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/*
84 * If ksoftirqd is scheduled, we do not want to process pending softirqs
85 * right now. Let ksoftirqd handle this at its own rate, to get fairness,
86 * unless we're doing some of the synchronous softirqs.
87 */
88#define SOFTIRQ_NOW_MASK ((1 << HI_SOFTIRQ) | (1 << TASKLET_SOFTIRQ))
89static bool ksoftirqd_running(unsigned long pending)
90{
91 struct task_struct *tsk = __this_cpu_read(ksoftirqd);
92
93 if (pending & SOFTIRQ_NOW_MASK)
94 return false;
95 return tsk && task_is_running(tsk) && !__kthread_should_park(tsk);
96}
97
98#ifdef CONFIG_TRACE_IRQFLAGS
99DEFINE_PER_CPU(int, hardirqs_enabled);
100DEFINE_PER_CPU(int, hardirq_context);
101EXPORT_PER_CPU_SYMBOL_GPL(hardirqs_enabled);
102EXPORT_PER_CPU_SYMBOL_GPL(hardirq_context);
103#endif
104
105/*
106 * SOFTIRQ_OFFSET usage:
107 *
108 * On !RT kernels 'count' is the preempt counter, on RT kernels this applies
109 * to a per CPU counter and to task::softirqs_disabled_cnt.
110 *
111 * - count is changed by SOFTIRQ_OFFSET on entering or leaving softirq
112 * processing.
113 *
114 * - count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
115 * on local_bh_disable or local_bh_enable.
116 *
117 * This lets us distinguish between whether we are currently processing
118 * softirq and whether we just have bh disabled.
119 */
120#ifdef CONFIG_PREEMPT_RT
121
122/*
123 * RT accounts for BH disabled sections in task::softirqs_disabled_cnt and
124 * also in per CPU softirq_ctrl::cnt. This is necessary to allow tasks in a
125 * softirq disabled section to be preempted.
126 *
127 * The per task counter is used for softirq_count(), in_softirq() and
128 * in_serving_softirqs() because these counts are only valid when the task
129 * holding softirq_ctrl::lock is running.
130 *
131 * The per CPU counter prevents pointless wakeups of ksoftirqd in case that
132 * the task which is in a softirq disabled section is preempted or blocks.
133 */
134struct softirq_ctrl {
135 local_lock_t lock;
136 int cnt;
137};
138
139static DEFINE_PER_CPU(struct softirq_ctrl, softirq_ctrl) = {
140 .lock = INIT_LOCAL_LOCK(softirq_ctrl.lock),
141};
142
143/**
144 * local_bh_blocked() - Check for idle whether BH processing is blocked
145 *
146 * Returns false if the per CPU softirq::cnt is 0 otherwise true.
147 *
148 * This is invoked from the idle task to guard against false positive
149 * softirq pending warnings, which would happen when the task which holds
150 * softirq_ctrl::lock was the only running task on the CPU and blocks on
151 * some other lock.
152 */
153bool local_bh_blocked(void)
154{
155 return __this_cpu_read(softirq_ctrl.cnt) != 0;
156}
157
158void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
159{
160 unsigned long flags;
161 int newcnt;
162
163 WARN_ON_ONCE(in_hardirq());
164
165 /* First entry of a task into a BH disabled section? */
166 if (!current->softirq_disable_cnt) {
167 if (preemptible()) {
168 local_lock(&softirq_ctrl.lock);
169 /* Required to meet the RCU bottomhalf requirements. */
170 rcu_read_lock();
171 } else {
172 DEBUG_LOCKS_WARN_ON(this_cpu_read(softirq_ctrl.cnt));
173 }
174 }
175
176 /*
177 * Track the per CPU softirq disabled state. On RT this is per CPU
178 * state to allow preemption of bottom half disabled sections.
179 */
180 newcnt = __this_cpu_add_return(softirq_ctrl.cnt, cnt);
181 /*
182 * Reflect the result in the task state to prevent recursion on the
183 * local lock and to make softirq_count() & al work.
