xref: /freebsd/sys/kern/sched_ule.c (revision 11484ad8a2b01049b3e4f25c0fae6041c2060629)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice unmodified, this list of conditions, and the following
12  *    disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 /*
30  * This file implements the ULE scheduler.  ULE supports independent CPU
31  * run queues and fine grain locking.  It has superior interactive
32  * performance under load even on uni-processor systems.
33  *
34  * etymology:
35  *   ULE is the last three letters in schedule.  It owes its name to a
36  * generic user created for a scheduling system by Paul Mikesell at
37  * Isilon Systems and a general lack of creativity on the part of the author.
38  */
39 
40 #include <sys/cdefs.h>
41 __FBSDID("$FreeBSD$");
42 
43 #include "opt_hwpmc_hooks.h"
44 #include "opt_sched.h"
45 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kdb.h>
49 #include <sys/kernel.h>
50 #include <sys/ktr.h>
51 #include <sys/limits.h>
52 #include <sys/lock.h>
53 #include <sys/mutex.h>
54 #include <sys/proc.h>
55 #include <sys/resource.h>
56 #include <sys/resourcevar.h>
57 #include <sys/sched.h>
58 #include <sys/sdt.h>
59 #include <sys/smp.h>
60 #include <sys/sx.h>
61 #include <sys/sysctl.h>
62 #include <sys/sysproto.h>
63 #include <sys/turnstile.h>
64 #include <sys/umtxvar.h>
65 #include <sys/vmmeter.h>
66 #include <sys/cpuset.h>
67 #include <sys/sbuf.h>
68 
69 #ifdef HWPMC_HOOKS
70 #include <sys/pmckern.h>
71 #endif
72 
73 #ifdef KDTRACE_HOOKS
74 #include <sys/dtrace_bsd.h>
75 int __read_mostly		dtrace_vtime_active;
76 dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
77 #endif
78 
79 #include <machine/cpu.h>
80 #include <machine/smp.h>
81 
82 #define	KTR_ULE	0
83 
84 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
85 #define	TDQ_NAME_LEN	(sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
86 #define	TDQ_LOADNAME_LEN	(sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
87 
88 /*
89  * Thread scheduler specific section.  All fields are protected
90  * by the thread lock.
91  */
92 struct td_sched {
93 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
94 	short		ts_flags;	/* TSF_* flags. */
95 	int		ts_cpu;		/* CPU that we have affinity for. */
96 	int		ts_rltick;	/* Real last tick, for affinity. */
97 	int		ts_slice;	/* Ticks of slice remaining. */
98 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
99 	u_int		ts_runtime;	/* Number of ticks we were running */
100 	int		ts_ltick;	/* Last tick that we were running on */
101 	int		ts_ftick;	/* First tick that we were running on */
102 	int		ts_ticks;	/* Tick count */
103 #ifdef KTR
104 	char		ts_name[TS_NAME_LEN];
105 #endif
106 };
107 /* flags kept in ts_flags */
108 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
109 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
110 
111 #define	THREAD_CAN_MIGRATE(td)	((td)->td_pinned == 0)
112 #define	THREAD_CAN_SCHED(td, cpu)	\
113     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
114 
115 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <=
116     sizeof(struct thread0_storage),
117     "increase struct thread0_storage.t0st_sched size");
118 
119 /*
120  * Priority ranges used for interactive and non-interactive timeshare
121  * threads.  The timeshare priorities are split up into four ranges.
122  * The first range handles interactive threads.  The last three ranges
123  * (NHALF, x, and NHALF) handle non-interactive threads with the outer
124  * ranges supporting nice values.
125  */
126 #define	PRI_TIMESHARE_RANGE	(PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
127 #define	PRI_INTERACT_RANGE	((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
128 #define	PRI_BATCH_RANGE		(PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
129 
130 #define	PRI_MIN_INTERACT	PRI_MIN_TIMESHARE
131 #define	PRI_MAX_INTERACT	(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
132 #define	PRI_MIN_BATCH		(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
133 #define	PRI_MAX_BATCH		PRI_MAX_TIMESHARE
134 
135 /*
136  * Cpu percentage computation macros and defines.
137  *
138  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
139  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
140  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
141  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
142  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
143  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
144  */
145 #define	SCHED_TICK_SECS		10
146 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
147 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
148 #define	SCHED_TICK_SHIFT	10
149 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
150 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
151 
152 /*
153  * These macros determine priorities for non-interactive threads.  They are
154  * assigned a priority based on their recent cpu utilization as expressed
155  * by the ratio of ticks to the tick total.  NHALF priorities at the start
156  * and end of the MIN to MAX timeshare range are only reachable with negative
157  * or positive nice respectively.
158  *
159  * PRI_RANGE:	Priority range for utilization dependent priorities.
160  * PRI_NRESV:	Number of nice values.
161  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
162  * PRI_NICE:	Determines the part of the priority inherited from nice.
163  */
164 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
165 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
166 #define	SCHED_PRI_MIN		(PRI_MIN_BATCH + SCHED_PRI_NHALF)
167 #define	SCHED_PRI_MAX		(PRI_MAX_BATCH - SCHED_PRI_NHALF)
168 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
169 #define	SCHED_PRI_TICKS(ts)						\
170     (SCHED_TICK_HZ((ts)) /						\
171     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
172 #define	SCHED_PRI_NICE(nice)	(nice)
173 
174 /*
175  * These determine the interactivity of a process.  Interactivity differs from
176  * cpu utilization in that it expresses the voluntary time slept vs time ran
177  * while cpu utilization includes all time not running.  This more accurately
178  * models the intent of the thread.
179  *
180  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
181  *		before throttling back.
182  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
183  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
184  * INTERACT_THRESH:	Threshold for placement on the current runq.
185  */
186 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
187 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
188 #define	SCHED_INTERACT_MAX	(100)
189 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
190 #define	SCHED_INTERACT_THRESH	(30)
191 
192 /*
193  * These parameters determine the slice behavior for batch work.
194  */
195 #define	SCHED_SLICE_DEFAULT_DIVISOR	10	/* ~94 ms, 12 stathz ticks. */
196 #define	SCHED_SLICE_MIN_DIVISOR		6	/* DEFAULT/MIN = ~16 ms. */
197 
198 /* Flags kept in td_flags. */
199 #define	TDF_PICKCPU	TDF_SCHED0	/* Thread should pick new CPU. */
200 #define	TDF_SLICEEND	TDF_SCHED2	/* Thread time slice is over. */
201 
202 /*
203  * tickincr:		Converts a stathz tick into a hz domain scaled by
204  *			the shift factor.  Without the shift the error rate
205  *			due to rounding would be unacceptably high.
206  * realstathz:		stathz is sometimes 0 and run off of hz.
207  * sched_slice:		Runtime of each thread before rescheduling.
208  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
209  */
210 static u_int __read_mostly sched_interact = SCHED_INTERACT_THRESH;
211 static int __read_mostly tickincr = 8 << SCHED_TICK_SHIFT;
212 static int __read_mostly realstathz = 127;	/* reset during boot. */
213 static int __read_mostly sched_slice = 10;	/* reset during boot. */
214 static int __read_mostly sched_slice_min = 1;	/* reset during boot. */
215 #ifdef PREEMPTION
216 #ifdef FULL_PREEMPTION
217 static int __read_mostly preempt_thresh = PRI_MAX_IDLE;
218 #else
219 static int __read_mostly preempt_thresh = PRI_MIN_KERN;
220 #endif
221 #else
222 static int __read_mostly preempt_thresh = 0;
223 #endif
224 static int __read_mostly static_boost = PRI_MIN_BATCH;
225 static int __read_mostly sched_idlespins = 10000;
226 static int __read_mostly sched_idlespinthresh = -1;
227 
228 /*
229  * tdq - per processor runqs and statistics.  A mutex synchronizes access to
230  * most fields.  Some fields are loaded or modified without the mutex.
231  *
232  * Locking protocols:
233  * (c)  constant after initialization
234  * (f)  flag, set with the tdq lock held, cleared on local CPU
235  * (l)  all accesses are CPU-local
236  * (ls) stores are performed by the local CPU, loads may be lockless
237  * (t)  all accesses are protected by the tdq mutex
238  * (ts) stores are serialized by the tdq mutex, loads may be lockless
239  */
240 struct tdq {
241 	/*
242 	 * Ordered to improve efficiency of cpu_search() and switch().
243 	 * tdq_lock is padded to avoid false sharing with tdq_load and
244 	 * tdq_cpu_idle.
