xref: /freebsd/sys/kern/sched_ule.c (revision 65b71c48d226f3ef3eb1c4452dccc678a7d2ac96)
1 /*-
2  * Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice unmodified, this list of conditions, and the following
10  *    disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25  */
26 
27 /*
28  * This file implements the ULE scheduler.  ULE supports independent CPU
29  * run queues and fine grain locking.  It has superior interactive
30  * performance under load even on uni-processor systems.
31  *
32  * etymology:
33  *   ULE is the last three letters in schedule.  It owes its name to a
34  * generic user created for a scheduling system by Paul Mikesell at
35  * Isilon Systems and a general lack of creativity on the part of the author.
36  */
37 
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
40 
41 #include "opt_hwpmc_hooks.h"
42 #include "opt_sched.h"
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/kdb.h>
47 #include <sys/kernel.h>
48 #include <sys/ktr.h>
49 #include <sys/limits.h>
50 #include <sys/lock.h>
51 #include <sys/mutex.h>
52 #include <sys/proc.h>
53 #include <sys/resource.h>
54 #include <sys/resourcevar.h>
55 #include <sys/sched.h>
56 #include <sys/sdt.h>
57 #include <sys/smp.h>
58 #include <sys/sx.h>
59 #include <sys/sysctl.h>
60 #include <sys/sysproto.h>
61 #include <sys/turnstile.h>
62 #include <sys/umtx.h>
63 #include <sys/vmmeter.h>
64 #include <sys/cpuset.h>
65 #include <sys/sbuf.h>
66 
67 #ifdef HWPMC_HOOKS
68 #include <sys/pmckern.h>
69 #endif
70 
71 #ifdef KDTRACE_HOOKS
72 #include <sys/dtrace_bsd.h>
73 int				dtrace_vtime_active;
74 dtrace_vtime_switch_func_t	dtrace_vtime_switch_func;
75 #endif
76 
77 #include <machine/cpu.h>
78 #include <machine/smp.h>
79 
80 #define	KTR_ULE	0
81 
82 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
83 #define	TDQ_NAME_LEN	(sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
84 #define	TDQ_LOADNAME_LEN	(sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
85 
86 /*
87  * Thread scheduler specific section.  All fields are protected
88  * by the thread lock.
89  */
90 struct td_sched {
91 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
92 	short		ts_flags;	/* TSF_* flags. */
93 	int		ts_cpu;		/* CPU that we have affinity for. */
94 	int		ts_rltick;	/* Real last tick, for affinity. */
95 	int		ts_slice;	/* Ticks of slice remaining. */
96 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
97 	u_int		ts_runtime;	/* Number of ticks we were running */
98 	int		ts_ltick;	/* Last tick that we were running on */
99 	int		ts_ftick;	/* First tick that we were running on */
100 	int		ts_ticks;	/* Tick count */
101 #ifdef KTR
102 	char		ts_name[TS_NAME_LEN];
103 #endif
104 };
105 /* flags kept in ts_flags */
106 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
107 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
108 
109 static struct td_sched td_sched0;
110 
111 #define	THREAD_CAN_MIGRATE(td)	((td)->td_pinned == 0)
112 #define	THREAD_CAN_SCHED(td, cpu)	\
113     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
114 
115 /*
116  * Priority ranges used for interactive and non-interactive timeshare
117  * threads.  The timeshare priorities are split up into four ranges.
118  * The first range handles interactive threads.  The last three ranges
119  * (NHALF, x, and NHALF) handle non-interactive threads with the outer
120  * ranges supporting nice values.
121  */
122 #define	PRI_TIMESHARE_RANGE	(PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
123 #define	PRI_INTERACT_RANGE	((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
124 #define	PRI_BATCH_RANGE		(PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
125 
126 #define	PRI_MIN_INTERACT	PRI_MIN_TIMESHARE
127 #define	PRI_MAX_INTERACT	(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
128 #define	PRI_MIN_BATCH		(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
129 #define	PRI_MAX_BATCH		PRI_MAX_TIMESHARE
130 
131 /*
132  * Cpu percentage computation macros and defines.
133  *
134  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
135  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
136  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
137  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
138  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
139  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
140  */
141 #define	SCHED_TICK_SECS		10
142 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
143 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
144 #define	SCHED_TICK_SHIFT	10
145 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
146 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
147 
148 /*
149  * These macros determine priorities for non-interactive threads.  They are
150  * assigned a priority based on their recent cpu utilization as expressed
151  * by the ratio of ticks to the tick total.  NHALF priorities at the start
152  * and end of the MIN to MAX timeshare range are only reachable with negative
153  * or positive nice respectively.
154  *
155  * PRI_RANGE:	Priority range for utilization dependent priorities.
156  * PRI_NRESV:	Number of nice values.
157  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
158  * PRI_NICE:	Determines the part of the priority inherited from nice.
159  */
160 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
161 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
162 #define	SCHED_PRI_MIN		(PRI_MIN_BATCH + SCHED_PRI_NHALF)
163 #define	SCHED_PRI_MAX		(PRI_MAX_BATCH - SCHED_PRI_NHALF)
164 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
165 #define	SCHED_PRI_TICKS(ts)						\
166     (SCHED_TICK_HZ((ts)) /						\
167     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
168 #define	SCHED_PRI_NICE(nice)	(nice)
169 
170 /*
171  * These determine the interactivity of a process.  Interactivity differs from
172  * cpu utilization in that it expresses the voluntary time slept vs time ran
173  * while cpu utilization includes all time not running.  This more accurately
174  * models the intent of the thread.
175  *
176  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
177  *		before throttling back.
178  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
179  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
180  * INTERACT_THRESH:	Threshold for placement on the current runq.
181  */
182 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
183 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
184 #define	SCHED_INTERACT_MAX	(100)
185 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
186 #define	SCHED_INTERACT_THRESH	(30)
187 
188 /*
189  * These parameters determine the slice behavior for batch work.
190  */
191 #define	SCHED_SLICE_DEFAULT_DIVISOR	10	/* ~94 ms, 12 stathz ticks. */
192 #define	SCHED_SLICE_MIN_DIVISOR		6	/* DEFAULT/MIN = ~16 ms. */
193 
194 /* Flags kept in td_flags. */
195 #define	TDF_SLICEEND	TDF_SCHED2	/* Thread time slice is over. */
196 
197 /*
198  * tickincr:		Converts a stathz tick into a hz domain scaled by
199  *			the shift factor.  Without the shift the error rate
200  *			due to rounding would be unacceptably high.
201  * realstathz:		stathz is sometimes 0 and run off of hz.
202  * sched_slice:		Runtime of each thread before rescheduling.
203  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
204  */
205 static int sched_interact = SCHED_INTERACT_THRESH;
206 static int tickincr = 8 << SCHED_TICK_SHIFT;
207 static int realstathz = 127;	/* reset during boot. */
208 static int sched_slice = 10;	/* reset during boot. */
209 static int sched_slice_min = 1;	/* reset during boot. */
210 #ifdef PREEMPTION
211 #ifdef FULL_PREEMPTION
212 static int preempt_thresh = PRI_MAX_IDLE;
213 #else
214 static int preempt_thresh = PRI_MIN_KERN;
215 #endif
216 #else
217 static int preempt_thresh = 0;
218 #endif
219 static int static_boost = PRI_MIN_BATCH;
220 static int sched_idlespins = 10000;
221 static int sched_idlespinthresh = -1;
222 
223 /*
224  * tdq - per processor runqs and statistics.  All fields are protected by the
225  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
226  * locking in sched_pickcpu();
227  */
228 struct tdq {
229 	/*
230 	 * Ordered to improve efficiency of cpu_search() and switch().
231 	 * tdq_lock is padded to avoid false sharing with tdq_load and
232 	 * tdq_cpu_idle.
233 	 */
234 	struct mtx_padalign tdq_lock;		/* run queue lock. */
235 	struct cpu_group *tdq_cg;		/* Pointer to cpu topology. */
236 	volatile int	tdq_load;		/* Aggregate load. */
237 	volatile int	tdq_cpu_idle;		/* cpu_idle() is active. */
238 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
239 	int		tdq_transferable;	/* Transferable thread count. */
240 	short		tdq_switchcnt;		/* Switches this tick. */
241 	short		tdq_oldswitchcnt;	/* Switches last tick. */
242 	u_char		tdq_lowpri;		/* Lowest priority thread. */
243 	u_char		tdq_ipipending;		/* IPI pending. */
244 	u_char		tdq_idx;		/* Current insert index. */
245 	u_char		tdq_ridx;		/* Current removal index. */
246 	struct runq	tdq_realtime;		/* real-time run queue. */
247 	struct runq	tdq_timeshare;		/* timeshare run queue. */
248 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
249 	char		tdq_name[TDQ_NAME_LEN];
250 #ifdef KTR
251 	char		tdq_loadname[TDQ_LOADNAME_LEN];
252 #endif
253 } __aligned(64);
254 
255 /* Idle thread states and config. */
256 #define	TDQ_RUNNING	1
257 #define	TDQ_IDLE	2
258 
259 #ifdef SMP
260 struct cpu_group *cpu_top;		/* CPU topology */
261 
262 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
263 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
264 
265 /*
266  * Run-time tunables.
267  */
268 static int rebalance = 1;
269 static int balance_interval = 128;	/* Default set in sched_initticks(). */
270 static int affinity;
271 static int steal_idle = 1;
272 static int steal_thresh = 2;
273 
274 /*
275  * One thread queue per processor.
276  */
277 static struct tdq	tdq_cpu[MAXCPU];
278 static struct tdq	*balance_tdq;
279 static int balance_ticks;
280 static DPCPU_DEFINE(uint32_t, randomval);
281 
282 #define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
283 #define	TDQ_CPU(x)	(&tdq_cpu[(x)])
284 #define	TDQ_ID(x)	((int)((x) - tdq_cpu))
285 #else	/* !SMP */
286 static struct tdq	tdq_cpu;
287 
288 #define	TDQ_ID(x)	(0)
289 #define	TDQ_SELF()	(&tdq_cpu)
290 #define	TDQ_CPU(x)	(&tdq_cpu)
291 #endif
292 
293 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
294 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
295 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
296 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
297 #define	TDQ_LOCKPTR(t)		((struct mtx *)(&(t)->tdq_lock))
298 
299 static void sched_priority(struct thread *);
300 static void sched_thread_priority(struct thread *, u_char);
301 static int sched_interact_score(struct thread *);
302 static void sched_interact_update(struct thread *);
303 static void sched_interact_fork(struct thread *);
304 static void sched_pctcpu_update(struct td_sched *, int);
305 
306 /* Operations on per processor queues */
307 static struct thread *tdq_choose(struct tdq *);
308 static void tdq_setup(struct tdq *);
309 static void tdq_load_add(struct tdq *, struct thread *);
310 static void tdq_load_rem(struct tdq *, struct thread *);
311 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
312 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
313 static inline int sched_shouldpreempt(int, int, int);
314 void tdq_print(int cpu);
315 static void runq_print(struct runq *rq);
316 static void tdq_add(struct tdq *, struct thread *, int);
317 #ifdef SMP
318 static int tdq_move(struct tdq *, struct tdq *);
319 static int tdq_idled(struct tdq *);
320 static void tdq_notify(struct tdq *, struct thread *);
321 static struct thread *tdq_steal(struct tdq *, int);
322 static struct thread *runq_steal(struct runq *, int);
323 static int sched_pickcpu(struct thread *, int);
324 static void sched_balance(void);
325 static int sched_balance_pair(struct tdq *, struct tdq *);
326 static inline struct tdq *sched_setcpu(struct thread *, int, int);
327 static inline void thread_unblock_switch(struct thread *, struct mtx *);
328 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
329 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
330 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
331     struct cpu_group *cg, int indent);
332 #endif
333 
334 static void sched_setup(void *dummy);
335 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
336 
337 static void sched_initticks(void *dummy);
338 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
339     NULL);
340 
341 SDT_PROVIDER_DEFINE(sched);
342 
343 SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *",
344     "struct proc *", "uint8_t");
345 SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *",
346     "struct proc *", "void *");
347 SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *",
348     "struct proc *", "void *", "int");
349 SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *",
350     "struct proc *", "uint8_t", "struct thread *");
351 SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int");
352 SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *",
353     "struct proc *");
354 SDT_PROBE_DEFINE(sched, , , on__cpu);
355 SDT_PROBE_DEFINE(sched, , , remain__cpu);
356 SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *",
357     "struct proc *");
358 
359 /*
360  * Print the threads waiting on a run-queue.
