genirq: Provide forced interrupt threading
[linux-2.6.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/module.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/tick.h>
27
28 #define CREATE_TRACE_POINTS
29 #include <trace/events/irq.h>
30
31 #include <asm/irq.h>
32 /*
33    - No shared variables, all the data are CPU local.
34    - If a softirq needs serialization, let it serialize itself
35      by its own spinlocks.
36    - Even if softirq is serialized, only local cpu is marked for
37      execution. Hence, we get something sort of weak cpu binding.
38      Though it is still not clear, will it result in better locality
39      or will not.
40
41    Examples:
42    - NET RX softirq. It is multithreaded and does not require
43      any global serialization.
44    - NET TX softirq. It kicks software netdevice queues, hence
45      it is logically serialized per device, but this serialization
46      is invisible to common code.
47    - Tasklets: serialized wrt itself.
48  */
49
50 #ifndef __ARCH_IRQ_STAT
51 irq_cpustat_t irq_stat[NR_CPUS] ____cacheline_aligned;
52 EXPORT_SYMBOL(irq_stat);
53 #endif
54
55 static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
56
57 static DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
58
59 char *softirq_to_name[NR_SOFTIRQS] = {
60         "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "BLOCK_IOPOLL",
61         "TASKLET", "SCHED", "HRTIMER",  "RCU"
62 };
63
64 /*
65  * we cannot loop indefinitely here to avoid userspace starvation,
66  * but we also don't want to introduce a worst case 1/HZ latency
67  * to the pending events, so lets the scheduler to balance
68  * the softirq load for us.
69  */
70 static void wakeup_softirqd(void)
71 {
72         /* Interrupts are disabled: no need to stop preemption */
73         struct task_struct *tsk = __this_cpu_read(ksoftirqd);
74
75         if (tsk && tsk->state != TASK_RUNNING)
76                 wake_up_process(tsk);
77 }
78
79 /*
80  * preempt_count and SOFTIRQ_OFFSET usage:
81  * - preempt_count is changed by SOFTIRQ_OFFSET on entering or leaving
82  *   softirq processing.
83  * - preempt_count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
84  *   on local_bh_disable or local_bh_enable.
85  * This lets us distinguish between whether we are currently processing
86  * softirq and whether we just have bh disabled.
87  */
88
89 /*
90  * This one is for softirq.c-internal use,
91  * where hardirqs are disabled legitimately:
92  */
93 #ifdef CONFIG_TRACE_IRQFLAGS
94 static void __local_bh_disable(unsigned long ip, unsigned int cnt)
95 {
96         unsigned long flags;
97
98         WARN_ON_ONCE(in_irq());
99
100         raw_local_irq_save(flags);
101         /*
102          * The preempt tracer hooks into add_preempt_count and will break
103          * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
104          * is set and before current->softirq_enabled is cleared.
105          * We must manually increment preempt_count here and manually
106          * call the trace_preempt_off later.
