Merge branch 'timers-nohz-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[linux-3.10.git] / kernel / softirq.c
1 /*
2  *      linux/kernel/softirq.c
3  *
4  *      Copyright (C) 1992 Linus Torvalds
5  *
6  *      Distribute under GPLv2.
7  *
8  *      Rewritten. Old one was good in 2.2, but in 2.3 it was immoral. --ANK (990903)
9  *
10  *      Remote softirq infrastructure is by Jens Axboe.
11  */
12
13 #include <linux/export.h>
14 #include <linux/kernel_stat.h>
15 #include <linux/interrupt.h>
16 #include <linux/init.h>
17 #include <linux/mm.h>
18 #include <linux/notifier.h>
19 #include <linux/percpu.h>
20 #include <linux/cpu.h>
21 #include <linux/freezer.h>
22 #include <linux/kthread.h>
23 #include <linux/rcupdate.h>
24 #include <linux/ftrace.h>
25 #include <linux/smp.h>
26 #include <linux/smpboot.h>
27 #include <linux/tick.h>
28
29 #define CREATE_TRACE_POINTS
30 #include <trace/events/irq.h>
31
32 #include <asm/irq.h>
33 /*
34    - No shared variables, all the data are CPU local.
35    - If a softirq needs serialization, let it serialize itself
36      by its own spinlocks.
37    - Even if softirq is serialized, only local cpu is marked for
38      execution. Hence, we get something sort of weak cpu binding.
39      Though it is still not clear, will it result in better locality
40      or will not.
41
42    Examples:
43    - NET RX softirq. It is multithreaded and does not require
44      any global serialization.
45    - NET TX softirq. It kicks software netdevice queues, hence
46      it is logically serialized per device, but this serialization
47      is invisible to common code.
48    - Tasklets: serialized wrt itself.
49  */
50
51 #ifndef __ARCH_IRQ_STAT
52 irq_cpustat_t irq_stat[NR_CPUS] ____cacheline_aligned;
53 EXPORT_SYMBOL(irq_stat);
54 #endif
55
56 static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
57
58 DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
59
60 char *softirq_to_name[NR_SOFTIRQS] = {
61         "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "BLOCK_IOPOLL",
62         "TASKLET", "SCHED", "HRTIMER", "RCU"
63 };
64
65 /*
66  * we cannot loop indefinitely here to avoid userspace starvation,
67  * but we also don't want to introduce a worst case 1/HZ latency
68  * to the pending events, so lets the scheduler to balance
69  * the softirq load for us.
70  */
71 static void wakeup_softirqd(void)
72 {
73         /* Interrupts are disabled: no need to stop preemption */
74         struct task_struct *tsk = __this_cpu_read(ksoftirqd);
75
76         if (tsk && tsk->state != TASK_RUNNING)
77                 wake_up_process(tsk);
78 }
79
80 /*
81  * preempt_count and SOFTIRQ_OFFSET usage:
82  * - preempt_count is changed by SOFTIRQ_OFFSET on entering or leaving
83  *   softirq processing.
84  * - preempt_count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
85  *   on local_bh_disable or local_bh_enable.
86  * This lets us distinguish between whether we are currently processing
87  * softirq and whether we just have bh disabled.
88  */
89
90 /*
91  * This one is for softirq.c-internal use,
92  * where hardirqs are disabled legitimately:
93  */
94 #ifdef CONFIG_TRACE_IRQFLAGS
95 static void __local_bh_disable(unsigned long ip, unsigned int cnt)
96 {
97         unsigned long flags;
98
99         WARN_ON_ONCE(in_irq());
100
101         raw_local_irq_save(flags);
102         /*
103          * The preempt tracer hooks into add_preempt_count and will break
104          * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
105          * is set and before current->softirq_enabled is cleared.
106          * We must manually increment preempt_count here and manually
107          * call the trace_preempt_off later.