361  */
362 static void
363 runq_print(struct runq *rq)
364 {
365 	struct rqhead *rqh;
366 	struct thread *td;
367 	int pri;
368 	int j;
369 	int i;
370 
371 	for (i = 0; i < RQB_LEN; i++) {
372 		printf("\t\trunq bits %d 0x%zx\n",
373 		    i, rq->rq_status.rqb_bits[i]);
374 		for (j = 0; j < RQB_BPW; j++)
375 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
376 				pri = j + (i << RQB_L2BPW);
377 				rqh = &rq->rq_queues[pri];
378 				TAILQ_FOREACH(td, rqh, td_runq) {
379 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
380 					    td, td->td_name, td->td_priority,
381 					    td->td_rqindex, pri);
382 				}
383 			}
384 	}
385 }
386 
387 /*
388  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
389  */
390 void
391 tdq_print(int cpu)
392 {
393 	struct tdq *tdq;
394 
395 	tdq = TDQ_CPU(cpu);
396 
397 	printf("tdq %d:\n", TDQ_ID(tdq));
398 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
399 	printf("\tLock name:      %s\n", tdq->tdq_name);
400 	printf("\tload:           %d\n", tdq->tdq_load);
401 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
402 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
403 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
404 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
405 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
406 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
407 	printf("\trealtime runq:\n");
408 	runq_print(&tdq->tdq_realtime);
409 	printf("\ttimeshare runq:\n");
410 	runq_print(&tdq->tdq_timeshare);
411 	printf("\tidle runq:\n");
412 	runq_print(&tdq->tdq_idle);
413 }
414 
415 static inline int
416 sched_shouldpreempt(int pri, int cpri, int remote)
417 {
418 	/*
419 	 * If the new priority is not better than the current priority there is
420 	 * nothing to do.
421 	 */
422 	if (pri >= cpri)
423 		return (0);
424 	/*
425 	 * Always preempt idle.
426 	 */
427 	if (cpri >= PRI_MIN_IDLE)
428 		return (1);
429 	/*
430 	 * If preemption is disabled don't preempt others.
431 	 */
432 	if (preempt_thresh == 0)
433 		return (0);
434 	/*
435 	 * Preempt if we exceed the threshold.
436 	 */
437 	if (pri <= preempt_thresh)
438 		return (1);
439 	/*
440 	 * If we're interactive or better and there is non-interactive
441 	 * or worse running preempt only remote processors.
442 	 */
443 	if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
444 		return (1);
445 	return (0);
446 }
447 
448 /*
449  * Add a thread to the actual run-queue.  Keeps transferable counts up to
450  * date with what is actually on the run-queue.  Selects the correct
451  * queue position for timeshare threads.
452  */
453 static __inline void
454 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
455 {
456 	struct td_sched *ts;
457 	u_char pri;
458 
459 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
460 	THREAD_LOCK_ASSERT(td, MA_OWNED);
461 
462 	pri = td->td_priority;
463 	ts = td->td_sched;
464 	TD_SET_RUNQ(td);
465 	if (THREAD_CAN_MIGRATE(td)) {
466 		tdq->tdq_transferable++;
467 		ts->ts_flags |= TSF_XFERABLE;
468 	}
469 	if (pri < PRI_MIN_BATCH) {
470 		ts->ts_runq = &tdq->tdq_realtime;
471 	} else if (pri <= PRI_MAX_BATCH) {
472 		ts->ts_runq = &tdq->tdq_timeshare;
473 		KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
474 			("Invalid priority %d on timeshare runq", pri));
475 		/*
476 		 * This queue contains only priorities between MIN and MAX
477 		 * realtime.  Use the whole queue to represent these values.
478 		 */
479 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
480 			pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
481 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
482 			/*
483 			 * This effectively shortens the queue by one so we
484 			 * can have a one slot difference between idx and
485 			 * ridx while we wait for threads to drain.
486 			 */
487 			if (tdq->tdq_ridx != tdq->tdq_idx &&
488 			    pri == tdq->tdq_ridx)
489 				pri = (unsigned char)(pri - 1) % RQ_NQS;
490 		} else
491 			pri = tdq->tdq_ridx;
492 		runq_add_pri(ts->ts_runq, td, pri, flags);
493 		return;
494 	} else
495 		ts->ts_runq = &tdq->tdq_idle;
496 	runq_add(ts->ts_runq, td, flags);
497 }
498 
499 /*
500  * Remove a thread from a run-queue.  This typically happens when a thread
501  * is selected to run.  Running threads are not on the queue and the
502  * transferable count does not reflect them.
503  */
504 static __inline void
505 tdq_runq_rem(struct tdq *tdq, struct thread *td)
506 {
507 	struct td_sched *ts;
508 
509 	ts = td->td_sched;
510 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
511 	KASSERT(ts->ts_runq != NULL,
512 	    ("tdq_runq_remove: thread %p null ts_runq", td));
513 	if (ts->ts_flags & TSF_XFERABLE) {
514 		tdq->tdq_transferable--;
515 		ts->ts_flags &= ~TSF_XFERABLE;
516 	}
517 	if (ts->ts_runq == &tdq->tdq_timeshare) {
518 		if (tdq->tdq_idx != tdq->tdq_ridx)
519 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
520 		else
521 			runq_remove_idx(ts->ts_runq, td, NULL);
522 	} else
523 		runq_remove(ts->ts_runq, td);
524 }
525 
526 /*
527  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
528  * for this thread to the referenced thread queue.
529  */
530 static void
531 tdq_load_add(struct tdq *tdq, struct thread *td)
532 {
533 
534 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
535 	THREAD_LOCK_ASSERT(td, MA_OWNED);
536 
537 	tdq->tdq_load++;
538 	if ((td->td_flags & TDF_NOLOAD) == 0)
539 		tdq->tdq_sysload++;
540 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
541 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
542 }
543 
544 /*
545  * Remove the load from a thread that is transitioning to a sleep state or
546  * exiting.
547  */
548 static void
549 tdq_load_rem(struct tdq *tdq, struct thread *td)
550 {
551 
552 	THREAD_LOCK_ASSERT(td, MA_OWNED);
553 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
554 	KASSERT(tdq->tdq_load != 0,
555 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
556 
557 	tdq->tdq_load--;
558 	if ((td->td_flags & TDF_NOLOAD) == 0)
559 		tdq->tdq_sysload--;
560 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
561 	SDT_PROBE2(sched, , , load__change, (int)TDQ_ID(tdq), tdq->tdq_load);
562 }
563 
564 /*
565  * Bound timeshare latency by decreasing slice size as load increases.  We
566  * consider the maximum latency as the sum of the threads waiting to run
567  * aside from curthread and target no more than sched_slice latency but
568  * no less than sched_slice_min runtime.
569  */
570 static inline int
571 tdq_slice(struct tdq *tdq)
572 {
573 	int load;
574 
575 	/*
576 	 * It is safe to use sys_load here because this is called from
577 	 * contexts where timeshare threads are running and so there
578 	 * cannot be higher priority load in the system.
579 	 */
580 	load = tdq->tdq_sysload - 1;
581 	if (load >= SCHED_SLICE_MIN_DIVISOR)
582 		return (sched_slice_min);
583 	if (load <= 1)
584 		return (sched_slice);
585 	return (sched_slice / load);
586 }
587 
588 /*
589  * Set lowpri to its exact value by searching the run-queue and
590  * evaluating curthread.  curthread may be passed as an optimization.
591  */
592 static void
593 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
594 {
595 	struct thread *td;
596 
597 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
598 	if (ctd == NULL)
599 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
600 	td = tdq_choose(tdq);
601 	if (td == NULL || td->td_priority > ctd->td_priority)
602 		tdq->tdq_lowpri = ctd->td_priority;
603 	else
604 		tdq->tdq_lowpri = td->td_priority;
605 }
606 
607 #ifdef SMP
608 /*
609  * We need some randomness. Implement a classic Linear Congruential
610  * Generator X_{n+1}=(aX_n+c) mod m. These values are optimized for
611  * m = 2^32, a = 69069 and c = 5. We only return the upper 16 bits
612  * of the random state (in the low bits of our answer) to keep
613  * the maximum randomness.
614  */
615 static uint32_t
616 sched_random(void)
617 {
618 	uint32_t *rndptr;
619 
620 	rndptr = DPCPU_PTR(randomval);
621 	*rndptr = *rndptr * 69069 + 5;
622 
623 	return (*rndptr >> 16);
624 }
625 
626 struct cpu_search {
627 	cpuset_t cs_mask;
628 	u_int	cs_prefer;
629 	int	cs_pri;		/* Min priority for low. */
630 	int	cs_limit;	/* Max load for low, min load for high. */
631 	int	cs_cpu;
632 	int	cs_load;
633 };
634 
635 #define	CPU_SEARCH_LOWEST	0x1
636 #define	CPU_SEARCH_HIGHEST	0x2
637 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
638 
639 #define	CPUSET_FOREACH(cpu, mask)				\
640 	for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++)		\
641 		if (CPU_ISSET(cpu, &mask))
642 
643 static __always_inline int cpu_search(const struct cpu_group *cg,
644     struct cpu_search *low, struct cpu_search *high, const int match);
645 int __noinline cpu_search_lowest(const struct cpu_group *cg,
646     struct cpu_search *low);
647 int __noinline cpu_search_highest(const struct cpu_group *cg,
648     struct cpu_search *high);
649 int __noinline cpu_search_both(const struct cpu_group *cg,
650     struct cpu_search *low, struct cpu_search *high);
651 
652 /*
653  * Search the tree of cpu_groups for the lowest or highest loaded cpu
654  * according to the match argument.  This routine actually compares the
655  * load on all paths through the tree and finds the least loaded cpu on
656  * the least loaded path, which may differ from the least loaded cpu in
657  * the system.  This balances work among caches and busses.
658  *
659  * This inline is instantiated in three forms below using constants for the
660  * match argument.  It is reduced to the minimum set for each case.  It is
661  * also recursive to the depth of the tree.
