xref: /freebsd/sys/kern/sched_ule.c (revision 3a92d97ff0f22d21608e1c19b83104c4937523b6)
1 /*-
2  * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice unmodified, this list of conditions, and the following
10  *    disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25  */
26 
27 /*
28  * This file implements the ULE scheduler.  ULE supports independent CPU
29  * run queues and fine grain locking.  It has superior interactive
30  * performance under load even on uni-processor systems.
31  *
32  * etymology:
33  *   ULE is the last three letters in schedule.  It owes its name to a
34  * generic user created for a scheduling system by Paul Mikesell at
35  * Isilon Systems and a general lack of creativity on the part of the author.
36  */
37 
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
40 
41 #include "opt_hwpmc_hooks.h"
42 #include "opt_sched.h"
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/kdb.h>
47 #include <sys/kernel.h>
48 #include <sys/ktr.h>
49 #include <sys/lock.h>
50 #include <sys/mutex.h>
51 #include <sys/proc.h>
52 #include <sys/resource.h>
53 #include <sys/resourcevar.h>
54 #include <sys/sched.h>
55 #include <sys/sdt.h>
56 #include <sys/smp.h>
57 #include <sys/sx.h>
58 #include <sys/sysctl.h>
59 #include <sys/sysproto.h>
60 #include <sys/turnstile.h>
61 #include <sys/umtx.h>
62 #include <sys/vmmeter.h>
63 #include <sys/cpuset.h>
64 #include <sys/sbuf.h>
65 
66 #ifdef HWPMC_HOOKS
67 #include <sys/pmckern.h>
68 #endif
69 
70 #ifdef KDTRACE_HOOKS
71 #include <sys/dtrace_bsd.h>
72 int				dtrace_vtime_active;
73 dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
74 #endif
75 
76 #include <machine/cpu.h>
77 #include <machine/smp.h>
78 
79 #define	KTR_ULE	0
80 
81 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
82 #define	TDQ_NAME_LEN	(sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
83 #define	TDQ_LOADNAME_LEN	(sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
84 
85 /*
86  * Thread scheduler specific section.  All fields are protected
87  * by the thread lock.
88  */
89 struct td_sched {
90 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
91 	short		ts_flags;	/* TSF_* flags. */
92 	u_char		ts_cpu;		/* CPU that we have affinity for. */
93 	int		ts_rltick;	/* Real last tick, for affinity. */
94 	int		ts_slice;	/* Ticks of slice remaining. */
95 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
96 	u_int		ts_runtime;	/* Number of ticks we were running */
97 	int		ts_ltick;	/* Last tick that we were running on */
98 	int		ts_ftick;	/* First tick that we were running on */
99 	int		ts_ticks;	/* Tick count */
100 #ifdef KTR
101 	char		ts_name[TS_NAME_LEN];
102 #endif
103 };
104 /* flags kept in ts_flags */
105 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
106 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
107 
108 static struct td_sched td_sched0;
109 
110 #define	THREAD_CAN_MIGRATE(td)	((td)->td_pinned == 0)
111 #define	THREAD_CAN_SCHED(td, cpu)	\
112     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
113 
114 /*
115  * Priority ranges used for interactive and non-interactive timeshare
116  * threads.  The timeshare priorities are split up into four ranges.
117  * The first range handles interactive threads.  The last three ranges
118  * (NHALF, x, and NHALF) handle non-interactive threads with the outer
119  * ranges supporting nice values.
120  */
121 #define	PRI_TIMESHARE_RANGE	(PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
122 #define	PRI_INTERACT_RANGE	((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
123 #define	PRI_BATCH_RANGE		(PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
124 
125 #define	PRI_MIN_INTERACT	PRI_MIN_TIMESHARE
126 #define	PRI_MAX_INTERACT	(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
127 #define	PRI_MIN_BATCH		(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
128 #define	PRI_MAX_BATCH		PRI_MAX_TIMESHARE
129 
130 /*
131  * Cpu percentage computation macros and defines.
132  *
133  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
134  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
135  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
136  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
137  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
138  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
139  */
140 #define	SCHED_TICK_SECS		10
141 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
142 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
143 #define	SCHED_TICK_SHIFT	10
144 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
145 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
146 
147 /*
148  * These macros determine priorities for non-interactive threads.  They are
149  * assigned a priority based on their recent cpu utilization as expressed
150  * by the ratio of ticks to the tick total.  NHALF priorities at the start
151  * and end of the MIN to MAX timeshare range are only reachable with negative
152  * or positive nice respectively.
153  *
154  * PRI_RANGE:	Priority range for utilization dependent priorities.
155  * PRI_NRESV:	Number of nice values.
156  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
157  * PRI_NICE:	Determines the part of the priority inherited from nice.
158  */
159 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
160 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
161 #define	SCHED_PRI_MIN		(PRI_MIN_BATCH + SCHED_PRI_NHALF)
162 #define	SCHED_PRI_MAX		(PRI_MAX_BATCH - SCHED_PRI_NHALF)
163 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
164 #define	SCHED_PRI_TICKS(ts)						\
165     (SCHED_TICK_HZ((ts)) /						\
166     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
167 #define	SCHED_PRI_NICE(nice)	(nice)
168 
169 /*
170  * These determine the interactivity of a process.  Interactivity differs from
171  * cpu utilization in that it expresses the voluntary time slept vs time ran
172  * while cpu utilization includes all time not running.  This more accurately
173  * models the intent of the thread.
174  *
175  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
176  *		before throttling back.
177  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
178  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
179  * INTERACT_THRESH:	Threshold for placement on the current runq.
180  */
181 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
182 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
183 #define	SCHED_INTERACT_MAX	(100)
184 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
185 #define	SCHED_INTERACT_THRESH	(30)
186 
187 /*
188  * These parameters determine the slice behavior for batch work.
189  */
190 #define	SCHED_SLICE_DEFAULT_DIVISOR	10	/* ~94 ms, 12 stathz ticks. */
191 #define	SCHED_SLICE_MIN_DIVISOR		6	/* DEFAULT/MIN = ~16 ms. */
192 
193 /* Flags kept in td_flags. */
194 #define	TDF_SLICEEND	TDF_SCHED2	/* Thread time slice is over. */
195 
196 /*
197  * tickincr:		Converts a stathz tick into a hz domain scaled by
198  *			the shift factor.  Without the shift the error rate
199  *			due to rounding would be unacceptably high.
200  * realstathz:		stathz is sometimes 0 and run off of hz.
201  * sched_slice:		Runtime of each thread before rescheduling.
202  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
203  */
204 static int sched_interact = SCHED_INTERACT_THRESH;
205 static int tickincr = 8 << SCHED_TICK_SHIFT;
206 static int realstathz = 127;	/* reset during boot. */
207 static int sched_slice = 10;	/* reset during boot. */
208 static int sched_slice_min = 1;	/* reset during boot. */
209 #ifdef PREEMPTION
210 #ifdef FULL_PREEMPTION
211 static int preempt_thresh = PRI_MAX_IDLE;
212 #else
213 static int preempt_thresh = PRI_MIN_KERN;
214 #endif
215 #else
216 static int preempt_thresh = 0;
217 #endif
218 static int static_boost = PRI_MIN_BATCH;
219 static int sched_idlespins = 10000;
220 static int sched_idlespinthresh = -1;
221 
222 /*
223  * tdq - per processor runqs and statistics.  All fields are protected by the
224  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
225  * locking in sched_pickcpu();
226  */
227 struct tdq {
228 	/*
229 	 * Ordered to improve efficiency of cpu_search() and switch().
230 	 * tdq_lock is padded to avoid false sharing with tdq_load and
231 	 * tdq_cpu_idle.
232 	 */
233 	struct mtx_padalign tdq_lock;		/* run queue lock. */
234 	struct cpu_group *tdq_cg;		/* Pointer to cpu topology. */
235 	volatile int	tdq_load;		/* Aggregate load. */
236 	volatile int	tdq_cpu_idle;		/* cpu_idle() is active. */
237 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
238 	int		tdq_transferable;	/* Transferable thread count. */
239 	short		tdq_switchcnt;		/* Switches this tick. */
240 	short		tdq_oldswitchcnt;	/* Switches last tick. */
241 	u_char		tdq_lowpri;		/* Lowest priority thread. */
242 	u_char		tdq_ipipending;		/* IPI pending. */
243 	u_char		tdq_idx;		/* Current insert index. */
244 	u_char		tdq_ridx;		/* Current removal index. */
245 	struct runq	tdq_realtime;		/* real-time run queue. */
246 	struct runq	tdq_timeshare;		/* timeshare run queue. */
247 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
248 	char		tdq_name[TDQ_NAME_LEN];
249 #ifdef KTR
250 	char		tdq_loadname[TDQ_LOADNAME_LEN];
251 #endif
252 } __aligned(64);
253 
254 /* Idle thread states and config. */
255 #define	TDQ_RUNNING	1
256 #define	TDQ_IDLE	2
257 
258 #ifdef SMP
259 struct cpu_group *cpu_top;		/* CPU topology */
260 
261 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
262 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
263 
264 /*
265  * Run-time tunables.
266  */
267 static int rebalance = 1;
268 static int balance_interval = 128;	/* Default set in sched_initticks(). */
269 static int affinity;
270 static int steal_idle = 1;
271 static int steal_thresh = 2;
272 
273 /*
274  * One thread queue per processor.
275  */
276 static struct tdq	tdq_cpu[MAXCPU];
277 static struct tdq	*balance_tdq;
278 static int balance_ticks;
279 static DPCPU_DEFINE(uint32_t, randomval);
280 
281 #define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
282 #define	TDQ_CPU(x)	(&tdq_cpu[(x)])
283 #define	TDQ_ID(x)	((int)((x) - tdq_cpu))
284 #else	/* !SMP */
285 static struct tdq	tdq_cpu;
286 
287 #define	TDQ_ID(x)	(0)
288 #define	TDQ_SELF()	(&tdq_cpu)
289 #define	TDQ_CPU(x)	(&tdq_cpu)
290 #endif
291 
292 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
293 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
294 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
295 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
296 #define	TDQ_LOCKPTR(t)		((struct mtx *)(&(t)->tdq_lock))
297 
298 static void sched_priority(struct thread *);
299 static void sched_thread_priority(struct thread *, u_char);
300 static int sched_interact_score(struct thread *);
301 static void sched_interact_update(struct thread *);
302 static void sched_interact_fork(struct thread *);
303 static void sched_pctcpu_update(struct td_sched *, int);
304 
305 /* Operations on per processor queues */
306 static struct thread *tdq_choose(struct tdq *);
307 static void tdq_setup(struct tdq *);
308 static void tdq_load_add(struct tdq *, struct thread *);
309 static void tdq_load_rem(struct tdq *, struct thread *);
310 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
311 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
312 static inline int sched_shouldpreempt(int, int, int);
313 void tdq_print(int cpu);
314 static void runq_print(struct runq *rq);
315 static void tdq_add(struct tdq *, struct thread *, int);
316 #ifdef SMP
317 static int tdq_move(struct tdq *, struct tdq *);
318 static int tdq_idled(struct tdq *);
319 static void tdq_notify(struct tdq *, struct thread *);
320 static struct thread *tdq_steal(struct tdq *, int);
321 static struct thread *runq_steal(struct runq *, int);
322 static int sched_pickcpu(struct thread *, int);
323 static void sched_balance(void);
324 static int sched_balance_pair(struct tdq *, struct tdq *);
325 static inline struct tdq *sched_setcpu(struct thread *, int, int);
326 static inline void thread_unblock_switch(struct thread *, struct mtx *);
327 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
328 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
329 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
330     struct cpu_group *cg, int indent);
331 #endif
332 
333 static void sched_setup(void *dummy);
334 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
335 
336 static void sched_initticks(void *dummy);
337 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
338     NULL);
339 
340 SDT_PROVIDER_DEFINE(sched);
341 
342 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
343     "struct proc *", "uint8_t");
344 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
345     "struct proc *", "void *");
346 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
347     "struct proc *", "void *", "int");
348 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
349     "struct proc *", "uint8_t", "struct thread *");
350 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
351 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
352     "struct proc *");
353 SDT_PROBE_DEFINE(sched, , , on__cpu);
354 SDT_PROBE_DEFINE(sched, , , remain__cpu);
355 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
356     "struct proc *");
357 
358 /*
359  * Print the threads waiting on a run-queue.
