Loading...
1/*
2 * linux/kernel/softirq.c
3 *
4 * Copyright (C) 1992 Linus Torvalds
5 *
6 * Distribute under GPLv2.
7 *
8 * Rewritten. Old one was good in 2.2, but in 2.3 it was immoral. --ANK (990903)
9 *
10 * Remote softirq infrastructure is by Jens Axboe.
11 */
12
13#include <linux/export.h>
14#include <linux/kernel_stat.h>
15#include <linux/interrupt.h>
16#include <linux/init.h>
17#include <linux/mm.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/tick.h>
27
28#define CREATE_TRACE_POINTS
29#include <trace/events/irq.h>
30
31#include <asm/irq.h>
32/*
33 - No shared variables, all the data are CPU local.
34 - If a softirq needs serialization, let it serialize itself
35 by its own spinlocks.
36 - Even if softirq is serialized, only local cpu is marked for
37 execution. Hence, we get something sort of weak cpu binding.
38 Though it is still not clear, will it result in better locality
39 or will not.
40
41 Examples:
42 - NET RX softirq. It is multithreaded and does not require
43 any global serialization.
44 - NET TX softirq. It kicks software netdevice queues, hence
45 it is logically serialized per device, but this serialization
46 is invisible to common code.
47 - Tasklets: serialized wrt itself.
48 */
49
50#ifndef __ARCH_IRQ_STAT
51irq_cpustat_t irq_stat[NR_CPUS] ____cacheline_aligned;
52EXPORT_SYMBOL(irq_stat);
53#endif
54
55static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
56
57DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
58
59char *softirq_to_name[NR_SOFTIRQS] = {
60 "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "BLOCK_IOPOLL",
61 "TASKLET", "SCHED", "HRTIMER", "RCU"
62};
63
64/*
65 * we cannot loop indefinitely here to avoid userspace starvation,
66 * but we also don't want to introduce a worst case 1/HZ latency
67 * to the pending events, so lets the scheduler to balance
68 * the softirq load for us.
69 */
70static void wakeup_softirqd(void)
71{
72 /* Interrupts are disabled: no need to stop preemption */
73 struct task_struct *tsk = __this_cpu_read(ksoftirqd);
74
75 if (tsk && tsk->state != TASK_RUNNING)
76 wake_up_process(tsk);
77}
78
79/*
80 * preempt_count and SOFTIRQ_OFFSET usage:
81 * - preempt_count is changed by SOFTIRQ_OFFSET on entering or leaving
82 * softirq processing.
83 * - preempt_count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
84 * on local_bh_disable or local_bh_enable.
85 * This lets us distinguish between whether we are currently processing
86 * softirq and whether we just have bh disabled.
87 */
88
89/*
90 * This one is for softirq.c-internal use,
91 * where hardirqs are disabled legitimately:
92 */
93#ifdef CONFIG_TRACE_IRQFLAGS
94static void __local_bh_disable(unsigned long ip, unsigned int cnt)
95{
96 unsigned long flags;
97
98 WARN_ON_ONCE(in_irq());
99
100 raw_local_irq_save(flags);
101 /*
102 * The preempt tracer hooks into add_preempt_count and will break
103 * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
104 * is set and before current->softirq_enabled is cleared.
105 * We must manually increment preempt_count here and manually
106 * call the trace_preempt_off later.
107 */
108 preempt_count() += cnt;
109 /*
110 * Were softirqs turned off above:
111 */
112 if (softirq_count() == cnt)
113 trace_softirqs_off(ip);
114 raw_local_irq_restore(flags);
115
116 if (preempt_count() == cnt)
117 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
118}
119#else /* !CONFIG_TRACE_IRQFLAGS */
120static inline void __local_bh_disable(unsigned long ip, unsigned int cnt)
121{
122 add_preempt_count(cnt);
123 barrier();
124}
125#endif /* CONFIG_TRACE_IRQFLAGS */
126
127void local_bh_disable(void)
128{
129 __local_bh_disable((unsigned long)__builtin_return_address(0),
130 SOFTIRQ_DISABLE_OFFSET);
131}
132
133EXPORT_SYMBOL(local_bh_disable);
134
135static void __local_bh_enable(unsigned int cnt)
136{
137 WARN_ON_ONCE(in_irq());
138 WARN_ON_ONCE(!irqs_disabled());
139
140 if (softirq_count() == cnt)
141 trace_softirqs_on((unsigned long)__builtin_return_address(0));
142 sub_preempt_count(cnt);
143}
144
145/*
146 * Special-case - softirqs can safely be enabled in
147 * cond_resched_softirq(), or by __do_softirq(),
148 * without processing still-pending softirqs:
149 */
150void _local_bh_enable(void)
151{
152 __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
153}
154
155EXPORT_SYMBOL(_local_bh_enable);
156
157static inline void _local_bh_enable_ip(unsigned long ip)
158{
159 WARN_ON_ONCE(in_irq() || irqs_disabled());
160#ifdef CONFIG_TRACE_IRQFLAGS
161 local_irq_disable();
162#endif
163 /*
164 * Are softirqs going to be turned on now:
165 */
166 if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
167 trace_softirqs_on(ip);
168 /*
169 * Keep preemption disabled until we are done with
170 * softirq processing:
171 */
172 sub_preempt_count(SOFTIRQ_DISABLE_OFFSET - 1);
173
174 if (unlikely(!in_interrupt() && local_softirq_pending()))
175 do_softirq();
176
177 dec_preempt_count();
178#ifdef CONFIG_TRACE_IRQFLAGS
179 local_irq_enable();
180#endif
181 preempt_check_resched();
182}
183
184void local_bh_enable(void)
185{
186 _local_bh_enable_ip((unsigned long)__builtin_return_address(0));
187}
188EXPORT_SYMBOL(local_bh_enable);
189
190void local_bh_enable_ip(unsigned long ip)
191{
192 _local_bh_enable_ip(ip);
193}
194EXPORT_SYMBOL(local_bh_enable_ip);
195
196/*
197 * We restart softirq processing MAX_SOFTIRQ_RESTART times,
198 * and we fall back to softirqd after that.
199 *
200 * This number has been established via experimentation.
201 * The two things to balance is latency against fairness -
202 * we want to handle softirqs as soon as possible, but they
203 * should not be able to lock up the box.
