xref: /freebsd/sys/kern/sched_ule.c (revision 445ed7b40948c160f2f7d363d2d0ae1ffac4aabd)
1 /*-
2  * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice unmodified, this list of conditions, and the following
10  *    disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25  */
26 
27 /*
28  * This file implements the ULE scheduler.  ULE supports independent CPU
29  * run queues and fine grain locking.  It has superior interactive
30  * performance under load even on uni-processor systems.
31  *
32  * etymology:
33  *   ULE is the last three letters in schedule.  It owes its name to a
34  * generic user created for a scheduling system by Paul Mikesell at
35  * Isilon Systems and a general lack of creativity on the part of the author.
36  */
37 
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
40 
41 #include "opt_hwpmc_hooks.h"
42 #include "opt_sched.h"
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/kdb.h>
47 #include <sys/kernel.h>
48 #include <sys/ktr.h>
49 #include <sys/limits.h>
50 #include <sys/lock.h>
51 #include <sys/mutex.h>
52 #include <sys/proc.h>
53 #include <sys/resource.h>
54 #include <sys/resourcevar.h>
55 #include <sys/sched.h>
56 #include <sys/sdt.h>
57 #include <sys/smp.h>
58 #include <sys/sx.h>
59 #include <sys/sysctl.h>
60 #include <sys/sysproto.h>
61 #include <sys/turnstile.h>
62 #include <sys/umtx.h>
63 #include <sys/vmmeter.h>
64 #include <sys/cpuset.h>
65 #include <sys/sbuf.h>
66 
67 #ifdef HWPMC_HOOKS
68 #include <sys/pmckern.h>
69 #endif
70 
71 #ifdef KDTRACE_HOOKS
72 #include <sys/dtrace_bsd.h>
73 int				dtrace_vtime_active;
74 dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
75 #endif
76 
77 #include <machine/cpu.h>
78 #include <machine/smp.h>
79 
80 #define	KTR_ULE	0
81 
82 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
83 #define	TDQ_NAME_LEN	(sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
84 #define	TDQ_LOADNAME_LEN	(sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
85 
86 /*
87  * Thread scheduler specific section.  All fields are protected
88  * by the thread lock.
89  */
90 struct td_sched {
91 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
92 	short		ts_flags;	/* TSF_* flags. */
93 	int		ts_cpu;		/* CPU that we have affinity for. */
94 	int		ts_rltick;	/* Real last tick, for affinity. */
95 	int		ts_slice;	/* Ticks of slice remaining. */
96 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
97 	u_int		ts_runtime;	/* Number of ticks we were running */
98 	int		ts_ltick;	/* Last tick that we were running on */
99 	int		ts_ftick;	/* First tick that we were running on */
100 	int		ts_ticks;	/* Tick count */
101 #ifdef KTR
102 	char		ts_name[TS_NAME_LEN];
103 #endif
104 };
105 /* flags kept in ts_flags */
106 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
107 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
108 
109 static struct td_sched td_sched0;
110 
111 #define	THREAD_CAN_MIGRATE(td)	((td)->td_pinned == 0)
112 #define	THREAD_CAN_SCHED(td, cpu)	\
113     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
114 
115 /*
116  * Priority ranges used for interactive and non-interactive timeshare
117  * threads.  The timeshare priorities are split up into four ranges.
118  * The first range handles interactive threads.  The last three ranges
119  * (NHALF, x, and NHALF) handle non-interactive threads with the outer
120  * ranges supporting nice values.
121  */
122 #define	PRI_TIMESHARE_RANGE	(PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
123 #define	PRI_INTERACT_RANGE	((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
124 #define	PRI_BATCH_RANGE		(PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
125 
126 #define	PRI_MIN_INTERACT	PRI_MIN_TIMESHARE
127 #define	PRI_MAX_INTERACT	(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
128 #define	PRI_MIN_BATCH		(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
129 #define	PRI_MAX_BATCH		PRI_MAX_TIMESHARE
130 
131 /*
132  * Cpu percentage computation macros and defines.
133  *
134  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
135  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
136  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
137  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
138  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
139  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
140  */
141 #define	SCHED_TICK_SECS		10
142 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
143 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
144 #define	SCHED_TICK_SHIFT	10
145 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
146 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
147 
148 /*
149  * These macros determine priorities for non-interactive threads.  They are
150  * assigned a priority based on their recent cpu utilization as expressed
151  * by the ratio of ticks to the tick total.  NHALF priorities at the start
152  * and end of the MIN to MAX timeshare range are only reachable with negative
153  * or positive nice respectively.
154  *
155  * PRI_RANGE:	Priority range for utilization dependent priorities.
156  * PRI_NRESV:	Number of nice values.
157  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
158  * PRI_NICE:	Determines the part of the priority inherited from nice.
159  */
160 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
161 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
162 #define	SCHED_PRI_MIN		(PRI_MIN_BATCH + SCHED_PRI_NHALF)
163 #define	SCHED_PRI_MAX		(PRI_MAX_BATCH - SCHED_PRI_NHALF)
164 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
165 #define	SCHED_PRI_TICKS(ts)						\
166     (SCHED_TICK_HZ((ts)) /						\
167     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
168 #define	SCHED_PRI_NICE(nice)	(nice)
169 
170 /*
171  * These determine the interactivity of a process.  Interactivity differs from
172  * cpu utilization in that it expresses the voluntary time slept vs time ran
173  * while cpu utilization includes all time not running.  This more accurately
174  * models the intent of the thread.
175  *
176  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
177  *		before throttling back.
178  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
179  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
180  * INTERACT_THRESH:	Threshold for placement on the current runq.
181  */
182 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
183 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
184 #define	SCHED_INTERACT_MAX	(100)
185 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
186 #define	SCHED_INTERACT_THRESH	(30)
187 
188 /*
189  * These parameters determine the slice behavior for batch work.
190  */
191 #define	SCHED_SLICE_DEFAULT_DIVISOR	10	/* ~94 ms, 12 stathz ticks. */
192 #define	SCHED_SLICE_MIN_DIVISOR		6	/* DEFAULT/MIN = ~16 ms. */
193 
194 /* Flags kept in td_flags. */
195 #define	TDF_SLICEEND	TDF_SCHED2	/* Thread time slice is over. */
196 
197 /*
198  * tickincr:		Converts a stathz tick into a hz domain scaled by
199  *			the shift factor.  Without the shift the error rate
200  *			due to rounding would be unacceptably high.
201  * realstathz:		stathz is sometimes 0 and run off of hz.
202  * sched_slice:		Runtime of each thread before rescheduling.
203  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
204  */
205 static int sched_interact = SCHED_INTERACT_THRESH;
206 static int tickincr = 8 << SCHED_TICK_SHIFT;
207 static int realstathz = 127;	/* reset during boot. */
208 static int sched_slice = 10;	/* reset during boot. */
209 static int sched_slice_min = 1;	/* reset during boot. */
210 #ifdef PREEMPTION
211 #ifdef FULL_PREEMPTION
212 static int preempt_thresh = PRI_MAX_IDLE;
213 #else
214 static int preempt_thresh = PRI_MIN_KERN;
215 #endif
216 #else
217 static int preempt_thresh = 0;
218 #endif
219 static int static_boost = PRI_MIN_BATCH;
220 static int sched_idlespins = 10000;
221 static int sched_idlespinthresh = -1;
222 
223 /*
224  * tdq - per processor runqs and statistics.  All fields are protected by the
225  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
226  * locking in sched_pickcpu();
227  */
228 struct tdq {
229 	/*
230 	 * Ordered to improve efficiency of cpu_search() and switch().
231 	 * tdq_lock is padded to avoid false sharing with tdq_load and
232 	 * tdq_cpu_idle.
233 	 */
234 	struct mtx_padalign tdq_lock;		/* run queue lock. */
235 	struct cpu_group *tdq_cg;		/* Pointer to cpu topology. */
236 	volatile int	tdq_load;		/* Aggregate load. */
237 	volatile int	tdq_cpu_idle;		/* cpu_idle() is active. */
238 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
239 	int		tdq_transferable;	/* Transferable thread count. */
240 	short		tdq_switchcnt;		/* Switches this tick. */
241 	short		tdq_oldswitchcnt;	/* Switches last tick. */
242 	u_char		tdq_lowpri;		/* Lowest priority thread. */
243 	u_char		tdq_ipipending;		/* IPI pending. */
244 	u_char		tdq_idx;		/* Current insert index. */
245 	u_char		tdq_ridx;		/* Current removal index. */
246 	struct runq	tdq_realtime;		/* real-time run queue. */
247 	struct runq	tdq_timeshare;		/* timeshare run queue. */
248 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
249 	char		tdq_name[TDQ_NAME_LEN];
250 #ifdef KTR
251 	char		tdq_loadname[TDQ_LOADNAME_LEN];
252 #endif
253 } __aligned(64);
254 
255 /* Idle thread states and config. */
256 #define	TDQ_RUNNING	1
257 #define	TDQ_IDLE	2
258 
259 #ifdef SMP
260 struct cpu_group *cpu_top;		/* CPU topology */
261 
262 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
263 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
264 
265 /*
266  * Run-time tunables.