245 	 */
246 	struct mtx_padalign tdq_lock;	/* run queue lock. */
247 	struct cpu_group *tdq_cg;	/* (c) Pointer to cpu topology. */
248 	struct thread	*tdq_curthread;	/* (t) Current executing thread. */
249 	int		tdq_load;	/* (ts) Aggregate load. */
250 	int		tdq_sysload;	/* (ts) For loadavg, !ITHD load. */
251 	int		tdq_cpu_idle;	/* (ls) cpu_idle() is active. */
252 	int		tdq_transferable; /* (ts) Transferable thread count. */
253 	short		tdq_switchcnt;	/* (l) Switches this tick. */
254 	short		tdq_oldswitchcnt; /* (l) Switches last tick. */
255 	u_char		tdq_lowpri;	/* (ts) Lowest priority thread. */
256 	u_char		tdq_owepreempt;	/* (f) Remote preemption pending. */
257 	u_char		tdq_idx;	/* (t) Current insert index. */
258 	u_char		tdq_ridx;	/* (t) Current removal index. */
259 	int		tdq_id;		/* (c) cpuid. */
260 	struct runq	tdq_realtime;	/* (t) real-time run queue. */
261 	struct runq	tdq_timeshare;	/* (t) timeshare run queue. */
262 	struct runq	tdq_idle;	/* (t) Queue of IDLE threads. */
263 	char		tdq_name[TDQ_NAME_LEN];
264 #ifdef KTR
265 	char		tdq_loadname[TDQ_LOADNAME_LEN];
266 #endif
267 };
268 
269 /* Idle thread states and config. */
270 #define	TDQ_RUNNING	1
271 #define	TDQ_IDLE	2
272 
273 /* Lockless accessors. */
274 #define	TDQ_LOAD(tdq)		atomic_load_int(&(tdq)->tdq_load)
275 #define	TDQ_TRANSFERABLE(tdq)	atomic_load_int(&(tdq)->tdq_transferable)
276 #define	TDQ_SWITCHCNT(tdq)	(atomic_load_short(&(tdq)->tdq_switchcnt) + \
277 				 atomic_load_short(&(tdq)->tdq_oldswitchcnt))
278 #define	TDQ_SWITCHCNT_INC(tdq)	(atomic_store_short(&(tdq)->tdq_switchcnt, \
279 				 atomic_load_short(&(tdq)->tdq_switchcnt) + 1))
280 
281 #ifdef SMP
282 struct cpu_group __read_mostly *cpu_top;		/* CPU topology */
283 
284 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
285 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
286 
287 /*
288  * Run-time tunables.
289  */
290 static int rebalance = 1;
291 static int balance_interval = 128;	/* Default set in sched_initticks(). */
292 static int __read_mostly affinity;
293 static int __read_mostly steal_idle = 1;
294 static int __read_mostly steal_thresh = 2;
295 static int __read_mostly always_steal = 0;
296 static int __read_mostly trysteal_limit = 2;
297 
298 /*
299  * One thread queue per processor.
300  */
301 static struct tdq __read_mostly *balance_tdq;
302 static int balance_ticks;
303 DPCPU_DEFINE_STATIC(struct tdq, tdq);
304 DPCPU_DEFINE_STATIC(uint32_t, randomval);
305 
306 #define	TDQ_SELF()	((struct tdq *)PCPU_GET(sched))
307 #define	TDQ_CPU(x)	(DPCPU_ID_PTR((x), tdq))
308 #define	TDQ_ID(x)	((x)->tdq_id)
309 #else	/* !SMP */
310 static struct tdq	tdq_cpu;
311 
312 #define	TDQ_ID(x)	(0)
313 #define	TDQ_SELF()	(&tdq_cpu)
314 #define	TDQ_CPU(x)	(&tdq_cpu)
315 #endif
316 
317 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
318 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
319 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
320 #define	TDQ_TRYLOCK(t)		mtx_trylock_spin(TDQ_LOCKPTR((t)))
321 #define	TDQ_TRYLOCK_FLAGS(t, f)	mtx_trylock_spin_flags(TDQ_LOCKPTR((t)), (f))
322 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
323 #define	TDQ_LOCKPTR(t)		((struct mtx *)(&(t)->tdq_lock))
324 
325 static void sched_setpreempt(int);
326 static void sched_priority(struct thread *);
327 static void sched_thread_priority(struct thread *, u_char);
328 static int sched_interact_score(struct thread *);
329 static void sched_interact_update(struct thread *);
330 static void sched_interact_fork(struct thread *);
331 static void sched_pctcpu_update(struct td_sched *, int);
332 
333 /* Operations on per processor queues */
334 static struct thread *tdq_choose(struct tdq *);
335 static void tdq_setup(struct tdq *, int i);
336 static void tdq_load_add(struct tdq *, struct thread *);
337 static void tdq_load_rem(struct tdq *, struct thread *);
338 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
339 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
340 static inline int sched_shouldpreempt(int, int, int);
341 static void tdq_print(int cpu);
342 static void runq_print(struct runq *rq);
343 static int tdq_add(struct tdq *, struct thread *, int);
344 #ifdef SMP
345 static int tdq_move(struct tdq *, struct tdq *);
346 static int tdq_idled(struct tdq *);
347 static void tdq_notify(struct tdq *, int lowpri);
348 static struct thread *tdq_steal(struct tdq *, int);
349 static struct thread *runq_steal(struct runq *, int);
350 static int sched_pickcpu(struct thread *, int);
351 static void sched_balance(void);
352 static bool sched_balance_pair(struct tdq *, struct tdq *);
353 static inline struct tdq *sched_setcpu(struct thread *, int, int);
354 static inline void thread_unblock_switch(struct thread *, struct mtx *);
355 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
356 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
357     struct cpu_group *cg, int indent);
358 #endif
359 
360 static void sched_setup(void *dummy);
361 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
362 
363 static void sched_initticks(void *dummy);
364 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
365     NULL);
366 
367 SDT_PROVIDER_DEFINE(sched);
368 
369 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
370     "struct proc *", "uint8_t");
371 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
372     "struct proc *", "void *");
373 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
374     "struct proc *", "void *", "int");
375 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
376     "struct proc *", "uint8_t", "struct thread *");
377 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
378 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
379     "struct proc *");
380 SDT_PROBE_DEFINE(sched, , , on__cpu);
381 SDT_PROBE_DEFINE(sched, , , remain__cpu);
382 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
383     "struct proc *");
384 
385 /*
386  * Print the threads waiting on a run-queue.
387  */
388 static void
389 runq_print(struct runq *rq)
390 {
391 	struct rqhead *rqh;
392 	struct thread *td;
393 	int pri;
394 	int j;
395 	int i;
396 
397 	for (i = 0; i < RQB_LEN; i++) {
398 		printf("\t\trunq bits %d 0x%zx\n",
399 		    i, rq->rq_status.rqb_bits[i]);
400 		for (j = 0; j < RQB_BPW; j++)
401 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
402 				pri = j + (i << RQB_L2BPW);
403 				rqh = &rq->rq_queues[pri];
404 				TAILQ_FOREACH(td, rqh, td_runq) {
405 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
406 					    td, td->td_name, td->td_priority,
407 					    td->td_rqindex, pri);
408 				}
409 			}
410 	}
411 }
412 
413 /*
414  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
415  */
416 static void __unused
417 tdq_print(int cpu)
418 {
419 	struct tdq *tdq;
420 
421 	tdq = TDQ_CPU(cpu);
422 
423 	printf("tdq %d:\n", TDQ_ID(tdq));
424 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
425 	printf("\tLock name:      %s\n", tdq->tdq_name);
426 	printf("\tload:           %d\n", tdq->tdq_load);
427 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
428 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
429 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
430 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
431 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
432 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
433 	printf("\trealtime runq:\n");
434 	runq_print(&tdq->tdq_realtime);
435 	printf("\ttimeshare runq:\n");
436 	runq_print(&tdq->tdq_timeshare);
437 	printf("\tidle runq:\n");
438 	runq_print(&tdq->tdq_idle);
439 }
440 
441 static inline int
442 sched_shouldpreempt(int pri, int cpri, int remote)
443 {
444 	/*
445 	 * If the new priority is not better than the current priority there is
446 	 * nothing to do.
447 	 */
448 	if (pri >= cpri)
449 		return (0);
450 	/*
451 	 * Always preempt idle.
452 	 */
453 	if (cpri >= PRI_MIN_IDLE)
454 		return (1);
455 	/*
456 	 * If preemption is disabled don't preempt others.
457 	 */
458 	if (preempt_thresh == 0)
459 		return (0);
460 	/*
461 	 * Preempt if we exceed the threshold.
462 	 */
463 	if (pri <= preempt_thresh)
464 		return (1);
465 	/*
466 	 * If we're interactive or better and there is non-interactive
467 	 * or worse running preempt only remote processors.
468 	 */
469 	if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
470 		return (1);
471 	return (0);
472 }
473 
474 /*
475  * Add a thread to the actual run-queue.  Keeps transferable counts up to
476  * date with what is actually on the run-queue.  Selects the correct
477  * queue position for timeshare threads.
478  */
479 static __inline void
480 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
481 {
482 	struct td_sched *ts;
483 	u_char pri;
484 
485 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
486 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
487 
488 	pri = td->td_priority;
489 	ts = td_get_sched(td);
490 	TD_SET_RUNQ(td);
491 	if (THREAD_CAN_MIGRATE(td)) {
492 		tdq->tdq_transferable++;
493 		ts->ts_flags |= TSF_XFERABLE;
494 	}
495 	if (pri < PRI_MIN_BATCH) {
496 		ts->ts_runq = &tdq->tdq_realtime;
497 	} else if (pri <= PRI_MAX_BATCH) {
498 		ts->ts_runq = &tdq->tdq_timeshare;
499 		KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
500 			("Invalid priority %d on timeshare runq", pri));
501 		/*
502 		 * This queue contains only priorities between MIN and MAX
503 		 * batch.  Use the whole queue to represent these values.
504 		 */
505 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
506 			pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
507 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
508 			/*
509 			 * This effectively shortens the queue by one so we
510 			 * can have a one slot difference between idx and
511 			 * ridx while we wait for threads to drain.
512 			 */
513 			if (tdq->tdq_ridx != tdq->tdq_idx &&
514 			    pri == tdq->tdq_ridx)
515 				pri = (unsigned char)(pri - 1) % RQ_NQS;
516 		} else
517 			pri = tdq->tdq_ridx;
518 		runq_add_pri(ts->ts_runq, td, pri, flags);
519 		return;
520 	} else
521 		ts->ts_runq = &tdq->tdq_idle;
522 	runq_add(ts->ts_runq, td, flags);
523 }
524 
525 /*
526  * Remove a thread from a run-queue.  This typically happens when a thread
527  * is selected to run.  Running threads are not on the queue and the
528  * transferable count does not reflect them.
