg->cg_parent;
2072 			if (parent == NULL) {
2073 				TDQ_LOCK(tdq);
2074 				break;
2075 			}
2076 			if (parent->cg_children == 2) {
2077 				if (cg == &parent->cg_child[0])
2078 					cg = &parent->cg_child[1];
2079 				else
2080 					cg = &parent->cg_child[0];
2081 				goup = 1;
2082 			} else
2083 				cg = parent;
2084 			continue;
2085 		}
2086 		steal = TDQ_CPU(cpu);
2087 		/*
2088 		 * The data returned by sched_highest() is stale and
2089 		 * the chosen CPU no longer has an eligible thread.
2090 		 * At this point unconditionally exit the loop to bound
2091 		 * the time spent in the critcal section.
2092 		 */
2093 		if (TDQ_LOAD(steal) < steal_thresh ||
2094 		    TDQ_TRANSFERABLE(steal) == 0)
2095 			continue;
2096 		/*
2097 		 * Try to lock both queues. If we are assigned a thread while
2098 		 * waited for the lock, switch to it now instead of stealing.
2099 		 * If we can't get the lock, then somebody likely got there
2100 		 * first.
2101 		 */
2102 		TDQ_LOCK(tdq);
2103 		if (tdq->tdq_load > 0)
2104 			break;
2105 		if (TDQ_TRYLOCK_FLAGS(steal, MTX_DUPOK) == 0)
2106 			break;
2107 		/*
2108 		 * The data returned by sched_highest() is stale and
2109                  * the chosen CPU no longer has an eligible thread.
2110 		 */
2111 		if (TDQ_LOAD(steal) < steal_thresh ||
2112 		    TDQ_TRANSFERABLE(steal) == 0) {
2113 			TDQ_UNLOCK(steal);
2114 			break;
2115 		}
2116 		/*
2117 		 * If we fail to acquire one due to affinity restrictions,
2118 		 * bail out and let the idle thread to a more complete search
2119 		 * outside of a critical section.
2120 		 */
2121 		if (tdq_move(steal, tdq) == -1) {
2122 			TDQ_UNLOCK(steal);
2123 			break;
2124 		}
2125 		TDQ_UNLOCK(steal);
2126 		break;
2127 	}
2128 	spinlock_exit();
2129 }
2130 #endif
2131 
2132 /*
2133  * Handle migration from sched_switch().  This happens only for
2134  * cpu binding.
2135  */
2136 static struct mtx *
2137 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
2138 {
2139 	struct tdq *tdn;
2140 #ifdef SMP
2141 	int lowpri;
2142 #endif
2143 
2144 	KASSERT(THREAD_CAN_MIGRATE(td) ||
2145 	    (td_get_sched(td)->ts_flags & TSF_BOUND) != 0,
2146 	    ("Thread %p shouldn't migrate", td));
2147 	KASSERT(!CPU_ABSENT(td_get_sched(td)->ts_cpu), ("sched_switch_migrate: "
2148 	    "thread %s queued on absent CPU %d.", td->td_name,
2149 	    td_get_sched(td)->ts_cpu));
2150 	tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
2151 #ifdef SMP
2152 	tdq_load_rem(tdq, td);
2153 	/*
2154 	 * Do the lock dance required to avoid LOR.  We have an
2155 	 * extra spinlock nesting from sched_switch() which will
2156 	 * prevent preemption while we're holding neither run-queue lock.
2157 	 */
2158 	TDQ_UNLOCK(tdq);
2159 	TDQ_LOCK(tdn);
2160 	lowpri = tdq_add(tdn, td, flags);
2161 	tdq_notify(tdn, lowpri);
2162 	TDQ_UNLOCK(tdn);
2163 	TDQ_LOCK(tdq);
2164 #endif
2165 	return (TDQ_LOCKPTR(tdn));
2166 }
2167 
2168 /*
2169  * thread_lock_unblock() that does not assume td_lock is blocked.
2170  */
2171 static inline void
2172 thread_unblock_switch(struct thread *td, struct mtx *mtx)
2173 {
2174 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
2175 	    (uintptr_t)mtx);
2176 }
2177 
2178 /*
2179  * Switch threads.  This function has to handle threads coming in while
2180  * blocked for some reason, running, or idle.  It also must deal with
2181  * migrating a thread from one queue to another as running threads may
2182  * be assigned elsewhere via binding.