360  */
361 static void
362 runq_print(struct runq *rq)
363 {
364 	struct rqhead *rqh;
365 	struct thread *td;
366 	int pri;
367 	int j;
368 	int i;
369 
370 	for (i = 0; i < RQB_LEN; i++) {
371 		printf("\t\trunq bits %d 0x%zx\n",
372 		    i, rq->rq_status.rqb_bits[i]);
373 		for (j = 0; j < RQB_BPW; j++)
374 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
375 				pri = j + (i << RQB_L2BPW);
376 				rqh = &rq->rq_queues[pri];
377 				TAILQ_FOREACH(td, rqh, td_runq) {
378 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
379 					    td, td->td_name, td->td_priority,
380 					    td->td_rqindex, pri);
381 				}
382 			}
383 	}
384 }
385 
386 /*
387  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
388  */
389 void
390 tdq_print(int cpu)
391 {
392 	struct tdq *tdq;
393 
394 	tdq = TDQ_CPU(cpu);
395 
396 	printf("tdq %d:\n", TDQ_ID(tdq));
397 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
398 	printf("\tLock name:      %s\n", tdq->tdq_name);
399 	printf("\tload:           %d\n", tdq->tdq_load);
400 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
401 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
402 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
403 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
404 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
405 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
406 	printf("\trealtime runq:\n");
407 	runq_print(&tdq->tdq_realtime);
408 	printf("\ttimeshare runq:\n");
409 	runq_print(&tdq->tdq_timeshare);
410 	printf("\tidle runq:\n");
411 	runq_print(&tdq->tdq_idle);
412 }
413 
414 static inline int
415 sched_shouldpreempt(int pri, int cpri, int remote)
416 {
417 	/*
418 	 * If the new priority is not better than the current priority there is
419 	 * nothing to do.
420 	 */
421 	if (pri >= cpri)
422 		return (0);
423 	/*
424 	 * Always preempt idle.
425 	 */
426 	if (cpri >= PRI_MIN_IDLE)
427 		return (1);
428 	/*
429 	 * If preemption is disabled don't preempt others.
430 	 */
431 	if (preempt_thresh == 0)
432 		return (0);
433 	/*
434 	 * Preempt if we exceed the threshold.
435 	 */
436 	if (pri <= preempt_thresh)
437 		return (1);
438 	/*
439 	 * If we're interactive or better and there is non-interactive
440 	 * or worse running preempt only remote processors.
441 	 */
442 	if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
443 		return (1);
444 	return (0);
445 }
446 
447 /*
448  * Add a thread to the actual run-queue.  Keeps transferable counts up to
449  * date with what is actually on the run-queue.  Selects the correct
450  * queue position for timeshare threads.
451  */
452 static __inline void
453 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
454 {
455 	struct td_sched *ts;
456 	u_char pri;
457 
458 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
459 	THREAD_LOCK_ASSERT(td, MA_OWNED);
460 
461 	pri = td->td_priority;
462 	ts = td->td_sched;
463 	TD_SET_RUNQ(td);
464 	if (THREAD_CAN_MIGRATE(td)) {
465 		tdq->tdq_transferable++;
466 		ts->ts_flags |= TSF_XFERABLE;
467 	}
468 	if (pri < PRI_MIN_BATCH) {
469 		ts->ts_runq = &tdq->tdq_realtime;
470 	} else if (pri <= PRI_MAX_BATCH) {
471 		ts->ts_runq = &tdq->tdq_timeshare;
472 		KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
473 			("Invalid priority %d on timeshare runq", pri));
474 		/*
475 		 * This queue contains only priorities between MIN and MAX
476 		 * realtime.  Use the whole queue to represent these values.
477 		 */
478 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
479 			pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
480 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
481 			/*
482 			 * This effectively shortens the queue by one so we
483 			 * can have a one slot difference between idx and
484 			 * ridx while we wait for threads to drain.
485 			 */
486 			if (tdq->tdq_ridx != tdq->tdq_idx &&
487 			    pri == tdq->tdq_ridx)
488 				pri = (unsigned char)(pri - 1) % RQ_NQS;
489 		} else
490 			pri = tdq->tdq_ridx;
491 		runq_add_pri(ts->ts_runq, td, pri, flags);
492 		return;
493 	} else
494 		ts->ts_runq = &tdq->tdq_idle;
495 	runq_add(ts->ts_runq, td, flags);
496 }
497 
498 /*
499  * Remove a thread from a run-queue.  This typically happens when a thread
500  * is selected to run.  Running threads are not on the queue and the
501  * transferable count does not reflect them.
502  */
503 static __inline void
504 tdq_runq_rem(struct tdq *tdq, struct thread *td)
505 {
506 	struct td_sched *ts;
507 
508 	ts = td->td_sched;
509 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
510 	KASSERT(ts->ts_runq != NULL,
511 	    ("tdq_runq_remove: thread %p null ts_runq", td));
512 	if (ts->ts_flags & TSF_XFERABLE) {
513 		tdq->tdq_transferable--;
514 		ts->ts_flags &= ~TSF_XFERABLE;
515 	}
516 	if (ts->ts_runq == &tdq->tdq_timeshare) {
517 		if (tdq->tdq_idx != tdq->tdq_ridx)
518 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
519 		else
520 			runq_remove_idx(ts->ts_runq, td, NULL);
521 	} else
522 		runq_remove(ts->ts_runq, td);
523 }
524 
525 /*
526  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
527  * for this thread to the referenced thread queue.
528  */
529 static void
530 tdq_load_add(struct tdq *tdq, struct thread *td)
531 {
532 
533 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
534 	THREAD_LOCK_ASSERT(td, MA_OWNED);
535 
536 	tdq->tdq_load++;
537 	if ((td->td_flags & TDF_NOLOAD) == 0)
538 		tdq->tdq_sysload++;
539 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
540 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
541 }
542 
543 /*
544  * Remove the load from a thread that is transitioning to a sleep state or
545  * exiting.
546  */
547 static void
548 tdq_load_rem(struct tdq *tdq, struct thread *td)
549 {
550 
551 	THREAD_LOCK_ASSERT(td, MA_OWNED);
552 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
553 	KASSERT(tdq->tdq_load != 0,
554 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
555 
556 	tdq->tdq_load--;
557 	if ((td->td_flags & TDF_NOLOAD) == 0)
558 		tdq->tdq_sysload--;
559 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
560 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
561 }
562 
563 /*
564  * Bound timeshare latency by decreasing slice size as load increases.  We
565  * consider the maximum latency as the sum of the threads waiting to run
566  * aside from curthread and target no more than sched_slice latency but
567  * no less than sched_slice_min runtime.
568  */
569 static inline int
570 tdq_slice(struct tdq *tdq)
571 {
572 	int load;
573 
574 	/*
575 	 * It is safe to use sys_load here because this is called from
576 	 * contexts where timeshare threads are running and so there
577 	 * cannot be higher priority load in the system.
578 	 */
579 	load = tdq->tdq_sysload - 1;
580 	if (load >= SCHED_SLICE_MIN_DIVISOR)
581 		return (sched_slice_min);
582 	if (load <= 1)
583 		return (sched_slice);
584 	return (sched_slice / load);
585 }
586 
587 /*
588  * Set lowpri to its exact value by searching the run-queue and
589  * evaluating curthread.  curthread may be passed as an optimization.
590  */
591 static void
592 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
593 {
594 	struct thread *td;
595 
596 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
597 	if (ctd == NULL)
598 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
599 	td = tdq_choose(tdq);
600 	if (td == NULL || td->td_priority > ctd->td_priority)
601 		tdq->tdq_lowpri = ctd->td_priority;
602 	else
603 		tdq->tdq_lowpri = td->td_priority;
604 }
605 
606 #ifdef SMP
607 struct cpu_search {
608 	cpuset_t cs_mask;
609 	u_int	cs_prefer;
610 	int	cs_pri;		/* Min priority for low. */
611 	int	cs_limit;	/* Max load for low, min load for high. */
612 	int	cs_cpu;
613 	int	cs_load;
614 };
615 
616 #define	CPU_SEARCH_LOWEST	0x1
617 #define	CPU_SEARCH_HIGHEST	0x2
618 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
619 
620 #define	CPUSET_FOREACH(cpu, mask)				\
621 	for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++)		\
622 		if (CPU_ISSET(cpu, &mask))
623 
624 static __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low,
625     struct cpu_search *high, const int match);
626 int cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low);
627 int cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high);
628 int cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
629     struct cpu_search *high);
630 
631 /*
632  * Search the tree of cpu_groups for the lowest or highest loaded cpu
633  * according to the match argument.  This routine actually compares the
634  * load on all paths through the tree and finds the least loaded cpu on
635  * the least loaded path, which may differ from the least loaded cpu in
636  * the system.  This balances work among caches and busses.
637  *
638  * This inline is instantiated in three forms below using constants for the
639  * match argument.  It is reduced to the minimum set for each case.  It is
640  * also recursive to the depth of the tree.
