xref: /freebsd/sys/kern/sched_ule.c (revision 864c53ead899f7838cd2e1cca3b485a4a82f5cdc)
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 __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low,
626     struct cpu_search *high, const int match);
627 int cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low);
628 int cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high);
629 int cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
630     struct cpu_search *high);
631 
632 /*
633  * Search the tree of cpu_groups for the lowest or highest loaded cpu
634  * according to the match argument.  This routine actually compares the
635  * load on all paths through the tree and finds the least loaded cpu on
636  * the least loaded path, which may differ from the least loaded cpu in
637  * the system.  This balances work among caches and busses.
638  *
639  * This inline is instantiated in three forms below using constants for the
640  * match argument.  It is reduced to the minimum set for each case.  It is
641  * also recursive to the depth of the tree.
642  */
643 static __inline int
644 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
645     struct cpu_search *high, const int match)
646 {
647 	struct cpu_search lgroup;
648 	struct cpu_search hgroup;
649 	cpuset_t cpumask;
650 	struct cpu_group *child;
651 	struct tdq *tdq;
652 	int cpu, i, hload, lload, load, total, rnd, *rndptr;
653 
654 	total = 0;
655 	cpumask = cg->cg_mask;
656 	if (match & CPU_SEARCH_LOWEST) {
657 		lload = INT_MAX;
658 		lgroup = *low;
659 	}
660 	if (match & CPU_SEARCH_HIGHEST) {
661 		hload = INT_MIN;
662 		hgroup = *high;
663 	}
664 
665 	/* Iterate through the child CPU groups and then remaining CPUs. */
666 	for (i = cg->cg_children, cpu = mp_maxid; ; ) {
667 		if (i == 0) {
668 #ifdef HAVE_INLINE_FFSL
669 			cpu = CPU_FFS(&cpumask) - 1;
670 #else
671 			while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
672 				cpu--;
673 #endif
674 			if (cpu < 0)
675 				break;
676 			child = NULL;
677 		} else
678 			child = &cg->cg_child[i - 1];
679 
680 		if (match & CPU_SEARCH_LOWEST)
681 			lgroup.cs_cpu = -1;
682 		if (match & CPU_SEARCH_HIGHEST)
683 			hgroup.cs_cpu = -1;
684 		if (child) {			/* Handle child CPU group. */
685 			CPU_NAND(&cpumask, &child->cg_mask);
686 			switch (match) {
687 			case CPU_SEARCH_LOWEST:
688 				load = cpu_search_lowest(child, &lgroup);
689 				break;
690 			case CPU_SEARCH_HIGHEST:
691 				load = cpu_search_highest(child, &hgroup);
692 				break;
693 			case CPU_SEARCH_BOTH:
694 				load = cpu_search_both(child, &lgroup, &hgroup);
695 				break;
696 			}
697 		} else {			/* Handle child CPU. */
698 			CPU_CLR(cpu, &cpumask);
699 			tdq = TDQ_CPU(cpu);
700 			load = tdq->tdq_load * 256;
701 			rndptr = DPCPU_PTR(randomval);
702 			rnd = (*rndptr = *rndptr * 69069 + 5) >> 26;
703 			if (match & CPU_SEARCH_LOWEST) {
704 				if (cpu == low->cs_prefer)
705 					load -= 64;
706 				/* If that CPU is allowed and get data. */
707 				if (tdq->tdq_lowpri > lgroup.cs_pri &&
708 				    tdq->tdq_load <= lgroup.cs_limit &&
709 				    CPU_ISSET(cpu, &lgroup.cs_mask)) {
710 					lgroup.cs_cpu = cpu;
711 					lgroup.cs_load = load - rnd;
712 				}
713 			}
714 			if (match & CPU_SEARCH_HIGHEST)
715 				if (tdq->tdq_load >= hgroup.cs_limit &&
716 				    tdq->tdq_transferable &&
717 				    CPU_ISSET(cpu, &hgroup.cs_mask)) {
718 					hgroup.cs_cpu = cpu;
719 					hgroup.cs_load = load - rnd;
720 				}
721 		}
722 		total += load;
723 
724 		/* We have info about child item. Compare it. */
725 		if (match & CPU_SEARCH_LOWEST) {
726 			if (lgroup.cs_cpu >= 0 &&
727 			    (load < lload ||
728 			     (load == lload && lgroup.cs_load < low->cs_load))) {
729 				lload = load;
730 				low->cs_cpu = lgroup.cs_cpu;
731 				low->cs_load = lgroup.cs_load;
732 			}
733 		}
734 		if (match & CPU_SEARCH_HIGHEST)
735 			if (hgroup.cs_cpu >= 0 &&
736 			    (load > hload ||
737 			     (load == hload && hgroup.cs_load > high->cs_load))) {
738 				hload = load;
739 				high->cs_cpu = hgroup.cs_cpu;
740 				high->cs_load = hgroup.cs_load;
741 			}
742 		if (child) {
743 			i--;
744 			if (i == 0 && CPU_EMPTY(&cpumask))
745 				break;
746 		}
747 #ifndef HAVE_INLINE_FFSL
748 		else
749 			cpu--;
750 #endif
751 	}
752 	return (total);
753 }
754 
755 /*
756  * cpu_search instantiations must pass constants to maintain the inline
757  * optimization.
758  */
759 int
760 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
761 {
762 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
763 }
764 
765 int
766 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
767 {
768 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
769 }
770 
771 int
772 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
773     struct cpu_search *high)
774 {
775 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
776 }
777 
778 /*
779  * Find the cpu with the least load via the least loaded path that has a
780  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
781  * acceptable.
782  */
783 static inline int
784 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
785     int prefer)
786 {
787 	struct cpu_search low;
788 
789 	low.cs_cpu = -1;
790 	low.cs_prefer = prefer;
791 	low.cs_mask = mask;
792 	low.cs_pri = pri;
793 	low.cs_limit = maxload;
794 	cpu_search_lowest(cg, &low);
795 	return low.cs_cpu;
796 }
797 
798 /*
799  * Find the cpu with the highest load via the highest loaded path.
800  */
801 static inline int
802 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
803 {
804 	struct cpu_search high;
805 
806 	high.cs_cpu = -1;
807 	high.cs_mask = mask;
808 	high.cs_limit = minload;
809 	cpu_search_highest(cg, &high);
810 	return high.cs_cpu;
811 }
812 
813 static void
814 sched_balance_group(struct cpu_group *cg)
815 {
816 	cpuset_t hmask, lmask;
817 	int high, low, anylow;
818 
819 	CPU_FILL(&hmask);
820 	for (;;) {
821 		high = sched_highest(cg, hmask, 1);
822 		/* Stop if there is no more CPU with transferrable threads. */
823 		if (high == -1)
824 			break;
825 		CPU_CLR(high, &hmask);
826 		CPU_COPY(&hmask, &lmask);
827 		/* Stop if there is no more CPU left for low. */
828 		if (CPU_EMPTY(&lmask))
829 			break;
830 		anylow = 1;
831 nextlow:
832 		low = sched_lowest(cg, lmask, -1,
833 		    TDQ_CPU(high)->tdq_load - 1, high);
834 		/* Stop if we looked well and found no less loaded CPU. */
835 		if (anylow && low == -1)
836 			break;
837 		/* Go to next high if we found no less loaded CPU. */
838 		if (low == -1)
839 			continue;
840 		/* Transfer thread from high to low. */
841 		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
842 			/* CPU that got thread can no longer be a donor. */
843 			CPU_CLR(low, &hmask);
844 		} else {
845 			/*
846 			 * If failed, then there is no threads on high
847 			 * that can run on this low. Drop low from low
848 			 * mask and look for different one.
849 			 */
850 			CPU_CLR(low, &lmask);
851 			anylow = 0;
852 			goto nextlow;
853 		}
854 	}
855 }
856 
857 static void
858 sched_balance(void)
859 {
860 	struct tdq *tdq;
861 
862 	/*
863 	 * Select a random time between .5 * balance_interval and
864 	 * 1.5 * balance_interval.
865 	 */
866 	balance_ticks = max(balance_interval / 2, 1);
867 	balance_ticks += random() % balance_interval;
868 	if (smp_started == 0 || rebalance == 0)
869 		return;
870 	tdq = TDQ_SELF();
871 	TDQ_UNLOCK(tdq);
872 	sched_balance_group(cpu_top);
873 	TDQ_LOCK(tdq);
874 }
875 
876 /*
877  * Lock two thread queues using their address to maintain lock order.
878  */
879 static void
880 tdq_lock_pair(struct tdq *one, struct tdq *two)
881 {
882 	if (one < two) {
883 		TDQ_LOCK(one);
884 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
885 	} else {
886 		TDQ_LOCK(two);
887 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
888 	}
889 }
890 
891 /*
892  * Unlock two thread queues.  Order is not important here.
