xref: /freebsd/sys/kern/sched_ule.c (revision 137a344c6341d1469432e9deb3a25593f96672ad)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice unmodified, this list of conditions, and the following
12  *    disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 /*
30  * This file implements the ULE scheduler.  ULE supports independent CPU
31  * run queues and fine grain locking.  It has superior interactive
32  * performance under load even on uni-processor systems.
33  *
34  * etymology:
35  *   ULE is the last three letters in schedule.  It owes its name to a
36  * generic user created for a scheduling system by Paul Mikesell at
37  * Isilon Systems and a general lack of creativity on the part of the author.
38  */
39 
40 #include <sys/cdefs.h>
41 __FBSDID("$FreeBSD$");
42 
43 #include "opt_hwpmc_hooks.h"
44 #include "opt_sched.h"
45 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kdb.h>
49 #include <sys/kernel.h>
50 #include <sys/ktr.h>
51 #include <sys/limits.h>
52 #include <sys/lock.h>
53 #include <sys/mutex.h>
54 #include <sys/proc.h>
55 #include <sys/resource.h>
56 #include <sys/resourcevar.h>
57 #include <sys/sched.h>
58 #include <sys/sdt.h>
59 #include <sys/smp.h>
60 #include <sys/sx.h>
61 #include <sys/sysctl.h>
62 #include <sys/sysproto.h>
63 #include <sys/turnstile.h>
64 #include <sys/umtx.h>
65 #include <sys/vmmeter.h>
66 #include <sys/cpuset.h>
67 #include <sys/sbuf.h>
68 
69 #ifdef HWPMC_HOOKS
70 #include <sys/pmckern.h>
71 #endif
72 
73 #ifdef KDTRACE_HOOKS
74 #include <sys/dtrace_bsd.h>
75 int				dtrace_vtime_active;
76 dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
77 #endif
78 
79 #include <machine/cpu.h>
80 #include <machine/smp.h>
81 
82 #define	KTR_ULE	0
83 
84 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
85 #define	TDQ_NAME_LEN	(sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
86 #define	TDQ_LOADNAME_LEN	(sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
87 
88 /*
89  * Thread scheduler specific section.  All fields are protected
90  * by the thread lock.
91  */
92 struct td_sched {
93 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
94 	short		ts_flags;	/* TSF_* flags. */
95 	int		ts_cpu;		/* CPU that we have affinity for. */
96 	int		ts_rltick;	/* Real last tick, for affinity. */
97 	int		ts_slice;	/* Ticks of slice remaining. */
98 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
99 	u_int		ts_runtime;	/* Number of ticks we were running */
100 	int		ts_ltick;	/* Last tick that we were running on */
101 	int		ts_ftick;	/* First tick that we were running on */
102 	int		ts_ticks;	/* Tick count */
103 #ifdef KTR
104 	char		ts_name[TS_NAME_LEN];
105 #endif
106 };
107 /* flags kept in ts_flags */
108 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
109 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
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 _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <=
116     sizeof(struct thread0_storage),
117     "increase struct thread0_storage.t0st_sched size");
118 
119 /*
120  * Priority ranges used for interactive and non-interactive timeshare
121  * threads.  The timeshare priorities are split up into four ranges.
122  * The first range handles interactive threads.  The last three ranges
123  * (NHALF, x, and NHALF) handle non-interactive threads with the outer
124  * ranges supporting nice values.
125  */
126 #define	PRI_TIMESHARE_RANGE	(PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
127 #define	PRI_INTERACT_RANGE	((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
128 #define	PRI_BATCH_RANGE		(PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
129 
130 #define	PRI_MIN_INTERACT	PRI_MIN_TIMESHARE
131 #define	PRI_MAX_INTERACT	(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
132 #define	PRI_MIN_BATCH		(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
133 #define	PRI_MAX_BATCH		PRI_MAX_TIMESHARE
134 
135 /*
136  * Cpu percentage computation macros and defines.
137  *
138  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
139  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
140  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
141  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
142  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
143  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
144  */
145 #define	SCHED_TICK_SECS		10
146 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
147 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
148 #define	SCHED_TICK_SHIFT	10
149 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
150 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
151 
152 /*
153  * These macros determine priorities for non-interactive threads.  They are
154  * assigned a priority based on their recent cpu utilization as expressed
155  * by the ratio of ticks to the tick total.  NHALF priorities at the start
156  * and end of the MIN to MAX timeshare range are only reachable with negative
157  * or positive nice respectively.
158  *
159  * PRI_RANGE:	Priority range for utilization dependent priorities.
160  * PRI_NRESV:	Number of nice values.
161  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
162  * PRI_NICE:	Determines the part of the priority inherited from nice.
163  */
164 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
165 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
166 #define	SCHED_PRI_MIN		(PRI_MIN_BATCH + SCHED_PRI_NHALF)
167 #define	SCHED_PRI_MAX		(PRI_MAX_BATCH - SCHED_PRI_NHALF)
168 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
169 #define	SCHED_PRI_TICKS(ts)						\
170     (SCHED_TICK_HZ((ts)) /						\
171     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
172 #define	SCHED_PRI_NICE(nice)	(nice)
173 
174 /*
175  * These determine the interactivity of a process.  Interactivity differs from
176  * cpu utilization in that it expresses the voluntary time slept vs time ran
177  * while cpu utilization includes all time not running.  This more accurately
178  * models the intent of the thread.
179  *
180  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
181  *		before throttling back.
182  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
183  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
184  * INTERACT_THRESH:	Threshold for placement on the current runq.
185  */
186 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
187 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
188 #define	SCHED_INTERACT_MAX	(100)
189 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
190 #define	SCHED_INTERACT_THRESH	(30)
191 
192 /*
193  * These parameters determine the slice behavior for batch work.
194  */
195 #define	SCHED_SLICE_DEFAULT_DIVISOR	10	/* ~94 ms, 12 stathz ticks. */
196 #define	SCHED_SLICE_MIN_DIVISOR		6	/* DEFAULT/MIN = ~16 ms. */
197 
198 /* Flags kept in td_flags. */
199 #define	TDF_SLICEEND	TDF_SCHED2	/* Thread time slice is over. */
200 
201 /*
202  * tickincr:		Converts a stathz tick into a hz domain scaled by
203  *			the shift factor.  Without the shift the error rate
204  *			due to rounding would be unacceptably high.
205  * realstathz:		stathz is sometimes 0 and run off of hz.
206  * sched_slice:		Runtime of each thread before rescheduling.
207  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
208  */
209 static int sched_interact = SCHED_INTERACT_THRESH;
210 static int tickincr = 8 << SCHED_TICK_SHIFT;
211 static int realstathz = 127;	/* reset during boot. */
212 static int sched_slice = 10;	/* reset during boot. */
213 static int sched_slice_min = 1;	/* reset during boot. */
214 #ifdef PREEMPTION
215 #ifdef FULL_PREEMPTION
216 static int preempt_thresh = PRI_MAX_IDLE;
217 #else
218 static int preempt_thresh = PRI_MIN_KERN;
219 #endif
220 #else
221 static int preempt_thresh = 0;
222 #endif
223 static int static_boost = PRI_MIN_BATCH;
224 static int sched_idlespins = 10000;
225 static int sched_idlespinthresh = -1;
226 
227 /*
228  * tdq - per processor runqs and statistics.  All fields are protected by the
229  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
230  * locking in sched_pickcpu();
231  */
232 struct tdq {
233 	/*
234 	 * Ordered to improve efficiency of cpu_search() and switch().
235 	 * tdq_lock is padded to avoid false sharing with tdq_load and
236 	 * tdq_cpu_idle.
237 	 */
238 	struct mtx_padalign tdq_lock;		/* run queue lock. */
239 	struct cpu_group *tdq_cg;		/* Pointer to cpu topology. */
240 	volatile int	tdq_load;		/* Aggregate load. */
241 	volatile int	tdq_cpu_idle;		/* cpu_idle() is active. */
242 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
243 	int		tdq_transferable;	/* Transferable thread count. */
244 	short		tdq_switchcnt;		/* Switches this tick. */
245 	short		tdq_oldswitchcnt;	/* Switches last tick. */
246 	u_char		tdq_lowpri;		/* Lowest priority thread. */
247 	u_char		tdq_ipipending;		/* IPI pending. */
248 	u_char		tdq_idx;		/* Current insert index. */
249 	u_char		tdq_ridx;		/* Current removal index. */
250 	struct runq	tdq_realtime;		/* real-time run queue. */
251 	struct runq	tdq_timeshare;		/* timeshare run queue. */
252 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
253 	char		tdq_name[TDQ_NAME_LEN];
254 #ifdef KTR
255 	char		tdq_loadname[TDQ_LOADNAME_LEN];
256 #endif
257 } __aligned(64);
258 
259 /* Idle thread states and config. */
260 #define	TDQ_RUNNING	1
261 #define	TDQ_IDLE	2
262 
263 #ifdef SMP
264 struct cpu_group *cpu_top;		/* CPU topology */
265 
266 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
267 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
268 
269 /*
270  * Run-time tunables.
271  */
272 static int rebalance = 1;
273 static int balance_interval = 128;	/* Default set in sched_initticks(). */
274 static int affinity;
275 static int steal_idle = 1;
276 static int steal_thresh = 2;
277 
278 /*
279  * One thread queue per processor.
280  */
281 static struct tdq	tdq_cpu[MAXCPU];
282 static struct tdq	*balance_tdq;
283 static int balance_ticks;
284 static DPCPU_DEFINE(uint32_t, randomval);
285 
286 #define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
287 #define	TDQ_CPU(x)	(&tdq_cpu[(x)])
288 #define	TDQ_ID(x)	((int)((x) - tdq_cpu))
289 #else	/* !SMP */
290 static struct tdq	tdq_cpu;
291 
292 #define	TDQ_ID(x)	(0)
293 #define	TDQ_SELF()	(&tdq_cpu)
294 #define	TDQ_CPU(x)	(&tdq_cpu)
295 #endif
296 
297 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
298 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
299 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
300 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
301 #define	TDQ_LOCKPTR(t)		((struct mtx *)(&(t)->tdq_lock))
302 
303 static void sched_priority(struct thread *);
304 static void sched_thread_priority(struct thread *, u_char);
305 static int sched_interact_score(struct thread *);
306 static void sched_interact_update(struct thread *);
307 static void sched_interact_fork(struct thread *);
308 static void sched_pctcpu_update(struct td_sched *, int);
309 
310 /* Operations on per processor queues */
311 static struct thread *tdq_choose(struct tdq *);
312 static void tdq_setup(struct tdq *);
313 static void tdq_load_add(struct tdq *, struct thread *);
314 static void tdq_load_rem(struct tdq *, struct thread *);
315 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
316 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
317 static inline int sched_shouldpreempt(int, int, int);
318 void tdq_print(int cpu);
319 static void runq_print(struct runq *rq);
320 static void tdq_add(struct tdq *, struct thread *, int);
321 #ifdef SMP
322 static int tdq_move(struct tdq *, struct tdq *);
323 static int tdq_idled(struct tdq *);
324 static void tdq_notify(struct tdq *, struct thread *);
325 static struct thread *tdq_steal(struct tdq *, int);
326 static struct thread *runq_steal(struct runq *, int);
327 static int sched_pickcpu(struct thread *, int);
328 static void sched_balance(void);
329 static int sched_balance_pair(struct tdq *, struct tdq *);
330 static inline struct tdq *sched_setcpu(struct thread *, int, int);
331 static inline void thread_unblock_switch(struct thread *, struct mtx *);
332 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
333 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
334 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
335     struct cpu_group *cg, int indent);
336 #endif
337 
338 static void sched_setup(void *dummy);
339 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
340 
341 static void sched_initticks(void *dummy);
342 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
343     NULL);
344 
345 SDT_PROVIDER_DEFINE(sched);
346 
347 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
348     "struct proc *", "uint8_t");
349 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
350     "struct proc *", "void *");
351 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
352     "struct proc *", "void *", "int");
353 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
354     "struct proc *", "uint8_t", "struct thread *");
355 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
356 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
357     "struct proc *");
358 SDT_PROBE_DEFINE(sched, , , on__cpu);
359 SDT_PROBE_DEFINE(sched, , , remain__cpu);
360 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
361     "struct proc *");
362 
363 /*
364  * Print the threads waiting on a run-queue.
