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((unsigned long)__builtin_return_address(0),
131 SOFTIRQ_DISABLE_OFFSET
);
134 EXPORT_SYMBOL(local_bh_disable
);
136 static void __local_bh_enable(unsigned int cnt
)
138 WARN_ON_ONCE(in_irq());
139 WARN_ON_ONCE(!irqs_disabled());
141 if (softirq_count() == cnt
)
142 trace_softirqs_on((unsigned long)__builtin_return_address(0));
143 sub_preempt_count(cnt
);
147 * Special-case - softirqs can safely be enabled in
148 * cond_resched_softirq(), or by __do_softirq(),
149 * without processing still-pending softirqs:
151 void _local_bh_enable(void)
153 __local_bh_enable(SOFTIRQ_DISABLE_OFFSET
);
156 EXPORT_SYMBOL(_local_bh_enable
);
158 static inline void _local_bh_enable_ip(unsigned long ip
)
160 WARN_ON_ONCE(in_irq() || irqs_disabled());
161 #ifdef CONFIG_TRACE_IRQFLAGS
165 * Are softirqs going to be turned on now:
167 if (softirq_count() == SOFTIRQ_DISABLE_OFFSET
)
168 trace_softirqs_on(ip
);
170 * Keep preemption disabled until we are done with
171 * softirq processing:
173 sub_preempt_count(SOFTIRQ_DISABLE_OFFSET
- 1);
175 if (unlikely(!in_interrupt() && local_softirq_pending()))
179 #ifdef CONFIG_TRACE_IRQFLAGS
182 preempt_check_resched();
185 void local_bh_enable(void)
187 _local_bh_enable_ip((unsigned long)__builtin_return_address(0));
189 EXPORT_SYMBOL(local_bh_enable
);
191 void local_bh_enable_ip(unsigned long ip
)
193 _local_bh_enable_ip(ip
);
195 EXPORT_SYMBOL(local_bh_enable_ip
);
198 * We restart softirq processing for at most 2 ms,
199 * and if need_resched() is not set.
201 * These limits have been established via experimentation.
202 * The two things to balance is latency against fairness -
203 * we want to handle softirqs as soon as possible, but they
204 * should not be able to lock up the box.
206 #define MAX_SOFTIRQ_TIME msecs_to_jiffies(2)
208 asmlinkage
void __do_softirq(void)
210 struct softirq_action
*h
;
212 unsigned long end
= jiffies
+ MAX_SOFTIRQ_TIME
;
214 unsigned long old_flags
= current
->flags
;
217 * Mask out PF_MEMALLOC s current task context is borrowed for the
218 * softirq. A softirq handled such as network RX might set PF_MEMALLOC
219 * again if the socket is related to swap
221 current
->flags
&= ~PF_MEMALLOC
;
223 pending
= local_softirq_pending();
224 account_irq_enter_time(current
);
226 __local_bh_disable((unsigned long)__builtin_return_address(0),
228 lockdep_softirq_enter();
230 cpu
= smp_processor_id();
232 /* Reset the pending bitmask before enabling irqs */
233 set_softirq_pending(0);
241 unsigned int vec_nr
= h
- softirq_vec
;
242 int prev_count
= preempt_count();
244 kstat_incr_softirqs_this_cpu(vec_nr
);
246 trace_softirq_entry(vec_nr
);
248 trace_softirq_exit(vec_nr
);
249 if (unlikely(prev_count
!= preempt_count())) {
250 printk(KERN_ERR
"huh, entered softirq %u %s %p"
251 "with preempt_count %08x,"
252 " exited with %08x?\n", vec_nr
,
253 softirq_to_name
[vec_nr
], h
->action
,
254 prev_count
, preempt_count());
255 preempt_count() = prev_count
;
266 pending
= local_softirq_pending();
268 if (time_before(jiffies
, end
) && !need_resched())
274 lockdep_softirq_exit();
276 account_irq_exit_time(current
);
277 __local_bh_enable(SOFTIRQ_OFFSET
);
278 tsk_restore_flags(current
, old_flags
, PF_MEMALLOC
);
281 #ifndef __ARCH_HAS_DO_SOFTIRQ
283 asmlinkage
void do_softirq(void)
291 local_irq_save(flags
);
293 pending
= local_softirq_pending();
298 local_irq_restore(flags
);
304 * Enter an interrupt context.
