xref: /freebsd/sys/kern/sched_ule.c (revision 686bcb5c14aba6e67524be84e125bfdd3514db9e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice unmodified, this list of conditions, and the following
12  *    disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 /*
30  * This file implements the ULE scheduler.  ULE supports independent CPU
31  * run queues and fine grain locking.  It has superior interactive
32  * performance under load even on uni-processor systems.
33  *
34  * etymology:
35  *   ULE is the last three letters in schedule.  It owes its name to a
36  * generic user created for a scheduling system by Paul Mikesell at
37  * Isilon Systems and a general lack of creativity on the part of the author.
38  */
39 
40 #include <sys/cdefs.h>
41 __FBSDID("$FreeBSD$");
42 
43 #include "opt_hwpmc_hooks.h"
44 #include "opt_sched.h"
45 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kdb.h>
49 #include <sys/kernel.h>
50 #include <sys/ktr.h>
51 #include <sys/limits.h>
52 #include <sys/lock.h>
53 #include <sys/mutex.h>
54 #include <sys/proc.h>
55 #include <sys/resource.h>
56 #include <sys/resourcevar.h>
57 #include <sys/sched.h>
58 #include <sys/sdt.h>
59 #include <sys/smp.h>
60 #include <sys/sx.h>
61 #include <sys/sysctl.h>
62 #include <sys/sysproto.h>
63 #include <sys/turnstile.h>
64 #include <sys/umtx.h>
65 #include <sys/vmmeter.h>
66 #include <sys/cpuset.h>
67 #include <sys/sbuf.h>
68 
69 #ifdef HWPMC_HOOKS
70 #include <sys/pmckern.h>
71 #endif
72 
73 #ifdef KDTRACE_HOOKS
74 #include <sys/dtrace_bsd.h>
75 int __read_mostly		dtrace_vtime_active;
76 dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
77 #endif
78 
79 #include <machine/cpu.h>
80 #include <machine/smp.h>
81 
82 #define	KTR_ULE	0
83 
84 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
85 #define	TDQ_NAME_LEN	(sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
86 #define	TDQ_LOADNAME_LEN	(sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
87 
88 /*
89  * Thread scheduler specific section.  All fields are protected
90  * by the thread lock.
91  */
92 struct td_sched {
93 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
94 	short		ts_flags;	/* TSF_* flags. */
95 	int		ts_cpu;		/* CPU that we have affinity for. */
96 	int		ts_rltick;	/* Real last tick, for affinity. */
97 	int		ts_slice;	/* Ticks of slice remaining. */
98 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
99 	u_int		ts_runtime;	/* Number of ticks we were running */
100 	int		ts_ltick;	/* Last tick that we were running on */
101 	int		ts_ftick;	/* First tick that we were running on */
102 	int		ts_ticks;	/* Tick count */
103 #ifdef KTR
104 	char		ts_name[TS_NAME_LEN];
105 #endif
106 };
107 /* flags kept in ts_flags */
108 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
109 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
110 
111 #define	THREAD_CAN_MIGRATE(td)	((td)->td_pinned == 0)
112 #define	THREAD_CAN_SCHED(td, cpu)	\
113     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
114 
115 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <=
116     sizeof(struct thread0_storage),
117     "increase struct thread0_storage.t0st_sched size");
118 
119 /*
120  * Priority ranges used for interactive and non-interactive timeshare
121  * threads.  The timeshare priorities are split up into four ranges.
122  * The first range handles interactive threads.  The last three ranges
123  * (NHALF, x, and NHALF) handle non-interactive threads with the outer
124  * ranges supporting nice values.
125  */
126 #define	PRI_TIMESHARE_RANGE	(PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
127 #define	PRI_INTERACT_RANGE	((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
128 #define	PRI_BATCH_RANGE		(PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
129 
130 #define	PRI_MIN_INTERACT	PRI_MIN_TIMESHARE
131 #define	PRI_MAX_INTERACT	(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
132 #define	PRI_MIN_BATCH		(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
133 #define	PRI_MAX_BATCH		PRI_MAX_TIMESHARE
134 
135 /*
136  * Cpu percentage computation macros and defines.
137  *
138  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
139  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
140  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
141  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
142  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
143  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
144  */
145 #define	SCHED_TICK_SECS		10
146 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
147 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
148 #define	SCHED_TICK_SHIFT	10
149 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
150 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
151 
152 /*
153  * These macros determine priorities for non-interactive threads.  They are
154  * assigned a priority based on their recent cpu utilization as expressed
155  * by the ratio of ticks to the tick total.  NHALF priorities at the start
156  * and end of the MIN to MAX timeshare range are only reachable with negative
157  * or positive nice respectively.
158  *
159  * PRI_RANGE:	Priority range for utilization dependent priorities.
160  * PRI_NRESV:	Number of nice values.
161  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
162  * PRI_NICE:	Determines the part of the priority inherited from nice.
163  */
164 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
165 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
166 #define	SCHED_PRI_MIN		(PRI_MIN_BATCH + SCHED_PRI_NHALF)
167 #define	SCHED_PRI_MAX		(PRI_MAX_BATCH - SCHED_PRI_NHALF)
168 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
169 #define	SCHED_PRI_TICKS(ts)						\
170     (SCHED_TICK_HZ((ts)) /						\
171     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
172 #define	SCHED_PRI_NICE(nice)	(nice)
173 
174 /*
175  * These determine the interactivity of a process.  Interactivity differs from
176  * cpu utilization in that it expresses the voluntary time slept vs time ran
177  * while cpu utilization includes all time not running.  This more accurately
178  * models the intent of the thread.
179  *
180  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
181  *		before throttling back.
182  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
183  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
184  * INTERACT_THRESH:	Threshold for placement on the current runq.
185  */
186 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
187 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
188 #define	SCHED_INTERACT_MAX	(100)
189 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
190 #define	SCHED_INTERACT_THRESH	(30)
191 
192 /*
193  * These parameters determine the slice behavior for batch work.
194  */
195 #define	SCHED_SLICE_DEFAULT_DIVISOR	10	/* ~94 ms, 12 stathz ticks. */
196 #define	SCHED_SLICE_MIN_DIVISOR		6	/* DEFAULT/MIN = ~16 ms. */
197 
198 /* Flags kept in td_flags. */
199 #define	TDF_SLICEEND	TDF_SCHED2	/* Thread time slice is over. */
200 
201 /*
202  * tickincr:		Converts a stathz tick into a hz domain scaled by
203  *			the shift factor.  Without the shift the error rate
204  *			due to rounding would be unacceptably high.
205  * realstathz:		stathz is sometimes 0 and run off of hz.
206  * sched_slice:		Runtime of each thread before rescheduling.
207  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
208  */
209 static int __read_mostly sched_interact = SCHED_INTERACT_THRESH;
210 static int __read_mostly tickincr = 8 << SCHED_TICK_SHIFT;
211 static int __read_mostly realstathz = 127;	/* reset during boot. */
212 static int __read_mostly sched_slice = 10;	/* reset during boot. */
213 static int __read_mostly sched_slice_min = 1;	/* reset during boot. */
214 #ifdef PREEMPTION
215 #ifdef FULL_PREEMPTION
216 static int __read_mostly preempt_thresh = PRI_MAX_IDLE;
217 #else
218 static int __read_mostly preempt_thresh = PRI_MIN_KERN;
219 #endif
220 #else
221 static int __read_mostly preempt_thresh = 0;
222 #endif
223 static int __read_mostly static_boost = PRI_MIN_BATCH;
224 static int __read_mostly sched_idlespins = 10000;
225 static int __read_mostly sched_idlespinthresh = -1;
226 
227 /*
228  * tdq - per processor runqs and statistics.  All fields are protected by the
229  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
230  * locking in sched_pickcpu();
231  */
232 struct tdq {
233 	/*
234 	 * Ordered to improve efficiency of cpu_search() and switch().
235 	 * tdq_lock is padded to avoid false sharing with tdq_load and
236 	 * tdq_cpu_idle.
237 	 */
238 	struct mtx_padalign tdq_lock;		/* run queue lock. */
239 	struct cpu_group *tdq_cg;		/* Pointer to cpu topology. */
240 	volatile int	tdq_load;		/* Aggregate load. */
241 	volatile int	tdq_cpu_idle;		/* cpu_idle() is active. */
242 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
243 	volatile int	tdq_transferable;	/* Transferable thread count. */
244 	volatile short	tdq_switchcnt;		/* Switches this tick. */
245 	volatile short	tdq_oldswitchcnt;	/* Switches last tick. */
246 	u_char		tdq_lowpri;		/* Lowest priority thread. */
247 	u_char		tdq_owepreempt;		/* Remote preemption pending. */
248 	u_char		tdq_idx;		/* Current insert index. */
249 	u_char		tdq_ridx;		/* Current removal index. */
250 	int		tdq_id;			/* cpuid. */
251 	struct runq	tdq_realtime;		/* real-time run queue. */
252 	struct runq	tdq_timeshare;		/* timeshare run queue. */
253 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
254 	char		tdq_name[TDQ_NAME_LEN];
255 #ifdef KTR
256 	char		tdq_loadname[TDQ_LOADNAME_LEN];
257 #endif
258 } __aligned(64);
259 
260 /* Idle thread states and config. */
261 #define	TDQ_RUNNING	1
262 #define	TDQ_IDLE	2
263 
264 #ifdef SMP
265 struct cpu_group __read_mostly *cpu_top;		/* CPU topology */
266 
267 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
268 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
269 
270 /*
271  * Run-time tunables.
272  */
273 static int rebalance = 1;
274 static int balance_interval = 128;	/* Default set in sched_initticks(). */
275 static int __read_mostly affinity;
276 static int __read_mostly steal_idle = 1;
277 static int __read_mostly steal_thresh = 2;
278 static int __read_mostly always_steal = 0;
279 static int __read_mostly trysteal_limit = 2;
280 
281 /*
282  * One thread queue per processor.
283  */
284 static struct tdq __read_mostly *balance_tdq;
285 static int balance_ticks;
286 DPCPU_DEFINE_STATIC(struct tdq, tdq);
287 DPCPU_DEFINE_STATIC(uint32_t, randomval);
288 
289 #define	TDQ_SELF()	((struct tdq *)PCPU_GET(sched))
290 #define	TDQ_CPU(x)	(DPCPU_ID_PTR((x), tdq))
291 #define	TDQ_ID(x)	((x)->tdq_id)
292 #else	/* !SMP */
293 static struct tdq	tdq_cpu;
294 
295 #define	TDQ_ID(x)	(0)
296 #define	TDQ_SELF()	(&tdq_cpu)
297 #define	TDQ_CPU(x)	(&tdq_cpu)
298 #endif
299 
300 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
301 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
302 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
303 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
304 #define	TDQ_LOCKPTR(t)		((struct mtx *)(&(t)->tdq_lock))
305 
306 static void sched_priority(struct thread *);
307 static void sched_thread_priority(struct thread *, u_char);
308 static int sched_interact_score(struct thread *);
309 static void sched_interact_update(struct thread *);
310 static void sched_interact_fork(struct thread *);
311 static void sched_pctcpu_update(struct td_sched *, int);
312 
313 /* Operations on per processor queues */
314 static struct thread *tdq_choose(struct tdq *);
315 static void tdq_setup(struct tdq *, int i);
316 static void tdq_load_add(struct tdq *, struct thread *);
317 static void tdq_load_rem(struct tdq *, struct thread *);
318 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
319 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
320 static inline int sched_shouldpreempt(int, int, int);
321 void tdq_print(int cpu);
322 static void runq_print(struct runq *rq);
323 static void tdq_add(struct tdq *, struct thread *, int);
324 #ifdef SMP
325 static struct thread *tdq_move(struct tdq *, struct tdq *);
326 static int tdq_idled(struct tdq *);
327 static void tdq_notify(struct tdq *, struct thread *);
328 static struct thread *tdq_steal(struct tdq *, int);
329 static struct thread *runq_steal(struct runq *, int);
330 static int sched_pickcpu(struct thread *, int);
331 static void sched_balance(void);
332 static int sched_balance_pair(struct tdq *, struct tdq *);
333 static inline struct tdq *sched_setcpu(struct thread *, int, int);
334 static inline void thread_unblock_switch(struct thread *, struct mtx *);
335 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
336 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
337     struct cpu_group *cg, int indent);
338 #endif
339 
340 static void sched_setup(void *dummy);
341 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
342 
343 static void sched_initticks(void *dummy);
344 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
345     NULL);
346 
347 SDT_PROVIDER_DEFINE(sched);
348 
349 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
350     "struct proc *", "uint8_t");
351 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
352     "struct proc *", "void *");
353 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
354     "struct proc *", "void *", "int");
355 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
356     "struct proc *", "uint8_t", "struct thread *");
357 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
358 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
359     "struct proc *");
360 SDT_PROBE_DEFINE(sched, , , on__cpu);
361 SDT_PROBE_DEFINE(sched, , , remain__cpu);
362 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
363     "struct proc *");
364 
365 /*
366  * Print the threads waiting on a run-queue.