529  */
530 static __inline void
531 tdq_runq_rem(struct tdq *tdq, struct thread *td)
532 {
533 	struct td_sched *ts;
534 
535 	ts = td_get_sched(td);
536 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
537 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
538 	KASSERT(ts->ts_runq != NULL,
539 	    ("tdq_runq_remove: thread %p null ts_runq", td));
540 	if (ts->ts_flags & TSF_XFERABLE) {
541 		tdq->tdq_transferable--;
542 		ts->ts_flags &= ~TSF_XFERABLE;
543 	}
544 	if (ts->ts_runq == &tdq->tdq_timeshare) {
545 		if (tdq->tdq_idx != tdq->tdq_ridx)
546 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
547 		else
548 			runq_remove_idx(ts->ts_runq, td, NULL);
549 	} else
550 		runq_remove(ts->ts_runq, td);
551 }
552 
553 /*
554  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
555  * for this thread to the referenced thread queue.
556  */
557 static void
558 tdq_load_add(struct tdq *tdq, struct thread *td)
559 {
560 
561 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
562 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
563 
564 	tdq->tdq_load++;
565 	if ((td->td_flags & TDF_NOLOAD) == 0)
566 		tdq->tdq_sysload++;
567 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
568 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
569 }
570 
571 /*
572  * Remove the load from a thread that is transitioning to a sleep state or
573  * exiting.
574  */
575 static void
576 tdq_load_rem(struct tdq *tdq, struct thread *td)
577 {
578 
579 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
580 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
581 	KASSERT(tdq->tdq_load != 0,
582 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
583 
584 	tdq->tdq_load--;
585 	if ((td->td_flags & TDF_NOLOAD) == 0)
586 		tdq->tdq_sysload--;
587 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
588 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
589 }
590 
591 /*
592  * Bound timeshare latency by decreasing slice size as load increases.  We
593  * consider the maximum latency as the sum of the threads waiting to run
594  * aside from curthread and target no more than sched_slice latency but
595  * no less than sched_slice_min runtime.
596  */
597 static inline int
598 tdq_slice(struct tdq *tdq)
599 {
600 	int load;
601 
602 	/*
603 	 * It is safe to use sys_load here because this is called from
604 	 * contexts where timeshare threads are running and so there
605 	 * cannot be higher priority load in the system.
606 	 */
607 	load = tdq->tdq_sysload - 1;
608 	if (load >= SCHED_SLICE_MIN_DIVISOR)
609 		return (sched_slice_min);
610 	if (load <= 1)
611 		return (sched_slice);
612 	return (sched_slice / load);
613 }
614 
615 /*
616  * Set lowpri to its exact value by searching the run-queue and
617  * evaluating curthread.  curthread may be passed as an optimization.
618  */
619 static void
620 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
621 {
622 	struct thread *td;
623 
624 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
625 	if (ctd == NULL)
626 		ctd = tdq->tdq_curthread;
627 	td = tdq_choose(tdq);
628 	if (td == NULL || td->td_priority > ctd->td_priority)
629 		tdq->tdq_lowpri = ctd->td_priority;
630 	else
631 		tdq->tdq_lowpri = td->td_priority;
632 }
633 
634 #ifdef SMP
635 /*
636  * We need some randomness. Implement a classic Linear Congruential
637  * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
638  * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
639  * of the random state (in the low bits of our answer) to keep
640  * the maximum randomness.
641  */
642 static uint32_t
643 sched_random(void)
644 {
645 	uint32_t *rndptr;
646 
647 	rndptr = DPCPU_PTR(randomval);
648 	*rndptr = *rndptr * 69069 + 5;
649 
650 	return (*rndptr >> 16);
651 }
652 
653 struct cpu_search {
654 	cpuset_t *cs_mask;	/* The mask of allowed CPUs to choose from. */
655 	int	cs_prefer;	/* Prefer this CPU and groups including it. */
656 	int	cs_running;	/* The thread is now running at cs_prefer. */
657 	int	cs_pri;		/* Min priority for low. */
658 	int	cs_load;	/* Max load for low, min load for high. */
659 	int	cs_trans;	/* Min transferable load for high. */
660 };
661 
662 struct cpu_search_res {
663 	int	csr_cpu;	/* The best CPU found. */
664 	int	csr_load;	/* The load of cs_cpu. */
665 };
666 
667 /*
668  * Search the tree of cpu_groups for the lowest or highest loaded CPU.
669  * These routines actually compare the load on all paths through the tree
670  * and find the least loaded cpu on the least loaded path, which may differ
671  * from the least loaded cpu in the system.  This balances work among caches
672  * and buses.
673  */
674 static int
675 cpu_search_lowest(const struct cpu_group *cg, const struct cpu_search *s,
676     struct cpu_search_res *r)
677 {
678 	struct cpu_search_res lr;
679 	struct tdq *tdq;
680 	int c, bload, l, load, p, total;
681 
682 	total = 0;
683 	bload = INT_MAX;
684 	r->csr_cpu = -1;
685 
686 	/* Loop through children CPU groups if there are any. */
687 	if (cg->cg_children > 0) {
688 		for (c = cg->cg_children - 1; c >= 0; c--) {
689 			load = cpu_search_lowest(&cg->cg_child[c], s, &lr);
690 			total += load;
691 
692 			/*
693 			 * When balancing do not prefer SMT groups with load >1.
694 			 * It allows round-robin between SMT groups with equal
695 			 * load within parent group for more fair scheduling.
696 			 */
697 			if (__predict_false(s->cs_running) &&
698 			    (cg->cg_child[c].cg_flags & CG_FLAG_THREAD) &&
699 			    load >= 128 && (load & 128) != 0)
700 				load += 128;
701 
702 			if (lr.csr_cpu >= 0 && (load < bload ||
703 			    (load == bload && lr.csr_load < r->csr_load))) {
704 				bload = load;
705 				r->csr_cpu = lr.csr_cpu;
706 				r->csr_load = lr.csr_load;
707 			}
708 		}
709 		return (total);
710 	}
711 
712 	/* Loop through children CPUs otherwise. */
713 	for (c = cg->cg_last; c >= cg->cg_first; c--) {
714 		if (!CPU_ISSET(c, &cg->cg_mask))
715 			continue;
716 		tdq = TDQ_CPU(c);
717 		l = TDQ_LOAD(tdq);
718 		if (c == s->cs_prefer) {
719 			if (__predict_false(s->cs_running))
720 				l--;
721 			p = 128;
722 		} else
723 			p = 0;
724 		load = l * 256;
725 		total += load - p;
726 
727 		/*
728 		 * Check this CPU is acceptable.
729 		 * If the threads is already on the CPU, don't look on the TDQ
730 		 * priority, since it can be the priority of the thread itself.
731 		 */
732 		if (l > s->cs_load ||
733 		    (atomic_load_char(&tdq->tdq_lowpri) <= s->cs_pri &&
734 		     (!s->cs_running || c != s->cs_prefer)) ||
735 		    !CPU_ISSET(c, s->cs_mask))
736 			continue;
737 
738 		/*
739 		 * When balancing do not prefer CPUs with load > 1.
740 		 * It allows round-robin between CPUs with equal load
741 		 * within the CPU group for more fair scheduling.
742 		 */
743 		if (__predict_false(s->cs_running) && l > 0)
744 			p = 0;
745 
746 		load -= sched_random() % 128;
747 		if (bload > load - p) {
748 			bload = load - p;
749 			r->csr_cpu = c;
750 			r->csr_load = load;
751 		}
752 	}
753 	return (total);
754 }
755 
756 static int
757 cpu_search_highest(const struct cpu_group *cg, const struct cpu_search *s,
758     struct cpu_search_res *r)
759 {
760 	struct cpu_search_res lr;
761 	struct tdq *tdq;
762 	int c, bload, l, load, total;
763 
764 	total = 0;
765 	bload = INT_MIN;
766 	r->csr_cpu = -1;
767 
768 	/* Loop through children CPU groups if there are any. */
769 	if (cg->cg_children > 0) {
770 		for (c = cg->cg_children - 1; c >= 0; c--) {
771 			load = cpu_search_highest(&cg->cg_child[c], s, &lr);
772 			total += load;
773 			if (lr.csr_cpu >= 0 && (load > bload ||
774 			    (load == bload && lr.csr_load > r->csr_load))) {
775 				bload = load;
776 				r->csr_cpu = lr.csr_cpu;
777 				r->csr_load = lr.csr_load;
778 			}
779 		}
780 		return (total);
781 	}
782 
783 	/* Loop through children CPUs otherwise. */
784 	for (c = cg->cg_last; c >= cg->cg_first; c--) {
785 		if (!CPU_ISSET(c, &cg->cg_mask))
786 			continue;
787 		tdq = TDQ_CPU(c);
788 		l = TDQ_LOAD(tdq);
789 		load = l * 256;
790 		total += load;
791 
792 		/*
793 		 * Check this CPU is acceptable.
794 		 */
795 		if (l < s->cs_load || TDQ_TRANSFERABLE(tdq) < s->cs_trans ||
796 		    !CPU_ISSET(c, s->cs_mask))
797 			continue;
798 
799 		load -= sched_random() % 256;
800 		if (load > bload) {
801 			bload = load;
802 			r->csr_cpu = c;
803 		}
804 	}
805 	r->csr_load = bload;
806 	return (total);
807 }
808 
809 /*
810  * Find the cpu with the least load via the least loaded path that has a
811  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
812  * acceptable.
