xref: /freebsd/sys/kern/sched_ule.c (revision 4ec234c813eed05c166859bba82c882e40826eb9)
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, "struct thread *",
347     "struct proc *", "uint8_t");
348 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
349     "struct proc *", "void *");
350 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
351     "struct proc *", "void *", "int");
352 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
353     "struct proc *", "uint8_t", "struct thread *");
354 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
355 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
356     "struct proc *");
357 SDT_PROBE_DEFINE(sched, , , on__cpu);
358 SDT_PROBE_DEFINE(sched, , , remain__cpu);
359 SDT_PROBE_DEFINE2(sched, , , 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 static void
817 sched_balance_group(struct cpu_group *cg)
818 {
819 	cpuset_t hmask, lmask;
820 	int high, low, anylow;
821 
822 	CPU_FILL(&hmask);
823 	for (;;) {
824 		high = sched_highest(cg, hmask, 1);
825 		/* Stop if there is no more CPU with transferrable threads. */
826 		if (high == -1)
827 			break;
828 		CPU_CLR(high, &hmask);
829 		CPU_COPY(&hmask, &lmask);
830 		/* Stop if there is no more CPU left for low. */
831 		if (CPU_EMPTY(&lmask))
832 			break;
833 		anylow = 1;
834 nextlow:
835 		low = sched_lowest(cg, lmask, -1,
836 		    TDQ_CPU(high)->tdq_load - 1, high);
837 		/* Stop if we looked well and found no less loaded CPU. */
838 		if (anylow && low == -1)
839 			break;
840 		/* Go to next high if we found no less loaded CPU. */
841 		if (low == -1)
842 			continue;
843 		/* Transfer thread from high to low. */
844 		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
845 			/* CPU that got thread can no longer be a donor. */
846 			CPU_CLR(low, &hmask);
847 		} else {
848 			/*
849 			 * If failed, then there is no threads on high
850 			 * that can run on this low. Drop low from low
851 			 * mask and look for different one.
852 			 */
853 			CPU_CLR(low, &lmask);
854 			anylow = 0;
855 			goto nextlow;
856 		}
857 	}
858 }
859 
860 static void
861 sched_balance(void)
862 {
863 	struct tdq *tdq;
864 
865 	/*
866 	 * Select a random time between .5 * balance_interval and
867 	 * 1.5 * balance_interval.
868 	 */
869 	balance_ticks = max(balance_interval / 2, 1);
870 	balance_ticks += random() % balance_interval;
871 	if (smp_started == 0 || rebalance == 0)
872 		return;
873 	tdq = TDQ_SELF();
874 	TDQ_UNLOCK(tdq);
875 	sched_balance_group(cpu_top);
876 	TDQ_LOCK(tdq);
877 }
878 
879 /*
880  * Lock two thread queues using their address to maintain lock order.
881  */
882 static void
883 tdq_lock_pair(struct tdq *one, struct tdq *two)
884 {
885 	if (one < two) {
886 		TDQ_LOCK(one);
887 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
888 	} else {
889 		TDQ_LOCK(two);
890 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
891 	}
892 }
893 
894 /*
895  * Unlock two thread queues.  Order is not important here.
896  */
897 static void
898 tdq_unlock_pair(struct tdq *one, struct tdq *two)
899 {
900 	TDQ_UNLOCK(one);
901 	TDQ_UNLOCK(two);
902 }
903 
904 /*
905  * Transfer load between two imbalanced thread queues.
906  */
907 static int
908 sched_balance_pair(struct tdq *high, struct tdq *low)
909 {
910 	int moved;
911 	int cpu;
912 
913 	tdq_lock_pair(high, low);
914 	moved = 0;
915 	/*
916 	 * Determine what the imbalance is and then adjust that to how many
917 	 * threads we actually have to give up (transferable).
918 	 */
919 	if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
920 	    (moved = tdq_move(high, low)) > 0) {
921 		/*
922 		 * In case the target isn't the current cpu IPI it to force a
923 		 * reschedule with the new workload.
924 		 */
925 		cpu = TDQ_ID(low);
926 		if (cpu != PCPU_GET(cpuid))
927 			ipi_cpu(cpu, IPI_PREEMPT);
928 	}
929 	tdq_unlock_pair(high, low);
930 	return (moved);
931 }
932 
933 /*
934  * Move a thread from one thread queue to another.
935  */
936 static int
937 tdq_move(struct tdq *from, struct tdq *to)
938 {
939 	struct td_sched *ts;
940 	struct thread *td;
941 	struct tdq *tdq;
942 	int cpu;
943 
944 	TDQ_LOCK_ASSERT(from, MA_OWNED);
945 	TDQ_LOCK_ASSERT(to, MA_OWNED);
946 
947 	tdq = from;
948 	cpu = TDQ_ID(to);
949 	td = tdq_steal(tdq, cpu);
950 	if (td == NULL)
951 		return (0);
952 	ts = td->td_sched;
953 	/*
954 	 * Although the run queue is locked the thread may be blocked.  Lock
955 	 * it to clear this and acquire the run-queue lock.
956 	 */
957 	thread_lock(td);
958 	/* Drop recursive lock on from acquired via thread_lock(). */
959 	TDQ_UNLOCK(from);
960 	sched_rem(td);
961 	ts->ts_cpu = cpu;
962 	td->td_lock = TDQ_LOCKPTR(to);
963 	tdq_add(to, td, SRQ_YIELDING);
964 	return (1);
965 }
966 
967 /*
968  * This tdq has idled.  Try to steal a thread from another cpu and switch
969  * to it.
970  */
971 static int
972 tdq_idled(struct tdq *tdq)
973 {
974 	struct cpu_group *cg;
975 	struct tdq *steal;
976 	cpuset_t mask;
977 	int thresh;
978 	int cpu;
979 
980 	if (smp_started == 0 || steal_idle == 0)
981 		return (1);
982 	CPU_FILL(&mask);
983 	CPU_CLR(PCPU_GET(cpuid), &mask);
984 	/* We don't want to be preempted while we're iterating. */
985 	spinlock_enter();
986 	for (cg = tdq->tdq_cg; cg != NULL; ) {
987 		if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
988 			thresh = steal_thresh;
989 		else
990 			thresh = 1;
991 		cpu = sched_highest(cg, mask, thresh);
992 		if (cpu == -1) {
993 			cg = cg->cg_parent;
994 			continue;
995 		}
996 		steal = TDQ_CPU(cpu);
997 		CPU_CLR(cpu, &mask);
998 		tdq_lock_pair(tdq, steal);
999 		if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
1000 			tdq_unlock_pair(tdq, steal);
1001 			continue;
1002 		}
1003 		/*
1004 		 * If a thread was added while interrupts were disabled don't
1005 		 * steal one here.  If we fail to acquire one due to affinity
1006 		 * restrictions loop again with this cpu removed from the
1007 		 * set.
1008 		 */
1009 		if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
1010 			tdq_unlock_pair(tdq, steal);
1011 			continue;
1012 		}
1013 		spinlock_exit();
1014 		TDQ_UNLOCK(steal);
1015 		mi_switch(SW_VOL | SWT_IDLE, NULL);
1016 		thread_unlock(curthread);
1017 
1018 		return (0);
1019 	}
1020 	spinlock_exit();
1021 	return (1);
1022 }
1023 
1024 /*
1025  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
1026  */
1027 static void
1028 tdq_notify(struct tdq *tdq, struct thread *td)
1029 {
1030 	struct thread *ctd;
1031 	int pri;
1032 	int cpu;
1033 
1034 	if (tdq->tdq_ipipending)
1035 		return;
1036 	cpu = td->td_sched->ts_cpu;
1037 	pri = td->td_priority;
1038 	ctd = pcpu_find(cpu)->pc_curthread;
1039 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1040 		return;
1041 	if (TD_IS_IDLETHREAD(ctd)) {
1042 		/*
1043 		 * If the MD code has an idle wakeup routine try that before
1044 		 * falling back to IPI.
1045 		 */
1046 		if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1047 			return;
1048 	}
1049 	tdq->tdq_ipipending = 1;
1050 	ipi_cpu(cpu, IPI_PREEMPT);
1051 }
1052 
1053 /*
1054  * Steals load from a timeshare queue.  Honors the rotating queue head
1055  * index.
