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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
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>
28 #include <linux/irq.h>
30 #define CREATE_TRACE_POINTS
31 #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 const char * const softirq_to_name
[NR_SOFTIRQS
] = {
61 "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "IRQ_POLL",
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 * If ksoftirqd is scheduled, we do not want to process pending softirqs
82 * right now. Let ksoftirqd handle this at its own rate, to get fairness.
84 static bool ksoftirqd_running(void)
86 struct task_struct
*tsk
= __this_cpu_read(ksoftirqd
);
88 return tsk
&& (tsk
->state
== TASK_RUNNING
);
92 * preempt_count and SOFTIRQ_OFFSET usage:
93 * - preempt_count is changed by SOFTIRQ_OFFSET on entering or leaving
95 * - preempt_count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
96 * on local_bh_disable or local_bh_enable.
97 * This lets us distinguish between whether we are currently processing
98 * softirq and whether we just have bh disabled.
102 * This one is for softirq.c-internal use,
103 * where hardirqs are disabled legitimately:
105 #ifdef CONFIG_TRACE_IRQFLAGS
106 void __local_bh_disable_ip(unsigned long ip
, unsigned int cnt
)
110 WARN_ON_ONCE(in_irq());
112 raw_local_irq_save(flags
);
114 * The preempt tracer hooks into preempt_count_add and will break
115 * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
116 * is set and before current->softirq_enabled is cleared.
117 * We must manually increment preempt_count here and manually
118 * call the trace_preempt_off later.
120 __preempt_count_add(cnt
);
122 * Were softirqs turned off above:
124 if (softirq_count() == (cnt
& SOFTIRQ_MASK
))
125 trace_softirqs_off(ip
);
126 raw_local_irq_restore(flags
);
128 if (preempt_count() == cnt
) {
129 #ifdef CONFIG_DEBUG_PREEMPT
130 current
->preempt_disable_ip
= get_lock_parent_ip();
132 trace_preempt_off(CALLER_ADDR0
, get_lock_parent_ip());
135 EXPORT_SYMBOL(__local_bh_disable_ip
);
136 #endif /* CONFIG_TRACE_IRQFLAGS */
138 static void __local_bh_enable(unsigned int cnt
)
140 lockdep_assert_irqs_disabled();
142 if (softirq_count() == (cnt
& SOFTIRQ_MASK
))
143 trace_softirqs_on(_RET_IP_
);
144 preempt_count_sub(cnt
);
148 * Special-case - softirqs can safely be enabled in
149 * cond_resched_softirq(), or by __do_softirq(),
150 * without processing still-pending softirqs:
152 void _local_bh_enable(void)
154 WARN_ON_ONCE(in_irq());
155 __local_bh_enable(SOFTIRQ_DISABLE_OFFSET
);
157 EXPORT_SYMBOL(_local_bh_enable
);
159 void __local_bh_enable_ip(unsigned long ip
, unsigned int cnt
)
161 WARN_ON_ONCE(in_irq());
162 lockdep_assert_irqs_enabled();
163 #ifdef CONFIG_TRACE_IRQFLAGS
167 * Are softirqs going to be turned on now:
169 if (softirq_count() == SOFTIRQ_DISABLE_OFFSET
)
170 trace_softirqs_on(ip
);
172 * Keep preemption disabled until we are done with
173 * softirq processing:
175 preempt_count_sub(cnt
- 1);
177 if (unlikely(!in_interrupt() && local_softirq_pending())) {
179 * Run softirq if any pending. And do it in its own stack
180 * as we may be calling this deep in a task call stack already.
186 #ifdef CONFIG_TRACE_IRQFLAGS
189 preempt_check_resched();
191 EXPORT_SYMBOL(__local_bh_enable_ip
);
194 * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times,
195 * but break the loop if need_resched() is set or after 2 ms.
