xref: /freebsd/sys/kern/sched_ule.c (revision f95c930e6701aa675377bc5871ea490dd565eeba)
1 /*-
2  * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice unmodified, this list of conditions, and the following
10  *    disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25  */
26 
27 /*
28  * This file implements the ULE scheduler.  ULE supports independent CPU
29  * run queues and fine grain locking.  It has superior interactive
30  * performance under load even on uni-processor systems.
31  *
32  * etymology:
33  *   ULE is the last three letters in schedule.  It owes its name to a
34  * generic user created for a scheduling system by Paul Mikesell at
35  * Isilon Systems and a general lack of creativity on the part of the author.
36  */
37 
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
40 
41 #include "opt_hwpmc_hooks.h"
42 #include "opt_sched.h"
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/kdb.h>
47 #include <sys/kernel.h>
48 #include <sys/ktr.h>
49 #include <sys/limits.h>
50 #include <sys/lock.h>
51 #include <sys/mutex.h>
52 #include <sys/proc.h>
53 #include <sys/resource.h>
54 #include <sys/resourcevar.h>
55 #include <sys/sched.h>
56 #include <sys/sdt.h>
57 #include <sys/smp.h>
58 #include <sys/sx.h>
59 #include <sys/sysctl.h>
60 #include <sys/sysproto.h>
61 #include <sys/turnstile.h>
62 #include <sys/umtx.h>
63 #include <sys/vmmeter.h>
64 #include <sys/cpuset.h>
65 #include <sys/sbuf.h>
66 
67 #ifdef HWPMC_HOOKS
68 #include <sys/pmckern.h>
69 #endif
70 
71 #ifdef KDTRACE_HOOKS
72 #include <sys/dtrace_bsd.h>
73 int				dtrace_vtime_active;
74 dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
75 #endif
76 
77 #include <machine/cpu.h>
78 #include <machine/smp.h>
79 
80 #define	KTR_ULE	0
81 
82 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
83 #define	TDQ_NAME_LEN	(sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
84 #define	TDQ_LOADNAME_LEN	(sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
85 
86 /*
87  * Thread scheduler specific section.  All fields are protected
88  * by the thread lock.
89  */
90 struct td_sched {
91 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
92 	short		ts_flags;	/* TSF_* flags. */
93 	int		ts_cpu;		/* CPU that we have affinity for. */
94 	int		ts_rltick;	/* Real last tick, for affinity. */
95 	int		ts_slice;	/* Ticks of slice remaining. */
96 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
97 	u_int		ts_runtime;	/* Number of ticks we were running */
98 	int		ts_ltick;	/* Last tick that we were running on */
99 	int		ts_ftick;	/* First tick that we were running on */
100 	int		ts_ticks;	/* Tick count */
101 #ifdef KTR
102 	char		ts_name[TS_NAME_LEN];
103 #endif
104 };
105 /* flags kept in ts_flags */
106 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
107 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
108 
109 static struct td_sched td_sched0;
110 
111 #define	THREAD_CAN_MIGRATE(td)	((td)->td_pinned == 0)
112 #define	THREAD_CAN_SCHED(td, cpu)	\
113     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
114 
115 /*
116  * Priority ranges used for interactive and non-interactive timeshare
117  * threads.  The timeshare priorities are split up into four ranges.
118  * The first range handles interactive threads.  The last three ranges
119  * (NHALF, x, and NHALF) handle non-interactive threads with the outer
120  * ranges supporting nice values.
121  */
122 #define	PRI_TIMESHARE_RANGE	(PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
123 #define	PRI_INTERACT_RANGE	((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
124 #define	PRI_BATCH_RANGE		(PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
125 
126 #define	PRI_MIN_INTERACT	PRI_MIN_TIMESHARE
127 #define	PRI_MAX_INTERACT	(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
128 #define	PRI_MIN_BATCH		(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
129 #define	PRI_MAX_BATCH		PRI_MAX_TIMESHARE
130 
131 /*
132  * Cpu percentage computation macros and defines.
133  *
134  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
135  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
136  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
137  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
138  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
139  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
140  */
141 #define	SCHED_TICK_SECS		10
142 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
143 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
144 #define	SCHED_TICK_SHIFT	10
145 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
146 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
147 
148 /*
149  * These macros determine priorities for non-interactive threads.  They are
150  * assigned a priority based on their recent cpu utilization as expressed
151  * by the ratio of ticks to the tick total.  NHALF priorities at the start
152  * and end of the MIN to MAX timeshare range are only reachable with negative
153  * or positive nice respectively.
154  *
155  * PRI_RANGE:	Priority range for utilization dependent priorities.
156  * PRI_NRESV:	Number of nice values.
157  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
158  * PRI_NICE:	Determines the part of the priority inherited from nice.
159  */
160 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
161 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
162 #define	SCHED_PRI_MIN		(PRI_MIN_BATCH + SCHED_PRI_NHALF)
163 #define	SCHED_PRI_MAX		(PRI_MAX_BATCH - SCHED_PRI_NHALF)
164 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
165 #define	SCHED_PRI_TICKS(ts)						\
166     (SCHED_TICK_HZ((ts)) /						\
167     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
168 #define	SCHED_PRI_NICE(nice)	(nice)
169 
170 /*
171  * These determine the interactivity of a process.  Interactivity differs from
172  * cpu utilization in that it expresses the voluntary time slept vs time ran
173  * while cpu utilization includes all time not running.  This more accurately
174  * models the intent of the thread.
175  *
176  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
177  *		before throttling back.
178  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
179  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
180  * INTERACT_THRESH:	Threshold for placement on the current runq.
181  */
182 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
183 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
184 #define	SCHED_INTERACT_MAX	(100)
185 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
186 #define	SCHED_INTERACT_THRESH	(30)
187 
188 /*
189  * These parameters determine the slice behavior for batch work.
190  */
191 #define	SCHED_SLICE_DEFAULT_DIVISOR	10	/* ~94 ms, 12 stathz ticks. */
192 #define	SCHED_SLICE_MIN_DIVISOR		6	/* DEFAULT/MIN = ~16 ms. */
193 
194 /* Flags kept in td_flags. */
195 #define	TDF_SLICEEND	TDF_SCHED2	/* Thread time slice is over. */
196 
197 /*
198  * tickincr:		Converts a stathz tick into a hz domain scaled by
199  *			the shift factor.  Without the shift the error rate
200  *			due to rounding would be unacceptably high.
201  * realstathz:		stathz is sometimes 0 and run off of hz.
202  * sched_slice:		Runtime of each thread before rescheduling.
203  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
204  */
205 static int sched_interact = SCHED_INTERACT_THRESH;
206 static int tickincr = 8 << SCHED_TICK_SHIFT;
207 static int realstathz = 127;	/* reset during boot. */
208 static int sched_slice = 10;	/* reset during boot. */
209 static int sched_slice_min = 1;	/* reset during boot. */
210 #ifdef PREEMPTION
211 #ifdef FULL_PREEMPTION
212 static int preempt_thresh = PRI_MAX_IDLE;
213 #else
214 static int preempt_thresh = PRI_MIN_KERN;
215 #endif
216 #else
217 static int preempt_thresh = 0;
218 #endif
219 static int static_boost = PRI_MIN_BATCH;
220 static int sched_idlespins = 10000;
221 static int sched_idlespinthresh = -1;
222 
223 /*
224  * tdq - per processor runqs and statistics.  All fields are protected by the
225  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
226  * locking in sched_pickcpu();
227  */
228 struct tdq {
229 	/*
230 	 * Ordered to improve efficiency of cpu_search() and switch().
231 	 * tdq_lock is padded to avoid false sharing with tdq_load and
232 	 * tdq_cpu_idle.
233 	 */
234 	struct mtx_padalign tdq_lock;		/* run queue lock. */
235 	struct cpu_group *tdq_cg;		/* Pointer to cpu topology. */
236 	volatile int	tdq_load;		/* Aggregate load. */
237 	volatile int	tdq_cpu_idle;		/* cpu_idle() is active. */
238 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
239 	int		tdq_transferable;	/* Transferable thread count. */
240 	short		tdq_switchcnt;		/* Switches this tick. */
241 	short		tdq_oldswitchcnt;	/* Switches last tick. */
242 	u_char		tdq_lowpri;		/* Lowest priority thread. */
243 	u_char		tdq_ipipending;		/* IPI pending. */
244 	u_char		tdq_idx;		/* Current insert index. */
245 	u_char		tdq_ridx;		/* Current removal index. */
246 	struct runq	tdq_realtime;		/* real-time run queue. */
247 	struct runq	tdq_timeshare;		/* timeshare run queue. */
248 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
249 	char		tdq_name[TDQ_NAME_LEN];
250 #ifdef KTR
251 	char		tdq_loadname[TDQ_LOADNAME_LEN];
252 #endif
253 } __aligned(64);
254 
255 /* Idle thread states and config. */
256 #define	TDQ_RUNNING	1
257 #define	TDQ_IDLE	2
258 
259 #ifdef SMP
260 struct cpu_group *cpu_top;		/* CPU topology */
261 
262 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
263 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
264 
265 /*
266  * Run-time tunables.
267  */
268 static int rebalance = 1;
269 static int balance_interval = 128;	/* Default set in sched_initticks(). */
270 static int affinity;
271 static int steal_idle = 1;
272 static int steal_thresh = 2;
273 
274 /*
275  * One thread queue per processor.
276  */
277 static struct tdq	tdq_cpu[MAXCPU];
278 static struct tdq	*balance_tdq;
279 static int balance_ticks;
280 static DPCPU_DEFINE(uint32_t, randomval);
281 
282 #define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
283 #define	TDQ_CPU(x)	(&tdq_cpu[(x)])
284 #define	TDQ_ID(x)	((int)((x) - tdq_cpu))
285 #else	/* !SMP */
286 static struct tdq	tdq_cpu;
287 
288 #define	TDQ_ID(x)	(0)
289 #define	TDQ_SELF()	(&tdq_cpu)
290 #define	TDQ_CPU(x)	(&tdq_cpu)
291 #endif
292 
293 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
294 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
295 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
296 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
297 #define	TDQ_LOCKPTR(t)		((struct mtx *)(&(t)->tdq_lock))
298 
299 static void sched_priority(struct thread *);
300 static void sched_thread_priority(struct thread *, u_char);
301 static int sched_interact_score(struct thread *);
302 static void sched_interact_update(struct thread *);
303 static void sched_interact_fork(struct thread *);
304 static void sched_pctcpu_update(struct td_sched *, int);
305 
306 /* Operations on per processor queues */
307 static struct thread *tdq_choose(struct tdq *);
308 static void tdq_setup(struct tdq *);
309 static void tdq_load_add(struct tdq *, struct thread *);
310 static void tdq_load_rem(struct tdq *, struct thread *);
311 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
312 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
313 static inline int sched_shouldpreempt(int, int, int);
314 void tdq_print(int cpu);
315 static void runq_print(struct runq *rq);
316 static void tdq_add(struct tdq *, struct thread *, int);
317 #ifdef SMP
318 static int tdq_move(struct tdq *, struct tdq *);
319 static int tdq_idled(struct tdq *);
320 static void tdq_notify(struct tdq *, struct thread *);
321 static struct thread *tdq_steal(struct tdq *, int);
322 static struct thread *runq_steal(struct runq *, int);
323 static int sched_pickcpu(struct thread *, int);
324 static void sched_balance(void);
325 static int sched_balance_pair(struct tdq *, struct tdq *);
326 static inline struct tdq *sched_setcpu(struct thread *, int, int);
327 static inline void thread_unblock_switch(struct thread *, struct mtx *);
328 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
329 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
330 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
331     struct cpu_group *cg, int indent);
332 #endif
333 
334 static void sched_setup(void *dummy);
335 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
336 
337 static void sched_initticks(void *dummy);
338 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
339     NULL);
340 
341 SDT_PROVIDER_DEFINE(sched);
342 
343 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
344     "struct proc *", "uint8_t");
345 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
346     "struct proc *", "void *");
347 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
348     "struct proc *", "void *", "int");
349 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
350     "struct proc *", "uint8_t", "struct thread *");
351 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
352 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
353     "struct proc *");
354 SDT_PROBE_DEFINE(sched, , , on__cpu);
355 SDT_PROBE_DEFINE(sched, , , remain__cpu);
356 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
357     "struct proc *");
358 
359 #ifdef SMP
360 /*
361  * We need some randomness. Implement the classic Linear Congruential
362  * generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
363  * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
364  * of the random state (in the low bits of our answer) to return
365  * the maximum randomness.
