FreeBSD/Linux Kernel Cross Reference
sys/kern/sched_ule.c
1 /*-
2 * Copyright (c) 2002-2003, 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 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD: src/sys/kern/sched_ule.c,v 1.121.2.14 2005/02/04 20:13:21 jeff Exp $");
29
30 #include <opt_sched.h>
31
32 #define kse td_sched
33
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/kdb.h>
37 #include <sys/kernel.h>
38 #include <sys/ktr.h>
39 #include <sys/lock.h>
40 #include <sys/mutex.h>
41 #include <sys/proc.h>
42 #include <sys/resource.h>
43 #include <sys/resourcevar.h>
44 #include <sys/sched.h>
45 #include <sys/smp.h>
46 #include <sys/sx.h>
47 #include <sys/sysctl.h>
48 #include <sys/sysproto.h>
49 #include <sys/vmmeter.h>
50 #ifdef KTRACE
51 #include <sys/uio.h>
52 #include <sys/ktrace.h>
53 #endif
54
55 #include <machine/cpu.h>
56 #include <machine/smp.h>
57
58 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
59 /* XXX This is bogus compatability crap for ps */
60 static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
61 SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
62
63 static void sched_setup(void *dummy);
64 SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL)
65
66 static SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
67
68 SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ule", 0,
69 "Scheduler name");
70
71 static int slice_min = 1;
72 SYSCTL_INT(_kern_sched, OID_AUTO, slice_min, CTLFLAG_RW, &slice_min, 0, "");
73
74 static int slice_max = 10;
75 SYSCTL_INT(_kern_sched, OID_AUTO, slice_max, CTLFLAG_RW, &slice_max, 0, "");
76
77 int realstathz;
78 int tickincr = 1;
79
80 /*
81 * The schedulable entity that can be given a context to run.
82 * A process may have several of these. Probably one per processor
83 * but posibly a few more. In this universe they are grouped
84 * with a KSEG that contains the priority and niceness
85 * for the group.
86 */
87 struct kse {
88 TAILQ_ENTRY(kse) ke_procq; /* (j/z) Run queue. */
89 int ke_flags; /* (j) KEF_* flags. */
90 struct thread *ke_thread; /* (*) Active associated thread. */
91 fixpt_t ke_pctcpu; /* (j) %cpu during p_swtime. */
92 char ke_rqindex; /* (j) Run queue index. */
93 enum {
94 KES_THREAD = 0x0, /* slaved to thread state */
95 KES_ONRUNQ
96 } ke_state; /* (j) thread sched specific status. */
97 int ke_slptime;
98 int ke_slice;
99 struct runq *ke_runq;
100 u_char ke_cpu; /* CPU that we have affinity for. */
101 /* The following variables are only used for pctcpu calculation */
102 int ke_ltick; /* Last tick that we were running on */
103 int ke_ftick; /* First tick that we were running on */
104 int ke_ticks; /* Tick count */
105
106 };
107
108
109 #define td_kse td_sched
110 #define td_slptime td_kse->ke_slptime
111 #define ke_proc ke_thread->td_proc
112 #define ke_ksegrp ke_thread->td_ksegrp
113
114 /* flags kept in ke_flags */
115 #define KEF_SCHED0 0x00001 /* For scheduler-specific use. */
116 #define KEF_SCHED1 0x00002 /* For scheduler-specific use. */
117 #define KEF_SCHED2 0x00004 /* For scheduler-specific use. */
118 #define KEF_SCHED3 0x00008 /* For scheduler-specific use. */
119 #define KEF_SCHED4 0x00010
120 #define KEF_SCHED5 0x00020
121 #define KEF_DIDRUN 0x02000 /* Thread actually ran. */
122 #define KEF_EXIT 0x04000 /* Thread is being killed. */
123
124 /*
125 * These datastructures are allocated within their parent datastructure but
126 * are scheduler specific.
127 */
128
129 #define ke_assign ke_procq.tqe_next
130
131 #define KEF_ASSIGNED KEF_SCHED0 /* Thread is being migrated. */
132 #define KEF_BOUND KEF_SCHED1 /* Thread can not migrate. */
133 #define KEF_XFERABLE KEF_SCHED2 /* Thread was added as transferable. */
134 #define KEF_HOLD KEF_SCHED3 /* Thread is temporarily bound. */
135 #define KEF_REMOVED KEF_SCHED4 /* Thread was removed while ASSIGNED */
136 #define KEF_PRIOELEV KEF_SCHED5 /* Thread has had its prio elevated. */
137
138 struct kg_sched {
139 struct thread *skg_last_assigned; /* (j) Last thread assigned to */
140 /* the system scheduler */
141 int skg_slptime; /* Number of ticks we vol. slept */
142 int skg_runtime; /* Number of ticks we were running */
143 int skg_avail_opennings; /* (j) Num unfilled slots in group.*/
144 int skg_concurrency; /* (j) Num threads requested in group.*/
145 };
146 #define kg_last_assigned kg_sched->skg_last_assigned
147 #define kg_avail_opennings kg_sched->skg_avail_opennings
148 #define kg_concurrency kg_sched->skg_concurrency
149 #define kg_runtime kg_sched->skg_runtime
150 #define kg_slptime kg_sched->skg_slptime
151
152 #define SLOT_RELEASE(kg) \
153 do { \
154 kg->kg_avail_opennings++; \
155 CTR3(KTR_RUNQ, "kg %p(%d) Slot released (->%d)", \
156 kg, \
157 kg->kg_concurrency, \
158 kg->kg_avail_opennings); \
159 /*KASSERT((kg->kg_avail_opennings <= kg->kg_concurrency), \
160 ("slots out of whack")); */ \
161 } while (0)
162
163 #define SLOT_USE(kg) \
164 do { \
165 kg->kg_avail_opennings--; \
166 CTR3(KTR_RUNQ, "kg %p(%d) Slot used (->%d)", \
167 kg, \
168 kg->kg_concurrency, \
169 kg->kg_avail_opennings); \
170 /*KASSERT((kg->kg_avail_opennings >= 0), \
171 ("slots out of whack"));*/ \
172 } while (0)
173
174 static struct kse kse0;
175 static struct kg_sched kg_sched0;
176
177 /*
178 * The priority is primarily determined by the interactivity score. Thus, we
179 * give lower(better) priorities to kse groups that use less CPU. The nice
180 * value is then directly added to this to allow nice to have some effect
181 * on latency.
182 *
183 * PRI_RANGE: Total priority range for timeshare threads.
184 * PRI_NRESV: Number of nice values.
185 * PRI_BASE: The start of the dynamic range.
186 */
187 #define SCHED_PRI_RANGE (PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE + 1)
188 #define SCHED_PRI_NRESV ((PRIO_MAX - PRIO_MIN) + 1)
189 #define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
190 #define SCHED_PRI_BASE (PRI_MIN_TIMESHARE)
191 #define SCHED_PRI_INTERACT(score) \
192 ((score) * SCHED_PRI_RANGE / SCHED_INTERACT_MAX)
193
194 /*
195 * These determine the interactivity of a process.
196 *
197 * SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
198 * before throttling back.
199 * SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
200 * INTERACT_MAX: Maximum interactivity value. Smaller is better.
201 * INTERACT_THRESH: Threshhold for placement on the current runq.
202 */
203 #define SCHED_SLP_RUN_MAX ((hz * 5) << 10)
204 #define SCHED_SLP_RUN_FORK ((hz / 2) << 10)
205 #define SCHED_INTERACT_MAX (100)
206 #define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
207 #define SCHED_INTERACT_THRESH (30)
208
209 /*
210 * These parameters and macros determine the size of the time slice that is
211 * granted to each thread.
212 *
213 * SLICE_MIN: Minimum time slice granted, in units of ticks.
214 * SLICE_MAX: Maximum time slice granted.
215 * SLICE_RANGE: Range of available time slices scaled by hz.
216 * SLICE_SCALE: The number slices granted per val in the range of [0, max].
217 * SLICE_NICE: Determine the amount of slice granted to a scaled nice.
218 * SLICE_NTHRESH: The nice cutoff point for slice assignment.
219 */
220 #define SCHED_SLICE_MIN (slice_min)
221 #define SCHED_SLICE_MAX (slice_max)
222 #define SCHED_SLICE_INTERACTIVE (slice_max)
223 #define SCHED_SLICE_NTHRESH (SCHED_PRI_NHALF - 1)
224 #define SCHED_SLICE_RANGE (SCHED_SLICE_MAX - SCHED_SLICE_MIN + 1)
225 #define SCHED_SLICE_SCALE(val, max) (((val) * SCHED_SLICE_RANGE) / (max))
226 #define SCHED_SLICE_NICE(nice) \
227 (SCHED_SLICE_MAX - SCHED_SLICE_SCALE((nice), SCHED_SLICE_NTHRESH))
228
229 /*
230 * This macro determines whether or not the thread belongs on the current or
231 * next run queue.