196 * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in
197 * certain cases, such as stop_machine(), jiffies may cease to
198 * increment and so we need the MAX_SOFTIRQ_RESTART limit as
199 * well to make sure we eventually return from this method.
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)
207 #define MAX_SOFTIRQ_RESTART 10
209 #ifdef CONFIG_TRACE_IRQFLAGS
211 * When we run softirqs from irq_exit() and thus on the hardirq stack we need
212 * to keep the lockdep irq context tracking as tight as possible in order to
213 * not miss-qualify lock contexts and miss possible deadlocks.
216 static inline bool lockdep_softirq_start(void)
218 bool in_hardirq
= false;
220 if (trace_hardirq_context(current
)) {
222 trace_hardirq_exit();
225 lockdep_softirq_enter();
230 static inline void lockdep_softirq_end(bool in_hardirq
)
232 lockdep_softirq_exit();
235 trace_hardirq_enter();
238 static inline bool lockdep_softirq_start(void) { return false; }
239 static inline void lockdep_softirq_end(bool in_hardirq
) { }
242 asmlinkage __visible
void __softirq_entry
__do_softirq(void)
244 unsigned long end
= jiffies
+ MAX_SOFTIRQ_TIME
;
245 unsigned long old_flags
= current
->flags
;
246 int max_restart
= MAX_SOFTIRQ_RESTART
;
247 struct softirq_action
*h
;
253 * Mask out PF_MEMALLOC s current task context is borrowed for the
254 * softirq. A softirq handled such as network RX might set PF_MEMALLOC
255 * again if the socket is related to swap
257 current
->flags
&= ~PF_MEMALLOC
;
259 pending
= local_softirq_pending();
260 account_irq_enter_time(current
);
262 __local_bh_disable_ip(_RET_IP_
, SOFTIRQ_OFFSET
);
263 in_hardirq
= lockdep_softirq_start();
266 /* Reset the pending bitmask before enabling irqs */
267 set_softirq_pending(0);
273 while ((softirq_bit
= ffs(pending
))) {
277 h
+= softirq_bit
- 1;
279 vec_nr
= h
- softirq_vec
;
280 prev_count
= preempt_count();
282 kstat_incr_softirqs_this_cpu(vec_nr
);
284 trace_softirq_entry(vec_nr
);
286 trace_softirq_exit(vec_nr
);
287 if (unlikely(prev_count
!= preempt_count())) {
288 pr_err("huh, entered softirq %u %s %p with preempt_count %08x, exited with %08x?\n",
289 vec_nr
, softirq_to_name
[vec_nr
], h
->action
,
290 prev_count
, preempt_count());
291 preempt_count_set(prev_count
);
294 pending
>>= softirq_bit
;
300 pending
= local_softirq_pending();
302 if (time_before(jiffies
, end
) && !need_resched() &&
309 lockdep_softirq_end(in_hardirq
);
310 account_irq_exit_time(current
);
311 __local_bh_enable(SOFTIRQ_OFFSET
);
312 WARN_ON_ONCE(in_interrupt());
313 current_restore_flags(old_flags
, PF_MEMALLOC
);
316 asmlinkage __visible
void do_softirq(void)
324 local_irq_save(flags
);
326 pending
= local_softirq_pending();
328 if (pending
&& !ksoftirqd_running())
329 do_softirq_own_stack();
331 local_irq_restore(flags
);
335 * Enter an interrupt context.
340 if (is_idle_task(current
) && !in_interrupt()) {
342 * Prevent raise_softirq from needlessly waking up ksoftirqd
343 * here, as softirq will be serviced on return from interrupt.
353 static inline void invoke_softirq(void)
355 if (ksoftirqd_running())
358 if (!force_irqthreads
) {
359 #ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
361 * We can safely execute softirq on the current stack if
362 * it is the irq stack, because it should be near empty
368 * Otherwise, irq_exit() is called on the task stack that can
369 * be potentially deep already. So call softirq in its own stack
370 * to prevent from any overrun.