1056  */
1057 static struct thread *
1058 runq_steal_from(struct runq *rq, int cpu, u_char start)
1059 {
1060 	struct rqbits *rqb;
1061 	struct rqhead *rqh;
1062 	struct thread *td, *first;
1063 	int bit;
1064 	int pri;
1065 	int i;
1066 
1067 	rqb = &rq->rq_status;
1068 	bit = start & (RQB_BPW -1);
1069 	pri = 0;
1070 	first = NULL;
1071 again:
1072 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1073 		if (rqb->rqb_bits[i] == 0)
1074 			continue;
1075 		if (bit != 0) {
1076 			for (pri = bit; pri < RQB_BPW; pri++)
1077 				if (rqb->rqb_bits[i] & (1ul << pri))
1078 					break;
1079 			if (pri >= RQB_BPW)
1080 				continue;
1081 		} else
1082 			pri = RQB_FFS(rqb->rqb_bits[i]);
1083 		pri += (i << RQB_L2BPW);
1084 		rqh = &rq->rq_queues[pri];
1085 		TAILQ_FOREACH(td, rqh, td_runq) {
1086 			if (first && THREAD_CAN_MIGRATE(td) &&
1087 			    THREAD_CAN_SCHED(td, cpu))
1088 				return (td);
1089 			first = td;
1090 		}
1091 	}
1092 	if (start != 0) {
1093 		start = 0;
1094 		goto again;
1095 	}
1096 
1097 	if (first && THREAD_CAN_MIGRATE(first) &&
1098 	    THREAD_CAN_SCHED(first, cpu))
1099 		return (first);
1100 	return (NULL);
1101 }
1102 
1103 /*
1104  * Steals load from a standard linear queue.
1105  */
1106 static struct thread *
1107 runq_steal(struct runq *rq, int cpu)
1108 {
1109 	struct rqhead *rqh;
1110 	struct rqbits *rqb;
1111 	struct thread *td;
1112 	int word;
1113 	int bit;
1114 
1115 	rqb = &rq->rq_status;
1116 	for (word = 0; word < RQB_LEN; word++) {
1117 		if (rqb->rqb_bits[word] == 0)
1118 			continue;
1119 		for (bit = 0; bit < RQB_BPW; bit++) {
1120 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1121 				continue;
1122 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1123 			TAILQ_FOREACH(td, rqh, td_runq)
1124 				if (THREAD_CAN_MIGRATE(td) &&
1125 				    THREAD_CAN_SCHED(td, cpu))
1126 					return (td);
1127 		}
1128 	}
1129 	return (NULL);
1130 }
1131 
1132 /*
1133  * Attempt to steal a thread in priority order from a thread queue.
1134  */
1135 static struct thread *
1136 tdq_steal(struct tdq *tdq, int cpu)
1137 {
1138 	struct thread *td;
1139 
1140 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1141 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1142 		return (td);
1143 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1144 	    cpu, tdq->tdq_ridx)) != NULL)
1145 		return (td);
1146 	return (runq_steal(&tdq->tdq_idle, cpu));
1147 }
1148 
1149 /*
1150  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1151  * current lock and returns with the assigned queue locked.
1152  */
1153 static inline struct tdq *
1154 sched_setcpu(struct thread *td, int cpu, int flags)
1155 {
1156 
1157 	struct tdq *tdq;
1158 
1159 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1160 	tdq = TDQ_CPU(cpu);
1161 	td->td_sched->ts_cpu = cpu;
1162 	/*
1163 	 * If the lock matches just return the queue.
1164 	 */
1165 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1166 		return (tdq);
1167 #ifdef notyet
1168 	/*
1169 	 * If the thread isn't running its lockptr is a
1170 	 * turnstile or a sleepqueue.  We can just lock_set without
1171 	 * blocking.
1172 	 */
1173 	if (TD_CAN_RUN(td)) {
1174 		TDQ_LOCK(tdq);
1175 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1176 		return (tdq);
1177 	}
1178 #endif
1179 	/*
1180 	 * The hard case, migration, we need to block the thread first to
1181 	 * prevent order reversals with other cpus locks.
1182 	 */
1183 	spinlock_enter();
1184 	thread_lock_block(td);
1185 	TDQ_LOCK(tdq);
1186 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1187 	spinlock_exit();
1188 	return (tdq);
1189 }
1190 
1191 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1192 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1193 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1194 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1195 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1196 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1197 
1198 static int
1199 sched_pickcpu(struct thread *td, int flags)
1200 {
1201 	struct cpu_group *cg, *ccg;
1202 	struct td_sched *ts;
1203 	struct tdq *tdq;
1204 	cpuset_t mask;
1205 	int cpu, pri, self;
1206 
1207 	self = PCPU_GET(cpuid);
1208 	ts = td->td_sched;
1209 	if (smp_started == 0)
1210 		return (self);
1211 	/*
1212 	 * Don't migrate a running thread from sched_switch().
1213 	 */
1214 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1215 		return (ts->ts_cpu);
1216 	/*
1217 	 * Prefer to run interrupt threads on the processors that generate
1218 	 * the interrupt.
1219 	 */
1220 	pri = td->td_priority;
1221 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1222 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1223 		SCHED_STAT_INC(pickcpu_intrbind);
1224 		ts->ts_cpu = self;
1225 		if (TDQ_CPU(self)->tdq_lowpri > pri) {
1226 			SCHED_STAT_INC(pickcpu_affinity);
1227 			return (ts->ts_cpu);
1228 		}
1229 	}
1230 	/*
1231 	 * If the thread can run on the last cpu and the affinity has not
1232 	 * expired or it is idle run it there.
1233 	 */
1234 	tdq = TDQ_CPU(ts->ts_cpu);
1235 	cg = tdq->tdq_cg;
1236 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1237 	    tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1238 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1239 		if (cg->cg_flags & CG_FLAG_THREAD) {
1240 			CPUSET_FOREACH(cpu, cg->cg_mask) {
1241 				if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1242 					break;
1243 			}
1244 		} else
1245 			cpu = INT_MAX;
1246 		if (cpu > mp_maxid) {
1247 			SCHED_STAT_INC(pickcpu_idle_affinity);
1248 			return (ts->ts_cpu);
1249 		}
1250 	}
1251 	/*
1252 	 * Search for the last level cache CPU group in the tree.
1253 	 * Skip caches with expired affinity time and SMT groups.
1254 	 * Affinity to higher level caches will be handled less aggressively.
1255 	 */
1256 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1257 		if (cg->cg_flags & CG_FLAG_THREAD)
1258 			continue;
1259 		if (!SCHED_AFFINITY(ts, cg->cg_level))
1260 			continue;
1261 		ccg = cg;
1262 	}
1263 	if (ccg != NULL)
1264 		cg = ccg;
1265 	cpu = -1;
1266 	/* Search the group for the less loaded idle CPU we can run now. */
1267 	mask = td->td_cpuset->cs_mask;
1268 	if (cg != NULL && cg != cpu_top &&
1269 	    CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1270 		cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1271 		    INT_MAX, ts->ts_cpu);
1272 	/* Search globally for the less loaded CPU we can run now. */
1273 	if (cpu == -1)
1274 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1275 	/* Search globally for the less loaded CPU. */
1276 	if (cpu == -1)
1277 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1278 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1279 	/*
1280 	 * Compare the lowest loaded cpu to current cpu.
1281 	 */
1282 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1283 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1284 	    TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1285 		SCHED_STAT_INC(pickcpu_local);
1286 		cpu = self;
1287 	} else
1288 		SCHED_STAT_INC(pickcpu_lowest);
1289 	if (cpu != ts->ts_cpu)
1290 		SCHED_STAT_INC(pickcpu_migration);
1291 	return (cpu);
1292 }
1293 #endif
1294 
1295 /*
1296  * Pick the highest priority task we have and return it.