893  */
894 static void
895 tdq_unlock_pair(struct tdq *one, struct tdq *two)
896 {
897 	TDQ_UNLOCK(one);
898 	TDQ_UNLOCK(two);
899 }
900 
901 /*
902  * Transfer load between two imbalanced thread queues.
903  */
904 static int
905 sched_balance_pair(struct tdq *high, struct tdq *low)
906 {
907 	int moved;
908 	int cpu;
909 
910 	tdq_lock_pair(high, low);
911 	moved = 0;
912 	/*
913 	 * Determine what the imbalance is and then adjust that to how many
914 	 * threads we actually have to give up (transferable).
915 	 */
916 	if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
917 	    (moved = tdq_move(high, low)) > 0) {
918 		/*
919 		 * In case the target isn't the current cpu IPI it to force a
920 		 * reschedule with the new workload.
921 		 */
922 		cpu = TDQ_ID(low);
923 		if (cpu != PCPU_GET(cpuid))
924 			ipi_cpu(cpu, IPI_PREEMPT);
925 	}
926 	tdq_unlock_pair(high, low);
927 	return (moved);
928 }
929 
930 /*
931  * Move a thread from one thread queue to another.
932  */
933 static int
934 tdq_move(struct tdq *from, struct tdq *to)
935 {
936 	struct td_sched *ts;
937 	struct thread *td;
938 	struct tdq *tdq;
939 	int cpu;
940 
941 	TDQ_LOCK_ASSERT(from, MA_OWNED);
942 	TDQ_LOCK_ASSERT(to, MA_OWNED);
943 
944 	tdq = from;
945 	cpu = TDQ_ID(to);
946 	td = tdq_steal(tdq, cpu);
947 	if (td == NULL)
948 		return (0);
949 	ts = td->td_sched;
950 	/*
951 	 * Although the run queue is locked the thread may be blocked.  Lock
952 	 * it to clear this and acquire the run-queue lock.
953 	 */
954 	thread_lock(td);
955 	/* Drop recursive lock on from acquired via thread_lock(). */
956 	TDQ_UNLOCK(from);
957 	sched_rem(td);
958 	ts->ts_cpu = cpu;
959 	td->td_lock = TDQ_LOCKPTR(to);
960 	tdq_add(to, td, SRQ_YIELDING);
961 	return (1);
962 }
963 
964 /*
965  * This tdq has idled.  Try to steal a thread from another cpu and switch
966  * to it.
967  */
968 static int
969 tdq_idled(struct tdq *tdq)
970 {
971 	struct cpu_group *cg;
972 	struct tdq *steal;
973 	cpuset_t mask;
974 	int thresh;
975 	int cpu;
976 
977 	if (smp_started == 0 || steal_idle == 0)
978 		return (1);
979 	CPU_FILL(&mask);
980 	CPU_CLR(PCPU_GET(cpuid), &mask);
981 	/* We don't want to be preempted while we're iterating. */
982 	spinlock_enter();
983 	for (cg = tdq->tdq_cg; cg != NULL; ) {
984 		if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
985 			thresh = steal_thresh;
986 		else
987 			thresh = 1;
988 		cpu = sched_highest(cg, mask, thresh);
989 		if (cpu == -1) {
990 			cg = cg->cg_parent;
991 			continue;
992 		}
993 		steal = TDQ_CPU(cpu);
994 		CPU_CLR(cpu, &mask);
995 		tdq_lock_pair(tdq, steal);
996 		if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
997 			tdq_unlock_pair(tdq, steal);
998 			continue;
999 		}
1000 		/*
1001 		 * If a thread was added while interrupts were disabled don't
1002 		 * steal one here.  If we fail to acquire one due to affinity
1003 		 * restrictions loop again with this cpu removed from the
1004 		 * set.
1005 		 */
1006 		if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
1007 			tdq_unlock_pair(tdq, steal);
1008 			continue;
1009 		}
1010 		spinlock_exit();
1011 		TDQ_UNLOCK(steal);
1012 		mi_switch(SW_VOL | SWT_IDLE, NULL);
1013 		thread_unlock(curthread);
1014 
1015 		return (0);
1016 	}
1017 	spinlock_exit();
1018 	return (1);
1019 }
1020 
1021 /*
1022  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
1023  */
1024 static void
1025 tdq_notify(struct tdq *tdq, struct thread *td)
1026 {
1027 	struct thread *ctd;
1028 	int pri;
1029 	int cpu;
1030 
1031 	if (tdq->tdq_ipipending)
1032 		return;
1033 	cpu = td->td_sched->ts_cpu;
1034 	pri = td->td_priority;
1035 	ctd = pcpu_find(cpu)->pc_curthread;
1036 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1037 		return;
1038 	if (TD_IS_IDLETHREAD(ctd)) {
1039 		/*
1040 		 * If the MD code has an idle wakeup routine try that before
1041 		 * falling back to IPI.
1042 		 */
1043 		if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1044 			return;
1045 	}
1046 	tdq->tdq_ipipending = 1;
1047 	ipi_cpu(cpu, IPI_PREEMPT);
1048 }
1049 
1050 /*
1051  * Steals load from a timeshare queue.  Honors the rotating queue head
1052  * index.
1053  */
1054 static struct thread *
1055 runq_steal_from(struct runq *rq, int cpu, u_char start)
1056 {
1057 	struct rqbits *rqb;
1058 	struct rqhead *rqh;
1059 	struct thread *td, *first;
1060 	int bit;
1061 	int i;
1062 
1063 	rqb = &rq->rq_status;
1064 	bit = start & (RQB_BPW -1);
1065 	first = NULL;
1066 again:
1067 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1068 		if (rqb->rqb_bits[i] == 0)
1069 			continue;
1070 		if (bit == 0)
1071 			bit = RQB_FFS(rqb->rqb_bits[i]);
1072 		for (; bit < RQB_BPW; bit++) {
1073 			if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1074 				continue;
1075 			rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1076 			TAILQ_FOREACH(td, rqh, td_runq) {
1077 				if (first && THREAD_CAN_MIGRATE(td) &&
1078 				    THREAD_CAN_SCHED(td, cpu))
1079 					return (td);
1080 				first = td;
1081 			}
1082 		}
1083 	}
1084 	if (start != 0) {
1085 		start = 0;
1086 		goto again;
1087 	}
1088 
1089 	if (first && THREAD_CAN_MIGRATE(first) &&
1090 	    THREAD_CAN_SCHED(first, cpu))
1091 		return (first);
1092 	return (NULL);
1093 }
1094 
1095 /*
1096  * Steals load from a standard linear queue.
1097  */
1098 static struct thread *
1099 runq_steal(struct runq *rq, int cpu)
1100 {
1101 	struct rqhead *rqh;
1102 	struct rqbits *rqb;
1103 	struct thread *td;
1104 	int word;
1105 	int bit;
1106 
1107 	rqb = &rq->rq_status;
1108 	for (word = 0; word < RQB_LEN; word++) {
1109 		if (rqb->rqb_bits[word] == 0)
1110 			continue;
1111 		for (bit = 0; bit < RQB_BPW; bit++) {
1112 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1113 				continue;
1114 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1115 			TAILQ_FOREACH(td, rqh, td_runq)
1116 				if (THREAD_CAN_MIGRATE(td) &&
1117 				    THREAD_CAN_SCHED(td, cpu))
1118 					return (td);
1119 		}
1120 	}
1121 	return (NULL);
1122 }
1123 
1124 /*
1125  * Attempt to steal a thread in priority order from a thread queue.
1126  */
1127 static struct thread *
1128 tdq_steal(struct tdq *tdq, int cpu)
1129 {
1130 	struct thread *td;
1131 
1132 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1133 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1134 		return (td);
1135 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1136 	    cpu, tdq->tdq_ridx)) != NULL)
1137 		return (td);
1138 	return (runq_steal(&tdq->tdq_idle, cpu));
1139 }
1140 
1141 /*
1142  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1143  * current lock and returns with the assigned queue locked.
1144  */
1145 static inline struct tdq *
1146 sched_setcpu(struct thread *td, int cpu, int flags)
1147 {
1148 
1149 	struct tdq *tdq;
1150 
1151 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1152 	tdq = TDQ_CPU(cpu);
1153 	td->td_sched->ts_cpu = cpu;
1154 	/*
1155 	 * If the lock matches just return the queue.
1156 	 */
1157 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1158 		return (tdq);
1159 #ifdef notyet
1160 	/*
1161 	 * If the thread isn't running its lockptr is a
1162 	 * turnstile or a sleepqueue.  We can just lock_set without
1163 	 * blocking.
