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)
11 #include <linux/export.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/interrupt.h>
14 #include <linux/init.h>
16 #include <linux/notifier.h>
17 #include <linux/percpu.h>
18 #include <linux/cpu.h>
19 #include <linux/freezer.h>
20 #include <linux/kthread.h>
21 #include <linux/rcupdate.h>
22 #include <linux/ftrace.h>
23 #include <linux/smp.h>
24 #include <linux/smpboot.h>
25 #include <linux/tick.h>
27 #define CREATE_TRACE_POINTS
28 #include <trace/events/irq.h>
31 - No shared variables, all the data are CPU local.
32 - If a softirq needs serialization, let it serialize itself
34 - Even if softirq is serialized, only local cpu is marked for
35 execution. Hence, we get something sort of weak cpu binding.
36 Though it is still not clear, will it result in better locality
40 - NET RX softirq. It is multithreaded and does not require
41 any global serialization.
42 - NET TX softirq. It kicks software netdevice queues, hence
43 it is logically serialized per device, but this serialization
44 is invisible to common code.
45 - Tasklets: serialized wrt itself.
48 #ifndef __ARCH_IRQ_STAT
49 irq_cpustat_t irq_stat[NR_CPUS] ____cacheline_aligned;
50 EXPORT_SYMBOL(irq_stat);
53 static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
55 DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
57 char *softirq_to_name[NR_SOFTIRQS] = {
58 "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "BLOCK_IOPOLL",
59 "TASKLET", "SCHED", "HRTIMER", "RCU"
63 * we cannot loop indefinitely here to avoid userspace starvation,
64 * but we also don't want to introduce a worst case 1/HZ latency
65 * to the pending events, so lets the scheduler to balance
66 * the softirq load for us.
68 static void wakeup_softirqd(void)
70 /* Interrupts are disabled: no need to stop preemption */
71 struct task_struct *tsk = __this_cpu_read(ksoftirqd);
73 if (tsk && tsk->state != TASK_RUNNING)
78 * preempt_count and SOFTIRQ_OFFSET usage:
79 * - preempt_count is changed by SOFTIRQ_OFFSET on entering or leaving
81 * - preempt_count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
82 * on local_bh_disable or local_bh_enable.
83 * This lets us distinguish between whether we are currently processing
84 * softirq and whether we just have bh disabled.
88 * This one is for softirq.c-internal use,
89 * where hardirqs are disabled legitimately:
91 #ifdef CONFIG_TRACE_IRQFLAGS
92 static void __local_bh_disable(unsigned long ip, unsigned int cnt)
96 WARN_ON_ONCE(in_irq());
98 raw_local_irq_save(flags);
100 * The preempt tracer hooks into preempt_count_add and will break
101 * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
102 * is set and before current->softirq_enabled is cleared.
103 * We must manually increment preempt_count here and manually
104 * call the trace_preempt_off later.
106 __preempt_count_add(cnt);
108 * Were softirqs turned off above:
110 if (softirq_count() == cnt)
111 trace_softirqs_off(ip);
112 raw_local_irq_restore(flags);
114 if (preempt_count() == cnt)
115 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
117 #else /* !CONFIG_TRACE_IRQFLAGS */
118 static inline void __local_bh_disable(unsigned long ip, unsigned int cnt)
120 preempt_count_add(cnt);
123 #endif /* CONFIG_TRACE_IRQFLAGS */
125 void local_bh_disable(void)
127 __local_bh_disable(_RET_IP_, SOFTIRQ_DISABLE_OFFSET);
130 EXPORT_SYMBOL(local_bh_disable);
132 static void __local_bh_enable(unsigned int cnt)
134 WARN_ON_ONCE(!irqs_disabled());
136 if (softirq_count() == cnt)
137 trace_softirqs_on(_RET_IP_);
138 preempt_count_sub(cnt);
142 * Special-case - softirqs can safely be enabled in
143 * cond_resched_softirq(), or by __do_softirq(),
144 * without processing still-pending softirqs:
146 void _local_bh_enable(void)
148 WARN_ON_ONCE(in_irq());
149 __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
152 EXPORT_SYMBOL(_local_bh_enable);
154 static inline void _local_bh_enable_ip(unsigned long ip)
156 WARN_ON_ONCE(in_irq() || irqs_disabled());
157 #ifdef CONFIG_TRACE_IRQFLAGS
161 * Are softirqs going to be turned on now:
163 if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
164 trace_softirqs_on(ip);
166 * Keep preemption disabled until we are done with
167 * softirq processing:
169 preempt_count_sub(SOFTIRQ_DISABLE_OFFSET - 1);
171 if (unlikely(!in_interrupt() && local_softirq_pending())) {
173 * Run softirq if any pending. And do it in its own stack
174 * as we may be calling this deep in a task call stack already.