1297  */
1298 static struct thread *
1299 tdq_choose(struct tdq *tdq)
1300 {
1301 	struct thread *td;
1302 
1303 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1304 	td = runq_choose(&tdq->tdq_realtime);
1305 	if (td != NULL)
1306 		return (td);
1307 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1308 	if (td != NULL) {
1309 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1310 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1311 		    td->td_priority));
1312 		return (td);
1313 	}
1314 	td = runq_choose(&tdq->tdq_idle);
1315 	if (td != NULL) {
1316 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1317 		    ("tdq_choose: Invalid priority on idle queue %d",
1318 		    td->td_priority));
1319 		return (td);
1320 	}
1321 
1322 	return (NULL);
1323 }
1324 
1325 /*
1326  * Initialize a thread queue.
1327  */
1328 static void
1329 tdq_setup(struct tdq *tdq)
1330 {
1331 
1332 	if (bootverbose)
1333 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1334 	runq_init(&tdq->tdq_realtime);
1335 	runq_init(&tdq->tdq_timeshare);
1336 	runq_init(&tdq->tdq_idle);
1337 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1338 	    "sched lock %d", (int)TDQ_ID(tdq));
1339 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1340 	    MTX_SPIN | MTX_RECURSE);
1341 #ifdef KTR
1342 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1343 	    "CPU %d load", (int)TDQ_ID(tdq));
1344 #endif
1345 }
1346 
1347 #ifdef SMP
1348 static void
1349 sched_setup_smp(void)
1350 {
1351 	struct tdq *tdq;
1352 	int i;
1353 
1354 	cpu_top = smp_topo();
1355 	CPU_FOREACH(i) {
1356 		tdq = TDQ_CPU(i);
1357 		tdq_setup(tdq);
1358 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1359 		if (tdq->tdq_cg == NULL)
1360 			panic("Can't find cpu group for %d\n", i);
1361 	}
1362 	balance_tdq = TDQ_SELF();
1363 	sched_balance();
1364 }
1365 #endif
1366 
1367 /*
1368  * Setup the thread queues and initialize the topology based on MD
1369  * information.
1370  */
1371 static void
1372 sched_setup(void *dummy)
1373 {
1374 	struct tdq *tdq;
1375 
1376 	tdq = TDQ_SELF();
1377 #ifdef SMP
1378 	sched_setup_smp();
1379 #else
1380 	tdq_setup(tdq);
1381 #endif
1382 
1383 	/* Add thread0's load since it's running. */
1384 	TDQ_LOCK(tdq);
1385 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1386 	tdq_load_add(tdq, &thread0);
1387 	tdq->tdq_lowpri = thread0.td_priority;
1388 	TDQ_UNLOCK(tdq);
1389 }
1390 
1391 /*
1392  * This routine determines time constants after stathz and hz are setup.
1393  */
1394 /* ARGSUSED */
1395 static void
1396 sched_initticks(void *dummy)
1397 {
1398 	int incr;
1399 
1400 	realstathz = stathz ? stathz : hz;
1401 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1402 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1403 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1404 	    realstathz);
1405 
1406 	/*
1407 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1408 	 * hz not being evenly divisible by stathz on all platforms.
1409 	 */
1410 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1411 	/*
1412 	 * This does not work for values of stathz that are more than
1413 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1414 	 */
1415 	if (incr == 0)
1416 		incr = 1;
1417 	tickincr = incr;
1418 #ifdef SMP
1419 	/*
1420 	 * Set the default balance interval now that we know
1421 	 * what realstathz is.
1422 	 */
1423 	balance_interval = realstathz;
1424 	affinity = SCHED_AFFINITY_DEFAULT;
1425 #endif
1426 	if (sched_idlespinthresh < 0)
1427 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1428 }
1429 
1430 
1431 /*
1432  * This is the core of the interactivity algorithm.  Determines a score based
1433  * on past behavior.  It is the ratio of sleep time to run time scaled to
1434  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1435  * differs from the cpu usage because it does not account for time spent
1436  * waiting on a run-queue.  Would be prettier if we had floating point.
1437  */
1438 static int
1439 sched_interact_score(struct thread *td)
1440 {
1441 	struct td_sched *ts;
1442 	int div;
1443 
1444 	ts = td->td_sched;
1445 	/*
1446 	 * The score is only needed if this is likely to be an interactive
1447 	 * task.  Don't go through the expense of computing it if there's
1448 	 * no chance.
1449 	 */
1450 	if (sched_interact <= SCHED_INTERACT_HALF &&
1451 		ts->ts_runtime >= ts->ts_slptime)
1452 			return (SCHED_INTERACT_HALF);
1453 
1454 	if (ts->ts_runtime > ts->ts_slptime) {
1455 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1456 		return (SCHED_INTERACT_HALF +
1457 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1458 	}
1459 	if (ts->ts_slptime > ts->ts_runtime) {
1460 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1461 		return (ts->ts_runtime / div);
1462 	}
1463 	/* runtime == slptime */
1464 	if (ts->ts_runtime)
1465 		return (SCHED_INTERACT_HALF);
1466 
1467 	/*
1468 	 * This can happen if slptime and runtime are 0.
1469 	 */
1470 	return (0);
1471 
1472 }
1473 
1474 /*
1475  * Scale the scheduling priority according to the "interactivity" of this
1476  * process.
1477  */
1478 static void
1479 sched_priority(struct thread *td)
1480 {
1481 	int score;
1482 	int pri;
1483 
1484 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1485 		return;
1486 	/*
1487 	 * If the score is interactive we place the thread in the realtime
1488 	 * queue with a priority that is less than kernel and interrupt
1489 	 * priorities.  These threads are not subject to nice restrictions.
1490 	 *
1491 	 * Scores greater than this are placed on the normal timeshare queue
1492 	 * where the priority is partially decided by the most recent cpu
1493 	 * utilization and the rest is decided by nice value.
1494 	 *
1495 	 * The nice value of the process has a linear effect on the calculated
1496 	 * score.  Negative nice values make it easier for a thread to be
1497 	 * considered interactive.
1498 	 */
1499 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1500 	if (score < sched_interact) {
1501 		pri = PRI_MIN_INTERACT;
1502 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1503 		    sched_interact) * score;
1504 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1505 		    ("sched_priority: invalid interactive priority %d score %d",
1506 		    pri, score));
1507 	} else {
1508 		pri = SCHED_PRI_MIN;
1509 		if (td->td_sched->ts_ticks)
1510 			pri += min(SCHED_PRI_TICKS(td->td_sched),
1511 			    SCHED_PRI_RANGE - 1);
1512 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1513 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1514 		    ("sched_priority: invalid priority %d: nice %d, "
1515 		    "ticks %d ftick %d ltick %d tick pri %d",
1516 		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1517 		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1518 		    SCHED_PRI_TICKS(td->td_sched)));
1519 	}
1520 	sched_user_prio(td, pri);
1521 
1522 	return;
1523 }
1524 
1525 /*
1526  * This routine enforces a maximum limit on the amount of scheduling history
1527  * kept.  It is called after either the slptime or runtime is adjusted.  This
1528  * function is ugly due to integer math.
1529  */
1530 static void
1531 sched_interact_update(struct thread *td)
1532 {
1533 	struct td_sched *ts;
1534 	u_int sum;
1535 
1536 	ts = td->td_sched;
1537 	sum = ts->ts_runtime + ts->ts_slptime;
1538 	if (sum < SCHED_SLP_RUN_MAX)
1539 		return;
1540 	/*
1541 	 * This only happens from two places:
1542 	 * 1) We have added an unusual amount of run time from fork_exit.
1543 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1544 	 */
1545 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1546 		if (ts->ts_runtime > ts->ts_slptime) {
1547 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1548 			ts->ts_slptime = 1;
1549 		} else {
1550 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1551 			ts->ts_runtime = 1;
1552 		}
1553 		return;
1554 	}
1555 	/*
1556 	 * If we have exceeded by more than 1/5th then the algorithm below
1557 	 * will not bring us back into range.  Dividing by two here forces
1558 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1559 	 */
1560 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1561 		ts->ts_runtime /= 2;
1562 		ts->ts_slptime /= 2;
1563 		return;
1564 	}
1565 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1566 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1567 }
1568 
1569 /*
1570  * Scale back the interactivity history when a child thread is created.  The
1571  * history is inherited from the parent but the thread may behave totally
1572  * differently.  For example, a shell spawning a compiler process.  We want
1573  * to learn that the compiler is behaving badly very quickly.