641  */
642 static __inline int
643 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
644     struct cpu_search *high, const int match)
645 {
646 	struct cpu_search lgroup;
647 	struct cpu_search hgroup;
648 	cpuset_t cpumask;
649 	struct cpu_group *child;
650 	struct tdq *tdq;
651 	int cpu, i, hload, lload, load, total, rnd, *rndptr;
652 
653 	total = 0;
654 	cpumask = cg->cg_mask;
655 	if (match & CPU_SEARCH_LOWEST) {
656 		lload = INT_MAX;
657 		lgroup = *low;
658 	}
659 	if (match & CPU_SEARCH_HIGHEST) {
660 		hload = INT_MIN;
661 		hgroup = *high;
662 	}
663 
664 	/* Iterate through the child CPU groups and then remaining CPUs. */
665 	for (i = cg->cg_children, cpu = mp_maxid; ; ) {
666 		if (i == 0) {
667 #ifdef HAVE_INLINE_FFSL
668 			cpu = CPU_FFS(&cpumask) - 1;
669 #else
670 			while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
671 				cpu--;
672 #endif
673 			if (cpu < 0)
674 				break;
675 			child = NULL;
676 		} else
677 			child = &cg->cg_child[i - 1];
678 
679 		if (match & CPU_SEARCH_LOWEST)
680 			lgroup.cs_cpu = -1;
681 		if (match & CPU_SEARCH_HIGHEST)
682 			hgroup.cs_cpu = -1;
683 		if (child) {			/* Handle child CPU group. */
684 			CPU_NAND(&cpumask, &child->cg_mask);
685 			switch (match) {
686 			case CPU_SEARCH_LOWEST:
687 				load = cpu_search_lowest(child, &lgroup);
688 				break;
689 			case CPU_SEARCH_HIGHEST:
690 				load = cpu_search_highest(child, &hgroup);
691 				break;
692 			case CPU_SEARCH_BOTH:
693 				load = cpu_search_both(child, &lgroup, &hgroup);
694 				break;
695 			}
696 		} else {			/* Handle child CPU. */
697 			CPU_CLR(cpu, &cpumask);
698 			tdq = TDQ_CPU(cpu);
699 			load = tdq->tdq_load * 256;
700 			rndptr = DPCPU_PTR(randomval);
701 			rnd = (*rndptr = *rndptr * 69069 + 5) >> 26;
702 			if (match & CPU_SEARCH_LOWEST) {
703 				if (cpu == low->cs_prefer)
704 					load -= 64;
705 				/* If that CPU is allowed and get data. */
706 				if (tdq->tdq_lowpri > lgroup.cs_pri &&
707 				    tdq->tdq_load <= lgroup.cs_limit &&
708 				    CPU_ISSET(cpu, &lgroup.cs_mask)) {
709 					lgroup.cs_cpu = cpu;
710 					lgroup.cs_load = load - rnd;
711 				}
712 			}
713 			if (match & CPU_SEARCH_HIGHEST)
714 				if (tdq->tdq_load >= hgroup.cs_limit &&
715 				    tdq->tdq_transferable &&
716 				    CPU_ISSET(cpu, &hgroup.cs_mask)) {
717 					hgroup.cs_cpu = cpu;
718 					hgroup.cs_load = load - rnd;
719 				}
720 		}
721 		total += load;
722 
723 		/* We have info about child item. Compare it. */
724 		if (match & CPU_SEARCH_LOWEST) {
725 			if (lgroup.cs_cpu >= 0 &&
726 			    (load < lload ||
727 			     (load == lload && lgroup.cs_load < low->cs_load))) {
728 				lload = load;
729 				low->cs_cpu = lgroup.cs_cpu;
730 				low->cs_load = lgroup.cs_load;
731 			}
732 		}
733 		if (match & CPU_SEARCH_HIGHEST)
734 			if (hgroup.cs_cpu >= 0 &&
735 			    (load > hload ||
736 			     (load == hload && hgroup.cs_load > high->cs_load))) {
737 				hload = load;
738 				high->cs_cpu = hgroup.cs_cpu;
739 				high->cs_load = hgroup.cs_load;
740 			}
741 		if (child) {
742 			i--;
743 			if (i == 0 && CPU_EMPTY(&cpumask))
744 				break;
745 		}
746 #ifndef HAVE_INLINE_FFSL
747 		else
748 			cpu--;
749 #endif
750 	}
751 	return (total);
752 }
753 
754 /*
755  * cpu_search instantiations must pass constants to maintain the inline
756  * optimization.
757  */
758 int
759 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
760 {
761 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
762 }
763 
764 int
765 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
766 {
767 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
768 }
769 
770 int
771 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
772     struct cpu_search *high)
773 {
774 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
775 }
776 
777 /*
778  * Find the cpu with the least load via the least loaded path that has a
779  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
780  * acceptable.
781  */
782 static inline int
783 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
784     int prefer)
785 {
786 	struct cpu_search low;
787 
788 	low.cs_cpu = -1;
789 	low.cs_prefer = prefer;
790 	low.cs_mask = mask;
791 	low.cs_pri = pri;
792 	low.cs_limit = maxload;
793 	cpu_search_lowest(cg, &low);
794 	return low.cs_cpu;
795 }
796 
797 /*
798  * Find the cpu with the highest load via the highest loaded path.
799  */
800 static inline int
801 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
802 {
803 	struct cpu_search high;
804 
805 	high.cs_cpu = -1;
806 	high.cs_mask = mask;
807 	high.cs_limit = minload;
808 	cpu_search_highest(cg, &high);
809 	return high.cs_cpu;
810 }
811 
812 static void
813 sched_balance_group(struct cpu_group *cg)
814 {
815 	cpuset_t hmask, lmask;
816 	int high, low, anylow;
817 
818 	CPU_FILL(&hmask);
819 	for (;;) {
820 		high = sched_highest(cg, hmask, 1);
821 		/* Stop if there is no more CPU with transferrable threads. */
822 		if (high == -1)
823 			break;
824 		CPU_CLR(high, &hmask);
825 		CPU_COPY(&hmask, &lmask);
826 		/* Stop if there is no more CPU left for low. */
827 		if (CPU_EMPTY(&lmask))
828 			break;
829 		anylow = 1;
830 nextlow:
831 		low = sched_lowest(cg, lmask, -1,
832 		    TDQ_CPU(high)->tdq_load - 1, high);
833 		/* Stop if we looked well and found no less loaded CPU. */
834 		if (anylow && low == -1)
835 			break;
836 		/* Go to next high if we found no less loaded CPU. */
837 		if (low == -1)
838 			continue;
839 		/* Transfer thread from high to low. */
840 		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
841 			/* CPU that got thread can no longer be a donor. */
842 			CPU_CLR(low, &hmask);
843 		} else {
844 			/*
845 			 * If failed, then there is no threads on high
846 			 * that can run on this low. Drop low from low
847 			 * mask and look for different one.
848 			 */
849 			CPU_CLR(low, &lmask);
850 			anylow = 0;
851 			goto nextlow;
852 		}
853 	}
854 }
855 
856 static void
857 sched_balance(void)
858 {
859 	struct tdq *tdq;
860 
861 	/*
862 	 * Select a random time between .5 * balance_interval and
863 	 * 1.5 * balance_interval.
864 	 */
865 	balance_ticks = max(balance_interval / 2, 1);
866 	balance_ticks += random() % balance_interval;
867 	if (smp_started == 0 || rebalance == 0)
868 		return;
869 	tdq = TDQ_SELF();
870 	TDQ_UNLOCK(tdq);
871 	sched_balance_group(cpu_top);
872 	TDQ_LOCK(tdq);
873 }
874 
875 /*
876  * Lock two thread queues using their address to maintain lock order.
877  */
878 static void
879 tdq_lock_pair(struct tdq *one, struct tdq *two)
880 {
881 	if (one < two) {
882 		TDQ_LOCK(one);
883 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
884 	} else {
885 		TDQ_LOCK(two);
886 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
887 	}
888 }
889 
890 /*
891  * Unlock two thread queues.  Order is not important here.
892  */
893 static void
894 tdq_unlock_pair(struct tdq *one, struct tdq *two)
895 {
896 	TDQ_UNLOCK(one);
897 	TDQ_UNLOCK(two);
898 }
899 
900 /*
901  * Transfer load between two imbalanced thread queues.
902  */
903 static int
904 sched_balance_pair(struct tdq *high, struct tdq *low)
905 {
906 	int moved;
907 	int cpu;
908 
909 	tdq_lock_pair(high, low);
910 	moved = 0;
911 	/*
912 	 * Determine what the imbalance is and then adjust that to how many
913 	 * threads we actually have to give up (transferable).
914 	 */
915 	if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
916 	    (moved = tdq_move(high, low)) > 0) {
917 		/*
918 		 * In case the target isn't the current cpu IPI it to force a
919 		 * reschedule with the new workload.
920 		 */
921 		cpu = TDQ_ID(low);
922 		if (cpu != PCPU_GET(cpuid))
923 			ipi_cpu(cpu, IPI_PREEMPT);
924 	}
925 	tdq_unlock_pair(high, low);
926 	return (moved);
927 }
928 
929 /*
930  * Move a thread from one thread queue to another.
931  */
932 static int
933 tdq_move(struct tdq *from, struct tdq *to)
934 {
935 	struct td_sched *ts;
936 	struct thread *td;
937 	struct tdq *tdq;
938 	int cpu;
939 
940 	TDQ_LOCK_ASSERT(from, MA_OWNED);
941 	TDQ_LOCK_ASSERT(to, MA_OWNED);
942 
943 	tdq = from;
944 	cpu = TDQ_ID(to);
945 	td = tdq_steal(tdq, cpu);
946 	if (td == NULL)
947 		return (0);
948 	ts = td->td_sched;
949 	/*
950 	 * Although the run queue is locked the thread may be blocked.  Lock
951 	 * it to clear this and acquire the run-queue lock.
952 	 */
953 	thread_lock(td);
954 	/* Drop recursive lock on from acquired via thread_lock(). */
955 	TDQ_UNLOCK(from);
956 	sched_rem(td);
957 	ts->ts_cpu = cpu;
958 	td->td_lock = TDQ_LOCKPTR(to);
959 	tdq_add(to, td, SRQ_YIELDING);
960 	return (1);
961 }
962 
963 /*
964  * This tdq has idled.  Try to steal a thread from another cpu and switch
965  * to it.
966  */
967 static int
968 tdq_idled(struct tdq *tdq)
969 {
970 	struct cpu_group *cg;
971 	struct tdq *steal;
972 	cpuset_t mask;
973 	int thresh;
974 	int cpu;
975 
976 	if (smp_started == 0 || steal_idle == 0)
977 		return (1);
978 	CPU_FILL(&mask);
979 	CPU_CLR(PCPU_GET(cpuid), &mask);
980 	/* We don't want to be preempted while we're iterating. */
981 	spinlock_enter();
982 	for (cg = tdq->tdq_cg; cg != NULL; ) {
983 		if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
984 			thresh = steal_thresh;
985 		else
986 			thresh = 1;
987 		cpu = sched_highest(cg, mask, thresh);
988 		if (cpu == -1) {
989 			cg = cg->cg_parent;
990 			continue;
991 		}
992 		steal = TDQ_CPU(cpu);
993 		CPU_CLR(cpu, &mask);
994 		tdq_lock_pair(tdq, steal);
995 		if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
996 			tdq_unlock_pair(tdq, steal);
997 			continue;
998 		}
999 		/*
1000 		 * If a thread was added while interrupts were disabled don't
1001 		 * steal one here.  If we fail to acquire one due to affinity
1002 		 * restrictions loop again with this cpu removed from the
1003 		 * set.
1004 		 */
1005 		if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
1006 			tdq_unlock_pair(tdq, steal);
1007 			continue;
1008 		}
1009 		spinlock_exit();
1010 		TDQ_UNLOCK(steal);
1011 		mi_switch(SW_VOL | SWT_IDLE, NULL);
1012 		thread_unlock(curthread);
1013 
1014 		return (0);
1015 	}
1016 	spinlock_exit();
1017 	return (1);
1018 }
1019 
1020 /*
1021  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
1022  */
1023 static void
1024 tdq_notify(struct tdq *tdq, struct thread *td)
1025 {
1026 	struct thread *ctd;
1027 	int pri;
1028 	int cpu;
1029 
1030 	if (tdq->tdq_ipipending)
1031 		return;
1032 	cpu = td->td_sched->ts_cpu;
1033 	pri = td->td_priority;
1034 	ctd = pcpu_find(cpu)->pc_curthread;
1035 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1036 		return;
1037 	if (TD_IS_IDLETHREAD(ctd)) {
1038 		/*
1039 		 * If the MD code has an idle wakeup routine try that before
1040 		 * falling back to IPI.