372 do_softirq_own_stack();
379 static inline void tick_irq_exit(void)
381 #ifdef CONFIG_NO_HZ_COMMON
382 int cpu
= smp_processor_id();
384 /* Make sure that timer wheel updates are propagated */
385 if ((idle_cpu(cpu
) && !need_resched()) || tick_nohz_full_cpu(cpu
)) {
387 tick_nohz_irq_exit();
393 * Exit an interrupt context. Process softirqs if needed and possible:
397 #ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
400 lockdep_assert_irqs_disabled();
402 account_irq_exit_time(current
);
403 preempt_count_sub(HARDIRQ_OFFSET
);
404 if (!in_interrupt() && local_softirq_pending())
409 trace_hardirq_exit(); /* must be last! */
413 * This function must run with irqs disabled!
415 inline void raise_softirq_irqoff(unsigned int nr
)
417 __raise_softirq_irqoff(nr
);
420 * If we're in an interrupt or softirq, we're done
421 * (this also catches softirq-disabled code). We will
422 * actually run the softirq once we return from
423 * the irq or softirq.
425 * Otherwise we wake up ksoftirqd to make sure we
426 * schedule the softirq soon.
432 void raise_softirq(unsigned int nr
)
436 local_irq_save(flags
);
437 raise_softirq_irqoff(nr
);
438 local_irq_restore(flags
);
441 void __raise_softirq_irqoff(unsigned int nr
)
443 trace_softirq_raise(nr
);
444 or_softirq_pending(1UL << nr
);
447 void open_softirq(int nr
, void (*action
)(struct softirq_action
*))
449 softirq_vec
[nr
].action
= action
;
455 struct tasklet_head
{
456 struct tasklet_struct
*head
;
457 struct tasklet_struct
**tail
;
460 static DEFINE_PER_CPU(struct tasklet_head
, tasklet_vec
);
461 static DEFINE_PER_CPU(struct tasklet_head
, tasklet_hi_vec
);
463 void __tasklet_schedule(struct tasklet_struct
*t
)
467 local_irq_save(flags
);
469 *__this_cpu_read(tasklet_vec
.tail
) = t
;
470 __this_cpu_write(tasklet_vec
.tail
, &(t
->next
));
471 raise_softirq_irqoff(TASKLET_SOFTIRQ
);
472 local_irq_restore(flags
);
474 EXPORT_SYMBOL(__tasklet_schedule
);
476 void __tasklet_hi_schedule(struct tasklet_struct
*t
)
480 local_irq_save(flags
);
482 *__this_cpu_read(tasklet_hi_vec
.tail
) = t
;
483 __this_cpu_write(tasklet_hi_vec
.tail
, &(t
->next
));
484 raise_softirq_irqoff(HI_SOFTIRQ
);
485 local_irq_restore(flags
);
487 EXPORT_SYMBOL(__tasklet_hi_schedule
);
489 static __latent_entropy
void tasklet_action(struct softirq_action
*a
)
491 struct tasklet_struct
*list
;
494 list
= __this_cpu_read(tasklet_vec
.head
);
495 __this_cpu_write(tasklet_vec
.head
, NULL
);
496 __this_cpu_write(tasklet_vec
.tail
, this_cpu_ptr(&tasklet_vec
.head
));
500 struct tasklet_struct
*t
= list
;
504 if (tasklet_trylock(t
)) {
505 if (!atomic_read(&t
->count
)) {
506 if (!test_and_clear_bit(TASKLET_STATE_SCHED
,
518 *__this_cpu_read(tasklet_vec
.tail
) = t
;
519 __this_cpu_write(tasklet_vec
.tail
, &(t
->next
));
520 __raise_softirq_irqoff(TASKLET_SOFTIRQ
);
525 static __latent_entropy
void tasklet_hi_action(struct softirq_action
*a
)
527 struct tasklet_struct
*list
;
530 list
= __this_cpu_read(tasklet_hi_vec
.head
);
531 __this_cpu_write(tasklet_hi_vec
.head
, NULL
);
532 __this_cpu_write(tasklet_hi_vec
