xref: /freebsd/sys/kern/sched_ule.c (revision b70d2a2aa5003027b422e62435ab5bb9390d543c)
1 /*-
2  * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice unmodified, this list of conditions, and the following
10  *    disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25  */
26 
27 /*
28  * This file implements the ULE scheduler.  ULE supports independent CPU
29  * run queues and fine grain locking.  It has superior interactive
30  * performance under load even on uni-processor systems.
31  *
32  * etymology:
33  *   ULE is the last three letters in schedule.  It owes its name to a
34  * generic user created for a scheduling system by Paul Mikesell at
35  * Isilon Systems and a general lack of creativity on the part of the author.
36  */
37 
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
40 
41 #include "opt_hwpmc_hooks.h"
42 #include "opt_sched.h"
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/kdb.h>
47 #include <sys/kernel.h>
48 #include <sys/ktr.h>
49 #include <sys/limits.h>
50 #include <sys/lock.h>
51 #include <sys/mutex.h>
52 #include <sys/proc.h>
53 #include <sys/resource.h>
54 #include <sys/resourcevar.h>
55 #include <sys/sched.h>
56 #include <sys/sdt.h>
57 #include <sys/smp.h>
58 #include <sys/sx.h>
59 #include <sys/sysctl.h>
60 #include <sys/sysproto.h>
61 #include <sys/turnstile.h>
62 #include <sys/umtx.h>
63 #include <sys/vmmeter.h>
64 #include <sys/cpuset.h>
65 #include <sys/sbuf.h>
66 
67 #ifdef HWPMC_HOOKS
68 #include <sys/pmckern.h>
69 #endif
70 
71 #ifdef KDTRACE_HOOKS
72 #include <sys/dtrace_bsd.h>
73 int				dtrace_vtime_active;
74 dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
75 #endif
76 
77 #include <machine/cpu.h>
78 #include <machine/smp.h>
79 
80 #define	KTR_ULE	0
81 
82 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
83 #define	TDQ_NAME_LEN	(sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
84 #define	TDQ_LOADNAME_LEN	(sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
85 
86 /*
87  * Thread scheduler specific section.  All fields are protected
88  * by the thread lock.
89  */
90 struct td_sched {
91 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
92 	short		ts_flags;	/* TSF_* flags. */
93 	int		ts_cpu;		/* CPU that we have affinity for. */
94 	int		ts_rltick;	/* Real last tick, for affinity. */
95 	int		ts_slice;	/* Ticks of slice remaining. */
96 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
97 	u_int		ts_runtime;	/* Number of ticks we were running */
98 	int		ts_ltick;	/* Last tick that we were running on */
99 	int		ts_ftick;	/* First tick that we were running on */
100 	int		ts_ticks;	/* Tick count */
101 #ifdef KTR
102 	char		ts_name[TS_NAME_LEN];
103 #endif
104 };
105 /* flags kept in ts_flags */
106 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
107 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
108 
109 #define	THREAD_CAN_MIGRATE(td)	((td)->td_pinned == 0)
110 #define	THREAD_CAN_SCHED(td, cpu)	\
111     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
112 
113 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <=
114     sizeof(struct thread0_storage),
115     "increase struct thread0_storage.t0st_sched size");
116 
117 /*
118  * Priority ranges used for interactive and non-interactive timeshare
119  * threads.  The timeshare priorities are split up into four ranges.
120  * The first range handles interactive threads.  The last three ranges
121  * (NHALF, x, and NHALF) handle non-interactive threads with the outer
122  * ranges supporting nice values.
123  */
124 #define	PRI_TIMESHARE_RANGE	(PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
125 #define	PRI_INTERACT_RANGE	((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
126 #define	PRI_BATCH_RANGE		(PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
127 
128 #define	PRI_MIN_INTERACT	PRI_MIN_TIMESHARE
129 #define	PRI_MAX_INTERACT	(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
130 #define	PRI_MIN_BATCH		(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
131 #define	PRI_MAX_BATCH		PRI_MAX_TIMESHARE
132 
133 /*
134  * Cpu percentage computation macros and defines.
135  *
136  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
137  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
138  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
139  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
140  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
141  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
142  */
143 #define	SCHED_TICK_SECS		10
144 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
145 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
146 #define	SCHED_TICK_SHIFT	10
147 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
148 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
149 
150 /*
151  * These macros determine priorities for non-interactive threads.  They are
152  * assigned a priority based on their recent cpu utilization as expressed
153  * by the ratio of ticks to the tick total.  NHALF priorities at the start
154  * and end of the MIN to MAX timeshare range are only reachable with negative
155  * or positive nice respectively.
156  *
157  * PRI_RANGE:	Priority range for utilization dependent priorities.
158  * PRI_NRESV:	Number of nice values.
159  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
160  * PRI_NICE:	Determines the part of the priority inherited from nice.
161  */
162 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
163 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
164 #define	SCHED_PRI_MIN		(PRI_MIN_BATCH + SCHED_PRI_NHALF)
165 #define	SCHED_PRI_MAX		(PRI_MAX_BATCH - SCHED_PRI_NHALF)
166 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
167 #define	SCHED_PRI_TICKS(ts)						\
168     (SCHED_TICK_HZ((ts)) /						\
169     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
170 #define	SCHED_PRI_NICE(nice)	(nice)
171 
172 /*
173  * These determine the interactivity of a process.  Interactivity differs from
174  * cpu utilization in that it expresses the voluntary time slept vs time ran
175  * while cpu utilization includes all time not running.  This more accurately
176  * models the intent of the thread.
177  *
178  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
179  *		before throttling back.
180  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
181  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
182  * INTERACT_THRESH:	Threshold for placement on the current runq.
183  */
184 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
185 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
186 #define	SCHED_INTERACT_MAX	(100)
187 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
188 #define	SCHED_INTERACT_THRESH	(30)
189 
190 /*
191  * These parameters determine the slice behavior for batch work.
192  */
193 #define	SCHED_SLICE_DEFAULT_DIVISOR	10	/* ~94 ms, 12 stathz ticks. */
194 #define	SCHED_SLICE_MIN_DIVISOR		6	/* DEFAULT/MIN = ~16 ms. */
195 
196 /* Flags kept in td_flags. */
197 #define	TDF_SLICEEND	TDF_SCHED2	/* Thread time slice is over. */
198 
199 /*
200  * tickincr:		Converts a stathz tick into a hz domain scaled by
201  *			the shift factor.  Without the shift the error rate
202  *			due to rounding would be unacceptably high.
203  * realstathz:		stathz is sometimes 0 and run off of hz.
204  * sched_slice:		Runtime of each thread before rescheduling.
205  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
206  */
207 static int sched_interact = SCHED_INTERACT_THRESH;
208 static int tickincr = 8 << SCHED_TICK_SHIFT;
209 static int realstathz = 127;	/* reset during boot. */
210 static int sched_slice = 10;	/* reset during boot. */
211 static int sched_slice_min = 1;	/* reset during boot. */
212 #ifdef PREEMPTION
213 #ifdef FULL_PREEMPTION
214 static int preempt_thresh = PRI_MAX_IDLE;
215 #else
216 static int preempt_thresh = PRI_MIN_KERN;
217 #endif
218 #else
219 static int preempt_thresh = 0;
220 #endif
221 static int static_boost = PRI_MIN_BATCH;
222 static int sched_idlespins = 10000;
223 static int sched_idlespinthresh = -1;
224 
225 /*
226  * tdq - per processor runqs and statistics.  All fields are protected by the
227  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
228  * locking in sched_pickcpu();
229  */
230 struct tdq {
231 	/*
232 	 * Ordered to improve efficiency of cpu_search() and switch().
233 	 * tdq_lock is padded to avoid false sharing with tdq_load and
234 	 * tdq_cpu_idle.
235 	 */
236 	struct mtx_padalign tdq_lock;		/* run queue lock. */
237 	struct cpu_group *tdq_cg;		/* Pointer to cpu topology. */
238 	volatile int	tdq_load;		/* Aggregate load. */
239 	volatile int	tdq_cpu_idle;		/* cpu_idle() is active. */
240 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
241 	int		tdq_transferable;	/* Transferable thread count. */
242 	short		tdq_switchcnt;		/* Switches this tick. */
243 	short		tdq_oldswitchcnt;	/* Switches last tick. */
244 	u_char		tdq_lowpri;		/* Lowest priority thread. */
245 	u_char		tdq_ipipending;		/* IPI pending. */
246 	u_char		tdq_idx;		/* Current insert index. */
247 	u_char		tdq_ridx;		/* Current removal index. */
248 	struct runq	tdq_realtime;		/* real-time run queue. */
249 	struct runq	tdq_timeshare;		/* timeshare run queue. */
250 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
251 	char		tdq_name[TDQ_NAME_LEN];
252 #ifdef KTR
253 	char		tdq_loadname[TDQ_LOADNAME_LEN];
254 #endif
255 } __aligned(64);
256 
257 /* Idle thread states and config. */
258 #define	TDQ_RUNNING	1
259 #define	TDQ_IDLE	2
260 
261 #ifdef SMP
262 struct cpu_group *cpu_top;		/* CPU topology */
263 
264 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
265 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
266 
267 /*
268  * Run-time tunables.
269  */
270 static int rebalance = 1;
271 static int balance_interval = 128;	/* Default set in sched_initticks(). */
272 static int affinity;
273 static int steal_idle = 1;
274 static int steal_thresh = 2;
275 
276 /*
277  * One thread queue per processor.
278  */
279 static struct tdq	tdq_cpu[MAXCPU];
280 static struct tdq	*balance_tdq;
281 static int balance_ticks;
282 static DPCPU_DEFINE(uint32_t, randomval);
283 
284 #define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
285 #define	TDQ_CPU(x)	(&tdq_cpu[(x)])
286 #define	TDQ_ID(x)	((int)((x) - tdq_cpu))
287 #else	/* !SMP */
288 static struct tdq	tdq_cpu;
289 
290 #define	TDQ_ID(x)	(0)
291 #define	TDQ_SELF()	(&tdq_cpu)
292 #define	TDQ_CPU(x)	(&tdq_cpu)
293 #endif
294 
295 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
296 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
297 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
298 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
299 #define	TDQ_LOCKPTR(t)		((struct mtx *)(&(t)->tdq_lock))
300 
301 static void sched_priority(struct thread *);
302 static void sched_thread_priority(struct thread *, u_char);
303 static int sched_interact_score(struct thread *);
304 static void sched_interact_update(struct thread *);
305 static void sched_interact_fork(struct thread *);
306 static void sched_pctcpu_update(struct td_sched *, int);
307 
308 /* Operations on per processor queues */
309 static struct thread *tdq_choose(struct tdq *);
310 static void tdq_setup(struct tdq *);
311 static void tdq_load_add(struct tdq *, struct thread *);
312 static void tdq_load_rem(struct tdq *, struct thread *);
313 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
314 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
315 static inline int sched_shouldpreempt(int, int, int);
316 void tdq_print(int cpu);
317 static void runq_print(struct runq *rq);
318 static void tdq_add(struct tdq *, struct thread *, int);
319 #ifdef SMP
320 static int tdq_move(struct tdq *, struct tdq *);
321 static int tdq_idled(struct tdq *);
322 static void tdq_notify(struct tdq *, struct thread *);
323 static struct thread *tdq_steal(struct tdq *, int);
324 static struct thread *runq_steal(struct runq *, int);
325 static int sched_pickcpu(struct thread *, int);
326 static void sched_balance(void);
327 static int sched_balance_pair(struct tdq *, struct tdq *);
328 static inline struct tdq *sched_setcpu(struct thread *, int, int);
329 static inline void thread_unblock_switch(struct thread *, struct mtx *);
330 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
331 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
332 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
333     struct cpu_group *cg, int indent);
334 #endif
335 
336 static void sched_setup(void *dummy);
337 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
338 
339 static void sched_initticks(void *dummy);
340 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
341     NULL);
342 
343 SDT_PROVIDER_DEFINE(sched);
344 
345 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
346     "struct proc *", "uint8_t");
347 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
348     "struct proc *", "void *");
349 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
350     "struct proc *", "void *", "int");
351 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
352     "struct proc *", "uint8_t", "struct thread *");
353 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
354 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
355     "struct proc *");
356 SDT_PROBE_DEFINE(sched, , , on__cpu);
357 SDT_PROBE_DEFINE(sched, , , remain__cpu);
358 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
359     "struct proc *");
360 
361 /*
362  * Print the threads waiting on a run-queue.