184 */
185 current->softirq_disable_cnt = newcnt;
186
187 if (IS_ENABLED(CONFIG_TRACE_IRQFLAGS) && newcnt == cnt) {
188 raw_local_irq_save(flags);
189 lockdep_softirqs_off(ip);
190 raw_local_irq_restore(flags);
191 }
192}
193EXPORT_SYMBOL(__local_bh_disable_ip);
194
195static void __local_bh_enable(unsigned int cnt, bool unlock)
196{
197 unsigned long flags;
198 int newcnt;
199
200 DEBUG_LOCKS_WARN_ON(current->softirq_disable_cnt !=
201 this_cpu_read(softirq_ctrl.cnt));
202
203 if (IS_ENABLED(CONFIG_TRACE_IRQFLAGS) && softirq_count() == cnt) {
204 raw_local_irq_save(flags);
205 lockdep_softirqs_on(_RET_IP_);
206 raw_local_irq_restore(flags);
207 }
208
209 newcnt = __this_cpu_sub_return(softirq_ctrl.cnt, cnt);
210 current->softirq_disable_cnt = newcnt;
211
212 if (!newcnt && unlock) {
213 rcu_read_unlock();
214 local_unlock(&softirq_ctrl.lock);
215 }
216}
217
218void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
219{
220 bool preempt_on = preemptible();
221 unsigned long flags;
222 u32 pending;
223 int curcnt;
224
225 WARN_ON_ONCE(in_irq());
226 lockdep_assert_irqs_enabled();
227
228 local_irq_save(flags);
229 curcnt = __this_cpu_read(softirq_ctrl.cnt);
230
231 /*
232 * If this is not reenabling soft interrupts, no point in trying to
233 * run pending ones.
234 */
235 if (curcnt != cnt)
236 goto out;
237
238 pending = local_softirq_pending();
239 if (!pending || ksoftirqd_running(pending))
240 goto out;
241
242 /*
243 * If this was called from non preemptible context, wake up the
244 * softirq daemon.
245 */
246 if (!preempt_on) {
247 wakeup_softirqd();
248 goto out;
249 }
250
251 /*
252 * Adjust softirq count to SOFTIRQ_OFFSET which makes
253 * in_serving_softirq() become true.
254 */
255 cnt = SOFTIRQ_OFFSET;
256 __local_bh_enable(cnt, false);
257 __do_softirq();
258
259out:
260 __local_bh_enable(cnt, preempt_on);
261 local_irq_restore(flags);
262}
263EXPORT_SYMBOL(__local_bh_enable_ip);
264
265/*
266 * Invoked from ksoftirqd_run() outside of the interrupt disabled section
267 * to acquire the per CPU local lock for reentrancy protection.
268 */
269static inline void ksoftirqd_run_begin(void)
270{
271 __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
272 local_irq_disable();
273}
274
275/* Counterpart to ksoftirqd_run_begin() */
276static inline void ksoftirqd_run_end(void)
277{
278 __local_bh_enable(SOFTIRQ_OFFSET, true);
279 WARN_ON_ONCE(in_interrupt());
280 local_irq_enable();
281}
282
283static inline void softirq_handle_begin(void) { }
284static inline void softirq_handle_end(void) { }
285
286static inline bool should_wake_ksoftirqd(void)
287{
288 return !this_cpu_read(softirq_ctrl.cnt);
289}
290
291static inline void invoke_softirq(void)
292{
293 if (should_wake_ksoftirqd())
294 wakeup_softirqd();
295}
296
297#else /* CONFIG_PREEMPT_RT */
298
299/*
300 * This one is for softirq.c-internal use, where hardirqs are disabled
301 * legitimately:
302 */
303#ifdef CONFIG_TRACE_IRQFLAGS
304void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
305{
306 unsigned long flags;
307
308 WARN_ON_ONCE(in_irq());
309
310 raw_local_irq_save(flags);
311 /*
312 * The preempt tracer hooks into preempt_count_add and will break
313 * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
314 * is set and before current->softirq_enabled is cleared.
315 * We must manually increment preempt_count here and manually
316 * call the trace_preempt_off later.