366  */
367 static uint32_t
368 sched_random(void)
369 {
370 	uint32_t *rndptr;
371 
372 	rndptr = DPCPU_PTR(randomval);
373 	*rndptr = *rndptr * 69069 + 5;
374 
375 	return (*rndptr >> 16);
376 }
377 #endif
378 
379 /*
380  * Print the threads waiting on a run-queue.
381  */
382 static void
383 runq_print(struct runq *rq)
384 {
385 	struct rqhead *rqh;
386 	struct thread *td;
387 	int pri;
388 	int j;
389 	int i;
390 
391 	for (i = 0; i < RQB_LEN; i++) {
392 		printf("\t\trunq bits %d 0x%zx\n",
393 		    i, rq->rq_status.rqb_bits[i]);
394 		for (j = 0; j < RQB_BPW; j++)
395 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
396 				pri = j + (i << RQB_L2BPW);
397 				rqh = &rq->rq_queues[pri];
398 				TAILQ_FOREACH(td, rqh, td_runq) {
399 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
400 					    td, td->td_name, td->td_priority,
401 					    td->td_rqindex, pri);
402 				}
403 			}
404 	}
405 }
406 
407 /*
408  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
409  */
410 void
411 tdq_print(int cpu)
412 {
413 	struct tdq *tdq;
414 
415 	tdq = TDQ_CPU(cpu);
416 
417 	printf("tdq %d:\n", TDQ_ID(tdq));
418 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
419 	printf("\tLock name:      %s\n", tdq->tdq_name);
420 	printf("\tload:           %d\n", tdq->tdq_load);
421 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
422 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
423 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
424 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
425 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
426 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
427 	printf("\trealtime runq:\n");
428 	runq_print(&tdq->tdq_realtime);
429 	printf("\ttimeshare runq:\n");
430 	runq_print(&tdq->tdq_timeshare);
431 	printf("\tidle runq:\n");
432 	runq_print(&tdq->tdq_idle);
433 }
434 
435 static inline int
436 sched_shouldpreempt(int pri, int cpri, int remote)
437 {
438 	/*
439 	 * If the new priority is not better than the current priority there is
440 	 * nothing to do.
441 	 */
442 	if (pri >= cpri)
443 		return (0);
444 	/*
445 	 * Always preempt idle.
446 	 */
447 	if (cpri >= PRI_MIN_IDLE)
448 		return (1);
449 	/*
450 	 * If preemption is disabled don't preempt others.
451 	 */
452 	if (preempt_thresh == 0)
453 		return (0);
454 	/*
455 	 * Preempt if we exceed the threshold.
456 	 */
457 	if (pri <= preempt_thresh)
458 		return (1);
459 	/*
460 	 * If we're interactive or better and there is non-interactive
461 	 * or worse running preempt only remote processors.
462 	 */
463 	if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
464 		return (1);
465 	return (0);
466 }
467 
468 /*
469  * Add a thread to the actual run-queue.  Keeps transferable counts up to
470  * date with what is actually on the run-queue.  Selects the correct
471  * queue position for timeshare threads.
472  */
473 static __inline void
474 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
475 {
476 	struct td_sched *ts;
477 	u_char pri;
478 
479 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
480 	THREAD_LOCK_ASSERT(td, MA_OWNED);
481 
482 	pri = td->td_priority;
483 	ts = td->td_sched;
484 	TD_SET_RUNQ(td);
485 	if (THREAD_CAN_MIGRATE(td)) {
486 		tdq->tdq_transferable++;
487 		ts->ts_flags |= TSF_XFERABLE;
488 	}
489 	if (pri < PRI_MIN_BATCH) {
490 		ts->ts_runq = &tdq->tdq_realtime;
491 	} else if (pri <= PRI_MAX_BATCH) {
492 		ts->ts_runq = &tdq->tdq_timeshare;
493 		KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
494 			("Invalid priority %d on timeshare runq", pri));
495 		/*
496 		 * This queue contains only priorities between MIN and MAX
497 		 * realtime.  Use the whole queue to represent these values.
498 		 */
499 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
500 			pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
501 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
502 			/*
503 			 * This effectively shortens the queue by one so we
504 			 * can have a one slot difference between idx and
505 			 * ridx while we wait for threads to drain.
506 			 */
507 			if (tdq->tdq_ridx != tdq->tdq_idx &&
508 			    pri == tdq->tdq_ridx)
509 				pri = (unsigned char)(pri - 1) % RQ_NQS;
510 		} else
511 			pri = tdq->tdq_ridx;
512 		runq_add_pri(ts->ts_runq, td, pri, flags);
513 		return;
514 	} else
515 		ts->ts_runq = &tdq->tdq_idle;
516 	runq_add(ts->ts_runq, td, flags);
517 }
518 
519 /*
520  * Remove a thread from a run-queue.  This typically happens when a thread
521  * is selected to run.  Running threads are not on the queue and the
522  * transferable count does not reflect them.
523  */
524 static __inline void
525 tdq_runq_rem(struct tdq *tdq, struct thread *td)
526 {
527 	struct td_sched *ts;
528 
529 	ts = td->td_sched;
530 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
531 	KASSERT(ts->ts_runq != NULL,
532 	    ("tdq_runq_remove: thread %p null ts_runq", td));
533 	if (ts->ts_flags & TSF_XFERABLE) {
534 		tdq->tdq_transferable--;
535 		ts->ts_flags &= ~TSF_XFERABLE;
536 	}
537 	if (ts->ts_runq == &tdq->tdq_timeshare) {
538 		if (tdq->tdq_idx != tdq->tdq_ridx)
539 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
540 		else
541 			runq_remove_idx(ts->ts_runq, td, NULL);
542 	} else
543 		runq_remove(ts->ts_runq, td);
544 }
545 
546 /*
547  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
548  * for this thread to the referenced thread queue.
549  */
550 static void
551 tdq_load_add(struct tdq *tdq, struct thread *td)
552 {
553 
554 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
555 	THREAD_LOCK_ASSERT(td, MA_OWNED);
556 
557 	tdq->tdq_load++;
558 	if ((td->td_flags & TDF_NOLOAD) == 0)
559 		tdq->tdq_sysload++;
560 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
561 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
562 }
563 
564 /*
565  * Remove the load from a thread that is transitioning to a sleep state or
566  * exiting.
567  */
568 static void
569 tdq_load_rem(struct tdq *tdq, struct thread *td)
570 {
571 
572 	THREAD_LOCK_ASSERT(td, MA_OWNED);
573 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
574 	KASSERT(tdq->tdq_load != 0,
575 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
576 
577 	tdq->tdq_load--;
578 	if ((td->td_flags & TDF_NOLOAD) == 0)
579 		tdq->tdq_sysload--;
580 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
581 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
582 }
583 
584 /*
585  * Bound timeshare latency by decreasing slice size as load increases.  We
586  * consider the maximum latency as the sum of the threads waiting to run
587  * aside from curthread and target no more than sched_slice latency but
588  * no less than sched_slice_min runtime.
589  */
590 static inline int
591 tdq_slice(struct tdq *tdq)
592 {
593 	int load;
594 
595 	/*
596 	 * It is safe to use sys_load here because this is called from
597 	 * contexts where timeshare threads are running and so there
598 	 * cannot be higher priority load in the system.
599 	 */
600 	load = tdq->tdq_sysload - 1;
601 	if (load >= SCHED_SLICE_MIN_DIVISOR)
602 		return (sched_slice_min);
603 	if (load <= 1)
604 		return (sched_slice);
605 	return (sched_slice / load);
606 }
607 
608 /*
609  * Set lowpri to its exact value by searching the run-queue and
610  * evaluating curthread.  curthread may be passed as an optimization.
611  */
612 static void
613 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
614 {
615 	struct thread *td;
616 
617 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
618 	if (ctd == NULL)
619 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
620 	td = tdq_choose(tdq);
621 	if (td == NULL || td->td_priority > ctd->td_priority)
622 		tdq->tdq_lowpri = ctd->td_priority;
623 	else
624 		tdq->tdq_lowpri = td->td_priority;
625 }
626 
627 #ifdef SMP
628 struct cpu_search {
629 	cpuset_t cs_mask;
630 	u_int	cs_prefer;
631 	int	cs_pri;		/* Min priority for low. */
632 	int	cs_limit;	/* Max load for low, min load for high. */
633 	int	cs_cpu;
634 	int	cs_load;
635 };
636 
637 #define	CPU_SEARCH_LOWEST	0x1
638 #define	CPU_SEARCH_HIGHEST	0x2
639 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
640 
641 #define	CPUSET_FOREACH(cpu, mask)				\
642 	for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++)		\
643 		if (CPU_ISSET(cpu, &mask))
644 
645 static __always_inline int cpu_search(const struct cpu_group *cg,
646     struct cpu_search *low, struct cpu_search *high, const int match);
647 int __noinline cpu_search_lowest(const struct cpu_group *cg,
648     struct cpu_search *low);
649 int __noinline cpu_search_highest(const struct cpu_group *cg,
650     struct cpu_search *high);
651 int __noinline cpu_search_both(const struct cpu_group *cg,
652     struct cpu_search *low, struct cpu_search *high);
653 
654 /*
655  * Search the tree of cpu_groups for the lowest or highest loaded cpu
656  * according to the match argument.  This routine actually compares the
657  * load on all paths through the tree and finds the least loaded cpu on
658  * the least loaded path, which may differ from the least loaded cpu in
659  * the system.  This balances work among caches and busses.
660  *
661  * This inline is instantiated in three forms below using constants for the
662  * match argument.  It is reduced to the minimum set for each case.  It is
663  * also recursive to the depth of the tree.