813  */
814 static inline int
815 sched_lowest(const struct cpu_group *cg, cpuset_t *mask, int pri, int maxload,
816     int prefer, int running)
817 {
818 	struct cpu_search s;
819 	struct cpu_search_res r;
820 
821 	s.cs_prefer = prefer;
822 	s.cs_running = running;
823 	s.cs_mask = mask;
824 	s.cs_pri = pri;
825 	s.cs_load = maxload;
826 	cpu_search_lowest(cg, &s, &r);
827 	return (r.csr_cpu);
828 }
829 
830 /*
831  * Find the cpu with the highest load via the highest loaded path.
832  */
833 static inline int
834 sched_highest(const struct cpu_group *cg, cpuset_t *mask, int minload,
835     int mintrans)
836 {
837 	struct cpu_search s;
838 	struct cpu_search_res r;
839 
840 	s.cs_mask = mask;
841 	s.cs_load = minload;
842 	s.cs_trans = mintrans;
843 	cpu_search_highest(cg, &s, &r);
844 	return (r.csr_cpu);
845 }
846 
847 static void
848 sched_balance_group(struct cpu_group *cg)
849 {
850 	struct tdq *tdq;
851 	struct thread *td;
852 	cpuset_t hmask, lmask;
853 	int high, low, anylow;
854 
855 	CPU_FILL(&hmask);
856 	for (;;) {
857 		high = sched_highest(cg, &hmask, 1, 0);
858 		/* Stop if there is no more CPU with transferrable threads. */
859 		if (high == -1)
860 			break;
861 		CPU_CLR(high, &hmask);
862 		CPU_COPY(&hmask, &lmask);
863 		/* Stop if there is no more CPU left for low. */
864 		if (CPU_EMPTY(&lmask))
865 			break;
866 		tdq = TDQ_CPU(high);
867 		if (TDQ_LOAD(tdq) == 1) {
868 			/*
869 			 * There is only one running thread.  We can't move
870 			 * it from here, so tell it to pick new CPU by itself.
871 			 */
872 			TDQ_LOCK(tdq);
873 			td = tdq->tdq_curthread;
874 			if ((td->td_flags & TDF_IDLETD) == 0 &&
875 			    THREAD_CAN_MIGRATE(td)) {
876 				td->td_flags |= TDF_NEEDRESCHED | TDF_PICKCPU;
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_int(&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 user priority, does not effect current running priority.
1967  */
1968 void
1969 sched_user_prio(struct thread *td, u_char prio)
1970 {
1971 
1972 	td->td_base_user_pri = prio;
1973 	if (td->td_lend_user_pri <= prio)
1974 		return;
1975 	td->td_user_pri = prio;
1976 }
1977 
1978 void
1979 sched_lend_user_prio(struct thread *td, u_char prio)
1980 {
1981 
1982 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1983 	td->td_lend_user_pri = prio;
1984 	td->td_user_pri = min(prio, td->td_base_user_pri);
1985 	if (td->td_priority > td->td_user_pri)
1986 		sched_prio(td, td->td_user_pri);
1987 	else if (td->td_priority != td->td_user_pri)
1988 		td->td_flags |= TDF_NEEDRESCHED;
1989 }
1990 
1991 /*
1992  * Like the above but first check if there is anything to do.
1993  */
1994 void
1995 sched_lend_user_prio_cond(struct thread *td, u_char prio)
1996 {
1997 
1998 	if (td->td_lend_user_pri != prio)
1999 		goto lend;
2000 	if (td->td_user_pri != min(prio, td->td_base_user_pri))
2001 		goto lend;
2002 	if (td->td_priority != td->td_user_pri)
2003 		goto lend;
2004 	return;
2005 
2006 lend:
2007 	thread_lock(td);
2008 	sched_lend_user_prio(td, prio);
2009 	thread_unlock(td);
2010 }
2011 
2012 #ifdef SMP
2013 /*
2014  * This tdq is about to idle.  Try to steal a thread from another CPU before
2015  * choosing the idle thread.
2016  */
2017 static void
2018 tdq_trysteal(struct tdq *tdq)
2019 {
2020 	struct cpu_group *cg, *parent;
2021 	struct tdq *steal;
2022 	cpuset_t mask;
2023 	int cpu, i, goup;
2024 
2025 	if (smp_started == 0 || steal_idle == 0 || trysteal_limit == 0 ||
2026 	    tdq->tdq_cg == NULL)
2027 		return;
2028 	CPU_FILL(&mask);
2029 	CPU_CLR(PCPU_GET(cpuid), &mask);
2030 	/* We don't want to be preempted while we're iterating. */
2031 	spinlock_enter();
2032 	TDQ_UNLOCK(tdq);
2033 	for (i = 1, cg = tdq->tdq_cg, goup = 0; ; ) {
2034 		cpu = sched_highest(cg, &mask, steal_thresh, 1);
2035 		/*
2036 		 * If a thread was added while interrupts were disabled don't
2037 		 * steal one here.
2038 		 */
2039 		if (TDQ_LOAD(tdq) > 0) {
2040 			TDQ_LOCK(tdq);
2041 			break;
2042 		}
2043 
2044 		/*
2045 		 * We found no CPU to steal from in this group.  Escalate to
2046 		 * the parent and repeat.  But if parent has only two children
2047 		 * groups we can avoid searching this group again by searching
2048 		 * the other one specifically and then escalating two levels.
2049 		 */
2050 		if (cpu == -1) {
2051 			if (goup) {
2052 				cg = cg->cg_parent;
2053 				goup = 0;
2054 			}
2055 			if (++i > trysteal_limit) {
2056 				TDQ_LOCK(tdq);
2057 				break;
2058 			}
2059 			parent = cg->cg_parent;
2060 			if (parent == NULL) {
2061 				TDQ_LOCK(tdq);
2062 				break;
2063 			}
2064 			if (parent->cg_children == 2) {
2065 				if (cg == &parent->cg_child[0])
2066 					cg = &parent->cg_child[1];
2067 				else
2068 					cg = &parent->cg_child[0];
2069 				goup = 1;
2070 			} else
2071 				cg = parent;
2072 			continue;
2073 		}
2074 		steal = TDQ_CPU(cpu);
2075 		/*
2076 		 * The data returned by sched_highest() is stale and
2077 		 * the chosen CPU no longer has an eligible thread.
2078 		 * At this point unconditionally exit the loop to bound
2079 		 * the time spent in the critcal section.
2080 		 */
2081 		if (TDQ_LOAD(steal) < steal_thresh ||
2082 		    TDQ_TRANSFERABLE(steal) == 0)
2083 			continue;
2084 		/*
2085 		 * Try to lock both queues. If we are assigned a thread while
2086 		 * waited for the lock, switch to it now instead of stealing.
2087 		 * If we can't get the lock, then somebody likely got there
2088 		 * first.
2089 		 */
2090 		TDQ_LOCK(tdq);
2091 		if (tdq->tdq_load > 0)
2092 			break;
2093 		if (TDQ_TRYLOCK_FLAGS(steal, MTX_DUPOK) == 0)
2094 			break;
2095 		/*
2096 		 * The data returned by sched_highest() is stale and
2097                  * the chosen CPU no longer has an eligible thread.
2098 		 */
2099 		if (TDQ_LOAD(steal) < steal_thresh ||
2100 		    TDQ_TRANSFERABLE(steal) == 0) {
2101 			TDQ_UNLOCK(steal);
2102 			break;
2103 		}
2104 		/*
2105 		 * If we fail to acquire one due to affinity restrictions,
2106 		 * bail out and let the idle thread to a more complete search
2107 		 * outside of a critical section.
2108 		 */
2109 		if (tdq_move(steal, tdq) == -1) {
2110 			TDQ_UNLOCK(steal);
2111 			break;
2112 		}
2113 		TDQ_UNLOCK(steal);
2114 		break;
2115 	}
2116 	spinlock_exit();
2117 }
2118 #endif
2119 
2120 /*
2121  * Handle migration from sched_switch().  This happens only for
2122  * cpu binding.
2123  */
2124 static struct mtx *
2125 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
2126 {
2127 	struct tdq *tdn;
2128 	int lowpri;
2129 
2130 	KASSERT(THREAD_CAN_MIGRATE(td) ||
2131 	    (td_get_sched(td)->ts_flags & TSF_BOUND) != 0,
2132 	    ("Thread %p shouldn't migrate", td));
2133 	KASSERT(!CPU_ABSENT(td_get_sched(td)->ts_cpu), ("sched_switch_migrate: "
2134 	    "thread %s queued on absent CPU %d.", td->td_name,
2135 	    td_get_sched(td)->ts_cpu));
2136 	tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
2137 #ifdef SMP
2138 	tdq_load_rem(tdq, td);
2139 	/*
2140 	 * Do the lock dance required to avoid LOR.  We have an
2141 	 * extra spinlock nesting from sched_switch() which will
2142 	 * prevent preemption while we're holding neither run-queue lock.
2143 	 */
2144 	TDQ_UNLOCK(tdq);
2145 	TDQ_LOCK(tdn);
2146 	lowpri = tdq_add(tdn, td, flags);
2147 	tdq_notify(tdn, lowpri);
2148 	TDQ_UNLOCK(tdn);
2149 	TDQ_LOCK(tdq);
2150 #endif
2151 	return (TDQ_LOCKPTR(tdn));
2152 }
2153 
2154 /*
2155  * thread_lock_unblock() that does not assume td_lock is blocked.
