2 * linux/kernel/softirq.c
4 * Copyright (C) 1992 Linus Torvalds
6 * Distribute under GPLv2.
8 * Rewritten. Old one was good in 2.2, but in 2.3 it was immoral. --ANK (990903)
10 * Remote softirq infrastructure is by Jens Axboe.
13 #include <linux/export.h>
14 #include <linux/kernel_stat.h>
15 #include <linux/interrupt.h>
16 #include <linux/init.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>
29 #define CREATE_TRACE_POINTS
30 #include <trace/events/irq.h>
34 - No shared variables, all the data are CPU local.
35 - If a softirq needs serialization, let it serialize itself
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
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.
51 #ifndef __ARCH_IRQ_STAT
52 irq_cpustat_t irq_stat[NR_CPUS] ____cacheline_aligned;
53 EXPORT_SYMBOL(irq_stat);
56 static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
58 DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
60 char *softirq_to_name[NR_SOFTIRQS] = {
61 "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "BLOCK_IOPOLL",
62 "TASKLET", "SCHED", "HRTIMER", "RCU"
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.
71 static void wakeup_softirqd(void)
73 /* Interrupts are disabled: no need to stop preemption */
74 struct task_struct *tsk = __this_cpu_read(ksoftirqd);
76 if (tsk && tsk->state != TASK_RUNNING)
81 * preempt_count and SOFTIRQ_OFFSET usage:
82 * - preempt_count is changed by SOFTIRQ_OFFSET on entering or leaving
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.
91 * This one is for softirq.c-internal use,
92 * where hardirqs are disabled legitimately:
94 #ifdef CONFIG_TRACE_IRQFLAGS
95 static void __local_bh_disable(unsigned long ip, unsigned int cnt)
99 WARN_ON_ONCE(in_irq());
101 raw_local_irq_save(flags);
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.
109 preempt_count() += cnt;
111 * Were softirqs turned off above:
113 if (softirq_count() == cnt)
114 trace_softirqs_off(ip);
115 raw_local_irq_restore(flags);
117 if (preempt_count() == cnt)
118 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
120 #else /* !CONFIG_TRACE_IRQFLAGS */
121 static inline void __local_bh_disable(unsigned long ip, unsigned int cnt)
123 add_preempt_count(cnt);
126 #endif /* CONFIG_TRACE_IRQFLAGS */
128 void local_bh_disable(void)
130 __local_bh_disable(_RET_IP_, SOFTIRQ_DISABLE_OFFSET);
133 EXPORT_SYMBOL(local_bh_disable);
135 static void __local_bh_enable(unsigned int cnt)
137 WARN_ON_ONCE(in_irq());
138 WARN_ON_ONCE(!irqs_disabled());
140 if (softirq_count() == cnt)
141 trace_softirqs_on(_RET_IP_);
142 sub_preempt_count(cnt);
146 * Special-case - softirqs can safely be enabled in
147 * cond_resched_softirq(), or by __do_softirq(),
148 * without processing still-pending softirqs:
150 void _local_bh_enable(void)
152 __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
155 EXPORT_SYMBOL(_local_bh_enable);
157 static inline void _local_bh_enable_ip(unsigned long ip)
159 WARN_ON_ONCE(in_irq() || irqs_disabled());
160 #ifdef CONFIG_TRACE_IRQFLAGS
164 * Are softirqs going to be turned on now:
166 if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
167 trace_softirqs_on(ip);
169 * Keep preemption disabled until we are done with
170 * softirq processing:
172 sub_preempt_count(SOFTIRQ_DISABLE_OFFSET - 1);
174 if (unlikely(!in_interrupt() && local_softirq_pending()))
178 #ifdef CONFIG_TRACE_IRQFLAGS
181 preempt_check_resched();
184 void local_bh_enable(void)
186 _local_bh_enable_ip(_RET_IP_);
188 EXPORT_SYMBOL(local_bh_enable);
190 void local_bh_enable_ip(unsigned long ip)
192 _local_bh_enable_ip(ip);
194 EXPORT_SYMBOL(local_bh_enable_ip);
197 * We restart softirq processing for at most 2 ms,
198 * and if need_resched() is not set.