367  */
368 static void
369 runq_print(struct runq *rq)
370 {
371 	struct rqhead *rqh;
372 	struct thread *td;
373 	int pri;
374 	int j;
375 	int i;
376 
377 	for (i = 0; i < RQB_LEN; i++) {
378 		printf("\t\trunq bits %d 0x%zx\n",
379 		    i, rq->rq_status.rqb_bits[i]);
380 		for (j = 0; j < RQB_BPW; j++)
381 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
382 				pri = j + (i << RQB_L2BPW);
383 				rqh = &rq->rq_queues[pri];
384 				TAILQ_FOREACH(td, rqh, td_runq) {
385 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
386 					    td, td->td_name, td->td_priority,
387 					    td->td_rqindex, pri);
388 				}
389 			}
390 	}
391 }
392 
393 /*
394  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
395  */
396 void
397 tdq_print(int cpu)
398 {
399 	struct tdq *tdq;
400 
401 	tdq = TDQ_CPU(cpu);
402 
403 	printf("tdq %d:\n", TDQ_ID(tdq));
404 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
405 	printf("\tLock name:      %s\n", tdq->tdq_name);
406 	printf("\tload:           %d\n", tdq->tdq_load);
407 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
408 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
409 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
410 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
411 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
412 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
413 	printf("\trealtime runq:\n");
414 	runq_print(&tdq->tdq_realtime);
415 	printf("\ttimeshare runq:\n");
416 	runq_print(&tdq->tdq_timeshare);
417 	printf("\tidle runq:\n");
418 	runq_print(&tdq->tdq_idle);
419 }
420 
421 static inline int
422 sched_shouldpreempt(int pri, int cpri, int remote)
423 {
424 	/*
425 	 * If the new priority is not better than the current priority there is
426 	 * nothing to do.
427 	 */
428 	if (pri >= cpri)
429 		return (0);
430 	/*
431 	 * Always preempt idle.
432 	 */
433 	if (cpri >= PRI_MIN_IDLE)
434 		return (1);
435 	/*
436 	 * If preemption is disabled don't preempt others.
437 	 */
438 	if (preempt_thresh == 0)
439 		return (0);
440 	/*
441 	 * Preempt if we exceed the threshold.
442 	 */
443 	if (pri <= preempt_thresh)
444 		return (1);
445 	/*
446 	 * If we're interactive or better and there is non-interactive
447 	 * or worse running preempt only remote processors.
448 	 */
449 	if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
450 		return (1);
451 	return (0);
452 }
453 
454 /*
455  * Add a thread to the actual run-queue.  Keeps transferable counts up to
456  * date with what is actually on the run-queue.  Selects the correct
457  * queue position for timeshare threads.
458  */
459 static __inline void
460 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
461 {
462 	struct td_sched *ts;
463 	u_char pri;
464 
465 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
466 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
467 
468 	pri = td->td_priority;
469 	ts = td_get_sched(td);
470 	TD_SET_RUNQ(td);
471 	if (THREAD_CAN_MIGRATE(td)) {
472 		tdq->tdq_transferable++;
473 		ts->ts_flags |= TSF_XFERABLE;
474 	}
475 	if (pri < PRI_MIN_BATCH) {
476 		ts->ts_runq = &tdq->tdq_realtime;
477 	} else if (pri <= PRI_MAX_BATCH) {
478 		ts->ts_runq = &tdq->tdq_timeshare;
479 		KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
480 			("Invalid priority %d on timeshare runq", pri));
481 		/*
482 		 * This queue contains only priorities between MIN and MAX
483 		 * realtime.  Use the whole queue to represent these values.
484 		 */
485 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
486 			pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
487 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
488 			/*
489 			 * This effectively shortens the queue by one so we
490 			 * can have a one slot difference between idx and
491 			 * ridx while we wait for threads to drain.
492 			 */
493 			if (tdq->tdq_ridx != tdq->tdq_idx &&
494 			    pri == tdq->tdq_ridx)
495 				pri = (unsigned char)(pri - 1) % RQ_NQS;
496 		} else
497 			pri = tdq->tdq_ridx;
498 		runq_add_pri(ts->ts_runq, td, pri, flags);
499 		return;
500 	} else
501 		ts->ts_runq = &tdq->tdq_idle;
502 	runq_add(ts->ts_runq, td, flags);
503 }
504 
505 /*
506  * Remove a thread from a run-queue.  This typically happens when a thread
507  * is selected to run.  Running threads are not on the queue and the
508  * transferable count does not reflect them.
509  */
510 static __inline void
511 tdq_runq_rem(struct tdq *tdq, struct thread *td)
512 {
513 	struct td_sched *ts;
514 
515 	ts = td_get_sched(td);
516 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
517 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
518 	KASSERT(ts->ts_runq != NULL,
519 	    ("tdq_runq_remove: thread %p null ts_runq", td));
520 	if (ts->ts_flags & TSF_XFERABLE) {
521 		tdq->tdq_transferable--;
522 		ts->ts_flags &= ~TSF_XFERABLE;
523 	}
524 	if (ts->ts_runq == &tdq->tdq_timeshare) {
525 		if (tdq->tdq_idx != tdq->tdq_ridx)
526 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
527 		else
528 			runq_remove_idx(ts->ts_runq, td, NULL);
529 	} else
530 		runq_remove(ts->ts_runq, td);
531 }
532 
533 /*
534  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
535  * for this thread to the referenced thread queue.
536  */
537 static void
538 tdq_load_add(struct tdq *tdq, struct thread *td)
539 {
540 
541 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
542 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
543 
544 	tdq->tdq_load++;
545 	if ((td->td_flags & TDF_NOLOAD) == 0)
546 		tdq->tdq_sysload++;
547 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
548 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
549 }
550 
551 /*
552  * Remove the load from a thread that is transitioning to a sleep state or
553  * exiting.
554  */
555 static void
556 tdq_load_rem(struct tdq *tdq, struct thread *td)
557 {
558 
559 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
560 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
561 	KASSERT(tdq->tdq_load != 0,
562 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
563 
564 	tdq->tdq_load--;
565 	if ((td->td_flags & TDF_NOLOAD) == 0)
566 		tdq->tdq_sysload--;
567 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
568 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
569 }
570 
571 /*
572  * Bound timeshare latency by decreasing slice size as load increases.  We
573  * consider the maximum latency as the sum of the threads waiting to run
574  * aside from curthread and target no more than sched_slice latency but
575  * no less than sched_slice_min runtime.
576  */
577 static inline int
578 tdq_slice(struct tdq *tdq)
579 {
580 	int load;
581 
582 	/*
583 	 * It is safe to use sys_load here because this is called from
584 	 * contexts where timeshare threads are running and so there
585 	 * cannot be higher priority load in the system.
586 	 */
587 	load = tdq->tdq_sysload - 1;
588 	if (load >= SCHED_SLICE_MIN_DIVISOR)
589 		return (sched_slice_min);
590 	if (load <= 1)
591 		return (sched_slice);
592 	return (sched_slice / load);
593 }
594 
595 /*
596  * Set lowpri to its exact value by searching the run-queue and
597  * evaluating curthread.  curthread may be passed as an optimization.
598  */
599 static void
600 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
601 {
602 	struct thread *td;
603 
604 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
605 	if (ctd == NULL)
606 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
607 	td = tdq_choose(tdq);
608 	if (td == NULL || td->td_priority > ctd->td_priority)
609 		tdq->tdq_lowpri = ctd->td_priority;
610 	else
611 		tdq->tdq_lowpri = td->td_priority;
612 }
613 
614 #ifdef SMP
615 /*
616  * We need some randomness. Implement a classic Linear Congruential
617  * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
618  * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
619  * of the random state (in the low bits of our answer) to keep
620  * the maximum randomness.
621  */
622 static uint32_t
623 sched_random(void)
624 {
625 	uint32_t *rndptr;
626 
627 	rndptr = DPCPU_PTR(randomval);
628 	*rndptr = *rndptr * 69069 + 5;
629 
630 	return (*rndptr >> 16);
631 }
632 
633 struct cpu_search {
634 	cpuset_t cs_mask;
635 	u_int	cs_prefer;
636 	int	cs_pri;		/* Min priority for low. */
637 	int	cs_limit;	/* Max load for low, min load for high. */
638 	int	cs_cpu;
639 	int	cs_load;
640 };
641 
642 #define	CPU_SEARCH_LOWEST	0x1
643 #define	CPU_SEARCH_HIGHEST	0x2
644 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
645 
646 static __always_inline int cpu_search(const struct cpu_group *cg,
647     struct cpu_search *low, struct cpu_search *high, const int match);
648 int __noinline cpu_search_lowest(const struct cpu_group *cg,
649     struct cpu_search *low);
650 int __noinline cpu_search_highest(const struct cpu_group *cg,
651     struct cpu_search *high);
652 int __noinline cpu_search_both(const struct cpu_group *cg,
653     struct cpu_search *low, struct cpu_search *high);
654 
655 /*
656  * Search the tree of cpu_groups for the lowest or highest loaded cpu
657  * according to the match argument.  This routine actually compares the
658  * load on all paths through the tree and finds the least loaded cpu on
659  * the least loaded path, which may differ from the least loaded cpu in
660  * the system.  This balances work among caches and buses.
661  *
662  * This inline is instantiated in three forms below using constants for the
663  * match argument.  It is reduced to the minimum set for each case.  It is
664  * also recursive to the depth of the tree.
665  */
666 static __always_inline int
667 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
668     struct cpu_search *high, const int match)
669 {
670 	struct cpu_search lgroup;
671 	struct cpu_search hgroup;
672 	cpuset_t cpumask;
673 	struct cpu_group *child;
674 	struct tdq *tdq;
675 	int cpu, i, hload, lload, load, total, rnd;
676 
677 	total = 0;
678 	cpumask = cg->cg_mask;
679 	if (match & CPU_SEARCH_LOWEST) {
680 		lload = INT_MAX;
681 		lgroup = *low;
682 	}
683 	if (match & CPU_SEARCH_HIGHEST) {
684 		hload = INT_MIN;
685 		hgroup = *high;
686 	}
687 
688 	/* Iterate through the child CPU groups and then remaining CPUs. */
689 	for (i = cg->cg_children, cpu = mp_maxid; ; ) {
690 		if (i == 0) {
691 #ifdef HAVE_INLINE_FFSL
692 			cpu = CPU_FFS(&cpumask) - 1;
693 #else
694 			while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
695 				cpu--;
696 #endif
697 			if (cpu < 0)
698 				break;
699 			child = NULL;
700 		} else
701 			child = &cg->cg_child[i - 1];
702 
703 		if (match & CPU_SEARCH_LOWEST)
704 			lgroup.cs_cpu = -1;
705 		if (match & CPU_SEARCH_HIGHEST)
706 			hgroup.cs_cpu = -1;
707 		if (child) {			/* Handle child CPU group. */
708 			CPU_ANDNOT(&cpumask, &child->cg_mask);
709 			switch (match) {
710 			case CPU_SEARCH_LOWEST:
711 				load = cpu_search_lowest(child, &lgroup);
712 				break;
713 			case CPU_SEARCH_HIGHEST:
714 				load = cpu_search_highest(child, &hgroup);
715 				break;
716 			case CPU_SEARCH_BOTH:
717 				load = cpu_search_both(child, &lgroup, &hgroup);
718 				break;
719 			}
720 		} else {			/* Handle child CPU. */
721 			CPU_CLR(cpu, &cpumask);
722 			tdq = TDQ_CPU(cpu);
723 			load = tdq->tdq_load * 256;
724 			rnd = sched_random() % 32;
725 			if (match & CPU_SEARCH_LOWEST) {
726 				if (cpu == low->cs_prefer)
727 					load -= 64;
728 				/* If that CPU is allowed and get data. */
729 				if (tdq->tdq_lowpri > lgroup.cs_pri &&
730 				    tdq->tdq_load <= lgroup.cs_limit &&
731 				    CPU_ISSET(cpu, &lgroup.cs_mask)) {
732 					lgroup.cs_cpu = cpu;
733 					lgroup.cs_load = load - rnd;
734 				}
735 			}
736 			if (match & CPU_SEARCH_HIGHEST)
737 				if (tdq->tdq_load >= hgroup.cs_limit &&
738 				    tdq->tdq_transferable &&
739 				    CPU_ISSET(cpu, &hgroup.cs_mask)) {
740 					hgroup.cs_cpu = cpu;
741 					hgroup.cs_load = load - rnd;
742 				}
743 		}
744 		total += load;
745 
746 		/* We have info about child item. Compare it. */
747 		if (match & CPU_SEARCH_LOWEST) {
748 			if (lgroup.cs_cpu >= 0 &&
749 			    (load < lload ||
750 			     (load == lload && lgroup.cs_load < low->cs_load))) {
751 				lload = load;
752 				low->cs_cpu = lgroup.cs_cpu;
753 				low->cs_load = lgroup.cs_load;
754 			}
755 		}
756 		if (match & CPU_SEARCH_HIGHEST)
757 			if (hgroup.cs_cpu >= 0 &&
758 			    (load > hload ||
759 			     (load == hload && hgroup.cs_load > high->cs_load))) {
760 				hload = load;
761 				high->cs_cpu = hgroup.cs_cpu;
762 				high->cs_load = hgroup.cs_load;
763 			}
764 		if (child) {
765 			i--;
766 			if (i == 0 && CPU_EMPTY(&cpumask))
767 				break;
768 		}
769 #ifndef HAVE_INLINE_FFSL
770 		else
771 			cpu--;
772 #endif
773 	}
774 	return (total);
775 }
776 
777 /*
778  * cpu_search instantiations must pass constants to maintain the inline
779  * optimization.