107          */
108         preempt_count() += cnt;
109         /*
110          * Were softirqs turned off above:
111          */
112         if (softirq_count() == cnt)
113                 trace_softirqs_off(ip);
114         raw_local_irq_restore(flags);
115
116         if (preempt_count() == cnt)
117                 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
118 }
119 #else /* !CONFIG_TRACE_IRQFLAGS */
120 static inline void __local_bh_disable(unsigned long ip, unsigned int cnt)
121 {
122         add_preempt_count(cnt);
123         barrier();
124 }
125 #endif /* CONFIG_TRACE_IRQFLAGS */
126
127 void local_bh_disable(void)
128 {
129         __local_bh_disable((unsigned long)__builtin_return_address(0),
130                                 SOFTIRQ_DISABLE_OFFSET);
131 }
132
133 EXPORT_SYMBOL(local_bh_disable);
134
135 static void __local_bh_enable(unsigned int cnt)
136 {
137         WARN_ON_ONCE(in_irq());
138         WARN_ON_ONCE(!irqs_disabled());
139
140         if (softirq_count() == cnt)
141                 trace_softirqs_on((unsigned long)__builtin_return_address(0));
142         sub_preempt_count(cnt);
143 }
144
145 /*
146  * Special-case - softirqs can safely be enabled in
147  * cond_resched_softirq(), or by __do_softirq(),
148  * without processing still-pending softirqs:
149  */
150 void _local_bh_enable(void)
151 {
152         __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
153 }
154
155 EXPORT_SYMBOL(_local_bh_enable);
156
157 static inline void _local_bh_enable_ip(unsigned long ip)
158 {
159         WARN_ON_ONCE(in_irq() || irqs_disabled());
160 #ifdef CONFIG_TRACE_IRQFLAGS
161         local_irq_disable();
162 #endif
163         /*
164          * Are softirqs going to be turned on now:
165          */
166         if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
167                 trace_softirqs_on(ip);
168         /*
169          * Keep preemption disabled until we are done with
170          * softirq processing:
171          */
172         sub_preempt_count(SOFTIRQ_DISABLE_OFFSET - 1);
173
174         if (unlikely(!in_interrupt() && local_softirq_pending()))
175                 do_softirq();
176
177         dec_preempt_count();
178 #ifdef CONFIG_TRACE_IRQFLAGS
179         local_irq_enable();
180 #endif
181         preempt_check_resched();
182 }
183
184 void local_bh_enable(void)
185 {
186         _local_bh_enable_ip((unsigned long)__builtin_return_address(0));
187 }
188 EXPORT_SYMBOL(local_bh_enable);
189
190 void local_bh_enable_ip(unsigned long ip)
191 {
192         _local_bh_enable_ip(ip);
193 }
194 EXPORT_SYMBOL(local_bh_enable_ip);
195
196 /*
197  * We restart softirq processing MAX_SOFTIRQ_RESTART times,
198  * and we fall back to softirqd after that.
199  *
200  * This number has been established via experimentation.
201  * The two things to balance is latency against fairness -
202  * we want to handle softirqs as soon as possible, but they
203  * should not be able to lock up the box.
204  */
205 #define MAX_SOFTIRQ_RESTART 10
206
207 asmlinkage void __do_softirq(void)
208 {
209         struct softirq_action *h;
210         __u32 pending;
211         int max_restart = MAX_SOFTIRQ_RESTART;
212         int cpu;
213
214         pending = local_softirq_pending();
215         account_system_vtime(current);
216
217         __local_bh_disable((unsigned long)__builtin_return_address(0),
218                                 SOFTIRQ_OFFSET);
219         lockdep_softirq_enter();
220
221         cpu = smp_processor_id();
222 restart:
223         /* Reset the pending bitmask before enabling irqs */
224         set_softirq_pending(0);
225
226         local_irq_enable();
227
228         h = softirq_vec;
229
230         do {
231                 if (pending & 1) {
232                         unsigned int vec_nr = h - softirq_vec;
233                         int prev_count = preempt_count();
234
235                         kstat_incr_softirqs_this_cpu(vec_nr);
236
237                         trace_softirq_entry(vec_nr);
238                         h->action(h);
239                         trace_softirq_exit(vec_nr);
240                         if (unlikely(prev_count != preempt_count())) {
241                                 printk(KERN_ERR "huh, entered softirq %u %s %p"
242                                        "with preempt_count %08x,"
243                                        " exited with %08x?\n", vec_nr,
244                                        softirq_to_name[vec_nr], h->action,
245                                        prev_count, preempt_count());
246                                 preempt_count() = prev_count;
247                         }
248
249                         rcu_bh_qs(cpu);
250                 }
251                 h++;
252                 pending >>= 1;
253         } while (pending);
254
255         local_irq_disable();
256
257         pending = local_softirq_pending();
258         if (pending && --max_restart)
259                 goto restart;
260
261         if (pending)
262                 wakeup_softirqd();
263
264         lockdep_softirq_exit();
265
266         account_system_vtime(current);
267         __local_bh_enable(SOFTIRQ_OFFSET);
268 }
269
270 #ifndef __ARCH_HAS_DO_SOFTIRQ
271
272 asmlinkage void do_softirq(void)
273 {
274         __u32 pending;
275         unsigned long flags;
276
277         if (in_interrupt())
278                 return;
279
280         local_irq_save(flags);
281
282         pending = local_softirq_pending();
283
284         if (pending)
285                 __do_softirq();
286
287         local_irq_restore(flags);
288 }
289
290 #endif
291
292 /*
293  * Enter an interrupt context.
