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 #ifdef CONFIG_TRACE_IRQFLAGS
216 * When we run softirqs from irq_exit() and thus on the hardirq stack we need
217 * to keep the lockdep irq context tracking as tight as possible in order to
218 * not miss-qualify lock contexts and miss possible deadlocks.
221 static inline bool lockdep_softirq_start(void)
223 bool in_hardirq = false;
225 if (trace_hardirq_context(current)) {
227 trace_hardirq_exit();
230 lockdep_softirq_enter();
235 static inline void lockdep_softirq_end(bool in_hardirq)
237 lockdep_softirq_exit();
240 trace_hardirq_enter();
243 static inline bool lockdep_softirq_start(void) { return false; }
244 static inline void lockdep_softirq_end(bool in_hardirq) { }
247 asmlinkage void __do_softirq(void)
249 unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
250 unsigned long old_flags = current->flags;
251 int max_restart = MAX_SOFTIRQ_RESTART;
252 struct softirq_action *h;
258 * Mask out PF_MEMALLOC s current task context is borrowed for the
259 * softirq. A softirq handled such as network RX might set PF_MEMALLOC
260 * again if the socket is related to swap
262 current->flags &= ~PF_MEMALLOC;
264 pending = local_softirq_pending();
265 account_irq_enter_time(current);
267 __local_bh_disable(_RET_IP_, SOFTIRQ_OFFSET);
268 in_hardirq = lockdep_softirq_start();
270 cpu = smp_processor_id();
272 /* Reset the pending bitmask before enabling irqs */
273 set_softirq_pending(0);
281 unsigned int vec_nr = h - softirq_vec;
282 int prev_count = preempt_count();
284 kstat_incr_softirqs_this_cpu(vec_nr);
286 trace_softirq_entry(vec_nr);
288 trace_softirq_exit(vec_nr);
289 if (unlikely(prev_count != preempt_count())) {
290 printk(KERN_ERR "huh, entered softirq %u %s %p"
291 "with preempt_count %08x,"
292 " exited with %08x?\n", vec_nr,
293 softirq_to_name[vec_nr], h->action,
294 prev_count, preempt_count());
295 preempt_count_set(prev_count);
306 pending = local_softirq_pending();
308 if (time_before(jiffies, end) && !need_resched() &&
315 lockdep_softirq_end(in_hardirq);
316 account_irq_exit_time(current);
317 __local_bh_enable(SOFTIRQ_OFFSET);
318 WARN_ON_ONCE(in_interrupt());
319 tsk_restore_flags(current, old_flags, PF_MEMALLOC);
322 asmlinkage void do_softirq(void)
330 local_irq_save(flags);
332 pending = local_softirq_pending();
335 do_softirq_own_stack();
337 local_irq_restore(flags);
341 * Enter an interrupt context.
345 int cpu = smp_processor_id();
348 if (is_idle_task(current) && !in_interrupt()) {
350 * Prevent raise_softirq from needlessly waking up ksoftirqd
351 * here, as softirq will be serviced on return from interrupt.
354 tick_check_idle(cpu);
361 static inline void invoke_softirq(void)
363 if (!force_irqthreads) {
364 #ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
366 * We can safely execute softirq on the current stack if
367 * it is the irq stack, because it should be near empty
373 * Otherwise, irq_exit() is called on the task stack that can
374 * be potentially deep already. So call softirq in its own stack
375 * to prevent from any overrun.
377 do_softirq_own_stack();
384 static inline void tick_irq_exit(void)
386 #ifdef CONFIG_NO_HZ_COMMON
387 int cpu = smp_processor_id();
389 /* Make sure that timer wheel updates are propagated */
390 if ((idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
392 tick_nohz_irq_exit();
398 * Exit an interrupt context. Process softirqs if needed and possible:
402 #ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
405 WARN_ON_ONCE(!irqs_disabled());
408 account_irq_exit_time(current);
409 preempt_count_sub(HARDIRQ_OFFSET);
410 if (!in_interrupt() && local_softirq_pending())
415 trace_hardirq_exit(); /* must be last! */
419 * This function must run with irqs disabled!
421 inline void raise_softirq_irqoff(unsigned int nr)
423 __raise_softirq_irqoff(nr);
426 * If we're in an interrupt or softirq, we're done
427 * (this also catches softirq-disabled code). We will
428 * actually run the softirq once we return from
429 * the irq or softirq.