.tail
, this_cpu_ptr(&tasklet_hi_vec
.head
));
536 struct tasklet_struct
*t
= list
;
540 if (tasklet_trylock(t
)) {
541 if (!atomic_read(&t
->count
)) {
542 if (!test_and_clear_bit(TASKLET_STATE_SCHED
,
554 *__this_cpu_read(tasklet_hi_vec
.tail
) = t
;
555 __this_cpu_write(tasklet_hi_vec
.tail
, &(t
->next
));
556 __raise_softirq_irqoff(HI_SOFTIRQ
);
561 void tasklet_init(struct tasklet_struct
*t
,
562 void (*func
)(unsigned long), unsigned long data
)
566 atomic_set(&t
->count
, 0);
570 EXPORT_SYMBOL(tasklet_init
);
572 void tasklet_kill(struct tasklet_struct
*t
)
575 pr_notice("Attempt to kill tasklet from interrupt\n");
577 while (test_and_set_bit(TASKLET_STATE_SCHED
, &t
->state
)) {
580 } while (test_bit(TASKLET_STATE_SCHED
, &t
->state
));
582 tasklet_unlock_wait(t
);
583 clear_bit(TASKLET_STATE_SCHED
, &t
->state
);
585 EXPORT_SYMBOL(tasklet_kill
);
592 * The trampoline is called when the hrtimer expires. It schedules a tasklet
593 * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
594 * hrtimer callback, but from softirq context.
596 static enum hrtimer_restart
__hrtimer_tasklet_trampoline(struct hrtimer
*timer
)
598 struct tasklet_hrtimer
*ttimer
=
599 container_of(timer
, struct tasklet_hrtimer
, timer
);
601 tasklet_hi_schedule(&ttimer
->tasklet
);
602 return HRTIMER_NORESTART
;
606 * Helper function which calls the hrtimer callback from
607 * tasklet/softirq context
609 static void __tasklet_hrtimer_trampoline(unsigned long data
)
611 struct tasklet_hrtimer
*ttimer
= (void *)data
;
612 enum hrtimer_restart restart
;
614 restart
= ttimer
->function(&ttimer
->timer
);
615 if (restart
!= HRTIMER_NORESTART
)
616 hrtimer_restart(&ttimer
->timer
);
620 * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
621 * @ttimer: tasklet_hrtimer which is initialized
622 * @function: hrtimer callback function which gets called from softirq context
623 * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
624 * @mode: hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
626 void tasklet_hrtimer_init(struct tasklet_hrtimer
*ttimer
,
627 enum hrtimer_restart (*function
)(struct hrtimer
*),
628 clockid_t which_clock
, enum hrtimer_mode mode
)
630 hrtimer_init(&ttimer
->timer
, which_clock
, mode
);
631 ttimer
->timer
.function
= __hrtimer_tasklet_trampoline
;
632 tasklet_init(&ttimer
->tasklet
, __tasklet_hrtimer_trampoline
,
633 (unsigned long)ttimer
);
634 ttimer
->function
= function
;
636 EXPORT_SYMBOL_GPL(tasklet_hrtimer_init
);
638 void __init
softirq_init(void)
642 for_each_possible_cpu(cpu
) {
643 per_cpu(tasklet_vec
, cpu
).tail
=
644 &per_cpu(tasklet_vec
, cpu
).head
;
645 per_cpu(tasklet_hi_vec
, cpu
).tail
=
646 &per_cpu(tasklet_hi_vec
, cpu
).head
;
649 open_softirq(TASKLET_SOFTIRQ
, tasklet_action
);
650 open_softirq(HI_SOFTIRQ
, tasklet_hi_action
);
653 static int ksoftirqd_should_run(unsigned int cpu
)
655 return local_softirq_pending();
658 static void run_ksoftirqd(unsigned int cpu
)
661 if (local_softirq_pending()) {
663 * We can safely run softirq on inline stack, as we are not deep
664 * in the task stack here.