267  */
268 static int rebalance = 1;
269 static int balance_interval = 128;	/* Default set in sched_initticks(). */
270 static int affinity;
271 static int steal_idle = 1;
272 static int steal_thresh = 2;
273 
274 /*
275  * One thread queue per processor.
276  */
277 static struct tdq	tdq_cpu[MAXCPU];
278 static struct tdq	*balance_tdq;
279 static int balance_ticks;
280 static DPCPU_DEFINE(uint32_t, randomval);
281 
282 #define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
283 #define	TDQ_CPU(x)	(&tdq_cpu[(x)])
284 #define	TDQ_ID(x)	((int)((x) - tdq_cpu))
285 #else	/* !SMP */
286 static struct tdq	tdq_cpu;
287 
288 #define	TDQ_ID(x)	(0)
289 #define	TDQ_SELF()	(&tdq_cpu)
290 #define	TDQ_CPU(x)	(&tdq_cpu)
291 #endif
292 
293 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
294 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
295 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
296 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
297 #define	TDQ_LOCKPTR(t)		((struct mtx *)(&(t)->tdq_lock))
298 
299 static void sched_priority(struct thread *);
300 static void sched_thread_priority(struct thread *, u_char);
301 static int sched_interact_score(struct thread *);
302 static void sched_interact_update(struct thread *);
303 static void sched_interact_fork(struct thread *);
304 static void sched_pctcpu_update(struct td_sched *, int);
305 
306 /* Operations on per processor queues */
307 static struct thread *tdq_choose(struct tdq *);
308 static void tdq_setup(struct tdq *);
309 static void tdq_load_add(struct tdq *, struct thread *);
310 static void tdq_load_rem(struct tdq *, struct thread *);
311 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
312 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
313 static inline int sched_shouldpreempt(int, int, int);
314 void tdq_print(int cpu);
315 static void runq_print(struct runq *rq);
316 static void tdq_add(struct tdq *, struct thread *, int);
317 #ifdef SMP
318 static int tdq_move(struct tdq *, struct tdq *);
319 static int tdq_idled(struct tdq *);
320 static void tdq_notify(struct tdq *, struct thread *);
321 static struct thread *tdq_steal(struct tdq *, int);
322 static struct thread *runq_steal(struct runq *, int);
323 static int sched_pickcpu(struct thread *, int);
324 static void sched_balance(void);
325 static int sched_balance_pair(struct tdq *, struct tdq *);
326 static inline struct tdq *sched_setcpu(struct thread *, int, int);
327 static inline void thread_unblock_switch(struct thread *, struct mtx *);
328 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
329 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
330 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
331     struct cpu_group *cg, int indent);
332 #endif
333 
334 static void sched_setup(void *dummy);
335 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
336 
337 static void sched_initticks(void *dummy);
338 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
339     NULL);
340 
341 SDT_PROVIDER_DEFINE(sched);
342 
343 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
344     "struct proc *", "uint8_t");
345 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
346     "struct proc *", "void *");
347 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
348     "struct proc *", "void *", "int");
349 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
350     "struct proc *", "uint8_t", "struct thread *");
351 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
352 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
353     "struct proc *");
354 SDT_PROBE_DEFINE(sched, , , on__cpu);
355 SDT_PROBE_DEFINE(sched, , , remain__cpu);
356 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
357     "struct proc *");
358 
359 /*
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 
1041 	/*
1042 	 * Make sure that our caller's earlier update to tdq_load is
1043 	 * globally visible before we read tdq_cpu_idle.  Idle thread
1044 	 * accesses both of them without locks, and the order is important.
1045 	 */
1046 	mb();
1047 
1048 	if (TD_IS_IDLETHREAD(ctd)) {
1049 		/*
1050 		 * If the MD code has an idle wakeup routine try that before
1051 		 * falling back to IPI.
1052 		 */
1053 		if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1054 			return;
1055 	}
1056 	tdq->tdq_ipipending = 1;
1057 	ipi_cpu(cpu, IPI_PREEMPT);
1058 }
1059 
1060 /*
1061  * Steals load from a timeshare queue.  Honors the rotating queue head
1062  * index.
1063  */
1064 static struct thread *
1065 runq_steal_from(struct runq *rq, int cpu, u_char start)
1066 {
1067 	struct rqbits *rqb;
1068 	struct rqhead *rqh;
1069 	struct thread *td, *first;
1070 	int bit;
1071 	int i;
1072 
1073 	rqb = &rq->rq_status;
1074 	bit = start & (RQB_BPW -1);
1075 	first = NULL;
1076 again:
1077 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1078 		if (rqb->rqb_bits[i] == 0)
1079 			continue;
1080 		if (bit == 0)
1081 			bit = RQB_FFS(rqb->rqb_bits[i]);
1082 		for (; bit < RQB_BPW; bit++) {
1083 			if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1084 				continue;
1085 			rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1086 			TAILQ_FOREACH(td, rqh, td_runq) {
1087 				if (first && THREAD_CAN_MIGRATE(td) &&
1088 				    THREAD_CAN_SCHED(td, cpu))
1089 					return (td);
1090 				first = td;
1091 			}
1092 		}
1093 	}
1094 	if (start != 0) {
1095 		start = 0;
1096 		goto again;
1097 	}
1098 
1099 	if (first && THREAD_CAN_MIGRATE(first) &&
1100 	    THREAD_CAN_SCHED(first, cpu))
1101 		return (first);
1102 	return (NULL);
1103 }
1104 
1105 /*
1106  * Steals load from a standard linear queue.
1107  */
1108 static struct thread *
1109 runq_steal(struct runq *rq, int cpu)
1110 {
1111 	struct rqhead *rqh;
1112 	struct rqbits *rqb;
1113 	struct thread *td;
1114 	int word;
1115 	int bit;
1116 
1117 	rqb = &rq->rq_status;
1118 	for (word = 0; word < RQB_LEN; word++) {
1119 		if (rqb->rqb_bits[word] == 0)
1120 			continue;
1121 		for (bit = 0; bit < RQB_BPW; bit++) {
1122 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1123 				continue;
1124 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1125 			TAILQ_FOREACH(td, rqh, td_runq)
1126 				if (THREAD_CAN_MIGRATE(td) &&
1127 				    THREAD_CAN_SCHED(td, cpu))
1128 					return (td);
1129 		}
1130 	}
1131 	return (NULL);
1132 }
1133 
1134 /*
1135  * Attempt to steal a thread in priority order from a thread queue.
1136  */
1137 static struct thread *
1138 tdq_steal(struct tdq *tdq, int cpu)
1139 {
1140 	struct thread *td;
1141 
1142 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1143 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1144 		return (td);
1145 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1146 	    cpu, tdq->tdq_ridx)) != NULL)
1147 		return (td);
1148 	return (runq_steal(&tdq->tdq_idle, cpu));
1149 }
1150 
1151 /*
1152  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1153  * current lock and returns with the assigned queue locked.
1154  */
1155 static inline struct tdq *
1156 sched_setcpu(struct thread *td, int cpu, int flags)
1157 {
1158 
1159 	struct tdq *tdq;
1160 
1161 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1162 	tdq = TDQ_CPU(cpu);
1163 	td->td_sched->ts_cpu = cpu;
1164 	/*
1165 	 * If the lock matches just return the queue.
1166 	 */
1167 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1168 		return (tdq);
1169 #ifdef notyet
1170 	/*
1171 	 * If the thread isn't running its lockptr is a
1172 	 * turnstile or a sleepqueue.  We can just lock_set without
1173 	 * blocking.
1174 	 */
1175 	if (TD_CAN_RUN(td)) {
1176 		TDQ_LOCK(tdq);
1177 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1178 		return (tdq);
1179 	}
1180 #endif
1181 	/*
1182 	 * The hard case, migration, we need to block the thread first to
1183 	 * prevent order reversals with other cpus locks.
1184 	 */
1185 	spinlock_enter();
1186 	thread_lock_block(td);
1187 	TDQ_LOCK(tdq);
1188 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1189 	spinlock_exit();
1190 	return (tdq);
1191 }
1192 
1193 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1194 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1195 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1196 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1197 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1198 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1199 
1200 static int
1201 sched_pickcpu(struct thread *td, int flags)
1202 {
1203 	struct cpu_group *cg, *ccg;
1204 	struct td_sched *ts;
1205 	struct tdq *tdq;
1206 	cpuset_t mask;
1207 	int cpu, pri, self;
1208 
1209 	self = PCPU_GET(cpuid);
1210 	ts = td->td_sched;
1211 	if (smp_started == 0)
1212 		return (self);
1213 	/*
1214 	 * Don't migrate a running thread from sched_switch().