1164 	 */
1165 	if (TD_CAN_RUN(td)) {
1166 		TDQ_LOCK(tdq);
1167 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1168 		return (tdq);
1169 	}
1170 #endif
1171 	/*
1172 	 * The hard case, migration, we need to block the thread first to
1173 	 * prevent order reversals with other cpus locks.
1174 	 */
1175 	spinlock_enter();
1176 	thread_lock_block(td);
1177 	TDQ_LOCK(tdq);
1178 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1179 	spinlock_exit();
1180 	return (tdq);
1181 }
1182 
1183 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1184 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1185 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1186 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1187 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1188 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1189 
1190 static int
1191 sched_pickcpu(struct thread *td, int flags)
1192 {
1193 	struct cpu_group *cg, *ccg;
1194 	struct td_sched *ts;
1195 	struct tdq *tdq;
1196 	cpuset_t mask;
1197 	int cpu, pri, self;
1198 
1199 	self = PCPU_GET(cpuid);
1200 	ts = td->td_sched;
1201 	if (smp_started == 0)
1202 		return (self);
1203 	/*
1204 	 * Don't migrate a running thread from sched_switch().
1205 	 */
1206 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1207 		return (ts->ts_cpu);
1208 	/*
1209 	 * Prefer to run interrupt threads on the processors that generate
1210 	 * the interrupt.
1211 	 */
1212 	pri = td->td_priority;
1213 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1214 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1215 		SCHED_STAT_INC(pickcpu_intrbind);
1216 		ts->ts_cpu = self;
1217 		if (TDQ_CPU(self)->tdq_lowpri > pri) {
1218 			SCHED_STAT_INC(pickcpu_affinity);
1219 			return (ts->ts_cpu);
1220 		}
1221 	}
1222 	/*
1223 	 * If the thread can run on the last cpu and the affinity has not
1224 	 * expired or it is idle run it there.
1225 	 */
1226 	tdq = TDQ_CPU(ts->ts_cpu);
1227 	cg = tdq->tdq_cg;
1228 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1229 	    tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1230 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1231 		if (cg->cg_flags & CG_FLAG_THREAD) {
1232 			CPUSET_FOREACH(cpu, cg->cg_mask) {
1233 				if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1234 					break;
1235 			}
1236 		} else
1237 			cpu = INT_MAX;
1238 		if (cpu > mp_maxid) {
1239 			SCHED_STAT_INC(pickcpu_idle_affinity);
1240 			return (ts->ts_cpu);
1241 		}
1242 	}
1243 	/*
1244 	 * Search for the last level cache CPU group in the tree.
1245 	 * Skip caches with expired affinity time and SMT groups.
1246 	 * Affinity to higher level caches will be handled less aggressively.
1247 	 */
1248 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1249 		if (cg->cg_flags & CG_FLAG_THREAD)
1250 			continue;
1251 		if (!SCHED_AFFINITY(ts, cg->cg_level))
1252 			continue;
1253 		ccg = cg;
1254 	}
1255 	if (ccg != NULL)
1256 		cg = ccg;
1257 	cpu = -1;
1258 	/* Search the group for the less loaded idle CPU we can run now. */
1259 	mask = td->td_cpuset->cs_mask;
1260 	if (cg != NULL && cg != cpu_top &&
1261 	    CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1262 		cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1263 		    INT_MAX, ts->ts_cpu);
1264 	/* Search globally for the less loaded CPU we can run now. */
1265 	if (cpu == -1)
1266 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1267 	/* Search globally for the less loaded CPU. */
1268 	if (cpu == -1)
1269 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1270 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1271 	/*
1272 	 * Compare the lowest loaded cpu to current cpu.
1273 	 */
1274 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1275 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1276 	    TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1277 		SCHED_STAT_INC(pickcpu_local);
1278 		cpu = self;
1279 	} else
1280 		SCHED_STAT_INC(pickcpu_lowest);
1281 	if (cpu != ts->ts_cpu)
1282 		SCHED_STAT_INC(pickcpu_migration);
1283 	return (cpu);
1284 }
1285 #endif
1286 
1287 /*
1288  * Pick the highest priority task we have and return it.
1289  */
1290 static struct thread *
1291 tdq_choose(struct tdq *tdq)
1292 {
1293 	struct thread *td;
1294 
1295 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1296 	td = runq_choose(&tdq->tdq_realtime);
1297 	if (td != NULL)
1298 		return (td);
1299 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1300 	if (td != NULL) {
1301 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1302 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1303 		    td->td_priority));
1304 		return (td);
1305 	}
1306 	td = runq_choose(&tdq->tdq_idle);
1307 	if (td != NULL) {
1308 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1309 		    ("tdq_choose: Invalid priority on idle queue %d",
1310 		    td->td_priority));
1311 		return (td);
1312 	}
1313 
1314 	return (NULL);
1315 }
1316 
1317 /*
1318  * Initialize a thread queue.
1319  */
1320 static void
1321 tdq_setup(struct tdq *tdq)
1322 {
1323 
1324 	if (bootverbose)
1325 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1326 	runq_init(&tdq->tdq_realtime);
1327 	runq_init(&tdq->tdq_timeshare);
1328 	runq_init(&tdq->tdq_idle);
1329 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1330 	    "sched lock %d", (int)TDQ_ID(tdq));
1331 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1332 	    MTX_SPIN | MTX_RECURSE);
1333 #ifdef KTR
1334 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1335 	    "CPU %d load", (int)TDQ_ID(tdq));
1336 #endif
1337 }
1338 
1339 #ifdef SMP
1340 static void
1341 sched_setup_smp(void)
1342 {
1343 	struct tdq *tdq;
1344 	int i;
1345 
1346 	cpu_top = smp_topo();
1347 	CPU_FOREACH(i) {
1348 		tdq = TDQ_CPU(i);
1349 		tdq_setup(tdq);
1350 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1351 		if (tdq->tdq_cg == NULL)
1352 			panic("Can't find cpu group for %d\n", i);
1353 	}
1354 	balance_tdq = TDQ_SELF();
1355 	sched_balance();
1356 }
1357 #endif
1358 
1359 /*
1360  * Setup the thread queues and initialize the topology based on MD
1361  * information.
1362  */
1363 static void
1364 sched_setup(void *dummy)
1365 {
1366 	struct tdq *tdq;
1367 
1368 	tdq = TDQ_SELF();
1369 #ifdef SMP
1370 	sched_setup_smp();
1371 #else
1372 	tdq_setup(tdq);
1373 #endif
1374 
1375 	/* Add thread0's load since it's running. */
1376 	TDQ_LOCK(tdq);
1377 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1378 	tdq_load_add(tdq, &thread0);
1379 	tdq->tdq_lowpri = thread0.td_priority;
1380 	TDQ_UNLOCK(tdq);
1381 }
1382 
1383 /*
1384  * This routine determines time constants after stathz and hz are setup.
1385  */
1386 /* ARGSUSED */
1387 static void
1388 sched_initticks(void *dummy)
1389 {
1390 	int incr;
1391 
1392 	realstathz = stathz ? stathz : hz;
1393 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1394 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1395 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1396 	    realstathz);
1397 
1398 	/*
1399 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1400 	 * hz not being evenly divisible by stathz on all platforms.
1401 	 */
1402 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1403 	/*
1404 	 * This does not work for values of stathz that are more than
1405 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1406 	 */
1407 	if (incr == 0)
1408 		incr = 1;
1409 	tickincr = incr;
1410 #ifdef SMP
1411 	/*
1412 	 * Set the default balance interval now that we know
1413 	 * what realstathz is.
1414 	 */
1415 	balance_interval = realstathz;
1416 	affinity = SCHED_AFFINITY_DEFAULT;
1417 #endif
1418 	if (sched_idlespinthresh < 0)
1419 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1420 }
1421 
1422 
1423 /*
1424  * This is the core of the interactivity algorithm.  Determines a score based
1425  * on past behavior.  It is the ratio of sleep time to run time scaled to
1426  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1427  * differs from the cpu usage because it does not account for time spent
1428  * waiting on a run-queue.  Would be prettier if we had floating point.
1429  */
1430 static int
1431 sched_interact_score(struct thread *td)
1432 {
1433 	struct td_sched *ts;
1434 	int div;
1435 
1436 	ts = td->td_sched;
1437 	/*
1438 	 * The score is only needed if this is likely to be an interactive
1439 	 * task.  Don't go through the expense of computing it if there's
1440 	 * no chance.
1441 	 */
1442 	if (sched_interact <= SCHED_INTERACT_HALF &&
1443 		ts->ts_runtime >= ts->ts_slptime)
1444 			return (SCHED_INTERACT_HALF);
1445 
1446 	if (ts->ts_runtime > ts->ts_slptime) {
1447 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1448 		return (SCHED_INTERACT_HALF +
1449 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1450 	}
1451 	if (ts->ts_slptime > ts->ts_runtime) {
1452 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1453 		return (ts->ts_runtime / div);
1454 	}
1455 	/* runtime == slptime */
1456 	if (ts->ts_runtime)
1457 		return (SCHED_INTERACT_HALF);
1458 
1459 	/*
1460 	 * This can happen if slptime and runtime are 0.