780  */
781 int
782 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
783 {
784 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
785 }
786 
787 int
788 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
789 {
790 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
791 }
792 
793 int
794 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
795     struct cpu_search *high)
796 {
797 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
798 }
799 
800 /*
801  * Find the cpu with the least load via the least loaded path that has a
802  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
803  * acceptable.
804  */
805 static inline int
806 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
807     int prefer)
808 {
809 	struct cpu_search low;
810 
811 	low.cs_cpu = -1;
812 	low.cs_prefer = prefer;
813 	low.cs_mask = mask;
814 	low.cs_pri = pri;
815 	low.cs_limit = maxload;
816 	cpu_search_lowest(cg, &low);
817 	return low.cs_cpu;
818 }
819 
820 /*
821  * Find the cpu with the highest load via the highest loaded path.
822  */
823 static inline int
824 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
825 {
826 	struct cpu_search high;
827 
828 	high.cs_cpu = -1;
829 	high.cs_mask = mask;
830 	high.cs_limit = minload;
831 	cpu_search_highest(cg, &high);
832 	return high.cs_cpu;
833 }
834 
835 static void
836 sched_balance_group(struct cpu_group *cg)
837 {
838 	struct tdq *tdq;
839 	cpuset_t hmask, lmask;
840 	int high, low, anylow;
841 
842 	CPU_FILL(&hmask);
843 	for (;;) {
844 		high = sched_highest(cg, hmask, 2);
845 		/* Stop if there is no more CPU with transferrable threads. */
846 		if (high == -1)
847 			break;
848 		CPU_CLR(high, &hmask);
849 		CPU_COPY(&hmask, &lmask);
850 		/* Stop if there is no more CPU left for low. */
851 		if (CPU_EMPTY(&lmask))
852 			break;
853 		anylow = 1;
854 		tdq = TDQ_CPU(high);
855 nextlow:
856 		low = sched_lowest(cg, lmask, -1, tdq->tdq_load - 1, high);
857 		/* Stop if we looked well and found no less loaded CPU. */
858 		if (anylow && low == -1)
859 			break;
860 		/* Go to next high if we found no less loaded CPU. */
861 		if (low == -1)
862 			continue;
863 		/* Transfer thread from high to low. */
864 		if (sched_balance_pair(tdq, TDQ_CPU(low))) {
865 			/* CPU that got thread can no longer be a donor. */
866 			CPU_CLR(low, &hmask);
867 		} else {
868 			/*
869 			 * If failed, then there is no threads on high
870 			 * that can run on this low. Drop low from low
871 			 * mask and look for different one.
872 			 */
873 			CPU_CLR(low, &lmask);
874 			anylow = 0;
875 			goto nextlow;
876 		}
877 	}
878 }
879 
880 static void
881 sched_balance(void)
882 {
883 	struct tdq *tdq;
884 
885 	balance_ticks = max(balance_interval / 2, 1) +
886 	    (sched_random() % balance_interval);
887 	tdq = TDQ_SELF();
888 	TDQ_UNLOCK(tdq);
889 	sched_balance_group(cpu_top);
890 	TDQ_LOCK(tdq);
891 }
892 
893 /*
894  * Lock two thread queues using their address to maintain lock order.
895  */
896 static void
897 tdq_lock_pair(struct tdq *one, struct tdq *two)
898 {
899 	if (one < two) {
900 		TDQ_LOCK(one);
901 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
902 	} else {
903 		TDQ_LOCK(two);
904 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
905 	}
906 }
907 
908 /*
909  * Unlock two thread queues.  Order is not important here.
910  */
911 static void
912 tdq_unlock_pair(struct tdq *one, struct tdq *two)
913 {
914 	TDQ_UNLOCK(one);
915 	TDQ_UNLOCK(two);
916 }
917 
918 /*
919  * Transfer load between two imbalanced thread queues.
920  */
921 static int
922 sched_balance_pair(struct tdq *high, struct tdq *low)
923 {
924 	struct thread *td;
925 	int cpu;
926 
927 	tdq_lock_pair(high, low);
928 	td = NULL;
929 	/*
930 	 * Transfer a thread from high to low.
931 	 */
932 	if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
933 	    (td = tdq_move(high, low)) != NULL) {
934 		/*
935 		 * In case the target isn't the current cpu notify it of the
936 		 * new load, possibly sending an IPI to force it to reschedule.
937 		 */
938 		cpu = TDQ_ID(low);
939 		if (cpu != PCPU_GET(cpuid))
940 			tdq_notify(low, td);
941 	}
942 	tdq_unlock_pair(high, low);
943 	return (td != NULL);
944 }
945 
946 /*
947  * Move a thread from one thread queue to another.
948  */
949 static struct thread *
950 tdq_move(struct tdq *from, struct tdq *to)
951 {
952 	struct thread *td;
953 	struct tdq *tdq;
954 	int cpu;
955 
956 	TDQ_LOCK_ASSERT(from, MA_OWNED);
957 	TDQ_LOCK_ASSERT(to, MA_OWNED);
958 
959 	tdq = from;
960 	cpu = TDQ_ID(to);
961 	td = tdq_steal(tdq, cpu);
962 	if (td == NULL)
963 		return (NULL);
964 
965 	/*
966 	 * Although the run queue is locked the thread may be
967 	 * blocked.  We can not set the lock until it is unblocked.
968 	 */
969 	thread_lock_block_wait(td);
970 	sched_rem(td);
971 	THREAD_LOCKPTR_ASSERT(td, TDQ_LOCKPTR(from));
972 	td->td_lock = TDQ_LOCKPTR(to);
973 	td_get_sched(td)->ts_cpu = cpu;
974 	tdq_add(to, td, SRQ_YIELDING);
975 
976 	return (td);
977 }
978 
979 /*
980  * This tdq has idled.  Try to steal a thread from another cpu and switch
981  * to it.
982  */
983 static int
984 tdq_idled(struct tdq *tdq)
985 {
986 	struct cpu_group *cg;
987 	struct tdq *steal;
988 	cpuset_t mask;
989 	int cpu, switchcnt;
990 
991 	if (smp_started == 0 || steal_idle == 0 || tdq->tdq_cg == NULL)
992 		return (1);
993 	CPU_FILL(&mask);
994 	CPU_CLR(PCPU_GET(cpuid), &mask);
995     restart:
996 	switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
997 	for (cg = tdq->tdq_cg; ; ) {
998 		cpu = sched_highest(cg, mask, steal_thresh);
999 		/*
1000 		 * We were assigned a thread but not preempted.  Returning
1001 		 * 0 here will cause our caller to switch to it.
1002 		 */
1003 		if (tdq->tdq_load)
1004 			return (0);
1005 		if (cpu == -1) {
1006 			cg = cg->cg_parent;
1007 			if (cg == NULL)
1008 				return (1);
1009 			continue;
1010 		}
1011 		steal = TDQ_CPU(cpu);
1012 		/*
1013 		 * The data returned by sched_highest() is stale and
1014 		 * the chosen CPU no longer has an eligible thread.
1015 		 *
1016 		 * Testing this ahead of tdq_lock_pair() only catches
1017 		 * this situation about 20% of the time on an 8 core
1018 		 * 16 thread Ryzen 7, but it still helps performance.
1019 		 */
1020 		if (steal->tdq_load < steal_thresh ||
1021 		    steal->tdq_transferable == 0)
1022 			goto restart;
1023 		tdq_lock_pair(tdq, steal);
1024 		/*
1025 		 * We were assigned a thread while waiting for the locks.
1026 		 * Switch to it now instead of stealing a thread.
1027 		 */
1028 		if (tdq->tdq_load)
1029 			break;
1030 		/*
1031 		 * The data returned by sched_highest() is stale and
1032 		 * the chosen CPU no longer has an eligible thread, or
1033 		 * we were preempted and the CPU loading info may be out
1034 		 * of date.  The latter is rare.  In either case restart
1035 		 * the search.
1036 		 */
1037 		if (steal->tdq_load < steal_thresh ||
1038 		    steal->tdq_transferable == 0 ||
1039 		    switchcnt != tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt) {
1040 			tdq_unlock_pair(tdq, steal);
1041 			goto restart;
1042 		}
1043 		/*
1044 		 * Steal the thread and switch to it.
1045 		 */
1046 		if (tdq_move(steal, tdq) != NULL)
1047 			break;
1048 		/*
1049 		 * We failed to acquire a thread even though it looked
1050 		 * like one was available.  This could be due to affinity
1051 		 * restrictions or for other reasons.  Loop again after
1052 		 * removing this CPU from the set.  The restart logic
1053 		 * above does not restore this CPU to the set due to the
1054 		 * likelyhood of failing here again.
1055 		 */
1056 		CPU_CLR(cpu, &mask);
1057 		tdq_unlock_pair(tdq, steal);
1058 	}
1059 	TDQ_UNLOCK(steal);
1060 	mi_switch(SW_VOL | SWT_IDLE);
1061 	return (0);
1062 }
1063 
1064 /*
1065  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
1066  */
1067 static void
1068 tdq_notify(struct tdq *tdq, struct thread *td)
1069 {
1070 	struct thread *ctd;
1071 	int pri;
1072 	int cpu;
1073 
1074 	if (tdq->tdq_owepreempt)
1075 		return;
1076 	cpu = td_get_sched(td)->ts_cpu;
1077 	pri = td->td_priority;
1078 	ctd = pcpu_find(cpu)->pc_curthread;
1079 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1080 		return;
1081 
1082 	/*
1083 	 * Make sure that our caller's earlier update to tdq_load is
1084 	 * globally visible before we read tdq_cpu_idle.  Idle thread
1085 	 * accesses both of them without locks, and the order is important.
1086 	 */
1087 	atomic_thread_fence_seq_cst();
1088 
1089 	if (TD_IS_IDLETHREAD(ctd)) {
1090 		/*
1091 		 * If the MD code has an idle wakeup routine try that before
1092 		 * falling back to IPI.
1093 		 */
1094 		if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1095 			return;
1096 	}
1097 
1098 	/*
1099 	 * The run queues have been updated, so any switch on the remote CPU
1100 	 * will satisfy the preemption request.
1101 	 */
1102 	tdq->tdq_owepreempt = 1;
1103 	ipi_cpu(cpu, IPI_PREEMPT);
1104 }
1105 
1106 /*
1107  * Steals load from a timeshare queue.  Honors the rotating queue head
1108  * index.
1109  */
1110 static struct thread *
1111 runq_steal_from(struct runq *rq, int cpu, u_char start)
1112 {
1113 	struct rqbits *rqb;
1114 	struct rqhead *rqh;
1115 	struct thread *td, *first;
1116 	int bit;
1117 	int i;
1118 
1119 	rqb = &rq->rq_status;
1120 	bit = start & (RQB_BPW -1);
1121 	first = NULL;
1122 again:
1123 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1124 		if (rqb->rqb_bits[i] == 0)
1125 			continue;
1126 		if (bit == 0)
1127 			bit = RQB_FFS(rqb->rqb_bits[i]);
1128 		for (; bit < RQB_BPW; bit++) {
1129 			if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1130 				continue;
1131 			rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1132 			TAILQ_FOREACH(td, rqh, td_runq) {
1133 				if (first && THREAD_CAN_MIGRATE(td) &&
1134 				    THREAD_CAN_SCHED(td, cpu))
1135 					return (td);
1136 				first = td;
1137 			}
1138 		}
1139 	}
1140 	if (start != 0) {
1141 		start = 0;
1142 		goto again;
1143 	}
1144 
1145 	if (first && THREAD_CAN_MIGRATE(first) &&
1146 	    THREAD_CAN_SCHED(first, cpu))
1147 		return (first);
1148 	return (NULL);
1149 }
1150 
1151 /*
1152  * Steals load from a standard linear queue.