2156  */
2157 static inline void
2158 thread_unblock_switch(struct thread *td, struct mtx *mtx)
2159 {
2160 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
2161 	    (uintptr_t)mtx);
2162 }
2163 
2164 /*
2165  * Switch threads.  This function has to handle threads coming in while
2166  * blocked for some reason, running, or idle.  It also must deal with
2167  * migrating a thread from one queue to another as running threads may
2168  * be assigned elsewhere via binding.
2169  */
2170 void
2171 sched_switch(struct thread *td, int flags)
2172 {
2173 	struct thread *newtd;
2174 	struct tdq *tdq;
2175 	struct td_sched *ts;
2176 	struct mtx *mtx;
2177 	int srqflag;
2178 	int cpuid, preempted;
2179 #ifdef SMP
2180 	int pickcpu;
2181 #endif
2182 
2183 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2184 
2185 	cpuid = PCPU_GET(cpuid);
2186 	tdq = TDQ_SELF();
2187 	ts = td_get_sched(td);
2188 	sched_pctcpu_update(ts, 1);
2189 #ifdef SMP
2190 	pickcpu = (td->td_flags & TDF_PICKCPU) != 0;
2191 	if (pickcpu)
2192 		ts->ts_rltick = ticks - affinity * MAX_CACHE_LEVELS;
2193 	else
2194 		ts->ts_rltick = ticks;
2195 #endif
2196 	td->td_lastcpu = td->td_oncpu;
2197 	preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
2198 	    (flags & SW_PREEMPT) != 0;
2199 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_PICKCPU | TDF_SLICEEND);
2200 	td->td_owepreempt = 0;
2201 	atomic_store_char(&tdq->tdq_owepreempt, 0);
2202 	if (!TD_IS_IDLETHREAD(td))
2203 		TDQ_SWITCHCNT_INC(tdq);
2204 
2205 	/*
2206 	 * Always block the thread lock so we can drop the tdq lock early.
2207 	 */
2208 	mtx = thread_lock_block(td);
2209 	spinlock_enter();
2210 	if (TD_IS_IDLETHREAD(td)) {
2211 		MPASS(mtx == TDQ_LOCKPTR(tdq));
2212 		TD_SET_CAN_RUN(td);
2213 	} else if (TD_IS_RUNNING(td)) {
2214 		MPASS(mtx == TDQ_LOCKPTR(tdq));
2215 		srqflag = preempted ?
2216 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
2217 		    SRQ_OURSELF|SRQ_YIELDING;
2218 #ifdef SMP
2219 		if (THREAD_CAN_MIGRATE(td) && (!THREAD_CAN_SCHED(td, ts->ts_cpu)
2220 		    || pickcpu))
2221 			ts->ts_cpu = sched_pickcpu(td, 0);
2222 #endif
2223 		if (ts->ts_cpu == cpuid)
2224 			tdq_runq_add(tdq, td, srqflag);
2225 		else
2226 			mtx = sched_switch_migrate(tdq, td, srqflag);
2227 	} else {
2228 		/* This thread must be going to sleep. */
2229 		if (mtx != TDQ_LOCKPTR(tdq)) {
2230 			mtx_unlock_spin(mtx);
2231 			TDQ_LOCK(tdq);
2232 		}
2233 		tdq_load_rem(tdq, td);
2234 #ifdef SMP
2235 		if (tdq->tdq_load == 0)
2236 			tdq_trysteal(tdq);
2237 #endif
2238 	}
2239 
2240 #if (KTR_COMPILE & KTR_SCHED) != 0
2241 	if (TD_IS_IDLETHREAD(td))
2242 		KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
2243 		    "prio:%d", td->td_priority);
2244 	else
2245 		KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
2246 		    "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
2247 		    "lockname:\"%s\"", td->td_lockname);
2248 #endif
2249 
2250 	/*
2251 	 * We enter here with the thread blocked and assigned to the
2252 	 * appropriate cpu run-queue or sleep-queue and with the current
2253 	 * thread-queue locked.
2254 	 */
2255 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2256 	MPASS(td == tdq->tdq_curthread);
2257 	newtd = choosethread();
2258 	sched_pctcpu_update(td_get_sched(newtd), 0);
2259 	TDQ_UNLOCK(tdq);
2260 
2261 	/*
2262 	 * Call the MD code to switch contexts if necessary.
2263 	 */
2264 	if (td != newtd) {
2265 #ifdef	HWPMC_HOOKS
2266 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2267 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
2268 #endif
2269 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
2270 
2271 #ifdef KDTRACE_HOOKS
2272 		/*
2273 		 * If DTrace has set the active vtime enum to anything
2274 		 * other than INACTIVE (0), then it should have set the
2275 		 * function to call.
2276 		 */
2277 		if (dtrace_vtime_active)
2278 			(*dtrace_vtime_switch_func)(newtd);
2279 #endif
2280 		td->td_oncpu = NOCPU;
2281 		cpu_switch(td, newtd, mtx);
2282 		cpuid = td->td_oncpu = PCPU_GET(cpuid);
2283 
2284 		SDT_PROBE0(sched, , , on__cpu);
2285 #ifdef	HWPMC_HOOKS
2286 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2287 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
2288 #endif
2289 	} else {
2290 		thread_unblock_switch(td, mtx);
2291 		SDT_PROBE0(sched, , , remain__cpu);
2292 	}
2293 	KASSERT(curthread->td_md.md_spinlock_count == 1,
2294 	    ("invalid count %d", curthread->td_md.md_spinlock_count));
2295 
2296 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2297 	    "prio:%d", td->td_priority);
2298 }
2299 
2300 /*
2301  * Adjust thread priorities as a result of a nice request.
2302  */
2303 void
2304 sched_nice(struct proc *p, int nice)
2305 {
2306 	struct thread *td;
2307 
2308 	PROC_LOCK_ASSERT(p, MA_OWNED);
2309 
2310 	p->p_nice = nice;
2311 	FOREACH_THREAD_IN_PROC(p, td) {
2312 		thread_lock(td);
2313 		sched_priority(td);
2314 		sched_prio(td, td->td_base_user_pri);
2315 		thread_unlock(td);
2316 	}
2317 }
2318 
2319 /*
2320  * Record the sleep time for the interactivity scorer.
2321  */
2322 void
2323 sched_sleep(struct thread *td, int prio)
2324 {
2325 
2326 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2327 
2328 	td->td_slptick = ticks;
2329 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2330 		td->td_flags |= TDF_CANSWAP;
2331 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2332 		return;
2333 	if (static_boost == 1 && prio)
2334 		sched_prio(td, prio);
2335 	else if (static_boost && td->td_priority > static_boost)
2336 		sched_prio(td, static_boost);
2337 }
2338 
2339 /*
2340  * Schedule a thread to resume execution and record how long it voluntarily
2341  * slept.  We also update the pctcpu, interactivity, and priority.
2342  *
2343  * Requires the thread lock on entry, drops on exit.
2344  */
2345 void
2346 sched_wakeup(struct thread *td, int srqflags)
2347 {
2348 	struct td_sched *ts;
2349 	int slptick;
2350 
2351 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2352 	ts = td_get_sched(td);
2353 	td->td_flags &= ~TDF_CANSWAP;
2354 
2355 	/*
2356 	 * If we slept for more than a tick update our interactivity and
2357 	 * priority.
2358 	 */
2359 	slptick = td->td_slptick;
2360 	td->td_slptick = 0;
2361 	if (slptick && slptick != ticks) {
2362 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2363 		sched_interact_update(td);
2364 		sched_pctcpu_update(ts, 0);
2365 	}
2366 	/*
2367 	 * Reset the slice value since we slept and advanced the round-robin.
2368 	 */
2369 	ts->ts_slice = 0;
2370 	sched_add(td, SRQ_BORING | srqflags);
2371 }
2372 
2373 /*
2374  * Penalize the parent for creating a new child and initialize the child's
2375  * priority.
2376  */
2377 void
2378 sched_fork(struct thread *td, struct thread *child)
2379 {
2380 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2381 	sched_pctcpu_update(td_get_sched(td), 1);
2382 	sched_fork_thread(td, child);
2383 	/*
2384 	 * Penalize the parent and child for forking.
2385 	 */
2386 	sched_interact_fork(child);
2387 	sched_priority(child);
2388 	td_get_sched(td)->ts_runtime += tickincr;
2389 	sched_interact_update(td);
2390 	sched_priority(td);
2391 }
2392 
2393 /*
2394  * Fork a new thread, may be within the same process.
2395  */
2396 void
2397 sched_fork_thread(struct thread *td, struct thread *child)
2398 {
2399 	struct td_sched *ts;
2400 	struct td_sched *ts2;
2401 	struct tdq *tdq;
2402 
2403 	tdq = TDQ_SELF();
2404 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2405 	/*
2406 	 * Initialize child.
2407 	 */
2408 	ts = td_get_sched(td);
2409 	ts2 = td_get_sched(child);
2410 	child->td_oncpu = NOCPU;
2411 	child->td_lastcpu = NOCPU;
2412 	child->td_lock = TDQ_LOCKPTR(tdq);
2413 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2414 	child->td_domain.dr_policy = td->td_cpuset->cs_domain;
2415 	ts2->ts_cpu = ts->ts_cpu;
2416 	ts2->ts_flags = 0;
2417 	/*
2418 	 * Grab our parents cpu estimation information.