2183  */
2184 void
2185 sched_switch(struct thread *td, int flags)
2186 {
2187 	struct thread *newtd;
2188 	struct tdq *tdq;
2189 	struct td_sched *ts;
2190 	struct mtx *mtx;
2191 	int srqflag;
2192 	int cpuid, preempted;
2193 #ifdef SMP
2194 	int pickcpu;
2195 #endif
2196 
2197 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2198 
2199 	cpuid = PCPU_GET(cpuid);
2200 	tdq = TDQ_SELF();
2201 	ts = td_get_sched(td);
2202 	sched_pctcpu_update(ts, 1);
2203 #ifdef SMP
2204 	pickcpu = (td->td_flags & TDF_PICKCPU) != 0;
2205 	if (pickcpu)
2206 		ts->ts_rltick = ticks - affinity * MAX_CACHE_LEVELS;
2207 	else
2208 		ts->ts_rltick = ticks;
2209 #endif
2210 	td->td_lastcpu = td->td_oncpu;
2211 	preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
2212 	    (flags & SW_PREEMPT) != 0;
2213 	td->td_flags &= ~(TDF_PICKCPU | TDF_SLICEEND);
2214 	ast_unsched_locked(td, TDA_SCHED);
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 			ast_sched_locked(td, TDA_SCHED);
2649 			td->td_flags |= TDF_SLICEEND;
2650 		}
2651 	}
2652 }
2653 
2654 u_int
2655 sched_estcpu(struct thread *td __unused)
2656 {
2657 
2658 	return (0);
2659 }
2660 
2661 /*
2662  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2663  * cooperative idle threads.
2664  */
2665 int
2666 sched_runnable(void)
2667 {
2668 	struct tdq *tdq;
2669 	int load;
2670 
2671 	load = 1;
2672 
2673 	tdq = TDQ_SELF();
2674 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2675 		if (TDQ_LOAD(tdq) > 0)
2676 			goto out;
2677 	} else
2678 		if (TDQ_LOAD(tdq) - 1 > 0)
2679 			goto out;
2680 	load = 0;
2681 out:
2682 	return (load);
2683 }
2684 
2685 /*
2686  * Choose the highest priority thread to run.  The thread is removed from
2687  * the run-queue while running however the load remains.
2688  */
2689 struct thread *
2690 sched_choose(void)
2691 {
2692 	struct thread *td;
2693 	struct tdq *tdq;
2694 
2695 	tdq = TDQ_SELF();
2696 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2697 	td = tdq_choose(tdq);
2698 	if (td != NULL) {
2699 		tdq_runq_rem(tdq, td);
2700 		tdq->tdq_lowpri = td->td_priority;
2701 	} else {
2702 		tdq->tdq_lowpri = PRI_MAX_IDLE;
2703 		td = PCPU_GET(idlethread);
2704 	}
2705 	tdq->tdq_curthread = td;
2706 	return (td);
2707 }
2708 
2709 /*
2710  * Set owepreempt if the currently running thread has lower priority than "pri".
2711  * Preemption never happens directly in ULE, we always request it once we exit a
2712  * critical section.
2713  */
2714 static void
2715 sched_setpreempt(int pri)
2716 {
2717 	struct thread *ctd;
2718 	int cpri;
2719 
2720 	ctd = curthread;
2721 	THREAD_LOCK_ASSERT(ctd, MA_OWNED);
2722 
2723 	cpri = ctd->td_priority;
2724 	if (pri < cpri)
2725 		ast_sched_locked(ctd, TDA_SCHED);
2726 	if (KERNEL_PANICKED() || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2727 		return;
2728 	if (!sched_shouldpreempt(pri, cpri, 0))
2729 		return;
2730 	ctd->td_owepreempt = 1;
2731 }
2732 
2733 /*
2734  * Add a thread to a thread queue.  Select the appropriate runq and add the
2735  * thread to it.  This is the internal function called when the tdq is
2736  * predetermined.