363  */
364 static void
365 runq_print(struct runq *rq)
366 {
367 	struct rqhead *rqh;
368 	struct thread *td;
369 	int pri;
370 	int j;
371 	int i;
372 
373 	for (i = 0; i < RQB_LEN; i++) {
374 		printf("\t\trunq bits %d 0x%zx\n",
375 		    i, rq->rq_status.rqb_bits[i]);
376 		for (j = 0; j < RQB_BPW; j++)
377 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
378 				pri = j + (i << RQB_L2BPW);
379 				rqh = &rq->rq_queues[pri];
380 				TAILQ_FOREACH(td, rqh, td_runq) {
381 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
382 					    td, td->td_name, td->td_priority,
383 					    td->td_rqindex, pri);
384 				}
385 			}
386 	}
387 }
388 
389 /*
390  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
391  */
392 void
393 tdq_print(int cpu)
394 {
395 	struct tdq *tdq;
396 
397 	tdq = TDQ_CPU(cpu);
398 
399 	printf("tdq %d:\n", TDQ_ID(tdq));
400 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
401 	printf("\tLock name:      %s\n", tdq->tdq_name);
402 	printf("\tload:           %d\n", tdq->tdq_load);
403 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
404 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
405 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
406 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
407 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
408 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
409 	printf("\trealtime runq:\n");
410 	runq_print(&tdq->tdq_realtime);
411 	printf("\ttimeshare runq:\n");
412 	runq_print(&tdq->tdq_timeshare);
413 	printf("\tidle runq:\n");
414 	runq_print(&tdq->tdq_idle);
415 }
416 
417 static inline int
418 sched_shouldpreempt(int pri, int cpri, int remote)
419 {
420 	/*
421 	 * If the new priority is not better than the current priority there is
422 	 * nothing to do.
423 	 */
424 	if (pri >= cpri)
425 		return (0);
426 	/*
427 	 * Always preempt idle.
428 	 */
429 	if (cpri >= PRI_MIN_IDLE)
430 		return (1);
431 	/*
432 	 * If preemption is disabled don't preempt others.
433 	 */
434 	if (preempt_thresh == 0)
435 		return (0);
436 	/*
437 	 * Preempt if we exceed the threshold.
438 	 */
439 	if (pri <= preempt_thresh)
440 		return (1);
441 	/*
442 	 * If we're interactive or better and there is non-interactive
443 	 * or worse running preempt only remote processors.
444 	 */
445 	if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
446 		return (1);
447 	return (0);
448 }
449 
450 /*
451  * Add a thread to the actual run-queue.  Keeps transferable counts up to
452  * date with what is actually on the run-queue.  Selects the correct
453  * queue position for timeshare threads.
454  */
455 static __inline void
456 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
457 {
458 	struct td_sched *ts;
459 	u_char pri;
460 
461 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
462 	THREAD_LOCK_ASSERT(td, MA_OWNED);
463 
464 	pri = td->td_priority;
465 	ts = td_get_sched(td);
466 	TD_SET_RUNQ(td);
467 	if (THREAD_CAN_MIGRATE(td)) {
468 		tdq->tdq_transferable++;
469 		ts->ts_flags |= TSF_XFERABLE;
470 	}
471 	if (pri < PRI_MIN_BATCH) {
472 		ts->ts_runq = &tdq->tdq_realtime;
473 	} else if (pri <= PRI_MAX_BATCH) {
474 		ts->ts_runq = &tdq->tdq_timeshare;
475 		KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
476 			("Invalid priority %d on timeshare runq", pri));
477 		/*
478 		 * This queue contains only priorities between MIN and MAX
479 		 * realtime.  Use the whole queue to represent these values.
480 		 */
481 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
482 			pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
483 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
484 			/*
485 			 * This effectively shortens the queue by one so we
486 			 * can have a one slot difference between idx and
487 			 * ridx while we wait for threads to drain.
488 			 */
489 			if (tdq->tdq_ridx != tdq->tdq_idx &&
490 			    pri == tdq->tdq_ridx)
491 				pri = (unsigned char)(pri - 1) % RQ_NQS;
492 		} else
493 			pri = tdq->tdq_ridx;
494 		runq_add_pri(ts->ts_runq, td, pri, flags);
495 		return;
496 	} else
497 		ts->ts_runq = &tdq->tdq_idle;
498 	runq_add(ts->ts_runq, td, flags);
499 }
500 
501 /*
502  * Remove a thread from a run-queue.  This typically happens when a thread
503  * is selected to run.  Running threads are not on the queue and the
504  * transferable count does not reflect them.
505  */
506 static __inline void
507 tdq_runq_rem(struct tdq *tdq, struct thread *td)
508 {
509 	struct td_sched *ts;
510 
511 	ts = td_get_sched(td);
512 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
513 	KASSERT(ts->ts_runq != NULL,
514 	    ("tdq_runq_remove: thread %p null ts_runq", td));
515 	if (ts->ts_flags & TSF_XFERABLE) {
516 		tdq->tdq_transferable--;
517 		ts->ts_flags &= ~TSF_XFERABLE;
518 	}
519 	if (ts->ts_runq == &tdq->tdq_timeshare) {
520 		if (tdq->tdq_idx != tdq->tdq_ridx)
521 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
522 		else
523 			runq_remove_idx(ts->ts_runq, td, NULL);
524 	} else
525 		runq_remove(ts->ts_runq, td);
526 }
527 
528 /*
529  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
530  * for this thread to the referenced thread queue.
531  */
532 static void
533 tdq_load_add(struct tdq *tdq, struct thread *td)
534 {
535 
536 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
537 	THREAD_LOCK_ASSERT(td, MA_OWNED);
538 
539 	tdq->tdq_load++;
540 	if ((td->td_flags & TDF_NOLOAD) == 0)
541 		tdq->tdq_sysload++;
542 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
543 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
544 }
545 
546 /*
547  * Remove the load from a thread that is transitioning to a sleep state or
548  * exiting.
549  */
550 static void
551 tdq_load_rem(struct tdq *tdq, struct thread *td)
552 {
553 
554 	THREAD_LOCK_ASSERT(td, MA_OWNED);
555 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
556 	KASSERT(tdq->tdq_load != 0,
557 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
558 
559 	tdq->tdq_load--;
560 	if ((td->td_flags & TDF_NOLOAD) == 0)
561 		tdq->tdq_sysload--;
562 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
563 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
564 }
565 
566 /*
567  * Bound timeshare latency by decreasing slice size as load increases.  We
568  * consider the maximum latency as the sum of the threads waiting to run
569  * aside from curthread and target no more than sched_slice latency but
570  * no less than sched_slice_min runtime.
571  */
572 static inline int
573 tdq_slice(struct tdq *tdq)
574 {
575 	int load;
576 
577 	/*
578 	 * It is safe to use sys_load here because this is called from
579 	 * contexts where timeshare threads are running and so there
580 	 * cannot be higher priority load in the system.
581 	 */
582 	load = tdq->tdq_sysload - 1;
583 	if (load >= SCHED_SLICE_MIN_DIVISOR)
584 		return (sched_slice_min);
585 	if (load <= 1)
586 		return (sched_slice);
587 	return (sched_slice / load);
588 }
589 
590 /*
591  * Set lowpri to its exact value by searching the run-queue and
592  * evaluating curthread.  curthread may be passed as an optimization.
593  */
594 static void
595 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
596 {
597 	struct thread *td;
598 
599 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
600 	if (ctd == NULL)
601 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
602 	td = tdq_choose(tdq);
603 	if (td == NULL || td->td_priority > ctd->td_priority)
604 		tdq->tdq_lowpri = ctd->td_priority;
605 	else
606 		tdq->tdq_lowpri = td->td_priority;
607 }
608 
609 #ifdef SMP
610 /*
611  * We need some randomness. Implement a classic Linear Congruential
612  * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
613  * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
614  * of the random state (in the low bits of our answer) to keep
615  * the maximum randomness.
616  */
617 static uint32_t
618 sched_random(void)
619 {
620 	uint32_t *rndptr;
621 
622 	rndptr = DPCPU_PTR(randomval);
623 	*rndptr = *rndptr * 69069 + 5;
624 
625 	return (*rndptr >> 16);
626 }
627 
628 struct cpu_search {
629 	cpuset_t cs_mask;
630 	u_int	cs_prefer;
631 	int	cs_pri;		/* Min priority for low. */
632 	int	cs_limit;	/* Max load for low, min load for high. */
633 	int	cs_cpu;
634 	int	cs_load;
635 };
636 
637 #define	CPU_SEARCH_LOWEST	0x1
638 #define	CPU_SEARCH_HIGHEST	0x2
639 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
640 
641 #define	CPUSET_FOREACH(cpu, mask)				\
642 	for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++)		\
643 		if (CPU_ISSET(cpu, &mask))
644 
645 static __always_inline int cpu_search(const struct cpu_group *cg,
646     struct cpu_search *low, struct cpu_search *high, const int match);
647 int __noinline cpu_search_lowest(const struct cpu_group *cg,
648     struct cpu_search *low);
649 int __noinline cpu_search_highest(const struct cpu_group *cg,
650     struct cpu_search *high);
651 int __noinline cpu_search_both(const struct cpu_group *cg,
652     struct cpu_search *low, struct cpu_search *high);
653 
654 /*
655  * Search the tree of cpu_groups for the lowest or highest loaded cpu
656  * according to the match argument.  This routine actually compares the
657  * load on all paths through the tree and finds the least loaded cpu on
658  * the least loaded path, which may differ from the least loaded cpu in
659  * the system.  This balances work among caches and buses.
660  *
661  * This inline is instantiated in three forms below using constants for the
662  * match argument.  It is reduced to the minimum set for each case.  It is
663  * also recursive to the depth of the tree.