204 */
205#define MAX_SOFTIRQ_RESTART 10
206
207asmlinkage void __do_softirq(void)
208{
209 struct softirq_action *h;
210 __u32 pending;
211 int max_restart = MAX_SOFTIRQ_RESTART;
212 int cpu;
213
214 pending = local_softirq_pending();
215 account_system_vtime(current);
216
217 __local_bh_disable((unsigned long)__builtin_return_address(0),
218 SOFTIRQ_OFFSET);
219 lockdep_softirq_enter();
220
221 cpu = smp_processor_id();
222restart:
223 /* Reset the pending bitmask before enabling irqs */
224 set_softirq_pending(0);
225
226 local_irq_enable();
227
228 h = softirq_vec;
229
230 do {
231 if (pending & 1) {
232 unsigned int vec_nr = h - softirq_vec;
233 int prev_count = preempt_count();
234
235 kstat_incr_softirqs_this_cpu(vec_nr);
236
237 trace_softirq_entry(vec_nr);
238 h->action(h);
239 trace_softirq_exit(vec_nr);
240 if (unlikely(prev_count != preempt_count())) {
241 printk(KERN_ERR "huh, entered softirq %u %s %p"
242 "with preempt_count %08x,"
243 " exited with %08x?\n", vec_nr,
244 softirq_to_name[vec_nr], h->action,
245 prev_count, preempt_count());
246 preempt_count() = prev_count;
247 }
248
249 rcu_bh_qs(cpu);
250 }
251 h++;
252 pending >>= 1;
253 } while (pending);
254
255 local_irq_disable();
256
257 pending = local_softirq_pending();
258 if (pending && --max_restart)
259 goto restart;
260
261 if (pending)
262 wakeup_softirqd();
263
264 lockdep_softirq_exit();
265
266 account_system_vtime(current);
267 __local_bh_enable(SOFTIRQ_OFFSET);
268}
269
270#ifndef __ARCH_HAS_DO_SOFTIRQ
271
272asmlinkage void do_softirq(void)
273{
274 __u32 pending;
275 unsigned long flags;
276
277 if (in_interrupt())
278 return;
279
280 local_irq_save(flags);
281
282 pending = local_softirq_pending();
283
284 if (pending)
285 __do_softirq();
286
287 local_irq_restore(flags);
288}
289
290#endif
291
292/*
293 * Enter an interrupt context.
294 */
295void irq_enter(void)
296{
297 int cpu = smp_processor_id();
298
299 rcu_irq_enter();
300 if (is_idle_task(current) && !in_interrupt()) {
301 /*
302 * Prevent raise_softirq from needlessly waking up ksoftirqd
303 * here, as softirq will be serviced on return from interrupt.
304 */
305 local_bh_disable();
306 tick_check_idle(cpu);
307 _local_bh_enable();
308 }
309
310 __irq_enter();
311}
312
313static inline void invoke_softirq(void)
314{
315 if (!force_irqthreads) {
316#ifdef __ARCH_IRQ_EXIT_IRQS_DISABLED
317 __do_softirq();
318#else
319 do_softirq();
320#endif
321 } else {
322 __local_bh_disable((unsigned long)__builtin_return_address(0),
323 SOFTIRQ_OFFSET);
324 wakeup_softirqd();
325 __local_bh_enable(SOFTIRQ_OFFSET);
326 }
327}
328
329/*
330 * Exit an interrupt context. Process softirqs if needed and possible:
331 */
332void irq_exit(void)
333{
334 account_system_vtime(current);
335 trace_hardirq_exit();
336 sub_preempt_count(IRQ_EXIT_OFFSET);
337 if (!in_interrupt() && local_softirq_pending())
338 invoke_softirq();
339
340#ifdef CONFIG_NO_HZ
341 /* Make sure that timer wheel updates are propagated */
342 if (idle_cpu(smp_processor_id()) && !in_interrupt() && !need_resched())
343 tick_nohz_irq_exit();
344#endif
345 rcu_irq_exit();
346 sched_preempt_enable_no_resched();
347}
348
349/*
350 * This function must run with irqs disabled!
351 */
352inline void raise_softirq_irqoff(unsigned int nr)
353{
354 __raise_softirq_irqoff(nr);
355
356 /*
357 * If we're in an interrupt or softirq, we're done
358 * (this also catches softirq-disabled code). We will
359 * actually run the softirq once we return from
360 * the irq or softirq.
361 *
362 * Otherwise we wake up ksoftirqd to make sure we
363 * schedule the softirq soon.
364 */
365 if (!in_interrupt())
366 wakeup_softirqd();
367}
368
369void raise_softirq(unsigned int nr)
370{
371 unsigned long flags;
372
373 local_irq_save(flags);
374 raise_softirq_irqoff(nr);
375 local_irq_restore(flags);
376}
377
378void __raise_softirq_irqoff(unsigned int nr)
379{
380 trace_softirq_raise(nr);
381 or_softirq_pending(1UL << nr);
382}
383
384void open_softirq(int nr, void (*action)(struct softirq_action *))
385{
386 softirq_vec[nr].action = action;
387}
388
389/*
390 * Tasklets
391 */
392struct tasklet_head
393{
394 struct tasklet_struct *head;
395 struct tasklet_struct **tail;
396};
397
398static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
399static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
400
401void __tasklet_schedule(struct tasklet_struct *t)
402{
403 unsigned long flags;
404
405 local_irq_save(flags);
406 t->next = NULL;
407 *__this_cpu_read(tasklet_vec.tail) = t;
408 __this_cpu_write(tasklet_vec.tail, &(t->next));
409 raise_softirq_irqoff(TASKLET_SOFTIRQ);
410 local_irq_restore(flags);
411}
412
413EXPORT_SYMBOL(__tasklet_schedule);
414
415void __tasklet_hi_schedule(struct tasklet_struct *t)
416{
417 unsigned long flags;
418
419 local_irq_save(flags);
420 t->next = NULL;
421 *__this_cpu_read(tasklet_hi_vec.tail) = t;
422 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
423 raise_softirq_irqoff(HI_SOFTIRQ);
424 local_irq_restore(flags);
425}
426
427EXPORT_SYMBOL(__tasklet_hi_schedule);
428
429void __tasklet_hi_schedule_first(struct tasklet_struct *t)
430{
431 BUG_ON(!irqs_disabled());
432
433 t->next = __this_cpu_read(tasklet_hi_vec.head);
434 __this_cpu_write(tasklet_hi_vec.head, t);
435 __raise_softirq_irqoff(HI_SOFTIRQ);
436}
437
438EXPORT_SYMBOL(__tasklet_hi_schedule_first);
439
440static void tasklet_action(struct softirq_action *a)
441{
442 struct tasklet_struct *list;
443
444 local_irq_disable();
445 list = __this_cpu_read(tasklet_vec.head);
446 __this_cpu_write(tasklet_vec.head, NULL);
447 __this_cpu_write(tasklet_vec.tail, &__get_cpu_var(tasklet_vec).head);
448 local_irq_enable();
449
450 while (list) {
451 struct tasklet_struct *t = list;
452
453 list = list->next;
454
455 if (tasklet_trylock(t)) {
456 if (!atomic_read(&t->count)) {
457 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
458 BUG();
459 t->func(t->data);
460 tasklet_unlock(t);
461 continue;
462 }
463 tasklet_unlock(t);
464 }
465
466 local_irq_disable();
467 t->next = NULL;
468 *__this_cpu_read(tasklet_vec.tail) = t;
469 __this_cpu_write(tasklet_vec.tail, &(t->next));
470 __raise_softirq_irqoff(TASKLET_SOFTIRQ);
471 local_irq_enable();
472 }
473}
474
475static void tasklet_hi_action(struct softirq_action *a)
476{
477 struct tasklet_struct *list;
478
479 local_irq_disable();
480 list = __this_cpu_read(tasklet_hi_vec.head);
481 __this_cpu_write(tasklet_hi_vec.head, NULL);
482 __this_cpu_write(tasklet_hi_vec.tail, &__get_cpu_var(tasklet_hi_vec).head);
483 local_irq_enable();
484
485 while (list) {
486 struct tasklet_struct *t = list;
487
488 list = list->next;
489
490 if (tasklet_trylock(t)) {
491 if (!atomic_read(&t->count)) {
492 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
493 BUG();
494 t->func(t->data);
495 tasklet_unlock(t);
496 continue;
497 }
498 tasklet_unlock(t);
499 }
500
501 local_irq_disable();
502 t->next = NULL;
503 *__this_cpu_read(tasklet_hi_vec.tail) = t;
504 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
505 __raise_softirq_irqoff(HI_SOFTIRQ);
506 local_irq_enable();
507 }
508}
509
510
511void tasklet_init(struct tasklet_struct *t,
512 void (*func)(unsigned long), unsigned long data)
513{
514 t->next = NULL;
515 t->state = 0;
516 atomic_set(&t->count, 0);
517 t->func = func;
518 t->data = data;
519}
520
521EXPORT_SYMBOL(tasklet_init);
522
523void tasklet_kill(struct tasklet_struct *t)
524{
525 if (in_interrupt())
526 printk("Attempt to kill tasklet from interrupt\n");
527
528 while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
529 do {
530 yield();
531 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
532 }
533 tasklet_unlock_wait(t);
534 clear_bit(TASKLET_STATE_SCHED, &t->state);
535}
536
537EXPORT_SYMBOL(tasklet_kill);
538
539/*
540 * tasklet_hrtimer
541 */
542
543/*
544 * The trampoline is called when the hrtimer expires. It schedules a tasklet
545 * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
546 * hrtimer callback, but from softirq context.