1461 	 */
1462 	return (0);
1463 
1464 }
1465 
1466 /*
1467  * Scale the scheduling priority according to the "interactivity" of this
1468  * process.
1469  */
1470 static void
1471 sched_priority(struct thread *td)
1472 {
1473 	int score;
1474 	int pri;
1475 
1476 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1477 		return;
1478 	/*
1479 	 * If the score is interactive we place the thread in the realtime
1480 	 * queue with a priority that is less than kernel and interrupt
1481 	 * priorities.  These threads are not subject to nice restrictions.
1482 	 *
1483 	 * Scores greater than this are placed on the normal timeshare queue
1484 	 * where the priority is partially decided by the most recent cpu
1485 	 * utilization and the rest is decided by nice value.
1486 	 *
1487 	 * The nice value of the process has a linear effect on the calculated
1488 	 * score.  Negative nice values make it easier for a thread to be
1489 	 * considered interactive.
1490 	 */
1491 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1492 	if (score < sched_interact) {
1493 		pri = PRI_MIN_INTERACT;
1494 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1495 		    sched_interact) * score;
1496 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1497 		    ("sched_priority: invalid interactive priority %d score %d",
1498 		    pri, score));
1499 	} else {
1500 		pri = SCHED_PRI_MIN;
1501 		if (td->td_sched->ts_ticks)
1502 			pri += min(SCHED_PRI_TICKS(td->td_sched),
1503 			    SCHED_PRI_RANGE - 1);
1504 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1505 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1506 		    ("sched_priority: invalid priority %d: nice %d, "
1507 		    "ticks %d ftick %d ltick %d tick pri %d",
1508 		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1509 		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1510 		    SCHED_PRI_TICKS(td->td_sched)));
1511 	}
1512 	sched_user_prio(td, pri);
1513 
1514 	return;
1515 }
1516 
1517 /*
1518  * This routine enforces a maximum limit on the amount of scheduling history
1519  * kept.  It is called after either the slptime or runtime is adjusted.  This
1520  * function is ugly due to integer math.
1521  */
1522 static void
1523 sched_interact_update(struct thread *td)
1524 {
1525 	struct td_sched *ts;
1526 	u_int sum;
1527 
1528 	ts = td->td_sched;
1529 	sum = ts->ts_runtime + ts->ts_slptime;
1530 	if (sum < SCHED_SLP_RUN_MAX)
1531 		return;
1532 	/*
1533 	 * This only happens from two places:
1534 	 * 1) We have added an unusual amount of run time from fork_exit.
1535 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1536 	 */
1537 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1538 		if (ts->ts_runtime > ts->ts_slptime) {
1539 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1540 			ts->ts_slptime = 1;
1541 		} else {
1542 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1543 			ts->ts_runtime = 1;
1544 		}
1545 		return;
1546 	}
1547 	/*
1548 	 * If we have exceeded by more than 1/5th then the algorithm below
1549 	 * will not bring us back into range.  Dividing by two here forces
1550 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1551 	 */
1552 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1553 		ts->ts_runtime /= 2;
1554 		ts->ts_slptime /= 2;
1555 		return;
1556 	}
1557 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1558 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1559 }
1560 
1561 /*
1562  * Scale back the interactivity history when a child thread is created.  The
1563  * history is inherited from the parent but the thread may behave totally
1564  * differently.  For example, a shell spawning a compiler process.  We want
1565  * to learn that the compiler is behaving badly very quickly.
1566  */
1567 static void
1568 sched_interact_fork(struct thread *td)
1569 {
1570 	int ratio;
1571 	int sum;
1572 
1573 	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1574 	if (sum > SCHED_SLP_RUN_FORK) {
1575 		ratio = sum / SCHED_SLP_RUN_FORK;
1576 		td->td_sched->ts_runtime /= ratio;
1577 		td->td_sched->ts_slptime /= ratio;
1578 	}
1579 }
1580 
1581 /*
1582  * Called from proc0_init() to setup the scheduler fields.
1583  */
1584 void
1585 schedinit(void)
1586 {
1587 
1588 	/*
1589 	 * Set up the scheduler specific parts of proc0.
1590 	 */
1591 	proc0.p_sched = NULL; /* XXX */
1592 	thread0.td_sched = &td_sched0;
1593 	td_sched0.ts_ltick = ticks;
1594 	td_sched0.ts_ftick = ticks;
1595 	td_sched0.ts_slice = 0;
1596 }
1597 
1598 /*
1599  * This is only somewhat accurate since given many processes of the same
1600  * priority they will switch when their slices run out, which will be
1601  * at most sched_slice stathz ticks.
1602  */
1603 int
1604 sched_rr_interval(void)
1605 {
1606 
1607 	/* Convert sched_slice from stathz to hz. */
1608 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1609 }
1610 
1611 /*
1612  * Update the percent cpu tracking information when it is requested or
1613  * the total history exceeds the maximum.  We keep a sliding history of
1614  * tick counts that slowly decays.  This is less precise than the 4BSD
1615  * mechanism since it happens with less regular and frequent events.
1616  */
1617 static void
1618 sched_pctcpu_update(struct td_sched *ts, int run)
1619 {
1620 	int t = ticks;
1621 
1622 	if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1623 		ts->ts_ticks = 0;
1624 		ts->ts_ftick = t - SCHED_TICK_TARG;
1625 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1626 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1627 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1628 		ts->ts_ftick = t - SCHED_TICK_TARG;
1629 	}
1630 	if (run)
1631 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1632 	ts->ts_ltick = t;
1633 }
1634 
1635 /*
1636  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1637  * if necessary.  This is the back-end for several priority related
1638  * functions.
1639  */
1640 static void
1641 sched_thread_priority(struct thread *td, u_char prio)
1642 {
1643 	struct td_sched *ts;
1644 	struct tdq *tdq;
1645 	int oldpri;
1646 
1647 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1648 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1649 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1650 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1651 	if (td != curthread && prio < td->td_priority) {
1652 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1653 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1654 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1655 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1656 		    curthread);
1657 	}
1658 	ts = td->td_sched;
1659 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1660 	if (td->td_priority == prio)
1661 		return;
1662 	/*
1663 	 * If the priority has been elevated due to priority
1664 	 * propagation, we may have to move ourselves to a new
1665 	 * queue.  This could be optimized to not re-add in some
1666 	 * cases.
1667 	 */
1668 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1669 		sched_rem(td);
1670 		td->td_priority = prio;
1671 		sched_add(td, SRQ_BORROWING);
1672 		return;
1673 	}
1674 	/*
1675 	 * If the thread is currently running we may have to adjust the lowpri
1676 	 * information so other cpus are aware of our current priority.
1677 	 */
1678 	if (TD_IS_RUNNING(td)) {
1679 		tdq = TDQ_CPU(ts->ts_cpu);
1680 		oldpri = td->td_priority;
1681 		td->td_priority = prio;
1682 		if (prio < tdq->tdq_lowpri)
1683 			tdq->tdq_lowpri = prio;
1684 		else if (tdq->tdq_lowpri == oldpri)
1685 			tdq_setlowpri(tdq, td);
1686 		return;
1687 	}
1688 	td->td_priority = prio;
1689 }
1690 
1691 /*
1692  * Update a thread's priority when it is lent another thread's
1693  * priority.
1694  */
1695 void
1696 sched_lend_prio(struct thread *td, u_char prio)
1697 {
1698 
1699 	td->td_flags |= TDF_BORROWING;
1700 	sched_thread_priority(td, prio);
1701 }
1702 
1703 /*
1704  * Restore a thread's priority when priority propagation is
1705  * over.  The prio argument is the minimum priority the thread
1706  * needs to have to satisfy other possible priority lending
1707  * requests.  If the thread's regular priority is less
1708  * important than prio, the thread will keep a priority boost
1709  * of prio.
1710  */
1711 void
1712 sched_unlend_prio(struct thread *td, u_char prio)
1713 {
1714 	u_char base_pri;
1715 
1716 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1717 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1718 		base_pri = td->td_user_pri;
1719 	else
1720 		base_pri = td->td_base_pri;
1721 	if (prio >= base_pri) {
1722 		td->td_flags &= ~TDF_BORROWING;
1723 		sched_thread_priority(td, base_pri);
1724 	} else
1725 		sched_lend_prio(td, prio);
1726 }
1727 
1728 /*
1729  * Standard entry for setting the priority to an absolute value.