1153  */
1154 static struct thread *
1155 runq_steal(struct runq *rq, int cpu)
1156 {
1157 	struct rqhead *rqh;
1158 	struct rqbits *rqb;
1159 	struct thread *td;
1160 	int word;
1161 	int bit;
1162 
1163 	rqb = &rq->rq_status;
1164 	for (word = 0; word < RQB_LEN; word++) {
1165 		if (rqb->rqb_bits[word] == 0)
1166 			continue;
1167 		for (bit = 0; bit < RQB_BPW; bit++) {
1168 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1169 				continue;
1170 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1171 			TAILQ_FOREACH(td, rqh, td_runq)
1172 				if (THREAD_CAN_MIGRATE(td) &&
1173 				    THREAD_CAN_SCHED(td, cpu))
1174 					return (td);
1175 		}
1176 	}
1177 	return (NULL);
1178 }
1179 
1180 /*
1181  * Attempt to steal a thread in priority order from a thread queue.
1182  */
1183 static struct thread *
1184 tdq_steal(struct tdq *tdq, int cpu)
1185 {
1186 	struct thread *td;
1187 
1188 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1189 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1190 		return (td);
1191 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1192 	    cpu, tdq->tdq_ridx)) != NULL)
1193 		return (td);
1194 	return (runq_steal(&tdq->tdq_idle, cpu));
1195 }
1196 
1197 /*
1198  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1199  * current lock and returns with the assigned queue locked.
1200  */
1201 static inline struct tdq *
1202 sched_setcpu(struct thread *td, int cpu, int flags)
1203 {
1204 
1205 	struct tdq *tdq;
1206 	struct mtx *mtx;
1207 
1208 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1209 	tdq = TDQ_CPU(cpu);
1210 	td_get_sched(td)->ts_cpu = cpu;
1211 	/*
1212 	 * If the lock matches just return the queue.
1213 	 */
1214 	if (td->td_lock == TDQ_LOCKPTR(tdq)) {
1215 		KASSERT((flags & SRQ_HOLD) == 0,
1216 		    ("sched_setcpu: Invalid lock for SRQ_HOLD"));
1217 		return (tdq);
1218 	}
1219 
1220 	/*
1221 	 * The hard case, migration, we need to block the thread first to
1222 	 * prevent order reversals with other cpus locks.
1223 	 */
1224 	spinlock_enter();
1225 	mtx = thread_lock_block(td);
1226 	if ((flags & SRQ_HOLD) == 0)
1227 		mtx_unlock_spin(mtx);
1228 	TDQ_LOCK(tdq);
1229 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1230 	spinlock_exit();
1231 	return (tdq);
1232 }
1233 
1234 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1235 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1236 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1237 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1238 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1239 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1240 
1241 static int
1242 sched_pickcpu(struct thread *td, int flags)
1243 {
1244 	struct cpu_group *cg, *ccg;
1245 	struct td_sched *ts;
1246 	struct tdq *tdq;
1247 	cpuset_t mask;
1248 	int cpu, pri, self, intr;
1249 
1250 	self = PCPU_GET(cpuid);
1251 	ts = td_get_sched(td);
1252 	KASSERT(!CPU_ABSENT(ts->ts_cpu), ("sched_pickcpu: Start scheduler on "
1253 	    "absent CPU %d for thread %s.", ts->ts_cpu, td->td_name));
1254 	if (smp_started == 0)
1255 		return (self);
1256 	/*
1257 	 * Don't migrate a running thread from sched_switch().
1258 	 */
1259 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1260 		return (ts->ts_cpu);
1261 	/*
1262 	 * Prefer to run interrupt threads on the processors that generate
1263 	 * the interrupt.
1264 	 */
1265 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1266 	    curthread->td_intr_nesting_level) {
1267 		tdq = TDQ_SELF();
1268 		if (tdq->tdq_lowpri >= PRI_MIN_IDLE) {
1269 			SCHED_STAT_INC(pickcpu_idle_affinity);
1270 			return (self);
1271 		}
1272 		ts->ts_cpu = self;
1273 		intr = 1;
1274 		cg = tdq->tdq_cg;
1275 		goto llc;
1276 	} else {
1277 		intr = 0;
1278 		tdq = TDQ_CPU(ts->ts_cpu);
1279 		cg = tdq->tdq_cg;
1280 	}
1281 	/*
1282 	 * If the thread can run on the last cpu and the affinity has not
1283 	 * expired and it is idle, run it there.
1284 	 */
1285 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1286 	    tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1287 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1288 		if (cg->cg_flags & CG_FLAG_THREAD) {
1289 			/* Check all SMT threads for being idle. */
1290 			for (cpu = CPU_FFS(&cg->cg_mask) - 1; ; cpu++) {
1291 				if (CPU_ISSET(cpu, &cg->cg_mask) &&
1292 				    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1293 					break;
1294 				if (cpu >= mp_maxid) {
1295 					SCHED_STAT_INC(pickcpu_idle_affinity);
1296 					return (ts->ts_cpu);
1297 				}
1298 			}
1299 		} else {
1300 			SCHED_STAT_INC(pickcpu_idle_affinity);
1301 			return (ts->ts_cpu);
1302 		}
1303 	}
1304 llc:
1305 	/*
1306 	 * Search for the last level cache CPU group in the tree.
1307 	 * Skip SMT, identical groups and caches with expired affinity.
1308 	 * Interrupt threads affinity is explicit and never expires.
1309 	 */
1310 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1311 		if (cg->cg_flags & CG_FLAG_THREAD)
1312 			continue;
1313 		if (cg->cg_children == 1 || cg->cg_count == 1)
1314 			continue;
1315 		if (cg->cg_level == CG_SHARE_NONE ||
1316 		    (!intr && !SCHED_AFFINITY(ts, cg->cg_level)))
1317 			continue;
1318 		ccg = cg;
1319 	}
1320 	/* Found LLC shared by all CPUs, so do a global search. */
1321 	if (ccg == cpu_top)
1322 		ccg = NULL;
1323 	cpu = -1;
1324 	mask = td->td_cpuset->cs_mask;
1325 	pri = td->td_priority;
1326 	/*
1327 	 * Try hard to keep interrupts within found LLC.  Search the LLC for
1328 	 * the least loaded CPU we can run now.  For NUMA systems it should
1329 	 * be within target domain, and it also reduces scheduling overhead.
1330 	 */
1331 	if (ccg != NULL && intr) {
1332 		cpu = sched_lowest(ccg, mask, pri, INT_MAX, ts->ts_cpu);
1333 		if (cpu >= 0)
1334 			SCHED_STAT_INC(pickcpu_intrbind);
1335 	} else
1336 	/* Search the LLC for the least loaded idle CPU we can run now. */
1337 	if (ccg != NULL) {
1338 		cpu = sched_lowest(ccg, mask, max(pri, PRI_MAX_TIMESHARE),
1339 		    INT_MAX, ts->ts_cpu);
1340 		if (cpu >= 0)
1341 			SCHED_STAT_INC(pickcpu_affinity);
1342 	}
1343 	/* Search globally for the least loaded CPU we can run now. */
1344 	if (cpu < 0) {
1345 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1346 		if (cpu >= 0)
1347 			SCHED_STAT_INC(pickcpu_lowest);
1348 	}
1349 	/* Search globally for the least loaded CPU. */
1350 	if (cpu < 0) {
1351 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1352 		if (cpu >= 0)
1353 			SCHED_STAT_INC(pickcpu_lowest);
1354 	}
1355 	KASSERT(cpu >= 0, ("sched_pickcpu: Failed to find a cpu."));
1356 	KASSERT(!CPU_ABSENT(cpu), ("sched_pickcpu: Picked absent CPU %d.", cpu));
1357 	/*
1358 	 * Compare the lowest loaded cpu to current cpu.
1359 	 */
1360 	tdq = TDQ_CPU(cpu);
1361 	if (THREAD_CAN_SCHED(td, self) && TDQ_SELF()->tdq_lowpri > pri &&
1362 	    tdq->tdq_lowpri < PRI_MIN_IDLE &&
1363 	    TDQ_SELF()->tdq_load <= tdq->tdq_load + 1) {
1364 		SCHED_STAT_INC(pickcpu_local);
1365 		cpu = self;
1366 	}
1367 	if (cpu != ts->ts_cpu)
1368 		SCHED_STAT_INC(pickcpu_migration);
1369 	return (cpu);
1370 }
1371 #endif
1372 
1373 /*
1374  * Pick the highest priority task we have and return it.
1375  */
1376 static struct thread *
1377 tdq_choose(struct tdq *tdq)
1378 {
1379 	struct thread *td;
1380 
1381 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1382 	td = runq_choose(&tdq->tdq_realtime);
1383 	if (td != NULL)
1384 		return (td);
1385 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1386 	if (td != NULL) {
1387 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1388 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1389 		    td->td_priority));
1390 		return (td);
1391 	}
1392 	td = runq_choose(&tdq->tdq_idle);
1393 	if (td != NULL) {
1394 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1395 		    ("tdq_choose: Invalid priority on idle queue %d",
1396 		    td->td_priority));
1397 		return (td);
1398 	}
1399 
1400 	return (NULL);
1401 }
1402 
1403 /*
1404  * Initialize a thread queue.
1405  */
1406 static void
1407 tdq_setup(struct tdq *tdq, int id)
1408 {
1409 
1410 	if (bootverbose)
1411 		printf("ULE: setup cpu %d\n", id);
1412 	runq_init(&tdq->tdq_realtime);
1413 	runq_init(&tdq->tdq_timeshare);
1414 	runq_init(&tdq->tdq_idle);
1415 	tdq->tdq_id = id;
1416 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1417 	    "sched lock %d", (int)TDQ_ID(tdq));
1418 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock", MTX_SPIN);
1419 #ifdef KTR
1420 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1421 	    "CPU %d load", (int)TDQ_ID(tdq));
1422 #endif
1423 }
1424 
1425 #ifdef SMP
1426 static void
1427 sched_setup_smp(void)
1428 {
1429 	struct tdq *tdq;
1430 	int i;
1431 
1432 	cpu_top = smp_topo();
1433 	CPU_FOREACH(i) {
1434 		tdq = DPCPU_ID_PTR(i, tdq);
1435 		tdq_setup(tdq, i);
1436 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1437 		if (tdq->tdq_cg == NULL)
1438 			panic("Can't find cpu group for %d\n", i);
1439 	}
1440 	PCPU_SET(sched, DPCPU_PTR(tdq));
1441 	balance_tdq = TDQ_SELF();
1442 }
1443 #endif
1444 
1445 /*
1446  * Setup the thread queues and initialize the topology based on MD
1447  * information.
1448  */
1449 static void
1450 sched_setup(void *dummy)
1451 {
1452 	struct tdq *tdq;
1453 
1454 #ifdef SMP
1455 	sched_setup_smp();
1456 #else
1457 	tdq_setup(TDQ_SELF(), 0);
1458 #endif
1459 	tdq = TDQ_SELF();
1460 
1461 	/* Add thread0's load since it's running. */
1462 	TDQ_LOCK(tdq);
1463 	thread0.td_lock = TDQ_LOCKPTR(tdq);
1464 	tdq_load_add(tdq, &thread0);
1465 	tdq->tdq_lowpri = thread0.td_priority;
1466 	TDQ_UNLOCK(tdq);
1467 }
1468 
1469 /*
1470  * This routine determines time constants after stathz and hz are setup.
1471  */
1472 /* ARGSUSED */
1473 static void
1474 sched_initticks(void *dummy)
1475 {
1476 	int incr;
1477 
1478 	realstathz = stathz ? stathz : hz;
1479 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1480 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1481 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1482 	    realstathz);
1483 
1484 	/*
1485 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1486 	 * hz not being evenly divisible by stathz on all platforms.
1487 	 */
1488 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1489 	/*
1490 	 * This does not work for values of stathz that are more than
1491 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1492 	 */
1493 	if (incr == 0)
1494 		incr = 1;
1495 	tickincr = incr;
1496 #ifdef SMP
1497 	/*
1498 	 * Set the default balance interval now that we know
1499 	 * what realstathz is.
1500 	 */
1501 	balance_interval = realstathz;
1502 	balance_ticks = balance_interval;
1503 	affinity = SCHED_AFFINITY_DEFAULT;
1504 #endif
1505 	if (sched_idlespinthresh < 0)
1506 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1507 }
1508 
1509 
1510 /*
1511  * This is the core of the interactivity algorithm.  Determines a score based
1512  * on past behavior.  It is the ratio of sleep time to run time scaled to
1513  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1514  * differs from the cpu usage because it does not account for time spent
1515  * waiting on a run-queue.  Would be prettier if we had floating point.
1516  *
1517  * When a thread's sleep time is greater than its run time the
1518  * calculation is:
1519  *
1520  *                           scaling factor
1521  * interactivity score =  ---------------------
1522  *                        sleep time / run time
1523  *
1524  *
1525  * When a thread's run time is greater than its sleep time the
1526  * calculation is:
1527  *
1528  *                           scaling factor
1529  * interactivity score =  ---------------------    + scaling factor
1530  *                        run time / sleep time
1531  */
1532 static int
1533 sched_interact_score(struct thread *td)
1534 {
1535 	struct td_sched *ts;
1536 	int div;
1537 
1538 	ts = td_get_sched(td);
1539 	/*
1540 	 * The score is only needed if this is likely to be an interactive
1541 	 * task.  Don't go through the expense of computing it if there's
1542 	 * no chance.