365  */
366 static void
367 runq_print(struct runq *rq)
368 {
369 	struct rqhead *rqh;
370 	struct thread *td;
371 	int pri;
372 	int j;
373 	int i;
374 
375 	for (i = 0; i < RQB_LEN; i++) {
376 		printf("\t\trunq bits %d 0x%zx\n",
377 		    i, rq->rq_status.rqb_bits[i]);
378 		for (j = 0; j < RQB_BPW; j++)
379 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
380 				pri = j + (i << RQB_L2BPW);
381 				rqh = &rq->rq_queues[pri];
382 				TAILQ_FOREACH(td, rqh, td_runq) {
383 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
384 					    td, td->td_name, td->td_priority,
385 					    td->td_rqindex, pri);
386 				}
387 			}
388 	}
389 }
390 
391 /*
392  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
393  */
394 void
395 tdq_print(int cpu)
396 {
397 	struct tdq *tdq;
398 
399 	tdq = TDQ_CPU(cpu);
400 
401 	printf("tdq %d:\n", TDQ_ID(tdq));
402 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
403 	printf("\tLock name:      %s\n", tdq->tdq_name);
404 	printf("\tload:           %d\n", tdq->tdq_load);
405 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
406 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
407 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
408 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
409 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
410 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
411 	printf("\trealtime runq:\n");
412 	runq_print(&tdq->tdq_realtime);
413 	printf("\ttimeshare runq:\n");
414 	runq_print(&tdq->tdq_timeshare);
415 	printf("\tidle runq:\n");
416 	runq_print(&tdq->tdq_idle);
417 }
418 
419 static inline int
420 sched_shouldpreempt(int pri, int cpri, int remote)
421 {
422 	/*
423 	 * If the new priority is not better than the current priority there is
424 	 * nothing to do.
425 	 */
426 	if (pri >= cpri)
427 		return (0);
428 	/*
429 	 * Always preempt idle.
430 	 */
431 	if (cpri >= PRI_MIN_IDLE)
432 		return (1);
433 	/*
434 	 * If preemption is disabled don't preempt others.
435 	 */
436 	if (preempt_thresh == 0)
437 		return (0);
438 	/*
439 	 * Preempt if we exceed the threshold.
440 	 */
441 	if (pri <= preempt_thresh)
442 		return (1);
443 	/*
444 	 * If we're interactive or better and there is non-interactive
445 	 * or worse running preempt only remote processors.
446 	 */
447 	if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
448 		return (1);
449 	return (0);
450 }
451 
452 /*
453  * Add a thread to the actual run-queue.  Keeps transferable counts up to
454  * date with what is actually on the run-queue.  Selects the correct
455  * queue position for timeshare threads.
456  */
457 static __inline void
458 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
459 {
460 	struct td_sched *ts;
461 	u_char pri;
462 
463 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
464 	THREAD_LOCK_ASSERT(td, MA_OWNED);
465 
466 	pri = td->td_priority;
467 	ts = td_get_sched(td);
468 	TD_SET_RUNQ(td);
469 	if (THREAD_CAN_MIGRATE(td)) {
470 		tdq->tdq_transferable++;
471 		ts->ts_flags |= TSF_XFERABLE;
472 	}
473 	if (pri < PRI_MIN_BATCH) {
474 		ts->ts_runq = &tdq->tdq_realtime;
475 	} else if (pri <= PRI_MAX_BATCH) {
476 		ts->ts_runq = &tdq->tdq_timeshare;
477 		KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
478 			("Invalid priority %d on timeshare runq", pri));
479 		/*
480 		 * This queue contains only priorities between MIN and MAX
481 		 * realtime.  Use the whole queue to represent these values.
482 		 */
483 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
484 			pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
485 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
486 			/*
487 			 * This effectively shortens the queue by one so we
488 			 * can have a one slot difference between idx and
489 			 * ridx while we wait for threads to drain.
490 			 */
491 			if (tdq->tdq_ridx != tdq->tdq_idx &&
492 			    pri == tdq->tdq_ridx)
493 				pri = (unsigned char)(pri - 1) % RQ_NQS;
494 		} else
495 			pri = tdq->tdq_ridx;
496 		runq_add_pri(ts->ts_runq, td, pri, flags);
497 		return;
498 	} else
499 		ts->ts_runq = &tdq->tdq_idle;
500 	runq_add(ts->ts_runq, td, flags);
501 }
502 
503 /*
504  * Remove a thread from a run-queue.  This typically happens when a thread
505  * is selected to run.  Running threads are not on the queue and the
506  * transferable count does not reflect them.
507  */
508 static __inline void
509 tdq_runq_rem(struct tdq *tdq, struct thread *td)
510 {
511 	struct td_sched *ts;
512 
513 	ts = td_get_sched(td);
514 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
515 	KASSERT(ts->ts_runq != NULL,
516 	    ("tdq_runq_remove: thread %p null ts_runq", td));
517 	if (ts->ts_flags & TSF_XFERABLE) {
518 		tdq->tdq_transferable--;
519 		ts->ts_flags &= ~TSF_XFERABLE;
520 	}
521 	if (ts->ts_runq == &tdq->tdq_timeshare) {
522 		if (tdq->tdq_idx != tdq->tdq_ridx)
523 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
524 		else
525 			runq_remove_idx(ts->ts_runq, td, NULL);
526 	} else
527 		runq_remove(ts->ts_runq, td);
528 }
529 
530 /*
531  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
532  * for this thread to the referenced thread queue.
533  */
534 static void
535 tdq_load_add(struct tdq *tdq, struct thread *td)
536 {
537 
538 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
539 	THREAD_LOCK_ASSERT(td, MA_OWNED);
540 
541 	tdq->tdq_load++;
542 	if ((td->td_flags & TDF_NOLOAD) == 0)
543 		tdq->tdq_sysload++;
544 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
545 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
546 }
547 
548 /*
549  * Remove the load from a thread that is transitioning to a sleep state or
550  * exiting.
551  */
552 static void
553 tdq_load_rem(struct tdq *tdq, struct thread *td)
554 {
555 
556 	THREAD_LOCK_ASSERT(td, MA_OWNED);
557 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
558 	KASSERT(tdq->tdq_load != 0,
559 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
560 
561 	tdq->tdq_load--;
562 	if ((td->td_flags & TDF_NOLOAD) == 0)
563 		tdq->tdq_sysload--;
564 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
565 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
566 }
567 
568 /*
569  * Bound timeshare latency by decreasing slice size as load increases.  We
570  * consider the maximum latency as the sum of the threads waiting to run
571  * aside from curthread and target no more than sched_slice latency but
572  * no less than sched_slice_min runtime.
573  */
574 static inline int
575 tdq_slice(struct tdq *tdq)
576 {
577 	int load;
578 
579 	/*
580 	 * It is safe to use sys_load here because this is called from
581 	 * contexts where timeshare threads are running and so there
582 	 * cannot be higher priority load in the system.
583 	 */
584 	load = tdq->tdq_sysload - 1;
585 	if (load >= SCHED_SLICE_MIN_DIVISOR)
586 		return (sched_slice_min);
587 	if (load <= 1)
588 		return (sched_slice);
589 	return (sched_slice / load);
590 }
591 
592 /*
593  * Set lowpri to its exact value by searching the run-queue and
594  * evaluating curthread.  curthread may be passed as an optimization.
595  */
596 static void
597 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
598 {
599 	struct thread *td;
600 
601 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
602 	if (ctd == NULL)
603 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
604 	td = tdq_choose(tdq);
605 	if (td == NULL || td->td_priority > ctd->td_priority)
606 		tdq->tdq_lowpri = ctd->td_priority;
607 	else
608 		tdq->tdq_lowpri = td->td_priority;
609 }
610 
611 #ifdef SMP
612 /*
613  * We need some randomness. Implement a classic Linear Congruential
614  * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
615  * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
616  * of the random state (in the low bits of our answer) to keep
617  * the maximum randomness.
618  */
619 static uint32_t
620 sched_random(void)
621 {
622 	uint32_t *rndptr;
623 
624 	rndptr = DPCPU_PTR(randomval);
625 	*rndptr = *rndptr * 69069 + 5;
626 
627 	return (*rndptr >> 16);
628 }
629 
630 struct cpu_search {
631 	cpuset_t cs_mask;
632 	u_int	cs_prefer;
633 	int	cs_pri;		/* Min priority for low. */
634 	int	cs_limit;	/* Max load for low, min load for high. */
635 	int	cs_cpu;
636 	int	cs_load;
637 };
638 
639 #define	CPU_SEARCH_LOWEST	0x1
640 #define	CPU_SEARCH_HIGHEST	0x2
641 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
642 
643 #define	CPUSET_FOREACH(cpu, mask)				\
644 	for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++)		\
645 		if (CPU_ISSET(cpu, &mask))
646 
647 static __always_inline int cpu_search(const struct cpu_group *cg,
648     struct cpu_search *low, struct cpu_search *high, const int match);
649 int __noinline cpu_search_lowest(const struct cpu_group *cg,
650     struct cpu_search *low);
651 int __noinline cpu_search_highest(const struct cpu_group *cg,
652     struct cpu_search *high);
653 int __noinline cpu_search_both(const struct cpu_group *cg,
654     struct cpu_search *low, struct cpu_search *high);
655 
656 /*
657  * Search the tree of cpu_groups for the lowest or highest loaded cpu
658  * according to the match argument.  This routine actually compares the
659  * load on all paths through the tree and finds the least loaded cpu on
660  * the least loaded path, which may differ from the least loaded cpu in
661  * the system.  This balances work among caches and buses.
662  *
663  * This inline is instantiated in three forms below using constants for the
664  * match argument.  It is reduced to the minimum set for each case.  It is
665  * also recursive to the depth of the tree.