664  */
665 static __always_inline int
666 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
667     struct cpu_search *high, const int match)
668 {
669 	struct cpu_search lgroup;
670 	struct cpu_search hgroup;
671 	cpuset_t cpumask;
672 	struct cpu_group *child;
673 	struct tdq *tdq;
674 	int cpu, i, hload, lload, load, total, rnd;
675 
676 	total = 0;
677 	cpumask = cg->cg_mask;
678 	if (match & CPU_SEARCH_LOWEST) {
679 		lload = INT_MAX;
680 		lgroup = *low;
681 	}
682 	if (match & CPU_SEARCH_HIGHEST) {
683 		hload = INT_MIN;
684 		hgroup = *high;
685 	}
686 
687 	/* Iterate through the child CPU groups and then remaining CPUs. */
688 	for (i = cg->cg_children, cpu = mp_maxid; ; ) {
689 		if (i == 0) {
690 #ifdef HAVE_INLINE_FFSL
691 			cpu = CPU_FFS(&cpumask) - 1;
692 #else
693 			while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
694 				cpu--;
695 #endif
696 			if (cpu < 0)
697 				break;
698 			child = NULL;
699 		} else
700 			child = &cg->cg_child[i - 1];
701 
702 		if (match & CPU_SEARCH_LOWEST)
703 			lgroup.cs_cpu = -1;
704 		if (match & CPU_SEARCH_HIGHEST)
705 			hgroup.cs_cpu = -1;
706 		if (child) {			/* Handle child CPU group. */
707 			CPU_NAND(&cpumask, &child->cg_mask);
708 			switch (match) {
709 			case CPU_SEARCH_LOWEST:
710 				load = cpu_search_lowest(child, &lgroup);
711 				break;
712 			case CPU_SEARCH_HIGHEST:
713 				load = cpu_search_highest(child, &hgroup);
714 				break;
715 			case CPU_SEARCH_BOTH:
716 				load = cpu_search_both(child, &lgroup, &hgroup);
717 				break;
718 			}
719 		} else {			/* Handle child CPU. */
720 			CPU_CLR(cpu, &cpumask);
721 			tdq = TDQ_CPU(cpu);
722 			load = tdq->tdq_load * 256;
723 			rnd = sched_random() % 32;
724 			if (match & CPU_SEARCH_LOWEST) {
725 				if (cpu == low->cs_prefer)
726 					load -= 64;
727 				/* If that CPU is allowed and get data. */
728 				if (tdq->tdq_lowpri > lgroup.cs_pri &&
729 				    tdq->tdq_load <= lgroup.cs_limit &&
730 				    CPU_ISSET(cpu, &lgroup.cs_mask)) {
731 					lgroup.cs_cpu = cpu;
732 					lgroup.cs_load = load - rnd;
733 				}
734 			}
735 			if (match & CPU_SEARCH_HIGHEST)
736 				if (tdq->tdq_load >= hgroup.cs_limit &&
737 				    tdq->tdq_transferable &&
738 				    CPU_ISSET(cpu, &hgroup.cs_mask)) {
739 					hgroup.cs_cpu = cpu;
740 					hgroup.cs_load = load - rnd;
741 				}
742 		}
743 		total += load;
744 
745 		/* We have info about child item. Compare it. */
746 		if (match & CPU_SEARCH_LOWEST) {
747 			if (lgroup.cs_cpu >= 0 &&
748 			    (load < lload ||
749 			     (load == lload && lgroup.cs_load < low->cs_load))) {
750 				lload = load;
751 				low->cs_cpu = lgroup.cs_cpu;
752 				low->cs_load = lgroup.cs_load;
753 			}
754 		}
755 		if (match & CPU_SEARCH_HIGHEST)
756 			if (hgroup.cs_cpu >= 0 &&
757 			    (load > hload ||
758 			     (load == hload && hgroup.cs_load > high->cs_load))) {
759 				hload = load;
760 				high->cs_cpu = hgroup.cs_cpu;
761 				high->cs_load = hgroup.cs_load;
762 			}
763 		if (child) {
764 			i--;
765 			if (i == 0 && CPU_EMPTY(&cpumask))
766 				break;
767 		}
768 #ifndef HAVE_INLINE_FFSL
769 		else
770 			cpu--;
771 #endif
772 	}
773 	return (total);
774 }
775 
776 /*
777  * cpu_search instantiations must pass constants to maintain the inline
778  * optimization.
779  */
780 int
781 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
782 {
783 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
784 }
785 
786 int
787 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
788 {
789 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
790 }
791 
792 int
793 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
794     struct cpu_search *high)
795 {
796 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
797 }
798 
799 /*
800  * Find the cpu with the least load via the least loaded path that has a
801  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
802  * acceptable.
803  */
804 static inline int
805 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
806     int prefer)
807 {
808 	struct cpu_search low;
809 
810 	low.cs_cpu = -1;
811 	low.cs_prefer = prefer;
812 	low.cs_mask = mask;
813 	low.cs_pri = pri;
814 	low.cs_limit = maxload;
815 	cpu_search_lowest(cg, &low);
816 	return low.cs_cpu;
817 }
818 
819 /*
820  * Find the cpu with the highest load via the highest loaded path.
821  */
822 static inline int
823 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
824 {
825 	struct cpu_search high;
826 
827 	high.cs_cpu = -1;
828 	high.cs_mask = mask;
829 	high.cs_limit = minload;
830 	cpu_search_highest(cg, &high);
831 	return high.cs_cpu;
832 }
833 
834 static void
835 sched_balance_group(struct cpu_group *cg)
836 {
837 	cpuset_t hmask, lmask;
838 	int high, low, anylow;
839 
840 	CPU_FILL(&hmask);
841 	for (;;) {
842 		high = sched_highest(cg, hmask, 1);
843 		/* Stop if there is no more CPU with transferrable threads. */
844 		if (high == -1)
845 			break;
846 		CPU_CLR(high, &hmask);
847 		CPU_COPY(&hmask, &lmask);
848 		/* Stop if there is no more CPU left for low. */
849 		if (CPU_EMPTY(&lmask))
850 			break;
851 		anylow = 1;
852 nextlow:
853 		low = sched_lowest(cg, lmask, -1,
854 		    TDQ_CPU(high)->tdq_load - 1, high);
855 		/* Stop if we looked well and found no less loaded CPU. */
856 		if (anylow && low == -1)
857 			break;
858 		/* Go to next high if we found no less loaded CPU. */
859 		if (low == -1)
860 			continue;
861 		/* Transfer thread from high to low. */
862 		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
863 			/* CPU that got thread can no longer be a donor. */
864 			CPU_CLR(low, &hmask);
865 		} else {
866 			/*
867 			 * If failed, then there is no threads on high
868 			 * that can run on this low. Drop low from low
869 			 * mask and look for different one.
870 			 */
871 			CPU_CLR(low, &lmask);
872 			anylow = 0;
873 			goto nextlow;
874 		}
875 	}
876 }
877 
878 static void
879 sched_balance(void)
880 {
881 	struct tdq *tdq;
882 
883 	if (smp_started == 0 || rebalance == 0)
884 		return;
885 
886 	balance_ticks = max(balance_interval / 2, 1) +
887 	    (sched_random() % balance_interval);
888 	tdq = TDQ_SELF();
889 	TDQ_UNLOCK(tdq);
890 	sched_balance_group(cpu_top);
891 	TDQ_LOCK(tdq);
892 }
893 
894 /*
895  * Lock two thread queues using their address to maintain lock order.
896  */
897 static void
898 tdq_lock_pair(struct tdq *one, struct tdq *two)
899 {
900 	if (one < two) {
901 		TDQ_LOCK(one);
902 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
903 	} else {
904 		TDQ_LOCK(two);
905 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
906 	}
907 }
908 
909 /*
910  * Unlock two thread queues.  Order is not important here.
911  */
912 static void
913 tdq_unlock_pair(struct tdq *one, struct tdq *two)
914 {
915 	TDQ_UNLOCK(one);
916 	TDQ_UNLOCK(two);
917 }
918 
919 /*
920  * Transfer load between two imbalanced thread queues.
921  */
922 static int
923 sched_balance_pair(struct tdq *high, struct tdq *low)
924 {
925 	int moved;
926 	int cpu;
927 
928 	tdq_lock_pair(high, low);
929 	moved = 0;
930 	/*
931 	 * Determine what the imbalance is and then adjust that to how many
932 	 * threads we actually have to give up (transferable).
933 	 */
934 	if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
935 	    (moved = tdq_move(high, low)) > 0) {
936 		/*
937 		 * In case the target isn't the current cpu IPI it to force a
938 		 * reschedule with the new workload.
939 		 */
940 		cpu = TDQ_ID(low);
941 		if (cpu != PCPU_GET(cpuid))
942 			ipi_cpu(cpu, IPI_PREEMPT);
943 	}
944 	tdq_unlock_pair(high, low);
945 	return (moved);
946 }
947 
948 /*
949  * Move a thread from one thread queue to another.
950  */
951 static int
952 tdq_move(struct tdq *from, struct tdq *to)
953 {
954 	struct td_sched *ts;
955 	struct thread *td;
956 	struct tdq *tdq;
957 	int cpu;
958 
959 	TDQ_LOCK_ASSERT(from, MA_OWNED);
960 	TDQ_LOCK_ASSERT(to, MA_OWNED);
961 
962 	tdq = from;
963 	cpu = TDQ_ID(to);
964 	td = tdq_steal(tdq, cpu);
965 	if (td == NULL)
966 		return (0);
967 	ts = td_get_sched(td);
968 	/*
969 	 * Although the run queue is locked the thread may be blocked.  Lock
970 	 * it to clear this and acquire the run-queue lock.
971 	 */
972 	thread_lock(td);
973 	/* Drop recursive lock on from acquired via thread_lock(). */
974 	TDQ_UNLOCK(from);
975 	sched_rem(td);
976 	ts->ts_cpu = cpu;
977 	td->td_lock = TDQ_LOCKPTR(to);
978 	tdq_add(to, td, SRQ_YIELDING);
979 	return (1);
980 }
981 
982 /*
983  * This tdq has idled.  Try to steal a thread from another cpu and switch
984  * to it.
985  */
986 static int
987 tdq_idled(struct tdq *tdq)
988 {
989 	struct cpu_group *cg;
990 	struct tdq *steal;
991 	cpuset_t mask;
992 	int thresh;
993 	int cpu;
994 
995 	if (smp_started == 0 || steal_idle == 0)
996 		return (1);
997 	CPU_FILL(&mask);
998 	CPU_CLR(PCPU_GET(cpuid), &mask);
999 	/* We don't want to be preempted while we're iterating. */
1000 	spinlock_enter();
1001 	for (cg = tdq->tdq_cg; cg != NULL; ) {
1002 		if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
1003 			thresh = steal_thresh;
1004 		else
1005 			thresh = 1;
1006 		cpu = sched_highest(cg, mask, thresh);
1007 		if (cpu == -1) {
1008 			cg = cg->cg_parent;
1009 			continue;
1010 		}
1011 		steal = TDQ_CPU(cpu);
1012 		CPU_CLR(cpu, &mask);
1013 		tdq_lock_pair(tdq, steal);
1014 		if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
1015 			tdq_unlock_pair(tdq, steal);
1016 			continue;
1017 		}
1018 		/*
1019 		 * If a thread was added while interrupts were disabled don't
1020 		 * steal one here.  If we fail to acquire one due to affinity
1021 		 * restrictions loop again with this cpu removed from the
1022 		 * set.
1023 		 */
1024 		if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
1025 			tdq_unlock_pair(tdq, steal);
1026 			continue;
1027 		}
1028 		spinlock_exit();
1029 		TDQ_UNLOCK(steal);
1030 		mi_switch(SW_VOL | SWT_IDLE, NULL);
1031 		thread_unlock(curthread);
1032 
1033 		return (0);
1034 	}
1035 	spinlock_exit();
1036 	return (1);
1037 }
1038 
1039 /*
1040  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
1041  */
1042 static void
1043 tdq_notify(struct tdq *tdq, struct thread *td)
1044 {
1045 	struct thread *ctd;
1046 	int pri;
1047 	int cpu;
1048 
1049 	if (tdq->tdq_ipipending)
1050 		return;
1051 	cpu = td_get_sched(td)->ts_cpu;
1052 	pri = td->td_priority;
1053 	ctd = pcpu_find(cpu)->pc_curthread;
1054 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1055 		return;
1056 
1057 	/*
1058 	 * Make sure that our caller's earlier update to tdq_load is
1059 	 * globally visible before we read tdq_cpu_idle.  Idle thread
1060 	 * accesses both of them without locks, and the order is important.