1574  */
1575 static void
1576 sched_interact_fork(struct thread *td)
1577 {
1578 	int ratio;
1579 	int sum;
1580 
1581 	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1582 	if (sum > SCHED_SLP_RUN_FORK) {
1583 		ratio = sum / SCHED_SLP_RUN_FORK;
1584 		td->td_sched->ts_runtime /= ratio;
1585 		td->td_sched->ts_slptime /= ratio;
1586 	}
1587 }
1588 
1589 /*
1590  * Called from proc0_init() to setup the scheduler fields.
1591  */
1592 void
1593 schedinit(void)
1594 {
1595 
1596 	/*
1597 	 * Set up the scheduler specific parts of proc0.
1598 	 */
1599 	proc0.p_sched = NULL; /* XXX */
1600 	thread0.td_sched = &td_sched0;
1601 	td_sched0.ts_ltick = ticks;
1602 	td_sched0.ts_ftick = ticks;
1603 	td_sched0.ts_slice = 0;
1604 }
1605 
1606 /*
1607  * This is only somewhat accurate since given many processes of the same
1608  * priority they will switch when their slices run out, which will be
1609  * at most sched_slice stathz ticks.
1610  */
1611 int
1612 sched_rr_interval(void)
1613 {
1614 
1615 	/* Convert sched_slice from stathz to hz. */
1616 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1617 }
1618 
1619 /*
1620  * Update the percent cpu tracking information when it is requested or
1621  * the total history exceeds the maximum.  We keep a sliding history of
1622  * tick counts that slowly decays.  This is less precise than the 4BSD
1623  * mechanism since it happens with less regular and frequent events.
1624  */
1625 static void
1626 sched_pctcpu_update(struct td_sched *ts, int run)
1627 {
1628 	int t = ticks;
1629 
1630 	if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1631 		ts->ts_ticks = 0;
1632 		ts->ts_ftick = t - SCHED_TICK_TARG;
1633 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1634 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1635 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1636 		ts->ts_ftick = t - SCHED_TICK_TARG;
1637 	}
1638 	if (run)
1639 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1640 	ts->ts_ltick = t;
1641 }
1642 
1643 /*
1644  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1645  * if necessary.  This is the back-end for several priority related
1646  * functions.
1647  */
1648 static void
1649 sched_thread_priority(struct thread *td, u_char prio)
1650 {
1651 	struct td_sched *ts;
1652 	struct tdq *tdq;
1653 	int oldpri;
1654 
1655 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1656 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1657 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1658 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1659 	if (td != curthread && prio < td->td_priority) {
1660 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1661 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1662 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1663 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1664 		    curthread);
1665 	}
1666 	ts = td->td_sched;
1667 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1668 	if (td->td_priority == prio)
1669 		return;
1670 	/*
1671 	 * If the priority has been elevated due to priority
1672 	 * propagation, we may have to move ourselves to a new
1673 	 * queue.  This could be optimized to not re-add in some
1674 	 * cases.
1675 	 */
1676 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1677 		sched_rem(td);
1678 		td->td_priority = prio;
1679 		sched_add(td, SRQ_BORROWING);
1680 		return;
1681 	}
1682 	/*
1683 	 * If the thread is currently running we may have to adjust the lowpri
1684 	 * information so other cpus are aware of our current priority.
1685 	 */
1686 	if (TD_IS_RUNNING(td)) {
1687 		tdq = TDQ_CPU(ts->ts_cpu);
1688 		oldpri = td->td_priority;
1689 		td->td_priority = prio;
1690 		if (prio < tdq->tdq_lowpri)
1691 			tdq->tdq_lowpri = prio;
1692 		else if (tdq->tdq_lowpri == oldpri)
1693 			tdq_setlowpri(tdq, td);
1694 		return;
1695 	}
1696 	td->td_priority = prio;
1697 }
1698 
1699 /*
1700  * Update a thread's priority when it is lent another thread's
1701  * priority.
1702  */
1703 void
1704 sched_lend_prio(struct thread *td, u_char prio)
1705 {
1706 
1707 	td->td_flags |= TDF_BORROWING;
1708 	sched_thread_priority(td, prio);
1709 }
1710 
1711 /*
1712  * Restore a thread's priority when priority propagation is
1713  * over.  The prio argument is the minimum priority the thread
1714  * needs to have to satisfy other possible priority lending
1715  * requests.  If the thread's regular priority is less
1716  * important than prio, the thread will keep a priority boost
1717  * of prio.
1718  */
1719 void
1720 sched_unlend_prio(struct thread *td, u_char prio)
1721 {
1722 	u_char base_pri;
1723 
1724 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1725 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1726 		base_pri = td->td_user_pri;
1727 	else
1728 		base_pri = td->td_base_pri;
1729 	if (prio >= base_pri) {
1730 		td->td_flags &= ~TDF_BORROWING;
1731 		sched_thread_priority(td, base_pri);
1732 	} else
1733 		sched_lend_prio(td, prio);
1734 }
1735 
1736 /*
1737  * Standard entry for setting the priority to an absolute value.
1738  */
1739 void
1740 sched_prio(struct thread *td, u_char prio)
1741 {
1742 	u_char oldprio;
1743 
1744 	/* First, update the base priority. */
1745 	td->td_base_pri = prio;
1746 
1747 	/*
1748 	 * If the thread is borrowing another thread's priority, don't
1749 	 * ever lower the priority.
1750 	 */
1751 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1752 		return;
1753 
1754 	/* Change the real priority. */
1755 	oldprio = td->td_priority;
1756 	sched_thread_priority(td, prio);
1757 
1758 	/*
1759 	 * If the thread is on a turnstile, then let the turnstile update
1760 	 * its state.
1761 	 */
1762 	if (TD_ON_LOCK(td) && oldprio != prio)
1763 		turnstile_adjust(td, oldprio);
1764 }
1765 
1766 /*
1767  * Set the base user priority, does not effect current running priority.
1768  */
1769 void
1770 sched_user_prio(struct thread *td, u_char prio)
1771 {
1772 
1773 	td->td_base_user_pri = prio;
1774 	if (td->td_lend_user_pri <= prio)
1775 		return;
1776 	td->td_user_pri = prio;
1777 }
1778 
1779 void
1780 sched_lend_user_prio(struct thread *td, u_char prio)
1781 {
1782 
1783 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1784 	td->td_lend_user_pri = prio;
1785 	td->td_user_pri = min(prio, td->td_base_user_pri);
1786 	if (td->td_priority > td->td_user_pri)
1787 		sched_prio(td, td->td_user_pri);
1788 	else if (td->td_priority != td->td_user_pri)
1789 		td->td_flags |= TDF_NEEDRESCHED;
1790 }
1791 
1792 /*
1793  * Handle migration from sched_switch().  This happens only for
1794  * cpu binding.
1795  */
1796 static struct mtx *
1797 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1798 {
1799 	struct tdq *tdn;
1800 
1801 	tdn = TDQ_CPU(td->td_sched->ts_cpu);
1802 #ifdef SMP
1803 	tdq_load_rem(tdq, td);
1804 	/*
1805 	 * Do the lock dance required to avoid LOR.  We grab an extra
1806 	 * spinlock nesting to prevent preemption while we're
1807 	 * not holding either run-queue lock.
1808 	 */
1809 	spinlock_enter();
1810 	thread_lock_block(td);	/* This releases the lock on tdq. */
1811 
1812 	/*
1813 	 * Acquire both run-queue locks before placing the thread on the new
1814 	 * run-queue to avoid deadlocks created by placing a thread with a
1815 	 * blocked lock on the run-queue of a remote processor.  The deadlock
1816 	 * occurs when a third processor attempts to lock the two queues in
1817 	 * question while the target processor is spinning with its own
1818 	 * run-queue lock held while waiting for the blocked lock to clear.
