xref: /freebsd/sys/kern/sched_ule.c (revision 595e514d0df2bac5b813d35f83e32875dbf16a83)
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_kdtrace.h"
43 #include "opt_sched.h"
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/kdb.h>
48 #include <sys/kernel.h>
49 #include <sys/ktr.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 #if defined(__powerpc__) && defined(BOOKE_E500)
81 #error "This architecture is not currently compatible with ULE"
82 #endif
83 
84 #define	KTR_ULE	0
85 
86 #define	TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX)))
87 #define	TDQ_NAME_LEN	(sizeof("sched lock ") + sizeof(__XSTRING(MAXCPU)))
88 #define	TDQ_LOADNAME_LEN	(sizeof("CPU ") + sizeof(__XSTRING(MAXCPU)) - 1 + sizeof(" load"))
89 
90 /*
91  * Thread scheduler specific section.  All fields are protected
92  * by the thread lock.
93  */
94 struct td_sched {
95 	struct runq	*ts_runq;	/* Run-queue we're queued on. */
96 	short		ts_flags;	/* TSF_* flags. */
97 	u_char		ts_cpu;		/* CPU that we have affinity for. */
98 	int		ts_rltick;	/* Real last tick, for affinity. */
99 	int		ts_slice;	/* Ticks of slice remaining. */
100 	u_int		ts_slptime;	/* Number of ticks we vol. slept */
101 	u_int		ts_runtime;	/* Number of ticks we were running */
102 	int		ts_ltick;	/* Last tick that we were running on */
103 	int		ts_ftick;	/* First tick that we were running on */
104 	int		ts_ticks;	/* Tick count */
105 #ifdef KTR
106 	char		ts_name[TS_NAME_LEN];
107 #endif
108 };
109 /* flags kept in ts_flags */
110 #define	TSF_BOUND	0x0001		/* Thread can not migrate. */
111 #define	TSF_XFERABLE	0x0002		/* Thread was added as transferable. */
112 
113 static struct td_sched td_sched0;
114 
115 #define	THREAD_CAN_MIGRATE(td)	((td)->td_pinned == 0)
116 #define	THREAD_CAN_SCHED(td, cpu)	\
117     CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask)
118 
119 /*
120  * Priority ranges used for interactive and non-interactive timeshare
121  * threads.  The timeshare priorities are split up into four ranges.
122  * The first range handles interactive threads.  The last three ranges
123  * (NHALF, x, and NHALF) handle non-interactive threads with the outer
124  * ranges supporting nice values.
125  */
126 #define	PRI_TIMESHARE_RANGE	(PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
127 #define	PRI_INTERACT_RANGE	((PRI_TIMESHARE_RANGE - SCHED_PRI_NRESV) / 2)
128 #define	PRI_BATCH_RANGE		(PRI_TIMESHARE_RANGE - PRI_INTERACT_RANGE)
129 
130 #define	PRI_MIN_INTERACT	PRI_MIN_TIMESHARE
131 #define	PRI_MAX_INTERACT	(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE - 1)
132 #define	PRI_MIN_BATCH		(PRI_MIN_TIMESHARE + PRI_INTERACT_RANGE)
133 #define	PRI_MAX_BATCH		PRI_MAX_TIMESHARE
134 
135 /*
136  * Cpu percentage computation macros and defines.
137  *
138  * SCHED_TICK_SECS:	Number of seconds to average the cpu usage across.
139  * SCHED_TICK_TARG:	Number of hz ticks to average the cpu usage across.
140  * SCHED_TICK_MAX:	Maximum number of ticks before scaling back.
141  * SCHED_TICK_SHIFT:	Shift factor to avoid rounding away results.
142  * SCHED_TICK_HZ:	Compute the number of hz ticks for a given ticks count.
143  * SCHED_TICK_TOTAL:	Gives the amount of time we've been recording ticks.
144  */
145 #define	SCHED_TICK_SECS		10
146 #define	SCHED_TICK_TARG		(hz * SCHED_TICK_SECS)
147 #define	SCHED_TICK_MAX		(SCHED_TICK_TARG + hz)
148 #define	SCHED_TICK_SHIFT	10
149 #define	SCHED_TICK_HZ(ts)	((ts)->ts_ticks >> SCHED_TICK_SHIFT)
150 #define	SCHED_TICK_TOTAL(ts)	(max((ts)->ts_ltick - (ts)->ts_ftick, hz))
151 
152 /*
153  * These macros determine priorities for non-interactive threads.  They are
154  * assigned a priority based on their recent cpu utilization as expressed
155  * by the ratio of ticks to the tick total.  NHALF priorities at the start
156  * and end of the MIN to MAX timeshare range are only reachable with negative
157  * or positive nice respectively.
158  *
159  * PRI_RANGE:	Priority range for utilization dependent priorities.
160  * PRI_NRESV:	Number of nice values.
161  * PRI_TICKS:	Compute a priority in PRI_RANGE from the ticks count and total.
162  * PRI_NICE:	Determines the part of the priority inherited from nice.
163  */
164 #define	SCHED_PRI_NRESV		(PRIO_MAX - PRIO_MIN)
165 #define	SCHED_PRI_NHALF		(SCHED_PRI_NRESV / 2)
166 #define	SCHED_PRI_MIN		(PRI_MIN_BATCH + SCHED_PRI_NHALF)
167 #define	SCHED_PRI_MAX		(PRI_MAX_BATCH - SCHED_PRI_NHALF)
168 #define	SCHED_PRI_RANGE		(SCHED_PRI_MAX - SCHED_PRI_MIN + 1)
169 #define	SCHED_PRI_TICKS(ts)						\
170     (SCHED_TICK_HZ((ts)) /						\
171     (roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
172 #define	SCHED_PRI_NICE(nice)	(nice)
173 
174 /*
175  * These determine the interactivity of a process.  Interactivity differs from
176  * cpu utilization in that it expresses the voluntary time slept vs time ran
177  * while cpu utilization includes all time not running.  This more accurately
178  * models the intent of the thread.
179  *
180  * SLP_RUN_MAX:	Maximum amount of sleep time + run time we'll accumulate
181  *		before throttling back.
182  * SLP_RUN_FORK:	Maximum slp+run time to inherit at fork time.
183  * INTERACT_MAX:	Maximum interactivity value.  Smaller is better.
184  * INTERACT_THRESH:	Threshold for placement on the current runq.
185  */
186 #define	SCHED_SLP_RUN_MAX	((hz * 5) << SCHED_TICK_SHIFT)
187 #define	SCHED_SLP_RUN_FORK	((hz / 2) << SCHED_TICK_SHIFT)
188 #define	SCHED_INTERACT_MAX	(100)
189 #define	SCHED_INTERACT_HALF	(SCHED_INTERACT_MAX / 2)
190 #define	SCHED_INTERACT_THRESH	(30)
191 
192 /*
193  * These parameters determine the slice behavior for batch work.
194  */
195 #define	SCHED_SLICE_DEFAULT_DIVISOR	10	/* ~94 ms, 12 stathz ticks. */
196 #define	SCHED_SLICE_MIN_DIVISOR		6	/* DEFAULT/MIN = ~16 ms. */
197 
198 /* Flags kept in td_flags. */
199 #define	TDF_SLICEEND	TDF_SCHED2	/* Thread time slice is over. */
200 
201 /*
202  * tickincr:		Converts a stathz tick into a hz domain scaled by
203  *			the shift factor.  Without the shift the error rate
204  *			due to rounding would be unacceptably high.
205  * realstathz:		stathz is sometimes 0 and run off of hz.
206  * sched_slice:		Runtime of each thread before rescheduling.
207  * preempt_thresh:	Priority threshold for preemption and remote IPIs.
208  */
209 static int sched_interact = SCHED_INTERACT_THRESH;
210 static int tickincr = 8 << SCHED_TICK_SHIFT;
211 static int realstathz = 127;	/* reset during boot. */
212 static int sched_slice = 10;	/* reset during boot. */
213 static int sched_slice_min = 1;	/* reset during boot. */
214 #ifdef PREEMPTION
215 #ifdef FULL_PREEMPTION
216 static int preempt_thresh = PRI_MAX_IDLE;
217 #else
218 static int preempt_thresh = PRI_MIN_KERN;
219 #endif
220 #else
221 static int preempt_thresh = 0;
222 #endif
223 static int static_boost = PRI_MIN_BATCH;
224 static int sched_idlespins = 10000;
225 static int sched_idlespinthresh = -1;
226 
227 /*
228  * tdq - per processor runqs and statistics.  All fields are protected by the
229  * tdq_lock.  The load and lowpri may be accessed without to avoid excess
230  * locking in sched_pickcpu();
231  */
232 struct tdq {
233 	/*
234 	 * Ordered to improve efficiency of cpu_search() and switch().
235 	 * tdq_lock is padded to avoid false sharing with tdq_load and
236 	 * tdq_cpu_idle.
237 	 */
238 	struct mtx_padalign tdq_lock;		/* run queue lock. */
239 	struct cpu_group *tdq_cg;		/* Pointer to cpu topology. */
240 	volatile int	tdq_load;		/* Aggregate load. */
241 	volatile int	tdq_cpu_idle;		/* cpu_idle() is active. */
242 	int		tdq_sysload;		/* For loadavg, !ITHD load. */
243 	int		tdq_transferable;	/* Transferable thread count. */
244 	short		tdq_switchcnt;		/* Switches this tick. */
245 	short		tdq_oldswitchcnt;	/* Switches last tick. */
246 	u_char		tdq_lowpri;		/* Lowest priority thread. */
247 	u_char		tdq_ipipending;		/* IPI pending. */
248 	u_char		tdq_idx;		/* Current insert index. */
249 	u_char		tdq_ridx;		/* Current removal index. */
250 	struct runq	tdq_realtime;		/* real-time run queue. */
251 	struct runq	tdq_timeshare;		/* timeshare run queue. */
252 	struct runq	tdq_idle;		/* Queue of IDLE threads. */
253 	char		tdq_name[TDQ_NAME_LEN];
254 #ifdef KTR
255 	char		tdq_loadname[TDQ_LOADNAME_LEN];
256 #endif
257 } __aligned(64);
258 
259 /* Idle thread states and config. */
260 #define	TDQ_RUNNING	1
261 #define	TDQ_IDLE	2
262 
263 #ifdef SMP
264 struct cpu_group *cpu_top;		/* CPU topology */
265 
266 #define	SCHED_AFFINITY_DEFAULT	(max(1, hz / 1000))
267 #define	SCHED_AFFINITY(ts, t)	((ts)->ts_rltick > ticks - ((t) * affinity))
268 
269 /*
270  * Run-time tunables.
271  */
272 static int rebalance = 1;
273 static int balance_interval = 128;	/* Default set in sched_initticks(). */
274 static int affinity;
275 static int steal_idle = 1;
276 static int steal_thresh = 2;
277 
278 /*
279  * One thread queue per processor.
280  */
281 static struct tdq	tdq_cpu[MAXCPU];
282 static struct tdq	*balance_tdq;
283 static int balance_ticks;
284 static DPCPU_DEFINE(uint32_t, randomval);
285 
286 #define	TDQ_SELF()	(&tdq_cpu[PCPU_GET(cpuid)])
287 #define	TDQ_CPU(x)	(&tdq_cpu[(x)])
288 #define	TDQ_ID(x)	((int)((x) - tdq_cpu))
289 #else	/* !SMP */
290 static struct tdq	tdq_cpu;
291 
292 #define	TDQ_ID(x)	(0)
293 #define	TDQ_SELF()	(&tdq_cpu)
294 #define	TDQ_CPU(x)	(&tdq_cpu)
295 #endif
296 
297 #define	TDQ_LOCK_ASSERT(t, type)	mtx_assert(TDQ_LOCKPTR((t)), (type))
298 #define	TDQ_LOCK(t)		mtx_lock_spin(TDQ_LOCKPTR((t)))
299 #define	TDQ_LOCK_FLAGS(t, f)	mtx_lock_spin_flags(TDQ_LOCKPTR((t)), (f))
300 #define	TDQ_UNLOCK(t)		mtx_unlock_spin(TDQ_LOCKPTR((t)))
301 #define	TDQ_LOCKPTR(t)		((struct mtx *)(&(t)->tdq_lock))
302 
303 static void sched_priority(struct thread *);
304 static void sched_thread_priority(struct thread *, u_char);
305 static int sched_interact_score(struct thread *);
306 static void sched_interact_update(struct thread *);
307 static void sched_interact_fork(struct thread *);
308 static void sched_pctcpu_update(struct td_sched *, int);
309 
310 /* Operations on per processor queues */
311 static struct thread *tdq_choose(struct tdq *);
312 static void tdq_setup(struct tdq *);
313 static void tdq_load_add(struct tdq *, struct thread *);
314 static void tdq_load_rem(struct tdq *, struct thread *);
315 static __inline void tdq_runq_add(struct tdq *, struct thread *, int);
316 static __inline void tdq_runq_rem(struct tdq *, struct thread *);
317 static inline int sched_shouldpreempt(int, int, int);
318 void tdq_print(int cpu);
319 static void runq_print(struct runq *rq);
320 static void tdq_add(struct tdq *, struct thread *, int);
321 #ifdef SMP
322 static int tdq_move(struct tdq *, struct tdq *);
323 static int tdq_idled(struct tdq *);
324 static void tdq_notify(struct tdq *, struct thread *);
325 static struct thread *tdq_steal(struct tdq *, int);
326 static struct thread *runq_steal(struct runq *, int);
327 static int sched_pickcpu(struct thread *, int);
328 static void sched_balance(void);
329 static int sched_balance_pair(struct tdq *, struct tdq *);
330 static inline struct tdq *sched_setcpu(struct thread *, int, int);
331 static inline void thread_unblock_switch(struct thread *, struct mtx *);
332 static struct mtx *sched_switch_migrate(struct tdq *, struct thread *, int);
333 static int sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS);
334 static int sysctl_kern_sched_topology_spec_internal(struct sbuf *sb,
335     struct cpu_group *cg, int indent);
336 #endif
337 
338 static void sched_setup(void *dummy);
339 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL);
340 
341 static void sched_initticks(void *dummy);
342 SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks,
343     NULL);
344 
345 SDT_PROVIDER_DEFINE(sched);
346 
347 SDT_PROBE_DEFINE3(sched, , , change_pri, change-pri, "struct thread *",
348     "struct proc *", "uint8_t");
349 SDT_PROBE_DEFINE3(sched, , , dequeue, dequeue, "struct thread *",
350     "struct proc *", "void *");
351 SDT_PROBE_DEFINE4(sched, , , enqueue, enqueue, "struct thread *",
352     "struct proc *", "void *", "int");
353 SDT_PROBE_DEFINE4(sched, , , lend_pri, lend-pri, "struct thread *",
354     "struct proc *", "uint8_t", "struct thread *");
355 SDT_PROBE_DEFINE2(sched, , , load_change, load-change, "int", "int");
356 SDT_PROBE_DEFINE2(sched, , , off_cpu, off-cpu, "struct thread *",
357     "struct proc *");
358 SDT_PROBE_DEFINE(sched, , , on_cpu, on-cpu);
359 SDT_PROBE_DEFINE(sched, , , remain_cpu, remain-cpu);
360 SDT_PROBE_DEFINE2(sched, , , surrender, surrender, "struct thread *",
361     "struct proc *");
362 
363 /*
364  * Print the threads waiting on a run-queue.
