xref: /freebsd/sys/kern/sched_ule.c (revision 1a61beb0549e05b33df31380e427d90f6e46ff7e)
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 pri;
1061 	int i;
1062 
1063 	rqb = &rq->rq_status;
1064 	bit = start & (RQB_BPW -1);
1065 	pri = 0;
1066 	first = NULL;
1067 again:
1068 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1069 		if (rqb->rqb_bits[i] == 0)
1070 			continue;
1071 		if (bit != 0) {
1072 			for (pri = bit; pri < RQB_BPW; pri++)
1073 				if (rqb->rqb_bits[i] & (1ul << pri))
1074 					break;
1075 			if (pri >= RQB_BPW)
1076 				continue;
1077 		} else
1078 			pri = RQB_FFS(rqb->rqb_bits[i]);
1079 		pri += (i << RQB_L2BPW);
1080 		rqh = &rq->rq_queues[pri];
1081 		TAILQ_FOREACH(td, rqh, td_runq) {
1082 			if (first && THREAD_CAN_MIGRATE(td) &&
1083 			    THREAD_CAN_SCHED(td, cpu))
1084 				return (td);
1085 			first = td;
1086 		}
1087 	}
1088 	if (start != 0) {
1089 		start = 0;
1090 		goto again;
1091 	}
1092 
1093 	if (first && THREAD_CAN_MIGRATE(first) &&
1094 	    THREAD_CAN_SCHED(first, cpu))
1095 		return (first);
1096 	return (NULL);
1097 }
1098 
1099 /*
1100  * Steals load from a standard linear queue.
1101  */
1102 static struct thread *
1103 runq_steal(struct runq *rq, int cpu)
1104 {
1105 	struct rqhead *rqh;
1106 	struct rqbits *rqb;
1107 	struct thread *td;
1108 	int word;
1109 	int bit;
1110 
1111 	rqb = &rq->rq_status;
1112 	for (word = 0; word < RQB_LEN; word++) {
1113 		if (rqb->rqb_bits[word] == 0)
1114 			continue;
1115 		for (bit = 0; bit < RQB_BPW; bit++) {
1116 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1117 				continue;
1118 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1119 			TAILQ_FOREACH(td, rqh, td_runq)
1120 				if (THREAD_CAN_MIGRATE(td) &&
1121 				    THREAD_CAN_SCHED(td, cpu))
1122 					return (td);
1123 		}
1124 	}
1125 	return (NULL);
1126 }
1127 
1128 /*
1129  * Attempt to steal a thread in priority order from a thread queue.
1130  */
1131 static struct thread *
1132 tdq_steal(struct tdq *tdq, int cpu)
1133 {
1134 	struct thread *td;
1135 
1136 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1137 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1138 		return (td);
1139 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1140 	    cpu, tdq->tdq_ridx)) != NULL)
1141 		return (td);
1142 	return (runq_steal(&tdq->tdq_idle, cpu));
1143 }
1144 
1145 /*
1146  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1147  * current lock and returns with the assigned queue locked.
1148  */
1149 static inline struct tdq *
1150 sched_setcpu(struct thread *td, int cpu, int flags)
1151 {
1152 
1153 	struct tdq *tdq;
1154 
1155 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1156 	tdq = TDQ_CPU(cpu);
1157 	td->td_sched->ts_cpu = cpu;
1158 	/*
1159 	 * If the lock matches just return the queue.
1160 	 */
1161 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1162 		return (tdq);
1163 #ifdef notyet
1164 	/*
1165 	 * If the thread isn't running its lockptr is a
1166 	 * turnstile or a sleepqueue.  We can just lock_set without
1167 	 * blocking.
1168 	 */
1169 	if (TD_CAN_RUN(td)) {
1170 		TDQ_LOCK(tdq);
1171 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1172 		return (tdq);
1173 	}
1174 #endif
1175 	/*
1176 	 * The hard case, migration, we need to block the thread first to
1177 	 * prevent order reversals with other cpus locks.
1178 	 */
1179 	spinlock_enter();
1180 	thread_lock_block(td);
1181 	TDQ_LOCK(tdq);
1182 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1183 	spinlock_exit();
1184 	return (tdq);
1185 }
1186 
1187 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1188 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1189 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1190 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1191 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1192 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1193 
1194 static int
1195 sched_pickcpu(struct thread *td, int flags)
1196 {
1197 	struct cpu_group *cg, *ccg;
1198 	struct td_sched *ts;
1199 	struct tdq *tdq;
1200 	cpuset_t mask;
1201 	int cpu, pri, self;
1202 
1203 	self = PCPU_GET(cpuid);
1204 	ts = td->td_sched;
1205 	if (smp_started == 0)
1206 		return (self);
1207 	/*
1208 	 * Don't migrate a running thread from sched_switch().
1209 	 */
1210 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1211 		return (ts->ts_cpu);
1212 	/*
1213 	 * Prefer to run interrupt threads on the processors that generate
1214 	 * the interrupt.
1215 	 */
1216 	pri = td->td_priority;
1217 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1218 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1219 		SCHED_STAT_INC(pickcpu_intrbind);
1220 		ts->ts_cpu = self;
1221 		if (TDQ_CPU(self)->tdq_lowpri > pri) {
1222 			SCHED_STAT_INC(pickcpu_affinity);
1223 			return (ts->ts_cpu);
1224 		}
1225 	}
1226 	/*
1227 	 * If the thread can run on the last cpu and the affinity has not
1228 	 * expired or it is idle run it there.
1229 	 */
1230 	tdq = TDQ_CPU(ts->ts_cpu);
1231 	cg = tdq->tdq_cg;
1232 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1233 	    tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1234 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1235 		if (cg->cg_flags & CG_FLAG_THREAD) {
1236 			CPUSET_FOREACH(cpu, cg->cg_mask) {
1237 				if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1238 					break;
1239 			}
1240 		} else
1241 			cpu = INT_MAX;
1242 		if (cpu > mp_maxid) {
1243 			SCHED_STAT_INC(pickcpu_idle_affinity);
1244 			return (ts->ts_cpu);
1245 		}
1246 	}
1247 	/*
1248 	 * Search for the last level cache CPU group in the tree.
1249 	 * Skip caches with expired affinity time and SMT groups.
1250 	 * Affinity to higher level caches will be handled less aggressively.
1251 	 */
1252 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1253 		if (cg->cg_flags & CG_FLAG_THREAD)
1254 			continue;
1255 		if (!SCHED_AFFINITY(ts, cg->cg_level))
1256 			continue;
1257 		ccg = cg;
1258 	}
1259 	if (ccg != NULL)
1260 		cg = ccg;
1261 	cpu = -1;
1262 	/* Search the group for the less loaded idle CPU we can run now. */
1263 	mask = td->td_cpuset->cs_mask;
1264 	if (cg != NULL && cg != cpu_top &&
1265 	    CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1266 		cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1267 		    INT_MAX, ts->ts_cpu);
1268 	/* Search globally for the less loaded CPU we can run now. */
1269 	if (cpu == -1)
1270 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1271 	/* Search globally for the less loaded CPU. */
1272 	if (cpu == -1)
1273 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1274 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1275 	/*
1276 	 * Compare the lowest loaded cpu to current cpu.