2737  */
2738 static int
2739 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2740 {
2741 	int lowpri;
2742 
2743 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2744 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
2745 	KASSERT((td->td_inhibitors == 0),
2746 	    ("sched_add: trying to run inhibited thread"));
2747 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2748 	    ("sched_add: bad thread state"));
2749 	KASSERT(td->td_flags & TDF_INMEM,
2750 	    ("sched_add: thread swapped out"));
2751 
2752 	lowpri = tdq->tdq_lowpri;
2753 	if (td->td_priority < lowpri)
2754 		tdq->tdq_lowpri = td->td_priority;
2755 	tdq_runq_add(tdq, td, flags);
2756 	tdq_load_add(tdq, td);
2757 	return (lowpri);
2758 }
2759 
2760 /*
2761  * Select the target thread queue and add a thread to it.  Request
2762  * preemption or IPI a remote processor if required.
2763  *
2764  * Requires the thread lock on entry, drops on exit.
2765  */
2766 void
2767 sched_add(struct thread *td, int flags)
2768 {
2769 	struct tdq *tdq;
2770 #ifdef SMP
2771 	int cpu, lowpri;
2772 #endif
2773 
2774 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2775 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2776 	    sched_tdname(curthread));
2777 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2778 	    KTR_ATTR_LINKED, sched_tdname(td));
2779 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2780 	    flags & SRQ_PREEMPTED);
2781 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2782 	/*
2783 	 * Recalculate the priority before we select the target cpu or
2784 	 * run-queue.
2785 	 */
2786 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2787 		sched_priority(td);
2788 #ifdef SMP
2789 	/*
2790 	 * Pick the destination cpu and if it isn't ours transfer to the
2791 	 * target cpu.
2792 	 */
2793 	cpu = sched_pickcpu(td, flags);
2794 	tdq = sched_setcpu(td, cpu, flags);
2795 	lowpri = tdq_add(tdq, td, flags);
2796 	if (cpu != PCPU_GET(cpuid))
2797 		tdq_notify(tdq, lowpri);
2798 	else if (!(flags & SRQ_YIELDING))
2799 		sched_setpreempt(td->td_priority);
2800 #else
2801 	tdq = TDQ_SELF();
2802 	/*
2803 	 * Now that the thread is moving to the run-queue, set the lock
2804 	 * to the scheduler's lock.
2805 	 */
2806 	if (td->td_lock != TDQ_LOCKPTR(tdq)) {
2807 		TDQ_LOCK(tdq);
2808 		if ((flags & SRQ_HOLD) != 0)
2809 			td->td_lock = TDQ_LOCKPTR(tdq);
2810 		else
2811 			thread_lock_set(td, TDQ_LOCKPTR(tdq));
2812 	}
2813 	(void)tdq_add(tdq, td, flags);
2814 	if (!(flags & SRQ_YIELDING))
2815 		sched_setpreempt(td->td_priority);
2816 #endif
2817 	if (!(flags & SRQ_HOLDTD))
2818 		thread_unlock(td);
2819 }
2820 
2821 /*
2822  * Remove a thread from a run-queue without running it.  This is used
2823  * when we're stealing a thread from a remote queue.  Otherwise all threads
2824  * exit by calling sched_exit_thread() and sched_throw() themselves.
2825  */
2826 void
2827 sched_rem(struct thread *td)
2828 {
2829 	struct tdq *tdq;
2830 
2831 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2832 	    "prio:%d", td->td_priority);
2833 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2834 	tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2835 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2836 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2837 	KASSERT(TD_ON_RUNQ(td),
2838 	    ("sched_rem: thread not on run queue"));
2839 	tdq_runq_rem(tdq, td);
2840 	tdq_load_rem(tdq, td);
2841 	TD_SET_CAN_RUN(td);
2842 	if (td->td_priority == tdq->tdq_lowpri)
2843 		tdq_setlowpri(tdq, NULL);
2844 }
2845 
2846 /*
2847  * Fetch cpu utilization information.  Updates on demand.