666  */
667 static __always_inline int
668 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
669     struct cpu_search *high, const int match)
670 {
671 	struct cpu_search lgroup;
672 	struct cpu_search hgroup;
673 	cpuset_t cpumask;
674 	struct cpu_group *child;
675 	struct tdq *tdq;
676 	int cpu, i, hload, lload, load, total, rnd;
677 
678 	total = 0;
679 	cpumask = cg->cg_mask;
680 	if (match & CPU_SEARCH_LOWEST) {
681 		lload = INT_MAX;
682 		lgroup = *low;
683 	}
684 	if (match & CPU_SEARCH_HIGHEST) {
685 		hload = INT_MIN;
686 		hgroup = *high;
687 	}
688 
689 	/* Iterate through the child CPU groups and then remaining CPUs. */
690 	for (i = cg->cg_children, cpu = mp_maxid; ; ) {
691 		if (i == 0) {
692 #ifdef HAVE_INLINE_FFSL
693 			cpu = CPU_FFS(&cpumask) - 1;
694 #else
695 			while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
696 				cpu--;
697 #endif
698 			if (cpu < 0)
699 				break;
700 			child = NULL;
701 		} else
702 			child = &cg->cg_child[i - 1];
703 
704 		if (match & CPU_SEARCH_LOWEST)
705 			lgroup.cs_cpu = -1;
706 		if (match & CPU_SEARCH_HIGHEST)
707 			hgroup.cs_cpu = -1;
708 		if (child) {			/* Handle child CPU group. */
709 			CPU_NAND(&cpumask, &child->cg_mask);
710 			switch (match) {
711 			case CPU_SEARCH_LOWEST:
712 				load = cpu_search_lowest(child, &lgroup);
713 				break;
714 			case CPU_SEARCH_HIGHEST:
715 				load = cpu_search_highest(child, &hgroup);
716 				break;
717 			case CPU_SEARCH_BOTH:
718 				load = cpu_search_both(child, &lgroup, &hgroup);
719 				break;
720 			}
721 		} else {			/* Handle child CPU. */
722 			CPU_CLR(cpu, &cpumask);
723 			tdq = TDQ_CPU(cpu);
724 			load = tdq->tdq_load * 256;
725 			rnd = sched_random() % 32;
726 			if (match & CPU_SEARCH_LOWEST) {
727 				if (cpu == low->cs_prefer)
728 					load -= 64;
729 				/* If that CPU is allowed and get data. */
730 				if (tdq->tdq_lowpri > lgroup.cs_pri &&
731 				    tdq->tdq_load <= lgroup.cs_limit &&
732 				    CPU_ISSET(cpu, &lgroup.cs_mask)) {
733 					lgroup.cs_cpu = cpu;
734 					lgroup.cs_load = load - rnd;
735 				}
736 			}
737 			if (match & CPU_SEARCH_HIGHEST)
738 				if (tdq->tdq_load >= hgroup.cs_limit &&
739 				    tdq->tdq_transferable &&
740 				    CPU_ISSET(cpu, &hgroup.cs_mask)) {
741 					hgroup.cs_cpu = cpu;
742 					hgroup.cs_load = load - rnd;
743 				}
744 		}
745 		total += load;
746 
747 		/* We have info about child item. Compare it. */
748 		if (match & CPU_SEARCH_LOWEST) {
749 			if (lgroup.cs_cpu >= 0 &&
750 			    (load < lload ||
751 			     (load == lload && lgroup.cs_load < low->cs_load))) {
752 				lload = load;
753 				low->cs_cpu = lgroup.cs_cpu;
754 				low->cs_load = lgroup.cs_load;
755 			}
756 		}
757 		if (match & CPU_SEARCH_HIGHEST)
758 			if (hgroup.cs_cpu >= 0 &&
759 			    (load > hload ||
760 			     (load == hload && hgroup.cs_load > high->cs_load))) {
761 				hload = load;
762 				high->cs_cpu = hgroup.cs_cpu;
763 				high->cs_load = hgroup.cs_load;
764 			}
765 		if (child) {
766 			i--;
767 			if (i == 0 && CPU_EMPTY(&cpumask))
768 				break;
769 		}
770 #ifndef HAVE_INLINE_FFSL
771 		else
772 			cpu--;
773 #endif
774 	}
775 	return (total);
776 }
777 
778 /*
779  * cpu_search instantiations must pass constants to maintain the inline
780  * optimization.
781  */
782 int
783 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
784 {
785 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
786 }
787 
788 int
789 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
790 {
791 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
792 }
793 
794 int
795 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
796     struct cpu_search *high)
797 {
798 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
799 }
800 
801 /*
802  * Find the cpu with the least load via the least loaded path that has a
803  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
804  * acceptable.
805  */
806 static inline int
807 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
808     int prefer)
809 {
810 	struct cpu_search low;
811 
812 	low.cs_cpu = -1;
813 	low.cs_prefer = prefer;
814 	low.cs_mask = mask;
815 	low.cs_pri = pri;
816 	low.cs_limit = maxload;
817 	cpu_search_lowest(cg, &low);
818 	return low.cs_cpu;
819 }
820 
821 /*
822  * Find the cpu with the highest load via the highest loaded path.
823  */
824 static inline int
825 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
826 {
827 	struct cpu_search high;
828 
829 	high.cs_cpu = -1;
830 	high.cs_mask = mask;
831 	high.cs_limit = minload;
832 	cpu_search_highest(cg, &high);
833 	return high.cs_cpu;
834 }
835 
836 static void
837 sched_balance_group(struct cpu_group *cg)
838 {
839 	cpuset_t hmask, lmask;
840 	int high, low, anylow;
841 
842 	CPU_FILL(&hmask);
843 	for (;;) {
844 		high = sched_highest(cg, hmask, 1);
845 		/* Stop if there is no more CPU with transferrable threads. */
846 		if (high == -1)
847 			break;
848 		CPU_CLR(high, &hmask);
849 		CPU_COPY(&hmask, &lmask);
850 		/* Stop if there is no more CPU left for low. */
851 		if (CPU_EMPTY(&lmask))
852 			break;
853 		anylow = 1;
854 nextlow:
855 		low = sched_lowest(cg, lmask, -1,
856 		    TDQ_CPU(high)->tdq_load - 1, high);
857 		/* Stop if we looked well and found no less loaded CPU. */
858 		if (anylow && low == -1)
859 			break;
860 		/* Go to next high if we found no less loaded CPU. */
861 		if (low == -1)
862 			continue;
863 		/* Transfer thread from high to low. */
864 		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
865 			/* CPU that got thread can no longer be a donor. */
866 			CPU_CLR(low, &hmask);
867 		} else {
868 			/*
869 			 * If failed, then there is no threads on high
870 			 * that can run on this low. Drop low from low
871 			 * mask and look for different one.
872 			 */
873 			CPU_CLR(low, &lmask);
874 			anylow = 0;
875 			goto nextlow;
876 		}
877 	}
878 }
879 
880 static void
881 sched_balance(void)
882 {
883 	struct tdq *tdq;
884 
885 	if (smp_started == 0 || rebalance == 0)
886 		return;
887 
888 	balance_ticks = max(balance_interval / 2, 1) +
889 	    (sched_random() % balance_interval);
890 	tdq = TDQ_SELF();
891 	TDQ_UNLOCK(tdq);
892 	sched_balance_group(cpu_top);
893 	TDQ_LOCK(tdq);
894 }
895 
896 /*
897  * Lock two thread queues using their address to maintain lock order.
898  */
899 static void
900 tdq_lock_pair(struct tdq *one, struct tdq *two)
901 {
902 	if (one < two) {
903 		TDQ_LOCK(one);
904 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
905 	} else {
906 		TDQ_LOCK(two);
907 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
908 	}
909 }
910 
911 /*
912  * Unlock two thread queues.  Order is not important here.
913  */
914 static void
915 tdq_unlock_pair(struct tdq *one, struct tdq *two)
916 {
917 	TDQ_UNLOCK(one);
918 	TDQ_UNLOCK(two);
919 }
920 
921 /*
922  * Transfer load between two imbalanced thread queues.
923  */
924 static int
925 sched_balance_pair(struct tdq *high, struct tdq *low)
926 {
927 	int moved;
928 	int cpu;
929 
930 	tdq_lock_pair(high, low);
931 	moved = 0;
932 	/*
933 	 * Determine what the imbalance is and then adjust that to how many
934 	 * threads we actually have to give up (transferable).
935 	 */
936 	if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
937 	    (moved = tdq_move(high, low)) > 0) {
938 		/*
939 		 * In case the target isn't the current cpu IPI it to force a
940 		 * reschedule with the new workload.
941 		 */
942 		cpu = TDQ_ID(low);
943 		if (cpu != PCPU_GET(cpuid))
944 			ipi_cpu(cpu, IPI_PREEMPT);
945 	}
946 	tdq_unlock_pair(high, low);
947 	return (moved);
948 }
949 
950 /*
951  * Move a thread from one thread queue to another.
952  */
953 static int
954 tdq_move(struct tdq *from, struct tdq *to)
955 {
956 	struct td_sched *ts;
957 	struct thread *td;
958 	struct tdq *tdq;
959 	int cpu;
960 
961 	TDQ_LOCK_ASSERT(from, MA_OWNED);
962 	TDQ_LOCK_ASSERT(to, MA_OWNED);
963 
964 	tdq = from;
965 	cpu = TDQ_ID(to);
966 	td = tdq_steal(tdq, cpu);
967 	if (td == NULL)
968 		return (0);
969 	ts = td_get_sched(td);
970 	/*
971 	 * Although the run queue is locked the thread may be blocked.  Lock
972 	 * it to clear this and acquire the run-queue lock.
973 	 */
974 	thread_lock(td);
975 	/* Drop recursive lock on from acquired via thread_lock(). */
976 	TDQ_UNLOCK(from);
977 	sched_rem(td);
978 	ts->ts_cpu = cpu;
979 	td->td_lock = TDQ_LOCKPTR(to);
980 	tdq_add(to, td, SRQ_YIELDING);
981 	return (1);
982 }
983 
984 /*
985  * This tdq has idled.  Try to steal a thread from another cpu and switch
986  * to it.
987  */
988 static int
989 tdq_idled(struct tdq *tdq)
990 {
991 	struct cpu_group *cg;
992 	struct tdq *steal;
993 	cpuset_t mask;
994 	int thresh;
995 	int cpu;
996 
997 	if (smp_started == 0 || steal_idle == 0)
998 		return (1);
999 	CPU_FILL(&mask);
1000 	CPU_CLR(PCPU_GET(cpuid), &mask);
1001 	/* We don't want to be preempted while we're iterating. */
1002 	spinlock_enter();
1003 	for (cg = tdq->tdq_cg; cg != NULL; ) {
1004 		if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
1005 			thresh = steal_thresh;
1006 		else
1007 			thresh = 1;
1008 		cpu = sched_highest(cg, mask, thresh);
1009 		if (cpu == -1) {
1010 			cg = cg->cg_parent;
1011 			continue;
1012 		}
1013 		steal = TDQ_CPU(cpu);
1014 		CPU_CLR(cpu, &mask);
1015 		tdq_lock_pair(tdq, steal);
1016 		if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
1017 			tdq_unlock_pair(tdq, steal);
1018 			continue;
1019 		}
1020 		/*
1021 		 * If a thread was added while interrupts were disabled don't
1022 		 * steal one here.  If we fail to acquire one due to affinity
1023 		 * restrictions loop again with this cpu removed from the
1024 		 * set.
1025 		 */
1026 		if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
1027 			tdq_unlock_pair(tdq, steal);
1028 			continue;
1029 		}
1030 		spinlock_exit();
1031 		TDQ_UNLOCK(steal);
1032 		mi_switch(SW_VOL | SWT_IDLE, NULL);
1033 		thread_unlock(curthread);
1034 
1035 		return (0);
1036 	}
1037 	spinlock_exit();
1038 	return (1);
1039 }
1040 
1041 /*
1042  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
1043  */
1044 static void
1045 tdq_notify(struct tdq *tdq, struct thread *td)
1046 {
1047 	struct thread *ctd;
1048 	int pri;
1049 	int cpu;
1050 
1051 	if (tdq->tdq_ipipending)
1052 		return;
1053 	cpu = td_get_sched(td)->ts_cpu;
1054 	pri = td->td_priority;
1055 	ctd = pcpu_find(cpu)->pc_curthread;
1056 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1057 		return;
1058 
1059 	/*
1060 	 * Make sure that our caller's earlier update to tdq_load is
1061 	 * globally visible before we read tdq_cpu_idle.  Idle thread
1062 	 * accesses both of them without locks, and the order is important.
1063 	 */
1064 	atomic_thread_fence_seq_cst();
1065 
1066 	if (TD_IS_IDLETHREAD(ctd)) {
1067 		/*
1068 		 * If the MD code has an idle wakeup routine try that before
1069 		 * falling back to IPI.
1070 		 */
1071 		if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1072 			return;
1073 	}
1074 	tdq->tdq_ipipending = 1;
1075 	ipi_cpu(cpu, IPI_PREEMPT);
1076 }
1077 
1078 /*
1079  * Steals load from a timeshare queue.  Honors the rotating queue head
1080  * index.
