Merge branch 'linux-2.6' into for-2.6.22
[linux-2.6] / kernel / softirq.c
1 /*
2  *      linux/kernel/softirq.c
3  *
4  *      Copyright (C) 1992 Linus Torvalds
5  *
6  * Rewritten. Old one was good in 2.2, but in 2.3 it was immoral. --ANK (990903)
7  */
8
9 #include <linux/module.h>
10 #include <linux/kernel_stat.h>
11 #include <linux/interrupt.h>
12 #include <linux/init.h>
13 #include <linux/mm.h>
14 #include <linux/notifier.h>
15 #include <linux/percpu.h>
16 #include <linux/cpu.h>
17 #include <linux/kthread.h>
18 #include <linux/rcupdate.h>
19 #include <linux/smp.h>
20 #include <linux/tick.h>
21
22 #include <asm/irq.h>
23 /*
24    - No shared variables, all the data are CPU local.
25    - If a softirq needs serialization, let it serialize itself
26      by its own spinlocks.
27    - Even if softirq is serialized, only local cpu is marked for
28      execution. Hence, we get something sort of weak cpu binding.
29      Though it is still not clear, will it result in better locality
30      or will not.
31
32    Examples:
33    - NET RX softirq. It is multithreaded and does not require
34      any global serialization.
35    - NET TX softirq. It kicks software netdevice queues, hence
36      it is logically serialized per device, but this serialization
37      is invisible to common code.
38    - Tasklets: serialized wrt itself.
39  */
40
41 #ifndef __ARCH_IRQ_STAT
42 irq_cpustat_t irq_stat[NR_CPUS] ____cacheline_aligned;
43 EXPORT_SYMBOL(irq_stat);
44 #endif
45
46 static struct softirq_action softirq_vec[32] __cacheline_aligned_in_smp;
47
48 static DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
49
50 /*
51  * we cannot loop indefinitely here to avoid userspace starvation,
52  * but we also don't want to introduce a worst case 1/HZ latency
53  * to the pending events, so lets the scheduler to balance
54  * the softirq load for us.
55  */
56 static inline void wakeup_softirqd(void)
57 {
58         /* Interrupts are disabled: no need to stop preemption */
59         struct task_struct *tsk = __get_cpu_var(ksoftirqd);
60
61         if (tsk && tsk->state != TASK_RUNNING)
62                 wake_up_process(tsk);
63 }
64
65 /*
66  * This one is for softirq.c-internal use,
67  * where hardirqs are disabled legitimately:
68  */
69 #ifdef CONFIG_TRACE_IRQFLAGS
70 static void __local_bh_disable(unsigned long ip)
71 {
72         unsigned long flags;
73
74         WARN_ON_ONCE(in_irq());
75
76         raw_local_irq_save(flags);
77         add_preempt_count(SOFTIRQ_OFFSET);
78         /*
79          * Were softirqs turned off above:
80          */
81         if (softirq_count() == SOFTIRQ_OFFSET)
82                 trace_softirqs_off(ip);
83         raw_local_irq_restore(flags);
84 }
85 #else /* !CONFIG_TRACE_IRQFLAGS */
86 static inline void __local_bh_disable(unsigned long ip)
87 {
88         add_preempt_count(SOFTIRQ_OFFSET);
89         barrier();
90 }
91 #endif /* CONFIG_TRACE_IRQFLAGS */
92
93 void local_bh_disable(void)
94 {
95         __local_bh_disable((unsigned long)__builtin_return_address(0));
96 }
97
98 EXPORT_SYMBOL(local_bh_disable);
99
100 void __local_bh_enable(void)
101 {
102         WARN_ON_ONCE(in_irq());
103
104         /*
105          * softirqs should never be enabled by __local_bh_enable(),
106          * it always nests inside local_bh_enable() sections:
107          */
108         WARN_ON_ONCE(softirq_count() == SOFTIRQ_OFFSET);
109
110         sub_preempt_count(SOFTIRQ_OFFSET);
111 }
112 EXPORT_SYMBOL_GPL(__local_bh_enable);
113
114 /*
115  * Special-case - softirqs can safely be enabled in
116  * cond_resched_softirq(), or by __do_softirq(),