662  */
663 static __always_inline int
664 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
665     struct cpu_search *high, const int match)
666 {
667 	struct cpu_search lgroup;
668 	struct cpu_search hgroup;
669 	cpuset_t cpumask;
670 	struct cpu_group *child;
671 	struct tdq *tdq;
672 	int cpu, i, hload, lload, load, total, rnd;
673 
674 	total = 0;
675 	cpumask = cg->cg_mask;
676 	if (match & CPU_SEARCH_LOWEST) {
677 		lload = INT_MAX;
678 		lgroup = *low;
679 	}
680 	if (match & CPU_SEARCH_HIGHEST) {
681 		hload = INT_MIN;
682 		hgroup = *high;
683 	}
684 
685 	/* Iterate through the child CPU groups and then remaining CPUs. */
686 	for (i = cg->cg_children, cpu = mp_maxid; ; ) {
687 		if (i == 0) {
688 #ifdef HAVE_INLINE_FFSL
689 			cpu = CPU_FFS(&cpumask) - 1;
690 #else
691 			while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
692 				cpu--;
693 #endif
694 			if (cpu < 0)
695 				break;
696 			child = NULL;
697 		} else
698 			child = &cg->cg_child[i - 1];
699 
700 		if (match & CPU_SEARCH_LOWEST)
701 			lgroup.cs_cpu = -1;
702 		if (match & CPU_SEARCH_HIGHEST)
703 			hgroup.cs_cpu = -1;
704 		if (child) {			/* Handle child CPU group. */
705 			CPU_NAND(&cpumask, &child->cg_mask);
706 			switch (match) {
707 			case CPU_SEARCH_LOWEST:
708 				load = cpu_search_lowest(child, &lgroup);
709 				break;
710 			case CPU_SEARCH_HIGHEST:
711 				load = cpu_search_highest(child, &hgroup);
712 				break;
713 			case CPU_SEARCH_BOTH:
714 				load = cpu_search_both(child, &lgroup, &hgroup);
715 				break;
716 			}
717 		} else {			/* Handle child CPU. */
718 			CPU_CLR(cpu, &cpumask);
719 			tdq = TDQ_CPU(cpu);
720 			load = tdq->tdq_load * 256;
721 			rnd = sched_random() % 32;
722 			if (match & CPU_SEARCH_LOWEST) {
723 				if (cpu == low->cs_prefer)
724 					load -= 64;
725 				/* If that CPU is allowed and get data. */
726 				if (tdq->tdq_lowpri > lgroup.cs_pri &&
727 				    tdq->tdq_load <= lgroup.cs_limit &&
728 				    CPU_ISSET(cpu, &lgroup.cs_mask)) {
729 					lgroup.cs_cpu = cpu;
730 					lgroup.cs_load = load - rnd;
731 				}
732 			}
733 			if (match & CPU_SEARCH_HIGHEST)
734 				if (tdq->tdq_load >= hgroup.cs_limit &&
735 				    tdq->tdq_transferable &&
736 				    CPU_ISSET(cpu, &hgroup.cs_mask)) {
737 					hgroup.cs_cpu = cpu;
738 					hgroup.cs_load = load - rnd;
739 				}
740 		}
741 		total += load;
742 
743 		/* We have info about child item. Compare it. */
744 		if (match & CPU_SEARCH_LOWEST) {
745 			if (lgroup.cs_cpu >= 0 &&
746 			    (load < lload ||
747 			     (load == lload && lgroup.cs_load < low->cs_load))) {
748 				lload = load;
749 				low->cs_cpu = lgroup.cs_cpu;
750 				low->cs_load = lgroup.cs_load;
751 			}
752 		}
753 		if (match & CPU_SEARCH_HIGHEST)
754 			if (hgroup.cs_cpu >= 0 &&
755 			    (load > hload ||
756 			     (load == hload && hgroup.cs_load > high->cs_load))) {
757 				hload = load;
758 				high->cs_cpu = hgroup.cs_cpu;
759 				high->cs_load = hgroup.cs_load;
760 			}
761 		if (child) {
762 			i--;
763 			if (i == 0 && CPU_EMPTY(&cpumask))
764 				break;
765 		}
766 #ifndef HAVE_INLINE_FFSL
767 		else
768 			cpu--;
769 #endif
770 	}
771 	return (total);
772 }
773 
774 /*
775  * cpu_search instantiations must pass constants to maintain the inline
776  * optimization.
777  */
778 int
779 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
780 {
781 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
782 }
783 
784 int
785 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
786 {
787 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
788 }
789 
790 int
791 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
792     struct cpu_search *high)
793 {
794 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
795 }
796 
797 /*
798  * Find the cpu with the least load via the least loaded path that has a
799  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
800  * acceptable.
801  */
802 static inline int
803 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
804     int prefer)
805 {
806 	struct cpu_search low;
807 
808 	low.cs_cpu = -1;
809 	low.cs_prefer = prefer;
810 	low.cs_mask = mask;
811 	low.cs_pri = pri;
812 	low.cs_limit = maxload;
813 	cpu_search_lowest(cg, &low);
814 	return low.cs_cpu;
815 }
816 
817 /*
818  * Find the cpu with the highest load via the highest loaded path.
819  */
820 static inline int
821 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
822 {
823 	struct cpu_search high;
824 
825 	high.cs_cpu = -1;
826 	high.cs_mask = mask;
827 	high.cs_limit = minload;
828 	cpu_search_highest(cg, &high);
829 	return high.cs_cpu;
830 }
831 
832 static void
833 sched_balance_group(struct cpu_group *cg)
834 {
835 	cpuset_t hmask, lmask;
836 	int high, low, anylow;
837 
838 	CPU_FILL(&hmask);
839 	for (;;) {
840 		high = sched_highest(cg, hmask, 1);
841 		/* Stop if there is no more CPU with transferrable threads. */
842 		if (high == -1)
843 			break;
844 		CPU_CLR(high, &hmask);
845 		CPU_COPY(&hmask, &lmask);
846 		/* Stop if there is no more CPU left for low. */
847 		if (CPU_EMPTY(&lmask))
848 			break;
849 		anylow = 1;
850 nextlow:
851 		low = sched_lowest(cg, lmask, -1,
852 		    TDQ_CPU(high)->tdq_load - 1, high);
853 		/* Stop if we looked well and found no less loaded CPU. */
854 		if (anylow && low == -1)
855 			break;
856 		/* Go to next high if we found no less loaded CPU. */
857 		if (low == -1)
858 			continue;
859 		/* Transfer thread from high to low. */
860 		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
861 			/* CPU that got thread can no longer be a donor. */
862 			CPU_CLR(low, &hmask);
863 		} else {
864 			/*
865 			 * If failed, then there is no threads on high
866 			 * that can run on this low. Drop low from low
867 			 * mask and look for different one.
868 			 */
869 			CPU_CLR(low, &lmask);
870 			anylow = 0;
871 			goto nextlow;
872 		}
873 	}
874 }
875 
876 static void
877 sched_balance(void)
878 {
879 	struct tdq *tdq;
880 
881 	if (smp_started == 0 || rebalance == 0)
882 		return;
883 
884 	balance_ticks = max(balance_interval / 2, 1) +
885 	    (sched_random() % balance_interval);
886 	tdq = TDQ_SELF();
887 	TDQ_UNLOCK(tdq);
888 	sched_balance_group(cpu_top);
889 	TDQ_LOCK(tdq);
890 }
891 
892 /*
893  * Lock two thread queues using their address to maintain lock order.
894  */
895 static void
896 tdq_lock_pair(struct tdq *one, struct tdq *two)
897 {
898 	if (one < two) {
899 		TDQ_LOCK(one);
900 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
901 	} else {
902 		TDQ_LOCK(two);
903 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
904 	}
905 }
906 
907 /*
908  * Unlock two thread queues.  Order is not important here.
909  */
910 static void
911 tdq_unlock_pair(struct tdq *one, struct tdq *two)
912 {
913 	TDQ_UNLOCK(one);
914 	TDQ_UNLOCK(two);
915 }
916 
917 /*
918  * Transfer load between two imbalanced thread queues.
919  */
920 static int
921 sched_balance_pair(struct tdq *high, struct tdq *low)
922 {
923 	int moved;
924 	int cpu;
925 
926 	tdq_lock_pair(high, low);
927 	moved = 0;
928 	/*
929 	 * Determine what the imbalance is and then adjust that to how many
930 	 * threads we actually have to give up (transferable).
931 	 */
932 	if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
933 	    (moved = tdq_move(high, low)) > 0) {
934 		/*
935 		 * In case the target isn't the current cpu IPI it to force a
936 		 * reschedule with the new workload.
937 		 */
938 		cpu = TDQ_ID(low);
939 		if (cpu != PCPU_GET(cpuid))
940 			ipi_cpu(cpu, IPI_PREEMPT);
941 	}
942 	tdq_unlock_pair(high, low);
943 	return (moved);
944 }
945 
946 /*
947  * Move a thread from one thread queue to another.
948  */
949 static int
950 tdq_move(struct tdq *from, struct tdq *to)
951 {
952 	struct td_sched *ts;
953 	struct thread *td;
954 	struct tdq *tdq;
955 	int cpu;
956 
957 	TDQ_LOCK_ASSERT(from, MA_OWNED);
958 	TDQ_LOCK_ASSERT(to, MA_OWNED);
959 
960 	tdq = from;
961 	cpu = TDQ_ID(to);
962 	td = tdq_steal(tdq, cpu);
963 	if (td == NULL)
964 		return (0);
965 	ts = td->td_sched;
966 	/*
967 	 * Although the run queue is locked the thread may be blocked.  Lock
968 	 * it to clear this and acquire the run-queue lock.
969 	 */
970 	thread_lock(td);
971 	/* Drop recursive lock on from acquired via thread_lock(). */
972 	TDQ_UNLOCK(from);
973 	sched_rem(td);
974 	ts->ts_cpu = cpu;
975 	td->td_lock = TDQ_LOCKPTR(to);
976 	tdq_add(to, td, SRQ_YIELDING);
977 	return (1);
978 }
979 
980 /*
981  * This tdq has idled.  Try to steal a thread from another cpu and switch
982  * to it.
983  */
984 static int
985 tdq_idled(struct tdq *tdq)
986 {
987 	struct cpu_group *cg;
988 	struct tdq *steal;
989 	cpuset_t mask;
990 	int thresh;
991 	int cpu;
992 
993 	if (smp_started == 0 || steal_idle == 0)
994 		return (1);
995 	CPU_FILL(&mask);
996 	CPU_CLR(PCPU_GET(cpuid), &mask);
997 	/* We don't want to be preempted while we're iterating. */
998 	spinlock_enter();
999 	for (cg = tdq->tdq_cg; cg != NULL; ) {
1000 		if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
1001 			thresh = steal_thresh;
1002 		else
1003 			thresh = 1;
1004 		cpu = sched_highest(cg, mask, thresh);
1005 		if (cpu == -1) {
1006 			cg = cg->cg_parent;
1007 			continue;
1008 		}
1009 		steal = TDQ_CPU(cpu);
1010 		CPU_CLR(cpu, &mask);
1011 		tdq_lock_pair(tdq, steal);
1012 		if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
1013 			tdq_unlock_pair(tdq, steal);
1014 			continue;
1015 		}
1016 		/*
1017 		 * If a thread was added while interrupts were disabled don't
1018 		 * steal one here.  If we fail to acquire one due to affinity
1019 		 * restrictions loop again with this cpu removed from the
1020 		 * set.
1021 		 */
1022 		if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
1023 			tdq_unlock_pair(tdq, steal);
1024 			continue;
1025 		}
1026 		spinlock_exit();
1027 		TDQ_UNLOCK(steal);
1028 		mi_switch(SW_VOL | SWT_IDLE, NULL);
1029 		thread_unlock(curthread);
1030 
1031 		return (0);
1032 	}
1033 	spinlock_exit();
1034 	return (1);
1035 }
1036 
1037 /*
1038  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
1039  */
1040 static void
1041 tdq_notify(struct tdq *tdq, struct thread *td)
1042 {
1043 	struct thread *ctd;
1044 	int pri;
1045 	int cpu;
1046 
1047 	if (tdq->tdq_ipipending)
1048 		return;
1049 	cpu = td->td_sched->ts_cpu;
1050 	pri = td->td_priority;
1051 	ctd = pcpu_find(cpu)->pc_curthread;
1052 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1053 		return;
1054 
1055 	/*
1056 	 * Make sure that our caller's earlier update to tdq_load is
1057 	 * globally visible before we read tdq_cpu_idle.  Idle thread
1058 	 * accesses both of them without locks, and the order is important.