180 #ifdef CONFIG_TRACE_IRQFLAGS
183 preempt_check_resched();
186 void local_bh_enable(void)
188 _local_bh_enable_ip(_RET_IP_);
190 EXPORT_SYMBOL(local_bh_enable);
192 void local_bh_enable_ip(unsigned long ip)
194 _local_bh_enable_ip(ip);
196 EXPORT_SYMBOL(local_bh_enable_ip);
199 * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times,
200 * but break the loop if need_resched() is set or after 2 ms.
201 * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in
202 * certain cases, such as stop_machine(), jiffies may cease to
203 * increment and so we need the MAX_SOFTIRQ_RESTART limit as
204 * well to make sure we eventually return from this method.
206 * These limits have been established via experimentation.
207 * The two things to balance is latency against fairness -
208 * we want to handle softirqs as soon as possible, but they
209 * should not be able to lock up the box.
211 #define MAX_SOFTIRQ_TIME msecs_to_jiffies(2)
212 #define MAX_SOFTIRQ_RESTART 10
214 asmlinkage void __do_softirq(void)
216 struct softirq_action *h;
218 unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
220 unsigned long old_flags = current->flags;
221 int max_restart = MAX_SOFTIRQ_RESTART;
224 * Mask out PF_MEMALLOC s current task context is borrowed for the
225 * softirq. A softirq handled such as network RX might set PF_MEMALLOC
226 * again if the socket is related to swap
228 current->flags &= ~PF_MEMALLOC;
230 pending = local_softirq_pending();
231 account_irq_enter_time(current);
233 __local_bh_disable(_RET_IP_, SOFTIRQ_OFFSET);
234 lockdep_softirq_enter();
236 cpu = smp_processor_id();
238 /* Reset the pending bitmask before enabling irqs */
239 set_softirq_pending(0);
247 unsigned int vec_nr = h - softirq_vec;
248 int prev_count = preempt_count();
250 kstat_incr_softirqs_this_cpu(vec_nr);
252 trace_softirq_entry(vec_nr);
254 trace_softirq_exit(vec_nr);
255 if (unlikely(prev_count != preempt_count())) {
256 printk(KERN_ERR "huh, entered softirq %u %s %p"
257 "with preempt_count %08x,"
258 " exited with %08x?\n", vec_nr,
259 softirq_to_name[vec_nr], h->action,
260 prev_count, preempt_count());
261 preempt_count_set(prev_count);
272 pending = local_softirq_pending();
274 if (time_before(jiffies, end) && !need_resched() &&
281 lockdep_softirq_exit();
283 account_irq_exit_time(current);
284 __local_bh_enable(SOFTIRQ_OFFSET);
285 WARN_ON_ONCE(in_interrupt());
286 tsk_restore_flags(current, old_flags, PF_MEMALLOC);
291 asmlinkage void do_softirq(void)
299 local_irq_save(flags);
301 pending = local_softirq_pending();
304 do_softirq_own_stack();
306 local_irq_restore(flags);
310 * Enter an interrupt context.
315 if (is_idle_task(current) && !in_interrupt()) {
317 * Prevent raise_softirq from needlessly waking up ksoftirqd
318 * here, as softirq will be serviced on return from interrupt.
328 static inline void invoke_softirq(void)
330 if (!force_irqthreads) {
331 #ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
333 * We can safely execute softirq on the current stack if
334 * it is the irq stack, because it should be near empty
340 * Otherwise, irq_exit() is called on the task stack that can
341 * be potentially deep already. So call softirq in its own stack
342 * to prevent from any overrun.
344 do_softirq_own_stack();
351 static inline void tick_irq_exit(void)
353 #ifdef CONFIG_NO_HZ_COMMON
354 int cpu = smp_processor_id();
356 /* Make sure that timer wheel updates are propagated */
357 if ((idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
359 tick_nohz_irq_exit();
365 * Exit an interrupt context. Process softirqs if needed and possible:
369 #ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
372 WARN_ON_ONCE(!irqs_disabled());
375 account_irq_exit_time(current);
376 trace_hardirq_exit();
377 preempt_count_sub(HARDIRQ_OFFSET);
378 if (!in_interrupt() && local_softirq_pending())
386 * This function must run with irqs disabled!
