Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next
[~shefty/rdma-dev.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 #ifdef __ARCH_IRQ_EXIT_IRQS_DISABLED
328                 __do_softirq();
329 #else
330                 do_softirq();
331 #endif
332         } else {
333                 __local_bh_disable((unsigned long)__builtin_return_address(0),
334                                 SOFTIRQ_OFFSET);
335                 wakeup_softirqd();
336                 __local_bh_enable(SOFTIRQ_OFFSET);
337         }
338 }
339
340 /*
341  * Exit an interrupt context. Process softirqs if needed and possible:
342  */
343 void irq_exit(void)
344 {
345         account_irq_exit_time(current);
346         trace_hardirq_exit();
347         sub_preempt_count(IRQ_EXIT_OFFSET);
348         if (!in_interrupt() && local_softirq_pending())
349                 invoke_softirq();
350
351 #ifdef CONFIG_NO_HZ
352         /* Make sure that timer wheel updates are propagated */
353         if (idle_cpu(smp_processor_id()) && !in_interrupt() && !need_resched())
354                 tick_nohz_irq_exit();
355 #endif
356         rcu_irq_exit();
357         sched_preempt_enable_no_resched();
358 }
359
360 /*
361  * This function must run with irqs disabled!
362  */
363 inline void raise_softirq_irqoff(unsigned int nr)
364 {
365         __raise_softirq_irqoff(nr);
366
367         /*
368          * If we're in an interrupt or softirq, we're done
369          * (this also catches softirq-disabled code). We will
370          * actually run the softirq once we return from
371          * the irq or softirq.
372          *
373          * Otherwise we wake up ksoftirqd to make sure we
374          * schedule the softirq soon.
375          */
376         if (!in_interrupt())
377                 wakeup_softirqd();
378 }
379
380 void raise_softirq(unsigned int nr)
381 {
382         unsigned long flags;
383
384         local_irq_save(flags);
385         raise_softirq_irqoff(nr);
386         local_irq_restore(flags);
387 }
388
389 void __raise_softirq_irqoff(unsigned int nr)
390 {
391         trace_softirq_raise(nr);
392         or_softirq_pending(1UL << nr);
393 }
394
395 void open_softirq(int nr, void (*action)(struct softirq_action *))
396 {
397         softirq_vec[nr].action = action;
398 }
399
400 /*
401  * Tasklets
402  */
403 struct tasklet_head
404 {
405         struct tasklet_struct *head;
406         struct tasklet_struct **tail;
407 };
408
409 static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
410 static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
411
412 void __tasklet_schedule(struct tasklet_struct *t)
413 {
414         unsigned long flags;
415
416         local_irq_save(flags);
417         t->next = NULL;
418         *__this_cpu_read(tasklet_vec.tail) = t;
419         __this_cpu_write(tasklet_vec.tail, &(t->next));
420         raise_softirq_irqoff(TASKLET_SOFTIRQ);
421         local_irq_restore(flags);
422 }
423
424 EXPORT_SYMBOL(__tasklet_schedule);
425
426 void __tasklet_hi_schedule(struct tasklet_struct *t)
427 {
428         unsigned long flags;
429
430         local_irq_save(flags);
431         t->next = NULL;
432         *__this_cpu_read(tasklet_hi_vec.tail) = t;
433         __this_cpu_write(tasklet_hi_vec.tail,  &(t->next));
434         raise_softirq_irqoff(HI_SOFTIRQ);
435         local_irq_restore(flags);
436 }
437
438 EXPORT_SYMBOL(__tasklet_hi_schedule);
439
440 void __tasklet_hi_schedule_first(struct tasklet_struct *t)
441 {
442         BUG_ON(!irqs_disabled());
443
444         t->next = __this_cpu_read(tasklet_hi_vec.head);
445         __this_cpu_write(tasklet_hi_vec.head, t);
446         __raise_softirq_irqoff(HI_SOFTIRQ);
447 }
448
449 EXPORT_SYMBOL(__tasklet_hi_schedule_first);
450
451 static void tasklet_action(struct softirq_action *a)
452 {
453         struct tasklet_struct *list;
454
455         local_irq_disable();
456         list = __this_cpu_read(tasklet_vec.head);
457         __this_cpu_write(tasklet_vec.head, NULL);
458         __this_cpu_write(tasklet_vec.tail, &__get_cpu_var(tasklet_vec).head);
459         local_irq_enable();
460
461         while (list) {
462                 struct tasklet_struct *t = list;
463
464                 list = list->next;
465
466                 if (tasklet_trylock(t)) {
467                         if (!atomic_read(&t->count)) {
468                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
469                                         BUG();
470                                 t->func(t->data);
471                                 tasklet_unlock(t);
472                                 continue;
473                         }
474                         tasklet_unlock(t);
475                 }
476
477                 local_irq_disable();
478                 t->next = NULL;
479                 *__this_cpu_read(tasklet_vec.tail) = t;
480                 __this_cpu_write(tasklet_vec.tail, &(t->next));