365  */
366 static void
367 runq_print(struct runq *rq)
368 {
369 	struct rqhead *rqh;
370 	struct thread *td;
371 	int pri;
372 	int j;
373 	int i;
374 
375 	for (i = 0; i < RQB_LEN; i++) {
376 		printf("\t\trunq bits %d 0x%zx\n",
377 		    i, rq->rq_status.rqb_bits[i]);
378 		for (j = 0; j < RQB_BPW; j++)
379 			if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
380 				pri = j + (i << RQB_L2BPW);
381 				rqh = &rq->rq_queues[pri];
382 				TAILQ_FOREACH(td, rqh, td_runq) {
383 					printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
384 					    td, td->td_name, td->td_priority,
385 					    td->td_rqindex, pri);
386 				}
387 			}
388 	}
389 }
390 
391 /*
392  * Print the status of a per-cpu thread queue.  Should be a ddb show cmd.
393  */
394 void
395 tdq_print(int cpu)
396 {
397 	struct tdq *tdq;
398 
399 	tdq = TDQ_CPU(cpu);
400 
401 	printf("tdq %d:\n", TDQ_ID(tdq));
402 	printf("\tlock            %p\n", TDQ_LOCKPTR(tdq));
403 	printf("\tLock name:      %s\n", tdq->tdq_name);
404 	printf("\tload:           %d\n", tdq->tdq_load);
405 	printf("\tswitch cnt:     %d\n", tdq->tdq_switchcnt);
406 	printf("\told switch cnt: %d\n", tdq->tdq_oldswitchcnt);
407 	printf("\ttimeshare idx:  %d\n", tdq->tdq_idx);
408 	printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
409 	printf("\tload transferable: %d\n", tdq->tdq_transferable);
410 	printf("\tlowest priority:   %d\n", tdq->tdq_lowpri);
411 	printf("\trealtime runq:\n");
412 	runq_print(&tdq->tdq_realtime);
413 	printf("\ttimeshare runq:\n");
414 	runq_print(&tdq->tdq_timeshare);
415 	printf("\tidle runq:\n");
416 	runq_print(&tdq->tdq_idle);
417 }
418 
419 static inline int
420 sched_shouldpreempt(int pri, int cpri, int remote)
421 {
422 	/*
423 	 * If the new priority is not better than the current priority there is
424 	 * nothing to do.
425 	 */
426 	if (pri >= cpri)
427 		return (0);
428 	/*
429 	 * Always preempt idle.
430 	 */
431 	if (cpri >= PRI_MIN_IDLE)
432 		return (1);
433 	/*
434 	 * If preemption is disabled don't preempt others.
435 	 */
436 	if (preempt_thresh == 0)
437 		return (0);
438 	/*
439 	 * Preempt if we exceed the threshold.
440 	 */
441 	if (pri <= preempt_thresh)
442 		return (1);
443 	/*
444 	 * If we're interactive or better and there is non-interactive
445 	 * or worse running preempt only remote processors.
446 	 */
447 	if (remote && pri <= PRI_MAX_INTERACT && cpri > PRI_MAX_INTERACT)
448 		return (1);
449 	return (0);
450 }
451 
452 /*
453  * Add a thread to the actual run-queue.  Keeps transferable counts up to
454  * date with what is actually on the run-queue.  Selects the correct
455  * queue position for timeshare threads.
456  */
457 static __inline void
458 tdq_runq_add(struct tdq *tdq, struct thread *td, int flags)
459 {
460 	struct td_sched *ts;
461 	u_char pri;
462 
463 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
464 	THREAD_LOCK_ASSERT(td, MA_OWNED);
465 
466 	pri = td->td_priority;
467 	ts = td->td_sched;
468 	TD_SET_RUNQ(td);
469 	if (THREAD_CAN_MIGRATE(td)) {
470 		tdq->tdq_transferable++;
471 		ts->ts_flags |= TSF_XFERABLE;
472 	}
473 	if (pri < PRI_MIN_BATCH) {
474 		ts->ts_runq = &tdq->tdq_realtime;
475 	} else if (pri <= PRI_MAX_BATCH) {
476 		ts->ts_runq = &tdq->tdq_timeshare;
477 		KASSERT(pri <= PRI_MAX_BATCH && pri >= PRI_MIN_BATCH,
478 			("Invalid priority %d on timeshare runq", pri));
479 		/*
480 		 * This queue contains only priorities between MIN and MAX
481 		 * realtime.  Use the whole queue to represent these values.
482 		 */
483 		if ((flags & (SRQ_BORROWING|SRQ_PREEMPTED)) == 0) {
484 			pri = RQ_NQS * (pri - PRI_MIN_BATCH) / PRI_BATCH_RANGE;
485 			pri = (pri + tdq->tdq_idx) % RQ_NQS;
486 			/*
487 			 * This effectively shortens the queue by one so we
488 			 * can have a one slot difference between idx and
489 			 * ridx while we wait for threads to drain.
490 			 */
491 			if (tdq->tdq_ridx != tdq->tdq_idx &&
492 			    pri == tdq->tdq_ridx)
493 				pri = (unsigned char)(pri - 1) % RQ_NQS;
494 		} else
495 			pri = tdq->tdq_ridx;
496 		runq_add_pri(ts->ts_runq, td, pri, flags);
497 		return;
498 	} else
499 		ts->ts_runq = &tdq->tdq_idle;
500 	runq_add(ts->ts_runq, td, flags);
501 }
502 
503 /*
504  * Remove a thread from a run-queue.  This typically happens when a thread
505  * is selected to run.  Running threads are not on the queue and the
506  * transferable count does not reflect them.
507  */
508 static __inline void
509 tdq_runq_rem(struct tdq *tdq, struct thread *td)
510 {
511 	struct td_sched *ts;
512 
513 	ts = td->td_sched;
514 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
515 	KASSERT(ts->ts_runq != NULL,
516 	    ("tdq_runq_remove: thread %p null ts_runq", td));
517 	if (ts->ts_flags & TSF_XFERABLE) {
518 		tdq->tdq_transferable--;
519 		ts->ts_flags &= ~TSF_XFERABLE;
520 	}
521 	if (ts->ts_runq == &tdq->tdq_timeshare) {
522 		if (tdq->tdq_idx != tdq->tdq_ridx)
523 			runq_remove_idx(ts->ts_runq, td, &tdq->tdq_ridx);
524 		else
525 			runq_remove_idx(ts->ts_runq, td, NULL);
526 	} else
527 		runq_remove(ts->ts_runq, td);
528 }
529 
530 /*
531  * Load is maintained for all threads RUNNING and ON_RUNQ.  Add the load
532  * for this thread to the referenced thread queue.
533  */
534 static void
535 tdq_load_add(struct tdq *tdq, struct thread *td)
536 {
537 
538 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
539 	THREAD_LOCK_ASSERT(td, MA_OWNED);
540 
541 	tdq->tdq_load++;
542 	if ((td->td_flags & TDF_NOLOAD) == 0)
543 		tdq->tdq_sysload++;
544 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
545 	SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load);
546 }
547 
548 /*
549  * Remove the load from a thread that is transitioning to a sleep state or
550  * exiting.
551  */
552 static void
553 tdq_load_rem(struct tdq *tdq, struct thread *td)
554 {
555 
556 	THREAD_LOCK_ASSERT(td, MA_OWNED);
557 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
558 	KASSERT(tdq->tdq_load != 0,
559 	    ("tdq_load_rem: Removing with 0 load on queue %d", TDQ_ID(tdq)));
560 
561 	tdq->tdq_load--;
562 	if ((td->td_flags & TDF_NOLOAD) == 0)
563 		tdq->tdq_sysload--;
564 	KTR_COUNTER0(KTR_SCHED, "load", tdq->tdq_loadname, tdq->tdq_load);
565 	SDT_PROBE2(sched, , , load_change, (int)TDQ_ID(tdq), tdq->tdq_load);
566 }
567 
568 /*
569  * Bound timeshare latency by decreasing slice size as load increases.  We
570  * consider the maximum latency as the sum of the threads waiting to run
571  * aside from curthread and target no more than sched_slice latency but
572  * no less than sched_slice_min runtime.
573  */
574 static inline int
575 tdq_slice(struct tdq *tdq)
576 {
577 	int load;
578 
579 	/*
580 	 * It is safe to use sys_load here because this is called from
581 	 * contexts where timeshare threads are running and so there
582 	 * cannot be higher priority load in the system.
583 	 */
584 	load = tdq->tdq_sysload - 1;
585 	if (load >= SCHED_SLICE_MIN_DIVISOR)
586 		return (sched_slice_min);
587 	if (load <= 1)
588 		return (sched_slice);
589 	return (sched_slice / load);
590 }
591 
592 /*
593  * Set lowpri to its exact value by searching the run-queue and
594  * evaluating curthread.  curthread may be passed as an optimization.
595  */
596 static void
597 tdq_setlowpri(struct tdq *tdq, struct thread *ctd)
598 {
599 	struct thread *td;
600 
601 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
602 	if (ctd == NULL)
603 		ctd = pcpu_find(TDQ_ID(tdq))->pc_curthread;
604 	td = tdq_choose(tdq);
605 	if (td == NULL || td->td_priority > ctd->td_priority)
606 		tdq->tdq_lowpri = ctd->td_priority;
607 	else
608 		tdq->tdq_lowpri = td->td_priority;
609 }
610 
611 #ifdef SMP
612 struct cpu_search {
613 	cpuset_t cs_mask;
614 	u_int	cs_prefer;
615 	int	cs_pri;		/* Min priority for low. */
616 	int	cs_limit;	/* Max load for low, min load for high. */
617 	int	cs_cpu;
618 	int	cs_load;
619 };
620 
621 #define	CPU_SEARCH_LOWEST	0x1
622 #define	CPU_SEARCH_HIGHEST	0x2
623 #define	CPU_SEARCH_BOTH		(CPU_SEARCH_LOWEST|CPU_SEARCH_HIGHEST)
624 
625 #define	CPUSET_FOREACH(cpu, mask)				\
626 	for ((cpu) = 0; (cpu) <= mp_maxid; (cpu)++)		\
627 		if (CPU_ISSET(cpu, &mask))
628 
629 static __inline int cpu_search(const struct cpu_group *cg, struct cpu_search *low,
630     struct cpu_search *high, const int match);
631 int cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low);
632 int cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high);
633 int cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
634     struct cpu_search *high);
635 
636 /*
637  * Search the tree of cpu_groups for the lowest or highest loaded cpu
638  * according to the match argument.  This routine actually compares the
639  * load on all paths through the tree and finds the least loaded cpu on
640  * the least loaded path, which may differ from the least loaded cpu in
641  * the system.  This balances work among caches and busses.
642  *
643  * This inline is instantiated in three forms below using constants for the
644  * match argument.  It is reduced to the minimum set for each case.  It is
645  * also recursive to the depth of the tree.