117  * without processing still-pending softirqs:
118  */
119 void _local_bh_enable(void)
120 {
121         WARN_ON_ONCE(in_irq());
122         WARN_ON_ONCE(!irqs_disabled());
123
124         if (softirq_count() == SOFTIRQ_OFFSET)
125                 trace_softirqs_on((unsigned long)__builtin_return_address(0));
126         sub_preempt_count(SOFTIRQ_OFFSET);
127 }
128
129 EXPORT_SYMBOL(_local_bh_enable);
130
131 void local_bh_enable(void)
132 {
133 #ifdef CONFIG_TRACE_IRQFLAGS
134         unsigned long flags;
135
136         WARN_ON_ONCE(in_irq());
137 #endif
138         WARN_ON_ONCE(irqs_disabled());
139
140 #ifdef CONFIG_TRACE_IRQFLAGS
141         local_irq_save(flags);
142 #endif
143         /*
144          * Are softirqs going to be turned on now:
145          */
146         if (softirq_count() == SOFTIRQ_OFFSET)
147                 trace_softirqs_on((unsigned long)__builtin_return_address(0));
148         /*
149          * Keep preemption disabled until we are done with
150          * softirq processing:
151          */
152         sub_preempt_count(SOFTIRQ_OFFSET - 1);
153
154         if (unlikely(!in_interrupt() && local_softirq_pending()))
155                 do_softirq();
156
157         dec_preempt_count();
158 #ifdef CONFIG_TRACE_IRQFLAGS
159         local_irq_restore(flags);
160 #endif
161         preempt_check_resched();
162 }
163 EXPORT_SYMBOL(local_bh_enable);
164
165 void local_bh_enable_ip(unsigned long ip)
166 {
167 #ifdef CONFIG_TRACE_IRQFLAGS
168         unsigned long flags;
169
170         WARN_ON_ONCE(in_irq());
171
172         local_irq_save(flags);
173 #endif
174         /*
175          * Are softirqs going to be turned on now:
176          */
177         if (softirq_count() == SOFTIRQ_OFFSET)
178                 trace_softirqs_on(ip);
179         /*
180          * Keep preemption disabled until we are done with
181          * softirq processing:
182          */
183         sub_preempt_count(SOFTIRQ_OFFSET - 1);
184
185         if (unlikely(!in_interrupt() && local_softirq_pending()))
186                 do_softirq();
187
188         dec_preempt_count();
189 #ifdef CONFIG_TRACE_IRQFLAGS
190         local_irq_restore(flags);
191 #endif
192         preempt_check_resched();
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         trace_softirq_enter();
219
220         cpu = smp_processor_id();
221 restart:
222         /* Reset the pending bitmask before enabling irqs */
223         set_softirq_pending(0);
224
225         local_irq_enable();
226
227         h = softirq_vec;
228
229         do {
230                 if (pending & 1) {
231                         h->action(h);
232                         rcu_bh_qsctr_inc(cpu);
233                 }
234                 h++;
235                 pending >>= 1;
236         } while (pending);
237
238         local_irq_disable();
239
240         pending = local_softirq_pending();
241         if (pending && --max_restart)
242                 goto restart;
243
244         if (pending)
245                 wakeup_softirqd();
246
247         trace_softirq_exit();
248
249         account_system_vtime(current);
250         _local_bh_enable();
251 }
252
253 #ifndef __ARCH_HAS_DO_SOFTIRQ
254
255 asmlinkage void do_softirq(void)
256 {
257         __u32 pending;
258         unsigned long flags;
259
260         if (in_interrupt())
261                 return;
262
263         local_irq_save(flags);
264
265         pending = local_softirq_pending();
266
267         if (pending)
268                 __do_softirq();
269
270         local_irq_restore(flags);
271 }
272
273 EXPORT_SYMBOL(do_softirq);
274
275 #endif
276
277 /*
278  * Enter an interrupt context.