547 */
548static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer)
549{
550 struct tasklet_hrtimer *ttimer =
551 container_of(timer, struct tasklet_hrtimer, timer);
552
553 tasklet_hi_schedule(&ttimer->tasklet);
554 return HRTIMER_NORESTART;
555}
556
557/*
558 * Helper function which calls the hrtimer callback from
559 * tasklet/softirq context
560 */
561static void __tasklet_hrtimer_trampoline(unsigned long data)
562{
563 struct tasklet_hrtimer *ttimer = (void *)data;
564 enum hrtimer_restart restart;
565
566 restart = ttimer->function(&ttimer->timer);
567 if (restart != HRTIMER_NORESTART)
568 hrtimer_restart(&ttimer->timer);
569}
570
571/**
572 * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
573 * @ttimer: tasklet_hrtimer which is initialized
574 * @function: hrtimer callback function which gets called from softirq context
575 * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
576 * @mode: hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
577 */
578void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer,
579 enum hrtimer_restart (*function)(struct hrtimer *),
580 clockid_t which_clock, enum hrtimer_mode mode)
581{
582 hrtimer_init(&ttimer->timer, which_clock, mode);
583 ttimer->timer.function = __hrtimer_tasklet_trampoline;
584 tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline,
585 (unsigned long)ttimer);
586 ttimer->function = function;
587}
588EXPORT_SYMBOL_GPL(tasklet_hrtimer_init);
589
590/*
591 * Remote softirq bits
592 */
593
594DEFINE_PER_CPU(struct list_head [NR_SOFTIRQS], softirq_work_list);
595EXPORT_PER_CPU_SYMBOL(softirq_work_list);
596
597static void __local_trigger(struct call_single_data *cp, int softirq)
598{
599 struct list_head *head = &__get_cpu_var(softirq_work_list[softirq]);
600
601 list_add_tail(&cp->list, head);
602
603 /* Trigger the softirq only if the list was previously empty. */
604 if (head->next == &cp->list)
605 raise_softirq_irqoff(softirq);
606}
607
608#ifdef CONFIG_USE_GENERIC_SMP_HELPERS
609static void remote_softirq_receive(void *data)
610{
611 struct call_single_data *cp = data;
612 unsigned long flags;
613 int softirq;
614
615 softirq = cp->priv;
616
617 local_irq_save(flags);
618 __local_trigger(cp, softirq);
619 local_irq_restore(flags);
620}
621
622static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
623{
624 if (cpu_online(cpu)) {
625 cp->func = remote_softirq_receive;
626 cp->info = cp;
627 cp->flags = 0;
628 cp->priv = softirq;
629
630 __smp_call_function_single(cpu, cp, 0);
631 return 0;
632 }
633 return 1;
634}
635#else /* CONFIG_USE_GENERIC_SMP_HELPERS */
636static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
637{
638 return 1;
639}
640#endif
641
642/**
643 * __send_remote_softirq - try to schedule softirq work on a remote cpu
644 * @cp: private SMP call function data area
645 * @cpu: the remote cpu
646 * @this_cpu: the currently executing cpu
647 * @softirq: the softirq for the work
648 *
649 * Attempt to schedule softirq work on a remote cpu. If this cannot be
650 * done, the work is instead queued up on the local cpu.
651 *
652 * Interrupts must be disabled.
653 */
654void __send_remote_softirq(struct call_single_data *cp, int cpu, int this_cpu, int softirq)
655{
656 if (cpu == this_cpu || __try_remote_softirq(cp, cpu, softirq))
657 __local_trigger(cp, softirq);
658}
659EXPORT_SYMBOL(__send_remote_softirq);
660
661/**
662 * send_remote_softirq - try to schedule softirq work on a remote cpu
663 * @cp: private SMP call function data area
664 * @cpu: the remote cpu
665 * @softirq: the softirq for the work
666 *
667 * Like __send_remote_softirq except that disabling interrupts and
668 * computing the current cpu is done for the caller.
669 */
670void send_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
671{
672 unsigned long flags;
673 int this_cpu;
674
675 local_irq_save(flags);
676 this_cpu = smp_processor_id();
677 __send_remote_softirq(cp, cpu, this_cpu, softirq);
678 local_irq_restore(flags);
679}
680EXPORT_SYMBOL(send_remote_softirq);
681
682static int __cpuinit remote_softirq_cpu_notify(struct notifier_block *self,
683 unsigned long action, void *hcpu)
684{
685 /*
686 * If a CPU goes away, splice its entries to the current CPU
687 * and trigger a run of the softirq
688 */
689 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
690 int cpu = (unsigned long) hcpu;
691 int i;
692
693 local_irq_disable();
694 for (i = 0; i < NR_SOFTIRQS; i++) {
695 struct list_head *head = &per_cpu(softirq_work_list[i], cpu);
696 struct list_head *local_head;
697
698 if (list_empty(head))
699 continue;
700
701 local_head = &__get_cpu_var(softirq_work_list[i]);
702 list_splice_init(head, local_head);
703 raise_softirq_irqoff(i);
704 }
705 local_irq_enable();
706 }
707
708 return NOTIFY_OK;
709}
710
711static struct notifier_block __cpuinitdata remote_softirq_cpu_notifier = {
712 .notifier_call = remote_softirq_cpu_notify,
713};
714
715void __init softirq_init(void)
716{
717 int cpu;
718
719 for_each_possible_cpu(cpu) {
720 int i;
721
722 per_cpu(tasklet_vec, cpu).tail =
723 &per_cpu(tasklet_vec, cpu).head;
724 per_cpu(tasklet_hi_vec, cpu).tail =
725 &per_cpu(tasklet_hi_vec, cpu).head;
726 for (i = 0; i < NR_SOFTIRQS; i++)
727 INIT_LIST_HEAD(&per_cpu(softirq_work_list[i], cpu));
728 }
729
730 register_hotcpu_notifier(&remote_softirq_cpu_notifier);
731
732 open_softirq(TASKLET_SOFTIRQ, tasklet_action);
733 open_softirq(HI_SOFTIRQ, tasklet_hi_action);
734}
735
736static int run_ksoftirqd(void * __bind_cpu)
737{
738 set_current_state(TASK_INTERRUPTIBLE);
739
740 while (!kthread_should_stop()) {
741 preempt_disable();
742 if (!local_softirq_pending()) {
743 schedule_preempt_disabled();
744 }
745
746 __set_current_state(TASK_RUNNING);
747
748 while (local_softirq_pending()) {
749 /* Preempt disable stops cpu going offline.