481                 __raise_softirq_irqoff(TASKLET_SOFTIRQ);
482                 local_irq_enable();
483         }
484 }
485
486 static void tasklet_hi_action(struct softirq_action *a)
487 {
488         struct tasklet_struct *list;
489
490         local_irq_disable();
491         list = __this_cpu_read(tasklet_hi_vec.head);
492         __this_cpu_write(tasklet_hi_vec.head, NULL);
493         __this_cpu_write(tasklet_hi_vec.tail, &__get_cpu_var(tasklet_hi_vec).head);
494         local_irq_enable();
495
496         while (list) {
497                 struct tasklet_struct *t = list;
498
499                 list = list->next;
500
501                 if (tasklet_trylock(t)) {
502                         if (!atomic_read(&t->count)) {
503                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
504                                         BUG();
505                                 t->func(t->data);
506                                 tasklet_unlock(t);
507                                 continue;
508                         }
509                         tasklet_unlock(t);
510                 }
511
512                 local_irq_disable();
513                 t->next = NULL;
514                 *__this_cpu_read(tasklet_hi_vec.tail) = t;
515                 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
516                 __raise_softirq_irqoff(HI_SOFTIRQ);
517                 local_irq_enable();
518         }
519 }
520
521
522 void tasklet_init(struct tasklet_struct *t,
523                   void (*func)(unsigned long), unsigned long data)
524 {
525         t->next = NULL;
526         t->state = 0;
527         atomic_set(&t->count, 0);
528         t->func = func;
529         t->data = data;
530 }
531
532 EXPORT_SYMBOL(tasklet_init);
533
534 void tasklet_kill(struct tasklet_struct *t)
535 {
536         if (in_interrupt())
537                 printk("Attempt to kill tasklet from interrupt\n");
538
539         while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
540                 do {
541                         yield();
542                 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
543         }
544         tasklet_unlock_wait(t);
545         clear_bit(TASKLET_STATE_SCHED, &t->state);
546 }
547
548 EXPORT_SYMBOL(tasklet_kill);
549
550 /*
551  * tasklet_hrtimer
552  */
553
554 /*
555  * The trampoline is called when the hrtimer expires. It schedules a tasklet
556  * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
557  * hrtimer callback, but from softirq context.
558  */
559 static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer)
560 {
561         struct tasklet_hrtimer *ttimer =
562                 container_of(timer, struct tasklet_hrtimer, timer);
563
564         tasklet_hi_schedule(&ttimer->tasklet);
565         return HRTIMER_NORESTART;
566 }
567
568 /*
569  * Helper function which calls the hrtimer callback from
570  * tasklet/softirq context
571  */
572 static void __tasklet_hrtimer_trampoline(unsigned long data)
573 {
574         struct tasklet_hrtimer *ttimer = (void *)data;
575         enum hrtimer_restart restart;
576
577         restart = ttimer->function(&ttimer->timer);
578         if (restart != HRTIMER_NORESTART)
579                 hrtimer_restart(&ttimer->timer);
580 }
581
582 /**
583  * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
584  * @ttimer:      tasklet_hrtimer which is initialized
585  * @function:    hrtimer callback function which gets called from softirq context
586  * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
587  * @mode:        hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
588  */
589 void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer,
590                           enum hrtimer_restart (*function)(struct hrtimer *),
591                           clockid_t which_clock, enum hrtimer_mode mode)
592 {
593         hrtimer_init(&ttimer->timer, which_clock, mode);
594         ttimer->timer.function = __hrtimer_tasklet_trampoline;
595         tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline,
596                      (unsigned long)ttimer);
597         ttimer->function = function;
598 }
599 EXPORT_SYMBOL_GPL(tasklet_hrtimer_init);
600
601 /*
602  * Remote softirq bits
603  */
604
605 DEFINE_PER_CPU(struct list_head [NR_SOFTIRQS], softirq_work_list);
606 EXPORT_PER_CPU_SYMBOL(softirq_work_list);
607
608 static void __local_trigger(struct call_single_data *cp, int softirq)
609 {
610         struct list_head *head = &__get_cpu_var(softirq_work_list[softirq]);
611
612         list_add_tail(&cp->list, head);
613
614         /* Trigger the softirq only if the list was previously empty.  */
615         if (head->next == &cp->list)
616                 raise_softirq_irqoff(softirq);
617 }
618
619 #ifdef CONFIG_USE_GENERIC_SMP_HELPERS
620 static void remote_softirq_receive(void *data)
621 {
