xref: /freebsd/sys/kern/sched_ule.c (revision a0e793cbf1951d07fc47a0d9ea389d7dacba5213)
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 	u_char		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 /*
360  * Print the threads waiting on a run-queue.
361  */
362 static void
363 runq_print(struct runq *rq)
364 {
365 	struct rqhead *rqh;
366 	struct thread *td;
367 	int pri;
368 	int j;
369 	int i;
370 
371 	for (i = 0; i < RQB_LEN; i++) {
372 		printf("\t\trunq bits %d 0x%zx\n",
373 		    i, rq->rq_status.rqb_bits[i]);
374 		for (j = 0; j < RQB_BPW; j++)
375 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
376 				pri = j + (i << RQB_L2BPW);
377 				rqh = &rq->rq_queues[pri];
378 				TAILQ_FOREACH(td, rqh, td_runq) {
379 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
380 					    td, td->td_name, td->td_priority,
381 					    td->td_rqindex, pri);
382 				}
383 			}
384 	}
385 }
386 
387 /*
388  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
389  */
390 void
391 tdq_print(int cpu)
392 {
393 	struct tdq *tdq;
394 
395 	tdq = TDQ_CPU(cpu);
396 
397 	printf("tdq %d:\n", TDQ_ID(tdq));
398 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
399 	printf("\tLock name:      %s\n", tdq->tdq_name);
400 	printf("\tload:           %d\n", tdq->tdq_load);
401 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
402 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
403 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
404 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
405 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
406 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
407 	printf("\trealtime runq:\n");
408 	runq_print(&tdq->tdq_realtime);
409 	printf("\ttimeshare runq:\n");
410 	runq_print(&tdq->tdq_timeshare);
411 	printf("\tidle runq:\n");
412 	runq_print(&tdq->tdq_idle);
413 }
414 
415 static inline int
416 sched_shouldpreempt(int pri, int cpri, int remote)
417 {
418 	/*
419 	 * If the new priority is not better than the current priority there is
420 	 * nothing to do.
421 	 */
422 	if (pri >= cpri)
423 		return (0);
424 	/*
425 	 * Always preempt idle.
426 	 */
427 	if (cpri >= PRI_MIN_IDLE)
428 		return (1);
429 	/*
430 	 * If preemption is disabled don't preempt others.
431 	 */
432 	if (preempt_thresh == 0)
433 		return (0);
434 	/*
435 	 * Preempt if we exceed the threshold.
436 	 */
437 	if (pri <= preempt_thresh)
438 		return (1);
439 	/*
440 	 * If we're interactive or better and there is non-interactive
441 	 * or worse running preempt only remote processors.
442 	 */
443 	if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
444 		return (1);
445 	return (0);
446 }
447 
448 /*
449  * Add a thread to the actual run-queue.  Keeps transferable counts up to
450  * date with what is actually on the run-queue.  Selects the correct
451  * queue position for timeshare threads.
452  */
453 static __inline void
454 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
455 {
456 	struct td_sched *ts;
457 	u_char pri;
458 
459 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
460 	THREAD_LOCK_ASSERT(td, MA_OWNED);
461 
462 	pri = td->td_priority;
463 	ts = td->td_sched;
464 	TD_SET_RUNQ(td);
465 	if (THREAD_CAN_MIGRATE(td)) {
466 		tdq->tdq_transferable++;
467 		ts->ts_flags |= TSF_XFERABLE;
468 	}
469 	if (pri < PRI_MIN_BATCH) {
470 		ts->ts_runq = &tdq->tdq_realtime;
471 	} else if (pri <= PRI_MAX_BATCH) {
472 		ts->ts_runq = &tdq->tdq_timeshare;
473 		KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
474 			("Invalid priority %d on timeshare runq", pri));
475 		/*
476 		 * This queue contains only priorities between MIN and MAX
477 		 * realtime.  Use the whole queue to represent these values.
478 		 */
479 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
480 			pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
481 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
482 			/*
483 			 * This effectively shortens the queue by one so we
484 			 * can have a one slot difference between idx and
485 			 * ridx while we wait for threads to drain.
486 			 */
487 			if (tdq->tdq_ridx != tdq->tdq_idx &&
488 			    pri == tdq->tdq_ridx)
489 				pri = (unsigned char)(pri - 1) % RQ_NQS;
490 		} else
491 			pri = tdq->tdq_ridx;
492 		runq_add_pri(ts->ts_runq, td, pri, flags);
493 		return;
494 	} else
495 		ts->ts_runq = &tdq->tdq_idle;
496 	runq_add(ts->ts_runq, td, flags);
497 }
498 
499 /*
500  * Remove a thread from a run-queue.  This typically happens when a thread
501  * is selected to run.  Running threads are not on the queue and the
502  * transferable count does not reflect them.
503  */
504 static __inline void
505 tdq_runq_rem(struct tdq *tdq, struct thread *td)
506 {
507 	struct td_sched *ts;
508 
509 	ts = td->td_sched;
510 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
511 	KASSERT(ts->ts_runq != NULL,
512 	    ("tdq_runq_remove: thread %p null ts_runq", td));
513 	if (ts->ts_flags & TSF_XFERABLE) {
514 		tdq->tdq_transferable--;
515 		ts->ts_flags &= ~TSF_XFERABLE;
516 	}
517 	if (ts->ts_runq == &tdq->tdq_timeshare) {
518 		if (tdq->tdq_idx != tdq->tdq_ridx)
519 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
520 		else
521 			runq_remove_idx(ts->ts_runq, td, NULL);
522 	} else
523 		runq_remove(ts->ts_runq, td);
524 }
525 
526 /*
527  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
528  * for this thread to the referenced thread queue.
529  */
530 static void
531 tdq_load_add(struct tdq *tdq, struct thread *td)
532 {
533 
534 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
535 	THREAD_LOCK_ASSERT(td, MA_OWNED);
536 
537 	tdq->tdq_load++;
538 	if ((td->td_flags & TDF_NOLOAD) == 0)
539 		tdq->tdq_sysload++;
540 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
541 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
542 }
543 
544 /*
545  * Remove the load from a thread that is transitioning to a sleep state or
546  * exiting.
547  */
548 static void
549 tdq_load_rem(struct tdq *tdq, struct thread *td)
550 {
551 
552 	THREAD_LOCK_ASSERT(td, MA_OWNED);
553 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
554 	KASSERT(tdq->tdq_load != 0,
555 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
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  * Bound timeshare latency by decreasing slice size as load increases.  We
566  * consider the maximum latency as the sum of the threads waiting to run
567  * aside from curthread and target no more than sched_slice latency but
568  * no less than sched_slice_min runtime.
569  */
570 static inline int
571 tdq_slice(struct tdq *tdq)
572 {
573 	int load;
574 
575 	/*
576 	 * It is safe to use sys_load here because this is called from
577 	 * contexts where timeshare threads are running and so there
578 	 * cannot be higher priority load in the system.
579 	 */
580 	load = tdq->tdq_sysload - 1;
581 	if (load >= SCHED_SLICE_MIN_DIVISOR)
582 		return (sched_slice_min);
583 	if (load <= 1)
584 		return (sched_slice);
585 	return (sched_slice / load);
586 }
587 
588 /*
589  * Set lowpri to its exact value by searching the run-queue and
590  * evaluating curthread.  curthread may be passed as an optimization.
591  */
592 static void
593 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
594 {
595 	struct thread *td;
596 
597 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
598 	if (ctd == NULL)
599 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
600 	td = tdq_choose(tdq);
601 	if (td == NULL || td->td_priority > ctd->td_priority)
602 		tdq->tdq_lowpri = ctd->td_priority;
603 	else
604 		tdq->tdq_lowpri = td->td_priority;
605 }
606 
607 #ifdef SMP
608 struct cpu_search {
609 	cpuset_t cs_mask;
610 	u_int	cs_prefer;
611 	int	cs_pri;		/* Min priority for low. */
612 	int	cs_limit;	/* Max load for low, min load for high. */
613 	int	cs_cpu;
614 	int	cs_load;
615 };
616 
617 #define	CPU_SEARCH_LOWEST	0x1
618 #define	CPU_SEARCH_HIGHEST	0x2
619 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
620 
621 #define	CPUSET_FOREACH(cpu, mask)				\
622 	for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++)		\
623 		if (CPU_ISSET(cpu, &mask))
624 
625 static __always_inline int cpu_search(const struct cpu_group *cg,
626     struct cpu_search *low, struct cpu_search *high, const int match);
627 int __noinline cpu_search_lowest(const struct cpu_group *cg,
628     struct cpu_search *low);
629 int __noinline cpu_search_highest(const struct cpu_group *cg,
630     struct cpu_search *high);
631 int __noinline cpu_search_both(const struct cpu_group *cg,
632     struct cpu_search *low, struct cpu_search *high);
633 
634 /*
635  * Search the tree of cpu_groups for the lowest or highest loaded cpu
636  * according to the match argument.  This routine actually compares the
637  * load on all paths through the tree and finds the least loaded cpu on
638  * the least loaded path, which may differ from the least loaded cpu in
639  * the system.  This balances work among caches and busses.
640  *
641  * This inline is instantiated in three forms below using constants for the
642  * match argument.  It is reduced to the minimum set for each case.  It is
643  * also recursive to the depth of the tree.