1041 		 */
1042 		if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1043 			return;
1044 	}
1045 	tdq->tdq_ipipending = 1;
1046 	ipi_cpu(cpu, IPI_PREEMPT);
1047 }
1048 
1049 /*
1050  * Steals load from a timeshare queue.  Honors the rotating queue head
1051  * index.
1052  */
1053 static struct thread *
1054 runq_steal_from(struct runq *rq, int cpu, u_char start)
1055 {
1056 	struct rqbits *rqb;
1057 	struct rqhead *rqh;
1058 	struct thread *td, *first;
1059 	int bit;
1060 	int i;
1061 
1062 	rqb = &rq->rq_status;
1063 	bit = start & (RQB_BPW -1);
1064 	first = NULL;
1065 again:
1066 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1067 		if (rqb->rqb_bits[i] == 0)
1068 			continue;
1069 		if (bit == 0)
1070 			bit = RQB_FFS(rqb->rqb_bits[i]);
1071 		for (; bit < RQB_BPW; bit++) {
1072 			if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1073 				continue;
1074 			rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1075 			TAILQ_FOREACH(td, rqh, td_runq) {
1076 				if (first && THREAD_CAN_MIGRATE(td) &&
1077 				    THREAD_CAN_SCHED(td, cpu))
1078 					return (td);
1079 				first = td;
1080 			}
1081 		}
1082 	}
1083 	if (start != 0) {
1084 		start = 0;
1085 		goto again;
1086 	}
1087 
1088 	if (first && THREAD_CAN_MIGRATE(first) &&
1089 	    THREAD_CAN_SCHED(first, cpu))
1090 		return (first);
1091 	return (NULL);
1092 }
1093 
1094 /*
1095  * Steals load from a standard linear queue.
1096  */
1097 static struct thread *
1098 runq_steal(struct runq *rq, int cpu)
1099 {
1100 	struct rqhead *rqh;
1101 	struct rqbits *rqb;
1102 	struct thread *td;
1103 	int word;
1104 	int bit;
1105 
1106 	rqb = &rq->rq_status;
1107 	for (word = 0; word < RQB_LEN; word++) {
1108 		if (rqb->rqb_bits[word] == 0)
1109 			continue;
1110 		for (bit = 0; bit < RQB_BPW; bit++) {
1111 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1112 				continue;
1113 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1114 			TAILQ_FOREACH(td, rqh, td_runq)
1115 				if (THREAD_CAN_MIGRATE(td) &&
1116 				    THREAD_CAN_SCHED(td, cpu))
1117 					return (td);
1118 		}
1119 	}
1120 	return (NULL);
1121 }
1122 
1123 /*
1124  * Attempt to steal a thread in priority order from a thread queue.
1125  */
1126 static struct thread *
1127 tdq_steal(struct tdq *tdq, int cpu)
1128 {
1129 	struct thread *td;
1130 
1131 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1132 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1133 		return (td);
1134 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1135 	    cpu, tdq->tdq_ridx)) != NULL)
1136 		return (td);
1137 	return (runq_steal(&tdq->tdq_idle, cpu));
1138 }
1139 
1140 /*
1141  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1142  * current lock and returns with the assigned queue locked.
1143  */
1144 static inline struct tdq *
1145 sched_setcpu(struct thread *td, int cpu, int flags)
1146 {
1147 
1148 	struct tdq *tdq;
1149 
1150 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1151 	tdq = TDQ_CPU(cpu);
1152 	td->td_sched->ts_cpu = cpu;
1153 	/*
1154 	 * If the lock matches just return the queue.
1155 	 */
1156 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1157 		return (tdq);
1158 #ifdef notyet
1159 	/*
1160 	 * If the thread isn't running its lockptr is a
1161 	 * turnstile or a sleepqueue.  We can just lock_set without
1162 	 * blocking.
1163 	 */
1164 	if (TD_CAN_RUN(td)) {
1165 		TDQ_LOCK(tdq);
1166 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1167 		return (tdq);
1168 	}
1169 #endif
1170 	/*
1171 	 * The hard case, migration, we need to block the thread first to
1172 	 * prevent order reversals with other cpus locks.
1173 	 */
1174 	spinlock_enter();
1175 	thread_lock_block(td);
1176 	TDQ_LOCK(tdq);
1177 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1178 	spinlock_exit();
1179 	return (tdq);
1180 }
1181 
1182 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1183 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1184 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1185 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1186 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1187 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1188 
1189 static int
1190 sched_pickcpu(struct thread *td, int flags)
1191 {
1192 	struct cpu_group *cg, *ccg;
1193 	struct td_sched *ts;
1194 	struct tdq *tdq;
1195 	cpuset_t mask;
1196 	int cpu, pri, self;
1197 
1198 	self = PCPU_GET(cpuid);
1199 	ts = td->td_sched;
1200 	if (smp_started == 0)
1201 		return (self);
1202 	/*
1203 	 * Don't migrate a running thread from sched_switch().
1204 	 */
1205 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1206 		return (ts->ts_cpu);
1207 	/*
1208 	 * Prefer to run interrupt threads on the processors that generate
1209 	 * the interrupt.
1210 	 */
1211 	pri = td->td_priority;
1212 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1213 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1214 		SCHED_STAT_INC(pickcpu_intrbind);
1215 		ts->ts_cpu = self;
1216 		if (TDQ_CPU(self)->tdq_lowpri > pri) {
1217 			SCHED_STAT_INC(pickcpu_affinity);
1218 			return (ts->ts_cpu);
1219 		}
1220 	}
1221 	/*
1222 	 * If the thread can run on the last cpu and the affinity has not
1223 	 * expired or it is idle run it there.
1224 	 */
1225 	tdq = TDQ_CPU(ts->ts_cpu);
1226 	cg = tdq->tdq_cg;
1227 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1228 	    tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1229 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1230 		if (cg->cg_flags & CG_FLAG_THREAD) {
1231 			CPUSET_FOREACH(cpu, cg->cg_mask) {
1232 				if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1233 					break;
1234 			}
1235 		} else
1236 			cpu = INT_MAX;
1237 		if (cpu > mp_maxid) {
1238 			SCHED_STAT_INC(pickcpu_idle_affinity);
1239 			return (ts->ts_cpu);
1240 		}
1241 	}
1242 	/*
1243 	 * Search for the last level cache CPU group in the tree.
1244 	 * Skip caches with expired affinity time and SMT groups.
1245 	 * Affinity to higher level caches will be handled less aggressively.
1246 	 */
1247 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1248 		if (cg->cg_flags & CG_FLAG_THREAD)
1249 			continue;
1250 		if (!SCHED_AFFINITY(ts, cg->cg_level))
1251 			continue;
1252 		ccg = cg;
1253 	}
1254 	if (ccg != NULL)
1255 		cg = ccg;
1256 	cpu = -1;
1257 	/* Search the group for the less loaded idle CPU we can run now. */
1258 	mask = td->td_cpuset->cs_mask;
1259 	if (cg != NULL && cg != cpu_top &&
1260 	    CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1261 		cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1262 		    INT_MAX, ts->ts_cpu);
1263 	/* Search globally for the less loaded CPU we can run now. */
1264 	if (cpu == -1)
1265 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1266 	/* Search globally for the less loaded CPU. */
1267 	if (cpu == -1)
1268 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1269 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1270 	/*
1271 	 * Compare the lowest loaded cpu to current cpu.
1272 	 */
1273 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1274 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1275 	    TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1276 		SCHED_STAT_INC(pickcpu_local);
1277 		cpu = self;
1278 	} else
1279 		SCHED_STAT_INC(pickcpu_lowest);
1280 	if (cpu != ts->ts_cpu)
1281 		SCHED_STAT_INC(pickcpu_migration);
1282 	return (cpu);
1283 }
1284 #endif
1285 
1286 /*
1287  * Pick the highest priority task we have and return it.
1288  */
1289 static struct thread *
1290 tdq_choose(struct tdq *tdq)
1291 {
1292 	struct thread *td;
1293 
1294 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1295 	td = runq_choose(&tdq->tdq_realtime);
1296 	if (td != NULL)
1297 		return (td);
1298 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1299 	if (td != NULL) {
1300 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1301 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1302 		    td->td_priority));
1303 		return (td);
1304 	}
1305 	td = runq_choose(&tdq->tdq_idle);
1306 	if (td != NULL) {
1307 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1308 		    ("tdq_choose: Invalid priority on idle queue %d",
1309 		    td->td_priority));
1310 		return (td);
1311 	}
1312 
1313 	return (NULL);
1314 }
1315 
1316 /*
1317  * Initialize a thread queue.
1318  */
1319 static void
1320 tdq_setup(struct tdq *tdq)
1321 {
1322 
1323 	if (bootverbose)
1324 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1325 	runq_init(&tdq->tdq_realtime);
1326 	runq_init(&tdq->tdq_timeshare);
1327 	runq_init(&tdq->tdq_idle);
1328 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1329 	    "sched lock %d", (int)TDQ_ID(tdq));
1330 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1331 	    MTX_SPIN | MTX_RECURSE);
1332 #ifdef KTR
1333 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1334 	    "CPU %d load", (int)TDQ_ID(tdq));
1335 #endif
1336 }
1337 
1338 #ifdef SMP
1339 static void
1340 sched_setup_smp(void)
1341 {
1342 	struct tdq *tdq;
1343 	int i;
1344 
1345 	cpu_top = smp_topo();
1346 	CPU_FOREACH(i) {
1347 		tdq = TDQ_CPU(i);
1348 		tdq_setup(tdq);
1349 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1350 		if (tdq->tdq_cg == NULL)
1351 			panic("Can't find cpu group for %d\n", i);
1352 	}
1353 	balance_tdq = TDQ_SELF();
1354 	sched_balance();
1355 }
1356 #endif
1357 
1358 /*
1359  * Setup the thread queues and initialize the topology based on MD
1360  * information.
1361  */
1362 static void
1363 sched_setup(void *dummy)
1364 {
1365 	struct tdq *tdq;
1366 
1367 	tdq = TDQ_SELF();
1368 #ifdef SMP
1369 	sched_setup_smp();
1370 #else
1371 	tdq_setup(tdq);
1372 #endif
1373 
1374 	/* Add thread0's load since it's running. */
1375 	TDQ_LOCK(tdq);
1376 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1377 	tdq_load_add(tdq, &thread0);
1378 	tdq->tdq_lowpri = thread0.td_priority;
1379 	TDQ_UNLOCK(tdq);
1380 }
1381 
1382 /*
1383  * This routine determines time constants after stathz and hz are setup.
1384  */
1385 /* ARGSUSED */
1386 static void
1387 sched_initticks(void *dummy)
1388 {
1389 	int incr;
1390 
1391 	realstathz = stathz ? stathz : hz;
1392 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1393 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1394 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1395 	    realstathz);
1396 
1397 	/*
1398 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1399 	 * hz not being evenly divisible by stathz on all platforms.
1400 	 */
1401 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1402 	/*
1403 	 * This does not work for values of stathz that are more than
1404 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1405 	 */
1406 	if (incr == 0)
1407 		incr = 1;
1408 	tickincr = incr;
1409 #ifdef SMP
1410 	/*
1411 	 * Set the default balance interval now that we know
1412 	 * what realstathz is.