750 If already offline, we'll be on wrong CPU:
751 don't process */
752 if (cpu_is_offline((long)__bind_cpu))
753 goto wait_to_die;
754 local_irq_disable();
755 if (local_softirq_pending())
756 __do_softirq();
757 local_irq_enable();
758 sched_preempt_enable_no_resched();
759 cond_resched();
760 preempt_disable();
761 rcu_note_context_switch((long)__bind_cpu);
762 }
763 preempt_enable();
764 set_current_state(TASK_INTERRUPTIBLE);
765 }
766 __set_current_state(TASK_RUNNING);
767 return 0;
768
769wait_to_die:
770 preempt_enable();
771 /* Wait for kthread_stop */
772 set_current_state(TASK_INTERRUPTIBLE);
773 while (!kthread_should_stop()) {
774 schedule();
775 set_current_state(TASK_INTERRUPTIBLE);
776 }
777 __set_current_state(TASK_RUNNING);
778 return 0;
779}
780
781#ifdef CONFIG_HOTPLUG_CPU
782/*
783 * tasklet_kill_immediate is called to remove a tasklet which can already be
784 * scheduled for execution on @cpu.
785 *
786 * Unlike tasklet_kill, this function removes the tasklet
787 * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
788 *
789 * When this function is called, @cpu must be in the CPU_DEAD state.
790 */
791void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
792{
793 struct tasklet_struct **i;
794
795 BUG_ON(cpu_online(cpu));
796 BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
797
798 if (!test_bit(TASKLET_STATE_SCHED, &t->state))
799 return;
800
801 /* CPU is dead, so no lock needed. */
802 for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
803 if (*i == t) {
804 *i = t->next;
805 /* If this was the tail element, move the tail ptr */
806 if (*i == NULL)
807 per_cpu(tasklet_vec, cpu).tail = i;
808 return;
809 }
810 }
811 BUG();
812}
813
814static void takeover_tasklets(unsigned int cpu)
815{
816 /* CPU is dead, so no lock needed. */
817 local_irq_disable();
818
819 /* Find end, append list for that CPU. */
820 if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
821 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
822 this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
823 per_cpu(tasklet_vec, cpu).head = NULL;
824 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
825 }
826 raise_softirq_irqoff(TASKLET_SOFTIRQ);
827
828 if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
829 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
830 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
831 per_cpu(tasklet_hi_vec, cpu).head = NULL;
832 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
833 }
834 raise_softirq_irqoff(HI_SOFTIRQ);
835
836 local_irq_enable();
837}
838#endif /* CONFIG_HOTPLUG_CPU */
839
840static int __cpuinit cpu_callback(struct notifier_block *nfb,
841 unsigned long action,
842 void *hcpu)
843{
844 int hotcpu = (unsigned long)hcpu;
845 struct task_struct *p;
846
847 switch (action) {
848 case CPU_UP_PREPARE:
849 case CPU_UP_PREPARE_FROZEN:
850 p = kthread_create_on_node(run_ksoftirqd,
851 hcpu,
852 cpu_to_node(hotcpu),
853 "ksoftirqd/%d", hotcpu);
854 if (IS_ERR(p)) {
855 printk("ksoftirqd for %i failed\n", hotcpu);
856 return notifier_from_errno(PTR_ERR(p));
857 }
858 kthread_bind(p, hotcpu);
859 per_cpu(ksoftirqd, hotcpu) = p;
860 break;
861 case CPU_ONLINE:
862 case CPU_ONLINE_FROZEN:
863 wake_up_process(per_cpu(ksoftirqd, hotcpu));
864 break;
865#ifdef CONFIG_HOTPLUG_CPU
866 case CPU_UP_CANCELED:
867 case CPU_UP_CANCELED_FROZEN:
868 if (!per_cpu(ksoftirqd, hotcpu))
869 break;
870 /* Unbind so it can run. Fall thru. */
871 kthread_bind(per_cpu(ksoftirqd, hotcpu),
872 cpumask_any(cpu_online_mask));
873 case CPU_DEAD:
874 case CPU_DEAD_FROZEN: {
875 static const struct sched_param param = {
876 .sched_priority = MAX_RT_PRIO-1
877 };
878
879 p = per_cpu(ksoftirqd, hotcpu);
880 per_cpu(ksoftirqd, hotcpu) = NULL;
881 sched_setscheduler_nocheck(p, SCHED_FIFO, ¶m);
882 kthread_stop(p);
883 takeover_tasklets(hotcpu);
884 break;
885 }
886#endif /* CONFIG_HOTPLUG_CPU */
887 }
888 return NOTIFY_OK;
889}
890
891static struct notifier_block __cpuinitdata cpu_nfb = {
892 .notifier_call = cpu_callback
893};
894
895static __init int spawn_ksoftirqd(void)
896{
897 void *cpu = (void *)(long)smp_processor_id();
898 int err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
899
900 BUG_ON(err != NOTIFY_OK);
901 cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
902 register_cpu_notifier(&cpu_nfb);
903 return 0;
904}
905early_initcall(spawn_ksoftirqd);
906
907/*
908 * [ These __weak aliases are kept in a separate compilation unit, so that
909 * GCC does not inline them incorrectly. ]
910 */
911
912int __init __weak early_irq_init(void)
913{
914 return 0;
915}
916
917#ifdef CONFIG_GENERIC_HARDIRQS
918int __init __weak arch_probe_nr_irqs(void)
919{
920 return NR_IRQS_LEGACY;
921}
922
923int __init __weak arch_early_irq_init(void)
924{
925 return 0;
926}
927#endif
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/kernel/softirq.c
4 *
5 * Copyright (C) 1992 Linus Torvalds
6 *
7 * Rewritten. Old one was good in 2.2, but in 2.3 it was immoral. --ANK (990903)
8 */
9
10#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
12#include <linux/export.h>
13#include <linux/kernel_stat.h>
14#include <linux/interrupt.h>
15#include <linux/init.h>
16#include <linux/local_lock.h>
17#include <linux/mm.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>
29#include <linux/wait_bit.h>
30
31#include <asm/softirq_stack.h>
32
33#define CREATE_TRACE_POINTS
34#include <trace/events/irq.h>
35
36/*
37 - No shared variables, all the data are CPU local.
38 - If a softirq needs serialization, let it serialize itself
39 by its own spinlocks.
40 - Even if softirq is serialized, only local cpu is marked for
41 execution. Hence, we get something sort of weak cpu binding.
42 Though it is still not clear, will it result in better locality
43 or will not.