622         struct call_single_data *cp = data;
623         unsigned long flags;
624         int softirq;
625
626         softirq = cp->priv;
627
628         local_irq_save(flags);
629         __local_trigger(cp, softirq);
630         local_irq_restore(flags);
631 }
632
633 static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
634 {
635         if (cpu_online(cpu)) {
636                 cp->func = remote_softirq_receive;
637                 cp->info = cp;
638                 cp->flags = 0;
639                 cp->priv = softirq;
640
641                 __smp_call_function_single(cpu, cp, 0);
642                 return 0;
643         }
644         return 1;
645 }
646 #else /* CONFIG_USE_GENERIC_SMP_HELPERS */
647 static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
648 {
649         return 1;
650 }
651 #endif
652
653 /**
654  * __send_remote_softirq - try to schedule softirq work on a remote cpu
655  * @cp: private SMP call function data area
656  * @cpu: the remote cpu
657  * @this_cpu: the currently executing cpu
658  * @softirq: the softirq for the work
659  *
660  * Attempt to schedule softirq work on a remote cpu.  If this cannot be
661  * done, the work is instead queued up on the local cpu.
662  *
663  * Interrupts must be disabled.
664  */
665 void __send_remote_softirq(struct call_single_data *cp, int cpu, int this_cpu, int softirq)
666 {
667         if (cpu == this_cpu || __try_remote_softirq(cp, cpu, softirq))
668                 __local_trigger(cp, softirq);
669 }
670 EXPORT_SYMBOL(__send_remote_softirq);
671
672 /**
673  * send_remote_softirq - try to schedule softirq work on a remote cpu
674  * @cp: private SMP call function data area
675  * @cpu: the remote cpu
676  * @softirq: the softirq for the work
677  *
678  * Like __send_remote_softirq except that disabling interrupts and
679  * computing the current cpu is done for the caller.
680  */
681 void send_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
682 {
683         unsigned long flags;
684         int this_cpu;
685
686         local_irq_save(flags);
687         this_cpu = smp_processor_id();
688         __send_remote_softirq(cp, cpu, this_cpu, softirq);
689         local_irq_restore(flags);
690 }
691 EXPORT_SYMBOL(send_remote_softirq);
692
693 static int __cpuinit remote_softirq_cpu_notify(struct notifier_block *self,
694                                                unsigned long action, void *hcpu)
695 {
696         /*
697          * If a CPU goes away, splice its entries to the current CPU
698          * and trigger a run of the softirq
699          */
700         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
701                 int cpu = (unsigned long) hcpu;
702                 int i;
703
704                 local_irq_disable();
705                 for (i = 0; i < NR_SOFTIRQS; i++) {
706                         struct list_head *head = &per_cpu(softirq_work_list[i], cpu);
707                         struct list_head *local_head;
708
709                         if (list_empty(head))
710                                 continue;
711
712                         local_head = &__get_cpu_var(softirq_work_list[i]);
713                         list_splice_init(head, local_head);
714                         raise_softirq_irqoff(i);
715                 }
716                 local_irq_enable();
717         }
718
719         return NOTIFY_OK;
720 }
721
722 static struct notifier_block __cpuinitdata remote_softirq_cpu_notifier = {
723         .notifier_call  = remote_softirq_cpu_notify,
724 };
725
726 void __init softirq_init(void)
727 {
728         int cpu;
729
730         for_each_possible_cpu(cpu) {
731                 int i;
732
733                 per_cpu(tasklet_vec, cpu).tail =
734                         &per_cpu(tasklet_vec, cpu).head;
735                 per_cpu(tasklet_hi_vec, cpu).tail =
736                         &per_cpu(tasklet_hi_vec, cpu).head;
737                 for (i = 0; i < NR_SOFTIRQS; i++)
738                         INIT_LIST_HEAD(&per_cpu(softirq_work_list[i], cpu));
739         }
740
741         register_hotcpu_notifier(&remote_softirq_cpu_notifier);
742
743         open_softirq(TASKLET_SOFTIRQ, tasklet_action);
744         open_softirq(HI_SOFTIRQ, tasklet_hi_action);
745 }
746
747 static int ksoftirqd_should_run(unsigned int cpu)
748 {
749         return local_softirq_pending();
750 }
751
752 static void run_ksoftirqd(unsigned int cpu)
753 {
754         local_irq_disable();
755         if (local_softirq_pending()) {
756                 __do_softirq();
757                 rcu_note_context_switch(cpu);
758                 local_irq_enable();
759                 cond_resched();
760                 return;
761         }
762         local_irq_enable();
763 }
764
765 #ifdef CONFIG_HOTPLUG_CPU
766 /*
767  * tasklet_kill_immediate is called to remove a tasklet which can already be
768  * scheduled for execution on @cpu.