1413 	 */
1414 	balance_interval = realstathz;
1415 	affinity = SCHED_AFFINITY_DEFAULT;
1416 #endif
1417 	if (sched_idlespinthresh < 0)
1418 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1419 }
1420 
1421 
1422 /*
1423  * This is the core of the interactivity algorithm.  Determines a score based
1424  * on past behavior.  It is the ratio of sleep time to run time scaled to
1425  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1426  * differs from the cpu usage because it does not account for time spent
1427  * waiting on a run-queue.  Would be prettier if we had floating point.
1428  */
1429 static int
1430 sched_interact_score(struct thread *td)
1431 {
1432 	struct td_sched *ts;
1433 	int div;
1434 
1435 	ts = td->td_sched;
1436 	/*
1437 	 * The score is only needed if this is likely to be an interactive
1438 	 * task.  Don't go through the expense of computing it if there's
1439 	 * no chance.
1440 	 */
1441 	if (sched_interact <= SCHED_INTERACT_HALF &&
1442 		ts->ts_runtime >= ts->ts_slptime)
1443 			return (SCHED_INTERACT_HALF);
1444 
1445 	if (ts->ts_runtime > ts->ts_slptime) {
1446 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1447 		return (SCHED_INTERACT_HALF +
1448 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1449 	}
1450 	if (ts->ts_slptime > ts->ts_runtime) {
1451 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1452 		return (ts->ts_runtime / div);
1453 	}
1454 	/* runtime == slptime */
1455 	if (ts->ts_runtime)
1456 		return (SCHED_INTERACT_HALF);
1457 
1458 	/*
1459 	 * This can happen if slptime and runtime are 0.
1460 	 */
1461 	return (0);
1462 
1463 }
1464 
1465 /*
1466  * Scale the scheduling priority according to the "interactivity" of this
1467  * process.
1468  */
1469 static void
1470 sched_priority(struct thread *td)
1471 {
1472 	int score;
1473 	int pri;
1474 
1475 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1476 		return;
1477 	/*
1478 	 * If the score is interactive we place the thread in the realtime
1479 	 * queue with a priority that is less than kernel and interrupt
1480 	 * priorities.  These threads are not subject to nice restrictions.
1481 	 *
1482 	 * Scores greater than this are placed on the normal timeshare queue
1483 	 * where the priority is partially decided by the most recent cpu
1484 	 * utilization and the rest is decided by nice value.
1485 	 *
1486 	 * The nice value of the process has a linear effect on the calculated
1487 	 * score.  Negative nice values make it easier for a thread to be
1488 	 * considered interactive.
1489 	 */
1490 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1491 	if (score < sched_interact) {
1492 		pri = PRI_MIN_INTERACT;
1493 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1494 		    sched_interact) * score;
1495 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1496 		    ("sched_priority: invalid interactive priority %d score %d",
1497 		    pri, score));
1498 	} else {
1499 		pri = SCHED_PRI_MIN;
1500 		if (td->td_sched->ts_ticks)
1501 			pri += min(SCHED_PRI_TICKS(td->td_sched),
1502 			    SCHED_PRI_RANGE - 1);
1503 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1504 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1505 		    ("sched_priority: invalid priority %d: nice %d, "
1506 		    "ticks %d ftick %d ltick %d tick pri %d",
1507 		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1508 		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1509 		    SCHED_PRI_TICKS(td->td_sched)));
1510 	}
1511 	sched_user_prio(td, pri);
1512 
1513 	return;
1514 }
1515 
1516 /*
1517  * This routine enforces a maximum limit on the amount of scheduling history
1518  * kept.  It is called after either the slptime or runtime is adjusted.  This
1519  * function is ugly due to integer math.
1520  */
1521 static void
1522 sched_interact_update(struct thread *td)
1523 {
1524 	struct td_sched *ts;
1525 	u_int sum;
1526 
1527 	ts = td->td_sched;
1528 	sum = ts->ts_runtime + ts->ts_slptime;
1529 	if (sum < SCHED_SLP_RUN_MAX)
1530 		return;
1531 	/*
1532 	 * This only happens from two places:
1533 	 * 1) We have added an unusual amount of run time from fork_exit.
1534 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1535 	 */
1536 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1537 		if (ts->ts_runtime > ts->ts_slptime) {
1538 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1539 			ts->ts_slptime = 1;
1540 		} else {
1541 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1542 			ts->ts_runtime = 1;
1543 		}
1544 		return;
1545 	}
1546 	/*
1547 	 * If we have exceeded by more than 1/5th then the algorithm below
1548 	 * will not bring us back into range.  Dividing by two here forces
1549 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1550 	 */
1551 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1552 		ts->ts_runtime /= 2;
1553 		ts->ts_slptime /= 2;
1554 		return;
1555 	}
1556 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1557 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1558 }
1559 
1560 /*
1561  * Scale back the interactivity history when a child thread is created.  The
1562  * history is inherited from the parent but the thread may behave totally
1563  * differently.  For example, a shell spawning a compiler process.  We want
1564  * to learn that the compiler is behaving badly very quickly.
1565  */
1566 static void
1567 sched_interact_fork(struct thread *td)
1568 {
1569 	int ratio;
1570 	int sum;
1571 
1572 	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1573 	if (sum > SCHED_SLP_RUN_FORK) {
1574 		ratio = sum / SCHED_SLP_RUN_FORK;
1575 		td->td_sched->ts_runtime /= ratio;
1576 		td->td_sched->ts_slptime /= ratio;
1577 	}
1578 }
1579 
1580 /*
1581  * Called from proc0_init() to setup the scheduler fields.
1582  */
1583 void
1584 schedinit(void)
1585 {
1586 
1587 	/*
1588 	 * Set up the scheduler specific parts of proc0.
1589 	 */
1590 	proc0.p_sched = NULL; /* XXX */
1591 	thread0.td_sched = &td_sched0;
1592 	td_sched0.ts_ltick = ticks;
1593 	td_sched0.ts_ftick = ticks;
1594 	td_sched0.ts_slice = 0;
1595 }
1596 
1597 /*
1598  * This is only somewhat accurate since given many processes of the same
1599  * priority they will switch when their slices run out, which will be
1600  * at most sched_slice stathz ticks.
1601  */
1602 int
1603 sched_rr_interval(void)
1604 {
1605 
1606 	/* Convert sched_slice from stathz to hz. */
1607 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1608 }
1609 
1610 /*
1611  * Update the percent cpu tracking information when it is requested or
1612  * the total history exceeds the maximum.  We keep a sliding history of
1613  * tick counts that slowly decays.  This is less precise than the 4BSD
1614  * mechanism since it happens with less regular and frequent events.
1615  */
1616 static void
1617 sched_pctcpu_update(struct td_sched *ts, int run)
1618 {
1619 	int t = ticks;
1620 
1621 	if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1622 		ts->ts_ticks = 0;
1623 		ts->ts_ftick = t - SCHED_TICK_TARG;
1624 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1625 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1626 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1627 		ts->ts_ftick = t - SCHED_TICK_TARG;
1628 	}
1629 	if (run)
1630 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1631 	ts->ts_ltick = t;
1632 }
1633 
1634 /*
1635  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1636  * if necessary.  This is the back-end for several priority related
1637  * functions.
1638  */
1639 static void
1640 sched_thread_priority(struct thread *td, u_char prio)
1641 {
1642 	struct td_sched *ts;
1643 	struct tdq *tdq;
1644 	int oldpri;
1645 
1646 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1647 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1648 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1649 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1650 	if (td != curthread && prio < td->td_priority) {
1651 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1652 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1653 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1654 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1655 		    curthread);
1656 	}
1657 	ts = td->td_sched;
1658 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1659 	if (td->td_priority == prio)
1660 		return;
1661 	/*
1662 	 * If the priority has been elevated due to priority
1663 	 * propagation, we may have to move ourselves to a new
1664 	 * queue.  This could be optimized to not re-add in some
1665 	 * cases.
1666 	 */
1667 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1668 		sched_rem(td);
1669 		td->td_priority = prio;
1670 		sched_add(td, SRQ_BORROWING);
1671 		return;
1672 	}
1673 	/*
1674 	 * If the thread is currently running we may have to adjust the lowpri
1675 	 * information so other cpus are aware of our current priority.
1676 	 */
1677 	if (TD_IS_RUNNING(td)) {
1678 		tdq = TDQ_CPU(ts->ts_cpu);
1679 		oldpri = td->td_priority;
1680 		td->td_priority = prio;
1681 		if (prio < tdq->tdq_lowpri)
1682 			tdq->tdq_lowpri = prio;
1683 		else if (tdq->tdq_lowpri == oldpri)
1684 			tdq_setlowpri(tdq, td);
1685 		return;
1686 	}
1687 	td->td_priority = prio;
1688 }
1689 
1690 /*
1691  * Update a thread's priority when it is lent another thread's
1692  * priority.
1693  */
1694 void
1695 sched_lend_prio(struct thread *td, u_char prio)
1696 {
1697 
1698 	td->td_flags |= TDF_BORROWING;
1699 	sched_thread_priority(td, prio);
1700 }
1701 
1702 /*
1703  * Restore a thread's priority when priority propagation is
1704  * over.  The prio argument is the minimum priority the thread
1705  * needs to have to satisfy other possible priority lending
1706  * requests.  If the thread's regular priority is less
1707  * important than prio, the thread will keep a priority boost
1708  * of prio.
1709  */
1710 void
1711 sched_unlend_prio(struct thread *td, u_char prio)
1712 {
1713 	u_char base_pri;
1714 
1715 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1716 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1717 		base_pri = td->td_user_pri;
1718 	else
1719 		base_pri = td->td_base_pri;
1720 	if (prio >= base_pri) {
1721 		td->td_flags &= ~TDF_BORROWING;
1722 		sched_thread_priority(td, base_pri);
1723 	} else
1724 		sched_lend_prio(td, prio);
1725 }
1726 
1727 /*
1728  * Standard entry for setting the priority to an absolute value.
1729  */
1730 void
1731 sched_prio(struct thread *td, u_char prio)
1732 {
1733 	u_char oldprio;
1734 
1735 	/* First, update the base priority. */
1736 	td->td_base_pri = prio;
1737 
1738 	/*
1739 	 * If the thread is borrowing another thread's priority, don't
1740 	 * ever lower the priority.
1741 	 */
1742 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1743 		return;
1744 
1745 	/* Change the real priority. */
1746 	oldprio = td->td_priority;
1747 	sched_thread_priority(td, prio);
1748 
1749 	/*
1750 	 * If the thread is on a turnstile, then let the turnstile update
1751 	 * its state.
1752 	 */
1753 	if (TD_ON_LOCK(td) && oldprio != prio)
1754 		turnstile_adjust(td, oldprio);
1755 }
1756 
1757 /*
1758  * Set the base user priority, does not effect current running priority.