44
45 Examples:
46 - NET RX softirq. It is multithreaded and does not require
47 any global serialization.
48 - NET TX softirq. It kicks software netdevice queues, hence
49 it is logically serialized per device, but this serialization
50 is invisible to common code.
51 - Tasklets: serialized wrt itself.
52 */
53
54#ifndef __ARCH_IRQ_STAT
55DEFINE_PER_CPU_ALIGNED(irq_cpustat_t, irq_stat);
56EXPORT_PER_CPU_SYMBOL(irq_stat);
57#endif
58
59static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
60
61DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
62
63const char * const softirq_to_name[NR_SOFTIRQS] = {
64 "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "IRQ_POLL",
65 "TASKLET", "SCHED", "HRTIMER", "RCU"
66};
67
68/*
69 * we cannot loop indefinitely here to avoid userspace starvation,
70 * but we also don't want to introduce a worst case 1/HZ latency
71 * to the pending events, so lets the scheduler to balance
72 * the softirq load for us.
73 */
74static void wakeup_softirqd(void)
75{
76 /* Interrupts are disabled: no need to stop preemption */
77 struct task_struct *tsk = __this_cpu_read(ksoftirqd);
78
79 if (tsk)
80 wake_up_process(tsk);
81}
82
83#ifdef CONFIG_TRACE_IRQFLAGS
84DEFINE_PER_CPU(int, hardirqs_enabled);
85DEFINE_PER_CPU(int, hardirq_context);
86EXPORT_PER_CPU_SYMBOL_GPL(hardirqs_enabled);
87EXPORT_PER_CPU_SYMBOL_GPL(hardirq_context);
88#endif
89
90/*
91 * SOFTIRQ_OFFSET usage:
92 *
93 * On !RT kernels 'count' is the preempt counter, on RT kernels this applies
94 * to a per CPU counter and to task::softirqs_disabled_cnt.
95 *
96 * - count is changed by SOFTIRQ_OFFSET on entering or leaving softirq
97 * processing.
98 *
99 * - count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
100 * on local_bh_disable or local_bh_enable.
101 *
102 * This lets us distinguish between whether we are currently processing
103 * softirq and whether we just have bh disabled.
104 */
105#ifdef CONFIG_PREEMPT_RT
106
107/*
108 * RT accounts for BH disabled sections in task::softirqs_disabled_cnt and
109 * also in per CPU softirq_ctrl::cnt. This is necessary to allow tasks in a
110 * softirq disabled section to be preempted.
111 *
112 * The per task counter is used for softirq_count(), in_softirq() and
113 * in_serving_softirqs() because these counts are only valid when the task
114 * holding softirq_ctrl::lock is running.
115 *
116 * The per CPU counter prevents pointless wakeups of ksoftirqd in case that
117 * the task which is in a softirq disabled section is preempted or blocks.
118 */
119struct softirq_ctrl {
120 local_lock_t lock;
121 int cnt;
122};
123
124static DEFINE_PER_CPU(struct softirq_ctrl, softirq_ctrl) = {
125 .lock = INIT_LOCAL_LOCK(softirq_ctrl.lock),
126};
127
128/**
129 * local_bh_blocked() - Check for idle whether BH processing is blocked
130 *
131 * Returns false if the per CPU softirq::cnt is 0 otherwise true.
132 *
133 * This is invoked from the idle task to guard against false positive
134 * softirq pending warnings, which would happen when the task which holds
135 * softirq_ctrl::lock was the only running task on the CPU and blocks on
136 * some other lock.
137 */
138bool local_bh_blocked(void)
139{
140 return __this_cpu_read(softirq_ctrl.cnt) != 0;
141}
142
143void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
144{
145 unsigned long flags;
146 int newcnt;
147
148 WARN_ON_ONCE(in_hardirq());
149
150 /* First entry of a task into a BH disabled section? */
151 if (!current->softirq_disable_cnt) {
152 if (preemptible()) {
153 local_lock(&softirq_ctrl.lock);
154 /* Required to meet the RCU bottomhalf requirements. */
155 rcu_read_lock();
156 } else {
157 DEBUG_LOCKS_WARN_ON(this_cpu_read(softirq_ctrl.cnt));
158 }
159 }
160
161 /*
162 * Track the per CPU softirq disabled state. On RT this is per CPU
163 * state to allow preemption of bottom half disabled sections.
164 */
165 newcnt = __this_cpu_add_return(softirq_ctrl.cnt, cnt);
166 /*
167 * Reflect the result in the task state to prevent recursion on the
168 * local lock and to make softirq_count() & al work.
169 */
170 current->softirq_disable_cnt = newcnt;
171
172 if (IS_ENABLED(CONFIG_TRACE_IRQFLAGS) && newcnt == cnt) {
173 raw_local_irq_save(flags);
174 lockdep_softirqs_off(ip);
175 raw_local_irq_restore(flags);
176 }
177}
178EXPORT_SYMBOL(__local_bh_disable_ip);
179
180static void __local_bh_enable(unsigned int cnt, bool unlock)
181{
182 unsigned long flags;
183 int newcnt;
184
185 DEBUG_LOCKS_WARN_ON(current->softirq_disable_cnt !=
186 this_cpu_read(softirq_ctrl.cnt));
187
188 if (IS_ENABLED(CONFIG_TRACE_IRQFLAGS) && softirq_count() == cnt) {
189 raw_local_irq_save(flags);
190 lockdep_softirqs_on(_RET_IP_);
191 raw_local_irq_restore(flags);
192 }
193
194 newcnt = __this_cpu_sub_return(softirq_ctrl.cnt, cnt);
195 current->softirq_disable_cnt = newcnt;
196
197 if (!newcnt && unlock) {
198 rcu_read_unlock();
199 local_unlock(&softirq_ctrl.lock);
200 }
201}
202
203void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
204{
205 bool preempt_on = preemptible();
206 unsigned long flags;
207 u32 pending;
208 int curcnt;
209
210 WARN_ON_ONCE(in_hardirq());
211 lockdep_assert_irqs_enabled();
212
213 local_irq_save(flags);
214 curcnt = __this_cpu_read(softirq_ctrl.cnt);
215
216 /*
217 * If this is not reenabling soft interrupts, no point in trying to
218 * run pending ones.
219 */
220 if (curcnt != cnt)
221 goto out;
222
223 pending = local_softirq_pending();
224 if (!pending)
225 goto out;
226
227 /*
228 * If this was called from non preemptible context, wake up the
229 * softirq daemon.
230 */
231 if (!preempt_on) {
232 wakeup_softirqd();
233 goto out;
234 }
235
236 /*
237 * Adjust softirq count to SOFTIRQ_OFFSET which makes
238 * in_serving_softirq() become true.
239 */
240 cnt = SOFTIRQ_OFFSET;
241 __local_bh_enable(cnt, false);
242 __do_softirq();
243
244out:
245 __local_bh_enable(cnt, preempt_on);
246 local_irq_restore(flags);
247}
248EXPORT_SYMBOL(__local_bh_enable_ip);
249
250/*
251 * Invoked from ksoftirqd_run() outside of the interrupt disabled section
252 * to acquire the per CPU local lock for reentrancy protection.