108          */
109         preempt_count() += cnt;
110         /*
111          * Were softirqs turned off above:
112          */
113         if (softirq_count() == cnt)
114                 trace_softirqs_off(ip);
115         raw_local_irq_restore(flags);
116
117         if (preempt_count() == cnt)
118                 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
119 }
120 #else /* !CONFIG_TRACE_IRQFLAGS */
121 static inline void __local_bh_disable(unsigned long ip, unsigned int cnt)
122 {
123         add_preempt_count(cnt);
124         barrier();
125 }
126 #endif /* CONFIG_TRACE_IRQFLAGS */
127
128 void local_bh_disable(void)
129 {
130         __local_bh_disable((unsigned long)__builtin_return_address(0),
131                                 SOFTIRQ_DISABLE_OFFSET);
132 }
133
134 EXPORT_SYMBOL(local_bh_disable);
135
136 static void __local_bh_enable(unsigned int cnt)
137 {
138         WARN_ON_ONCE(in_irq());
139         WARN_ON_ONCE(!irqs_disabled());
140
141         if (softirq_count() == cnt)
142                 trace_softirqs_on((unsigned long)__builtin_return_address(0));
143         sub_preempt_count(cnt);
144 }
145
146 /*
147  * Special-case - softirqs can safely be enabled in
148  * cond_resched_softirq(), or by __do_softirq(),
149  * without processing still-pending softirqs:
150  */
151 void _local_bh_enable(void)
152 {
153         __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
154 }
155
156 EXPORT_SYMBOL(_local_bh_enable);
157
158 static inline void _local_bh_enable_ip(unsigned long ip)
159 {
160         WARN_ON_ONCE(in_irq() || irqs_disabled());
161 #ifdef CONFIG_TRACE_IRQFLAGS
162         local_irq_disable();
163 #endif
164         /*
165          * Are softirqs going to be turned on now:
166          */
167         if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
168                 trace_softirqs_on(ip);
169         /*
170          * Keep preemption disabled until we are done with
171          * softirq processing:
172          */
173         sub_preempt_count(SOFTIRQ_DISABLE_OFFSET - 1);
174
175         if (unlikely(!in_interrupt() && local_softirq_pending()))
176                 do_softirq();
177
178         dec_preempt_count();
179 #ifdef CONFIG_TRACE_IRQFLAGS
180         local_irq_enable();
181 #endif
182         preempt_check_resched();
183 }
184
185 void local_bh_enable(void)
186 {
187         _local_bh_enable_ip((unsigned long)__builtin_return_address(0));
188 }
189 EXPORT_SYMBOL(local_bh_enable);
190
191 void local_bh_enable_ip(unsigned long ip)
192 {
193         _local_bh_enable_ip(ip);
194 }
195 EXPORT_SYMBOL(local_bh_enable_ip);
196
197 /*
198  * We restart softirq processing for at most 2 ms,
199  * and if need_resched() is not set.
200  *
201  * These limits have been established via experimentation.
202  * The two things to balance is latency against fairness -
203  * we want to handle softirqs as soon as possible, but they
204  * should not be able to lock up the box.