1730  */
1731 void
1732 sched_prio(struct thread *td, u_char prio)
1733 {
1734 	u_char oldprio;
1735 
1736 	/* First, update the base priority. */
1737 	td->td_base_pri = prio;
1738 
1739 	/*
1740 	 * If the thread is borrowing another thread's priority, don't
1741 	 * ever lower the priority.
1742 	 */
1743 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1744 		return;
1745 
1746 	/* Change the real priority. */
1747 	oldprio = td->td_priority;
1748 	sched_thread_priority(td, prio);
1749 
1750 	/*
1751 	 * If the thread is on a turnstile, then let the turnstile update
1752 	 * its state.
1753 	 */
1754 	if (TD_ON_LOCK(td) && oldprio != prio)
1755 		turnstile_adjust(td, oldprio);
1756 }
1757 
1758 /*
1759  * Set the base user priority, does not effect current running priority.
1760  */
1761 void
1762 sched_user_prio(struct thread *td, u_char prio)
1763 {
1764 
1765 	td->td_base_user_pri = prio;
1766 	if (td->td_lend_user_pri <= prio)
1767 		return;
1768 	td->td_user_pri = prio;
1769 }
1770 
1771 void
1772 sched_lend_user_prio(struct thread *td, u_char prio)
1773 {
1774 
1775 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1776 	td->td_lend_user_pri = prio;
1777 	td->td_user_pri = min(prio, td->td_base_user_pri);
1778 	if (td->td_priority > td->td_user_pri)
1779 		sched_prio(td, td->td_user_pri);
1780 	else if (td->td_priority != td->td_user_pri)
1781 		td->td_flags |= TDF_NEEDRESCHED;
1782 }
1783 
1784 /*
1785  * Handle migration from sched_switch().  This happens only for
1786  * cpu binding.
1787  */
1788 static struct mtx *
1789 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1790 {
1791 	struct tdq *tdn;
1792 
1793 	tdn = TDQ_CPU(td->td_sched->ts_cpu);
1794 #ifdef SMP
1795 	tdq_load_rem(tdq, td);
1796 	/*
1797 	 * Do the lock dance required to avoid LOR.  We grab an extra
1798 	 * spinlock nesting to prevent preemption while we're
1799 	 * not holding either run-queue lock.
1800 	 */
1801 	spinlock_enter();
1802 	thread_lock_block(td);	/* This releases the lock on tdq. */
1803 
1804 	/*
1805 	 * Acquire both run-queue locks before placing the thread on the new
1806 	 * run-queue to avoid deadlocks created by placing a thread with a
1807 	 * blocked lock on the run-queue of a remote processor.  The deadlock
1808 	 * occurs when a third processor attempts to lock the two queues in
1809 	 * question while the target processor is spinning with its own
1810 	 * run-queue lock held while waiting for the blocked lock to clear.
1811 	 */
1812 	tdq_lock_pair(tdn, tdq);
1813 	tdq_add(tdn, td, flags);
1814 	tdq_notify(tdn, td);
1815 	TDQ_UNLOCK(tdn);
1816 	spinlock_exit();
1817 #endif
1818 	return (TDQ_LOCKPTR(tdn));
1819 }
1820 
1821 /*
1822  * Variadic version of thread_lock_unblock() that does not assume td_lock
1823  * is blocked.
1824  */
1825 static inline void
1826 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1827 {
1828 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1829 	    (uintptr_t)mtx);
1830 }
1831 
1832 /*
1833  * Switch threads.  This function has to handle threads coming in while
1834  * blocked for some reason, running, or idle.  It also must deal with
1835  * migrating a thread from one queue to another as running threads may
1836  * be assigned elsewhere via binding.
1837  */
1838 void
1839 sched_switch(struct thread *td, struct thread *newtd, int flags)
1840 {
1841 	struct tdq *tdq;
1842 	struct td_sched *ts;
1843 	struct mtx *mtx;
1844 	int srqflag;
1845 	int cpuid, preempted;
1846 
1847 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1848 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1849 
1850 	cpuid = PCPU_GET(cpuid);
1851 	tdq = TDQ_CPU(cpuid);
1852 	ts = td->td_sched;
1853 	mtx = td->td_lock;
1854 	sched_pctcpu_update(ts, 1);
1855 	ts->ts_rltick = ticks;
1856 	td->td_lastcpu = td->td_oncpu;
1857 	td->td_oncpu = NOCPU;
1858 	preempted = !(td->td_flags & TDF_SLICEEND);
1859 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1860 	td->td_owepreempt = 0;
1861 	if (!TD_IS_IDLETHREAD(td))
1862 		tdq->tdq_switchcnt++;
1863 	/*
1864 	 * The lock pointer in an idle thread should never change.  Reset it
1865 	 * to CAN_RUN as well.
1866 	 */
1867 	if (TD_IS_IDLETHREAD(td)) {
1868 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1869 		TD_SET_CAN_RUN(td);
1870 	} else if (TD_IS_RUNNING(td)) {
1871 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1872 		srqflag = preempted ?
1873 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1874 		    SRQ_OURSELF|SRQ_YIELDING;
1875 #ifdef SMP
1876 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1877 			ts->ts_cpu = sched_pickcpu(td, 0);
1878 #endif
1879 		if (ts->ts_cpu == cpuid)
1880 			tdq_runq_add(tdq, td, srqflag);
1881 		else {
1882 			KASSERT(THREAD_CAN_MIGRATE(td) ||
1883 			    (ts->ts_flags & TSF_BOUND) != 0,
1884 			    ("Thread %p shouldn't migrate", td));
1885 			mtx = sched_switch_migrate(tdq, td, srqflag);
1886 		}
1887 	} else {
1888 		/* This thread must be going to sleep. */
1889 		TDQ_LOCK(tdq);
1890 		mtx = thread_lock_block(td);
1891 		tdq_load_rem(tdq, td);
1892 	}
1893 	/*
1894 	 * We enter here with the thread blocked and assigned to the
1895 	 * appropriate cpu run-queue or sleep-queue and with the current
1896 	 * thread-queue locked.
1897 	 */
1898 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1899 	newtd = choosethread();
1900 	/*
1901 	 * Call the MD code to switch contexts if necessary.
1902 	 */
1903 	if (td != newtd) {
1904 #ifdef	HWPMC_HOOKS
1905 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1906 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1907 #endif
1908 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1909 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1910 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1911 		sched_pctcpu_update(newtd->td_sched, 0);
1912 
1913 #ifdef KDTRACE_HOOKS
1914 		/*
1915 		 * If DTrace has set the active vtime enum to anything
1916 		 * other than INACTIVE (0), then it should have set the
1917 		 * function to call.
1918 		 */
1919 		if (dtrace_vtime_active)
1920 			(*dtrace_vtime_switch_func)(newtd);
1921 #endif
1922 
1923 		cpu_switch(td, newtd, mtx);
1924 		/*
1925 		 * We may return from cpu_switch on a different cpu.  However,
1926 		 * we always return with td_lock pointing to the current cpu's
1927 		 * run queue lock.
1928 		 */
1929 		cpuid = PCPU_GET(cpuid);
1930 		tdq = TDQ_CPU(cpuid);
1931 		lock_profile_obtain_lock_success(
1932 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1933 
1934 		SDT_PROBE0(sched, , , on__cpu);
1935 #ifdef	HWPMC_HOOKS
1936 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1937 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1938 #endif
1939 	} else {
1940 		thread_unblock_switch(td, mtx);
1941 		SDT_PROBE0(sched, , , remain__cpu);
1942 	}
1943 	/*
1944 	 * Assert that all went well and return.
1945 	 */
1946 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1947 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1948 	td->td_oncpu = cpuid;
1949 }
1950 
1951 /*
1952  * Adjust thread priorities as a result of a nice request.
1953  */
1954 void
1955 sched_nice(struct proc *p, int nice)
1956 {
1957 	struct thread *td;
1958 
1959 	PROC_LOCK_ASSERT(p, MA_OWNED);
1960 
1961 	p->p_nice = nice;
1962 	FOREACH_THREAD_IN_PROC(p, td) {
1963 		thread_lock(td);
1964 		sched_priority(td);
1965 		sched_prio(td, td->td_base_user_pri);
1966 		thread_unlock(td);
1967 	}
1968 }
1969 
1970 /*
1971  * Record the sleep time for the interactivity scorer.
1972  */
1973 void
1974 sched_sleep(struct thread *td, int prio)
1975 {
1976 
1977 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1978 
1979 	td->td_slptick = ticks;
1980 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1981 		td->td_flags |= TDF_CANSWAP;
1982 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1983 		return;
1984 	if (static_boost == 1 && prio)
1985 		sched_prio(td, prio);
1986 	else if (static_boost && td->td_priority > static_boost)
1987 		sched_prio(td, static_boost);
1988 }
1989 
1990 /*
1991  * Schedule a thread to resume execution and record how long it voluntarily
1992  * slept.  We also update the pctcpu, interactivity, and priority.