253 */
254static inline void ksoftirqd_run_begin(void)
255{
256 __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
257 local_irq_disable();
258}
259
260/* Counterpart to ksoftirqd_run_begin() */
261static inline void ksoftirqd_run_end(void)
262{
263 __local_bh_enable(SOFTIRQ_OFFSET, true);
264 WARN_ON_ONCE(in_interrupt());
265 local_irq_enable();
266}
267
268static inline void softirq_handle_begin(void) { }
269static inline void softirq_handle_end(void) { }
270
271static inline bool should_wake_ksoftirqd(void)
272{
273 return !this_cpu_read(softirq_ctrl.cnt);
274}
275
276static inline void invoke_softirq(void)
277{
278 if (should_wake_ksoftirqd())
279 wakeup_softirqd();
280}
281
282/*
283 * flush_smp_call_function_queue() can raise a soft interrupt in a function
284 * call. On RT kernels this is undesired and the only known functionality
285 * in the block layer which does this is disabled on RT. If soft interrupts
286 * get raised which haven't been raised before the flush, warn so it can be
287 * investigated.
288 */
289void do_softirq_post_smp_call_flush(unsigned int was_pending)
290{
291 if (WARN_ON_ONCE(was_pending != local_softirq_pending()))
292 invoke_softirq();
293}
294
295#else /* CONFIG_PREEMPT_RT */
296
297/*
298 * This one is for softirq.c-internal use, where hardirqs are disabled
299 * legitimately:
300 */
301#ifdef CONFIG_TRACE_IRQFLAGS
302void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
303{
304 unsigned long flags;
305
306 WARN_ON_ONCE(in_hardirq());
307
308 raw_local_irq_save(flags);
309 /*
310 * The preempt tracer hooks into preempt_count_add and will break
311 * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
312 * is set and before current->softirq_enabled is cleared.
313 * We must manually increment preempt_count here and manually
314 * call the trace_preempt_off later.
315 */
316 __preempt_count_add(cnt);
317 /*
318 * Were softirqs turned off above:
319 */
320 if (softirq_count() == (cnt & SOFTIRQ_MASK))
321 lockdep_softirqs_off(ip);
322 raw_local_irq_restore(flags);
323
324 if (preempt_count() == cnt) {
325#ifdef CONFIG_DEBUG_PREEMPT
326 current->preempt_disable_ip = get_lock_parent_ip();
327#endif
328 trace_preempt_off(CALLER_ADDR0, get_lock_parent_ip());
329 }
330}
331EXPORT_SYMBOL(__local_bh_disable_ip);
332#endif /* CONFIG_TRACE_IRQFLAGS */
333
334static void __local_bh_enable(unsigned int cnt)
335{
336 lockdep_assert_irqs_disabled();
337
338 if (preempt_count() == cnt)
339 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
340
341 if (softirq_count() == (cnt & SOFTIRQ_MASK))
342 lockdep_softirqs_on(_RET_IP_);
343
344 __preempt_count_sub(cnt);
345}
346
347/*
348 * Special-case - softirqs can safely be enabled by __do_softirq(),
349 * without processing still-pending softirqs:
350 */
351void _local_bh_enable(void)
352{
353 WARN_ON_ONCE(in_hardirq());
354 __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
355}
356EXPORT_SYMBOL(_local_bh_enable);
357
358void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
359{
360 WARN_ON_ONCE(in_hardirq());
361 lockdep_assert_irqs_enabled();
362#ifdef CONFIG_TRACE_IRQFLAGS
363 local_irq_disable();
364#endif
365 /*
366 * Are softirqs going to be turned on now:
367 */
368 if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
369 lockdep_softirqs_on(ip);
370 /*
371 * Keep preemption disabled until we are done with
372 * softirq processing:
373 */
374 __preempt_count_sub(cnt - 1);
375
376 if (unlikely(!in_interrupt() && local_softirq_pending())) {
377 /*
378 * Run softirq if any pending. And do it in its own stack
379 * as we may be calling this deep in a task call stack already.
380 */
381 do_softirq();
382 }
383
384 preempt_count_dec();
385#ifdef CONFIG_TRACE_IRQFLAGS
386 local_irq_enable();
387#endif
388 preempt_check_resched();
389}
390EXPORT_SYMBOL(__local_bh_enable_ip);
391
392static inline void softirq_handle_begin(void)
393{
394 __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
395}
396
397static inline void softirq_handle_end(void)
398{
399 __local_bh_enable(SOFTIRQ_OFFSET);
400 WARN_ON_ONCE(in_interrupt());
401}
402
403static inline void ksoftirqd_run_begin(void)
404{
405 local_irq_disable();
406}
407
408static inline void ksoftirqd_run_end(void)
409{
410 local_irq_enable();
411}
412
413static inline bool should_wake_ksoftirqd(void)
414{
415 return true;
416}
417
418static inline void invoke_softirq(void)
419{
420 if (!force_irqthreads() || !__this_cpu_read(ksoftirqd)) {
421#ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
422 /*
423 * We can safely execute softirq on the current stack if
424 * it is the irq stack, because it should be near empty
425 * at this stage.
426 */
427 __do_softirq();
428#else
429 /*
430 * Otherwise, irq_exit() is called on the task stack that can
431 * be potentially deep already. So call softirq in its own stack
432 * to prevent from any overrun.
433 */
434 do_softirq_own_stack();
435#endif
436 } else {
437 wakeup_softirqd();
438 }
439}
440
441asmlinkage __visible void do_softirq(void)
442{
443 __u32 pending;
444 unsigned long flags;
445
446 if (in_interrupt())
447 return;
448
449 local_irq_save(flags);
450
451 pending = local_softirq_pending();
452
453 if (pending)
454 do_softirq_own_stack();
455
456 local_irq_restore(flags);
457}
458
459#endif /* !CONFIG_PREEMPT_RT */
460
461/*
462 * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times,
463 * but break the loop if need_resched() is set or after 2 ms.
464 * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in
465 * certain cases, such as stop_machine(), jiffies may cease to
466 * increment and so we need the MAX_SOFTIRQ_RESTART limit as
467 * well to make sure we eventually return from this method.
468 *
469 * These limits have been established via experimentation.
470 * The two things to balance is latency against fairness -
471 * we want to handle softirqs as soon as possible, but they
472 * should not be able to lock up the box.
473 */
474#define MAX_SOFTIRQ_TIME msecs_to_jiffies(2)
475#define MAX_SOFTIRQ_RESTART 10
476
477#ifdef CONFIG_TRACE_IRQFLAGS
478/*
479 * When we run softirqs from irq_exit() and thus on the hardirq stack we need
480 * to keep the lockdep irq context tracking as tight as possible in order to
481 * not miss-qualify lock contexts and miss possible deadlocks.
482 */
483
484static inline bool lockdep_softirq_start(void)
485{
486 bool in_hardirq = false;
487
488 if (lockdep_hardirq_context()) {
489 in_hardirq = true;
490 lockdep_hardirq_exit();
491 }
492
493 lockdep_softirq_enter();
494
495 return in_hardirq;
496}
497
498static inline void lockdep_softirq_end(bool in_hardirq)
499{
500 lockdep_softirq_exit();
501
502 if (in_hardirq)
503 lockdep_hardirq_enter();
504}
505#else
506static inline bool lockdep_softirq_start(void) { return false; }
507static inline void lockdep_softirq_end(bool in_hardirq) { }
508#endif
509
510asmlinkage __visible void __softirq_entry __do_softirq(void)
511{
512 unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
513 unsigned long old_flags = current->flags;
514 int max_restart = MAX_SOFTIRQ_RESTART;
515 struct softirq_action *h;
516 bool in_hardirq;
517 __u32 pending;
518 int softirq_bit;
519
520 /*
521 * Mask out PF_MEMALLOC as the current task context is borrowed for the
522 * softirq. A softirq handled, such as network RX, might set PF_MEMALLOC
523 * again if the socket is related to swapping.