200 * These limits have 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.
205 #define MAX_SOFTIRQ_TIME msecs_to_jiffies(2)
207 asmlinkage void __do_softirq(void)
209 struct softirq_action *h;
211 unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
213 unsigned long old_flags = current->flags;
216 * Mask out PF_MEMALLOC s current task context is borrowed for the
217 * softirq. A softirq handled such as network RX might set PF_MEMALLOC
218 * again if the socket is related to swap
220 current->flags &= ~PF_MEMALLOC;
222 pending = local_softirq_pending();
223 account_irq_enter_time(current);
225 __local_bh_disable(_RET_IP_, SOFTIRQ_OFFSET);
226 lockdep_softirq_enter();
228 cpu = smp_processor_id();
230 /* Reset the pending bitmask before enabling irqs */
231 set_softirq_pending(0);
239 unsigned int vec_nr = h - softirq_vec;
240 int prev_count = preempt_count();
242 kstat_incr_softirqs_this_cpu(vec_nr);
244 trace_softirq_entry(vec_nr);
246 trace_softirq_exit(vec_nr);
247 if (unlikely(prev_count != preempt_count())) {
248 printk(KERN_ERR "huh, entered softirq %u %s %p"
249 "with preempt_count %08x,"
250 " exited with %08x?\n", vec_nr,
251 softirq_to_name[vec_nr], h->action,
252 prev_count, preempt_count());
253 preempt_count() = prev_count;
264 pending = local_softirq_pending();
266 if (time_before(jiffies, end) && !need_resched())
272 lockdep_softirq_exit();
274 account_irq_exit_time(current);
275 __local_bh_enable(SOFTIRQ_OFFSET);
276 tsk_restore_flags(current, old_flags, PF_MEMALLOC);
279 #ifndef __ARCH_HAS_DO_SOFTIRQ
281 asmlinkage void do_softirq(void)
289 local_irq_save(flags);
291 pending = local_softirq_pending();
296 local_irq_restore(flags);
302 * Enter an interrupt context.
306 int cpu = smp_processor_id();
309 if (is_idle_task(current) && !in_interrupt()) {
311 * Prevent raise_softirq from needlessly waking up ksoftirqd
312 * here, as softirq will be serviced on return from interrupt.
315 tick_check_idle(cpu);
322 static inline void invoke_softirq(void)
324 if (!force_irqthreads)
330 static inline void tick_irq_exit(void)
332 #ifdef CONFIG_NO_HZ_COMMON
333 int cpu = smp_processor_id();
335 /* Make sure that timer wheel updates are propagated */
336 if ((idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
338 tick_nohz_irq_exit();
344 * Exit an interrupt context. Process softirqs if needed and possible:
348 #ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
351 WARN_ON_ONCE(!irqs_disabled());
354 account_irq_exit_time(current);
355 trace_hardirq_exit();
356 sub_preempt_count(HARDIRQ_OFFSET);
357 if (!in_interrupt() && local_softirq_pending())
365 * This function must run with irqs disabled!
367 inline void raise_softirq_irqoff(unsigned int nr)
369 __raise_softirq_irqoff(nr);
372 * If we're in an interrupt or softirq, we're done
373 * (this also catches softirq-disabled code). We will
374 * actually run the softirq once we return from
375 * the irq or softirq.
377 * Otherwise we wake up ksoftirqd to make sure we
378 * schedule the softirq soon.