644  */
645 static __always_inline int
646 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
647     struct cpu_search *high, const int match)
648 {
649 	struct cpu_search lgroup;
650 	struct cpu_search hgroup;
651 	cpuset_t cpumask;
652 	struct cpu_group *child;
653 	struct tdq *tdq;
654 	int cpu, i, hload, lload, load, total, rnd, *rndptr;
655 
656 	total = 0;
657 	cpumask = cg->cg_mask;
658 	if (match & CPU_SEARCH_LOWEST) {
659 		lload = INT_MAX;
660 		lgroup = *low;
661 	}
662 	if (match & CPU_SEARCH_HIGHEST) {
663 		hload = INT_MIN;
664 		hgroup = *high;
665 	}
666 
667 	/* Iterate through the child CPU groups and then remaining CPUs. */
668 	for (i = cg->cg_children, cpu = mp_maxid; ; ) {
669 		if (i == 0) {
670 #ifdef HAVE_INLINE_FFSL
671 			cpu = CPU_FFS(&cpumask) - 1;
672 #else
673 			while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
674 				cpu--;
675 #endif
676 			if (cpu < 0)
677 				break;
678 			child = NULL;
679 		} else
680 			child = &cg->cg_child[i - 1];
681 
682 		if (match & CPU_SEARCH_LOWEST)
683 			lgroup.cs_cpu = -1;
684 		if (match & CPU_SEARCH_HIGHEST)
685 			hgroup.cs_cpu = -1;
686 		if (child) {			/* Handle child CPU group. */
687 			CPU_NAND(&cpumask, &child->cg_mask);
688 			switch (match) {
689 			case CPU_SEARCH_LOWEST:
690 				load = cpu_search_lowest(child, &lgroup);
691 				break;
692 			case CPU_SEARCH_HIGHEST:
693 				load = cpu_search_highest(child, &hgroup);
694 				break;
695 			case CPU_SEARCH_BOTH:
696 				load = cpu_search_both(child, &lgroup, &hgroup);
697 				break;
698 			}
699 		} else {			/* Handle child CPU. */
700 			CPU_CLR(cpu, &cpumask);
701 			tdq = TDQ_CPU(cpu);
702 			load = tdq->tdq_load * 256;
703 			rndptr = DPCPU_PTR(randomval);
704 			rnd = (*rndptr = *rndptr * 69069 + 5) >> 26;
705 			if (match & CPU_SEARCH_LOWEST) {
706 				if (cpu == low->cs_prefer)
707 					load -= 64;
708 				/* If that CPU is allowed and get data. */
709 				if (tdq->tdq_lowpri > lgroup.cs_pri &&
710 				    tdq->tdq_load <= lgroup.cs_limit &&
711 				    CPU_ISSET(cpu, &lgroup.cs_mask)) {
712 					lgroup.cs_cpu = cpu;
713 					lgroup.cs_load = load - rnd;
714 				}
715 			}
716 			if (match & CPU_SEARCH_HIGHEST)
717 				if (tdq->tdq_load >= hgroup.cs_limit &&
718 				    tdq->tdq_transferable &&
719 				    CPU_ISSET(cpu, &hgroup.cs_mask)) {
720 					hgroup.cs_cpu = cpu;
721 					hgroup.cs_load = load - rnd;
722 				}
723 		}
724 		total += load;
725 
726 		/* We have info about child item. Compare it. */
727 		if (match & CPU_SEARCH_LOWEST) {
728 			if (lgroup.cs_cpu >= 0 &&
729 			    (load < lload ||
730 			     (load == lload && lgroup.cs_load < low->cs_load))) {
731 				lload = load;
732 				low->cs_cpu = lgroup.cs_cpu;
733 				low->cs_load = lgroup.cs_load;
734 			}
735 		}
736 		if (match & CPU_SEARCH_HIGHEST)
737 			if (hgroup.cs_cpu >= 0 &&
738 			    (load > hload ||
739 			     (load == hload && hgroup.cs_load > high->cs_load))) {
740 				hload = load;
741 				high->cs_cpu = hgroup.cs_cpu;
742 				high->cs_load = hgroup.cs_load;
743 			}
744 		if (child) {
745 			i--;
746 			if (i == 0 && CPU_EMPTY(&cpumask))
747 				break;
748 		}
749 #ifndef HAVE_INLINE_FFSL
750 		else
751 			cpu--;
752 #endif
753 	}
754 	return (total);
755 }
756 
757 /*
758  * cpu_search instantiations must pass constants to maintain the inline
759  * optimization.
760  */
761 int
762 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
763 {
764 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
765 }
766 
767 int
768 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
769 {
770 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
771 }
772 
773 int
774 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
775     struct cpu_search *high)
776 {
777 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
778 }
779 
780 /*
781  * Find the cpu with the least load via the least loaded path that has a
782  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
783  * acceptable.
784  */
785 static inline int
786 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
787     int prefer)
788 {
789 	struct cpu_search low;
790 
791 	low.cs_cpu = -1;
792 	low.cs_prefer = prefer;
793 	low.cs_mask = mask;
794 	low.cs_pri = pri;
795 	low.cs_limit = maxload;
796 	cpu_search_lowest(cg, &low);
797 	return low.cs_cpu;
798 }
799 
800 /*
801  * Find the cpu with the highest load via the highest loaded path.
802  */
803 static inline int
804 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
805 {
806 	struct cpu_search high;
807 
808 	high.cs_cpu = -1;
809 	high.cs_mask = mask;
810 	high.cs_limit = minload;
811 	cpu_search_highest(cg, &high);
812 	return high.cs_cpu;
813 }
814 
815 static void
816 sched_balance_group(struct cpu_group *cg)
817 {
818 	cpuset_t hmask, lmask;
819 	int high, low, anylow;
820 
821 	CPU_FILL(&hmask);
822 	for (;;) {
823 		high = sched_highest(cg, hmask, 1);
824 		/* Stop if there is no more CPU with transferrable threads. */
825 		if (high == -1)
826 			break;
827 		CPU_CLR(high, &hmask);
828 		CPU_COPY(&hmask, &lmask);
829 		/* Stop if there is no more CPU left for low. */
830 		if (CPU_EMPTY(&lmask))
831 			break;
832 		anylow = 1;
833 nextlow:
834 		low = sched_lowest(cg, lmask, -1,
835 		    TDQ_CPU(high)->tdq_load - 1, high);
836 		/* Stop if we looked well and found no less loaded CPU. */
837 		if (anylow && low == -1)
838 			break;
839 		/* Go to next high if we found no less loaded CPU. */
840 		if (low == -1)
841 			continue;
842 		/* Transfer thread from high to low. */
843 		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
844 			/* CPU that got thread can no longer be a donor. */
845 			CPU_CLR(low, &hmask);
846 		} else {
847 			/*
848 			 * If failed, then there is no threads on high
849 			 * that can run on this low. Drop low from low
850 			 * mask and look for different one.
851 			 */
852 			CPU_CLR(low, &lmask);
853 			anylow = 0;
854 			goto nextlow;
855 		}
856 	}
857 }
858 
859 static void
860 sched_balance(void)
861 {
862 	struct tdq *tdq;
863 
864 	/*
865 	 * Select a random time between .5 * balance_interval and
866 	 * 1.5 * balance_interval.
867 	 */
868 	balance_ticks = max(balance_interval / 2, 1);
869 	balance_ticks += random() % balance_interval;
870 	if (smp_started == 0 || rebalance == 0)
871 		return;
872 	tdq = TDQ_SELF();
873 	TDQ_UNLOCK(tdq);
874 	sched_balance_group(cpu_top);
875 	TDQ_LOCK(tdq);
876 }
877 
878 /*
879  * Lock two thread queues using their address to maintain lock order.
880  */
881 static void
882 tdq_lock_pair(struct tdq *one, struct tdq *two)
883 {
884 	if (one < two) {
885 		TDQ_LOCK(one);
886 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
887 	} else {
888 		TDQ_LOCK(two);
889 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
890 	}
891 }
892 
893 /*
894  * Unlock two thread queues.  Order is not important here.
895  */
896 static void
897 tdq_unlock_pair(struct tdq *one, struct tdq *two)
898 {
899 	TDQ_UNLOCK(one);
900 	TDQ_UNLOCK(two);
901 }
902 
903 /*
904  * Transfer load between two imbalanced thread queues.
905  */
906 static int
907 sched_balance_pair(struct tdq *high, struct tdq *low)
908 {
909 	int moved;
910 	int cpu;
911 
912 	tdq_lock_pair(high, low);
913 	moved = 0;
914 	/*
915 	 * Determine what the imbalance is and then adjust that to how many
916 	 * threads we actually have to give up (transferable).
917 	 */
918 	if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
919 	    (moved = tdq_move(high, low)) > 0) {
920 		/*
921 		 * In case the target isn't the current cpu IPI it to force a
922 		 * reschedule with the new workload.
923 		 */
924 		cpu = TDQ_ID(low);
925 		if (cpu != PCPU_GET(cpuid))
926 			ipi_cpu(cpu, IPI_PREEMPT);
927 	}
928 	tdq_unlock_pair(high, low);
929 	return (moved);
930 }
931 
932 /*
933  * Move a thread from one thread queue to another.
934  */
935 static int
936 tdq_move(struct tdq *from, struct tdq *to)
937 {
938 	struct td_sched *ts;
939 	struct thread *td;
940 	struct tdq *tdq;
941 	int cpu;
942 
943 	TDQ_LOCK_ASSERT(from, MA_OWNED);
944 	TDQ_LOCK_ASSERT(to, MA_OWNED);
945 
946 	tdq = from;
947 	cpu = TDQ_ID(to);
948 	td = tdq_steal(tdq, cpu);
949 	if (td == NULL)
950 		return (0);
951 	ts = td->td_sched;
952 	/*
953 	 * Although the run queue is locked the thread may be blocked.  Lock
954 	 * it to clear this and acquire the run-queue lock.
955 	 */
956 	thread_lock(td);
957 	/* Drop recursive lock on from acquired via thread_lock(). */
958 	TDQ_UNLOCK(from);
959 	sched_rem(td);
960 	ts->ts_cpu = cpu;
961 	td->td_lock = TDQ_LOCKPTR(to);
962 	tdq_add(to, td, SRQ_YIELDING);
963 	return (1);
964 }
965 
966 /*
967  * This tdq has idled.  Try to steal a thread from another cpu and switch
968  * to it.