524 */
525 current->flags &= ~PF_MEMALLOC;
526
527 pending = local_softirq_pending();
528
529 softirq_handle_begin();
530 in_hardirq = lockdep_softirq_start();
531 account_softirq_enter(current);
532
533restart:
534 /* Reset the pending bitmask before enabling irqs */
535 set_softirq_pending(0);
536
537 local_irq_enable();
538
539 h = softirq_vec;
540
541 while ((softirq_bit = ffs(pending))) {
542 unsigned int vec_nr;
543 int prev_count;
544
545 h += softirq_bit - 1;
546
547 vec_nr = h - softirq_vec;
548 prev_count = preempt_count();
549
550 kstat_incr_softirqs_this_cpu(vec_nr);
551
552 trace_softirq_entry(vec_nr);
553 h->action(h);
554 trace_softirq_exit(vec_nr);
555 if (unlikely(prev_count != preempt_count())) {
556 pr_err("huh, entered softirq %u %s %p with preempt_count %08x, exited with %08x?\n",
557 vec_nr, softirq_to_name[vec_nr], h->action,
558 prev_count, preempt_count());
559 preempt_count_set(prev_count);
560 }
561 h++;
562 pending >>= softirq_bit;
563 }
564
565 if (!IS_ENABLED(CONFIG_PREEMPT_RT) &&
566 __this_cpu_read(ksoftirqd) == current)
567 rcu_softirq_qs();
568
569 local_irq_disable();
570
571 pending = local_softirq_pending();
572 if (pending) {
573 if (time_before(jiffies, end) && !need_resched() &&
574 --max_restart)
575 goto restart;
576
577 wakeup_softirqd();
578 }
579
580 account_softirq_exit(current);
581 lockdep_softirq_end(in_hardirq);
582 softirq_handle_end();
583 current_restore_flags(old_flags, PF_MEMALLOC);
584}
585
586/**
587 * irq_enter_rcu - Enter an interrupt context with RCU watching
588 */
589void irq_enter_rcu(void)
590{
591 __irq_enter_raw();
592
593 if (tick_nohz_full_cpu(smp_processor_id()) ||
594 (is_idle_task(current) && (irq_count() == HARDIRQ_OFFSET)))
595 tick_irq_enter();
596
597 account_hardirq_enter(current);
598}
599
600/**
601 * irq_enter - Enter an interrupt context including RCU update
602 */
603void irq_enter(void)
604{
605 ct_irq_enter();
606 irq_enter_rcu();
607}
608
609static inline void tick_irq_exit(void)
610{
611#ifdef CONFIG_NO_HZ_COMMON
612 int cpu = smp_processor_id();
613
614 /* Make sure that timer wheel updates are propagated */
615 if ((sched_core_idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
616 if (!in_hardirq())
617 tick_nohz_irq_exit();
618 }
619#endif
620}
621
622static inline void __irq_exit_rcu(void)
623{
624#ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
625 local_irq_disable();
626#else
627 lockdep_assert_irqs_disabled();
628#endif
629 account_hardirq_exit(current);
630 preempt_count_sub(HARDIRQ_OFFSET);
631 if (!in_interrupt() && local_softirq_pending())
632 invoke_softirq();
633
634 tick_irq_exit();
635}
636
637/**
638 * irq_exit_rcu() - Exit an interrupt context without updating RCU
639 *
640 * Also processes softirqs if needed and possible.
641 */
642void irq_exit_rcu(void)
643{
644 __irq_exit_rcu();
645 /* must be last! */
646 lockdep_hardirq_exit();
647}
648
649/**
650 * irq_exit - Exit an interrupt context, update RCU and lockdep
651 *
652 * Also processes softirqs if needed and possible.
653 */
654void irq_exit(void)
655{
656 __irq_exit_rcu();
657 ct_irq_exit();
658 /* must be last! */
659 lockdep_hardirq_exit();
660}
661
662/*
663 * This function must run with irqs disabled!
664 */
665inline void raise_softirq_irqoff(unsigned int nr)
666{
667 __raise_softirq_irqoff(nr);
668
669 /*
670 * If we're in an interrupt or softirq, we're done
671 * (this also catches softirq-disabled code). We will
672 * actually run the softirq once we return from
673 * the irq or softirq.
674 *
675 * Otherwise we wake up ksoftirqd to make sure we
676 * schedule the softirq soon.
677 */
678 if (!in_interrupt() && should_wake_ksoftirqd())
679 wakeup_softirqd();
680}
681
682void raise_softirq(unsigned int nr)
683{
684 unsigned long flags;
685
686 local_irq_save(flags);
687 raise_softirq_irqoff(nr);
688 local_irq_restore(flags);
689}
690
691void __raise_softirq_irqoff(unsigned int nr)
692{
693 lockdep_assert_irqs_disabled();
694 trace_softirq_raise(nr);
695 or_softirq_pending(1UL << nr);
696}
697
698void open_softirq(int nr, void (*action)(struct softirq_action *))
699{
700 softirq_vec[nr].action = action;
701}
702
703/*
704 * Tasklets
705 */
706struct tasklet_head {
707 struct tasklet_struct *head;
708 struct tasklet_struct **tail;
709};
710
711static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
712static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
713
714static void __tasklet_schedule_common(struct tasklet_struct *t,
715 struct tasklet_head __percpu *headp,
716 unsigned int softirq_nr)
717{
718 struct tasklet_head *head;
719 unsigned long flags;
720
721 local_irq_save(flags);
722 head = this_cpu_ptr(headp);
723 t->next = NULL;
724 *head->tail = t;
725 head->tail = &(t->next);
726 raise_softirq_irqoff(softirq_nr);
727 local_irq_restore(flags);
728}
729
730void __tasklet_schedule(struct tasklet_struct *t)
731{
732 __tasklet_schedule_common(t, &tasklet_vec,
733 TASKLET_SOFTIRQ);
734}
735EXPORT_SYMBOL(__tasklet_schedule);
736
737void __tasklet_hi_schedule(struct tasklet_struct *t)
738{
739 __tasklet_schedule_common(t, &tasklet_hi_vec,
740 HI_SOFTIRQ);
741}
742EXPORT_SYMBOL(__tasklet_hi_schedule);
743
744static bool tasklet_clear_sched(struct tasklet_struct *t)
745{
746 if (test_and_clear_bit(TASKLET_STATE_SCHED, &t->state)) {
747 wake_up_var(&t->state);
748 return true;
749 }
750
751 WARN_ONCE(1, "tasklet SCHED state not set: %s %pS\n",
752 t->use_callback ? "callback" : "func",
753 t->use_callback ? (void *)t->callback : (void *)t->func);
754
755 return false;
756}
757
758static void tasklet_action_common(struct softirq_action *a,
759 struct tasklet_head *tl_head,
760 unsigned int softirq_nr)
761{
762 struct tasklet_struct *list;
763
764 local_irq_disable();
765 list = tl_head->head;
766 tl_head->head = NULL;
767 tl_head->tail = &tl_head->head;
768 local_irq_enable();
769
770 while (list) {
771 struct tasklet_struct *t = list;