317 */
318 __preempt_count_add(cnt);
319 /*
320 * Were softirqs turned off above:
321 */
322 if (softirq_count() == (cnt & SOFTIRQ_MASK))
323 lockdep_softirqs_off(ip);
324 raw_local_irq_restore(flags);
325
326 if (preempt_count() == cnt) {
327#ifdef CONFIG_DEBUG_PREEMPT
328 current->preempt_disable_ip = get_lock_parent_ip();
329#endif
330 trace_preempt_off(CALLER_ADDR0, get_lock_parent_ip());
331 }
332}
333EXPORT_SYMBOL(__local_bh_disable_ip);
334#endif /* CONFIG_TRACE_IRQFLAGS */
335
336static void __local_bh_enable(unsigned int cnt)
337{
338 lockdep_assert_irqs_disabled();
339
340 if (preempt_count() == cnt)
341 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
342
343 if (softirq_count() == (cnt & SOFTIRQ_MASK))
344 lockdep_softirqs_on(_RET_IP_);
345
346 __preempt_count_sub(cnt);
347}
348
349/*
350 * Special-case - softirqs can safely be enabled by __do_softirq(),
351 * without processing still-pending softirqs:
352 */
353void _local_bh_enable(void)
354{
355 WARN_ON_ONCE(in_irq());
356 __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
357}
358EXPORT_SYMBOL(_local_bh_enable);
359
360void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
361{
362 WARN_ON_ONCE(in_irq());
363 lockdep_assert_irqs_enabled();
364#ifdef CONFIG_TRACE_IRQFLAGS
365 local_irq_disable();
366#endif
367 /*
368 * Are softirqs going to be turned on now:
369 */
370 if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
371 lockdep_softirqs_on(ip);
372 /*
373 * Keep preemption disabled until we are done with
374 * softirq processing:
375 */
376 __preempt_count_sub(cnt - 1);
377
378 if (unlikely(!in_interrupt() && local_softirq_pending())) {
379 /*
380 * Run softirq if any pending. And do it in its own stack
381 * as we may be calling this deep in a task call stack already.
382 */
383 do_softirq();
384 }
385
386 preempt_count_dec();
387#ifdef CONFIG_TRACE_IRQFLAGS
388 local_irq_enable();
389#endif
390 preempt_check_resched();
391}
392EXPORT_SYMBOL(__local_bh_enable_ip);
393
394static inline void softirq_handle_begin(void)
395{
396 __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
397}
398
399static inline void softirq_handle_end(void)
400{
401 __local_bh_enable(SOFTIRQ_OFFSET);
402 WARN_ON_ONCE(in_interrupt());
403}
404
405static inline void ksoftirqd_run_begin(void)
406{
407 local_irq_disable();
408}
409
410static inline void ksoftirqd_run_end(void)
411{
412 local_irq_enable();
413}
414
415static inline bool should_wake_ksoftirqd(void)
416{
417 return true;
418}
419
420static inline void invoke_softirq(void)
421{
422 if (ksoftirqd_running(local_softirq_pending()))
423 return;
424
425 if (!force_irqthreads || !__this_cpu_read(ksoftirqd)) {
426#ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
427 /*
428 * We can safely execute softirq on the current stack if
429 * it is the irq stack, because it should be near empty
430 * at this stage.
431 */
432 __do_softirq();
433#else
434 /*
435 * Otherwise, irq_exit() is called on the task stack that can
436 * be potentially deep already. So call softirq in its own stack
437 * to prevent from any overrun.
438 */
439 do_softirq_own_stack();
440#endif
441 } else {
442 wakeup_softirqd();
443 }
444}
445
446asmlinkage __visible void do_softirq(void)
447{
448 __u32 pending;
449 unsigned long flags;
450
451 if (in_interrupt())
452 return;
453
454 local_irq_save(flags);
455
456 pending = local_softirq_pending();
457
458 if (pending && !ksoftirqd_running(pending))
459 do_softirq_own_stack();
460
461 local_irq_restore(flags);
462}
463
464#endif /* !CONFIG_PREEMPT_RT */
465
466/*
467 * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times,
468 * but break the loop if need_resched() is set or after 2 ms.
469 * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in
470 * certain cases, such as stop_machine(), jiffies may cease to
471 * increment and so we need the MAX_SOFTIRQ_RESTART limit as
472 * well to make sure we eventually return from this method.
473 *
474 * These limits have been established via experimentation.
475 * The two things to balance is latency against fairness -
476 * we want to handle softirqs as soon as possible, but they
477 * should not be able to lock up the box.