969  */
970 static int
971 tdq_idled(struct tdq *tdq)
972 {
973 	struct cpu_group *cg;
974 	struct tdq *steal;
975 	cpuset_t mask;
976 	int thresh;
977 	int cpu;
978 
979 	if (smp_started == 0 || steal_idle == 0)
980 		return (1);
981 	CPU_FILL(&mask);
982 	CPU_CLR(PCPU_GET(cpuid), &mask);
983 	/* We don't want to be preempted while we're iterating. */
984 	spinlock_enter();
985 	for (cg = tdq->tdq_cg; cg != NULL; ) {
986 		if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
987 			thresh = steal_thresh;
988 		else
989 			thresh = 1;
990 		cpu = sched_highest(cg, mask, thresh);
991 		if (cpu == -1) {
992 			cg = cg->cg_parent;
993 			continue;
994 		}
995 		steal = TDQ_CPU(cpu);
996 		CPU_CLR(cpu, &mask);
997 		tdq_lock_pair(tdq, steal);
998 		if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
999 			tdq_unlock_pair(tdq, steal);
1000 			continue;
1001 		}
1002 		/*
1003 		 * If a thread was added while interrupts were disabled don't
1004 		 * steal one here.  If we fail to acquire one due to affinity
1005 		 * restrictions loop again with this cpu removed from the
1006 		 * set.
1007 		 */
1008 		if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
1009 			tdq_unlock_pair(tdq, steal);
1010 			continue;
1011 		}
1012 		spinlock_exit();
1013 		TDQ_UNLOCK(steal);
1014 		mi_switch(SW_VOL | SWT_IDLE, NULL);
1015 		thread_unlock(curthread);
1016 
1017 		return (0);
1018 	}
1019 	spinlock_exit();
1020 	return (1);
1021 }
1022 
1023 /*
1024  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
1025  */
1026 static void
1027 tdq_notify(struct tdq *tdq, struct thread *td)
1028 {
1029 	struct thread *ctd;
1030 	int pri;
1031 	int cpu;
1032 
1033 	if (tdq->tdq_ipipending)
1034 		return;
1035 	cpu = td->td_sched->ts_cpu;
1036 	pri = td->td_priority;
1037 	ctd = pcpu_find(cpu)->pc_curthread;
1038 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1039 		return;
1040 	if (TD_IS_IDLETHREAD(ctd)) {
1041 		/*
1042 		 * If the MD code has an idle wakeup routine try that before
1043 		 * falling back to IPI.
1044 		 */
1045 		if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1046 			return;
1047 	}
1048 	tdq->tdq_ipipending = 1;
1049 	ipi_cpu(cpu, IPI_PREEMPT);
1050 }
1051 
1052 /*
1053  * Steals load from a timeshare queue.  Honors the rotating queue head
1054  * index.
1055  */
1056 static struct thread *
1057 runq_steal_from(struct runq *rq, int cpu, u_char start)
1058 {
1059 	struct rqbits *rqb;
1060 	struct rqhead *rqh;
1061 	struct thread *td, *first;
1062 	int bit;
1063 	int i;
1064 
1065 	rqb = &rq->rq_status;
1066 	bit = start & (RQB_BPW -1);
1067 	first = NULL;
1068 again:
1069 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1070 		if (rqb->rqb_bits[i] == 0)
1071 			continue;
1072 		if (bit == 0)
1073 			bit = RQB_FFS(rqb->rqb_bits[i]);
1074 		for (; bit < RQB_BPW; bit++) {
1075 			if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1076 				continue;
1077 			rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1078 			TAILQ_FOREACH(td, rqh, td_runq) {
1079 				if (first && THREAD_CAN_MIGRATE(td) &&
1080 				    THREAD_CAN_SCHED(td, cpu))
1081 					return (td);
1082 				first = td;
1083 			}
1084 		}
1085 	}
1086 	if (start != 0) {
1087 		start = 0;
1088 		goto again;
1089 	}
1090 
1091 	if (first && THREAD_CAN_MIGRATE(first) &&
1092 	    THREAD_CAN_SCHED(first, cpu))
1093 		return (first);
1094 	return (NULL);
1095 }
1096 
1097 /*
1098  * Steals load from a standard linear queue.
1099  */
1100 static struct thread *
1101 runq_steal(struct runq *rq, int cpu)
1102 {
1103 	struct rqhead *rqh;
1104 	struct rqbits *rqb;
1105 	struct thread *td;
1106 	int word;
1107 	int bit;
1108 
1109 	rqb = &rq->rq_status;
1110 	for (word = 0; word < RQB_LEN; word++) {
1111 		if (rqb->rqb_bits[word] == 0)
1112 			continue;
1113 		for (bit = 0; bit < RQB_BPW; bit++) {
1114 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1115 				continue;
1116 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1117 			TAILQ_FOREACH(td, rqh, td_runq)
1118 				if (THREAD_CAN_MIGRATE(td) &&
1119 				    THREAD_CAN_SCHED(td, cpu))
1120 					return (td);
1121 		}
1122 	}
1123 	return (NULL);
1124 }
1125 
1126 /*
1127  * Attempt to steal a thread in priority order from a thread queue.
1128  */
1129 static struct thread *
1130 tdq_steal(struct tdq *tdq, int cpu)
1131 {
1132 	struct thread *td;
1133 
1134 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1135 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1136 		return (td);
1137 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1138 	    cpu, tdq->tdq_ridx)) != NULL)
1139 		return (td);
1140 	return (runq_steal(&tdq->tdq_idle, cpu));
1141 }
1142 
1143 /*
1144  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1145  * current lock and returns with the assigned queue locked.
1146  */
1147 static inline struct tdq *
1148 sched_setcpu(struct thread *td, int cpu, int flags)
1149 {
1150 
1151 	struct tdq *tdq;
1152 
1153 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1154 	tdq = TDQ_CPU(cpu);
1155 	td->td_sched->ts_cpu = cpu;
1156 	/*
1157 	 * If the lock matches just return the queue.
1158 	 */
1159 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1160 		return (tdq);
1161 #ifdef notyet
1162 	/*
1163 	 * If the thread isn't running its lockptr is a
1164 	 * turnstile or a sleepqueue.  We can just lock_set without
1165 	 * blocking.
1166 	 */
1167 	if (TD_CAN_RUN(td)) {
1168 		TDQ_LOCK(tdq);
1169 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1170 		return (tdq);
1171 	}
1172 #endif
1173 	/*
1174 	 * The hard case, migration, we need to block the thread first to
1175 	 * prevent order reversals with other cpus locks.
1176 	 */
1177 	spinlock_enter();
1178 	thread_lock_block(td);
1179 	TDQ_LOCK(tdq);
1180 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1181 	spinlock_exit();
1182 	return (tdq);
1183 }
1184 
1185 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1186 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1187 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1188 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1189 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1190 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1191 
1192 static int
1193 sched_pickcpu(struct thread *td, int flags)
1194 {
1195 	struct cpu_group *cg, *ccg;
1196 	struct td_sched *ts;
1197 	struct tdq *tdq;
1198 	cpuset_t mask;
1199 	int cpu, pri, self;
1200 
1201 	self = PCPU_GET(cpuid);
1202 	ts = td->td_sched;
1203 	if (smp_started == 0)
1204 		return (self);
1205 	/*
1206 	 * Don't migrate a running thread from sched_switch().
1207 	 */
1208 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1209 		return (ts->ts_cpu);
1210 	/*
1211 	 * Prefer to run interrupt threads on the processors that generate
1212 	 * the interrupt.
1213 	 */
1214 	pri = td->td_priority;
1215 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1216 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1217 		SCHED_STAT_INC(pickcpu_intrbind);
1218 		ts->ts_cpu = self;
1219 		if (TDQ_CPU(self)->tdq_lowpri > pri) {
1220 			SCHED_STAT_INC(pickcpu_affinity);
1221 			return (ts->ts_cpu);
1222 		}
1223 	}
1224 	/*
1225 	 * If the thread can run on the last cpu and the affinity has not
1226 	 * expired or it is idle run it there.
1227 	 */
1228 	tdq = TDQ_CPU(ts->ts_cpu);
1229 	cg = tdq->tdq_cg;
1230 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1231 	    tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1232 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1233 		if (cg->cg_flags & CG_FLAG_THREAD) {
1234 			CPUSET_FOREACH(cpu, cg->cg_mask) {
1235 				if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1236 					break;
1237 			}
1238 		} else
1239 			cpu = INT_MAX;
1240 		if (cpu > mp_maxid) {
1241 			SCHED_STAT_INC(pickcpu_idle_affinity);
1242 			return (ts->ts_cpu);
1243 		}
1244 	}
1245 	/*
1246 	 * Search for the last level cache CPU group in the tree.
1247 	 * Skip caches with expired affinity time and SMT groups.
1248 	 * Affinity to higher level caches will be handled less aggressively.
1249 	 */
1250 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1251 		if (cg->cg_flags & CG_FLAG_THREAD)
1252 			continue;
1253 		if (!SCHED_AFFINITY(ts, cg->cg_level))
1254 			continue;
1255 		ccg = cg;
1256 	}
1257 	if (ccg != NULL)
1258 		cg = ccg;
1259 	cpu = -1;
1260 	/* Search the group for the less loaded idle CPU we can run now. */
1261 	mask = td->td_cpuset->cs_mask;
1262 	if (cg != NULL && cg != cpu_top &&
1263 	    CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1264 		cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1265 		    INT_MAX, ts->ts_cpu);
1266 	/* Search globally for the less loaded CPU we can run now. */
1267 	if (cpu == -1)
1268 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1269 	/* Search globally for the less loaded CPU. */
1270 	if (cpu == -1)
1271 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1272 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1273 	/*
1274 	 * Compare the lowest loaded cpu to current cpu.