772
773 list = list->next;
774
775 if (tasklet_trylock(t)) {
776 if (!atomic_read(&t->count)) {
777 if (tasklet_clear_sched(t)) {
778 if (t->use_callback) {
779 trace_tasklet_entry(t, t->callback);
780 t->callback(t);
781 trace_tasklet_exit(t, t->callback);
782 } else {
783 trace_tasklet_entry(t, t->func);
784 t->func(t->data);
785 trace_tasklet_exit(t, t->func);
786 }
787 }
788 tasklet_unlock(t);
789 continue;
790 }
791 tasklet_unlock(t);
792 }
793
794 local_irq_disable();
795 t->next = NULL;
796 *tl_head->tail = t;
797 tl_head->tail = &t->next;
798 __raise_softirq_irqoff(softirq_nr);
799 local_irq_enable();
800 }
801}
802
803static __latent_entropy void tasklet_action(struct softirq_action *a)
804{
805 tasklet_action_common(a, this_cpu_ptr(&tasklet_vec), TASKLET_SOFTIRQ);
806}
807
808static __latent_entropy void tasklet_hi_action(struct softirq_action *a)
809{
810 tasklet_action_common(a, this_cpu_ptr(&tasklet_hi_vec), HI_SOFTIRQ);
811}
812
813void tasklet_setup(struct tasklet_struct *t,
814 void (*callback)(struct tasklet_struct *))
815{
816 t->next = NULL;
817 t->state = 0;
818 atomic_set(&t->count, 0);
819 t->callback = callback;
820 t->use_callback = true;
821 t->data = 0;
822}
823EXPORT_SYMBOL(tasklet_setup);
824
825void tasklet_init(struct tasklet_struct *t,
826 void (*func)(unsigned long), unsigned long data)
827{
828 t->next = NULL;
829 t->state = 0;
830 atomic_set(&t->count, 0);
831 t->func = func;
832 t->use_callback = false;
833 t->data = data;
834}
835EXPORT_SYMBOL(tasklet_init);
836
837#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
838/*
839 * Do not use in new code. Waiting for tasklets from atomic contexts is
840 * error prone and should be avoided.
841 */
842void tasklet_unlock_spin_wait(struct tasklet_struct *t)
843{
844 while (test_bit(TASKLET_STATE_RUN, &(t)->state)) {
845 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
846 /*
847 * Prevent a live lock when current preempted soft
848 * interrupt processing or prevents ksoftirqd from
849 * running. If the tasklet runs on a different CPU
850 * then this has no effect other than doing the BH
851 * disable/enable dance for nothing.
852 */
853 local_bh_disable();
854 local_bh_enable();
855 } else {
856 cpu_relax();
857 }
858 }
859}
860EXPORT_SYMBOL(tasklet_unlock_spin_wait);
861#endif
862
863void tasklet_kill(struct tasklet_struct *t)
864{
865 if (in_interrupt())
866 pr_notice("Attempt to kill tasklet from interrupt\n");
867
868 while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state))
869 wait_var_event(&t->state, !test_bit(TASKLET_STATE_SCHED, &t->state));
870
871 tasklet_unlock_wait(t);
872 tasklet_clear_sched(t);
873}
874EXPORT_SYMBOL(tasklet_kill);
875
876#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
877void tasklet_unlock(struct tasklet_struct *t)
878{
879 smp_mb__before_atomic();
880 clear_bit(TASKLET_STATE_RUN, &t->state);
881 smp_mb__after_atomic();
882 wake_up_var(&t->state);
883}
884EXPORT_SYMBOL_GPL(tasklet_unlock);
885
886void tasklet_unlock_wait(struct tasklet_struct *t)
887{
888 wait_var_event(&t->state, !test_bit(TASKLET_STATE_RUN, &t->state));
889}
890EXPORT_SYMBOL_GPL(tasklet_unlock_wait);
891#endif
892
893void __init softirq_init(void)
894{
895 int cpu;
896
897 for_each_possible_cpu(cpu) {
898 per_cpu(tasklet_vec, cpu).tail =
899 &per_cpu(tasklet_vec, cpu).head;
900 per_cpu(tasklet_hi_vec, cpu).tail =
901 &per_cpu(tasklet_hi_vec, cpu).head;
902 }
903
904 open_softirq(TASKLET_SOFTIRQ, tasklet_action);
905 open_softirq(HI_SOFTIRQ, tasklet_hi_action);
906}
907
908static int ksoftirqd_should_run(unsigned int cpu)
909{
910 return local_softirq_pending();
911}
912
913static void run_ksoftirqd(unsigned int cpu)
914{
915 ksoftirqd_run_begin();
916 if (local_softirq_pending()) {
917 /*
918 * We can safely run softirq on inline stack, as we are not deep
919 * in the task stack here.
920 */
921 __do_softirq();
922 ksoftirqd_run_end();
923 cond_resched();
924 return;
925 }
926 ksoftirqd_run_end();
927}
928
929#ifdef CONFIG_HOTPLUG_CPU
930static int takeover_tasklets(unsigned int cpu)
931{
932 /* CPU is dead, so no lock needed. */
933 local_irq_disable();
934
935 /* Find end, append list for that CPU. */
936 if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
937 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
938 __this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
939 per_cpu(tasklet_vec, cpu).head = NULL;
940 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
941 }
942 raise_softirq_irqoff(TASKLET_SOFTIRQ);
943
944 if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
945 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
946 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
947 per_cpu(tasklet_hi_vec, cpu).head = NULL;
948 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
949 }
950 raise_softirq_irqoff(HI_SOFTIRQ);
951
952 local_irq_enable();
953 return 0;
954}
955#else
956#define takeover_tasklets NULL
957#endif /* CONFIG_HOTPLUG_CPU */
958
959static struct smp_hotplug_thread softirq_threads = {
960 .store = &ksoftirqd,
961 .thread_should_run = ksoftirqd_should_run,
962 .thread_fn = run_ksoftirqd,
963 .thread_comm = "ksoftirqd/%u",
964};
965
966static __init int spawn_ksoftirqd(void)
967{
968 cpuhp_setup_state_nocalls(CPUHP_SOFTIRQ_DEAD, "softirq:dead", NULL,
969 takeover_tasklets);
970 BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
971
972 return 0;
973}
974early_initcall(spawn_ksoftirqd);
975
976/*
977 * [ These __weak aliases are kept in a separate compilation unit, so that
978 * GCC does not inline them incorrectly. ]
979 */
980
981int __init __weak early_irq_init(void)
982{
983 return 0;
984}
985
986int __init __weak arch_probe_nr_irqs(void)
987{
988 return NR_IRQS_LEGACY;
989}
990
991int __init __weak arch_early_irq_init(void)
992{
993 return 0;
994}
995
996unsigned int __weak arch_dynirq_lower_bound(unsigned int from)
997{
998 return from;
999}