1215 	 */
1216 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1217 		return (ts->ts_cpu);
1218 	/*
1219 	 * Prefer to run interrupt threads on the processors that generate
1220 	 * the interrupt.
1221 	 */
1222 	pri = td->td_priority;
1223 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1224 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1225 		SCHED_STAT_INC(pickcpu_intrbind);
1226 		ts->ts_cpu = self;
1227 		if (TDQ_CPU(self)->tdq_lowpri > pri) {
1228 			SCHED_STAT_INC(pickcpu_affinity);
1229 			return (ts->ts_cpu);
1230 		}
1231 	}
1232 	/*
1233 	 * If the thread can run on the last cpu and the affinity has not
1234 	 * expired or it is idle run it there.
1235 	 */
1236 	tdq = TDQ_CPU(ts->ts_cpu);
1237 	cg = tdq->tdq_cg;
1238 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1239 	    tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1240 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1241 		if (cg->cg_flags & CG_FLAG_THREAD) {
1242 			CPUSET_FOREACH(cpu, cg->cg_mask) {
1243 				if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1244 					break;
1245 			}
1246 		} else
1247 			cpu = INT_MAX;
1248 		if (cpu > mp_maxid) {
1249 			SCHED_STAT_INC(pickcpu_idle_affinity);
1250 			return (ts->ts_cpu);
1251 		}
1252 	}
1253 	/*
1254 	 * Search for the last level cache CPU group in the tree.
1255 	 * Skip caches with expired affinity time and SMT groups.
1256 	 * Affinity to higher level caches will be handled less aggressively.
1257 	 */
1258 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1259 		if (cg->cg_flags & CG_FLAG_THREAD)
1260 			continue;
1261 		if (!SCHED_AFFINITY(ts, cg->cg_level))
1262 			continue;
1263 		ccg = cg;
1264 	}
1265 	if (ccg != NULL)
1266 		cg = ccg;
1267 	cpu = -1;
1268 	/* Search the group for the less loaded idle CPU we can run now. */
1269 	mask = td->td_cpuset->cs_mask;
1270 	if (cg != NULL && cg != cpu_top &&
1271 	    CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1272 		cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1273 		    INT_MAX, ts->ts_cpu);
1274 	/* Search globally for the less loaded CPU we can run now. */
1275 	if (cpu == -1)
1276 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1277 	/* Search globally for the less loaded CPU. */
1278 	if (cpu == -1)
1279 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1280 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1281 	/*
1282 	 * Compare the lowest loaded cpu to current cpu.
1283 	 */
1284 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1285 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1286 	    TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1287 		SCHED_STAT_INC(pickcpu_local);
1288 		cpu = self;
1289 	} else
1290 		SCHED_STAT_INC(pickcpu_lowest);
1291 	if (cpu != ts->ts_cpu)
1292 		SCHED_STAT_INC(pickcpu_migration);
1293 	return (cpu);
1294 }
1295 #endif
1296 
1297 /*
1298  * Pick the highest priority task we have and return it.
1299  */
1300 static struct thread *
1301 tdq_choose(struct tdq *tdq)
1302 {
1303 	struct thread *td;
1304 
1305 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1306 	td = runq_choose(&tdq->tdq_realtime);
1307 	if (td != NULL)
1308 		return (td);
1309 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1310 	if (td != NULL) {
1311 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1312 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1313 		    td->td_priority));
1314 		return (td);
1315 	}
1316 	td = runq_choose(&tdq->tdq_idle);
1317 	if (td != NULL) {
1318 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1319 		    ("tdq_choose: Invalid priority on idle queue %d",
1320 		    td->td_priority));
1321 		return (td);
1322 	}
1323 
1324 	return (NULL);
1325 }
1326 
1327 /*
1328  * Initialize a thread queue.
1329  */
1330 static void
1331 tdq_setup(struct tdq *tdq)
1332 {
1333 
1334 	if (bootverbose)
1335 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1336 	runq_init(&tdq->tdq_realtime);
1337 	runq_init(&tdq->tdq_timeshare);
1338 	runq_init(&tdq->tdq_idle);
1339 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1340 	    "sched lock %d", (int)TDQ_ID(tdq));
1341 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1342 	    MTX_SPIN | MTX_RECURSE);
1343 #ifdef KTR
1344 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1345 	    "CPU %d load", (int)TDQ_ID(tdq));
1346 #endif
1347 }
1348 
1349 #ifdef SMP
1350 static void
1351 sched_setup_smp(void)
1352 {
1353 	struct tdq *tdq;
1354 	int i;
1355 
1356 	cpu_top = smp_topo();
1357 	CPU_FOREACH(i) {
1358 		tdq = TDQ_CPU(i);
1359 		tdq_setup(tdq);
1360 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1361 		if (tdq->tdq_cg == NULL)
1362 			panic("Can't find cpu group for %d\n", i);
1363 	}
1364 	balance_tdq = TDQ_SELF();
1365 	sched_balance();
1366 }
1367 #endif
1368 
1369 /*
1370  * Setup the thread queues and initialize the topology based on MD
1371  * information.
1372  */
1373 static void
1374 sched_setup(void *dummy)
1375 {
1376 	struct tdq *tdq;
1377 
1378 	tdq = TDQ_SELF();
1379 #ifdef SMP
1380 	sched_setup_smp();
1381 #else
1382 	tdq_setup(tdq);
1383 #endif
1384 
1385 	/* Add thread0's load since it's running. */
1386 	TDQ_LOCK(tdq);
1387 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1388 	tdq_load_add(tdq, &thread0);
1389 	tdq->tdq_lowpri = thread0.td_priority;
1390 	TDQ_UNLOCK(tdq);
1391 }
1392 
1393 /*
1394  * This routine determines time constants after stathz and hz are setup.
1395  */
1396 /* ARGSUSED */
1397 static void
1398 sched_initticks(void *dummy)
1399 {
1400 	int incr;
1401 
1402 	realstathz = stathz ? stathz : hz;
1403 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1404 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1405 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1406 	    realstathz);
1407 
1408 	/*
1409 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1410 	 * hz not being evenly divisible by stathz on all platforms.
1411 	 */
1412 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1413 	/*
1414 	 * This does not work for values of stathz that are more than
1415 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1416 	 */
1417 	if (incr == 0)
1418 		incr = 1;
1419 	tickincr = incr;
1420 #ifdef SMP
1421 	/*
1422 	 * Set the default balance interval now that we know
1423 	 * what realstathz is.
1424 	 */
1425 	balance_interval = realstathz;
1426 	affinity = SCHED_AFFINITY_DEFAULT;
1427 #endif
1428 	if (sched_idlespinthresh < 0)
1429 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1430 }
1431 
1432 
1433 /*
1434  * This is the core of the interactivity algorithm.  Determines a score based
1435  * on past behavior.  It is the ratio of sleep time to run time scaled to
1436  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1437  * differs from the cpu usage because it does not account for time spent
1438  * waiting on a run-queue.  Would be prettier if we had floating point.
1439  */
1440 static int
1441 sched_interact_score(struct thread *td)
1442 {
1443 	struct td_sched *ts;
1444 	int div;
1445 
1446 	ts = td->td_sched;
1447 	/*
1448 	 * The score is only needed if this is likely to be an interactive
1449 	 * task.  Don't go through the expense of computing it if there's
1450 	 * no chance.
1451 	 */
1452 	if (sched_interact <= SCHED_INTERACT_HALF &&
1453 		ts->ts_runtime >= ts->ts_slptime)
1454 			return (SCHED_INTERACT_HALF);
1455 
1456 	if (ts->ts_runtime > ts->ts_slptime) {
1457 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1458 		return (SCHED_INTERACT_HALF +
1459 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1460 	}
1461 	if (ts->ts_slptime > ts->ts_runtime) {
1462 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1463 		return (ts->ts_runtime / div);
1464 	}
1465 	/* runtime == slptime */
1466 	if (ts->ts_runtime)
1467 		return (SCHED_INTERACT_HALF);
1468 
1469 	/*
1470 	 * This can happen if slptime and runtime are 0.
1471 	 */
1472 	return (0);
1473 
1474 }
1475 
1476 /*
1477  * Scale the scheduling priority according to the "interactivity" of this
1478  * process.