1275 	 */
1276 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1277 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1278 	    TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1279 		SCHED_STAT_INC(pickcpu_local);
1280 		cpu = self;
1281 	} else
1282 		SCHED_STAT_INC(pickcpu_lowest);
1283 	if (cpu != ts->ts_cpu)
1284 		SCHED_STAT_INC(pickcpu_migration);
1285 	return (cpu);
1286 }
1287 #endif
1288 
1289 /*
1290  * Pick the highest priority task we have and return it.
1291  */
1292 static struct thread *
1293 tdq_choose(struct tdq *tdq)
1294 {
1295 	struct thread *td;
1296 
1297 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1298 	td = runq_choose(&tdq->tdq_realtime);
1299 	if (td != NULL)
1300 		return (td);
1301 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1302 	if (td != NULL) {
1303 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1304 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1305 		    td->td_priority));
1306 		return (td);
1307 	}
1308 	td = runq_choose(&tdq->tdq_idle);
1309 	if (td != NULL) {
1310 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1311 		    ("tdq_choose: Invalid priority on idle queue %d",
1312 		    td->td_priority));
1313 		return (td);
1314 	}
1315 
1316 	return (NULL);
1317 }
1318 
1319 /*
1320  * Initialize a thread queue.
1321  */
1322 static void
1323 tdq_setup(struct tdq *tdq)
1324 {
1325 
1326 	if (bootverbose)
1327 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1328 	runq_init(&tdq->tdq_realtime);
1329 	runq_init(&tdq->tdq_timeshare);
1330 	runq_init(&tdq->tdq_idle);
1331 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1332 	    "sched lock %d", (int)TDQ_ID(tdq));
1333 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1334 	    MTX_SPIN | MTX_RECURSE);
1335 #ifdef KTR
1336 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1337 	    "CPU %d load", (int)TDQ_ID(tdq));
1338 #endif
1339 }
1340 
1341 #ifdef SMP
1342 static void
1343 sched_setup_smp(void)
1344 {
1345 	struct tdq *tdq;
1346 	int i;
1347 
1348 	cpu_top = smp_topo();
1349 	CPU_FOREACH(i) {
1350 		tdq = TDQ_CPU(i);
1351 		tdq_setup(tdq);
1352 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1353 		if (tdq->tdq_cg == NULL)
1354 			panic("Can't find cpu group for %d\n", i);
1355 	}
1356 	balance_tdq = TDQ_SELF();
1357 	sched_balance();
1358 }
1359 #endif
1360 
1361 /*
1362  * Setup the thread queues and initialize the topology based on MD
1363  * information.
1364  */
1365 static void
1366 sched_setup(void *dummy)
1367 {
1368 	struct tdq *tdq;
1369 
1370 	tdq = TDQ_SELF();
1371 #ifdef SMP
1372 	sched_setup_smp();
1373 #else
1374 	tdq_setup(tdq);
1375 #endif
1376 
1377 	/* Add thread0's load since it's running. */
1378 	TDQ_LOCK(tdq);
1379 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1380 	tdq_load_add(tdq, &thread0);
1381 	tdq->tdq_lowpri = thread0.td_priority;
1382 	TDQ_UNLOCK(tdq);
1383 }
1384 
1385 /*
1386  * This routine determines time constants after stathz and hz are setup.
1387  */
1388 /* ARGSUSED */
1389 static void
1390 sched_initticks(void *dummy)
1391 {
1392 	int incr;
1393 
1394 	realstathz = stathz ? stathz : hz;
1395 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1396 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1397 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1398 	    realstathz);
1399 
1400 	/*
1401 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1402 	 * hz not being evenly divisible by stathz on all platforms.
1403 	 */
1404 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1405 	/*
1406 	 * This does not work for values of stathz that are more than
1407 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1408 	 */
1409 	if (incr == 0)
1410 		incr = 1;
1411 	tickincr = incr;
1412 #ifdef SMP
1413 	/*
1414 	 * Set the default balance interval now that we know
1415 	 * what realstathz is.
1416 	 */
1417 	balance_interval = realstathz;
1418 	affinity = SCHED_AFFINITY_DEFAULT;
1419 #endif
1420 	if (sched_idlespinthresh < 0)
1421 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1422 }
1423 
1424 
1425 /*
1426  * This is the core of the interactivity algorithm.  Determines a score based
1427  * on past behavior.  It is the ratio of sleep time to run time scaled to
1428  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1429  * differs from the cpu usage because it does not account for time spent
1430  * waiting on a run-queue.  Would be prettier if we had floating point.
1431  */
1432 static int
1433 sched_interact_score(struct thread *td)
1434 {
1435 	struct td_sched *ts;
1436 	int div;
1437 
1438 	ts = td->td_sched;
1439 	/*
1440 	 * The score is only needed if this is likely to be an interactive
1441 	 * task.  Don't go through the expense of computing it if there's
1442 	 * no chance.
1443 	 */
1444 	if (sched_interact <= SCHED_INTERACT_HALF &&
1445 		ts->ts_runtime >= ts->ts_slptime)
1446 			return (SCHED_INTERACT_HALF);
1447 
1448 	if (ts->ts_runtime > ts->ts_slptime) {
1449 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1450 		return (SCHED_INTERACT_HALF +
1451 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1452 	}
1453 	if (ts->ts_slptime > ts->ts_runtime) {
1454 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1455 		return (ts->ts_runtime / div);
1456 	}
1457 	/* runtime == slptime */
1458 	if (ts->ts_runtime)
1459 		return (SCHED_INTERACT_HALF);
1460 
1461 	/*
1462 	 * This can happen if slptime and runtime are 0.
1463 	 */
1464 	return (0);
1465 
1466 }
1467 
1468 /*
1469  * Scale the scheduling priority according to the "interactivity" of this
1470  * process.
1471  */
1472 static void
1473 sched_priority(struct thread *td)
1474 {
1475 	int score;
1476 	int pri;
1477 
1478 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1479 		return;
1480 	/*
1481 	 * If the score is interactive we place the thread in the realtime
1482 	 * queue with a priority that is less than kernel and interrupt
1483 	 * priorities.  These threads are not subject to nice restrictions.
1484 	 *
1485 	 * Scores greater than this are placed on the normal timeshare queue
1486 	 * where the priority is partially decided by the most recent cpu
1487 	 * utilization and the rest is decided by nice value.
1488 	 *
1489 	 * The nice value of the process has a linear effect on the calculated
1490 	 * score.  Negative nice values make it easier for a thread to be
1491 	 * considered interactive.
1492 	 */
1493 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1494 	if (score < sched_interact) {
1495 		pri = PRI_MIN_INTERACT;
1496 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1497 		    sched_interact) * score;
1498 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1499 		    ("sched_priority: invalid interactive priority %d score %d",
1500 		    pri, score));
1501 	} else {
1502 		pri = SCHED_PRI_MIN;
1503 		if (td->td_sched->ts_ticks)
1504 			pri += min(SCHED_PRI_TICKS(td->td_sched),
1505 			    SCHED_PRI_RANGE - 1);
1506 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1507 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1508 		    ("sched_priority: invalid priority %d: nice %d, "
1509 		    "ticks %d ftick %d ltick %d tick pri %d",
1510 		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1511 		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1512 		    SCHED_PRI_TICKS(td->td_sched)));
1513 	}
1514 	sched_user_prio(td, pri);
1515 
1516 	return;
1517 }
1518 
1519 /*
1520  * This routine enforces a maximum limit on the amount of scheduling history
1521  * kept.  It is called after either the slptime or runtime is adjusted.  This
1522  * function is ugly due to integer math.
1523  */
1524 static void
1525 sched_interact_update(struct thread *td)
1526 {
1527 	struct td_sched *ts;
1528 	u_int sum;
1529 
1530 	ts = td->td_sched;
1531 	sum = ts->ts_runtime + ts->ts_slptime;
1532 	if (sum < SCHED_SLP_RUN_MAX)
1533 		return;
1534 	/*
1535 	 * This only happens from two places:
1536 	 * 1) We have added an unusual amount of run time from fork_exit.
1537 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1538 	 */
1539 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1540 		if (ts->ts_runtime > ts->ts_slptime) {
1541 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1542 			ts->ts_slptime = 1;
1543 		} else {
1544 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1545 			ts->ts_runtime = 1;
1546 		}
1547 		return;
1548 	}
1549 	/*
1550 	 * If we have exceeded by more than 1/5th then the algorithm below
1551 	 * will not bring us back into range.  Dividing by two here forces
1552 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1553 	 */
1554 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1555 		ts->ts_runtime /= 2;
1556 		ts->ts_slptime /= 2;
1557 		return;
1558 	}
1559 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1560 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1561 }
1562 
1563 /*
1564  * Scale back the interactivity history when a child thread is created.  The
1565  * history is inherited from the parent but the thread may behave totally
1566  * differently.  For example, a shell spawning a compiler process.  We want
1567  * to learn that the compiler is behaving badly very quickly.
1568  */
1569 static void
1570 sched_interact_fork(struct thread *td)
1571 {
1572 	int ratio;
1573 	int sum;
1574 
1575 	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1576 	if (sum > SCHED_SLP_RUN_FORK) {
1577 		ratio = sum / SCHED_SLP_RUN_FORK;
1578 		td->td_sched->ts_runtime /= ratio;
1579 		td->td_sched->ts_slptime /= ratio;
1580 	}
1581 }
1582 
1583 /*
1584  * Called from proc0_init() to setup the scheduler fields.
1585  */
1586 void
1587 schedinit(void)
1588 {
1589 
1590 	/*
1591 	 * Set up the scheduler specific parts of proc0.
