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 #ifdef CONFIG_TRACE_IRQFLAGS
193 * When we run softirqs from irq_exit() and thus on the hardirq stack we need
194 * to keep the lockdep irq context tracking as tight as possible in order to
195 * not miss-qualify lock contexts and miss possible deadlocks.
198 static inline bool lockdep_softirq_start(void)
200 bool in_hardirq = false;
202 if (trace_hardirq_context(current)) {
204 trace_hardirq_exit();
207 lockdep_softirq_enter();
212 static inline void lockdep_softirq_end(bool in_hardirq)
214 lockdep_softirq_exit();
217 trace_hardirq_enter();
220 static inline bool lockdep_softirq_start(void) { return false; }
221 static inline void lockdep_softirq_end(bool in_hardirq) { }
224 asmlinkage void __do_softirq(void)
226 unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
227 unsigned long old_flags = current->flags;
228 int max_restart = MAX_SOFTIRQ_RESTART;
229 struct softirq_action *h;
235 * Mask out PF_MEMALLOC s current task context is borrowed for the
236 * softirq. A softirq handled such as network RX might set PF_MEMALLOC
237 * again if the socket is related to swap
239 current->flags &= ~PF_MEMALLOC;
241 pending = local_softirq_pending();
242 account_irq_enter_time(current);
244 __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
245 in_hardirq = lockdep_softirq_start();
247 cpu = smp_processor_id();
249 /* Reset the pending bitmask before enabling irqs */
250 set_softirq_pending(0);
258 unsigned int vec_nr = h - softirq_vec;
259 int prev_count = preempt_count();
261 kstat_incr_softirqs_this_cpu(vec_nr);
263 trace_softirq_entry(vec_nr);
265 trace_softirq_exit(vec_nr);
266 if (unlikely(prev_count != preempt_count())) {
267 printk(KERN_ERR "huh, entered softirq %u %s %p"
268 "with preempt_count %08x,"
269 " exited with %08x?\n", vec_nr,
270 softirq_to_name[vec_nr], h->action,
271 prev_count, preempt_count());
272 preempt_count_set(prev_count);
283 pending = local_softirq_pending();
285 if (time_before(jiffies, end) && !need_resched() &&
292 lockdep_softirq_end(in_hardirq);
293 account_irq_exit_time(current);
294 __local_bh_enable(SOFTIRQ_OFFSET);
295 WARN_ON_ONCE(in_interrupt());
296 tsk_restore_flags(current, old_flags, PF_MEMALLOC);
299 asmlinkage void do_softirq(void)
307 local_irq_save(flags);
309 pending = local_softirq_pending();
312 do_softirq_own_stack();
314 local_irq_restore(flags);
318 * Enter an interrupt context.
322 int cpu = smp_processor_id();
325 if (is_idle_task(current) && !in_interrupt()) {
327 * Prevent raise_softirq from needlessly waking up ksoftirqd
328 * here, as softirq will be serviced on return from interrupt.
331 tick_check_idle(cpu);
338 static inline void invoke_softirq(void)
340 if (!force_irqthreads) {
341 #ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
343 * We can safely execute softirq on the current stack if
344 * it is the irq stack, because it should be near empty
350 * Otherwise, irq_exit() is called on the task stack that can
351 * be potentially deep already. So call softirq in its own stack
352 * to prevent from any overrun.
354 do_softirq_own_stack();
361 static inline void tick_irq_exit(void)
363 #ifdef CONFIG_NO_HZ_COMMON
364 int cpu = smp_processor_id();
366 /* Make sure that timer wheel updates are propagated */
367 if ((idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
369 tick_nohz_irq_exit();
375 * Exit an interrupt context. Process softirqs if needed and possible:
379 #ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
382 WARN_ON_ONCE(!irqs_disabled());
385 account_irq_exit_time(current);
386 preempt_count_sub(HARDIRQ_OFFSET);
387 if (!in_interrupt() && local_softirq_pending())
392 trace_hardirq_exit(); /* must be last! */
396 * This function must run with irqs disabled!
398 inline void raise_softirq_irqoff(unsigned int nr)
400 __raise_softirq_irqoff(nr);
403 * If we're in an interrupt or softirq, we're done
404 * (this also catches softirq-disabled code). We will
405 * actually run the softirq once we return from
406 * the irq or softirq.