294  */
295 void irq_enter(void)
296 {
297         int cpu = smp_processor_id();
298
299         rcu_irq_enter();
300         if (idle_cpu(cpu) && !in_interrupt()) {
301                 /*
302                  * Prevent raise_softirq from needlessly waking up ksoftirqd
303                  * here, as softirq will be serviced on return from interrupt.
304                  */
305                 local_bh_disable();
306                 tick_check_idle(cpu);
307                 _local_bh_enable();
308         }
309
310         __irq_enter();
311 }
312
313 #ifdef __ARCH_IRQ_EXIT_IRQS_DISABLED
314 static inline void invoke_softirq(void)
315 {
316         if (!force_irqthreads)
317                 __do_softirq();
318         else
319                 wakeup_softirqd();
320 }
321 #else
322 static inline void invoke_softirq(void)
323 {
324         if (!force_irqthreads)
325                 do_softirq();
326         else
327                 wakeup_softirqd();
328 }
329 #endif
330
331 /*
332  * Exit an interrupt context. Process softirqs if needed and possible:
333  */
334 void irq_exit(void)
335 {
336         account_system_vtime(current);
337         trace_hardirq_exit();
338         sub_preempt_count(IRQ_EXIT_OFFSET);
339         if (!in_interrupt() && local_softirq_pending())
340                 invoke_softirq();
341
342         rcu_irq_exit();
343 #ifdef CONFIG_NO_HZ
344         /* Make sure that timer wheel updates are propagated */
345         if (idle_cpu(smp_processor_id()) && !in_interrupt() && !need_resched())
346                 tick_nohz_stop_sched_tick(0);
347 #endif
348         preempt_enable_no_resched();
349 }
350
351 /*
352  * This function must run with irqs disabled!
353  */
354 inline void raise_softirq_irqoff(unsigned int nr)
355 {
356         __raise_softirq_irqoff(nr);
357
358         /*
359          * If we're in an interrupt or softirq, we're done
360          * (this also catches softirq-disabled code). We will
361          * actually run the softirq once we return from
362          * the irq or softirq.
363          *
364          * Otherwise we wake up ksoftirqd to make sure we
365          * schedule the softirq soon.
366          */
367         if (!in_interrupt())
368                 wakeup_softirqd();
369 }
370
371 void raise_softirq(unsigned int nr)
372 {
373         unsigned long flags;
374
375         local_irq_save(flags);
376         raise_softirq_irqoff(nr);
377         local_irq_restore(flags);
378 }
379
380 void open_softirq(int nr, void (*action)(struct softirq_action *))
381 {
382         softirq_vec[nr].action = action;
383 }
384
385 /*
386  * Tasklets
387  */
388 struct tasklet_head
389 {
390         struct tasklet_struct *head;
391         struct tasklet_struct **tail;
392 };
393
394 static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
395 static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
396
397 void __tasklet_schedule(struct tasklet_struct *t)
398 {
399         unsigned long flags;
400
401         local_irq_save(flags);
402         t->next = NULL;
403         *__this_cpu_read(tasklet_vec.tail) = t;
404         __this_cpu_write(tasklet_vec.tail, &(t->next));
405         raise_softirq_irqoff(TASKLET_SOFTIRQ);
406         local_irq_restore(flags);
407 }
408
409 EXPORT_SYMBOL(__tasklet_schedule);
410
411 void __tasklet_hi_schedule(struct tasklet_struct *t)
412 {
413         unsigned long flags;
414
415         local_irq_save(flags);
416         t->next = NULL;
417         *__this_cpu_read(tasklet_hi_vec.tail) = t;
418         __this_cpu_write(tasklet_hi_vec.tail,  &(t->next));
419         raise_softirq_irqoff(HI_SOFTIRQ);
420         local_irq_restore(flags);
421 }
422
423 EXPORT_SYMBOL(__tasklet_hi_schedule);
424
425 void __tasklet_hi_schedule_first(struct tasklet_struct *t)
426 {
427         BUG_ON(!irqs_disabled());
428
429         t->next = __this_cpu_read(tasklet_hi_vec.head);
