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 void __local_bh_disable_ip(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 & SOFTIRQ_MASK))
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 EXPORT_SYMBOL(__local_bh_disable_ip);
118 #endif /* CONFIG_TRACE_IRQFLAGS */
120 static void __local_bh_enable(unsigned int cnt)
122 WARN_ON_ONCE(!irqs_disabled());
124 if (softirq_count() == (cnt & SOFTIRQ_MASK))
125 trace_softirqs_on(_RET_IP_);
126 preempt_count_sub(cnt);
130 * Special-case - softirqs can safely be enabled in
131 * cond_resched_softirq(), or by __do_softirq(),
132 * without processing still-pending softirqs:
134 void _local_bh_enable(void)
136 WARN_ON_ONCE(in_irq());
137 __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
140 EXPORT_SYMBOL(_local_bh_enable);
142 void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
144 WARN_ON_ONCE(in_irq() || irqs_disabled());
145 #ifdef CONFIG_TRACE_IRQFLAGS
149 * Are softirqs going to be turned on now:
151 if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
152 trace_softirqs_on(ip);
154 * Keep preemption disabled until we are done with
155 * softirq processing:
157 preempt_count_sub(cnt - 1);
159 if (unlikely(!in_interrupt() && local_softirq_pending())) {
161 * Run softirq if any pending. And do it in its own stack
162 * as we may be calling this deep in a task call stack already.
168 #ifdef CONFIG_TRACE_IRQFLAGS
171 preempt_check_resched();
173 EXPORT_SYMBOL(__local_bh_enable_ip);
176 * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times,
177 * but break the loop if need_resched() is set or after 2 ms.
178 * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in
179 * certain cases, such as stop_machine(), jiffies may cease to
180 * increment and so we need the MAX_SOFTIRQ_RESTART limit as
181 * well to make sure we eventually return from this method.
183 * These limits have been established via experimentation.
184 * The two things to balance is latency against fairness -
185 * we want to handle softirqs as soon as possible, but they
186 * should not be able to lock up the box.
188 #define MAX_SOFTIRQ_TIME msecs_to_jiffies(2)
189 #define MAX_SOFTIRQ_RESTART 10
191 asmlinkage void __do_softirq(void)
193 struct softirq_action *h;
195 unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
197 unsigned long old_flags = current->flags;
198 int max_restart = MAX_SOFTIRQ_RESTART;
201 * Mask out PF_MEMALLOC s current task context is borrowed for the
202 * softirq. A softirq handled such as network RX might set PF_MEMALLOC
203 * again if the socket is related to swap
205 current->flags &= ~PF_MEMALLOC;
207 pending = local_softirq_pending();
208 account_irq_enter_time(current);
210 __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
211 lockdep_softirq_enter();
213 cpu = smp_processor_id();
215 /* Reset the pending bitmask before enabling irqs */
216 set_softirq_pending(0);
224 unsigned int vec_nr = h - softirq_vec;
225 int prev_count = preempt_count();
227 kstat_incr_softirqs_this_cpu(vec_nr);
229 trace_softirq_entry(vec_nr);
231 trace_softirq_exit(vec_nr);
232 if (unlikely(prev_count != preempt_count())) {
233 printk(KERN_ERR "huh, entered softirq %u %s %p"
234 "with preempt_count %08x,"
235 " exited with %08x?\n", vec_nr,
236 softirq_to_name[vec_nr], h->action,
237 prev_count, preempt_count());
238 preempt_count_set(prev_count);
249 pending = local_softirq_pending();
251 if (time_before(jiffies, end) && !need_resched() &&
258 lockdep_softirq_exit();
260 account_irq_exit_time(current);
261 __local_bh_enable(SOFTIRQ_OFFSET);
262 WARN_ON_ONCE(in_interrupt());
263 tsk_restore_flags(current, old_flags, PF_MEMALLOC);
268 asmlinkage void do_softirq(void)
276 local_irq_save(flags);
278 pending = local_softirq_pending();
281 do_softirq_own_stack();
283 local_irq_restore(flags);
287 * Enter an interrupt context.