205  */
206 #define MAX_SOFTIRQ_TIME  msecs_to_jiffies(2)
207
208 asmlinkage void __do_softirq(void)
209 {
210         struct softirq_action *h;
211         __u32 pending;
212         unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
213         int cpu;
214         unsigned long old_flags = current->flags;
215
216         /*
217          * Mask out PF_MEMALLOC s current task context is borrowed for the
218          * softirq. A softirq handled such as network RX might set PF_MEMALLOC
219          * again if the socket is related to swap
220          */
221         current->flags &= ~PF_MEMALLOC;
222
223         pending = local_softirq_pending();
224         account_irq_enter_time(current);
225
226         __local_bh_disable((unsigned long)__builtin_return_address(0),
227                                 SOFTIRQ_OFFSET);
228         lockdep_softirq_enter();
229
230         cpu = smp_processor_id();
231 restart:
232         /* Reset the pending bitmask before enabling irqs */
233         set_softirq_pending(0);
234
235         local_irq_enable();
236
237         h = softirq_vec;
238
239         do {
240                 if (pending & 1) {
241                         unsigned int vec_nr = h - softirq_vec;
242                         int prev_count = preempt_count();
243
244                         kstat_incr_softirqs_this_cpu(vec_nr);
245
246                         trace_softirq_entry(vec_nr);
247                         h->action(h);
248                         trace_softirq_exit(vec_nr);
249                         if (unlikely(prev_count != preempt_count())) {
250                                 printk(KERN_ERR "huh, entered softirq %u %s %p"
251                                        "with preempt_count %08x,"
252                                        " exited with %08x?\n", vec_nr,
253                                        softirq_to_name[vec_nr], h->action,
254                                        prev_count, preempt_count());
255                                 preempt_count() = prev_count;
256                         }
257
258                         rcu_bh_qs(cpu);
259                 }
260                 h++;
261                 pending >>= 1;
262         } while (pending);
263
264         local_irq_disable();
265
266         pending = local_softirq_pending();
267         if (pending) {
268                 if (time_before(jiffies, end) && !need_resched())
269                         goto restart;
270
271                 wakeup_softirqd();
272         }
273
274         lockdep_softirq_exit();
275
276         account_irq_exit_time(current);
277         __local_bh_enable(SOFTIRQ_OFFSET);
278         tsk_restore_flags(current, old_flags, PF_MEMALLOC);
279 }
280
281 #ifndef __ARCH_HAS_DO_SOFTIRQ
282
283 asmlinkage void do_softirq(void)
284 {
285         __u32 pending;
286         unsigned long flags;
287
288         if (in_interrupt())
289                 return;
290
291         local_irq_save(flags);
292
293         pending = local_softirq_pending();
294
295         if (pending)
296                 __do_softirq();
297
298         local_irq_restore(flags);
299 }
300
301 #endif
302
303 /*
304  * Enter an interrupt context.
305  */
306 void irq_enter(void)
307 {
308         int cpu = smp_processor_id();
309
310         rcu_irq_enter();
311         if (is_idle_task(current) && !in_interrupt()) {
312                 /*
313                  * Prevent raise_softirq from needlessly waking up ksoftirqd
314                  * here, as softirq will be serviced on return from interrupt.
315                  */
316                 local_bh_disable();
317                 tick_check_idle(cpu);
318                 _local_bh_enable();
319         }
320
321         __irq_enter();
322 }
323
324 static inline void invoke_softirq(void)
325 {
326         if (!force_irqthreads)
327                 __do_softirq();
328         else
329                 wakeup_softirqd();
330 }
331
332 static inline void tick_irq_exit(void)
333 {
334 #ifdef CONFIG_NO_HZ_COMMON
335         int cpu = smp_processor_id();
336
337         /* Make sure that timer wheel updates are propagated */
338         if ((idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
339                 if (!in_interrupt())
340                         tick_nohz_irq_exit();
341         }
342 #endif
343 }
344
345 /*
346  * Exit an interrupt context. Process softirqs if needed and possible:
347  */
348 void irq_exit(void)
349 {
350 #ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
351         local_irq_disable();
352 #else
353         WARN_ON_ONCE(!irqs_disabled());
354 #endif
355
356         account_irq_exit_time(current);
357         trace_hardirq_exit();
358         sub_preempt_count(HARDIRQ_OFFSET);
359         if (!in_interrupt() && local_softirq_pending())
360                 invoke_softirq();
361
362         tick_irq_exit();
363         rcu_irq_exit();
364 }
365
366 /*
367  * This function must run with irqs disabled!
368  */
369 inline void raise_softirq_irqoff(unsigned int nr)
370 {
371         __raise_softirq_irqoff(nr);
372
373         /*
374          * If we're in an interrupt or softirq, we're done
375          * (this also catches softirq-disabled code). We will
376          * actually run the softirq once we return from
377          * the irq or softirq.