1061 	 */
1062 	atomic_thread_fence_seq_cst();
1063 
1064 	if (TD_IS_IDLETHREAD(ctd)) {
1065 		/*
1066 		 * If the MD code has an idle wakeup routine try that before
1067 		 * falling back to IPI.
1068 		 */
1069 		if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1070 			return;
1071 	}
1072 	tdq->tdq_ipipending = 1;
1073 	ipi_cpu(cpu, IPI_PREEMPT);
1074 }
1075 
1076 /*
1077  * Steals load from a timeshare queue.  Honors the rotating queue head
1078  * index.
1079  */
1080 static struct thread *
1081 runq_steal_from(struct runq *rq, int cpu, u_char start)
1082 {
1083 	struct rqbits *rqb;
1084 	struct rqhead *rqh;
1085 	struct thread *td, *first;
1086 	int bit;
1087 	int i;
1088 
1089 	rqb = &rq->rq_status;
1090 	bit = start & (RQB_BPW -1);
1091 	first = NULL;
1092 again:
1093 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1094 		if (rqb->rqb_bits[i] == 0)
1095 			continue;
1096 		if (bit == 0)
1097 			bit = RQB_FFS(rqb->rqb_bits[i]);
1098 		for (; bit < RQB_BPW; bit++) {
1099 			if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1100 				continue;
1101 			rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1102 			TAILQ_FOREACH(td, rqh, td_runq) {
1103 				if (first && THREAD_CAN_MIGRATE(td) &&
1104 				    THREAD_CAN_SCHED(td, cpu))
1105 					return (td);
1106 				first = td;
1107 			}
1108 		}
1109 	}
1110 	if (start != 0) {
1111 		start = 0;
1112 		goto again;
1113 	}
1114 
1115 	if (first && THREAD_CAN_MIGRATE(first) &&
1116 	    THREAD_CAN_SCHED(first, cpu))
1117 		return (first);
1118 	return (NULL);
1119 }
1120 
1121 /*
1122  * Steals load from a standard linear queue.
1123  */
1124 static struct thread *
1125 runq_steal(struct runq *rq, int cpu)
1126 {
1127 	struct rqhead *rqh;
1128 	struct rqbits *rqb;
1129 	struct thread *td;
1130 	int word;
1131 	int bit;
1132 
1133 	rqb = &rq->rq_status;
1134 	for (word = 0; word < RQB_LEN; word++) {
1135 		if (rqb->rqb_bits[word] == 0)
1136 			continue;
1137 		for (bit = 0; bit < RQB_BPW; bit++) {
1138 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1139 				continue;
1140 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1141 			TAILQ_FOREACH(td, rqh, td_runq)
1142 				if (THREAD_CAN_MIGRATE(td) &&
1143 				    THREAD_CAN_SCHED(td, cpu))
1144 					return (td);
1145 		}
1146 	}
1147 	return (NULL);
1148 }
1149 
1150 /*
1151  * Attempt to steal a thread in priority order from a thread queue.
1152  */
1153 static struct thread *
1154 tdq_steal(struct tdq *tdq, int cpu)
1155 {
1156 	struct thread *td;
1157 
1158 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1159 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1160 		return (td);
1161 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1162 	    cpu, tdq->tdq_ridx)) != NULL)
1163 		return (td);
1164 	return (runq_steal(&tdq->tdq_idle, cpu));
1165 }
1166 
1167 /*
1168  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1169  * current lock and returns with the assigned queue locked.
1170  */
1171 static inline struct tdq *
1172 sched_setcpu(struct thread *td, int cpu, int flags)
1173 {
1174 
1175 	struct tdq *tdq;
1176 
1177 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1178 	tdq = TDQ_CPU(cpu);
1179 	td_get_sched(td)->ts_cpu = cpu;
1180 	/*
1181 	 * If the lock matches just return the queue.
1182 	 */
1183 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1184 		return (tdq);
1185 #ifdef notyet
1186 	/*
1187 	 * If the thread isn't running its lockptr is a
1188 	 * turnstile or a sleepqueue.  We can just lock_set without
1189 	 * blocking.
1190 	 */
1191 	if (TD_CAN_RUN(td)) {
1192 		TDQ_LOCK(tdq);
1193 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1194 		return (tdq);
1195 	}
1196 #endif
1197 	/*
1198 	 * The hard case, migration, we need to block the thread first to
1199 	 * prevent order reversals with other cpus locks.
1200 	 */
1201 	spinlock_enter();
1202 	thread_lock_block(td);
1203 	TDQ_LOCK(tdq);
1204 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1205 	spinlock_exit();
1206 	return (tdq);
1207 }
1208 
1209 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1210 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1211 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1212 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1213 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1214 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1215 
1216 static int
1217 sched_pickcpu(struct thread *td, int flags)
1218 {
1219 	struct cpu_group *cg, *ccg;
1220 	struct td_sched *ts;
1221 	struct tdq *tdq;
1222 	cpuset_t mask;
1223 	int cpu, pri, self;
1224 
1225 	self = PCPU_GET(cpuid);
1226 	ts = td_get_sched(td);
1227 	if (smp_started == 0)
1228 		return (self);
1229 	/*
1230 	 * Don't migrate a running thread from sched_switch().
1231 	 */
1232 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1233 		return (ts->ts_cpu);
1234 	/*
1235 	 * Prefer to run interrupt threads on the processors that generate
1236 	 * the interrupt.
1237 	 */
1238 	pri = td->td_priority;
1239 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1240 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1241 		SCHED_STAT_INC(pickcpu_intrbind);
1242 		ts->ts_cpu = self;
1243 		if (TDQ_CPU(self)->tdq_lowpri > pri) {
1244 			SCHED_STAT_INC(pickcpu_affinity);
1245 			return (ts->ts_cpu);
1246 		}
1247 	}
1248 	/*
1249 	 * If the thread can run on the last cpu and the affinity has not
1250 	 * expired or it is idle run it there.
1251 	 */
1252 	tdq = TDQ_CPU(ts->ts_cpu);
1253 	cg = tdq->tdq_cg;
1254 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1255 	    tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1256 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1257 		if (cg->cg_flags & CG_FLAG_THREAD) {
1258 			CPUSET_FOREACH(cpu, cg->cg_mask) {
1259 				if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1260 					break;
1261 			}
1262 		} else
1263 			cpu = INT_MAX;
1264 		if (cpu > mp_maxid) {
1265 			SCHED_STAT_INC(pickcpu_idle_affinity);
1266 			return (ts->ts_cpu);
1267 		}
1268 	}
1269 	/*
1270 	 * Search for the last level cache CPU group in the tree.
1271 	 * Skip caches with expired affinity time and SMT groups.
1272 	 * Affinity to higher level caches will be handled less aggressively.
1273 	 */
1274 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1275 		if (cg->cg_flags & CG_FLAG_THREAD)
1276 			continue;
1277 		if (!SCHED_AFFINITY(ts, cg->cg_level))
1278 			continue;
1279 		ccg = cg;
1280 	}
1281 	if (ccg != NULL)
1282 		cg = ccg;
1283 	cpu = -1;
1284 	/* Search the group for the less loaded idle CPU we can run now. */
1285 	mask = td->td_cpuset->cs_mask;
1286 	if (cg != NULL && cg != cpu_top &&
1287 	    CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1288 		cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1289 		    INT_MAX, ts->ts_cpu);
1290 	/* Search globally for the less loaded CPU we can run now. */
1291 	if (cpu == -1)
1292 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1293 	/* Search globally for the less loaded CPU. */
1294 	if (cpu == -1)
1295 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1296 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1297 	/*
1298 	 * Compare the lowest loaded cpu to current cpu.
1299 	 */
1300 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1301 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1302 	    TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1303 		SCHED_STAT_INC(pickcpu_local);
1304 		cpu = self;
1305 	} else
1306 		SCHED_STAT_INC(pickcpu_lowest);
1307 	if (cpu != ts->ts_cpu)
1308 		SCHED_STAT_INC(pickcpu_migration);
1309 	return (cpu);
1310 }
1311 #endif
1312 
1313 /*
1314  * Pick the highest priority task we have and return it.
1315  */
1316 static struct thread *
1317 tdq_choose(struct tdq *tdq)
1318 {
1319 	struct thread *td;
1320 
1321 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1322 	td = runq_choose(&tdq->tdq_realtime);
1323 	if (td != NULL)
1324 		return (td);
1325 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1326 	if (td != NULL) {
1327 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1328 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1329 		    td->td_priority));
1330 		return (td);
1331 	}
1332 	td = runq_choose(&tdq->tdq_idle);
1333 	if (td != NULL) {
1334 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1335 		    ("tdq_choose: Invalid priority on idle queue %d",
1336 		    td->td_priority));
1337 		return (td);
1338 	}
1339 
1340 	return (NULL);
1341 }
1342 
1343 /*
1344  * Initialize a thread queue.
1345  */
1346 static void
1347 tdq_setup(struct tdq *tdq)
1348 {
1349 
1350 	if (bootverbose)
1351 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1352 	runq_init(&tdq->tdq_realtime);
1353 	runq_init(&tdq->tdq_timeshare);
1354 	runq_init(&tdq->tdq_idle);
1355 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1356 	    "sched lock %d", (int)TDQ_ID(tdq));
1357 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1358 	    MTX_SPIN | MTX_RECURSE);
1359 #ifdef KTR
1360 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1361 	    "CPU %d load", (int)TDQ_ID(tdq));
1362 #endif
1363 }
1364 
1365 #ifdef SMP
1366 static void
1367 sched_setup_smp(void)
1368 {
1369 	struct tdq *tdq;
1370 	int i;
1371 
1372 	cpu_top = smp_topo();
1373 	CPU_FOREACH(i) {
1374 		tdq = TDQ_CPU(i);
1375 		tdq_setup(tdq);
1376 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1377 		if (tdq->tdq_cg == NULL)
1378 			panic("Can't find cpu group for %d\n", i);
1379 	}
1380 	balance_tdq = TDQ_SELF();
1381 	sched_balance();
1382 }
1383 #endif
1384 
1385 /*
1386  * Setup the thread queues and initialize the topology based on MD
1387  * information.
1388  */
1389 static void
1390 sched_setup(void *dummy)
1391 {
1392 	struct tdq *tdq;
1393 
1394 	tdq = TDQ_SELF();
1395 #ifdef SMP
1396 	sched_setup_smp();
1397 #else
1398 	tdq_setup(tdq);
1399 #endif
1400 
1401 	/* Add thread0's load since it's running. */
1402 	TDQ_LOCK(tdq);
1403 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1404 	tdq_load_add(tdq, &thread0);
1405 	tdq->tdq_lowpri = thread0.td_priority;
1406 	TDQ_UNLOCK(tdq);
1407 }
1408 
1409 /*
1410  * This routine determines time constants after stathz and hz are setup.
1411  */
1412 /* ARGSUSED */
1413 static void
1414 sched_initticks(void *dummy)
1415 {
1416 	int incr;
1417 
1418 	realstathz = stathz ? stathz : hz;
1419 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1420 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1421 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1422 	    realstathz);
1423 
1424 	/*
1425 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1426 	 * hz not being evenly divisible by stathz on all platforms.
1427 	 */
1428 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1429 	/*
1430 	 * This does not work for values of stathz that are more than
1431 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1432 	 */
1433 	if (incr == 0)
1434 		incr = 1;
1435 	tickincr = incr;
1436 #ifdef SMP
1437 	/*
1438 	 * Set the default balance interval now that we know
1439 	 * what realstathz is.