1543 	 */
1544 	if (sched_interact <= SCHED_INTERACT_HALF &&
1545 		ts->ts_runtime >= ts->ts_slptime)
1546 			return (SCHED_INTERACT_HALF);
1547 
1548 	if (ts->ts_runtime > ts->ts_slptime) {
1549 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1550 		return (SCHED_INTERACT_HALF +
1551 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1552 	}
1553 	if (ts->ts_slptime > ts->ts_runtime) {
1554 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1555 		return (ts->ts_runtime / div);
1556 	}
1557 	/* runtime == slptime */
1558 	if (ts->ts_runtime)
1559 		return (SCHED_INTERACT_HALF);
1560 
1561 	/*
1562 	 * This can happen if slptime and runtime are 0.
1563 	 */
1564 	return (0);
1565 
1566 }
1567 
1568 /*
1569  * Scale the scheduling priority according to the "interactivity" of this
1570  * process.
1571  */
1572 static void
1573 sched_priority(struct thread *td)
1574 {
1575 	int score;
1576 	int pri;
1577 
1578 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1579 		return;
1580 	/*
1581 	 * If the score is interactive we place the thread in the realtime
1582 	 * queue with a priority that is less than kernel and interrupt
1583 	 * priorities.  These threads are not subject to nice restrictions.
1584 	 *
1585 	 * Scores greater than this are placed on the normal timeshare queue
1586 	 * where the priority is partially decided by the most recent cpu
1587 	 * utilization and the rest is decided by nice value.
1588 	 *
1589 	 * The nice value of the process has a linear effect on the calculated
1590 	 * score.  Negative nice values make it easier for a thread to be
1591 	 * considered interactive.
1592 	 */
1593 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1594 	if (score < sched_interact) {
1595 		pri = PRI_MIN_INTERACT;
1596 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1597 		    sched_interact) * score;
1598 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1599 		    ("sched_priority: invalid interactive priority %d score %d",
1600 		    pri, score));
1601 	} else {
1602 		pri = SCHED_PRI_MIN;
1603 		if (td_get_sched(td)->ts_ticks)
1604 			pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1605 			    SCHED_PRI_RANGE - 1);
1606 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1607 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1608 		    ("sched_priority: invalid priority %d: nice %d, "
1609 		    "ticks %d ftick %d ltick %d tick pri %d",
1610 		    pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1611 		    td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1612 		    SCHED_PRI_TICKS(td_get_sched(td))));
1613 	}
1614 	sched_user_prio(td, pri);
1615 
1616 	return;
1617 }
1618 
1619 /*
1620  * This routine enforces a maximum limit on the amount of scheduling history
1621  * kept.  It is called after either the slptime or runtime is adjusted.  This
1622  * function is ugly due to integer math.
1623  */
1624 static void
1625 sched_interact_update(struct thread *td)
1626 {
1627 	struct td_sched *ts;
1628 	u_int sum;
1629 
1630 	ts = td_get_sched(td);
1631 	sum = ts->ts_runtime + ts->ts_slptime;
1632 	if (sum < SCHED_SLP_RUN_MAX)
1633 		return;
1634 	/*
1635 	 * This only happens from two places:
1636 	 * 1) We have added an unusual amount of run time from fork_exit.
1637 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1638 	 */
1639 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1640 		if (ts->ts_runtime > ts->ts_slptime) {
1641 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1642 			ts->ts_slptime = 1;
1643 		} else {
1644 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1645 			ts->ts_runtime = 1;
1646 		}
1647 		return;
1648 	}
1649 	/*
1650 	 * If we have exceeded by more than 1/5th then the algorithm below
1651 	 * will not bring us back into range.  Dividing by two here forces
1652 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1653 	 */
1654 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1655 		ts->ts_runtime /= 2;
1656 		ts->ts_slptime /= 2;
1657 		return;
1658 	}
1659 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1660 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1661 }
1662 
1663 /*
1664  * Scale back the interactivity history when a child thread is created.  The
1665  * history is inherited from the parent but the thread may behave totally
1666  * differently.  For example, a shell spawning a compiler process.  We want
1667  * to learn that the compiler is behaving badly very quickly.
1668  */
1669 static void
1670 sched_interact_fork(struct thread *td)
1671 {
1672 	struct td_sched *ts;
1673 	int ratio;
1674 	int sum;
1675 
1676 	ts = td_get_sched(td);
1677 	sum = ts->ts_runtime + ts->ts_slptime;
1678 	if (sum > SCHED_SLP_RUN_FORK) {
1679 		ratio = sum / SCHED_SLP_RUN_FORK;
1680 		ts->ts_runtime /= ratio;
1681 		ts->ts_slptime /= ratio;
1682 	}
1683 }
1684 
1685 /*
1686  * Called from proc0_init() to setup the scheduler fields.
1687  */
1688 void
1689 schedinit(void)
1690 {
1691 	struct td_sched *ts0;
1692 
1693 	/*
1694 	 * Set up the scheduler specific parts of thread0.
1695 	 */
1696 	ts0 = td_get_sched(&thread0);
1697 	ts0->ts_ltick = ticks;
1698 	ts0->ts_ftick = ticks;
1699 	ts0->ts_slice = 0;
1700 	ts0->ts_cpu = curcpu;	/* set valid CPU number */
1701 }
1702 
1703 /*
1704  * This is only somewhat accurate since given many processes of the same
1705  * priority they will switch when their slices run out, which will be
1706  * at most sched_slice stathz ticks.
1707  */
1708 int
1709 sched_rr_interval(void)
1710 {
1711 
1712 	/* Convert sched_slice from stathz to hz. */
1713 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1714 }
1715 
1716 /*
1717  * Update the percent cpu tracking information when it is requested or
1718  * the total history exceeds the maximum.  We keep a sliding history of
1719  * tick counts that slowly decays.  This is less precise than the 4BSD
1720  * mechanism since it happens with less regular and frequent events.
1721  */
1722 static void
1723 sched_pctcpu_update(struct td_sched *ts, int run)
1724 {
1725 	int t = ticks;
1726 
1727 	/*
1728 	 * The signed difference may be negative if the thread hasn't run for
1729 	 * over half of the ticks rollover period.
1730 	 */
1731 	if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) {
1732 		ts->ts_ticks = 0;
1733 		ts->ts_ftick = t - SCHED_TICK_TARG;
1734 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1735 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1736 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1737 		ts->ts_ftick = t - SCHED_TICK_TARG;
1738 	}
1739 	if (run)
1740 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1741 	ts->ts_ltick = t;
1742 }
1743 
1744 /*
1745  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1746  * if necessary.  This is the back-end for several priority related
1747  * functions.
1748  */
1749 static void
1750 sched_thread_priority(struct thread *td, u_char prio)
1751 {
1752 	struct td_sched *ts;
1753 	struct tdq *tdq;
1754 	int oldpri;
1755 
1756 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1757 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1758 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1759 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1760 	if (td != curthread && prio < td->td_priority) {
1761 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1762 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1763 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1764 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1765 		    curthread);
1766 	}
1767 	ts = td_get_sched(td);
1768 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1769 	if (td->td_priority == prio)
1770 		return;
1771 	/*
1772 	 * If the priority has been elevated due to priority
1773 	 * propagation, we may have to move ourselves to a new
1774 	 * queue.  This could be optimized to not re-add in some
1775 	 * cases.
1776 	 */
1777 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1778 		sched_rem(td);
1779 		td->td_priority = prio;
1780 		sched_add(td, SRQ_BORROWING | SRQ_HOLDTD);
1781 		return;
1782 	}
1783 	/*
1784 	 * If the thread is currently running we may have to adjust the lowpri
1785 	 * information so other cpus are aware of our current priority.
1786 	 */
1787 	if (TD_IS_RUNNING(td)) {
1788 		tdq = TDQ_CPU(ts->ts_cpu);
1789 		oldpri = td->td_priority;
1790 		td->td_priority = prio;
1791 		if (prio < tdq->tdq_lowpri)
1792 			tdq->tdq_lowpri = prio;
1793 		else if (tdq->tdq_lowpri == oldpri)
1794 			tdq_setlowpri(tdq, td);
1795 		return;
1796 	}
1797 	td->td_priority = prio;
1798 }
1799 
1800 /*
1801  * Update a thread's priority when it is lent another thread's
1802  * priority.
1803  */
1804 void
1805 sched_lend_prio(struct thread *td, u_char prio)
1806 {
1807 
1808 	td->td_flags |= TDF_BORROWING;
1809 	sched_thread_priority(td, prio);
1810 }
1811 
1812 /*
1813  * Restore a thread's priority when priority propagation is
1814  * over.  The prio argument is the minimum priority the thread
1815  * needs to have to satisfy other possible priority lending
1816  * requests.  If the thread's regular priority is less
1817  * important than prio, the thread will keep a priority boost
1818  * of prio.
1819  */
1820 void
1821 sched_unlend_prio(struct thread *td, u_char prio)
1822 {
1823 	u_char base_pri;
1824 
1825 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1826 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1827 		base_pri = td->td_user_pri;
1828 	else
1829 		base_pri = td->td_base_pri;
1830 	if (prio >= base_pri) {
1831 		td->td_flags &= ~TDF_BORROWING;
1832 		sched_thread_priority(td, base_pri);
1833 	} else
1834 		sched_lend_prio(td, prio);
1835 }
1836 
1837 /*
1838  * Standard entry for setting the priority to an absolute value.
1839  */
1840 void
1841 sched_prio(struct thread *td, u_char prio)
1842 {
1843 	u_char oldprio;
1844 
1845 	/* First, update the base priority. */
1846 	td->td_base_pri = prio;
1847 
1848 	/*
1849 	 * If the thread is borrowing another thread's priority, don't
1850 	 * ever lower the priority.
1851 	 */
1852 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1853 		return;
1854 
1855 	/* Change the real priority. */
1856 	oldprio = td->td_priority;
1857 	sched_thread_priority(td, prio);
1858 
1859 	/*
1860 	 * If the thread is on a turnstile, then let the turnstile update
1861 	 * its state.
1862 	 */
1863 	if (TD_ON_LOCK(td) && oldprio != prio)
1864 		turnstile_adjust(td, oldprio);
1865 }
1866 
1867 /*
1868  * Set the base user priority, does not effect current running priority.
1869  */
1870 void
1871 sched_user_prio(struct thread *td, u_char prio)
1872 {
1873 
1874 	td->td_base_user_pri = prio;
1875 	if (td->td_lend_user_pri <= prio)
1876 		return;
1877 	td->td_user_pri = prio;
1878 }
1879 
1880 void
1881 sched_lend_user_prio(struct thread *td, u_char prio)
1882 {
1883 
1884 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1885 	td->td_lend_user_pri = prio;
1886 	td->td_user_pri = min(prio, td->td_base_user_pri);
1887 	if (td->td_priority > td->td_user_pri)
1888 		sched_prio(td, td->td_user_pri);
1889 	else if (td->td_priority != td->td_user_pri)
1890 		td->td_flags |= TDF_NEEDRESCHED;
1891 }
1892 
1893 /*
1894  * Like the above but first check if there is anything to do.
1895  */
1896 void
1897 sched_lend_user_prio_cond(struct thread *td, u_char prio)
1898 {
1899 
1900 	if (td->td_lend_user_pri != prio)
1901 		goto lend;
1902 	if (td->td_user_pri != min(prio, td->td_base_user_pri))
1903 		goto lend;
1904 	if (td->td_priority >= td->td_user_pri)
1905 		goto lend;
1906 	return;
1907 
1908 lend:
1909 	thread_lock(td);
1910 	sched_lend_user_prio(td, prio);
1911 	thread_unlock(td);
1912 }
1913 
1914 #ifdef SMP
1915 /*
1916  * This tdq is about to idle.  Try to steal a thread from another CPU before
1917  * choosing the idle thread.
1918  */
1919 static void
1920 tdq_trysteal(struct tdq *tdq)
1921 {
1922 	struct cpu_group *cg;
1923 	struct tdq *steal;
1924 	cpuset_t mask;
1925 	int cpu, i;
1926 
1927 	if (smp_started == 0 || trysteal_limit == 0 || tdq->tdq_cg == NULL)
1928 		return;
1929 	CPU_FILL(&mask);
1930 	CPU_CLR(PCPU_GET(cpuid), &mask);
1931 	/* We don't want to be preempted while we're iterating. */
1932 	spinlock_enter();
1933 	TDQ_UNLOCK(tdq);
1934 	for (i = 1, cg = tdq->tdq_cg; ; ) {
1935 		cpu = sched_highest(cg, mask, steal_thresh);
1936 		/*
1937 		 * If a thread was added while interrupts were disabled don't
1938 		 * steal one here.