430         __this_cpu_write(tasklet_hi_vec.head, t);
431         __raise_softirq_irqoff(HI_SOFTIRQ);
432 }
433
434 EXPORT_SYMBOL(__tasklet_hi_schedule_first);
435
436 static void tasklet_action(struct softirq_action *a)
437 {
438         struct tasklet_struct *list;
439
440         local_irq_disable();
441         list = __this_cpu_read(tasklet_vec.head);
442         __this_cpu_write(tasklet_vec.head, NULL);
443         __this_cpu_write(tasklet_vec.tail, &__get_cpu_var(tasklet_vec).head);
444         local_irq_enable();
445
446         while (list) {
447                 struct tasklet_struct *t = list;
448
449                 list = list->next;
450
451                 if (tasklet_trylock(t)) {
452                         if (!atomic_read(&t->count)) {
453                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
454                                         BUG();
455                                 t->func(t->data);
456                                 tasklet_unlock(t);
457                                 continue;
458                         }
459                         tasklet_unlock(t);
460                 }
461
462                 local_irq_disable();
463                 t->next = NULL;
464                 *__this_cpu_read(tasklet_vec.tail) = t;
465                 __this_cpu_write(tasklet_vec.tail, &(t->next));
466                 __raise_softirq_irqoff(TASKLET_SOFTIRQ);
467                 local_irq_enable();
468         }
469 }
470
471 static void tasklet_hi_action(struct softirq_action *a)
472 {
473         struct tasklet_struct *list;
474
475         local_irq_disable();
476         list = __this_cpu_read(tasklet_hi_vec.head);
477         __this_cpu_write(tasklet_hi_vec.head, NULL);
478         __this_cpu_write(tasklet_hi_vec.tail, &__get_cpu_var(tasklet_hi_vec).head);
479         local_irq_enable();
480
481         while (list) {
482                 struct tasklet_struct *t = list;
483
484                 list = list->next;
485
486                 if (tasklet_trylock(t)) {
487                         if (!atomic_read(&t->count)) {
488                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
489                                         BUG();
490                                 t->func(t->data);
491                                 tasklet_unlock(t);
492                                 continue;
493                         }
494                         tasklet_unlock(t);
495                 }
496
497                 local_irq_disable();
498                 t->next = NULL;
499                 *__this_cpu_read(tasklet_hi_vec.tail) = t;
500                 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
501                 __raise_softirq_irqoff(HI_SOFTIRQ);
502                 local_irq_enable();
503         }
504 }
505
506
507 void tasklet_init(struct tasklet_struct *t,
508                   void (*func)(unsigned long), unsigned long data)
509 {
510         t->next = NULL;
511         t->state = 0;
512         atomic_set(&t->count, 0);
513         t->func = func;
514         t->data = data;
515 }
516
517 EXPORT_SYMBOL(tasklet_init);
518
519 void tasklet_kill(struct tasklet_struct *t)
520 {
521         if (in_interrupt())
522                 printk("Attempt to kill tasklet from interrupt\n");
523
524         while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
525                 do {
526                         yield();
527                 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
528         }
529         tasklet_unlock_wait(t);
530         clear_bit(TASKLET_STATE_SCHED, &t->state);
531 }
532
533 EXPORT_SYMBOL(tasklet_kill);
534
535 /*
536  * tasklet_hrtimer
537  */
538
539 /*
540  * The trampoline is called when the hrtimer expires. It schedules a tasklet
541  * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
542  * hrtimer callback, but from softirq context.