1759  */
1760 void
1761 sched_user_prio(struct thread *td, u_char prio)
1762 {
1763 
1764 	td->td_base_user_pri = prio;
1765 	if (td->td_lend_user_pri <= prio)
1766 		return;
1767 	td->td_user_pri = prio;
1768 }
1769 
1770 void
1771 sched_lend_user_prio(struct thread *td, u_char prio)
1772 {
1773 
1774 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1775 	td->td_lend_user_pri = prio;
1776 	td->td_user_pri = min(prio, td->td_base_user_pri);
1777 	if (td->td_priority > td->td_user_pri)
1778 		sched_prio(td, td->td_user_pri);
1779 	else if (td->td_priority != td->td_user_pri)
1780 		td->td_flags |= TDF_NEEDRESCHED;
1781 }
1782 
1783 /*
1784  * Handle migration from sched_switch().  This happens only for
1785  * cpu binding.
1786  */
1787 static struct mtx *
1788 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1789 {
1790 	struct tdq *tdn;
1791 
1792 	tdn = TDQ_CPU(td->td_sched->ts_cpu);
1793 #ifdef SMP
1794 	tdq_load_rem(tdq, td);
1795 	/*
1796 	 * Do the lock dance required to avoid LOR.  We grab an extra
1797 	 * spinlock nesting to prevent preemption while we're
1798 	 * not holding either run-queue lock.
1799 	 */
1800 	spinlock_enter();
1801 	thread_lock_block(td);	/* This releases the lock on tdq. */
1802 
1803 	/*
1804 	 * Acquire both run-queue locks before placing the thread on the new
1805 	 * run-queue to avoid deadlocks created by placing a thread with a
1806 	 * blocked lock on the run-queue of a remote processor.  The deadlock
1807 	 * occurs when a third processor attempts to lock the two queues in
1808 	 * question while the target processor is spinning with its own
1809 	 * run-queue lock held while waiting for the blocked lock to clear.
1810 	 */
1811 	tdq_lock_pair(tdn, tdq);
1812 	tdq_add(tdn, td, flags);
1813 	tdq_notify(tdn, td);
1814 	TDQ_UNLOCK(tdn);
1815 	spinlock_exit();
1816 #endif
1817 	return (TDQ_LOCKPTR(tdn));
1818 }
1819 
1820 /*
1821  * Variadic version of thread_lock_unblock() that does not assume td_lock
1822  * is blocked.
1823  */
1824 static inline void
1825 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1826 {
1827 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1828 	    (uintptr_t)mtx);
1829 }
1830 
1831 /*
1832  * Switch threads.  This function has to handle threads coming in while
1833  * blocked for some reason, running, or idle.  It also must deal with
1834  * migrating a thread from one queue to another as running threads may
1835  * be assigned elsewhere via binding.
1836  */
1837 void
1838 sched_switch(struct thread *td, struct thread *newtd, int flags)
1839 {
1840 	struct tdq *tdq;
1841 	struct td_sched *ts;
1842 	struct mtx *mtx;
1843 	int srqflag;
1844 	int cpuid, preempted;
1845 
1846 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1847 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1848 
1849 	cpuid = PCPU_GET(cpuid);
1850 	tdq = TDQ_CPU(cpuid);
1851 	ts = td->td_sched;
1852 	mtx = td->td_lock;
1853 	sched_pctcpu_update(ts, 1);
1854 	ts->ts_rltick = ticks;
1855 	td->td_lastcpu = td->td_oncpu;
1856 	td->td_oncpu = NOCPU;
1857 	preempted = !(td->td_flags & TDF_SLICEEND);
1858 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1859 	td->td_owepreempt = 0;
1860 	if (!TD_IS_IDLETHREAD(td))
1861 		tdq->tdq_switchcnt++;
1862 	/*
1863 	 * The lock pointer in an idle thread should never change.  Reset it
1864 	 * to CAN_RUN as well.
1865 	 */
1866 	if (TD_IS_IDLETHREAD(td)) {
1867 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1868 		TD_SET_CAN_RUN(td);
1869 	} else if (TD_IS_RUNNING(td)) {
1870 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1871 		srqflag = preempted ?
1872 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1873 		    SRQ_OURSELF|SRQ_YIELDING;
1874 #ifdef SMP
1875 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1876 			ts->ts_cpu = sched_pickcpu(td, 0);
1877 #endif
1878 		if (ts->ts_cpu == cpuid)
1879 			tdq_runq_add(tdq, td, srqflag);
1880 		else {
1881 			KASSERT(THREAD_CAN_MIGRATE(td) ||
1882 			    (ts->ts_flags & TSF_BOUND) != 0,
1883 			    ("Thread %p shouldn't migrate", td));
1884 			mtx = sched_switch_migrate(tdq, td, srqflag);
1885 		}
1886 	} else {
1887 		/* This thread must be going to sleep. */
1888 		TDQ_LOCK(tdq);
1889 		mtx = thread_lock_block(td);
1890 		tdq_load_rem(tdq, td);
1891 	}
1892 	/*
1893 	 * We enter here with the thread blocked and assigned to the
1894 	 * appropriate cpu run-queue or sleep-queue and with the current
1895 	 * thread-queue locked.
1896 	 */
1897 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1898 	newtd = choosethread();
1899 	/*
1900 	 * Call the MD code to switch contexts if necessary.
1901 	 */
1902 	if (td != newtd) {
1903 #ifdef	HWPMC_HOOKS
1904 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1905 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1906 #endif
1907 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1908 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1909 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1910 		sched_pctcpu_update(newtd->td_sched, 0);
1911 
1912 #ifdef KDTRACE_HOOKS
1913 		/*
1914 		 * If DTrace has set the active vtime enum to anything
1915 		 * other than INACTIVE (0), then it should have set the
1916 		 * function to call.
1917 		 */
1918 		if (dtrace_vtime_active)
1919 			(*dtrace_vtime_switch_func)(newtd);
1920 #endif
1921 
1922 		cpu_switch(td, newtd, mtx);
1923 		/*
1924 		 * We may return from cpu_switch on a different cpu.  However,
1925 		 * we always return with td_lock pointing to the current cpu's
1926 		 * run queue lock.
1927 		 */
1928 		cpuid = PCPU_GET(cpuid);
1929 		tdq = TDQ_CPU(cpuid);
1930 		lock_profile_obtain_lock_success(
1931 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1932 
1933 		SDT_PROBE0(sched, , , on__cpu);
1934 #ifdef	HWPMC_HOOKS
1935 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1936 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1937 #endif
1938 	} else {
1939 		thread_unblock_switch(td, mtx);
1940 		SDT_PROBE0(sched, , , remain__cpu);
1941 	}
1942 	/*
1943 	 * Assert that all went well and return.
1944 	 */
1945 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1946 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1947 	td->td_oncpu = cpuid;
1948 }
1949 
1950 /*
1951  * Adjust thread priorities as a result of a nice request.
1952  */
1953 void
1954 sched_nice(struct proc *p, int nice)
1955 {
1956 	struct thread *td;
1957 
1958 	PROC_LOCK_ASSERT(p, MA_OWNED);
1959 
1960 	p->p_nice = nice;
1961 	FOREACH_THREAD_IN_PROC(p, td) {
1962 		thread_lock(td);
1963 		sched_priority(td);
1964 		sched_prio(td, td->td_base_user_pri);
1965 		thread_unlock(td);
1966 	}
1967 }
1968 
1969 /*
1970  * Record the sleep time for the interactivity scorer.
1971  */
1972 void
1973 sched_sleep(struct thread *td, int prio)
1974 {
1975 
1976 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1977 
1978 	td->td_slptick = ticks;
1979 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1980 		td->td_flags |= TDF_CANSWAP;
1981 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1982 		return;
1983 	if (static_boost == 1 && prio)
1984 		sched_prio(td, prio);
1985 	else if (static_boost && td->td_priority > static_boost)
1986 		sched_prio(td, static_boost);
1987 }
1988 
1989 /*
1990  * Schedule a thread to resume execution and record how long it voluntarily
1991  * slept.  We also update the pctcpu, interactivity, and priority.
1992  */
1993 void
1994 sched_wakeup(struct thread *td)
1995 {
1996 	struct td_sched *ts;
1997 	int slptick;
1998 
1999 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2000 	ts = td->td_sched;
2001 	td->td_flags &= ~TDF_CANSWAP;
2002 	/*
2003 	 * If we slept for more than a tick update our interactivity and
2004 	 * priority.
2005 	 */
2006 	slptick = td->td_slptick;
2007 	td->td_slptick = 0;
2008 	if (slptick && slptick != ticks) {
2009 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2010 		sched_interact_update(td);
2011 		sched_pctcpu_update(ts, 0);
2012 	}
2013 	/*
2014 	 * Reset the slice value since we slept and advanced the round-robin.
2015 	 */
2016 	ts->ts_slice = 0;
2017 	sched_add(td, SRQ_BORING);
2018 }
2019 
2020 /*
2021  * Penalize the parent for creating a new child and initialize the child's
2022  * priority.
2023  */
2024 void
2025 sched_fork(struct thread *td, struct thread *child)
2026 {
2027 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2028 	sched_pctcpu_update(td->td_sched, 1);
2029 	sched_fork_thread(td, child);
2030 	/*
2031 	 * Penalize the parent and child for forking.
2032 	 */
2033 	sched_interact_fork(child);
2034 	sched_priority(child);
2035 	td->td_sched->ts_runtime += tickincr;
2036 	sched_interact_update(td);
2037 	sched_priority(td);
2038 }
2039 
2040 /*
2041  * Fork a new thread, may be within the same process.
2042  */
2043 void
2044 sched_fork_thread(struct thread *td, struct thread *child)
2045 {
2046 	struct td_sched *ts;
2047 	struct td_sched *ts2;
2048 	struct tdq *tdq;
2049 
2050 	tdq = TDQ_SELF();
2051 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2052 	/*
2053 	 * Initialize child.
2054 	 */
2055 	ts = td->td_sched;
2056 	ts2 = child->td_sched;
2057 	child->td_lock = TDQ_LOCKPTR(tdq);
2058 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2059 	ts2->ts_cpu = ts->ts_cpu;
2060 	ts2->ts_flags = 0;
2061 	/*
2062 	 * Grab our parents cpu estimation information.
2063 	 */
2064 	ts2->ts_ticks = ts->ts_ticks;
2065 	ts2->ts_ltick = ts->ts_ltick;
2066 	ts2->ts_ftick = ts->ts_ftick;
2067 	/*
2068 	 * Do not inherit any borrowed priority from the parent.
2069 	 */
2070 	child->td_priority = child->td_base_pri;
2071 	/*
2072 	 * And update interactivity score.
2073 	 */
2074 	ts2->ts_slptime = ts->ts_slptime;
2075 	ts2->ts_runtime = ts->ts_runtime;
2076 	/* Attempt to quickly learn interactivity. */
2077 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2078 #ifdef KTR
2079 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2080 #endif
2081 }
2082 
2083 /*
2084  * Adjust the priority class of a thread.
2085  */
2086 void
2087 sched_class(struct thread *td, int class)
2088 {
2089 
2090 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2091 	if (td->td_pri_class == class)
2092 		return;
2093 	td->td_pri_class = class;
2094 }
2095 
2096 /*
2097  * Return some of the child's priority and interactivity to the parent.
2098  */
2099 void
2100 sched_exit(struct proc *p, struct thread *child)
2101 {
2102 	struct thread *td;
2103 
2104 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2105 	    "prio:%d", child->td_priority);
2106 	PROC_LOCK_ASSERT(p, MA_OWNED);
2107 	td = FIRST_THREAD_IN_PROC(p);
2108 	sched_exit_thread(td, child);
2109 }
2110 
2111 /*
2112  * Penalize another thread for the time spent on this one.  This helps to
2113  * worsen the priority and interactivity of processes which schedule batch
2114  * jobs such as make.  This has little effect on the make process itself but
2115  * causes new processes spawned by it to receive worse scores immediately.