1479  */
1480 static void
1481 sched_priority(struct thread *td)
1482 {
1483 	int score;
1484 	int pri;
1485 
1486 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1487 		return;
1488 	/*
1489 	 * If the score is interactive we place the thread in the realtime
1490 	 * queue with a priority that is less than kernel and interrupt
1491 	 * priorities.  These threads are not subject to nice restrictions.
1492 	 *
1493 	 * Scores greater than this are placed on the normal timeshare queue
1494 	 * where the priority is partially decided by the most recent cpu
1495 	 * utilization and the rest is decided by nice value.
1496 	 *
1497 	 * The nice value of the process has a linear effect on the calculated
1498 	 * score.  Negative nice values make it easier for a thread to be
1499 	 * considered interactive.
1500 	 */
1501 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1502 	if (score < sched_interact) {
1503 		pri = PRI_MIN_INTERACT;
1504 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1505 		    sched_interact) * score;
1506 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1507 		    ("sched_priority: invalid interactive priority %d score %d",
1508 		    pri, score));
1509 	} else {
1510 		pri = SCHED_PRI_MIN;
1511 		if (td->td_sched->ts_ticks)
1512 			pri += min(SCHED_PRI_TICKS(td->td_sched),
1513 			    SCHED_PRI_RANGE - 1);
1514 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1515 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1516 		    ("sched_priority: invalid priority %d: nice %d, "
1517 		    "ticks %d ftick %d ltick %d tick pri %d",
1518 		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1519 		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1520 		    SCHED_PRI_TICKS(td->td_sched)));
1521 	}
1522 	sched_user_prio(td, pri);
1523 
1524 	return;
1525 }
1526 
1527 /*
1528  * This routine enforces a maximum limit on the amount of scheduling history
1529  * kept.  It is called after either the slptime or runtime is adjusted.  This
1530  * function is ugly due to integer math.
1531  */
1532 static void
1533 sched_interact_update(struct thread *td)
1534 {
1535 	struct td_sched *ts;
1536 	u_int sum;
1537 
1538 	ts = td->td_sched;
1539 	sum = ts->ts_runtime + ts->ts_slptime;
1540 	if (sum < SCHED_SLP_RUN_MAX)
1541 		return;
1542 	/*
1543 	 * This only happens from two places:
1544 	 * 1) We have added an unusual amount of run time from fork_exit.
1545 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1546 	 */
1547 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1548 		if (ts->ts_runtime > ts->ts_slptime) {
1549 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1550 			ts->ts_slptime = 1;
1551 		} else {
1552 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1553 			ts->ts_runtime = 1;
1554 		}
1555 		return;
1556 	}
1557 	/*
1558 	 * If we have exceeded by more than 1/5th then the algorithm below
1559 	 * will not bring us back into range.  Dividing by two here forces
1560 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1561 	 */
1562 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1563 		ts->ts_runtime /= 2;
1564 		ts->ts_slptime /= 2;
1565 		return;
1566 	}
1567 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1568 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1569 }
1570 
1571 /*
1572  * Scale back the interactivity history when a child thread is created.  The
1573  * history is inherited from the parent but the thread may behave totally
1574  * differently.  For example, a shell spawning a compiler process.  We want
1575  * to learn that the compiler is behaving badly very quickly.
1576  */
1577 static void
1578 sched_interact_fork(struct thread *td)
1579 {
1580 	int ratio;
1581 	int sum;
1582 
1583 	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1584 	if (sum > SCHED_SLP_RUN_FORK) {
1585 		ratio = sum / SCHED_SLP_RUN_FORK;
1586 		td->td_sched->ts_runtime /= ratio;
1587 		td->td_sched->ts_slptime /= ratio;
1588 	}
1589 }
1590 
1591 /*
1592  * Called from proc0_init() to setup the scheduler fields.
1593  */
1594 void
1595 schedinit(void)
1596 {
1597 
1598 	/*
1599 	 * Set up the scheduler specific parts of proc0.
1600 	 */
1601 	proc0.p_sched = NULL; /* XXX */
1602 	thread0.td_sched = &td_sched0;
1603 	td_sched0.ts_ltick = ticks;
1604 	td_sched0.ts_ftick = ticks;
1605 	td_sched0.ts_slice = 0;
1606 }
1607 
1608 /*
1609  * This is only somewhat accurate since given many processes of the same
1610  * priority they will switch when their slices run out, which will be
1611  * at most sched_slice stathz ticks.
1612  */
1613 int
1614 sched_rr_interval(void)
1615 {
1616 
1617 	/* Convert sched_slice from stathz to hz. */
1618 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1619 }
1620 
1621 /*
1622  * Update the percent cpu tracking information when it is requested or
1623  * the total history exceeds the maximum.  We keep a sliding history of
1624  * tick counts that slowly decays.  This is less precise than the 4BSD
1625  * mechanism since it happens with less regular and frequent events.
1626  */
1627 static void
1628 sched_pctcpu_update(struct td_sched *ts, int run)
1629 {
1630 	int t = ticks;
1631 
1632 	if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1633 		ts->ts_ticks = 0;
1634 		ts->ts_ftick = t - SCHED_TICK_TARG;
1635 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1636 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1637 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1638 		ts->ts_ftick = t - SCHED_TICK_TARG;
1639 	}
1640 	if (run)
1641 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1642 	ts->ts_ltick = t;
1643 }
1644 
1645 /*
1646  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1647  * if necessary.  This is the back-end for several priority related
1648  * functions.
1649  */
1650 static void
1651 sched_thread_priority(struct thread *td, u_char prio)
1652 {
1653 	struct td_sched *ts;
1654 	struct tdq *tdq;
1655 	int oldpri;
1656 
1657 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1658 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1659 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1660 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1661 	if (td != curthread && prio < td->td_priority) {
1662 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1663 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1664 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1665 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1666 		    curthread);
1667 	}
1668 	ts = td->td_sched;
1669 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1670 	if (td->td_priority == prio)
1671 		return;
1672 	/*
1673 	 * If the priority has been elevated due to priority
1674 	 * propagation, we may have to move ourselves to a new
1675 	 * queue.  This could be optimized to not re-add in some
1676 	 * cases.
1677 	 */
1678 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1679 		sched_rem(td);
1680 		td->td_priority = prio;
1681 		sched_add(td, SRQ_BORROWING);
1682 		return;
1683 	}
1684 	/*
1685 	 * If the thread is currently running we may have to adjust the lowpri
1686 	 * information so other cpus are aware of our current priority.
1687 	 */
1688 	if (TD_IS_RUNNING(td)) {
1689 		tdq = TDQ_CPU(ts->ts_cpu);
1690 		oldpri = td->td_priority;
1691 		td->td_priority = prio;
1692 		if (prio < tdq->tdq_lowpri)
1693 			tdq->tdq_lowpri = prio;
1694 		else if (tdq->tdq_lowpri == oldpri)
1695 			tdq_setlowpri(tdq, td);
1696 		return;
1697 	}
1698 	td->td_priority = prio;
1699 }
1700 
1701 /*
1702  * Update a thread's priority when it is lent another thread's
1703  * priority.
1704  */
1705 void
1706 sched_lend_prio(struct thread *td, u_char prio)
1707 {
1708 
1709 	td->td_flags |= TDF_BORROWING;
1710 	sched_thread_priority(td, prio);
1711 }
1712 
1713 /*
1714  * Restore a thread's priority when priority propagation is
1715  * over.  The prio argument is the minimum priority the thread
1716  * needs to have to satisfy other possible priority lending
1717  * requests.  If the thread's regular priority is less
1718  * important than prio, the thread will keep a priority boost
1719  * of prio.
1720  */
1721 void
1722 sched_unlend_prio(struct thread *td, u_char prio)
1723 {
1724 	u_char base_pri;
1725 
1726 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1727 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1728 		base_pri = td->td_user_pri;
1729 	else
1730 		base_pri = td->td_base_pri;
1731 	if (prio >= base_pri) {
1732 		td->td_flags &= ~TDF_BORROWING;
1733 		sched_thread_priority(td, base_pri);
1734 	} else
1735 		sched_lend_prio(td, prio);
1736 }
1737 
1738 /*
1739  * Standard entry for setting the priority to an absolute value.
1740  */
1741 void
1742 sched_prio(struct thread *td, u_char prio)
1743 {
1744 	u_char oldprio;
1745 
1746 	/* First, update the base priority. */
1747 	td->td_base_pri = prio;
1748 
1749 	/*
1750 	 * If the thread is borrowing another thread's priority, don't
1751 	 * ever lower the priority.
1752 	 */
1753 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1754 		return;
1755 
1756 	/* Change the real priority. */
1757 	oldprio = td->td_priority;
1758 	sched_thread_priority(td, prio);
1759 
1760 	/*
1761 	 * If the thread is on a turnstile, then let the turnstile update
1762 	 * its state.