543  */
544 static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer)
545 {
546         struct tasklet_hrtimer *ttimer =
547                 container_of(timer, struct tasklet_hrtimer, timer);
548
549         tasklet_hi_schedule(&ttimer->tasklet);
550         return HRTIMER_NORESTART;
551 }
552
553 /*
554  * Helper function which calls the hrtimer callback from
555  * tasklet/softirq context
556  */
557 static void __tasklet_hrtimer_trampoline(unsigned long data)
558 {
559         struct tasklet_hrtimer *ttimer = (void *)data;
560         enum hrtimer_restart restart;
561
562         restart = ttimer->function(&ttimer->timer);
563         if (restart != HRTIMER_NORESTART)
564                 hrtimer_restart(&ttimer->timer);
565 }
566
567 /**
568  * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
569  * @ttimer:      tasklet_hrtimer which is initialized
570  * @function:    hrtimer callback funtion which gets called from softirq context
571  * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
572  * @mode:        hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
573  */
574 void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer,
575                           enum hrtimer_restart (*function)(struct hrtimer *),
576                           clockid_t which_clock, enum hrtimer_mode mode)
577 {
578         hrtimer_init(&ttimer->timer, which_clock, mode);
579         ttimer->timer.function = __hrtimer_tasklet_trampoline;
580         tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline,
581                      (unsigned long)ttimer);
582         ttimer->function = function;
583 }
584 EXPORT_SYMBOL_GPL(tasklet_hrtimer_init);
585
586 /*
587  * Remote softirq bits
588  */
589
590 DEFINE_PER_CPU(struct list_head [NR_SOFTIRQS], softirq_work_list);
591 EXPORT_PER_CPU_SYMBOL(softirq_work_list);
592
593 static void __local_trigger(struct call_single_data *cp, int softirq)
594 {
595         struct list_head *head = &__get_cpu_var(softirq_work_list[softirq]);
596
597         list_add_tail(&cp->list, head);
598
599         /* Trigger the softirq only if the list was previously empty.  */
600         if (head->next == &cp->list)
601                 raise_softirq_irqoff(softirq);
602 }
603
604 #ifdef CONFIG_USE_GENERIC_SMP_HELPERS
605 static void remote_softirq_receive(void *data)
606 {
607         struct call_single_data *cp = data;
608         unsigned long flags;
609         int softirq;
610
611         softirq = cp->priv;
612
613         local_irq_save(flags);
614         __local_trigger(cp, softirq);
615         local_irq_restore(flags);
616 }
617
618 static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
619 {
620         if (cpu_online(cpu)) {
621                 cp->func = remote_softirq_receive;
622                 cp->info = cp;
623                 cp->flags = 0;
624                 cp->priv = softirq;
625
626                 __smp_call_function_single(cpu, cp, 0);
627                 return 0;
628         }
629         return 1;
630 }
631 #else /* CONFIG_USE_GENERIC_SMP_HELPERS */
632 static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
633 {
634         return 1;
635 }
636 #endif
637
638 /**
639  * __send_remote_softirq - try to schedule softirq work on a remote cpu
640  * @cp: private SMP call function data area
641  * @cpu: the remote cpu
642  * @this_cpu: the currently executing cpu
643  * @softirq: the softirq for the work
644  *
645  * Attempt to schedule softirq work on a remote cpu.  If this cannot be
646  * done, the work is instead queued up on the local cpu.
647  *
648  * Interrupts must be disabled.
649  */
650 void __send_remote_softirq(struct call_single_data *cp, int cpu, int this_cpu, int softirq)
651 {
652         if (cpu == this_cpu || __try_remote_softirq(cp, cpu, softirq))
653                 __local_trigger(cp, softirq);
654 }
655 EXPORT_SYMBOL(__send_remote_softirq);
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  * @softirq: the softirq for the work
662  *
663  * Like __send_remote_softirq except that disabling interrupts and
664  * computing the current cpu is done for the caller.