384 void raise_softirq(unsigned int nr)
388 local_irq_save(flags);
389 raise_softirq_irqoff(nr);
390 local_irq_restore(flags);
393 void __raise_softirq_irqoff(unsigned int nr)
395 trace_softirq_raise(nr);
396 or_softirq_pending(1UL << nr);
399 void open_softirq(int nr, void (*action)(struct softirq_action *))
401 softirq_vec[nr].action = action;
409 struct tasklet_struct *head;
410 struct tasklet_struct **tail;
413 static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
414 static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
416 void __tasklet_schedule(struct tasklet_struct *t)
420 local_irq_save(flags);
422 *__this_cpu_read(tasklet_vec.tail) = t;
423 __this_cpu_write(tasklet_vec.tail, &(t->next));
424 raise_softirq_irqoff(TASKLET_SOFTIRQ);
425 local_irq_restore(flags);
428 EXPORT_SYMBOL(__tasklet_schedule);
430 void __tasklet_hi_schedule(struct tasklet_struct *t)
434 local_irq_save(flags);
436 *__this_cpu_read(tasklet_hi_vec.tail) = t;
437 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
438 raise_softirq_irqoff(HI_SOFTIRQ);
439 local_irq_restore(flags);
442 EXPORT_SYMBOL(__tasklet_hi_schedule);
444 void __tasklet_hi_schedule_first(struct tasklet_struct *t)
446 BUG_ON(!irqs_disabled());
448 t->next = __this_cpu_read(tasklet_hi_vec.head);
449 __this_cpu_write(tasklet_hi_vec.head, t);
450 __raise_softirq_irqoff(HI_SOFTIRQ);
453 EXPORT_SYMBOL(__tasklet_hi_schedule_first);
455 static void tasklet_action(struct softirq_action *a)
457 struct tasklet_struct *list;
460 list = __this_cpu_read(tasklet_vec.head);
461 __this_cpu_write(tasklet_vec.head, NULL);
462 __this_cpu_write(tasklet_vec.tail, &__get_cpu_var(tasklet_vec).head);
466 struct tasklet_struct *t = list;
470 if (tasklet_trylock(t)) {
471 if (!atomic_read(&t->count)) {
472 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
483 *__this_cpu_read(tasklet_vec.tail) = t;
484 __this_cpu_write(tasklet_vec.tail, &(t->next));
485 __raise_softirq_irqoff(TASKLET_SOFTIRQ);
490 static void tasklet_hi_action(struct softirq_action *a)
492 struct tasklet_struct *list;
495 list = __this_cpu_read(tasklet_hi_vec.head);
496 __this_cpu_write(tasklet_hi_vec.head, NULL);
497 __this_cpu_write(tasklet_hi_vec.tail, &__get_cpu_var(tasklet_hi_vec).head);
501 struct tasklet_struct *t = list;
505 if (tasklet_trylock(t)) {
506 if (!atomic_read(&t->count)) {
507 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
518 *__this_cpu_read(tasklet_hi_vec.tail) = t;
519 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
520 __raise_softirq_irqoff(HI_SOFTIRQ);
526 void tasklet_init(struct tasklet_struct *t,
527 void (*func)(unsigned long), unsigned long data)
531 atomic_set(&t->count, 0);
536 EXPORT_SYMBOL(tasklet_init);
538 void tasklet_kill(struct tasklet_struct *t)
541 printk("Attempt to kill tasklet from interrupt\n");
543 while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
546 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
548 tasklet_unlock_wait(t);
549 clear_bit(TASKLET_STATE_SCHED, &t->state);
552 EXPORT_SYMBOL(tasklet_kill);
559 * The trampoline is called when the hrtimer expires. It schedules a tasklet
560 * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
561 * hrtimer callback, but from softirq context.