1277 	 */
1278 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1279 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1280 	    TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1281 		SCHED_STAT_INC(pickcpu_local);
1282 		cpu = self;
1283 	} else
1284 		SCHED_STAT_INC(pickcpu_lowest);
1285 	if (cpu != ts->ts_cpu)
1286 		SCHED_STAT_INC(pickcpu_migration);
1287 	return (cpu);
1288 }
1289 #endif
1290 
1291 /*
1292  * Pick the highest priority task we have and return it.
1293  */
1294 static struct thread *
1295 tdq_choose(struct tdq *tdq)
1296 {
1297 	struct thread *td;
1298 
1299 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1300 	td = runq_choose(&tdq->tdq_realtime);
1301 	if (td != NULL)
1302 		return (td);
1303 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1304 	if (td != NULL) {
1305 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1306 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1307 		    td->td_priority));
1308 		return (td);
1309 	}
1310 	td = runq_choose(&tdq->tdq_idle);
1311 	if (td != NULL) {
1312 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1313 		    ("tdq_choose: Invalid priority on idle queue %d",
1314 		    td->td_priority));
1315 		return (td);
1316 	}
1317 
1318 	return (NULL);
1319 }
1320 
1321 /*
1322  * Initialize a thread queue.
1323  */
1324 static void
1325 tdq_setup(struct tdq *tdq)
1326 {
1327 
1328 	if (bootverbose)
1329 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1330 	runq_init(&tdq->tdq_realtime);
1331 	runq_init(&tdq->tdq_timeshare);
1332 	runq_init(&tdq->tdq_idle);
1333 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1334 	    "sched lock %d", (int)TDQ_ID(tdq));
1335 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1336 	    MTX_SPIN | MTX_RECURSE);
1337 #ifdef KTR
1338 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1339 	    "CPU %d load", (int)TDQ_ID(tdq));
1340 #endif
1341 }
1342 
1343 #ifdef SMP
1344 static void
1345 sched_setup_smp(void)
1346 {
1347 	struct tdq *tdq;
1348 	int i;
1349 
1350 	cpu_top = smp_topo();
1351 	CPU_FOREACH(i) {
1352 		tdq = TDQ_CPU(i);
1353 		tdq_setup(tdq);
1354 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1355 		if (tdq->tdq_cg == NULL)
1356 			panic("Can't find cpu group for %d\n", i);
1357 	}
1358 	balance_tdq = TDQ_SELF();
1359 	sched_balance();
1360 }
1361 #endif
1362 
1363 /*
1364  * Setup the thread queues and initialize the topology based on MD
1365  * information.
1366  */
1367 static void
1368 sched_setup(void *dummy)
1369 {
1370 	struct tdq *tdq;
1371 
1372 	tdq = TDQ_SELF();
1373 #ifdef SMP
1374 	sched_setup_smp();
1375 #else
1376 	tdq_setup(tdq);
1377 #endif
1378 
1379 	/* Add thread0's load since it's running. */
1380 	TDQ_LOCK(tdq);
1381 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1382 	tdq_load_add(tdq, &thread0);
1383 	tdq->tdq_lowpri = thread0.td_priority;
1384 	TDQ_UNLOCK(tdq);
1385 }
1386 
1387 /*
1388  * This routine determines time constants after stathz and hz are setup.
1389  */
1390 /* ARGSUSED */
1391 static void
1392 sched_initticks(void *dummy)
1393 {
1394 	int incr;
1395 
1396 	realstathz = stathz ? stathz : hz;
1397 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1398 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1399 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1400 	    realstathz);
1401 
1402 	/*
1403 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1404 	 * hz not being evenly divisible by stathz on all platforms.
1405 	 */
1406 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1407 	/*
1408 	 * This does not work for values of stathz that are more than
1409 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1410 	 */
1411 	if (incr == 0)
1412 		incr = 1;
1413 	tickincr = incr;
1414 #ifdef SMP
1415 	/*
1416 	 * Set the default balance interval now that we know
1417 	 * what realstathz is.
1418 	 */
1419 	balance_interval = realstathz;
1420 	affinity = SCHED_AFFINITY_DEFAULT;
1421 #endif
1422 	if (sched_idlespinthresh < 0)
1423 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1424 }
1425 
1426 
1427 /*
1428  * This is the core of the interactivity algorithm.  Determines a score based
1429  * on past behavior.  It is the ratio of sleep time to run time scaled to
1430  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1431  * differs from the cpu usage because it does not account for time spent
1432  * waiting on a run-queue.  Would be prettier if we had floating point.
1433  */
1434 static int
1435 sched_interact_score(struct thread *td)
1436 {
1437 	struct td_sched *ts;
1438 	int div;
1439 
1440 	ts = td->td_sched;
1441 	/*
1442 	 * The score is only needed if this is likely to be an interactive
1443 	 * task.  Don't go through the expense of computing it if there's
1444 	 * no chance.
1445 	 */
1446 	if (sched_interact <= SCHED_INTERACT_HALF &&
1447 		ts->ts_runtime >= ts->ts_slptime)
1448 			return (SCHED_INTERACT_HALF);
1449 
1450 	if (ts->ts_runtime > ts->ts_slptime) {
1451 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1452 		return (SCHED_INTERACT_HALF +
1453 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1454 	}
1455 	if (ts->ts_slptime > ts->ts_runtime) {
1456 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1457 		return (ts->ts_runtime / div);
1458 	}
1459 	/* runtime == slptime */
1460 	if (ts->ts_runtime)
1461 		return (SCHED_INTERACT_HALF);
1462 
1463 	/*
1464 	 * This can happen if slptime and runtime are 0.
1465 	 */
1466 	return (0);
1467 
1468 }
1469 
1470 /*
1471  * Scale the scheduling priority according to the "interactivity" of this
1472  * process.
1473  */
1474 static void
1475 sched_priority(struct thread *td)
1476 {
1477 	int score;
1478 	int pri;
1479 
1480 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1481 		return;
1482 	/*
1483 	 * If the score is interactive we place the thread in the realtime
1484 	 * queue with a priority that is less than kernel and interrupt
1485 	 * priorities.  These threads are not subject to nice restrictions.
1486 	 *
1487 	 * Scores greater than this are placed on the normal timeshare queue
1488 	 * where the priority is partially decided by the most recent cpu
1489 	 * utilization and the rest is decided by nice value.
1490 	 *
1491 	 * The nice value of the process has a linear effect on the calculated
1492 	 * score.  Negative nice values make it easier for a thread to be
1493 	 * considered interactive.