431 * Otherwise we wake up ksoftirqd to make sure we
432 * schedule the softirq soon.
438 void raise_softirq(unsigned int nr)
442 local_irq_save(flags);
443 raise_softirq_irqoff(nr);
444 local_irq_restore(flags);
447 void __raise_softirq_irqoff(unsigned int nr)
449 trace_softirq_raise(nr);
450 or_softirq_pending(1UL << nr);
453 void open_softirq(int nr, void (*action)(struct softirq_action *))
455 softirq_vec[nr].action = action;
463 struct tasklet_struct *head;
464 struct tasklet_struct **tail;
467 static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
468 static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
470 void __tasklet_schedule(struct tasklet_struct *t)
474 local_irq_save(flags);
476 *__this_cpu_read(tasklet_vec.tail) = t;
477 __this_cpu_write(tasklet_vec.tail, &(t->next));
478 raise_softirq_irqoff(TASKLET_SOFTIRQ);
479 local_irq_restore(flags);
482 EXPORT_SYMBOL(__tasklet_schedule);
484 void __tasklet_hi_schedule(struct tasklet_struct *t)
488 local_irq_save(flags);
490 *__this_cpu_read(tasklet_hi_vec.tail) = t;
491 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
492 raise_softirq_irqoff(HI_SOFTIRQ);
493 local_irq_restore(flags);
496 EXPORT_SYMBOL(__tasklet_hi_schedule);
498 void __tasklet_hi_schedule_first(struct tasklet_struct *t)
500 BUG_ON(!irqs_disabled());
502 t->next = __this_cpu_read(tasklet_hi_vec.head);
503 __this_cpu_write(tasklet_hi_vec.head, t);
504 __raise_softirq_irqoff(HI_SOFTIRQ);
507 EXPORT_SYMBOL(__tasklet_hi_schedule_first);
509 static void tasklet_action(struct softirq_action *a)
511 struct tasklet_struct *list;
514 list = __this_cpu_read(tasklet_vec.head);
515 __this_cpu_write(tasklet_vec.head, NULL);
516 __this_cpu_write(tasklet_vec.tail, &__get_cpu_var(tasklet_vec).head);
520 struct tasklet_struct *t = list;
524 if (tasklet_trylock(t)) {
525 if (!atomic_read(&t->count)) {
526 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
537 *__this_cpu_read(tasklet_vec.tail) = t;
538 __this_cpu_write(tasklet_vec.tail, &(t->next));
539 __raise_softirq_irqoff(TASKLET_SOFTIRQ);
544 static void tasklet_hi_action(struct softirq_action *a)
546 struct tasklet_struct *list;
549 list = __this_cpu_read(tasklet_hi_vec.head);
550 __this_cpu_write(tasklet_hi_vec.head, NULL);
551 __this_cpu_write(tasklet_hi_vec.tail, &__get_cpu_var(tasklet_hi_vec).head);
555 struct tasklet_struct *t = list;
559 if (tasklet_trylock(t)) {
560 if (!atomic_read(&t->count)) {
561 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
572 *__this_cpu_read(tasklet_hi_vec.tail) = t;
573 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
574 __raise_softirq_irqoff(HI_SOFTIRQ);
580 void tasklet_init(struct tasklet_struct *t,
581 void (*func)(unsigned long), unsigned long data)
585 atomic_set(&t->count, 0);
590 EXPORT_SYMBOL(tasklet_init);
592 void tasklet_kill(struct tasklet_struct *t)
595 printk("Attempt to kill tasklet from interrupt\n");
597 while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
600 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
602 tasklet_unlock_wait(t);
603 clear_bit(TASKLET_STATE_SCHED, &t->state);
606 EXPORT_SYMBOL(tasklet_kill);
613 * The trampoline is called when the hrtimer expires. It schedules a tasklet
614 * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
615 * hrtimer callback, but from softirq context.
