xref: /freebsd/sys/kern/sched_ule.c (revision 84dfba8d183d31e3412639ecb4b8ad4433cf7e80)
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/lock.h>
50 #include <sys/mutex.h>
51 #include <sys/proc.h>
52 #include <sys/resource.h>
53 #include <sys/resourcevar.h>
54 #include <sys/sched.h>
55 #include <sys/sdt.h>
56 #include <sys/smp.h>
57 #include <sys/sx.h>
58 #include <sys/sysctl.h>
59 #include <sys/sysproto.h>
60 #include <sys/turnstile.h>
61 #include <sys/umtx.h>
62 #include <sys/vmmeter.h>
63 #include <sys/cpuset.h>
64 #include <sys/sbuf.h>
65 
66 #ifdef HWPMC_HOOKS
67 #include <sys/pmckern.h>
68 #endif
69 
70 #ifdef KDTRACE_HOOKS
71 #include <sys/dtrace_bsd.h>
72 int				dtrace_vtime_active;
73 dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
74 #endif
75 
76 #include <machine/cpu.h>
77 #include <machine/smp.h>
78 
79 #if defined(__powerpc__) && defined(BOOKE_E500)
80 #error "This architecture is not currently compatible with ULE"
81 #endif
82 
83 #define	KTR_ULE	0
84 
85 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
86 #define	TDQ_NAME_LEN	(sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
87 #define	TDQ_LOADNAME_LEN	(sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
88 
89 /*
90  * Thread scheduler specific section.  All fields are protected
91  * by the thread lock.
92  */
93 struct td_sched {
94 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
95 	short		ts_flags;	/* TSF_* flags. */
96 	u_char		ts_cpu;		/* CPU that we have affinity for. */
97 	int		ts_rltick;	/* Real last tick, for affinity. */
98 	int		ts_slice;	/* Ticks of slice remaining. */
99 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
100 	u_int		ts_runtime;	/* Number of ticks we were running */
101 	int		ts_ltick;	/* Last tick that we were running on */
102 	int		ts_ftick;	/* First tick that we were running on */
103 	int		ts_ticks;	/* Tick count */
104 #ifdef KTR
105 	char		ts_name[TS_NAME_LEN];
106 #endif
107 };
108 /* flags kept in ts_flags */
109 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
110 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
111 
112 static struct td_sched td_sched0;
113 
114 #define	THREAD_CAN_MIGRATE(td)	((td)->td_pinned == 0)
115 #define	THREAD_CAN_SCHED(td, cpu)	\
116     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
117 
118 /*
119  * Priority ranges used for interactive and non-interactive timeshare
120  * threads.  The timeshare priorities are split up into four ranges.
121  * The first range handles interactive threads.  The last three ranges
122  * (NHALF, x, and NHALF) handle non-interactive threads with the outer
123  * ranges supporting nice values.
124  */
125 #define	PRI_TIMESHARE_RANGE	(PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
126 #define	PRI_INTERACT_RANGE	((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
127 #define	PRI_BATCH_RANGE		(PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
128 
129 #define	PRI_MIN_INTERACT	PRI_MIN_TIMESHARE
130 #define	PRI_MAX_INTERACT	(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
131 #define	PRI_MIN_BATCH		(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
132 #define	PRI_MAX_BATCH		PRI_MAX_TIMESHARE
133 
134 /*
135  * Cpu percentage computation macros and defines.
136  *
137  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
138  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
139  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
140  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
141  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
142  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
143  */
144 #define	SCHED_TICK_SECS		10
145 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
146 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
147 #define	SCHED_TICK_SHIFT	10
148 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
149 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
150 
151 /*
152  * These macros determine priorities for non-interactive threads.  They are
153  * assigned a priority based on their recent cpu utilization as expressed
154  * by the ratio of ticks to the tick total.  NHALF priorities at the start
155  * and end of the MIN to MAX timeshare range are only reachable with negative
156  * or positive nice respectively.
157  *
158  * PRI_RANGE:	Priority range for utilization dependent priorities.
159  * PRI_NRESV:	Number of nice values.
160  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
161  * PRI_NICE:	Determines the part of the priority inherited from nice.
162  */
163 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
164 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
165 #define	SCHED_PRI_MIN		(PRI_MIN_BATCH + SCHED_PRI_NHALF)
166 #define	SCHED_PRI_MAX		(PRI_MAX_BATCH - SCHED_PRI_NHALF)
167 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
168 #define	SCHED_PRI_TICKS(ts)						\
169     (SCHED_TICK_HZ((ts)) /						\
170     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
171 #define	SCHED_PRI_NICE(nice)	(nice)
172 
173 /*
174  * These determine the interactivity of a process.  Interactivity differs from
175  * cpu utilization in that it expresses the voluntary time slept vs time ran
176  * while cpu utilization includes all time not running.  This more accurately
177  * models the intent of the thread.
178  *
179  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
180  *		before throttling back.
181  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
182  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
183  * INTERACT_THRESH:	Threshold for placement on the current runq.
184  */
185 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
186 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
187 #define	SCHED_INTERACT_MAX	(100)
188 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
189 #define	SCHED_INTERACT_THRESH	(30)
190 
191 /*
192  * These parameters determine the slice behavior for batch work.
193  */
194 #define	SCHED_SLICE_DEFAULT_DIVISOR	10	/* ~94 ms, 12 stathz ticks. */
195 #define	SCHED_SLICE_MIN_DIVISOR		6	/* DEFAULT/MIN = ~16 ms. */
196 
197 /* Flags kept in td_flags. */
198 #define	TDF_SLICEEND	TDF_SCHED2	/* Thread time slice is over. */
199 
200 /*
201  * tickincr:		Converts a stathz tick into a hz domain scaled by
202  *			the shift factor.  Without the shift the error rate
203  *			due to rounding would be unacceptably high.
204  * realstathz:		stathz is sometimes 0 and run off of hz.
205  * sched_slice:		Runtime of each thread before rescheduling.
206  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
207  */
208 static int sched_interact = SCHED_INTERACT_THRESH;
209 static int tickincr = 8 << SCHED_TICK_SHIFT;
210 static int realstathz = 127;	/* reset during boot. */
211 static int sched_slice = 10;	/* reset during boot. */
212 static int sched_slice_min = 1;	/* reset during boot. */
213 #ifdef PREEMPTION
214 #ifdef FULL_PREEMPTION
215 static int preempt_thresh = PRI_MAX_IDLE;
216 #else
217 static int preempt_thresh = PRI_MIN_KERN;
218 #endif
219 #else
220 static int preempt_thresh = 0;
221 #endif
222 static int static_boost = PRI_MIN_BATCH;
223 static int sched_idlespins = 10000;
224 static int sched_idlespinthresh = -1;
225 
226 /*
227  * tdq - per processor runqs and statistics.  All fields are protected by the
228  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
229  * locking in sched_pickcpu();
230  */
231 struct tdq {
232 	/*
233 	 * Ordered to improve efficiency of cpu_search() and switch().
234 	 * tdq_lock is padded to avoid false sharing with tdq_load and
235 	 * tdq_cpu_idle.
236 	 */
237 	struct mtx_padalign tdq_lock;		/* run queue lock. */
238 	struct cpu_group *tdq_cg;		/* Pointer to cpu topology. */
239 	volatile int	tdq_load;		/* Aggregate load. */
240 	volatile int	tdq_cpu_idle;		/* cpu_idle() is active. */
241 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
242 	int		tdq_transferable;	/* Transferable thread count. */
243 	short		tdq_switchcnt;		/* Switches this tick. */
244 	short		tdq_oldswitchcnt;	/* Switches last tick. */
245 	u_char		tdq_lowpri;		/* Lowest priority thread. */
246 	u_char		tdq_ipipending;		/* IPI pending. */
247 	u_char		tdq_idx;		/* Current insert index. */
248 	u_char		tdq_ridx;		/* Current removal index. */
249 	struct runq	tdq_realtime;		/* real-time run queue. */
250 	struct runq	tdq_timeshare;		/* timeshare run queue. */
251 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
252 	char		tdq_name[TDQ_NAME_LEN];
253 #ifdef KTR
254 	char		tdq_loadname[TDQ_LOADNAME_LEN];
255 #endif
256 } __aligned(64);
257 
258 /* Idle thread states and config. */
259 #define	TDQ_RUNNING	1
260 #define	TDQ_IDLE	2
261 
262 #ifdef SMP
263 struct cpu_group *cpu_top;		/* CPU topology */
264 
265 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
266 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
267 
268 /*
269  * Run-time tunables.
270  */
271 static int rebalance = 1;
272 static int balance_interval = 128;	/* Default set in sched_initticks(). */
273 static int affinity;
274 static int steal_idle = 1;
275 static int steal_thresh = 2;
276 
277 /*
278  * One thread queue per processor.
279  */
280 static struct tdq	tdq_cpu[MAXCPU];
281 static struct tdq	*balance_tdq;
282 static int balance_ticks;
283 static DPCPU_DEFINE(uint32_t, randomval);
284 
285 #define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
286 #define	TDQ_CPU(x)	(&tdq_cpu[(x)])
287 #define	TDQ_ID(x)	((int)((x) - tdq_cpu))
288 #else	/* !SMP */
289 static struct tdq	tdq_cpu;
290 
291 #define	TDQ_ID(x)	(0)
292 #define	TDQ_SELF()	(&tdq_cpu)
293 #define	TDQ_CPU(x)	(&tdq_cpu)
294 #endif
295 
296 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
297 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
298 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
299 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
300 #define	TDQ_LOCKPTR(t)		((struct mtx *)(&(t)->tdq_lock))
301 
302 static void sched_priority(struct thread *);
303 static void sched_thread_priority(struct thread *, u_char);
304 static int sched_interact_score(struct thread *);
305 static void sched_interact_update(struct thread *);
306 static void sched_interact_fork(struct thread *);
307 static void sched_pctcpu_update(struct td_sched *, int);
308 
309 /* Operations on per processor queues */
310 static struct thread *tdq_choose(struct tdq *);
311 static void tdq_setup(struct tdq *);
312 static void tdq_load_add(struct tdq *, struct thread *);
313 static void tdq_load_rem(struct tdq *, struct thread *);
314 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
315 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
316 static inline int sched_shouldpreempt(int, int, int);
317 void tdq_print(int cpu);
318 static void runq_print(struct runq *rq);
319 static void tdq_add(struct tdq *, struct thread *, int);
320 #ifdef SMP
321 static int tdq_move(struct tdq *, struct tdq *);
322 static int tdq_idled(struct tdq *);
323 static void tdq_notify(struct tdq *, struct thread *);
324 static struct thread *tdq_steal(struct tdq *, int);
325 static struct thread *runq_steal(struct runq *, int);
326 static int sched_pickcpu(struct thread *, int);
327 static void sched_balance(void);
328 static int sched_balance_pair(struct tdq *, struct tdq *);
329 static inline struct tdq *sched_setcpu(struct thread *, int, int);
330 static inline void thread_unblock_switch(struct thread *, struct mtx *);
331 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
332 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
333 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
334     struct cpu_group *cg, int indent);
335 #endif
336 
337 static void sched_setup(void *dummy);
338 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
339 
340 static void sched_initticks(void *dummy);
341 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
342     NULL);
343 
344 SDT_PROVIDER_DEFINE(sched);
345 
346 SDT_PROBE_DEFINE3(sched, , , change_pri, change-pri, "struct thread *",
347     "struct proc *", "uint8_t");
348 SDT_PROBE_DEFINE3(sched, , , dequeue, dequeue, "struct thread *",
349     "struct proc *", "void *");
350 SDT_PROBE_DEFINE4(sched, , , enqueue, enqueue, "struct thread *",
351     "struct proc *", "void *", "int");
352 SDT_PROBE_DEFINE4(sched, , , lend_pri, lend-pri, "struct thread *",
353     "struct proc *", "uint8_t", "struct thread *");
354 SDT_PROBE_DEFINE2(sched, , , load_change, load-change, "int", "int");
355 SDT_PROBE_DEFINE2(sched, , , off_cpu, off-cpu, "struct thread *",
356     "struct proc *");
357 SDT_PROBE_DEFINE(sched, , , on_cpu, on-cpu);
358 SDT_PROBE_DEFINE(sched, , , remain_cpu, remain-cpu);
359 SDT_PROBE_DEFINE2(sched, , , surrender, surrender, "struct thread *",
360     "struct proc *");
361 
362 /*
363  * Print the threads waiting on a run-queue.