1059 	 */
1060 	atomic_thread_fence_seq_cst();
1061 
1062 	if (TD_IS_IDLETHREAD(ctd)) {
1063 		/*
1064 		 * If the MD code has an idle wakeup routine try that before
1065 		 * falling back to IPI.
1066 		 */
1067 		if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1068 			return;
1069 	}
1070 	tdq->tdq_ipipending = 1;
1071 	ipi_cpu(cpu, IPI_PREEMPT);
1072 }
1073 
1074 /*
1075  * Steals load from a timeshare queue.  Honors the rotating queue head
1076  * index.
1077  */
1078 static struct thread *
1079 runq_steal_from(struct runq *rq, int cpu, u_char start)
1080 {
1081 	struct rqbits *rqb;
1082 	struct rqhead *rqh;
1083 	struct thread *td, *first;
1084 	int bit;
1085 	int i;
1086 
1087 	rqb = &rq->rq_status;
1088 	bit = start & (RQB_BPW -1);
1089 	first = NULL;
1090 again:
1091 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1092 		if (rqb->rqb_bits[i] == 0)
1093 			continue;
1094 		if (bit == 0)
1095 			bit = RQB_FFS(rqb->rqb_bits[i]);
1096 		for (; bit < RQB_BPW; bit++) {
1097 			if ((rqb->rqb_bits[i] & (1ul << bit)) == 0)
1098 				continue;
1099 			rqh = &rq->rq_queues[bit + (i << RQB_L2BPW)];
1100 			TAILQ_FOREACH(td, rqh, td_runq) {
1101 				if (first && THREAD_CAN_MIGRATE(td) &&
1102 				    THREAD_CAN_SCHED(td, cpu))
1103 					return (td);
1104 				first = td;
1105 			}
1106 		}
1107 	}
1108 	if (start != 0) {
1109 		start = 0;
1110 		goto again;
1111 	}
1112 
1113 	if (first && THREAD_CAN_MIGRATE(first) &&
1114 	    THREAD_CAN_SCHED(first, cpu))
1115 		return (first);
1116 	return (NULL);
1117 }
1118 
1119 /*
1120  * Steals load from a standard linear queue.
1121  */
1122 static struct thread *
1123 runq_steal(struct runq *rq, int cpu)
1124 {
1125 	struct rqhead *rqh;
1126 	struct rqbits *rqb;
1127 	struct thread *td;
1128 	int word;
1129 	int bit;
1130 
1131 	rqb = &rq->rq_status;
1132 	for (word = 0; word < RQB_LEN; word++) {
1133 		if (rqb->rqb_bits[word] == 0)
1134 			continue;
1135 		for (bit = 0; bit < RQB_BPW; bit++) {
1136 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1137 				continue;
1138 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1139 			TAILQ_FOREACH(td, rqh, td_runq)
1140 				if (THREAD_CAN_MIGRATE(td) &&
1141 				    THREAD_CAN_SCHED(td, cpu))
1142 					return (td);
1143 		}
1144 	}
1145 	return (NULL);
1146 }
1147 
1148 /*
1149  * Attempt to steal a thread in priority order from a thread queue.
1150  */
1151 static struct thread *
1152 tdq_steal(struct tdq *tdq, int cpu)
1153 {
1154 	struct thread *td;
1155 
1156 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1157 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1158 		return (td);
1159 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1160 	    cpu, tdq->tdq_ridx)) != NULL)
1161 		return (td);
1162 	return (runq_steal(&tdq->tdq_idle, cpu));
1163 }
1164 
1165 /*
1166  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1167  * current lock and returns with the assigned queue locked.
1168  */
1169 static inline struct tdq *
1170 sched_setcpu(struct thread *td, int cpu, int flags)
1171 {
1172 
1173 	struct tdq *tdq;
1174 
1175 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1176 	tdq = TDQ_CPU(cpu);
1177 	td->td_sched->ts_cpu = cpu;
1178 	/*
1179 	 * If the lock matches just return the queue.
1180 	 */
1181 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1182 		return (tdq);
1183 #ifdef notyet
1184 	/*
1185 	 * If the thread isn't running its lockptr is a
1186 	 * turnstile or a sleepqueue.  We can just lock_set without
1187 	 * blocking.
1188 	 */
1189 	if (TD_CAN_RUN(td)) {
1190 		TDQ_LOCK(tdq);
1191 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1192 		return (tdq);
1193 	}
1194 #endif
1195 	/*
1196 	 * The hard case, migration, we need to block the thread first to
1197 	 * prevent order reversals with other cpus locks.
1198 	 */
1199 	spinlock_enter();
1200 	thread_lock_block(td);
1201 	TDQ_LOCK(tdq);
1202 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1203 	spinlock_exit();
1204 	return (tdq);
1205 }
1206 
1207 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1208 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1209 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1210 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1211 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1212 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1213 
1214 static int
1215 sched_pickcpu(struct thread *td, int flags)
1216 {
1217 	struct cpu_group *cg, *ccg;
1218 	struct td_sched *ts;
1219 	struct tdq *tdq;
1220 	cpuset_t mask;
1221 	int cpu, pri, self;
1222 
1223 	self = PCPU_GET(cpuid);
1224 	ts = td->td_sched;
1225 	if (smp_started == 0)
1226 		return (self);
1227 	/*
1228 	 * Don't migrate a running thread from sched_switch().
1229 	 */
1230 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1231 		return (ts->ts_cpu);
1232 	/*
1233 	 * Prefer to run interrupt threads on the processors that generate
1234 	 * the interrupt.
1235 	 */
1236 	pri = td->td_priority;
1237 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1238 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1239 		SCHED_STAT_INC(pickcpu_intrbind);
1240 		ts->ts_cpu = self;
1241 		if (TDQ_CPU(self)->tdq_lowpri > pri) {
1242 			SCHED_STAT_INC(pickcpu_affinity);
1243 			return (ts->ts_cpu);
1244 		}
1245 	}
1246 	/*
1247 	 * If the thread can run on the last cpu and the affinity has not
1248 	 * expired or it is idle run it there.
1249 	 */
1250 	tdq = TDQ_CPU(ts->ts_cpu);
1251 	cg = tdq->tdq_cg;
1252 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1253 	    tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1254 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1255 		if (cg->cg_flags & CG_FLAG_THREAD) {
1256 			CPUSET_FOREACH(cpu, cg->cg_mask) {
1257 				if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1258 					break;
1259 			}
1260 		} else
1261 			cpu = INT_MAX;
1262 		if (cpu > mp_maxid) {
1263 			SCHED_STAT_INC(pickcpu_idle_affinity);
1264 			return (ts->ts_cpu);
1265 		}
1266 	}
1267 	/*
1268 	 * Search for the last level cache CPU group in the tree.
1269 	 * Skip caches with expired affinity time and SMT groups.
1270 	 * Affinity to higher level caches will be handled less aggressively.
1271 	 */
1272 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1273 		if (cg->cg_flags & CG_FLAG_THREAD)
1274 			continue;
1275 		if (!SCHED_AFFINITY(ts, cg->cg_level))
1276 			continue;
1277 		ccg = cg;
1278 	}
1279 	if (ccg != NULL)
1280 		cg = ccg;
1281 	cpu = -1;
1282 	/* Search the group for the less loaded idle CPU we can run now. */
1283 	mask = td->td_cpuset->cs_mask;
1284 	if (cg != NULL && cg != cpu_top &&
1285 	    CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1286 		cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1287 		    INT_MAX, ts->ts_cpu);
1288 	/* Search globally for the less loaded CPU we can run now. */
1289 	if (cpu == -1)
1290 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1291 	/* Search globally for the less loaded CPU. */
1292 	if (cpu == -1)
1293 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1294 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1295 	/*
1296 	 * Compare the lowest loaded cpu to current cpu.
1297 	 */
1298 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1299 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1300 	    TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1301 		SCHED_STAT_INC(pickcpu_local);
1302 		cpu = self;
1303 	} else
1304 		SCHED_STAT_INC(pickcpu_lowest);
1305 	if (cpu != ts->ts_cpu)
1306 		SCHED_STAT_INC(pickcpu_migration);
1307 	return (cpu);
1308 }
1309 #endif
1310 
1311 /*
1312  * Pick the highest priority task we have and return it.
1313  */
1314 static struct thread *
1315 tdq_choose(struct tdq *tdq)
1316 {
1317 	struct thread *td;
1318 
1319 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1320 	td = runq_choose(&tdq->tdq_realtime);
1321 	if (td != NULL)
1322 		return (td);
1323 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1324 	if (td != NULL) {
1325 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1326 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1327 		    td->td_priority));
1328 		return (td);
1329 	}
1330 	td = runq_choose(&tdq->tdq_idle);
1331 	if (td != NULL) {
1332 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1333 		    ("tdq_choose: Invalid priority on idle queue %d",
1334 		    td->td_priority));
1335 		return (td);
1336 	}
1337 
1338 	return (NULL);
1339 }
1340 
1341 /*
1342  * Initialize a thread queue.
1343  */
1344 static void
1345 tdq_setup(struct tdq *tdq)
1346 {
1347 
1348 	if (bootverbose)
1349 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1350 	runq_init(&tdq->tdq_realtime);
1351 	runq_init(&tdq->tdq_timeshare);
1352 	runq_init(&tdq->tdq_idle);
1353 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1354 	    "sched lock %d", (int)TDQ_ID(tdq));
1355 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1356 	    MTX_SPIN | MTX_RECURSE);
1357 #ifdef KTR
1358 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1359 	    "CPU %d load", (int)TDQ_ID(tdq));
1360 #endif
1361 }
1362 
1363 #ifdef SMP
1364 static void
1365 sched_setup_smp(void)
1366 {
1367 	struct tdq *tdq;
1368 	int i;
1369 
1370 	cpu_top = smp_topo();
1371 	CPU_FOREACH(i) {
1372 		tdq = TDQ_CPU(i);
1373 		tdq_setup(tdq);
1374 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1375 		if (tdq->tdq_cg == NULL)
1376 			panic("Can't find cpu group for %d\n", i);
1377 	}
1378 	balance_tdq = TDQ_SELF();
1379 	sched_balance();
1380 }
1381 #endif
1382 
1383 /*
1384  * Setup the thread queues and initialize the topology based on MD
1385  * information.
1386  */
1387 static void
1388 sched_setup(void *dummy)
1389 {
1390 	struct tdq *tdq;
1391 
1392 	tdq = TDQ_SELF();
1393 #ifdef SMP
1394 	sched_setup_smp();
1395 #else
1396 	tdq_setup(tdq);
1397 #endif
1398 
1399 	/* Add thread0's load since it's running. */
1400 	TDQ_LOCK(tdq);
1401 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1402 	tdq_load_add(tdq, &thread0);
1403 	tdq->tdq_lowpri = thread0.td_priority;
1404 	TDQ_UNLOCK(tdq);
1405 }
1406 
1407 /*
1408  * This routine determines time constants after stathz and hz are setup.
1409  */
1410 /* ARGSUSED */
1411 static void
1412 sched_initticks(void *dummy)
1413 {
1414 	int incr;
1415 
1416 	realstathz = stathz ? stathz : hz;
1417 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1418 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1419 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1420 	    realstathz);
1421 
1422 	/*
1423 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1424 	 * hz not being evenly divisible by stathz on all platforms.