664  */
665 static __always_inline int
666 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
667     struct cpu_search *high, const int match)
668 {
669 	struct cpu_search lgroup;
670 	struct cpu_search hgroup;
671 	cpuset_t cpumask;
672 	struct cpu_group *child;
673 	struct tdq *tdq;
674 	int cpu, i, hload, lload, load, total, rnd;
675 
676 	total = 0;
677 	cpumask = cg->cg_mask;
678 	if (match & CPU_SEARCH_LOWEST) {
679 		lload = INT_MAX;
680 		lgroup = *low;
681 	}
682 	if (match & CPU_SEARCH_HIGHEST) {
683 		hload = INT_MIN;
684 		hgroup = *high;
685 	}
686 
687 	/* Iterate through the child CPU groups and then remaining CPUs. */
688 	for (i = cg->cg_children, cpu = mp_maxid; ; ) {
689 		if (i == 0) {
690 #ifdef HAVE_INLINE_FFSL
691 			cpu = CPU_FFS(&cpumask) - 1;
692 #else
693 			while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
694 				cpu--;
695 #endif
696 			if (cpu < 0)
697 				break;
698 			child = NULL;
699 		} else
700 			child = &cg->cg_child[i - 1];
701 
702 		if (match & CPU_SEARCH_LOWEST)
703 			lgroup.cs_cpu = -1;
704 		if (match & CPU_SEARCH_HIGHEST)
705 			hgroup.cs_cpu = -1;
706 		if (child) {			/* Handle child CPU group. */
707 			CPU_NAND(&cpumask, &child->cg_mask);
708 			switch (match) {
709 			case CPU_SEARCH_LOWEST:
710 				load = cpu_search_lowest(child, &lgroup);
711 				break;
712 			case CPU_SEARCH_HIGHEST:
713 				load = cpu_search_highest(child, &hgroup);
714 				break;
715 			case CPU_SEARCH_BOTH:
716 				load = cpu_search_both(child, &lgroup, &hgroup);
717 				break;
718 			}
719 		} else {			/* Handle child CPU. */
720 			CPU_CLR(cpu, &cpumask);
721 			tdq = TDQ_CPU(cpu);
722 			load = tdq->tdq_load * 256;
723 			rnd = sched_random() % 32;
724 			if (match & CPU_SEARCH_LOWEST) {
725 				if (cpu == low->cs_prefer)
726 					load -= 64;
727 				/* If that CPU is allowed and get data. */
728 				if (tdq->tdq_lowpri > lgroup.cs_pri &&
729 				    tdq->tdq_load <= lgroup.cs_limit &&
730 				    CPU_ISSET(cpu, &lgroup.cs_mask)) {
731 					lgroup.cs_cpu = cpu;
732 					lgroup.cs_load = load - rnd;
733 				}
734 			}
735 			if (match & CPU_SEARCH_HIGHEST)
736 				if (tdq->tdq_load >= hgroup.cs_limit &&
737 				    tdq->tdq_transferable &&
738 				    CPU_ISSET(cpu, &hgroup.cs_mask)) {
739 					hgroup.cs_cpu = cpu;
740 					hgroup.cs_load = load - rnd;
741 				}
742 		}
743 		total += load;
744 
745 		/* We have info about child item. Compare it. */
746 		if (match & CPU_SEARCH_LOWEST) {
747 			if (lgroup.cs_cpu >= 0 &&
748 			    (load < lload ||
749 			     (load == lload && lgroup.cs_load < low->cs_load))) {
750 				lload = load;
751 				low->cs_cpu = lgroup.cs_cpu;
752 				low->cs_load = lgroup.cs_load;
753 			}
754 		}
755 		if (match & CPU_SEARCH_HIGHEST)
756 			if (hgroup.cs_cpu >= 0 &&
757 			    (load > hload ||
758 			     (load == hload && hgroup.cs_load > high->cs_load))) {
759 				hload = load;
760 				high->cs_cpu = hgroup.cs_cpu;
761 				high->cs_load = hgroup.cs_load;
762 			}
763 		if (child) {
764 			i--;
765 			if (i == 0 && CPU_EMPTY(&cpumask))
766 				break;
767 		}
768 #ifndef HAVE_INLINE_FFSL
769 		else
770 			cpu--;
771 #endif
772 	}
773 	return (total);
774 }
775 
776 /*
777  * cpu_search instantiations must pass constants to maintain the inline
778  * optimization.
779  */
780 int
781 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
782 {
783 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
784 }
785 
786 int
787 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
788 {
789 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
790 }
791 
792 int
793 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
794     struct cpu_search *high)
795 {
796 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
797 }
798 
799 /*
800  * Find the cpu with the least load via the least loaded path that has a
801  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
802  * acceptable.
803  */
804 static inline int
805 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
806     int prefer)
807 {
808 	struct cpu_search low;
809 
810 	low.cs_cpu = -1;
811 	low.cs_prefer = prefer;
812 	low.cs_mask = mask;
813 	low.cs_pri = pri;
814 	low.cs_limit = maxload;
815 	cpu_search_lowest(cg, &low);
816 	return low.cs_cpu;
817 }
818 
819 /*
820  * Find the cpu with the highest load via the highest loaded path.
821  */
822 static inline int
823 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
824 {
825 	struct cpu_search high;
826 
827 	high.cs_cpu = -1;
828 	high.cs_mask = mask;
829 	high.cs_limit = minload;
830 	cpu_search_highest(cg, &high);
831 	return high.cs_cpu;
832 }
833 
834 static void
835 sched_balance_group(struct cpu_group *cg)
836 {
837 	cpuset_t hmask, lmask;
838 	int high, low, anylow;
839 
840 	CPU_FILL(&hmask);
841 	for (;;) {
842 		high = sched_highest(cg, hmask, 1);
843 		/* Stop if there is no more CPU with transferrable threads. */
844 		if (high == -1)
845 			break;
846 		CPU_CLR(high, &hmask);
847 		CPU_COPY(&hmask, &lmask);
848 		/* Stop if there is no more CPU left for low. */
849 		if (CPU_EMPTY(&lmask))
850 			break;
851 		anylow = 1;
852 nextlow:
853 		low = sched_lowest(cg, lmask, -1,
854 		    TDQ_CPU(high)->tdq_load - 1, high);
855 		/* Stop if we looked well and found no less loaded CPU. */
856 		if (anylow && low == -1)
857 			break;
858 		/* Go to next high if we found no less loaded CPU. */
859 		if (low == -1)
860 			continue;
861 		/* Transfer thread from high to low. */
862 		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
863 			/* CPU that got thread can no longer be a donor. */
864 			CPU_CLR(low, &hmask);
865 		} else {
866 			/*
867 			 * If failed, then there is no threads on high
868 			 * that can run on this low. Drop low from low
869 			 * mask and look for different one.
870 			 */
871 			CPU_CLR(low, &lmask);
872 			anylow = 0;
873 			goto nextlow;
874 		}
875 	}
876 }
877 
878 static void
879 sched_balance(void)
880 {
881 	struct tdq *tdq;
882 
883 	if (smp_started == 0 || rebalance == 0)
884 		return;
885 
886 	balance_ticks = max(balance_interval / 2, 1) +
887 	    (sched_random() % balance_interval);
888 	tdq = TDQ_SELF();
889 	TDQ_UNLOCK(tdq);
890 	sched_balance_group(cpu_top);
891 	TDQ_LOCK(tdq);
892 }
893 
894 /*
895  * Lock two thread queues using their address to maintain lock order.
896  */
897 static void
898 tdq_lock_pair(struct tdq *one, struct tdq *two)
899 {
900 	if (one < two) {
901 		TDQ_LOCK(one);
902 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
903 	} else {
904 		TDQ_LOCK(two);
905 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
906 	}
907 }
908 
909 /*
910  * Unlock two thread queues.  Order is not important here.
911  */
912 static void
913 tdq_unlock_pair(struct tdq *one, struct tdq *two)
914 {
915 	TDQ_UNLOCK(one);
916 	TDQ_UNLOCK(two);
917 }
918 
919 /*
920  * Transfer load between two imbalanced thread queues.
921  */
922 static int
923 sched_balance_pair(struct tdq *high, struct tdq *low)
924 {
925 	int moved;
926 	int cpu;
927 
928 	tdq_lock_pair(high, low);
929 	moved = 0;
930 	/*
931 	 * Determine what the imbalance is and then adjust that to how many
932 	 * threads we actually have to give up (transferable).
933 	 */
934 	if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
935 	    (moved = tdq_move(high, low)) > 0) {
936 		/*
937 		 * In case the target isn't the current cpu IPI it to force a
938 		 * reschedule with the new workload.
939 		 */
940 		cpu = TDQ_ID(low);
941 		if (cpu != PCPU_GET(cpuid))
942 			ipi_cpu(cpu, IPI_PREEMPT);
943 	}
944 	tdq_unlock_pair(high, low);
945 	return (moved);
946 }
947 
948 /*
949  * Move a thread from one thread queue to another.
950  */
951 static int
952 tdq_move(struct tdq *from, struct tdq *to)
953 {
954 	struct td_sched *ts;
955 	struct thread *td;
956 	struct tdq *tdq;
957 	int cpu;
958 
959 	TDQ_LOCK_ASSERT(from, MA_OWNED);
960 	TDQ_LOCK_ASSERT(to, MA_OWNED);
961 
962 	tdq = from;
963 	cpu = TDQ_ID(to);
964 	td = tdq_steal(tdq, cpu);
965 	if (td == NULL)
966 		return (0);
967 	ts = td->td_sched;
968 	/*
969 	 * Although the run queue is locked the thread may be blocked.  Lock
970 	 * it to clear this and acquire the run-queue lock.
971 	 */
972 	thread_lock(td);
973 	/* Drop recursive lock on from acquired via thread_lock(). */
974 	TDQ_UNLOCK(from);
975 	sched_rem(td);
976 	ts->ts_cpu = cpu;
977 	td->td_lock = TDQ_LOCKPTR(to);
978 	tdq_add(to, td, SRQ_YIELDING);
979 	return (1);
980 }
981 
982 /*
983  * This tdq has idled.  Try to steal a thread from another cpu and switch
984  * to it.
985  */
986 static int
987 tdq_idled(struct tdq *tdq)
988 {
989 	struct cpu_group *cg;
990 	struct tdq *steal;
991 	cpuset_t mask;
992 	int thresh;
993 	int cpu;
994 
995 	if (smp_started == 0 || steal_idle == 0)
996 		return (1);
997 	CPU_FILL(&mask);
998 	CPU_CLR(PCPU_GET(cpuid), &mask);
999 	/* We don't want to be preempted while we're iterating. */
1000 	spinlock_enter();
1001 	for (cg = tdq->tdq_cg; cg != NULL; ) {
1002 		if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
1003 			thresh = steal_thresh;
1004 		else
1005 			thresh = 1;
1006 		cpu = sched_highest(cg, mask, thresh);
1007 		if (cpu == -1) {
1008 			cg = cg->cg_parent;
1009 			continue;
1010 		}
1011 		steal = TDQ_CPU(cpu);
1012 		CPU_CLR(cpu, &mask);
1013 		tdq_lock_pair(tdq, steal);
1014 		if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
1015 			tdq_unlock_pair(tdq, steal);
1016 			continue;
1017 		}
1018 		/*
1019 		 * If a thread was added while interrupts were disabled don't
1020 		 * steal one here.  If we fail to acquire one due to affinity
1021 		 * restrictions loop again with this cpu removed from the
1022 		 * set.