364  */
365 static void
366 runq_print(struct runq *rq)
367 {
368 	struct rqhead *rqh;
369 	struct thread *td;
370 	int pri;
371 	int j;
372 	int i;
373 
374 	for (i = 0; i < RQB_LEN; i++) {
375 		printf("\t\trunq bits %d 0x%zx\n",
376 		    i, rq->rq_status.rqb_bits[i]);
377 		for (j = 0; j < RQB_BPW; j++)
378 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
379 				pri = j + (i << RQB_L2BPW);
380 				rqh = &rq->rq_queues[pri];
381 				TAILQ_FOREACH(td, rqh, td_runq) {
382 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
383 					    td, td->td_name, td->td_priority,
384 					    td->td_rqindex, pri);
385 				}
386 			}
387 	}
388 }
389 
390 /*
391  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
392  */
393 void
394 tdq_print(int cpu)
395 {
396 	struct tdq *tdq;
397 
398 	tdq = TDQ_CPU(cpu);
399 
400 	printf("tdq %d:\n", TDQ_ID(tdq));
401 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
402 	printf("\tLock name:      %s\n", tdq->tdq_name);
403 	printf("\tload:           %d\n", tdq->tdq_load);
404 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
405 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
406 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
407 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
408 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
409 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
410 	printf("\trealtime runq:\n");
411 	runq_print(&tdq->tdq_realtime);
412 	printf("\ttimeshare runq:\n");
413 	runq_print(&tdq->tdq_timeshare);
414 	printf("\tidle runq:\n");
415 	runq_print(&tdq->tdq_idle);
416 }
417 
418 static inline int
419 sched_shouldpreempt(int pri, int cpri, int remote)
420 {
421 	/*
422 	 * If the new priority is not better than the current priority there is
423 	 * nothing to do.
424 	 */
425 	if (pri >= cpri)
426 		return (0);
427 	/*
428 	 * Always preempt idle.
429 	 */
430 	if (cpri >= PRI_MIN_IDLE)
431 		return (1);
432 	/*
433 	 * If preemption is disabled don't preempt others.
434 	 */
435 	if (preempt_thresh == 0)
436 		return (0);
437 	/*
438 	 * Preempt if we exceed the threshold.
439 	 */
440 	if (pri <= preempt_thresh)
441 		return (1);
442 	/*
443 	 * If we're interactive or better and there is non-interactive
444 	 * or worse running preempt only remote processors.
445 	 */
446 	if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
447 		return (1);
448 	return (0);
449 }
450 
451 /*
452  * Add a thread to the actual run-queue.  Keeps transferable counts up to
453  * date with what is actually on the run-queue.  Selects the correct
454  * queue position for timeshare threads.
455  */
456 static __inline void
457 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
458 {
459 	struct td_sched *ts;
460 	u_char pri;
461 
462 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
463 	THREAD_LOCK_ASSERT(td, MA_OWNED);
464 
465 	pri = td->td_priority;
466 	ts = td->td_sched;
467 	TD_SET_RUNQ(td);
468 	if (THREAD_CAN_MIGRATE(td)) {
469 		tdq->tdq_transferable++;
470 		ts->ts_flags |= TSF_XFERABLE;
471 	}
472 	if (pri < PRI_MIN_BATCH) {
473 		ts->ts_runq = &tdq->tdq_realtime;
474 	} else if (pri <= PRI_MAX_BATCH) {
475 		ts->ts_runq = &tdq->tdq_timeshare;
476 		KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
477 			("Invalid priority %d on timeshare runq", pri));
478 		/*
479 		 * This queue contains only priorities between MIN and MAX
480 		 * realtime.  Use the whole queue to represent these values.
481 		 */
482 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
483 			pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
484 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
485 			/*
486 			 * This effectively shortens the queue by one so we
487 			 * can have a one slot difference between idx and
488 			 * ridx while we wait for threads to drain.
489 			 */
490 			if (tdq->tdq_ridx != tdq->tdq_idx &&
491 			    pri == tdq->tdq_ridx)
492 				pri = (unsigned char)(pri - 1) % RQ_NQS;
493 		} else
494 			pri = tdq->tdq_ridx;
495 		runq_add_pri(ts->ts_runq, td, pri, flags);
496 		return;
497 	} else
498 		ts->ts_runq = &tdq->tdq_idle;
499 	runq_add(ts->ts_runq, td, flags);
500 }
501 
502 /*
503  * Remove a thread from a run-queue.  This typically happens when a thread
504  * is selected to run.  Running threads are not on the queue and the
505  * transferable count does not reflect them.
506  */
507 static __inline void
508 tdq_runq_rem(struct tdq *tdq, struct thread *td)
509 {
510 	struct td_sched *ts;
511 
512 	ts = td->td_sched;
513 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
514 	KASSERT(ts->ts_runq != NULL,
515 	    ("tdq_runq_remove: thread %p null ts_runq", td));
516 	if (ts->ts_flags & TSF_XFERABLE) {
517 		tdq->tdq_transferable--;
518 		ts->ts_flags &= ~TSF_XFERABLE;
519 	}
520 	if (ts->ts_runq == &tdq->tdq_timeshare) {
521 		if (tdq->tdq_idx != tdq->tdq_ridx)
522 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
523 		else
524 			runq_remove_idx(ts->ts_runq, td, NULL);
525 	} else
526 		runq_remove(ts->ts_runq, td);
527 }
528 
529 /*
530  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
531  * for this thread to the referenced thread queue.
532  */
533 static void
534 tdq_load_add(struct tdq *tdq, struct thread *td)
535 {
536 
537 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
538 	THREAD_LOCK_ASSERT(td, MA_OWNED);
539 
540 	tdq->tdq_load++;
541 	if ((td->td_flags & TDF_NOLOAD) == 0)
542 		tdq->tdq_sysload++;
543 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
544 	SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load);
545 }
546 
547 /*
548  * Remove the load from a thread that is transitioning to a sleep state or
549  * exiting.
550  */
551 static void
552 tdq_load_rem(struct tdq *tdq, struct thread *td)
553 {
554 
555 	THREAD_LOCK_ASSERT(td, MA_OWNED);
556 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
557 	KASSERT(tdq->tdq_load != 0,
558 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
559 
560 	tdq->tdq_load--;
561 	if ((td->td_flags & TDF_NOLOAD) == 0)
562 		tdq->tdq_sysload--;
563 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
564 	SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load);
565 }
566 
567 /*
568  * Bound timeshare latency by decreasing slice size as load increases.  We
569  * consider the maximum latency as the sum of the threads waiting to run
570  * aside from curthread and target no more than sched_slice latency but
571  * no less than sched_slice_min runtime.
572  */
573 static inline int
574 tdq_slice(struct tdq *tdq)
575 {
576 	int load;
577 
578 	/*
579 	 * It is safe to use sys_load here because this is called from
580 	 * contexts where timeshare threads are running and so there
581 	 * cannot be higher priority load in the system.
582 	 */
583 	load = tdq->tdq_sysload - 1;
584 	if (load >= SCHED_SLICE_MIN_DIVISOR)
585 		return (sched_slice_min);
586 	if (load <= 1)
587 		return (sched_slice);
588 	return (sched_slice / load);
589 }
590 
591 /*
592  * Set lowpri to its exact value by searching the run-queue and
593  * evaluating curthread.  curthread may be passed as an optimization.
594  */
595 static void
596 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
597 {
598 	struct thread *td;
599 
600 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
601 	if (ctd == NULL)
602 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
603 	td = tdq_choose(tdq);
604 	if (td == NULL || td->td_priority > ctd->td_priority)
605 		tdq->tdq_lowpri = ctd->td_priority;
606 	else
607 		tdq->tdq_lowpri = td->td_priority;
608 }
609 
610 #ifdef SMP
611 struct cpu_search {
612 	cpuset_t cs_mask;
613 	u_int	cs_prefer;
614 	int	cs_pri;		/* Min priority for low. */
615 	int	cs_limit;	/* Max load for low, min load for high. */
616 	int	cs_cpu;
617 	int	cs_load;
618 };
619 
620 #define	CPU_SEARCH_LOWEST	0x1
621 #define	CPU_SEARCH_HIGHEST	0x2
622 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
623 
624 #define	CPUSET_FOREACH(cpu, mask)				\
625 	for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++)		\
626 		if (CPU_ISSET(cpu, &mask))
627 
628 static __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low,
629     struct cpu_search *high, const int match);
630 int cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low);
631 int cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high);
632 int cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
633     struct cpu_search *high);
634 
635 /*
636  * Search the tree of cpu_groups for the lowest or highest loaded cpu
637  * according to the match argument.  This routine actually compares the
638  * load on all paths through the tree and finds the least loaded cpu on
639  * the least loaded path, which may differ from the least loaded cpu in
640  * the system.  This balances work among caches and busses.
641  *
642  * This inline is instantiated in three forms below using constants for the
643  * match argument.  It is reduced to the minimum set for each case.  It is
644  * also recursive to the depth of the tree.