1592 	 */
1593 	proc0.p_sched = NULL; /* XXX */
1594 	thread0.td_sched = &td_sched0;
1595 	td_sched0.ts_ltick = ticks;
1596 	td_sched0.ts_ftick = ticks;
1597 	td_sched0.ts_slice = 0;
1598 }
1599 
1600 /*
1601  * This is only somewhat accurate since given many processes of the same
1602  * priority they will switch when their slices run out, which will be
1603  * at most sched_slice stathz ticks.
1604  */
1605 int
1606 sched_rr_interval(void)
1607 {
1608 
1609 	/* Convert sched_slice from stathz to hz. */
1610 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1611 }
1612 
1613 /*
1614  * Update the percent cpu tracking information when it is requested or
1615  * the total history exceeds the maximum.  We keep a sliding history of
1616  * tick counts that slowly decays.  This is less precise than the 4BSD
1617  * mechanism since it happens with less regular and frequent events.
1618  */
1619 static void
1620 sched_pctcpu_update(struct td_sched *ts, int run)
1621 {
1622 	int t = ticks;
1623 
1624 	if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1625 		ts->ts_ticks = 0;
1626 		ts->ts_ftick = t - SCHED_TICK_TARG;
1627 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1628 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1629 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1630 		ts->ts_ftick = t - SCHED_TICK_TARG;
1631 	}
1632 	if (run)
1633 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1634 	ts->ts_ltick = t;
1635 }
1636 
1637 /*
1638  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1639  * if necessary.  This is the back-end for several priority related
1640  * functions.
1641  */
1642 static void
1643 sched_thread_priority(struct thread *td, u_char prio)
1644 {
1645 	struct td_sched *ts;
1646 	struct tdq *tdq;
1647 	int oldpri;
1648 
1649 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1650 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1651 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1652 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1653 	if (td != curthread && prio < td->td_priority) {
1654 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1655 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1656 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1657 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1658 		    curthread);
1659 	}
1660 	ts = td->td_sched;
1661 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1662 	if (td->td_priority == prio)
1663 		return;
1664 	/*
1665 	 * If the priority has been elevated due to priority
1666 	 * propagation, we may have to move ourselves to a new
1667 	 * queue.  This could be optimized to not re-add in some
1668 	 * cases.
1669 	 */
1670 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1671 		sched_rem(td);
1672 		td->td_priority = prio;
1673 		sched_add(td, SRQ_BORROWING);
1674 		return;
1675 	}
1676 	/*
1677 	 * If the thread is currently running we may have to adjust the lowpri
1678 	 * information so other cpus are aware of our current priority.
1679 	 */
1680 	if (TD_IS_RUNNING(td)) {
1681 		tdq = TDQ_CPU(ts->ts_cpu);
1682 		oldpri = td->td_priority;
1683 		td->td_priority = prio;
1684 		if (prio < tdq->tdq_lowpri)
1685 			tdq->tdq_lowpri = prio;
1686 		else if (tdq->tdq_lowpri == oldpri)
1687 			tdq_setlowpri(tdq, td);
1688 		return;
1689 	}
1690 	td->td_priority = prio;
1691 }
1692 
1693 /*
1694  * Update a thread's priority when it is lent another thread's
1695  * priority.
1696  */
1697 void
1698 sched_lend_prio(struct thread *td, u_char prio)
1699 {
1700 
1701 	td->td_flags |= TDF_BORROWING;
1702 	sched_thread_priority(td, prio);
1703 }
1704 
1705 /*
1706  * Restore a thread's priority when priority propagation is
1707  * over.  The prio argument is the minimum priority the thread
1708  * needs to have to satisfy other possible priority lending
1709  * requests.  If the thread's regular priority is less
1710  * important than prio, the thread will keep a priority boost
1711  * of prio.
1712  */
1713 void
1714 sched_unlend_prio(struct thread *td, u_char prio)
1715 {
1716 	u_char base_pri;
1717 
1718 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1719 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1720 		base_pri = td->td_user_pri;
1721 	else
1722 		base_pri = td->td_base_pri;
1723 	if (prio >= base_pri) {
1724 		td->td_flags &= ~TDF_BORROWING;
1725 		sched_thread_priority(td, base_pri);
1726 	} else
1727 		sched_lend_prio(td, prio);
1728 }
1729 
1730 /*
1731  * Standard entry for setting the priority to an absolute value.
1732  */
1733 void
1734 sched_prio(struct thread *td, u_char prio)
1735 {
1736 	u_char oldprio;
1737 
1738 	/* First, update the base priority. */
1739 	td->td_base_pri = prio;
1740 
1741 	/*
1742 	 * If the thread is borrowing another thread's priority, don't
1743 	 * ever lower the priority.
1744 	 */
1745 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1746 		return;
1747 
1748 	/* Change the real priority. */
1749 	oldprio = td->td_priority;
1750 	sched_thread_priority(td, prio);
1751 
1752 	/*
1753 	 * If the thread is on a turnstile, then let the turnstile update
1754 	 * its state.
1755 	 */
1756 	if (TD_ON_LOCK(td) && oldprio != prio)
1757 		turnstile_adjust(td, oldprio);
1758 }
1759 
1760 /*
1761  * Set the base user priority, does not effect current running priority.
1762  */
1763 void
1764 sched_user_prio(struct thread *td, u_char prio)
1765 {
1766 
1767 	td->td_base_user_pri = prio;
1768 	if (td->td_lend_user_pri <= prio)
1769 		return;
1770 	td->td_user_pri = prio;
1771 }
1772 
1773 void
1774 sched_lend_user_prio(struct thread *td, u_char prio)
1775 {
1776 
1777 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1778 	td->td_lend_user_pri = prio;
1779 	td->td_user_pri = min(prio, td->td_base_user_pri);
1780 	if (td->td_priority > td->td_user_pri)
1781 		sched_prio(td, td->td_user_pri);
1782 	else if (td->td_priority != td->td_user_pri)
1783 		td->td_flags |= TDF_NEEDRESCHED;
1784 }
1785 
1786 /*
1787  * Handle migration from sched_switch().  This happens only for
1788  * cpu binding.
1789  */
1790 static struct mtx *
1791 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1792 {
1793 	struct tdq *tdn;
1794 
1795 	tdn = TDQ_CPU(td->td_sched->ts_cpu);
1796 #ifdef SMP
1797 	tdq_load_rem(tdq, td);
1798 	/*
1799 	 * Do the lock dance required to avoid LOR.  We grab an extra
1800 	 * spinlock nesting to prevent preemption while we're
1801 	 * not holding either run-queue lock.
1802 	 */
1803 	spinlock_enter();
1804 	thread_lock_block(td);	/* This releases the lock on tdq. */
1805 
1806 	/*
1807 	 * Acquire both run-queue locks before placing the thread on the new
1808 	 * run-queue to avoid deadlocks created by placing a thread with a
1809 	 * blocked lock on the run-queue of a remote processor.  The deadlock
1810 	 * occurs when a third processor attempts to lock the two queues in
1811 	 * question while the target processor is spinning with its own
1812 	 * run-queue lock held while waiting for the blocked lock to clear.
1813 	 */
1814 	tdq_lock_pair(tdn, tdq);
1815 	tdq_add(tdn, td, flags);
1816 	tdq_notify(tdn, td);
1817 	TDQ_UNLOCK(tdn);
1818 	spinlock_exit();
1819 #endif
1820 	return (TDQ_LOCKPTR(tdn));
1821 }
1822 
1823 /*
1824  * Variadic version of thread_lock_unblock() that does not assume td_lock
1825  * is blocked.
1826  */
1827 static inline void
1828 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1829 {
1830 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1831 	    (uintptr_t)mtx);
1832 }
1833 
1834 /*
1835  * Switch threads.  This function has to handle threads coming in while
1836  * blocked for some reason, running, or idle.  It also must deal with
1837  * migrating a thread from one queue to another as running threads may
1838  * be assigned elsewhere via binding.
1839  */
1840 void
1841 sched_switch(struct thread *td, struct thread *newtd, int flags)
1842 {
1843 	struct tdq *tdq;
1844 	struct td_sched *ts;
1845 	struct mtx *mtx;
1846 	int srqflag;
1847 	int cpuid, preempted;
1848 
1849 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1850 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1851 
1852 	cpuid = PCPU_GET(cpuid);
1853 	tdq = TDQ_CPU(cpuid);
1854 	ts = td->td_sched;
1855 	mtx = td->td_lock;
1856 	sched_pctcpu_update(ts, 1);
1857 	ts->ts_rltick = ticks;
1858 	td->td_lastcpu = td->td_oncpu;
1859 	td->td_oncpu = NOCPU;
1860 	preempted = !(td->td_flags & TDF_SLICEEND);
1861 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1862 	td->td_owepreempt = 0;
1863 	if (!TD_IS_IDLETHREAD(td))
1864 		tdq->tdq_switchcnt++;
1865 	/*
1866 	 * The lock pointer in an idle thread should never change.  Reset it
1867 	 * to CAN_RUN as well.
1868 	 */
1869 	if (TD_IS_IDLETHREAD(td)) {
1870 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1871 		TD_SET_CAN_RUN(td);
1872 	} else if (TD_IS_RUNNING(td)) {
1873 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1874 		srqflag = preempted ?
1875 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1876 		    SRQ_OURSELF|SRQ_YIELDING;
1877 #ifdef SMP
1878 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1879 			ts->ts_cpu = sched_pickcpu(td, 0);
1880 #endif
1881 		if (ts->ts_cpu == cpuid)
1882 			tdq_runq_add(tdq, td, srqflag);
1883 		else {
1884 			KASSERT(THREAD_CAN_MIGRATE(td) ||
1885 			    (ts->ts_flags & TSF_BOUND) != 0,
1886 			    ("Thread %p shouldn't migrate", td));
1887 			mtx = sched_switch_migrate(tdq, td, srqflag);
1888 		}
1889 	} else {
1890 		/* This thread must be going to sleep. */
1891 		TDQ_LOCK(tdq);
1892 		mtx = thread_lock_block(td);
1893 		tdq_load_rem(tdq, td);
1894 	}
1895 	/*
1896 	 * We enter here with the thread blocked and assigned to the
1897 	 * appropriate cpu run-queue or sleep-queue and with the current
1898 	 * thread-queue locked.