1081  */
1082 static struct thread *
1083 runq_steal_from(struct runq *rq, int cpu, u_char start)
1084 {
1085 	struct rqbits *rqb;
1086 	struct rqhead *rqh;
1087 	struct thread *td, *first;
1088 	int bit;
1089 	int i;
1090 
1091 	rqb = &rq->rq_status;
1092 	bit = start & (RQB_BPW -1);
1093 	first = NULL;
1094 again:
1095 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1096 		if (rqb->rqb_bits[i] == 0)
1097 			continue;
1098 		if (bit == 0)
1099 			bit = RQB_FFS(rqb->rqb_bits[i]);
1100 		for (; bit < RQB_BPW; bit++) {
1101 			if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1102 				continue;
1103 			rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1104 			TAILQ_FOREACH(td, rqh, td_runq) {
1105 				if (first && THREAD_CAN_MIGRATE(td) &&
1106 				    THREAD_CAN_SCHED(td, cpu))
1107 					return (td);
1108 				first = td;
1109 			}
1110 		}
1111 	}
1112 	if (start != 0) {
1113 		start = 0;
1114 		goto again;
1115 	}
1116 
1117 	if (first && THREAD_CAN_MIGRATE(first) &&
1118 	    THREAD_CAN_SCHED(first, cpu))
1119 		return (first);
1120 	return (NULL);
1121 }
1122 
1123 /*
1124  * Steals load from a standard linear queue.
1125  */
1126 static struct thread *
1127 runq_steal(struct runq *rq, int cpu)
1128 {
1129 	struct rqhead *rqh;
1130 	struct rqbits *rqb;
1131 	struct thread *td;
1132 	int word;
1133 	int bit;
1134 
1135 	rqb = &rq->rq_status;
1136 	for (word = 0; word < RQB_LEN; word++) {
1137 		if (rqb->rqb_bits[word] == 0)
1138 			continue;
1139 		for (bit = 0; bit < RQB_BPW; bit++) {
1140 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1141 				continue;
1142 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1143 			TAILQ_FOREACH(td, rqh, td_runq)
1144 				if (THREAD_CAN_MIGRATE(td) &&
1145 				    THREAD_CAN_SCHED(td, cpu))
1146 					return (td);
1147 		}
1148 	}
1149 	return (NULL);
1150 }
1151 
1152 /*
1153  * Attempt to steal a thread in priority order from a thread queue.
1154  */
1155 static struct thread *
1156 tdq_steal(struct tdq *tdq, int cpu)
1157 {
1158 	struct thread *td;
1159 
1160 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1161 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1162 		return (td);
1163 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1164 	    cpu, tdq->tdq_ridx)) != NULL)
1165 		return (td);
1166 	return (runq_steal(&tdq->tdq_idle, cpu));
1167 }
1168 
1169 /*
1170  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1171  * current lock and returns with the assigned queue locked.
1172  */
1173 static inline struct tdq *
1174 sched_setcpu(struct thread *td, int cpu, int flags)
1175 {
1176 
1177 	struct tdq *tdq;
1178 
1179 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1180 	tdq = TDQ_CPU(cpu);
1181 	td_get_sched(td)->ts_cpu = cpu;
1182 	/*
1183 	 * If the lock matches just return the queue.
1184 	 */
1185 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1186 		return (tdq);
1187 #ifdef notyet
1188 	/*
1189 	 * If the thread isn't running its lockptr is a
1190 	 * turnstile or a sleepqueue.  We can just lock_set without
1191 	 * blocking.
1192 	 */
1193 	if (TD_CAN_RUN(td)) {
1194 		TDQ_LOCK(tdq);
1195 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1196 		return (tdq);
1197 	}
1198 #endif
1199 	/*
1200 	 * The hard case, migration, we need to block the thread first to
1201 	 * prevent order reversals with other cpus locks.
1202 	 */
1203 	spinlock_enter();
1204 	thread_lock_block(td);
1205 	TDQ_LOCK(tdq);
1206 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1207 	spinlock_exit();
1208 	return (tdq);
1209 }
1210 
1211 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1212 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1213 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1214 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1215 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1216 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1217 
1218 static int
1219 sched_pickcpu(struct thread *td, int flags)
1220 {
1221 	struct cpu_group *cg, *ccg;
1222 	struct td_sched *ts;
1223 	struct tdq *tdq;
1224 	cpuset_t mask;
1225 	int cpu, pri, self;
1226 
1227 	self = PCPU_GET(cpuid);
1228 	ts = td_get_sched(td);
1229 	KASSERT(!CPU_ABSENT(ts->ts_cpu), ("sched_pickcpu: Start scheduler on "
1230 	    "absent CPU %d for thread %s.", ts->ts_cpu, td->td_name));
1231 	if (smp_started == 0)
1232 		return (self);
1233 	/*
1234 	 * Don't migrate a running thread from sched_switch().
1235 	 */
1236 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1237 		return (ts->ts_cpu);
1238 	/*
1239 	 * Prefer to run interrupt threads on the processors that generate
1240 	 * the interrupt.
1241 	 */
1242 	pri = td->td_priority;
1243 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1244 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1245 		SCHED_STAT_INC(pickcpu_intrbind);
1246 		ts->ts_cpu = self;
1247 		if (TDQ_CPU(self)->tdq_lowpri > pri) {
1248 			SCHED_STAT_INC(pickcpu_affinity);
1249 			return (ts->ts_cpu);
1250 		}
1251 	}
1252 	/*
1253 	 * If the thread can run on the last cpu and the affinity has not
1254 	 * expired or it is idle run it there.
1255 	 */
1256 	tdq = TDQ_CPU(ts->ts_cpu);
1257 	cg = tdq->tdq_cg;
1258 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1259 	    tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1260 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1261 		if (cg->cg_flags & CG_FLAG_THREAD) {
1262 			CPUSET_FOREACH(cpu, cg->cg_mask) {
1263 				if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1264 					break;
1265 			}
1266 		} else
1267 			cpu = INT_MAX;
1268 		if (cpu > mp_maxid) {
1269 			SCHED_STAT_INC(pickcpu_idle_affinity);
1270 			return (ts->ts_cpu);
1271 		}
1272 	}
1273 	/*
1274 	 * Search for the last level cache CPU group in the tree.
1275 	 * Skip caches with expired affinity time and SMT groups.
1276 	 * Affinity to higher level caches will be handled less aggressively.
1277 	 */
1278 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1279 		if (cg->cg_flags & CG_FLAG_THREAD)
1280 			continue;
1281 		if (!SCHED_AFFINITY(ts, cg->cg_level))
1282 			continue;
1283 		ccg = cg;
1284 	}
1285 	if (ccg != NULL)
1286 		cg = ccg;
1287 	cpu = -1;
1288 	/* Search the group for the less loaded idle CPU we can run now. */
1289 	mask = td->td_cpuset->cs_mask;
1290 	if (cg != NULL && cg != cpu_top &&
1291 	    CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1292 		cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1293 		    INT_MAX, ts->ts_cpu);
1294 	/* Search globally for the less loaded CPU we can run now. */
1295 	if (cpu == -1)
1296 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1297 	/* Search globally for the less loaded CPU. */
1298 	if (cpu == -1)
1299 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1300 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1301 	KASSERT(!CPU_ABSENT(cpu), ("sched_pickcpu: Picked absent CPU %d.", cpu));
1302 	/*
1303 	 * Compare the lowest loaded cpu to current cpu.
1304 	 */
1305 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1306 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1307 	    TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1308 		SCHED_STAT_INC(pickcpu_local);
1309 		cpu = self;
1310 	} else
1311 		SCHED_STAT_INC(pickcpu_lowest);
1312 	if (cpu != ts->ts_cpu)
1313 		SCHED_STAT_INC(pickcpu_migration);
1314 	return (cpu);
1315 }
1316 #endif
1317 
1318 /*
1319  * Pick the highest priority task we have and return it.
1320  */
1321 static struct thread *
1322 tdq_choose(struct tdq *tdq)
1323 {
1324 	struct thread *td;
1325 
1326 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1327 	td = runq_choose(&tdq->tdq_realtime);
1328 	if (td != NULL)
1329 		return (td);
1330 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1331 	if (td != NULL) {
1332 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1333 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1334 		    td->td_priority));
1335 		return (td);
1336 	}
1337 	td = runq_choose(&tdq->tdq_idle);
1338 	if (td != NULL) {
1339 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1340 		    ("tdq_choose: Invalid priority on idle queue %d",
1341 		    td->td_priority));
1342 		return (td);
1343 	}
1344 
1345 	return (NULL);
1346 }
1347 
1348 /*
1349  * Initialize a thread queue.
1350  */
1351 static void
1352 tdq_setup(struct tdq *tdq)
1353 {
1354 
1355 	if (bootverbose)
1356 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1357 	runq_init(&tdq->tdq_realtime);
1358 	runq_init(&tdq->tdq_timeshare);
1359 	runq_init(&tdq->tdq_idle);
1360 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1361 	    "sched lock %d", (int)TDQ_ID(tdq));
1362 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1363 	    MTX_SPIN | MTX_RECURSE);
1364 #ifdef KTR
1365 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1366 	    "CPU %d load", (int)TDQ_ID(tdq));
1367 #endif
1368 }
1369 
1370 #ifdef SMP
1371 static void
1372 sched_setup_smp(void)
1373 {
1374 	struct tdq *tdq;
1375 	int i;
1376 
1377 	cpu_top = smp_topo();
1378 	CPU_FOREACH(i) {
1379 		tdq = TDQ_CPU(i);
1380 		tdq_setup(tdq);
1381 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1382 		if (tdq->tdq_cg == NULL)
1383 			panic("Can't find cpu group for %d\n", i);
1384 	}
1385 	balance_tdq = TDQ_SELF();
1386 	sched_balance();
1387 }
1388 #endif
1389 
1390 /*
1391  * Setup the thread queues and initialize the topology based on MD
1392  * information.
1393  */
1394 static void
1395 sched_setup(void *dummy)
1396 {
1397 	struct tdq *tdq;
1398 
1399 	tdq = TDQ_SELF();
1400 #ifdef SMP
1401 	sched_setup_smp();
1402 #else
1403 	tdq_setup(tdq);
1404 #endif
1405 
1406 	/* Add thread0's load since it's running. */
1407 	TDQ_LOCK(tdq);
1408 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1409 	tdq_load_add(tdq, &thread0);
1410 	tdq->tdq_lowpri = thread0.td_priority;
1411 	TDQ_UNLOCK(tdq);
1412 }
1413 
1414 /*
1415  * This routine determines time constants after stathz and hz are setup.
1416  */
1417 /* ARGSUSED */
1418 static void
1419 sched_initticks(void *dummy)
1420 {
1421 	int incr;
1422 
1423 	realstathz = stathz ? stathz : hz;
1424 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1425 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1426 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1427 	    realstathz);
1428 
1429 	/*
1430 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1431 	 * hz not being evenly divisible by stathz on all platforms.
1432 	 */
1433 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1434 	/*
1435 	 * This does not work for values of stathz that are more than
1436 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1437 	 */
1438 	if (incr == 0)
1439 		incr = 1;
1440 	tickincr = incr;
1441 #ifdef SMP
1442 	/*
1443 	 * Set the default balance interval now that we know
1444 	 * what realstathz is.
1445 	 */
1446 	balance_interval = realstathz;
1447 	affinity = SCHED_AFFINITY_DEFAULT;
1448 #endif
1449 	if (sched_idlespinthresh < 0)
1450 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1451 }
1452 
1453 
1454 /*
1455  * This is the core of the interactivity algorithm.  Determines a score based
1456  * on past behavior.  It is the ratio of sleep time to run time scaled to
1457  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1458  * differs from the cpu usage because it does not account for time spent
1459  * waiting on a run-queue.  Would be prettier if we had floating point.