1023 		 */
1024 		if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
1025 			tdq_unlock_pair(tdq, steal);
1026 			continue;
1027 		}
1028 		spinlock_exit();
1029 		TDQ_UNLOCK(steal);
1030 		mi_switch(SW_VOL | SWT_IDLE, NULL);
1031 		thread_unlock(curthread);
1032 
1033 		return (0);
1034 	}
1035 	spinlock_exit();
1036 	return (1);
1037 }
1038 
1039 /*
1040  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
1041  */
1042 static void
1043 tdq_notify(struct tdq *tdq, struct thread *td)
1044 {
1045 	struct thread *ctd;
1046 	int pri;
1047 	int cpu;
1048 
1049 	if (tdq->tdq_ipipending)
1050 		return;
1051 	cpu = td->td_sched->ts_cpu;
1052 	pri = td->td_priority;
1053 	ctd = pcpu_find(cpu)->pc_curthread;
1054 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1055 		return;
1056 
1057 	/*
1058 	 * Make sure that our caller's earlier update to tdq_load is
1059 	 * globally visible before we read tdq_cpu_idle.  Idle thread
1060 	 * accesses both of them without locks, and the order is important.
1061 	 */
1062 	mb();
1063 
1064 	if (TD_IS_IDLETHREAD(ctd)) {
1065 		/*
1066 		 * If the MD code has an idle wakeup routine try that before
1067 		 * falling back to IPI.
1068 		 */
1069 		if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1070 			return;
1071 	}
1072 	tdq->tdq_ipipending = 1;
1073 	ipi_cpu(cpu, IPI_PREEMPT);
1074 }
1075 
1076 /*
1077  * Steals load from a timeshare queue.  Honors the rotating queue head
1078  * index.
1079  */
1080 static struct thread *
1081 runq_steal_from(struct runq *rq, int cpu, u_char start)
1082 {
1083 	struct rqbits *rqb;
1084 	struct rqhead *rqh;
1085 	struct thread *td, *first;
1086 	int bit;
1087 	int i;
1088 
1089 	rqb = &rq->rq_status;
1090 	bit = start & (RQB_BPW -1);
1091 	first = NULL;
1092 again:
1093 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1094 		if (rqb->rqb_bits[i] == 0)
1095 			continue;
1096 		if (bit == 0)
1097 			bit = RQB_FFS(rqb->rqb_bits[i]);
1098 		for (; bit < RQB_BPW; bit++) {
1099 			if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1100 				continue;
1101 			rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1102 			TAILQ_FOREACH(td, rqh, td_runq) {
1103 				if (first && THREAD_CAN_MIGRATE(td) &&
1104 				    THREAD_CAN_SCHED(td, cpu))
1105 					return (td);
1106 				first = td;
1107 			}
1108 		}
1109 	}
1110 	if (start != 0) {
1111 		start = 0;
1112 		goto again;
1113 	}
1114 
1115 	if (first && THREAD_CAN_MIGRATE(first) &&
1116 	    THREAD_CAN_SCHED(first, cpu))
1117 		return (first);
1118 	return (NULL);
1119 }
1120 
1121 /*
1122  * Steals load from a standard linear queue.
1123  */
1124 static struct thread *
1125 runq_steal(struct runq *rq, int cpu)
1126 {
1127 	struct rqhead *rqh;
1128 	struct rqbits *rqb;
1129 	struct thread *td;
1130 	int word;
1131 	int bit;
1132 
1133 	rqb = &rq->rq_status;
1134 	for (word = 0; word < RQB_LEN; word++) {
1135 		if (rqb->rqb_bits[word] == 0)
1136 			continue;
1137 		for (bit = 0; bit < RQB_BPW; bit++) {
1138 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1139 				continue;
1140 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1141 			TAILQ_FOREACH(td, rqh, td_runq)
1142 				if (THREAD_CAN_MIGRATE(td) &&
1143 				    THREAD_CAN_SCHED(td, cpu))
1144 					return (td);
1145 		}
1146 	}
1147 	return (NULL);
1148 }
1149 
1150 /*
1151  * Attempt to steal a thread in priority order from a thread queue.
1152  */
1153 static struct thread *
1154 tdq_steal(struct tdq *tdq, int cpu)
1155 {
1156 	struct thread *td;
1157 
1158 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1159 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1160 		return (td);
1161 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1162 	    cpu, tdq->tdq_ridx)) != NULL)
1163 		return (td);
1164 	return (runq_steal(&tdq->tdq_idle, cpu));
1165 }
1166 
1167 /*
1168  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1169  * current lock and returns with the assigned queue locked.
1170  */
1171 static inline struct tdq *
1172 sched_setcpu(struct thread *td, int cpu, int flags)
1173 {
1174 
1175 	struct tdq *tdq;
1176 
1177 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1178 	tdq = TDQ_CPU(cpu);
1179 	td->td_sched->ts_cpu = cpu;
1180 	/*
1181 	 * If the lock matches just return the queue.
1182 	 */
1183 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1184 		return (tdq);
1185 #ifdef notyet
1186 	/*
1187 	 * If the thread isn't running its lockptr is a
1188 	 * turnstile or a sleepqueue.  We can just lock_set without
1189 	 * blocking.
1190 	 */
1191 	if (TD_CAN_RUN(td)) {
1192 		TDQ_LOCK(tdq);
1193 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1194 		return (tdq);
1195 	}
1196 #endif
1197 	/*
1198 	 * The hard case, migration, we need to block the thread first to
1199 	 * prevent order reversals with other cpus locks.
1200 	 */
1201 	spinlock_enter();
1202 	thread_lock_block(td);
1203 	TDQ_LOCK(tdq);
1204 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1205 	spinlock_exit();
1206 	return (tdq);
1207 }
1208 
1209 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1210 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1211 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1212 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1213 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1214 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1215 
1216 static int
1217 sched_pickcpu(struct thread *td, int flags)
1218 {
1219 	struct cpu_group *cg, *ccg;
1220 	struct td_sched *ts;
1221 	struct tdq *tdq;
1222 	cpuset_t mask;
1223 	int cpu, pri, self;
1224 
1225 	self = PCPU_GET(cpuid);
1226 	ts = td->td_sched;
1227 	if (smp_started == 0)
1228 		return (self);
1229 	/*
1230 	 * Don't migrate a running thread from sched_switch().
1231 	 */
1232 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1233 		return (ts->ts_cpu);
1234 	/*
1235 	 * Prefer to run interrupt threads on the processors that generate
1236 	 * the interrupt.
1237 	 */
1238 	pri = td->td_priority;
1239 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1240 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1241 		SCHED_STAT_INC(pickcpu_intrbind);
1242 		ts->ts_cpu = self;
1243 		if (TDQ_CPU(self)->tdq_lowpri > pri) {
1244 			SCHED_STAT_INC(pickcpu_affinity);
1245 			return (ts->ts_cpu);
1246 		}
1247 	}
1248 	/*
1249 	 * If the thread can run on the last cpu and the affinity has not
1250 	 * expired or it is idle run it there.
1251 	 */
1252 	tdq = TDQ_CPU(ts->ts_cpu);
1253 	cg = tdq->tdq_cg;
1254 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1255 	    tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1256 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1257 		if (cg->cg_flags & CG_FLAG_THREAD) {
1258 			CPUSET_FOREACH(cpu, cg->cg_mask) {
1259 				if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1260 					break;
1261 			}
1262 		} else
1263 			cpu = INT_MAX;
1264 		if (cpu > mp_maxid) {
1265 			SCHED_STAT_INC(pickcpu_idle_affinity);
1266 			return (ts->ts_cpu);
1267 		}
1268 	}
1269 	/*
1270 	 * Search for the last level cache CPU group in the tree.
1271 	 * Skip caches with expired affinity time and SMT groups.
1272 	 * Affinity to higher level caches will be handled less aggressively.
1273 	 */
1274 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1275 		if (cg->cg_flags & CG_FLAG_THREAD)
1276 			continue;
1277 		if (!SCHED_AFFINITY(ts, cg->cg_level))
1278 			continue;
1279 		ccg = cg;
1280 	}
1281 	if (ccg != NULL)
1282 		cg = ccg;
1283 	cpu = -1;
1284 	/* Search the group for the less loaded idle CPU we can run now. */
1285 	mask = td->td_cpuset->cs_mask;
1286 	if (cg != NULL && cg != cpu_top &&
1287 	    CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1288 		cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1289 		    INT_MAX, ts->ts_cpu);
1290 	/* Search globally for the less loaded CPU we can run now. */
1291 	if (cpu == -1)
1292 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1293 	/* Search globally for the less loaded CPU. */
1294 	if (cpu == -1)
1295 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1296 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1297 	/*
1298 	 * Compare the lowest loaded cpu to current cpu.
1299 	 */
1300 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1301 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1302 	    TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1303 		SCHED_STAT_INC(pickcpu_local);
1304 		cpu = self;
1305 	} else
1306 		SCHED_STAT_INC(pickcpu_lowest);
1307 	if (cpu != ts->ts_cpu)
1308 		SCHED_STAT_INC(pickcpu_migration);
1309 	return (cpu);
1310 }
1311 #endif
1312 
1313 /*
1314  * Pick the highest priority task we have and return it.
1315  */
1316 static struct thread *
1317 tdq_choose(struct tdq *tdq)
1318 {
1319 	struct thread *td;
1320 
1321 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1322 	td = runq_choose(&tdq->tdq_realtime);
1323 	if (td != NULL)
1324 		return (td);
1325 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1326 	if (td != NULL) {
1327 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1328 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1329 		    td->td_priority));
1330 		return (td);
1331 	}
1332 	td = runq_choose(&tdq->tdq_idle);
1333 	if (td != NULL) {
1334 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1335 		    ("tdq_choose: Invalid priority on idle queue %d",
1336 		    td->td_priority));
1337 		return (td);
1338 	}
1339 
1340 	return (NULL);
1341 }
1342 
1343 /*
1344  * Initialize a thread queue.
1345  */
1346 static void
1347 tdq_setup(struct tdq *tdq)
1348 {
1349 
1350 	if (bootverbose)
1351 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1352 	runq_init(&tdq->tdq_realtime);
1353 	runq_init(&tdq->tdq_timeshare);
1354 	runq_init(&tdq->tdq_idle);
1355 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1356 	    "sched lock %d", (int)TDQ_ID(tdq));
1357 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1358 	    MTX_SPIN | MTX_RECURSE);
1359 #ifdef KTR
1360 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1361 	    "CPU %d load", (int)TDQ_ID(tdq));
1362 #endif
1363 }
1364 
1365 #ifdef SMP
1366 static void
1367 sched_setup_smp(void)
1368 {
1369 	struct tdq *tdq;
1370 	int i;
1371 
1372 	cpu_top = smp_topo();
1373 	CPU_FOREACH(i) {
1374 		tdq = TDQ_CPU(i);
1375 		tdq_setup(tdq);
1376 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1377 		if (tdq->tdq_cg == NULL)
1378 			panic("Can't find cpu group for %d\n", i);
1379 	}
1380 	balance_tdq = TDQ_SELF();
1381 	sched_balance();
1382 }
1383 #endif
1384 
1385 /*
1386  * Setup the thread queues and initialize the topology based on MD
1387  * information.
1388  */
1389 static void
1390 sched_setup(void *dummy)
1391 {
1392 	struct tdq *tdq;
1393 
1394 	tdq = TDQ_SELF();
1395 #ifdef SMP
1396 	sched_setup_smp();
1397 #else
1398 	tdq_setup(tdq);
1399 #endif
1400 
1401 	/* Add thread0's load since it's running. */
1402 	TDQ_LOCK(tdq);
1403 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1404 	tdq_load_add(tdq, &thread0);
1405 	tdq->tdq_lowpri = thread0.td_priority;
1406 	TDQ_UNLOCK(tdq);
1407 }
1408 
1409 /*
1410  * This routine determines time constants after stathz and hz are setup.