232 */
233 #define SCHED_INTERACTIVE(kg) \
234 (sched_interact_score(kg) < SCHED_INTERACT_THRESH)
235 #define SCHED_CURR(kg, ke) \
236 ((ke->ke_flags & KEF_PRIOELEV) || SCHED_INTERACTIVE(kg))
237
238 /*
239 * Cpu percentage computation macros and defines.
240 *
241 * SCHED_CPU_TIME: Number of seconds to average the cpu usage across.
242 * SCHED_CPU_TICKS: Number of hz ticks to average the cpu usage across.
243 */
244
245 #define SCHED_CPU_TIME 10
246 #define SCHED_CPU_TICKS (hz * SCHED_CPU_TIME)
247
248 /*
249 * kseq - per processor runqs and statistics.
250 */
251 struct kseq {
252 struct runq ksq_idle; /* Queue of IDLE threads. */
253 struct runq ksq_timeshare[2]; /* Run queues for !IDLE. */
254 struct runq *ksq_next; /* Next timeshare queue. */
255 struct runq *ksq_curr; /* Current queue. */
256 int ksq_load_timeshare; /* Load for timeshare. */
257 int ksq_load; /* Aggregate load. */
258 short ksq_nice[SCHED_PRI_NRESV]; /* KSEs in each nice bin. */
259 short ksq_nicemin; /* Least nice. */
260 #ifdef SMP
261 int ksq_transferable;
262 LIST_ENTRY(kseq) ksq_siblings; /* Next in kseq group. */
263 struct kseq_group *ksq_group; /* Our processor group. */
264 volatile struct kse *ksq_assigned; /* assigned by another CPU. */
265 #else
266 int ksq_sysload; /* For loadavg, !ITHD load. */
267 #endif
268 };
269
270 #ifdef SMP
271 /*
272 * kseq groups are groups of processors which can cheaply share threads. When
273 * one processor in the group goes idle it will check the runqs of the other
274 * processors in its group prior to halting and waiting for an interrupt.
275 * These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA.
276 * In a numa environment we'd want an idle bitmap per group and a two tiered
277 * load balancer.
278 */
279 struct kseq_group {
280 int ksg_cpus; /* Count of CPUs in this kseq group. */
281 cpumask_t ksg_cpumask; /* Mask of cpus in this group. */
282 cpumask_t ksg_idlemask; /* Idle cpus in this group. */
283 cpumask_t ksg_mask; /* Bit mask for first cpu. */
284 int ksg_load; /* Total load of this group. */
285 int ksg_transferable; /* Transferable load of this group. */
286 LIST_HEAD(, kseq) ksg_members; /* Linked list of all members. */
287 };
288 #endif
289
290 /*
291 * One kse queue per processor.
292 */
293 #ifdef SMP
294 static cpumask_t kseq_idle;
295 static int ksg_maxid;
296 static struct kseq kseq_cpu[MAXCPU];
297 static struct kseq_group kseq_groups[MAXCPU];
298 static int bal_tick;
299 static int gbal_tick;
300
301 #define KSEQ_SELF() (&kseq_cpu[PCPU_GET(cpuid)])
302 #define KSEQ_CPU(x) (&kseq_cpu[(x)])
303 #define KSEQ_ID(x) ((x) - kseq_cpu)
304 #define KSEQ_GROUP(x) (&kseq_groups[(x)])
305 #else /* !SMP */
306 static struct kseq kseq_cpu;
307
308 #define KSEQ_SELF() (&kseq_cpu)
309 #define KSEQ_CPU(x) (&kseq_cpu)
310 #endif
311
312 static void slot_fill(struct ksegrp *kg);
313 static struct kse *sched_choose(void); /* XXX Should be thread * */
314 static void sched_add_internal(struct thread *td, int preemptive);
315 static void sched_slice(struct kse *ke);
316 static void sched_priority(struct ksegrp *kg);
317 static int sched_interact_score(struct ksegrp *kg);
318 static void sched_interact_update(struct ksegrp *kg);
319 static void sched_interact_fork(struct ksegrp *kg);
320 static void sched_pctcpu_update(struct kse *ke);
321
322 /* Operations on per processor queues */
323 static struct kse * kseq_choose(struct kseq *kseq);
324 static void kseq_setup(struct kseq *kseq);
325 static void kseq_load_add(struct kseq *kseq, struct kse *ke);
326 static void kseq_load_rem(struct kseq *kseq, struct kse *ke);
327 static __inline void kseq_runq_add(struct kseq *kseq, struct kse *ke);
328 static __inline void kseq_runq_rem(struct kseq *kseq, struct kse *ke);
329 static void kseq_nice_add(struct kseq *kseq, int nice);
330 static void kseq_nice_rem(struct kseq *kseq, int nice);
331 void kseq_print(int cpu);
332 #ifdef SMP
333 static int kseq_transfer(struct kseq *ksq, struct kse *ke, int class);
334 static struct kse *runq_steal(struct runq *rq);
335 static void sched_balance(void);
336 static void sched_balance_groups(void);
337 static void sched_balance_group(struct kseq_group *ksg);
338 static void sched_balance_pair(struct kseq *high, struct kseq *low);
339 static void kseq_move(struct kseq *from, int cpu);
340 static int kseq_idled(struct kseq *kseq);
341 static void kseq_notify(struct kse *ke, int cpu);
342 static void kseq_assign(struct kseq *);
343 static struct kse *kseq_steal(struct kseq *kseq, int stealidle);
344 /*
345 * On P4 Xeons the round-robin interrupt delivery is broken. As a result of
346 * this, we can't pin interrupts to the cpu that they were delivered to,
347 * otherwise all ithreads only run on CPU 0.
348 */
349 #ifdef __i386__
350 #define KSE_CAN_MIGRATE(ke, class) \
351 ((ke)->ke_thread->td_pinned == 0 && ((ke)->ke_flags & KEF_BOUND) == 0)
352 #else /* !__i386__ */
353 #define KSE_CAN_MIGRATE(ke, class) \
354 ((class) != PRI_ITHD && (ke)->ke_thread->td_pinned == 0 && \
355 ((ke)->ke_flags & KEF_BOUND) == 0)
356 #endif /* !__i386__ */
357 #endif
358
359 void
360 kseq_print(int cpu)
361 {
362 struct kseq *kseq;
363 int i;
364
365 kseq = KSEQ_CPU(cpu);
366
367 printf("kseq:\n");
368 printf("\tload: %d\n", kseq->ksq_load);
369 printf("\tload TIMESHARE: %d\n", kseq->ksq_load_timeshare);
370 #ifdef SMP
371 printf("\tload transferable: %d\n", kseq->ksq_transferable);
372 #endif
373 printf("\tnicemin:\t%d\n", kseq->ksq_nicemin);
374 printf("\tnice counts:\n");
375 for (i = 0; i < SCHED_PRI_NRESV; i++)
376 if (kseq->ksq_nice[i])
377 printf("\t\t%d = %d\n",
378 i - SCHED_PRI_NHALF, kseq->ksq_nice[i]);
379 }
380
381 static __inline void
382 kseq_runq_add(struct kseq *kseq, struct kse *ke)
383 {
384 #ifdef SMP
385 if (KSE_CAN_MIGRATE(ke, PRI_BASE(ke->ke_ksegrp->kg_pri_class))) {
386 kseq->ksq_transferable++;
387 kseq->ksq_group->ksg_transferable++;
388 ke->ke_flags |= KEF_XFERABLE;
389 }
390 #endif
391 runq_add(ke->ke_runq, ke, 0);
392 }
393
394 static __inline void
395 kseq_runq_rem(struct kseq *kseq, struct kse *ke)
396 {
397 #ifdef SMP
398 if (ke->ke_flags & KEF_XFERABLE) {
399 kseq->ksq_transferable--;
400 kseq->ksq_group->ksg_transferable--;
401 ke->ke_flags &= ~KEF_XFERABLE;
402 }
403 #endif
404 runq_remove(ke->ke_runq, ke);
405 }
406
407 static void
408 kseq_load_add(struct kseq *kseq, struct kse *ke)
409 {
410 int class;
411 mtx_assert(&sched_lock, MA_OWNED);
412 class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
413 if (class == PRI_TIMESHARE)
414 kseq->ksq_load_timeshare++;
415 kseq->ksq_load++;
416 CTR1(KTR_SCHED, "load: %d", kseq->ksq_load);
417 if (class != PRI_ITHD && (ke->ke_proc->p_flag & P_NOLOAD) == 0)
418 #ifdef SMP
419 kseq->ksq_group->ksg_load++;
420 #else
421 kseq->ksq_sysload++;
422 #endif
423 if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE)
424 kseq_nice_add(kseq, ke->ke_proc->p_nice);
425 }
426
427 static void
428 kseq_load_rem(struct kseq *kseq, struct kse *ke)
429 {
430 int class;
431 mtx_assert(&sched_lock, MA_OWNED);
432 class = PRI_BASE(ke->ke_ksegrp->kg_pri_class);
433 if (class == PRI_TIMESHARE)
434 kseq->ksq_load_timeshare--;
435 if (class != PRI_ITHD && (ke->ke_proc->p_flag & P_NOLOAD) == 0)
436 #ifdef SMP
437 kseq->ksq_group->ksg_load--;
438 #else
439 kseq->ksq_sysload--;
440 #endif
441 kseq->ksq_load--;
442 CTR1(KTR_SCHED, "load: %d", kseq->ksq_load);
443 ke->ke_runq = NULL;
444 if (ke->ke_ksegrp->kg_pri_class == PRI_TIMESHARE)
445 kseq_nice_rem(kseq, ke->ke_proc->p_nice);
446 }
447
448 static void
449 kseq_nice_add(struct kseq *kseq, int nice)
450 {
451 mtx_assert(&sched_lock, MA_OWNED);
452 /* Normalize to zero. */
453 kseq->ksq_nice[nice + SCHED_PRI_NHALF]++;
454 if (nice < kseq->ksq_nicemin || kseq->ksq_load_timeshare == 1)
455 kseq->ksq_nicemin = nice;
456 }
457
458 static void
459 kseq_nice_rem(struct kseq *kseq, int nice)
460 {
461 int n;
462
463 mtx_assert(&sched_lock, MA_OWNED);
464 /* Normalize to zero. */
465 n = nice + SCHED_PRI_NHALF;
466 kseq->ksq_nice[n]--;
467 KASSERT(kseq->ksq_nice[n] >= 0, ("Negative nice count."));
468
469 /*
470 * If this wasn't the smallest nice value or there are more in
471 * this bucket we can just return. Otherwise we have to recalculate
472 * the smallest nice.