769  *
770  * Unlike tasklet_kill, this function removes the tasklet
771  * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
772  *
773  * When this function is called, @cpu must be in the CPU_DEAD state.
774  */
775 void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
776 {
777         struct tasklet_struct **i;
778
779         BUG_ON(cpu_online(cpu));
780         BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
781
782         if (!test_bit(TASKLET_STATE_SCHED, &t->state))
783                 return;
784
785         /* CPU is dead, so no lock needed. */
786         for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
787                 if (*i == t) {
788                         *i = t->next;
789                         /* If this was the tail element, move the tail ptr */
790                         if (*i == NULL)
791                                 per_cpu(tasklet_vec, cpu).tail = i;
792                         return;
793                 }
794         }
795         BUG();
796 }
797
798 static void takeover_tasklets(unsigned int cpu)
799 {
800         /* CPU is dead, so no lock needed. */
801         local_irq_disable();
802
803         /* Find end, append list for that CPU. */
804         if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
805                 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
806                 this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
807                 per_cpu(tasklet_vec, cpu).head = NULL;
808                 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
809         }
810         raise_softirq_irqoff(TASKLET_SOFTIRQ);
811
812         if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
813                 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
814                 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
815                 per_cpu(tasklet_hi_vec, cpu).head = NULL;
816                 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
817         }
818         raise_softirq_irqoff(HI_SOFTIRQ);
819
820         local_irq_enable();
821 }
822 #endif /* CONFIG_HOTPLUG_CPU */
823
824 static int __cpuinit cpu_callback(struct notifier_block *nfb,
825                                   unsigned long action,
826                                   void *hcpu)
827 {
828         switch (action) {
829 #ifdef CONFIG_HOTPLUG_CPU
830         case CPU_DEAD:
831         case CPU_DEAD_FROZEN:
832                 takeover_tasklets((unsigned long)hcpu);
833                 break;
834 #endif /* CONFIG_HOTPLUG_CPU */
835         }
836         return NOTIFY_OK;
837 }
838
839 static struct notifier_block __cpuinitdata cpu_nfb = {
840         .notifier_call = cpu_callback
841 };
842
843 static struct smp_hotplug_thread softirq_threads = {
844         .store                  = &ksoftirqd,
845         .thread_should_run      = ksoftirqd_should_run,
846         .thread_fn              = run_ksoftirqd,
847         .thread_comm            = "ksoftirqd/%u",
848 };
849
850 static __init int spawn_ksoftirqd(void)
851 {
852         register_cpu_notifier(&cpu_nfb);
853
854         BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
855
856         return 0;
857 }
858 early_initcall(spawn_ksoftirqd);
859
860 /*
861  * [ These __weak aliases are kept in a separate compilation unit, so that
862  *   GCC does not inline them incorrectly. ]
863  */
864
865 int __init __weak early_irq_init(void)
866 {
867         return 0;
868 }
869
870 #ifdef CONFIG_GENERIC_HARDIRQS
871 int __init __weak arch_probe_nr_irqs(void)
872 {
873         return NR_IRQS_LEGACY;
874 }
875
876 int __init __weak arch_early_irq_init(void)
877 {
878         return 0;
879 }
880 #endif