617 static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer)
619 struct tasklet_hrtimer *ttimer =
620 container_of(timer, struct tasklet_hrtimer, timer);
622 tasklet_hi_schedule(&ttimer->tasklet);
623 return HRTIMER_NORESTART;
627 * Helper function which calls the hrtimer callback from
628 * tasklet/softirq context
630 static void __tasklet_hrtimer_trampoline(unsigned long data)
632 struct tasklet_hrtimer *ttimer = (void *)data;
633 enum hrtimer_restart restart;
635 restart = ttimer->function(&ttimer->timer);
636 if (restart != HRTIMER_NORESTART)
637 hrtimer_restart(&ttimer->timer);
641 * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
642 * @ttimer: tasklet_hrtimer which is initialized
643 * @function: hrtimer callback function which gets called from softirq context
644 * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
645 * @mode: hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
647 void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer,
648 enum hrtimer_restart (*function)(struct hrtimer *),
649 clockid_t which_clock, enum hrtimer_mode mode)
651 hrtimer_init(&ttimer->timer, which_clock, mode);
652 ttimer->timer.function = __hrtimer_tasklet_trampoline;
653 tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline,
654 (unsigned long)ttimer);
655 ttimer->function = function;
657 EXPORT_SYMBOL_GPL(tasklet_hrtimer_init);
659 void __init softirq_init(void)
663 for_each_possible_cpu(cpu) {
664 per_cpu(tasklet_vec, cpu).tail =
665 &per_cpu(tasklet_vec, cpu).head;
666 per_cpu(tasklet_hi_vec, cpu).tail =
667 &per_cpu(tasklet_hi_vec, cpu).head;
670 open_softirq(TASKLET_SOFTIRQ, tasklet_action);
671 open_softirq(HI_SOFTIRQ, tasklet_hi_action);
674 static int ksoftirqd_should_run(unsigned int cpu)
676 return local_softirq_pending();
679 static void run_ksoftirqd(unsigned int cpu)
682 if (local_softirq_pending()) {
684 * We can safely run softirq on inline stack, as we are not deep
685 * in the task stack here.
688 rcu_note_context_switch(cpu);
696 #ifdef CONFIG_HOTPLUG_CPU
698 * tasklet_kill_immediate is called to remove a tasklet which can already be
699 * scheduled for execution on @cpu.
701 * Unlike tasklet_kill, this function removes the tasklet
702 * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
704 * When this function is called, @cpu must be in the CPU_DEAD state.
706 void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
708 struct tasklet_struct **i;
710 BUG_ON(cpu_online(cpu));
711 BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
713 if (!test_bit(TASKLET_STATE_SCHED, &t->state))
716 /* CPU is dead, so no lock needed. */
717 for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
720 /* If this was the tail element, move the tail ptr */
722 per_cpu(tasklet_vec, cpu).tail = i;
729 static void takeover_tasklets(unsigned int cpu)
731 /* CPU is dead, so no lock needed. */
734 /* Find end, append list for that CPU. */
735 if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
736 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
737 this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
738 per_cpu(tasklet_vec, cpu).head = NULL;
739 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
741 raise_softirq_irqoff(TASKLET_SOFTIRQ);
743 if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
744 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
745 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
746 per_cpu(tasklet_hi_vec, cpu).head = NULL;
747 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
749 raise_softirq_irqoff(HI_SOFTIRQ);
753 #endif /* CONFIG_HOTPLUG_CPU */
755 static int cpu_callback(struct notifier_block *nfb,
756 unsigned long action,
760 #ifdef CONFIG_HOTPLUG_CPU
762 case CPU_DEAD_FROZEN:
763 takeover_tasklets((unsigned long)hcpu);
765 #endif /* CONFIG_HOTPLUG_CPU */
770 static struct notifier_block cpu_nfb = {
771 .notifier_call = cpu_callback
774 static struct smp_hotplug_thread softirq_threads = {
776 .thread_should_run = ksoftirqd_should_run,
777 .thread_fn = run_ksoftirqd,
778 .thread_comm = "ksoftirqd/%u",
781 static __init int spawn_ksoftirqd(void)
783 register_cpu_notifier(&cpu_nfb);
785 BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
789 early_initcall(spawn_ksoftirqd);
792 * [ These __weak aliases are kept in a separate compilation unit, so that
793 * GCC does not inline them incorrectly. ]
796 int __init __weak early_irq_init(void)
801 int __init __weak arch_probe_nr_irqs(void)
803 return NR_IRQS_LEGACY;
806 int __init __weak arch_early_irq_init(void)