1939 		 */
1940 		if (tdq->tdq_load > 0) {
1941 			TDQ_LOCK(tdq);
1942 			break;
1943 		}
1944 		if (cpu == -1) {
1945 			i++;
1946 			cg = cg->cg_parent;
1947 			if (cg == NULL || i > trysteal_limit) {
1948 				TDQ_LOCK(tdq);
1949 				break;
1950 			}
1951 			continue;
1952 		}
1953 		steal = TDQ_CPU(cpu);
1954 		/*
1955 		 * The data returned by sched_highest() is stale and
1956                  * the chosen CPU no longer has an eligible thread.
1957 		 */
1958 		if (steal->tdq_load < steal_thresh ||
1959 		    steal->tdq_transferable == 0)
1960 			continue;
1961 		tdq_lock_pair(tdq, steal);
1962 		/*
1963 		 * If we get to this point, unconditonally exit the loop
1964 		 * to bound the time spent in the critcal section.
1965 		 *
1966 		 * If a thread was added while interrupts were disabled don't
1967 		 * steal one here.
1968 		 */
1969 		if (tdq->tdq_load > 0) {
1970 			TDQ_UNLOCK(steal);
1971 			break;
1972 		}
1973 		/*
1974 		 * The data returned by sched_highest() is stale and
1975                  * the chosen CPU no longer has an eligible thread.
1976 		 */
1977 		if (steal->tdq_load < steal_thresh ||
1978 		    steal->tdq_transferable == 0) {
1979 			TDQ_UNLOCK(steal);
1980 			break;
1981 		}
1982 		/*
1983 		 * If we fail to acquire one due to affinity restrictions,
1984 		 * bail out and let the idle thread to a more complete search
1985 		 * outside of a critical section.
1986 		 */
1987 		if (tdq_move(steal, tdq) == NULL) {
1988 			TDQ_UNLOCK(steal);
1989 			break;
1990 		}
1991 		TDQ_UNLOCK(steal);
1992 		break;
1993 	}
1994 	spinlock_exit();
1995 }
1996 #endif
1997 
1998 /*
1999  * Handle migration from sched_switch().  This happens only for
2000  * cpu binding.
2001  */
2002 static struct mtx *
2003 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
2004 {
2005 	struct tdq *tdn;
2006 
2007 	KASSERT(THREAD_CAN_MIGRATE(td) ||
2008 	    (td_get_sched(td)->ts_flags & TSF_BOUND) != 0,
2009 	    ("Thread %p shouldn't migrate", td));
2010 	KASSERT(!CPU_ABSENT(td_get_sched(td)->ts_cpu), ("sched_switch_migrate: "
2011 	    "thread %s queued on absent CPU %d.", td->td_name,
2012 	    td_get_sched(td)->ts_cpu));
2013 	tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
2014 #ifdef SMP
2015 	tdq_load_rem(tdq, td);
2016 	/*
2017 	 * Do the lock dance required to avoid LOR.  We have an
2018 	 * extra spinlock nesting from sched_switch() which will
2019 	 * prevent preemption while we're holding neither run-queue lock.
2020 	 */
2021 	TDQ_UNLOCK(tdq);
2022 	TDQ_LOCK(tdn);
2023 	tdq_add(tdn, td, flags);
2024 	tdq_notify(tdn, td);
2025 	TDQ_UNLOCK(tdn);
2026 	TDQ_LOCK(tdq);
2027 #endif
2028 	return (TDQ_LOCKPTR(tdn));
2029 }
2030 
2031 /*
2032  * thread_lock_unblock() that does not assume td_lock is blocked.
2033  */
2034 static inline void
2035 thread_unblock_switch(struct thread *td, struct mtx *mtx)
2036 {
2037 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
2038 	    (uintptr_t)mtx);
2039 }
2040 
2041 /*
2042  * Switch threads.  This function has to handle threads coming in while
2043  * blocked for some reason, running, or idle.  It also must deal with
2044  * migrating a thread from one queue to another as running threads may
2045  * be assigned elsewhere via binding.
2046  */
2047 void
2048 sched_switch(struct thread *td, int flags)
2049 {
2050 	struct thread *newtd;
2051 	struct tdq *tdq;
2052 	struct td_sched *ts;
2053 	struct mtx *mtx;
2054 	int srqflag;
2055 	int cpuid, preempted;
2056 
2057 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2058 
2059 	cpuid = PCPU_GET(cpuid);
2060 	tdq = TDQ_SELF();
2061 	ts = td_get_sched(td);
2062 	sched_pctcpu_update(ts, 1);
2063 	ts->ts_rltick = ticks;
2064 	td->td_lastcpu = td->td_oncpu;
2065 	preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
2066 	    (flags & SW_PREEMPT) != 0;
2067 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
2068 	td->td_owepreempt = 0;
2069 	tdq->tdq_owepreempt = 0;
2070 	if (!TD_IS_IDLETHREAD(td))
2071 		tdq->tdq_switchcnt++;
2072 
2073 	/*
2074 	 * Always block the thread lock so we can drop the tdq lock early.
2075 	 */
2076 	mtx = thread_lock_block(td);
2077 	spinlock_enter();
2078 	if (TD_IS_IDLETHREAD(td)) {
2079 		MPASS(mtx == TDQ_LOCKPTR(tdq));
2080 		TD_SET_CAN_RUN(td);
2081 	} else if (TD_IS_RUNNING(td)) {
2082 		MPASS(mtx == TDQ_LOCKPTR(tdq));
2083 		srqflag = preempted ?
2084 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
2085 		    SRQ_OURSELF|SRQ_YIELDING;
2086 #ifdef SMP
2087 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
2088 			ts->ts_cpu = sched_pickcpu(td, 0);
2089 #endif
2090 		if (ts->ts_cpu == cpuid)
2091 			tdq_runq_add(tdq, td, srqflag);
2092 		else
2093 			mtx = sched_switch_migrate(tdq, td, srqflag);
2094 	} else {
2095 		/* This thread must be going to sleep. */
2096 		if (mtx != TDQ_LOCKPTR(tdq)) {
2097 			mtx_unlock_spin(mtx);
2098 			TDQ_LOCK(tdq);
2099 		}
2100 		tdq_load_rem(tdq, td);
2101 #ifdef SMP
2102 		if (tdq->tdq_load == 0)
2103 			tdq_trysteal(tdq);
2104 #endif
2105 	}
2106 
2107 #if (KTR_COMPILE & KTR_SCHED) != 0
2108 	if (TD_IS_IDLETHREAD(td))
2109 		KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
2110 		    "prio:%d", td->td_priority);
2111 	else
2112 		KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
2113 		    "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
2114 		    "lockname:\"%s\"", td->td_lockname);
2115 #endif
2116 
2117 	/*
2118 	 * We enter here with the thread blocked and assigned to the
2119 	 * appropriate cpu run-queue or sleep-queue and with the current
2120 	 * thread-queue locked.
2121 	 */
2122 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2123 	newtd = choosethread();
2124 	sched_pctcpu_update(td_get_sched(newtd), 0);
2125 	TDQ_UNLOCK(tdq);
2126 
2127 	/*
2128 	 * Call the MD code to switch contexts if necessary.
2129 	 */
2130 	if (td != newtd) {
2131 #ifdef	HWPMC_HOOKS
2132 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2133 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
2134 #endif
2135 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
2136 
2137 #ifdef KDTRACE_HOOKS
2138 		/*
2139 		 * If DTrace has set the active vtime enum to anything
2140 		 * other than INACTIVE (0), then it should have set the
2141 		 * function to call.
2142 		 */
2143 		if (dtrace_vtime_active)
2144 			(*dtrace_vtime_switch_func)(newtd);
2145 #endif
2146 		td->td_oncpu = NOCPU;
2147 		cpu_switch(td, newtd, mtx);
2148 		cpuid = td->td_oncpu = PCPU_GET(cpuid);
2149 
2150 		SDT_PROBE0(sched, , , on__cpu);
2151 #ifdef	HWPMC_HOOKS
2152 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2153 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
2154 #endif
2155 	} else {
2156 		thread_unblock_switch(td, mtx);
2157 		SDT_PROBE0(sched, , , remain__cpu);
2158 	}
2159 	KASSERT(curthread->td_md.md_spinlock_count == 1,
2160 	    ("invalid count %d", curthread->td_md.md_spinlock_count));
2161 
2162 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2163 	    "prio:%d", td->td_priority);
2164 }
2165 
2166 /*
2167  * Adjust thread priorities as a result of a nice request.
2168  */
2169 void
2170 sched_nice(struct proc *p, int nice)
2171 {
2172 	struct thread *td;
2173 
2174 	PROC_LOCK_ASSERT(p, MA_OWNED);
2175 
2176 	p->p_nice = nice;
2177 	FOREACH_THREAD_IN_PROC(p, td) {
2178 		thread_lock(td);
2179 		sched_priority(td);
2180 		sched_prio(td, td->td_base_user_pri);
2181 		thread_unlock(td);
2182 	}
2183 }
2184 
2185 /*
2186  * Record the sleep time for the interactivity scorer.
2187  */
2188 void
2189 sched_sleep(struct thread *td, int prio)
2190 {
2191 
2192 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2193 
2194 	td->td_slptick = ticks;
2195 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2196 		td->td_flags |= TDF_CANSWAP;
2197 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2198 		return;
2199 	if (static_boost == 1 && prio)
2200 		sched_prio(td, prio);
2201 	else if (static_boost && td->td_priority > static_boost)
2202 		sched_prio(td, static_boost);
2203 }
2204 
2205 /*
2206  * Schedule a thread to resume execution and record how long it voluntarily
2207  * slept.  We also update the pctcpu, interactivity, and priority.
2208  *
2209  * Requires the thread lock on entry, drops on exit.
2210  */
2211 void
2212 sched_wakeup(struct thread *td, int srqflags)
2213 {
2214 	struct td_sched *ts;
2215 	int slptick;
2216 
2217 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2218 	ts = td_get_sched(td);
2219 	td->td_flags &= ~TDF_CANSWAP;
2220 
2221 	/*
2222 	 * If we slept for more than a tick update our interactivity and
2223 	 * priority.
2224 	 */
2225 	slptick = td->td_slptick;
2226 	td->td_slptick = 0;
2227 	if (slptick && slptick != ticks) {
2228 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2229 		sched_interact_update(td);
2230 		sched_pctcpu_update(ts, 0);
2231 	}
2232 	/*
2233 	 * Reset the slice value since we slept and advanced the round-robin.
2234 	 */
2235 	ts->ts_slice = 0;
2236 	sched_add(td, SRQ_BORING | srqflags);
2237 }
2238 
2239 /*
2240  * Penalize the parent for creating a new child and initialize the child's
2241  * priority.
2242  */
2243 void
2244 sched_fork(struct thread *td, struct thread *child)
2245 {
2246 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2247 	sched_pctcpu_update(td_get_sched(td), 1);
2248 	sched_fork_thread(td, child);
2249 	/*
2250 	 * Penalize the parent and child for forking.
2251 	 */
2252 	sched_interact_fork(child);
2253 	sched_priority(child);
2254 	td_get_sched(td)->ts_runtime += tickincr;
2255 	sched_interact_update(td);
2256 	sched_priority(td);
2257 }
2258 
2259 /*
2260  * Fork a new thread, may be within the same process.
2261  */
2262 void
2263 sched_fork_thread(struct thread *td, struct thread *child)
2264 {
2265 	struct td_sched *ts;
2266 	struct td_sched *ts2;
2267 	struct tdq *tdq;
2268 
2269 	tdq = TDQ_SELF();
2270 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2271 	/*
2272 	 * Initialize child.
2273 	 */
2274 	ts = td_get_sched(td);
2275 	ts2 = td_get_sched(child);
2276 	child->td_oncpu = NOCPU;
2277 	child->td_lastcpu = NOCPU;
2278 	child->td_lock = TDQ_LOCKPTR(tdq);
2279 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2280 	child->td_domain.dr_policy = td->td_cpuset->cs_domain;
2281 	ts2->ts_cpu = ts->ts_cpu;
2282 	ts2->ts_flags = 0;
2283 	/*
2284 	 * Grab our parents cpu estimation information.
2285 	 */
2286 	ts2->ts_ticks = ts->ts_ticks;
2287 	ts2->ts_ltick = ts->ts_ltick;
2288 	ts2->ts_ftick = ts->ts_ftick;
2289 	/*
2290 	 * Do not inherit any borrowed priority from the parent.
2291 	 */
2292 	child->td_priority = child->td_base_pri;
2293 	/*
2294 	 * And update interactivity score.
2295 	 */
2296 	ts2->ts_slptime = ts->ts_slptime;
2297 	ts2->ts_runtime = ts->ts_runtime;
2298 	/* Attempt to quickly learn interactivity. */
2299 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2300 #ifdef KTR
2301 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2302 #endif
2303 }
2304 
2305 /*
2306  * Adjust the priority class of a thread.