478 */
479#define MAX_SOFTIRQ_TIME msecs_to_jiffies(2)
480#define MAX_SOFTIRQ_RESTART 10
481
482#ifdef CONFIG_TRACE_IRQFLAGS
483/*
484 * When we run softirqs from irq_exit() and thus on the hardirq stack we need
485 * to keep the lockdep irq context tracking as tight as possible in order to
486 * not miss-qualify lock contexts and miss possible deadlocks.
487 */
488
489static inline bool lockdep_softirq_start(void)
490{
491 bool in_hardirq = false;
492
493 if (lockdep_hardirq_context()) {
494 in_hardirq = true;
495 lockdep_hardirq_exit();
496 }
497
498 lockdep_softirq_enter();
499
500 return in_hardirq;
501}
502
503static inline void lockdep_softirq_end(bool in_hardirq)
504{
505 lockdep_softirq_exit();
506
507 if (in_hardirq)
508 lockdep_hardirq_enter();
509}
510#else
511static inline bool lockdep_softirq_start(void) { return false; }
512static inline void lockdep_softirq_end(bool in_hardirq) { }
513#endif
514
515asmlinkage __visible void __softirq_entry __do_softirq(void)
516{
517 unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
518 unsigned long old_flags = current->flags;
519 int max_restart = MAX_SOFTIRQ_RESTART;
520 struct softirq_action *h;
521 bool in_hardirq;
522 __u32 pending;
523 int softirq_bit;
524
525 /*
526 * Mask out PF_MEMALLOC as the current task context is borrowed for the
527 * softirq. A softirq handled, such as network RX, might set PF_MEMALLOC
528 * again if the socket is related to swapping.
529 */
530 current->flags &= ~PF_MEMALLOC;
531
532 pending = local_softirq_pending();
533
534 softirq_handle_begin();
535 in_hardirq = lockdep_softirq_start();
536 account_softirq_enter(current);
537
538restart:
539 /* Reset the pending bitmask before enabling irqs */
540 set_softirq_pending(0);
541
542 local_irq_enable();
543
544 h = softirq_vec;
545
546 while ((softirq_bit = ffs(pending))) {
547 unsigned int vec_nr;
548 int prev_count;
549
550 h += softirq_bit - 1;
551
552 vec_nr = h - softirq_vec;
553 prev_count = preempt_count();
554
555 kstat_incr_softirqs_this_cpu(vec_nr);
556
557 trace_softirq_entry(vec_nr);
558 h->action(h);
559 trace_softirq_exit(vec_nr);
560 if (unlikely(prev_count != preempt_count())) {
561 pr_err("huh, entered softirq %u %s %p with preempt_count %08x, exited with %08x?\n",
562 vec_nr, softirq_to_name[vec_nr], h->action,
563 prev_count, preempt_count());
564 preempt_count_set(prev_count);
565 }
566 h++;
567 pending >>= softirq_bit;
568 }
569
570 if (!IS_ENABLED(CONFIG_PREEMPT_RT) &&
571 __this_cpu_read(ksoftirqd) == current)
572 rcu_softirq_qs();
573
574 local_irq_disable();
575
576 pending = local_softirq_pending();
577 if (pending) {
578 if (time_before(jiffies, end) && !need_resched() &&
579 --max_restart)
580 goto restart;
581
582 wakeup_softirqd();
583 }
584
585 account_softirq_exit(current);
586 lockdep_softirq_end(in_hardirq);
587 softirq_handle_end();
588 current_restore_flags(old_flags, PF_MEMALLOC);
589}
590
591/**
592 * irq_enter_rcu - Enter an interrupt context with RCU watching
593 */
594void irq_enter_rcu(void)
595{
596 __irq_enter_raw();
597
598 if (is_idle_task(current) && (irq_count() == HARDIRQ_OFFSET))
599 tick_irq_enter();
600
601 account_hardirq_enter(current);
602}
603
604/**
605 * irq_enter - Enter an interrupt context including RCU update
606 */
607void irq_enter(void)
608{
609 rcu_irq_enter();
610 irq_enter_rcu();
611}
612
613static inline void tick_irq_exit(void)
614{
615#ifdef CONFIG_NO_HZ_COMMON
616 int cpu = smp_processor_id();
617
618 /* Make sure that timer wheel updates are propagated */
619 if ((idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
620 if (!in_irq())
621 tick_nohz_irq_exit();
622 }
623#endif
624}
625
626static inline void __irq_exit_rcu(void)
627{
628#ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
629 local_irq_disable();
630#else
631 lockdep_assert_irqs_disabled();
632#endif
633 account_hardirq_exit(current);
634 preempt_count_sub(HARDIRQ_OFFSET);
635 if (!in_interrupt() && local_softirq_pending())
636 invoke_softirq();
637
638 tick_irq_exit();
639}
640
641/**
642 * irq_exit_rcu() - Exit an interrupt context without updating RCU
643 *
644 * Also processes softirqs if needed and possible.