1425 	 */
1426 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1427 	/*
1428 	 * This does not work for values of stathz that are more than
1429 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1430 	 */
1431 	if (incr == 0)
1432 		incr = 1;
1433 	tickincr = incr;
1434 #ifdef SMP
1435 	/*
1436 	 * Set the default balance interval now that we know
1437 	 * what realstathz is.
1438 	 */
1439 	balance_interval = realstathz;
1440 	affinity = SCHED_AFFINITY_DEFAULT;
1441 #endif
1442 	if (sched_idlespinthresh < 0)
1443 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1444 }
1445 
1446 
1447 /*
1448  * This is the core of the interactivity algorithm.  Determines a score based
1449  * on past behavior.  It is the ratio of sleep time to run time scaled to
1450  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1451  * differs from the cpu usage because it does not account for time spent
1452  * waiting on a run-queue.  Would be prettier if we had floating point.
1453  *
1454  * When a thread's sleep time is greater than its run time the
1455  * calculation is:
1456  *
1457  *                           scaling factor
1458  * interactivity score =  ---------------------
1459  *                        sleep time / run time
1460  *
1461  *
1462  * When a thread's run time is greater than its sleep time the
1463  * calculation is:
1464  *
1465  *                           scaling factor
1466  * interactivity score =  ---------------------    + scaling factor
1467  *                        run time / sleep time
1468  */
1469 static int
1470 sched_interact_score(struct thread *td)
1471 {
1472 	struct td_sched *ts;
1473 	int div;
1474 
1475 	ts = td->td_sched;
1476 	/*
1477 	 * The score is only needed if this is likely to be an interactive
1478 	 * task.  Don't go through the expense of computing it if there's
1479 	 * no chance.
1480 	 */
1481 	if (sched_interact <= SCHED_INTERACT_HALF &&
1482 		ts->ts_runtime >= ts->ts_slptime)
1483 			return (SCHED_INTERACT_HALF);
1484 
1485 	if (ts->ts_runtime > ts->ts_slptime) {
1486 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1487 		return (SCHED_INTERACT_HALF +
1488 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1489 	}
1490 	if (ts->ts_slptime > ts->ts_runtime) {
1491 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1492 		return (ts->ts_runtime / div);
1493 	}
1494 	/* runtime == slptime */
1495 	if (ts->ts_runtime)
1496 		return (SCHED_INTERACT_HALF);
1497 
1498 	/*
1499 	 * This can happen if slptime and runtime are 0.
1500 	 */
1501 	return (0);
1502 
1503 }
1504 
1505 /*
1506  * Scale the scheduling priority according to the "interactivity" of this
1507  * process.
1508  */
1509 static void
1510 sched_priority(struct thread *td)
1511 {
1512 	int score;
1513 	int pri;
1514 
1515 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1516 		return;
1517 	/*
1518 	 * If the score is interactive we place the thread in the realtime
1519 	 * queue with a priority that is less than kernel and interrupt
1520 	 * priorities.  These threads are not subject to nice restrictions.
1521 	 *
1522 	 * Scores greater than this are placed on the normal timeshare queue
1523 	 * where the priority is partially decided by the most recent cpu
1524 	 * utilization and the rest is decided by nice value.
1525 	 *
1526 	 * The nice value of the process has a linear effect on the calculated
1527 	 * score.  Negative nice values make it easier for a thread to be
1528 	 * considered interactive.
1529 	 */
1530 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1531 	if (score < sched_interact) {
1532 		pri = PRI_MIN_INTERACT;
1533 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1534 		    sched_interact) * score;
1535 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1536 		    ("sched_priority: invalid interactive priority %d score %d",
1537 		    pri, score));
1538 	} else {
1539 		pri = SCHED_PRI_MIN;
1540 		if (td->td_sched->ts_ticks)
1541 			pri += min(SCHED_PRI_TICKS(td->td_sched),
1542 			    SCHED_PRI_RANGE - 1);
1543 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1544 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1545 		    ("sched_priority: invalid priority %d: nice %d, "
1546 		    "ticks %d ftick %d ltick %d tick pri %d",
1547 		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1548 		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1549 		    SCHED_PRI_TICKS(td->td_sched)));
1550 	}
1551 	sched_user_prio(td, pri);
1552 
1553 	return;
1554 }
1555 
1556 /*
1557  * This routine enforces a maximum limit on the amount of scheduling history
1558  * kept.  It is called after either the slptime or runtime is adjusted.  This
1559  * function is ugly due to integer math.
1560  */
1561 static void
1562 sched_interact_update(struct thread *td)
1563 {
1564 	struct td_sched *ts;
1565 	u_int sum;
1566 
1567 	ts = td->td_sched;
1568 	sum = ts->ts_runtime + ts->ts_slptime;
1569 	if (sum < SCHED_SLP_RUN_MAX)
1570 		return;
1571 	/*
1572 	 * This only happens from two places:
1573 	 * 1) We have added an unusual amount of run time from fork_exit.
1574 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1575 	 */
1576 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1577 		if (ts->ts_runtime > ts->ts_slptime) {
1578 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1579 			ts->ts_slptime = 1;
1580 		} else {
1581 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1582 			ts->ts_runtime = 1;
1583 		}
1584 		return;
1585 	}
1586 	/*
1587 	 * If we have exceeded by more than 1/5th then the algorithm below
1588 	 * will not bring us back into range.  Dividing by two here forces
1589 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1590 	 */
1591 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1592 		ts->ts_runtime /= 2;
1593 		ts->ts_slptime /= 2;
1594 		return;
1595 	}
1596 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1597 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1598 }
1599 
1600 /*
1601  * Scale back the interactivity history when a child thread is created.  The
1602  * history is inherited from the parent but the thread may behave totally
1603  * differently.  For example, a shell spawning a compiler process.  We want
1604  * to learn that the compiler is behaving badly very quickly.
1605  */
1606 static void
1607 sched_interact_fork(struct thread *td)
1608 {
1609 	int ratio;
1610 	int sum;
1611 
1612 	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1613 	if (sum > SCHED_SLP_RUN_FORK) {
1614 		ratio = sum / SCHED_SLP_RUN_FORK;
1615 		td->td_sched->ts_runtime /= ratio;
1616 		td->td_sched->ts_slptime /= ratio;
1617 	}
1618 }
1619 
1620 /*
1621  * Called from proc0_init() to setup the scheduler fields.
1622  */
1623 void
1624 schedinit(void)
1625 {
1626 
1627 	/*
1628 	 * Set up the scheduler specific parts of proc0.
1629 	 */
1630 	proc0.p_sched = NULL; /* XXX */
1631 	thread0.td_sched = &td_sched0;
1632 	td_sched0.ts_ltick = ticks;
1633 	td_sched0.ts_ftick = ticks;
1634 	td_sched0.ts_slice = 0;
1635 }
1636 
1637 /*
1638  * This is only somewhat accurate since given many processes of the same
1639  * priority they will switch when their slices run out, which will be
1640  * at most sched_slice stathz ticks.
1641  */
1642 int
1643 sched_rr_interval(void)
1644 {
1645 
1646 	/* Convert sched_slice from stathz to hz. */
1647 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1648 }
1649 
1650 /*
1651  * Update the percent cpu tracking information when it is requested or
1652  * the total history exceeds the maximum.  We keep a sliding history of
1653  * tick counts that slowly decays.  This is less precise than the 4BSD
1654  * mechanism since it happens with less regular and frequent events.
1655  */
1656 static void
1657 sched_pctcpu_update(struct td_sched *ts, int run)
1658 {
1659 	int t = ticks;
1660 
1661 	if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1662 		ts->ts_ticks = 0;
1663 		ts->ts_ftick = t - SCHED_TICK_TARG;
1664 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1665 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1666 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1667 		ts->ts_ftick = t - SCHED_TICK_TARG;
1668 	}
1669 	if (run)
1670 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1671 	ts->ts_ltick = t;
1672 }
1673 
1674 /*
1675  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1676  * if necessary.  This is the back-end for several priority related
1677  * functions.
1678  */
1679 static void
1680 sched_thread_priority(struct thread *td, u_char prio)
1681 {
1682 	struct td_sched *ts;
1683 	struct tdq *tdq;
1684 	int oldpri;
1685 
1686 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1687 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1688 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1689 	SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio);
1690 	if (td != curthread && prio < td->td_priority) {
1691 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1692 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1693 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1694 		SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio,
1695 		    curthread);
1696 	}
1697 	ts = td->td_sched;
1698 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1699 	if (td->td_priority == prio)
1700 		return;
1701 	/*
1702 	 * If the priority has been elevated due to priority
1703 	 * propagation, we may have to move ourselves to a new
1704 	 * queue.  This could be optimized to not re-add in some
1705 	 * cases.
1706 	 */
1707 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1708 		sched_rem(td);
1709 		td->td_priority = prio;
1710 		sched_add(td, SRQ_BORROWING);
1711 		return;
1712 	}
1713 	/*
1714 	 * If the thread is currently running we may have to adjust the lowpri
1715 	 * information so other cpus are aware of our current priority.
1716 	 */
1717 	if (TD_IS_RUNNING(td)) {
1718 		tdq = TDQ_CPU(ts->ts_cpu);
1719 		oldpri = td->td_priority;
1720 		td->td_priority = prio;
1721 		if (prio < tdq->tdq_lowpri)
1722 			tdq->tdq_lowpri = prio;
1723 		else if (tdq->tdq_lowpri == oldpri)
1724 			tdq_setlowpri(tdq, td);
1725 		return;
1726 	}
1727 	td->td_priority = prio;
1728 }
1729 
1730 /*
1731  * Update a thread's priority when it is lent another thread's
1732  * priority.
1733  */
1734 void
1735 sched_lend_prio(struct thread *td, u_char prio)
1736 {
1737 
1738 	td->td_flags |= TDF_BORROWING;
1739 	sched_thread_priority(td, prio);
1740 }
1741 
1742 /*
1743  * Restore a thread's priority when priority propagation is
1744  * over.  The prio argument is the minimum priority the thread
1745  * needs to have to satisfy other possible priority lending
1746  * requests.  If the thread's regular priority is less
1747  * important than prio, the thread will keep a priority boost
1748  * of prio.
1749  */
1750 void
1751 sched_unlend_prio(struct thread *td, u_char prio)
1752 {
1753 	u_char base_pri;
1754 
1755 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1756 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1757 		base_pri = td->td_user_pri;
1758 	else
1759 		base_pri = td->td_base_pri;
1760 	if (prio >= base_pri) {
1761 		td->td_flags &= ~TDF_BORROWING;
1762 		sched_thread_priority(td, base_pri);
1763 	} else
1764 		sched_lend_prio(td, prio);
1765 }
1766 
1767 /*
1768  * Standard entry for setting the priority to an absolute value.
1769  */
1770 void
1771 sched_prio(struct thread *td, u_char prio)
1772 {
1773 	u_char oldprio;
1774 
1775 	/* First, update the base priority. */
1776 	td->td_base_pri = prio;
1777 
1778 	/*
1779 	 * If the thread is borrowing another thread's priority, don't
1780 	 * ever lower the priority.
1781 	 */
1782 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1783 		return;
1784 
1785 	/* Change the real priority. */
1786 	oldprio = td->td_priority;
1787 	sched_thread_priority(td, prio);
1788 
1789 	/*
1790 	 * If the thread is on a turnstile, then let the turnstile update
1791 	 * its state.
1792 	 */
1793 	if (TD_ON_LOCK(td) && oldprio != prio)
1794 		turnstile_adjust(td, oldprio);
1795 }
1796 
1797 /*
1798  * Set the base user priority, does not effect current running priority.