1460  *
1461  * When a thread's sleep time is greater than its run time the
1462  * calculation is:
1463  *
1464  *                           scaling factor
1465  * interactivity score =  ---------------------
1466  *                        sleep time / run time
1467  *
1468  *
1469  * When a thread's run time is greater than its sleep time the
1470  * calculation is:
1471  *
1472  *                           scaling factor
1473  * interactivity score =  ---------------------    + scaling factor
1474  *                        run time / sleep time
1475  */
1476 static int
1477 sched_interact_score(struct thread *td)
1478 {
1479 	struct td_sched *ts;
1480 	int div;
1481 
1482 	ts = td_get_sched(td);
1483 	/*
1484 	 * The score is only needed if this is likely to be an interactive
1485 	 * task.  Don't go through the expense of computing it if there's
1486 	 * no chance.
1487 	 */
1488 	if (sched_interact <= SCHED_INTERACT_HALF &&
1489 		ts->ts_runtime >= ts->ts_slptime)
1490 			return (SCHED_INTERACT_HALF);
1491 
1492 	if (ts->ts_runtime > ts->ts_slptime) {
1493 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1494 		return (SCHED_INTERACT_HALF +
1495 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1496 	}
1497 	if (ts->ts_slptime > ts->ts_runtime) {
1498 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1499 		return (ts->ts_runtime / div);
1500 	}
1501 	/* runtime == slptime */
1502 	if (ts->ts_runtime)
1503 		return (SCHED_INTERACT_HALF);
1504 
1505 	/*
1506 	 * This can happen if slptime and runtime are 0.
1507 	 */
1508 	return (0);
1509 
1510 }
1511 
1512 /*
1513  * Scale the scheduling priority according to the "interactivity" of this
1514  * process.
1515  */
1516 static void
1517 sched_priority(struct thread *td)
1518 {
1519 	int score;
1520 	int pri;
1521 
1522 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1523 		return;
1524 	/*
1525 	 * If the score is interactive we place the thread in the realtime
1526 	 * queue with a priority that is less than kernel and interrupt
1527 	 * priorities.  These threads are not subject to nice restrictions.
1528 	 *
1529 	 * Scores greater than this are placed on the normal timeshare queue
1530 	 * where the priority is partially decided by the most recent cpu
1531 	 * utilization and the rest is decided by nice value.
1532 	 *
1533 	 * The nice value of the process has a linear effect on the calculated
1534 	 * score.  Negative nice values make it easier for a thread to be
1535 	 * considered interactive.
1536 	 */
1537 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1538 	if (score < sched_interact) {
1539 		pri = PRI_MIN_INTERACT;
1540 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1541 		    sched_interact) * score;
1542 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1543 		    ("sched_priority: invalid interactive priority %d score %d",
1544 		    pri, score));
1545 	} else {
1546 		pri = SCHED_PRI_MIN;
1547 		if (td_get_sched(td)->ts_ticks)
1548 			pri += min(SCHED_PRI_TICKS(td_get_sched(td)),
1549 			    SCHED_PRI_RANGE - 1);
1550 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1551 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1552 		    ("sched_priority: invalid priority %d: nice %d, "
1553 		    "ticks %d ftick %d ltick %d tick pri %d",
1554 		    pri, td->td_proc->p_nice, td_get_sched(td)->ts_ticks,
1555 		    td_get_sched(td)->ts_ftick, td_get_sched(td)->ts_ltick,
1556 		    SCHED_PRI_TICKS(td_get_sched(td))));
1557 	}
1558 	sched_user_prio(td, pri);
1559 
1560 	return;
1561 }
1562 
1563 /*
1564  * This routine enforces a maximum limit on the amount of scheduling history
1565  * kept.  It is called after either the slptime or runtime is adjusted.  This
1566  * function is ugly due to integer math.
1567  */
1568 static void
1569 sched_interact_update(struct thread *td)
1570 {
1571 	struct td_sched *ts;
1572 	u_int sum;
1573 
1574 	ts = td_get_sched(td);
1575 	sum = ts->ts_runtime + ts->ts_slptime;
1576 	if (sum < SCHED_SLP_RUN_MAX)
1577 		return;
1578 	/*
1579 	 * This only happens from two places:
1580 	 * 1) We have added an unusual amount of run time from fork_exit.
1581 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1582 	 */
1583 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1584 		if (ts->ts_runtime > ts->ts_slptime) {
1585 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1586 			ts->ts_slptime = 1;
1587 		} else {
1588 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1589 			ts->ts_runtime = 1;
1590 		}
1591 		return;
1592 	}
1593 	/*
1594 	 * If we have exceeded by more than 1/5th then the algorithm below
1595 	 * will not bring us back into range.  Dividing by two here forces
1596 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1597 	 */
1598 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1599 		ts->ts_runtime /= 2;
1600 		ts->ts_slptime /= 2;
1601 		return;
1602 	}
1603 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1604 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1605 }
1606 
1607 /*
1608  * Scale back the interactivity history when a child thread is created.  The
1609  * history is inherited from the parent but the thread may behave totally
1610  * differently.  For example, a shell spawning a compiler process.  We want
1611  * to learn that the compiler is behaving badly very quickly.
1612  */
1613 static void
1614 sched_interact_fork(struct thread *td)
1615 {
1616 	struct td_sched *ts;
1617 	int ratio;
1618 	int sum;
1619 
1620 	ts = td_get_sched(td);
1621 	sum = ts->ts_runtime + ts->ts_slptime;
1622 	if (sum > SCHED_SLP_RUN_FORK) {
1623 		ratio = sum / SCHED_SLP_RUN_FORK;
1624 		ts->ts_runtime /= ratio;
1625 		ts->ts_slptime /= ratio;
1626 	}
1627 }
1628 
1629 /*
1630  * Called from proc0_init() to setup the scheduler fields.
1631  */
1632 void
1633 schedinit(void)
1634 {
1635 	struct td_sched *ts0;
1636 
1637 	/*
1638 	 * Set up the scheduler specific parts of thread0.
1639 	 */
1640 	ts0 = td_get_sched(&thread0);
1641 	ts0->ts_ltick = ticks;
1642 	ts0->ts_ftick = ticks;
1643 	ts0->ts_slice = 0;
1644 	ts0->ts_cpu = curcpu;	/* set valid CPU number */
1645 }
1646 
1647 /*
1648  * This is only somewhat accurate since given many processes of the same
1649  * priority they will switch when their slices run out, which will be
1650  * at most sched_slice stathz ticks.
1651  */
1652 int
1653 sched_rr_interval(void)
1654 {
1655 
1656 	/* Convert sched_slice from stathz to hz. */
1657 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1658 }
1659 
1660 /*
1661  * Update the percent cpu tracking information when it is requested or
1662  * the total history exceeds the maximum.  We keep a sliding history of
1663  * tick counts that slowly decays.  This is less precise than the 4BSD
1664  * mechanism since it happens with less regular and frequent events.
1665  */
1666 static void
1667 sched_pctcpu_update(struct td_sched *ts, int run)
1668 {
1669 	int t = ticks;
1670 
1671 	/*
1672 	 * The signed difference may be negative if the thread hasn't run for
1673 	 * over half of the ticks rollover period.
1674 	 */
1675 	if ((u_int)(t - ts->ts_ltick) >= SCHED_TICK_TARG) {
1676 		ts->ts_ticks = 0;
1677 		ts->ts_ftick = t - SCHED_TICK_TARG;
1678 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1679 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1680 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1681 		ts->ts_ftick = t - SCHED_TICK_TARG;
1682 	}
1683 	if (run)
1684 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1685 	ts->ts_ltick = t;
1686 }
1687 
1688 /*
1689  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1690  * if necessary.  This is the back-end for several priority related
1691  * functions.
1692  */
1693 static void
1694 sched_thread_priority(struct thread *td, u_char prio)
1695 {
1696 	struct td_sched *ts;
1697 	struct tdq *tdq;
1698 	int oldpri;
1699 
1700 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1701 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1702 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1703 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1704 	if (td != curthread && prio < td->td_priority) {
1705 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1706 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1707 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1708 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1709 		    curthread);
1710 	}
1711 	ts = td_get_sched(td);
1712 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1713 	if (td->td_priority == prio)
1714 		return;
1715 	/*
1716 	 * If the priority has been elevated due to priority
1717 	 * propagation, we may have to move ourselves to a new
1718 	 * queue.  This could be optimized to not re-add in some
1719 	 * cases.
1720 	 */
1721 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1722 		sched_rem(td);
1723 		td->td_priority = prio;
1724 		sched_add(td, SRQ_BORROWING);
1725 		return;
1726 	}
1727 	/*
1728 	 * If the thread is currently running we may have to adjust the lowpri
1729 	 * information so other cpus are aware of our current priority.
1730 	 */
1731 	if (TD_IS_RUNNING(td)) {
1732 		tdq = TDQ_CPU(ts->ts_cpu);
1733 		oldpri = td->td_priority;
1734 		td->td_priority = prio;
1735 		if (prio < tdq->tdq_lowpri)
1736 			tdq->tdq_lowpri = prio;
1737 		else if (tdq->tdq_lowpri == oldpri)
1738 			tdq_setlowpri(tdq, td);
1739 		return;
1740 	}
1741 	td->td_priority = prio;
1742 }
1743 
1744 /*
1745  * Update a thread's priority when it is lent another thread's
1746  * priority.
1747  */
1748 void
1749 sched_lend_prio(struct thread *td, u_char prio)
1750 {
1751 
1752 	td->td_flags |= TDF_BORROWING;
1753 	sched_thread_priority(td, prio);
1754 }
1755 
1756 /*
1757  * Restore a thread's priority when priority propagation is
1758  * over.  The prio argument is the minimum priority the thread
1759  * needs to have to satisfy other possible priority lending
1760  * requests.  If the thread's regular priority is less
1761  * important than prio, the thread will keep a priority boost
1762  * of prio.
1763  */
1764 void
1765 sched_unlend_prio(struct thread *td, u_char prio)
1766 {
1767 	u_char base_pri;
1768 
1769 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1770 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1771 		base_pri = td->td_user_pri;
1772 	else
1773 		base_pri = td->td_base_pri;
1774 	if (prio >= base_pri) {
1775 		td->td_flags &= ~TDF_BORROWING;
1776 		sched_thread_priority(td, base_pri);
1777 	} else
1778 		sched_lend_prio(td, prio);
1779 }
1780 
1781 /*
1782  * Standard entry for setting the priority to an absolute value.
1783  */
1784 void
1785 sched_prio(struct thread *td, u_char prio)
1786 {
1787 	u_char oldprio;
1788 
1789 	/* First, update the base priority. */
1790 	td->td_base_pri = prio;
1791 
1792 	/*
1793 	 * If the thread is borrowing another thread's priority, don't
1794 	 * ever lower the priority.
1795 	 */
1796 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1797 		return;
1798 
1799 	/* Change the real priority. */
1800 	oldprio = td->td_priority;
1801 	sched_thread_priority(td, prio);
1802 
1803 	/*
1804 	 * If the thread is on a turnstile, then let the turnstile update
1805 	 * its state.
1806 	 */
1807 	if (TD_ON_LOCK(td) && oldprio != prio)
1808 		turnstile_adjust(td, oldprio);
1809 }
1810 
1811 /*
1812  * Set the base user priority, does not effect current running priority.
1813  */
1814 void
1815 sched_user_prio(struct thread *td, u_char prio)
1816 {
1817 
1818 	td->td_base_user_pri = prio;
1819 	if (td->td_lend_user_pri <= prio)
1820 		return;
1821 	td->td_user_pri = prio;
1822 }
1823 
1824 void
1825 sched_lend_user_prio(struct thread *td, u_char prio)
1826 {
1827 
1828 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1829 	td->td_lend_user_pri = prio;
1830 	td->td_user_pri = min(prio, td->td_base_user_pri);
1831 	if (td->td_priority > td->td_user_pri)
1832 		sched_prio(td, td->td_user_pri);
1833 	else if (td->td_priority != td->td_user_pri)
1834 		td->td_flags |= TDF_NEEDRESCHED;
1835 }
1836 
1837 /*
1838  * Handle migration from sched_switch().  This happens only for
1839  * cpu binding.