1763 	 */
1764 	if (TD_ON_LOCK(td) && oldprio != prio)
1765 		turnstile_adjust(td, oldprio);
1766 }
1767 
1768 /*
1769  * Set the base user priority, does not effect current running priority.
1770  */
1771 void
1772 sched_user_prio(struct thread *td, u_char prio)
1773 {
1774 
1775 	td->td_base_user_pri = prio;
1776 	if (td->td_lend_user_pri <= prio)
1777 		return;
1778 	td->td_user_pri = prio;
1779 }
1780 
1781 void
1782 sched_lend_user_prio(struct thread *td, u_char prio)
1783 {
1784 
1785 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1786 	td->td_lend_user_pri = prio;
1787 	td->td_user_pri = min(prio, td->td_base_user_pri);
1788 	if (td->td_priority > td->td_user_pri)
1789 		sched_prio(td, td->td_user_pri);
1790 	else if (td->td_priority != td->td_user_pri)
1791 		td->td_flags |= TDF_NEEDRESCHED;
1792 }
1793 
1794 /*
1795  * Handle migration from sched_switch().  This happens only for
1796  * cpu binding.
1797  */
1798 static struct mtx *
1799 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1800 {
1801 	struct tdq *tdn;
1802 
1803 	tdn = TDQ_CPU(td->td_sched->ts_cpu);
1804 #ifdef SMP
1805 	tdq_load_rem(tdq, td);
1806 	/*
1807 	 * Do the lock dance required to avoid LOR.  We grab an extra
1808 	 * spinlock nesting to prevent preemption while we're
1809 	 * not holding either run-queue lock.
1810 	 */
1811 	spinlock_enter();
1812 	thread_lock_block(td);	/* This releases the lock on tdq. */
1813 
1814 	/*
1815 	 * Acquire both run-queue locks before placing the thread on the new
1816 	 * run-queue to avoid deadlocks created by placing a thread with a
1817 	 * blocked lock on the run-queue of a remote processor.  The deadlock
1818 	 * occurs when a third processor attempts to lock the two queues in
1819 	 * question while the target processor is spinning with its own
1820 	 * run-queue lock held while waiting for the blocked lock to clear.
1821 	 */
1822 	tdq_lock_pair(tdn, tdq);
1823 	tdq_add(tdn, td, flags);
1824 	tdq_notify(tdn, td);
1825 	TDQ_UNLOCK(tdn);
1826 	spinlock_exit();
1827 #endif
1828 	return (TDQ_LOCKPTR(tdn));
1829 }
1830 
1831 /*
1832  * Variadic version of thread_lock_unblock() that does not assume td_lock
1833  * is blocked.
1834  */
1835 static inline void
1836 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1837 {
1838 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1839 	    (uintptr_t)mtx);
1840 }
1841 
1842 /*
1843  * Switch threads.  This function has to handle threads coming in while
1844  * blocked for some reason, running, or idle.  It also must deal with
1845  * migrating a thread from one queue to another as running threads may
1846  * be assigned elsewhere via binding.
1847  */
1848 void
1849 sched_switch(struct thread *td, struct thread *newtd, int flags)
1850 {
1851 	struct tdq *tdq;
1852 	struct td_sched *ts;
1853 	struct mtx *mtx;
1854 	int srqflag;
1855 	int cpuid, preempted;
1856 
1857 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1858 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1859 
1860 	cpuid = PCPU_GET(cpuid);
1861 	tdq = TDQ_CPU(cpuid);
1862 	ts = td->td_sched;
1863 	mtx = td->td_lock;
1864 	sched_pctcpu_update(ts, 1);
1865 	ts->ts_rltick = ticks;
1866 	td->td_lastcpu = td->td_oncpu;
1867 	td->td_oncpu = NOCPU;
1868 	preempted = !((td->td_flags & TDF_SLICEEND) ||
1869 	    (flags & SWT_RELINQUISH));
1870 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1871 	td->td_owepreempt = 0;
1872 	if (!TD_IS_IDLETHREAD(td))
1873 		tdq->tdq_switchcnt++;
1874 	/*
1875 	 * The lock pointer in an idle thread should never change.  Reset it
1876 	 * to CAN_RUN as well.
1877 	 */
1878 	if (TD_IS_IDLETHREAD(td)) {
1879 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1880 		TD_SET_CAN_RUN(td);
1881 	} else if (TD_IS_RUNNING(td)) {
1882 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1883 		srqflag = preempted ?
1884 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1885 		    SRQ_OURSELF|SRQ_YIELDING;
1886 #ifdef SMP
1887 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1888 			ts->ts_cpu = sched_pickcpu(td, 0);
1889 #endif
1890 		if (ts->ts_cpu == cpuid)
1891 			tdq_runq_add(tdq, td, srqflag);
1892 		else {
1893 			KASSERT(THREAD_CAN_MIGRATE(td) ||
1894 			    (ts->ts_flags & TSF_BOUND) != 0,
1895 			    ("Thread %p shouldn't migrate", td));
1896 			mtx = sched_switch_migrate(tdq, td, srqflag);
1897 		}
1898 	} else {
1899 		/* This thread must be going to sleep. */
1900 		TDQ_LOCK(tdq);
1901 		mtx = thread_lock_block(td);
1902 		tdq_load_rem(tdq, td);
1903 	}
1904 	/*
1905 	 * We enter here with the thread blocked and assigned to the
1906 	 * appropriate cpu run-queue or sleep-queue and with the current
1907 	 * thread-queue locked.
1908 	 */
1909 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1910 	newtd = choosethread();
1911 	/*
1912 	 * Call the MD code to switch contexts if necessary.
1913 	 */
1914 	if (td != newtd) {
1915 #ifdef	HWPMC_HOOKS
1916 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1917 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1918 #endif
1919 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1920 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1921 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1922 		sched_pctcpu_update(newtd->td_sched, 0);
1923 
1924 #ifdef KDTRACE_HOOKS
1925 		/*
1926 		 * If DTrace has set the active vtime enum to anything
1927 		 * other than INACTIVE (0), then it should have set the
1928 		 * function to call.
1929 		 */
1930 		if (dtrace_vtime_active)
1931 			(*dtrace_vtime_switch_func)(newtd);
1932 #endif
1933 
1934 		cpu_switch(td, newtd, mtx);
1935 		/*
1936 		 * We may return from cpu_switch on a different cpu.  However,
1937 		 * we always return with td_lock pointing to the current cpu's
1938 		 * run queue lock.
1939 		 */
1940 		cpuid = PCPU_GET(cpuid);
1941 		tdq = TDQ_CPU(cpuid);
1942 		lock_profile_obtain_lock_success(
1943 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1944 
1945 		SDT_PROBE0(sched, , , on__cpu);
1946 #ifdef	HWPMC_HOOKS
1947 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1948 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1949 #endif
1950 	} else {
1951 		thread_unblock_switch(td, mtx);
1952 		SDT_PROBE0(sched, , , remain__cpu);
1953 	}
1954 	/*
1955 	 * Assert that all went well and return.
1956 	 */
1957 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1958 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1959 	td->td_oncpu = cpuid;
1960 }
1961 
1962 /*
1963  * Adjust thread priorities as a result of a nice request.
1964  */
1965 void
1966 sched_nice(struct proc *p, int nice)
1967 {
1968 	struct thread *td;
1969 
1970 	PROC_LOCK_ASSERT(p, MA_OWNED);
1971 
1972 	p->p_nice = nice;
1973 	FOREACH_THREAD_IN_PROC(p, td) {
1974 		thread_lock(td);
1975 		sched_priority(td);
1976 		sched_prio(td, td->td_base_user_pri);
1977 		thread_unlock(td);
1978 	}
1979 }
1980 
1981 /*
1982  * Record the sleep time for the interactivity scorer.
1983  */
1984 void
1985 sched_sleep(struct thread *td, int prio)
1986 {
1987 
1988 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1989 
1990 	td->td_slptick = ticks;
1991 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1992 		td->td_flags |= TDF_CANSWAP;
1993 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1994 		return;
1995 	if (static_boost == 1 && prio)
1996 		sched_prio(td, prio);
1997 	else if (static_boost && td->td_priority > static_boost)
1998 		sched_prio(td, static_boost);
1999 }
2000 
2001 /*
2002  * Schedule a thread to resume execution and record how long it voluntarily
2003  * slept.  We also update the pctcpu, interactivity, and priority.
2004  */
2005 void
2006 sched_wakeup(struct thread *td)
2007 {
2008 	struct td_sched *ts;
2009 	int slptick;
2010 
2011 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2012 	ts = td->td_sched;
2013 	td->td_flags &= ~TDF_CANSWAP;
2014 	/*
2015 	 * If we slept for more than a tick update our interactivity and
2016 	 * priority.