473 */
474 if (nice != kseq->ksq_nicemin ||
475 kseq->ksq_nice[n] != 0 ||
476 kseq->ksq_load_timeshare == 0)
477 return;
478
479 for (; n < SCHED_PRI_NRESV; n++)
480 if (kseq->ksq_nice[n]) {
481 kseq->ksq_nicemin = n - SCHED_PRI_NHALF;
482 return;
483 }
484 }
485
486 #ifdef SMP
487 /*
488 * sched_balance is a simple CPU load balancing algorithm. It operates by
489 * finding the least loaded and most loaded cpu and equalizing their load
490 * by migrating some processes.
491 *
492 * Dealing only with two CPUs at a time has two advantages. Firstly, most
493 * installations will only have 2 cpus. Secondly, load balancing too much at
494 * once can have an unpleasant effect on the system. The scheduler rarely has
495 * enough information to make perfect decisions. So this algorithm chooses
496 * algorithm simplicity and more gradual effects on load in larger systems.
497 *
498 * It could be improved by considering the priorities and slices assigned to
499 * each task prior to balancing them. There are many pathological cases with
500 * any approach and so the semi random algorithm below may work as well as any.
501 *
502 */
503 static void
504 sched_balance(void)
505 {
506 struct kseq_group *high;
507 struct kseq_group *low;
508 struct kseq_group *ksg;
509 int cnt;
510 int i;
511
512 if (smp_started == 0)
513 goto out;
514 low = high = NULL;
515 i = random() % (ksg_maxid + 1);
516 for (cnt = 0; cnt <= ksg_maxid; cnt++) {
517 ksg = KSEQ_GROUP(i);
518 /*
519 * Find the CPU with the highest load that has some
520 * threads to transfer.
521 */
522 if ((high == NULL || ksg->ksg_load > high->ksg_load)
523 && ksg->ksg_transferable)
524 high = ksg;
525 if (low == NULL || ksg->ksg_load < low->ksg_load)
526 low = ksg;
527 if (++i > ksg_maxid)
528 i = 0;
529 }
530 if (low != NULL && high != NULL && high != low)
531 sched_balance_pair(LIST_FIRST(&high->ksg_members),
532 LIST_FIRST(&low->ksg_members));
533 out:
534 bal_tick = ticks + (random() % (hz * 2));
535 }
536
537 static void
538 sched_balance_groups(void)
539 {
540 int i;
541
542 mtx_assert(&sched_lock, MA_OWNED);
543 if (smp_started)
544 for (i = 0; i <= ksg_maxid; i++)
545 sched_balance_group(KSEQ_GROUP(i));
546 gbal_tick = ticks + (random() % (hz * 2));
547 }
548
549 static void
550 sched_balance_group(struct kseq_group *ksg)
551 {
552 struct kseq *kseq;
553 struct kseq *high;
554 struct kseq *low;
555 int load;
556
557 if (ksg->ksg_transferable == 0)
558 return;
559 low = NULL;
560 high = NULL;
561 LIST_FOREACH(kseq, &ksg->ksg_members, ksq_siblings) {
562 load = kseq->ksq_load;
563 if (high == NULL || load > high->ksq_load)
564 high = kseq;
565 if (low == NULL || load < low->ksq_load)
566 low = kseq;
567 }
568 if (high != NULL && low != NULL && high != low)
569 sched_balance_pair(high, low);
570 }
571
572 static void
573 sched_balance_pair(struct kseq *high, struct kseq *low)
574 {
575 int transferable;
576 int high_load;
577 int low_load;
578 int move;
579 int diff;
580 int i;
581
582 /*
583 * If we're transfering within a group we have to use this specific
584 * kseq's transferable count, otherwise we can steal from other members
585 * of the group.
586 */
587 if (high->ksq_group == low->ksq_group) {
588 transferable = high->ksq_transferable;
589 high_load = high->ksq_load;
590 low_load = low->ksq_load;
591 } else {
592 transferable = high->ksq_group->ksg_transferable;
593 high_load = high->ksq_group->ksg_load;
594 low_load = low->ksq_group->ksg_load;
595 }
596 if (transferable == 0)
597 return;
598 /*
599 * Determine what the imbalance is and then adjust that to how many
600 * kses we actually have to give up (transferable).
601 */
602 diff = high_load - low_load;
603 move = diff / 2;
604 if (diff & 0x1)
605 move++;
606 move = min(move, transferable);
607 for (i = 0; i < move; i++)
608 kseq_move(high, KSEQ_ID(low));
609 return;
610 }
611
612 static void
613 kseq_move(struct kseq *from, int cpu)
614 {
615 struct kseq *kseq;
616 struct kseq *to;
617 struct kse *ke;
618
619 kseq = from;
620 to = KSEQ_CPU(cpu);
621 ke = kseq_steal(kseq, 1);
622 if (ke == NULL) {
623 struct kseq_group *ksg;
624
625 ksg = kseq->ksq_group;
626 LIST_FOREACH(kseq, &ksg->ksg_members, ksq_siblings) {
627 if (kseq == from || kseq->ksq_transferable == 0)
628 continue;
629 ke = kseq_steal(kseq, 1);
630 break;
631 }
632 if (ke == NULL)
633 panic("kseq_move: No KSEs available with a "
634 "transferable count of %d\n",
635 ksg->ksg_transferable);
636 }
637 if (kseq == to)
638 return;
639 ke->ke_state = KES_THREAD;
640 kseq_runq_rem(kseq, ke);
641 kseq_load_rem(kseq, ke);
642 kseq_notify(ke, cpu);
643 }
644
645 static int
646 kseq_idled(struct kseq *kseq)
647 {
648 struct kseq_group *ksg;
649 struct kseq *steal;
650 struct kse *ke;
651
652 ksg = kseq->ksq_group;
653 /*
654 * If we're in a cpu group, try and steal kses from another cpu in
655 * the group before idling.
656 */
657 if (ksg->ksg_cpus > 1 && ksg->ksg_transferable) {
658 LIST_FOREACH(steal, &ksg->ksg_members, ksq_siblings) {
659 if (steal == kseq || steal->ksq_transferable == 0)
660 continue;
661 ke = kseq_steal(steal, 0);
662 if (ke == NULL)
663 continue;
664 ke->ke_state = KES_THREAD;
665 kseq_runq_rem(steal, ke);
666 kseq_load_rem(steal, ke);
667 ke->ke_cpu = PCPU_GET(cpuid);
668 sched_add_internal(ke->ke_thread, 0);
669 return (0);
670 }
671 }
672 /*
673 * We only set the idled bit when all of the cpus in the group are
674 * idle. Otherwise we could get into a situation where a KSE bounces
675 * back and forth between two idle cores on seperate physical CPUs.