1899 	 */
1900 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1901 	newtd = choosethread();
1902 	/*
1903 	 * Call the MD code to switch contexts if necessary.
1904 	 */
1905 	if (td != newtd) {
1906 #ifdef	HWPMC_HOOKS
1907 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1908 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1909 #endif
1910 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1911 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1912 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1913 		sched_pctcpu_update(newtd->td_sched, 0);
1914 
1915 #ifdef KDTRACE_HOOKS
1916 		/*
1917 		 * If DTrace has set the active vtime enum to anything
1918 		 * other than INACTIVE (0), then it should have set the
1919 		 * function to call.
1920 		 */
1921 		if (dtrace_vtime_active)
1922 			(*dtrace_vtime_switch_func)(newtd);
1923 #endif
1924 
1925 		cpu_switch(td, newtd, mtx);
1926 		/*
1927 		 * We may return from cpu_switch on a different cpu.  However,
1928 		 * we always return with td_lock pointing to the current cpu's
1929 		 * run queue lock.
1930 		 */
1931 		cpuid = PCPU_GET(cpuid);
1932 		tdq = TDQ_CPU(cpuid);
1933 		lock_profile_obtain_lock_success(
1934 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1935 
1936 		SDT_PROBE0(sched, , , on__cpu);
1937 #ifdef	HWPMC_HOOKS
1938 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1939 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1940 #endif
1941 	} else {
1942 		thread_unblock_switch(td, mtx);
1943 		SDT_PROBE0(sched, , , remain__cpu);
1944 	}
1945 	/*
1946 	 * Assert that all went well and return.
1947 	 */
1948 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1949 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1950 	td->td_oncpu = cpuid;
1951 }
1952 
1953 /*
1954  * Adjust thread priorities as a result of a nice request.
1955  */
1956 void
1957 sched_nice(struct proc *p, int nice)
1958 {
1959 	struct thread *td;
1960 
1961 	PROC_LOCK_ASSERT(p, MA_OWNED);
1962 
1963 	p->p_nice = nice;
1964 	FOREACH_THREAD_IN_PROC(p, td) {
1965 		thread_lock(td);
1966 		sched_priority(td);
1967 		sched_prio(td, td->td_base_user_pri);
1968 		thread_unlock(td);
1969 	}
1970 }
1971 
1972 /*
1973  * Record the sleep time for the interactivity scorer.
1974  */
1975 void
1976 sched_sleep(struct thread *td, int prio)
1977 {
1978 
1979 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1980 
1981 	td->td_slptick = ticks;
1982 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1983 		td->td_flags |= TDF_CANSWAP;
1984 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1985 		return;
1986 	if (static_boost == 1 && prio)
1987 		sched_prio(td, prio);
1988 	else if (static_boost && td->td_priority > static_boost)
1989 		sched_prio(td, static_boost);
1990 }
1991 
1992 /*
1993  * Schedule a thread to resume execution and record how long it voluntarily
1994  * slept.  We also update the pctcpu, interactivity, and priority.
1995  */
1996 void
1997 sched_wakeup(struct thread *td)
1998 {
1999 	struct td_sched *ts;
2000 	int slptick;
2001 
2002 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2003 	ts = td->td_sched;
2004 	td->td_flags &= ~TDF_CANSWAP;
2005 	/*
2006 	 * If we slept for more than a tick update our interactivity and
2007 	 * priority.
2008 	 */
2009 	slptick = td->td_slptick;
2010 	td->td_slptick = 0;
2011 	if (slptick && slptick != ticks) {
2012 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2013 		sched_interact_update(td);
2014 		sched_pctcpu_update(ts, 0);
2015 	}
2016 	/*
2017 	 * Reset the slice value since we slept and advanced the round-robin.
2018 	 */
2019 	ts->ts_slice = 0;
2020 	sched_add(td, SRQ_BORING);
2021 }
2022 
2023 /*
2024  * Penalize the parent for creating a new child and initialize the child's
2025  * priority.
2026  */
2027 void
2028 sched_fork(struct thread *td, struct thread *child)
2029 {
2030 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2031 	sched_pctcpu_update(td->td_sched, 1);
2032 	sched_fork_thread(td, child);
2033 	/*
2034 	 * Penalize the parent and child for forking.
2035 	 */
2036 	sched_interact_fork(child);
2037 	sched_priority(child);
2038 	td->td_sched->ts_runtime += tickincr;
2039 	sched_interact_update(td);
2040 	sched_priority(td);
2041 }
2042 
2043 /*
2044  * Fork a new thread, may be within the same process.
2045  */
2046 void
2047 sched_fork_thread(struct thread *td, struct thread *child)
2048 {
2049 	struct td_sched *ts;
2050 	struct td_sched *ts2;
2051 	struct tdq *tdq;
2052 
2053 	tdq = TDQ_SELF();
2054 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2055 	/*
2056 	 * Initialize child.
2057 	 */
2058 	ts = td->td_sched;
2059 	ts2 = child->td_sched;
2060 	child->td_lock = TDQ_LOCKPTR(tdq);
2061 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2062 	ts2->ts_cpu = ts->ts_cpu;
2063 	ts2->ts_flags = 0;
2064 	/*
2065 	 * Grab our parents cpu estimation information.
2066 	 */
2067 	ts2->ts_ticks = ts->ts_ticks;
2068 	ts2->ts_ltick = ts->ts_ltick;
2069 	ts2->ts_ftick = ts->ts_ftick;
2070 	/*
2071 	 * Do not inherit any borrowed priority from the parent.
2072 	 */
2073 	child->td_priority = child->td_base_pri;
2074 	/*
2075 	 * And update interactivity score.
2076 	 */
2077 	ts2->ts_slptime = ts->ts_slptime;
2078 	ts2->ts_runtime = ts->ts_runtime;
2079 	/* Attempt to quickly learn interactivity. */
2080 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2081 #ifdef KTR
2082 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2083 #endif
2084 }
2085 
2086 /*
2087  * Adjust the priority class of a thread.
2088  */
2089 void
2090 sched_class(struct thread *td, int class)
2091 {
2092 
2093 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2094 	if (td->td_pri_class == class)
2095 		return;
2096 	td->td_pri_class = class;
2097 }
2098 
2099 /*
2100  * Return some of the child's priority and interactivity to the parent.
2101  */
2102 void
2103 sched_exit(struct proc *p, struct thread *child)
2104 {
2105 	struct thread *td;
2106 
2107 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2108 	    "prio:%d", child->td_priority);
2109 	PROC_LOCK_ASSERT(p, MA_OWNED);
2110 	td = FIRST_THREAD_IN_PROC(p);
2111 	sched_exit_thread(td, child);
2112 }
2113 
2114 /*
2115  * Penalize another thread for the time spent on this one.  This helps to
2116  * worsen the priority and interactivity of processes which schedule batch
2117  * jobs such as make.  This has little effect on the make process itself but
2118  * causes new processes spawned by it to receive worse scores immediately.
2119  */
2120 void
2121 sched_exit_thread(struct thread *td, struct thread *child)
2122 {
2123 
2124 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2125 	    "prio:%d", child->td_priority);
2126 	/*
2127 	 * Give the child's runtime to the parent without returning the
2128 	 * sleep time as a penalty to the parent.  This causes shells that
2129 	 * launch expensive things to mark their children as expensive.
2130 	 */
2131 	thread_lock(td);
2132 	td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2133 	sched_interact_update(td);
2134 	sched_priority(td);
2135 	thread_unlock(td);
2136 }
2137 
2138 void
2139 sched_preempt(struct thread *td)
2140 {
2141 	struct tdq *tdq;
2142 
2143 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2144 
2145 	thread_lock(td);
2146 	tdq = TDQ_SELF();
2147 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2148 	tdq->tdq_ipipending = 0;
2149 	if (td->td_priority > tdq->tdq_lowpri) {
2150 		int flags;
2151 
2152 		flags = SW_INVOL | SW_PREEMPT;
2153 		if (td->td_critnest > 1)
2154 			td->td_owepreempt = 1;
2155 		else if (TD_IS_IDLETHREAD(td))
2156 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2157 		else
2158 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2159 	}
2160 	thread_unlock(td);
2161 }
2162 
2163 /*
2164  * Fix priorities on return to user-space.  Priorities may be elevated due
2165  * to static priorities in msleep() or similar.
2166  */
2167 void
2168 sched_userret(struct thread *td)
2169 {
2170 	/*
2171 	 * XXX we cheat slightly on the locking here to avoid locking in
2172 	 * the usual case.  Setting td_priority here is essentially an
2173 	 * incomplete workaround for not setting it properly elsewhere.
2174 	 * Now that some interrupt handlers are threads, not setting it
2175 	 * properly elsewhere can clobber it in the window between setting
2176 	 * it here and returning to user mode, so don't waste time setting
2177 	 * it perfectly here.
2178 	 */
2179 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2180 	    ("thread with borrowed priority returning to userland"));
2181 	if (td->td_priority != td->td_user_pri) {
2182 		thread_lock(td);
2183 		td->td_priority = td->td_user_pri;
2184 		td->td_base_pri = td->td_user_pri;
2185 		tdq_setlowpri(TDQ_SELF(), td);
2186 		thread_unlock(td);
2187         }
2188 }
2189 
2190 /*
2191  * Handle a stathz tick.  This is really only relevant for timeshare
2192  * threads.
2193  */
2194 void
2195 sched_clock(struct thread *td)
2196 {
2197 	struct tdq *tdq;
2198 	struct td_sched *ts;
2199 
2200 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2201 	tdq = TDQ_SELF();
2202 #ifdef SMP
2203 	/*
2204 	 * We run the long term load balancer infrequently on the first cpu.