2017 	 */
2018 	slptick = td->td_slptick;
2019 	td->td_slptick = 0;
2020 	if (slptick && slptick != ticks) {
2021 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2022 		sched_interact_update(td);
2023 		sched_pctcpu_update(ts, 0);
2024 	}
2025 	/*
2026 	 * Reset the slice value since we slept and advanced the round-robin.
2027 	 */
2028 	ts->ts_slice = 0;
2029 	sched_add(td, SRQ_BORING);
2030 }
2031 
2032 /*
2033  * Penalize the parent for creating a new child and initialize the child's
2034  * priority.
2035  */
2036 void
2037 sched_fork(struct thread *td, struct thread *child)
2038 {
2039 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2040 	sched_pctcpu_update(td->td_sched, 1);
2041 	sched_fork_thread(td, child);
2042 	/*
2043 	 * Penalize the parent and child for forking.
2044 	 */
2045 	sched_interact_fork(child);
2046 	sched_priority(child);
2047 	td->td_sched->ts_runtime += tickincr;
2048 	sched_interact_update(td);
2049 	sched_priority(td);
2050 }
2051 
2052 /*
2053  * Fork a new thread, may be within the same process.
2054  */
2055 void
2056 sched_fork_thread(struct thread *td, struct thread *child)
2057 {
2058 	struct td_sched *ts;
2059 	struct td_sched *ts2;
2060 	struct tdq *tdq;
2061 
2062 	tdq = TDQ_SELF();
2063 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2064 	/*
2065 	 * Initialize child.
2066 	 */
2067 	ts = td->td_sched;
2068 	ts2 = child->td_sched;
2069 	child->td_lock = TDQ_LOCKPTR(tdq);
2070 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2071 	ts2->ts_cpu = ts->ts_cpu;
2072 	ts2->ts_flags = 0;
2073 	/*
2074 	 * Grab our parents cpu estimation information.
2075 	 */
2076 	ts2->ts_ticks = ts->ts_ticks;
2077 	ts2->ts_ltick = ts->ts_ltick;
2078 	ts2->ts_ftick = ts->ts_ftick;
2079 	/*
2080 	 * Do not inherit any borrowed priority from the parent.
2081 	 */
2082 	child->td_priority = child->td_base_pri;
2083 	/*
2084 	 * And update interactivity score.
2085 	 */
2086 	ts2->ts_slptime = ts->ts_slptime;
2087 	ts2->ts_runtime = ts->ts_runtime;
2088 	/* Attempt to quickly learn interactivity. */
2089 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2090 #ifdef KTR
2091 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2092 #endif
2093 }
2094 
2095 /*
2096  * Adjust the priority class of a thread.
2097  */
2098 void
2099 sched_class(struct thread *td, int class)
2100 {
2101 
2102 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2103 	if (td->td_pri_class == class)
2104 		return;
2105 	td->td_pri_class = class;
2106 }
2107 
2108 /*
2109  * Return some of the child's priority and interactivity to the parent.
2110  */
2111 void
2112 sched_exit(struct proc *p, struct thread *child)
2113 {
2114 	struct thread *td;
2115 
2116 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2117 	    "prio:%d", child->td_priority);
2118 	PROC_LOCK_ASSERT(p, MA_OWNED);
2119 	td = FIRST_THREAD_IN_PROC(p);
2120 	sched_exit_thread(td, child);
2121 }
2122 
2123 /*
2124  * Penalize another thread for the time spent on this one.  This helps to
2125  * worsen the priority and interactivity of processes which schedule batch
2126  * jobs such as make.  This has little effect on the make process itself but
2127  * causes new processes spawned by it to receive worse scores immediately.
2128  */
2129 void
2130 sched_exit_thread(struct thread *td, struct thread *child)
2131 {
2132 
2133 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2134 	    "prio:%d", child->td_priority);
2135 	/*
2136 	 * Give the child's runtime to the parent without returning the
2137 	 * sleep time as a penalty to the parent.  This causes shells that
2138 	 * launch expensive things to mark their children as expensive.
2139 	 */
2140 	thread_lock(td);
2141 	td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2142 	sched_interact_update(td);
2143 	sched_priority(td);
2144 	thread_unlock(td);
2145 }
2146 
2147 void
2148 sched_preempt(struct thread *td)
2149 {
2150 	struct tdq *tdq;
2151 
2152 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2153 
2154 	thread_lock(td);
2155 	tdq = TDQ_SELF();
2156 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2157 	tdq->tdq_ipipending = 0;
2158 	if (td->td_priority > tdq->tdq_lowpri) {
2159 		int flags;
2160 
2161 		flags = SW_INVOL | SW_PREEMPT;
2162 		if (td->td_critnest > 1)
2163 			td->td_owepreempt = 1;
2164 		else if (TD_IS_IDLETHREAD(td))
2165 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2166 		else
2167 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2168 	}
2169 	thread_unlock(td);
2170 }
2171 
2172 /*
2173  * Fix priorities on return to user-space.  Priorities may be elevated due
2174  * to static priorities in msleep() or similar.
2175  */
2176 void
2177 sched_userret(struct thread *td)
2178 {
2179 	/*
2180 	 * XXX we cheat slightly on the locking here to avoid locking in
2181 	 * the usual case.  Setting td_priority here is essentially an
2182 	 * incomplete workaround for not setting it properly elsewhere.
2183 	 * Now that some interrupt handlers are threads, not setting it
2184 	 * properly elsewhere can clobber it in the window between setting
2185 	 * it here and returning to user mode, so don't waste time setting
2186 	 * it perfectly here.
2187 	 */
2188 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2189 	    ("thread with borrowed priority returning to userland"));
2190 	if (td->td_priority != td->td_user_pri) {
2191 		thread_lock(td);
2192 		td->td_priority = td->td_user_pri;
2193 		td->td_base_pri = td->td_user_pri;
2194 		tdq_setlowpri(TDQ_SELF(), td);
2195 		thread_unlock(td);
2196         }
2197 }
2198 
2199 /*
2200  * Handle a stathz tick.  This is really only relevant for timeshare
2201  * threads.
2202  */
2203 void
2204 sched_clock(struct thread *td)
2205 {
2206 	struct tdq *tdq;
2207 	struct td_sched *ts;
2208 
2209 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2210 	tdq = TDQ_SELF();
2211 #ifdef SMP
2212 	/*
2213 	 * We run the long term load balancer infrequently on the first cpu.
2214 	 */
2215 	if (balance_tdq == tdq) {
2216 		if (balance_ticks && --balance_ticks == 0)
2217 			sched_balance();
2218 	}
2219 #endif
2220 	/*
2221 	 * Save the old switch count so we have a record of the last ticks
2222 	 * activity.   Initialize the new switch count based on our load.
2223 	 * If there is some activity seed it to reflect that.
2224 	 */
2225 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2226 	tdq->tdq_switchcnt = tdq->tdq_load;
2227 	/*
2228 	 * Advance the insert index once for each tick to ensure that all
2229 	 * threads get a chance to run.
2230 	 */
2231 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2232 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2233 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2234 			tdq->tdq_ridx = tdq->tdq_idx;
2235 	}
2236 	ts = td->td_sched;
2237 	sched_pctcpu_update(ts, 1);
2238 	if (td->td_pri_class & PRI_FIFO_BIT)
2239 		return;
2240 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2241 		/*
2242 		 * We used a tick; charge it to the thread so
2243 		 * that we can compute our interactivity.
2244 		 */
2245 		td->td_sched->ts_runtime += tickincr;
2246 		sched_interact_update(td);
2247 		sched_priority(td);
2248 	}
2249 
2250 	/*
2251 	 * Force a context switch if the current thread has used up a full
2252 	 * time slice (default is 100ms).
2253 	 */
2254 	if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2255 		ts->ts_slice = 0;
2256 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2257 	}
2258 }
2259 
2260 /*
2261  * Called once per hz tick.
2262  */
2263 void
2264 sched_tick(int cnt)
2265 {
2266 
2267 }
2268 
2269 /*
2270  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2271  * cooperative idle threads.
2272  */
2273 int
2274 sched_runnable(void)
2275 {
2276 	struct tdq *tdq;
2277 	int load;
2278 
2279 	load = 1;
2280 
2281 	tdq = TDQ_SELF();
2282 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2283 		if (tdq->tdq_load > 0)
2284 			goto out;
2285 	} else
2286 		if (tdq->tdq_load - 1 > 0)
2287 			goto out;
2288 	load = 0;
2289 out:
2290 	return (load);
2291 }
2292 
2293 /*
2294  * Choose the highest priority thread to run.  The thread is removed from
2295  * the run-queue while running however the load remains.  For SMP we set
2296  * the tdq in the global idle bitmask if it idles here.