645  */
646 static __inline int
647 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
648     struct cpu_search *high, const int match)
649 {
650 	struct cpu_search lgroup;
651 	struct cpu_search hgroup;
652 	cpuset_t cpumask;
653 	struct cpu_group *child;
654 	struct tdq *tdq;
655 	int cpu, i, hload, lload, load, total, rnd, *rndptr;
656 
657 	total = 0;
658 	cpumask = cg->cg_mask;
659 	if (match & CPU_SEARCH_LOWEST) {
660 		lload = INT_MAX;
661 		lgroup = *low;
662 	}
663 	if (match & CPU_SEARCH_HIGHEST) {
664 		hload = INT_MIN;
665 		hgroup = *high;
666 	}
667 
668 	/* Iterate through the child CPU groups and then remaining CPUs. */
669 	for (i = cg->cg_children, cpu = mp_maxid; ; ) {
670 		if (i == 0) {
671 #ifdef HAVE_INLINE_FFSL
672 			cpu = CPU_FFS(&cpumask) - 1;
673 #else
674 			while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
675 				cpu--;
676 #endif
677 			if (cpu < 0)
678 				break;
679 			child = NULL;
680 		} else
681 			child = &cg->cg_child[i - 1];
682 
683 		if (match & CPU_SEARCH_LOWEST)
684 			lgroup.cs_cpu = -1;
685 		if (match & CPU_SEARCH_HIGHEST)
686 			hgroup.cs_cpu = -1;
687 		if (child) {			/* Handle child CPU group. */
688 			CPU_NAND(&cpumask, &child->cg_mask);
689 			switch (match) {
690 			case CPU_SEARCH_LOWEST:
691 				load = cpu_search_lowest(child, &lgroup);
692 				break;
693 			case CPU_SEARCH_HIGHEST:
694 				load = cpu_search_highest(child, &hgroup);
695 				break;
696 			case CPU_SEARCH_BOTH:
697 				load = cpu_search_both(child, &lgroup, &hgroup);
698 				break;
699 			}
700 		} else {			/* Handle child CPU. */
701 			CPU_CLR(cpu, &cpumask);
702 			tdq = TDQ_CPU(cpu);
703 			load = tdq->tdq_load * 256;
704 			rndptr = DPCPU_PTR(randomval);
705 			rnd = (*rndptr = *rndptr * 69069 + 5) >> 26;
706 			if (match & CPU_SEARCH_LOWEST) {
707 				if (cpu == low->cs_prefer)
708 					load -= 64;
709 				/* If that CPU is allowed and get data. */
710 				if (tdq->tdq_lowpri > lgroup.cs_pri &&
711 				    tdq->tdq_load <= lgroup.cs_limit &&
712 				    CPU_ISSET(cpu, &lgroup.cs_mask)) {
713 					lgroup.cs_cpu = cpu;
714 					lgroup.cs_load = load - rnd;
715 				}
716 			}
717 			if (match & CPU_SEARCH_HIGHEST)
718 				if (tdq->tdq_load >= hgroup.cs_limit &&
719 				    tdq->tdq_transferable &&
720 				    CPU_ISSET(cpu, &hgroup.cs_mask)) {
721 					hgroup.cs_cpu = cpu;
722 					hgroup.cs_load = load - rnd;
723 				}
724 		}
725 		total += load;
726 
727 		/* We have info about child item. Compare it. */
728 		if (match & CPU_SEARCH_LOWEST) {
729 			if (lgroup.cs_cpu >= 0 &&
730 			    (load < lload ||
731 			     (load == lload && lgroup.cs_load < low->cs_load))) {
732 				lload = load;
733 				low->cs_cpu = lgroup.cs_cpu;
734 				low->cs_load = lgroup.cs_load;
735 			}
736 		}
737 		if (match & CPU_SEARCH_HIGHEST)
738 			if (hgroup.cs_cpu >= 0 &&
739 			    (load > hload ||
740 			     (load == hload && hgroup.cs_load > high->cs_load))) {
741 				hload = load;
742 				high->cs_cpu = hgroup.cs_cpu;
743 				high->cs_load = hgroup.cs_load;
744 			}
745 		if (child) {
746 			i--;
747 			if (i == 0 && CPU_EMPTY(&cpumask))
748 				break;
749 		}
750 #ifndef HAVE_INLINE_FFSL
751 		else
752 			cpu--;
753 #endif
754 	}
755 	return (total);
756 }
757 
758 /*
759  * cpu_search instantiations must pass constants to maintain the inline
760  * optimization.
761  */
762 int
763 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
764 {
765 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
766 }
767 
768 int
769 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
770 {
771 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
772 }
773 
774 int
775 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
776     struct cpu_search *high)
777 {
778 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
779 }
780 
781 /*
782  * Find the cpu with the least load via the least loaded path that has a
783  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
784  * acceptable.
785  */
786 static inline int
787 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
788     int prefer)
789 {
790 	struct cpu_search low;
791 
792 	low.cs_cpu = -1;
793 	low.cs_prefer = prefer;
794 	low.cs_mask = mask;
795 	low.cs_pri = pri;
796 	low.cs_limit = maxload;
797 	cpu_search_lowest(cg, &low);
798 	return low.cs_cpu;
799 }
800 
801 /*
802  * Find the cpu with the highest load via the highest loaded path.
803  */
804 static inline int
805 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
806 {
807 	struct cpu_search high;
808 
809 	high.cs_cpu = -1;
810 	high.cs_mask = mask;
811 	high.cs_limit = minload;
812 	cpu_search_highest(cg, &high);
813 	return high.cs_cpu;
814 }
815 
816 /*
817  * Simultaneously find the highest and lowest loaded cpu reachable via
818  * cg.
819  */
820 static inline void
821 sched_both(const struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu)
822 {
823 	struct cpu_search high;
824 	struct cpu_search low;
825 
826 	low.cs_cpu = -1;
827 	low.cs_prefer = -1;
828 	low.cs_pri = -1;
829 	low.cs_limit = INT_MAX;
830 	low.cs_mask = mask;
831 	high.cs_cpu = -1;
832 	high.cs_limit = -1;
833 	high.cs_mask = mask;
834 	cpu_search_both(cg, &low, &high);
835 	*lowcpu = low.cs_cpu;
836 	*highcpu = high.cs_cpu;
837 	return;
838 }
839 
840 static void
841 sched_balance_group(struct cpu_group *cg)
842 {
843 	cpuset_t hmask, lmask;
844 	int high, low, anylow;
845 
846 	CPU_FILL(&hmask);
847 	for (;;) {
848 		high = sched_highest(cg, hmask, 1);
849 		/* Stop if there is no more CPU with transferrable threads. */
850 		if (high == -1)
851 			break;
852 		CPU_CLR(high, &hmask);
853 		CPU_COPY(&hmask, &lmask);
854 		/* Stop if there is no more CPU left for low. */
855 		if (CPU_EMPTY(&lmask))
856 			break;
857 		anylow = 1;
858 nextlow:
859 		low = sched_lowest(cg, lmask, -1,
860 		    TDQ_CPU(high)->tdq_load - 1, high);
861 		/* Stop if we looked well and found no less loaded CPU. */
862 		if (anylow && low == -1)
863 			break;
864 		/* Go to next high if we found no less loaded CPU. */
865 		if (low == -1)
866 			continue;
867 		/* Transfer thread from high to low. */
868 		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
869 			/* CPU that got thread can no longer be a donor. */
870 			CPU_CLR(low, &hmask);
871 		} else {
872 			/*
873 			 * If failed, then there is no threads on high
874 			 * that can run on this low. Drop low from low
875 			 * mask and look for different one.
876 			 */
877 			CPU_CLR(low, &lmask);
878 			anylow = 0;
879 			goto nextlow;
880 		}
881 	}
882 }
883 
884 static void
885 sched_balance(void)
886 {
887 	struct tdq *tdq;
888 
889 	/*
890 	 * Select a random time between .5 * balance_interval and
891 	 * 1.5 * balance_interval.
892 	 */
893 	balance_ticks = max(balance_interval / 2, 1);
894 	balance_ticks += random() % balance_interval;
895 	if (smp_started == 0 || rebalance == 0)
896 		return;
897 	tdq = TDQ_SELF();
898 	TDQ_UNLOCK(tdq);
899 	sched_balance_group(cpu_top);
900 	TDQ_LOCK(tdq);
901 }
902 
903 /*
904  * Lock two thread queues using their address to maintain lock order.
905  */
906 static void
907 tdq_lock_pair(struct tdq *one, struct tdq *two)
908 {
909 	if (one < two) {
910 		TDQ_LOCK(one);
911 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
912 	} else {
913 		TDQ_LOCK(two);
914 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
915 	}
916 }
917 
918 /*
919  * Unlock two thread queues.  Order is not important here.
920  */
921 static void
922 tdq_unlock_pair(struct tdq *one, struct tdq *two)
923 {
924 	TDQ_UNLOCK(one);
925 	TDQ_UNLOCK(two);
926 }
927 
928 /*
929  * Transfer load between two imbalanced thread queues.
930  */
931 static int
932 sched_balance_pair(struct tdq *high, struct tdq *low)
933 {
934 	int moved;
935 	int cpu;
936 
937 	tdq_lock_pair(high, low);
938 	moved = 0;
939 	/*
940 	 * Determine what the imbalance is and then adjust that to how many
941 	 * threads we actually have to give up (transferable).
942 	 */
943 	if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
944 	    (moved = tdq_move(high, low)) > 0) {
945 		/*
946 		 * In case the target isn't the current cpu IPI it to force a
947 		 * reschedule with the new workload.
948 		 */
949 		cpu = TDQ_ID(low);
950 		if (cpu != PCPU_GET(cpuid))
951 			ipi_cpu(cpu, IPI_PREEMPT);
952 	}
953 	tdq_unlock_pair(high, low);
954 	return (moved);
955 }
956 
957 /*
958  * Move a thread from one thread queue to another.
959  */
960 static int
961 tdq_move(struct tdq *from, struct tdq *to)
962 {
963 	struct td_sched *ts;
964 	struct thread *td;
965 	struct tdq *tdq;
966 	int cpu;
967 
968 	TDQ_LOCK_ASSERT(from, MA_OWNED);
969 	TDQ_LOCK_ASSERT(to, MA_OWNED);
970 
971 	tdq = from;
972 	cpu = TDQ_ID(to);
973 	td = tdq_steal(tdq, cpu);
974 	if (td == NULL)
975 		return (0);
976 	ts = td->td_sched;
977 	/*
978 	 * Although the run queue is locked the thread may be blocked.  Lock
979 	 * it to clear this and acquire the run-queue lock.
980 	 */
981 	thread_lock(td);
982 	/* Drop recursive lock on from acquired via thread_lock(). */
983 	TDQ_UNLOCK(from);
984 	sched_rem(td);
985 	ts->ts_cpu = cpu;
986 	td->td_lock = TDQ_LOCKPTR(to);
987 	tdq_add(to, td, SRQ_YIELDING);
988 	return (1);
989 }
990 
991 /*
992  * This tdq has idled.  Try to steal a thread from another cpu and switch
993  * to it.
994  */
995 static int
996 tdq_idled(struct tdq *tdq)
997 {
998 	struct cpu_group *cg;
999 	struct tdq *steal;
1000 	cpuset_t mask;
1001 	int thresh;
1002 	int cpu;
1003 
1004 	if (smp_started == 0 || steal_idle == 0)
1005 		return (1);
1006 	CPU_FILL(&mask);
1007 	CPU_CLR(PCPU_GET(cpuid), &mask);
1008 	/* We don't want to be preempted while we're iterating. */
1009 	spinlock_enter();
1010 	for (cg = tdq->tdq_cg; cg != NULL; ) {
1011 		if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
1012 			thresh = steal_thresh;
1013 		else
1014 			thresh = 1;
1015 		cpu = sched_highest(cg, mask, thresh);
1016 		if (cpu == -1) {
1017 			cg = cg->cg_parent;
1018 			continue;
1019 		}
1020 		steal = TDQ_CPU(cpu);
1021 		CPU_CLR(cpu, &mask);
1022 		tdq_lock_pair(tdq, steal);
1023 		if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
1024 			tdq_unlock_pair(tdq, steal);
1025 			continue;
1026 		}
1027 		/*
1028 		 * If a thread was added while interrupts were disabled don't
1029 		 * steal one here.  If we fail to acquire one due to affinity
1030 		 * restrictions loop again with this cpu removed from the
1031 		 * set.