1440 	 */
1441 	balance_interval = realstathz;
1442 	affinity = SCHED_AFFINITY_DEFAULT;
1443 #endif
1444 	if (sched_idlespinthresh < 0)
1445 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1446 }
1447 
1448 
1449 /*
1450  * This is the core of the interactivity algorithm.  Determines a score based
1451  * on past behavior.  It is the ratio of sleep time to run time scaled to
1452  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1453  * differs from the cpu usage because it does not account for time spent
1454  * waiting on a run-queue.  Would be prettier if we had floating point.
1455  *
1456  * When a thread's sleep time is greater than its run time the
1457  * calculation is:
1458  *
1459  *                           scaling factor
1460  * interactivity score =  ---------------------
1461  *                        sleep time / run time
1462  *
1463  *
1464  * When a thread's run time is greater than its sleep time the
1465  * calculation is:
1466  *
1467  *                           scaling factor
1468  * interactivity score =  ---------------------    + scaling factor
1469  *                        run time / sleep time
1470  */
1471 static int
1472 sched_interact_score(struct thread *td)
1473 {
1474 	struct td_sched *ts;
1475 	int div;
1476 
1477 	ts = td_get_sched(td);
1478 	/*
1479 	 * The score is only needed if this is likely to be an interactive
1480 	 * task.  Don't go through the expense of computing it if there's
1481 	 * no chance.
1482 	 */
1483 	if (sched_interact <= SCHED_INTERACT_HALF &&
1484 		ts->ts_runtime >= ts->ts_slptime)
1485 			return (SCHED_INTERACT_HALF);
1486 
1487 	if (ts->ts_runtime > ts->ts_slptime) {
1488 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1489 		return (SCHED_INTERACT_HALF +
1490 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1491 	}
1492 	if (ts->ts_slptime > ts->ts_runtime) {
1493 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1494 		return (ts->ts_runtime / div);
1495 	}
1496 	/* runtime == slptime */
1497 	if (ts->ts_runtime)
1498 		return (SCHED_INTERACT_HALF);
1499 
1500 	/*
1501 	 * This can happen if slptime and runtime are 0.
1502 	 */
1503 	return (0);
1504 
1505 }
1506 
1507 /*
1508  * Scale the scheduling priority according to the "interactivity" of this
1509  * process.
1510  */
1511 static void
1512 sched_priority(struct thread *td)
1513 {
1514 	int score;
1515 	int pri;
1516 
1517 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1518 		return;
1519 	/*
1520 	 * If the score is interactive we place the thread in the realtime
1521 	 * queue with a priority that is less than kernel and interrupt
1522 	 * priorities.  These threads are not subject to nice restrictions.
1523 	 *
1524 	 * Scores greater than this are placed on the normal timeshare queue
1525 	 * where the priority is partially decided by the most recent cpu
1526 	 * utilization and the rest is decided by nice value.
1527 	 *
1528 	 * The nice value of the process has a linear effect on the calculated
1529 	 * score.  Negative nice values make it easier for a thread to be
1530 	 * considered interactive.
1531 	 */
1532 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1533 	if (score < sched_interact) {
1534 		pri = PRI_MIN_INTERACT;
1535 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1536 		    sched_interact) * score;
1537 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1538 		    ("sched_priority: invalid interactive priority %d score %d",
1539 		    pri, score));
1540 	} else {
1541 		pri = SCHED_PRI_MIN;
1542 		if (td_get_sched(td)->ts_ticks)
1543 			pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1544 			    SCHED_PRI_RANGE - 1);
1545 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1546 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1547 		    ("sched_priority: invalid priority %d: nice %d, "
1548 		    "ticks %d ftick %d ltick %d tick pri %d",
1549 		    pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1550 		    td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1551 		    SCHED_PRI_TICKS(td_get_sched(td))));
1552 	}
1553 	sched_user_prio(td, pri);
1554 
1555 	return;
1556 }
1557 
1558 /*
1559  * This routine enforces a maximum limit on the amount of scheduling history
1560  * kept.  It is called after either the slptime or runtime is adjusted.  This
1561  * function is ugly due to integer math.
1562  */
1563 static void
1564 sched_interact_update(struct thread *td)
1565 {
1566 	struct td_sched *ts;
1567 	u_int sum;
1568 
1569 	ts = td_get_sched(td);
1570 	sum = ts->ts_runtime + ts->ts_slptime;
1571 	if (sum < SCHED_SLP_RUN_MAX)
1572 		return;
1573 	/*
1574 	 * This only happens from two places:
1575 	 * 1) We have added an unusual amount of run time from fork_exit.
1576 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1577 	 */
1578 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1579 		if (ts->ts_runtime > ts->ts_slptime) {
1580 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1581 			ts->ts_slptime = 1;
1582 		} else {
1583 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1584 			ts->ts_runtime = 1;
1585 		}
1586 		return;
1587 	}
1588 	/*
1589 	 * If we have exceeded by more than 1/5th then the algorithm below
1590 	 * will not bring us back into range.  Dividing by two here forces
1591 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1592 	 */
1593 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1594 		ts->ts_runtime /= 2;
1595 		ts->ts_slptime /= 2;
1596 		return;
1597 	}
1598 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1599 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1600 }
1601 
1602 /*
1603  * Scale back the interactivity history when a child thread is created.  The
1604  * history is inherited from the parent but the thread may behave totally
1605  * differently.  For example, a shell spawning a compiler process.  We want
1606  * to learn that the compiler is behaving badly very quickly.
1607  */
1608 static void
1609 sched_interact_fork(struct thread *td)
1610 {
1611 	struct td_sched *ts;
1612 	int ratio;
1613 	int sum;
1614 
1615 	ts = td_get_sched(td);
1616 	sum = ts->ts_runtime + ts->ts_slptime;
1617 	if (sum > SCHED_SLP_RUN_FORK) {
1618 		ratio = sum / SCHED_SLP_RUN_FORK;
1619 		ts->ts_runtime /= ratio;
1620 		ts->ts_slptime /= ratio;
1621 	}
1622 }
1623 
1624 /*
1625  * Called from proc0_init() to setup the scheduler fields.
1626  */
1627 void
1628 schedinit(void)
1629 {
1630 	struct td_sched *ts0;
1631 
1632 	/*
1633 	 * Set up the scheduler specific parts of thread0.
1634 	 */
1635 	ts0 = td_get_sched(&thread0);
1636 	ts0->ts_ltick = ticks;
1637 	ts0->ts_ftick = ticks;
1638 	ts0->ts_slice = 0;
1639 }
1640 
1641 /*
1642  * This is only somewhat accurate since given many processes of the same
1643  * priority they will switch when their slices run out, which will be
1644  * at most sched_slice stathz ticks.
1645  */
1646 int
1647 sched_rr_interval(void)
1648 {
1649 
1650 	/* Convert sched_slice from stathz to hz. */
1651 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1652 }
1653 
1654 /*
1655  * Update the percent cpu tracking information when it is requested or
1656  * the total history exceeds the maximum.  We keep a sliding history of
1657  * tick counts that slowly decays.  This is less precise than the 4BSD
1658  * mechanism since it happens with less regular and frequent events.
1659  */
1660 static void
1661 sched_pctcpu_update(struct td_sched *ts, int run)
1662 {
1663 	int t = ticks;
1664 
1665 	/*
1666 	 * The signed difference may be negative if the thread hasn't run for
1667 	 * over half of the ticks rollover period.
1668 	 */
1669 	if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) {
1670 		ts->ts_ticks = 0;
1671 		ts->ts_ftick = t - SCHED_TICK_TARG;
1672 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1673 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1674 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1675 		ts->ts_ftick = t - SCHED_TICK_TARG;
1676 	}
1677 	if (run)
1678 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1679 	ts->ts_ltick = t;
1680 }
1681 
1682 /*
1683  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1684  * if necessary.  This is the back-end for several priority related
1685  * functions.
1686  */
1687 static void
1688 sched_thread_priority(struct thread *td, u_char prio)
1689 {
1690 	struct td_sched *ts;
1691 	struct tdq *tdq;
1692 	int oldpri;
1693 
1694 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1695 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1696 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1697 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1698 	if (td != curthread && prio < td->td_priority) {
1699 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1700 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1701 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1702 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1703 		    curthread);
1704 	}
1705 	ts = td_get_sched(td);
1706 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1707 	if (td->td_priority == prio)
1708 		return;
1709 	/*
1710 	 * If the priority has been elevated due to priority
1711 	 * propagation, we may have to move ourselves to a new
1712 	 * queue.  This could be optimized to not re-add in some
1713 	 * cases.
1714 	 */
1715 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1716 		sched_rem(td);
1717 		td->td_priority = prio;
1718 		sched_add(td, SRQ_BORROWING);
1719 		return;
1720 	}
1721 	/*
1722 	 * If the thread is currently running we may have to adjust the lowpri
1723 	 * information so other cpus are aware of our current priority.
1724 	 */
1725 	if (TD_IS_RUNNING(td)) {
1726 		tdq = TDQ_CPU(ts->ts_cpu);
1727 		oldpri = td->td_priority;
1728 		td->td_priority = prio;
1729 		if (prio < tdq->tdq_lowpri)
1730 			tdq->tdq_lowpri = prio;
1731 		else if (tdq->tdq_lowpri == oldpri)
1732 			tdq_setlowpri(tdq, td);
1733 		return;
1734 	}
1735 	td->td_priority = prio;
1736 }
1737 
1738 /*
1739  * Update a thread's priority when it is lent another thread's
1740  * priority.
1741  */
1742 void
1743 sched_lend_prio(struct thread *td, u_char prio)
1744 {
1745 
1746 	td->td_flags |= TDF_BORROWING;
1747 	sched_thread_priority(td, prio);
1748 }
1749 
1750 /*
1751  * Restore a thread's priority when priority propagation is
1752  * over.  The prio argument is the minimum priority the thread
1753  * needs to have to satisfy other possible priority lending
1754  * requests.  If the thread's regular priority is less
1755  * important than prio, the thread will keep a priority boost
1756  * of prio.
1757  */
1758 void
1759 sched_unlend_prio(struct thread *td, u_char prio)
1760 {
1761 	u_char base_pri;
1762 
1763 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1764 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1765 		base_pri = td->td_user_pri;
1766 	else
1767 		base_pri = td->td_base_pri;
1768 	if (prio >= base_pri) {
1769 		td->td_flags &= ~TDF_BORROWING;
1770 		sched_thread_priority(td, base_pri);
1771 	} else
1772 		sched_lend_prio(td, prio);
1773 }
1774 
1775 /*
1776  * Standard entry for setting the priority to an absolute value.
1777  */
1778 void
1779 sched_prio(struct thread *td, u_char prio)
1780 {
1781 	u_char oldprio;
1782 
1783 	/* First, update the base priority. */
1784 	td->td_base_pri = prio;
1785 
1786 	/*
1787 	 * If the thread is borrowing another thread's priority, don't
1788 	 * ever lower the priority.
1789 	 */
1790 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1791 		return;
1792 
1793 	/* Change the real priority. */
1794 	oldprio = td->td_priority;
1795 	sched_thread_priority(td, prio);
1796 
1797 	/*
1798 	 * If the thread is on a turnstile, then let the turnstile update
1799 	 * its state.
1800 	 */
1801 	if (TD_ON_LOCK(td) && oldprio != prio)
1802 		turnstile_adjust(td, oldprio);
1803 }
1804 
1805 /*
1806  * Set the base user priority, does not effect current running priority.