674 #ifdef CONFIG_HOTPLUG_CPU
676 * tasklet_kill_immediate is called to remove a tasklet which can already be
677 * scheduled for execution on @cpu.
679 * Unlike tasklet_kill, this function removes the tasklet
680 * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
682 * When this function is called, @cpu must be in the CPU_DEAD state.
684 void tasklet_kill_immediate(struct tasklet_struct
*t
, unsigned int cpu
)
686 struct tasklet_struct
**i
;
688 BUG_ON(cpu_online(cpu
));
689 BUG_ON(test_bit(TASKLET_STATE_RUN
, &t
->state
));
691 if (!test_bit(TASKLET_STATE_SCHED
, &t
->state
))
694 /* CPU is dead, so no lock needed. */
695 for (i
= &per_cpu(tasklet_vec
, cpu
).head
; *i
; i
= &(*i
)->next
) {
698 /* If this was the tail element, move the tail ptr */
700 per_cpu(tasklet_vec
, cpu
).tail
= i
;
707 static int takeover_tasklets(unsigned int cpu
)
709 /* CPU is dead, so no lock needed. */
712 /* Find end, append list for that CPU. */
713 if (&per_cpu(tasklet_vec
, cpu
).head
!= per_cpu(tasklet_vec
, cpu
).tail
) {
714 *__this_cpu_read(tasklet_vec
.tail
) = per_cpu(tasklet_vec
, cpu
).head
;
715 this_cpu_write(tasklet_vec
.tail
, per_cpu(tasklet_vec
, cpu
).tail
);
716 per_cpu(tasklet_vec
, cpu
).head
= NULL
;
717 per_cpu(tasklet_vec
, cpu
).tail
= &per_cpu(tasklet_vec
, cpu
).head
;
719 raise_softirq_irqoff(TASKLET_SOFTIRQ
);
721 if (&per_cpu(tasklet_hi_vec
, cpu
).head
!= per_cpu(tasklet_hi_vec
, cpu
).tail
) {
722 *__this_cpu_read(tasklet_hi_vec
.tail
) = per_cpu(tasklet_hi_vec
, cpu
).head
;
723 __this_cpu_write(tasklet_hi_vec
.tail
, per_cpu(tasklet_hi_vec
, cpu
).tail
);
724 per_cpu(tasklet_hi_vec
, cpu
).head
= NULL
;
725 per_cpu(tasklet_hi_vec
, cpu
).tail
= &per_cpu(tasklet_hi_vec
, cpu
).head
;
727 raise_softirq_irqoff(HI_SOFTIRQ
);
733 #define takeover_tasklets NULL
734 #endif /* CONFIG_HOTPLUG_CPU */
736 static struct smp_hotplug_thread softirq_threads
= {
738 .thread_should_run
= ksoftirqd_should_run
,
739 .thread_fn
= run_ksoftirqd
,
740 .thread_comm
= "ksoftirqd/%u",
743 static __init
int spawn_ksoftirqd(void)
745 cpuhp_setup_state_nocalls(CPUHP_SOFTIRQ_DEAD
, "softirq:dead", NULL
,
747 BUG_ON(smpboot_register_percpu_thread(&softirq_threads
));
751 early_initcall(spawn_ksoftirqd
);
754 * [ These __weak aliases are kept in a separate compilation unit, so that
755 * GCC does not inline them incorrectly. ]
758 int __init __weak
early_irq_init(void)
763 int __init __weak
arch_probe_nr_irqs(void)
765 return NR_IRQS_LEGACY
;
768 int __init __weak
arch_early_irq_init(void)
773 unsigned int __weak
arch_dynirq_lower_bound(unsigned int from
)