1032 		 */
1033 		if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
1034 			tdq_unlock_pair(tdq, steal);
1035 			continue;
1036 		}
1037 		spinlock_exit();
1038 		TDQ_UNLOCK(steal);
1039 		mi_switch(SW_VOL | SWT_IDLE, NULL);
1040 		thread_unlock(curthread);
1041 
1042 		return (0);
1043 	}
1044 	spinlock_exit();
1045 	return (1);
1046 }
1047 
1048 /*
1049  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
1050  */
1051 static void
1052 tdq_notify(struct tdq *tdq, struct thread *td)
1053 {
1054 	struct thread *ctd;
1055 	int pri;
1056 	int cpu;
1057 
1058 	if (tdq->tdq_ipipending)
1059 		return;
1060 	cpu = td->td_sched->ts_cpu;
1061 	pri = td->td_priority;
1062 	ctd = pcpu_find(cpu)->pc_curthread;
1063 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1064 		return;
1065 	if (TD_IS_IDLETHREAD(ctd)) {
1066 		/*
1067 		 * If the MD code has an idle wakeup routine try that before
1068 		 * falling back to IPI.
1069 		 */
1070 		if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1071 			return;
1072 	}
1073 	tdq->tdq_ipipending = 1;
1074 	ipi_cpu(cpu, IPI_PREEMPT);
1075 }
1076 
1077 /*
1078  * Steals load from a timeshare queue.  Honors the rotating queue head
1079  * index.
1080  */
1081 static struct thread *
1082 runq_steal_from(struct runq *rq, int cpu, u_char start)
1083 {
1084 	struct rqbits *rqb;
1085 	struct rqhead *rqh;
1086 	struct thread *td, *first;
1087 	int bit;
1088 	int pri;
1089 	int i;
1090 
1091 	rqb = &rq->rq_status;
1092 	bit = start & (RQB_BPW -1);
1093 	pri = 0;
1094 	first = NULL;
1095 again:
1096 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1097 		if (rqb->rqb_bits[i] == 0)
1098 			continue;
1099 		if (bit != 0) {
1100 			for (pri = bit; pri < RQB_BPW; pri++)
1101 				if (rqb->rqb_bits[i] & (1ul << pri))
1102 					break;
1103 			if (pri >= RQB_BPW)
1104 				continue;
1105 		} else
1106 			pri = RQB_FFS(rqb->rqb_bits[i]);
1107 		pri += (i << RQB_L2BPW);
1108 		rqh = &rq->rq_queues[pri];
1109 		TAILQ_FOREACH(td, rqh, td_runq) {
1110 			if (first && THREAD_CAN_MIGRATE(td) &&
1111 			    THREAD_CAN_SCHED(td, cpu))
1112 				return (td);
1113 			first = td;
1114 		}
1115 	}
1116 	if (start != 0) {
1117 		start = 0;
1118 		goto again;
1119 	}
1120 
1121 	if (first && THREAD_CAN_MIGRATE(first) &&
1122 	    THREAD_CAN_SCHED(first, cpu))
1123 		return (first);
1124 	return (NULL);
1125 }
1126 
1127 /*
1128  * Steals load from a standard linear queue.
1129  */
1130 static struct thread *
1131 runq_steal(struct runq *rq, int cpu)
1132 {
1133 	struct rqhead *rqh;
1134 	struct rqbits *rqb;
1135 	struct thread *td;
1136 	int word;
1137 	int bit;
1138 
1139 	rqb = &rq->rq_status;
1140 	for (word = 0; word < RQB_LEN; word++) {
1141 		if (rqb->rqb_bits[word] == 0)
1142 			continue;
1143 		for (bit = 0; bit < RQB_BPW; bit++) {
1144 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1145 				continue;
1146 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1147 			TAILQ_FOREACH(td, rqh, td_runq)
1148 				if (THREAD_CAN_MIGRATE(td) &&
1149 				    THREAD_CAN_SCHED(td, cpu))
1150 					return (td);
1151 		}
1152 	}
1153 	return (NULL);
1154 }
1155 
1156 /*
1157  * Attempt to steal a thread in priority order from a thread queue.
1158  */
1159 static struct thread *
1160 tdq_steal(struct tdq *tdq, int cpu)
1161 {
1162 	struct thread *td;
1163 
1164 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1165 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1166 		return (td);
1167 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1168 	    cpu, tdq->tdq_ridx)) != NULL)
1169 		return (td);
1170 	return (runq_steal(&tdq->tdq_idle, cpu));
1171 }
1172 
1173 /*
1174  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1175  * current lock and returns with the assigned queue locked.
1176  */
1177 static inline struct tdq *
1178 sched_setcpu(struct thread *td, int cpu, int flags)
1179 {
1180 
1181 	struct tdq *tdq;
1182 
1183 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1184 	tdq = TDQ_CPU(cpu);
1185 	td->td_sched->ts_cpu = cpu;
1186 	/*
1187 	 * If the lock matches just return the queue.
1188 	 */
1189 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1190 		return (tdq);
1191 #ifdef notyet
1192 	/*
1193 	 * If the thread isn't running its lockptr is a
1194 	 * turnstile or a sleepqueue.  We can just lock_set without
1195 	 * blocking.
1196 	 */
1197 	if (TD_CAN_RUN(td)) {
1198 		TDQ_LOCK(tdq);
1199 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1200 		return (tdq);
1201 	}
1202 #endif
1203 	/*
1204 	 * The hard case, migration, we need to block the thread first to
1205 	 * prevent order reversals with other cpus locks.
1206 	 */
1207 	spinlock_enter();
1208 	thread_lock_block(td);
1209 	TDQ_LOCK(tdq);
1210 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1211 	spinlock_exit();
1212 	return (tdq);
1213 }
1214 
1215 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1216 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1217 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1218 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1219 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1220 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1221 
1222 static int
1223 sched_pickcpu(struct thread *td, int flags)
1224 {
1225 	struct cpu_group *cg, *ccg;
1226 	struct td_sched *ts;
1227 	struct tdq *tdq;
1228 	cpuset_t mask;
1229 	int cpu, pri, self;
1230 
1231 	self = PCPU_GET(cpuid);
1232 	ts = td->td_sched;
1233 	if (smp_started == 0)
1234 		return (self);
1235 	/*
1236 	 * Don't migrate a running thread from sched_switch().
1237 	 */
1238 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1239 		return (ts->ts_cpu);
1240 	/*
1241 	 * Prefer to run interrupt threads on the processors that generate
1242 	 * the interrupt.
1243 	 */
1244 	pri = td->td_priority;
1245 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1246 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1247 		SCHED_STAT_INC(pickcpu_intrbind);
1248 		ts->ts_cpu = self;
1249 		if (TDQ_CPU(self)->tdq_lowpri > pri) {
1250 			SCHED_STAT_INC(pickcpu_affinity);
1251 			return (ts->ts_cpu);
1252 		}
1253 	}
1254 	/*
1255 	 * If the thread can run on the last cpu and the affinity has not
1256 	 * expired or it is idle run it there.
1257 	 */
1258 	tdq = TDQ_CPU(ts->ts_cpu);
1259 	cg = tdq->tdq_cg;
1260 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1261 	    tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1262 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1263 		if (cg->cg_flags & CG_FLAG_THREAD) {
1264 			CPUSET_FOREACH(cpu, cg->cg_mask) {
1265 				if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1266 					break;
1267 			}
1268 		} else
1269 			cpu = INT_MAX;
1270 		if (cpu > mp_maxid) {
1271 			SCHED_STAT_INC(pickcpu_idle_affinity);
1272 			return (ts->ts_cpu);
1273 		}
1274 	}
1275 	/*
1276 	 * Search for the last level cache CPU group in the tree.
1277 	 * Skip caches with expired affinity time and SMT groups.
1278 	 * Affinity to higher level caches will be handled less aggressively.
1279 	 */
1280 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1281 		if (cg->cg_flags & CG_FLAG_THREAD)
1282 			continue;
1283 		if (!SCHED_AFFINITY(ts, cg->cg_level))
1284 			continue;
1285 		ccg = cg;
1286 	}
1287 	if (ccg != NULL)
1288 		cg = ccg;
1289 	cpu = -1;
1290 	/* Search the group for the less loaded idle CPU we can run now. */
1291 	mask = td->td_cpuset->cs_mask;
1292 	if (cg != NULL && cg != cpu_top &&
1293 	    CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1294 		cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1295 		    INT_MAX, ts->ts_cpu);
1296 	/* Search globally for the less loaded CPU we can run now. */
1297 	if (cpu == -1)
1298 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1299 	/* Search globally for the less loaded CPU. */
1300 	if (cpu == -1)
1301 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1302 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1303 	/*
1304 	 * Compare the lowest loaded cpu to current cpu.
1305 	 */
1306 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1307 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1308 	    TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1309 		SCHED_STAT_INC(pickcpu_local);
1310 		cpu = self;
1311 	} else
1312 		SCHED_STAT_INC(pickcpu_lowest);
1313 	if (cpu != ts->ts_cpu)
1314 		SCHED_STAT_INC(pickcpu_migration);
1315 	return (cpu);
1316 }
1317 #endif
1318 
1319 /*
1320  * Pick the highest priority task we have and return it.
1321  */
1322 static struct thread *
1323 tdq_choose(struct tdq *tdq)
1324 {
1325 	struct thread *td;
1326 
1327 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1328 	td = runq_choose(&tdq->tdq_realtime);
1329 	if (td != NULL)
1330 		return (td);
1331 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1332 	if (td != NULL) {
1333 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1334 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1335 		    td->td_priority));
1336 		return (td);
1337 	}
1338 	td = runq_choose(&tdq->tdq_idle);
1339 	if (td != NULL) {
1340 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1341 		    ("tdq_choose: Invalid priority on idle queue %d",
1342 		    td->td_priority));
1343 		return (td);
1344 	}
1345 
1346 	return (NULL);
1347 }
1348 
1349 /*
1350  * Initialize a thread queue.
1351  */
1352 static void
1353 tdq_setup(struct tdq *tdq)
1354 {
1355 
1356 	if (bootverbose)
1357 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1358 	runq_init(&tdq->tdq_realtime);
1359 	runq_init(&tdq->tdq_timeshare);
1360 	runq_init(&tdq->tdq_idle);
1361 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1362 	    "sched lock %d", (int)TDQ_ID(tdq));
1363 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1364 	    MTX_SPIN | MTX_RECURSE);
1365 #ifdef KTR
1366 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1367 	    "CPU %d load", (int)TDQ_ID(tdq));
1368 #endif
1369 }
1370 
1371 #ifdef SMP
1372 static void
1373 sched_setup_smp(void)
1374 {
1375 	struct tdq *tdq;
1376 	int i;
1377 
1378 	cpu_top = smp_topo();
1379 	CPU_FOREACH(i) {
1380 		tdq = TDQ_CPU(i);
1381 		tdq_setup(tdq);
1382 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1383 		if (tdq->tdq_cg == NULL)
1384 			panic("Can't find cpu group for %d\n", i);
1385 	}
1386 	balance_tdq = TDQ_SELF();
1387 	sched_balance();
1388 }
1389 #endif
1390 
1391 /*
1392  * Setup the thread queues and initialize the topology based on MD
1393  * information.
1394  */
1395 static void
1396 sched_setup(void *dummy)
1397 {
1398 	struct tdq *tdq;
1399 
1400 	tdq = TDQ_SELF();
1401 #ifdef SMP
1402 	sched_setup_smp();
1403 #else
1404 	tdq_setup(tdq);
1405 #endif
1406 
1407 	/* Add thread0's load since it's running. */
1408 	TDQ_LOCK(tdq);
1409 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1410 	tdq_load_add(tdq, &thread0);
1411 	tdq->tdq_lowpri = thread0.td_priority;
1412 	TDQ_UNLOCK(tdq);
1413 }
1414 
1415 /*
1416  * This routine determines time constants after stathz and hz are setup.