646  */
647 static __inline int
648 cpu_search(const struct cpu_group *cg, struct cpu_search *low,
649     struct cpu_search *high, const int match)
650 {
651 	struct cpu_search lgroup;
652 	struct cpu_search hgroup;
653 	cpuset_t cpumask;
654 	struct cpu_group *child;
655 	struct tdq *tdq;
656 	int cpu, i, hload, lload, load, total, rnd, *rndptr;
657 
658 	total = 0;
659 	cpumask = cg->cg_mask;
660 	if (match & CPU_SEARCH_LOWEST) {
661 		lload = INT_MAX;
662 		lgroup = *low;
663 	}
664 	if (match & CPU_SEARCH_HIGHEST) {
665 		hload = INT_MIN;
666 		hgroup = *high;
667 	}
668 
669 	/* Iterate through the child CPU groups and then remaining CPUs. */
670 	for (i = cg->cg_children, cpu = mp_maxid; i >= 0; ) {
671 		if (i == 0) {
672 			while (cpu >= 0 && !CPU_ISSET(cpu, &cpumask))
673 				cpu--;
674 			if (cpu < 0)
675 				break;
676 			child = NULL;
677 		} else
678 			child = &cg->cg_child[i - 1];
679 
680 		if (match & CPU_SEARCH_LOWEST)
681 			lgroup.cs_cpu = -1;
682 		if (match & CPU_SEARCH_HIGHEST)
683 			hgroup.cs_cpu = -1;
684 		if (child) {			/* Handle child CPU group. */
685 			CPU_NAND(&cpumask, &child->cg_mask);
686 			switch (match) {
687 			case CPU_SEARCH_LOWEST:
688 				load = cpu_search_lowest(child, &lgroup);
689 				break;
690 			case CPU_SEARCH_HIGHEST:
691 				load = cpu_search_highest(child, &hgroup);
692 				break;
693 			case CPU_SEARCH_BOTH:
694 				load = cpu_search_both(child, &lgroup, &hgroup);
695 				break;
696 			}
697 		} else {			/* Handle child CPU. */
698 			tdq = TDQ_CPU(cpu);
699 			load = tdq->tdq_load * 256;
700 			rndptr = DPCPU_PTR(randomval);
701 			rnd = (*rndptr = *rndptr * 69069 + 5) >> 26;
702 			if (match & CPU_SEARCH_LOWEST) {
703 				if (cpu == low->cs_prefer)
704 					load -= 64;
705 				/* If that CPU is allowed and get data. */
706 				if (tdq->tdq_lowpri > lgroup.cs_pri &&
707 				    tdq->tdq_load <= lgroup.cs_limit &&
708 				    CPU_ISSET(cpu, &lgroup.cs_mask)) {
709 					lgroup.cs_cpu = cpu;
710 					lgroup.cs_load = load - rnd;
711 				}
712 			}
713 			if (match & CPU_SEARCH_HIGHEST)
714 				if (tdq->tdq_load >= hgroup.cs_limit &&
715 				    tdq->tdq_transferable &&
716 				    CPU_ISSET(cpu, &hgroup.cs_mask)) {
717 					hgroup.cs_cpu = cpu;
718 					hgroup.cs_load = load - rnd;
719 				}
720 		}
721 		total += load;
722 
723 		/* We have info about child item. Compare it. */
724 		if (match & CPU_SEARCH_LOWEST) {
725 			if (lgroup.cs_cpu >= 0 &&
726 			    (load < lload ||
727 			     (load == lload && lgroup.cs_load < low->cs_load))) {
728 				lload = load;
729 				low->cs_cpu = lgroup.cs_cpu;
730 				low->cs_load = lgroup.cs_load;
731 			}
732 		}
733 		if (match & CPU_SEARCH_HIGHEST)
734 			if (hgroup.cs_cpu >= 0 &&
735 			    (load > hload ||
736 			     (load == hload && hgroup.cs_load > high->cs_load))) {
737 				hload = load;
738 				high->cs_cpu = hgroup.cs_cpu;
739 				high->cs_load = hgroup.cs_load;
740 			}
741 		if (child) {
742 			i--;
743 			if (i == 0 && CPU_EMPTY(&cpumask))
744 				break;
745 		} else
746 			cpu--;
747 	}
748 	return (total);
749 }
750 
751 /*
752  * cpu_search instantiations must pass constants to maintain the inline
753  * optimization.
754  */
755 int
756 cpu_search_lowest(const struct cpu_group *cg, struct cpu_search *low)
757 {
758 	return cpu_search(cg, low, NULL, CPU_SEARCH_LOWEST);
759 }
760 
761 int
762 cpu_search_highest(const struct cpu_group *cg, struct cpu_search *high)
763 {
764 	return cpu_search(cg, NULL, high, CPU_SEARCH_HIGHEST);
765 }
766 
767 int
768 cpu_search_both(const struct cpu_group *cg, struct cpu_search *low,
769     struct cpu_search *high)
770 {
771 	return cpu_search(cg, low, high, CPU_SEARCH_BOTH);
772 }
773 
774 /*
775  * Find the cpu with the least load via the least loaded path that has a
776  * lowpri greater than pri  pri.  A pri of -1 indicates any priority is
777  * acceptable.
778  */
779 static inline int
780 sched_lowest(const struct cpu_group *cg, cpuset_t mask, int pri, int maxload,
781     int prefer)
782 {
783 	struct cpu_search low;
784 
785 	low.cs_cpu = -1;
786 	low.cs_prefer = prefer;
787 	low.cs_mask = mask;
788 	low.cs_pri = pri;
789 	low.cs_limit = maxload;
790 	cpu_search_lowest(cg, &low);
791 	return low.cs_cpu;
792 }
793 
794 /*
795  * Find the cpu with the highest load via the highest loaded path.
796  */
797 static inline int
798 sched_highest(const struct cpu_group *cg, cpuset_t mask, int minload)
799 {
800 	struct cpu_search high;
801 
802 	high.cs_cpu = -1;
803 	high.cs_mask = mask;
804 	high.cs_limit = minload;
805 	cpu_search_highest(cg, &high);
806 	return high.cs_cpu;
807 }
808 
809 /*
810  * Simultaneously find the highest and lowest loaded cpu reachable via
811  * cg.
812  */
813 static inline void
814 sched_both(const struct cpu_group *cg, cpuset_t mask, int *lowcpu, int *highcpu)
815 {
816 	struct cpu_search high;
817 	struct cpu_search low;
818 
819 	low.cs_cpu = -1;
820 	low.cs_prefer = -1;
821 	low.cs_pri = -1;
822 	low.cs_limit = INT_MAX;
823 	low.cs_mask = mask;
824 	high.cs_cpu = -1;
825 	high.cs_limit = -1;
826 	high.cs_mask = mask;
827 	cpu_search_both(cg, &low, &high);
828 	*lowcpu = low.cs_cpu;
829 	*highcpu = high.cs_cpu;
830 	return;
831 }
832 
833 static void
834 sched_balance_group(struct cpu_group *cg)
835 {
836 	cpuset_t hmask, lmask;
837 	int high, low, anylow;
838 
839 	CPU_FILL(&hmask);
840 	for (;;) {
841 		high = sched_highest(cg, hmask, 1);
842 		/* Stop if there is no more CPU with transferrable threads. */
843 		if (high == -1)
844 			break;
845 		CPU_CLR(high, &hmask);
846 		CPU_COPY(&hmask, &lmask);
847 		/* Stop if there is no more CPU left for low. */
848 		if (CPU_EMPTY(&lmask))
849 			break;
850 		anylow = 1;
851 nextlow:
852 		low = sched_lowest(cg, lmask, -1,
853 		    TDQ_CPU(high)->tdq_load - 1, high);
854 		/* Stop if we looked well and found no less loaded CPU. */
855 		if (anylow && low == -1)
856 			break;
857 		/* Go to next high if we found no less loaded CPU. */
858 		if (low == -1)
859 			continue;
860 		/* Transfer thread from high to low. */
861 		if (sched_balance_pair(TDQ_CPU(high), TDQ_CPU(low))) {
862 			/* CPU that got thread can no longer be a donor. */
863 			CPU_CLR(low, &hmask);
864 		} else {
865 			/*
866 			 * If failed, then there is no threads on high
867 			 * that can run on this low. Drop low from low
868 			 * mask and look for different one.
869 			 */
870 			CPU_CLR(low, &lmask);
871 			anylow = 0;
872 			goto nextlow;
873 		}
874 	}
875 }
876 
877 static void
878 sched_balance(void)
879 {
880 	struct tdq *tdq;
881 
882 	/*
883 	 * Select a random time between .5 * balance_interval and
884 	 * 1.5 * balance_interval.
885 	 */
886 	balance_ticks = max(balance_interval / 2, 1);
887 	balance_ticks += random() % balance_interval;
888 	if (smp_started == 0 || rebalance == 0)
889 		return;
890 	tdq = TDQ_SELF();
891 	TDQ_UNLOCK(tdq);
892 	sched_balance_group(cpu_top);
893 	TDQ_LOCK(tdq);
894 }
895 
896 /*
897  * Lock two thread queues using their address to maintain lock order.
898  */
899 static void
900 tdq_lock_pair(struct tdq *one, struct tdq *two)
901 {
902 	if (one < two) {
903 		TDQ_LOCK(one);
904 		TDQ_LOCK_FLAGS(two, MTX_DUPOK);
905 	} else {
906 		TDQ_LOCK(two);
907 		TDQ_LOCK_FLAGS(one, MTX_DUPOK);
908 	}
909 }
910 
911 /*
912  * Unlock two thread queues.  Order is not important here.
913  */
914 static void
915 tdq_unlock_pair(struct tdq *one, struct tdq *two)
916 {
917 	TDQ_UNLOCK(one);
918 	TDQ_UNLOCK(two);
919 }
920 
921 /*
922  * Transfer load between two imbalanced thread queues.
923  */
924 static int
925 sched_balance_pair(struct tdq *high, struct tdq *low)
926 {
927 	int moved;
928 	int cpu;
929 
930 	tdq_lock_pair(high, low);
931 	moved = 0;
932 	/*
933 	 * Determine what the imbalance is and then adjust that to how many
934 	 * threads we actually have to give up (transferable).
935 	 */
936 	if (high->tdq_transferable != 0 && high->tdq_load > low->tdq_load &&
937 	    (moved = tdq_move(high, low)) > 0) {
938 		/*
939 		 * In case the target isn't the current cpu IPI it to force a
940 		 * reschedule with the new workload.
941 		 */
942 		cpu = TDQ_ID(low);
943 		if (cpu != PCPU_GET(cpuid))
944 			ipi_cpu(cpu, IPI_PREEMPT);
945 	}
946 	tdq_unlock_pair(high, low);
947 	return (moved);
948 }
949 
950 /*
951  * Move a thread from one thread queue to another.
952  */
953 static int
954 tdq_move(struct tdq *from, struct tdq *to)
955 {
956 	struct td_sched *ts;
957 	struct thread *td;
958 	struct tdq *tdq;
959 	int cpu;
960 
961 	TDQ_LOCK_ASSERT(from, MA_OWNED);
962 	TDQ_LOCK_ASSERT(to, MA_OWNED);
963 
964 	tdq = from;
965 	cpu = TDQ_ID(to);
966 	td = tdq_steal(tdq, cpu);
967 	if (td == NULL)
968 		return (0);
969 	ts = td->td_sched;
970 	/*
971 	 * Although the run queue is locked the thread may be blocked.  Lock
972 	 * it to clear this and acquire the run-queue lock.
973 	 */
974 	thread_lock(td);
975 	/* Drop recursive lock on from acquired via thread_lock(). */
976 	TDQ_UNLOCK(from);
977 	sched_rem(td);
978 	ts->ts_cpu = cpu;
979 	td->td_lock = TDQ_LOCKPTR(to);
980 	tdq_add(to, td, SRQ_YIELDING);
981 	return (1);
982 }
983 
984 /*
985  * This tdq has idled.  Try to steal a thread from another cpu and switch
986  * to it.
987  */
988 static int
989 tdq_idled(struct tdq *tdq)
990 {
991 	struct cpu_group *cg;
992 	struct tdq *steal;
993 	cpuset_t mask;
994 	int thresh;
995 	int cpu;
996 
997 	if (smp_started == 0 || steal_idle == 0)
998 		return (1);
999 	CPU_FILL(&mask);
1000 	CPU_CLR(PCPU_GET(cpuid), &mask);
1001 	/* We don't want to be preempted while we're iterating. */
1002 	spinlock_enter();
1003 	for (cg = tdq->tdq_cg; cg != NULL; ) {
1004 		if ((cg->cg_flags & CG_FLAG_THREAD) == 0)
1005 			thresh = steal_thresh;
1006 		else
1007 			thresh = 1;
1008 		cpu = sched_highest(cg, mask, thresh);
1009 		if (cpu == -1) {
1010 			cg = cg->cg_parent;
1011 			continue;
1012 		}
1013 		steal = TDQ_CPU(cpu);
1014 		CPU_CLR(cpu, &mask);
1015 		tdq_lock_pair(tdq, steal);
1016 		if (steal->tdq_load < thresh || steal->tdq_transferable == 0) {
1017 			tdq_unlock_pair(tdq, steal);
1018 			continue;
1019 		}
1020 		/*
1021 		 * If a thread was added while interrupts were disabled don't
1022 		 * steal one here.  If we fail to acquire one due to affinity
1023 		 * restrictions loop again with this cpu removed from the
1024 		 * set.
1025 		 */
1026 		if (tdq->tdq_load == 0 && tdq_move(steal, tdq) == 0) {
1027 			tdq_unlock_pair(tdq, steal);
1028 			continue;
1029 		}
1030 		spinlock_exit();
1031 		TDQ_UNLOCK(steal);
1032 		mi_switch(SW_VOL | SWT_IDLE, NULL);
1033 		thread_unlock(curthread);
1034 
1035 		return (0);
1036 	}
1037 	spinlock_exit();
1038 	return (1);
1039 }
1040 
1041 /*
1042  * Notify a remote cpu of new work.  Sends an IPI if criteria are met.
1043  */
1044 static void
1045 tdq_notify(struct tdq *tdq, struct thread *td)
1046 {
1047 	struct thread *ctd;
1048 	int pri;
1049 	int cpu;
1050 
1051 	if (tdq->tdq_ipipending)
1052 		return;
1053 	cpu = td->td_sched->ts_cpu;
1054 	pri = td->td_priority;
1055 	ctd = pcpu_find(cpu)->pc_curthread;
1056 	if (!sched_shouldpreempt(pri, ctd->td_priority, 1))
1057 		return;
1058 	if (TD_IS_IDLETHREAD(ctd)) {
1059 		/*
1060 		 * If the MD code has an idle wakeup routine try that before
1061 		 * falling back to IPI.
1062 		 */
1063 		if (!tdq->tdq_cpu_idle || cpu_idle_wakeup(cpu))
1064 			return;
1065 	}
1066 	tdq->tdq_ipipending = 1;
1067 	ipi_cpu(cpu, IPI_PREEMPT);
1068 }
1069 
1070 /*
1071  * Steals load from a timeshare queue.  Honors the rotating queue head
1072  * index.