2297  */
2298 struct thread *
2299 sched_choose(void)
2300 {
2301 	struct thread *td;
2302 	struct tdq *tdq;
2303 
2304 	tdq = TDQ_SELF();
2305 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2306 	td = tdq_choose(tdq);
2307 	if (td) {
2308 		tdq_runq_rem(tdq, td);
2309 		tdq->tdq_lowpri = td->td_priority;
2310 		return (td);
2311 	}
2312 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2313 	return (PCPU_GET(idlethread));
2314 }
2315 
2316 /*
2317  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2318  * we always request it once we exit a critical section.
2319  */
2320 static inline void
2321 sched_setpreempt(struct thread *td)
2322 {
2323 	struct thread *ctd;
2324 	int cpri;
2325 	int pri;
2326 
2327 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2328 
2329 	ctd = curthread;
2330 	pri = td->td_priority;
2331 	cpri = ctd->td_priority;
2332 	if (pri < cpri)
2333 		ctd->td_flags |= TDF_NEEDRESCHED;
2334 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2335 		return;
2336 	if (!sched_shouldpreempt(pri, cpri, 0))
2337 		return;
2338 	ctd->td_owepreempt = 1;
2339 }
2340 
2341 /*
2342  * Add a thread to a thread queue.  Select the appropriate runq and add the
2343  * thread to it.  This is the internal function called when the tdq is
2344  * predetermined.
2345  */
2346 void
2347 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2348 {
2349 
2350 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2351 	KASSERT((td->td_inhibitors == 0),
2352 	    ("sched_add: trying to run inhibited thread"));
2353 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2354 	    ("sched_add: bad thread state"));
2355 	KASSERT(td->td_flags & TDF_INMEM,
2356 	    ("sched_add: thread swapped out"));
2357 
2358 	if (td->td_priority < tdq->tdq_lowpri)
2359 		tdq->tdq_lowpri = td->td_priority;
2360 	tdq_runq_add(tdq, td, flags);
2361 	tdq_load_add(tdq, td);
2362 }
2363 
2364 /*
2365  * Select the target thread queue and add a thread to it.  Request
2366  * preemption or IPI a remote processor if required.
2367  */
2368 void
2369 sched_add(struct thread *td, int flags)
2370 {
2371 	struct tdq *tdq;
2372 #ifdef SMP
2373 	int cpu;
2374 #endif
2375 
2376 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2377 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2378 	    sched_tdname(curthread));
2379 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2380 	    KTR_ATTR_LINKED, sched_tdname(td));
2381 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2382 	    flags & SRQ_PREEMPTED);
2383 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2384 	/*
2385 	 * Recalculate the priority before we select the target cpu or
2386 	 * run-queue.
2387 	 */
2388 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2389 		sched_priority(td);
2390 #ifdef SMP
2391 	/*
2392 	 * Pick the destination cpu and if it isn't ours transfer to the
2393 	 * target cpu.
2394 	 */
2395 	cpu = sched_pickcpu(td, flags);
2396 	tdq = sched_setcpu(td, cpu, flags);
2397 	tdq_add(tdq, td, flags);
2398 	if (cpu != PCPU_GET(cpuid)) {
2399 		tdq_notify(tdq, td);
2400 		return;
2401 	}
2402 #else
2403 	tdq = TDQ_SELF();
2404 	TDQ_LOCK(tdq);
2405 	/*
2406 	 * Now that the thread is moving to the run-queue, set the lock
2407 	 * to the scheduler's lock.
2408 	 */
2409 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2410 	tdq_add(tdq, td, flags);
2411 #endif
2412 	if (!(flags & SRQ_YIELDING))
2413 		sched_setpreempt(td);
2414 }
2415 
2416 /*
2417  * Remove a thread from a run-queue without running it.  This is used
2418  * when we're stealing a thread from a remote queue.  Otherwise all threads
2419  * exit by calling sched_exit_thread() and sched_throw() themselves.
2420  */
2421 void
2422 sched_rem(struct thread *td)
2423 {
2424 	struct tdq *tdq;
2425 
2426 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2427 	    "prio:%d", td->td_priority);
2428 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2429 	tdq = TDQ_CPU(td->td_sched->ts_cpu);
2430 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2431 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2432 	KASSERT(TD_ON_RUNQ(td),
2433 	    ("sched_rem: thread not on run queue"));
2434 	tdq_runq_rem(tdq, td);
2435 	tdq_load_rem(tdq, td);
2436 	TD_SET_CAN_RUN(td);
2437 	if (td->td_priority == tdq->tdq_lowpri)
2438 		tdq_setlowpri(tdq, NULL);
2439 }
2440 
2441 /*
2442  * Fetch cpu utilization information.  Updates on demand.
2443  */
2444 fixpt_t
2445 sched_pctcpu(struct thread *td)
2446 {
2447 	fixpt_t pctcpu;
2448 	struct td_sched *ts;
2449 
2450 	pctcpu = 0;
2451 	ts = td->td_sched;
2452 	if (ts == NULL)
2453 		return (0);
2454 
2455 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2456 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2457 	if (ts->ts_ticks) {
2458 		int rtick;
2459 
2460 		/* How many rtick per second ? */
2461 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2462 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2463 	}
2464 
2465 	return (pctcpu);
2466 }
2467 
2468 /*
2469  * Enforce affinity settings for a thread.  Called after adjustments to
2470  * cpumask.
2471  */
2472 void
2473 sched_affinity(struct thread *td)
2474 {
2475 #ifdef SMP
2476 	struct td_sched *ts;
2477 
2478 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2479 	ts = td->td_sched;
2480 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2481 		return;
2482 	if (TD_ON_RUNQ(td)) {
2483 		sched_rem(td);
2484 		sched_add(td, SRQ_BORING);
2485 		return;
2486 	}
2487 	if (!TD_IS_RUNNING(td))
2488 		return;
2489 	/*
2490 	 * Force a switch before returning to userspace.  If the
2491 	 * target thread is not running locally send an ipi to force
2492 	 * the issue.
2493 	 */
2494 	td->td_flags |= TDF_NEEDRESCHED;
2495 	if (td != curthread)
2496 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2497 #endif
2498 }
2499 
2500 /*
2501  * Bind a thread to a target cpu.
2502  */
2503 void
2504 sched_bind(struct thread *td, int cpu)
2505 {
2506 	struct td_sched *ts;
2507 
2508 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2509 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2510 	ts = td->td_sched;
2511 	if (ts->ts_flags & TSF_BOUND)
2512 		sched_unbind(td);
2513 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2514 	ts->ts_flags |= TSF_BOUND;
2515 	sched_pin();
2516 	if (PCPU_GET(cpuid) == cpu)
2517 		return;
2518 	ts->ts_cpu = cpu;
2519 	/* When we return from mi_switch we'll be on the correct cpu. */
2520 	mi_switch(SW_VOL, NULL);
2521 }
2522 
2523 /*
2524  * Release a bound thread.
2525  */
2526 void
2527 sched_unbind(struct thread *td)
2528 {
2529 	struct td_sched *ts;
2530 
2531 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2532 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2533 	ts = td->td_sched;
2534 	if ((ts->ts_flags & TSF_BOUND) == 0)
2535 		return;
2536 	ts->ts_flags &= ~TSF_BOUND;
2537 	sched_unpin();
2538 }
2539 
2540 int
2541 sched_is_bound(struct thread *td)
2542 {
2543 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2544 	return (td->td_sched->ts_flags & TSF_BOUND);
2545 }
2546 
2547 /*
2548  * Basic yield call.
2549  */
2550 void
2551 sched_relinquish(struct thread *td)
2552 {
2553 	thread_lock(td);
2554 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2555 	thread_unlock(td);
2556 }
2557 
2558 /*
2559  * Return the total system load.
2560  */
2561 int
2562 sched_load(void)
2563 {
2564 #ifdef SMP
2565 	int total;
2566 	int i;
2567 
2568 	total = 0;
2569 	CPU_FOREACH(i)
2570 		total += TDQ_CPU(i)->tdq_sysload;
2571 	return (total);
2572 #else
2573 	return (TDQ_SELF()->tdq_sysload);
2574 #endif
2575 }
2576 
2577 int
2578 sched_sizeof_proc(void)
2579 {
2580 	return (sizeof(struct proc));
2581 }
2582 
2583 int
2584 sched_sizeof_thread(void)
2585 {
2586 	return (sizeof(struct thread) + sizeof(struct td_sched));
2587 }
2588 
2589 #ifdef SMP
2590 #define	TDQ_IDLESPIN(tdq)						\
2591     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2592 #else
2593 #define	TDQ_IDLESPIN(tdq)	1
2594 #endif
2595 
2596 /*
2597  * The actual idle process.