1840  */
1841 static struct mtx *
1842 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1843 {
1844 	struct tdq *tdn;
1845 
1846 	KASSERT(!CPU_ABSENT(td_get_sched(td)->ts_cpu), ("sched_switch_migrate: "
1847 	    "thread %s queued on absent CPU %d.", td->td_name,
1848 	    td_get_sched(td)->ts_cpu));
1849 	tdn = TDQ_CPU(td_get_sched(td)->ts_cpu);
1850 #ifdef SMP
1851 	tdq_load_rem(tdq, td);
1852 	/*
1853 	 * Do the lock dance required to avoid LOR.  We grab an extra
1854 	 * spinlock nesting to prevent preemption while we're
1855 	 * not holding either run-queue lock.
1856 	 */
1857 	spinlock_enter();
1858 	thread_lock_block(td);	/* This releases the lock on tdq. */
1859 
1860 	/*
1861 	 * Acquire both run-queue locks before placing the thread on the new
1862 	 * run-queue to avoid deadlocks created by placing a thread with a
1863 	 * blocked lock on the run-queue of a remote processor.  The deadlock
1864 	 * occurs when a third processor attempts to lock the two queues in
1865 	 * question while the target processor is spinning with its own
1866 	 * run-queue lock held while waiting for the blocked lock to clear.
1867 	 */
1868 	tdq_lock_pair(tdn, tdq);
1869 	tdq_add(tdn, td, flags);
1870 	tdq_notify(tdn, td);
1871 	TDQ_UNLOCK(tdn);
1872 	spinlock_exit();
1873 #endif
1874 	return (TDQ_LOCKPTR(tdn));
1875 }
1876 
1877 /*
1878  * Variadic version of thread_lock_unblock() that does not assume td_lock
1879  * is blocked.
1880  */
1881 static inline void
1882 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1883 {
1884 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1885 	    (uintptr_t)mtx);
1886 }
1887 
1888 /*
1889  * Switch threads.  This function has to handle threads coming in while
1890  * blocked for some reason, running, or idle.  It also must deal with
1891  * migrating a thread from one queue to another as running threads may
1892  * be assigned elsewhere via binding.
1893  */
1894 void
1895 sched_switch(struct thread *td, struct thread *newtd, int flags)
1896 {
1897 	struct tdq *tdq;
1898 	struct td_sched *ts;
1899 	struct mtx *mtx;
1900 	int srqflag;
1901 	int cpuid, preempted;
1902 
1903 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1904 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1905 
1906 	cpuid = PCPU_GET(cpuid);
1907 	tdq = TDQ_CPU(cpuid);
1908 	ts = td_get_sched(td);
1909 	mtx = td->td_lock;
1910 	sched_pctcpu_update(ts, 1);
1911 	ts->ts_rltick = ticks;
1912 	td->td_lastcpu = td->td_oncpu;
1913 	td->td_oncpu = NOCPU;
1914 	preempted = (td->td_flags & TDF_SLICEEND) == 0 &&
1915 	    (flags & SW_PREEMPT) != 0;
1916 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1917 	td->td_owepreempt = 0;
1918 	if (!TD_IS_IDLETHREAD(td))
1919 		tdq->tdq_switchcnt++;
1920 	/*
1921 	 * The lock pointer in an idle thread should never change.  Reset it
1922 	 * to CAN_RUN as well.
1923 	 */
1924 	if (TD_IS_IDLETHREAD(td)) {
1925 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1926 		TD_SET_CAN_RUN(td);
1927 	} else if (TD_IS_RUNNING(td)) {
1928 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1929 		srqflag = preempted ?
1930 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1931 		    SRQ_OURSELF|SRQ_YIELDING;
1932 #ifdef SMP
1933 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1934 			ts->ts_cpu = sched_pickcpu(td, 0);
1935 #endif
1936 		if (ts->ts_cpu == cpuid)
1937 			tdq_runq_add(tdq, td, srqflag);
1938 		else {
1939 			KASSERT(THREAD_CAN_MIGRATE(td) ||
1940 			    (ts->ts_flags & TSF_BOUND) != 0,
1941 			    ("Thread %p shouldn't migrate", td));
1942 			mtx = sched_switch_migrate(tdq, td, srqflag);
1943 		}
1944 	} else {
1945 		/* This thread must be going to sleep. */
1946 		TDQ_LOCK(tdq);
1947 		mtx = thread_lock_block(td);
1948 		tdq_load_rem(tdq, td);
1949 	}
1950 
1951 #if (KTR_COMPILE & KTR_SCHED) != 0
1952 	if (TD_IS_IDLETHREAD(td))
1953 		KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle",
1954 		    "prio:%d", td->td_priority);
1955 	else
1956 		KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td),
1957 		    "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg,
1958 		    "lockname:\"%s\"", td->td_lockname);
1959 #endif
1960 
1961 	/*
1962 	 * We enter here with the thread blocked and assigned to the
1963 	 * appropriate cpu run-queue or sleep-queue and with the current
1964 	 * thread-queue locked.
1965 	 */
1966 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1967 	newtd = choosethread();
1968 	/*
1969 	 * Call the MD code to switch contexts if necessary.
1970 	 */
1971 	if (td != newtd) {
1972 #ifdef	HWPMC_HOOKS
1973 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1974 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1975 #endif
1976 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1977 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1978 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1979 		sched_pctcpu_update(td_get_sched(newtd), 0);
1980 
1981 #ifdef KDTRACE_HOOKS
1982 		/*
1983 		 * If DTrace has set the active vtime enum to anything
1984 		 * other than INACTIVE (0), then it should have set the
1985 		 * function to call.
1986 		 */
1987 		if (dtrace_vtime_active)
1988 			(*dtrace_vtime_switch_func)(newtd);
1989 #endif
1990 
1991 		cpu_switch(td, newtd, mtx);
1992 		/*
1993 		 * We may return from cpu_switch on a different cpu.  However,
1994 		 * we always return with td_lock pointing to the current cpu's
1995 		 * run queue lock.
1996 		 */
1997 		cpuid = PCPU_GET(cpuid);
1998 		tdq = TDQ_CPU(cpuid);
1999 		lock_profile_obtain_lock_success(
2000 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2001 
2002 		SDT_PROBE0(sched, , , on__cpu);
2003 #ifdef	HWPMC_HOOKS
2004 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
2005 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
2006 #endif
2007 	} else {
2008 		thread_unblock_switch(td, mtx);
2009 		SDT_PROBE0(sched, , , remain__cpu);
2010 	}
2011 
2012 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2013 	    "prio:%d", td->td_priority);
2014 
2015 	/*
2016 	 * Assert that all went well and return.
2017 	 */
2018 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
2019 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2020 	td->td_oncpu = cpuid;
2021 }
2022 
2023 /*
2024  * Adjust thread priorities as a result of a nice request.
2025  */
2026 void
2027 sched_nice(struct proc *p, int nice)
2028 {
2029 	struct thread *td;
2030 
2031 	PROC_LOCK_ASSERT(p, MA_OWNED);
2032 
2033 	p->p_nice = nice;
2034 	FOREACH_THREAD_IN_PROC(p, td) {
2035 		thread_lock(td);
2036 		sched_priority(td);
2037 		sched_prio(td, td->td_base_user_pri);
2038 		thread_unlock(td);
2039 	}
2040 }
2041 
2042 /*
2043  * Record the sleep time for the interactivity scorer.
2044  */
2045 void
2046 sched_sleep(struct thread *td, int prio)
2047 {
2048 
2049 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2050 
2051 	td->td_slptick = ticks;
2052 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2053 		td->td_flags |= TDF_CANSWAP;
2054 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2055 		return;
2056 	if (static_boost == 1 && prio)
2057 		sched_prio(td, prio);
2058 	else if (static_boost && td->td_priority > static_boost)
2059 		sched_prio(td, static_boost);
2060 }
2061 
2062 /*
2063  * Schedule a thread to resume execution and record how long it voluntarily
2064  * slept.  We also update the pctcpu, interactivity, and priority.
2065  */
2066 void
2067 sched_wakeup(struct thread *td)
2068 {
2069 	struct td_sched *ts;
2070 	int slptick;
2071 
2072 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2073 	ts = td_get_sched(td);
2074 	td->td_flags &= ~TDF_CANSWAP;
2075 	/*
2076 	 * If we slept for more than a tick update our interactivity and
2077 	 * priority.
2078 	 */
2079 	slptick = td->td_slptick;
2080 	td->td_slptick = 0;
2081 	if (slptick && slptick != ticks) {
2082 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2083 		sched_interact_update(td);
2084 		sched_pctcpu_update(ts, 0);
2085 	}
2086 	/*
2087 	 * Reset the slice value since we slept and advanced the round-robin.
2088 	 */
2089 	ts->ts_slice = 0;
2090 	sched_add(td, SRQ_BORING);
2091 }
2092 
2093 /*
2094  * Penalize the parent for creating a new child and initialize the child's
2095  * priority.
2096  */
2097 void
2098 sched_fork(struct thread *td, struct thread *child)
2099 {
2100 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2101 	sched_pctcpu_update(td_get_sched(td), 1);
2102 	sched_fork_thread(td, child);
2103 	/*
2104 	 * Penalize the parent and child for forking.
2105 	 */
2106 	sched_interact_fork(child);
2107 	sched_priority(child);
2108 	td_get_sched(td)->ts_runtime += tickincr;
2109 	sched_interact_update(td);
2110 	sched_priority(td);
2111 }
2112 
2113 /*
2114  * Fork a new thread, may be within the same process.
2115  */
2116 void
2117 sched_fork_thread(struct thread *td, struct thread *child)
2118 {
2119 	struct td_sched *ts;
2120 	struct td_sched *ts2;
2121 	struct tdq *tdq;
2122 
2123 	tdq = TDQ_SELF();
2124 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2125 	/*
2126 	 * Initialize child.
2127 	 */
2128 	ts = td_get_sched(td);
2129 	ts2 = td_get_sched(child);
2130 	child->td_oncpu = NOCPU;
2131 	child->td_lastcpu = NOCPU;
2132 	child->td_lock = TDQ_LOCKPTR(tdq);
2133 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2134 	child->td_domain.dr_policy = td->td_cpuset->cs_domain;
2135 	ts2->ts_cpu = ts->ts_cpu;
2136 	ts2->ts_flags = 0;
2137 	/*
2138 	 * Grab our parents cpu estimation information.
2139 	 */
2140 	ts2->ts_ticks = ts->ts_ticks;
2141 	ts2->ts_ltick = ts->ts_ltick;
2142 	ts2->ts_ftick = ts->ts_ftick;
2143 	/*
2144 	 * Do not inherit any borrowed priority from the parent.
2145 	 */
2146 	child->td_priority = child->td_base_pri;
2147 	/*
2148 	 * And update interactivity score.
2149 	 */
2150 	ts2->ts_slptime = ts->ts_slptime;
2151 	ts2->ts_runtime = ts->ts_runtime;
2152 	/* Attempt to quickly learn interactivity. */
2153 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2154 #ifdef KTR
2155 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2156 #endif
2157 }
2158 
2159 /*
2160  * Adjust the priority class of a thread.
2161  */
2162 void
2163 sched_class(struct thread *td, int class)
2164 {
2165 
2166 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2167 	if (td->td_pri_class == class)
2168 		return;
2169 	td->td_pri_class = class;
2170 }
2171 
2172 /*
2173  * Return some of the child's priority and interactivity to the parent.