1073  */
1074 static struct thread *
1075 runq_steal_from(struct runq *rq, int cpu, u_char start)
1076 {
1077 	struct rqbits *rqb;
1078 	struct rqhead *rqh;
1079 	struct thread *td, *first;
1080 	int bit;
1081 	int pri;
1082 	int i;
1083 
1084 	rqb = &rq->rq_status;
1085 	bit = start & (RQB_BPW -1);
1086 	pri = 0;
1087 	first = NULL;
1088 again:
1089 	for (i = RQB_WORD(start); i < RQB_LEN; bit = 0, i++) {
1090 		if (rqb->rqb_bits[i] == 0)
1091 			continue;
1092 		if (bit != 0) {
1093 			for (pri = bit; pri < RQB_BPW; pri++)
1094 				if (rqb->rqb_bits[i] & (1ul << pri))
1095 					break;
1096 			if (pri >= RQB_BPW)
1097 				continue;
1098 		} else
1099 			pri = RQB_FFS(rqb->rqb_bits[i]);
1100 		pri += (i << RQB_L2BPW);
1101 		rqh = &rq->rq_queues[pri];
1102 		TAILQ_FOREACH(td, rqh, td_runq) {
1103 			if (first && THREAD_CAN_MIGRATE(td) &&
1104 			    THREAD_CAN_SCHED(td, cpu))
1105 				return (td);
1106 			first = td;
1107 		}
1108 	}
1109 	if (start != 0) {
1110 		start = 0;
1111 		goto again;
1112 	}
1113 
1114 	if (first && THREAD_CAN_MIGRATE(first) &&
1115 	    THREAD_CAN_SCHED(first, cpu))
1116 		return (first);
1117 	return (NULL);
1118 }
1119 
1120 /*
1121  * Steals load from a standard linear queue.
1122  */
1123 static struct thread *
1124 runq_steal(struct runq *rq, int cpu)
1125 {
1126 	struct rqhead *rqh;
1127 	struct rqbits *rqb;
1128 	struct thread *td;
1129 	int word;
1130 	int bit;
1131 
1132 	rqb = &rq->rq_status;
1133 	for (word = 0; word < RQB_LEN; word++) {
1134 		if (rqb->rqb_bits[word] == 0)
1135 			continue;
1136 		for (bit = 0; bit < RQB_BPW; bit++) {
1137 			if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
1138 				continue;
1139 			rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
1140 			TAILQ_FOREACH(td, rqh, td_runq)
1141 				if (THREAD_CAN_MIGRATE(td) &&
1142 				    THREAD_CAN_SCHED(td, cpu))
1143 					return (td);
1144 		}
1145 	}
1146 	return (NULL);
1147 }
1148 
1149 /*
1150  * Attempt to steal a thread in priority order from a thread queue.
1151  */
1152 static struct thread *
1153 tdq_steal(struct tdq *tdq, int cpu)
1154 {
1155 	struct thread *td;
1156 
1157 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1158 	if ((td = runq_steal(&tdq->tdq_realtime, cpu)) != NULL)
1159 		return (td);
1160 	if ((td = runq_steal_from(&tdq->tdq_timeshare,
1161 	    cpu, tdq->tdq_ridx)) != NULL)
1162 		return (td);
1163 	return (runq_steal(&tdq->tdq_idle, cpu));
1164 }
1165 
1166 /*
1167  * Sets the thread lock and ts_cpu to match the requested cpu.  Unlocks the
1168  * current lock and returns with the assigned queue locked.
1169  */
1170 static inline struct tdq *
1171 sched_setcpu(struct thread *td, int cpu, int flags)
1172 {
1173 
1174 	struct tdq *tdq;
1175 
1176 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1177 	tdq = TDQ_CPU(cpu);
1178 	td->td_sched->ts_cpu = cpu;
1179 	/*
1180 	 * If the lock matches just return the queue.
1181 	 */
1182 	if (td->td_lock == TDQ_LOCKPTR(tdq))
1183 		return (tdq);
1184 #ifdef notyet
1185 	/*
1186 	 * If the thread isn't running its lockptr is a
1187 	 * turnstile or a sleepqueue.  We can just lock_set without
1188 	 * blocking.
1189 	 */
1190 	if (TD_CAN_RUN(td)) {
1191 		TDQ_LOCK(tdq);
1192 		thread_lock_set(td, TDQ_LOCKPTR(tdq));
1193 		return (tdq);
1194 	}
1195 #endif
1196 	/*
1197 	 * The hard case, migration, we need to block the thread first to
1198 	 * prevent order reversals with other cpus locks.
1199 	 */
1200 	spinlock_enter();
1201 	thread_lock_block(td);
1202 	TDQ_LOCK(tdq);
1203 	thread_lock_unblock(td, TDQ_LOCKPTR(tdq));
1204 	spinlock_exit();
1205 	return (tdq);
1206 }
1207 
1208 SCHED_STAT_DEFINE(pickcpu_intrbind, "Soft interrupt binding");
1209 SCHED_STAT_DEFINE(pickcpu_idle_affinity, "Picked idle cpu based on affinity");
1210 SCHED_STAT_DEFINE(pickcpu_affinity, "Picked cpu based on affinity");
1211 SCHED_STAT_DEFINE(pickcpu_lowest, "Selected lowest load");
1212 SCHED_STAT_DEFINE(pickcpu_local, "Migrated to current cpu");
1213 SCHED_STAT_DEFINE(pickcpu_migration, "Selection may have caused migration");
1214 
1215 static int
1216 sched_pickcpu(struct thread *td, int flags)
1217 {
1218 	struct cpu_group *cg, *ccg;
1219 	struct td_sched *ts;
1220 	struct tdq *tdq;
1221 	cpuset_t mask;
1222 	int cpu, pri, self;
1223 
1224 	self = PCPU_GET(cpuid);
1225 	ts = td->td_sched;
1226 	if (smp_started == 0)
1227 		return (self);
1228 	/*
1229 	 * Don't migrate a running thread from sched_switch().
1230 	 */
1231 	if ((flags & SRQ_OURSELF) || !THREAD_CAN_MIGRATE(td))
1232 		return (ts->ts_cpu);
1233 	/*
1234 	 * Prefer to run interrupt threads on the processors that generate
1235 	 * the interrupt.
1236 	 */
1237 	pri = td->td_priority;
1238 	if (td->td_priority <= PRI_MAX_ITHD && THREAD_CAN_SCHED(td, self) &&
1239 	    curthread->td_intr_nesting_level && ts->ts_cpu != self) {
1240 		SCHED_STAT_INC(pickcpu_intrbind);
1241 		ts->ts_cpu = self;
1242 		if (TDQ_CPU(self)->tdq_lowpri > pri) {
1243 			SCHED_STAT_INC(pickcpu_affinity);
1244 			return (ts->ts_cpu);
1245 		}
1246 	}
1247 	/*
1248 	 * If the thread can run on the last cpu and the affinity has not
1249 	 * expired or it is idle run it there.
1250 	 */
1251 	tdq = TDQ_CPU(ts->ts_cpu);
1252 	cg = tdq->tdq_cg;
1253 	if (THREAD_CAN_SCHED(td, ts->ts_cpu) &&
1254 	    tdq->tdq_lowpri >= PRI_MIN_IDLE &&
1255 	    SCHED_AFFINITY(ts, CG_SHARE_L2)) {
1256 		if (cg->cg_flags & CG_FLAG_THREAD) {
1257 			CPUSET_FOREACH(cpu, cg->cg_mask) {
1258 				if (TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE)
1259 					break;
1260 			}
1261 		} else
1262 			cpu = INT_MAX;
1263 		if (cpu > mp_maxid) {
1264 			SCHED_STAT_INC(pickcpu_idle_affinity);
1265 			return (ts->ts_cpu);
1266 		}
1267 	}
1268 	/*
1269 	 * Search for the last level cache CPU group in the tree.
1270 	 * Skip caches with expired affinity time and SMT groups.
1271 	 * Affinity to higher level caches will be handled less aggressively.
1272 	 */
1273 	for (ccg = NULL; cg != NULL; cg = cg->cg_parent) {
1274 		if (cg->cg_flags & CG_FLAG_THREAD)
1275 			continue;
1276 		if (!SCHED_AFFINITY(ts, cg->cg_level))
1277 			continue;
1278 		ccg = cg;
1279 	}
1280 	if (ccg != NULL)
1281 		cg = ccg;
1282 	cpu = -1;
1283 	/* Search the group for the less loaded idle CPU we can run now. */
1284 	mask = td->td_cpuset->cs_mask;
1285 	if (cg != NULL && cg != cpu_top &&
1286 	    CPU_CMP(&cg->cg_mask, &cpu_top->cg_mask) != 0)
1287 		cpu = sched_lowest(cg, mask, max(pri, PRI_MAX_TIMESHARE),
1288 		    INT_MAX, ts->ts_cpu);
1289 	/* Search globally for the less loaded CPU we can run now. */
1290 	if (cpu == -1)
1291 		cpu = sched_lowest(cpu_top, mask, pri, INT_MAX, ts->ts_cpu);
1292 	/* Search globally for the less loaded CPU. */
1293 	if (cpu == -1)
1294 		cpu = sched_lowest(cpu_top, mask, -1, INT_MAX, ts->ts_cpu);
1295 	KASSERT(cpu != -1, ("sched_pickcpu: Failed to find a cpu."));
1296 	/*
1297 	 * Compare the lowest loaded cpu to current cpu.
1298 	 */
1299 	if (THREAD_CAN_SCHED(td, self) && TDQ_CPU(self)->tdq_lowpri > pri &&
1300 	    TDQ_CPU(cpu)->tdq_lowpri < PRI_MIN_IDLE &&
1301 	    TDQ_CPU(self)->tdq_load <= TDQ_CPU(cpu)->tdq_load + 1) {
1302 		SCHED_STAT_INC(pickcpu_local);
1303 		cpu = self;
1304 	} else
1305 		SCHED_STAT_INC(pickcpu_lowest);
1306 	if (cpu != ts->ts_cpu)
1307 		SCHED_STAT_INC(pickcpu_migration);
1308 	return (cpu);
1309 }
1310 #endif
1311 
1312 /*
1313  * Pick the highest priority task we have and return it.
1314  */
1315 static struct thread *
1316 tdq_choose(struct tdq *tdq)
1317 {
1318 	struct thread *td;
1319 
1320 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
1321 	td = runq_choose(&tdq->tdq_realtime);
1322 	if (td != NULL)
1323 		return (td);
1324 	td = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
1325 	if (td != NULL) {
1326 		KASSERT(td->td_priority >= PRI_MIN_BATCH,
1327 		    ("tdq_choose: Invalid priority on timeshare queue %d",
1328 		    td->td_priority));
1329 		return (td);
1330 	}
1331 	td = runq_choose(&tdq->tdq_idle);
1332 	if (td != NULL) {
1333 		KASSERT(td->td_priority >= PRI_MIN_IDLE,
1334 		    ("tdq_choose: Invalid priority on idle queue %d",
1335 		    td->td_priority));
1336 		return (td);
1337 	}
1338 
1339 	return (NULL);
1340 }
1341 
1342 /*
1343  * Initialize a thread queue.
1344  */
1345 static void
1346 tdq_setup(struct tdq *tdq)
1347 {
1348 
1349 	if (bootverbose)
1350 		printf("ULE: setup cpu %d\n", TDQ_ID(tdq));
1351 	runq_init(&tdq->tdq_realtime);
1352 	runq_init(&tdq->tdq_timeshare);
1353 	runq_init(&tdq->tdq_idle);
1354 	snprintf(tdq->tdq_name, sizeof(tdq->tdq_name),
1355 	    "sched lock %d", (int)TDQ_ID(tdq));
1356 	mtx_init(&tdq->tdq_lock, tdq->tdq_name, "sched lock",
1357 	    MTX_SPIN | MTX_RECURSE);
1358 #ifdef KTR
1359 	snprintf(tdq->tdq_loadname, sizeof(tdq->tdq_loadname),
1360 	    "CPU %d load", (int)TDQ_ID(tdq));
1361 #endif
1362 }
1363 
1364 #ifdef SMP
1365 static void
1366 sched_setup_smp(void)
1367 {
1368 	struct tdq *tdq;
1369 	int i;
1370 
1371 	cpu_top = smp_topo();
1372 	CPU_FOREACH(i) {
1373 		tdq = TDQ_CPU(i);
1374 		tdq_setup(tdq);
1375 		tdq->tdq_cg = smp_topo_find(cpu_top, i);
1376 		if (tdq->tdq_cg == NULL)
1377 			panic("Can't find cpu group for %d\n", i);
1378 	}
1379 	balance_tdq = TDQ_SELF();
1380 	sched_balance();
1381 }
1382 #endif
1383 
1384 /*
1385  * Setup the thread queues and initialize the topology based on MD
1386  * information.
1387  */
1388 static void
1389 sched_setup(void *dummy)
1390 {
1391 	struct tdq *tdq;
1392 
1393 	tdq = TDQ_SELF();
1394 #ifdef SMP
1395 	sched_setup_smp();
1396 #else
1397 	tdq_setup(tdq);
1398 #endif
1399 
1400 	/* Add thread0's load since it's running. */
1401 	TDQ_LOCK(tdq);
1402 	thread0.td_lock = TDQ_LOCKPTR(TDQ_SELF());
1403 	tdq_load_add(tdq, &thread0);
1404 	tdq->tdq_lowpri = thread0.td_priority;
1405 	TDQ_UNLOCK(tdq);
1406 }
1407 
1408 /*
1409  * This routine determines time constants after stathz and hz are setup.