2307  */
2308 void
2309 sched_class(struct thread *td, int class)
2310 {
2311 
2312 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2313 	if (td->td_pri_class == class)
2314 		return;
2315 	td->td_pri_class = class;
2316 }
2317 
2318 /*
2319  * Return some of the child's priority and interactivity to the parent.
2320  */
2321 void
2322 sched_exit(struct proc *p, struct thread *child)
2323 {
2324 	struct thread *td;
2325 
2326 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2327 	    "prio:%d", child->td_priority);
2328 	PROC_LOCK_ASSERT(p, MA_OWNED);
2329 	td = FIRST_THREAD_IN_PROC(p);
2330 	sched_exit_thread(td, child);
2331 }
2332 
2333 /*
2334  * Penalize another thread for the time spent on this one.  This helps to
2335  * worsen the priority and interactivity of processes which schedule batch
2336  * jobs such as make.  This has little effect on the make process itself but
2337  * causes new processes spawned by it to receive worse scores immediately.
2338  */
2339 void
2340 sched_exit_thread(struct thread *td, struct thread *child)
2341 {
2342 
2343 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2344 	    "prio:%d", child->td_priority);
2345 	/*
2346 	 * Give the child's runtime to the parent without returning the
2347 	 * sleep time as a penalty to the parent.  This causes shells that
2348 	 * launch expensive things to mark their children as expensive.
2349 	 */
2350 	thread_lock(td);
2351 	td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2352 	sched_interact_update(td);
2353 	sched_priority(td);
2354 	thread_unlock(td);
2355 }
2356 
2357 void
2358 sched_preempt(struct thread *td)
2359 {
2360 	struct tdq *tdq;
2361 	int flags;
2362 
2363 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2364 
2365 	thread_lock(td);
2366 	tdq = TDQ_SELF();
2367 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2368 	if (td->td_priority > tdq->tdq_lowpri) {
2369 		if (td->td_critnest == 1) {
2370 			flags = SW_INVOL | SW_PREEMPT;
2371 			flags |= TD_IS_IDLETHREAD(td) ? SWT_REMOTEWAKEIDLE :
2372 			    SWT_REMOTEPREEMPT;
2373 			mi_switch(flags);
2374 			/* Switch dropped thread lock. */
2375 			return;
2376 		}
2377 		td->td_owepreempt = 1;
2378 	} else {
2379 		tdq->tdq_owepreempt = 0;
2380 	}
2381 	thread_unlock(td);
2382 }
2383 
2384 /*
2385  * Fix priorities on return to user-space.  Priorities may be elevated due
2386  * to static priorities in msleep() or similar.
2387  */
2388 void
2389 sched_userret_slowpath(struct thread *td)
2390 {
2391 
2392 	thread_lock(td);
2393 	td->td_priority = td->td_user_pri;
2394 	td->td_base_pri = td->td_user_pri;
2395 	tdq_setlowpri(TDQ_SELF(), td);
2396 	thread_unlock(td);
2397 }
2398 
2399 /*
2400  * Handle a stathz tick.  This is really only relevant for timeshare
2401  * threads.
2402  */
2403 void
2404 sched_clock(struct thread *td, int cnt)
2405 {
2406 	struct tdq *tdq;
2407 	struct td_sched *ts;
2408 
2409 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2410 	tdq = TDQ_SELF();
2411 #ifdef SMP
2412 	/*
2413 	 * We run the long term load balancer infrequently on the first cpu.
2414 	 */
2415 	if (balance_tdq == tdq && smp_started != 0 && rebalance != 0 &&
2416 	    balance_ticks != 0) {
2417 		balance_ticks -= cnt;
2418 		if (balance_ticks <= 0)
2419 			sched_balance();
2420 	}
2421 #endif
2422 	/*
2423 	 * Save the old switch count so we have a record of the last ticks
2424 	 * activity.   Initialize the new switch count based on our load.
2425 	 * If there is some activity seed it to reflect that.
2426 	 */
2427 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2428 	tdq->tdq_switchcnt = tdq->tdq_load;
2429 	/*
2430 	 * Advance the insert index once for each tick to ensure that all
2431 	 * threads get a chance to run.
2432 	 */
2433 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2434 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2435 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2436 			tdq->tdq_ridx = tdq->tdq_idx;
2437 	}
2438 	ts = td_get_sched(td);
2439 	sched_pctcpu_update(ts, 1);
2440 	if ((td->td_pri_class & PRI_FIFO_BIT) || TD_IS_IDLETHREAD(td))
2441 		return;
2442 
2443 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2444 		/*
2445 		 * We used a tick; charge it to the thread so
2446 		 * that we can compute our interactivity.
2447 		 */
2448 		td_get_sched(td)->ts_runtime += tickincr * cnt;
2449 		sched_interact_update(td);
2450 		sched_priority(td);
2451 	}
2452 
2453 	/*
2454 	 * Force a context switch if the current thread has used up a full
2455 	 * time slice (default is 100ms).
2456 	 */
2457 	ts->ts_slice += cnt;
2458 	if (ts->ts_slice >= tdq_slice(tdq)) {
2459 		ts->ts_slice = 0;
2460 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2461 	}
2462 }
2463 
2464 u_int
2465 sched_estcpu(struct thread *td __unused)
2466 {
2467 
2468 	return (0);
2469 }
2470 
2471 /*
2472  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2473  * cooperative idle threads.
2474  */
2475 int
2476 sched_runnable(void)
2477 {
2478 	struct tdq *tdq;
2479 	int load;
2480 
2481 	load = 1;
2482 
2483 	tdq = TDQ_SELF();
2484 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2485 		if (tdq->tdq_load > 0)
2486 			goto out;
2487 	} else
2488 		if (tdq->tdq_load - 1 > 0)
2489 			goto out;
2490 	load = 0;
2491 out:
2492 	return (load);
2493 }
2494 
2495 /*
2496  * Choose the highest priority thread to run.  The thread is removed from
2497  * the run-queue while running however the load remains.  For SMP we set
2498  * the tdq in the global idle bitmask if it idles here.
2499  */
2500 struct thread *
2501 sched_choose(void)
2502 {
2503 	struct thread *td;
2504 	struct tdq *tdq;
2505 
2506 	tdq = TDQ_SELF();
2507 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2508 	td = tdq_choose(tdq);
2509 	if (td) {
2510 		tdq_runq_rem(tdq, td);
2511 		tdq->tdq_lowpri = td->td_priority;
2512 		return (td);
2513 	}
2514 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2515 	return (PCPU_GET(idlethread));
2516 }
2517 
2518 /*
2519  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2520  * we always request it once we exit a critical section.
2521  */
2522 static inline void
2523 sched_setpreempt(struct thread *td)
2524 {
2525 	struct thread *ctd;
2526 	int cpri;
2527 	int pri;
2528 
2529 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2530 
2531 	ctd = curthread;
2532 	pri = td->td_priority;
2533 	cpri = ctd->td_priority;
2534 	if (pri < cpri)
2535 		ctd->td_flags |= TDF_NEEDRESCHED;
2536 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2537 		return;
2538 	if (!sched_shouldpreempt(pri, cpri, 0))
2539 		return;
2540 	ctd->td_owepreempt = 1;
2541 }
2542 
2543 /*
2544  * Add a thread to a thread queue.  Select the appropriate runq and add the
2545  * thread to it.  This is the internal function called when the tdq is
2546  * predetermined.
2547  */
2548 void
2549 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2550 {
2551 
2552 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2553 	THREAD_LOCK_BLOCKED_ASSERT(td, MA_OWNED);
2554 	KASSERT((td->td_inhibitors == 0),
2555 	    ("sched_add: trying to run inhibited thread"));
2556 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2557 	    ("sched_add: bad thread state"));
2558 	KASSERT(td->td_flags & TDF_INMEM,
2559 	    ("sched_add: thread swapped out"));
2560 
2561 	if (td->td_priority < tdq->tdq_lowpri)
2562 		tdq->tdq_lowpri = td->td_priority;
2563 	tdq_runq_add(tdq, td, flags);
2564 	tdq_load_add(tdq, td);
2565 }
2566 
2567 /*
2568  * Select the target thread queue and add a thread to it.  Request
2569  * preemption or IPI a remote processor if required.
2570  *
2571  * Requires the thread lock on entry, drops on exit.
2572  */
2573 void
2574 sched_add(struct thread *td, int flags)
2575 {
2576 	struct tdq *tdq;
2577 #ifdef SMP
2578 	int cpu;
2579 #endif
2580 
2581 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2582 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2583 	    sched_tdname(curthread));
2584 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2585 	    KTR_ATTR_LINKED, sched_tdname(td));
2586 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2587 	    flags & SRQ_PREEMPTED);
2588 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2589 	/*
2590 	 * Recalculate the priority before we select the target cpu or
2591 	 * run-queue.
2592 	 */
2593 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2594 		sched_priority(td);
2595 #ifdef SMP
2596 	/*
2597 	 * Pick the destination cpu and if it isn't ours transfer to the
2598 	 * target cpu.
2599 	 */
2600 	cpu = sched_pickcpu(td, flags);
2601 	tdq = sched_setcpu(td, cpu, flags);
2602 	tdq_add(tdq, td, flags);
2603 	if (cpu != PCPU_GET(cpuid))
2604 		tdq_notify(tdq, td);
2605 	else if (!(flags & SRQ_YIELDING))
2606 		sched_setpreempt(td);
2607 #else
2608 	tdq = TDQ_SELF();
2609 	TDQ_LOCK(tdq);
2610 	/*
2611 	 * Now that the thread is moving to the run-queue, set the lock
2612 	 * to the scheduler's lock.
2613 	 */
2614 	if ((flags & SRQ_HOLD) != 0)
2615 		td->td_lock = TDQ_LOCKPTR(tdq);
2616 	else
2617 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
2618 	tdq_add(tdq, td, flags);
2619 	if (!(flags & SRQ_YIELDING))
2620 		sched_setpreempt(td);
2621 #endif
2622 	if (!(flags & SRQ_HOLDTD))
2623 		thread_unlock(td);
2624 }
2625 
2626 /*
2627  * Remove a thread from a run-queue without running it.  This is used
2628  * when we're stealing a thread from a remote queue.  Otherwise all threads
2629  * exit by calling sched_exit_thread() and sched_throw() themselves.
2630  */
2631 void
2632 sched_rem(struct thread *td)
2633 {
2634 	struct tdq *tdq;
2635 
2636 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2637 	    "prio:%d", td->td_priority);
2638 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2639 	tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2640 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2641 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2642 	KASSERT(TD_ON_RUNQ(td),
2643 	    ("sched_rem: thread not on run queue"));
2644 	tdq_runq_rem(tdq, td);
2645 	tdq_load_rem(tdq, td);
2646 	TD_SET_CAN_RUN(td);
2647 	if (td->td_priority == tdq->tdq_lowpri)
2648 		tdq_setlowpri(tdq, NULL);
2649 }
2650 
2651 /*
2652  * Fetch cpu utilization information.  Updates on demand.
2653  */
2654 fixpt_t
2655 sched_pctcpu(struct thread *td)
2656 {
2657 	fixpt_t pctcpu;
2658 	struct td_sched *ts;
2659 
2660 	pctcpu = 0;
2661 	ts = td_get_sched(td);
2662 
2663 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2664 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2665 	if (ts->ts_ticks) {
2666 		int rtick;
2667 
2668 		/* How many rtick per second ? */
2669 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2670 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2671 	}
2672 
2673 	return (pctcpu);
2674 }
2675 
2676 /*
2677  * Enforce affinity settings for a thread.  Called after adjustments to
2678  * cpumask.
2679  */
2680 void
2681 sched_affinity(struct thread *td)
2682 {
2683 #ifdef SMP
2684 	struct td_sched *ts;
2685 
2686 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2687 	ts = td_get_sched(td);
2688 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2689 		return;
2690 	if (TD_ON_RUNQ(td)) {
2691 		sched_rem(td);
2692 		sched_add(td, SRQ_BORING);
2693 		return;
2694 	}
2695 	if (!TD_IS_RUNNING(td))
2696 		return;
2697 	/*
2698 	 * Force a switch before returning to userspace.  If the
2699 	 * target thread is not running locally send an ipi to force
2700 	 * the issue.
2701 	 */
2702 	td->td_flags |= TDF_NEEDRESCHED;
2703 	if (td != curthread)
2704 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2705 #endif
2706 }
2707 
2708 /*
2709  * Bind a thread to a target cpu.