2848  */
2849 fixpt_t
2850 sched_pctcpu(struct thread *td)
2851 {
2852 	fixpt_t pctcpu;
2853 	struct td_sched *ts;
2854 
2855 	pctcpu = 0;
2856 	ts = td_get_sched(td);
2857 
2858 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2859 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2860 	if (ts->ts_ticks) {
2861 		int rtick;
2862 
2863 		/* How many rtick per second ? */
2864 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2865 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2866 	}
2867 
2868 	return (pctcpu);
2869 }
2870 
2871 /*
2872  * Enforce affinity settings for a thread.  Called after adjustments to
2873  * cpumask.
2874  */
2875 void
2876 sched_affinity(struct thread *td)
2877 {
2878 #ifdef SMP
2879 	struct td_sched *ts;
2880 
2881 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2882 	ts = td_get_sched(td);
2883 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2884 		return;
2885 	if (TD_ON_RUNQ(td)) {
2886 		sched_rem(td);
2887 		sched_add(td, SRQ_BORING | SRQ_HOLDTD);
2888 		return;
2889 	}
2890 	if (!TD_IS_RUNNING(td))
2891 		return;
2892 	/*
2893 	 * Force a switch before returning to userspace.  If the
2894 	 * target thread is not running locally send an ipi to force
2895 	 * the issue.
2896 	 */
2897 	ast_sched_locked(td, TDA_SCHED);
2898 	if (td != curthread)
2899 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2900 #endif
2901 }
2902 
2903 /*
2904  * Bind a thread to a target cpu.
2905  */
2906 void
2907 sched_bind(struct thread *td, int cpu)
2908 {
2909 	struct td_sched *ts;
2910 
2911 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2912 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2913 	ts = td_get_sched(td);
2914 	if (ts->ts_flags & TSF_BOUND)
2915 		sched_unbind(td);
2916 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2917 	ts->ts_flags |= TSF_BOUND;
2918 	sched_pin();
2919 	if (PCPU_GET(cpuid) == cpu)
2920 		return;
2921 	ts->ts_cpu = cpu;
2922 	/* When we return from mi_switch we'll be on the correct cpu. */
2923 	mi_switch(SW_VOL | SWT_BIND);
2924 	thread_lock(td);
2925 }
2926 
2927 /*
2928  * Release a bound thread.
2929  */
2930 void
2931 sched_unbind(struct thread *td)
2932 {
2933 	struct td_sched *ts;
2934 
2935 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2936 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2937 	ts = td_get_sched(td);
2938 	if ((ts->ts_flags & TSF_BOUND) == 0)
2939 		return;
2940 	ts->ts_flags &= ~TSF_BOUND;
2941 	sched_unpin();
2942 }
2943 
2944 int
2945 sched_is_bound(struct thread *td)
2946 {
2947 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2948 	return (td_get_sched(td)->ts_flags & TSF_BOUND);
2949 }
2950 
2951 /*
2952  * Basic yield call.
2953  */
2954 void
2955 sched_relinquish(struct thread *td)
2956 {
2957 	thread_lock(td);
2958 	mi_switch(SW_VOL | SWT_RELINQUISH);
2959 }
2960 
2961 /*
2962  * Return the total system load.
2963  */
2964 int
2965 sched_load(void)
2966 {
2967 #ifdef SMP
2968 	int total;
2969 	int i;
2970 
2971 	total = 0;
2972 	CPU_FOREACH(i)
2973 		total += atomic_load_int(&TDQ_CPU(i)->tdq_sysload);
2974 	return (total);
2975 #else
2976 	return (atomic_load_int(&TDQ_SELF()->tdq_sysload));
2977 #endif
2978 }
2979 
2980 int
2981 sched_sizeof_proc(void)
2982 {
2983 	return (sizeof(struct proc));
2984 }
2985 
2986 int
2987 sched_sizeof_thread(void)
2988 {
2989 	return (sizeof(struct thread) + sizeof(struct td_sched));
2990 }
2991 
2992 #ifdef SMP
2993 #define	TDQ_IDLESPIN(tdq)						\
2994     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2995 #else
2996 #define	TDQ_IDLESPIN(tdq)	1
2997 #endif
2998 
2999 /*
3000  * The actual idle process.