676 */
677 ksg->ksg_idlemask |= PCPU_GET(cpumask);
678 if (ksg->ksg_idlemask != ksg->ksg_cpumask)
679 return (1);
680 atomic_set_int(&kseq_idle, ksg->ksg_mask);
681 return (1);
682 }
683
684 static void
685 kseq_assign(struct kseq *kseq)
686 {
687 struct kse *nke;
688 struct kse *ke;
689
690 do {
691 *(volatile struct kse **)&ke = kseq->ksq_assigned;
692 } while(!atomic_cmpset_ptr(&kseq->ksq_assigned, ke, NULL));
693 for (; ke != NULL; ke = nke) {
694 nke = ke->ke_assign;
695 ke->ke_flags &= ~KEF_ASSIGNED;
696 SLOT_RELEASE(ke->ke_thread->td_ksegrp);
697 sched_add_internal(ke->ke_thread, 0);
698 }
699 }
700
701 static void
702 kseq_notify(struct kse *ke, int cpu)
703 {
704 struct kseq *kseq;
705 struct thread *td;
706 struct pcpu *pcpu;
707 int prio;
708
709 ke->ke_cpu = cpu;
710 ke->ke_flags |= KEF_ASSIGNED;
711 SLOT_USE(ke->ke_thread->td_ksegrp);
712 prio = ke->ke_thread->td_priority;
713
714 kseq = KSEQ_CPU(cpu);
715
716 /*
717 * Place a KSE on another cpu's queue and force a resched.
718 */
719 do {
720 *(volatile struct kse **)&ke->ke_assign = kseq->ksq_assigned;
721 } while(!atomic_cmpset_ptr(&kseq->ksq_assigned, ke->ke_assign, ke));
722 /*
723 * Without sched_lock we could lose a race where we set NEEDRESCHED
724 * on a thread that is switched out before the IPI is delivered. This
725 * would lead us to miss the resched. This will be a problem once
726 * sched_lock is pushed down.
727 */
728 pcpu = pcpu_find(cpu);
729 td = pcpu->pc_curthread;
730 if (ke->ke_thread->td_priority < td->td_priority ||
731 td == pcpu->pc_idlethread) {
732 td->td_flags |= TDF_NEEDRESCHED;
733 ipi_selected(1 << cpu, IPI_AST);
734 }
735 }
736
737 static struct kse *
738 runq_steal(struct runq *rq)
739 {
740 struct rqhead *rqh;
741 struct rqbits *rqb;
742 struct kse *ke;
743 int word;
744 int bit;
745
746 mtx_assert(&sched_lock, MA_OWNED);
747 rqb = &rq->rq_status;
748 for (word = 0; word < RQB_LEN; word++) {
749 if (rqb->rqb_bits[word] == 0)
750 continue;
751 for (bit = 0; bit < RQB_BPW; bit++) {
752 if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
753 continue;
754 rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
755 TAILQ_FOREACH(ke, rqh, ke_procq) {
756 if (KSE_CAN_MIGRATE(ke,
757 PRI_BASE(ke->ke_ksegrp->kg_pri_class)))
758 return (ke);
759 }
760 }
761 }
762 return (NULL);
763 }
764
765 static struct kse *
766 kseq_steal(struct kseq *kseq, int stealidle)
767 {
768 struct kse *ke;
769
770 /*
771 * Steal from next first to try to get a non-interactive task that
772 * may not have run for a while.
773 */
774 if ((ke = runq_steal(kseq->ksq_next)) != NULL)
775 return (ke);
776 if ((ke = runq_steal(kseq->ksq_curr)) != NULL)
777 return (ke);
778 if (stealidle)
779 return (runq_steal(&kseq->ksq_idle));
780 return (NULL);
781 }
782
783 int
784 kseq_transfer(struct kseq *kseq, struct kse *ke, int class)
785 {
786 struct kseq_group *ksg;
787 int cpu;
788
789 if (smp_started == 0)
790 return (0);
791 cpu = 0;
792 /*
793 * If our load exceeds a certain threshold we should attempt to
794 * reassign this thread. The first candidate is the cpu that
795 * originally ran the thread. If it is idle, assign it there,
796 * otherwise, pick an idle cpu.
797 *
798 * The threshold at which we start to reassign kses has a large impact
799 * on the overall performance of the system. Tuned too high and
800 * some CPUs may idle. Too low and there will be excess migration
801 * and context switches.
802 */
803 ksg = kseq->ksq_group;
804 if (ksg->ksg_load > ksg->ksg_cpus && kseq_idle) {
805 ksg = KSEQ_CPU(ke->ke_cpu)->ksq_group;
806 if (kseq_idle & ksg->ksg_mask) {
807 cpu = ffs(ksg->ksg_idlemask);
808 if (cpu)
809 goto migrate;
810 }
811 /*
812 * Multiple cpus could find this bit simultaneously
813 * but the race shouldn't be terrible.
814 */
815 cpu = ffs(kseq_idle);
816 if (cpu)
817 goto migrate;
818 }
819 /*
820 * If another cpu in this group has idled, assign a thread over
821 * to them after checking to see if there are idled groups.
822 */
823 ksg = kseq->ksq_group;
824 if (ksg->ksg_idlemask) {
825 cpu = ffs(ksg->ksg_idlemask);
826 if (cpu)
827 goto migrate;
828 }
829 /*
830 * No new CPU was found.
831 */
832 return (0);
833 migrate:
834 /*
835 * Now that we've found an idle CPU, migrate the thread.
836 */
837 cpu--;
838 ke->ke_runq = NULL;
839 kseq_notify(ke, cpu);
840
841 return (1);
842 }
843
844 #endif /* SMP */
845
846 /*
847 * Pick the highest priority task we have and return it.
848 */
849
850 static struct kse *
851 kseq_choose(struct kseq *kseq)
852 {
853 struct runq *swap;
854 struct kse *ke;
855 int nice;
856
857 mtx_assert(&sched_lock, MA_OWNED);
858 swap = NULL;
859
860 for (;;) {
861 ke = runq_choose(kseq->ksq_curr);
862 if (ke == NULL) {
863 /*
864 * We already swapped once and didn't get anywhere.
865 */
866 if (swap)
867 break;
868 swap = kseq->ksq_curr;
869 kseq->ksq_curr = kseq->ksq_next;
870 kseq->ksq_next = swap;
871 continue;
872 }
873 /*
874 * If we encounter a slice of 0 the kse is in a
875 * TIMESHARE kse group and its nice was too far out
876 * of the range that receives slices.
877 */
878 nice = ke->ke_proc->p_nice + (0 - kseq->ksq_nicemin);
879 if (ke->ke_slice == 0 || (nice > SCHED_SLICE_NTHRESH &&
880 ke->ke_proc->p_nice != 0)) {
881 runq_remove(ke->ke_runq, ke);
882 sched_slice(ke);
883 ke->ke_runq = kseq->ksq_next;
884 runq_add(ke->ke_runq, ke, 0);
885 continue;
886 }
887 return (ke);
888 }
889
890 return (runq_choose(&kseq->ksq_idle));
891 }
892
893 static void
894 kseq_setup(struct kseq *kseq)
895 {
896 runq_init(&kseq->ksq_timeshare[0]);
897 runq_init(&kseq->ksq_timeshare[1]);
898 runq_init(&kseq->ksq_idle);
899 kseq->ksq_curr = &kseq->ksq_timeshare[0];
900 kseq->ksq_next = &kseq->ksq_timeshare[1];
901 kseq->ksq_load = 0;
902 kseq->ksq_load_timeshare = 0;
903 }
904
905 static void
906 sched_setup(void *dummy)
907 {
908 #ifdef SMP
909 int balance_groups;
910 int i;
911 #endif
912
913 slice_min = (hz/100); /* 10ms */
914 slice_max = (hz/7); /* ~140ms */
915
916 #ifdef SMP
917 balance_groups = 0;
918 /*
919 * Initialize the kseqs.
920 */
921 for (i = 0; i < MAXCPU; i++) {
922 struct kseq *ksq;
923
924 ksq = &kseq_cpu[i];
925 ksq->ksq_assigned = NULL;
926 kseq_setup(&kseq_cpu[i]);
927 }
928 if (smp_topology == NULL) {
929 struct kseq_group *ksg;
930 struct kseq *ksq;
931
932 for (i = 0; i < MAXCPU; i++) {
933 ksq = &kseq_cpu[i];
934 ksg = &kseq_groups[i];
935 /*
936 * Setup a kseq group with one member.
937 */
938 ksq->ksq_transferable = 0;
939 ksq->ksq_group = ksg;
940 ksg->ksg_cpus = 1;
941 ksg->ksg_idlemask = 0;
942 ksg->ksg_cpumask = ksg->ksg_mask = 1 << i;
943 ksg->ksg_load = 0;
944 ksg->ksg_transferable = 0;
945 LIST_INIT(&ksg->ksg_members);
946 LIST_INSERT_HEAD(&ksg->ksg_members, ksq, ksq_siblings);
947 }
948 } else {
949 struct kseq_group *ksg;
950 struct cpu_group *cg;
951 int j;
952
953 for (i = 0; i < smp_topology->ct_count; i++) {
954 cg = &smp_topology->ct_group[i];
955 ksg = &kseq_groups[i];
956 /*
957 * Initialize the group.