1819 	 */
1820 	tdq_lock_pair(tdn, tdq);
1821 	tdq_add(tdn, td, flags);
1822 	tdq_notify(tdn, td);
1823 	TDQ_UNLOCK(tdn);
1824 	spinlock_exit();
1825 #endif
1826 	return (TDQ_LOCKPTR(tdn));
1827 }
1828 
1829 /*
1830  * Variadic version of thread_lock_unblock() that does not assume td_lock
1831  * is blocked.
1832  */
1833 static inline void
1834 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1835 {
1836 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1837 	    (uintptr_t)mtx);
1838 }
1839 
1840 /*
1841  * Switch threads.  This function has to handle threads coming in while
1842  * blocked for some reason, running, or idle.  It also must deal with
1843  * migrating a thread from one queue to another as running threads may
1844  * be assigned elsewhere via binding.
1845  */
1846 void
1847 sched_switch(struct thread *td, struct thread *newtd, int flags)
1848 {
1849 	struct tdq *tdq;
1850 	struct td_sched *ts;
1851 	struct mtx *mtx;
1852 	int srqflag;
1853 	int cpuid, preempted;
1854 
1855 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1856 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1857 
1858 	cpuid = PCPU_GET(cpuid);
1859 	tdq = TDQ_CPU(cpuid);
1860 	ts = td->td_sched;
1861 	mtx = td->td_lock;
1862 	sched_pctcpu_update(ts, 1);
1863 	ts->ts_rltick = ticks;
1864 	td->td_lastcpu = td->td_oncpu;
1865 	td->td_oncpu = NOCPU;
1866 	preempted = !(td->td_flags & TDF_SLICEEND);
1867 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1868 	td->td_owepreempt = 0;
1869 	if (!TD_IS_IDLETHREAD(td))
1870 		tdq->tdq_switchcnt++;
1871 	/*
1872 	 * The lock pointer in an idle thread should never change.  Reset it
1873 	 * to CAN_RUN as well.
1874 	 */
1875 	if (TD_IS_IDLETHREAD(td)) {
1876 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1877 		TD_SET_CAN_RUN(td);
1878 	} else if (TD_IS_RUNNING(td)) {
1879 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1880 		srqflag = preempted ?
1881 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1882 		    SRQ_OURSELF|SRQ_YIELDING;
1883 #ifdef SMP
1884 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1885 			ts->ts_cpu = sched_pickcpu(td, 0);
1886 #endif
1887 		if (ts->ts_cpu == cpuid)
1888 			tdq_runq_add(tdq, td, srqflag);
1889 		else {
1890 			KASSERT(THREAD_CAN_MIGRATE(td) ||
1891 			    (ts->ts_flags & TSF_BOUND) != 0,
1892 			    ("Thread %p shouldn't migrate", td));
1893 			mtx = sched_switch_migrate(tdq, td, srqflag);
1894 		}
1895 	} else {
1896 		/* This thread must be going to sleep. */
1897 		TDQ_LOCK(tdq);
1898 		mtx = thread_lock_block(td);
1899 		tdq_load_rem(tdq, td);
1900 	}
1901 	/*
1902 	 * We enter here with the thread blocked and assigned to the
1903 	 * appropriate cpu run-queue or sleep-queue and with the current
1904 	 * thread-queue locked.
1905 	 */
1906 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1907 	newtd = choosethread();
1908 	/*
1909 	 * Call the MD code to switch contexts if necessary.
1910 	 */
1911 	if (td != newtd) {
1912 #ifdef	HWPMC_HOOKS
1913 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1914 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1915 #endif
1916 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1917 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1918 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1919 		sched_pctcpu_update(newtd->td_sched, 0);
1920 
1921 #ifdef KDTRACE_HOOKS
1922 		/*
1923 		 * If DTrace has set the active vtime enum to anything
1924 		 * other than INACTIVE (0), then it should have set the
1925 		 * function to call.
1926 		 */
1927 		if (dtrace_vtime_active)
1928 			(*dtrace_vtime_switch_func)(newtd);
1929 #endif
1930 
1931 		cpu_switch(td, newtd, mtx);
1932 		/*
1933 		 * We may return from cpu_switch on a different cpu.  However,
1934 		 * we always return with td_lock pointing to the current cpu's
1935 		 * run queue lock.
1936 		 */
1937 		cpuid = PCPU_GET(cpuid);
1938 		tdq = TDQ_CPU(cpuid);
1939 		lock_profile_obtain_lock_success(
1940 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1941 
1942 		SDT_PROBE0(sched, , , on__cpu);
1943 #ifdef	HWPMC_HOOKS
1944 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1945 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1946 #endif
1947 	} else {
1948 		thread_unblock_switch(td, mtx);
1949 		SDT_PROBE0(sched, , , remain__cpu);
1950 	}
1951 	/*
1952 	 * Assert that all went well and return.
1953 	 */
1954 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1955 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1956 	td->td_oncpu = cpuid;
1957 }
1958 
1959 /*
1960  * Adjust thread priorities as a result of a nice request.
1961  */
1962 void
1963 sched_nice(struct proc *p, int nice)
1964 {
1965 	struct thread *td;
1966 
1967 	PROC_LOCK_ASSERT(p, MA_OWNED);
1968 
1969 	p->p_nice = nice;
1970 	FOREACH_THREAD_IN_PROC(p, td) {
1971 		thread_lock(td);
1972 		sched_priority(td);
1973 		sched_prio(td, td->td_base_user_pri);
1974 		thread_unlock(td);
1975 	}
1976 }
1977 
1978 /*
1979  * Record the sleep time for the interactivity scorer.
1980  */
1981 void
1982 sched_sleep(struct thread *td, int prio)
1983 {
1984 
1985 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1986 
1987 	td->td_slptick = ticks;
1988 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1989 		td->td_flags |= TDF_CANSWAP;
1990 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1991 		return;
1992 	if (static_boost == 1 && prio)
1993 		sched_prio(td, prio);
1994 	else if (static_boost && td->td_priority > static_boost)
1995 		sched_prio(td, static_boost);
1996 }
1997 
1998 /*
1999  * Schedule a thread to resume execution and record how long it voluntarily
2000  * slept.  We also update the pctcpu, interactivity, and priority.
2001  */
2002 void
2003 sched_wakeup(struct thread *td)
2004 {
2005 	struct td_sched *ts;
2006 	int slptick;
2007 
2008 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2009 	ts = td->td_sched;
2010 	td->td_flags &= ~TDF_CANSWAP;
2011 	/*
2012 	 * If we slept for more than a tick update our interactivity and
2013 	 * priority.
2014 	 */
2015 	slptick = td->td_slptick;
2016 	td->td_slptick = 0;
2017 	if (slptick && slptick != ticks) {
2018 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2019 		sched_interact_update(td);
2020 		sched_pctcpu_update(ts, 0);
2021 	}
2022 	/*
2023 	 * Reset the slice value since we slept and advanced the round-robin.
2024 	 */
2025 	ts->ts_slice = 0;
2026 	sched_add(td, SRQ_BORING);
2027 }
2028 
2029 /*
2030  * Penalize the parent for creating a new child and initialize the child's
2031  * priority.
2032  */
2033 void
2034 sched_fork(struct thread *td, struct thread *child)
2035 {
2036 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2037 	sched_pctcpu_update(td->td_sched, 1);
2038 	sched_fork_thread(td, child);
2039 	/*
2040 	 * Penalize the parent and child for forking.
2041 	 */
2042 	sched_interact_fork(child);
2043 	sched_priority(child);
2044 	td->td_sched->ts_runtime += tickincr;
2045 	sched_interact_update(td);
2046 	sched_priority(td);
2047 }
2048 
2049 /*
2050  * Fork a new thread, may be within the same process.
2051  */
2052 void
2053 sched_fork_thread(struct thread *td, struct thread *child)
2054 {
2055 	struct td_sched *ts;
2056 	struct td_sched *ts2;
2057 	struct tdq *tdq;
2058 
2059 	tdq = TDQ_SELF();
2060 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2061 	/*
2062 	 * Initialize child.
2063 	 */
2064 	ts = td->td_sched;
2065 	ts2 = child->td_sched;
2066 	child->td_lock = TDQ_LOCKPTR(tdq);
2067 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2068 	ts2->ts_cpu = ts->ts_cpu;
2069 	ts2->ts_flags = 0;
2070 	/*
2071 	 * Grab our parents cpu estimation information.