378          *
379          * Otherwise we wake up ksoftirqd to make sure we
380          * schedule the softirq soon.
381          */
382         if (!in_interrupt())
383                 wakeup_softirqd();
384 }
385
386 void raise_softirq(unsigned int nr)
387 {
388         unsigned long flags;
389
390         local_irq_save(flags);
391         raise_softirq_irqoff(nr);
392         local_irq_restore(flags);
393 }
394
395 void __raise_softirq_irqoff(unsigned int nr)
396 {
397         trace_softirq_raise(nr);
398         or_softirq_pending(1UL << nr);
399 }
400
401 void open_softirq(int nr, void (*action)(struct softirq_action *))
402 {
403         softirq_vec[nr].action = action;
404 }
405
406 /*
407  * Tasklets
408  */
409 struct tasklet_head
410 {
411         struct tasklet_struct *head;
412         struct tasklet_struct **tail;
413 };
414
415 static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
416 static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
417
418 void __tasklet_schedule(struct tasklet_struct *t)
419 {
420         unsigned long flags;
421
422         local_irq_save(flags);
423         t->next = NULL;
424         *__this_cpu_read(tasklet_vec.tail) = t;
425         __this_cpu_write(tasklet_vec.tail, &(t->next));
426         raise_softirq_irqoff(TASKLET_SOFTIRQ);
427         local_irq_restore(flags);
428 }
429
430 EXPORT_SYMBOL(__tasklet_schedule);
431
432 void __tasklet_hi_schedule(struct tasklet_struct *t)
433 {
434         unsigned long flags;
435
436         local_irq_save(flags);
437         t->next = NULL;
438         *__this_cpu_read(tasklet_hi_vec.tail) = t;
439         __this_cpu_write(tasklet_hi_vec.tail,  &(t->next));
440         raise_softirq_irqoff(HI_SOFTIRQ);
441         local_irq_restore(flags);
442 }
443
444 EXPORT_SYMBOL(__tasklet_hi_schedule);
445
446 void __tasklet_hi_schedule_first(struct tasklet_struct *t)
447 {
448         BUG_ON(!irqs_disabled());
449
450         t->next = __this_cpu_read(tasklet_hi_vec.head);
451         __this_cpu_write(tasklet_hi_vec.head, t);
452         __raise_softirq_irqoff(HI_SOFTIRQ);
453 }
454
455 EXPORT_SYMBOL(__tasklet_hi_schedule_first);
456
457 static void tasklet_action(struct softirq_action *a)
458 {
459         struct tasklet_struct *list;
460
461         local_irq_disable();
462         list = __this_cpu_read(tasklet_vec.head);
463         __this_cpu_write(tasklet_vec.head, NULL);
464         __this_cpu_write(tasklet_vec.tail, &__get_cpu_var(tasklet_vec).head);
465         local_irq_enable();
466
467         while (list) {
468                 struct tasklet_struct *t = list;
469
470                 list = list->next;
471
472                 if (tasklet_trylock(t)) {
473                         if (!atomic_read(&t->count)) {
474                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
475                                         BUG();
476                                 t->func(t->data);
477                                 tasklet_unlock(t);
478                                 continue;
479                         }
480                         tasklet_unlock(t);
481                 }
482
483                 local_irq_disable();
484                 t->next = NULL;
485                 *__this_cpu_read(tasklet_vec.tail) = t;
486                 __this_cpu_write(tasklet_vec.tail, &(t->next));
487                 __raise_softirq_irqoff(TASKLET_SOFTIRQ);
488                 local_irq_enable();
489         }
490 }
491
492 static void tasklet_hi_action(struct softirq_action *a)
493 {
494         struct tasklet_struct *list;
495
496         local_irq_disable();
497         list = __this_cpu_read(tasklet_hi_vec.head);
498         __this_cpu_write(tasklet_hi_vec.head, NULL);
499         __this_cpu_write(tasklet_hi_vec.tail, &__get_cpu_var(tasklet_hi_vec).head);
500         local_irq_enable();
501
502         while (list) {
503                 struct tasklet_struct *t = list;
504
505                 list = list->next;
506
507                 if (tasklet_trylock(t)) {
508                         if (!atomic_read(&t->count)) {
509                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
510                                         BUG();
511                                 t->func(t->data);
512                                 tasklet_unlock(t);
513                                 continue;
514                         }
515                         tasklet_unlock(t);
516                 }
517
518                 local_irq_disable();
519                 t->next = NULL;
520                 *__this_cpu_read(tasklet_hi_vec.tail) = t;
521                 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
522                 __raise_softirq_irqoff(HI_SOFTIRQ);