958 */
959 ksg->ksg_idlemask = 0;
960 ksg->ksg_load = 0;
961 ksg->ksg_transferable = 0;
962 ksg->ksg_cpus = cg->cg_count;
963 ksg->ksg_cpumask = cg->cg_mask;
964 LIST_INIT(&ksg->ksg_members);
965 /*
966 * Find all of the group members and add them.
967 */
968 for (j = 0; j < MAXCPU; j++) {
969 if ((cg->cg_mask & (1 << j)) != 0) {
970 if (ksg->ksg_mask == 0)
971 ksg->ksg_mask = 1 << j;
972 kseq_cpu[j].ksq_transferable = 0;
973 kseq_cpu[j].ksq_group = ksg;
974 LIST_INSERT_HEAD(&ksg->ksg_members,
975 &kseq_cpu[j], ksq_siblings);
976 }
977 }
978 if (ksg->ksg_cpus > 1)
979 balance_groups = 1;
980 }
981 ksg_maxid = smp_topology->ct_count - 1;
982 }
983 /*
984 * Stagger the group and global load balancer so they do not
985 * interfere with each other.
986 */
987 bal_tick = ticks + hz;
988 if (balance_groups)
989 gbal_tick = ticks + (hz / 2);
990 #else
991 kseq_setup(KSEQ_SELF());
992 #endif
993 mtx_lock_spin(&sched_lock);
994 kseq_load_add(KSEQ_SELF(), &kse0);
995 mtx_unlock_spin(&sched_lock);
996 }
997
998 /*
999 * Scale the scheduling priority according to the "interactivity" of this
1000 * process.
1001 */
1002 static void
1003 sched_priority(struct ksegrp *kg)
1004 {
1005 int pri;
1006
1007 if (kg->kg_pri_class != PRI_TIMESHARE)
1008 return;
1009
1010 pri = SCHED_PRI_INTERACT(sched_interact_score(kg));
1011 pri += SCHED_PRI_BASE;
1012 pri += kg->kg_proc->p_nice;
1013
1014 if (pri > PRI_MAX_TIMESHARE)
1015 pri = PRI_MAX_TIMESHARE;
1016 else if (pri < PRI_MIN_TIMESHARE)
1017 pri = PRI_MIN_TIMESHARE;
1018
1019 kg->kg_user_pri = pri;
1020
1021 return;
1022 }
1023
1024 /*
1025 * Calculate a time slice based on the properties of the kseg and the runq
1026 * that we're on. This is only for PRI_TIMESHARE ksegrps.
1027 */
1028 static void
1029 sched_slice(struct kse *ke)
1030 {
1031 struct kseq *kseq;
1032 struct ksegrp *kg;
1033
1034 kg = ke->ke_ksegrp;
1035 kseq = KSEQ_CPU(ke->ke_cpu);
1036
1037 if (ke->ke_flags & KEF_PRIOELEV) {
1038 ke->ke_slice = SCHED_SLICE_MIN;
1039 return;
1040 }
1041
1042 /*
1043 * Rationale:
1044 * KSEs in interactive ksegs get a minimal slice so that we
1045 * quickly notice if it abuses its advantage.
1046 *
1047 * KSEs in non-interactive ksegs are assigned a slice that is
1048 * based on the ksegs nice value relative to the least nice kseg
1049 * on the run queue for this cpu.
1050 *
1051 * If the KSE is less nice than all others it gets the maximum
1052 * slice and other KSEs will adjust their slice relative to
1053 * this when they first expire.
1054 *
1055 * There is 20 point window that starts relative to the least
1056 * nice kse on the run queue. Slice size is determined by
1057 * the kse distance from the last nice ksegrp.
1058 *
1059 * If the kse is outside of the window it will get no slice
1060 * and will be reevaluated each time it is selected on the
1061 * run queue. The exception to this is nice 0 ksegs when
1062 * a nice -20 is running. They are always granted a minimum
1063 * slice.
1064 */
1065 if (!SCHED_INTERACTIVE(kg)) {
1066 int nice;
1067
1068 nice = kg->kg_proc->p_nice + (0 - kseq->ksq_nicemin);
1069 if (kseq->ksq_load_timeshare == 0 ||
1070 kg->kg_proc->p_nice < kseq->ksq_nicemin)
1071 ke->ke_slice = SCHED_SLICE_MAX;
1072 else if (nice <= SCHED_SLICE_NTHRESH)
1073 ke->ke_slice = SCHED_SLICE_NICE(nice);
1074 else if (kg->kg_proc->p_nice == 0)
1075 ke->ke_slice = SCHED_SLICE_MIN;
1076 else
1077 ke->ke_slice = 0;
1078 } else
1079 ke->ke_slice = SCHED_SLICE_INTERACTIVE;
1080
1081 return;
1082 }
1083
1084 /*
1085 * This routine enforces a maximum limit on the amount of scheduling history
1086 * kept. It is called after either the slptime or runtime is adjusted.
1087 * This routine will not operate correctly when slp or run times have been
1088 * adjusted to more than double their maximum.
1089 */
1090 static void
1091 sched_interact_update(struct ksegrp *kg)
1092 {
1093 int sum;
1094
1095 sum = kg->kg_runtime + kg->kg_slptime;
1096 if (sum < SCHED_SLP_RUN_MAX)
1097 return;
1098 /*
1099 * If we have exceeded by more than 1/5th then the algorithm below
1100 * will not bring us back into range. Dividing by two here forces
1101 * us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
1102 */
1103 if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
1104 kg->kg_runtime /= 2;
1105 kg->kg_slptime /= 2;
1106 return;
1107 }
1108 kg->kg_runtime = (kg->kg_runtime / 5) * 4;
1109 kg->kg_slptime = (kg->kg_slptime / 5) * 4;
1110 }
1111
1112 static void
1113 sched_interact_fork(struct ksegrp *kg)
1114 {
1115 int ratio;
1116 int sum;
1117
1118 sum = kg->kg_runtime + kg->kg_slptime;
1119 if (sum > SCHED_SLP_RUN_FORK) {
1120 ratio = sum / SCHED_SLP_RUN_FORK;
1121 kg->kg_runtime /= ratio;
1122 kg->kg_slptime /= ratio;
1123 }
1124 }
1125
1126 static int
1127 sched_interact_score(struct ksegrp *kg)
1128 {
1129 int div;
1130
1131 if (kg->kg_runtime > kg->kg_slptime) {
1132 div = max(1, kg->kg_runtime / SCHED_INTERACT_HALF);
1133 return (SCHED_INTERACT_HALF +
1134 (SCHED_INTERACT_HALF - (kg->kg_slptime / div)));
1135 } if (kg->kg_slptime > kg->kg_runtime) {
1136 div = max(1, kg->kg_slptime / SCHED_INTERACT_HALF);
1137 return (kg->kg_runtime / div);
1138 }
1139
1140 /*
1141 * This can happen if slptime and runtime are 0.
1142 */
1143 return (0);
1144
1145 }
1146
1147 /*
1148 * Very early in the boot some setup of scheduler-specific
1149 * parts of proc0 and of soem scheduler resources needs to be done.
1150 * Called from:
1151 * proc0_init()
1152 */
1153 void
1154 schedinit(void)
1155 {
1156 /*
1157 * Set up the scheduler specific parts of proc0.
1158 */
1159 proc0.p_sched = NULL; /* XXX */
1160 ksegrp0.kg_sched = &kg_sched0;
1161 thread0.td_sched = &kse0;
1162 kse0.ke_thread = &thread0;
1163 kse0.ke_state = KES_THREAD;
1164 kg_sched0.skg_concurrency = 1;
1165 kg_sched0.skg_avail_opennings = 0; /* we are already running */
1166 }
1167
1168 /*
1169 * This is only somewhat accurate since given many processes of the same
1170 * priority they will switch when their slices run out, which will be
1171 * at most SCHED_SLICE_MAX.
1172 */
1173 int
1174 sched_rr_interval(void)
1175 {
1176 return (SCHED_SLICE_MAX);
1177 }
1178
1179 static void
1180 sched_pctcpu_update(struct kse *ke)
1181 {
1182 /*
1183 * Adjust counters and watermark for pctcpu calc.
1184 */
1185 if (ke->ke_ltick > ticks - SCHED_CPU_TICKS) {
1186 /*
1187 * Shift the tick count out so that the divide doesn't
1188 * round away our results.