1494 	 */
1495 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1496 	if (score < sched_interact) {
1497 		pri = PRI_MIN_INTERACT;
1498 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1499 		    sched_interact) * score;
1500 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1501 		    ("sched_priority: invalid interactive priority %d score %d",
1502 		    pri, score));
1503 	} else {
1504 		pri = SCHED_PRI_MIN;
1505 		if (td->td_sched->ts_ticks)
1506 			pri += min(SCHED_PRI_TICKS(td->td_sched),
1507 			    SCHED_PRI_RANGE - 1);
1508 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1509 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1510 		    ("sched_priority: invalid priority %d: nice %d, "
1511 		    "ticks %d ftick %d ltick %d tick pri %d",
1512 		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1513 		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1514 		    SCHED_PRI_TICKS(td->td_sched)));
1515 	}
1516 	sched_user_prio(td, pri);
1517 
1518 	return;
1519 }
1520 
1521 /*
1522  * This routine enforces a maximum limit on the amount of scheduling history
1523  * kept.  It is called after either the slptime or runtime is adjusted.  This
1524  * function is ugly due to integer math.
1525  */
1526 static void
1527 sched_interact_update(struct thread *td)
1528 {
1529 	struct td_sched *ts;
1530 	u_int sum;
1531 
1532 	ts = td->td_sched;
1533 	sum = ts->ts_runtime + ts->ts_slptime;
1534 	if (sum < SCHED_SLP_RUN_MAX)
1535 		return;
1536 	/*
1537 	 * This only happens from two places:
1538 	 * 1) We have added an unusual amount of run time from fork_exit.
1539 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1540 	 */
1541 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1542 		if (ts->ts_runtime > ts->ts_slptime) {
1543 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1544 			ts->ts_slptime = 1;
1545 		} else {
1546 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1547 			ts->ts_runtime = 1;
1548 		}
1549 		return;
1550 	}
1551 	/*
1552 	 * If we have exceeded by more than 1/5th then the algorithm below
1553 	 * will not bring us back into range.  Dividing by two here forces
1554 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1555 	 */
1556 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1557 		ts->ts_runtime /= 2;
1558 		ts->ts_slptime /= 2;
1559 		return;
1560 	}
1561 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1562 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1563 }
1564 
1565 /*
1566  * Scale back the interactivity history when a child thread is created.  The
1567  * history is inherited from the parent but the thread may behave totally
1568  * differently.  For example, a shell spawning a compiler process.  We want
1569  * to learn that the compiler is behaving badly very quickly.
1570  */
1571 static void
1572 sched_interact_fork(struct thread *td)
1573 {
1574 	int ratio;
1575 	int sum;
1576 
1577 	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1578 	if (sum > SCHED_SLP_RUN_FORK) {
1579 		ratio = sum / SCHED_SLP_RUN_FORK;
1580 		td->td_sched->ts_runtime /= ratio;
1581 		td->td_sched->ts_slptime /= ratio;
1582 	}
1583 }
1584 
1585 /*
1586  * Called from proc0_init() to setup the scheduler fields.
1587  */
1588 void
1589 schedinit(void)
1590 {
1591 
1592 	/*
1593 	 * Set up the scheduler specific parts of proc0.
1594 	 */
1595 	proc0.p_sched = NULL; /* XXX */
1596 	thread0.td_sched = &td_sched0;
1597 	td_sched0.ts_ltick = ticks;
1598 	td_sched0.ts_ftick = ticks;
1599 	td_sched0.ts_slice = 0;
1600 }
1601 
1602 /*
1603  * This is only somewhat accurate since given many processes of the same
1604  * priority they will switch when their slices run out, which will be
1605  * at most sched_slice stathz ticks.
1606  */
1607 int
1608 sched_rr_interval(void)
1609 {
1610 
1611 	/* Convert sched_slice from stathz to hz. */
1612 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1613 }
1614 
1615 /*
1616  * Update the percent cpu tracking information when it is requested or
1617  * the total history exceeds the maximum.  We keep a sliding history of
1618  * tick counts that slowly decays.  This is less precise than the 4BSD
1619  * mechanism since it happens with less regular and frequent events.
1620  */
1621 static void
1622 sched_pctcpu_update(struct td_sched *ts, int run)
1623 {
1624 	int t = ticks;
1625 
1626 	if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1627 		ts->ts_ticks = 0;
1628 		ts->ts_ftick = t - SCHED_TICK_TARG;
1629 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1630 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1631 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1632 		ts->ts_ftick = t - SCHED_TICK_TARG;
1633 	}
1634 	if (run)
1635 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1636 	ts->ts_ltick = t;
1637 }
1638 
1639 /*
1640  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1641  * if necessary.  This is the back-end for several priority related
1642  * functions.
1643  */
1644 static void
1645 sched_thread_priority(struct thread *td, u_char prio)
1646 {
1647 	struct td_sched *ts;
1648 	struct tdq *tdq;
1649 	int oldpri;
1650 
1651 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1652 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1653 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1654 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1655 	if (td != curthread && prio < td->td_priority) {
1656 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1657 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1658 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1659 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1660 		    curthread);
1661 	}
1662 	ts = td->td_sched;
1663 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1664 	if (td->td_priority == prio)
1665 		return;
1666 	/*
1667 	 * If the priority has been elevated due to priority
1668 	 * propagation, we may have to move ourselves to a new
1669 	 * queue.  This could be optimized to not re-add in some
1670 	 * cases.
1671 	 */
1672 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1673 		sched_rem(td);
1674 		td->td_priority = prio;
1675 		sched_add(td, SRQ_BORROWING);
1676 		return;
1677 	}
1678 	/*
1679 	 * If the thread is currently running we may have to adjust the lowpri
1680 	 * information so other cpus are aware of our current priority.
1681 	 */
1682 	if (TD_IS_RUNNING(td)) {
1683 		tdq = TDQ_CPU(ts->ts_cpu);
1684 		oldpri = td->td_priority;
1685 		td->td_priority = prio;
1686 		if (prio < tdq->tdq_lowpri)
1687 			tdq->tdq_lowpri = prio;
1688 		else if (tdq->tdq_lowpri == oldpri)
1689 			tdq_setlowpri(tdq, td);
1690 		return;
1691 	}
1692 	td->td_priority = prio;
1693 }
1694 
1695 /*
1696  * Update a thread's priority when it is lent another thread's
1697  * priority.
1698  */
1699 void
1700 sched_lend_prio(struct thread *td, u_char prio)
1701 {
1702 
1703 	td->td_flags |= TDF_BORROWING;
1704 	sched_thread_priority(td, prio);
1705 }
1706 
1707 /*
1708  * Restore a thread's priority when priority propagation is
1709  * over.  The prio argument is the minimum priority the thread
1710  * needs to have to satisfy other possible priority lending
1711  * requests.  If the thread's regular priority is less
1712  * important than prio, the thread will keep a priority boost
1713  * of prio.
1714  */
1715 void
1716 sched_unlend_prio(struct thread *td, u_char prio)
1717 {
1718 	u_char base_pri;
1719 
1720 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1721 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1722 		base_pri = td->td_user_pri;
1723 	else
1724 		base_pri = td->td_base_pri;
1725 	if (prio >= base_pri) {
1726 		td->td_flags &= ~TDF_BORROWING;
1727 		sched_thread_priority(td, base_pri);
1728 	} else
1729 		sched_lend_prio(td, prio);
1730 }
1731 
1732 /*
1733  * Standard entry for setting the priority to an absolute value.