2419 	 */
2420 	ts2->ts_ticks = ts->ts_ticks;
2421 	ts2->ts_ltick = ts->ts_ltick;
2422 	ts2->ts_ftick = ts->ts_ftick;
2423 	/*
2424 	 * Do not inherit any borrowed priority from the parent.
2425 	 */
2426 	child->td_priority = child->td_base_pri;
2427 	/*
2428 	 * And update interactivity score.
2429 	 */
2430 	ts2->ts_slptime = ts->ts_slptime;
2431 	ts2->ts_runtime = ts->ts_runtime;
2432 	/* Attempt to quickly learn interactivity. */
2433 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2434 #ifdef KTR
2435 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2436 #endif
2437 }
2438 
2439 /*
2440  * Adjust the priority class of a thread.
2441  */
2442 void
2443 sched_class(struct thread *td, int class)
2444 {
2445 
2446 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2447 	if (td->td_pri_class == class)
2448 		return;
2449 	td->td_pri_class = class;
2450 }
2451 
2452 /*
2453  * Return some of the child's priority and interactivity to the parent.
2454  */
2455 void
2456 sched_exit(struct proc *p, struct thread *child)
2457 {
2458 	struct thread *td;
2459 
2460 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2461 	    "prio:%d", child->td_priority);
2462 	PROC_LOCK_ASSERT(p, MA_OWNED);
2463 	td = FIRST_THREAD_IN_PROC(p);
2464 	sched_exit_thread(td, child);
2465 }
2466 
2467 /*
2468  * Penalize another thread for the time spent on this one.  This helps to
2469  * worsen the priority and interactivity of processes which schedule batch
2470  * jobs such as make.  This has little effect on the make process itself but
2471  * causes new processes spawned by it to receive worse scores immediately.
2472  */
2473 void
2474 sched_exit_thread(struct thread *td, struct thread *child)
2475 {
2476 
2477 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2478 	    "prio:%d", child->td_priority);
2479 	/*
2480 	 * Give the child's runtime to the parent without returning the
2481 	 * sleep time as a penalty to the parent.  This causes shells that
2482 	 * launch expensive things to mark their children as expensive.
2483 	 */
2484 	thread_lock(td);
2485 	td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2486 	sched_interact_update(td);
2487 	sched_priority(td);
2488 	thread_unlock(td);
2489 }
2490 
2491 void
2492 sched_preempt(struct thread *td)
2493 {
2494 	struct tdq *tdq;
2495 	int flags;
2496 
2497 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2498 
2499 	thread_lock(td);
2500 	tdq = TDQ_SELF();
2501 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2502 	if (td->td_priority > tdq->tdq_lowpri) {
2503 		if (td->td_critnest == 1) {
2504 			flags = SW_INVOL | SW_PREEMPT;
2505 			flags |= TD_IS_IDLETHREAD(td) ? SWT_REMOTEWAKEIDLE :
2506 			    SWT_REMOTEPREEMPT;
2507 			mi_switch(flags);
2508 			/* Switch dropped thread lock. */
2509 			return;
2510 		}
2511 		td->td_owepreempt = 1;
2512 	} else {
2513 		tdq->tdq_owepreempt = 0;
2514 	}
2515 	thread_unlock(td);
2516 }
2517 
2518 /*
2519  * Fix priorities on return to user-space.  Priorities may be elevated due
2520  * to static priorities in msleep() or similar.
2521  */
2522 void
2523 sched_userret_slowpath(struct thread *td)
2524 {
2525 
2526 	thread_lock(td);
2527 	td->td_priority = td->td_user_pri;
2528 	td->td_base_pri = td->td_user_pri;
2529 	tdq_setlowpri(TDQ_SELF(), td);
2530 	thread_unlock(td);
2531 }
2532 
2533 /*
2534  * Handle a stathz tick.  This is really only relevant for timeshare
2535  * threads.
2536  */
2537 void
2538 sched_clock(struct thread *td, int cnt)
2539 {
2540 	struct tdq *tdq;
2541 	struct td_sched *ts;
2542 
2543 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2544 	tdq = TDQ_SELF();
2545 #ifdef SMP
2546 	/*
2547 	 * We run the long term load balancer infrequently on the first cpu.
2548 	 */
2549 	if (balance_tdq == tdq && smp_started != 0 && rebalance != 0 &&
2550 	    balance_ticks != 0) {
2551 		balance_ticks -= cnt;
2552 		if (balance_ticks <= 0)
2553 			sched_balance();
2554 	}
2555 #endif
2556 	/*
2557 	 * Save the old switch count so we have a record of the last ticks
2558 	 * activity.   Initialize the new switch count based on our load.
2559 	 * If there is some activity seed it to reflect that.
2560 	 */
2561 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2562 	tdq->tdq_switchcnt = tdq->tdq_load;
2563 
2564 	/*
2565 	 * Advance the insert index once for each tick to ensure that all
2566 	 * threads get a chance to run.
2567 	 */
2568 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2569 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2570 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2571 			tdq->tdq_ridx = tdq->tdq_idx;
2572 	}
2573 	ts = td_get_sched(td);
2574 	sched_pctcpu_update(ts, 1);
2575 	if ((td->td_pri_class & PRI_FIFO_BIT) || TD_IS_IDLETHREAD(td))
2576 		return;
2577 
2578 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2579 		/*
2580 		 * We used a tick; charge it to the thread so
2581 		 * that we can compute our interactivity.
2582 		 */
2583 		td_get_sched(td)->ts_runtime += tickincr * cnt;
2584 		sched_interact_update(td);
2585 		sched_priority(td);
2586 	}
2587 
2588 	/*
2589 	 * Force a context switch if the current thread has used up a full
2590 	 * time slice (default is 100ms).
2591 	 */
2592 	ts->ts_slice += cnt;
2593 	if (ts->ts_slice >= tdq_slice(tdq)) {
2594 		ts->ts_slice = 0;
2595 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2596 	}
2597 }
2598 
2599 u_int
2600 sched_estcpu(struct thread *td __unused)
2601 {
2602 
2603 	return (0);
2604 }
2605 
2606 /*
2607  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2608  * cooperative idle threads.
2609  */
2610 int
2611 sched_runnable(void)
2612 {
2613 	struct tdq *tdq;
2614 	int load;
2615 
2616 	load = 1;
2617 
2618 	tdq = TDQ_SELF();
2619 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2620 		if (TDQ_LOAD(tdq) > 0)
2621 			goto out;
2622 	} else
2623 		if (TDQ_LOAD(tdq) - 1 > 0)
2624 			goto out;
2625 	load = 0;
2626 out:
2627 	return (load);
2628 }
2629 
2630 /*
2631  * Choose the highest priority thread to run.  The thread is removed from
2632  * the run-queue while running however the load remains.
2633  */
2634 struct thread *
2635 sched_choose(void)
2636 {
2637 	struct thread *td;
2638 	struct tdq *tdq;
2639 
2640 	tdq = TDQ_SELF();
2641 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2642 	td = tdq_choose(tdq);
2643 	if (td != NULL) {
2644 		tdq_runq_rem(tdq, td);
2645 		tdq->tdq_lowpri = td->td_priority;
2646 	} else {
2647 		tdq->tdq_lowpri = PRI_MAX_IDLE;
2648 		td = PCPU_GET(idlethread);
2649 	}
2650 	tdq->tdq_curthread = td;
2651 	return (td);
2652 }
2653 
2654 /*
2655  * Set owepreempt if the currently running thread has lower priority than "pri".
2656  * Preemption never happens directly in ULE, we always request it once we exit a
2657  * critical section.
2658  */
2659 static void
2660 sched_setpreempt(int pri)
2661 {
2662 	struct thread *ctd;
2663 	int cpri;
2664 
2665 	ctd = curthread;
2666 	THREAD_LOCK_ASSERT(ctd, MA_OWNED);
2667 
2668 	cpri = ctd->td_priority;
2669 	if (pri < cpri)
2670 		ctd->td_flags |= TDF_NEEDRESCHED;
2671 	if (KERNEL_PANICKED() || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2672 		return;
2673 	if (!sched_shouldpreempt(pri, cpri, 0))
2674 		return;
2675 	ctd->td_owepreempt = 1;
2676 }
2677 
2678 /*
2679  * Add a thread to a thread queue.  Select the appropriate runq and add the
2680  * thread to it.  This is the internal function called when the tdq is
2681  * predetermined.
2682  */
2683 static int
2684 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2685 {
2686 	int lowpri;
2687 
2688 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2689 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
2690 	KASSERT((td->td_inhibitors == 0),
2691 	    ("sched_add: trying to run inhibited thread"));
2692 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2693 	    ("sched_add: bad thread state"));
2694 	KASSERT(td->td_flags & TDF_INMEM,
2695 	    ("sched_add: thread swapped out"));
2696 
2697 	lowpri = tdq->tdq_lowpri;
2698 	if (td->td_priority < lowpri)
2699 		tdq->tdq_lowpri = td->td_priority;
2700 	tdq_runq_add(tdq, td, flags);
2701 	tdq_load_add(tdq, td);
2702 	return (lowpri);
2703 }
2704 
2705 /*
2706  * Select the target thread queue and add a thread to it.  Request
2707  * preemption or IPI a remote processor if required.
2708  *
2709  * Requires the thread lock on entry, drops on exit.
2710  */
2711 void
2712 sched_add(struct thread *td, int flags)
2713 {
2714 	struct tdq *tdq;
2715 #ifdef SMP
2716 	int cpu, lowpri;
2717 #endif
2718 
2719 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2720 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2721 	    sched_tdname(curthread));
2722 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2723 	    KTR_ATTR_LINKED, sched_tdname(td));
2724 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2725 	    flags & SRQ_PREEMPTED);
2726 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2727 	/*
2728 	 * Recalculate the priority before we select the target cpu or
2729 	 * run-queue.