1417  */
1418 /* ARGSUSED */
1419 static void
1420 sched_initticks(void *dummy)
1421 {
1422 	int incr;
1423 
1424 	realstathz = stathz ? stathz : hz;
1425 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1426 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1427 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1428 	    realstathz);
1429 
1430 	/*
1431 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1432 	 * hz not being evenly divisible by stathz on all platforms.
1433 	 */
1434 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1435 	/*
1436 	 * This does not work for values of stathz that are more than
1437 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1438 	 */
1439 	if (incr == 0)
1440 		incr = 1;
1441 	tickincr = incr;
1442 #ifdef SMP
1443 	/*
1444 	 * Set the default balance interval now that we know
1445 	 * what realstathz is.
1446 	 */
1447 	balance_interval = realstathz;
1448 	affinity = SCHED_AFFINITY_DEFAULT;
1449 #endif
1450 	if (sched_idlespinthresh < 0)
1451 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1452 }
1453 
1454 
1455 /*
1456  * This is the core of the interactivity algorithm.  Determines a score based
1457  * on past behavior.  It is the ratio of sleep time to run time scaled to
1458  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1459  * differs from the cpu usage because it does not account for time spent
1460  * waiting on a run-queue.  Would be prettier if we had floating point.
1461  */
1462 static int
1463 sched_interact_score(struct thread *td)
1464 {
1465 	struct td_sched *ts;
1466 	int div;
1467 
1468 	ts = td->td_sched;
1469 	/*
1470 	 * The score is only needed if this is likely to be an interactive
1471 	 * task.  Don't go through the expense of computing it if there's
1472 	 * no chance.
1473 	 */
1474 	if (sched_interact <= SCHED_INTERACT_HALF &&
1475 		ts->ts_runtime >= ts->ts_slptime)
1476 			return (SCHED_INTERACT_HALF);
1477 
1478 	if (ts->ts_runtime > ts->ts_slptime) {
1479 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1480 		return (SCHED_INTERACT_HALF +
1481 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1482 	}
1483 	if (ts->ts_slptime > ts->ts_runtime) {
1484 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1485 		return (ts->ts_runtime / div);
1486 	}
1487 	/* runtime == slptime */
1488 	if (ts->ts_runtime)
1489 		return (SCHED_INTERACT_HALF);
1490 
1491 	/*
1492 	 * This can happen if slptime and runtime are 0.
1493 	 */
1494 	return (0);
1495 
1496 }
1497 
1498 /*
1499  * Scale the scheduling priority according to the "interactivity" of this
1500  * process.
1501  */
1502 static void
1503 sched_priority(struct thread *td)
1504 {
1505 	int score;
1506 	int pri;
1507 
1508 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1509 		return;
1510 	/*
1511 	 * If the score is interactive we place the thread in the realtime
1512 	 * queue with a priority that is less than kernel and interrupt
1513 	 * priorities.  These threads are not subject to nice restrictions.
1514 	 *
1515 	 * Scores greater than this are placed on the normal timeshare queue
1516 	 * where the priority is partially decided by the most recent cpu
1517 	 * utilization and the rest is decided by nice value.
1518 	 *
1519 	 * The nice value of the process has a linear effect on the calculated
1520 	 * score.  Negative nice values make it easier for a thread to be
1521 	 * considered interactive.
1522 	 */
1523 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1524 	if (score < sched_interact) {
1525 		pri = PRI_MIN_INTERACT;
1526 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1527 		    sched_interact) * score;
1528 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1529 		    ("sched_priority: invalid interactive priority %d score %d",
1530 		    pri, score));
1531 	} else {
1532 		pri = SCHED_PRI_MIN;
1533 		if (td->td_sched->ts_ticks)
1534 			pri += min(SCHED_PRI_TICKS(td->td_sched),
1535 			    SCHED_PRI_RANGE);
1536 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1537 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1538 		    ("sched_priority: invalid priority %d: nice %d, "
1539 		    "ticks %d ftick %d ltick %d tick pri %d",
1540 		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1541 		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1542 		    SCHED_PRI_TICKS(td->td_sched)));
1543 	}
1544 	sched_user_prio(td, pri);
1545 
1546 	return;
1547 }
1548 
1549 /*
1550  * This routine enforces a maximum limit on the amount of scheduling history
1551  * kept.  It is called after either the slptime or runtime is adjusted.  This
1552  * function is ugly due to integer math.
1553  */
1554 static void
1555 sched_interact_update(struct thread *td)
1556 {
1557 	struct td_sched *ts;
1558 	u_int sum;
1559 
1560 	ts = td->td_sched;
1561 	sum = ts->ts_runtime + ts->ts_slptime;
1562 	if (sum < SCHED_SLP_RUN_MAX)
1563 		return;
1564 	/*
1565 	 * This only happens from two places:
1566 	 * 1) We have added an unusual amount of run time from fork_exit.
1567 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1568 	 */
1569 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1570 		if (ts->ts_runtime > ts->ts_slptime) {
1571 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1572 			ts->ts_slptime = 1;
1573 		} else {
1574 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1575 			ts->ts_runtime = 1;
1576 		}
1577 		return;
1578 	}
1579 	/*
1580 	 * If we have exceeded by more than 1/5th then the algorithm below
1581 	 * will not bring us back into range.  Dividing by two here forces
1582 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1583 	 */
1584 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1585 		ts->ts_runtime /= 2;
1586 		ts->ts_slptime /= 2;
1587 		return;
1588 	}
1589 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1590 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1591 }
1592 
1593 /*
1594  * Scale back the interactivity history when a child thread is created.  The
1595  * history is inherited from the parent but the thread may behave totally
1596  * differently.  For example, a shell spawning a compiler process.  We want
1597  * to learn that the compiler is behaving badly very quickly.
1598  */
1599 static void
1600 sched_interact_fork(struct thread *td)
1601 {
1602 	int ratio;
1603 	int sum;
1604 
1605 	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1606 	if (sum > SCHED_SLP_RUN_FORK) {
1607 		ratio = sum / SCHED_SLP_RUN_FORK;
1608 		td->td_sched->ts_runtime /= ratio;
1609 		td->td_sched->ts_slptime /= ratio;
1610 	}
1611 }
1612 
1613 /*
1614  * Called from proc0_init() to setup the scheduler fields.
1615  */
1616 void
1617 schedinit(void)
1618 {
1619 
1620 	/*
1621 	 * Set up the scheduler specific parts of proc0.
1622 	 */
1623 	proc0.p_sched = NULL; /* XXX */
1624 	thread0.td_sched = &td_sched0;
1625 	td_sched0.ts_ltick = ticks;
1626 	td_sched0.ts_ftick = ticks;
1627 	td_sched0.ts_slice = 0;
1628 }
1629 
1630 /*
1631  * This is only somewhat accurate since given many processes of the same
1632  * priority they will switch when their slices run out, which will be
1633  * at most sched_slice stathz ticks.
1634  */
1635 int
1636 sched_rr_interval(void)
1637 {
1638 
1639 	/* Convert sched_slice from stathz to hz. */
1640 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1641 }
1642 
1643 /*
1644  * Update the percent cpu tracking information when it is requested or
1645  * the total history exceeds the maximum.  We keep a sliding history of
1646  * tick counts that slowly decays.  This is less precise than the 4BSD
1647  * mechanism since it happens with less regular and frequent events.
1648  */
1649 static void
1650 sched_pctcpu_update(struct td_sched *ts, int run)
1651 {
1652 	int t = ticks;
1653 
1654 	if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1655 		ts->ts_ticks = 0;
1656 		ts->ts_ftick = t - SCHED_TICK_TARG;
1657 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1658 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1659 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1660 		ts->ts_ftick = t - SCHED_TICK_TARG;
1661 	}
1662 	if (run)
1663 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1664 	ts->ts_ltick = t;
1665 }
1666 
1667 /*
1668  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1669  * if necessary.  This is the back-end for several priority related
1670  * functions.
1671  */
1672 static void
1673 sched_thread_priority(struct thread *td, u_char prio)
1674 {
1675 	struct td_sched *ts;
1676 	struct tdq *tdq;
1677 	int oldpri;
1678 
1679 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1680 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1681 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1682 	SDT_PROBE3(sched, , , change_pri, td, td->td_proc, prio);
1683 	if (td != curthread && prio < td->td_priority) {
1684 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1685 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1686 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1687 		SDT_PROBE4(sched, , , lend_pri, td, td->td_proc, prio,
1688 		    curthread);
1689 	}
1690 	ts = td->td_sched;
1691 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1692 	if (td->td_priority == prio)
1693 		return;
1694 	/*
1695 	 * If the priority has been elevated due to priority
1696 	 * propagation, we may have to move ourselves to a new
1697 	 * queue.  This could be optimized to not re-add in some
1698 	 * cases.
1699 	 */
1700 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1701 		sched_rem(td);
1702 		td->td_priority = prio;
1703 		sched_add(td, SRQ_BORROWING);
1704 		return;
1705 	}
1706 	/*
1707 	 * If the thread is currently running we may have to adjust the lowpri
1708 	 * information so other cpus are aware of our current priority.
1709 	 */
1710 	if (TD_IS_RUNNING(td)) {
1711 		tdq = TDQ_CPU(ts->ts_cpu);
1712 		oldpri = td->td_priority;
1713 		td->td_priority = prio;
1714 		if (prio < tdq->tdq_lowpri)
1715 			tdq->tdq_lowpri = prio;
1716 		else if (tdq->tdq_lowpri == oldpri)
1717 			tdq_setlowpri(tdq, td);
1718 		return;
1719 	}
1720 	td->td_priority = prio;
1721 }
1722 
1723 /*
1724  * Update a thread's priority when it is lent another thread's
1725  * priority.
1726  */
1727 void
1728 sched_lend_prio(struct thread *td, u_char prio)
1729 {
1730 
1731 	td->td_flags |= TDF_BORROWING;
1732 	sched_thread_priority(td, prio);
1733 }
1734 
1735 /*
1736  * Restore a thread's priority when priority propagation is
1737  * over.  The prio argument is the minimum priority the thread
1738  * needs to have to satisfy other possible priority lending
1739  * requests.  If the thread's regular priority is less
1740  * important than prio, the thread will keep a priority boost
1741  * of prio.
1742  */
1743 void
1744 sched_unlend_prio(struct thread *td, u_char prio)
1745 {
1746 	u_char base_pri;
1747 
1748 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1749 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1750 		base_pri = td->td_user_pri;
1751 	else
1752 		base_pri = td->td_base_pri;
1753 	if (prio >= base_pri) {
1754 		td->td_flags &= ~TDF_BORROWING;
1755 		sched_thread_priority(td, base_pri);
1756 	} else
1757 		sched_lend_prio(td, prio);
1758 }
1759 
1760 /*
1761  * Standard entry for setting the priority to an absolute value.
1762  */
1763 void
1764 sched_prio(struct thread *td, u_char prio)
1765 {
1766 	u_char oldprio;
1767 
1768 	/* First, update the base priority. */
1769 	td->td_base_pri = prio;
1770 
1771 	/*
1772 	 * If the thread is borrowing another thread's priority, don't
1773 	 * ever lower the priority.
1774 	 */
1775 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1776 		return;
1777 
1778 	/* Change the real priority. */
1779 	oldprio = td->td_priority;
1780 	sched_thread_priority(td, prio);
1781 
1782 	/*
1783 	 * If the thread is on a turnstile, then let the turnstile update
1784 	 * its state.
1785 	 */
1786 	if (TD_ON_LOCK(td) && oldprio != prio)
1787 		turnstile_adjust(td, oldprio);
1788 }
1789 
1790 /*
1791  * Set the base user priority, does not effect current running priority.