1807  */
1808 void
1809 sched_user_prio(struct thread *td, u_char prio)
1810 {
1811 
1812 	td->td_base_user_pri = prio;
1813 	if (td->td_lend_user_pri <= prio)
1814 		return;
1815 	td->td_user_pri = prio;
1816 }
1817 
1818 void
1819 sched_lend_user_prio(struct thread *td, u_char prio)
1820 {
1821 
1822 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1823 	td->td_lend_user_pri = prio;
1824 	td->td_user_pri = min(prio, td->td_base_user_pri);
1825 	if (td->td_priority > td->td_user_pri)
1826 		sched_prio(td, td->td_user_pri);
1827 	else if (td->td_priority != td->td_user_pri)
1828 		td->td_flags |= TDF_NEEDRESCHED;
1829 }
1830 
1831 /*
1832  * Handle migration from sched_switch().  This happens only for
1833  * cpu binding.
1834  */
1835 static struct mtx *
1836 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1837 {
1838 	struct tdq *tdn;
1839 
1840 	tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
1841 #ifdef SMP
1842 	tdq_load_rem(tdq, td);
1843 	/*
1844 	 * Do the lock dance required to avoid LOR.  We grab an extra
1845 	 * spinlock nesting to prevent preemption while we're
1846 	 * not holding either run-queue lock.
1847 	 */
1848 	spinlock_enter();
1849 	thread_lock_block(td);	/* This releases the lock on tdq. */
1850 
1851 	/*
1852 	 * Acquire both run-queue locks before placing the thread on the new
1853 	 * run-queue to avoid deadlocks created by placing a thread with a
1854 	 * blocked lock on the run-queue of a remote processor.  The deadlock
1855 	 * occurs when a third processor attempts to lock the two queues in
1856 	 * question while the target processor is spinning with its own
1857 	 * run-queue lock held while waiting for the blocked lock to clear.
1858 	 */
1859 	tdq_lock_pair(tdn, tdq);
1860 	tdq_add(tdn, td, flags);
1861 	tdq_notify(tdn, td);
1862 	TDQ_UNLOCK(tdn);
1863 	spinlock_exit();
1864 #endif
1865 	return (TDQ_LOCKPTR(tdn));
1866 }
1867 
1868 /*
1869  * Variadic version of thread_lock_unblock() that does not assume td_lock
1870  * is blocked.
1871  */
1872 static inline void
1873 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1874 {
1875 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1876 	    (uintptr_t)mtx);
1877 }
1878 
1879 /*
1880  * Switch threads.  This function has to handle threads coming in while
1881  * blocked for some reason, running, or idle.  It also must deal with
1882  * migrating a thread from one queue to another as running threads may
1883  * be assigned elsewhere via binding.
1884  */
1885 void
1886 sched_switch(struct thread *td, struct thread *newtd, int flags)
1887 {
1888 	struct tdq *tdq;
1889 	struct td_sched *ts;
1890 	struct mtx *mtx;
1891 	int srqflag;
1892 	int cpuid, preempted;
1893 
1894 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1895 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1896 
1897 	cpuid = PCPU_GET(cpuid);
1898 	tdq = TDQ_CPU(cpuid);
1899 	ts = td_get_sched(td);
1900 	mtx = td->td_lock;
1901 	sched_pctcpu_update(ts, 1);
1902 	ts->ts_rltick = ticks;
1903 	td->td_lastcpu = td->td_oncpu;
1904 	td->td_oncpu = NOCPU;
1905 	preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
1906 	    (flags & SW_PREEMPT) != 0;
1907 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1908 	td->td_owepreempt = 0;
1909 	if (!TD_IS_IDLETHREAD(td))
1910 		tdq->tdq_switchcnt++;
1911 	/*
1912 	 * The lock pointer in an idle thread should never change.  Reset it
1913 	 * to CAN_RUN as well.
1914 	 */
1915 	if (TD_IS_IDLETHREAD(td)) {
1916 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1917 		TD_SET_CAN_RUN(td);
1918 	} else if (TD_IS_RUNNING(td)) {
1919 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1920 		srqflag = preempted ?
1921 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1922 		    SRQ_OURSELF|SRQ_YIELDING;
1923 #ifdef SMP
1924 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1925 			ts->ts_cpu = sched_pickcpu(td, 0);
1926 #endif
1927 		if (ts->ts_cpu == cpuid)
1928 			tdq_runq_add(tdq, td, srqflag);
1929 		else {
1930 			KASSERT(THREAD_CAN_MIGRATE(td) ||
1931 			    (ts->ts_flags & TSF_BOUND) != 0,
1932 			    ("Thread %p shouldn't migrate", td));
1933 			mtx = sched_switch_migrate(tdq, td, srqflag);
1934 		}
1935 	} else {
1936 		/* This thread must be going to sleep. */
1937 		TDQ_LOCK(tdq);
1938 		mtx = thread_lock_block(td);
1939 		tdq_load_rem(tdq, td);
1940 	}
1941 	/*
1942 	 * We enter here with the thread blocked and assigned to the
1943 	 * appropriate cpu run-queue or sleep-queue and with the current
1944 	 * thread-queue locked.
1945 	 */
1946 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1947 	newtd = choosethread();
1948 	/*
1949 	 * Call the MD code to switch contexts if necessary.
1950 	 */
1951 	if (td != newtd) {
1952 #ifdef	HWPMC_HOOKS
1953 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1954 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1955 #endif
1956 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1957 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1958 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1959 		sched_pctcpu_update(td_get_sched(newtd), 0);
1960 
1961 #ifdef KDTRACE_HOOKS
1962 		/*
1963 		 * If DTrace has set the active vtime enum to anything
1964 		 * other than INACTIVE (0), then it should have set the
1965 		 * function to call.
1966 		 */
1967 		if (dtrace_vtime_active)
1968 			(*dtrace_vtime_switch_func)(newtd);
1969 #endif
1970 
1971 		cpu_switch(td, newtd, mtx);
1972 		/*
1973 		 * We may return from cpu_switch on a different cpu.  However,
1974 		 * we always return with td_lock pointing to the current cpu's
1975 		 * run queue lock.
1976 		 */
1977 		cpuid = PCPU_GET(cpuid);
1978 		tdq = TDQ_CPU(cpuid);
1979 		lock_profile_obtain_lock_success(
1980 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1981 
1982 		SDT_PROBE0(sched, , , on__cpu);
1983 #ifdef	HWPMC_HOOKS
1984 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1985 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1986 #endif
1987 	} else {
1988 		thread_unblock_switch(td, mtx);
1989 		SDT_PROBE0(sched, , , remain__cpu);
1990 	}
1991 	/*
1992 	 * Assert that all went well and return.
1993 	 */
1994 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1995 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1996 	td->td_oncpu = cpuid;
1997 }
1998 
1999 /*
2000  * Adjust thread priorities as a result of a nice request.
2001  */
2002 void
2003 sched_nice(struct proc *p, int nice)
2004 {
2005 	struct thread *td;
2006 
2007 	PROC_LOCK_ASSERT(p, MA_OWNED);
2008 
2009 	p->p_nice = nice;
2010 	FOREACH_THREAD_IN_PROC(p, td) {
2011 		thread_lock(td);
2012 		sched_priority(td);
2013 		sched_prio(td, td->td_base_user_pri);
2014 		thread_unlock(td);
2015 	}
2016 }
2017 
2018 /*
2019  * Record the sleep time for the interactivity scorer.
2020  */
2021 void
2022 sched_sleep(struct thread *td, int prio)
2023 {
2024 
2025 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2026 
2027 	td->td_slptick = ticks;
2028 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2029 		td->td_flags |= TDF_CANSWAP;
2030 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2031 		return;
2032 	if (static_boost == 1 && prio)
2033 		sched_prio(td, prio);
2034 	else if (static_boost && td->td_priority > static_boost)
2035 		sched_prio(td, static_boost);
2036 }
2037 
2038 /*
2039  * Schedule a thread to resume execution and record how long it voluntarily
2040  * slept.  We also update the pctcpu, interactivity, and priority.
2041  */
2042 void
2043 sched_wakeup(struct thread *td)
2044 {
2045 	struct td_sched *ts;
2046 	int slptick;
2047 
2048 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2049 	ts = td_get_sched(td);
2050 	td->td_flags &= ~TDF_CANSWAP;
2051 	/*
2052 	 * If we slept for more than a tick update our interactivity and
2053 	 * priority.
2054 	 */
2055 	slptick = td->td_slptick;
2056 	td->td_slptick = 0;
2057 	if (slptick && slptick != ticks) {
2058 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2059 		sched_interact_update(td);
2060 		sched_pctcpu_update(ts, 0);
2061 	}
2062 	/*
2063 	 * Reset the slice value since we slept and advanced the round-robin.
2064 	 */
2065 	ts->ts_slice = 0;
2066 	sched_add(td, SRQ_BORING);
2067 }
2068 
2069 /*
2070  * Penalize the parent for creating a new child and initialize the child's
2071  * priority.
2072  */
2073 void
2074 sched_fork(struct thread *td, struct thread *child)
2075 {
2076 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2077 	sched_pctcpu_update(td_get_sched(td), 1);
2078 	sched_fork_thread(td, child);
2079 	/*
2080 	 * Penalize the parent and child for forking.
2081 	 */
2082 	sched_interact_fork(child);
2083 	sched_priority(child);
2084 	td_get_sched(td)->ts_runtime += tickincr;
2085 	sched_interact_update(td);
2086 	sched_priority(td);
2087 }
2088 
2089 /*
2090  * Fork a new thread, may be within the same process.
2091  */
2092 void
2093 sched_fork_thread(struct thread *td, struct thread *child)
2094 {
2095 	struct td_sched *ts;
2096 	struct td_sched *ts2;
2097 	struct tdq *tdq;
2098 
2099 	tdq = TDQ_SELF();
2100 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2101 	/*
2102 	 * Initialize child.
2103 	 */
2104 	ts = td_get_sched(td);
2105 	ts2 = td_get_sched(child);
2106 	child->td_oncpu = NOCPU;
2107 	child->td_lastcpu = NOCPU;
2108 	child->td_lock = TDQ_LOCKPTR(tdq);
2109 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2110 	ts2->ts_cpu = ts->ts_cpu;
2111 	ts2->ts_flags = 0;
2112 	/*
2113 	 * Grab our parents cpu estimation information.
2114 	 */
2115 	ts2->ts_ticks = ts->ts_ticks;
2116 	ts2->ts_ltick = ts->ts_ltick;
2117 	ts2->ts_ftick = ts->ts_ftick;
2118 	/*
2119 	 * Do not inherit any borrowed priority from the parent.
2120 	 */
2121 	child->td_priority = child->td_base_pri;
2122 	/*
2123 	 * And update interactivity score.
2124 	 */
2125 	ts2->ts_slptime = ts->ts_slptime;
2126 	ts2->ts_runtime = ts->ts_runtime;
2127 	/* Attempt to quickly learn interactivity. */
2128 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2129 #ifdef KTR
2130 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2131 #endif
2132 }
2133 
2134 /*
2135  * Adjust the priority class of a thread.
2136  */
2137 void
2138 sched_class(struct thread *td, int class)
2139 {
2140 
2141 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2142 	if (td->td_pri_class == class)
2143 		return;
2144 	td->td_pri_class = class;
2145 }
2146 
2147 /*
2148  * Return some of the child's priority and interactivity to the parent.