1993  */
1994 void
1995 sched_wakeup(struct thread *td)
1996 {
1997 	struct td_sched *ts;
1998 	int slptick;
1999 
2000 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2001 	ts = td->td_sched;
2002 	td->td_flags &= ~TDF_CANSWAP;
2003 	/*
2004 	 * If we slept for more than a tick update our interactivity and
2005 	 * priority.
2006 	 */
2007 	slptick = td->td_slptick;
2008 	td->td_slptick = 0;
2009 	if (slptick && slptick != ticks) {
2010 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2011 		sched_interact_update(td);
2012 		sched_pctcpu_update(ts, 0);
2013 	}
2014 	/*
2015 	 * Reset the slice value since we slept and advanced the round-robin.
2016 	 */
2017 	ts->ts_slice = 0;
2018 	sched_add(td, SRQ_BORING);
2019 }
2020 
2021 /*
2022  * Penalize the parent for creating a new child and initialize the child's
2023  * priority.
2024  */
2025 void
2026 sched_fork(struct thread *td, struct thread *child)
2027 {
2028 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2029 	sched_pctcpu_update(td->td_sched, 1);
2030 	sched_fork_thread(td, child);
2031 	/*
2032 	 * Penalize the parent and child for forking.
2033 	 */
2034 	sched_interact_fork(child);
2035 	sched_priority(child);
2036 	td->td_sched->ts_runtime += tickincr;
2037 	sched_interact_update(td);
2038 	sched_priority(td);
2039 }
2040 
2041 /*
2042  * Fork a new thread, may be within the same process.
2043  */
2044 void
2045 sched_fork_thread(struct thread *td, struct thread *child)
2046 {
2047 	struct td_sched *ts;
2048 	struct td_sched *ts2;
2049 	struct tdq *tdq;
2050 
2051 	tdq = TDQ_SELF();
2052 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2053 	/*
2054 	 * Initialize child.
2055 	 */
2056 	ts = td->td_sched;
2057 	ts2 = child->td_sched;
2058 	child->td_lock = TDQ_LOCKPTR(tdq);
2059 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2060 	ts2->ts_cpu = ts->ts_cpu;
2061 	ts2->ts_flags = 0;
2062 	/*
2063 	 * Grab our parents cpu estimation information.
2064 	 */
2065 	ts2->ts_ticks = ts->ts_ticks;
2066 	ts2->ts_ltick = ts->ts_ltick;
2067 	ts2->ts_ftick = ts->ts_ftick;
2068 	/*
2069 	 * Do not inherit any borrowed priority from the parent.
2070 	 */
2071 	child->td_priority = child->td_base_pri;
2072 	/*
2073 	 * And update interactivity score.
2074 	 */
2075 	ts2->ts_slptime = ts->ts_slptime;
2076 	ts2->ts_runtime = ts->ts_runtime;
2077 	/* Attempt to quickly learn interactivity. */
2078 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2079 #ifdef KTR
2080 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2081 #endif
2082 }
2083 
2084 /*
2085  * Adjust the priority class of a thread.
2086  */
2087 void
2088 sched_class(struct thread *td, int class)
2089 {
2090 
2091 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2092 	if (td->td_pri_class == class)
2093 		return;
2094 	td->td_pri_class = class;
2095 }
2096 
2097 /*
2098  * Return some of the child's priority and interactivity to the parent.
2099  */
2100 void
2101 sched_exit(struct proc *p, struct thread *child)
2102 {
2103 	struct thread *td;
2104 
2105 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2106 	    "prio:%d", child->td_priority);
2107 	PROC_LOCK_ASSERT(p, MA_OWNED);
2108 	td = FIRST_THREAD_IN_PROC(p);
2109 	sched_exit_thread(td, child);
2110 }
2111 
2112 /*
2113  * Penalize another thread for the time spent on this one.  This helps to
2114  * worsen the priority and interactivity of processes which schedule batch
2115  * jobs such as make.  This has little effect on the make process itself but
2116  * causes new processes spawned by it to receive worse scores immediately.
2117  */
2118 void
2119 sched_exit_thread(struct thread *td, struct thread *child)
2120 {
2121 
2122 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2123 	    "prio:%d", child->td_priority);
2124 	/*
2125 	 * Give the child's runtime to the parent without returning the
2126 	 * sleep time as a penalty to the parent.  This causes shells that
2127 	 * launch expensive things to mark their children as expensive.
2128 	 */
2129 	thread_lock(td);
2130 	td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2131 	sched_interact_update(td);
2132 	sched_priority(td);
2133 	thread_unlock(td);
2134 }
2135 
2136 void
2137 sched_preempt(struct thread *td)
2138 {
2139 	struct tdq *tdq;
2140 
2141 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2142 
2143 	thread_lock(td);
2144 	tdq = TDQ_SELF();
2145 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2146 	tdq->tdq_ipipending = 0;
2147 	if (td->td_priority > tdq->tdq_lowpri) {
2148 		int flags;
2149 
2150 		flags = SW_INVOL | SW_PREEMPT;
2151 		if (td->td_critnest > 1)
2152 			td->td_owepreempt = 1;
2153 		else if (TD_IS_IDLETHREAD(td))
2154 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2155 		else
2156 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2157 	}
2158 	thread_unlock(td);
2159 }
2160 
2161 /*
2162  * Fix priorities on return to user-space.  Priorities may be elevated due
2163  * to static priorities in msleep() or similar.
2164  */
2165 void
2166 sched_userret(struct thread *td)
2167 {
2168 	/*
2169 	 * XXX we cheat slightly on the locking here to avoid locking in
2170 	 * the usual case.  Setting td_priority here is essentially an
2171 	 * incomplete workaround for not setting it properly elsewhere.
2172 	 * Now that some interrupt handlers are threads, not setting it
2173 	 * properly elsewhere can clobber it in the window between setting
2174 	 * it here and returning to user mode, so don't waste time setting
2175 	 * it perfectly here.
2176 	 */
2177 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2178 	    ("thread with borrowed priority returning to userland"));
2179 	if (td->td_priority != td->td_user_pri) {
2180 		thread_lock(td);
2181 		td->td_priority = td->td_user_pri;
2182 		td->td_base_pri = td->td_user_pri;
2183 		tdq_setlowpri(TDQ_SELF(), td);
2184 		thread_unlock(td);
2185         }
2186 }
2187 
2188 /*
2189  * Handle a stathz tick.  This is really only relevant for timeshare
2190  * threads.
2191  */
2192 void
2193 sched_clock(struct thread *td)
2194 {
2195 	struct tdq *tdq;
2196 	struct td_sched *ts;
2197 
2198 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2199 	tdq = TDQ_SELF();
2200 #ifdef SMP
2201 	/*
2202 	 * We run the long term load balancer infrequently on the first cpu.
2203 	 */
2204 	if (balance_tdq == tdq) {
2205 		if (balance_ticks && --balance_ticks == 0)
2206 			sched_balance();
2207 	}
2208 #endif
2209 	/*
2210 	 * Save the old switch count so we have a record of the last ticks
2211 	 * activity.   Initialize the new switch count based on our load.
2212 	 * If there is some activity seed it to reflect that.
2213 	 */
2214 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2215 	tdq->tdq_switchcnt = tdq->tdq_load;
2216 	/*
2217 	 * Advance the insert index once for each tick to ensure that all
2218 	 * threads get a chance to run.
2219 	 */
2220 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2221 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2222 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2223 			tdq->tdq_ridx = tdq->tdq_idx;
2224 	}
2225 	ts = td->td_sched;
2226 	sched_pctcpu_update(ts, 1);
2227 	if (td->td_pri_class & PRI_FIFO_BIT)
2228 		return;
2229 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2230 		/*
2231 		 * We used a tick; charge it to the thread so
2232 		 * that we can compute our interactivity.
2233 		 */
2234 		td->td_sched->ts_runtime += tickincr;
2235 		sched_interact_update(td);
2236 		sched_priority(td);
2237 	}
2238 
2239 	/*
2240 	 * Force a context switch if the current thread has used up a full
2241 	 * time slice (default is 100ms).
2242 	 */
2243 	if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2244 		ts->ts_slice = 0;
2245 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2246 	}
2247 }
2248 
2249 /*
2250  * Called once per hz tick.
2251  */
2252 void
2253 sched_tick(int cnt)
2254 {
2255 
2256 }
2257 
2258 /*
2259  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2260  * cooperative idle threads.
2261  */
2262 int
2263 sched_runnable(void)
2264 {
2265 	struct tdq *tdq;
2266 	int load;
2267 
2268 	load = 1;
2269 
2270 	tdq = TDQ_SELF();
2271 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2272 		if (tdq->tdq_load > 0)
2273 			goto out;
2274 	} else
2275 		if (tdq->tdq_load - 1 > 0)
2276 			goto out;
2277 	load = 0;
2278 out:
2279 	return (load);
2280 }
2281 
2282 /*
2283  * Choose the highest priority thread to run.  The thread is removed from
2284  * the run-queue while running however the load remains.  For SMP we set
2285  * the tdq in the global idle bitmask if it idles here.