408 * Otherwise we wake up ksoftirqd to make sure we
409 * schedule the softirq soon.
415 void raise_softirq(unsigned int nr)
419 local_irq_save(flags);
420 raise_softirq_irqoff(nr);
421 local_irq_restore(flags);
424 void __raise_softirq_irqoff(unsigned int nr)
426 trace_softirq_raise(nr);
427 or_softirq_pending(1UL << nr);
430 void open_softirq(int nr, void (*action)(struct softirq_action *))
432 softirq_vec[nr].action = action;
440 struct tasklet_struct *head;
441 struct tasklet_struct **tail;
444 static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
445 static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
447 void __tasklet_schedule(struct tasklet_struct *t)
451 local_irq_save(flags);
453 *__this_cpu_read(tasklet_vec.tail) = t;
454 __this_cpu_write(tasklet_vec.tail, &(t->next));
455 raise_softirq_irqoff(TASKLET_SOFTIRQ);
456 local_irq_restore(flags);
459 EXPORT_SYMBOL(__tasklet_schedule);
461 void __tasklet_hi_schedule(struct tasklet_struct *t)
465 local_irq_save(flags);
467 *__this_cpu_read(tasklet_hi_vec.tail) = t;
468 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
469 raise_softirq_irqoff(HI_SOFTIRQ);
470 local_irq_restore(flags);
473 EXPORT_SYMBOL(__tasklet_hi_schedule);
475 void __tasklet_hi_schedule_first(struct tasklet_struct *t)
477 BUG_ON(!irqs_disabled());
479 t->next = __this_cpu_read(tasklet_hi_vec.head);
480 __this_cpu_write(tasklet_hi_vec.head, t);
481 __raise_softirq_irqoff(HI_SOFTIRQ);
484 EXPORT_SYMBOL(__tasklet_hi_schedule_first);
486 static void tasklet_action(struct softirq_action *a)
488 struct tasklet_struct *list;
491 list = __this_cpu_read(tasklet_vec.head);
492 __this_cpu_write(tasklet_vec.head, NULL);
493 __this_cpu_write(tasklet_vec.tail, &__get_cpu_var(tasklet_vec).head);
497 struct tasklet_struct *t = list;
501 if (tasklet_trylock(t)) {
502 if (!atomic_read(&t->count)) {
503 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
514 *__this_cpu_read(tasklet_vec.tail) = t;
515 __this_cpu_write(tasklet_vec.tail, &(t->next));
516 __raise_softirq_irqoff(TASKLET_SOFTIRQ);
521 static void tasklet_hi_action(struct softirq_action *a)
523 struct tasklet_struct *list;
526 list = __this_cpu_read(tasklet_hi_vec.head);
527 __this_cpu_write(tasklet_hi_vec.head, NULL);
528 __this_cpu_write(tasklet_hi_vec.tail, &__get_cpu_var(tasklet_hi_vec).head);
532 struct tasklet_struct *t = list;
536 if (tasklet_trylock(t)) {
537 if (!atomic_read(&t->count)) {
538 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
549 *__this_cpu_read(tasklet_hi_vec.tail) = t;
550 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
551 __raise_softirq_irqoff(HI_SOFTIRQ);
557 void tasklet_init(struct tasklet_struct *t,
558 void (*func)(unsigned long), unsigned long data)
562 atomic_set(&t->count, 0);
567 EXPORT_SYMBOL(tasklet_init);
569 void tasklet_kill(struct tasklet_struct *t)
572 printk("Attempt to kill tasklet from interrupt\n");
574 while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
577 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
579 tasklet_unlock_wait(t);
580 clear_bit(TASKLET_STATE_SCHED, &t->state);
583 EXPORT_SYMBOL(tasklet_kill);
590 * The trampoline is called when the hrtimer expires. It schedules a tasklet
591 * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
592 * hrtimer callback, but from softirq context.