523                 local_irq_enable();
524         }
525 }
526
527
528 void tasklet_init(struct tasklet_struct *t,
529                   void (*func)(unsigned long), unsigned long data)
530 {
531         t->next = NULL;
532         t->state = 0;
533         atomic_set(&t->count, 0);
534         t->func = func;
535         t->data = data;
536 }
537
538 EXPORT_SYMBOL(tasklet_init);
539
540 void tasklet_kill(struct tasklet_struct *t)
541 {
542         if (in_interrupt())
543                 printk("Attempt to kill tasklet from interrupt\n");
544
545         while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
546                 do {
547                         yield();
548                 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
549         }
550         tasklet_unlock_wait(t);
551         clear_bit(TASKLET_STATE_SCHED, &t->state);
552 }
553
554 EXPORT_SYMBOL(tasklet_kill);
555
556 /*
557  * tasklet_hrtimer
558  */
559
560 /*
561  * The trampoline is called when the hrtimer expires. It schedules a tasklet
562  * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
563  * hrtimer callback, but from softirq context.
564  */
565 static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer)
566 {
567         struct tasklet_hrtimer *ttimer =
568                 container_of(timer, struct tasklet_hrtimer, timer);
569
570         tasklet_hi_schedule(&ttimer->tasklet);
571         return HRTIMER_NORESTART;
572 }
573
574 /*
575  * Helper function which calls the hrtimer callback from
576  * tasklet/softirq context
577  */
578 static void __tasklet_hrtimer_trampoline(unsigned long data)
579 {
580         struct tasklet_hrtimer *ttimer = (void *)data;
581         enum hrtimer_restart restart;
582
583         restart = ttimer->function(&ttimer->timer);
584         if (restart != HRTIMER_NORESTART)
585                 hrtimer_restart(&ttimer->timer);
586 }
587
588 /**
589  * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
590  * @ttimer:      tasklet_hrtimer which is initialized
591  * @function:    hrtimer callback function which gets called from softirq context
592  * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
593  * @mode:        hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
594  */
595 void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer,
596                           enum hrtimer_restart (*function)(struct hrtimer *),
597                           clockid_t which_clock, enum hrtimer_mode mode)
598 {
599         hrtimer_init(&ttimer->timer, which_clock, mode);
600         ttimer->timer.function = __hrtimer_tasklet_trampoline;
601         tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline,
602                      (unsigned long)ttimer);
603         ttimer->function = function;
604 }
605 EXPORT_SYMBOL_GPL(tasklet_hrtimer_init);
606
607 /*
608  * Remote softirq bits
609  */
610
611 DEFINE_PER_CPU(struct list_head [NR_SOFTIRQS], softirq_work_list);
612 EXPORT_PER_CPU_SYMBOL(softirq_work_list);
613
614 static void __local_trigger(struct call_single_data *cp, int softirq)
615 {
616         struct list_head *head = &__get_cpu_var(softirq_work_list[softirq]);
617
618         list_add_tail(&cp->list, head);
619
620         /* Trigger the softirq only if the list was previously empty.  */
621         if (head->next == &cp->list)
622                 raise_softirq_irqoff(softirq);
623 }
624
625 #ifdef CONFIG_USE_GENERIC_SMP_HELPERS
626 static void remote_softirq_receive(void *data)
627 {
628         struct call_single_data *cp = data;
629         unsigned long flags;
630         int softirq;
631
632         softirq = *(int *)cp->info;
633         local_irq_save(flags);
634         __local_trigger(cp, softirq);
635         local_irq_restore(flags);
636 }
637
638 static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
639 {
640         if (cpu_online(cpu)) {
641                 cp->func = remote_softirq_receive;
642                 cp->info = &softirq;
643                 cp->flags = 0;
644
645                 __smp_call_function_single(cpu, cp, 0);
646                 return 0;
647         }
648         return 1;
649 }
650 #else /* CONFIG_USE_GENERIC_SMP_HELPERS */
651 static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
652 {
653         return 1;
654 }
655 #endif
656
657 /**
658  * __send_remote_softirq - try to schedule softirq work on a remote cpu
659  * @cp: private SMP call function data area
660  * @cpu: the remote cpu
661  * @this_cpu: the currently executing cpu
662  * @softirq: the softirq for the work
663  *
664  * Attempt to schedule softirq work on a remote cpu.  If this cannot be
665  * done, the work is instead queued up on the local cpu.