3001  */
3002 void
3003 sched_idletd(void *dummy)
3004 {
3005 	struct thread *td;
3006 	struct tdq *tdq;
3007 	int oldswitchcnt, switchcnt;
3008 	int i;
3009 
3010 	mtx_assert(&Giant, MA_NOTOWNED);
3011 	td = curthread;
3012 	tdq = TDQ_SELF();
3013 	THREAD_NO_SLEEPING();
3014 	oldswitchcnt = -1;
3015 	for (;;) {
3016 		if (TDQ_LOAD(tdq)) {
3017 			thread_lock(td);
3018 			mi_switch(SW_VOL | SWT_IDLE);
3019 		}
3020 		switchcnt = TDQ_SWITCHCNT(tdq);
3021 #ifdef SMP
3022 		if (always_steal || switchcnt != oldswitchcnt) {
3023 			oldswitchcnt = switchcnt;
3024 			if (tdq_idled(tdq) == 0)
3025 				continue;
3026 		}
3027 		switchcnt = TDQ_SWITCHCNT(tdq);
3028 #else
3029 		oldswitchcnt = switchcnt;
3030 #endif
3031 		/*
3032 		 * If we're switching very frequently, spin while checking
3033 		 * for load rather than entering a low power state that
3034 		 * may require an IPI.  However, don't do any busy
3035 		 * loops while on SMT machines as this simply steals
3036 		 * cycles from cores doing useful work.
3037 		 */
3038 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
3039 			for (i = 0; i < sched_idlespins; i++) {
3040 				if (TDQ_LOAD(tdq))
3041 					break;
3042 				cpu_spinwait();
3043 			}
3044 		}
3045 
3046 		/* If there was context switch during spin, restart it. */
3047 		switchcnt = TDQ_SWITCHCNT(tdq);
3048 		if (TDQ_LOAD(tdq) != 0 || switchcnt != oldswitchcnt)
3049 			continue;
3050 
3051 		/* Run main MD idle handler. */
3052 		atomic_store_int(&tdq->tdq_cpu_idle, 1);
3053 		/*
3054 		 * Make sure that the tdq_cpu_idle update is globally visible
3055 		 * before cpu_idle() reads tdq_load.  The order is important
3056 		 * to avoid races with tdq_notify().
3057 		 */
3058 		atomic_thread_fence_seq_cst();
3059 		/*
3060 		 * Checking for again after the fence picks up assigned
3061 		 * threads often enough to make it worthwhile to do so in
3062 		 * order to avoid calling cpu_idle().
3063 		 */
3064 		if (TDQ_LOAD(tdq) != 0) {
3065 			atomic_store_int(&tdq->tdq_cpu_idle, 0);
3066 			continue;
3067 		}
3068 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
3069 		atomic_store_int(&tdq->tdq_cpu_idle, 0);
3070 
3071 		/*
3072 		 * Account thread-less hardware interrupts and
3073 		 * other wakeup reasons equal to context switches.
3074 		 */
3075 		switchcnt = TDQ_SWITCHCNT(tdq);
3076 		if (switchcnt != oldswitchcnt)
3077 			continue;
3078 		TDQ_SWITCHCNT_INC(tdq);
3079 		oldswitchcnt++;
3080 	}
3081 }
3082 
3083 /*
3084  * sched_throw_grab() chooses a thread from the queue to switch to
3085  * next.  It returns with the tdq lock dropped in a spinlock section to
3086  * keep interrupts disabled until the CPU is running in a proper threaded
3087  * context.
3088  */
3089 static struct thread *
3090 sched_throw_grab(struct tdq *tdq)
3091 {
3092 	struct thread *newtd;
3093 
3094 	newtd = choosethread();
3095 	spinlock_enter();
3096 	TDQ_UNLOCK(tdq);
3097 	KASSERT(curthread->td_md.md_spinlock_count == 1,
3098 	    ("invalid count %d", curthread->td_md.md_spinlock_count));
3099 	return (newtd);
3100 }
3101 
3102 /*
3103  * A CPU is entering for the first time.