1411  */
1412 /* ARGSUSED */
1413 static void
1414 sched_initticks(void *dummy)
1415 {
1416 	int incr;
1417 
1418 	realstathz = stathz ? stathz : hz;
1419 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1420 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1421 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1422 	    realstathz);
1423 
1424 	/*
1425 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1426 	 * hz not being evenly divisible by stathz on all platforms.
1427 	 */
1428 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1429 	/*
1430 	 * This does not work for values of stathz that are more than
1431 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1432 	 */
1433 	if (incr == 0)
1434 		incr = 1;
1435 	tickincr = incr;
1436 #ifdef SMP
1437 	/*
1438 	 * Set the default balance interval now that we know
1439 	 * what realstathz is.
1440 	 */
1441 	balance_interval = realstathz;
1442 	affinity = SCHED_AFFINITY_DEFAULT;
1443 #endif
1444 	if (sched_idlespinthresh < 0)
1445 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1446 }
1447 
1448 
1449 /*
1450  * This is the core of the interactivity algorithm.  Determines a score based
1451  * on past behavior.  It is the ratio of sleep time to run time scaled to
1452  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1453  * differs from the cpu usage because it does not account for time spent
1454  * waiting on a run-queue.  Would be prettier if we had floating point.
1455  */
1456 static int
1457 sched_interact_score(struct thread *td)
1458 {
1459 	struct td_sched *ts;
1460 	int div;
1461 
1462 	ts = td->td_sched;
1463 	/*
1464 	 * The score is only needed if this is likely to be an interactive
1465 	 * task.  Don't go through the expense of computing it if there's
1466 	 * no chance.
1467 	 */
1468 	if (sched_interact <= SCHED_INTERACT_HALF &&
1469 		ts->ts_runtime >= ts->ts_slptime)
1470 			return (SCHED_INTERACT_HALF);
1471 
1472 	if (ts->ts_runtime > ts->ts_slptime) {
1473 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1474 		return (SCHED_INTERACT_HALF +
1475 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1476 	}
1477 	if (ts->ts_slptime > ts->ts_runtime) {
1478 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1479 		return (ts->ts_runtime / div);
1480 	}
1481 	/* runtime == slptime */
1482 	if (ts->ts_runtime)
1483 		return (SCHED_INTERACT_HALF);
1484 
1485 	/*
1486 	 * This can happen if slptime and runtime are 0.
1487 	 */
1488 	return (0);
1489 
1490 }
1491 
1492 /*
1493  * Scale the scheduling priority according to the "interactivity" of this
1494  * process.
1495  */
1496 static void
1497 sched_priority(struct thread *td)
1498 {
1499 	int score;
1500 	int pri;
1501 
1502 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1503 		return;
1504 	/*
1505 	 * If the score is interactive we place the thread in the realtime
1506 	 * queue with a priority that is less than kernel and interrupt
1507 	 * priorities.  These threads are not subject to nice restrictions.
1508 	 *
1509 	 * Scores greater than this are placed on the normal timeshare queue
1510 	 * where the priority is partially decided by the most recent cpu
1511 	 * utilization and the rest is decided by nice value.
1512 	 *
1513 	 * The nice value of the process has a linear effect on the calculated
1514 	 * score.  Negative nice values make it easier for a thread to be
1515 	 * considered interactive.
1516 	 */
1517 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1518 	if (score < sched_interact) {
1519 		pri = PRI_MIN_INTERACT;
1520 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1521 		    sched_interact) * score;
1522 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1523 		    ("sched_priority: invalid interactive priority %d score %d",
1524 		    pri, score));
1525 	} else {
1526 		pri = SCHED_PRI_MIN;
1527 		if (td->td_sched->ts_ticks)
1528 			pri += min(SCHED_PRI_TICKS(td->td_sched),
1529 			    SCHED_PRI_RANGE - 1);
1530 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1531 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1532 		    ("sched_priority: invalid priority %d: nice %d, "
1533 		    "ticks %d ftick %d ltick %d tick pri %d",
1534 		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1535 		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1536 		    SCHED_PRI_TICKS(td->td_sched)));
1537 	}
1538 	sched_user_prio(td, pri);
1539 
1540 	return;
1541 }
1542 
1543 /*
1544  * This routine enforces a maximum limit on the amount of scheduling history
1545  * kept.  It is called after either the slptime or runtime is adjusted.  This
1546  * function is ugly due to integer math.
1547  */
1548 static void
1549 sched_interact_update(struct thread *td)
1550 {
1551 	struct td_sched *ts;
1552 	u_int sum;
1553 
1554 	ts = td->td_sched;
1555 	sum = ts->ts_runtime + ts->ts_slptime;
1556 	if (sum < SCHED_SLP_RUN_MAX)
1557 		return;
1558 	/*
1559 	 * This only happens from two places:
1560 	 * 1) We have added an unusual amount of run time from fork_exit.
1561 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1562 	 */
1563 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1564 		if (ts->ts_runtime > ts->ts_slptime) {
1565 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1566 			ts->ts_slptime = 1;
1567 		} else {
1568 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1569 			ts->ts_runtime = 1;
1570 		}
1571 		return;
1572 	}
1573 	/*
1574 	 * If we have exceeded by more than 1/5th then the algorithm below
1575 	 * will not bring us back into range.  Dividing by two here forces
1576 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1577 	 */
1578 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1579 		ts->ts_runtime /= 2;
1580 		ts->ts_slptime /= 2;
1581 		return;
1582 	}
1583 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1584 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1585 }
1586 
1587 /*
1588  * Scale back the interactivity history when a child thread is created.  The
1589  * history is inherited from the parent but the thread may behave totally
1590  * differently.  For example, a shell spawning a compiler process.  We want
1591  * to learn that the compiler is behaving badly very quickly.
1592  */
1593 static void
1594 sched_interact_fork(struct thread *td)
1595 {
1596 	int ratio;
1597 	int sum;
1598 
1599 	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1600 	if (sum > SCHED_SLP_RUN_FORK) {
1601 		ratio = sum / SCHED_SLP_RUN_FORK;
1602 		td->td_sched->ts_runtime /= ratio;
1603 		td->td_sched->ts_slptime /= ratio;
1604 	}
1605 }
1606 
1607 /*
1608  * Called from proc0_init() to setup the scheduler fields.
1609  */
1610 void
1611 schedinit(void)
1612 {
1613 
1614 	/*
1615 	 * Set up the scheduler specific parts of proc0.
1616 	 */
1617 	proc0.p_sched = NULL; /* XXX */
1618 	thread0.td_sched = &td_sched0;
1619 	td_sched0.ts_ltick = ticks;
1620 	td_sched0.ts_ftick = ticks;
1621 	td_sched0.ts_slice = 0;
1622 }
1623 
1624 /*
1625  * This is only somewhat accurate since given many processes of the same
1626  * priority they will switch when their slices run out, which will be
1627  * at most sched_slice stathz ticks.
1628  */
1629 int
1630 sched_rr_interval(void)
1631 {
1632 
1633 	/* Convert sched_slice from stathz to hz. */
1634 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1635 }
1636 
1637 /*
1638  * Update the percent cpu tracking information when it is requested or
1639  * the total history exceeds the maximum.  We keep a sliding history of
1640  * tick counts that slowly decays.  This is less precise than the 4BSD
1641  * mechanism since it happens with less regular and frequent events.
1642  */
1643 static void
1644 sched_pctcpu_update(struct td_sched *ts, int run)
1645 {
1646 	int t = ticks;
1647 
1648 	if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1649 		ts->ts_ticks = 0;
1650 		ts->ts_ftick = t - SCHED_TICK_TARG;
1651 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1652 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1653 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1654 		ts->ts_ftick = t - SCHED_TICK_TARG;
1655 	}
1656 	if (run)
1657 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1658 	ts->ts_ltick = t;
1659 }
1660 
1661 /*
1662  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1663  * if necessary.  This is the back-end for several priority related
1664  * functions.
1665  */
1666 static void
1667 sched_thread_priority(struct thread *td, u_char prio)
1668 {
1669 	struct td_sched *ts;
1670 	struct tdq *tdq;
1671 	int oldpri;
1672 
1673 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1674 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1675 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1676 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1677 	if (td != curthread && prio < td->td_priority) {
1678 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1679 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1680 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1681 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1682 		    curthread);
1683 	}
1684 	ts = td->td_sched;
1685 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1686 	if (td->td_priority == prio)
1687 		return;
1688 	/*
1689 	 * If the priority has been elevated due to priority
1690 	 * propagation, we may have to move ourselves to a new
1691 	 * queue.  This could be optimized to not re-add in some
1692 	 * cases.
1693 	 */
1694 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1695 		sched_rem(td);
1696 		td->td_priority = prio;
1697 		sched_add(td, SRQ_BORROWING);
1698 		return;
1699 	}
1700 	/*
1701 	 * If the thread is currently running we may have to adjust the lowpri
1702 	 * information so other cpus are aware of our current priority.
1703 	 */
1704 	if (TD_IS_RUNNING(td)) {
1705 		tdq = TDQ_CPU(ts->ts_cpu);
1706 		oldpri = td->td_priority;
1707 		td->td_priority = prio;
1708 		if (prio < tdq->tdq_lowpri)
1709 			tdq->tdq_lowpri = prio;
1710 		else if (tdq->tdq_lowpri == oldpri)
1711 			tdq_setlowpri(tdq, td);
1712 		return;
1713 	}
1714 	td->td_priority = prio;
1715 }
1716 
1717 /*
1718  * Update a thread's priority when it is lent another thread's
1719  * priority.
1720  */
1721 void
1722 sched_lend_prio(struct thread *td, u_char prio)
1723 {
1724 
1725 	td->td_flags |= TDF_BORROWING;
1726 	sched_thread_priority(td, prio);
1727 }
1728 
1729 /*
1730  * Restore a thread's priority when priority propagation is
1731  * over.  The prio argument is the minimum priority the thread
1732  * needs to have to satisfy other possible priority lending
1733  * requests.  If the thread's regular priority is less
1734  * important than prio, the thread will keep a priority boost
1735  * of prio.
1736  */
1737 void
1738 sched_unlend_prio(struct thread *td, u_char prio)
1739 {
1740 	u_char base_pri;
1741 
1742 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1743 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1744 		base_pri = td->td_user_pri;
1745 	else
1746 		base_pri = td->td_base_pri;
1747 	if (prio >= base_pri) {
1748 		td->td_flags &= ~TDF_BORROWING;
1749 		sched_thread_priority(td, base_pri);
1750 	} else
1751 		sched_lend_prio(td, prio);
1752 }
1753 
1754 /*
1755  * Standard entry for setting the priority to an absolute value.
1756  */
1757 void
1758 sched_prio(struct thread *td, u_char prio)
1759 {
1760 	u_char oldprio;
1761 
1762 	/* First, update the base priority. */
1763 	td->td_base_pri = prio;
1764 
1765 	/*
1766 	 * If the thread is borrowing another thread's priority, don't
1767 	 * ever lower the priority.
1768 	 */
1769 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1770 		return;
1771 
1772 	/* Change the real priority. */
1773 	oldprio = td->td_priority;
1774 	sched_thread_priority(td, prio);
1775 
1776 	/*
1777 	 * If the thread is on a turnstile, then let the turnstile update
1778 	 * its state.
1779 	 */
1780 	if (TD_ON_LOCK(td) && oldprio != prio)
1781 		turnstile_adjust(td, oldprio);
1782 }
1783 
1784 /*
1785  * Set the base user priority, does not effect current running priority.