1734  */
1735 void
1736 sched_prio(struct thread *td, u_char prio)
1737 {
1738 	u_char oldprio;
1739 
1740 	/* First, update the base priority. */
1741 	td->td_base_pri = prio;
1742 
1743 	/*
1744 	 * If the thread is borrowing another thread's priority, don't
1745 	 * ever lower the priority.
1746 	 */
1747 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1748 		return;
1749 
1750 	/* Change the real priority. */
1751 	oldprio = td->td_priority;
1752 	sched_thread_priority(td, prio);
1753 
1754 	/*
1755 	 * If the thread is on a turnstile, then let the turnstile update
1756 	 * its state.
1757 	 */
1758 	if (TD_ON_LOCK(td) && oldprio != prio)
1759 		turnstile_adjust(td, oldprio);
1760 }
1761 
1762 /*
1763  * Set the base user priority, does not effect current running priority.
1764  */
1765 void
1766 sched_user_prio(struct thread *td, u_char prio)
1767 {
1768 
1769 	td->td_base_user_pri = prio;
1770 	if (td->td_lend_user_pri <= prio)
1771 		return;
1772 	td->td_user_pri = prio;
1773 }
1774 
1775 void
1776 sched_lend_user_prio(struct thread *td, u_char prio)
1777 {
1778 
1779 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1780 	td->td_lend_user_pri = prio;
1781 	td->td_user_pri = min(prio, td->td_base_user_pri);
1782 	if (td->td_priority > td->td_user_pri)
1783 		sched_prio(td, td->td_user_pri);
1784 	else if (td->td_priority != td->td_user_pri)
1785 		td->td_flags |= TDF_NEEDRESCHED;
1786 }
1787 
1788 /*
1789  * Handle migration from sched_switch().  This happens only for
1790  * cpu binding.
1791  */
1792 static struct mtx *
1793 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1794 {
1795 	struct tdq *tdn;
1796 
1797 	tdn = TDQ_CPU(td->td_sched->ts_cpu);
1798 #ifdef SMP
1799 	tdq_load_rem(tdq, td);
1800 	/*
1801 	 * Do the lock dance required to avoid LOR.  We grab an extra
1802 	 * spinlock nesting to prevent preemption while we're
1803 	 * not holding either run-queue lock.
1804 	 */
1805 	spinlock_enter();
1806 	thread_lock_block(td);	/* This releases the lock on tdq. */
1807 
1808 	/*
1809 	 * Acquire both run-queue locks before placing the thread on the new
1810 	 * run-queue to avoid deadlocks created by placing a thread with a
1811 	 * blocked lock on the run-queue of a remote processor.  The deadlock
1812 	 * occurs when a third processor attempts to lock the two queues in
1813 	 * question while the target processor is spinning with its own
1814 	 * run-queue lock held while waiting for the blocked lock to clear.
1815 	 */
1816 	tdq_lock_pair(tdn, tdq);
1817 	tdq_add(tdn, td, flags);
1818 	tdq_notify(tdn, td);
1819 	TDQ_UNLOCK(tdn);
1820 	spinlock_exit();
1821 #endif
1822 	return (TDQ_LOCKPTR(tdn));
1823 }
1824 
1825 /*
1826  * Variadic version of thread_lock_unblock() that does not assume td_lock
1827  * is blocked.
1828  */
1829 static inline void
1830 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1831 {
1832 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1833 	    (uintptr_t)mtx);
1834 }
1835 
1836 /*
1837  * Switch threads.  This function has to handle threads coming in while
1838  * blocked for some reason, running, or idle.  It also must deal with
1839  * migrating a thread from one queue to another as running threads may
1840  * be assigned elsewhere via binding.
1841  */
1842 void
1843 sched_switch(struct thread *td, struct thread *newtd, int flags)
1844 {
1845 	struct tdq *tdq;
1846 	struct td_sched *ts;
1847 	struct mtx *mtx;
1848 	int srqflag;
1849 	int cpuid, preempted;
1850 
1851 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1852 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1853 
1854 	cpuid = PCPU_GET(cpuid);
1855 	tdq = TDQ_CPU(cpuid);
1856 	ts = td->td_sched;
1857 	mtx = td->td_lock;
1858 	sched_pctcpu_update(ts, 1);
1859 	ts->ts_rltick = ticks;
1860 	td->td_lastcpu = td->td_oncpu;
1861 	td->td_oncpu = NOCPU;
1862 	preempted = !(td->td_flags & TDF_SLICEEND);
1863 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1864 	td->td_owepreempt = 0;
1865 	if (!TD_IS_IDLETHREAD(td))
1866 		tdq->tdq_switchcnt++;
1867 	/*
1868 	 * The lock pointer in an idle thread should never change.  Reset it
1869 	 * to CAN_RUN as well.
1870 	 */
1871 	if (TD_IS_IDLETHREAD(td)) {
1872 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1873 		TD_SET_CAN_RUN(td);
1874 	} else if (TD_IS_RUNNING(td)) {
1875 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1876 		srqflag = preempted ?
1877 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1878 		    SRQ_OURSELF|SRQ_YIELDING;
1879 #ifdef SMP
1880 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1881 			ts->ts_cpu = sched_pickcpu(td, 0);
1882 #endif
1883 		if (ts->ts_cpu == cpuid)
1884 			tdq_runq_add(tdq, td, srqflag);
1885 		else {
1886 			KASSERT(THREAD_CAN_MIGRATE(td) ||
1887 			    (ts->ts_flags & TSF_BOUND) != 0,
1888 			    ("Thread %p shouldn't migrate", td));
1889 			mtx = sched_switch_migrate(tdq, td, srqflag);
1890 		}
1891 	} else {
1892 		/* This thread must be going to sleep. */
1893 		TDQ_LOCK(tdq);
1894 		mtx = thread_lock_block(td);
1895 		tdq_load_rem(tdq, td);
1896 	}
1897 	/*
1898 	 * We enter here with the thread blocked and assigned to the
1899 	 * appropriate cpu run-queue or sleep-queue and with the current
1900 	 * thread-queue locked.
1901 	 */
1902 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1903 	newtd = choosethread();
1904 	/*
1905 	 * Call the MD code to switch contexts if necessary.
1906 	 */
1907 	if (td != newtd) {
1908 #ifdef	HWPMC_HOOKS
1909 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1910 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1911 #endif
1912 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1913 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1914 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1915 		sched_pctcpu_update(newtd->td_sched, 0);
1916 
1917 #ifdef KDTRACE_HOOKS
1918 		/*
1919 		 * If DTrace has set the active vtime enum to anything
1920 		 * other than INACTIVE (0), then it should have set the
1921 		 * function to call.
1922 		 */
1923 		if (dtrace_vtime_active)
1924 			(*dtrace_vtime_switch_func)(newtd);
1925 #endif
1926 
1927 		cpu_switch(td, newtd, mtx);
1928 		/*
1929 		 * We may return from cpu_switch on a different cpu.  However,
1930 		 * we always return with td_lock pointing to the current cpu's
1931 		 * run queue lock.