2116  */
2117 void
2118 sched_exit_thread(struct thread *td, struct thread *child)
2119 {
2120 
2121 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2122 	    "prio:%d", child->td_priority);
2123 	/*
2124 	 * Give the child's runtime to the parent without returning the
2125 	 * sleep time as a penalty to the parent.  This causes shells that
2126 	 * launch expensive things to mark their children as expensive.
2127 	 */
2128 	thread_lock(td);
2129 	td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2130 	sched_interact_update(td);
2131 	sched_priority(td);
2132 	thread_unlock(td);
2133 }
2134 
2135 void
2136 sched_preempt(struct thread *td)
2137 {
2138 	struct tdq *tdq;
2139 
2140 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2141 
2142 	thread_lock(td);
2143 	tdq = TDQ_SELF();
2144 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2145 	tdq->tdq_ipipending = 0;
2146 	if (td->td_priority > tdq->tdq_lowpri) {
2147 		int flags;
2148 
2149 		flags = SW_INVOL | SW_PREEMPT;
2150 		if (td->td_critnest > 1)
2151 			td->td_owepreempt = 1;
2152 		else if (TD_IS_IDLETHREAD(td))
2153 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2154 		else
2155 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2156 	}
2157 	thread_unlock(td);
2158 }
2159 
2160 /*
2161  * Fix priorities on return to user-space.  Priorities may be elevated due
2162  * to static priorities in msleep() or similar.
2163  */
2164 void
2165 sched_userret(struct thread *td)
2166 {
2167 	/*
2168 	 * XXX we cheat slightly on the locking here to avoid locking in
2169 	 * the usual case.  Setting td_priority here is essentially an
2170 	 * incomplete workaround for not setting it properly elsewhere.
2171 	 * Now that some interrupt handlers are threads, not setting it
2172 	 * properly elsewhere can clobber it in the window between setting
2173 	 * it here and returning to user mode, so don't waste time setting
2174 	 * it perfectly here.
2175 	 */
2176 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2177 	    ("thread with borrowed priority returning to userland"));
2178 	if (td->td_priority != td->td_user_pri) {
2179 		thread_lock(td);
2180 		td->td_priority = td->td_user_pri;
2181 		td->td_base_pri = td->td_user_pri;
2182 		tdq_setlowpri(TDQ_SELF(), td);
2183 		thread_unlock(td);
2184         }
2185 }
2186 
2187 /*
2188  * Handle a stathz tick.  This is really only relevant for timeshare
2189  * threads.
2190  */
2191 void
2192 sched_clock(struct thread *td)
2193 {
2194 	struct tdq *tdq;
2195 	struct td_sched *ts;
2196 
2197 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2198 	tdq = TDQ_SELF();
2199 #ifdef SMP
2200 	/*
2201 	 * We run the long term load balancer infrequently on the first cpu.
2202 	 */
2203 	if (balance_tdq == tdq) {
2204 		if (balance_ticks && --balance_ticks == 0)
2205 			sched_balance();
2206 	}
2207 #endif
2208 	/*
2209 	 * Save the old switch count so we have a record of the last ticks
2210 	 * activity.   Initialize the new switch count based on our load.
2211 	 * If there is some activity seed it to reflect that.
2212 	 */
2213 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2214 	tdq->tdq_switchcnt = tdq->tdq_load;
2215 	/*
2216 	 * Advance the insert index once for each tick to ensure that all
2217 	 * threads get a chance to run.
2218 	 */
2219 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2220 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2221 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2222 			tdq->tdq_ridx = tdq->tdq_idx;
2223 	}
2224 	ts = td->td_sched;
2225 	sched_pctcpu_update(ts, 1);
2226 	if (td->td_pri_class & PRI_FIFO_BIT)
2227 		return;
2228 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2229 		/*
2230 		 * We used a tick; charge it to the thread so
2231 		 * that we can compute our interactivity.
2232 		 */
2233 		td->td_sched->ts_runtime += tickincr;
2234 		sched_interact_update(td);
2235 		sched_priority(td);
2236 	}
2237 
2238 	/*
2239 	 * Force a context switch if the current thread has used up a full
2240 	 * time slice (default is 100ms).
2241 	 */
2242 	if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2243 		ts->ts_slice = 0;
2244 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2245 	}
2246 }
2247 
2248 /*
2249  * Called once per hz tick.
2250  */
2251 void
2252 sched_tick(int cnt)
2253 {
2254 
2255 }
2256 
2257 /*
2258  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2259  * cooperative idle threads.
2260  */
2261 int
2262 sched_runnable(void)
2263 {
2264 	struct tdq *tdq;
2265 	int load;
2266 
2267 	load = 1;
2268 
2269 	tdq = TDQ_SELF();
2270 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2271 		if (tdq->tdq_load > 0)
2272 			goto out;
2273 	} else
2274 		if (tdq->tdq_load - 1 > 0)
2275 			goto out;
2276 	load = 0;
2277 out:
2278 	return (load);
2279 }
2280 
2281 /*
2282  * Choose the highest priority thread to run.  The thread is removed from
2283  * the run-queue while running however the load remains.  For SMP we set
2284  * the tdq in the global idle bitmask if it idles here.
2285  */
2286 struct thread *
2287 sched_choose(void)
2288 {
2289 	struct thread *td;
2290 	struct tdq *tdq;
2291 
2292 	tdq = TDQ_SELF();
2293 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2294 	td = tdq_choose(tdq);
2295 	if (td) {
2296 		tdq_runq_rem(tdq, td);
2297 		tdq->tdq_lowpri = td->td_priority;
2298 		return (td);
2299 	}
2300 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2301 	return (PCPU_GET(idlethread));
2302 }
2303 
2304 /*
2305  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2306  * we always request it once we exit a critical section.
2307  */
2308 static inline void
2309 sched_setpreempt(struct thread *td)
2310 {
2311 	struct thread *ctd;
2312 	int cpri;
2313 	int pri;
2314 
2315 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2316 
2317 	ctd = curthread;
2318 	pri = td->td_priority;
2319 	cpri = ctd->td_priority;
2320 	if (pri < cpri)
2321 		ctd->td_flags |= TDF_NEEDRESCHED;
2322 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2323 		return;
2324 	if (!sched_shouldpreempt(pri, cpri, 0))
2325 		return;
2326 	ctd->td_owepreempt = 1;
2327 }
2328 
2329 /*
2330  * Add a thread to a thread queue.  Select the appropriate runq and add the
2331  * thread to it.  This is the internal function called when the tdq is
2332  * predetermined.
2333  */
2334 void
2335 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2336 {
2337 
2338 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2339 	KASSERT((td->td_inhibitors == 0),
2340 	    ("sched_add: trying to run inhibited thread"));
2341 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2342 	    ("sched_add: bad thread state"));
2343 	KASSERT(td->td_flags & TDF_INMEM,
2344 	    ("sched_add: thread swapped out"));
2345 
2346 	if (td->td_priority < tdq->tdq_lowpri)
2347 		tdq->tdq_lowpri = td->td_priority;
2348 	tdq_runq_add(tdq, td, flags);
2349 	tdq_load_add(tdq, td);
2350 }
2351 
2352 /*
2353  * Select the target thread queue and add a thread to it.  Request
2354  * preemption or IPI a remote processor if required.
2355  */
2356 void
2357 sched_add(struct thread *td, int flags)
2358 {
2359 	struct tdq *tdq;
2360 #ifdef SMP
2361 	int cpu;
2362 #endif
2363 
2364 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2365 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2366 	    sched_tdname(curthread));
2367 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2368 	    KTR_ATTR_LINKED, sched_tdname(td));
2369 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2370 	    flags & SRQ_PREEMPTED);
2371 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2372 	/*
2373 	 * Recalculate the priority before we select the target cpu or
2374 	 * run-queue.
2375 	 */
2376 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2377 		sched_priority(td);
2378 #ifdef SMP
2379 	/*
2380 	 * Pick the destination cpu and if it isn't ours transfer to the
2381 	 * target cpu.
2382 	 */
2383 	cpu = sched_pickcpu(td, flags);
2384 	tdq = sched_setcpu(td, cpu, flags);
2385 	tdq_add(tdq, td, flags);
2386 	if (cpu != PCPU_GET(cpuid)) {
2387 		tdq_notify(tdq, td);
2388 		return;
2389 	}
2390 #else
2391 	tdq = TDQ_SELF();
2392 	TDQ_LOCK(tdq);
2393 	/*
2394 	 * Now that the thread is moving to the run-queue, set the lock
2395 	 * to the scheduler's lock.
2396 	 */
2397 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2398 	tdq_add(tdq, td, flags);
2399 #endif
2400 	if (!(flags & SRQ_YIELDING))
2401 		sched_setpreempt(td);
2402 }
2403 
2404 /*
2405  * Remove a thread from a run-queue without running it.  This is used
2406  * when we're stealing a thread from a remote queue.  Otherwise all threads
2407  * exit by calling sched_exit_thread() and sched_throw() themselves.
2408  */
2409 void
2410 sched_rem(struct thread *td)
2411 {
2412 	struct tdq *tdq;
2413 
2414 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2415 	    "prio:%d", td->td_priority);
2416 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2417 	tdq = TDQ_CPU(td->td_sched->ts_cpu);
2418 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2419 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2420 	KASSERT(TD_ON_RUNQ(td),
2421 	    ("sched_rem: thread not on run queue"));
2422 	tdq_runq_rem(tdq, td);
2423 	tdq_load_rem(tdq, td);
2424 	TD_SET_CAN_RUN(td);
2425 	if (td->td_priority == tdq->tdq_lowpri)
2426 		tdq_setlowpri(tdq, NULL);
2427 }
2428 
2429 /*
2430  * Fetch cpu utilization information.  Updates on demand.
2431  */
2432 fixpt_t
2433 sched_pctcpu(struct thread *td)
2434 {
2435 	fixpt_t pctcpu;
2436 	struct td_sched *ts;
2437 
2438 	pctcpu = 0;
2439 	ts = td->td_sched;
2440 	if (ts == NULL)
2441 		return (0);
2442 
2443 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2444 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2445 	if (ts->ts_ticks) {
2446 		int rtick;
2447 
2448 		/* How many rtick per second ? */
2449 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2450 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2451 	}
2452 
2453 	return (pctcpu);
2454 }
2455 
2456 /*
2457  * Enforce affinity settings for a thread.  Called after adjustments to
2458  * cpumask.
2459  */
2460 void
2461 sched_affinity(struct thread *td)
2462 {
2463 #ifdef SMP
2464 	struct td_sched *ts;
2465 
2466 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2467 	ts = td->td_sched;
2468 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2469 		return;
2470 	if (TD_ON_RUNQ(td)) {
2471 		sched_rem(td);
2472 		sched_add(td, SRQ_BORING);
2473 		return;
2474 	}
2475 	if (!TD_IS_RUNNING(td))
2476 		return;
2477 	/*
2478 	 * Force a switch before returning to userspace.  If the
2479 	 * target thread is not running locally send an ipi to force
2480 	 * the issue.