308 int cpu
= smp_processor_id();
311 if (is_idle_task(current
) && !in_interrupt()) {
313 * Prevent raise_softirq from needlessly waking up ksoftirqd
314 * here, as softirq will be serviced on return from interrupt.
317 tick_check_idle(cpu
);
324 static inline void invoke_softirq(void)
326 if (!force_irqthreads
)
333 * Exit an interrupt context. Process softirqs if needed and possible:
337 #ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
340 WARN_ON_ONCE(!irqs_disabled());
343 account_irq_exit_time(current
);
344 trace_hardirq_exit();
345 sub_preempt_count(HARDIRQ_OFFSET
);
346 if (!in_interrupt() && local_softirq_pending())
350 /* Make sure that timer wheel updates are propagated */
351 if (idle_cpu(smp_processor_id()) && !in_interrupt() && !need_resched())
352 tick_nohz_irq_exit();
358 * This function must run with irqs disabled!
360 inline void raise_softirq_irqoff(unsigned int nr
)
362 __raise_softirq_irqoff(nr
);
365 * If we're in an interrupt or softirq, we're done
366 * (this also catches softirq-disabled code). We will
367 * actually run the softirq once we return from
368 * the irq or softirq.
370 * Otherwise we wake up ksoftirqd to make sure we
371 * schedule the softirq soon.
377 void raise_softirq(unsigned int nr
)
381 local_irq_save(flags
);
382 raise_softirq_irqoff(nr
);
383 local_irq_restore(flags
);
386 void __raise_softirq_irqoff(unsigned int nr
)
388 trace_softirq_raise(nr
);
389 or_softirq_pending(1UL << nr
);
392 void open_softirq(int nr
, void (*action
)(struct softirq_action
*))
394 softirq_vec
[nr
].action
= action
;
402 struct tasklet_struct
*head
;
403 struct tasklet_struct
**tail
;
406 static DEFINE_PER_CPU(struct tasklet_head
, tasklet_vec
);
407 static DEFINE_PER_CPU(struct tasklet_head
, tasklet_hi_vec
);
409 void __tasklet_schedule(struct tasklet_struct
*t
)
413 local_irq_save(flags
);
415 *__this_cpu_read(tasklet_vec
.tail
) = t
;
416 __this_cpu_write(tasklet_vec
.tail
, &(t
->next
));
417 raise_softirq_irqoff(TASKLET_SOFTIRQ
);
418 local_irq_restore(flags
);
421 EXPORT_SYMBOL(__tasklet_schedule
);
423 void __tasklet_hi_schedule(struct tasklet_struct
*t
)
427 local_irq_save(flags
);
429 *__this_cpu_read(tasklet_hi_vec
.tail
) = t
;
430 __this_cpu_write(tasklet_hi_vec
.tail
, &(t
->next
));
431 raise_softirq_irqoff(HI_SOFTIRQ
);
432 local_irq_restore(flags
);
435 EXPORT_SYMBOL(__tasklet_hi_schedule
);
437 void __tasklet_hi_schedule_first(struct tasklet_struct
*t
)
439 BUG_ON(!irqs_disabled());
441 t
->next
= __this_cpu_read(tasklet_hi_vec
.head
);
442 __this_cpu_write(tasklet_hi_vec
.head
, t
);
443 __raise_softirq_irqoff(HI_SOFTIRQ
);
446 EXPORT_SYMBOL(__tasklet_hi_schedule_first
);
448 static void tasklet_action(struct softirq_action
*a
)
450 struct tasklet_struct
*list
;
453 list
= __this_cpu_read(tasklet_vec
.head
);
454 __this_cpu_write(tasklet_vec
.head
, NULL
);
455 __this_cpu_write(tasklet_vec
.tail
, &__get_cpu_var(tasklet_vec
).head
);
459 struct tasklet_struct
*t
= list
;
463 if (tasklet_trylock(t
)) {
464 if (!atomic_read(&t
->count
)) {
465 if (!test_and_clear_bit(TASKLET_STATE_SCHED
, &t
->state
))
476 *__this_cpu_read(tasklet_vec
.tail
) = t
;
477 __this_cpu_write(tasklet_vec
.tail
, &(t
->next
));
478 __raise_softirq_irqoff(TASKLET_SOFTIRQ
);
483 static void tasklet_hi_action(struct softirq_action