279  */
280 void irq_enter(void)
281 {
282         __irq_enter();
283 #ifdef CONFIG_NO_HZ
284         if (idle_cpu(smp_processor_id()))
285                 tick_nohz_update_jiffies();
286 #endif
287 }
288
289 #ifdef __ARCH_IRQ_EXIT_IRQS_DISABLED
290 # define invoke_softirq()       __do_softirq()
291 #else
292 # define invoke_softirq()       do_softirq()
293 #endif
294
295 /*
296  * Exit an interrupt context. Process softirqs if needed and possible:
297  */
298 void irq_exit(void)
299 {
300         account_system_vtime(current);
301         trace_hardirq_exit();
302         sub_preempt_count(IRQ_EXIT_OFFSET);
303         if (!in_interrupt() && local_softirq_pending())
304                 invoke_softirq();
305
306 #ifdef CONFIG_NO_HZ
307         /* Make sure that timer wheel updates are propagated */
308         if (!in_interrupt() && idle_cpu(smp_processor_id()) && !need_resched())
309                 tick_nohz_stop_sched_tick();
310 #endif
311         preempt_enable_no_resched();
312 }
313
314 /*
315  * This function must run with irqs disabled!
316  */
317 inline fastcall void raise_softirq_irqoff(unsigned int nr)
318 {
319         __raise_softirq_irqoff(nr);
320
321         /*
322          * If we're in an interrupt or softirq, we're done
323          * (this also catches softirq-disabled code). We will
324          * actually run the softirq once we return from
325          * the irq or softirq.
326          *
327          * Otherwise we wake up ksoftirqd to make sure we
328          * schedule the softirq soon.
329          */
330         if (!in_interrupt())
331                 wakeup_softirqd();
332 }
333
334 EXPORT_SYMBOL(raise_softirq_irqoff);
335
336 void fastcall raise_softirq(unsigned int nr)
337 {
338         unsigned long flags;
339
340         local_irq_save(flags);
341         raise_softirq_irqoff(nr);
342         local_irq_restore(flags);
343 }
344
345 void open_softirq(int nr, void (*action)(struct softirq_action*), void *data)
346 {
347         softirq_vec[nr].data = data;
348         softirq_vec[nr].action = action;
349 }
350
351 /* Tasklets */
352 struct tasklet_head
353 {
354         struct tasklet_struct *list;
355 };
356
357 /* Some compilers disobey section attribute on statics when not
358    initialized -- RR */
359 static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec) = { NULL };
360 static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec) = { NULL };
361
362 void fastcall __tasklet_schedule(struct tasklet_struct *t)
363 {
364         unsigned long flags;
365
366         local_irq_save(flags);
367         t->next = __get_cpu_var(tasklet_vec).list;
368         __get_cpu_var(tasklet_vec).list = t;
369         raise_softirq_irqoff(TASKLET_SOFTIRQ);
370         local_irq_restore(flags);
371 }
372
373 EXPORT_SYMBOL(__tasklet_schedule);
374
375 void fastcall __tasklet_hi_schedule(struct tasklet_struct *t)
376 {
377         unsigned long flags;
378
379         local_irq_save(flags);
380         t->next = __get_cpu_var(tasklet_hi_vec).list;
381         __get_cpu_var(tasklet_hi_vec).list = t;
382         raise_softirq_irqoff(HI_SOFTIRQ);
383         local_irq_restore(flags);
384 }
385
386 EXPORT_SYMBOL(__tasklet_hi_schedule);
387
388 static void tasklet_action(struct softirq_action *a)
389 {
390         struct tasklet_struct *list;
391
392         local_irq_disable();
393         list = __get_cpu_var(tasklet_vec).list;
394         __get_cpu_var(tasklet_vec).list = NULL;
395         local_irq_enable();
396
397         while (list) {
398                 struct tasklet_struct *t = list;
399
400                 list = list->next;
401
402                 if (tasklet_trylock(t)) {
403                         if (!atomic_read(&t->count)) {
404                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
405                                         BUG();
406                                 t->func(t->data);
407                                 tasklet_unlock(t);
408                                 continue;
409                         }
410                         tasklet_unlock(t);
411                 }
412
413                 local_irq_disable();
414                 t->next = __get_cpu_var(tasklet_vec).list;
415                 __get_cpu_var(tasklet_vec).list = t;
416                 __raise_softirq_irqoff(TASKLET_SOFTIRQ);