2072 	 */
2073 	ts2->ts_ticks = ts->ts_ticks;
2074 	ts2->ts_ltick = ts->ts_ltick;
2075 	ts2->ts_ftick = ts->ts_ftick;
2076 	/*
2077 	 * Do not inherit any borrowed priority from the parent.
2078 	 */
2079 	child->td_priority = child->td_base_pri;
2080 	/*
2081 	 * And update interactivity score.
2082 	 */
2083 	ts2->ts_slptime = ts->ts_slptime;
2084 	ts2->ts_runtime = ts->ts_runtime;
2085 	/* Attempt to quickly learn interactivity. */
2086 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2087 #ifdef KTR
2088 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2089 #endif
2090 }
2091 
2092 /*
2093  * Adjust the priority class of a thread.
2094  */
2095 void
2096 sched_class(struct thread *td, int class)
2097 {
2098 
2099 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2100 	if (td->td_pri_class == class)
2101 		return;
2102 	td->td_pri_class = class;
2103 }
2104 
2105 /*
2106  * Return some of the child's priority and interactivity to the parent.
2107  */
2108 void
2109 sched_exit(struct proc *p, struct thread *child)
2110 {
2111 	struct thread *td;
2112 
2113 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2114 	    "prio:%d", child->td_priority);
2115 	PROC_LOCK_ASSERT(p, MA_OWNED);
2116 	td = FIRST_THREAD_IN_PROC(p);
2117 	sched_exit_thread(td, child);
2118 }
2119 
2120 /*
2121  * Penalize another thread for the time spent on this one.  This helps to
2122  * worsen the priority and interactivity of processes which schedule batch
2123  * jobs such as make.  This has little effect on the make process itself but
2124  * causes new processes spawned by it to receive worse scores immediately.
2125  */
2126 void
2127 sched_exit_thread(struct thread *td, struct thread *child)
2128 {
2129 
2130 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2131 	    "prio:%d", child->td_priority);
2132 	/*
2133 	 * Give the child's runtime to the parent without returning the
2134 	 * sleep time as a penalty to the parent.  This causes shells that
2135 	 * launch expensive things to mark their children as expensive.
2136 	 */
2137 	thread_lock(td);
2138 	td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2139 	sched_interact_update(td);
2140 	sched_priority(td);
2141 	thread_unlock(td);
2142 }
2143 
2144 void
2145 sched_preempt(struct thread *td)
2146 {
2147 	struct tdq *tdq;
2148 
2149 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2150 
2151 	thread_lock(td);
2152 	tdq = TDQ_SELF();
2153 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2154 	tdq->tdq_ipipending = 0;
2155 	if (td->td_priority > tdq->tdq_lowpri) {
2156 		int flags;
2157 
2158 		flags = SW_INVOL | SW_PREEMPT;
2159 		if (td->td_critnest > 1)
2160 			td->td_owepreempt = 1;
2161 		else if (TD_IS_IDLETHREAD(td))
2162 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2163 		else
2164 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2165 	}
2166 	thread_unlock(td);
2167 }
2168 
2169 /*
2170  * Fix priorities on return to user-space.  Priorities may be elevated due
2171  * to static priorities in msleep() or similar.
2172  */
2173 void
2174 sched_userret(struct thread *td)
2175 {
2176 	/*
2177 	 * XXX we cheat slightly on the locking here to avoid locking in
2178 	 * the usual case.  Setting td_priority here is essentially an
2179 	 * incomplete workaround for not setting it properly elsewhere.
2180 	 * Now that some interrupt handlers are threads, not setting it
2181 	 * properly elsewhere can clobber it in the window between setting
2182 	 * it here and returning to user mode, so don't waste time setting
2183 	 * it perfectly here.
2184 	 */
2185 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2186 	    ("thread with borrowed priority returning to userland"));
2187 	if (td->td_priority != td->td_user_pri) {
2188 		thread_lock(td);
2189 		td->td_priority = td->td_user_pri;
2190 		td->td_base_pri = td->td_user_pri;
2191 		tdq_setlowpri(TDQ_SELF(), td);
2192 		thread_unlock(td);
2193         }
2194 }
2195 
2196 /*
2197  * Handle a stathz tick.  This is really only relevant for timeshare
2198  * threads.
2199  */
2200 void
2201 sched_clock(struct thread *td)
2202 {
2203 	struct tdq *tdq;
2204 	struct td_sched *ts;
2205 
2206 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2207 	tdq = TDQ_SELF();
2208 #ifdef SMP
2209 	/*
2210 	 * We run the long term load balancer infrequently on the first cpu.
2211 	 */
2212 	if (balance_tdq == tdq) {
2213 		if (balance_ticks && --balance_ticks == 0)
2214 			sched_balance();
2215 	}
2216 #endif
2217 	/*
2218 	 * Save the old switch count so we have a record of the last ticks
2219 	 * activity.   Initialize the new switch count based on our load.
2220 	 * If there is some activity seed it to reflect that.
2221 	 */
2222 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2223 	tdq->tdq_switchcnt = tdq->tdq_load;
2224 	/*
2225 	 * Advance the insert index once for each tick to ensure that all
2226 	 * threads get a chance to run.
2227 	 */
2228 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2229 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2230 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2231 			tdq->tdq_ridx = tdq->tdq_idx;
2232 	}
2233 	ts = td->td_sched;
2234 	sched_pctcpu_update(ts, 1);
2235 	if (td->td_pri_class & PRI_FIFO_BIT)
2236 		return;
2237 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2238 		/*
2239 		 * We used a tick; charge it to the thread so
2240 		 * that we can compute our interactivity.
2241 		 */
2242 		td->td_sched->ts_runtime += tickincr;
2243 		sched_interact_update(td);
2244 		sched_priority(td);
2245 	}
2246 
2247 	/*
2248 	 * Force a context switch if the current thread has used up a full
2249 	 * time slice (default is 100ms).
2250 	 */
2251 	if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2252 		ts->ts_slice = 0;
2253 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2254 	}
2255 }
2256 
2257 /*
2258  * Called once per hz tick.
2259  */
2260 void
2261 sched_tick(int cnt)
2262 {
2263 
2264 }
2265 
2266 /*
2267  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2268  * cooperative idle threads.
2269  */
2270 int
2271 sched_runnable(void)
2272 {
2273 	struct tdq *tdq;
2274 	int load;
2275 
2276 	load = 1;
2277 
2278 	tdq = TDQ_SELF();
2279 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2280 		if (tdq->tdq_load > 0)
2281 			goto out;
2282 	} else
2283 		if (tdq->tdq_load - 1 > 0)
2284 			goto out;
2285 	load = 0;
2286 out:
2287 	return (load);
2288 }
2289 
2290 /*
2291  * Choose the highest priority thread to run.  The thread is removed from
2292  * the run-queue while running however the load remains.  For SMP we set
2293  * the tdq in the global idle bitmask if it idles here.
2294  */
2295 struct thread *
2296 sched_choose(void)
2297 {
2298 	struct thread *td;
2299 	struct tdq *tdq;
2300 
2301 	tdq = TDQ_SELF();
2302 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2303 	td = tdq_choose(tdq);
2304 	if (td) {
2305 		tdq_runq_rem(tdq, td);
2306 		tdq->tdq_lowpri = td->td_priority;
2307 		return (td);
2308 	}
2309 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2310 	return (PCPU_GET(idlethread));
2311 }
2312 
2313 /*
2314  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2315  * we always request it once we exit a critical section.
2316  */
2317 static inline void
2318 sched_setpreempt(struct thread *td)
2319 {
2320 	struct thread *ctd;
2321 	int cpri;
2322 	int pri;
2323 
2324 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2325 
2326 	ctd = curthread;
2327 	pri = td->td_priority;
2328 	cpri = ctd->td_priority;
2329 	if (pri < cpri)
2330 		ctd->td_flags |= TDF_NEEDRESCHED;
2331 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2332 		return;
2333 	if (!sched_shouldpreempt(pri, cpri, 0))
2334 		return;
2335 	ctd->td_owepreempt = 1;
2336 }
2337 
2338 /*
2339  * Add a thread to a thread queue.  Select the appropriate runq and add the
2340  * thread to it.  This is the internal function called when the tdq is
2341  * predetermined.