666  *
667  * Interrupts must be disabled.
668  */
669 void __send_remote_softirq(struct call_single_data *cp, int cpu, int this_cpu, int softirq)
670 {
671         if (cpu == this_cpu || __try_remote_softirq(cp, cpu, softirq))
672                 __local_trigger(cp, softirq);
673 }
674 EXPORT_SYMBOL(__send_remote_softirq);
675
676 /**
677  * send_remote_softirq - try to schedule softirq work on a remote cpu
678  * @cp: private SMP call function data area
679  * @cpu: the remote cpu
680  * @softirq: the softirq for the work
681  *
682  * Like __send_remote_softirq except that disabling interrupts and
683  * computing the current cpu is done for the caller.
684  */
685 void send_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
686 {
687         unsigned long flags;
688         int this_cpu;
689
690         local_irq_save(flags);
691         this_cpu = smp_processor_id();
692         __send_remote_softirq(cp, cpu, this_cpu, softirq);
693         local_irq_restore(flags);
694 }
695 EXPORT_SYMBOL(send_remote_softirq);
696
697 static int __cpuinit remote_softirq_cpu_notify(struct notifier_block *self,
698                                                unsigned long action, void *hcpu)
699 {
700         /*
701          * If a CPU goes away, splice its entries to the current CPU
702          * and trigger a run of the softirq
703          */
704         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
705                 int cpu = (unsigned long) hcpu;
706                 int i;
707
708                 local_irq_disable();
709                 for (i = 0; i < NR_SOFTIRQS; i++) {
710                         struct list_head *head = &per_cpu(softirq_work_list[i], cpu);
711                         struct list_head *local_head;
712
713                         if (list_empty(head))
714                                 continue;
715
716                         local_head = &__get_cpu_var(softirq_work_list[i]);
717                         list_splice_init(head, local_head);
718                         raise_softirq_irqoff(i);
719                 }
720                 local_irq_enable();
721         }
722
723         return NOTIFY_OK;
724 }
725
726 static struct notifier_block __cpuinitdata remote_softirq_cpu_notifier = {
727         .notifier_call  = remote_softirq_cpu_notify,
728 };
729
730 void __init softirq_init(void)
731 {
732         int cpu;
733
734         for_each_possible_cpu(cpu) {
735                 int i;
736
737                 per_cpu(tasklet_vec, cpu).tail =
738                         &per_cpu(tasklet_vec, cpu).head;
739                 per_cpu(tasklet_hi_vec, cpu).tail =
740                         &per_cpu(tasklet_hi_vec, cpu).head;
741                 for (i = 0; i < NR_SOFTIRQS; i++)
742                         INIT_LIST_HEAD(&per_cpu(softirq_work_list[i], cpu));
743         }
744
745         register_hotcpu_notifier(&remote_softirq_cpu_notifier);
746
747         open_softirq(TASKLET_SOFTIRQ, tasklet_action);
748         open_softirq(HI_SOFTIRQ, tasklet_hi_action);
749 }
750
751 static int ksoftirqd_should_run(unsigned int cpu)
752 {
753         return local_softirq_pending();
754 }
755
756 static void run_ksoftirqd(unsigned int cpu)
757 {
758         local_irq_disable();
759         if (local_softirq_pending()) {
760                 __do_softirq();
761                 rcu_note_context_switch(cpu);
762                 local_irq_enable();
763                 cond_resched();
764                 return;
765         }
766         local_irq_enable();
767 }
768
769 #ifdef CONFIG_HOTPLUG_CPU
770 /*
771  * tasklet_kill_immediate is called to remove a tasklet which can already be
772  * scheduled for execution on @cpu.