417                 local_irq_enable();
418         }
419 }
420
421 static void tasklet_hi_action(struct softirq_action *a)
422 {
423         struct tasklet_struct *list;
424
425         local_irq_disable();
426         list = __get_cpu_var(tasklet_hi_vec).list;
427         __get_cpu_var(tasklet_hi_vec).list = NULL;
428         local_irq_enable();
429
430         while (list) {
431                 struct tasklet_struct *t = list;
432
433                 list = list->next;
434
435                 if (tasklet_trylock(t)) {
436                         if (!atomic_read(&t->count)) {
437                                 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
438                                         BUG();
439                                 t->func(t->data);
440                                 tasklet_unlock(t);
441                                 continue;
442                         }
443                         tasklet_unlock(t);
444                 }
445
446                 local_irq_disable();
447                 t->next = __get_cpu_var(tasklet_hi_vec).list;
448                 __get_cpu_var(tasklet_hi_vec).list = t;
449                 __raise_softirq_irqoff(HI_SOFTIRQ);
450                 local_irq_enable();
451         }
452 }
453
454
455 void tasklet_init(struct tasklet_struct *t,
456                   void (*func)(unsigned long), unsigned long data)
457 {
458         t->next = NULL;
459         t->state = 0;
460         atomic_set(&t->count, 0);
461         t->func = func;
462         t->data = data;
463 }
464
465 EXPORT_SYMBOL(tasklet_init);
466
467 void tasklet_kill(struct tasklet_struct *t)
468 {
469         if (in_interrupt())
470                 printk("Attempt to kill tasklet from interrupt\n");
471
472         while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
473                 do
474                         yield();
475                 while (test_bit(TASKLET_STATE_SCHED, &t->state));
476         }
477         tasklet_unlock_wait(t);
478         clear_bit(TASKLET_STATE_SCHED, &t->state);
479 }
480
481 EXPORT_SYMBOL(tasklet_kill);
482
483 void __init softirq_init(void)
484 {
485         open_softirq(TASKLET_SOFTIRQ, tasklet_action, NULL);
486         open_softirq(HI_SOFTIRQ, tasklet_hi_action, NULL);
487 }
488
489 static int ksoftirqd(void * __bind_cpu)
490 {
491         set_user_nice(current, 19);
492         current->flags |= PF_NOFREEZE;
493
494         set_current_state(TASK_INTERRUPTIBLE);
495
496         while (!kthread_should_stop()) {
497                 preempt_disable();
498                 if (!local_softirq_pending()) {
499                         preempt_enable_no_resched();
500                         schedule();
501                         preempt_disable();
502                 }
503
504                 __set_current_state(TASK_RUNNING);
505
506                 while (local_softirq_pending()) {
507                         /* Preempt disable stops cpu going offline.
508                            If already offline, we'll be on wrong CPU:
509                            don't process */
510                         if (cpu_is_offline((long)__bind_cpu))
511                                 goto wait_to_die;
512                         do_softirq();
513                         preempt_enable_no_resched();
514                         cond_resched();
515                         preempt_disable();
516                 }
517                 preempt_enable();
518                 set_current_state(TASK_INTERRUPTIBLE);
519         }
520         __set_current_state(TASK_RUNNING);
521         return 0;
522
523 wait_to_die:
524         preempt_enable();
525         /* Wait for kthread_stop */
526         set_current_state(TASK_INTERRUPTIBLE);
527         while (!kthread_should_stop()) {
528                 schedule();
529                 set_current_state(TASK_INTERRUPTIBLE);
530         }
531         __set_current_state(TASK_RUNNING);
532         return 0;
533 }
534
535 #ifdef CONFIG_HOTPLUG_CPU
536 /*
537  * tasklet_kill_immediate is called to remove a tasklet which can already be
538  * scheduled for execution on @cpu.
539  *
540  * Unlike tasklet_kill, this function removes the tasklet
541  * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
542  *
543  * When this function is called, @cpu must be in the CPU_DEAD state.