2598  */
2599 void
2600 sched_idletd(void *dummy)
2601 {
2602 	struct thread *td;
2603 	struct tdq *tdq;
2604 	int oldswitchcnt, switchcnt;
2605 	int i;
2606 
2607 	mtx_assert(&Giant, MA_NOTOWNED);
2608 	td = curthread;
2609 	tdq = TDQ_SELF();
2610 	THREAD_NO_SLEEPING();
2611 	oldswitchcnt = -1;
2612 	for (;;) {
2613 		if (tdq->tdq_load) {
2614 			thread_lock(td);
2615 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2616 			thread_unlock(td);
2617 		}
2618 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2619 #ifdef SMP
2620 		if (switchcnt != oldswitchcnt) {
2621 			oldswitchcnt = switchcnt;
2622 			if (tdq_idled(tdq) == 0)
2623 				continue;
2624 		}
2625 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2626 #else
2627 		oldswitchcnt = switchcnt;
2628 #endif
2629 		/*
2630 		 * If we're switching very frequently, spin while checking
2631 		 * for load rather than entering a low power state that
2632 		 * may require an IPI.  However, don't do any busy
2633 		 * loops while on SMT machines as this simply steals
2634 		 * cycles from cores doing useful work.
2635 		 */
2636 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2637 			for (i = 0; i < sched_idlespins; i++) {
2638 				if (tdq->tdq_load)
2639 					break;
2640 				cpu_spinwait();
2641 			}
2642 		}
2643 
2644 		/* If there was context switch during spin, restart it. */
2645 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2646 		if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2647 			continue;
2648 
2649 		/* Run main MD idle handler. */
2650 		tdq->tdq_cpu_idle = 1;
2651 		/*
2652 		 * Make sure that tdq_cpu_idle update is globally visible
2653 		 * before cpu_idle() read tdq_load.  The order is important
2654 		 * to avoid race with tdq_notify.
2655 		 */
2656 		mb();
2657 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2658 		tdq->tdq_cpu_idle = 0;
2659 
2660 		/*
2661 		 * Account thread-less hardware interrupts and
2662 		 * other wakeup reasons equal to context switches.
2663 		 */
2664 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2665 		if (switchcnt != oldswitchcnt)
2666 			continue;
2667 		tdq->tdq_switchcnt++;
2668 		oldswitchcnt++;
2669 	}
2670 }
2671 
2672 /*
2673  * A CPU is entering for the first time or a thread is exiting.
2674  */
2675 void
2676 sched_throw(struct thread *td)
2677 {
2678 	struct thread *newtd;
2679 	struct tdq *tdq;
2680 
2681 	tdq = TDQ_SELF();
2682 	if (td == NULL) {
2683 		/* Correct spinlock nesting and acquire the correct lock. */
2684 		TDQ_LOCK(tdq);
2685 		spinlock_exit();
2686 		PCPU_SET(switchtime, cpu_ticks());
2687 		PCPU_SET(switchticks, ticks);
2688 	} else {
2689 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2690 		tdq_load_rem(tdq, td);
2691 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2692 	}
2693 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2694 	newtd = choosethread();
2695 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2696 	cpu_throw(td, newtd);		/* doesn't return */
2697 }
2698 
2699 /*
2700  * This is called from fork_exit().  Just acquire the correct locks and
2701  * let fork do the rest of the work.
2702  */
2703 void
2704 sched_fork_exit(struct thread *td)
2705 {
2706 	struct tdq *tdq;
2707 	int cpuid;
2708 
2709 	/*
2710 	 * Finish setting up thread glue so that it begins execution in a
2711 	 * non-nested critical section with the scheduler lock held.
2712 	 */
2713 	cpuid = PCPU_GET(cpuid);
2714 	tdq = TDQ_CPU(cpuid);
2715 	if (TD_IS_IDLETHREAD(td))
2716 		td->td_lock = TDQ_LOCKPTR(tdq);
2717 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2718 	td->td_oncpu = cpuid;
2719 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2720 	lock_profile_obtain_lock_success(
2721 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2722 }
2723 
2724 /*
2725  * Create on first use to catch odd startup conditons.
2726  */
2727 char *
2728 sched_tdname(struct thread *td)
2729 {
2730 #ifdef KTR
2731 	struct td_sched *ts;
2732 
2733 	ts = td->td_sched;
2734 	if (ts->ts_name[0] == '\0')
2735 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2736 		    "%s tid %d", td->td_name, td->td_tid);
2737 	return (ts->ts_name);
2738 #else
2739 	return (td->td_name);
2740 #endif
2741 }
2742 
2743 #ifdef KTR
2744 void
2745 sched_clear_tdname(struct thread *td)
2746 {
2747 	struct td_sched *ts;
2748 
2749 	ts = td->td_sched;
2750 	ts->ts_name[0] = '\0';
2751 }
2752 #endif
2753 
2754 #ifdef SMP
2755 
2756 /*
2757  * Build the CPU topology dump string. Is recursively called to collect
2758  * the topology tree.
2759  */
2760 static int
2761 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2762     int indent)
2763 {
2764 	char cpusetbuf[CPUSETBUFSIZ];
2765 	int i, first;
2766 
2767 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2768 	    "", 1 + indent / 2, cg->cg_level);
2769 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2770 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2771 	first = TRUE;
2772 	for (i = 0; i < MAXCPU; i++) {
2773 		if (CPU_ISSET(i, &cg->cg_mask)) {
2774 			if (!first)
2775 				sbuf_printf(sb, ", ");
2776 			else
2777 				first = FALSE;
2778 			sbuf_printf(sb, "%d", i);
2779 		}
2780 	}
2781 	sbuf_printf(sb, "</cpu>\n");
2782 
2783 	if (cg->cg_flags != 0) {
2784 		sbuf_printf(sb, "%*s <flags>", indent, "");
2785 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2786 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2787 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2788 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2789 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2790 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2791 		sbuf_printf(sb, "</flags>\n");
2792 	}
2793 
2794 	if (cg->cg_children > 0) {
2795 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2796 		for (i = 0; i < cg->cg_children; i++)
2797 			sysctl_kern_sched_topology_spec_internal(sb,
2798 			    &cg->cg_child[i], indent+2);
2799 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2800 	}
2801 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2802 	return (0);
2803 }
2804 
2805 /*
2806  * Sysctl handler for retrieving topology dump. It's a wrapper for
2807  * the recursive sysctl_kern_smp_topology_spec_internal().
2808  */
2809 static int
2810 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2811 {
2812 	struct sbuf *topo;
2813 	int err;
2814 
2815 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2816 
2817 	topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2818 	if (topo == NULL)
2819 		return (ENOMEM);
2820 
2821 	sbuf_printf(topo, "<groups>\n");
2822 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2823 	sbuf_printf(topo, "</groups>\n");
2824 
2825 	if (err == 0) {
2826 		sbuf_finish(topo);
2827 		err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2828 	}
2829 	sbuf_delete(topo);
2830 	return (err);
2831 }
2832 
2833 #endif
2834 
2835 static int
2836 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2837 {
2838 	int error, new_val, period;
2839 
2840 	period = 1000000 / realstathz;
2841 	new_val = period * sched_slice;
2842 	error = sysctl_handle_int(oidp, &new_val, 0, req);
2843 	if (error != 0 || req->newptr == NULL)
2844 		return (error);
2845 	if (new_val <= 0)
2846 		return (EINVAL);
2847 	sched_slice = imax(1, (new_val + period / 2) / period);
2848 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
2849 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2850 	    realstathz);
2851 	return (0);
2852 }
2853 
2854 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2855 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2856     "Scheduler name");
2857 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2858     NULL, 0, sysctl_kern_quantum, "I",
2859     "Quantum for timeshare threads in microseconds");
2860 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2861     "Quantum for timeshare threads in stathz ticks");
2862 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2863     "Interactivity score threshold");
2864 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2865     &preempt_thresh, 0,
2866     "Maximal (lowest) priority for preemption");
2867 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2868     "Assign static kernel priorities to sleeping threads");
2869 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2870     "Number of times idle thread will spin waiting for new work");
2871 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2872     &sched_idlespinthresh, 0,
2873     "Threshold before we will permit idle thread spinning");
2874 #ifdef SMP
2875 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2876     "Number of hz ticks to keep thread affinity for");
2877 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2878     "Enables the long-term load balancer");
2879 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2880     &balance_interval, 0,
2881     "Average period in stathz ticks to run the long-term balancer");
2882 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2883     "Attempts to steal work from other cores before idling");
2884 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2885     "Minimum load on remote CPU before we'll steal");
2886 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2887     CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2888     "XML dump of detected CPU topology");
2889 #endif
2890 
2891 /* ps compat.  All cpu percentages from ULE are weighted. */
2892 static int ccpu = 0;
2893 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2894