1786  */
1787 void
1788 sched_user_prio(struct thread *td, u_char prio)
1789 {
1790 
1791 	td->td_base_user_pri = prio;
1792 	if (td->td_lend_user_pri <= prio)
1793 		return;
1794 	td->td_user_pri = prio;
1795 }
1796 
1797 void
1798 sched_lend_user_prio(struct thread *td, u_char prio)
1799 {
1800 
1801 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1802 	td->td_lend_user_pri = prio;
1803 	td->td_user_pri = min(prio, td->td_base_user_pri);
1804 	if (td->td_priority > td->td_user_pri)
1805 		sched_prio(td, td->td_user_pri);
1806 	else if (td->td_priority != td->td_user_pri)
1807 		td->td_flags |= TDF_NEEDRESCHED;
1808 }
1809 
1810 /*
1811  * Handle migration from sched_switch().  This happens only for
1812  * cpu binding.
1813  */
1814 static struct mtx *
1815 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1816 {
1817 	struct tdq *tdn;
1818 
1819 	tdn = TDQ_CPU(td->td_sched->ts_cpu);
1820 #ifdef SMP
1821 	tdq_load_rem(tdq, td);
1822 	/*
1823 	 * Do the lock dance required to avoid LOR.  We grab an extra
1824 	 * spinlock nesting to prevent preemption while we're
1825 	 * not holding either run-queue lock.
1826 	 */
1827 	spinlock_enter();
1828 	thread_lock_block(td);	/* This releases the lock on tdq. */
1829 
1830 	/*
1831 	 * Acquire both run-queue locks before placing the thread on the new
1832 	 * run-queue to avoid deadlocks created by placing a thread with a
1833 	 * blocked lock on the run-queue of a remote processor.  The deadlock
1834 	 * occurs when a third processor attempts to lock the two queues in
1835 	 * question while the target processor is spinning with its own
1836 	 * run-queue lock held while waiting for the blocked lock to clear.
1837 	 */
1838 	tdq_lock_pair(tdn, tdq);
1839 	tdq_add(tdn, td, flags);
1840 	tdq_notify(tdn, td);
1841 	TDQ_UNLOCK(tdn);
1842 	spinlock_exit();
1843 #endif
1844 	return (TDQ_LOCKPTR(tdn));
1845 }
1846 
1847 /*
1848  * Variadic version of thread_lock_unblock() that does not assume td_lock
1849  * is blocked.
1850  */
1851 static inline void
1852 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1853 {
1854 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1855 	    (uintptr_t)mtx);
1856 }
1857 
1858 /*
1859  * Switch threads.  This function has to handle threads coming in while
1860  * blocked for some reason, running, or idle.  It also must deal with
1861  * migrating a thread from one queue to another as running threads may
1862  * be assigned elsewhere via binding.
1863  */
1864 void
1865 sched_switch(struct thread *td, struct thread *newtd, int flags)
1866 {
1867 	struct tdq *tdq;
1868 	struct td_sched *ts;
1869 	struct mtx *mtx;
1870 	int srqflag;
1871 	int cpuid, preempted;
1872 
1873 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1874 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1875 
1876 	cpuid = PCPU_GET(cpuid);
1877 	tdq = TDQ_CPU(cpuid);
1878 	ts = td->td_sched;
1879 	mtx = td->td_lock;
1880 	sched_pctcpu_update(ts, 1);
1881 	ts->ts_rltick = ticks;
1882 	td->td_lastcpu = td->td_oncpu;
1883 	td->td_oncpu = NOCPU;
1884 	preempted = !((td->td_flags & TDF_SLICEEND) ||
1885 	    (flags & SWT_RELINQUISH));
1886 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1887 	td->td_owepreempt = 0;
1888 	if (!TD_IS_IDLETHREAD(td))
1889 		tdq->tdq_switchcnt++;
1890 	/*
1891 	 * The lock pointer in an idle thread should never change.  Reset it
1892 	 * to CAN_RUN as well.
1893 	 */
1894 	if (TD_IS_IDLETHREAD(td)) {
1895 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1896 		TD_SET_CAN_RUN(td);
1897 	} else if (TD_IS_RUNNING(td)) {
1898 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1899 		srqflag = preempted ?
1900 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1901 		    SRQ_OURSELF|SRQ_YIELDING;
1902 #ifdef SMP
1903 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1904 			ts->ts_cpu = sched_pickcpu(td, 0);
1905 #endif
1906 		if (ts->ts_cpu == cpuid)
1907 			tdq_runq_add(tdq, td, srqflag);
1908 		else {
1909 			KASSERT(THREAD_CAN_MIGRATE(td) ||
1910 			    (ts->ts_flags & TSF_BOUND) != 0,
1911 			    ("Thread %p shouldn't migrate", td));
1912 			mtx = sched_switch_migrate(tdq, td, srqflag);
1913 		}
1914 	} else {
1915 		/* This thread must be going to sleep. */
1916 		TDQ_LOCK(tdq);
1917 		mtx = thread_lock_block(td);
1918 		tdq_load_rem(tdq, td);
1919 	}
1920 	/*
1921 	 * We enter here with the thread blocked and assigned to the
1922 	 * appropriate cpu run-queue or sleep-queue and with the current
1923 	 * thread-queue locked.
1924 	 */
1925 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1926 	newtd = choosethread();
1927 	/*
1928 	 * Call the MD code to switch contexts if necessary.
1929 	 */
1930 	if (td != newtd) {
1931 #ifdef	HWPMC_HOOKS
1932 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1933 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1934 #endif
1935 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1936 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1937 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1938 		sched_pctcpu_update(newtd->td_sched, 0);
1939 
1940 #ifdef KDTRACE_HOOKS
1941 		/*
1942 		 * If DTrace has set the active vtime enum to anything
1943 		 * other than INACTIVE (0), then it should have set the
1944 		 * function to call.
1945 		 */
1946 		if (dtrace_vtime_active)
1947 			(*dtrace_vtime_switch_func)(newtd);
1948 #endif
1949 
1950 		cpu_switch(td, newtd, mtx);
1951 		/*
1952 		 * We may return from cpu_switch on a different cpu.  However,
1953 		 * we always return with td_lock pointing to the current cpu's
1954 		 * run queue lock.
1955 		 */
1956 		cpuid = PCPU_GET(cpuid);
1957 		tdq = TDQ_CPU(cpuid);
1958 		lock_profile_obtain_lock_success(
1959 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1960 
1961 		SDT_PROBE0(sched, , , on__cpu);
1962 #ifdef	HWPMC_HOOKS
1963 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1964 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1965 #endif
1966 	} else {
1967 		thread_unblock_switch(td, mtx);
1968 		SDT_PROBE0(sched, , , remain__cpu);
1969 	}
1970 	/*
1971 	 * Assert that all went well and return.
1972 	 */
1973 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1974 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1975 	td->td_oncpu = cpuid;
1976 }
1977 
1978 /*
1979  * Adjust thread priorities as a result of a nice request.
1980  */
1981 void
1982 sched_nice(struct proc *p, int nice)
1983 {
1984 	struct thread *td;
1985 
1986 	PROC_LOCK_ASSERT(p, MA_OWNED);
1987 
1988 	p->p_nice = nice;
1989 	FOREACH_THREAD_IN_PROC(p, td) {
1990 		thread_lock(td);
1991 		sched_priority(td);
1992 		sched_prio(td, td->td_base_user_pri);
1993 		thread_unlock(td);
1994 	}
1995 }
1996 
1997 /*
1998  * Record the sleep time for the interactivity scorer.
1999  */
2000 void
2001 sched_sleep(struct thread *td, int prio)
2002 {
2003 
2004 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2005 
2006 	td->td_slptick = ticks;
2007 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2008 		td->td_flags |= TDF_CANSWAP;
2009 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2010 		return;
2011 	if (static_boost == 1 && prio)
2012 		sched_prio(td, prio);
2013 	else if (static_boost && td->td_priority > static_boost)
2014 		sched_prio(td, static_boost);
2015 }
2016 
2017 /*
2018  * Schedule a thread to resume execution and record how long it voluntarily
2019  * slept.  We also update the pctcpu, interactivity, and priority.
2020  */
2021 void
2022 sched_wakeup(struct thread *td)
2023 {
2024 	struct td_sched *ts;
2025 	int slptick;
2026 
2027 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2028 	ts = td->td_sched;
2029 	td->td_flags &= ~TDF_CANSWAP;
2030 	/*
2031 	 * If we slept for more than a tick update our interactivity and
2032 	 * priority.
2033 	 */
2034 	slptick = td->td_slptick;
2035 	td->td_slptick = 0;
2036 	if (slptick && slptick != ticks) {
2037 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2038 		sched_interact_update(td);
2039 		sched_pctcpu_update(ts, 0);
2040 	}
2041 	/*
2042 	 * Reset the slice value since we slept and advanced the round-robin.
2043 	 */
2044 	ts->ts_slice = 0;
2045 	sched_add(td, SRQ_BORING);
2046 }
2047 
2048 /*
2049  * Penalize the parent for creating a new child and initialize the child's
2050  * priority.
2051  */
2052 void
2053 sched_fork(struct thread *td, struct thread *child)
2054 {
2055 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2056 	sched_pctcpu_update(td->td_sched, 1);
2057 	sched_fork_thread(td, child);
2058 	/*
2059 	 * Penalize the parent and child for forking.
2060 	 */
2061 	sched_interact_fork(child);
2062 	sched_priority(child);
2063 	td->td_sched->ts_runtime += tickincr;
2064 	sched_interact_update(td);
2065 	sched_priority(td);
2066 }
2067 
2068 /*
2069  * Fork a new thread, may be within the same process.
2070  */
2071 void
2072 sched_fork_thread(struct thread *td, struct thread *child)
2073 {
2074 	struct td_sched *ts;
2075 	struct td_sched *ts2;
2076 	struct tdq *tdq;
2077 
2078 	tdq = TDQ_SELF();
2079 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2080 	/*
2081 	 * Initialize child.
2082 	 */
2083 	ts = td->td_sched;
2084 	ts2 = child->td_sched;
2085 	child->td_lock = TDQ_LOCKPTR(tdq);
2086 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2087 	ts2->ts_cpu = ts->ts_cpu;
2088 	ts2->ts_flags = 0;
2089 	/*
2090 	 * Grab our parents cpu estimation information.
2091 	 */
2092 	ts2->ts_ticks = ts->ts_ticks;
2093 	ts2->ts_ltick = ts->ts_ltick;
2094 	ts2->ts_ftick = ts->ts_ftick;
2095 	/*
2096 	 * Do not inherit any borrowed priority from the parent.
2097 	 */
2098 	child->td_priority = child->td_base_pri;
2099 	/*
2100 	 * And update interactivity score.
2101 	 */
2102 	ts2->ts_slptime = ts->ts_slptime;
2103 	ts2->ts_runtime = ts->ts_runtime;
2104 	/* Attempt to quickly learn interactivity. */
2105 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2106 #ifdef KTR
2107 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2108 #endif
2109 }
2110 
2111 /*
2112  * Adjust the priority class of a thread.
2113  */
2114 void
2115 sched_class(struct thread *td, int class)
2116 {
2117 
2118 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2119 	if (td->td_pri_class == class)
2120 		return;
2121 	td->td_pri_class = class;
2122 }
2123 
2124 /*
2125  * Return some of the child's priority and interactivity to the parent.