2481 	 */
2482 	td->td_flags |= TDF_NEEDRESCHED;
2483 	if (td != curthread)
2484 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2485 #endif
2486 }
2487 
2488 /*
2489  * Bind a thread to a target cpu.
2490  */
2491 void
2492 sched_bind(struct thread *td, int cpu)
2493 {
2494 	struct td_sched *ts;
2495 
2496 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2497 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2498 	ts = td->td_sched;
2499 	if (ts->ts_flags & TSF_BOUND)
2500 		sched_unbind(td);
2501 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2502 	ts->ts_flags |= TSF_BOUND;
2503 	sched_pin();
2504 	if (PCPU_GET(cpuid) == cpu)
2505 		return;
2506 	ts->ts_cpu = cpu;
2507 	/* When we return from mi_switch we'll be on the correct cpu. */
2508 	mi_switch(SW_VOL, NULL);
2509 }
2510 
2511 /*
2512  * Release a bound thread.
2513  */
2514 void
2515 sched_unbind(struct thread *td)
2516 {
2517 	struct td_sched *ts;
2518 
2519 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2520 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2521 	ts = td->td_sched;
2522 	if ((ts->ts_flags & TSF_BOUND) == 0)
2523 		return;
2524 	ts->ts_flags &= ~TSF_BOUND;
2525 	sched_unpin();
2526 }
2527 
2528 int
2529 sched_is_bound(struct thread *td)
2530 {
2531 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2532 	return (td->td_sched->ts_flags & TSF_BOUND);
2533 }
2534 
2535 /*
2536  * Basic yield call.
2537  */
2538 void
2539 sched_relinquish(struct thread *td)
2540 {
2541 	thread_lock(td);
2542 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2543 	thread_unlock(td);
2544 }
2545 
2546 /*
2547  * Return the total system load.
2548  */
2549 int
2550 sched_load(void)
2551 {
2552 #ifdef SMP
2553 	int total;
2554 	int i;
2555 
2556 	total = 0;
2557 	CPU_FOREACH(i)
2558 		total += TDQ_CPU(i)->tdq_sysload;
2559 	return (total);
2560 #else
2561 	return (TDQ_SELF()->tdq_sysload);
2562 #endif
2563 }
2564 
2565 int
2566 sched_sizeof_proc(void)
2567 {
2568 	return (sizeof(struct proc));
2569 }
2570 
2571 int
2572 sched_sizeof_thread(void)
2573 {
2574 	return (sizeof(struct thread) + sizeof(struct td_sched));
2575 }
2576 
2577 #ifdef SMP
2578 #define	TDQ_IDLESPIN(tdq)						\
2579     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2580 #else
2581 #define	TDQ_IDLESPIN(tdq)	1
2582 #endif
2583 
2584 /*
2585  * The actual idle process.
2586  */
2587 void
2588 sched_idletd(void *dummy)
2589 {
2590 	struct thread *td;
2591 	struct tdq *tdq;
2592 	int oldswitchcnt, switchcnt;
2593 	int i;
2594 
2595 	mtx_assert(&Giant, MA_NOTOWNED);
2596 	td = curthread;
2597 	tdq = TDQ_SELF();
2598 	THREAD_NO_SLEEPING();
2599 	oldswitchcnt = -1;
2600 	for (;;) {
2601 		if (tdq->tdq_load) {
2602 			thread_lock(td);
2603 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2604 			thread_unlock(td);
2605 		}
2606 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2607 #ifdef SMP
2608 		if (switchcnt != oldswitchcnt) {
2609 			oldswitchcnt = switchcnt;
2610 			if (tdq_idled(tdq) == 0)
2611 				continue;
2612 		}
2613 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2614 #else
2615 		oldswitchcnt = switchcnt;
2616 #endif
2617 		/*
2618 		 * If we're switching very frequently, spin while checking
2619 		 * for load rather than entering a low power state that
2620 		 * may require an IPI.  However, don't do any busy
2621 		 * loops while on SMT machines as this simply steals
2622 		 * cycles from cores doing useful work.
2623 		 */
2624 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2625 			for (i = 0; i < sched_idlespins; i++) {
2626 				if (tdq->tdq_load)
2627 					break;
2628 				cpu_spinwait();
2629 			}
2630 		}
2631 
2632 		/* If there was context switch during spin, restart it. */
2633 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2634 		if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2635 			continue;
2636 
2637 		/* Run main MD idle handler. */
2638 		tdq->tdq_cpu_idle = 1;
2639 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2640 		tdq->tdq_cpu_idle = 0;
2641 
2642 		/*
2643 		 * Account thread-less hardware interrupts and
2644 		 * other wakeup reasons equal to context switches.
2645 		 */
2646 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2647 		if (switchcnt != oldswitchcnt)
2648 			continue;
2649 		tdq->tdq_switchcnt++;
2650 		oldswitchcnt++;
2651 	}
2652 }
2653 
2654 /*
2655  * A CPU is entering for the first time or a thread is exiting.
2656  */
2657 void
2658 sched_throw(struct thread *td)
2659 {
2660 	struct thread *newtd;
2661 	struct tdq *tdq;
2662 
2663 	tdq = TDQ_SELF();
2664 	if (td == NULL) {
2665 		/* Correct spinlock nesting and acquire the correct lock. */
2666 		TDQ_LOCK(tdq);
2667 		spinlock_exit();
2668 		PCPU_SET(switchtime, cpu_ticks());
2669 		PCPU_SET(switchticks, ticks);
2670 	} else {
2671 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2672 		tdq_load_rem(tdq, td);
2673 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2674 	}
2675 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2676 	newtd = choosethread();
2677 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2678 	cpu_throw(td, newtd);		/* doesn't return */
2679 }
2680 
2681 /*
2682  * This is called from fork_exit().  Just acquire the correct locks and
2683  * let fork do the rest of the work.
2684  */
2685 void
2686 sched_fork_exit(struct thread *td)
2687 {
2688 	struct td_sched *ts;
2689 	struct tdq *tdq;
2690 	int cpuid;
2691 
2692 	/*
2693 	 * Finish setting up thread glue so that it begins execution in a
2694 	 * non-nested critical section with the scheduler lock held.
2695 	 */
2696 	cpuid = PCPU_GET(cpuid);
2697 	tdq = TDQ_CPU(cpuid);
2698 	ts = td->td_sched;
2699 	if (TD_IS_IDLETHREAD(td))
2700 		td->td_lock = TDQ_LOCKPTR(tdq);
2701 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2702 	td->td_oncpu = cpuid;
2703 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2704 	lock_profile_obtain_lock_success(
2705 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2706 }
2707 
2708 /*
2709  * Create on first use to catch odd startup conditons.
2710  */
2711 char *
2712 sched_tdname(struct thread *td)
2713 {
2714 #ifdef KTR
2715 	struct td_sched *ts;
2716 
2717 	ts = td->td_sched;
2718 	if (ts->ts_name[0] == '\0')
2719 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2720 		    "%s tid %d", td->td_name, td->td_tid);
2721 	return (ts->ts_name);
2722 #else
2723 	return (td->td_name);
2724 #endif
2725 }
2726 
2727 #ifdef KTR
2728 void
2729 sched_clear_tdname(struct thread *td)
2730 {
2731 	struct td_sched *ts;
2732 
2733 	ts = td->td_sched;
2734 	ts->ts_name[0] = '\0';
2735 }
2736 #endif
2737 
2738 #ifdef SMP
2739 
2740 /*
2741  * Build the CPU topology dump string. Is recursively called to collect
2742  * the topology tree.
2743  */
2744 static int
2745 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2746     int indent)
2747 {
2748 	char cpusetbuf[CPUSETBUFSIZ];
2749 	int i, first;
2750 
2751 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2752 	    "", 1 + indent / 2, cg->cg_level);
2753 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2754 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2755 	first = TRUE;
2756 	for (i = 0; i < MAXCPU; i++) {
2757 		if (CPU_ISSET(i, &cg->cg_mask)) {
2758 			if (!first)
2759 				sbuf_printf(sb, ", ");
2760 			else
2761 				first = FALSE;
2762 			sbuf_printf(sb, "%d", i);
2763 		}
2764 	}
2765 	sbuf_printf(sb, "</cpu>\n");
2766 
2767 	if (cg->cg_flags != 0) {
2768 		sbuf_printf(sb, "%*s <flags>", indent, "");
2769 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2770 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2771 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2772 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2773 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2774 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2775 		sbuf_printf(sb, "</flags>\n");
2776 	}
2777 
2778 	if (cg->cg_children > 0) {
2779 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2780 		for (i = 0; i < cg->cg_children; i++)
2781 			sysctl_kern_sched_topology_spec_internal(sb,
2782 			    &cg->cg_child[i], indent+2);
2783 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2784 	}
2785 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2786 	return (0);
2787 }
2788 
2789 /*
2790  * Sysctl handler for retrieving topology dump. It's a wrapper for
2791  * the recursive sysctl_kern_smp_topology_spec_internal().
2792  */
2793 static int
2794 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2795 {
2796 	struct sbuf *topo;
2797 	int err;
2798 
2799 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2800 
2801 	topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2802 	if (topo == NULL)
2803 		return (ENOMEM);
2804 
2805 	sbuf_printf(topo, "<groups>\n");
2806 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2807 	sbuf_printf(topo, "</groups>\n");
2808 
2809 	if (err == 0) {
2810 		sbuf_finish(topo);
2811 		err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2812 	}
2813 	sbuf_delete(topo);
2814 	return (err);
2815 }
2816 
2817 #endif
2818 
2819 static int
2820 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2821 {
2822 	int error, new_val, period;
2823 
2824 	period = 1000000 / realstathz;
2825 	new_val = period * sched_slice;
2826 	error = sysctl_handle_int(oidp, &new_val, 0, req);
2827 	if (error != 0 || req->newptr == NULL)
2828 		return (error);
2829 	if (new_val <= 0)
2830 		return (EINVAL);
2831 	sched_slice = imax(1, (new_val + period / 2) / period);
2832 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
2833 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2834 	    realstathz);
2835 	return (0);
2836 }
2837 
2838 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2839 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2840     "Scheduler name");
2841 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2842     NULL, 0, sysctl_kern_quantum, "I",
2843     "Quantum for timeshare threads in microseconds");
2844 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2845     "Quantum for timeshare threads in stathz ticks");
2846 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2847     "Interactivity score threshold");
2848 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2849     &preempt_thresh, 0,
2850     "Maximal (lowest) priority for preemption");
2851 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2852     "Assign static kernel priorities to sleeping threads");
2853 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2854     "Number of times idle thread will spin waiting for new work");
2855 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2856     &sched_idlespinthresh, 0,
2857     "Threshold before we will permit idle thread spinning");
2858 #ifdef SMP
2859 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2860     "Number of hz ticks to keep thread affinity for");
2861 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2862     "Enables the long-term load balancer");
2863 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2864     &balance_interval, 0,
2865     "Average period in stathz ticks to run the long-term balancer");
2866 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2867     "Attempts to steal work from other cores before idling");
2868 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2869     "Minimum load on remote CPU before we'll steal");
2870 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2871     CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2872     "XML dump of detected CPU topology");
2873 #endif
2874 
2875 /* ps compat.  All cpu percentages from ULE are weighted. */
2876 static int ccpu = 0;
2877 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2878