388 inline void raise_softirq_irqoff(unsigned int nr)
390 __raise_softirq_irqoff(nr);
393 * If we're in an interrupt or softirq, we're done
394 * (this also catches softirq-disabled code). We will
395 * actually run the softirq once we return from
396 * the irq or softirq.
398 * Otherwise we wake up ksoftirqd to make sure we
399 * schedule the softirq soon.
405 void raise_softirq(unsigned int nr)
409 local_irq_save(flags);
410 raise_softirq_irqoff(nr);
411 local_irq_restore(flags);
414 void __raise_softirq_irqoff(unsigned int nr)
416 trace_softirq_raise(nr);
417 or_softirq_pending(1UL << nr);
420 void open_softirq(int nr, void (*action)(struct softirq_action *))
422 softirq_vec[nr].action = action;
430 struct tasklet_struct *head;
431 struct tasklet_struct **tail;
434 static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
435 static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
437 void __tasklet_schedule(struct tasklet_struct *t)
441 local_irq_save(flags);
443 *__this_cpu_read(tasklet_vec.tail) = t;
444 __this_cpu_write(tasklet_vec.tail, &(t->next));
445 raise_softirq_irqoff(TASKLET_SOFTIRQ);
446 local_irq_restore(flags);
449 EXPORT_SYMBOL(__tasklet_schedule);
451 void __tasklet_hi_schedule(struct tasklet_struct *t)
455 local_irq_save(flags);
457 *__this_cpu_read(tasklet_hi_vec.tail) = t;
458 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
459 raise_softirq_irqoff(HI_SOFTIRQ);
460 local_irq_restore(flags);
463 EXPORT_SYMBOL(__tasklet_hi_schedule);
465 void __tasklet_hi_schedule_first(struct tasklet_struct *t)
467 BUG_ON(!irqs_disabled());
469 t->next = __this_cpu_read(tasklet_hi_vec.head);
470 __this_cpu_write(tasklet_hi_vec.head, t);
471 __raise_softirq_irqoff(HI_SOFTIRQ);
474 EXPORT_SYMBOL(__tasklet_hi_schedule_first);
476 static void tasklet_action(struct softirq_action *a)
478 struct tasklet_struct *list;
481 list = __this_cpu_read(tasklet_vec.head);
482 __this_cpu_write(tasklet_vec.head, NULL);
483 __this_cpu_write(tasklet_vec.tail, &__get_cpu_var(tasklet_vec).head);
487 struct tasklet_struct *t = list;
491 if (tasklet_trylock(t)) {
492 if (!atomic_read(&t->count)) {
493 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
504 *__this_cpu_read(tasklet_vec.tail) = t;
505 __this_cpu_write(tasklet_vec.tail, &(t->next));
506 __raise_softirq_irqoff(TASKLET_SOFTIRQ);
511 static void tasklet_hi_action(struct softirq_action *a)
513 struct tasklet_struct *list;
516 list = __this_cpu_read(tasklet_hi_vec.head);
517 __this_cpu_write(tasklet_hi_vec.head, NULL);
518 __this_cpu_write(tasklet_hi_vec.tail, &__get_cpu_var(tasklet_hi_vec).head);
522 struct tasklet_struct *t = list;
526 if (tasklet_trylock(t)) {
527 if (!atomic_read(&t->count)) {
528 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
539 *__this_cpu_read(tasklet_hi_vec.tail) = t;
540 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
541 __raise_softirq_irqoff(HI_SOFTIRQ);
547 void tasklet_init(struct tasklet_struct *t,
548 void (*func)(unsigned long), unsigned long data)
552 atomic_set(&t->count, 0);
557 EXPORT_SYMBOL(tasklet_init);
559 void tasklet_kill(struct tasklet_struct *t)
562 printk("Attempt to kill tasklet from interrupt\n");
564 while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
567 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
569 tasklet_unlock_wait(t);
570 clear_bit(TASKLET_STATE_SCHED, &t->state);
573 EXPORT_SYMBOL(tasklet_kill);
580 * The trampoline is called when the hrtimer expires. It schedules a tasklet
581 * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
582 * hrtimer callback, but from softirq context.