563 static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer)
565 struct tasklet_hrtimer *ttimer =
566 container_of(timer, struct tasklet_hrtimer, timer);
568 tasklet_hi_schedule(&ttimer->tasklet);
569 return HRTIMER_NORESTART;
573 * Helper function which calls the hrtimer callback from
574 * tasklet/softirq context
576 static void __tasklet_hrtimer_trampoline(unsigned long data)
578 struct tasklet_hrtimer *ttimer = (void *)data;
579 enum hrtimer_restart restart;
581 restart = ttimer->function(&ttimer->timer);
582 if (restart != HRTIMER_NORESTART)
583 hrtimer_restart(&ttimer->timer);
587 * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
588 * @ttimer: tasklet_hrtimer which is initialized
589 * @function: hrtimer callback function which gets called from softirq context
590 * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
591 * @mode: hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
593 void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer,
594 enum hrtimer_restart (*function)(struct hrtimer *),
595 clockid_t which_clock, enum hrtimer_mode mode)
597 hrtimer_init(&ttimer->timer, which_clock, mode);
598 ttimer->timer.function = __hrtimer_tasklet_trampoline;
599 tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline,
600 (unsigned long)ttimer);
601 ttimer->function = function;
603 EXPORT_SYMBOL_GPL(tasklet_hrtimer_init);
606 * Remote softirq bits
609 DEFINE_PER_CPU(struct list_head [NR_SOFTIRQS], softirq_work_list);
610 EXPORT_PER_CPU_SYMBOL(softirq_work_list);
612 static void __local_trigger(struct call_single_data *cp, int softirq)
614 struct list_head *head = &__get_cpu_var(softirq_work_list[softirq]);
616 list_add_tail(&cp->list, head);
618 /* Trigger the softirq only if the list was previously empty. */
619 if (head->next == &cp->list)
620 raise_softirq_irqoff(softirq);
623 #ifdef CONFIG_USE_GENERIC_SMP_HELPERS
624 static void remote_softirq_receive(void *data)
626 struct call_single_data *cp = data;
630 softirq = *(int *)cp->info;
631 local_irq_save(flags);
632 __local_trigger(cp, softirq);
633 local_irq_restore(flags);
636 static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
638 if (cpu_online(cpu)) {
639 cp->func = remote_softirq_receive;
643 __smp_call_function_single(cpu, cp, 0);
648 #else /* CONFIG_USE_GENERIC_SMP_HELPERS */
649 static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
656 * __send_remote_softirq - try to schedule softirq work on a remote cpu
657 * @cp: private SMP call function data area
658 * @cpu: the remote cpu
659 * @this_cpu: the currently executing cpu
660 * @softirq: the softirq for the work
662 * Attempt to schedule softirq work on a remote cpu. If this cannot be
663 * done, the work is instead queued up on the local cpu.
665 * Interrupts must be disabled.
667 void __send_remote_softirq(struct call_single_data *cp, int cpu, int this_cpu, int softirq)
669 if (cpu == this_cpu || __try_remote_softirq(cp, cpu, softirq))
670 __local_trigger(cp, softirq);
672 EXPORT_SYMBOL(__send_remote_softirq);
675 * send_remote_softirq - try to schedule softirq work on a remote cpu
676 * @cp: private SMP call function data area
677 * @cpu: the remote cpu
678 * @softirq: the softirq for the work
680 * Like __send_remote_softirq except that disabling interrupts and
681 * computing the current cpu is done for the caller.