2710  */
2711 void
2712 sched_bind(struct thread *td, int cpu)
2713 {
2714 	struct td_sched *ts;
2715 
2716 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2717 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2718 	ts = td_get_sched(td);
2719 	if (ts->ts_flags & TSF_BOUND)
2720 		sched_unbind(td);
2721 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2722 	ts->ts_flags |= TSF_BOUND;
2723 	sched_pin();
2724 	if (PCPU_GET(cpuid) == cpu)
2725 		return;
2726 	ts->ts_cpu = cpu;
2727 	/* When we return from mi_switch we'll be on the correct cpu. */
2728 	mi_switch(SW_VOL);
2729 	thread_lock(td);
2730 }
2731 
2732 /*
2733  * Release a bound thread.
2734  */
2735 void
2736 sched_unbind(struct thread *td)
2737 {
2738 	struct td_sched *ts;
2739 
2740 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2741 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2742 	ts = td_get_sched(td);
2743 	if ((ts->ts_flags & TSF_BOUND) == 0)
2744 		return;
2745 	ts->ts_flags &= ~TSF_BOUND;
2746 	sched_unpin();
2747 }
2748 
2749 int
2750 sched_is_bound(struct thread *td)
2751 {
2752 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2753 	return (td_get_sched(td)->ts_flags & TSF_BOUND);
2754 }
2755 
2756 /*
2757  * Basic yield call.
2758  */
2759 void
2760 sched_relinquish(struct thread *td)
2761 {
2762 	thread_lock(td);
2763 	mi_switch(SW_VOL | SWT_RELINQUISH);
2764 }
2765 
2766 /*
2767  * Return the total system load.
2768  */
2769 int
2770 sched_load(void)
2771 {
2772 #ifdef SMP
2773 	int total;
2774 	int i;
2775 
2776 	total = 0;
2777 	CPU_FOREACH(i)
2778 		total += TDQ_CPU(i)->tdq_sysload;
2779 	return (total);
2780 #else
2781 	return (TDQ_SELF()->tdq_sysload);
2782 #endif
2783 }
2784 
2785 int
2786 sched_sizeof_proc(void)
2787 {
2788 	return (sizeof(struct proc));
2789 }
2790 
2791 int
2792 sched_sizeof_thread(void)
2793 {
2794 	return (sizeof(struct thread) + sizeof(struct td_sched));
2795 }
2796 
2797 #ifdef SMP
2798 #define	TDQ_IDLESPIN(tdq)						\
2799     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2800 #else
2801 #define	TDQ_IDLESPIN(tdq)	1
2802 #endif
2803 
2804 /*
2805  * The actual idle process.
2806  */
2807 void
2808 sched_idletd(void *dummy)
2809 {
2810 	struct thread *td;
2811 	struct tdq *tdq;
2812 	int oldswitchcnt, switchcnt;
2813 	int i;
2814 
2815 	mtx_assert(&Giant, MA_NOTOWNED);
2816 	td = curthread;
2817 	tdq = TDQ_SELF();
2818 	THREAD_NO_SLEEPING();
2819 	oldswitchcnt = -1;
2820 	for (;;) {
2821 		if (tdq->tdq_load) {
2822 			thread_lock(td);
2823 			mi_switch(SW_VOL | SWT_IDLE);
2824 		}
2825 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2826 #ifdef SMP
2827 		if (always_steal || switchcnt != oldswitchcnt) {
2828 			oldswitchcnt = switchcnt;
2829 			if (tdq_idled(tdq) == 0)
2830 				continue;
2831 		}
2832 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2833 #else
2834 		oldswitchcnt = switchcnt;
2835 #endif
2836 		/*
2837 		 * If we're switching very frequently, spin while checking
2838 		 * for load rather than entering a low power state that
2839 		 * may require an IPI.  However, don't do any busy
2840 		 * loops while on SMT machines as this simply steals
2841 		 * cycles from cores doing useful work.
2842 		 */
2843 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2844 			for (i = 0; i < sched_idlespins; i++) {
2845 				if (tdq->tdq_load)
2846 					break;
2847 				cpu_spinwait();
2848 			}
2849 		}
2850 
2851 		/* If there was context switch during spin, restart it. */
2852 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2853 		if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2854 			continue;
2855 
2856 		/* Run main MD idle handler. */
2857 		tdq->tdq_cpu_idle = 1;
2858 		/*
2859 		 * Make sure that tdq_cpu_idle update is globally visible
2860 		 * before cpu_idle() read tdq_load.  The order is important
2861 		 * to avoid race with tdq_notify.
2862 		 */
2863 		atomic_thread_fence_seq_cst();
2864 		/*
2865 		 * Checking for again after the fence picks up assigned
2866 		 * threads often enough to make it worthwhile to do so in
2867 		 * order to avoid calling cpu_idle().
2868 		 */
2869 		if (tdq->tdq_load != 0) {
2870 			tdq->tdq_cpu_idle = 0;
2871 			continue;
2872 		}
2873 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2874 		tdq->tdq_cpu_idle = 0;
2875 
2876 		/*
2877 		 * Account thread-less hardware interrupts and
2878 		 * other wakeup reasons equal to context switches.
2879 		 */
2880 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2881 		if (switchcnt != oldswitchcnt)
2882 			continue;
2883 		tdq->tdq_switchcnt++;
2884 		oldswitchcnt++;
2885 	}
2886 }
2887 
2888 /*
2889  * A CPU is entering for the first time or a thread is exiting.
2890  */
2891 void
2892 sched_throw(struct thread *td)
2893 {
2894 	struct thread *newtd;
2895 	struct tdq *tdq;
2896 
2897 	if (td == NULL) {
2898 #ifdef SMP
2899 		PCPU_SET(sched, DPCPU_PTR(tdq));
2900 #endif
2901 		/* Correct spinlock nesting and acquire the correct lock. */
2902 		tdq = TDQ_SELF();
2903 		TDQ_LOCK(tdq);
2904 		spinlock_exit();
2905 		PCPU_SET(switchtime, cpu_ticks());
2906 		PCPU_SET(switchticks, ticks);
2907 		PCPU_GET(idlethread)->td_lock = TDQ_LOCKPTR(tdq);
2908 	} else {
2909 		tdq = TDQ_SELF();
2910 		THREAD_LOCK_ASSERT(td, MA_OWNED);
2911 		THREAD_LOCKPTR_ASSERT(td, TDQ_LOCKPTR(tdq));
2912 		tdq_load_rem(tdq, td);
2913 		td->td_lastcpu = td->td_oncpu;
2914 		td->td_oncpu = NOCPU;
2915 	}
2916 	newtd = choosethread();
2917 	spinlock_enter();
2918 	TDQ_UNLOCK(tdq);
2919 	KASSERT(curthread->td_md.md_spinlock_count == 1,
2920 	    ("invalid count %d", curthread->td_md.md_spinlock_count));
2921 	cpu_throw(td, newtd);		/* doesn't return */
2922 }
2923 
2924 /*
2925  * This is called from fork_exit().  Just acquire the correct locks and
2926  * let fork do the rest of the work.
2927  */
2928 void
2929 sched_fork_exit(struct thread *td)
2930 {
2931 	struct tdq *tdq;
2932 	int cpuid;
2933 
2934 	/*
2935 	 * Finish setting up thread glue so that it begins execution in a
2936 	 * non-nested critical section with the scheduler lock held.
2937 	 */
2938 	KASSERT(curthread->td_md.md_spinlock_count == 1,
2939 	    ("invalid count %d", curthread->td_md.md_spinlock_count));
2940 	cpuid = PCPU_GET(cpuid);
2941 	tdq = TDQ_SELF();
2942 	TDQ_LOCK(tdq);
2943 	spinlock_exit();
2944 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2945 	td->td_oncpu = cpuid;
2946 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2947 	    "prio:%d", td->td_priority);
2948 	SDT_PROBE0(sched, , , on__cpu);
2949 }
2950 
2951 /*
2952  * Create on first use to catch odd startup conditons.
2953  */
2954 char *
2955 sched_tdname(struct thread *td)
2956 {
2957 #ifdef KTR
2958 	struct td_sched *ts;
2959 
2960 	ts = td_get_sched(td);
2961 	if (ts->ts_name[0] == '\0')
2962 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2963 		    "%s tid %d", td->td_name, td->td_tid);
2964 	return (ts->ts_name);
2965 #else
2966 	return (td->td_name);
2967 #endif
2968 }
2969 
2970 #ifdef KTR
2971 void
2972 sched_clear_tdname(struct thread *td)
2973 {
2974 	struct td_sched *ts;
2975 
2976 	ts = td_get_sched(td);
2977 	ts->ts_name[0] = '\0';
2978 }
2979 #endif
2980 
2981 #ifdef SMP
2982 
2983 /*
2984  * Build the CPU topology dump string. Is recursively called to collect
2985  * the topology tree.
2986  */
2987 static int
2988 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2989     int indent)
2990 {
2991 	char cpusetbuf[CPUSETBUFSIZ];
2992 	int i, first;
2993 
2994 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2995 	    "", 1 + indent / 2, cg->cg_level);
2996 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2997 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2998 	first = TRUE;
2999 	for (i = 0; i < MAXCPU; i++) {
3000 		if (CPU_ISSET(i, &cg->cg_mask)) {
3001 			if (!first)
3002 				sbuf_printf(sb, ", ");
3003 			else
3004 				first = FALSE;
3005 			sbuf_printf(sb, "%d", i);
3006 		}
3007 	}
3008 	sbuf_printf(sb, "</cpu>\n");
3009 
3010 	if (cg->cg_flags != 0) {
3011 		sbuf_printf(sb, "%*s <flags>", indent, "");
3012 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
3013 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
3014 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
3015 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
3016 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
3017 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
3018 		sbuf_printf(sb, "</flags>\n");
3019 	}
3020 
3021 	if (cg->cg_children > 0) {
3022 		sbuf_printf(sb, "%*s <children>\n", indent, "");
3023 		for (i = 0; i < cg->cg_children; i++)
3024 			sysctl_kern_sched_topology_spec_internal(sb,
3025 			    &cg->cg_child[i], indent+2);
3026 		sbuf_printf(sb, "%*s </children>\n", indent, "");
3027 	}
3028 	sbuf_printf(sb, "%*s</group>\n", indent, "");
3029 	return (0);
3030 }
3031 
3032 /*
3033  * Sysctl handler for retrieving topology dump. It's a wrapper for
3034  * the recursive sysctl_kern_smp_topology_spec_internal().
3035  */
3036 static int
3037 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
3038 {
3039 	struct sbuf *topo;
3040 	int err;
3041 
3042 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
3043 
3044 	topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
3045 	if (topo == NULL)
3046 		return (ENOMEM);
3047 
3048 	sbuf_printf(topo, "<groups>\n");
3049 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
3050 	sbuf_printf(topo, "</groups>\n");
3051 
3052 	if (err == 0) {
3053 		err = sbuf_finish(topo);
3054 	}
3055 	sbuf_delete(topo);
3056 	return (err);
3057 }
3058 
3059 #endif
3060 
3061 static int
3062 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
3063 {
3064 	int error, new_val, period;
3065 
3066 	period = 1000000 / realstathz;
3067 	new_val = period * sched_slice;
3068 	error = sysctl_handle_int(oidp, &new_val, 0, req);
3069 	if (error != 0 || req->newptr == NULL)
3070 		return (error);
3071 	if (new_val <= 0)
3072 		return (EINVAL);
3073 	sched_slice = imax(1, (new_val + period / 2) / period);
3074 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
3075 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
3076 	    realstathz);
3077 	return (0);
3078 }
3079 
3080 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
3081 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
3082     "Scheduler name");
3083 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
3084     NULL, 0, sysctl_kern_quantum, "I",
3085     "Quantum for timeshare threads in microseconds");
3086 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
3087     "Quantum for timeshare threads in stathz ticks");
3088 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
3089     "Interactivity score threshold");
3090 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
3091     &preempt_thresh, 0,
3092     "Maximal (lowest) priority for preemption");
3093 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
3094     "Assign static kernel priorities to sleeping threads");
3095 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
3096     "Number of times idle thread will spin waiting for new work");
3097 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
3098     &sched_idlespinthresh, 0,
3099     "Threshold before we will permit idle thread spinning");
3100 #ifdef SMP
3101 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
3102     "Number of hz ticks to keep thread affinity for");
3103 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
3104     "Enables the long-term load balancer");
3105 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
3106     &balance_interval, 0,
3107     "Average period in stathz ticks to run the long-term balancer");
3108 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
3109     "Attempts to steal work from other cores before idling");
3110 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
3111     "Minimum load on remote CPU before we'll steal");
3112 SYSCTL_INT(_kern_sched, OID_AUTO, trysteal_limit, CTLFLAG_RW, &trysteal_limit,
3113     0, "Topological distance limit for stealing threads in sched_switch()");
3114 SYSCTL_INT(_kern_sched, OID_AUTO, always_steal, CTLFLAG_RW, &always_steal, 0,
3115     "Always run the stealer from the idle thread");
3116 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
3117     CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
3118     "XML dump of detected CPU topology");
3119 #endif
3120 
3121 /* ps compat.  All cpu percentages from ULE are weighted. */
3122 static int ccpu = 0;
3123 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
3124