645 */
646void irq_exit_rcu(void)
647{
648 __irq_exit_rcu();
649 /* must be last! */
650 lockdep_hardirq_exit();
651}
652
653/**
654 * irq_exit - Exit an interrupt context, update RCU and lockdep
655 *
656 * Also processes softirqs if needed and possible.
657 */
658void irq_exit(void)
659{
660 __irq_exit_rcu();
661 rcu_irq_exit();
662 /* must be last! */
663 lockdep_hardirq_exit();
664}
665
666/*
667 * This function must run with irqs disabled!
668 */
669inline void raise_softirq_irqoff(unsigned int nr)
670{
671 __raise_softirq_irqoff(nr);
672
673 /*
674 * If we're in an interrupt or softirq, we're done
675 * (this also catches softirq-disabled code). We will
676 * actually run the softirq once we return from
677 * the irq or softirq.
678 *
679 * Otherwise we wake up ksoftirqd to make sure we
680 * schedule the softirq soon.
681 */
682 if (!in_interrupt() && should_wake_ksoftirqd())
683 wakeup_softirqd();
684}
685
686void raise_softirq(unsigned int nr)
687{
688 unsigned long flags;
689
690 local_irq_save(flags);
691 raise_softirq_irqoff(nr);
692 local_irq_restore(flags);
693}
694
695void __raise_softirq_irqoff(unsigned int nr)
696{
697 lockdep_assert_irqs_disabled();
698 trace_softirq_raise(nr);
699 or_softirq_pending(1UL << nr);
700}
701
702void open_softirq(int nr, void (*action)(struct softirq_action *))
703{
704 softirq_vec[nr].action = action;
705}
706
707/*
708 * Tasklets
709 */
710struct tasklet_head {
711 struct tasklet_struct *head;
712 struct tasklet_struct **tail;
713};
714
715static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
716static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
717
718static void __tasklet_schedule_common(struct tasklet_struct *t,
719 struct tasklet_head __percpu *headp,
720 unsigned int softirq_nr)
721{
722 struct tasklet_head *head;
723 unsigned long flags;
724
725 local_irq_save(flags);
726 head = this_cpu_ptr(headp);
727 t->next = NULL;
728 *head->tail = t;
729 head->tail = &(t->next);
730 raise_softirq_irqoff(softirq_nr);
731 local_irq_restore(flags);
732}
733
734void __tasklet_schedule(struct tasklet_struct *t)
735{
736 __tasklet_schedule_common(t, &tasklet_vec,
737 TASKLET_SOFTIRQ);
738}
739EXPORT_SYMBOL(__tasklet_schedule);
740
741void __tasklet_hi_schedule(struct tasklet_struct *t)
742{
743 __tasklet_schedule_common(t, &tasklet_hi_vec,
744 HI_SOFTIRQ);
745}
746EXPORT_SYMBOL(__tasklet_hi_schedule);
747
748static bool tasklet_clear_sched(struct tasklet_struct *t)
749{
750 if (test_and_clear_bit(TASKLET_STATE_SCHED, &t->state)) {
751 wake_up_var(&t->state);
752 return true;
753 }
754
755 WARN_ONCE(1, "tasklet SCHED state not set: %s %pS\n",
756 t->use_callback ? "callback" : "func",
757 t->use_callback ? (void *)t->callback : (void *)t->func);
758
759 return false;
760}
761
762static void tasklet_action_common(struct softirq_action *a,
763 struct tasklet_head *tl_head,
764 unsigned int softirq_nr)
765{
766 struct tasklet_struct *list;
767
768 local_irq_disable();
769 list = tl_head->head;
770 tl_head->head = NULL;
771 tl_head->tail = &tl_head->head;
772 local_irq_enable();
773
774 while (list) {
775 struct tasklet_struct *t = list;
776
777 list = list->next;
778
779 if (tasklet_trylock(t)) {
780 if (!atomic_read(&t->count)) {
781 if (tasklet_clear_sched(t)) {