1410  */
1411 /* ARGSUSED */
1412 static void
1413 sched_initticks(void *dummy)
1414 {
1415 	int incr;
1416 
1417 	realstathz = stathz ? stathz : hz;
1418 	sched_slice = realstathz / SCHED_SLICE_DEFAULT_DIVISOR;
1419 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
1420 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
1421 	    realstathz);
1422 
1423 	/*
1424 	 * tickincr is shifted out by 10 to avoid rounding errors due to
1425 	 * hz not being evenly divisible by stathz on all platforms.
1426 	 */
1427 	incr = (hz << SCHED_TICK_SHIFT) / realstathz;
1428 	/*
1429 	 * This does not work for values of stathz that are more than
1430 	 * 1 << SCHED_TICK_SHIFT * hz.  In practice this does not happen.
1431 	 */
1432 	if (incr == 0)
1433 		incr = 1;
1434 	tickincr = incr;
1435 #ifdef SMP
1436 	/*
1437 	 * Set the default balance interval now that we know
1438 	 * what realstathz is.
1439 	 */
1440 	balance_interval = realstathz;
1441 	affinity = SCHED_AFFINITY_DEFAULT;
1442 #endif
1443 	if (sched_idlespinthresh < 0)
1444 		sched_idlespinthresh = 2 * max(10000, 6 * hz) / realstathz;
1445 }
1446 
1447 
1448 /*
1449  * This is the core of the interactivity algorithm.  Determines a score based
1450  * on past behavior.  It is the ratio of sleep time to run time scaled to
1451  * a [0, 100] integer.  This is the voluntary sleep time of a process, which
1452  * differs from the cpu usage because it does not account for time spent
1453  * waiting on a run-queue.  Would be prettier if we had floating point.
1454  */
1455 static int
1456 sched_interact_score(struct thread *td)
1457 {
1458 	struct td_sched *ts;
1459 	int div;
1460 
1461 	ts = td->td_sched;
1462 	/*
1463 	 * The score is only needed if this is likely to be an interactive
1464 	 * task.  Don't go through the expense of computing it if there's
1465 	 * no chance.
1466 	 */
1467 	if (sched_interact <= SCHED_INTERACT_HALF &&
1468 		ts->ts_runtime >= ts->ts_slptime)
1469 			return (SCHED_INTERACT_HALF);
1470 
1471 	if (ts->ts_runtime > ts->ts_slptime) {
1472 		div = max(1, ts->ts_runtime / SCHED_INTERACT_HALF);
1473 		return (SCHED_INTERACT_HALF +
1474 		    (SCHED_INTERACT_HALF - (ts->ts_slptime / div)));
1475 	}
1476 	if (ts->ts_slptime > ts->ts_runtime) {
1477 		div = max(1, ts->ts_slptime / SCHED_INTERACT_HALF);
1478 		return (ts->ts_runtime / div);
1479 	}
1480 	/* runtime == slptime */
1481 	if (ts->ts_runtime)
1482 		return (SCHED_INTERACT_HALF);
1483 
1484 	/*
1485 	 * This can happen if slptime and runtime are 0.
1486 	 */
1487 	return (0);
1488 
1489 }
1490 
1491 /*
1492  * Scale the scheduling priority according to the "interactivity" of this
1493  * process.
1494  */
1495 static void
1496 sched_priority(struct thread *td)
1497 {
1498 	int score;
1499 	int pri;
1500 
1501 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
1502 		return;
1503 	/*
1504 	 * If the score is interactive we place the thread in the realtime
1505 	 * queue with a priority that is less than kernel and interrupt
1506 	 * priorities.  These threads are not subject to nice restrictions.
1507 	 *
1508 	 * Scores greater than this are placed on the normal timeshare queue
1509 	 * where the priority is partially decided by the most recent cpu
1510 	 * utilization and the rest is decided by nice value.
1511 	 *
1512 	 * The nice value of the process has a linear effect on the calculated
1513 	 * score.  Negative nice values make it easier for a thread to be
1514 	 * considered interactive.
1515 	 */
1516 	score = imax(0, sched_interact_score(td) + td->td_proc->p_nice);
1517 	if (score < sched_interact) {
1518 		pri = PRI_MIN_INTERACT;
1519 		pri += ((PRI_MAX_INTERACT - PRI_MIN_INTERACT + 1) /
1520 		    sched_interact) * score;
1521 		KASSERT(pri >= PRI_MIN_INTERACT && pri <= PRI_MAX_INTERACT,
1522 		    ("sched_priority: invalid interactive priority %d score %d",
1523 		    pri, score));
1524 	} else {
1525 		pri = SCHED_PRI_MIN;
1526 		if (td->td_sched->ts_ticks)
1527 			pri += min(SCHED_PRI_TICKS(td->td_sched),
1528 			    SCHED_PRI_RANGE);
1529 		pri += SCHED_PRI_NICE(td->td_proc->p_nice);
1530 		KASSERT(pri >= PRI_MIN_BATCH && pri <= PRI_MAX_BATCH,
1531 		    ("sched_priority: invalid priority %d: nice %d, "
1532 		    "ticks %d ftick %d ltick %d tick pri %d",
1533 		    pri, td->td_proc->p_nice, td->td_sched->ts_ticks,
1534 		    td->td_sched->ts_ftick, td->td_sched->ts_ltick,
1535 		    SCHED_PRI_TICKS(td->td_sched)));
1536 	}
1537 	sched_user_prio(td, pri);
1538 
1539 	return;
1540 }
1541 
1542 /*
1543  * This routine enforces a maximum limit on the amount of scheduling history
1544  * kept.  It is called after either the slptime or runtime is adjusted.  This
1545  * function is ugly due to integer math.
1546  */
1547 static void
1548 sched_interact_update(struct thread *td)
1549 {
1550 	struct td_sched *ts;
1551 	u_int sum;
1552 
1553 	ts = td->td_sched;
1554 	sum = ts->ts_runtime + ts->ts_slptime;
1555 	if (sum < SCHED_SLP_RUN_MAX)
1556 		return;
1557 	/*
1558 	 * This only happens from two places:
1559 	 * 1) We have added an unusual amount of run time from fork_exit.
1560 	 * 2) We have added an unusual amount of sleep time from sched_sleep().
1561 	 */
1562 	if (sum > SCHED_SLP_RUN_MAX * 2) {
1563 		if (ts->ts_runtime > ts->ts_slptime) {
1564 			ts->ts_runtime = SCHED_SLP_RUN_MAX;
1565 			ts->ts_slptime = 1;
1566 		} else {
1567 			ts->ts_slptime = SCHED_SLP_RUN_MAX;
1568 			ts->ts_runtime = 1;
1569 		}
1570 		return;
1571 	}
1572 	/*
1573 	 * If we have exceeded by more than 1/5th then the algorithm below
1574 	 * will not bring us back into range.  Dividing by two here forces
1575 	 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1576 	 */
1577 	if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1578 		ts->ts_runtime /= 2;
1579 		ts->ts_slptime /= 2;
1580 		return;
1581 	}
1582 	ts->ts_runtime = (ts->ts_runtime / 5) * 4;
1583 	ts->ts_slptime = (ts->ts_slptime / 5) * 4;
1584 }
1585 
1586 /*
1587  * Scale back the interactivity history when a child thread is created.  The
1588  * history is inherited from the parent but the thread may behave totally
1589  * differently.  For example, a shell spawning a compiler process.  We want
1590  * to learn that the compiler is behaving badly very quickly.
1591  */
1592 static void
1593 sched_interact_fork(struct thread *td)
1594 {
1595 	int ratio;
1596 	int sum;
1597 
1598 	sum = td->td_sched->ts_runtime + td->td_sched->ts_slptime;
1599 	if (sum > SCHED_SLP_RUN_FORK) {
1600 		ratio = sum / SCHED_SLP_RUN_FORK;
1601 		td->td_sched->ts_runtime /= ratio;
1602 		td->td_sched->ts_slptime /= ratio;
1603 	}
1604 }
1605 
1606 /*
1607  * Called from proc0_init() to setup the scheduler fields.
1608  */
1609 void
1610 schedinit(void)
1611 {
1612 
1613 	/*
1614 	 * Set up the scheduler specific parts of proc0.
1615 	 */
1616 	proc0.p_sched = NULL; /* XXX */
1617 	thread0.td_sched = &td_sched0;
1618 	td_sched0.ts_ltick = ticks;
1619 	td_sched0.ts_ftick = ticks;
1620 	td_sched0.ts_slice = 0;
1621 }
1622 
1623 /*
1624  * This is only somewhat accurate since given many processes of the same
1625  * priority they will switch when their slices run out, which will be
1626  * at most sched_slice stathz ticks.
1627  */
1628 int
1629 sched_rr_interval(void)
1630 {
1631 
1632 	/* Convert sched_slice from stathz to hz. */
1633 	return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz));
1634 }
1635 
1636 /*
1637  * Update the percent cpu tracking information when it is requested or
1638  * the total history exceeds the maximum.  We keep a sliding history of
1639  * tick counts that slowly decays.  This is less precise than the 4BSD
1640  * mechanism since it happens with less regular and frequent events.
1641  */
1642 static void
1643 sched_pctcpu_update(struct td_sched *ts, int run)
1644 {
1645 	int t = ticks;
1646 
1647 	if (t - ts->ts_ltick >= SCHED_TICK_TARG) {
1648 		ts->ts_ticks = 0;
1649 		ts->ts_ftick = t - SCHED_TICK_TARG;
1650 	} else if (t - ts->ts_ftick >= SCHED_TICK_MAX) {
1651 		ts->ts_ticks = (ts->ts_ticks / (ts->ts_ltick - ts->ts_ftick)) *
1652 		    (ts->ts_ltick - (t - SCHED_TICK_TARG));
1653 		ts->ts_ftick = t - SCHED_TICK_TARG;
1654 	}
1655 	if (run)
1656 		ts->ts_ticks += (t - ts->ts_ltick) << SCHED_TICK_SHIFT;
1657 	ts->ts_ltick = t;
1658 }
1659 
1660 /*
1661  * Adjust the priority of a thread.  Move it to the appropriate run-queue
1662  * if necessary.  This is the back-end for several priority related
1663  * functions.
1664  */
1665 static void
1666 sched_thread_priority(struct thread *td, u_char prio)
1667 {
1668 	struct td_sched *ts;
1669 	struct tdq *tdq;
1670 	int oldpri;
1671 
1672 	KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "prio",
1673 	    "prio:%d", td->td_priority, "new prio:%d", prio,
1674 	    KTR_ATTR_LINKED, sched_tdname(curthread));
1675 	SDT_PROBE3(sched, , , change_pri, td, td->td_proc, prio);
1676 	if (td != curthread && prio < td->td_priority) {
1677 		KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread),
1678 		    "lend prio", "prio:%d", td->td_priority, "new prio:%d",
1679 		    prio, KTR_ATTR_LINKED, sched_tdname(td));
1680 		SDT_PROBE4(sched, , , lend_pri, td, td->td_proc, prio,
1681 		    curthread);
1682 	}
1683 	ts = td->td_sched;
1684 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1685 	if (td->td_priority == prio)
1686 		return;
1687 	/*
1688 	 * If the priority has been elevated due to priority
1689 	 * propagation, we may have to move ourselves to a new
1690 	 * queue.  This could be optimized to not re-add in some
1691 	 * cases.
1692 	 */
1693 	if (TD_ON_RUNQ(td) && prio < td->td_priority) {
1694 		sched_rem(td);
1695 		td->td_priority = prio;
1696 		sched_add(td, SRQ_BORROWING);
1697 		return;
1698 	}
1699 	/*
1700 	 * If the thread is currently running we may have to adjust the lowpri
1701 	 * information so other cpus are aware of our current priority.
1702 	 */
1703 	if (TD_IS_RUNNING(td)) {
1704 		tdq = TDQ_CPU(ts->ts_cpu);
1705 		oldpri = td->td_priority;
1706 		td->td_priority = prio;
1707 		if (prio < tdq->tdq_lowpri)
1708 			tdq->tdq_lowpri = prio;
1709 		else if (tdq->tdq_lowpri == oldpri)
1710 			tdq_setlowpri(tdq, td);
1711 		return;
1712 	}
1713 	td->td_priority = prio;
1714 }
1715 
1716 /*
1717  * Update a thread's priority when it is lent another thread's
1718  * priority.
1719  */
1720 void
1721 sched_lend_prio(struct thread *td, u_char prio)
1722 {
1723 
1724 	td->td_flags |= TDF_BORROWING;
1725 	sched_thread_priority(td, prio);
1726 }
1727 
1728 /*
1729  * Restore a thread's priority when priority propagation is
1730  * over.  The prio argument is the minimum priority the thread
1731  * needs to have to satisfy other possible priority lending
1732  * requests.  If the thread's regular priority is less
1733  * important than prio, the thread will keep a priority boost
1734  * of prio.
1735  */
1736 void
1737 sched_unlend_prio(struct thread *td, u_char prio)
1738 {
1739 	u_char base_pri;
1740 
1741 	if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
1742 	    td->td_base_pri <= PRI_MAX_TIMESHARE)
1743 		base_pri = td->td_user_pri;
1744 	else
1745 		base_pri = td->td_base_pri;
1746 	if (prio >= base_pri) {
1747 		td->td_flags &= ~TDF_BORROWING;
1748 		sched_thread_priority(td, base_pri);
1749 	} else
1750 		sched_lend_prio(td, prio);
1751 }
1752 
1753 /*
1754  * Standard entry for setting the priority to an absolute value.
1755  */
1756 void
1757 sched_prio(struct thread *td, u_char prio)
1758 {
1759 	u_char oldprio;
1760 
1761 	/* First, update the base priority. */
1762 	td->td_base_pri = prio;
1763 
1764 	/*
1765 	 * If the thread is borrowing another thread's priority, don't
1766 	 * ever lower the priority.