1792  */
1793 void
1794 sched_user_prio(struct thread *td, u_char prio)
1795 {
1796 
1797 	td->td_base_user_pri = prio;
1798 	if (td->td_lend_user_pri <= prio)
1799 		return;
1800 	td->td_user_pri = prio;
1801 }
1802 
1803 void
1804 sched_lend_user_prio(struct thread *td, u_char prio)
1805 {
1806 
1807 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1808 	td->td_lend_user_pri = prio;
1809 	td->td_user_pri = min(prio, td->td_base_user_pri);
1810 	if (td->td_priority > td->td_user_pri)
1811 		sched_prio(td, td->td_user_pri);
1812 	else if (td->td_priority != td->td_user_pri)
1813 		td->td_flags |= TDF_NEEDRESCHED;
1814 }
1815 
1816 /*
1817  * Handle migration from sched_switch().  This happens only for
1818  * cpu binding.
1819  */
1820 static struct mtx *
1821 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1822 {
1823 	struct tdq *tdn;
1824 
1825 	tdn = TDQ_CPU(td->td_sched->ts_cpu);
1826 #ifdef SMP
1827 	tdq_load_rem(tdq, td);
1828 	/*
1829 	 * Do the lock dance required to avoid LOR.  We grab an extra
1830 	 * spinlock nesting to prevent preemption while we're
1831 	 * not holding either run-queue lock.
1832 	 */
1833 	spinlock_enter();
1834 	thread_lock_block(td);	/* This releases the lock on tdq. */
1835 
1836 	/*
1837 	 * Acquire both run-queue locks before placing the thread on the new
1838 	 * run-queue to avoid deadlocks created by placing a thread with a
1839 	 * blocked lock on the run-queue of a remote processor.  The deadlock
1840 	 * occurs when a third processor attempts to lock the two queues in
1841 	 * question while the target processor is spinning with its own
1842 	 * run-queue lock held while waiting for the blocked lock to clear.
1843 	 */
1844 	tdq_lock_pair(tdn, tdq);
1845 	tdq_add(tdn, td, flags);
1846 	tdq_notify(tdn, td);
1847 	TDQ_UNLOCK(tdn);
1848 	spinlock_exit();
1849 #endif
1850 	return (TDQ_LOCKPTR(tdn));
1851 }
1852 
1853 /*
1854  * Variadic version of thread_lock_unblock() that does not assume td_lock
1855  * is blocked.
1856  */
1857 static inline void
1858 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1859 {
1860 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1861 	    (uintptr_t)mtx);
1862 }
1863 
1864 /*
1865  * Switch threads.  This function has to handle threads coming in while
1866  * blocked for some reason, running, or idle.  It also must deal with
1867  * migrating a thread from one queue to another as running threads may
1868  * be assigned elsewhere via binding.
1869  */
1870 void
1871 sched_switch(struct thread *td, struct thread *newtd, int flags)
1872 {
1873 	struct tdq *tdq;
1874 	struct td_sched *ts;
1875 	struct mtx *mtx;
1876 	int srqflag;
1877 	int cpuid, preempted;
1878 
1879 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1880 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1881 
1882 	cpuid = PCPU_GET(cpuid);
1883 	tdq = TDQ_CPU(cpuid);
1884 	ts = td->td_sched;
1885 	mtx = td->td_lock;
1886 	sched_pctcpu_update(ts, 1);
1887 	ts->ts_rltick = ticks;
1888 	td->td_lastcpu = td->td_oncpu;
1889 	td->td_oncpu = NOCPU;
1890 	preempted = !(td->td_flags & TDF_SLICEEND);
1891 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1892 	td->td_owepreempt = 0;
1893 	if (!TD_IS_IDLETHREAD(td))
1894 		tdq->tdq_switchcnt++;
1895 	/*
1896 	 * The lock pointer in an idle thread should never change.  Reset it
1897 	 * to CAN_RUN as well.
1898 	 */
1899 	if (TD_IS_IDLETHREAD(td)) {
1900 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1901 		TD_SET_CAN_RUN(td);
1902 	} else if (TD_IS_RUNNING(td)) {
1903 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1904 		srqflag = preempted ?
1905 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1906 		    SRQ_OURSELF|SRQ_YIELDING;
1907 #ifdef SMP
1908 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1909 			ts->ts_cpu = sched_pickcpu(td, 0);
1910 #endif
1911 		if (ts->ts_cpu == cpuid)
1912 			tdq_runq_add(tdq, td, srqflag);
1913 		else {
1914 			KASSERT(THREAD_CAN_MIGRATE(td) ||
1915 			    (ts->ts_flags & TSF_BOUND) != 0,
1916 			    ("Thread %p shouldn't migrate", td));
1917 			mtx = sched_switch_migrate(tdq, td, srqflag);
1918 		}
1919 	} else {
1920 		/* This thread must be going to sleep. */
1921 		TDQ_LOCK(tdq);
1922 		mtx = thread_lock_block(td);
1923 		tdq_load_rem(tdq, td);
1924 	}
1925 	/*
1926 	 * We enter here with the thread blocked and assigned to the
1927 	 * appropriate cpu run-queue or sleep-queue and with the current
1928 	 * thread-queue locked.
1929 	 */
1930 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1931 	newtd = choosethread();
1932 	/*
1933 	 * Call the MD code to switch contexts if necessary.
1934 	 */
1935 	if (td != newtd) {
1936 #ifdef	HWPMC_HOOKS
1937 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1938 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1939 #endif
1940 		SDT_PROBE2(sched, , , off_cpu, newtd, newtd->td_proc);
1941 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1942 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1943 		sched_pctcpu_update(newtd->td_sched, 0);
1944 
1945 #ifdef KDTRACE_HOOKS
1946 		/*
1947 		 * If DTrace has set the active vtime enum to anything
1948 		 * other than INACTIVE (0), then it should have set the
1949 		 * function to call.
1950 		 */
1951 		if (dtrace_vtime_active)
1952 			(*dtrace_vtime_switch_func)(newtd);
1953 #endif
1954 
1955 		cpu_switch(td, newtd, mtx);
1956 		/*
1957 		 * We may return from cpu_switch on a different cpu.  However,
1958 		 * we always return with td_lock pointing to the current cpu's
1959 		 * run queue lock.
1960 		 */
1961 		cpuid = PCPU_GET(cpuid);
1962 		tdq = TDQ_CPU(cpuid);
1963 		lock_profile_obtain_lock_success(
1964 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1965 
1966 		SDT_PROBE0(sched, , , on_cpu);
1967 #ifdef	HWPMC_HOOKS
1968 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1969 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1970 #endif
1971 	} else {
1972 		thread_unblock_switch(td, mtx);
1973 		SDT_PROBE0(sched, , , remain_cpu);
1974 	}
1975 	/*
1976 	 * Assert that all went well and return.
1977 	 */
1978 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1979 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1980 	td->td_oncpu = cpuid;
1981 }
1982 
1983 /*
1984  * Adjust thread priorities as a result of a nice request.
1985  */
1986 void
1987 sched_nice(struct proc *p, int nice)
1988 {
1989 	struct thread *td;
1990 
1991 	PROC_LOCK_ASSERT(p, MA_OWNED);
1992 
1993 	p->p_nice = nice;
1994 	FOREACH_THREAD_IN_PROC(p, td) {
1995 		thread_lock(td);
1996 		sched_priority(td);
1997 		sched_prio(td, td->td_base_user_pri);
1998 		thread_unlock(td);
1999 	}
2000 }
2001 
2002 /*
2003  * Record the sleep time for the interactivity scorer.
2004  */
2005 void
2006 sched_sleep(struct thread *td, int prio)
2007 {
2008 
2009 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2010 
2011 	td->td_slptick = ticks;
2012 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2013 		td->td_flags |= TDF_CANSWAP;
2014 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2015 		return;
2016 	if (static_boost == 1 && prio)
2017 		sched_prio(td, prio);
2018 	else if (static_boost && td->td_priority > static_boost)
2019 		sched_prio(td, static_boost);
2020 }
2021 
2022 /*
2023  * Schedule a thread to resume execution and record how long it voluntarily
2024  * slept.  We also update the pctcpu, interactivity, and priority.
2025  */
2026 void
2027 sched_wakeup(struct thread *td)
2028 {
2029 	struct td_sched *ts;
2030 	int slptick;
2031 
2032 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2033 	ts = td->td_sched;
2034 	td->td_flags &= ~TDF_CANSWAP;
2035 	/*
2036 	 * If we slept for more than a tick update our interactivity and
2037 	 * priority.
2038 	 */
2039 	slptick = td->td_slptick;
2040 	td->td_slptick = 0;
2041 	if (slptick && slptick != ticks) {
2042 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2043 		sched_interact_update(td);
2044 		sched_pctcpu_update(ts, 0);
2045 	}
2046 	/*
2047 	 * Reset the slice value since we slept and advanced the round-robin.
2048 	 */
2049 	ts->ts_slice = 0;
2050 	sched_add(td, SRQ_BORING);
2051 }
2052 
2053 /*
2054  * Penalize the parent for creating a new child and initialize the child's
2055  * priority.
2056  */
2057 void
2058 sched_fork(struct thread *td, struct thread *child)
2059 {
2060 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2061 	sched_pctcpu_update(td->td_sched, 1);
2062 	sched_fork_thread(td, child);
2063 	/*
2064 	 * Penalize the parent and child for forking.
2065 	 */
2066 	sched_interact_fork(child);
2067 	sched_priority(child);
2068 	td->td_sched->ts_runtime += tickincr;
2069 	sched_interact_update(td);
2070 	sched_priority(td);
2071 }
2072 
2073 /*
2074  * Fork a new thread, may be within the same process.
2075  */
2076 void
2077 sched_fork_thread(struct thread *td, struct thread *child)
2078 {
2079 	struct td_sched *ts;
2080 	struct td_sched *ts2;
2081 	struct tdq *tdq;
2082 
2083 	tdq = TDQ_SELF();
2084 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2085 	/*
2086 	 * Initialize child.
2087 	 */
2088 	ts = td->td_sched;
2089 	ts2 = child->td_sched;
2090 	child->td_lock = TDQ_LOCKPTR(tdq);
2091 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2092 	ts2->ts_cpu = ts->ts_cpu;
2093 	ts2->ts_flags = 0;
2094 	/*
2095 	 * Grab our parents cpu estimation information.
2096 	 */
2097 	ts2->ts_ticks = ts->ts_ticks;
2098 	ts2->ts_ltick = ts->ts_ltick;
2099 	ts2->ts_ftick = ts->ts_ftick;
2100 	/*
2101 	 * Do not inherit any borrowed priority from the parent.
2102 	 */
2103 	child->td_priority = child->td_base_pri;
2104 	/*
2105 	 * And update interactivity score.
2106 	 */
2107 	ts2->ts_slptime = ts->ts_slptime;
2108 	ts2->ts_runtime = ts->ts_runtime;
2109 	/* Attempt to quickly learn interactivity. */
2110 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2111 #ifdef KTR
2112 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2113 #endif
2114 }
2115 
2116 /*
2117  * Adjust the priority class of a thread.
2118  */
2119 void
2120 sched_class(struct thread *td, int class)
2121 {
2122 
2123 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2124 	if (td->td_pri_class == class)
2125 		return;
2126 	td->td_pri_class = class;
2127 }
2128 
2129 /*
2130  * Return some of the child's priority and interactivity to the parent.