291 int cpu = smp_processor_id();
294 if (is_idle_task(current) && !in_interrupt()) {
296 * Prevent raise_softirq from needlessly waking up ksoftirqd
297 * here, as softirq will be serviced on return from interrupt.
300 tick_check_idle(cpu);
307 static inline void invoke_softirq(void)
309 if (!force_irqthreads) {
310 #ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
312 * We can safely execute softirq on the current stack if
313 * it is the irq stack, because it should be near empty
319 * Otherwise, irq_exit() is called on the task stack that can
320 * be potentially deep already. So call softirq in its own stack
321 * to prevent from any overrun.
323 do_softirq_own_stack();
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 preempt_count_sub(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);
605 void __init softirq_init(void)
609 for_each_possible_cpu(cpu) {
610 per_cpu(tasklet_vec, cpu).tail =
611 &per_cpu(tasklet_vec, cpu).head;
612 per_cpu(tasklet_hi_vec, cpu).tail =
613 &per_cpu(tasklet_hi_vec, cpu).head;
616 open_softirq(TASKLET_SOFTIRQ, tasklet_action);
617 open_softirq(HI_SOFTIRQ, tasklet_hi_action);
620 static int ksoftirqd_should_run(unsigned int cpu)
622 return local_softirq_pending();
625 static void run_ksoftirqd(unsigned int cpu)
628 if (local_softirq_pending()) {
630 * We can safely run softirq on inline stack, as we are not deep
631 * in the task stack here.
634 rcu_note_context_switch(cpu);
642 #ifdef CONFIG_HOTPLUG_CPU
644 * tasklet_kill_immediate is called to remove a tasklet which can already be
645 * scheduled for execution on @cpu.
647 * Unlike tasklet_kill, this function removes the tasklet
648 * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
650 * When this function is called, @cpu must be in the CPU_DEAD state.
652 void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
654 struct tasklet_struct **i;
656 BUG_ON(cpu_online(cpu));
657 BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
659 if (!test_bit(TASKLET_STATE_SCHED, &t->state))
662 /* CPU is dead, so no lock needed. */
663 for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
666 /* If this was the tail element, move the tail ptr */
668 per_cpu(tasklet_vec, cpu).tail = i;
675 static void takeover_tasklets(unsigned int cpu)
677 /* CPU is dead, so no lock needed. */
680 /* Find end, append list for that CPU. */
681 if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
682 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
683 this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
684 per_cpu(tasklet_vec, cpu).head = NULL;
685 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
687 raise_softirq_irqoff(TASKLET_SOFTIRQ);
689 if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
690 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
691 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
692 per_cpu(tasklet_hi_vec, cpu).head = NULL;
693 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
695 raise_softirq_irqoff(HI_SOFTIRQ);
699 #endif /* CONFIG_HOTPLUG_CPU */
701 static int cpu_callback(struct notifier_block *nfb,
702 unsigned long action,
706 #ifdef CONFIG_HOTPLUG_CPU
708 case CPU_DEAD_FROZEN:
709 takeover_tasklets((unsigned long)hcpu);
711 #endif /* CONFIG_HOTPLUG_CPU */
716 static struct notifier_block cpu_nfb = {
717 .notifier_call = cpu_callback
720 static struct smp_hotplug_thread softirq_threads = {
722 .thread_should_run = ksoftirqd_should_run,
723 .thread_fn = run_ksoftirqd,
724 .thread_comm = "ksoftirqd/%u",
727 static __init int spawn_ksoftirqd(void)
729 register_cpu_notifier(&cpu_nfb);
731 BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
735 early_initcall(spawn_ksoftirqd);
738 * [ These __weak aliases are kept in a separate compilation unit, so that
739 * GCC does not inline them incorrectly. ]
742 int __init __weak early_irq_init(void)
747 int __init __weak arch_probe_nr_irqs(void)
749 return NR_IRQS_LEGACY;
752 int __init __weak arch_early_irq_init(void)