2342  */
2343 void
2344 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2345 {
2346 
2347 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2348 	KASSERT((td->td_inhibitors == 0),
2349 	    ("sched_add: trying to run inhibited thread"));
2350 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2351 	    ("sched_add: bad thread state"));
2352 	KASSERT(td->td_flags & TDF_INMEM,
2353 	    ("sched_add: thread swapped out"));
2354 
2355 	if (td->td_priority < tdq->tdq_lowpri)
2356 		tdq->tdq_lowpri = td->td_priority;
2357 	tdq_runq_add(tdq, td, flags);
2358 	tdq_load_add(tdq, td);
2359 }
2360 
2361 /*
2362  * Select the target thread queue and add a thread to it.  Request
2363  * preemption or IPI a remote processor if required.
2364  */
2365 void
2366 sched_add(struct thread *td, int flags)
2367 {
2368 	struct tdq *tdq;
2369 #ifdef SMP
2370 	int cpu;
2371 #endif
2372 
2373 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2374 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2375 	    sched_tdname(curthread));
2376 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2377 	    KTR_ATTR_LINKED, sched_tdname(td));
2378 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2379 	    flags & SRQ_PREEMPTED);
2380 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2381 	/*
2382 	 * Recalculate the priority before we select the target cpu or
2383 	 * run-queue.
2384 	 */
2385 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2386 		sched_priority(td);
2387 #ifdef SMP
2388 	/*
2389 	 * Pick the destination cpu and if it isn't ours transfer to the
2390 	 * target cpu.
2391 	 */
2392 	cpu = sched_pickcpu(td, flags);
2393 	tdq = sched_setcpu(td, cpu, flags);
2394 	tdq_add(tdq, td, flags);
2395 	if (cpu != PCPU_GET(cpuid)) {
2396 		tdq_notify(tdq, td);
2397 		return;
2398 	}
2399 #else
2400 	tdq = TDQ_SELF();
2401 	TDQ_LOCK(tdq);
2402 	/*
2403 	 * Now that the thread is moving to the run-queue, set the lock
2404 	 * to the scheduler's lock.
2405 	 */
2406 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2407 	tdq_add(tdq, td, flags);
2408 #endif
2409 	if (!(flags & SRQ_YIELDING))
2410 		sched_setpreempt(td);
2411 }
2412 
2413 /*
2414  * Remove a thread from a run-queue without running it.  This is used
2415  * when we're stealing a thread from a remote queue.  Otherwise all threads
2416  * exit by calling sched_exit_thread() and sched_throw() themselves.
2417  */
2418 void
2419 sched_rem(struct thread *td)
2420 {
2421 	struct tdq *tdq;
2422 
2423 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2424 	    "prio:%d", td->td_priority);
2425 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2426 	tdq = TDQ_CPU(td->td_sched->ts_cpu);
2427 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2428 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2429 	KASSERT(TD_ON_RUNQ(td),
2430 	    ("sched_rem: thread not on run queue"));
2431 	tdq_runq_rem(tdq, td);
2432 	tdq_load_rem(tdq, td);
2433 	TD_SET_CAN_RUN(td);
2434 	if (td->td_priority == tdq->tdq_lowpri)
2435 		tdq_setlowpri(tdq, NULL);
2436 }
2437 
2438 /*
2439  * Fetch cpu utilization information.  Updates on demand.
2440  */
2441 fixpt_t
2442 sched_pctcpu(struct thread *td)
2443 {
2444 	fixpt_t pctcpu;
2445 	struct td_sched *ts;
2446 
2447 	pctcpu = 0;
2448 	ts = td->td_sched;
2449 	if (ts == NULL)
2450 		return (0);
2451 
2452 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2453 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2454 	if (ts->ts_ticks) {
2455 		int rtick;
2456 
2457 		/* How many rtick per second ? */
2458 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2459 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2460 	}
2461 
2462 	return (pctcpu);
2463 }
2464 
2465 /*
2466  * Enforce affinity settings for a thread.  Called after adjustments to
2467  * cpumask.
2468  */
2469 void
2470 sched_affinity(struct thread *td)
2471 {
2472 #ifdef SMP
2473 	struct td_sched *ts;
2474 
2475 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2476 	ts = td->td_sched;
2477 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2478 		return;
2479 	if (TD_ON_RUNQ(td)) {
2480 		sched_rem(td);
2481 		sched_add(td, SRQ_BORING);
2482 		return;
2483 	}
2484 	if (!TD_IS_RUNNING(td))
2485 		return;
2486 	/*
2487 	 * Force a switch before returning to userspace.  If the
2488 	 * target thread is not running locally send an ipi to force
2489 	 * the issue.
2490 	 */
2491 	td->td_flags |= TDF_NEEDRESCHED;
2492 	if (td != curthread)
2493 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2494 #endif
2495 }
2496 
2497 /*
2498  * Bind a thread to a target cpu.
2499  */
2500 void
2501 sched_bind(struct thread *td, int cpu)
2502 {
2503 	struct td_sched *ts;
2504 
2505 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2506 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2507 	ts = td->td_sched;
2508 	if (ts->ts_flags & TSF_BOUND)
2509 		sched_unbind(td);
2510 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2511 	ts->ts_flags |= TSF_BOUND;
2512 	sched_pin();
2513 	if (PCPU_GET(cpuid) == cpu)
2514 		return;
2515 	ts->ts_cpu = cpu;
2516 	/* When we return from mi_switch we'll be on the correct cpu. */
2517 	mi_switch(SW_VOL, NULL);
2518 }
2519 
2520 /*
2521  * Release a bound thread.
2522  */
2523 void
2524 sched_unbind(struct thread *td)
2525 {
2526 	struct td_sched *ts;
2527 
2528 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2529 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2530 	ts = td->td_sched;
2531 	if ((ts->ts_flags & TSF_BOUND) == 0)
2532 		return;
2533 	ts->ts_flags &= ~TSF_BOUND;
2534 	sched_unpin();
2535 }
2536 
2537 int
2538 sched_is_bound(struct thread *td)
2539 {
2540 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2541 	return (td->td_sched->ts_flags & TSF_BOUND);
2542 }
2543 
2544 /*
2545  * Basic yield call.
2546  */
2547 void
2548 sched_relinquish(struct thread *td)
2549 {
2550 	thread_lock(td);
2551 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2552 	thread_unlock(td);
2553 }
2554 
2555 /*
2556  * Return the total system load.
2557  */
2558 int
2559 sched_load(void)
2560 {
2561 #ifdef SMP
2562 	int total;
2563 	int i;
2564 
2565 	total = 0;
2566 	CPU_FOREACH(i)
2567 		total += TDQ_CPU(i)->tdq_sysload;
2568 	return (total);
2569 #else
2570 	return (TDQ_SELF()->tdq_sysload);
2571 #endif
2572 }
2573 
2574 int
2575 sched_sizeof_proc(void)
2576 {
2577 	return (sizeof(struct proc));
2578 }
2579 
2580 int
2581 sched_sizeof_thread(void)
2582 {
2583 	return (sizeof(struct thread) + sizeof(struct td_sched));
2584 }
2585 
2586 #ifdef SMP
2587 #define	TDQ_IDLESPIN(tdq)						\
2588     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2589 #else
2590 #define	TDQ_IDLESPIN(tdq)	1
2591 #endif
2592 
2593 /*
2594  * The actual idle process.