2131  */
2132 void
2133 sched_exit(struct proc *p, struct thread *child)
2134 {
2135 	struct thread *td;
2136 
2137 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2138 	    "prio:%d", child->td_priority);
2139 	PROC_LOCK_ASSERT(p, MA_OWNED);
2140 	td = FIRST_THREAD_IN_PROC(p);
2141 	sched_exit_thread(td, child);
2142 }
2143 
2144 /*
2145  * Penalize another thread for the time spent on this one.  This helps to
2146  * worsen the priority and interactivity of processes which schedule batch
2147  * jobs such as make.  This has little effect on the make process itself but
2148  * causes new processes spawned by it to receive worse scores immediately.
2149  */
2150 void
2151 sched_exit_thread(struct thread *td, struct thread *child)
2152 {
2153 
2154 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2155 	    "prio:%d", child->td_priority);
2156 	/*
2157 	 * Give the child's runtime to the parent without returning the
2158 	 * sleep time as a penalty to the parent.  This causes shells that
2159 	 * launch expensive things to mark their children as expensive.
2160 	 */
2161 	thread_lock(td);
2162 	td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2163 	sched_interact_update(td);
2164 	sched_priority(td);
2165 	thread_unlock(td);
2166 }
2167 
2168 void
2169 sched_preempt(struct thread *td)
2170 {
2171 	struct tdq *tdq;
2172 
2173 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2174 
2175 	thread_lock(td);
2176 	tdq = TDQ_SELF();
2177 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2178 	tdq->tdq_ipipending = 0;
2179 	if (td->td_priority > tdq->tdq_lowpri) {
2180 		int flags;
2181 
2182 		flags = SW_INVOL | SW_PREEMPT;
2183 		if (td->td_critnest > 1)
2184 			td->td_owepreempt = 1;
2185 		else if (TD_IS_IDLETHREAD(td))
2186 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2187 		else
2188 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2189 	}
2190 	thread_unlock(td);
2191 }
2192 
2193 /*
2194  * Fix priorities on return to user-space.  Priorities may be elevated due
2195  * to static priorities in msleep() or similar.
2196  */
2197 void
2198 sched_userret(struct thread *td)
2199 {
2200 	/*
2201 	 * XXX we cheat slightly on the locking here to avoid locking in
2202 	 * the usual case.  Setting td_priority here is essentially an
2203 	 * incomplete workaround for not setting it properly elsewhere.
2204 	 * Now that some interrupt handlers are threads, not setting it
2205 	 * properly elsewhere can clobber it in the window between setting
2206 	 * it here and returning to user mode, so don't waste time setting
2207 	 * it perfectly here.
2208 	 */
2209 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2210 	    ("thread with borrowed priority returning to userland"));
2211 	if (td->td_priority != td->td_user_pri) {
2212 		thread_lock(td);
2213 		td->td_priority = td->td_user_pri;
2214 		td->td_base_pri = td->td_user_pri;
2215 		tdq_setlowpri(TDQ_SELF(), td);
2216 		thread_unlock(td);
2217         }
2218 }
2219 
2220 /*
2221  * Handle a stathz tick.  This is really only relevant for timeshare
2222  * threads.
2223  */
2224 void
2225 sched_clock(struct thread *td)
2226 {
2227 	struct tdq *tdq;
2228 	struct td_sched *ts;
2229 
2230 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2231 	tdq = TDQ_SELF();
2232 #ifdef SMP
2233 	/*
2234 	 * We run the long term load balancer infrequently on the first cpu.
2235 	 */
2236 	if (balance_tdq == tdq) {
2237 		if (balance_ticks && --balance_ticks == 0)
2238 			sched_balance();
2239 	}
2240 #endif
2241 	/*
2242 	 * Save the old switch count so we have a record of the last ticks
2243 	 * activity.   Initialize the new switch count based on our load.
2244 	 * If there is some activity seed it to reflect that.
2245 	 */
2246 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2247 	tdq->tdq_switchcnt = tdq->tdq_load;
2248 	/*
2249 	 * Advance the insert index once for each tick to ensure that all
2250 	 * threads get a chance to run.
2251 	 */
2252 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2253 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2254 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2255 			tdq->tdq_ridx = tdq->tdq_idx;
2256 	}
2257 	ts = td->td_sched;
2258 	sched_pctcpu_update(ts, 1);
2259 	if (td->td_pri_class & PRI_FIFO_BIT)
2260 		return;
2261 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2262 		/*
2263 		 * We used a tick; charge it to the thread so
2264 		 * that we can compute our interactivity.
2265 		 */
2266 		td->td_sched->ts_runtime += tickincr;
2267 		sched_interact_update(td);
2268 		sched_priority(td);
2269 	}
2270 
2271 	/*
2272 	 * Force a context switch if the current thread has used up a full
2273 	 * time slice (default is 100ms).
2274 	 */
2275 	if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2276 		ts->ts_slice = 0;
2277 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2278 	}
2279 }
2280 
2281 /*
2282  * Called once per hz tick.
2283  */
2284 void
2285 sched_tick(int cnt)
2286 {
2287 
2288 }
2289 
2290 /*
2291  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2292  * cooperative idle threads.
2293  */
2294 int
2295 sched_runnable(void)
2296 {
2297 	struct tdq *tdq;
2298 	int load;
2299 
2300 	load = 1;
2301 
2302 	tdq = TDQ_SELF();
2303 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2304 		if (tdq->tdq_load > 0)
2305 			goto out;
2306 	} else
2307 		if (tdq->tdq_load - 1 > 0)
2308 			goto out;
2309 	load = 0;
2310 out:
2311 	return (load);
2312 }
2313 
2314 /*
2315  * Choose the highest priority thread to run.  The thread is removed from
2316  * the run-queue while running however the load remains.  For SMP we set
2317  * the tdq in the global idle bitmask if it idles here.
2318  */
2319 struct thread *
2320 sched_choose(void)
2321 {
2322 	struct thread *td;
2323 	struct tdq *tdq;
2324 
2325 	tdq = TDQ_SELF();
2326 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2327 	td = tdq_choose(tdq);
2328 	if (td) {
2329 		tdq_runq_rem(tdq, td);
2330 		tdq->tdq_lowpri = td->td_priority;
2331 		return (td);
2332 	}
2333 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2334 	return (PCPU_GET(idlethread));
2335 }
2336 
2337 /*
2338  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2339  * we always request it once we exit a critical section.
2340  */
2341 static inline void
2342 sched_setpreempt(struct thread *td)
2343 {
2344 	struct thread *ctd;
2345 	int cpri;
2346 	int pri;
2347 
2348 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2349 
2350 	ctd = curthread;
2351 	pri = td->td_priority;
2352 	cpri = ctd->td_priority;
2353 	if (pri < cpri)
2354 		ctd->td_flags |= TDF_NEEDRESCHED;
2355 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2356 		return;
2357 	if (!sched_shouldpreempt(pri, cpri, 0))
2358 		return;
2359 	ctd->td_owepreempt = 1;
2360 }
2361 
2362 /*
2363  * Add a thread to a thread queue.  Select the appropriate runq and add the
2364  * thread to it.  This is the internal function called when the tdq is
2365  * predetermined.
2366  */
2367 void
2368 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2369 {
2370 
2371 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2372 	KASSERT((td->td_inhibitors == 0),
2373 	    ("sched_add: trying to run inhibited thread"));
2374 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2375 	    ("sched_add: bad thread state"));
2376 	KASSERT(td->td_flags & TDF_INMEM,
2377 	    ("sched_add: thread swapped out"));
2378 
2379 	if (td->td_priority < tdq->tdq_lowpri)
2380 		tdq->tdq_lowpri = td->td_priority;
2381 	tdq_runq_add(tdq, td, flags);
2382 	tdq_load_add(tdq, td);
2383 }
2384 
2385 /*
2386  * Select the target thread queue and add a thread to it.  Request
2387  * preemption or IPI a remote processor if required.
2388  */
2389 void
2390 sched_add(struct thread *td, int flags)
2391 {
2392 	struct tdq *tdq;
2393 #ifdef SMP
2394 	int cpu;
2395 #endif
2396 
2397 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2398 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2399 	    sched_tdname(curthread));
2400 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2401 	    KTR_ATTR_LINKED, sched_tdname(td));
2402 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2403 	    flags & SRQ_PREEMPTED);
2404 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2405 	/*
2406 	 * Recalculate the priority before we select the target cpu or
2407 	 * run-queue.
2408 	 */
2409 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2410 		sched_priority(td);
2411 #ifdef SMP
2412 	/*
2413 	 * Pick the destination cpu and if it isn't ours transfer to the
2414 	 * target cpu.
2415 	 */
2416 	cpu = sched_pickcpu(td, flags);
2417 	tdq = sched_setcpu(td, cpu, flags);
2418 	tdq_add(tdq, td, flags);
2419 	if (cpu != PCPU_GET(cpuid)) {
2420 		tdq_notify(tdq, td);
2421 		return;
2422 	}
2423 #else
2424 	tdq = TDQ_SELF();
2425 	TDQ_LOCK(tdq);
2426 	/*
2427 	 * Now that the thread is moving to the run-queue, set the lock
2428 	 * to the scheduler's lock.
2429 	 */
2430 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2431 	tdq_add(tdq, td, flags);
2432 #endif
2433 	if (!(flags & SRQ_YIELDING))
2434 		sched_setpreempt(td);
2435 }
2436 
2437 /*
2438  * Remove a thread from a run-queue without running it.  This is used
2439  * when we're stealing a thread from a remote queue.  Otherwise all threads
2440  * exit by calling sched_exit_thread() and sched_throw() themselves.
2441  */
2442 void
2443 sched_rem(struct thread *td)
2444 {
2445 	struct tdq *tdq;
2446 
2447 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2448 	    "prio:%d", td->td_priority);
2449 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2450 	tdq = TDQ_CPU(td->td_sched->ts_cpu);
2451 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2452 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2453 	KASSERT(TD_ON_RUNQ(td),
2454 	    ("sched_rem: thread not on run queue"));
2455 	tdq_runq_rem(tdq, td);
2456 	tdq_load_rem(tdq, td);
2457 	TD_SET_CAN_RUN(td);
2458 	if (td->td_priority == tdq->tdq_lowpri)
2459 		tdq_setlowpri(tdq, NULL);
2460 }
2461 
2462 /*
2463  * Fetch cpu utilization information.  Updates on demand.
2464  */
2465 fixpt_t
2466 sched_pctcpu(struct thread *td)
2467 {
2468 	fixpt_t pctcpu;
2469 	struct td_sched *ts;
2470 
2471 	pctcpu = 0;
2472 	ts = td->td_sched;
2473 	if (ts == NULL)
2474 		return (0);
2475 
2476 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2477 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2478 	if (ts->ts_ticks) {
2479 		int rtick;
2480 
2481 		/* How many rtick per second ? */
2482 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2483 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2484 	}
2485 
2486 	return (pctcpu);
2487 }
2488 
2489 /*
2490  * Enforce affinity settings for a thread.  Called after adjustments to
2491  * cpumask.
2492  */
2493 void
2494 sched_affinity(struct thread *td)
2495 {
2496 #ifdef SMP
2497 	struct td_sched *ts;
2498 
2499 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2500 	ts = td->td_sched;
2501 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2502 		return;
2503 	if (TD_ON_RUNQ(td)) {
2504 		sched_rem(td);
2505 		sched_add(td, SRQ_BORING);
2506 		return;
2507 	}
2508 	if (!TD_IS_RUNNING(td))
2509 		return;
2510 	/*
2511 	 * Force a switch before returning to userspace.  If the
2512 	 * target thread is not running locally send an ipi to force
2513 	 * the issue.