584 static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer)
586 struct tasklet_hrtimer *ttimer =
587 container_of(timer, struct tasklet_hrtimer, timer);
589 tasklet_hi_schedule(&ttimer->tasklet);
590 return HRTIMER_NORESTART;
594 * Helper function which calls the hrtimer callback from
595 * tasklet/softirq context
597 static void __tasklet_hrtimer_trampoline(unsigned long data)
599 struct tasklet_hrtimer *ttimer = (void *)data;
600 enum hrtimer_restart restart;
602 restart = ttimer->function(&ttimer->timer);
603 if (restart != HRTIMER_NORESTART)
604 hrtimer_restart(&ttimer->timer);
608 * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
609 * @ttimer: tasklet_hrtimer which is initialized
610 * @function: hrtimer callback function which gets called from softirq context
611 * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
612 * @mode: hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
614 void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer,
615 enum hrtimer_restart (*function)(struct hrtimer *),
616 clockid_t which_clock, enum hrtimer_mode mode)
618 hrtimer_init(&ttimer->timer, which_clock, mode);
619 ttimer->timer.function = __hrtimer_tasklet_trampoline;
620 tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline,
621 (unsigned long)ttimer);
622 ttimer->function = function;
624 EXPORT_SYMBOL_GPL(tasklet_hrtimer_init);
626 void __init softirq_init(void)
630 for_each_possible_cpu(cpu) {
631 per_cpu(tasklet_vec, cpu).tail =
632 &per_cpu(tasklet_vec, cpu).head;
633 per_cpu(tasklet_hi_vec, cpu).tail =
634 &per_cpu(tasklet_hi_vec, cpu).head;
637 open_softirq(TASKLET_SOFTIRQ, tasklet_action);
638 open_softirq(HI_SOFTIRQ, tasklet_hi_action);
641 static int ksoftirqd_should_run(unsigned int cpu)
643 return local_softirq_pending();
646 static void run_ksoftirqd(unsigned int cpu)
649 if (local_softirq_pending()) {
651 * We can safely run softirq on inline stack, as we are not deep
652 * in the task stack here.
655 rcu_note_context_switch(cpu);
663 #ifdef CONFIG_HOTPLUG_CPU
665 * tasklet_kill_immediate is called to remove a tasklet which can already be
666 * scheduled for execution on @cpu.
668 * Unlike tasklet_kill, this function removes the tasklet
669 * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
671 * When this function is called, @cpu must be in the CPU_DEAD state.
673 void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
675 struct tasklet_struct **i;
677 BUG_ON(cpu_online(cpu));
678 BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
680 if (!test_bit(TASKLET_STATE_SCHED, &t->state))
683 /* CPU is dead, so no lock needed. */
684 for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
687 /* If this was the tail element, move the tail ptr */
689 per_cpu(tasklet_vec, cpu).tail = i;
696 static void takeover_tasklets(unsigned int cpu)
698 /* CPU is dead, so no lock needed. */
701 /* Find end, append list for that CPU. */
702 if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
703 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
704 this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
705 per_cpu(tasklet_vec, cpu).head = NULL;
706 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
708 raise_softirq_irqoff(TASKLET_SOFTIRQ);
710 if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
711 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
712 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
713 per_cpu(tasklet_hi_vec, cpu).head = NULL;
714 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
716 raise_softirq_irqoff(HI_SOFTIRQ);
720 #endif /* CONFIG_HOTPLUG_CPU */
722 static int cpu_callback(struct notifier_block *nfb,
723 unsigned long action,
727 #ifdef CONFIG_HOTPLUG_CPU
729 case CPU_DEAD_FROZEN:
730 takeover_tasklets((unsigned long)hcpu);
732 #endif /* CONFIG_HOTPLUG_CPU */
737 static struct notifier_block cpu_nfb = {
738 .notifier_call = cpu_callback
741 static struct smp_hotplug_thread softirq_threads = {
743 .thread_should_run = ksoftirqd_should_run,
744 .thread_fn = run_ksoftirqd,
745 .thread_comm = "ksoftirqd/%u",
748 static __init int spawn_ksoftirqd(void)
750 register_cpu_notifier(&cpu_nfb);
752 BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
756 early_initcall(spawn_ksoftirqd);
759 * [ These __weak aliases are kept in a separate compilation unit, so that
760 * GCC does not inline them incorrectly. ]
763 int __init __weak early_irq_init(void)
768 int __init __weak arch_probe_nr_irqs(void)
770 return NR_IRQS_LEGACY;
773 int __init __weak arch_early_irq_init(void)