773  *
774  * Unlike tasklet_kill, this function removes the tasklet
775  * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
776  *
777  * When this function is called, @cpu must be in the CPU_DEAD state.
778  */
779 void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
780 {
781         struct tasklet_struct **i;
782
783         BUG_ON(cpu_online(cpu));
784         BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
785
786         if (!test_bit(TASKLET_STATE_SCHED, &t->state))
787                 return;
788
789         /* CPU is dead, so no lock needed. */
790         for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
791                 if (*i == t) {
792                         *i = t->next;
793                         /* If this was the tail element, move the tail ptr */
794                         if (*i == NULL)
795                                 per_cpu(tasklet_vec, cpu).tail = i;
796                         return;
797                 }
798         }
799         BUG();
800 }
801
802 static void takeover_tasklets(unsigned int cpu)
803 {
804         /* CPU is dead, so no lock needed. */
805         local_irq_disable();
806
807         /* Find end, append list for that CPU. */
808         if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
809                 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
810                 this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
811                 per_cpu(tasklet_vec, cpu).head = NULL;
812                 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
813         }
814         raise_softirq_irqoff(TASKLET_SOFTIRQ);
815
816         if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
817                 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
818                 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
819                 per_cpu(tasklet_hi_vec, cpu).head = NULL;
820                 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
821         }
822         raise_softirq_irqoff(HI_SOFTIRQ);
823
824         local_irq_enable();
825 }
826 #endif /* CONFIG_HOTPLUG_CPU */
827
828 static int __cpuinit cpu_callback(struct notifier_block *nfb,
829                                   unsigned long action,
830                                   void *hcpu)
831 {
832         switch (action) {
833 #ifdef CONFIG_HOTPLUG_CPU
834         case CPU_DEAD:
835         case CPU_DEAD_FROZEN:
836                 takeover_tasklets((unsigned long)hcpu);
837                 break;
838 #endif /* CONFIG_HOTPLUG_CPU */
839         }
840         return NOTIFY_OK;
841 }
842
843 static struct notifier_block __cpuinitdata cpu_nfb = {
844         .notifier_call = cpu_callback
845 };
846
847 static struct smp_hotplug_thread softirq_threads = {
848         .store                  = &ksoftirqd,
849         .thread_should_run      = ksoftirqd_should_run,
850         .thread_fn              = run_ksoftirqd,
851         .thread_comm            = "ksoftirqd/%u",
852 };
853
854 static __init int spawn_ksoftirqd(void)
855 {
856         register_cpu_notifier(&cpu_nfb);
857
858         BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
859
860         return 0;
861 }
862 early_initcall(spawn_ksoftirqd);
863
864 /*
865  * [ These __weak aliases are kept in a separate compilation unit, so that
866  *   GCC does not inline them incorrectly. ]
867  */
868
869 int __init __weak early_irq_init(void)
870 {
871         return 0;
872 }
873
874 #ifdef CONFIG_GENERIC_HARDIRQS
875 int __init __weak arch_probe_nr_irqs(void)
876 {
877         return NR_IRQS_LEGACY;
878 }
879
880 int __init __weak arch_early_irq_init(void)
881 {
882         return 0;
883 }
884 #endif