1799  */
1800 void
1801 sched_user_prio(struct thread *td, u_char prio)
1802 {
1803 
1804 	td->td_base_user_pri = prio;
1805 	if (td->td_lend_user_pri <= prio)
1806 		return;
1807 	td->td_user_pri = prio;
1808 }
1809 
1810 void
1811 sched_lend_user_prio(struct thread *td, u_char prio)
1812 {
1813 
1814 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1815 	td->td_lend_user_pri = prio;
1816 	td->td_user_pri = min(prio, td->td_base_user_pri);
1817 	if (td->td_priority > td->td_user_pri)
1818 		sched_prio(td, td->td_user_pri);
1819 	else if (td->td_priority != td->td_user_pri)
1820 		td->td_flags |= TDF_NEEDRESCHED;
1821 }
1822 
1823 /*
1824  * Handle migration from sched_switch().  This happens only for
1825  * cpu binding.
1826  */
1827 static struct mtx *
1828 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1829 {
1830 	struct tdq *tdn;
1831 
1832 	tdn = TDQ_CPU(td->td_sched->ts_cpu);
1833 #ifdef SMP
1834 	tdq_load_rem(tdq, td);
1835 	/*
1836 	 * Do the lock dance required to avoid LOR.  We grab an extra
1837 	 * spinlock nesting to prevent preemption while we're
1838 	 * not holding either run-queue lock.
1839 	 */
1840 	spinlock_enter();
1841 	thread_lock_block(td);	/* This releases the lock on tdq. */
1842 
1843 	/*
1844 	 * Acquire both run-queue locks before placing the thread on the new
1845 	 * run-queue to avoid deadlocks created by placing a thread with a
1846 	 * blocked lock on the run-queue of a remote processor.  The deadlock
1847 	 * occurs when a third processor attempts to lock the two queues in
1848 	 * question while the target processor is spinning with its own
1849 	 * run-queue lock held while waiting for the blocked lock to clear.
1850 	 */
1851 	tdq_lock_pair(tdn, tdq);
1852 	tdq_add(tdn, td, flags);
1853 	tdq_notify(tdn, td);
1854 	TDQ_UNLOCK(tdn);
1855 	spinlock_exit();
1856 #endif
1857 	return (TDQ_LOCKPTR(tdn));
1858 }
1859 
1860 /*
1861  * Variadic version of thread_lock_unblock() that does not assume td_lock
1862  * is blocked.
1863  */
1864 static inline void
1865 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1866 {
1867 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1868 	    (uintptr_t)mtx);
1869 }
1870 
1871 /*
1872  * Switch threads.  This function has to handle threads coming in while
1873  * blocked for some reason, running, or idle.  It also must deal with
1874  * migrating a thread from one queue to another as running threads may
1875  * be assigned elsewhere via binding.
1876  */
1877 void
1878 sched_switch(struct thread *td, struct thread *newtd, int flags)
1879 {
1880 	struct tdq *tdq;
1881 	struct td_sched *ts;
1882 	struct mtx *mtx;
1883 	int srqflag;
1884 	int cpuid, preempted;
1885 
1886 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1887 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1888 
1889 	cpuid = PCPU_GET(cpuid);
1890 	tdq = TDQ_CPU(cpuid);
1891 	ts = td->td_sched;
1892 	mtx = td->td_lock;
1893 	sched_pctcpu_update(ts, 1);
1894 	ts->ts_rltick = ticks;
1895 	td->td_lastcpu = td->td_oncpu;
1896 	td->td_oncpu = NOCPU;
1897 	preempted = !((td->td_flags & TDF_SLICEEND) ||
1898 	    (flags & SWT_RELINQUISH));
1899 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1900 	td->td_owepreempt = 0;
1901 	if (!TD_IS_IDLETHREAD(td))
1902 		tdq->tdq_switchcnt++;
1903 	/*
1904 	 * The lock pointer in an idle thread should never change.  Reset it
1905 	 * to CAN_RUN as well.
1906 	 */
1907 	if (TD_IS_IDLETHREAD(td)) {
1908 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1909 		TD_SET_CAN_RUN(td);
1910 	} else if (TD_IS_RUNNING(td)) {
1911 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1912 		srqflag = preempted ?
1913 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1914 		    SRQ_OURSELF|SRQ_YIELDING;
1915 #ifdef SMP
1916 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1917 			ts->ts_cpu = sched_pickcpu(td, 0);
1918 #endif
1919 		if (ts->ts_cpu == cpuid)
1920 			tdq_runq_add(tdq, td, srqflag);
1921 		else {
1922 			KASSERT(THREAD_CAN_MIGRATE(td) ||
1923 			    (ts->ts_flags & TSF_BOUND) != 0,
1924 			    ("Thread %p shouldn't migrate", td));
1925 			mtx = sched_switch_migrate(tdq, td, srqflag);
1926 		}
1927 	} else {
1928 		/* This thread must be going to sleep. */
1929 		TDQ_LOCK(tdq);
1930 		mtx = thread_lock_block(td);
1931 		tdq_load_rem(tdq, td);
1932 	}
1933 	/*
1934 	 * We enter here with the thread blocked and assigned to the
1935 	 * appropriate cpu run-queue or sleep-queue and with the current
1936 	 * thread-queue locked.
1937 	 */
1938 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1939 	newtd = choosethread();
1940 	/*
1941 	 * Call the MD code to switch contexts if necessary.
1942 	 */
1943 	if (td != newtd) {
1944 #ifdef	HWPMC_HOOKS
1945 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1946 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1947 #endif
1948 		SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc);
1949 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1950 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1951 		sched_pctcpu_update(newtd->td_sched, 0);
1952 
1953 #ifdef KDTRACE_HOOKS
1954 		/*
1955 		 * If DTrace has set the active vtime enum to anything
1956 		 * other than INACTIVE (0), then it should have set the
1957 		 * function to call.
1958 		 */
1959 		if (dtrace_vtime_active)
1960 			(*dtrace_vtime_switch_func)(newtd);
1961 #endif
1962 
1963 		cpu_switch(td, newtd, mtx);
1964 		/*
1965 		 * We may return from cpu_switch on a different cpu.  However,
1966 		 * we always return with td_lock pointing to the current cpu's
1967 		 * run queue lock.
1968 		 */
1969 		cpuid = PCPU_GET(cpuid);
1970 		tdq = TDQ_CPU(cpuid);
1971 		lock_profile_obtain_lock_success(
1972 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1973 
1974 		SDT_PROBE0(sched, , , on__cpu);
1975 #ifdef	HWPMC_HOOKS
1976 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1977 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1978 #endif
1979 	} else {
1980 		thread_unblock_switch(td, mtx);
1981 		SDT_PROBE0(sched, , , remain__cpu);
1982 	}
1983 	/*
1984 	 * Assert that all went well and return.
1985 	 */
1986 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1987 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1988 	td->td_oncpu = cpuid;
1989 }
1990 
1991 /*
1992  * Adjust thread priorities as a result of a nice request.
1993  */
1994 void
1995 sched_nice(struct proc *p, int nice)
1996 {
1997 	struct thread *td;
1998 
1999 	PROC_LOCK_ASSERT(p, MA_OWNED);
2000 
2001 	p->p_nice = nice;
2002 	FOREACH_THREAD_IN_PROC(p, td) {
2003 		thread_lock(td);
2004 		sched_priority(td);
2005 		sched_prio(td, td->td_base_user_pri);
2006 		thread_unlock(td);
2007 	}
2008 }
2009 
2010 /*
2011  * Record the sleep time for the interactivity scorer.
2012  */
2013 void
2014 sched_sleep(struct thread *td, int prio)
2015 {
2016 
2017 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2018 
2019 	td->td_slptick = ticks;
2020 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2021 		td->td_flags |= TDF_CANSWAP;
2022 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2023 		return;
2024 	if (static_boost == 1 && prio)
2025 		sched_prio(td, prio);
2026 	else if (static_boost && td->td_priority > static_boost)
2027 		sched_prio(td, static_boost);
2028 }
2029 
2030 /*
2031  * Schedule a thread to resume execution and record how long it voluntarily
2032  * slept.  We also update the pctcpu, interactivity, and priority.
2033  */
2034 void
2035 sched_wakeup(struct thread *td)
2036 {
2037 	struct td_sched *ts;
2038 	int slptick;
2039 
2040 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2041 	ts = td->td_sched;
2042 	td->td_flags &= ~TDF_CANSWAP;
2043 	/*
2044 	 * If we slept for more than a tick update our interactivity and
2045 	 * priority.
2046 	 */
2047 	slptick = td->td_slptick;
2048 	td->td_slptick = 0;
2049 	if (slptick && slptick != ticks) {
2050 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2051 		sched_interact_update(td);
2052 		sched_pctcpu_update(ts, 0);
2053 	}
2054 	/*
2055 	 * Reset the slice value since we slept and advanced the round-robin.
2056 	 */
2057 	ts->ts_slice = 0;
2058 	sched_add(td, SRQ_BORING);
2059 }
2060 
2061 /*
2062  * Penalize the parent for creating a new child and initialize the child's
2063  * priority.
2064  */
2065 void
2066 sched_fork(struct thread *td, struct thread *child)
2067 {
2068 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2069 	sched_pctcpu_update(td->td_sched, 1);
2070 	sched_fork_thread(td, child);
2071 	/*
2072 	 * Penalize the parent and child for forking.
2073 	 */
2074 	sched_interact_fork(child);
2075 	sched_priority(child);
2076 	td->td_sched->ts_runtime += tickincr;
2077 	sched_interact_update(td);
2078 	sched_priority(td);
2079 }
2080 
2081 /*
2082  * Fork a new thread, may be within the same process.
2083  */
2084 void
2085 sched_fork_thread(struct thread *td, struct thread *child)
2086 {
2087 	struct td_sched *ts;
2088 	struct td_sched *ts2;
2089 	struct tdq *tdq;
2090 
2091 	tdq = TDQ_SELF();
2092 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2093 	/*
2094 	 * Initialize child.
2095 	 */
2096 	ts = td->td_sched;
2097 	ts2 = child->td_sched;
2098 	child->td_oncpu = NOCPU;
2099 	child->td_lastcpu = NOCPU;
2100 	child->td_lock = TDQ_LOCKPTR(tdq);
2101 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2102 	ts2->ts_cpu = ts->ts_cpu;
2103 	ts2->ts_flags = 0;
2104 	/*
2105 	 * Grab our parents cpu estimation information.
2106 	 */
2107 	ts2->ts_ticks = ts->ts_ticks;
2108 	ts2->ts_ltick = ts->ts_ltick;
2109 	ts2->ts_ftick = ts->ts_ftick;
2110 	/*
2111 	 * Do not inherit any borrowed priority from the parent.
2112 	 */
2113 	child->td_priority = child->td_base_pri;
2114 	/*
2115 	 * And update interactivity score.
2116 	 */
2117 	ts2->ts_slptime = ts->ts_slptime;
2118 	ts2->ts_runtime = ts->ts_runtime;
2119 	/* Attempt to quickly learn interactivity. */
2120 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2121 #ifdef KTR
2122 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2123 #endif
2124 }
2125 
2126 /*
2127  * Adjust the priority class of a thread.
2128  */
2129 void
2130 sched_class(struct thread *td, int class)
2131 {
2132 
2133 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2134 	if (td->td_pri_class == class)
2135 		return;
2136 	td->td_pri_class = class;
2137 }
2138 
2139 /*
2140  * Return some of the child's priority and interactivity to the parent.