683 void send_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
688 local_irq_save(flags);
689 this_cpu = smp_processor_id();
690 __send_remote_softirq(cp, cpu, this_cpu, softirq);
691 local_irq_restore(flags);
693 EXPORT_SYMBOL(send_remote_softirq);
695 static int __cpuinit remote_softirq_cpu_notify(struct notifier_block *self,
696 unsigned long action, void *hcpu)
699 * If a CPU goes away, splice its entries to the current CPU
700 * and trigger a run of the softirq
702 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
703 int cpu = (unsigned long) hcpu;
707 for (i = 0; i < NR_SOFTIRQS; i++) {
708 struct list_head *head = &per_cpu(softirq_work_list[i], cpu);
709 struct list_head *local_head;
711 if (list_empty(head))
714 local_head = &__get_cpu_var(softirq_work_list[i]);
715 list_splice_init(head, local_head);
716 raise_softirq_irqoff(i);
724 static struct notifier_block __cpuinitdata remote_softirq_cpu_notifier = {
725 .notifier_call = remote_softirq_cpu_notify,
728 void __init softirq_init(void)
732 for_each_possible_cpu(cpu) {
735 per_cpu(tasklet_vec, cpu).tail =
736 &per_cpu(tasklet_vec, cpu).head;
737 per_cpu(tasklet_hi_vec, cpu).tail =
738 &per_cpu(tasklet_hi_vec, cpu).head;
739 for (i = 0; i < NR_SOFTIRQS; i++)
740 INIT_LIST_HEAD(&per_cpu(softirq_work_list[i], cpu));
743 register_hotcpu_notifier(&remote_softirq_cpu_notifier);
745 open_softirq(TASKLET_SOFTIRQ, tasklet_action);
746 open_softirq(HI_SOFTIRQ, tasklet_hi_action);
749 static int ksoftirqd_should_run(unsigned int cpu)
751 return local_softirq_pending();
754 static void run_ksoftirqd(unsigned int cpu)
757 if (local_softirq_pending()) {
759 rcu_note_context_switch(cpu);
767 #ifdef CONFIG_HOTPLUG_CPU
769 * tasklet_kill_immediate is called to remove a tasklet which can already be
770 * scheduled for execution on @cpu.
772 * Unlike tasklet_kill, this function removes the tasklet
773 * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
775 * When this function is called, @cpu must be in the CPU_DEAD state.
777 void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
779 struct tasklet_struct **i;
781 BUG_ON(cpu_online(cpu));
782 BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
784 if (!test_bit(TASKLET_STATE_SCHED, &t->state))
787 /* CPU is dead, so no lock needed. */
788 for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
791 /* If this was the tail element, move the tail ptr */
793 per_cpu(tasklet_vec, cpu).tail = i;
800 static void takeover_tasklets(unsigned int cpu)
802 /* CPU is dead, so no lock needed. */
805 /* Find end, append list for that CPU. */
806 if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
807 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
808 this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
809 per_cpu(tasklet_vec, cpu).head = NULL;
810 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
812 raise_softirq_irqoff(TASKLET_SOFTIRQ);
814 if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
815 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
816 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
817 per_cpu(tasklet_hi_vec, cpu).head = NULL;
818 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
820 raise_softirq_irqoff(HI_SOFTIRQ);
824 #endif /* CONFIG_HOTPLUG_CPU */
826 static int __cpuinit cpu_callback(struct notifier_block *nfb,
827 unsigned long action,
831 #ifdef CONFIG_HOTPLUG_CPU
833 case CPU_DEAD_FROZEN:
834 takeover_tasklets((unsigned long)hcpu);
836 #endif /* CONFIG_HOTPLUG_CPU */
841 static struct notifier_block __cpuinitdata cpu_nfb = {
842 .notifier_call = cpu_callback
845 static struct smp_hotplug_thread softirq_threads = {
847 .thread_should_run = ksoftirqd_should_run,
848 .thread_fn = run_ksoftirqd,
849 .thread_comm = "ksoftirqd/%u",
852 static __init int spawn_ksoftirqd(void)
854 register_cpu_notifier(&cpu_nfb);
856 BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
860 early_initcall(spawn_ksoftirqd);
863 * [ These __weak aliases are kept in a separate compilation unit, so that
864 * GCC does not inline them incorrectly. ]
867 int __init __weak early_irq_init(void)
872 #ifdef CONFIG_GENERIC_HARDIRQS
873 int __init __weak arch_probe_nr_irqs(void)
875 return NR_IRQS_LEGACY;
878 int __init __weak arch_early_irq_init(void)