*a
)
485 struct tasklet_struct
*list
;
488 list
= __this_cpu_read(tasklet_hi_vec
.head
);
489 __this_cpu_write(tasklet_hi_vec
.head
, NULL
);
490 __this_cpu_write(tasklet_hi_vec
.tail
, &__get_cpu_var(tasklet_hi_vec
).head
);
494 struct tasklet_struct
*t
= list
;
498 if (tasklet_trylock(t
)) {
499 if (!atomic_read(&t
->count
)) {
500 if (!test_and_clear_bit(TASKLET_STATE_SCHED
, &t
->state
))
511 *__this_cpu_read(tasklet_hi_vec
.tail
) = t
;
512 __this_cpu_write(tasklet_hi_vec
.tail
, &(t
->next
));
513 __raise_softirq_irqoff(HI_SOFTIRQ
);
519 void tasklet_init(struct tasklet_struct
*t
,
520 void (*func
)(unsigned long), unsigned long data
)
524 atomic_set(&t
->count
, 0);
529 EXPORT_SYMBOL(tasklet_init
);
531 void tasklet_kill(struct tasklet_struct
*t
)
534 printk("Attempt to kill tasklet from interrupt\n");
536 while (test_and_set_bit(TASKLET_STATE_SCHED
, &t
->state
)) {
539 } while (test_bit(TASKLET_STATE_SCHED
, &t
->state
));
541 tasklet_unlock_wait(t
);
542 clear_bit(TASKLET_STATE_SCHED
, &t
->state
);
545 EXPORT_SYMBOL(tasklet_kill
);
552 * The trampoline is called when the hrtimer expires. It schedules a tasklet
553 * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
554 * hrtimer callback, but from softirq context.
556 static enum hrtimer_restart
__hrtimer_tasklet_trampoline(struct hrtimer
*timer
)
558 struct tasklet_hrtimer
*ttimer
=
559 container_of(timer
, struct tasklet_hrtimer
, timer
);
561 tasklet_hi_schedule(&ttimer
->tasklet
);
562 return HRTIMER_NORESTART
;
566 * Helper function which calls the hrtimer callback from
567 * tasklet/softirq context
569 static void __tasklet_hrtimer_trampoline(unsigned long data
)
571 struct tasklet_hrtimer
*ttimer
= (void *)data
;
572 enum hrtimer_restart restart
;
574 restart
= ttimer
->function(&ttimer
->timer
);
575 if (restart
!= HRTIMER_NORESTART
)
576 hrtimer_restart(&ttimer
->timer
);
580 * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
581 * @ttimer: tasklet_hrtimer which is initialized
582 * @function: hrtimer callback function which gets called from softirq context
583 * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
584 * @mode: hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
586 void tasklet_hrtimer_init(struct tasklet_hrtimer
*ttimer
,
587 enum hrtimer_restart (*function
)(struct hrtimer
*),
588 clockid_t which_clock
, enum hrtimer_mode mode
)
590 hrtimer_init(&ttimer
->timer
, which_clock
, mode
);
591 ttimer
->timer
.function
= __hrtimer_tasklet_trampoline
;
592 tasklet_init(&ttimer
->tasklet
, __tasklet_hrtimer_trampoline
,
593 (unsigned long)ttimer
);
594 ttimer
->function
= function
;
596 EXPORT_SYMBOL_GPL(tasklet_hrtimer_init
);
599 * Remote softirq bits
602 DEFINE_PER_CPU(struct list_head
[NR_SOFTIRQS
], softirq_work_list
);
603 EXPORT_PER_CPU_SYMBOL(softirq_work_list
);
605 static void __local_trigger(struct call_single_data
*cp
, int softirq
)
607 struct list_head
*head
= &__get_cpu_var(softirq_work_list
[softirq
]);
609 list_add_tail(&cp
->list
, head
);
611 /* Trigger the softirq only if the list was previously empty. */
612 if (head
->next
== &cp
->list
)
613 raise_softirq_irqoff(softirq
);
616 #ifdef CONFIG_USE_GENERIC_SMP_HELPERS
617 static void remote_softirq_receive(void *data
)
619 struct call_single_data
*cp
= data