1767 	 */
1768 	if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
1769 		return;
1770 
1771 	/* Change the real priority. */
1772 	oldprio = td->td_priority;
1773 	sched_thread_priority(td, prio);
1774 
1775 	/*
1776 	 * If the thread is on a turnstile, then let the turnstile update
1777 	 * its state.
1778 	 */
1779 	if (TD_ON_LOCK(td) && oldprio != prio)
1780 		turnstile_adjust(td, oldprio);
1781 }
1782 
1783 /*
1784  * Set the base user priority, does not effect current running priority.
1785  */
1786 void
1787 sched_user_prio(struct thread *td, u_char prio)
1788 {
1789 
1790 	td->td_base_user_pri = prio;
1791 	if (td->td_lend_user_pri <= prio)
1792 		return;
1793 	td->td_user_pri = prio;
1794 }
1795 
1796 void
1797 sched_lend_user_prio(struct thread *td, u_char prio)
1798 {
1799 
1800 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1801 	td->td_lend_user_pri = prio;
1802 	td->td_user_pri = min(prio, td->td_base_user_pri);
1803 	if (td->td_priority > td->td_user_pri)
1804 		sched_prio(td, td->td_user_pri);
1805 	else if (td->td_priority != td->td_user_pri)
1806 		td->td_flags |= TDF_NEEDRESCHED;
1807 }
1808 
1809 /*
1810  * Handle migration from sched_switch().  This happens only for
1811  * cpu binding.
1812  */
1813 static struct mtx *
1814 sched_switch_migrate(struct tdq *tdq, struct thread *td, int flags)
1815 {
1816 	struct tdq *tdn;
1817 
1818 	tdn = TDQ_CPU(td->td_sched->ts_cpu);
1819 #ifdef SMP
1820 	tdq_load_rem(tdq, td);
1821 	/*
1822 	 * Do the lock dance required to avoid LOR.  We grab an extra
1823 	 * spinlock nesting to prevent preemption while we're
1824 	 * not holding either run-queue lock.
1825 	 */
1826 	spinlock_enter();
1827 	thread_lock_block(td);	/* This releases the lock on tdq. */
1828 
1829 	/*
1830 	 * Acquire both run-queue locks before placing the thread on the new
1831 	 * run-queue to avoid deadlocks created by placing a thread with a
1832 	 * blocked lock on the run-queue of a remote processor.  The deadlock
1833 	 * occurs when a third processor attempts to lock the two queues in
1834 	 * question while the target processor is spinning with its own
1835 	 * run-queue lock held while waiting for the blocked lock to clear.
1836 	 */
1837 	tdq_lock_pair(tdn, tdq);
1838 	tdq_add(tdn, td, flags);
1839 	tdq_notify(tdn, td);
1840 	TDQ_UNLOCK(tdn);
1841 	spinlock_exit();
1842 #endif
1843 	return (TDQ_LOCKPTR(tdn));
1844 }
1845 
1846 /*
1847  * Variadic version of thread_lock_unblock() that does not assume td_lock
1848  * is blocked.
1849  */
1850 static inline void
1851 thread_unblock_switch(struct thread *td, struct mtx *mtx)
1852 {
1853 	atomic_store_rel_ptr((volatile uintptr_t *)&td->td_lock,
1854 	    (uintptr_t)mtx);
1855 }
1856 
1857 /*
1858  * Switch threads.  This function has to handle threads coming in while
1859  * blocked for some reason, running, or idle.  It also must deal with
1860  * migrating a thread from one queue to another as running threads may
1861  * be assigned elsewhere via binding.
1862  */
1863 void
1864 sched_switch(struct thread *td, struct thread *newtd, int flags)
1865 {
1866 	struct tdq *tdq;
1867 	struct td_sched *ts;
1868 	struct mtx *mtx;
1869 	int srqflag;
1870 	int cpuid, preempted;
1871 
1872 	THREAD_LOCK_ASSERT(td, MA_OWNED);
1873 	KASSERT(newtd == NULL, ("sched_switch: Unsupported newtd argument"));
1874 
1875 	cpuid = PCPU_GET(cpuid);
1876 	tdq = TDQ_CPU(cpuid);
1877 	ts = td->td_sched;
1878 	mtx = td->td_lock;
1879 	sched_pctcpu_update(ts, 1);
1880 	ts->ts_rltick = ticks;
1881 	td->td_lastcpu = td->td_oncpu;
1882 	td->td_oncpu = NOCPU;
1883 	preempted = !(td->td_flags & TDF_SLICEEND);
1884 	td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND);
1885 	td->td_owepreempt = 0;
1886 	if (!TD_IS_IDLETHREAD(td))
1887 		tdq->tdq_switchcnt++;
1888 	/*
1889 	 * The lock pointer in an idle thread should never change.  Reset it
1890 	 * to CAN_RUN as well.
1891 	 */
1892 	if (TD_IS_IDLETHREAD(td)) {
1893 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1894 		TD_SET_CAN_RUN(td);
1895 	} else if (TD_IS_RUNNING(td)) {
1896 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1897 		srqflag = preempted ?
1898 		    SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
1899 		    SRQ_OURSELF|SRQ_YIELDING;
1900 #ifdef SMP
1901 		if (THREAD_CAN_MIGRATE(td) && !THREAD_CAN_SCHED(td, ts->ts_cpu))
1902 			ts->ts_cpu = sched_pickcpu(td, 0);
1903 #endif
1904 		if (ts->ts_cpu == cpuid)
1905 			tdq_runq_add(tdq, td, srqflag);
1906 		else {
1907 			KASSERT(THREAD_CAN_MIGRATE(td) ||
1908 			    (ts->ts_flags & TSF_BOUND) != 0,
1909 			    ("Thread %p shouldn't migrate", td));
1910 			mtx = sched_switch_migrate(tdq, td, srqflag);
1911 		}
1912 	} else {
1913 		/* This thread must be going to sleep. */
1914 		TDQ_LOCK(tdq);
1915 		mtx = thread_lock_block(td);
1916 		tdq_load_rem(tdq, td);
1917 	}
1918 	/*
1919 	 * We enter here with the thread blocked and assigned to the
1920 	 * appropriate cpu run-queue or sleep-queue and with the current
1921 	 * thread-queue locked.
1922 	 */
1923 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
1924 	newtd = choosethread();
1925 	/*
1926 	 * Call the MD code to switch contexts if necessary.
1927 	 */
1928 	if (td != newtd) {
1929 #ifdef	HWPMC_HOOKS
1930 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1931 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
1932 #endif
1933 		SDT_PROBE2(sched, , , off_cpu, newtd, newtd->td_proc);
1934 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
1935 		TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
1936 		sched_pctcpu_update(newtd->td_sched, 0);
1937 
1938 #ifdef KDTRACE_HOOKS
1939 		/*
1940 		 * If DTrace has set the active vtime enum to anything
1941 		 * other than INACTIVE (0), then it should have set the
1942 		 * function to call.
1943 		 */
1944 		if (dtrace_vtime_active)
1945 			(*dtrace_vtime_switch_func)(newtd);
1946 #endif
1947 
1948 		cpu_switch(td, newtd, mtx);
1949 		/*
1950 		 * We may return from cpu_switch on a different cpu.  However,
1951 		 * we always return with td_lock pointing to the current cpu's
1952 		 * run queue lock.
1953 		 */
1954 		cpuid = PCPU_GET(cpuid);
1955 		tdq = TDQ_CPU(cpuid);
1956 		lock_profile_obtain_lock_success(
1957 		    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
1958 
1959 		SDT_PROBE0(sched, , , on_cpu);
1960 #ifdef	HWPMC_HOOKS
1961 		if (PMC_PROC_IS_USING_PMCS(td->td_proc))
1962 			PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
1963 #endif
1964 	} else {
1965 		thread_unblock_switch(td, mtx);
1966 		SDT_PROBE0(sched, , , remain_cpu);
1967 	}
1968 	/*
1969 	 * Assert that all went well and return.
1970 	 */
1971 	TDQ_LOCK_ASSERT(tdq, MA_OWNED|MA_NOTRECURSED);
1972 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
1973 	td->td_oncpu = cpuid;
1974 }
1975 
1976 /*
1977  * Adjust thread priorities as a result of a nice request.
1978  */
1979 void
1980 sched_nice(struct proc *p, int nice)
1981 {
1982 	struct thread *td;
1983 
1984 	PROC_LOCK_ASSERT(p, MA_OWNED);
1985 
1986 	p->p_nice = nice;
1987 	FOREACH_THREAD_IN_PROC(p, td) {
1988 		thread_lock(td);
1989 		sched_priority(td);
1990 		sched_prio(td, td->td_base_user_pri);
1991 		thread_unlock(td);
1992 	}
1993 }
1994 
1995 /*
1996  * Record the sleep time for the interactivity scorer.
1997  */
1998 void
1999 sched_sleep(struct thread *td, int prio)
2000 {
2001 
2002 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2003 
2004 	td->td_slptick = ticks;
2005 	if (TD_IS_SUSPENDED(td) || prio >= PSOCK)
2006 		td->td_flags |= TDF_CANSWAP;
2007 	if (PRI_BASE(td->td_pri_class) != PRI_TIMESHARE)
2008 		return;
2009 	if (static_boost == 1 && prio)
2010 		sched_prio(td, prio);
2011 	else if (static_boost && td->td_priority > static_boost)
2012 		sched_prio(td, static_boost);
2013 }
2014 
2015 /*
2016  * Schedule a thread to resume execution and record how long it voluntarily
2017  * slept.  We also update the pctcpu, interactivity, and priority.
2018  */
2019 void
2020 sched_wakeup(struct thread *td)
2021 {
2022 	struct td_sched *ts;
2023 	int slptick;
2024 
2025 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2026 	ts = td->td_sched;
2027 	td->td_flags &= ~TDF_CANSWAP;
2028 	/*
2029 	 * If we slept for more than a tick update our interactivity and
2030 	 * priority.
2031 	 */
2032 	slptick = td->td_slptick;
2033 	td->td_slptick = 0;
2034 	if (slptick && slptick != ticks) {
2035 		ts->ts_slptime += (ticks - slptick) << SCHED_TICK_SHIFT;
2036 		sched_interact_update(td);
2037 		sched_pctcpu_update(ts, 0);
2038 	}
2039 	/*
2040 	 * Reset the slice value since we slept and advanced the round-robin.
2041 	 */
2042 	ts->ts_slice = 0;
2043 	sched_add(td, SRQ_BORING);
2044 }
2045 
2046 /*
2047  * Penalize the parent for creating a new child and initialize the child's
2048  * priority.
2049  */
2050 void
2051 sched_fork(struct thread *td, struct thread *child)
2052 {
2053 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2054 	sched_pctcpu_update(td->td_sched, 1);
2055 	sched_fork_thread(td, child);
2056 	/*
2057 	 * Penalize the parent and child for forking.
2058 	 */
2059 	sched_interact_fork(child);
2060 	sched_priority(child);
2061 	td->td_sched->ts_runtime += tickincr;
2062 	sched_interact_update(td);
2063 	sched_priority(td);
2064 }
2065 
2066 /*
2067  * Fork a new thread, may be within the same process.
2068  */
2069 void
2070 sched_fork_thread(struct thread *td, struct thread *child)
2071 {
2072 	struct td_sched *ts;
2073 	struct td_sched *ts2;
2074 	struct tdq *tdq;
2075 
2076 	tdq = TDQ_SELF();
2077 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2078 	/*
2079 	 * Initialize child.
2080 	 */
2081 	ts = td->td_sched;
2082 	ts2 = child->td_sched;
2083 	child->td_lock = TDQ_LOCKPTR(tdq);
2084 	child->td_cpuset = cpuset_ref(td->td_cpuset);
2085 	ts2->ts_cpu = ts->ts_cpu;
2086 	ts2->ts_flags = 0;
2087 	/*
2088 	 * Grab our parents cpu estimation information.
2089 	 */
2090 	ts2->ts_ticks = ts->ts_ticks;
2091 	ts2->ts_ltick = ts->ts_ltick;
2092 	ts2->ts_ftick = ts->ts_ftick;
2093 	/*
2094 	 * Do not inherit any borrowed priority from the parent.
2095 	 */
2096 	child->td_priority = child->td_base_pri;
2097 	/*
2098 	 * And update interactivity score.
2099 	 */
2100 	ts2->ts_slptime = ts->ts_slptime;
2101 	ts2->ts_runtime = ts->ts_runtime;
2102 	/* Attempt to quickly learn interactivity. */
2103 	ts2->ts_slice = tdq_slice(tdq) - sched_slice_min;
2104 #ifdef KTR
2105 	bzero(ts2->ts_name, sizeof(ts2->ts_name));
2106 #endif
2107 }
2108 
2109 /*
2110  * Adjust the priority class of a thread.
2111  */
2112 void
2113 sched_class(struct thread *td, int class)
2114 {
2115 
2116 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2117 	if (td->td_pri_class == class)
2118 		return;
2119 	td->td_pri_class = class;
2120 }
2121 
2122 /*
2123  * Return some of the child's priority and interactivity to the parent.
2124  */
2125 void
2126 sched_exit(struct proc *p, struct thread *child)
2127 {
2128 	struct thread *td;
2129 
2130 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "proc exit",
2131 	    "prio:%d", child->td_priority);
2132 	PROC_LOCK_ASSERT(p, MA_OWNED);
2133 	td = FIRST_THREAD_IN_PROC(p);
2134 	sched_exit_thread(td, child);
2135 }
2136 
2137 /*
2138  * Penalize another thread for the time spent on this one.  This helps to
2139  * worsen the priority and interactivity of processes which schedule batch
2140  * jobs such as make.  This has little effect on the make process itself but
2141  * causes new processes spawned by it to receive worse scores immediately.