3104  */
3105 void
3106 sched_ap_entry(void)
3107 {
3108 	struct thread *newtd;
3109 	struct tdq *tdq;
3110 
3111 	tdq = TDQ_SELF();
3112 
3113 	/* This should have been setup in schedinit_ap(). */
3114 	THREAD_LOCKPTR_ASSERT(curthread, TDQ_LOCKPTR(tdq));
3115 
3116 	TDQ_LOCK(tdq);
3117 	/* Correct spinlock nesting. */
3118 	spinlock_exit();
3119 	PCPU_SET(switchtime, cpu_ticks());
3120 	PCPU_SET(switchticks, ticks);
3121 
3122 	newtd = sched_throw_grab(tdq);
3123 
3124 	/* doesn't return */
3125 	cpu_throw(NULL, newtd);
3126 }
3127 
3128 /*
3129  * A thread is exiting.
3130  */
3131 void
3132 sched_throw(struct thread *td)
3133 {
3134 	struct thread *newtd;
3135 	struct tdq *tdq;
3136 
3137 	tdq = TDQ_SELF();
3138 
3139 	MPASS(td != NULL);
3140 	THREAD_LOCK_ASSERT(td, MA_OWNED);
3141 	THREAD_LOCKPTR_ASSERT(td, TDQ_LOCKPTR(tdq));
3142 
3143 	tdq_load_rem(tdq, td);
3144 	td->td_lastcpu = td->td_oncpu;
3145 	td->td_oncpu = NOCPU;
3146 	thread_lock_block(td);
3147 
3148 	newtd = sched_throw_grab(tdq);
3149 
3150 	/* doesn't return */
3151 	cpu_switch(td, newtd, TDQ_LOCKPTR(tdq));
3152 }
3153 
3154 /*
3155  * This is called from fork_exit().  Just acquire the correct locks and
3156  * let fork do the rest of the work.
3157  */
3158 void
3159 sched_fork_exit(struct thread *td)
3160 {
3161 	struct tdq *tdq;
3162 	int cpuid;
3163 
3164 	/*
3165 	 * Finish setting up thread glue so that it begins execution in a
3166 	 * non-nested critical section with the scheduler lock held.
3167 	 */
3168 	KASSERT(curthread->td_md.md_spinlock_count == 1,
3169 	    ("invalid count %d", curthread->td_md.md_spinlock_count));
3170 	cpuid = PCPU_GET(cpuid);
3171 	tdq = TDQ_SELF();
3172 	TDQ_LOCK(tdq);
3173 	spinlock_exit();
3174 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
3175 	td->td_oncpu = cpuid;
3176 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
3177 	    "prio:%d", td->td_priority);
3178 	SDT_PROBE0(sched, , , on__cpu);
3179 }
3180 
3181 /*
3182  * Create on first use to catch odd startup conditions.
3183  */
3184 char *
3185 sched_tdname(struct thread *td)
3186 {
3187 #ifdef KTR
3188 	struct td_sched *ts;
3189 
3190 	ts = td_get_sched(td);
3191 	if (ts->ts_name[0] == '\0')
3192 		snprintf(ts->ts_name, sizeof(ts->ts_name),
3193 		    "%s tid %d", td->td_name, td->td_tid);
3194 	return (ts->ts_name);
3195 #else
3196 	return (td->td_name);
3197 #endif
3198 }
3199 
3200 #ifdef KTR
3201 void
3202 sched_clear_tdname(struct thread *td)
3203 {
3204 	struct td_sched *ts;
3205 
3206 	ts = td_get_sched(td);
3207 	ts->ts_name[0] = '\0';
3208 }
3209 #endif
3210 
3211 #ifdef SMP
3212 
3213 /*
3214  * Build the CPU topology dump string. Is recursively called to collect
3215  * the topology tree.