;
625 local_irq_save(flags
);
626 __local_trigger(cp
, softirq
);
627 local_irq_restore(flags
);
630 static int __try_remote_softirq(struct call_single_data
*cp
, int cpu
, int softirq
)
632 if (cpu_online(cpu
)) {
633 cp
->func
= remote_softirq_receive
;
638 __smp_call_function_single(cpu
, cp
, 0);
643 #else /* CONFIG_USE_GENERIC_SMP_HELPERS */
644 static int __try_remote_softirq(struct call_single_data
*cp
, int cpu
, int softirq
)
651 * __send_remote_softirq - try to schedule softirq work on a remote cpu
652 * @cp: private SMP call function data area
653 * @cpu: the remote cpu
654 * @this_cpu: the currently executing cpu
655 * @softirq: the softirq for the work
657 * Attempt to schedule softirq work on a remote cpu. If this cannot be
658 * done, the work is instead queued up on the local cpu.
660 * Interrupts must be disabled.
662 void __send_remote_softirq(struct call_single_data
*cp
, int cpu
, int this_cpu
, int softirq
)
664 if (cpu
== this_cpu
|| __try_remote_softirq(cp
, cpu
, softirq
))
665 __local_trigger(cp
, softirq
);
667 EXPORT_SYMBOL(__send_remote_softirq
);
670 * send_remote_softirq - try to schedule softirq work on a remote cpu
671 * @cp: private SMP call function data area
672 * @cpu: the remote cpu
673 * @softirq: the softirq for the work
675 * Like __send_remote_softirq except that disabling interrupts and
676 * computing the current cpu is done for the caller.
678 void send_remote_softirq(struct call_single_data
*cp
, int cpu
, int softirq
)
683 local_irq_save(flags
);
684 this_cpu
= smp_processor_id();
685 __send_remote_softirq(cp
, cpu
, this_cpu
, softirq
);
686 local_irq_restore(flags
);
688 EXPORT_SYMBOL(send_remote_softirq
);
690 static int __cpuinit
remote_softirq_cpu_notify(struct notifier_block
*self
,
691 unsigned long action
, void *hcpu
)
694 * If a CPU goes away, splice its entries to the current CPU
695 * and trigger a run of the softirq
697 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
698 int cpu
= (unsigned long) hcpu
;
702 for (i
= 0; i
< NR_SOFTIRQS
; i
++) {
703 struct list_head
*head
= &per_cpu(softirq_work_list
[i
], cpu
);
704 struct list_head
*local_head
;
706 if (list_empty(head
))
709 local_head
= &__get_cpu_var(softirq_work_list
[i
]);
710 list_splice_init(head
, local_head
);
711 raise_softirq_irqoff(i
);
719 static struct notifier_block __cpuinitdata remote_softirq_cpu_notifier
= {
720 .notifier_call
= remote_softirq_cpu_notify
,
723 void __init
softirq_init(void)
727 for_each_possible_cpu(cpu
) {
730 per_cpu(tasklet_vec
, cpu
).tail
=
731 &per_cpu(tasklet_vec
, cpu
).head
;
732 per_cpu(tasklet_hi_vec
, cpu
).tail
=
733 &per_cpu(tasklet_hi_vec
, cpu
).head
;
734 for (i
= 0; i
< NR_SOFTIRQS
; i
++)
735 INIT_LIST_HEAD(&per_cpu(softirq_work_list
[i
], cpu
));
738 register_hotcpu_notifier(&remote_softirq_cpu_notifier
);
740 open_softirq(TASKLET_SOFTIRQ
, tasklet_action
);
741 open_softirq(HI_SOFTIRQ
, tasklet_hi_action
);
744 static int ksoftirqd_should_run(unsigned int cpu
)
746 return local_softirq_pending();
749 static void run_ksoftirqd(unsigned int cpu
)
752 if (local_softirq_pending()) {
754 rcu_note_context_switch(cpu
);
762 #ifdef CONFIG_HOTPLUG_CPU
764 * tasklet_kill_immediate is called to remove a tasklet which can already be
765 * scheduled for execution on @cpu.