2142  */
2143 void
2144 sched_exit_thread(struct thread *td, struct thread *child)
2145 {
2146 
2147 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "thread exit",
2148 	    "prio:%d", child->td_priority);
2149 	/*
2150 	 * Give the child's runtime to the parent without returning the
2151 	 * sleep time as a penalty to the parent.  This causes shells that
2152 	 * launch expensive things to mark their children as expensive.
2153 	 */
2154 	thread_lock(td);
2155 	td->td_sched->ts_runtime += child->td_sched->ts_runtime;
2156 	sched_interact_update(td);
2157 	sched_priority(td);
2158 	thread_unlock(td);
2159 }
2160 
2161 void
2162 sched_preempt(struct thread *td)
2163 {
2164 	struct tdq *tdq;
2165 
2166 	SDT_PROBE2(sched, , , surrender, td, td->td_proc);
2167 
2168 	thread_lock(td);
2169 	tdq = TDQ_SELF();
2170 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2171 	tdq->tdq_ipipending = 0;
2172 	if (td->td_priority > tdq->tdq_lowpri) {
2173 		int flags;
2174 
2175 		flags = SW_INVOL | SW_PREEMPT;
2176 		if (td->td_critnest > 1)
2177 			td->td_owepreempt = 1;
2178 		else if (TD_IS_IDLETHREAD(td))
2179 			mi_switch(flags | SWT_REMOTEWAKEIDLE, NULL);
2180 		else
2181 			mi_switch(flags | SWT_REMOTEPREEMPT, NULL);
2182 	}
2183 	thread_unlock(td);
2184 }
2185 
2186 /*
2187  * Fix priorities on return to user-space.  Priorities may be elevated due
2188  * to static priorities in msleep() or similar.
2189  */
2190 void
2191 sched_userret(struct thread *td)
2192 {
2193 	/*
2194 	 * XXX we cheat slightly on the locking here to avoid locking in
2195 	 * the usual case.  Setting td_priority here is essentially an
2196 	 * incomplete workaround for not setting it properly elsewhere.
2197 	 * Now that some interrupt handlers are threads, not setting it
2198 	 * properly elsewhere can clobber it in the window between setting
2199 	 * it here and returning to user mode, so don't waste time setting
2200 	 * it perfectly here.
2201 	 */
2202 	KASSERT((td->td_flags & TDF_BORROWING) == 0,
2203 	    ("thread with borrowed priority returning to userland"));
2204 	if (td->td_priority != td->td_user_pri) {
2205 		thread_lock(td);
2206 		td->td_priority = td->td_user_pri;
2207 		td->td_base_pri = td->td_user_pri;
2208 		tdq_setlowpri(TDQ_SELF(), td);
2209 		thread_unlock(td);
2210         }
2211 }
2212 
2213 /*
2214  * Handle a stathz tick.  This is really only relevant for timeshare
2215  * threads.
2216  */
2217 void
2218 sched_clock(struct thread *td)
2219 {
2220 	struct tdq *tdq;
2221 	struct td_sched *ts;
2222 
2223 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2224 	tdq = TDQ_SELF();
2225 #ifdef SMP
2226 	/*
2227 	 * We run the long term load balancer infrequently on the first cpu.
2228 	 */
2229 	if (balance_tdq == tdq) {
2230 		if (balance_ticks && --balance_ticks == 0)
2231 			sched_balance();
2232 	}
2233 #endif
2234 	/*
2235 	 * Save the old switch count so we have a record of the last ticks
2236 	 * activity.   Initialize the new switch count based on our load.
2237 	 * If there is some activity seed it to reflect that.
2238 	 */
2239 	tdq->tdq_oldswitchcnt = tdq->tdq_switchcnt;
2240 	tdq->tdq_switchcnt = tdq->tdq_load;
2241 	/*
2242 	 * Advance the insert index once for each tick to ensure that all
2243 	 * threads get a chance to run.
2244 	 */
2245 	if (tdq->tdq_idx == tdq->tdq_ridx) {
2246 		tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
2247 		if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
2248 			tdq->tdq_ridx = tdq->tdq_idx;
2249 	}
2250 	ts = td->td_sched;
2251 	sched_pctcpu_update(ts, 1);
2252 	if (td->td_pri_class & PRI_FIFO_BIT)
2253 		return;
2254 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) {
2255 		/*
2256 		 * We used a tick; charge it to the thread so
2257 		 * that we can compute our interactivity.
2258 		 */
2259 		td->td_sched->ts_runtime += tickincr;
2260 		sched_interact_update(td);
2261 		sched_priority(td);
2262 	}
2263 
2264 	/*
2265 	 * Force a context switch if the current thread has used up a full
2266 	 * time slice (default is 100ms).
2267 	 */
2268 	if (!TD_IS_IDLETHREAD(td) && ++ts->ts_slice >= tdq_slice(tdq)) {
2269 		ts->ts_slice = 0;
2270 		td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND;
2271 	}
2272 }
2273 
2274 /*
2275  * Called once per hz tick.
2276  */
2277 void
2278 sched_tick(int cnt)
2279 {
2280 
2281 }
2282 
2283 /*
2284  * Return whether the current CPU has runnable tasks.  Used for in-kernel
2285  * cooperative idle threads.
2286  */
2287 int
2288 sched_runnable(void)
2289 {
2290 	struct tdq *tdq;
2291 	int load;
2292 
2293 	load = 1;
2294 
2295 	tdq = TDQ_SELF();
2296 	if ((curthread->td_flags & TDF_IDLETD) != 0) {
2297 		if (tdq->tdq_load > 0)
2298 			goto out;
2299 	} else
2300 		if (tdq->tdq_load - 1 > 0)
2301 			goto out;
2302 	load = 0;
2303 out:
2304 	return (load);
2305 }
2306 
2307 /*
2308  * Choose the highest priority thread to run.  The thread is removed from
2309  * the run-queue while running however the load remains.  For SMP we set
2310  * the tdq in the global idle bitmask if it idles here.
2311  */
2312 struct thread *
2313 sched_choose(void)
2314 {
2315 	struct thread *td;
2316 	struct tdq *tdq;
2317 
2318 	tdq = TDQ_SELF();
2319 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2320 	td = tdq_choose(tdq);
2321 	if (td) {
2322 		tdq_runq_rem(tdq, td);
2323 		tdq->tdq_lowpri = td->td_priority;
2324 		return (td);
2325 	}
2326 	tdq->tdq_lowpri = PRI_MAX_IDLE;
2327 	return (PCPU_GET(idlethread));
2328 }
2329 
2330 /*
2331  * Set owepreempt if necessary.  Preemption never happens directly in ULE,
2332  * we always request it once we exit a critical section.
2333  */
2334 static inline void
2335 sched_setpreempt(struct thread *td)
2336 {
2337 	struct thread *ctd;
2338 	int cpri;
2339 	int pri;
2340 
2341 	THREAD_LOCK_ASSERT(curthread, MA_OWNED);
2342 
2343 	ctd = curthread;
2344 	pri = td->td_priority;
2345 	cpri = ctd->td_priority;
2346 	if (pri < cpri)
2347 		ctd->td_flags |= TDF_NEEDRESCHED;
2348 	if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
2349 		return;
2350 	if (!sched_shouldpreempt(pri, cpri, 0))
2351 		return;
2352 	ctd->td_owepreempt = 1;
2353 }
2354 
2355 /*
2356  * Add a thread to a thread queue.  Select the appropriate runq and add the
2357  * thread to it.  This is the internal function called when the tdq is
2358  * predetermined.
2359  */
2360 void
2361 tdq_add(struct tdq *tdq, struct thread *td, int flags)
2362 {
2363 
2364 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2365 	KASSERT((td->td_inhibitors == 0),
2366 	    ("sched_add: trying to run inhibited thread"));
2367 	KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
2368 	    ("sched_add: bad thread state"));
2369 	KASSERT(td->td_flags & TDF_INMEM,
2370 	    ("sched_add: thread swapped out"));
2371 
2372 	if (td->td_priority < tdq->tdq_lowpri)
2373 		tdq->tdq_lowpri = td->td_priority;
2374 	tdq_runq_add(tdq, td, flags);
2375 	tdq_load_add(tdq, td);
2376 }
2377 
2378 /*
2379  * Select the target thread queue and add a thread to it.  Request
2380  * preemption or IPI a remote processor if required.
2381  */
2382 void
2383 sched_add(struct thread *td, int flags)
2384 {
2385 	struct tdq *tdq;
2386 #ifdef SMP
2387 	int cpu;
2388 #endif
2389 
2390 	KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add",
2391 	    "prio:%d", td->td_priority, KTR_ATTR_LINKED,
2392 	    sched_tdname(curthread));
2393 	KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup",
2394 	    KTR_ATTR_LINKED, sched_tdname(td));
2395 	SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL,
2396 	    flags & SRQ_PREEMPTED);
2397 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2398 	/*
2399 	 * Recalculate the priority before we select the target cpu or
2400 	 * run-queue.
2401 	 */
2402 	if (PRI_BASE(td->td_pri_class) == PRI_TIMESHARE)
2403 		sched_priority(td);
2404 #ifdef SMP
2405 	/*
2406 	 * Pick the destination cpu and if it isn't ours transfer to the
2407 	 * target cpu.
2408 	 */
2409 	cpu = sched_pickcpu(td, flags);
2410 	tdq = sched_setcpu(td, cpu, flags);
2411 	tdq_add(tdq, td, flags);
2412 	if (cpu != PCPU_GET(cpuid)) {
2413 		tdq_notify(tdq, td);
2414 		return;
2415 	}
2416 #else
2417 	tdq = TDQ_SELF();
2418 	TDQ_LOCK(tdq);
2419 	/*
2420 	 * Now that the thread is moving to the run-queue, set the lock
2421 	 * to the scheduler's lock.
2422 	 */
2423 	thread_lock_set(td, TDQ_LOCKPTR(tdq));
2424 	tdq_add(tdq, td, flags);
2425 #endif
2426 	if (!(flags & SRQ_YIELDING))
2427 		sched_setpreempt(td);
2428 }
2429 
2430 /*
2431  * Remove a thread from a run-queue without running it.  This is used
2432  * when we're stealing a thread from a remote queue.  Otherwise all threads
2433  * exit by calling sched_exit_thread() and sched_throw() themselves.
2434  */
2435 void
2436 sched_rem(struct thread *td)
2437 {
2438 	struct tdq *tdq;
2439 
2440 	KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "runq rem",
2441 	    "prio:%d", td->td_priority);
2442 	SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL);
2443 	tdq = TDQ_CPU(td->td_sched->ts_cpu);
2444 	TDQ_LOCK_ASSERT(tdq, MA_OWNED);
2445 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2446 	KASSERT(TD_ON_RUNQ(td),
2447 	    ("sched_rem: thread not on run queue"));
2448 	tdq_runq_rem(tdq, td);
2449 	tdq_load_rem(tdq, td);
2450 	TD_SET_CAN_RUN(td);
2451 	if (td->td_priority == tdq->tdq_lowpri)
2452 		tdq_setlowpri(tdq, NULL);
2453 }
2454 
2455 /*
2456  * Fetch cpu utilization information.  Updates on demand.
2457  */
2458 fixpt_t
2459 sched_pctcpu(struct thread *td)
2460 {
2461 	fixpt_t pctcpu;
2462 	struct td_sched *ts;
2463 
2464 	pctcpu = 0;
2465 	ts = td->td_sched;
2466 	if (ts == NULL)
2467 		return (0);
2468 
2469 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2470 	sched_pctcpu_update(ts, TD_IS_RUNNING(td));
2471 	if (ts->ts_ticks) {
2472 		int rtick;
2473 
2474 		/* How many rtick per second ? */
2475 		rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
2476 		pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
2477 	}
2478 
2479 	return (pctcpu);
2480 }
2481 
2482 /*
2483  * Enforce affinity settings for a thread.  Called after adjustments to
2484  * cpumask.
2485  */
2486 void
2487 sched_affinity(struct thread *td)
2488 {
2489 #ifdef SMP
2490 	struct td_sched *ts;
2491 
2492 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2493 	ts = td->td_sched;
2494 	if (THREAD_CAN_SCHED(td, ts->ts_cpu))
2495 		return;
2496 	if (TD_ON_RUNQ(td)) {
2497 		sched_rem(td);
2498 		sched_add(td, SRQ_BORING);
2499 		return;
2500 	}
2501 	if (!TD_IS_RUNNING(td))
2502 		return;
2503 	/*
2504 	 * Force a switch before returning to userspace.  If the
2505 	 * target thread is not running locally send an ipi to force
2506 	 * the issue.
2507 	 */
2508 	td->td_flags |= TDF_NEEDRESCHED;
2509 	if (td != curthread)
2510 		ipi_cpu(ts->ts_cpu, IPI_PREEMPT);
2511 #endif
2512 }
2513 
2514 /*
2515  * Bind a thread to a target cpu.
2516  */
2517 void
2518 sched_bind(struct thread *td, int cpu)
2519 {
2520 	struct td_sched *ts;
2521 
2522 	THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED);
2523 	KASSERT(td == curthread, ("sched_bind: can only bind curthread"));
2524 	ts = td->td_sched;
2525 	if (ts->ts_flags & TSF_BOUND)
2526 		sched_unbind(td);
2527 	KASSERT(THREAD_CAN_MIGRATE(td), ("%p must be migratable", td));
2528 	ts->ts_flags |= TSF_BOUND;
2529 	sched_pin();
2530 	if (PCPU_GET(cpuid) == cpu)
2531 		return;
2532 	ts->ts_cpu = cpu;
2533 	/* When we return from mi_switch we'll be on the correct cpu. */
2534 	mi_switch(SW_VOL, NULL);
2535 }
2536 
2537 /*
2538  * Release a bound thread.