2126  */
2127 void
2128 sched_exit(struct proc *p, struct thread *child)
2129 {
2130 	struct thread *td;
2131 
2132 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2133 	    "prio:%d", child->td_priority);
2134 	PROC_LOCK_ASSERT(p, MA_OWNED);
2135 	td = FIRST_THREAD_IN_PROC(p);
2136 	sched_exit_thread(td, child);
2137 }
2138 
2139 /*
2140  * Penalize another thread for the time spent on this one.  This helps to
2141  * worsen the priority and interactivity of processes which schedule batch
2142  * jobs such as make.  This has little effect on the make process itself but
2143  * causes new processes spawned by it to receive worse scores immediately.
2144  */
2145 void
2146 sched_exit_thread(struct thread *td, struct thread *child)
2147 {
2148 
2149 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2150 	    "prio:%d", child->td_priority);
2151 	/*
2152 	 * Give the child's runtime to the parent without returning the
2153 	 * sleep time as a penalty to the parent.  This causes shells that
2154 	 * launch expensive things to mark their children as expensive.
2155 	 */
2156 	thread_lock(td);
2157 	td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2158 	sched_interact_update(td);
2159 	sched_priority(td);
2160 	thread_unlock(td);
2161 }
2162 
2163 void
2164 sched_preempt(struct thread *td)
2165 {
2166 	struct tdq *tdq;
2167 
2168 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2169 
2170 	thread_lock(td);
2171 	tdq = TDQ_SELF();
2172 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2173 	tdq->tdq_ipipending = 0;
2174 	if (td->td_priority > tdq->tdq_lowpri) {
2175 		int flags;
2176 
2177 		flags = SW_INVOL | SW_PREEMPT;
2178 		if (td->td_critnest > 1)
2179 			td->td_owepreempt = 1;
2180 		else if (TD_IS_IDLETHREAD(td))
2181 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2182 		else
2183 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2184 	}
2185 	thread_unlock(td);
2186 }
2187 
2188 /*
2189  * Fix priorities on return to user-space.  Priorities may be elevated due
2190  * to static priorities in msleep() or similar.
2191  */
2192 void
2193 sched_userret(struct thread *td)
2194 {
2195 	/*
2196 	 * XXX we cheat slightly on the locking here to avoid locking in
2197 	 * the usual case.  Setting td_priority here is essentially an
2198 	 * incomplete workaround for not setting it properly elsewhere.
2199 	 * Now that some interrupt handlers are threads, not setting it
2200 	 * properly elsewhere can clobber it in the window between setting
2201 	 * it here and returning to user mode, so don't waste time setting
2202 	 * it perfectly here.
2203 	 */
2204 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2205 	    ("thread with borrowed priority returning to userland"));
2206 	if (td->td_priority != td->td_user_pri) {
2207 		thread_lock(td);
2208 		td->td_priority = td->td_user_pri;
2209 		td->td_base_pri = td->td_user_pri;
2210 		tdq_setlowpri(TDQ_SELF(), td);
2211 		thread_unlock(td);
2212         }
2213 }
2214 
2215 /*
2216  * Handle a stathz tick.  This is really only relevant for timeshare
2217  * threads.
2218  */
2219 void
2220 sched_clock(struct thread *td)
2221 {
2222 	struct tdq *tdq;
2223 	struct td_sched *ts;
2224 
2225 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2226 	tdq = TDQ_SELF();
2227 #ifdef SMP
2228 	/*
2229 	 * We run the long term load balancer infrequently on the first cpu.
2230 	 */
2231 	if (balance_tdq == tdq) {
2232 		if (balance_ticks && --balance_ticks == 0)
2233 			sched_balance();
2234 	}
2235 #endif
2236 	/*
2237 	 * Save the old switch count so we have a record of the last ticks
2238 	 * activity.   Initialize the new switch count based on our load.
2239 	 * If there is some activity seed it to reflect that.
2240 	 */
2241 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2242 	tdq->tdq_switchcnt = tdq->tdq_load;
2243 	/*
2244 	 * Advance the insert index once for each tick to ensure that all
2245 	 * threads get a chance to run.
2246 	 */
2247 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2248 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2249 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2250 			tdq->tdq_ridx = tdq->tdq_idx;
2251 	}
2252 	ts = td->td_sched;
2253 	sched_pctcpu_update(ts, 1);
2254 	if (td->td_pri_class & PRI_FIFO_BIT)
2255 		return;
2256 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2257 		/*
2258 		 * We used a tick; charge it to the thread so
2259 		 * that we can compute our interactivity.
2260 		 */
2261 		td->td_sched->ts_runtime += tickincr;
2262 		sched_interact_update(td);
2263 		sched_priority(td);
2264 	}
2265 
2266 	/*
2267 	 * Force a context switch if the current thread has used up a full
2268 	 * time slice (default is 100ms).
2269 	 */
2270 	if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2271 		ts->ts_slice = 0;
2272 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2273 	}
2274 }
2275 
2276 /*
2277  * Called once per hz tick.
2278  */
2279 void
2280 sched_tick(int cnt)
2281 {
2282 
2283 }
2284 
2285 /*
2286  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2287  * cooperative idle threads.
2288  */
2289 int
2290 sched_runnable(void)
2291 {
2292 	struct tdq *tdq;
2293 	int load;
2294 
2295 	load = 1;
2296 
2297 	tdq = TDQ_SELF();
2298 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2299 		if (tdq->tdq_load > 0)
2300 			goto out;
2301 	} else
2302 		if (tdq->tdq_load - 1 > 0)
2303 			goto out;
2304 	load = 0;
2305 out:
2306 	return (load);
2307 }
2308 
2309 /*
2310  * Choose the highest priority thread to run.  The thread is removed from
2311  * the run-queue while running however the load remains.  For SMP we set
2312  * the tdq in the global idle bitmask if it idles here.
2313  */
2314 struct thread *
2315 sched_choose(void)
2316 {
2317 	struct thread *td;
2318 	struct tdq *tdq;
2319 
2320 	tdq = TDQ_SELF();
2321 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2322 	td = tdq_choose(tdq);
2323 	if (td) {
2324 		tdq_runq_rem(tdq, td);
2325 		tdq->tdq_lowpri = td->td_priority;
2326 		return (td);
2327 	}
2328 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2329 	return (PCPU_GET(idlethread));
2330 }
2331 
2332 /*
2333  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2334  * we always request it once we exit a critical section.
2335  */
2336 static inline void
2337 sched_setpreempt(struct thread *td)
2338 {
2339 	struct thread *ctd;
2340 	int cpri;
2341 	int pri;
2342 
2343 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2344 
2345 	ctd = curthread;
2346 	pri = td->td_priority;
2347 	cpri = ctd->td_priority;
2348 	if (pri < cpri)
2349 		ctd->td_flags |= TDF_NEEDRESCHED;
2350 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2351 		return;
2352 	if (!sched_shouldpreempt(pri, cpri, 0))
2353 		return;
2354 	ctd->td_owepreempt = 1;
2355 }
2356 
2357 /*
2358  * Add a thread to a thread queue.  Select the appropriate runq and add the
2359  * thread to it.  This is the internal function called when the tdq is
2360  * predetermined.
2361  */
2362 void
2363 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2364 {
2365 
2366 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2367 	KASSERT((td->td_inhibitors == 0),
2368 	    ("sched_add: trying to run inhibited thread"));
2369 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2370 	    ("sched_add: bad thread state"));
2371 	KASSERT(td->td_flags & TDF_INMEM,
2372 	    ("sched_add: thread swapped out"));
2373 
2374 	if (td->td_priority < tdq->tdq_lowpri)
2375 		tdq->tdq_lowpri = td->td_priority;
2376 	tdq_runq_add(tdq, td, flags);
2377 	tdq_load_add(tdq, td);
2378 }
2379 
2380 /*
2381  * Select the target thread queue and add a thread to it.  Request
2382  * preemption or IPI a remote processor if required.
2383  */
2384 void
2385 sched_add(struct thread *td, int flags)
2386 {
2387 	struct tdq *tdq;
2388 #ifdef SMP
2389 	int cpu;
2390 #endif
2391 
2392 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2393 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2394 	    sched_tdname(curthread));
2395 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2396 	    KTR_ATTR_LINKED, sched_tdname(td));
2397 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2398 	    flags & SRQ_PREEMPTED);
2399 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2400 	/*
2401 	 * Recalculate the priority before we select the target cpu or
2402 	 * run-queue.
2403 	 */
2404 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2405 		sched_priority(td);
2406 #ifdef SMP
2407 	/*
2408 	 * Pick the destination cpu and if it isn't ours transfer to the
2409 	 * target cpu.
2410 	 */
2411 	cpu = sched_pickcpu(td, flags);
2412 	tdq = sched_setcpu(td, cpu, flags);
2413 	tdq_add(tdq, td, flags);
2414 	if (cpu != PCPU_GET(cpuid)) {
2415 		tdq_notify(tdq, td);
2416 		return;
2417 	}
2418 #else
2419 	tdq = TDQ_SELF();
2420 	TDQ_LOCK(tdq);
2421 	/*
2422 	 * Now that the thread is moving to the run-queue, set the lock
2423 	 * to the scheduler's lock.
2424 	 */
2425 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2426 	tdq_add(tdq, td, flags);
2427 #endif
2428 	if (!(flags & SRQ_YIELDING))
2429 		sched_setpreempt(td);
2430 }
2431 
2432 /*
2433  * Remove a thread from a run-queue without running it.  This is used
2434  * when we're stealing a thread from a remote queue.  Otherwise all threads
2435  * exit by calling sched_exit_thread() and sched_throw() themselves.
2436  */
2437 void
2438 sched_rem(struct thread *td)
2439 {
2440 	struct tdq *tdq;
2441 
2442 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2443 	    "prio:%d", td->td_priority);
2444 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2445 	tdq = TDQ_CPU(td->td_sched->ts_cpu);
2446 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2447 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2448 	KASSERT(TD_ON_RUNQ(td),
2449 	    ("sched_rem: thread not on run queue"));
2450 	tdq_runq_rem(tdq, td);
2451 	tdq_load_rem(tdq, td);
2452 	TD_SET_CAN_RUN(td);
2453 	if (td->td_priority == tdq->tdq_lowpri)
2454 		tdq_setlowpri(tdq, NULL);
2455 }
2456 
2457 /*
2458  * Fetch cpu utilization information.  Updates on demand.
2459  */
2460 fixpt_t
2461 sched_pctcpu(struct thread *td)
2462 {
2463 	fixpt_t pctcpu;
2464 	struct td_sched *ts;
2465 
2466 	pctcpu = 0;
2467 	ts = td->td_sched;
2468 	if (ts == NULL)
2469 		return (0);
2470 
2471 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2472 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2473 	if (ts->ts_ticks) {
2474 		int rtick;
2475 
2476 		/* How many rtick per second ? */
2477 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2478 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2479 	}
2480 
2481 	return (pctcpu);
2482 }
2483 
2484 /*
2485  * Enforce affinity settings for a thread.  Called after adjustments to
2486  * cpumask.
2487  */
2488 void
2489 sched_affinity(struct thread *td)
2490 {
2491 #ifdef SMP
2492 	struct td_sched *ts;
2493 
2494 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2495 	ts = td->td_sched;
2496 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2497 		return;
2498 	if (TD_ON_RUNQ(td)) {
2499 		sched_rem(td);
2500 		sched_add(td, SRQ_BORING);
2501 		return;
2502 	}
2503 	if (!TD_IS_RUNNING(td))
2504 		return;
2505 	/*
2506 	 * Force a switch before returning to userspace.  If the
2507 	 * target thread is not running locally send an ipi to force
2508 	 * the issue.