2730 	 */
2731 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2732 		sched_priority(td);
2733 #ifdef SMP
2734 	/*
2735 	 * Pick the destination cpu and if it isn't ours transfer to the
2736 	 * target cpu.
2737 	 */
2738 	cpu = sched_pickcpu(td, flags);
2739 	tdq = sched_setcpu(td, cpu, flags);
2740 	lowpri = tdq_add(tdq, td, flags);
2741 	if (cpu != PCPU_GET(cpuid))
2742 		tdq_notify(tdq, lowpri);
2743 	else if (!(flags & SRQ_YIELDING))
2744 		sched_setpreempt(td->td_priority);
2745 #else
2746 	tdq = TDQ_SELF();
2747 	/*
2748 	 * Now that the thread is moving to the run-queue, set the lock
2749 	 * to the scheduler's lock.
2750 	 */
2751 	if (td->td_lock != TDQ_LOCKPTR(tdq)) {
2752 		TDQ_LOCK(tdq);
2753 		if ((flags & SRQ_HOLD) != 0)
2754 			td->td_lock = TDQ_LOCKPTR(tdq);
2755 		else
2756 			thread_lock_set(td, TDQ_LOCKPTR(tdq));
2757 	}
2758 	(void)tdq_add(tdq, td, flags);
2759 	if (!(flags & SRQ_YIELDING))
2760 		sched_setpreempt(td->td_priority);
2761 #endif
2762 	if (!(flags & SRQ_HOLDTD))
2763 		thread_unlock(td);
2764 }
2765 
2766 /*
2767  * Remove a thread from a run-queue without running it.  This is used
2768  * when we're stealing a thread from a remote queue.  Otherwise all threads
2769  * exit by calling sched_exit_thread() and sched_throw() themselves.
2770  */
2771 void
2772 sched_rem(struct thread *td)
2773 {
2774 	struct tdq *tdq;
2775 
2776 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2777 	    "prio:%d", td->td_priority);
2778 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2779 	tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2780 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2781 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2782 	KASSERT(TD_ON_RUNQ(td),
2783 	    ("sched_rem: thread not on run queue"));
2784 	tdq_runq_rem(tdq, td);
2785 	tdq_load_rem(tdq, td);
2786 	TD_SET_CAN_RUN(td);
2787 	if (td->td_priority == tdq->tdq_lowpri)
2788 		tdq_setlowpri(tdq, NULL);
2789 }
2790 
2791 /*
2792  * Fetch cpu utilization information.  Updates on demand.
2793  */
2794 fixpt_t
2795 sched_pctcpu(struct thread *td)
2796 {
2797 	fixpt_t pctcpu;
2798 	struct td_sched *ts;
2799 
2800 	pctcpu = 0;
2801 	ts = td_get_sched(td);
2802 
2803 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2804 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2805 	if (ts->ts_ticks) {
2806 		int rtick;
2807 
2808 		/* How many rtick per second ? */
2809 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2810 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2811 	}
2812 
2813 	return (pctcpu);
2814 }
2815 
2816 /*
2817  * Enforce affinity settings for a thread.  Called after adjustments to
2818  * cpumask.
2819  */
2820 void
2821 sched_affinity(struct thread *td)
2822 {
2823 #ifdef SMP
2824 	struct td_sched *ts;
2825 
2826 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2827 	ts = td_get_sched(td);
2828 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2829 		return;
2830 	if (TD_ON_RUNQ(td)) {
2831 		sched_rem(td);
2832 		sched_add(td, SRQ_BORING | SRQ_HOLDTD);
2833 		return;
2834 	}
2835 	if (!TD_IS_RUNNING(td))
2836 		return;
2837 	/*
2838 	 * Force a switch before returning to userspace.  If the
2839 	 * target thread is not running locally send an ipi to force
2840 	 * the issue.
2841 	 */
2842 	td->td_flags |= TDF_NEEDRESCHED;
2843 	if (td != curthread)
2844 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2845 #endif
2846 }
2847 
2848 /*
2849  * Bind a thread to a target cpu.
2850  */
2851 void
2852 sched_bind(struct thread *td, int cpu)
2853 {
2854 	struct td_sched *ts;
2855 
2856 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2857 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2858 	ts = td_get_sched(td);
2859 	if (ts->ts_flags & TSF_BOUND)
2860 		sched_unbind(td);
2861 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2862 	ts->ts_flags |= TSF_BOUND;
2863 	sched_pin();
2864 	if (PCPU_GET(cpuid) == cpu)
2865 		return;
2866 	ts->ts_cpu = cpu;
2867 	/* When we return from mi_switch we'll be on the correct cpu. */
2868 	mi_switch(SW_VOL);
2869 	thread_lock(td);
2870 }
2871 
2872 /*
2873  * Release a bound thread.
2874  */
2875 void
2876 sched_unbind(struct thread *td)
2877 {
2878 	struct td_sched *ts;
2879 
2880 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2881 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2882 	ts = td_get_sched(td);
2883 	if ((ts->ts_flags & TSF_BOUND) == 0)
2884 		return;
2885 	ts->ts_flags &= ~TSF_BOUND;
2886 	sched_unpin();
2887 }
2888 
2889 int
2890 sched_is_bound(struct thread *td)
2891 {
2892 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2893 	return (td_get_sched(td)->ts_flags & TSF_BOUND);
2894 }
2895 
2896 /*
2897  * Basic yield call.
2898  */
2899 void
2900 sched_relinquish(struct thread *td)
2901 {
2902 	thread_lock(td);
2903 	mi_switch(SW_VOL | SWT_RELINQUISH);
2904 }
2905 
2906 /*
2907  * Return the total system load.
2908  */
2909 int
2910 sched_load(void)
2911 {
2912 #ifdef SMP
2913 	int total;
2914 	int i;
2915 
2916 	total = 0;
2917 	CPU_FOREACH(i)
2918 		total += atomic_load_int(&TDQ_CPU(i)->tdq_sysload);
2919 	return (total);
2920 #else
2921 	return (atomic_load_int(&TDQ_SELF()->tdq_sysload));
2922 #endif
2923 }
2924 
2925 int
2926 sched_sizeof_proc(void)
2927 {
2928 	return (sizeof(struct proc));
2929 }
2930 
2931 int
2932 sched_sizeof_thread(void)
2933 {
2934 	return (sizeof(struct thread) + sizeof(struct td_sched));
2935 }
2936 
2937 #ifdef SMP
2938 #define	TDQ_IDLESPIN(tdq)						\
2939     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2940 #else
2941 #define	TDQ_IDLESPIN(tdq)	1
2942 #endif
2943 
2944 /*
2945  * The actual idle process.
2946  */
2947 void
2948 sched_idletd(void *dummy)
2949 {
2950 	struct thread *td;
2951 	struct tdq *tdq;
2952 	int oldswitchcnt, switchcnt;
2953 	int i;
2954 
2955 	mtx_assert(&Giant, MA_NOTOWNED);
2956 	td = curthread;
2957 	tdq = TDQ_SELF();
2958 	THREAD_NO_SLEEPING();
2959 	oldswitchcnt = -1;
2960 	for (;;) {
2961 		if (TDQ_LOAD(tdq)) {
2962 			thread_lock(td);
2963 			mi_switch(SW_VOL | SWT_IDLE);
2964 		}
2965 		switchcnt = TDQ_SWITCHCNT(tdq);
2966 #ifdef SMP
2967 		if (always_steal || switchcnt != oldswitchcnt) {
2968 			oldswitchcnt = switchcnt;
2969 			if (tdq_idled(tdq) == 0)
2970 				continue;
2971 		}
2972 		switchcnt = TDQ_SWITCHCNT(tdq);
2973 #else
2974 		oldswitchcnt = switchcnt;
2975 #endif
2976 		/*
2977 		 * If we're switching very frequently, spin while checking
2978 		 * for load rather than entering a low power state that
2979 		 * may require an IPI.  However, don't do any busy
2980 		 * loops while on SMT machines as this simply steals
2981 		 * cycles from cores doing useful work.
2982 		 */
2983 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2984 			for (i = 0; i < sched_idlespins; i++) {
2985 				if (TDQ_LOAD(tdq))
2986 					break;
2987 				cpu_spinwait();
2988 			}
2989 		}
2990 
2991 		/* If there was context switch during spin, restart it. */
2992 		switchcnt = TDQ_SWITCHCNT(tdq);
2993 		if (TDQ_LOAD(tdq) != 0 || switchcnt != oldswitchcnt)
2994 			continue;
2995 
2996 		/* Run main MD idle handler. */
2997 		atomic_store_int(&tdq->tdq_cpu_idle, 1);
2998 		/*
2999 		 * Make sure that the tdq_cpu_idle update is globally visible
3000 		 * before cpu_idle() reads tdq_load.  The order is important
3001 		 * to avoid races with tdq_notify().
3002 		 */
3003 		atomic_thread_fence_seq_cst();
3004 		/*
3005 		 * Checking for again after the fence picks up assigned
3006 		 * threads often enough to make it worthwhile to do so in
3007 		 * order to avoid calling cpu_idle().
3008 		 */
3009 		if (TDQ_LOAD(tdq) != 0) {
3010 			atomic_store_int(&tdq->tdq_cpu_idle, 0);
3011 			continue;
3012 		}
3013 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
3014 		atomic_store_int(&tdq->tdq_cpu_idle, 0);
3015 
3016 		/*
3017 		 * Account thread-less hardware interrupts and
3018 		 * other wakeup reasons equal to context switches.