1932 		 */
1933 		cpuid = PCPU_GET(cpuid);
1934 		tdq = TDQ_CPU(cpuid);
1935 		lock_profile_obtain_lock_success(
1936 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1937 
1938 		SDT_PROBE0(sched, , , on__cpu);
1939 #ifdef	HWPMC_HOOKS
1940 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1941 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1942 #endif
1943 	} else {
1944 		thread_unblock_switch(td, mtx);
1945 		SDT_PROBE0(sched, , , remain__cpu);
1946 	}
1947 	/*
1948 	 * Assert that all went well and return.
1949 	 */
1950 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1951 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1952 	td->td_oncpu = cpuid;
1953 }
1954 
1955 /*
1956  * Adjust thread priorities as a result of a nice request.
1957  */
1958 void
1959 sched_nice(struct proc *p, int nice)
1960 {
1961 	struct thread *td;
1962 
1963 	PROC_LOCK_ASSERT(p, MA_OWNED);
1964 
1965 	p->p_nice = nice;
1966 	FOREACH_THREAD_IN_PROC(p, td) {
1967 		thread_lock(td);
1968 		sched_priority(td);
1969 		sched_prio(td, td->td_base_user_pri);
1970 		thread_unlock(td);
1971 	}
1972 }
1973 
1974 /*
1975  * Record the sleep time for the interactivity scorer.
1976  */
1977 void
1978 sched_sleep(struct thread *td, int prio)
1979 {
1980 
1981 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1982 
1983 	td->td_slptick = ticks;
1984 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
1985 		td->td_flags |= TDF_CANSWAP;
1986 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1987 		return;
1988 	if (static_boost == 1 && prio)
1989 		sched_prio(td, prio);
1990 	else if (static_boost && td->td_priority > static_boost)
1991 		sched_prio(td, static_boost);
1992 }
1993 
1994 /*
1995  * Schedule a thread to resume execution and record how long it voluntarily
1996  * slept.  We also update the pctcpu, interactivity, and priority.
1997  */
1998 void
1999 sched_wakeup(struct thread *td)
2000 {
2001 	struct td_sched *ts;
2002 	int slptick;
2003 
2004 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2005 	ts = td->td_sched;
2006 	td->td_flags &= ~TDF_CANSWAP;
2007 	/*
2008 	 * If we slept for more than a tick update our interactivity and
2009 	 * priority.
2010 	 */
2011 	slptick = td->td_slptick;
2012 	td->td_slptick = 0;
2013 	if (slptick && slptick != ticks) {
2014 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2015 		sched_interact_update(td);
2016 		sched_pctcpu_update(ts, 0);
2017 	}
2018 	/*
2019 	 * Reset the slice value since we slept and advanced the round-robin.
2020 	 */
2021 	ts->ts_slice = 0;
2022 	sched_add(td, SRQ_BORING);
2023 }
2024 
2025 /*
2026  * Penalize the parent for creating a new child and initialize the child's
2027  * priority.
2028  */
2029 void
2030 sched_fork(struct thread *td, struct thread *child)
2031 {
2032 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2033 	sched_pctcpu_update(td->td_sched, 1);
2034 	sched_fork_thread(td, child);
2035 	/*
2036 	 * Penalize the parent and child for forking.
2037 	 */
2038 	sched_interact_fork(child);
2039 	sched_priority(child);
2040 	td->td_sched->ts_runtime += tickincr;
2041 	sched_interact_update(td);
2042 	sched_priority(td);
2043 }
2044 
2045 /*
2046  * Fork a new thread, may be within the same process.
2047  */
2048 void
2049 sched_fork_thread(struct thread *td, struct thread *child)
2050 {
2051 	struct td_sched *ts;
2052 	struct td_sched *ts2;
2053 	struct tdq *tdq;
2054 
2055 	tdq = TDQ_SELF();
2056 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2057 	/*
2058 	 * Initialize child.
2059 	 */
2060 	ts = td->td_sched;
2061 	ts2 = child->td_sched;
2062 	child->td_lock = TDQ_LOCKPTR(tdq);
2063 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2064 	ts2->ts_cpu = ts->ts_cpu;
2065 	ts2->ts_flags = 0;
2066 	/*
2067 	 * Grab our parents cpu estimation information.
2068 	 */
2069 	ts2->ts_ticks = ts->ts_ticks;
2070 	ts2->ts_ltick = ts->ts_ltick;
2071 	ts2->ts_ftick = ts->ts_ftick;
2072 	/*
2073 	 * Do not inherit any borrowed priority from the parent.
2074 	 */
2075 	child->td_priority = child->td_base_pri;
2076 	/*
2077 	 * And update interactivity score.
2078 	 */
2079 	ts2->ts_slptime = ts->ts_slptime;
2080 	ts2->ts_runtime = ts->ts_runtime;
2081 	/* Attempt to quickly learn interactivity. */
2082 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2083 #ifdef KTR
2084 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2085 #endif
2086 }
2087 
2088 /*
2089  * Adjust the priority class of a thread.
2090  */
2091 void
2092 sched_class(struct thread *td, int class)
2093 {
2094 
2095 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2096 	if (td->td_pri_class == class)
2097 		return;
2098 	td->td_pri_class = class;
2099 }
2100 
2101 /*
2102  * Return some of the child's priority and interactivity to the parent.
2103  */
2104 void
2105 sched_exit(struct proc *p, struct thread *child)
2106 {
2107 	struct thread *td;
2108 
2109 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2110 	    "prio:%d", child->td_priority);
2111 	PROC_LOCK_ASSERT(p, MA_OWNED);
2112 	td = FIRST_THREAD_IN_PROC(p);
2113 	sched_exit_thread(td, child);
2114 }
2115 
2116 /*
2117  * Penalize another thread for the time spent on this one.  This helps to
2118  * worsen the priority and interactivity of processes which schedule batch
2119  * jobs such as make.  This has little effect on the make process itself but
2120  * causes new processes spawned by it to receive worse scores immediately.
2121  */
2122 void
2123 sched_exit_thread(struct thread *td, struct thread *child)
2124 {
2125 
2126 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2127 	    "prio:%d", child->td_priority);
2128 	/*
2129 	 * Give the child's runtime to the parent without returning the
2130 	 * sleep time as a penalty to the parent.  This causes shells that
2131 	 * launch expensive things to mark their children as expensive.
2132 	 */
2133 	thread_lock(td);
2134 	td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2135 	sched_interact_update(td);
2136 	sched_priority(td);
2137 	thread_unlock(td);
2138 }
2139 
2140 void
2141 sched_preempt(struct thread *td)
2142 {
2143 	struct tdq *tdq;
2144 
2145 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2146 
2147 	thread_lock(td);
2148 	tdq = TDQ_SELF();
2149 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2150 	tdq->tdq_ipipending = 0;
2151 	if (td->td_priority > tdq->tdq_lowpri) {
2152 		int flags;
2153 
2154 		flags = SW_INVOL | SW_PREEMPT;
2155 		if (td->td_critnest > 1)
2156 			td->td_owepreempt = 1;
2157 		else if (TD_IS_IDLETHREAD(td))
2158 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2159 		else
2160 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2161 	}
2162 	thread_unlock(td);
2163 }
2164 
2165 /*
2166  * Fix priorities on return to user-space.  Priorities may be elevated due
2167  * to static priorities in msleep() or similar.