2286  */
2287 struct thread *
2288 sched_choose(void)
2289 {
2290 	struct thread *td;
2291 	struct tdq *tdq;
2292 
2293 	tdq = TDQ_SELF();
2294 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2295 	td = tdq_choose(tdq);
2296 	if (td) {
2297 		tdq_runq_rem(tdq, td);
2298 		tdq->tdq_lowpri = td->td_priority;
2299 		return (td);
2300 	}
2301 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2302 	return (PCPU_GET(idlethread));
2303 }
2304 
2305 /*
2306  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2307  * we always request it once we exit a critical section.
2308  */
2309 static inline void
2310 sched_setpreempt(struct thread *td)
2311 {
2312 	struct thread *ctd;
2313 	int cpri;
2314 	int pri;
2315 
2316 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2317 
2318 	ctd = curthread;
2319 	pri = td->td_priority;
2320 	cpri = ctd->td_priority;
2321 	if (pri < cpri)
2322 		ctd->td_flags |= TDF_NEEDRESCHED;
2323 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2324 		return;
2325 	if (!sched_shouldpreempt(pri, cpri, 0))
2326 		return;
2327 	ctd->td_owepreempt = 1;
2328 }
2329 
2330 /*
2331  * Add a thread to a thread queue.  Select the appropriate runq and add the
2332  * thread to it.  This is the internal function called when the tdq is
2333  * predetermined.
2334  */
2335 void
2336 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2337 {
2338 
2339 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2340 	KASSERT((td->td_inhibitors == 0),
2341 	    ("sched_add: trying to run inhibited thread"));
2342 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2343 	    ("sched_add: bad thread state"));
2344 	KASSERT(td->td_flags & TDF_INMEM,
2345 	    ("sched_add: thread swapped out"));
2346 
2347 	if (td->td_priority < tdq->tdq_lowpri)
2348 		tdq->tdq_lowpri = td->td_priority;
2349 	tdq_runq_add(tdq, td, flags);
2350 	tdq_load_add(tdq, td);
2351 }
2352 
2353 /*
2354  * Select the target thread queue and add a thread to it.  Request
2355  * preemption or IPI a remote processor if required.
2356  */
2357 void
2358 sched_add(struct thread *td, int flags)
2359 {
2360 	struct tdq *tdq;
2361 #ifdef SMP
2362 	int cpu;
2363 #endif
2364 
2365 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2366 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2367 	    sched_tdname(curthread));
2368 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2369 	    KTR_ATTR_LINKED, sched_tdname(td));
2370 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2371 	    flags & SRQ_PREEMPTED);
2372 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2373 	/*
2374 	 * Recalculate the priority before we select the target cpu or
2375 	 * run-queue.
2376 	 */
2377 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2378 		sched_priority(td);
2379 #ifdef SMP
2380 	/*
2381 	 * Pick the destination cpu and if it isn't ours transfer to the
2382 	 * target cpu.
2383 	 */
2384 	cpu = sched_pickcpu(td, flags);
2385 	tdq = sched_setcpu(td, cpu, flags);
2386 	tdq_add(tdq, td, flags);
2387 	if (cpu != PCPU_GET(cpuid)) {
2388 		tdq_notify(tdq, td);
2389 		return;
2390 	}
2391 #else
2392 	tdq = TDQ_SELF();
2393 	TDQ_LOCK(tdq);
2394 	/*
2395 	 * Now that the thread is moving to the run-queue, set the lock
2396 	 * to the scheduler's lock.
2397 	 */
2398 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2399 	tdq_add(tdq, td, flags);
2400 #endif
2401 	if (!(flags & SRQ_YIELDING))
2402 		sched_setpreempt(td);
2403 }
2404 
2405 /*
2406  * Remove a thread from a run-queue without running it.  This is used
2407  * when we're stealing a thread from a remote queue.  Otherwise all threads
2408  * exit by calling sched_exit_thread() and sched_throw() themselves.
2409  */
2410 void
2411 sched_rem(struct thread *td)
2412 {
2413 	struct tdq *tdq;
2414 
2415 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2416 	    "prio:%d", td->td_priority);
2417 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2418 	tdq = TDQ_CPU(td->td_sched->ts_cpu);
2419 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2420 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2421 	KASSERT(TD_ON_RUNQ(td),
2422 	    ("sched_rem: thread not on run queue"));
2423 	tdq_runq_rem(tdq, td);
2424 	tdq_load_rem(tdq, td);
2425 	TD_SET_CAN_RUN(td);
2426 	if (td->td_priority == tdq->tdq_lowpri)
2427 		tdq_setlowpri(tdq, NULL);
2428 }
2429 
2430 /*
2431  * Fetch cpu utilization information.  Updates on demand.
2432  */
2433 fixpt_t
2434 sched_pctcpu(struct thread *td)
2435 {
2436 	fixpt_t pctcpu;
2437 	struct td_sched *ts;
2438 
2439 	pctcpu = 0;
2440 	ts = td->td_sched;
2441 	if (ts == NULL)
2442 		return (0);
2443 
2444 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2445 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2446 	if (ts->ts_ticks) {
2447 		int rtick;
2448 
2449 		/* How many rtick per second ? */
2450 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2451 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2452 	}
2453 
2454 	return (pctcpu);
2455 }
2456 
2457 /*
2458  * Enforce affinity settings for a thread.  Called after adjustments to
2459  * cpumask.
2460  */
2461 void
2462 sched_affinity(struct thread *td)
2463 {
2464 #ifdef SMP
2465 	struct td_sched *ts;
2466 
2467 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2468 	ts = td->td_sched;
2469 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2470 		return;
2471 	if (TD_ON_RUNQ(td)) {
2472 		sched_rem(td);
2473 		sched_add(td, SRQ_BORING);
2474 		return;
2475 	}
2476 	if (!TD_IS_RUNNING(td))
2477 		return;
2478 	/*
2479 	 * Force a switch before returning to userspace.  If the
2480 	 * target thread is not running locally send an ipi to force
2481 	 * the issue.
2482 	 */
2483 	td->td_flags |= TDF_NEEDRESCHED;
2484 	if (td != curthread)
2485 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2486 #endif
2487 }
2488 
2489 /*
2490  * Bind a thread to a target cpu.
2491  */
2492 void
2493 sched_bind(struct thread *td, int cpu)
2494 {
2495 	struct td_sched *ts;
2496 
2497 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2498 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2499 	ts = td->td_sched;
2500 	if (ts->ts_flags & TSF_BOUND)
2501 		sched_unbind(td);
2502 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2503 	ts->ts_flags |= TSF_BOUND;
2504 	sched_pin();
2505 	if (PCPU_GET(cpuid) == cpu)
2506 		return;
2507 	ts->ts_cpu = cpu;
2508 	/* When we return from mi_switch we'll be on the correct cpu. */
2509 	mi_switch(SW_VOL, NULL);
2510 }
2511 
2512 /*
2513  * Release a bound thread.
2514  */
2515 void
2516 sched_unbind(struct thread *td)
2517 {
2518 	struct td_sched *ts;
2519 
2520 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2521 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2522 	ts = td->td_sched;
2523 	if ((ts->ts_flags & TSF_BOUND) == 0)
2524 		return;
2525 	ts->ts_flags &= ~TSF_BOUND;
2526 	sched_unpin();
2527 }
2528 
2529 int
2530 sched_is_bound(struct thread *td)
2531 {
2532 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2533 	return (td->td_sched->ts_flags & TSF_BOUND);
2534 }
2535 
2536 /*
2537  * Basic yield call.
2538  */
2539 void
2540 sched_relinquish(struct thread *td)
2541 {
2542 	thread_lock(td);
2543 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2544 	thread_unlock(td);
2545 }
2546 
2547 /*
2548  * Return the total system load.
2549  */
2550 int
2551 sched_load(void)
2552 {
2553 #ifdef SMP
2554 	int total;
2555 	int i;
2556 
2557 	total = 0;
2558 	CPU_FOREACH(i)
2559 		total += TDQ_CPU(i)->tdq_sysload;
2560 	return (total);
2561 #else
2562 	return (TDQ_SELF()->tdq_sysload);
2563 #endif
2564 }
2565 
2566 int
2567 sched_sizeof_proc(void)
2568 {
2569 	return (sizeof(struct proc));
2570 }
2571 
2572 int
2573 sched_sizeof_thread(void)
2574 {
2575 	return (sizeof(struct thread) + sizeof(struct td_sched));
2576 }
2577 
2578 #ifdef SMP
2579 #define	TDQ_IDLESPIN(tdq)						\
2580     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2581 #else
2582 #define	TDQ_IDLESPIN(tdq)	1
2583 #endif
2584 
2585 /*
2586  * The actual idle process.