2174  */
2175 void
2176 sched_exit(struct proc *p, struct thread *child)
2177 {
2178 	struct thread *td;
2179 
2180 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2181 	    "prio:%d", child->td_priority);
2182 	PROC_LOCK_ASSERT(p, MA_OWNED);
2183 	td = FIRST_THREAD_IN_PROC(p);
2184 	sched_exit_thread(td, child);
2185 }
2186 
2187 /*
2188  * Penalize another thread for the time spent on this one.  This helps to
2189  * worsen the priority and interactivity of processes which schedule batch
2190  * jobs such as make.  This has little effect on the make process itself but
2191  * causes new processes spawned by it to receive worse scores immediately.
2192  */
2193 void
2194 sched_exit_thread(struct thread *td, struct thread *child)
2195 {
2196 
2197 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2198 	    "prio:%d", child->td_priority);
2199 	/*
2200 	 * Give the child's runtime to the parent without returning the
2201 	 * sleep time as a penalty to the parent.  This causes shells that
2202 	 * launch expensive things to mark their children as expensive.
2203 	 */
2204 	thread_lock(td);
2205 	td_get_sched(td)->ts_runtime += td_get_sched(child)->ts_runtime;
2206 	sched_interact_update(td);
2207 	sched_priority(td);
2208 	thread_unlock(td);
2209 }
2210 
2211 void
2212 sched_preempt(struct thread *td)
2213 {
2214 	struct tdq *tdq;
2215 
2216 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2217 
2218 	thread_lock(td);
2219 	tdq = TDQ_SELF();
2220 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2221 	tdq->tdq_ipipending = 0;
2222 	if (td->td_priority > tdq->tdq_lowpri) {
2223 		int flags;
2224 
2225 		flags = SW_INVOL | SW_PREEMPT;
2226 		if (td->td_critnest > 1)
2227 			td->td_owepreempt = 1;
2228 		else if (TD_IS_IDLETHREAD(td))
2229 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2230 		else
2231 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2232 	}
2233 	thread_unlock(td);
2234 }
2235 
2236 /*
2237  * Fix priorities on return to user-space.  Priorities may be elevated due
2238  * to static priorities in msleep() or similar.
2239  */
2240 void
2241 sched_userret(struct thread *td)
2242 {
2243 	/*
2244 	 * XXX we cheat slightly on the locking here to avoid locking in
2245 	 * the usual case.  Setting td_priority here is essentially an
2246 	 * incomplete workaround for not setting it properly elsewhere.
2247 	 * Now that some interrupt handlers are threads, not setting it
2248 	 * properly elsewhere can clobber it in the window between setting
2249 	 * it here and returning to user mode, so don't waste time setting
2250 	 * it perfectly here.
2251 	 */
2252 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2253 	    ("thread with borrowed priority returning to userland"));
2254 	if (td->td_priority != td->td_user_pri) {
2255 		thread_lock(td);
2256 		td->td_priority = td->td_user_pri;
2257 		td->td_base_pri = td->td_user_pri;
2258 		tdq_setlowpri(TDQ_SELF(), td);
2259 		thread_unlock(td);
2260         }
2261 }
2262 
2263 /*
2264  * Handle a stathz tick.  This is really only relevant for timeshare
2265  * threads.
2266  */
2267 void
2268 sched_clock(struct thread *td)
2269 {
2270 	struct tdq *tdq;
2271 	struct td_sched *ts;
2272 
2273 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2274 	tdq = TDQ_SELF();
2275 #ifdef SMP
2276 	/*
2277 	 * We run the long term load balancer infrequently on the first cpu.
2278 	 */
2279 	if (balance_tdq == tdq) {
2280 		if (balance_ticks && --balance_ticks == 0)
2281 			sched_balance();
2282 	}
2283 #endif
2284 	/*
2285 	 * Save the old switch count so we have a record of the last ticks
2286 	 * activity.   Initialize the new switch count based on our load.
2287 	 * If there is some activity seed it to reflect that.
2288 	 */
2289 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2290 	tdq->tdq_switchcnt = tdq->tdq_load;
2291 	/*
2292 	 * Advance the insert index once for each tick to ensure that all
2293 	 * threads get a chance to run.
2294 	 */
2295 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2296 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2297 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2298 			tdq->tdq_ridx = tdq->tdq_idx;
2299 	}
2300 	ts = td_get_sched(td);
2301 	sched_pctcpu_update(ts, 1);
2302 	if (td->td_pri_class & PRI_FIFO_BIT)
2303 		return;
2304 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2305 		/*
2306 		 * We used a tick; charge it to the thread so
2307 		 * that we can compute our interactivity.
2308 		 */
2309 		td_get_sched(td)->ts_runtime += tickincr;
2310 		sched_interact_update(td);
2311 		sched_priority(td);
2312 	}
2313 
2314 	/*
2315 	 * Force a context switch if the current thread has used up a full
2316 	 * time slice (default is 100ms).
2317 	 */
2318 	if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2319 		ts->ts_slice = 0;
2320 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2321 	}
2322 }
2323 
2324 u_int
2325 sched_estcpu(struct thread *td __unused)
2326 {
2327 
2328 	return (0);
2329 }
2330 
2331 /*
2332  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2333  * cooperative idle threads.
2334  */
2335 int
2336 sched_runnable(void)
2337 {
2338 	struct tdq *tdq;
2339 	int load;
2340 
2341 	load = 1;
2342 
2343 	tdq = TDQ_SELF();
2344 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2345 		if (tdq->tdq_load > 0)
2346 			goto out;
2347 	} else
2348 		if (tdq->tdq_load - 1 > 0)
2349 			goto out;
2350 	load = 0;
2351 out:
2352 	return (load);
2353 }
2354 
2355 /*
2356  * Choose the highest priority thread to run.  The thread is removed from
2357  * the run-queue while running however the load remains.  For SMP we set
2358  * the tdq in the global idle bitmask if it idles here.
2359  */
2360 struct thread *
2361 sched_choose(void)
2362 {
2363 	struct thread *td;
2364 	struct tdq *tdq;
2365 
2366 	tdq = TDQ_SELF();
2367 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2368 	td = tdq_choose(tdq);
2369 	if (td) {
2370 		tdq_runq_rem(tdq, td);
2371 		tdq->tdq_lowpri = td->td_priority;
2372 		return (td);
2373 	}
2374 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2375 	return (PCPU_GET(idlethread));
2376 }
2377 
2378 /*
2379  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2380  * we always request it once we exit a critical section.
2381  */
2382 static inline void
2383 sched_setpreempt(struct thread *td)
2384 {
2385 	struct thread *ctd;
2386 	int cpri;
2387 	int pri;
2388 
2389 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2390 
2391 	ctd = curthread;
2392 	pri = td->td_priority;
2393 	cpri = ctd->td_priority;
2394 	if (pri < cpri)
2395 		ctd->td_flags |= TDF_NEEDRESCHED;
2396 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2397 		return;
2398 	if (!sched_shouldpreempt(pri, cpri, 0))
2399 		return;
2400 	ctd->td_owepreempt = 1;
2401 }
2402 
2403 /*
2404  * Add a thread to a thread queue.  Select the appropriate runq and add the
2405  * thread to it.  This is the internal function called when the tdq is
2406  * predetermined.
2407  */
2408 void
2409 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2410 {
2411 
2412 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2413 	KASSERT((td->td_inhibitors == 0),
2414 	    ("sched_add: trying to run inhibited thread"));
2415 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2416 	    ("sched_add: bad thread state"));
2417 	KASSERT(td->td_flags & TDF_INMEM,
2418 	    ("sched_add: thread swapped out"));
2419 
2420 	if (td->td_priority < tdq->tdq_lowpri)
2421 		tdq->tdq_lowpri = td->td_priority;
2422 	tdq_runq_add(tdq, td, flags);
2423 	tdq_load_add(tdq, td);
2424 }
2425 
2426 /*
2427  * Select the target thread queue and add a thread to it.  Request
2428  * preemption or IPI a remote processor if required.
2429  */
2430 void
2431 sched_add(struct thread *td, int flags)
2432 {
2433 	struct tdq *tdq;
2434 #ifdef SMP
2435 	int cpu;
2436 #endif
2437 
2438 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2439 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2440 	    sched_tdname(curthread));
2441 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2442 	    KTR_ATTR_LINKED, sched_tdname(td));
2443 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2444 	    flags & SRQ_PREEMPTED);
2445 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2446 	/*
2447 	 * Recalculate the priority before we select the target cpu or
2448 	 * run-queue.
2449 	 */
2450 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2451 		sched_priority(td);
2452 #ifdef SMP
2453 	/*
2454 	 * Pick the destination cpu and if it isn't ours transfer to the
2455 	 * target cpu.
2456 	 */
2457 	cpu = sched_pickcpu(td, flags);
2458 	tdq = sched_setcpu(td, cpu, flags);
2459 	tdq_add(tdq, td, flags);
2460 	if (cpu != PCPU_GET(cpuid)) {
2461 		tdq_notify(tdq, td);
2462 		return;
2463 	}
2464 #else
2465 	tdq = TDQ_SELF();
2466 	TDQ_LOCK(tdq);
2467 	/*
2468 	 * Now that the thread is moving to the run-queue, set the lock
2469 	 * to the scheduler's lock.
2470 	 */
2471 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2472 	tdq_add(tdq, td, flags);
2473 #endif
2474 	if (!(flags & SRQ_YIELDING))
2475 		sched_setpreempt(td);
2476 }
2477 
2478 /*
2479  * Remove a thread from a run-queue without running it.  This is used
2480  * when we're stealing a thread from a remote queue.  Otherwise all threads
2481  * exit by calling sched_exit_thread() and sched_throw() themselves.
2482  */
2483 void
2484 sched_rem(struct thread *td)
2485 {
2486 	struct tdq *tdq;
2487 
2488 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2489 	    "prio:%d", td->td_priority);
2490 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2491 	tdq = TDQ_CPU(td_get_sched(td)->ts_cpu);
2492 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2493 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2494 	KASSERT(TD_ON_RUNQ(td),
2495 	    ("sched_rem: thread not on run queue"));
2496 	tdq_runq_rem(tdq, td);
2497 	tdq_load_rem(tdq, td);
2498 	TD_SET_CAN_RUN(td);
2499 	if (td->td_priority == tdq->tdq_lowpri)
2500 		tdq_setlowpri(tdq, NULL);
2501 }
2502 
2503 /*
2504  * Fetch cpu utilization information.  Updates on demand.
2505  */
2506 fixpt_t
2507 sched_pctcpu(struct thread *td)
2508 {
2509 	fixpt_t pctcpu;
2510 	struct td_sched *ts;
2511 
2512 	pctcpu = 0;
2513 	ts = td_get_sched(td);
2514 
2515 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2516 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2517 	if (ts->ts_ticks) {
2518 		int rtick;
2519 
2520 		/* How many rtick per second ? */
2521 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2522 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2523 	}
2524 
2525 	return (pctcpu);
2526 }
2527 
2528 /*
2529  * Enforce affinity settings for a thread.  Called after adjustments to
2530  * cpumask.
2531  */
2532 void
2533 sched_affinity(struct thread *td)
2534 {
2535 #ifdef SMP
2536 	struct td_sched *ts;
2537 
2538 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2539 	ts = td_get_sched(td);
2540 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2541 		return;
2542 	if (TD_ON_RUNQ(td)) {
2543 		sched_rem(td);
2544 		sched_add(td, SRQ_BORING);
2545 		return;
2546 	}
2547 	if (!TD_IS_RUNNING(td))
2548 		return;
2549 	/*
2550 	 * Force a switch before returning to userspace.  If the
2551 	 * target thread is not running locally send an ipi to force
2552 	 * the issue.