2168  */
2169 void
2170 sched_userret(struct thread *td)
2171 {
2172 	/*
2173 	 * XXX we cheat slightly on the locking here to avoid locking in
2174 	 * the usual case.  Setting td_priority here is essentially an
2175 	 * incomplete workaround for not setting it properly elsewhere.
2176 	 * Now that some interrupt handlers are threads, not setting it
2177 	 * properly elsewhere can clobber it in the window between setting
2178 	 * it here and returning to user mode, so don't waste time setting
2179 	 * it perfectly here.
2180 	 */
2181 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2182 	    ("thread with borrowed priority returning to userland"));
2183 	if (td->td_priority != td->td_user_pri) {
2184 		thread_lock(td);
2185 		td->td_priority = td->td_user_pri;
2186 		td->td_base_pri = td->td_user_pri;
2187 		tdq_setlowpri(TDQ_SELF(), td);
2188 		thread_unlock(td);
2189         }
2190 }
2191 
2192 /*
2193  * Handle a stathz tick.  This is really only relevant for timeshare
2194  * threads.
2195  */
2196 void
2197 sched_clock(struct thread *td)
2198 {
2199 	struct tdq *tdq;
2200 	struct td_sched *ts;
2201 
2202 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2203 	tdq = TDQ_SELF();
2204 #ifdef SMP
2205 	/*
2206 	 * We run the long term load balancer infrequently on the first cpu.
2207 	 */
2208 	if (balance_tdq == tdq) {
2209 		if (balance_ticks && --balance_ticks == 0)
2210 			sched_balance();
2211 	}
2212 #endif
2213 	/*
2214 	 * Save the old switch count so we have a record of the last ticks
2215 	 * activity.   Initialize the new switch count based on our load.
2216 	 * If there is some activity seed it to reflect that.
2217 	 */
2218 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2219 	tdq->tdq_switchcnt = tdq->tdq_load;
2220 	/*
2221 	 * Advance the insert index once for each tick to ensure that all
2222 	 * threads get a chance to run.
2223 	 */
2224 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2225 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2226 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2227 			tdq->tdq_ridx = tdq->tdq_idx;
2228 	}
2229 	ts = td->td_sched;
2230 	sched_pctcpu_update(ts, 1);
2231 	if (td->td_pri_class & PRI_FIFO_BIT)
2232 		return;
2233 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2234 		/*
2235 		 * We used a tick; charge it to the thread so
2236 		 * that we can compute our interactivity.
2237 		 */
2238 		td->td_sched->ts_runtime += tickincr;
2239 		sched_interact_update(td);
2240 		sched_priority(td);
2241 	}
2242 
2243 	/*
2244 	 * Force a context switch if the current thread has used up a full
2245 	 * time slice (default is 100ms).
2246 	 */
2247 	if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2248 		ts->ts_slice = 0;
2249 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2250 	}
2251 }
2252 
2253 /*
2254  * Called once per hz tick.
2255  */
2256 void
2257 sched_tick(int cnt)
2258 {
2259 
2260 }
2261 
2262 /*
2263  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2264  * cooperative idle threads.
2265  */
2266 int
2267 sched_runnable(void)
2268 {
2269 	struct tdq *tdq;
2270 	int load;
2271 
2272 	load = 1;
2273 
2274 	tdq = TDQ_SELF();
2275 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2276 		if (tdq->tdq_load > 0)
2277 			goto out;
2278 	} else
2279 		if (tdq->tdq_load - 1 > 0)
2280 			goto out;
2281 	load = 0;
2282 out:
2283 	return (load);
2284 }
2285 
2286 /*
2287  * Choose the highest priority thread to run.  The thread is removed from
2288  * the run-queue while running however the load remains.  For SMP we set
2289  * the tdq in the global idle bitmask if it idles here.
2290  */
2291 struct thread *
2292 sched_choose(void)
2293 {
2294 	struct thread *td;
2295 	struct tdq *tdq;
2296 
2297 	tdq = TDQ_SELF();
2298 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2299 	td = tdq_choose(tdq);
2300 	if (td) {
2301 		tdq_runq_rem(tdq, td);
2302 		tdq->tdq_lowpri = td->td_priority;
2303 		return (td);
2304 	}
2305 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2306 	return (PCPU_GET(idlethread));
2307 }
2308 
2309 /*
2310  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2311  * we always request it once we exit a critical section.
2312  */
2313 static inline void
2314 sched_setpreempt(struct thread *td)
2315 {
2316 	struct thread *ctd;
2317 	int cpri;
2318 	int pri;
2319 
2320 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2321 
2322 	ctd = curthread;
2323 	pri = td->td_priority;
2324 	cpri = ctd->td_priority;
2325 	if (pri < cpri)
2326 		ctd->td_flags |= TDF_NEEDRESCHED;
2327 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2328 		return;
2329 	if (!sched_shouldpreempt(pri, cpri, 0))
2330 		return;
2331 	ctd->td_owepreempt = 1;
2332 }
2333 
2334 /*
2335  * Add a thread to a thread queue.  Select the appropriate runq and add the
2336  * thread to it.  This is the internal function called when the tdq is
2337  * predetermined.
2338  */
2339 void
2340 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2341 {
2342 
2343 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2344 	KASSERT((td->td_inhibitors == 0),
2345 	    ("sched_add: trying to run inhibited thread"));
2346 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2347 	    ("sched_add: bad thread state"));
2348 	KASSERT(td->td_flags & TDF_INMEM,
2349 	    ("sched_add: thread swapped out"));
2350 
2351 	if (td->td_priority < tdq->tdq_lowpri)
2352 		tdq->tdq_lowpri = td->td_priority;
2353 	tdq_runq_add(tdq, td, flags);
2354 	tdq_load_add(tdq, td);
2355 }
2356 
2357 /*
2358  * Select the target thread queue and add a thread to it.  Request
2359  * preemption or IPI a remote processor if required.
2360  */
2361 void
2362 sched_add(struct thread *td, int flags)
2363 {
2364 	struct tdq *tdq;
2365 #ifdef SMP
2366 	int cpu;
2367 #endif
2368 
2369 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2370 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2371 	    sched_tdname(curthread));
2372 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2373 	    KTR_ATTR_LINKED, sched_tdname(td));
2374 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2375 	    flags & SRQ_PREEMPTED);
2376 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2377 	/*
2378 	 * Recalculate the priority before we select the target cpu or
2379 	 * run-queue.
2380 	 */
2381 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2382 		sched_priority(td);
2383 #ifdef SMP
2384 	/*
2385 	 * Pick the destination cpu and if it isn't ours transfer to the
2386 	 * target cpu.
2387 	 */
2388 	cpu = sched_pickcpu(td, flags);
2389 	tdq = sched_setcpu(td, cpu, flags);
2390 	tdq_add(tdq, td, flags);
2391 	if (cpu != PCPU_GET(cpuid)) {
2392 		tdq_notify(tdq, td);
2393 		return;
2394 	}
2395 #else
2396 	tdq = TDQ_SELF();
2397 	TDQ_LOCK(tdq);
2398 	/*
2399 	 * Now that the thread is moving to the run-queue, set the lock
2400 	 * to the scheduler's lock.