2539  */
2540 void
2541 sched_unbind(struct thread *td)
2542 {
2543 	struct td_sched *ts;
2544 
2545 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2546 	KASSERT(td == curthread, ("sched_unbind: can only bind curthread"));
2547 	ts = td->td_sched;
2548 	if ((ts->ts_flags & TSF_BOUND) == 0)
2549 		return;
2550 	ts->ts_flags &= ~TSF_BOUND;
2551 	sched_unpin();
2552 }
2553 
2554 int
2555 sched_is_bound(struct thread *td)
2556 {
2557 	THREAD_LOCK_ASSERT(td, MA_OWNED);
2558 	return (td->td_sched->ts_flags & TSF_BOUND);
2559 }
2560 
2561 /*
2562  * Basic yield call.
2563  */
2564 void
2565 sched_relinquish(struct thread *td)
2566 {
2567 	thread_lock(td);
2568 	mi_switch(SW_VOL | SWT_RELINQUISH, NULL);
2569 	thread_unlock(td);
2570 }
2571 
2572 /*
2573  * Return the total system load.
2574  */
2575 int
2576 sched_load(void)
2577 {
2578 #ifdef SMP
2579 	int total;
2580 	int i;
2581 
2582 	total = 0;
2583 	CPU_FOREACH(i)
2584 		total += TDQ_CPU(i)->tdq_sysload;
2585 	return (total);
2586 #else
2587 	return (TDQ_SELF()->tdq_sysload);
2588 #endif
2589 }
2590 
2591 int
2592 sched_sizeof_proc(void)
2593 {
2594 	return (sizeof(struct proc));
2595 }
2596 
2597 int
2598 sched_sizeof_thread(void)
2599 {
2600 	return (sizeof(struct thread) + sizeof(struct td_sched));
2601 }
2602 
2603 #ifdef SMP
2604 #define	TDQ_IDLESPIN(tdq)						\
2605     ((tdq)->tdq_cg != NULL && ((tdq)->tdq_cg->cg_flags & CG_FLAG_THREAD) == 0)
2606 #else
2607 #define	TDQ_IDLESPIN(tdq)	1
2608 #endif
2609 
2610 /*
2611  * The actual idle process.
2612  */
2613 void
2614 sched_idletd(void *dummy)
2615 {
2616 	struct thread *td;
2617 	struct tdq *tdq;
2618 	int oldswitchcnt, switchcnt;
2619 	int i;
2620 
2621 	mtx_assert(&Giant, MA_NOTOWNED);
2622 	td = curthread;
2623 	tdq = TDQ_SELF();
2624 	THREAD_NO_SLEEPING();
2625 	oldswitchcnt = -1;
2626 	for (;;) {
2627 		if (tdq->tdq_load) {
2628 			thread_lock(td);
2629 			mi_switch(SW_VOL | SWT_IDLE, NULL);
2630 			thread_unlock(td);
2631 		}
2632 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2633 #ifdef SMP
2634 		if (switchcnt != oldswitchcnt) {
2635 			oldswitchcnt = switchcnt;
2636 			if (tdq_idled(tdq) == 0)
2637 				continue;
2638 		}
2639 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2640 #else
2641 		oldswitchcnt = switchcnt;
2642 #endif
2643 		/*
2644 		 * If we're switching very frequently, spin while checking
2645 		 * for load rather than entering a low power state that
2646 		 * may require an IPI.  However, don't do any busy
2647 		 * loops while on SMT machines as this simply steals
2648 		 * cycles from cores doing useful work.
2649 		 */
2650 		if (TDQ_IDLESPIN(tdq) && switchcnt > sched_idlespinthresh) {
2651 			for (i = 0; i < sched_idlespins; i++) {
2652 				if (tdq->tdq_load)
2653 					break;
2654 				cpu_spinwait();
2655 			}
2656 		}
2657 
2658 		/* If there was context switch during spin, restart it. */
2659 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2660 		if (tdq->tdq_load != 0 || switchcnt != oldswitchcnt)
2661 			continue;
2662 
2663 		/* Run main MD idle handler. */
2664 		tdq->tdq_cpu_idle = 1;
2665 		cpu_idle(switchcnt * 4 > sched_idlespinthresh);
2666 		tdq->tdq_cpu_idle = 0;
2667 
2668 		/*
2669 		 * Account thread-less hardware interrupts and
2670 		 * other wakeup reasons equal to context switches.
2671 		 */
2672 		switchcnt = tdq->tdq_switchcnt + tdq->tdq_oldswitchcnt;
2673 		if (switchcnt != oldswitchcnt)
2674 			continue;
2675 		tdq->tdq_switchcnt++;
2676 		oldswitchcnt++;
2677 	}
2678 }
2679 
2680 /*
2681  * A CPU is entering for the first time or a thread is exiting.
2682  */
2683 void
2684 sched_throw(struct thread *td)
2685 {
2686 	struct thread *newtd;
2687 	struct tdq *tdq;
2688 
2689 	tdq = TDQ_SELF();
2690 	if (td == NULL) {
2691 		/* Correct spinlock nesting and acquire the correct lock. */
2692 		TDQ_LOCK(tdq);
2693 		spinlock_exit();
2694 		PCPU_SET(switchtime, cpu_ticks());
2695 		PCPU_SET(switchticks, ticks);
2696 	} else {
2697 		MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2698 		tdq_load_rem(tdq, td);
2699 		lock_profile_release_lock(&TDQ_LOCKPTR(tdq)->lock_object);
2700 	}
2701 	KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
2702 	newtd = choosethread();
2703 	TDQ_LOCKPTR(tdq)->mtx_lock = (uintptr_t)newtd;
2704 	cpu_throw(td, newtd);		/* doesn't return */
2705 }
2706 
2707 /*
2708  * This is called from fork_exit().  Just acquire the correct locks and
2709  * let fork do the rest of the work.
2710  */
2711 void
2712 sched_fork_exit(struct thread *td)
2713 {
2714 	struct td_sched *ts;
2715 	struct tdq *tdq;
2716 	int cpuid;
2717 
2718 	/*
2719 	 * Finish setting up thread glue so that it begins execution in a
2720 	 * non-nested critical section with the scheduler lock held.
2721 	 */
2722 	cpuid = PCPU_GET(cpuid);
2723 	tdq = TDQ_CPU(cpuid);
2724 	ts = td->td_sched;
2725 	if (TD_IS_IDLETHREAD(td))
2726 		td->td_lock = TDQ_LOCKPTR(tdq);
2727 	MPASS(td->td_lock == TDQ_LOCKPTR(tdq));
2728 	td->td_oncpu = cpuid;
2729 	TDQ_LOCK_ASSERT(tdq, MA_OWNED | MA_NOTRECURSED);
2730 	lock_profile_obtain_lock_success(
2731 	    &TDQ_LOCKPTR(tdq)->lock_object, 0, 0, __FILE__, __LINE__);
2732 }
2733 
2734 /*
2735  * Create on first use to catch odd startup conditons.
2736  */
2737 char *
2738 sched_tdname(struct thread *td)
2739 {
2740 #ifdef KTR
2741 	struct td_sched *ts;
2742 
2743 	ts = td->td_sched;
2744 	if (ts->ts_name[0] == '\0')
2745 		snprintf(ts->ts_name, sizeof(ts->ts_name),
2746 		    "%s tid %d", td->td_name, td->td_tid);
2747 	return (ts->ts_name);
2748 #else
2749 	return (td->td_name);
2750 #endif
2751 }
2752 
2753 #ifdef KTR
2754 void
2755 sched_clear_tdname(struct thread *td)
2756 {
2757 	struct td_sched *ts;
2758 
2759 	ts = td->td_sched;
2760 	ts->ts_name[0] = '\0';
2761 }
2762 #endif
2763 
2764 #ifdef SMP
2765 
2766 /*
2767  * Build the CPU topology dump string. Is recursively called to collect
2768  * the topology tree.
2769  */
2770 static int
2771 sysctl_kern_sched_topology_spec_internal(struct sbuf *sb, struct cpu_group *cg,
2772     int indent)
2773 {
2774 	char cpusetbuf[CPUSETBUFSIZ];
2775 	int i, first;
2776 
2777 	sbuf_printf(sb, "%*s<group level=\"%d\" cache-level=\"%d\">\n", indent,
2778 	    "", 1 + indent / 2, cg->cg_level);
2779 	sbuf_printf(sb, "%*s <cpu count=\"%d\" mask=\"%s\">", indent, "",
2780 	    cg->cg_count, cpusetobj_strprint(cpusetbuf, &cg->cg_mask));
2781 	first = TRUE;
2782 	for (i = 0; i < MAXCPU; i++) {
2783 		if (CPU_ISSET(i, &cg->cg_mask)) {
2784 			if (!first)
2785 				sbuf_printf(sb, ", ");
2786 			else
2787 				first = FALSE;
2788 			sbuf_printf(sb, "%d", i);
2789 		}
2790 	}
2791 	sbuf_printf(sb, "</cpu>\n");
2792 
2793 	if (cg->cg_flags != 0) {
2794 		sbuf_printf(sb, "%*s <flags>", indent, "");
2795 		if ((cg->cg_flags & CG_FLAG_HTT) != 0)
2796 			sbuf_printf(sb, "<flag name=\"HTT\">HTT group</flag>");
2797 		if ((cg->cg_flags & CG_FLAG_THREAD) != 0)
2798 			sbuf_printf(sb, "<flag name=\"THREAD\">THREAD group</flag>");
2799 		if ((cg->cg_flags & CG_FLAG_SMT) != 0)
2800 			sbuf_printf(sb, "<flag name=\"SMT\">SMT group</flag>");
2801 		sbuf_printf(sb, "</flags>\n");
2802 	}
2803 
2804 	if (cg->cg_children > 0) {
2805 		sbuf_printf(sb, "%*s <children>\n", indent, "");
2806 		for (i = 0; i < cg->cg_children; i++)
2807 			sysctl_kern_sched_topology_spec_internal(sb,
2808 			    &cg->cg_child[i], indent+2);
2809 		sbuf_printf(sb, "%*s </children>\n", indent, "");
2810 	}
2811 	sbuf_printf(sb, "%*s</group>\n", indent, "");
2812 	return (0);
2813 }
2814 
2815 /*
2816  * Sysctl handler for retrieving topology dump. It's a wrapper for
2817  * the recursive sysctl_kern_smp_topology_spec_internal().
2818  */
2819 static int
2820 sysctl_kern_sched_topology_spec(SYSCTL_HANDLER_ARGS)
2821 {
2822 	struct sbuf *topo;
2823 	int err;
2824 
2825 	KASSERT(cpu_top != NULL, ("cpu_top isn't initialized"));
2826 
2827 	topo = sbuf_new(NULL, NULL, 500, SBUF_AUTOEXTEND);
2828 	if (topo == NULL)
2829 		return (ENOMEM);
2830 
2831 	sbuf_printf(topo, "<groups>\n");
2832 	err = sysctl_kern_sched_topology_spec_internal(topo, cpu_top, 1);
2833 	sbuf_printf(topo, "</groups>\n");
2834 
2835 	if (err == 0) {
2836 		sbuf_finish(topo);
2837 		err = SYSCTL_OUT(req, sbuf_data(topo), sbuf_len(topo));
2838 	}
2839 	sbuf_delete(topo);
2840 	return (err);
2841 }
2842 
2843 #endif
2844 
2845 static int
2846 sysctl_kern_quantum(SYSCTL_HANDLER_ARGS)
2847 {
2848 	int error, new_val, period;
2849 
2850 	period = 1000000 / realstathz;
2851 	new_val = period * sched_slice;
2852 	error = sysctl_handle_int(oidp, &new_val, 0, req);
2853 	if (error != 0 || req->newptr == NULL)
2854 		return (error);
2855 	if (new_val <= 0)
2856 		return (EINVAL);
2857 	sched_slice = imax(1, (new_val + period / 2) / period);
2858 	sched_slice_min = sched_slice / SCHED_SLICE_MIN_DIVISOR;
2859 	hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) /
2860 	    realstathz);
2861 	return (0);
2862 }
2863 
2864 SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
2865 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ULE", 0,
2866     "Scheduler name");
2867 SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW,
2868     NULL, 0, sysctl_kern_quantum, "I",
2869     "Quantum for timeshare threads in microseconds");
2870 SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0,
2871     "Quantum for timeshare threads in stathz ticks");
2872 SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0,
2873     "Interactivity score threshold");
2874 SYSCTL_INT(_kern_sched, OID_AUTO, preempt_thresh, CTLFLAG_RW,
2875     &preempt_thresh, 0,
2876     "Maximal (lowest) priority for preemption");
2877 SYSCTL_INT(_kern_sched, OID_AUTO, static_boost, CTLFLAG_RW, &static_boost, 0,
2878     "Assign static kernel priorities to sleeping threads");
2879 SYSCTL_INT(_kern_sched, OID_AUTO, idlespins, CTLFLAG_RW, &sched_idlespins, 0,
2880     "Number of times idle thread will spin waiting for new work");
2881 SYSCTL_INT(_kern_sched, OID_AUTO, idlespinthresh, CTLFLAG_RW,
2882     &sched_idlespinthresh, 0,
2883     "Threshold before we will permit idle thread spinning");
2884 #ifdef SMP
2885 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
2886     "Number of hz ticks to keep thread affinity for");
2887 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
2888     "Enables the long-term load balancer");
2889 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
2890     &balance_interval, 0,
2891     "Average period in stathz ticks to run the long-term balancer");
2892 SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
2893     "Attempts to steal work from other cores before idling");
2894 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
2895     "Minimum load on remote CPU before we'll steal");
2896 SYSCTL_PROC(_kern_sched, OID_AUTO, topology_spec, CTLTYPE_STRING |
2897     CTLFLAG_RD, NULL, 0, sysctl_kern_sched_topology_spec, "A",
2898     "XML dump of detected CPU topology");
2899 #endif
2900 
2901 /* ps compat.  All cpu percentages from ULE are weighted. */
2902 static int ccpu = 0;
2903 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
2904