665  */
666 void send_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
667 {
668         unsigned long flags;
669         int this_cpu;
670
671         local_irq_save(flags);
672         this_cpu = smp_processor_id();
673         __send_remote_softirq(cp, cpu, this_cpu, softirq);
674         local_irq_restore(flags);
675 }
676 EXPORT_SYMBOL(send_remote_softirq);
677
678 static int __cpuinit remote_softirq_cpu_notify(struct notifier_block *self,
679                                                unsigned long action, void *hcpu)
680 {
681         /*
682          * If a CPU goes away, splice its entries to the current CPU
683          * and trigger a run of the softirq
684          */
685         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
686                 int cpu = (unsigned long) hcpu;
687                 int i;
688
689                 local_irq_disable();
690                 for (i = 0; i < NR_SOFTIRQS; i++) {
691                         struct list_head *head = &per_cpu(softirq_work_list[i], cpu);
692                         struct list_head *local_head;
693
694                         if (list_empty(head))
695                                 continue;
696
697                         local_head = &__get_cpu_var(softirq_work_list[i]);
698                         list_splice_init(head, local_head);
699                         raise_softirq_irqoff(i);
700                 }
701                 local_irq_enable();
702         }
703
704         return NOTIFY_OK;
705 }
706
707 static struct notifier_block __cpuinitdata remote_softirq_cpu_notifier = {
708         .notifier_call  = remote_softirq_cpu_notify,
709 };
710
711 void __init softirq_init(void)
712 {
713         int cpu;
714
715         for_each_possible_cpu(cpu) {
716                 int i;
717
718                 per_cpu(tasklet_vec, cpu).tail =
719                         &per_cpu(tasklet_vec, cpu).head;
720                 per_cpu(tasklet_hi_vec, cpu).tail =
721                         &per_cpu(tasklet_hi_vec, cpu).head;
722                 for (i = 0; i < NR_SOFTIRQS; i++)
723                         INIT_LIST_HEAD(&per_cpu(softirq_work_list[i], cpu));
724         }
725
726         register_hotcpu_notifier(&remote_softirq_cpu_notifier);
727
728         open_softirq(TASKLET_SOFTIRQ, tasklet_action);
729         open_softirq(HI_SOFTIRQ, tasklet_hi_action);
730 }
731
732 static int run_ksoftirqd(void * __bind_cpu)
733 {
734         set_current_state(TASK_INTERRUPTIBLE);
735
736         current->flags |= PF_KSOFTIRQD;
737         while (!kthread_should_stop()) {
738                 preempt_disable();
739                 if (!local_softirq_pending()) {
740                         preempt_enable_no_resched();
741                         schedule();
742                         preempt_disable();
743                 }
744
745                 __set_current_state(TASK_RUNNING);
746
747                 while (local_softirq_pending()) {
748                         /* Preempt disable stops cpu going offline.
749                            If already offline, we'll be on wrong CPU:
750                            don't process */
751                         if (cpu_is_offline((long)__bind_cpu))
752                                 goto wait_to_die;
753                         local_irq_disable();
754                         if (local_softirq_pending())
755                                 __do_softirq();
756                         local_irq_enable();
757                         preempt_enable_no_resched();
758                         cond_resched();
759                         preempt_disable();
760                         rcu_note_context_switch((long)__bind_cpu);
761                 }
762                 preempt_enable();
763                 set_current_state(TASK_INTERRUPTIBLE);
764         }
765         __set_current_state(TASK_RUNNING);
766         return 0;
767
768 wait_to_die:
769         preempt_enable();
770         /* Wait for kthread_stop */
771         set_current_state(TASK_INTERRUPTIBLE);
772         while (!kthread_should_stop()) {
773                 schedule();
774                 set_current_state(TASK_INTERRUPTIBLE);
775         }
776         __set_current_state(TASK_RUNNING);
777         return 0;
778 }
779
780 #ifdef CONFIG_HOTPLUG_CPU
781 /*
782  * tasklet_kill_immediate is called to remove a tasklet which can already be
783  * scheduled for execution on @cpu.
784  *
785  * Unlike tasklet_kill, this function removes the tasklet
786  * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
787  *
788  * When this function is called, @cpu must be in the CPU_DEAD state.