3216  */
3217 static int
3218 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
3219     int indent)
3220 {
3221 	char cpusetbuf[CPUSETBUFSIZ];
3222 	int i, first;
3223 
3224 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
3225 	    "", 1 + indent / 2, cg->cg_level);
3226 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
3227 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
3228 	first = TRUE;
3229 	for (i = cg->cg_first; i <= cg->cg_last; i++) {
3230 		if (CPU_ISSET(i, &cg->cg_mask)) {
3231 			if (!first)
3232 				sbuf_cat(sb, ", ");
3233 			else
3234 				first = FALSE;
3235 			sbuf_printf(sb, "%d", i);
3236 		}
3237 	}
3238 	sbuf_cat(sb, "</cpu>\n");
3239 
3240 	if (cg->cg_flags != 0) {
3241 		sbuf_printf(sb, "%*s <flags>", indent, "");
3242 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
3243 			sbuf_cat(sb, "<flag name=\"HTT\">HTT group</flag>");
3244 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
3245 			sbuf_cat(sb, "<flag name=\"THREAD\">THREAD group</flag>");
3246 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
3247 			sbuf_cat(sb, "<flag name=\"SMT\">SMT group</flag>");
3248 		if ((cg->cg_flags & CG_FLAG_NODE) != 0)
3249 			sbuf_cat(sb, "<flag name=\"NODE\">NUMA node</flag>");
3250 		sbuf_cat(sb, "</flags>\n");
3251 	}
3252 
3253 	if (cg->cg_children > 0) {
3254 		sbuf_printf(sb, "%*s <children>\n", indent, "");
3255 		for (i = 0; i < cg->cg_children; i++)
3256 			sysctl_kern_sched_topology_spec_internal(sb,
3257 			    &cg->cg_child[i], indent+2);
3258 		sbuf_printf(sb, "%*s </children>\n", indent, "");
3259 	}
3260 	sbuf_printf(sb, "%*s</group>\n", indent, "");
3261 	return (0);
3262 }
3263 
3264 /*
3265  * Sysctl handler for retrieving topology dump. It's a wrapper for
3266  * the recursive sysctl_kern_smp_topology_spec_internal().
3267  */
3268 static int
3269 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
3270 {
3271 	struct sbuf *topo;
3272 	int err;
3273 
3274 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
3275 
3276 	topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
3277 	if (topo == NULL)
3278 		return (ENOMEM);
3279 
3280 	sbuf_cat(topo, "<groups>\n");
3281 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
3282 	sbuf_cat(topo, "</groups>\n");
3283 
3284 	if (err == 0) {
3285 		err = sbuf_finish(topo);
3286 	}
3287 	sbuf_delete(topo);
3288 	return (err);
3289 }
3290 
3291 #endif
3292 
3293 static int
3294 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
3295 {
3296 	int error, new_val, period;
3297 
3298 	period = 1000000 / realstathz;
3299 	new_val = period * sched_slice;
3300 	error = sysctl_handle_int(oidp, &new_val, 0, req);
3301 	if (error != 0 || req->newptr == NULL)
3302 		return (error);
3303 	if (new_val <= 0)
3304 		return (EINVAL);
3305 	sched_slice = imax(1, (new_val + period / 2) / period);
3306 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
3307 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
3308 	    realstathz);
3309 	return (0);
3310 }
3311 
3312 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
3313     "Scheduler");
3314 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
3315     "Scheduler name");
3316 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum,
3317     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
3318     sysctl_kern_quantum, "I",
3319     "Quantum for timeshare threads in microseconds");
3320 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
3321     "Quantum for timeshare threads in stathz ticks");
3322 SYSCTL_UINT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
3323     "Interactivity score threshold");
3324 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
3325     &preempt_thresh, 0,
3326     "Maximal (lowest) priority for preemption");
3327 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
3328     "Assign static kernel priorities to sleeping threads");
3329 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
3330     "Number of times idle thread will spin waiting for new work");
3331 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
3332     &sched_idlespinthresh, 0,
3333     "Threshold before we will permit idle thread spinning");
3334 #ifdef SMP
3335 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
3336     "Number of hz ticks to keep thread affinity for");
3337 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
3338     "Enables the long-term load balancer");
3339 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
3340     &balance_interval, 0,
3341     "Average period in stathz ticks to run the long-term balancer");
3342 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
3343     "Attempts to steal work from other cores before idling");
3344 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
3345     "Minimum load on remote CPU before we'll steal");
3346 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit,
3347     0, "Topological distance limit for stealing threads in sched_switch()");
3348 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0,
3349     "Always run the stealer from the idle thread");
3350 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
3351     CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
3352     "XML dump of detected CPU topology");
3353 #endif
3354 
3355 /* ps compat.  All cpu percentages from ULE are weighted. */
3356 static int ccpu = 0;
3357 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0,
3358     "Decay factor used for updating %CPU in 4BSD scheduler");
3359