767 * Unlike tasklet_kill, this function removes the tasklet
768 * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
770 * When this function is called, @cpu must be in the CPU_DEAD state.
772 void tasklet_kill_immediate(struct tasklet_struct
*t
, unsigned int cpu
)
774 struct tasklet_struct
**i
;
776 BUG_ON(cpu_online(cpu
));
777 BUG_ON(test_bit(TASKLET_STATE_RUN
, &t
->state
));
779 if (!test_bit(TASKLET_STATE_SCHED
, &t
->state
))
782 /* CPU is dead, so no lock needed. */
783 for (i
= &per_cpu(tasklet_vec
, cpu
).head
; *i
; i
= &(*i
)->next
) {
786 /* If this was the tail element, move the tail ptr */
788 per_cpu(tasklet_vec
, cpu
).tail
= i
;
795 static void takeover_tasklets(unsigned int cpu
)
797 /* CPU is dead, so no lock needed. */
800 /* Find end, append list for that CPU. */
801 if (&per_cpu(tasklet_vec
, cpu
).head
!= per_cpu(tasklet_vec
, cpu
).tail
) {
802 *__this_cpu_read(tasklet_vec
.tail
) = per_cpu(tasklet_vec
, cpu
).head
;
803 this_cpu_write(tasklet_vec
.tail
, per_cpu(tasklet_vec
, cpu
).tail
);
804 per_cpu(tasklet_vec
, cpu
).head
= NULL
;
805 per_cpu(tasklet_vec
, cpu
).tail
= &per_cpu(tasklet_vec
, cpu
).head
;
807 raise_softirq_irqoff(TASKLET_SOFTIRQ
);
809 if (&per_cpu(tasklet_hi_vec
, cpu
).head
!= per_cpu(tasklet_hi_vec
, cpu
).tail
) {
810 *__this_cpu_read(tasklet_hi_vec
.tail
) = per_cpu(tasklet_hi_vec
, cpu
).head
;
811 __this_cpu_write(tasklet_hi_vec
.tail
, per_cpu(tasklet_hi_vec
, cpu
).tail
);
812 per_cpu(tasklet_hi_vec
, cpu
).head
= NULL
;
813 per_cpu(tasklet_hi_vec
, cpu
).tail
= &per_cpu(tasklet_hi_vec
, cpu
).head
;
815 raise_softirq_irqoff(HI_SOFTIRQ
);
819 #endif /* CONFIG_HOTPLUG_CPU */
821 static int __cpuinit
cpu_callback(struct notifier_block
*nfb
,
822 unsigned long action
,
826 #ifdef CONFIG_HOTPLUG_CPU
828 case CPU_DEAD_FROZEN
:
829 takeover_tasklets((unsigned long)hcpu
);
831 #endif /* CONFIG_HOTPLUG_CPU */
836 static struct notifier_block __cpuinitdata cpu_nfb
= {
837 .notifier_call
= cpu_callback
840 static struct smp_hotplug_thread softirq_threads
= {
842 .thread_should_run
= ksoftirqd_should_run
,
843 .thread_fn
= run_ksoftirqd
,
844 .thread_comm
= "ksoftirqd/%u",
847 static __init
int spawn_ksoftirqd(void)
849 register_cpu_notifier(&cpu_nfb
);
851 BUG_ON(smpboot_register_percpu_thread(&softirq_threads
));
855 early_initcall(spawn_ksoftirqd
);
858 * [ These __weak aliases are kept in a separate compilation unit, so that
859 * GCC does not inline them incorrectly. ]
862 int __init __weak
early_irq_init(void)
867 #ifdef CONFIG_GENERIC_HARDIRQS
868 int __init __weak
arch_probe_nr_irqs(void)
870 return NR_IRQS_LEGACY
;
873 int __init __weak
arch_early_irq_init(void)