2587  */
2588 void
2589 sched_idletd(void *dummy)
2590 {
2591 	struct thread *td;
2592 	struct tdq *tdq;
2593 	int oldswitchcnt, switchcnt;
2594 	int i;
2595 
2596 	mtx_assert(&Giant, MA_NOTOWNED);
2597 	td = curthread;
2598 	tdq = TDQ_SELF();
2599 	THREAD_NO_SLEEPING();
2600 	oldswitchcnt = -1;
2601 	for (;;) {
2602 		if (tdq->tdq_load) {
2603 			thread_lock(td);
2604 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2605 			thread_unlock(td);
2606 		}
2607 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2608 #ifdef SMP
2609 		if (switchcnt != oldswitchcnt) {
2610 			oldswitchcnt = switchcnt;
2611 			if (tdq_idled(tdq) == 0)
2612 				continue;
2613 		}
2614 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2615 #else
2616 		oldswitchcnt = switchcnt;
2617 #endif
2618 		/*
2619 		 * If we're switching very frequently, spin while checking
2620 		 * for load rather than entering a low power state that
2621 		 * may require an IPI.  However, don't do any busy
2622 		 * loops while on SMT machines as this simply steals
2623 		 * cycles from cores doing useful work.
2624 		 */
2625 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2626 			for (i = 0; i < sched_idlespins; i++) {
2627 				if (tdq->tdq_load)
2628 					break;
2629 				cpu_spinwait();
2630 			}
2631 		}
2632 
2633 		/* If there was context switch during spin, restart it. */
2634 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2635 		if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2636 			continue;
2637 
2638 		/* Run main MD idle handler. */
2639 		tdq->tdq_cpu_idle = 1;
2640 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2641 		tdq->tdq_cpu_idle = 0;
2642 
2643 		/*
2644 		 * Account thread-less hardware interrupts and
2645 		 * other wakeup reasons equal to context switches.
2646 		 */
2647 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2648 		if (switchcnt != oldswitchcnt)
2649 			continue;
2650 		tdq->tdq_switchcnt++;
2651 		oldswitchcnt++;
2652 	}
2653 }
2654 
2655 /*
2656  * A CPU is entering for the first time or a thread is exiting.
2657  */
2658 void
2659 sched_throw(struct thread *td)
2660 {
2661 	struct thread *newtd;
2662 	struct tdq *tdq;
2663 
2664 	tdq = TDQ_SELF();
2665 	if (td == NULL) {
2666 		/* Correct spinlock nesting and acquire the correct lock. */
2667 		TDQ_LOCK(tdq);
2668 		spinlock_exit();
2669 		PCPU_SET(switchtime, cpu_ticks());
2670 		PCPU_SET(switchticks, ticks);
2671 	} else {
2672 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2673 		tdq_load_rem(tdq, td);
2674 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2675 	}
2676 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2677 	newtd = choosethread();
2678 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2679 	cpu_throw(td, newtd);		/* doesn't return */
2680 }
2681 
2682 /*
2683  * This is called from fork_exit().  Just acquire the correct locks and
2684  * let fork do the rest of the work.
2685  */
2686 void
2687 sched_fork_exit(struct thread *td)
2688 {
2689 	struct tdq *tdq;
2690 	int cpuid;
2691 
2692 	/*
2693 	 * Finish setting up thread glue so that it begins execution in a
2694 	 * non-nested critical section with the scheduler lock held.
2695 	 */
2696 	cpuid = PCPU_GET(cpuid);
2697 	tdq = TDQ_CPU(cpuid);
2698 	if (TD_IS_IDLETHREAD(td))
2699 		td->td_lock = TDQ_LOCKPTR(tdq);
2700 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2701 	td->td_oncpu = cpuid;
2702 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2703 	lock_profile_obtain_lock_success(
2704 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2705 }
2706 
2707 /*
2708  * Create on first use to catch odd startup conditons.
2709  */
2710 char *
2711 sched_tdname(struct thread *td)
2712 {
2713 #ifdef KTR
2714 	struct td_sched *ts;
2715 
2716 	ts = td->td_sched;
2717 	if (ts->ts_name[0] == '\0')
2718 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2719 		    "%s tid %d", td->td_name, td->td_tid);
2720 	return (ts->ts_name);
2721 #else
2722 	return (td->td_name);
2723 #endif
2724 }
2725 
2726 #ifdef KTR
2727 void
2728 sched_clear_tdname(struct thread *td)
2729 {
2730 	struct td_sched *ts;
2731 
2732 	ts = td->td_sched;
2733 	ts->ts_name[0] = '\0';
2734 }
2735 #endif
2736 
2737 #ifdef SMP
2738 
2739 /*
2740  * Build the CPU topology dump string. Is recursively called to collect
2741  * the topology tree.
2742  */
2743 static int
2744 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2745     int indent)
2746 {
2747 	char cpusetbuf[CPUSETBUFSIZ];
2748 	int i, first;
2749 
2750 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2751 	    "", 1 + indent / 2, cg->cg_level);
2752 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2753 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2754 	first = TRUE;
2755 	for (i = 0; i < MAXCPU; i++) {
2756 		if (CPU_ISSET(i, &cg->cg_mask)) {
2757 			if (!first)
2758 				sbuf_printf(sb, ", ");
2759 			else
2760 				first = FALSE;
2761 			sbuf_printf(sb, "%d", i);
2762 		}
2763 	}
2764 	sbuf_printf(sb, "</cpu>\n");
2765 
2766 	if (cg->cg_flags != 0) {
2767 		sbuf_printf(sb, "%*s <flags>", indent, "");
2768 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2769 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2770 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2771 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2772 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2773 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2774 		sbuf_printf(sb, "</flags>\n");
2775 	}
2776 
2777 	if (cg->cg_children > 0) {
2778 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2779 		for (i = 0; i < cg->cg_children; i++)
2780 			sysctl_kern_sched_topology_spec_internal(sb,
2781 			    &cg->cg_child[i], indent+2);
2782 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2783 	}
2784 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2785 	return (0);
2786 }
2787 
2788 /*
2789  * Sysctl handler for retrieving topology dump. It's a wrapper for
2790  * the recursive sysctl_kern_smp_topology_spec_internal().
2791  */
2792 static int
2793 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2794 {
2795 	struct sbuf *topo;
2796 	int err;
2797 
2798 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2799 
2800 	topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2801 	if (topo == NULL)
2802 		return (ENOMEM);
2803 
2804 	sbuf_printf(topo, "<groups>\n");
2805 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2806 	sbuf_printf(topo, "</groups>\n");
2807 
2808 	if (err == 0) {
2809 		sbuf_finish(topo);
2810 		err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2811 	}
2812 	sbuf_delete(topo);
2813 	return (err);
2814 }
2815 
2816 #endif
2817 
2818 static int
2819 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2820 {
2821 	int error, new_val, period;
2822 
2823 	period = 1000000 / realstathz;
2824 	new_val = period * sched_slice;
2825 	error = sysctl_handle_int(oidp, &new_val, 0, req);
2826 	if (error != 0 || req->newptr == NULL)
2827 		return (error);
2828 	if (new_val <= 0)
2829 		return (EINVAL);
2830 	sched_slice = imax(1, (new_val + period / 2) / period);
2831 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
2832 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2833 	    realstathz);
2834 	return (0);
2835 }
2836 
2837 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2838 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2839     "Scheduler name");
2840 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2841     NULL, 0, sysctl_kern_quantum, "I",
2842     "Quantum for timeshare threads in microseconds");
2843 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2844     "Quantum for timeshare threads in stathz ticks");
2845 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2846     "Interactivity score threshold");
2847 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2848     &preempt_thresh, 0,
2849     "Maximal (lowest) priority for preemption");
2850 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2851     "Assign static kernel priorities to sleeping threads");
2852 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2853     "Number of times idle thread will spin waiting for new work");
2854 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2855     &sched_idlespinthresh, 0,
2856     "Threshold before we will permit idle thread spinning");
2857 #ifdef SMP
2858 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2859     "Number of hz ticks to keep thread affinity for");
2860 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2861     "Enables the long-term load balancer");
2862 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2863     &balance_interval, 0,
2864     "Average period in stathz ticks to run the long-term balancer");
2865 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2866     "Attempts to steal work from other cores before idling");
2867 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2868     "Minimum load on remote CPU before we'll steal");
2869 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2870     CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2871     "XML dump of detected CPU topology");
2872 #endif
2873 
2874 /* ps compat.  All cpu percentages from ULE are weighted. */
2875 static int ccpu = 0;
2876 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2877