2553 	 */
2554 	td->td_flags |= TDF_NEEDRESCHED;
2555 	if (td != curthread)
2556 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2557 #endif
2558 }
2559 
2560 /*
2561  * Bind a thread to a target cpu.
2562  */
2563 void
2564 sched_bind(struct thread *td, int cpu)
2565 {
2566 	struct td_sched *ts;
2567 
2568 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2569 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2570 	ts = td_get_sched(td);
2571 	if (ts->ts_flags & TSF_BOUND)
2572 		sched_unbind(td);
2573 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2574 	ts->ts_flags |= TSF_BOUND;
2575 	sched_pin();
2576 	if (PCPU_GET(cpuid) == cpu)
2577 		return;
2578 	ts->ts_cpu = cpu;
2579 	/* When we return from mi_switch we'll be on the correct cpu. */
2580 	mi_switch(SW_VOL, NULL);
2581 }
2582 
2583 /*
2584  * Release a bound thread.
2585  */
2586 void
2587 sched_unbind(struct thread *td)
2588 {
2589 	struct td_sched *ts;
2590 
2591 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2592 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2593 	ts = td_get_sched(td);
2594 	if ((ts->ts_flags & TSF_BOUND) == 0)
2595 		return;
2596 	ts->ts_flags &= ~TSF_BOUND;
2597 	sched_unpin();
2598 }
2599 
2600 int
2601 sched_is_bound(struct thread *td)
2602 {
2603 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2604 	return (td_get_sched(td)->ts_flags & TSF_BOUND);
2605 }
2606 
2607 /*
2608  * Basic yield call.
2609  */
2610 void
2611 sched_relinquish(struct thread *td)
2612 {
2613 	thread_lock(td);
2614 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2615 	thread_unlock(td);
2616 }
2617 
2618 /*
2619  * Return the total system load.
2620  */
2621 int
2622 sched_load(void)
2623 {
2624 #ifdef SMP
2625 	int total;
2626 	int i;
2627 
2628 	total = 0;
2629 	CPU_FOREACH(i)
2630 		total += TDQ_CPU(i)->tdq_sysload;
2631 	return (total);
2632 #else
2633 	return (TDQ_SELF()->tdq_sysload);
2634 #endif
2635 }
2636 
2637 int
2638 sched_sizeof_proc(void)
2639 {
2640 	return (sizeof(struct proc));
2641 }
2642 
2643 int
2644 sched_sizeof_thread(void)
2645 {
2646 	return (sizeof(struct thread) + sizeof(struct td_sched));
2647 }
2648 
2649 #ifdef SMP
2650 #define	TDQ_IDLESPIN(tdq)						\
2651     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2652 #else
2653 #define	TDQ_IDLESPIN(tdq)	1
2654 #endif
2655 
2656 /*
2657  * The actual idle process.
2658  */
2659 void
2660 sched_idletd(void *dummy)
2661 {
2662 	struct thread *td;
2663 	struct tdq *tdq;
2664 	int oldswitchcnt, switchcnt;
2665 	int i;
2666 
2667 	mtx_assert(&Giant, MA_NOTOWNED);
2668 	td = curthread;
2669 	tdq = TDQ_SELF();
2670 	THREAD_NO_SLEEPING();
2671 	oldswitchcnt = -1;
2672 	for (;;) {
2673 		if (tdq->tdq_load) {
2674 			thread_lock(td);
2675 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2676 			thread_unlock(td);
2677 		}
2678 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2679 #ifdef SMP
2680 		if (switchcnt != oldswitchcnt) {
2681 			oldswitchcnt = switchcnt;
2682 			if (tdq_idled(tdq) == 0)
2683 				continue;
2684 		}
2685 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2686 #else
2687 		oldswitchcnt = switchcnt;
2688 #endif
2689 		/*
2690 		 * If we're switching very frequently, spin while checking
2691 		 * for load rather than entering a low power state that
2692 		 * may require an IPI.  However, don't do any busy
2693 		 * loops while on SMT machines as this simply steals
2694 		 * cycles from cores doing useful work.
2695 		 */
2696 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2697 			for (i = 0; i < sched_idlespins; i++) {
2698 				if (tdq->tdq_load)
2699 					break;
2700 				cpu_spinwait();
2701 			}
2702 		}
2703 
2704 		/* If there was context switch during spin, restart it. */
2705 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2706 		if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2707 			continue;
2708 
2709 		/* Run main MD idle handler. */
2710 		tdq->tdq_cpu_idle = 1;
2711 		/*
2712 		 * Make sure that tdq_cpu_idle update is globally visible
2713 		 * before cpu_idle() read tdq_load.  The order is important
2714 		 * to avoid race with tdq_notify.
2715 		 */
2716 		atomic_thread_fence_seq_cst();
2717 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2718 		tdq->tdq_cpu_idle = 0;
2719 
2720 		/*
2721 		 * Account thread-less hardware interrupts and
2722 		 * other wakeup reasons equal to context switches.
2723 		 */
2724 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2725 		if (switchcnt != oldswitchcnt)
2726 			continue;
2727 		tdq->tdq_switchcnt++;
2728 		oldswitchcnt++;
2729 	}
2730 }
2731 
2732 /*
2733  * A CPU is entering for the first time or a thread is exiting.
2734  */
2735 void
2736 sched_throw(struct thread *td)
2737 {
2738 	struct thread *newtd;
2739 	struct tdq *tdq;
2740 
2741 	tdq = TDQ_SELF();
2742 	if (td == NULL) {
2743 		/* Correct spinlock nesting and acquire the correct lock. */
2744 		TDQ_LOCK(tdq);
2745 		spinlock_exit();
2746 		PCPU_SET(switchtime, cpu_ticks());
2747 		PCPU_SET(switchticks, ticks);
2748 	} else {
2749 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2750 		tdq_load_rem(tdq, td);
2751 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2752 		td->td_lastcpu = td->td_oncpu;
2753 		td->td_oncpu = NOCPU;
2754 	}
2755 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2756 	newtd = choosethread();
2757 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2758 	cpu_throw(td, newtd);		/* doesn't return */
2759 }
2760 
2761 /*
2762  * This is called from fork_exit().  Just acquire the correct locks and
2763  * let fork do the rest of the work.
2764  */
2765 void
2766 sched_fork_exit(struct thread *td)
2767 {
2768 	struct tdq *tdq;
2769 	int cpuid;
2770 
2771 	/*
2772 	 * Finish setting up thread glue so that it begins execution in a
2773 	 * non-nested critical section with the scheduler lock held.
2774 	 */
2775 	cpuid = PCPU_GET(cpuid);
2776 	tdq = TDQ_CPU(cpuid);
2777 	if (TD_IS_IDLETHREAD(td))
2778 		td->td_lock = TDQ_LOCKPTR(tdq);
2779 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2780 	td->td_oncpu = cpuid;
2781 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2782 	lock_profile_obtain_lock_success(
2783 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2784 
2785 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running",
2786 	    "prio:%d", td->td_priority);
2787 	SDT_PROBE0(sched, , , on__cpu);
2788 }
2789 
2790 /*
2791  * Create on first use to catch odd startup conditons.
2792  */
2793 char *
2794 sched_tdname(struct thread *td)
2795 {
2796 #ifdef KTR
2797 	struct td_sched *ts;
2798 
2799 	ts = td_get_sched(td);
2800 	if (ts->ts_name[0] == '\0')
2801 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2802 		    "%s tid %d", td->td_name, td->td_tid);
2803 	return (ts->ts_name);
2804 #else
2805 	return (td->td_name);
2806 #endif
2807 }
2808 
2809 #ifdef KTR
2810 void
2811 sched_clear_tdname(struct thread *td)
2812 {
2813 	struct td_sched *ts;
2814 
2815 	ts = td_get_sched(td);
2816 	ts->ts_name[0] = '\0';
2817 }
2818 #endif
2819 
2820 #ifdef SMP
2821 
2822 /*
2823  * Build the CPU topology dump string. Is recursively called to collect
2824  * the topology tree.
2825  */
2826 static int
2827 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2828     int indent)
2829 {
2830 	char cpusetbuf[CPUSETBUFSIZ];
2831 	int i, first;
2832 
2833 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2834 	    "", 1 + indent / 2, cg->cg_level);
2835 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2836 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2837 	first = TRUE;
2838 	for (i = 0; i < MAXCPU; i++) {
2839 		if (CPU_ISSET(i, &cg->cg_mask)) {
2840 			if (!first)
2841 				sbuf_printf(sb, ", ");
2842 			else
2843 				first = FALSE;
2844 			sbuf_printf(sb, "%d", i);
2845 		}
2846 	}
2847 	sbuf_printf(sb, "</cpu>\n");
2848 
2849 	if (cg->cg_flags != 0) {
2850 		sbuf_printf(sb, "%*s <flags>", indent, "");
2851 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2852 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2853 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2854 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2855 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2856 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2857 		sbuf_printf(sb, "</flags>\n");
2858 	}
2859 
2860 	if (cg->cg_children > 0) {
2861 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2862 		for (i = 0; i < cg->cg_children; i++)
2863 			sysctl_kern_sched_topology_spec_internal(sb,
2864 			    &cg->cg_child[i], indent+2);
2865 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2866 	}
2867 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2868 	return (0);
2869 }
2870 
2871 /*
2872  * Sysctl handler for retrieving topology dump. It's a wrapper for
2873  * the recursive sysctl_kern_smp_topology_spec_internal().
2874  */
2875 static int
2876 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2877 {
2878 	struct sbuf *topo;
2879 	int err;
2880 
2881 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2882 
2883 	topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
2884 	if (topo == NULL)
2885 		return (ENOMEM);
2886 
2887 	sbuf_printf(topo, "<groups>\n");
2888 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2889 	sbuf_printf(topo, "</groups>\n");
2890 
2891 	if (err == 0) {
2892 		err = sbuf_finish(topo);
2893 	}
2894 	sbuf_delete(topo);
2895 	return (err);
2896 }
2897 
2898 #endif
2899 
2900 static int
2901 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2902 {
2903 	int error, new_val, period;
2904 
2905 	period = 1000000 / realstathz;
2906 	new_val = period * sched_slice;
2907 	error = sysctl_handle_int(oidp, &new_val, 0, req);
2908 	if (error != 0 || req->newptr == NULL)
2909 		return (error);
2910 	if (new_val <= 0)
2911 		return (EINVAL);
2912 	sched_slice = imax(1, (new_val + period / 2) / period);
2913 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
2914 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2915 	    realstathz);
2916 	return (0);
2917 }
2918 
2919 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2920 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2921     "Scheduler name");
2922 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2923     NULL, 0, sysctl_kern_quantum, "I",
2924     "Quantum for timeshare threads in microseconds");
2925 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2926     "Quantum for timeshare threads in stathz ticks");
2927 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2928     "Interactivity score threshold");
2929 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2930     &preempt_thresh, 0,
2931     "Maximal (lowest) priority for preemption");
2932 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2933     "Assign static kernel priorities to sleeping threads");
2934 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2935     "Number of times idle thread will spin waiting for new work");
2936 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2937     &sched_idlespinthresh, 0,
2938     "Threshold before we will permit idle thread spinning");
2939 #ifdef SMP
2940 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2941     "Number of hz ticks to keep thread affinity for");
2942 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2943     "Enables the long-term load balancer");
2944 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2945     &balance_interval, 0,
2946     "Average period in stathz ticks to run the long-term balancer");
2947 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2948     "Attempts to steal work from other cores before idling");
2949 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2950     "Minimum load on remote CPU before we'll steal");
2951 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2952     CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2953     "XML dump of detected CPU topology");
2954 #endif
2955 
2956 /* ps compat.  All cpu percentages from ULE are weighted. */
2957 static int ccpu = 0;
2958 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2959