1189 */
1190 ke->ke_ticks <<= 10;
1191 ke->ke_ticks = (ke->ke_ticks / (ticks - ke->ke_ftick)) *
1192 SCHED_CPU_TICKS;
1193 ke->ke_ticks >>= 10;
1194 } else
1195 ke->ke_ticks = 0;
1196 ke->ke_ltick = ticks;
1197 ke->ke_ftick = ke->ke_ltick - SCHED_CPU_TICKS;
1198 }
1199
1200 void
1201 sched_prio(struct thread *td, u_char prio)
1202 {
1203 struct kse *ke;
1204
1205 CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)",
1206 td, td->td_proc->p_comm, td->td_priority, prio, curthread,
1207 curthread->td_proc->p_comm);
1208 ke = td->td_kse;
1209 mtx_assert(&sched_lock, MA_OWNED);
1210 if (TD_ON_RUNQ(td)) {
1211 /*
1212 * If the priority has been elevated due to priority
1213 * propagation, we may have to move ourselves to a new
1214 * queue. We still call adjustrunqueue below in case kse
1215 * needs to fix things up.
1216 */
1217 if (prio < td->td_priority && ke->ke_runq != NULL &&
1218 (ke->ke_flags & KEF_ASSIGNED) == 0 &&
1219 ke->ke_runq != KSEQ_CPU(ke->ke_cpu)->ksq_curr) {
1220 runq_remove(ke->ke_runq, ke);
1221 ke->ke_runq = KSEQ_CPU(ke->ke_cpu)->ksq_curr;
1222 runq_add(ke->ke_runq, ke, 0);
1223 }
1224 if (prio < td->td_priority)
1225 ke->ke_flags |= KEF_PRIOELEV;
1226 /*
1227 * Hold this kse on this cpu so that sched_prio() doesn't
1228 * cause excessive migration. We only want migration to
1229 * happen as the result of a wakeup.
1230 */
1231 ke->ke_flags |= KEF_HOLD;
1232 adjustrunqueue(td, prio);
1233 } else
1234 td->td_priority = prio;
1235 }
1236
1237 void
1238 sched_switch(struct thread *td, struct thread *newtd, int flags)
1239 {
1240 struct kse *ke;
1241
1242 mtx_assert(&sched_lock, MA_OWNED);
1243
1244 ke = td->td_kse;
1245
1246 td->td_lastcpu = td->td_oncpu;
1247 td->td_oncpu = NOCPU;
1248 td->td_flags &= ~TDF_NEEDRESCHED;
1249 td->td_pflags &= ~TDP_OWEPREEMPT;
1250
1251 /*
1252 * If the KSE has been assigned it may be in the process of switching
1253 * to the new cpu. This is the case in sched_bind().
1254 */
1255 if ((ke->ke_flags & KEF_ASSIGNED) == 0) {
1256 if (td == PCPU_GET(idlethread)) {
1257 TD_SET_CAN_RUN(td);
1258 } else {
1259 /* We are ending our run so make our slot available again */
1260 SLOT_RELEASE(td->td_ksegrp);
1261 if (TD_IS_RUNNING(td)) {
1262 kseq_load_rem(KSEQ_CPU(ke->ke_cpu), ke);
1263 /*
1264 * Don't allow the thread to migrate
1265 * from a preemption.
1266 */
1267 ke->ke_flags |= KEF_HOLD;
1268 setrunqueue(td, SRQ_OURSELF|SRQ_YIELDING);
1269 } else {
1270 if (ke->ke_runq) {
1271 kseq_load_rem(KSEQ_CPU(ke->ke_cpu), ke);
1272 } else if ((td->td_flags & TDF_IDLETD) == 0)
1273 kdb_backtrace();
1274 /*
1275 * We will not be on the run queue.
1276 * So we must be sleeping or similar.
1277 * Don't use the slot if we will need it
1278 * for newtd.
1279 */
1280 if ((td->td_proc->p_flag & P_HADTHREADS) &&
1281 (newtd == NULL ||
1282 newtd->td_ksegrp != td->td_ksegrp))
1283 slot_fill(td->td_ksegrp);
1284 }
1285 }
1286 }
1287 if (newtd != NULL) {
1288 /*
1289 * If we bring in a thread,
1290 * then account for it as if it had been added to the
1291 * run queue and then chosen.
1292 */
1293 newtd->td_kse->ke_flags |= KEF_DIDRUN;
1294 SLOT_USE(newtd->td_ksegrp);
1295 TD_SET_RUNNING(newtd);
1296 kseq_load_add(KSEQ_SELF(), newtd->td_kse);
1297 } else
1298 newtd = choosethread();
1299 if (td != newtd)
1300 cpu_switch(td, newtd);
1301 sched_lock.mtx_lock = (uintptr_t)td;
1302
1303 td->td_oncpu = PCPU_GET(cpuid);
1304 }
1305
1306 void
1307 sched_nice(struct proc *p, int nice)
1308 {
1309 struct ksegrp *kg;
1310 struct kse *ke;
1311 struct thread *td;
1312 struct kseq *kseq;
1313
1314 PROC_LOCK_ASSERT(p, MA_OWNED);
1315 mtx_assert(&sched_lock, MA_OWNED);
1316 /*
1317 * We need to adjust the nice counts for running KSEs.
1318 */
1319 FOREACH_KSEGRP_IN_PROC(p, kg) {
1320 if (kg->kg_pri_class == PRI_TIMESHARE) {
1321 FOREACH_THREAD_IN_GROUP(kg, td) {
1322 ke = td->td_kse;
1323 if (ke->ke_runq == NULL)
1324 continue;
1325 kseq = KSEQ_CPU(ke->ke_cpu);
1326 kseq_nice_rem(kseq, p->p_nice);
1327 kseq_nice_add(kseq, nice);
1328 }
1329 }
1330 }
1331 p->p_nice = nice;
1332 FOREACH_KSEGRP_IN_PROC(p, kg) {
1333 sched_priority(kg);
1334 FOREACH_THREAD_IN_GROUP(kg, td)
1335 td->td_flags |= TDF_NEEDRESCHED;
1336 }
1337 }
1338
1339 void
1340 sched_sleep(struct thread *td)
1341 {
1342 mtx_assert(&sched_lock, MA_OWNED);
1343
1344 td->td_slptime = ticks;
1345 td->td_base_pri = td->td_priority;
1346 }
1347
1348 void
1349 sched_wakeup(struct thread *td)
1350 {
1351 mtx_assert(&sched_lock, MA_OWNED);
1352
1353 /*
1354 * Let the kseg know how long we slept for. This is because process
1355 * interactivity behavior is modeled in the kseg.
1356 */
1357 if (td->td_slptime) {
1358 struct ksegrp *kg;
1359 int hzticks;
1360
1361 kg = td->td_ksegrp;
1362 hzticks = (ticks - td->td_slptime) << 10;
1363 if (hzticks >= SCHED_SLP_RUN_MAX) {
1364 kg->kg_slptime = SCHED_SLP_RUN_MAX;
1365 kg->kg_runtime = 1;
1366 } else {
1367 kg->kg_slptime += hzticks;
1368 sched_interact_update(kg);
1369 }
1370 sched_priority(kg);
1371 sched_slice(td->td_kse);
1372 td->td_slptime = 0;
1373 }
1374 setrunqueue(td, SRQ_BORING);
1375 }
1376
1377 /*
1378 * Penalize the parent for creating a new child and initialize the child's
1379 * priority.
1380 */
1381 void
1382 sched_fork(struct thread *td, struct thread *childtd)
1383 {
1384
1385 mtx_assert(&sched_lock, MA_OWNED);
1386
1387 sched_fork_ksegrp(td, childtd->td_ksegrp);
1388 sched_fork_thread(td, childtd);
1389 }
1390
1391 void
1392 sched_fork_ksegrp(struct thread *td, struct ksegrp *child)
1393 {
1394 struct ksegrp *kg = td->td_ksegrp;
1395 mtx_assert(&sched_lock, MA_OWNED);
1396
1397 child->kg_slptime = kg->kg_slptime;
1398 child->kg_runtime = kg->kg_runtime;
1399 child->kg_user_pri = kg->kg_user_pri;
1400 sched_interact_fork(child);
1401 kg->kg_runtime += tickincr << 10;
1402 sched_interact_update(kg);
1403 }
1404
1405 void
1406 sched_fork_thread(struct thread *td, struct thread *child)
1407 {
1408 struct kse *ke;
1409 struct kse *ke2;
1410
1411 sched_newthread(child);
1412 ke = td->td_kse;
1413 ke2 = child->td_kse;
1414 ke2->ke_slice = 1; /* Attempt to quickly learn interactivity. */
1415 ke2->ke_cpu = ke->ke_cpu;
1416 ke2->ke_runq = NULL;
1417
1418 /* Grab our parents cpu estimation information. */
1419 ke2->ke_ticks = ke->ke_ticks;
1420 ke2->ke_ltick = ke->ke_ltick;
1421 ke2->ke_ftick = ke->ke_ftick;
1422 }
1423
1424 void
1425 sched_class(struct ksegrp *kg, int class)
1426 {
1427 struct kseq *kseq;
1428 struct kse *ke;
1429 struct thread *td;
1430 int nclass;
1431 int oclass;
1432
1433 mtx_assert(&sched_lock, MA_OWNED);
1434 if (kg->kg_pri_class == class)
1435 return;
1436
1437 nclass = PRI_BASE(class);
1438 oclass = PRI_BASE(kg->kg_pri_class);
1439 FOREACH_THREAD_IN_GROUP(kg, td) {
1440 ke = td->td_kse;
1441 if ((ke->ke_state != KES_ONRUNQ &&
1442 ke->ke_state != KES_THREAD) || ke->ke_runq == NULL)
1443 continue;
1444 kseq = KSEQ_CPU(ke->ke_cpu);
1445
1446 #ifdef SMP
1447 /*
1448 * On SMP if we're on the RUNQ we must adjust the transferable
1449 * count because could be changing to or from an interrupt
1450 * class.