2149  */
2150 void
2151 sched_exit(struct proc *p, struct thread *child)
2152 {
2153 	struct thread *td;
2154 
2155 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2156 	    "prio:%d", child->td_priority);
2157 	PROC_LOCK_ASSERT(p, MA_OWNED);
2158 	td = FIRST_THREAD_IN_PROC(p);
2159 	sched_exit_thread(td, child);
2160 }
2161 
2162 /*
2163  * Penalize another thread for the time spent on this one.  This helps to
2164  * worsen the priority and interactivity of processes which schedule batch
2165  * jobs such as make.  This has little effect on the make process itself but
2166  * causes new processes spawned by it to receive worse scores immediately.
2167  */
2168 void
2169 sched_exit_thread(struct thread *td, struct thread *child)
2170 {
2171 
2172 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2173 	    "prio:%d", child->td_priority);
2174 	/*
2175 	 * Give the child's runtime to the parent without returning the
2176 	 * sleep time as a penalty to the parent.  This causes shells that
2177 	 * launch expensive things to mark their children as expensive.
2178 	 */
2179 	thread_lock(td);
2180 	td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2181 	sched_interact_update(td);
2182 	sched_priority(td);
2183 	thread_unlock(td);
2184 }
2185 
2186 void
2187 sched_preempt(struct thread *td)
2188 {
2189 	struct tdq *tdq;
2190 
2191 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2192 
2193 	thread_lock(td);
2194 	tdq = TDQ_SELF();
2195 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2196 	tdq->tdq_ipipending = 0;
2197 	if (td->td_priority > tdq->tdq_lowpri) {
2198 		int flags;
2199 
2200 		flags = SW_INVOL | SW_PREEMPT;
2201 		if (td->td_critnest > 1)
2202 			td->td_owepreempt = 1;
2203 		else if (TD_IS_IDLETHREAD(td))
2204 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2205 		else
2206 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2207 	}
2208 	thread_unlock(td);
2209 }
2210 
2211 /*
2212  * Fix priorities on return to user-space.  Priorities may be elevated due
2213  * to static priorities in msleep() or similar.
2214  */
2215 void
2216 sched_userret(struct thread *td)
2217 {
2218 	/*
2219 	 * XXX we cheat slightly on the locking here to avoid locking in
2220 	 * the usual case.  Setting td_priority here is essentially an
2221 	 * incomplete workaround for not setting it properly elsewhere.
2222 	 * Now that some interrupt handlers are threads, not setting it
2223 	 * properly elsewhere can clobber it in the window between setting
2224 	 * it here and returning to user mode, so don't waste time setting
2225 	 * it perfectly here.
2226 	 */
2227 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2228 	    ("thread with borrowed priority returning to userland"));
2229 	if (td->td_priority != td->td_user_pri) {
2230 		thread_lock(td);
2231 		td->td_priority = td->td_user_pri;
2232 		td->td_base_pri = td->td_user_pri;
2233 		tdq_setlowpri(TDQ_SELF(), td);
2234 		thread_unlock(td);
2235         }
2236 }
2237 
2238 /*
2239  * Handle a stathz tick.  This is really only relevant for timeshare
2240  * threads.
2241  */
2242 void
2243 sched_clock(struct thread *td)
2244 {
2245 	struct tdq *tdq;
2246 	struct td_sched *ts;
2247 
2248 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2249 	tdq = TDQ_SELF();
2250 #ifdef SMP
2251 	/*
2252 	 * We run the long term load balancer infrequently on the first cpu.
2253 	 */
2254 	if (balance_tdq == tdq) {
2255 		if (balance_ticks && --balance_ticks == 0)
2256 			sched_balance();
2257 	}
2258 #endif
2259 	/*
2260 	 * Save the old switch count so we have a record of the last ticks
2261 	 * activity.   Initialize the new switch count based on our load.
2262 	 * If there is some activity seed it to reflect that.
2263 	 */
2264 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2265 	tdq->tdq_switchcnt = tdq->tdq_load;
2266 	/*
2267 	 * Advance the insert index once for each tick to ensure that all
2268 	 * threads get a chance to run.
2269 	 */
2270 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2271 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2272 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2273 			tdq->tdq_ridx = tdq->tdq_idx;
2274 	}
2275 	ts = td_get_sched(td);
2276 	sched_pctcpu_update(ts, 1);
2277 	if (td->td_pri_class & PRI_FIFO_BIT)
2278 		return;
2279 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2280 		/*
2281 		 * We used a tick; charge it to the thread so
2282 		 * that we can compute our interactivity.
2283 		 */
2284 		td_get_sched(td)->ts_runtime += tickincr;
2285 		sched_interact_update(td);
2286 		sched_priority(td);
2287 	}
2288 
2289 	/*
2290 	 * Force a context switch if the current thread has used up a full
2291 	 * time slice (default is 100ms).
2292 	 */
2293 	if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2294 		ts->ts_slice = 0;
2295 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2296 	}
2297 }
2298 
2299 u_int
2300 sched_estcpu(struct thread *td __unused)
2301 {
2302 
2303 	return (0);
2304 }
2305 
2306 /*
2307  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2308  * cooperative idle threads.
2309  */
2310 int
2311 sched_runnable(void)
2312 {
2313 	struct tdq *tdq;
2314 	int load;
2315 
2316 	load = 1;
2317 
2318 	tdq = TDQ_SELF();
2319 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2320 		if (tdq->tdq_load > 0)
2321 			goto out;
2322 	} else
2323 		if (tdq->tdq_load - 1 > 0)
2324 			goto out;
2325 	load = 0;
2326 out:
2327 	return (load);
2328 }
2329 
2330 /*
2331  * Choose the highest priority thread to run.  The thread is removed from
2332  * the run-queue while running however the load remains.  For SMP we set
2333  * the tdq in the global idle bitmask if it idles here.
2334  */
2335 struct thread *
2336 sched_choose(void)
2337 {
2338 	struct thread *td;
2339 	struct tdq *tdq;
2340 
2341 	tdq = TDQ_SELF();
2342 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2343 	td = tdq_choose(tdq);
2344 	if (td) {
2345 		tdq_runq_rem(tdq, td);
2346 		tdq->tdq_lowpri = td->td_priority;
2347 		return (td);
2348 	}
2349 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2350 	return (PCPU_GET(idlethread));
2351 }
2352 
2353 /*
2354  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2355  * we always request it once we exit a critical section.
2356  */
2357 static inline void
2358 sched_setpreempt(struct thread *td)
2359 {
2360 	struct thread *ctd;
2361 	int cpri;
2362 	int pri;
2363 
2364 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2365 
2366 	ctd = curthread;
2367 	pri = td->td_priority;
2368 	cpri = ctd->td_priority;
2369 	if (pri < cpri)
2370 		ctd->td_flags |= TDF_NEEDRESCHED;
2371 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2372 		return;
2373 	if (!sched_shouldpreempt(pri, cpri, 0))
2374 		return;
2375 	ctd->td_owepreempt = 1;
2376 }
2377 
2378 /*
2379  * Add a thread to a thread queue.  Select the appropriate runq and add the
2380  * thread to it.  This is the internal function called when the tdq is
2381  * predetermined.
2382  */
2383 void
2384 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2385 {
2386 
2387 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2388 	KASSERT((td->td_inhibitors == 0),
2389 	    ("sched_add: trying to run inhibited thread"));
2390 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2391 	    ("sched_add: bad thread state"));
2392 	KASSERT(td->td_flags & TDF_INMEM,
2393 	    ("sched_add: thread swapped out"));
2394 
2395 	if (td->td_priority < tdq->tdq_lowpri)
2396 		tdq->tdq_lowpri = td->td_priority;
2397 	tdq_runq_add(tdq, td, flags);
2398 	tdq_load_add(tdq, td);
2399 }
2400 
2401 /*
2402  * Select the target thread queue and add a thread to it.  Request
2403  * preemption or IPI a remote processor if required.
2404  */
2405 void
2406 sched_add(struct thread *td, int flags)
2407 {
2408 	struct tdq *tdq;
2409 #ifdef SMP
2410 	int cpu;
2411 #endif
2412 
2413 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2414 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2415 	    sched_tdname(curthread));
2416 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2417 	    KTR_ATTR_LINKED, sched_tdname(td));
2418 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2419 	    flags & SRQ_PREEMPTED);
2420 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2421 	/*
2422 	 * Recalculate the priority before we select the target cpu or
2423 	 * run-queue.
2424 	 */
2425 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2426 		sched_priority(td);
2427 #ifdef SMP
2428 	/*
2429 	 * Pick the destination cpu and if it isn't ours transfer to the
2430 	 * target cpu.
2431 	 */
2432 	cpu = sched_pickcpu(td, flags);
2433 	tdq = sched_setcpu(td, cpu, flags);
2434 	tdq_add(tdq, td, flags);
2435 	if (cpu != PCPU_GET(cpuid)) {
2436 		tdq_notify(tdq, td);
2437 		return;
2438 	}
2439 #else
2440 	tdq = TDQ_SELF();
2441 	TDQ_LOCK(tdq);
2442 	/*
2443 	 * Now that the thread is moving to the run-queue, set the lock
2444 	 * to the scheduler's lock.
2445 	 */
2446 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2447 	tdq_add(tdq, td, flags);
2448 #endif
2449 	if (!(flags & SRQ_YIELDING))
2450 		sched_setpreempt(td);
2451 }
2452 
2453 /*
2454  * Remove a thread from a run-queue without running it.  This is used
2455  * when we're stealing a thread from a remote queue.  Otherwise all threads
2456  * exit by calling sched_exit_thread() and sched_throw() themselves.
2457  */
2458 void
2459 sched_rem(struct thread *td)
2460 {
2461 	struct tdq *tdq;
2462 
2463 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2464 	    "prio:%d", td->td_priority);
2465 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2466 	tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2467 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2468 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2469 	KASSERT(TD_ON_RUNQ(td),
2470 	    ("sched_rem: thread not on run queue"));
2471 	tdq_runq_rem(tdq, td);
2472 	tdq_load_rem(tdq, td);
2473 	TD_SET_CAN_RUN(td);
2474 	if (td->td_priority == tdq->tdq_lowpri)
2475 		tdq_setlowpri(tdq, NULL);
2476 }
2477 
2478 /*
2479  * Fetch cpu utilization information.  Updates on demand.
2480  */
2481 fixpt_t
2482 sched_pctcpu(struct thread *td)
2483 {
2484 	fixpt_t pctcpu;
2485 	struct td_sched *ts;
2486 
2487 	pctcpu = 0;
2488 	ts = td_get_sched(td);
2489 
2490 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2491 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2492 	if (ts->ts_ticks) {
2493 		int rtick;
2494 
2495 		/* How many rtick per second ? */
2496 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2497 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2498 	}
2499 
2500 	return (pctcpu);
2501 }
2502 
2503 /*
2504  * Enforce affinity settings for a thread.  Called after adjustments to
2505  * cpumask.
2506  */
2507 void
2508 sched_affinity(struct thread *td)
2509 {
2510 #ifdef SMP
2511 	struct td_sched *ts;
2512 
2513 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2514 	ts = td_get_sched(td);
2515 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2516 		return;
2517 	if (TD_ON_RUNQ(td)) {
2518 		sched_rem(td);
2519 		sched_add(td, SRQ_BORING);
2520 		return;
2521 	}
2522 	if (!TD_IS_RUNNING(td))
2523 		return;
2524 	/*
2525 	 * Force a switch before returning to userspace.  If the
2526 	 * target thread is not running locally send an ipi to force
2527 	 * the issue.