2514 	 */
2515 	td->td_flags |= TDF_NEEDRESCHED;
2516 	if (td != curthread)
2517 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2518 #endif
2519 }
2520 
2521 /*
2522  * Bind a thread to a target cpu.
2523  */
2524 void
2525 sched_bind(struct thread *td, int cpu)
2526 {
2527 	struct td_sched *ts;
2528 
2529 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2530 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2531 	ts = td->td_sched;
2532 	if (ts->ts_flags & TSF_BOUND)
2533 		sched_unbind(td);
2534 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2535 	ts->ts_flags |= TSF_BOUND;
2536 	sched_pin();
2537 	if (PCPU_GET(cpuid) == cpu)
2538 		return;
2539 	ts->ts_cpu = cpu;
2540 	/* When we return from mi_switch we'll be on the correct cpu. */
2541 	mi_switch(SW_VOL, NULL);
2542 }
2543 
2544 /*
2545  * Release a bound thread.
2546  */
2547 void
2548 sched_unbind(struct thread *td)
2549 {
2550 	struct td_sched *ts;
2551 
2552 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2553 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2554 	ts = td->td_sched;
2555 	if ((ts->ts_flags & TSF_BOUND) == 0)
2556 		return;
2557 	ts->ts_flags &= ~TSF_BOUND;
2558 	sched_unpin();
2559 }
2560 
2561 int
2562 sched_is_bound(struct thread *td)
2563 {
2564 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2565 	return (td->td_sched->ts_flags & TSF_BOUND);
2566 }
2567 
2568 /*
2569  * Basic yield call.
2570  */
2571 void
2572 sched_relinquish(struct thread *td)
2573 {
2574 	thread_lock(td);
2575 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2576 	thread_unlock(td);
2577 }
2578 
2579 /*
2580  * Return the total system load.
2581  */
2582 int
2583 sched_load(void)
2584 {
2585 #ifdef SMP
2586 	int total;
2587 	int i;
2588 
2589 	total = 0;
2590 	CPU_FOREACH(i)
2591 		total += TDQ_CPU(i)->tdq_sysload;
2592 	return (total);
2593 #else
2594 	return (TDQ_SELF()->tdq_sysload);
2595 #endif
2596 }
2597 
2598 int
2599 sched_sizeof_proc(void)
2600 {
2601 	return (sizeof(struct proc));
2602 }
2603 
2604 int
2605 sched_sizeof_thread(void)
2606 {
2607 	return (sizeof(struct thread) + sizeof(struct td_sched));
2608 }
2609 
2610 #ifdef SMP
2611 #define	TDQ_IDLESPIN(tdq)						\
2612     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2613 #else
2614 #define	TDQ_IDLESPIN(tdq)	1
2615 #endif
2616 
2617 /*
2618  * The actual idle process.
2619  */
2620 void
2621 sched_idletd(void *dummy)
2622 {
2623 	struct thread *td;
2624 	struct tdq *tdq;
2625 	int oldswitchcnt, switchcnt;
2626 	int i;
2627 
2628 	mtx_assert(&Giant, MA_NOTOWNED);
2629 	td = curthread;
2630 	tdq = TDQ_SELF();
2631 	THREAD_NO_SLEEPING();
2632 	oldswitchcnt = -1;
2633 	for (;;) {
2634 		if (tdq->tdq_load) {
2635 			thread_lock(td);
2636 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2637 			thread_unlock(td);
2638 		}
2639 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2640 #ifdef SMP
2641 		if (switchcnt != oldswitchcnt) {
2642 			oldswitchcnt = switchcnt;
2643 			if (tdq_idled(tdq) == 0)
2644 				continue;
2645 		}
2646 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2647 #else
2648 		oldswitchcnt = switchcnt;
2649 #endif
2650 		/*
2651 		 * If we're switching very frequently, spin while checking
2652 		 * for load rather than entering a low power state that
2653 		 * may require an IPI.  However, don't do any busy
2654 		 * loops while on SMT machines as this simply steals
2655 		 * cycles from cores doing useful work.
2656 		 */
2657 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2658 			for (i = 0; i < sched_idlespins; i++) {
2659 				if (tdq->tdq_load)
2660 					break;
2661 				cpu_spinwait();
2662 			}
2663 		}
2664 
2665 		/* If there was context switch during spin, restart it. */
2666 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2667 		if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2668 			continue;
2669 
2670 		/* Run main MD idle handler. */
2671 		tdq->tdq_cpu_idle = 1;
2672 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2673 		tdq->tdq_cpu_idle = 0;
2674 
2675 		/*
2676 		 * Account thread-less hardware interrupts and
2677 		 * other wakeup reasons equal to context switches.
2678 		 */
2679 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2680 		if (switchcnt != oldswitchcnt)
2681 			continue;
2682 		tdq->tdq_switchcnt++;
2683 		oldswitchcnt++;
2684 	}
2685 }
2686 
2687 /*
2688  * A CPU is entering for the first time or a thread is exiting.
2689  */
2690 void
2691 sched_throw(struct thread *td)
2692 {
2693 	struct thread *newtd;
2694 	struct tdq *tdq;
2695 
2696 	tdq = TDQ_SELF();
2697 	if (td == NULL) {
2698 		/* Correct spinlock nesting and acquire the correct lock. */
2699 		TDQ_LOCK(tdq);
2700 		spinlock_exit();
2701 		PCPU_SET(switchtime, cpu_ticks());
2702 		PCPU_SET(switchticks, ticks);
2703 	} else {
2704 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2705 		tdq_load_rem(tdq, td);
2706 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2707 	}
2708 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2709 	newtd = choosethread();
2710 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2711 	cpu_throw(td, newtd);		/* doesn't return */
2712 }
2713 
2714 /*
2715  * This is called from fork_exit().  Just acquire the correct locks and
2716  * let fork do the rest of the work.
2717  */
2718 void
2719 sched_fork_exit(struct thread *td)
2720 {
2721 	struct td_sched *ts;
2722 	struct tdq *tdq;
2723 	int cpuid;
2724 
2725 	/*
2726 	 * Finish setting up thread glue so that it begins execution in a
2727 	 * non-nested critical section with the scheduler lock held.
2728 	 */
2729 	cpuid = PCPU_GET(cpuid);
2730 	tdq = TDQ_CPU(cpuid);
2731 	ts = td->td_sched;
2732 	if (TD_IS_IDLETHREAD(td))
2733 		td->td_lock = TDQ_LOCKPTR(tdq);
2734 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2735 	td->td_oncpu = cpuid;
2736 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2737 	lock_profile_obtain_lock_success(
2738 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2739 }
2740 
2741 /*
2742  * Create on first use to catch odd startup conditons.
2743  */
2744 char *
2745 sched_tdname(struct thread *td)
2746 {
2747 #ifdef KTR
2748 	struct td_sched *ts;
2749 
2750 	ts = td->td_sched;
2751 	if (ts->ts_name[0] == '\0')
2752 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2753 		    "%s tid %d", td->td_name, td->td_tid);
2754 	return (ts->ts_name);
2755 #else
2756 	return (td->td_name);
2757 #endif
2758 }
2759 
2760 #ifdef KTR
2761 void
2762 sched_clear_tdname(struct thread *td)
2763 {
2764 	struct td_sched *ts;
2765 
2766 	ts = td->td_sched;
2767 	ts->ts_name[0] = '\0';
2768 }
2769 #endif
2770 
2771 #ifdef SMP
2772 
2773 /*
2774  * Build the CPU topology dump string. Is recursively called to collect
2775  * the topology tree.
2776  */
2777 static int
2778 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2779     int indent)
2780 {
2781 	char cpusetbuf[CPUSETBUFSIZ];
2782 	int i, first;
2783 
2784 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2785 	    "", 1 + indent / 2, cg->cg_level);
2786 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2787 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2788 	first = TRUE;
2789 	for (i = 0; i < MAXCPU; i++) {
2790 		if (CPU_ISSET(i, &cg->cg_mask)) {
2791 			if (!first)
2792 				sbuf_printf(sb, ", ");
2793 			else
2794 				first = FALSE;
2795 			sbuf_printf(sb, "%d", i);
2796 		}
2797 	}
2798 	sbuf_printf(sb, "</cpu>\n");
2799 
2800 	if (cg->cg_flags != 0) {
2801 		sbuf_printf(sb, "%*s <flags>", indent, "");
2802 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2803 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2804 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2805 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2806 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2807 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2808 		sbuf_printf(sb, "</flags>\n");
2809 	}
2810 
2811 	if (cg->cg_children > 0) {
2812 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2813 		for (i = 0; i < cg->cg_children; i++)
2814 			sysctl_kern_sched_topology_spec_internal(sb,
2815 			    &cg->cg_child[i], indent+2);
2816 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2817 	}
2818 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2819 	return (0);
2820 }
2821 
2822 /*
2823  * Sysctl handler for retrieving topology dump. It's a wrapper for
2824  * the recursive sysctl_kern_smp_topology_spec_internal().
2825  */
2826 static int
2827 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2828 {
2829 	struct sbuf *topo;
2830 	int err;
2831 
2832 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2833 
2834 	topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2835 	if (topo == NULL)
2836 		return (ENOMEM);
2837 
2838 	sbuf_printf(topo, "<groups>\n");
2839 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2840 	sbuf_printf(topo, "</groups>\n");
2841 
2842 	if (err == 0) {
2843 		sbuf_finish(topo);
2844 		err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2845 	}
2846 	sbuf_delete(topo);
2847 	return (err);
2848 }
2849 
2850 #endif
2851 
2852 static int
2853 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2854 {
2855 	int error, new_val, period;
2856 
2857 	period = 1000000 / realstathz;
2858 	new_val = period * sched_slice;
2859 	error = sysctl_handle_int(oidp, &new_val, 0, req);
2860 	if (error != 0 || req->newptr == NULL)
2861 		return (error);
2862 	if (new_val <= 0)
2863 		return (EINVAL);
2864 	sched_slice = imax(1, (new_val + period / 2) / period);
2865 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
2866 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2867 	    realstathz);
2868 	return (0);
2869 }
2870 
2871 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2872 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2873     "Scheduler name");
2874 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2875     NULL, 0, sysctl_kern_quantum, "I",
2876     "Quantum for timeshare threads in microseconds");
2877 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2878     "Quantum for timeshare threads in stathz ticks");
2879 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2880     "Interactivity score threshold");
2881 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2882     &preempt_thresh, 0,
2883     "Maximal (lowest) priority for preemption");
2884 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2885     "Assign static kernel priorities to sleeping threads");
2886 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2887     "Number of times idle thread will spin waiting for new work");
2888 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2889     &sched_idlespinthresh, 0,
2890     "Threshold before we will permit idle thread spinning");
2891 #ifdef SMP
2892 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2893     "Number of hz ticks to keep thread affinity for");
2894 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2895     "Enables the long-term load balancer");
2896 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2897     &balance_interval, 0,
2898     "Average period in stathz ticks to run the long-term balancer");
2899 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2900     "Attempts to steal work from other cores before idling");
2901 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2902     "Minimum load on remote CPU before we'll steal");
2903 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2904     CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2905     "XML dump of detected CPU topology");
2906 #endif
2907 
2908 /* ps compat.  All cpu percentages from ULE are weighted. */
2909 static int ccpu = 0;
2910 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2911