2205 	 */
2206 	if (balance_tdq == tdq) {
2207 		if (balance_ticks && --balance_ticks == 0)
2208 			sched_balance();
2209 	}
2210 #endif
2211 	/*
2212 	 * Save the old switch count so we have a record of the last ticks
2213 	 * activity.   Initialize the new switch count based on our load.
2214 	 * If there is some activity seed it to reflect that.
2215 	 */
2216 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2217 	tdq->tdq_switchcnt = tdq->tdq_load;
2218 	/*
2219 	 * Advance the insert index once for each tick to ensure that all
2220 	 * threads get a chance to run.
2221 	 */
2222 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2223 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2224 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2225 			tdq->tdq_ridx = tdq->tdq_idx;
2226 	}
2227 	ts = td->td_sched;
2228 	sched_pctcpu_update(ts, 1);
2229 	if (td->td_pri_class & PRI_FIFO_BIT)
2230 		return;
2231 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2232 		/*
2233 		 * We used a tick; charge it to the thread so
2234 		 * that we can compute our interactivity.
2235 		 */
2236 		td->td_sched->ts_runtime += tickincr;
2237 		sched_interact_update(td);
2238 		sched_priority(td);
2239 	}
2240 
2241 	/*
2242 	 * Force a context switch if the current thread has used up a full
2243 	 * time slice (default is 100ms).
2244 	 */
2245 	if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2246 		ts->ts_slice = 0;
2247 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2248 	}
2249 }
2250 
2251 /*
2252  * Called once per hz tick.
2253  */
2254 void
2255 sched_tick(int cnt)
2256 {
2257 
2258 }
2259 
2260 /*
2261  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2262  * cooperative idle threads.
2263  */
2264 int
2265 sched_runnable(void)
2266 {
2267 	struct tdq *tdq;
2268 	int load;
2269 
2270 	load = 1;
2271 
2272 	tdq = TDQ_SELF();
2273 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2274 		if (tdq->tdq_load > 0)
2275 			goto out;
2276 	} else
2277 		if (tdq->tdq_load - 1 > 0)
2278 			goto out;
2279 	load = 0;
2280 out:
2281 	return (load);
2282 }
2283 
2284 /*
2285  * Choose the highest priority thread to run.  The thread is removed from
2286  * the run-queue while running however the load remains.  For SMP we set
2287  * the tdq in the global idle bitmask if it idles here.
2288  */
2289 struct thread *
2290 sched_choose(void)
2291 {
2292 	struct thread *td;
2293 	struct tdq *tdq;
2294 
2295 	tdq = TDQ_SELF();
2296 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2297 	td = tdq_choose(tdq);
2298 	if (td) {
2299 		tdq_runq_rem(tdq, td);
2300 		tdq->tdq_lowpri = td->td_priority;
2301 		return (td);
2302 	}
2303 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2304 	return (PCPU_GET(idlethread));
2305 }
2306 
2307 /*
2308  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2309  * we always request it once we exit a critical section.
2310  */
2311 static inline void
2312 sched_setpreempt(struct thread *td)
2313 {
2314 	struct thread *ctd;
2315 	int cpri;
2316 	int pri;
2317 
2318 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2319 
2320 	ctd = curthread;
2321 	pri = td->td_priority;
2322 	cpri = ctd->td_priority;
2323 	if (pri < cpri)
2324 		ctd->td_flags |= TDF_NEEDRESCHED;
2325 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2326 		return;
2327 	if (!sched_shouldpreempt(pri, cpri, 0))
2328 		return;
2329 	ctd->td_owepreempt = 1;
2330 }
2331 
2332 /*
2333  * Add a thread to a thread queue.  Select the appropriate runq and add the
2334  * thread to it.  This is the internal function called when the tdq is
2335  * predetermined.
2336  */
2337 void
2338 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2339 {
2340 
2341 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2342 	KASSERT((td->td_inhibitors == 0),
2343 	    ("sched_add: trying to run inhibited thread"));
2344 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2345 	    ("sched_add: bad thread state"));
2346 	KASSERT(td->td_flags & TDF_INMEM,
2347 	    ("sched_add: thread swapped out"));
2348 
2349 	if (td->td_priority < tdq->tdq_lowpri)
2350 		tdq->tdq_lowpri = td->td_priority;
2351 	tdq_runq_add(tdq, td, flags);
2352 	tdq_load_add(tdq, td);
2353 }
2354 
2355 /*
2356  * Select the target thread queue and add a thread to it.  Request
2357  * preemption or IPI a remote processor if required.
2358  */
2359 void
2360 sched_add(struct thread *td, int flags)
2361 {
2362 	struct tdq *tdq;
2363 #ifdef SMP
2364 	int cpu;
2365 #endif
2366 
2367 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2368 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2369 	    sched_tdname(curthread));
2370 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2371 	    KTR_ATTR_LINKED, sched_tdname(td));
2372 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2373 	    flags & SRQ_PREEMPTED);
2374 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2375 	/*
2376 	 * Recalculate the priority before we select the target cpu or
2377 	 * run-queue.
2378 	 */
2379 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2380 		sched_priority(td);
2381 #ifdef SMP
2382 	/*
2383 	 * Pick the destination cpu and if it isn't ours transfer to the
2384 	 * target cpu.
2385 	 */
2386 	cpu = sched_pickcpu(td, flags);
2387 	tdq = sched_setcpu(td, cpu, flags);
2388 	tdq_add(tdq, td, flags);
2389 	if (cpu != PCPU_GET(cpuid)) {
2390 		tdq_notify(tdq, td);
2391 		return;
2392 	}
2393 #else
2394 	tdq = TDQ_SELF();
2395 	TDQ_LOCK(tdq);
2396 	/*
2397 	 * Now that the thread is moving to the run-queue, set the lock
2398 	 * to the scheduler's lock.
2399 	 */
2400 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2401 	tdq_add(tdq, td, flags);
2402 #endif
2403 	if (!(flags & SRQ_YIELDING))
2404 		sched_setpreempt(td);
2405 }
2406 
2407 /*
2408  * Remove a thread from a run-queue without running it.  This is used
2409  * when we're stealing a thread from a remote queue.  Otherwise all threads
2410  * exit by calling sched_exit_thread() and sched_throw() themselves.
2411  */
2412 void
2413 sched_rem(struct thread *td)
2414 {
2415 	struct tdq *tdq;
2416 
2417 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2418 	    "prio:%d", td->td_priority);
2419 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2420 	tdq = TDQ_CPU(td->td_sched->ts_cpu);
2421 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2422 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2423 	KASSERT(TD_ON_RUNQ(td),
2424 	    ("sched_rem: thread not on run queue"));
2425 	tdq_runq_rem(tdq, td);
2426 	tdq_load_rem(tdq, td);
2427 	TD_SET_CAN_RUN(td);
2428 	if (td->td_priority == tdq->tdq_lowpri)
2429 		tdq_setlowpri(tdq, NULL);
2430 }
2431 
2432 /*
2433  * Fetch cpu utilization information.  Updates on demand.
2434  */
2435 fixpt_t
2436 sched_pctcpu(struct thread *td)
2437 {
2438 	fixpt_t pctcpu;
2439 	struct td_sched *ts;
2440 
2441 	pctcpu = 0;
2442 	ts = td->td_sched;
2443 	if (ts == NULL)
2444 		return (0);
2445 
2446 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2447 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2448 	if (ts->ts_ticks) {
2449 		int rtick;
2450 
2451 		/* How many rtick per second ? */
2452 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2453 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2454 	}
2455 
2456 	return (pctcpu);
2457 }
2458 
2459 /*
2460  * Enforce affinity settings for a thread.  Called after adjustments to
2461  * cpumask.
2462  */
2463 void
2464 sched_affinity(struct thread *td)
2465 {
2466 #ifdef SMP
2467 	struct td_sched *ts;
2468 
2469 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2470 	ts = td->td_sched;
2471 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2472 		return;
2473 	if (TD_ON_RUNQ(td)) {
2474 		sched_rem(td);
2475 		sched_add(td, SRQ_BORING);
2476 		return;
2477 	}
2478 	if (!TD_IS_RUNNING(td))
2479 		return;
2480 	/*
2481 	 * Force a switch before returning to userspace.  If the
2482 	 * target thread is not running locally send an ipi to force
2483 	 * the issue.
2484 	 */
2485 	td->td_flags |= TDF_NEEDRESCHED;
2486 	if (td != curthread)
2487 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2488 #endif
2489 }
2490 
2491 /*
2492  * Bind a thread to a target cpu.
2493  */
2494 void
2495 sched_bind(struct thread *td, int cpu)
2496 {
2497 	struct td_sched *ts;
2498 
2499 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2500 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2501 	ts = td->td_sched;
2502 	if (ts->ts_flags & TSF_BOUND)
2503 		sched_unbind(td);
2504 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2505 	ts->ts_flags |= TSF_BOUND;
2506 	sched_pin();
2507 	if (PCPU_GET(cpuid) == cpu)
2508 		return;
2509 	ts->ts_cpu = cpu;
2510 	/* When we return from mi_switch we'll be on the correct cpu. */
2511 	mi_switch(SW_VOL, NULL);
2512 }
2513 
2514 /*
2515  * Release a bound thread.
2516  */
2517 void
2518 sched_unbind(struct thread *td)
2519 {
2520 	struct td_sched *ts;
2521 
2522 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2523 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2524 	ts = td->td_sched;
2525 	if ((ts->ts_flags & TSF_BOUND) == 0)
2526 		return;
2527 	ts->ts_flags &= ~TSF_BOUND;
2528 	sched_unpin();
2529 }
2530 
2531 int
2532 sched_is_bound(struct thread *td)
2533 {
2534 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2535 	return (td->td_sched->ts_flags & TSF_BOUND);
2536 }
2537 
2538 /*
2539  * Basic yield call.