2528 	 */
2529 	td->td_flags |= TDF_NEEDRESCHED;
2530 	if (td != curthread)
2531 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2532 #endif
2533 }
2534 
2535 /*
2536  * Bind a thread to a target cpu.
2537  */
2538 void
2539 sched_bind(struct thread *td, int cpu)
2540 {
2541 	struct td_sched *ts;
2542 
2543 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2544 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2545 	ts = td_get_sched(td);
2546 	if (ts->ts_flags & TSF_BOUND)
2547 		sched_unbind(td);
2548 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2549 	ts->ts_flags |= TSF_BOUND;
2550 	sched_pin();
2551 	if (PCPU_GET(cpuid) == cpu)
2552 		return;
2553 	ts->ts_cpu = cpu;
2554 	/* When we return from mi_switch we'll be on the correct cpu. */
2555 	mi_switch(SW_VOL, NULL);
2556 }
2557 
2558 /*
2559  * Release a bound thread.
2560  */
2561 void
2562 sched_unbind(struct thread *td)
2563 {
2564 	struct td_sched *ts;
2565 
2566 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2567 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2568 	ts = td_get_sched(td);
2569 	if ((ts->ts_flags & TSF_BOUND) == 0)
2570 		return;
2571 	ts->ts_flags &= ~TSF_BOUND;
2572 	sched_unpin();
2573 }
2574 
2575 int
2576 sched_is_bound(struct thread *td)
2577 {
2578 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2579 	return (td_get_sched(td)->ts_flags & TSF_BOUND);
2580 }
2581 
2582 /*
2583  * Basic yield call.
2584  */
2585 void
2586 sched_relinquish(struct thread *td)
2587 {
2588 	thread_lock(td);
2589 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2590 	thread_unlock(td);
2591 }
2592 
2593 /*
2594  * Return the total system load.
2595  */
2596 int
2597 sched_load(void)
2598 {
2599 #ifdef SMP
2600 	int total;
2601 	int i;
2602 
2603 	total = 0;
2604 	CPU_FOREACH(i)
2605 		total += TDQ_CPU(i)->tdq_sysload;
2606 	return (total);
2607 #else
2608 	return (TDQ_SELF()->tdq_sysload);
2609 #endif
2610 }
2611 
2612 int
2613 sched_sizeof_proc(void)
2614 {
2615 	return (sizeof(struct proc));
2616 }
2617 
2618 int
2619 sched_sizeof_thread(void)
2620 {
2621 	return (sizeof(struct thread) + sizeof(struct td_sched));
2622 }
2623 
2624 #ifdef SMP
2625 #define	TDQ_IDLESPIN(tdq)						\
2626     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2627 #else
2628 #define	TDQ_IDLESPIN(tdq)	1
2629 #endif
2630 
2631 /*
2632  * The actual idle process.
2633  */
2634 void
2635 sched_idletd(void *dummy)
2636 {
2637 	struct thread *td;
2638 	struct tdq *tdq;
2639 	int oldswitchcnt, switchcnt;
2640 	int i;
2641 
2642 	mtx_assert(&Giant, MA_NOTOWNED);
2643 	td = curthread;
2644 	tdq = TDQ_SELF();
2645 	THREAD_NO_SLEEPING();
2646 	oldswitchcnt = -1;
2647 	for (;;) {
2648 		if (tdq->tdq_load) {
2649 			thread_lock(td);
2650 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2651 			thread_unlock(td);
2652 		}
2653 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2654 #ifdef SMP
2655 		if (switchcnt != oldswitchcnt) {
2656 			oldswitchcnt = switchcnt;
2657 			if (tdq_idled(tdq) == 0)
2658 				continue;
2659 		}
2660 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2661 #else
2662 		oldswitchcnt = switchcnt;
2663 #endif
2664 		/*
2665 		 * If we're switching very frequently, spin while checking
2666 		 * for load rather than entering a low power state that
2667 		 * may require an IPI.  However, don't do any busy
2668 		 * loops while on SMT machines as this simply steals
2669 		 * cycles from cores doing useful work.
2670 		 */
2671 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2672 			for (i = 0; i < sched_idlespins; i++) {
2673 				if (tdq->tdq_load)
2674 					break;
2675 				cpu_spinwait();
2676 			}
2677 		}
2678 
2679 		/* If there was context switch during spin, restart it. */
2680 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2681 		if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2682 			continue;
2683 
2684 		/* Run main MD idle handler. */
2685 		tdq->tdq_cpu_idle = 1;
2686 		/*
2687 		 * Make sure that tdq_cpu_idle update is globally visible
2688 		 * before cpu_idle() read tdq_load.  The order is important
2689 		 * to avoid race with tdq_notify.
2690 		 */
2691 		atomic_thread_fence_seq_cst();
2692 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2693 		tdq->tdq_cpu_idle = 0;
2694 
2695 		/*
2696 		 * Account thread-less hardware interrupts and
2697 		 * other wakeup reasons equal to context switches.
2698 		 */
2699 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2700 		if (switchcnt != oldswitchcnt)
2701 			continue;
2702 		tdq->tdq_switchcnt++;
2703 		oldswitchcnt++;
2704 	}
2705 }
2706 
2707 /*
2708  * A CPU is entering for the first time or a thread is exiting.
2709  */
2710 void
2711 sched_throw(struct thread *td)
2712 {
2713 	struct thread *newtd;
2714 	struct tdq *tdq;
2715 
2716 	tdq = TDQ_SELF();
2717 	if (td == NULL) {
2718 		/* Correct spinlock nesting and acquire the correct lock. */
2719 		TDQ_LOCK(tdq);
2720 		spinlock_exit();
2721 		PCPU_SET(switchtime, cpu_ticks());
2722 		PCPU_SET(switchticks, ticks);
2723 	} else {
2724 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2725 		tdq_load_rem(tdq, td);
2726 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2727 		td->td_lastcpu = td->td_oncpu;
2728 		td->td_oncpu = NOCPU;
2729 	}
2730 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2731 	newtd = choosethread();
2732 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2733 	cpu_throw(td, newtd);		/* doesn't return */
2734 }
2735 
2736 /*
2737  * This is called from fork_exit().  Just acquire the correct locks and
2738  * let fork do the rest of the work.
2739  */
2740 void
2741 sched_fork_exit(struct thread *td)
2742 {
2743 	struct tdq *tdq;
2744 	int cpuid;
2745 
2746 	/*
2747 	 * Finish setting up thread glue so that it begins execution in a
2748 	 * non-nested critical section with the scheduler lock held.
2749 	 */
2750 	cpuid = PCPU_GET(cpuid);
2751 	tdq = TDQ_CPU(cpuid);
2752 	if (TD_IS_IDLETHREAD(td))
2753 		td->td_lock = TDQ_LOCKPTR(tdq);
2754 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2755 	td->td_oncpu = cpuid;
2756 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2757 	lock_profile_obtain_lock_success(
2758 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2759 }
2760 
2761 /*
2762  * Create on first use to catch odd startup conditons.
2763  */
2764 char *
2765 sched_tdname(struct thread *td)
2766 {
2767 #ifdef KTR
2768 	struct td_sched *ts;
2769 
2770 	ts = td_get_sched(td);
2771 	if (ts->ts_name[0] == '\0')
2772 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2773 		    "%s tid %d", td->td_name, td->td_tid);
2774 	return (ts->ts_name);
2775 #else
2776 	return (td->td_name);
2777 #endif
2778 }
2779 
2780 #ifdef KTR
2781 void
2782 sched_clear_tdname(struct thread *td)
2783 {
2784 	struct td_sched *ts;
2785 
2786 	ts = td_get_sched(td);
2787 	ts->ts_name[0] = '\0';
2788 }
2789 #endif
2790 
2791 #ifdef SMP
2792 
2793 /*
2794  * Build the CPU topology dump string. Is recursively called to collect
2795  * the topology tree.
2796  */
2797 static int
2798 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2799     int indent)
2800 {
2801 	char cpusetbuf[CPUSETBUFSIZ];
2802 	int i, first;
2803 
2804 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2805 	    "", 1 + indent / 2, cg->cg_level);
2806 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2807 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2808 	first = TRUE;
2809 	for (i = 0; i < MAXCPU; i++) {
2810 		if (CPU_ISSET(i, &cg->cg_mask)) {
2811 			if (!first)
2812 				sbuf_printf(sb, ", ");
2813 			else
2814 				first = FALSE;
2815 			sbuf_printf(sb, "%d", i);
2816 		}
2817 	}
2818 	sbuf_printf(sb, "</cpu>\n");
2819 
2820 	if (cg->cg_flags != 0) {
2821 		sbuf_printf(sb, "%*s <flags>", indent, "");
2822 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2823 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2824 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2825 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2826 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2827 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2828 		sbuf_printf(sb, "</flags>\n");
2829 	}
2830 
2831 	if (cg->cg_children > 0) {
2832 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2833 		for (i = 0; i < cg->cg_children; i++)
2834 			sysctl_kern_sched_topology_spec_internal(sb,
2835 			    &cg->cg_child[i], indent+2);
2836 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2837 	}
2838 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2839 	return (0);
2840 }
2841 
2842 /*
2843  * Sysctl handler for retrieving topology dump. It's a wrapper for
2844  * the recursive sysctl_kern_smp_topology_spec_internal().
2845  */
2846 static int
2847 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2848 {
2849 	struct sbuf *topo;
2850 	int err;
2851 
2852 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2853 
2854 	topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
2855 	if (topo == NULL)
2856 		return (ENOMEM);
2857 
2858 	sbuf_printf(topo, "<groups>\n");
2859 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2860 	sbuf_printf(topo, "</groups>\n");
2861 
2862 	if (err == 0) {
2863 		err = sbuf_finish(topo);
2864 	}
2865 	sbuf_delete(topo);
2866 	return (err);
2867 }
2868 
2869 #endif
2870 
2871 static int
2872 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2873 {
2874 	int error, new_val, period;
2875 
2876 	period = 1000000 / realstathz;
2877 	new_val = period * sched_slice;
2878 	error = sysctl_handle_int(oidp, &new_val, 0, req);
2879 	if (error != 0 || req->newptr == NULL)
2880 		return (error);
2881 	if (new_val <= 0)
2882 		return (EINVAL);
2883 	sched_slice = imax(1, (new_val + period / 2) / period);
2884 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
2885 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2886 	    realstathz);
2887 	return (0);
2888 }
2889 
2890 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2891 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2892     "Scheduler name");
2893 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2894     NULL, 0, sysctl_kern_quantum, "I",
2895     "Quantum for timeshare threads in microseconds");
2896 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2897     "Quantum for timeshare threads in stathz ticks");
2898 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2899     "Interactivity score threshold");
2900 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2901     &preempt_thresh, 0,
2902     "Maximal (lowest) priority for preemption");
2903 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2904     "Assign static kernel priorities to sleeping threads");
2905 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2906     "Number of times idle thread will spin waiting for new work");
2907 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2908     &sched_idlespinthresh, 0,
2909     "Threshold before we will permit idle thread spinning");
2910 #ifdef SMP
2911 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2912     "Number of hz ticks to keep thread affinity for");
2913 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2914     "Enables the long-term load balancer");
2915 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2916     &balance_interval, 0,
2917     "Average period in stathz ticks to run the long-term balancer");
2918 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2919     "Attempts to steal work from other cores before idling");
2920 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2921     "Minimum load on remote CPU before we'll steal");
2922 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2923     CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2924     "XML dump of detected CPU topology");
2925 #endif
2926 
2927 /* ps compat.  All cpu percentages from ULE are weighted. */
2928 static int ccpu = 0;
2929 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2930