2141  */
2142 void
2143 sched_exit(struct proc *p, struct thread *child)
2144 {
2145 	struct thread *td;
2146 
2147 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2148 	    "prio:%d", child->td_priority);
2149 	PROC_LOCK_ASSERT(p, MA_OWNED);
2150 	td = FIRST_THREAD_IN_PROC(p);
2151 	sched_exit_thread(td, child);
2152 }
2153 
2154 /*
2155  * Penalize another thread for the time spent on this one.  This helps to
2156  * worsen the priority and interactivity of processes which schedule batch
2157  * jobs such as make.  This has little effect on the make process itself but
2158  * causes new processes spawned by it to receive worse scores immediately.
2159  */
2160 void
2161 sched_exit_thread(struct thread *td, struct thread *child)
2162 {
2163 
2164 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2165 	    "prio:%d", child->td_priority);
2166 	/*
2167 	 * Give the child's runtime to the parent without returning the
2168 	 * sleep time as a penalty to the parent.  This causes shells that
2169 	 * launch expensive things to mark their children as expensive.
2170 	 */
2171 	thread_lock(td);
2172 	td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2173 	sched_interact_update(td);
2174 	sched_priority(td);
2175 	thread_unlock(td);
2176 }
2177 
2178 void
2179 sched_preempt(struct thread *td)
2180 {
2181 	struct tdq *tdq;
2182 
2183 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2184 
2185 	thread_lock(td);
2186 	tdq = TDQ_SELF();
2187 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2188 	tdq->tdq_ipipending = 0;
2189 	if (td->td_priority > tdq->tdq_lowpri) {
2190 		int flags;
2191 
2192 		flags = SW_INVOL | SW_PREEMPT;
2193 		if (td->td_critnest > 1)
2194 			td->td_owepreempt = 1;
2195 		else if (TD_IS_IDLETHREAD(td))
2196 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2197 		else
2198 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2199 	}
2200 	thread_unlock(td);
2201 }
2202 
2203 /*
2204  * Fix priorities on return to user-space.  Priorities may be elevated due
2205  * to static priorities in msleep() or similar.
2206  */
2207 void
2208 sched_userret(struct thread *td)
2209 {
2210 	/*
2211 	 * XXX we cheat slightly on the locking here to avoid locking in
2212 	 * the usual case.  Setting td_priority here is essentially an
2213 	 * incomplete workaround for not setting it properly elsewhere.
2214 	 * Now that some interrupt handlers are threads, not setting it
2215 	 * properly elsewhere can clobber it in the window between setting
2216 	 * it here and returning to user mode, so don't waste time setting
2217 	 * it perfectly here.
2218 	 */
2219 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2220 	    ("thread with borrowed priority returning to userland"));
2221 	if (td->td_priority != td->td_user_pri) {
2222 		thread_lock(td);
2223 		td->td_priority = td->td_user_pri;
2224 		td->td_base_pri = td->td_user_pri;
2225 		tdq_setlowpri(TDQ_SELF(), td);
2226 		thread_unlock(td);
2227         }
2228 }
2229 
2230 /*
2231  * Handle a stathz tick.  This is really only relevant for timeshare
2232  * threads.
2233  */
2234 void
2235 sched_clock(struct thread *td)
2236 {
2237 	struct tdq *tdq;
2238 	struct td_sched *ts;
2239 
2240 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2241 	tdq = TDQ_SELF();
2242 #ifdef SMP
2243 	/*
2244 	 * We run the long term load balancer infrequently on the first cpu.
2245 	 */
2246 	if (balance_tdq == tdq) {
2247 		if (balance_ticks && --balance_ticks == 0)
2248 			sched_balance();
2249 	}
2250 #endif
2251 	/*
2252 	 * Save the old switch count so we have a record of the last ticks
2253 	 * activity.   Initialize the new switch count based on our load.
2254 	 * If there is some activity seed it to reflect that.
2255 	 */
2256 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2257 	tdq->tdq_switchcnt = tdq->tdq_load;
2258 	/*
2259 	 * Advance the insert index once for each tick to ensure that all
2260 	 * threads get a chance to run.
2261 	 */
2262 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2263 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2264 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2265 			tdq->tdq_ridx = tdq->tdq_idx;
2266 	}
2267 	ts = td->td_sched;
2268 	sched_pctcpu_update(ts, 1);
2269 	if (td->td_pri_class & PRI_FIFO_BIT)
2270 		return;
2271 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2272 		/*
2273 		 * We used a tick; charge it to the thread so
2274 		 * that we can compute our interactivity.
2275 		 */
2276 		td->td_sched->ts_runtime += tickincr;
2277 		sched_interact_update(td);
2278 		sched_priority(td);
2279 	}
2280 
2281 	/*
2282 	 * Force a context switch if the current thread has used up a full
2283 	 * time slice (default is 100ms).
2284 	 */
2285 	if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2286 		ts->ts_slice = 0;
2287 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2288 	}
2289 }
2290 
2291 /*
2292  * Called once per hz tick.
2293  */
2294 void
2295 sched_tick(int cnt)
2296 {
2297 
2298 }
2299 
2300 /*
2301  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2302  * cooperative idle threads.
2303  */
2304 int
2305 sched_runnable(void)
2306 {
2307 	struct tdq *tdq;
2308 	int load;
2309 
2310 	load = 1;
2311 
2312 	tdq = TDQ_SELF();
2313 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2314 		if (tdq->tdq_load > 0)
2315 			goto out;
2316 	} else
2317 		if (tdq->tdq_load - 1 > 0)
2318 			goto out;
2319 	load = 0;
2320 out:
2321 	return (load);
2322 }
2323 
2324 /*
2325  * Choose the highest priority thread to run.  The thread is removed from
2326  * the run-queue while running however the load remains.  For SMP we set
2327  * the tdq in the global idle bitmask if it idles here.
2328  */
2329 struct thread *
2330 sched_choose(void)
2331 {
2332 	struct thread *td;
2333 	struct tdq *tdq;
2334 
2335 	tdq = TDQ_SELF();
2336 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2337 	td = tdq_choose(tdq);
2338 	if (td) {
2339 		tdq_runq_rem(tdq, td);
2340 		tdq->tdq_lowpri = td->td_priority;
2341 		return (td);
2342 	}
2343 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2344 	return (PCPU_GET(idlethread));
2345 }
2346 
2347 /*
2348  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2349  * we always request it once we exit a critical section.
2350  */
2351 static inline void
2352 sched_setpreempt(struct thread *td)
2353 {
2354 	struct thread *ctd;
2355 	int cpri;
2356 	int pri;
2357 
2358 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2359 
2360 	ctd = curthread;
2361 	pri = td->td_priority;
2362 	cpri = ctd->td_priority;
2363 	if (pri < cpri)
2364 		ctd->td_flags |= TDF_NEEDRESCHED;
2365 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2366 		return;
2367 	if (!sched_shouldpreempt(pri, cpri, 0))
2368 		return;
2369 	ctd->td_owepreempt = 1;
2370 }
2371 
2372 /*
2373  * Add a thread to a thread queue.  Select the appropriate runq and add the
2374  * thread to it.  This is the internal function called when the tdq is
2375  * predetermined.
2376  */
2377 void
2378 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2379 {
2380 
2381 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2382 	KASSERT((td->td_inhibitors == 0),
2383 	    ("sched_add: trying to run inhibited thread"));
2384 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2385 	    ("sched_add: bad thread state"));
2386 	KASSERT(td->td_flags & TDF_INMEM,
2387 	    ("sched_add: thread swapped out"));
2388 
2389 	if (td->td_priority < tdq->tdq_lowpri)
2390 		tdq->tdq_lowpri = td->td_priority;
2391 	tdq_runq_add(tdq, td, flags);
2392 	tdq_load_add(tdq, td);
2393 }
2394 
2395 /*
2396  * Select the target thread queue and add a thread to it.  Request
2397  * preemption or IPI a remote processor if required.
2398  */
2399 void
2400 sched_add(struct thread *td, int flags)
2401 {
2402 	struct tdq *tdq;
2403 #ifdef SMP
2404 	int cpu;
2405 #endif
2406 
2407 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2408 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2409 	    sched_tdname(curthread));
2410 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2411 	    KTR_ATTR_LINKED, sched_tdname(td));
2412 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2413 	    flags & SRQ_PREEMPTED);
2414 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2415 	/*
2416 	 * Recalculate the priority before we select the target cpu or
2417 	 * run-queue.
2418 	 */
2419 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2420 		sched_priority(td);
2421 #ifdef SMP
2422 	/*
2423 	 * Pick the destination cpu and if it isn't ours transfer to the
2424 	 * target cpu.
2425 	 */
2426 	cpu = sched_pickcpu(td, flags);
2427 	tdq = sched_setcpu(td, cpu, flags);
2428 	tdq_add(tdq, td, flags);
2429 	if (cpu != PCPU_GET(cpuid)) {
2430 		tdq_notify(tdq, td);
2431 		return;
2432 	}
2433 #else
2434 	tdq = TDQ_SELF();
2435 	TDQ_LOCK(tdq);
2436 	/*
2437 	 * Now that the thread is moving to the run-queue, set the lock
2438 	 * to the scheduler's lock.
2439 	 */
2440 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2441 	tdq_add(tdq, td, flags);
2442 #endif
2443 	if (!(flags & SRQ_YIELDING))
2444 		sched_setpreempt(td);
2445 }
2446 
2447 /*
2448  * Remove a thread from a run-queue without running it.  This is used
2449  * when we're stealing a thread from a remote queue.  Otherwise all threads
2450  * exit by calling sched_exit_thread() and sched_throw() themselves.
2451  */
2452 void
2453 sched_rem(struct thread *td)
2454 {
2455 	struct tdq *tdq;
2456 
2457 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2458 	    "prio:%d", td->td_priority);
2459 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2460 	tdq = TDQ_CPU(td->td_sched->ts_cpu);
2461 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2462 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2463 	KASSERT(TD_ON_RUNQ(td),
2464 	    ("sched_rem: thread not on run queue"));
2465 	tdq_runq_rem(tdq, td);
2466 	tdq_load_rem(tdq, td);
2467 	TD_SET_CAN_RUN(td);
2468 	if (td->td_priority == tdq->tdq_lowpri)
2469 		tdq_setlowpri(tdq, NULL);
2470 }
2471 
2472 /*
2473  * Fetch cpu utilization information.  Updates on demand.
2474  */
2475 fixpt_t
2476 sched_pctcpu(struct thread *td)
2477 {
2478 	fixpt_t pctcpu;
2479 	struct td_sched *ts;
2480 
2481 	pctcpu = 0;
2482 	ts = td->td_sched;
2483 	if (ts == NULL)
2484 		return (0);
2485 
2486 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2487 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2488 	if (ts->ts_ticks) {
2489 		int rtick;
2490 
2491 		/* How many rtick per second ? */
2492 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2493 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2494 	}
2495 
2496 	return (pctcpu);
2497 }
2498 
2499 /*
2500  * Enforce affinity settings for a thread.  Called after adjustments to
2501  * cpumask.
2502  */
2503 void
2504 sched_affinity(struct thread *td)
2505 {
2506 #ifdef SMP
2507 	struct td_sched *ts;
2508 
2509 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2510 	ts = td->td_sched;
2511 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2512 		return;
2513 	if (TD_ON_RUNQ(td)) {
2514 		sched_rem(td);
2515 		sched_add(td, SRQ_BORING);
2516 		return;
2517 	}
2518 	if (!TD_IS_RUNNING(td))
2519 		return;
2520 	/*
2521 	 * Force a switch before returning to userspace.  If the
2522 	 * target thread is not running locally send an ipi to force
2523 	 * the issue.