2595  */
2596 void
2597 sched_idletd(void *dummy)
2598 {
2599 	struct thread *td;
2600 	struct tdq *tdq;
2601 	int oldswitchcnt, switchcnt;
2602 	int i;
2603 
2604 	mtx_assert(&Giant, MA_NOTOWNED);
2605 	td = curthread;
2606 	tdq = TDQ_SELF();
2607 	THREAD_NO_SLEEPING();
2608 	oldswitchcnt = -1;
2609 	for (;;) {
2610 		if (tdq->tdq_load) {
2611 			thread_lock(td);
2612 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2613 			thread_unlock(td);
2614 		}
2615 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2616 #ifdef SMP
2617 		if (switchcnt != oldswitchcnt) {
2618 			oldswitchcnt = switchcnt;
2619 			if (tdq_idled(tdq) == 0)
2620 				continue;
2621 		}
2622 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2623 #else
2624 		oldswitchcnt = switchcnt;
2625 #endif
2626 		/*
2627 		 * If we're switching very frequently, spin while checking
2628 		 * for load rather than entering a low power state that
2629 		 * may require an IPI.  However, don't do any busy
2630 		 * loops while on SMT machines as this simply steals
2631 		 * cycles from cores doing useful work.
2632 		 */
2633 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2634 			for (i = 0; i < sched_idlespins; i++) {
2635 				if (tdq->tdq_load)
2636 					break;
2637 				cpu_spinwait();
2638 			}
2639 		}
2640 
2641 		/* If there was context switch during spin, restart it. */
2642 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2643 		if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2644 			continue;
2645 
2646 		/* Run main MD idle handler. */
2647 		tdq->tdq_cpu_idle = 1;
2648 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2649 		tdq->tdq_cpu_idle = 0;
2650 
2651 		/*
2652 		 * Account thread-less hardware interrupts and
2653 		 * other wakeup reasons equal to context switches.
2654 		 */
2655 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2656 		if (switchcnt != oldswitchcnt)
2657 			continue;
2658 		tdq->tdq_switchcnt++;
2659 		oldswitchcnt++;
2660 	}
2661 }
2662 
2663 /*
2664  * A CPU is entering for the first time or a thread is exiting.
2665  */
2666 void
2667 sched_throw(struct thread *td)
2668 {
2669 	struct thread *newtd;
2670 	struct tdq *tdq;
2671 
2672 	tdq = TDQ_SELF();
2673 	if (td == NULL) {
2674 		/* Correct spinlock nesting and acquire the correct lock. */
2675 		TDQ_LOCK(tdq);
2676 		spinlock_exit();
2677 		PCPU_SET(switchtime, cpu_ticks());
2678 		PCPU_SET(switchticks, ticks);
2679 	} else {
2680 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2681 		tdq_load_rem(tdq, td);
2682 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2683 	}
2684 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2685 	newtd = choosethread();
2686 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2687 	cpu_throw(td, newtd);		/* doesn't return */
2688 }
2689 
2690 /*
2691  * This is called from fork_exit().  Just acquire the correct locks and
2692  * let fork do the rest of the work.
2693  */
2694 void
2695 sched_fork_exit(struct thread *td)
2696 {
2697 	struct td_sched *ts;
2698 	struct tdq *tdq;
2699 	int cpuid;
2700 
2701 	/*
2702 	 * Finish setting up thread glue so that it begins execution in a
2703 	 * non-nested critical section with the scheduler lock held.
2704 	 */
2705 	cpuid = PCPU_GET(cpuid);
2706 	tdq = TDQ_CPU(cpuid);
2707 	ts = td->td_sched;
2708 	if (TD_IS_IDLETHREAD(td))
2709 		td->td_lock = TDQ_LOCKPTR(tdq);
2710 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2711 	td->td_oncpu = cpuid;
2712 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2713 	lock_profile_obtain_lock_success(
2714 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2715 }
2716 
2717 /*
2718  * Create on first use to catch odd startup conditons.
2719  */
2720 char *
2721 sched_tdname(struct thread *td)
2722 {
2723 #ifdef KTR
2724 	struct td_sched *ts;
2725 
2726 	ts = td->td_sched;
2727 	if (ts->ts_name[0] == '\0')
2728 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2729 		    "%s tid %d", td->td_name, td->td_tid);
2730 	return (ts->ts_name);
2731 #else
2732 	return (td->td_name);
2733 #endif
2734 }
2735 
2736 #ifdef KTR
2737 void
2738 sched_clear_tdname(struct thread *td)
2739 {
2740 	struct td_sched *ts;
2741 
2742 	ts = td->td_sched;
2743 	ts->ts_name[0] = '\0';
2744 }
2745 #endif
2746 
2747 #ifdef SMP
2748 
2749 /*
2750  * Build the CPU topology dump string. Is recursively called to collect
2751  * the topology tree.
2752  */
2753 static int
2754 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2755     int indent)
2756 {
2757 	char cpusetbuf[CPUSETBUFSIZ];
2758 	int i, first;
2759 
2760 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2761 	    "", 1 + indent / 2, cg->cg_level);
2762 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2763 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2764 	first = TRUE;
2765 	for (i = 0; i < MAXCPU; i++) {
2766 		if (CPU_ISSET(i, &cg->cg_mask)) {
2767 			if (!first)
2768 				sbuf_printf(sb, ", ");
2769 			else
2770 				first = FALSE;
2771 			sbuf_printf(sb, "%d", i);
2772 		}
2773 	}
2774 	sbuf_printf(sb, "</cpu>\n");
2775 
2776 	if (cg->cg_flags != 0) {
2777 		sbuf_printf(sb, "%*s <flags>", indent, "");
2778 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2779 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2780 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2781 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2782 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2783 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2784 		sbuf_printf(sb, "</flags>\n");
2785 	}
2786 
2787 	if (cg->cg_children > 0) {
2788 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2789 		for (i = 0; i < cg->cg_children; i++)
2790 			sysctl_kern_sched_topology_spec_internal(sb,
2791 			    &cg->cg_child[i], indent+2);
2792 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2793 	}
2794 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2795 	return (0);
2796 }
2797 
2798 /*
2799  * Sysctl handler for retrieving topology dump. It's a wrapper for
2800  * the recursive sysctl_kern_smp_topology_spec_internal().
2801  */
2802 static int
2803 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2804 {
2805 	struct sbuf *topo;
2806 	int err;
2807 
2808 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2809 
2810 	topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2811 	if (topo == NULL)
2812 		return (ENOMEM);
2813 
2814 	sbuf_printf(topo, "<groups>\n");
2815 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2816 	sbuf_printf(topo, "</groups>\n");
2817 
2818 	if (err == 0) {
2819 		sbuf_finish(topo);
2820 		err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2821 	}
2822 	sbuf_delete(topo);
2823 	return (err);
2824 }
2825 
2826 #endif
2827 
2828 static int
2829 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2830 {
2831 	int error, new_val, period;
2832 
2833 	period = 1000000 / realstathz;
2834 	new_val = period * sched_slice;
2835 	error = sysctl_handle_int(oidp, &new_val, 0, req);
2836 	if (error != 0 || req->newptr == NULL)
2837 		return (error);
2838 	if (new_val <= 0)
2839 		return (EINVAL);
2840 	sched_slice = imax(1, (new_val + period / 2) / period);
2841 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
2842 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2843 	    realstathz);
2844 	return (0);
2845 }
2846 
2847 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2848 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2849     "Scheduler name");
2850 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2851     NULL, 0, sysctl_kern_quantum, "I",
2852     "Quantum for timeshare threads in microseconds");
2853 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2854     "Quantum for timeshare threads in stathz ticks");
2855 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2856     "Interactivity score threshold");
2857 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2858     &preempt_thresh, 0,
2859     "Maximal (lowest) priority for preemption");
2860 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2861     "Assign static kernel priorities to sleeping threads");
2862 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2863     "Number of times idle thread will spin waiting for new work");
2864 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2865     &sched_idlespinthresh, 0,
2866     "Threshold before we will permit idle thread spinning");
2867 #ifdef SMP
2868 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2869     "Number of hz ticks to keep thread affinity for");
2870 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2871     "Enables the long-term load balancer");
2872 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2873     &balance_interval, 0,
2874     "Average period in stathz ticks to run the long-term balancer");
2875 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2876     "Attempts to steal work from other cores before idling");
2877 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2878     "Minimum load on remote CPU before we'll steal");
2879 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2880     CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2881     "XML dump of detected CPU topology");
2882 #endif
2883 
2884 /* ps compat.  All cpu percentages from ULE are weighted. */
2885 static int ccpu = 0;
2886 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2887