1451 */
1452 if (ke->ke_state == KES_ONRUNQ) {
1453 if (KSE_CAN_MIGRATE(ke, oclass)) {
1454 kseq->ksq_transferable--;
1455 kseq->ksq_group->ksg_transferable--;
1456 }
1457 if (KSE_CAN_MIGRATE(ke, nclass)) {
1458 kseq->ksq_transferable++;
1459 kseq->ksq_group->ksg_transferable++;
1460 }
1461 }
1462 #endif
1463 if (oclass == PRI_TIMESHARE) {
1464 kseq->ksq_load_timeshare--;
1465 kseq_nice_rem(kseq, kg->kg_proc->p_nice);
1466 }
1467 if (nclass == PRI_TIMESHARE) {
1468 kseq->ksq_load_timeshare++;
1469 kseq_nice_add(kseq, kg->kg_proc->p_nice);
1470 }
1471 }
1472
1473 kg->kg_pri_class = class;
1474 }
1475
1476 /*
1477 * Return some of the child's priority and interactivity to the parent.
1478 */
1479 void
1480 sched_exit(struct proc *p, struct thread *childtd)
1481 {
1482 mtx_assert(&sched_lock, MA_OWNED);
1483 sched_exit_ksegrp(FIRST_KSEGRP_IN_PROC(p), childtd);
1484 sched_exit_thread(NULL, childtd);
1485 }
1486
1487 void
1488 sched_exit_ksegrp(struct ksegrp *kg, struct thread *td)
1489 {
1490 /* kg->kg_slptime += td->td_ksegrp->kg_slptime; */
1491 kg->kg_runtime += td->td_ksegrp->kg_runtime;
1492 sched_interact_update(kg);
1493 }
1494
1495 void
1496 sched_exit_thread(struct thread *td, struct thread *childtd)
1497 {
1498 CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d",
1499 childtd, childtd->td_proc->p_comm, childtd->td_priority);
1500 kseq_load_rem(KSEQ_CPU(childtd->td_kse->ke_cpu), childtd->td_kse);
1501 }
1502
1503 void
1504 sched_clock(struct thread *td)
1505 {
1506 struct kseq *kseq;
1507 struct ksegrp *kg;
1508 struct kse *ke;
1509
1510 mtx_assert(&sched_lock, MA_OWNED);
1511 kseq = KSEQ_SELF();
1512 #ifdef SMP
1513 if (ticks == bal_tick)
1514 sched_balance();
1515 if (ticks == gbal_tick)
1516 sched_balance_groups();
1517 /*
1518 * We could have been assigned a non real-time thread without an
1519 * IPI.
1520 */
1521 if (kseq->ksq_assigned)
1522 kseq_assign(kseq); /* Potentially sets NEEDRESCHED */
1523 #endif
1524 /*
1525 * sched_setup() apparently happens prior to stathz being set. We
1526 * need to resolve the timers earlier in the boot so we can avoid
1527 * calculating this here.
1528 */
1529 if (realstathz == 0) {
1530 realstathz = stathz ? stathz : hz;
1531 tickincr = hz / realstathz;
1532 /*
1533 * XXX This does not work for values of stathz that are much
1534 * larger than hz.
1535 */
1536 if (tickincr == 0)
1537 tickincr = 1;
1538 }
1539
1540 ke = td->td_kse;
1541 kg = ke->ke_ksegrp;
1542
1543 /* Adjust ticks for pctcpu */
1544 ke->ke_ticks++;
1545 ke->ke_ltick = ticks;
1546
1547 /* Go up to one second beyond our max and then trim back down */
1548 if (ke->ke_ftick + SCHED_CPU_TICKS + hz < ke->ke_ltick)
1549 sched_pctcpu_update(ke);
1550
1551 if (td->td_flags & TDF_IDLETD)
1552 return;
1553 /*
1554 * We only do slicing code for TIMESHARE ksegrps.
1555 */
1556 if (kg->kg_pri_class != PRI_TIMESHARE)
1557 return;
1558 /*
1559 * We used a tick charge it to the ksegrp so that we can compute our
1560 * interactivity.
1561 */
1562 kg->kg_runtime += tickincr << 10;
1563 sched_interact_update(kg);
1564
1565 /*
1566 * We used up one time slice.
1567 */
1568 if (--ke->ke_slice > 0)
1569 return;
1570 /*
1571 * We're out of time, recompute priorities and requeue.
1572 */
1573 kseq_load_rem(kseq, ke);
1574 sched_priority(kg);
1575 sched_slice(ke);
1576 if (SCHED_CURR(kg, ke))
1577 ke->ke_runq = kseq->ksq_curr;
1578 else
1579 ke->ke_runq = kseq->ksq_next;
1580 kseq_load_add(kseq, ke);
1581 td->td_flags |= TDF_NEEDRESCHED;
1582 }
1583
1584 int
1585 sched_runnable(void)
1586 {
1587 struct kseq *kseq;
1588 int load;
1589
1590 load = 1;
1591
1592 kseq = KSEQ_SELF();
1593 #ifdef SMP
1594 if (kseq->ksq_assigned) {
1595 mtx_lock_spin(&sched_lock);
1596 kseq_assign(kseq);
1597 mtx_unlock_spin(&sched_lock);
1598 }
1599 #endif
1600 if ((curthread->td_flags & TDF_IDLETD) != 0) {
1601 if (kseq->ksq_load > 0)
1602 goto out;
1603 } else
1604 if (kseq->ksq_load - 1 > 0)
1605 goto out;
1606 load = 0;
1607 out:
1608 return (load);
1609 }
1610
1611 void
1612 sched_userret(struct thread *td)
1613 {
1614 struct ksegrp *kg;
1615 struct kse *ke;
1616
1617 kg = td->td_ksegrp;
1618 ke = td->td_kse;
1619
1620 if (td->td_priority != kg->kg_user_pri ||
1621 ke->ke_flags & KEF_PRIOELEV) {
1622 mtx_lock_spin(&sched_lock);
1623 td->td_priority = kg->kg_user_pri;
1624 if (ke->ke_flags & KEF_PRIOELEV) {
1625 ke->ke_flags &= ~KEF_PRIOELEV;
1626 sched_slice(ke);
1627 if (ke->ke_slice == 0)
1628 mi_switch(SW_INVOL, NULL);
1629 }
1630 mtx_unlock_spin(&sched_lock);
1631 }
1632 }
1633
1634 struct kse *
1635 sched_choose(void)
1636 {
1637 struct kseq *kseq;
1638 struct kse *ke;
1639
1640 mtx_assert(&sched_lock, MA_OWNED);
1641 kseq = KSEQ_SELF();
1642 #ifdef SMP
1643 restart:
1644 if (kseq->ksq_assigned)
1645 kseq_assign(kseq);
1646 #endif
1647 ke = kseq_choose(kseq);
1648 if (ke) {
1649 #ifdef SMP
1650 if (ke->ke_ksegrp->kg_pri_class == PRI_IDLE)
1651 if (kseq_idled(kseq) == 0)
1652 goto restart;
1653 #endif
1654 kseq_runq_rem(kseq, ke);
1655 ke->ke_state = KES_THREAD;
1656 return (ke);
1657 }
1658 #ifdef SMP
1659 if (kseq_idled(kseq) == 0)
1660 goto restart;
1661 #endif
1662 return (NULL);
1663 }
1664
1665 void
1666 sched_add(struct thread *td, int flags)
1667 {
1668
1669 /* let jeff work out how to map the flags better */
1670 /* I'm open to suggestions */
1671 if (flags & SRQ_YIELDING)
1672 /*
1673 * Preempting during switching can be bad JUJU
1674 * especially for KSE processes
1675 */
1676 sched_add_internal(td, 0);
1677 else
1678 sched_add_internal(td, 1);
1679 }
1680
1681 static void
1682 sched_add_internal(struct thread *td, int preemptive)
1683 {
1684 struct kseq *kseq;
1685 struct ksegrp *kg;
1686 struct kse *ke;
1687 #ifdef SMP
1688 int canmigrate;
1689 #endif
1690 int class;
1691
1692 CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
1693 td, td->td_proc->p_comm, td->td_priority, curthread,
1694 curthread->td_proc->p_comm);
1695 mtx_assert(&sched_lock, MA_OWNED);
1696 ke = td->td_kse;
1697 kg = td->td_ksegrp;
1698 if (ke->ke_flags & KEF_ASSIGNED) {
1699 if (ke->ke_flags & KEF_REMOVED) {
1700 SLOT_USE(ke->ke_ksegrp);
1701 ke->ke_flags &= ~KEF_REMOVED;
1702 }
1703 return;
1704 }
1705 kseq = KSEQ_SELF();
1706 KASSERT(ke->ke_state != KES_ONRUNQ,
1707 ("sched_add: kse %p (%s) already in run queue", ke,
1708 ke->ke_proc->p_comm));
1709 KASSERT(ke->ke_proc->p_sflag & PS_INMEM,
1710 ("sched_add: process swapped out"));
1711 KASSERT(ke->ke_runq == NULL,
1712 ("sched_add: KSE %p is still assigned to a run queue", ke));
1713
1714 class = PRI_BASE(kg->kg_pri_class);
1715 switch (class) {
1716 case PRI_ITHD:
1717 case PRI_REALTIME:
1718 ke->ke_runq = kseq->ksq_curr;
1719 ke->ke_slice = SCHED_SLICE_MAX;
1720 ke->ke_cpu = PCPU_GET(cpuid);
1721 break;
1722 case PRI_TIMESHARE:
1723 if (SCHED_CURR(kg, ke))
1724 ke->ke_runq = kseq->ksq_curr;
1725 else
1726 ke->ke_runq = kseq->ksq_next;
1727 break;
1728 case PRI_IDLE:
1729 /*
1730 * This is for priority prop.