2524 	 */
2525 	td->td_flags |= TDF_NEEDRESCHED;
2526 	if (td != curthread)
2527 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2528 #endif
2529 }
2530 
2531 /*
2532  * Bind a thread to a target cpu.
2533  */
2534 void
2535 sched_bind(struct thread *td, int cpu)
2536 {
2537 	struct td_sched *ts;
2538 
2539 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2540 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2541 	ts = td->td_sched;
2542 	if (ts->ts_flags & TSF_BOUND)
2543 		sched_unbind(td);
2544 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2545 	ts->ts_flags |= TSF_BOUND;
2546 	sched_pin();
2547 	if (PCPU_GET(cpuid) == cpu)
2548 		return;
2549 	ts->ts_cpu = cpu;
2550 	/* When we return from mi_switch we'll be on the correct cpu. */
2551 	mi_switch(SW_VOL, NULL);
2552 }
2553 
2554 /*
2555  * Release a bound thread.
2556  */
2557 void
2558 sched_unbind(struct thread *td)
2559 {
2560 	struct td_sched *ts;
2561 
2562 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2563 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2564 	ts = td->td_sched;
2565 	if ((ts->ts_flags & TSF_BOUND) == 0)
2566 		return;
2567 	ts->ts_flags &= ~TSF_BOUND;
2568 	sched_unpin();
2569 }
2570 
2571 int
2572 sched_is_bound(struct thread *td)
2573 {
2574 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2575 	return (td->td_sched->ts_flags & TSF_BOUND);
2576 }
2577 
2578 /*
2579  * Basic yield call.
2580  */
2581 void
2582 sched_relinquish(struct thread *td)
2583 {
2584 	thread_lock(td);
2585 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2586 	thread_unlock(td);
2587 }
2588 
2589 /*
2590  * Return the total system load.
2591  */
2592 int
2593 sched_load(void)
2594 {
2595 #ifdef SMP
2596 	int total;
2597 	int i;
2598 
2599 	total = 0;
2600 	CPU_FOREACH(i)
2601 		total += TDQ_CPU(i)->tdq_sysload;
2602 	return (total);
2603 #else
2604 	return (TDQ_SELF()->tdq_sysload);
2605 #endif
2606 }
2607 
2608 int
2609 sched_sizeof_proc(void)
2610 {
2611 	return (sizeof(struct proc));
2612 }
2613 
2614 int
2615 sched_sizeof_thread(void)
2616 {
2617 	return (sizeof(struct thread) + sizeof(struct td_sched));
2618 }
2619 
2620 #ifdef SMP
2621 #define	TDQ_IDLESPIN(tdq)						\
2622     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2623 #else
2624 #define	TDQ_IDLESPIN(tdq)	1
2625 #endif
2626 
2627 /*
2628  * The actual idle process.
2629  */
2630 void
2631 sched_idletd(void *dummy)
2632 {
2633 	struct thread *td;
2634 	struct tdq *tdq;
2635 	int oldswitchcnt, switchcnt;
2636 	int i;
2637 
2638 	mtx_assert(&Giant, MA_NOTOWNED);
2639 	td = curthread;
2640 	tdq = TDQ_SELF();
2641 	THREAD_NO_SLEEPING();
2642 	oldswitchcnt = -1;
2643 	for (;;) {
2644 		if (tdq->tdq_load) {
2645 			thread_lock(td);
2646 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2647 			thread_unlock(td);
2648 		}
2649 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2650 #ifdef SMP
2651 		if (switchcnt != oldswitchcnt) {
2652 			oldswitchcnt = switchcnt;
2653 			if (tdq_idled(tdq) == 0)
2654 				continue;
2655 		}
2656 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2657 #else
2658 		oldswitchcnt = switchcnt;
2659 #endif
2660 		/*
2661 		 * If we're switching very frequently, spin while checking
2662 		 * for load rather than entering a low power state that
2663 		 * may require an IPI.  However, don't do any busy
2664 		 * loops while on SMT machines as this simply steals
2665 		 * cycles from cores doing useful work.
2666 		 */
2667 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2668 			for (i = 0; i < sched_idlespins; i++) {
2669 				if (tdq->tdq_load)
2670 					break;
2671 				cpu_spinwait();
2672 			}
2673 		}
2674 
2675 		/* If there was context switch during spin, restart it. */
2676 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2677 		if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2678 			continue;
2679 
2680 		/* Run main MD idle handler. */
2681 		tdq->tdq_cpu_idle = 1;
2682 		/*
2683 		 * Make sure that tdq_cpu_idle update is globally visible
2684 		 * before cpu_idle() read tdq_load.  The order is important
2685 		 * to avoid race with tdq_notify.
2686 		 */
2687 		atomic_thread_fence_seq_cst();
2688 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2689 		tdq->tdq_cpu_idle = 0;
2690 
2691 		/*
2692 		 * Account thread-less hardware interrupts and
2693 		 * other wakeup reasons equal to context switches.
2694 		 */
2695 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2696 		if (switchcnt != oldswitchcnt)
2697 			continue;
2698 		tdq->tdq_switchcnt++;
2699 		oldswitchcnt++;
2700 	}
2701 }
2702 
2703 /*
2704  * A CPU is entering for the first time or a thread is exiting.
2705  */
2706 void
2707 sched_throw(struct thread *td)
2708 {
2709 	struct thread *newtd;
2710 	struct tdq *tdq;
2711 
2712 	tdq = TDQ_SELF();
2713 	if (td == NULL) {
2714 		/* Correct spinlock nesting and acquire the correct lock. */
2715 		TDQ_LOCK(tdq);
2716 		spinlock_exit();
2717 		PCPU_SET(switchtime, cpu_ticks());
2718 		PCPU_SET(switchticks, ticks);
2719 	} else {
2720 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2721 		tdq_load_rem(tdq, td);
2722 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2723 		td->td_lastcpu = td->td_oncpu;
2724 		td->td_oncpu = NOCPU;
2725 	}
2726 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2727 	newtd = choosethread();
2728 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2729 	cpu_throw(td, newtd);		/* doesn't return */
2730 }
2731 
2732 /*
2733  * This is called from fork_exit().  Just acquire the correct locks and
2734  * let fork do the rest of the work.
2735  */
2736 void
2737 sched_fork_exit(struct thread *td)
2738 {
2739 	struct tdq *tdq;
2740 	int cpuid;
2741 
2742 	/*
2743 	 * Finish setting up thread glue so that it begins execution in a
2744 	 * non-nested critical section with the scheduler lock held.
2745 	 */
2746 	cpuid = PCPU_GET(cpuid);
2747 	tdq = TDQ_CPU(cpuid);
2748 	if (TD_IS_IDLETHREAD(td))
2749 		td->td_lock = TDQ_LOCKPTR(tdq);
2750 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2751 	td->td_oncpu = cpuid;
2752 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2753 	lock_profile_obtain_lock_success(
2754 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2755 }
2756 
2757 /*
2758  * Create on first use to catch odd startup conditons.
2759  */
2760 char *
2761 sched_tdname(struct thread *td)
2762 {
2763 #ifdef KTR
2764 	struct td_sched *ts;
2765 
2766 	ts = td->td_sched;
2767 	if (ts->ts_name[0] == '\0')
2768 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2769 		    "%s tid %d", td->td_name, td->td_tid);
2770 	return (ts->ts_name);
2771 #else
2772 	return (td->td_name);
2773 #endif
2774 }
2775 
2776 #ifdef KTR
2777 void
2778 sched_clear_tdname(struct thread *td)
2779 {
2780 	struct td_sched *ts;
2781 
2782 	ts = td->td_sched;
2783 	ts->ts_name[0] = '\0';
2784 }
2785 #endif
2786 
2787 #ifdef SMP
2788 
2789 /*
2790  * Build the CPU topology dump string. Is recursively called to collect
2791  * the topology tree.
2792  */
2793 static int
2794 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2795     int indent)
2796 {
2797 	char cpusetbuf[CPUSETBUFSIZ];
2798 	int i, first;
2799 
2800 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2801 	    "", 1 + indent / 2, cg->cg_level);
2802 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2803 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2804 	first = TRUE;
2805 	for (i = 0; i < MAXCPU; i++) {
2806 		if (CPU_ISSET(i, &cg->cg_mask)) {
2807 			if (!first)
2808 				sbuf_printf(sb, ", ");
2809 			else
2810 				first = FALSE;
2811 			sbuf_printf(sb, "%d", i);
2812 		}
2813 	}
2814 	sbuf_printf(sb, "</cpu>\n");
2815 
2816 	if (cg->cg_flags != 0) {
2817 		sbuf_printf(sb, "%*s <flags>", indent, "");
2818 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2819 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2820 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2821 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2822 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2823 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2824 		sbuf_printf(sb, "</flags>\n");
2825 	}
2826 
2827 	if (cg->cg_children > 0) {
2828 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2829 		for (i = 0; i < cg->cg_children; i++)
2830 			sysctl_kern_sched_topology_spec_internal(sb,
2831 			    &cg->cg_child[i], indent+2);
2832 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2833 	}
2834 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2835 	return (0);
2836 }
2837 
2838 /*
2839  * Sysctl handler for retrieving topology dump. It's a wrapper for
2840  * the recursive sysctl_kern_smp_topology_spec_internal().
2841  */
2842 static int
2843 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2844 {
2845 	struct sbuf *topo;
2846 	int err;
2847 
2848 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2849 
2850 	topo = sbuf_new_for_sysctl(NULL, NULL, 512, req);
2851 	if (topo == NULL)
2852 		return (ENOMEM);
2853 
2854 	sbuf_printf(topo, "<groups>\n");
2855 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2856 	sbuf_printf(topo, "</groups>\n");
2857 
2858 	if (err == 0) {
2859 		err = sbuf_finish(topo);
2860 	}
2861 	sbuf_delete(topo);
2862 	return (err);
2863 }
2864 
2865 #endif
2866 
2867 static int
2868 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2869 {
2870 	int error, new_val, period;
2871 
2872 	period = 1000000 / realstathz;
2873 	new_val = period * sched_slice;
2874 	error = sysctl_handle_int(oidp, &new_val, 0, req);
2875 	if (error != 0 || req->newptr == NULL)
2876 		return (error);
2877 	if (new_val <= 0)
2878 		return (EINVAL);
2879 	sched_slice = imax(1, (new_val + period / 2) / period);
2880 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
2881 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2882 	    realstathz);
2883 	return (0);
2884 }
2885 
2886 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2887 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2888     "Scheduler name");
2889 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2890     NULL, 0, sysctl_kern_quantum, "I",
2891     "Quantum for timeshare threads in microseconds");
2892 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2893     "Quantum for timeshare threads in stathz ticks");
2894 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2895     "Interactivity score threshold");
2896 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2897     &preempt_thresh, 0,
2898     "Maximal (lowest) priority for preemption");
2899 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2900     "Assign static kernel priorities to sleeping threads");
2901 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2902     "Number of times idle thread will spin waiting for new work");
2903 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2904     &sched_idlespinthresh, 0,
2905     "Threshold before we will permit idle thread spinning");
2906 #ifdef SMP
2907 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2908     "Number of hz ticks to keep thread affinity for");
2909 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2910     "Enables the long-term load balancer");
2911 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2912     &balance_interval, 0,
2913     "Average period in stathz ticks to run the long-term balancer");
2914 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2915     "Attempts to steal work from other cores before idling");
2916 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2917     "Minimum load on remote CPU before we'll steal");
2918 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2919     CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2920     "XML dump of detected CPU topology");
2921 #endif
2922 
2923 /* ps compat.  All cpu percentages from ULE are weighted. */
2924 static int ccpu = 0;
2925 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2926