544  */
545 void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
546 {
547         struct tasklet_struct **i;
548
549         BUG_ON(cpu_online(cpu));
550         BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
551
552         if (!test_bit(TASKLET_STATE_SCHED, &t->state))
553                 return;
554
555         /* CPU is dead, so no lock needed. */
556         for (i = &per_cpu(tasklet_vec, cpu).list; *i; i = &(*i)->next) {
557                 if (*i == t) {
558                         *i = t->next;
559                         return;
560                 }
561         }
562         BUG();
563 }
564
565 static void takeover_tasklets(unsigned int cpu)
566 {
567         struct tasklet_struct **i;
568
569         /* CPU is dead, so no lock needed. */
570         local_irq_disable();
571
572         /* Find end, append list for that CPU. */
573         for (i = &__get_cpu_var(tasklet_vec).list; *i; i = &(*i)->next);
574         *i = per_cpu(tasklet_vec, cpu).list;
575         per_cpu(tasklet_vec, cpu).list = NULL;
576         raise_softirq_irqoff(TASKLET_SOFTIRQ);
577
578         for (i = &__get_cpu_var(tasklet_hi_vec).list; *i; i = &(*i)->next);
579         *i = per_cpu(tasklet_hi_vec, cpu).list;
580         per_cpu(tasklet_hi_vec, cpu).list = NULL;
581         raise_softirq_irqoff(HI_SOFTIRQ);
582
583         local_irq_enable();
584 }
585 #endif /* CONFIG_HOTPLUG_CPU */
586
587 static int __cpuinit cpu_callback(struct notifier_block *nfb,
588                                   unsigned long action,
589                                   void *hcpu)
590 {
591         int hotcpu = (unsigned long)hcpu;
592         struct task_struct *p;
593
594         switch (action) {
595         case CPU_UP_PREPARE:
596                 p = kthread_create(ksoftirqd, hcpu, "ksoftirqd/%d", hotcpu);
597                 if (IS_ERR(p)) {
598                         printk("ksoftirqd for %i failed\n", hotcpu);
599                         return NOTIFY_BAD;
600                 }
601                 kthread_bind(p, hotcpu);
602                 per_cpu(ksoftirqd, hotcpu) = p;
603                 break;
604         case CPU_ONLINE:
605                 wake_up_process(per_cpu(ksoftirqd, hotcpu));
606                 break;
607 #ifdef CONFIG_HOTPLUG_CPU
608         case CPU_UP_CANCELED:
609                 if (!per_cpu(ksoftirqd, hotcpu))
610                         break;
611                 /* Unbind so it can run.  Fall thru. */
612                 kthread_bind(per_cpu(ksoftirqd, hotcpu),
613                              any_online_cpu(cpu_online_map));
614         case CPU_DEAD:
615                 p = per_cpu(ksoftirqd, hotcpu);
616                 per_cpu(ksoftirqd, hotcpu) = NULL;
617                 kthread_stop(p);
618                 takeover_tasklets(hotcpu);
619                 break;
620 #endif /* CONFIG_HOTPLUG_CPU */
621         }
622         return NOTIFY_OK;
623 }
624
625 static struct notifier_block __cpuinitdata cpu_nfb = {
626         .notifier_call = cpu_callback
627 };
628
629 __init int spawn_ksoftirqd(void)
630 {
631         void *cpu = (void *)(long)smp_processor_id();
632         int err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
633
634         BUG_ON(err == NOTIFY_BAD);
635         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
636         register_cpu_notifier(&cpu_nfb);
637         return 0;
638 }
639
640 #ifdef CONFIG_SMP
641 /*
642  * Call a function on all processors
643  */
644 int on_each_cpu(void (*func) (void *info), void *info, int retry, int wait)
645 {
646         int ret = 0;
647
648         preempt_disable();
649         ret = smp_call_function(func, info, retry, wait);
650         local_irq_disable();
651         func(info);
652         local_irq_enable();
653         preempt_enable();
654         return ret;
655 }
656 EXPORT_SYMBOL(on_each_cpu);
657 #endif