789  */
790 void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
791 {
792         struct tasklet_struct **i;
793
794         BUG_ON(cpu_online(cpu));
795         BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
796
797         if (!test_bit(TASKLET_STATE_SCHED, &t->state))
798                 return;
799
800         /* CPU is dead, so no lock needed. */
801         for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
802                 if (*i == t) {
803                         *i = t->next;
804                         /* If this was the tail element, move the tail ptr */
805                         if (*i == NULL)
806                                 per_cpu(tasklet_vec, cpu).tail = i;
807                         return;
808                 }
809         }
810         BUG();
811 }
812
813 static void takeover_tasklets(unsigned int cpu)
814 {
815         /* CPU is dead, so no lock needed. */
816         local_irq_disable();
817
818         /* Find end, append list for that CPU. */
819         if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
820                 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
821                 this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
822                 per_cpu(tasklet_vec, cpu).head = NULL;
823                 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
824         }
825         raise_softirq_irqoff(TASKLET_SOFTIRQ);
826
827         if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
828                 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
829                 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
830                 per_cpu(tasklet_hi_vec, cpu).head = NULL;
831                 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
832         }
833         raise_softirq_irqoff(HI_SOFTIRQ);
834
835         local_irq_enable();
836 }
837 #endif /* CONFIG_HOTPLUG_CPU */
838
839 static int __cpuinit cpu_callback(struct notifier_block *nfb,
840                                   unsigned long action,
841                                   void *hcpu)
842 {
843         int hotcpu = (unsigned long)hcpu;
844         struct task_struct *p;
845
846         switch (action) {
847         case CPU_UP_PREPARE:
848         case CPU_UP_PREPARE_FROZEN:
849                 p = kthread_create(run_ksoftirqd, hcpu, "ksoftirqd/%d", hotcpu);
850                 if (IS_ERR(p)) {
851                         printk("ksoftirqd for %i failed\n", hotcpu);
852                         return notifier_from_errno(PTR_ERR(p));
853                 }
854                 kthread_bind(p, hotcpu);
855                 per_cpu(ksoftirqd, hotcpu) = p;
856                 break;
857         case CPU_ONLINE:
858         case CPU_ONLINE_FROZEN:
859                 wake_up_process(per_cpu(ksoftirqd, hotcpu));
860                 break;
861 #ifdef CONFIG_HOTPLUG_CPU
862         case CPU_UP_CANCELED:
863         case CPU_UP_CANCELED_FROZEN:
864                 if (!per_cpu(ksoftirqd, hotcpu))
865                         break;
866                 /* Unbind so it can run.  Fall thru. */
867                 kthread_bind(per_cpu(ksoftirqd, hotcpu),
868                              cpumask_any(cpu_online_mask));
869         case CPU_DEAD:
870         case CPU_DEAD_FROZEN: {
871                 static const struct sched_param param = {
872                         .sched_priority = MAX_RT_PRIO-1
873                 };
874
875                 p = per_cpu(ksoftirqd, hotcpu);
876                 per_cpu(ksoftirqd, hotcpu) = NULL;
877                 sched_setscheduler_nocheck(p, SCHED_FIFO, &param);
878                 kthread_stop(p);
879                 takeover_tasklets(hotcpu);
880                 break;
881         }
882 #endif /* CONFIG_HOTPLUG_CPU */
883         }
884         return NOTIFY_OK;
885 }
886
887 static struct notifier_block __cpuinitdata cpu_nfb = {
888         .notifier_call = cpu_callback
889 };
890
891 static __init int spawn_ksoftirqd(void)
892 {
893         void *cpu = (void *)(long)smp_processor_id();
894         int err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
895
896         BUG_ON(err != NOTIFY_OK);
897         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
898         register_cpu_notifier(&cpu_nfb);
899         return 0;
900 }
901 early_initcall(spawn_ksoftirqd);
902
903 /*
904  * [ These __weak aliases are kept in a separate compilation unit, so that
905  *   GCC does not inline them incorrectly. ]
906  */
907
908 int __init __weak early_irq_init(void)
909 {
910         return 0;
911 }
912
913 #ifdef CONFIG_GENERIC_HARDIRQS
914 int __init __weak arch_probe_nr_irqs(void)
915 {
916         return NR_IRQS_LEGACY;
917 }
918
919 int __init __weak arch_early_irq_init(void)
920 {
921         return 0;
922 }
923 #endif