1731 */
1732 if (ke->ke_thread->td_priority < PRI_MIN_IDLE)
1733 ke->ke_runq = kseq->ksq_curr;
1734 else
1735 ke->ke_runq = &kseq->ksq_idle;
1736 ke->ke_slice = SCHED_SLICE_MIN;
1737 break;
1738 default:
1739 panic("Unknown pri class.");
1740 break;
1741 }
1742 #ifdef SMP
1743 /*
1744 * Don't migrate running threads here. Force the long term balancer
1745 * to do it.
1746 */
1747 canmigrate = KSE_CAN_MIGRATE(ke, class);
1748 if (ke->ke_flags & KEF_HOLD) {
1749 ke->ke_flags &= ~KEF_HOLD;
1750 canmigrate = 0;
1751 }
1752 /*
1753 * If this thread is pinned or bound, notify the target cpu.
1754 */
1755 if (!canmigrate && ke->ke_cpu != PCPU_GET(cpuid) ) {
1756 ke->ke_runq = NULL;
1757 kseq_notify(ke, ke->ke_cpu);
1758 return;
1759 }
1760 /*
1761 * If we had been idle, clear our bit in the group and potentially
1762 * the global bitmap. If not, see if we should transfer this thread.
1763 */
1764 if ((class == PRI_TIMESHARE || class == PRI_REALTIME) &&
1765 (kseq->ksq_group->ksg_idlemask & PCPU_GET(cpumask)) != 0) {
1766 /*
1767 * Check to see if our group is unidling, and if so, remove it
1768 * from the global idle mask.
1769 */
1770 if (kseq->ksq_group->ksg_idlemask ==
1771 kseq->ksq_group->ksg_cpumask)
1772 atomic_clear_int(&kseq_idle, kseq->ksq_group->ksg_mask);
1773 /*
1774 * Now remove ourselves from the group specific idle mask.
1775 */
1776 kseq->ksq_group->ksg_idlemask &= ~PCPU_GET(cpumask);
1777 } else if (kseq->ksq_load > 1 && canmigrate)
1778 if (kseq_transfer(kseq, ke, class))
1779 return;
1780 ke->ke_cpu = PCPU_GET(cpuid);
1781 #endif
1782 /*
1783 * XXX With preemption this is not necessary.
1784 */
1785 if (td->td_priority < curthread->td_priority &&
1786 ke->ke_runq == kseq->ksq_curr)
1787 curthread->td_flags |= TDF_NEEDRESCHED;
1788 if (preemptive && maybe_preempt(td))
1789 return;
1790 SLOT_USE(td->td_ksegrp);
1791 ke->ke_state = KES_ONRUNQ;
1792
1793 kseq_runq_add(kseq, ke);
1794 kseq_load_add(kseq, ke);
1795 }
1796
1797 void
1798 sched_rem(struct thread *td)
1799 {
1800 struct kseq *kseq;
1801 struct kse *ke;
1802
1803 mtx_assert(&sched_lock, MA_OWNED);
1804 ke = td->td_kse;
1805 CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)",
1806 td, td->td_proc->p_comm, td->td_priority, curthread,
1807 curthread->td_proc->p_comm);
1808 /*
1809 * It is safe to just return here because sched_rem() is only ever
1810 * used in places where we're immediately going to add the
1811 * kse back on again. In that case it'll be added with the correct
1812 * thread and priority when the caller drops the sched_lock.
1813 */
1814 if (ke->ke_flags & KEF_ASSIGNED) {
1815 SLOT_RELEASE(td->td_ksegrp);
1816 ke->ke_flags |= KEF_REMOVED;
1817 return;
1818 }
1819 KASSERT((ke->ke_state == KES_ONRUNQ),
1820 ("sched_rem: KSE not on run queue"));
1821
1822 SLOT_RELEASE(td->td_ksegrp);
1823 ke->ke_state = KES_THREAD;
1824 kseq = KSEQ_CPU(ke->ke_cpu);
1825 kseq_runq_rem(kseq, ke);
1826 kseq_load_rem(kseq, ke);
1827 }
1828
1829 fixpt_t
1830 sched_pctcpu(struct thread *td)
1831 {
1832 fixpt_t pctcpu;
1833 struct kse *ke;
1834
1835 pctcpu = 0;
1836 ke = td->td_kse;
1837 if (ke == NULL)
1838 return (0);
1839
1840 mtx_lock_spin(&sched_lock);
1841 if (ke->ke_ticks) {
1842 int rtick;
1843
1844 /*
1845 * Don't update more frequently than twice a second. Allowing
1846 * this causes the cpu usage to decay away too quickly due to
1847 * rounding errors.
1848 */
1849 if (ke->ke_ftick + SCHED_CPU_TICKS < ke->ke_ltick ||
1850 ke->ke_ltick < (ticks - (hz / 2)))
1851 sched_pctcpu_update(ke);
1852 /* How many rtick per second ? */
1853 rtick = min(ke->ke_ticks / SCHED_CPU_TIME, SCHED_CPU_TICKS);
1854 pctcpu = (FSCALE * ((FSCALE * rtick)/realstathz)) >> FSHIFT;
1855 }
1856
1857 ke->ke_proc->p_swtime = ke->ke_ltick - ke->ke_ftick;
1858 mtx_unlock_spin(&sched_lock);
1859
1860 return (pctcpu);
1861 }
1862
1863 void
1864 sched_bind(struct thread *td, int cpu)
1865 {
1866 struct kse *ke;
1867
1868 mtx_assert(&sched_lock, MA_OWNED);
1869 ke = td->td_kse;
1870 ke->ke_flags |= KEF_BOUND;
1871 #ifdef SMP
1872 if (PCPU_GET(cpuid) == cpu)
1873 return;
1874 /* sched_rem without the runq_remove */
1875 ke->ke_state = KES_THREAD;
1876 kseq_load_rem(KSEQ_CPU(ke->ke_cpu), ke);
1877 kseq_notify(ke, cpu);
1878 /* When we return from mi_switch we'll be on the correct cpu. */
1879 mi_switch(SW_VOL, NULL);
1880 #endif
1881 }
1882
1883 void
1884 sched_unbind(struct thread *td)
1885 {
1886 mtx_assert(&sched_lock, MA_OWNED);
1887 td->td_kse->ke_flags &= ~KEF_BOUND;
1888 }
1889
1890 int
1891 sched_load(void)
1892 {
1893 #ifdef SMP
1894 int total;
1895 int i;
1896
1897 total = 0;
1898 for (i = 0; i <= ksg_maxid; i++)
1899 total += KSEQ_GROUP(i)->ksg_load;
1900 return (total);
1901 #else
1902 return (KSEQ_SELF()->ksq_sysload);
1903 #endif
1904 }
1905
1906 int
1907 sched_sizeof_ksegrp(void)
1908 {
1909 return (sizeof(struct ksegrp) + sizeof(struct kg_sched));
1910 }
1911
1912 int
1913 sched_sizeof_proc(void)
1914 {
1915 return (sizeof(struct proc));
1916 }
1917
1918 int
1919 sched_sizeof_thread(void)
1920 {
1921 return (sizeof(struct thread) + sizeof(struct td_sched));
1922 }
1923 #define KERN_SWITCH_INCLUDE 1
1924 #include "kern/kern_switch.c"
Cache object: f03fe06927cd9d486765d83e89c50409
|