594 static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer)
596 struct tasklet_hrtimer *ttimer =
597 container_of(timer, struct tasklet_hrtimer, timer);
599 tasklet_hi_schedule(&ttimer->tasklet);
600 return HRTIMER_NORESTART;
604 * Helper function which calls the hrtimer callback from
605 * tasklet/softirq context
607 static void __tasklet_hrtimer_trampoline(unsigned long data)
609 struct tasklet_hrtimer *ttimer = (void *)data;
610 enum hrtimer_restart restart;
612 restart = ttimer->function(&ttimer->timer);
613 if (restart != HRTIMER_NORESTART)
614 hrtimer_restart(&ttimer->timer);
618 * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
619 * @ttimer: tasklet_hrtimer which is initialized
620 * @function: hrtimer callback function which gets called from softirq context
621 * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
622 * @mode: hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
624 void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer,
625 enum hrtimer_restart (*function)(struct hrtimer *),
626 clockid_t which_clock, enum hrtimer_mode mode)
628 hrtimer_init(&ttimer->timer, which_clock, mode);
629 ttimer->timer.function = __hrtimer_tasklet_trampoline;
630 tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline,
631 (unsigned long)ttimer);
632 ttimer->function = function;
634 EXPORT_SYMBOL_GPL(tasklet_hrtimer_init);
636 void __init softirq_init(void)
640 for_each_possible_cpu(cpu) {
641 per_cpu(tasklet_vec, cpu).tail =
642 &per_cpu(tasklet_vec, cpu).head;
643 per_cpu(tasklet_hi_vec, cpu).tail =
644 &per_cpu(tasklet_hi_vec, cpu).head;
647 open_softirq(TASKLET_SOFTIRQ, tasklet_action);
648 open_softirq(HI_SOFTIRQ, tasklet_hi_action);
651 static int ksoftirqd_should_run(unsigned int cpu)
653 return local_softirq_pending();
656 static void run_ksoftirqd(unsigned int cpu)
659 if (local_softirq_pending()) {
661 * We can safely run softirq on inline stack, as we are not deep
662 * in the task stack here.
665 rcu_note_context_switch(cpu);
673 #ifdef CONFIG_HOTPLUG_CPU
675 * tasklet_kill_immediate is called to remove a tasklet which can already be
676 * scheduled for execution on @cpu.
678 * Unlike tasklet_kill, this function removes the tasklet
679 * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
681 * When this function is called, @cpu must be in the CPU_DEAD state.
683 void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
685 struct tasklet_struct **i;
687 BUG_ON(cpu_online(cpu));
688 BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
690 if (!test_bit(TASKLET_STATE_SCHED, &t->state))
693 /* CPU is dead, so no lock needed. */
694 for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
697 /* If this was the tail element, move the tail ptr */
699 per_cpu(tasklet_vec, cpu).tail = i;
706 static void takeover_tasklets(unsigned int cpu)
708 /* CPU is dead, so no lock needed. */
711 /* Find end, append list for that CPU. */
712 if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
713 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
714 this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
715 per_cpu(tasklet_vec, cpu).head = NULL;
716 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
718 raise_softirq_irqoff(TASKLET_SOFTIRQ);
720 if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
721 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
722 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
723 per_cpu(tasklet_hi_vec, cpu).head = NULL;
724 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
726 raise_softirq_irqoff(HI_SOFTIRQ);
730 #endif /* CONFIG_HOTPLUG_CPU */
732 static int cpu_callback(struct notifier_block *nfb,
733 unsigned long action,
737 #ifdef CONFIG_HOTPLUG_CPU
739 case CPU_DEAD_FROZEN:
740 takeover_tasklets((unsigned long)hcpu);
742 #endif /* CONFIG_HOTPLUG_CPU */
747 static struct notifier_block cpu_nfb = {
748 .notifier_call = cpu_callback
751 static struct smp_hotplug_thread softirq_threads = {
753 .thread_should_run = ksoftirqd_should_run,
754 .thread_fn = run_ksoftirqd,
755 .thread_comm = "ksoftirqd/%u",
758 static __init int spawn_ksoftirqd(void)
760 register_cpu_notifier(&cpu_nfb);
762 BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
766 early_initcall(spawn_ksoftirqd);
769 * [ These __weak aliases are kept in a separate compilation unit, so that
770 * GCC does not inline them incorrectly. ]
773 int __init __weak early_irq_init(void)
778 int __init __weak arch_probe_nr_irqs(void)
780 return NR_IRQS_LEGACY;
783 int __init __weak arch_early_irq_init(void)