2540  */
2541 void
2542 sched_relinquish(struct thread *td)
2543 {
2544 	thread_lock(td);
2545 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2546 	thread_unlock(td);
2547 }
2548 
2549 /*
2550  * Return the total system load.
2551  */
2552 int
2553 sched_load(void)
2554 {
2555 #ifdef SMP
2556 	int total;
2557 	int i;
2558 
2559 	total = 0;
2560 	CPU_FOREACH(i)
2561 		total += TDQ_CPU(i)->tdq_sysload;
2562 	return (total);
2563 #else
2564 	return (TDQ_SELF()->tdq_sysload);
2565 #endif
2566 }
2567 
2568 int
2569 sched_sizeof_proc(void)
2570 {
2571 	return (sizeof(struct proc));
2572 }
2573 
2574 int
2575 sched_sizeof_thread(void)
2576 {
2577 	return (sizeof(struct thread) + sizeof(struct td_sched));
2578 }
2579 
2580 #ifdef SMP
2581 #define	TDQ_IDLESPIN(tdq)						\
2582     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2583 #else
2584 #define	TDQ_IDLESPIN(tdq)	1
2585 #endif
2586 
2587 /*
2588  * The actual idle process.
2589  */
2590 void
2591 sched_idletd(void *dummy)
2592 {
2593 	struct thread *td;
2594 	struct tdq *tdq;
2595 	int oldswitchcnt, switchcnt;
2596 	int i;
2597 
2598 	mtx_assert(&Giant, MA_NOTOWNED);
2599 	td = curthread;
2600 	tdq = TDQ_SELF();
2601 	THREAD_NO_SLEEPING();
2602 	oldswitchcnt = -1;
2603 	for (;;) {
2604 		if (tdq->tdq_load) {
2605 			thread_lock(td);
2606 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2607 			thread_unlock(td);
2608 		}
2609 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2610 #ifdef SMP
2611 		if (switchcnt != oldswitchcnt) {
2612 			oldswitchcnt = switchcnt;
2613 			if (tdq_idled(tdq) == 0)
2614 				continue;
2615 		}
2616 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2617 #else
2618 		oldswitchcnt = switchcnt;
2619 #endif
2620 		/*
2621 		 * If we're switching very frequently, spin while checking
2622 		 * for load rather than entering a low power state that
2623 		 * may require an IPI.  However, don't do any busy
2624 		 * loops while on SMT machines as this simply steals
2625 		 * cycles from cores doing useful work.
2626 		 */
2627 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2628 			for (i = 0; i < sched_idlespins; i++) {
2629 				if (tdq->tdq_load)
2630 					break;
2631 				cpu_spinwait();
2632 			}
2633 		}
2634 
2635 		/* If there was context switch during spin, restart it. */
2636 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2637 		if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2638 			continue;
2639 
2640 		/* Run main MD idle handler. */
2641 		tdq->tdq_cpu_idle = 1;
2642 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2643 		tdq->tdq_cpu_idle = 0;
2644 
2645 		/*
2646 		 * Account thread-less hardware interrupts and
2647 		 * other wakeup reasons equal to context switches.
2648 		 */
2649 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2650 		if (switchcnt != oldswitchcnt)
2651 			continue;
2652 		tdq->tdq_switchcnt++;
2653 		oldswitchcnt++;
2654 	}
2655 }
2656 
2657 /*
2658  * A CPU is entering for the first time or a thread is exiting.
2659  */
2660 void
2661 sched_throw(struct thread *td)
2662 {
2663 	struct thread *newtd;
2664 	struct tdq *tdq;
2665 
2666 	tdq = TDQ_SELF();
2667 	if (td == NULL) {
2668 		/* Correct spinlock nesting and acquire the correct lock. */
2669 		TDQ_LOCK(tdq);
2670 		spinlock_exit();
2671 		PCPU_SET(switchtime, cpu_ticks());
2672 		PCPU_SET(switchticks, ticks);
2673 	} else {
2674 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2675 		tdq_load_rem(tdq, td);
2676 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2677 	}
2678 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2679 	newtd = choosethread();
2680 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2681 	cpu_throw(td, newtd);		/* doesn't return */
2682 }
2683 
2684 /*
2685  * This is called from fork_exit().  Just acquire the correct locks and
2686  * let fork do the rest of the work.
2687  */
2688 void
2689 sched_fork_exit(struct thread *td)
2690 {
2691 	struct tdq *tdq;
2692 	int cpuid;
2693 
2694 	/*
2695 	 * Finish setting up thread glue so that it begins execution in a
2696 	 * non-nested critical section with the scheduler lock held.
2697 	 */
2698 	cpuid = PCPU_GET(cpuid);
2699 	tdq = TDQ_CPU(cpuid);
2700 	if (TD_IS_IDLETHREAD(td))
2701 		td->td_lock = TDQ_LOCKPTR(tdq);
2702 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2703 	td->td_oncpu = cpuid;
2704 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2705 	lock_profile_obtain_lock_success(
2706 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2707 }
2708 
2709 /*
2710  * Create on first use to catch odd startup conditons.
2711  */
2712 char *
2713 sched_tdname(struct thread *td)
2714 {
2715 #ifdef KTR
2716 	struct td_sched *ts;
2717 
2718 	ts = td->td_sched;
2719 	if (ts->ts_name[0] == '\0')
2720 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2721 		    "%s tid %d", td->td_name, td->td_tid);
2722 	return (ts->ts_name);
2723 #else
2724 	return (td->td_name);
2725 #endif
2726 }
2727 
2728 #ifdef KTR
2729 void
2730 sched_clear_tdname(struct thread *td)
2731 {
2732 	struct td_sched *ts;
2733 
2734 	ts = td->td_sched;
2735 	ts->ts_name[0] = '\0';
2736 }
2737 #endif
2738 
2739 #ifdef SMP
2740 
2741 /*
2742  * Build the CPU topology dump string. Is recursively called to collect
2743  * the topology tree.
2744  */
2745 static int
2746 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2747     int indent)
2748 {
2749 	char cpusetbuf[CPUSETBUFSIZ];
2750 	int i, first;
2751 
2752 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2753 	    "", 1 + indent / 2, cg->cg_level);
2754 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2755 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2756 	first = TRUE;
2757 	for (i = 0; i < MAXCPU; i++) {
2758 		if (CPU_ISSET(i, &cg->cg_mask)) {
2759 			if (!first)
2760 				sbuf_printf(sb, ", ");
2761 			else
2762 				first = FALSE;
2763 			sbuf_printf(sb, "%d", i);
2764 		}
2765 	}
2766 	sbuf_printf(sb, "</cpu>\n");
2767 
2768 	if (cg->cg_flags != 0) {
2769 		sbuf_printf(sb, "%*s <flags>", indent, "");
2770 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2771 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2772 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2773 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2774 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2775 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2776 		sbuf_printf(sb, "</flags>\n");
2777 	}
2778 
2779 	if (cg->cg_children > 0) {
2780 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2781 		for (i = 0; i < cg->cg_children; i++)
2782 			sysctl_kern_sched_topology_spec_internal(sb,
2783 			    &cg->cg_child[i], indent+2);
2784 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2785 	}
2786 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2787 	return (0);
2788 }
2789 
2790 /*
2791  * Sysctl handler for retrieving topology dump. It's a wrapper for
2792  * the recursive sysctl_kern_smp_topology_spec_internal().
2793  */
2794 static int
2795 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2796 {
2797 	struct sbuf *topo;
2798 	int err;
2799 
2800 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2801 
2802 	topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2803 	if (topo == NULL)
2804 		return (ENOMEM);
2805 
2806 	sbuf_printf(topo, "<groups>\n");
2807 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2808 	sbuf_printf(topo, "</groups>\n");
2809 
2810 	if (err == 0) {
2811 		sbuf_finish(topo);
2812 		err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2813 	}
2814 	sbuf_delete(topo);
2815 	return (err);
2816 }
2817 
2818 #endif
2819 
2820 static int
2821 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2822 {
2823 	int error, new_val, period;
2824 
2825 	period = 1000000 / realstathz;
2826 	new_val = period * sched_slice;
2827 	error = sysctl_handle_int(oidp, &new_val, 0, req);
2828 	if (error != 0 || req->newptr == NULL)
2829 		return (error);
2830 	if (new_val <= 0)
2831 		return (EINVAL);
2832 	sched_slice = imax(1, (new_val + period / 2) / period);
2833 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
2834 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2835 	    realstathz);
2836 	return (0);
2837 }
2838 
2839 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2840 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2841     "Scheduler name");
2842 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2843     NULL, 0, sysctl_kern_quantum, "I",
2844     "Quantum for timeshare threads in microseconds");
2845 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2846     "Quantum for timeshare threads in stathz ticks");
2847 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2848     "Interactivity score threshold");
2849 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2850     &preempt_thresh, 0,
2851     "Maximal (lowest) priority for preemption");
2852 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2853     "Assign static kernel priorities to sleeping threads");
2854 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2855     "Number of times idle thread will spin waiting for new work");
2856 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2857     &sched_idlespinthresh, 0,
2858     "Threshold before we will permit idle thread spinning");
2859 #ifdef SMP
2860 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2861     "Number of hz ticks to keep thread affinity for");
2862 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2863     "Enables the long-term load balancer");
2864 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2865     &balance_interval, 0,
2866     "Average period in stathz ticks to run the long-term balancer");
2867 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2868     "Attempts to steal work from other cores before idling");
2869 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2870     "Minimum load on remote CPU before we'll steal");
2871 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2872     CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2873     "XML dump of detected CPU topology");
2874 #endif
2875 
2876 /* ps compat.  All cpu percentages from ULE are weighted. */
2877 static int ccpu = 0;
2878 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2879