2401 	 */
2402 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2403 	tdq_add(tdq, td, flags);
2404 #endif
2405 	if (!(flags & SRQ_YIELDING))
2406 		sched_setpreempt(td);
2407 }
2408 
2409 /*
2410  * Remove a thread from a run-queue without running it.  This is used
2411  * when we're stealing a thread from a remote queue.  Otherwise all threads
2412  * exit by calling sched_exit_thread() and sched_throw() themselves.
2413  */
2414 void
2415 sched_rem(struct thread *td)
2416 {
2417 	struct tdq *tdq;
2418 
2419 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2420 	    "prio:%d", td->td_priority);
2421 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2422 	tdq = TDQ_CPU(td->td_sched->ts_cpu);
2423 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2424 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2425 	KASSERT(TD_ON_RUNQ(td),
2426 	    ("sched_rem: thread not on run queue"));
2427 	tdq_runq_rem(tdq, td);
2428 	tdq_load_rem(tdq, td);
2429 	TD_SET_CAN_RUN(td);
2430 	if (td->td_priority == tdq->tdq_lowpri)
2431 		tdq_setlowpri(tdq, NULL);
2432 }
2433 
2434 /*
2435  * Fetch cpu utilization information.  Updates on demand.
2436  */
2437 fixpt_t
2438 sched_pctcpu(struct thread *td)
2439 {
2440 	fixpt_t pctcpu;
2441 	struct td_sched *ts;
2442 
2443 	pctcpu = 0;
2444 	ts = td->td_sched;
2445 	if (ts == NULL)
2446 		return (0);
2447 
2448 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2449 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2450 	if (ts->ts_ticks) {
2451 		int rtick;
2452 
2453 		/* How many rtick per second ? */
2454 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2455 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2456 	}
2457 
2458 	return (pctcpu);
2459 }
2460 
2461 /*
2462  * Enforce affinity settings for a thread.  Called after adjustments to
2463  * cpumask.
2464  */
2465 void
2466 sched_affinity(struct thread *td)
2467 {
2468 #ifdef SMP
2469 	struct td_sched *ts;
2470 
2471 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2472 	ts = td->td_sched;
2473 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2474 		return;
2475 	if (TD_ON_RUNQ(td)) {
2476 		sched_rem(td);
2477 		sched_add(td, SRQ_BORING);
2478 		return;
2479 	}
2480 	if (!TD_IS_RUNNING(td))
2481 		return;
2482 	/*
2483 	 * Force a switch before returning to userspace.  If the
2484 	 * target thread is not running locally send an ipi to force
2485 	 * the issue.
2486 	 */
2487 	td->td_flags |= TDF_NEEDRESCHED;
2488 	if (td != curthread)
2489 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2490 #endif
2491 }
2492 
2493 /*
2494  * Bind a thread to a target cpu.
2495  */
2496 void
2497 sched_bind(struct thread *td, int cpu)
2498 {
2499 	struct td_sched *ts;
2500 
2501 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2502 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2503 	ts = td->td_sched;
2504 	if (ts->ts_flags & TSF_BOUND)
2505 		sched_unbind(td);
2506 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2507 	ts->ts_flags |= TSF_BOUND;
2508 	sched_pin();
2509 	if (PCPU_GET(cpuid) == cpu)
2510 		return;
2511 	ts->ts_cpu = cpu;
2512 	/* When we return from mi_switch we'll be on the correct cpu. */
2513 	mi_switch(SW_VOL, NULL);
2514 }
2515 
2516 /*
2517  * Release a bound thread.
2518  */
2519 void
2520 sched_unbind(struct thread *td)
2521 {
2522 	struct td_sched *ts;
2523 
2524 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2525 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2526 	ts = td->td_sched;
2527 	if ((ts->ts_flags & TSF_BOUND) == 0)
2528 		return;
2529 	ts->ts_flags &= ~TSF_BOUND;
2530 	sched_unpin();
2531 }
2532 
2533 int
2534 sched_is_bound(struct thread *td)
2535 {
2536 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2537 	return (td->td_sched->ts_flags & TSF_BOUND);
2538 }
2539 
2540 /*
2541  * Basic yield call.
2542  */
2543 void
2544 sched_relinquish(struct thread *td)
2545 {
2546 	thread_lock(td);
2547 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2548 	thread_unlock(td);
2549 }
2550 
2551 /*
2552  * Return the total system load.
2553  */
2554 int
2555 sched_load(void)
2556 {
2557 #ifdef SMP
2558 	int total;
2559 	int i;
2560 
2561 	total = 0;
2562 	CPU_FOREACH(i)
2563 		total += TDQ_CPU(i)->tdq_sysload;
2564 	return (total);
2565 #else
2566 	return (TDQ_SELF()->tdq_sysload);
2567 #endif
2568 }
2569 
2570 int
2571 sched_sizeof_proc(void)
2572 {
2573 	return (sizeof(struct proc));
2574 }
2575 
2576 int
2577 sched_sizeof_thread(void)
2578 {
2579 	return (sizeof(struct thread) + sizeof(struct td_sched));
2580 }
2581 
2582 #ifdef SMP
2583 #define	TDQ_IDLESPIN(tdq)						\
2584     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2585 #else
2586 #define	TDQ_IDLESPIN(tdq)	1
2587 #endif
2588 
2589 /*
2590  * The actual idle process.
2591  */
2592 void
2593 sched_idletd(void *dummy)
2594 {
2595 	struct thread *td;
2596 	struct tdq *tdq;
2597 	int oldswitchcnt, switchcnt;
2598 	int i;
2599 
2600 	mtx_assert(&Giant, MA_NOTOWNED);
2601 	td = curthread;
2602 	tdq = TDQ_SELF();
2603 	THREAD_NO_SLEEPING();
2604 	oldswitchcnt = -1;
2605 	for (;;) {
2606 		if (tdq->tdq_load) {
2607 			thread_lock(td);
2608 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2609 			thread_unlock(td);
2610 		}
2611 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2612 #ifdef SMP
2613 		if (switchcnt != oldswitchcnt) {
2614 			oldswitchcnt = switchcnt;
2615 			if (tdq_idled(tdq) == 0)
2616 				continue;
2617 		}
2618 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2619 #else
2620 		oldswitchcnt = switchcnt;
2621 #endif
2622 		/*
2623 		 * If we're switching very frequently, spin while checking
2624 		 * for load rather than entering a low power state that
2625 		 * may require an IPI.  However, don't do any busy
2626 		 * loops while on SMT machines as this simply steals
2627 		 * cycles from cores doing useful work.
2628 		 */
2629 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2630 			for (i = 0; i < sched_idlespins; i++) {
2631 				if (tdq->tdq_load)
2632 					break;
2633 				cpu_spinwait();
2634 			}
2635 		}
2636 
2637 		/* If there was context switch during spin, restart it. */
2638 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2639 		if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2640 			continue;
2641 
2642 		/* Run main MD idle handler. */
2643 		tdq->tdq_cpu_idle = 1;
2644 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2645 		tdq->tdq_cpu_idle = 0;
2646 
2647 		/*
2648 		 * Account thread-less hardware interrupts and
2649 		 * other wakeup reasons equal to context switches.