3019 		 */
3020 		switchcnt = TDQ_SWITCHCNT(tdq);
3021 		if (switchcnt != oldswitchcnt)
3022 			continue;
3023 		TDQ_SWITCHCNT_INC(tdq);
3024 		oldswitchcnt++;
3025 	}
3026 }
3027 
3028 /*
3029  * sched_throw_grab() chooses a thread from the queue to switch to
3030  * next.  It returns with the tdq lock dropped in a spinlock section to
3031  * keep interrupts disabled until the CPU is running in a proper threaded
3032  * context.
3033  */
3034 static struct thread *
3035 sched_throw_grab(struct tdq *tdq)
3036 {
3037 	struct thread *newtd;
3038 
3039 	newtd = choosethread();
3040 	spinlock_enter();
3041 	TDQ_UNLOCK(tdq);
3042 	KASSERT(curthread->td_md.md_spinlock_count == 1,
3043 	    ("invalid count %d", curthread->td_md.md_spinlock_count));
3044 	return (newtd);
3045 }
3046 
3047 /*
3048  * A CPU is entering for the first time.
3049  */
3050 void
3051 sched_ap_entry(void)
3052 {
3053 	struct thread *newtd;
3054 	struct tdq *tdq;
3055 
3056 	tdq = TDQ_SELF();
3057 
3058 	/* This should have been setup in schedinit_ap(). */
3059 	THREAD_LOCKPTR_ASSERT(curthread, TDQ_LOCKPTR(tdq));
3060 
3061 	TDQ_LOCK(tdq);
3062 	/* Correct spinlock nesting. */
3063 	spinlock_exit();
3064 	PCPU_SET(switchtime, cpu_ticks());
3065 	PCPU_SET(switchticks, ticks);
3066 
3067 	newtd = sched_throw_grab(tdq);
3068 
3069 	/* doesn't return */
3070 	cpu_throw(NULL, newtd);
3071 }
3072 
3073 /*
3074  * A thread is exiting.
3075  */
3076 void
3077 sched_throw(struct thread *td)
3078 {
3079 	struct thread *newtd;
3080 	struct tdq *tdq;
3081 
3082 	tdq = TDQ_SELF();
3083 
3084 	MPASS(td != NULL);
3085 	THREAD_LOCK_ASSERT(td, MA_OWNED);
3086 	THREAD_LOCKPTR_ASSERT(td, TDQ_LOCKPTR(tdq));
3087 
3088 	tdq_load_rem(tdq, td);
3089 	td->td_lastcpu = td->td_oncpu;
3090 	td->td_oncpu = NOCPU;
3091 	thread_lock_block(td);
3092 
3093 	newtd = sched_throw_grab(tdq);
3094 
3095 	/* doesn't return */
3096 	cpu_switch(td, newtd, TDQ_LOCKPTR(tdq));
3097 }
3098 
3099 /*
3100  * This is called from fork_exit().  Just acquire the correct locks and
3101  * let fork do the rest of the work.
3102  */
3103 void
3104 sched_fork_exit(struct thread *td)
3105 {
3106 	struct tdq *tdq;
3107 	int cpuid;
3108 
3109 	/*
3110 	 * Finish setting up thread glue so that it begins execution in a
3111 	 * non-nested critical section with the scheduler lock held.
3112 	 */
3113 	KASSERT(curthread->td_md.md_spinlock_count == 1,
3114 	    ("invalid count %d", curthread->td_md.md_spinlock_count));
3115 	cpuid = PCPU_GET(cpuid);
3116 	tdq = TDQ_SELF();
3117 	TDQ_LOCK(tdq);
3118 	spinlock_exit();
3119 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
3120 	td->td_oncpu = cpuid;
3121 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
3122 	    "prio:%d", td->td_priority);
3123 	SDT_PROBE0(sched, , , on__cpu);
3124 }
3125 
3126 /*
3127  * Create on first use to catch odd startup conditions.
3128  */
3129 char *
3130 sched_tdname(struct thread *td)
3131 {
3132 #ifdef KTR
3133 	struct td_sched *ts;
3134 
3135 	ts = td_get_sched(td);
3136 	if (ts->ts_name[0] == '\0')
3137 		snprintf(ts->ts_name, sizeof(ts->ts_name),
3138 		    "%s tid %d", td->td_name, td->td_tid);
3139 	return (ts->ts_name);
3140 #else
3141 	return (td->td_name);
3142 #endif
3143 }
3144 
3145 #ifdef KTR
3146 void
3147 sched_clear_tdname(struct thread *td)
3148 {
3149 	struct td_sched *ts;
3150 
3151 	ts = td_get_sched(td);
3152 	ts->ts_name[0] = '\0';
3153 }
3154 #endif
3155 
3156 #ifdef SMP
3157 
3158 /*
3159  * Build the CPU topology dump string. Is recursively called to collect
3160  * the topology tree.
3161  */
3162 static int
3163 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
3164     int indent)
3165 {
3166 	char cpusetbuf[CPUSETBUFSIZ];
3167 	int i, first;
3168 
3169 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
3170 	    "", 1 + indent / 2, cg->cg_level);
3171 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
3172 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
3173 	first = TRUE;
3174 	for (i = cg->cg_first; i <= cg->cg_last; i++) {
3175 		if (CPU_ISSET(i, &cg->cg_mask)) {
3176 			if (!first)
3177 				sbuf_printf(sb, ", ");
3178 			else
3179 				first = FALSE;
3180 			sbuf_printf(sb, "%d", i);
3181 		}
3182 	}
3183 	sbuf_printf(sb, "</cpu>\n");
3184 
3185 	if (cg->cg_flags != 0) {
3186 		sbuf_printf(sb, "%*s <flags>", indent, "");
3187 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
3188 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
3189 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
3190 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
3191 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
3192 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
3193 		if ((cg->cg_flags & CG_FLAG_NODE) != 0)
3194 			sbuf_printf(sb, "<flag name=\"NODE\">NUMA node</flag>");
3195 		sbuf_printf(sb, "</flags>\n");
3196 	}
3197 
3198 	if (cg->cg_children > 0) {
3199 		sbuf_printf(sb, "%*s <children>\n", indent, "");
3200 		for (i = 0; i < cg->cg_children; i++)
3201 			sysctl_kern_sched_topology_spec_internal(sb,
3202 			    &cg->cg_child[i], indent+2);
3203 		sbuf_printf(sb, "%*s </children>\n", indent, "");
3204 	}
3205 	sbuf_printf(sb, "%*s</group>\n", indent, "");
3206 	return (0);
3207 }
3208 
3209 /*
3210  * Sysctl handler for retrieving topology dump. It's a wrapper for
3211  * the recursive sysctl_kern_smp_topology_spec_internal().
3212  */
3213 static int
3214 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
3215 {
3216 	struct sbuf *topo;
3217 	int err;
3218 
3219 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
3220 
3221 	topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
3222 	if (topo == NULL)
3223 		return (ENOMEM);
3224 
3225 	sbuf_printf(topo, "<groups>\n");
3226 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
3227 	sbuf_printf(topo, "</groups>\n");
3228 
3229 	if (err == 0) {
3230 		err = sbuf_finish(topo);
3231 	}
3232 	sbuf_delete(topo);
3233 	return (err);
3234 }
3235 
3236 #endif
3237 
3238 static int
3239 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
3240 {
3241 	int error, new_val, period;
3242 
3243 	period = 1000000 / realstathz;
3244 	new_val = period * sched_slice;
3245 	error = sysctl_handle_int(oidp, &new_val, 0, req);
3246 	if (error != 0 || req->newptr == NULL)
3247 		return (error);
3248 	if (new_val <= 0)
3249 		return (EINVAL);
3250 	sched_slice = imax(1, (new_val + period / 2) / period);
3251 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
3252 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
3253 	    realstathz);
3254 	return (0);
3255 }
3256 
3257 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
3258     "Scheduler");
3259 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
3260     "Scheduler name");
3261 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum,
3262     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0,
3263     sysctl_kern_quantum, "I",
3264     "Quantum for timeshare threads in microseconds");
3265 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
3266     "Quantum for timeshare threads in stathz ticks");
3267 SYSCTL_UINT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
3268     "Interactivity score threshold");
3269 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
3270     &preempt_thresh, 0,
3271     "Maximal (lowest) priority for preemption");
3272 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
3273     "Assign static kernel priorities to sleeping threads");
3274 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
3275     "Number of times idle thread will spin waiting for new work");
3276 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
3277     &sched_idlespinthresh, 0,
3278     "Threshold before we will permit idle thread spinning");
3279 #ifdef SMP
3280 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
3281     "Number of hz ticks to keep thread affinity for");
3282 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
3283     "Enables the long-term load balancer");
3284 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
3285     &balance_interval, 0,
3286     "Average period in stathz ticks to run the long-term balancer");
3287 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
3288     "Attempts to steal work from other cores before idling");
3289 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
3290     "Minimum load on remote CPU before we'll steal");
3291 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit,
3292     0, "Topological distance limit for stealing threads in sched_switch()");
3293 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0,
3294     "Always run the stealer from the idle thread");
3295 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
3296     CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
3297     "XML dump of detected CPU topology");
3298 #endif
3299 
3300 /* ps compat.  All cpu percentages from ULE are weighted. */
3301 static int ccpu = 0;
3302 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0,
3303     "Decay factor used for updating %CPU in 4BSD scheduler");
3304