2509 	 */
2510 	td->td_flags |= TDF_NEEDRESCHED;
2511 	if (td != curthread)
2512 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2513 #endif
2514 }
2515 
2516 /*
2517  * Bind a thread to a target cpu.
2518  */
2519 void
2520 sched_bind(struct thread *td, int cpu)
2521 {
2522 	struct td_sched *ts;
2523 
2524 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2525 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2526 	ts = td->td_sched;
2527 	if (ts->ts_flags & TSF_BOUND)
2528 		sched_unbind(td);
2529 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2530 	ts->ts_flags |= TSF_BOUND;
2531 	sched_pin();
2532 	if (PCPU_GET(cpuid) == cpu)
2533 		return;
2534 	ts->ts_cpu = cpu;
2535 	/* When we return from mi_switch we'll be on the correct cpu. */
2536 	mi_switch(SW_VOL, NULL);
2537 }
2538 
2539 /*
2540  * Release a bound thread.
2541  */
2542 void
2543 sched_unbind(struct thread *td)
2544 {
2545 	struct td_sched *ts;
2546 
2547 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2548 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2549 	ts = td->td_sched;
2550 	if ((ts->ts_flags & TSF_BOUND) == 0)
2551 		return;
2552 	ts->ts_flags &= ~TSF_BOUND;
2553 	sched_unpin();
2554 }
2555 
2556 int
2557 sched_is_bound(struct thread *td)
2558 {
2559 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2560 	return (td->td_sched->ts_flags & TSF_BOUND);
2561 }
2562 
2563 /*
2564  * Basic yield call.
2565  */
2566 void
2567 sched_relinquish(struct thread *td)
2568 {
2569 	thread_lock(td);
2570 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2571 	thread_unlock(td);
2572 }
2573 
2574 /*
2575  * Return the total system load.
2576  */
2577 int
2578 sched_load(void)
2579 {
2580 #ifdef SMP
2581 	int total;
2582 	int i;
2583 
2584 	total = 0;
2585 	CPU_FOREACH(i)
2586 		total += TDQ_CPU(i)->tdq_sysload;
2587 	return (total);
2588 #else
2589 	return (TDQ_SELF()->tdq_sysload);
2590 #endif
2591 }
2592 
2593 int
2594 sched_sizeof_proc(void)
2595 {
2596 	return (sizeof(struct proc));
2597 }
2598 
2599 int
2600 sched_sizeof_thread(void)
2601 {
2602 	return (sizeof(struct thread) + sizeof(struct td_sched));
2603 }
2604 
2605 #ifdef SMP
2606 #define	TDQ_IDLESPIN(tdq)						\
2607     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2608 #else
2609 #define	TDQ_IDLESPIN(tdq)	1
2610 #endif
2611 
2612 /*
2613  * The actual idle process.
2614  */
2615 void
2616 sched_idletd(void *dummy)
2617 {
2618 	struct thread *td;
2619 	struct tdq *tdq;
2620 	int oldswitchcnt, switchcnt;
2621 	int i;
2622 
2623 	mtx_assert(&Giant, MA_NOTOWNED);
2624 	td = curthread;
2625 	tdq = TDQ_SELF();
2626 	THREAD_NO_SLEEPING();
2627 	oldswitchcnt = -1;
2628 	for (;;) {
2629 		if (tdq->tdq_load) {
2630 			thread_lock(td);
2631 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2632 			thread_unlock(td);
2633 		}
2634 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2635 #ifdef SMP
2636 		if (switchcnt != oldswitchcnt) {
2637 			oldswitchcnt = switchcnt;
2638 			if (tdq_idled(tdq) == 0)
2639 				continue;
2640 		}
2641 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2642 #else
2643 		oldswitchcnt = switchcnt;
2644 #endif
2645 		/*
2646 		 * If we're switching very frequently, spin while checking
2647 		 * for load rather than entering a low power state that
2648 		 * may require an IPI.  However, don't do any busy
2649 		 * loops while on SMT machines as this simply steals
2650 		 * cycles from cores doing useful work.
2651 		 */
2652 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2653 			for (i = 0; i < sched_idlespins; i++) {
2654 				if (tdq->tdq_load)
2655 					break;
2656 				cpu_spinwait();
2657 			}
2658 		}
2659 
2660 		/* If there was context switch during spin, restart it. */
2661 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2662 		if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2663 			continue;
2664 
2665 		/* Run main MD idle handler. */
2666 		tdq->tdq_cpu_idle = 1;
2667 		/*
2668 		 * Make sure that tdq_cpu_idle update is globally visible
2669 		 * before cpu_idle() read tdq_load.  The order is important
2670 		 * to avoid race with tdq_notify.
2671 		 */
2672 		mb();
2673 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2674 		tdq->tdq_cpu_idle = 0;
2675 
2676 		/*
2677 		 * Account thread-less hardware interrupts and
2678 		 * other wakeup reasons equal to context switches.
2679 		 */
2680 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2681 		if (switchcnt != oldswitchcnt)
2682 			continue;
2683 		tdq->tdq_switchcnt++;
2684 		oldswitchcnt++;
2685 	}
2686 }
2687 
2688 /*
2689  * A CPU is entering for the first time or a thread is exiting.
2690  */
2691 void
2692 sched_throw(struct thread *td)
2693 {
2694 	struct thread *newtd;
2695 	struct tdq *tdq;
2696 
2697 	tdq = TDQ_SELF();
2698 	if (td == NULL) {
2699 		/* Correct spinlock nesting and acquire the correct lock. */
2700 		TDQ_LOCK(tdq);
2701 		spinlock_exit();
2702 		PCPU_SET(switchtime, cpu_ticks());
2703 		PCPU_SET(switchticks, ticks);
2704 	} else {
2705 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2706 		tdq_load_rem(tdq, td);
2707 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2708 	}
2709 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2710 	newtd = choosethread();
2711 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2712 	cpu_throw(td, newtd);		/* doesn't return */
2713 }
2714 
2715 /*
2716  * This is called from fork_exit().  Just acquire the correct locks and
2717  * let fork do the rest of the work.
2718  */
2719 void
2720 sched_fork_exit(struct thread *td)
2721 {
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 	if (TD_IS_IDLETHREAD(td))
2732 		td->td_lock = TDQ_LOCKPTR(tdq);
2733 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2734 	td->td_oncpu = cpuid;
2735 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2736 	lock_profile_obtain_lock_success(
2737 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2738 }
2739 
2740 /*
2741  * Create on first use to catch odd startup conditons.
2742  */
2743 char *
2744 sched_tdname(struct thread *td)
2745 {
2746 #ifdef KTR
2747 	struct td_sched *ts;
2748 
2749 	ts = td->td_sched;
2750 	if (ts->ts_name[0] == '\0')
2751 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2752 		    "%s tid %d", td->td_name, td->td_tid);
2753 	return (ts->ts_name);
2754 #else
2755 	return (td->td_name);
2756 #endif
2757 }
2758 
2759 #ifdef KTR
2760 void
2761 sched_clear_tdname(struct thread *td)
2762 {
2763 	struct td_sched *ts;
2764 
2765 	ts = td->td_sched;
2766 	ts->ts_name[0] = '\0';
2767 }
2768 #endif
2769 
2770 #ifdef SMP
2771 
2772 /*
2773  * Build the CPU topology dump string. Is recursively called to collect
2774  * the topology tree.
2775  */
2776 static int
2777 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2778     int indent)
2779 {
2780 	char cpusetbuf[CPUSETBUFSIZ];
2781 	int i, first;
2782 
2783 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2784 	    "", 1 + indent / 2, cg->cg_level);
2785 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2786 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2787 	first = TRUE;
2788 	for (i = 0; i < MAXCPU; i++) {
2789 		if (CPU_ISSET(i, &cg->cg_mask)) {
2790 			if (!first)
2791 				sbuf_printf(sb, ", ");
2792 			else
2793 				first = FALSE;
2794 			sbuf_printf(sb, "%d", i);
2795 		}
2796 	}
2797 	sbuf_printf(sb, "</cpu>\n");
2798 
2799 	if (cg->cg_flags != 0) {
2800 		sbuf_printf(sb, "%*s <flags>", indent, "");
2801 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2802 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2803 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2804 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2805 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2806 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2807 		sbuf_printf(sb, "</flags>\n");
2808 	}
2809 
2810 	if (cg->cg_children > 0) {
2811 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2812 		for (i = 0; i < cg->cg_children; i++)
2813 			sysctl_kern_sched_topology_spec_internal(sb,
2814 			    &cg->cg_child[i], indent+2);
2815 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2816 	}
2817 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2818 	return (0);
2819 }
2820 
2821 /*
2822  * Sysctl handler for retrieving topology dump. It's a wrapper for
2823  * the recursive sysctl_kern_smp_topology_spec_internal().
2824  */
2825 static int
2826 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2827 {
2828 	struct sbuf *topo;
2829 	int err;
2830 
2831 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2832 
2833 	topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
2834 	if (topo == NULL)
2835 		return (ENOMEM);
2836 
2837 	sbuf_printf(topo, "<groups>\n");
2838 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2839 	sbuf_printf(topo, "</groups>\n");
2840 
2841 	if (err == 0) {
2842 		err = sbuf_finish(topo);
2843 	}
2844 	sbuf_delete(topo);
2845 	return (err);
2846 }
2847 
2848 #endif
2849 
2850 static int
2851 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2852 {
2853 	int error, new_val, period;
2854 
2855 	period = 1000000 / realstathz;
2856 	new_val = period * sched_slice;
2857 	error = sysctl_handle_int(oidp, &new_val, 0, req);
2858 	if (error != 0 || req->newptr == NULL)
2859 		return (error);
2860 	if (new_val <= 0)
2861 		return (EINVAL);
2862 	sched_slice = imax(1, (new_val + period / 2) / period);
2863 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
2864 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2865 	    realstathz);
2866 	return (0);
2867 }
2868 
2869 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2870 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2871     "Scheduler name");
2872 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2873     NULL, 0, sysctl_kern_quantum, "I",
2874     "Quantum for timeshare threads in microseconds");
2875 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2876     "Quantum for timeshare threads in stathz ticks");
2877 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2878     "Interactivity score threshold");
2879 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2880     &preempt_thresh, 0,
2881     "Maximal (lowest) priority for preemption");
2882 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2883     "Assign static kernel priorities to sleeping threads");
2884 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2885     "Number of times idle thread will spin waiting for new work");
2886 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2887     &sched_idlespinthresh, 0,
2888     "Threshold before we will permit idle thread spinning");
2889 #ifdef SMP
2890 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2891     "Number of hz ticks to keep thread affinity for");
2892 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2893     "Enables the long-term load balancer");
2894 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2895     &balance_interval, 0,
2896     "Average period in stathz ticks to run the long-term balancer");
2897 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2898     "Attempts to steal work from other cores before idling");
2899 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2900     "Minimum load on remote CPU before we'll steal");
2901 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2902     CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2903     "XML dump of detected CPU topology");
2904 #endif
2905 
2906 /* ps compat.  All cpu percentages from ULE are weighted. */
2907 static int ccpu = 0;
2908 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2909