2650 		 */
2651 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2652 		if (switchcnt != oldswitchcnt)
2653 			continue;
2654 		tdq->tdq_switchcnt++;
2655 		oldswitchcnt++;
2656 	}
2657 }
2658 
2659 /*
2660  * A CPU is entering for the first time or a thread is exiting.
2661  */
2662 void
2663 sched_throw(struct thread *td)
2664 {
2665 	struct thread *newtd;
2666 	struct tdq *tdq;
2667 
2668 	tdq = TDQ_SELF();
2669 	if (td == NULL) {
2670 		/* Correct spinlock nesting and acquire the correct lock. */
2671 		TDQ_LOCK(tdq);
2672 		spinlock_exit();
2673 		PCPU_SET(switchtime, cpu_ticks());
2674 		PCPU_SET(switchticks, ticks);
2675 	} else {
2676 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2677 		tdq_load_rem(tdq, td);
2678 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2679 	}
2680 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2681 	newtd = choosethread();
2682 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2683 	cpu_throw(td, newtd);		/* doesn't return */
2684 }
2685 
2686 /*
2687  * This is called from fork_exit().  Just acquire the correct locks and
2688  * let fork do the rest of the work.
2689  */
2690 void
2691 sched_fork_exit(struct thread *td)
2692 {
2693 	struct td_sched *ts;
2694 	struct tdq *tdq;
2695 	int cpuid;
2696 
2697 	/*
2698 	 * Finish setting up thread glue so that it begins execution in a
2699 	 * non-nested critical section with the scheduler lock held.
2700 	 */
2701 	cpuid = PCPU_GET(cpuid);
2702 	tdq = TDQ_CPU(cpuid);
2703 	ts = td->td_sched;
2704 	if (TD_IS_IDLETHREAD(td))
2705 		td->td_lock = TDQ_LOCKPTR(tdq);
2706 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2707 	td->td_oncpu = cpuid;
2708 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2709 	lock_profile_obtain_lock_success(
2710 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2711 }
2712 
2713 /*
2714  * Create on first use to catch odd startup conditons.
2715  */
2716 char *
2717 sched_tdname(struct thread *td)
2718 {
2719 #ifdef KTR
2720 	struct td_sched *ts;
2721 
2722 	ts = td->td_sched;
2723 	if (ts->ts_name[0] == '\0')
2724 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2725 		    "%s tid %d", td->td_name, td->td_tid);
2726 	return (ts->ts_name);
2727 #else
2728 	return (td->td_name);
2729 #endif
2730 }
2731 
2732 #ifdef KTR
2733 void
2734 sched_clear_tdname(struct thread *td)
2735 {
2736 	struct td_sched *ts;
2737 
2738 	ts = td->td_sched;
2739 	ts->ts_name[0] = '\0';
2740 }
2741 #endif
2742 
2743 #ifdef SMP
2744 
2745 /*
2746  * Build the CPU topology dump string. Is recursively called to collect
2747  * the topology tree.
2748  */
2749 static int
2750 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2751     int indent)
2752 {
2753 	char cpusetbuf[CPUSETBUFSIZ];
2754 	int i, first;
2755 
2756 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2757 	    "", 1 + indent / 2, cg->cg_level);
2758 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2759 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2760 	first = TRUE;
2761 	for (i = 0; i < MAXCPU; i++) {
2762 		if (CPU_ISSET(i, &cg->cg_mask)) {
2763 			if (!first)
2764 				sbuf_printf(sb, ", ");
2765 			else
2766 				first = FALSE;
2767 			sbuf_printf(sb, "%d", i);
2768 		}
2769 	}
2770 	sbuf_printf(sb, "</cpu>\n");
2771 
2772 	if (cg->cg_flags != 0) {
2773 		sbuf_printf(sb, "%*s <flags>", indent, "");
2774 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2775 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2776 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2777 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2778 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2779 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2780 		sbuf_printf(sb, "</flags>\n");
2781 	}
2782 
2783 	if (cg->cg_children > 0) {
2784 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2785 		for (i = 0; i < cg->cg_children; i++)
2786 			sysctl_kern_sched_topology_spec_internal(sb,
2787 			    &cg->cg_child[i], indent+2);
2788 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2789 	}
2790 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2791 	return (0);
2792 }
2793 
2794 /*
2795  * Sysctl handler for retrieving topology dump. It's a wrapper for
2796  * the recursive sysctl_kern_smp_topology_spec_internal().
2797  */
2798 static int
2799 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2800 {
2801 	struct sbuf *topo;
2802 	int err;
2803 
2804 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2805 
2806 	topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2807 	if (topo == NULL)
2808 		return (ENOMEM);
2809 
2810 	sbuf_printf(topo, "<groups>\n");
2811 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2812 	sbuf_printf(topo, "</groups>\n");
2813 
2814 	if (err == 0) {
2815 		sbuf_finish(topo);
2816 		err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2817 	}
2818 	sbuf_delete(topo);
2819 	return (err);
2820 }
2821 
2822 #endif
2823 
2824 static int
2825 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2826 {
2827 	int error, new_val, period;
2828 
2829 	period = 1000000 / realstathz;
2830 	new_val = period * sched_slice;
2831 	error = sysctl_handle_int(oidp, &new_val, 0, req);
2832 	if (error != 0 || req->newptr == NULL)
2833 		return (error);
2834 	if (new_val <= 0)
2835 		return (EINVAL);
2836 	sched_slice = imax(1, (new_val + period / 2) / period);
2837 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
2838 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2839 	    realstathz);
2840 	return (0);
2841 }
2842 
2843 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2844 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2845     "Scheduler name");
2846 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2847     NULL, 0, sysctl_kern_quantum, "I",
2848     "Quantum for timeshare threads in microseconds");
2849 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2850     "Quantum for timeshare threads in stathz ticks");
2851 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2852     "Interactivity score threshold");
2853 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2854     &preempt_thresh, 0,
2855     "Maximal (lowest) priority for preemption");
2856 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2857     "Assign static kernel priorities to sleeping threads");
2858 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2859     "Number of times idle thread will spin waiting for new work");
2860 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2861     &sched_idlespinthresh, 0,
2862     "Threshold before we will permit idle thread spinning");
2863 #ifdef SMP
2864 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2865     "Number of hz ticks to keep thread affinity for");
2866 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2867     "Enables the long-term load balancer");
2868 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2869     &balance_interval, 0,
2870     "Average period in stathz ticks to run the long-term balancer");
2871 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2872     "Attempts to steal work from other cores before idling");
2873 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2874     "Minimum load on remote CPU before we'll steal");
2875 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2876     CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2877     "XML dump of detected CPU topology");
2878 #endif
2879 
2880 /* ps compat.  All cpu percentages from ULE are weighted. */
2881 static int ccpu = 0;
2882 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2883