782 if (t->use_callback)
783 t->callback(t);
784 else
785 t->func(t->data);
786 }
787 tasklet_unlock(t);
788 continue;
789 }
790 tasklet_unlock(t);
791 }
792
793 local_irq_disable();
794 t->next = NULL;
795 *tl_head->tail = t;
796 tl_head->tail = &t->next;
797 __raise_softirq_irqoff(softirq_nr);
798 local_irq_enable();
799 }
800}
801
802static __latent_entropy void tasklet_action(struct softirq_action *a)
803{
804 tasklet_action_common(a, this_cpu_ptr(&tasklet_vec), TASKLET_SOFTIRQ);
805}
806
807static __latent_entropy void tasklet_hi_action(struct softirq_action *a)
808{
809 tasklet_action_common(a, this_cpu_ptr(&tasklet_hi_vec), HI_SOFTIRQ);
810}
811
812void tasklet_setup(struct tasklet_struct *t,
813 void (*callback)(struct tasklet_struct *))
814{
815 t->next = NULL;
816 t->state = 0;
817 atomic_set(&t->count, 0);
818 t->callback = callback;
819 t->use_callback = true;
820 t->data = 0;
821}
822EXPORT_SYMBOL(tasklet_setup);
823
824void tasklet_init(struct tasklet_struct *t,
825 void (*func)(unsigned long), unsigned long data)
826{
827 t->next = NULL;
828 t->state = 0;
829 atomic_set(&t->count, 0);
830 t->func = func;
831 t->use_callback = false;
832 t->data = data;
833}
834EXPORT_SYMBOL(tasklet_init);
835
836#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
837/*
838 * Do not use in new code. Waiting for tasklets from atomic contexts is
839 * error prone and should be avoided.
840 */
841void tasklet_unlock_spin_wait(struct tasklet_struct *t)
842{
843 while (test_bit(TASKLET_STATE_RUN, &(t)->state)) {
844 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
845 /*
846 * Prevent a live lock when current preempted soft
847 * interrupt processing or prevents ksoftirqd from
848 * running. If the tasklet runs on a different CPU
849 * then this has no effect other than doing the BH
850 * disable/enable dance for nothing.
851 */
852 local_bh_disable();
853 local_bh_enable();
854 } else {
855 cpu_relax();
856 }
857 }
858}
859EXPORT_SYMBOL(tasklet_unlock_spin_wait);
860#endif
861
862void tasklet_kill(struct tasklet_struct *t)
863{
864 if (in_interrupt())
865 pr_notice("Attempt to kill tasklet from interrupt\n");
866
867 while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state))
868 wait_var_event(&t->state, !test_bit(TASKLET_STATE_SCHED, &t->state));
869
870 tasklet_unlock_wait(t);
871 tasklet_clear_sched(t);
872}
873EXPORT_SYMBOL(tasklet_kill);
874
875#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
876void tasklet_unlock(struct tasklet_struct *t)
877{
878 smp_mb__before_atomic();
879 clear_bit(TASKLET_STATE_RUN, &t->state);
880 smp_mb__after_atomic();
881 wake_up_var(&t->state);
882}
883EXPORT_SYMBOL_GPL(tasklet_unlock);
884
885void tasklet_unlock_wait(struct tasklet_struct *t)
886{
887 wait_var_event(&t->state, !test_bit(TASKLET_STATE_RUN, &t->state));
888}
889EXPORT_SYMBOL_GPL(tasklet_unlock_wait);
890#endif
891
892void __init softirq_init(void)
893{
894 int cpu;
895
896 for_each_possible_cpu(cpu) {
897 per_cpu(tasklet_vec, cpu).tail =
898 &per_cpu(tasklet_vec, cpu).head;
899 per_cpu(tasklet_hi_vec, cpu).tail =
900 &per_cpu(tasklet_hi_vec, cpu).head;
901 }
902
903 open_softirq(TASKLET_SOFTIRQ, tasklet_action);
904 open_softirq(HI_SOFTIRQ, tasklet_hi_action);
905}
906
907static int ksoftirqd_should_run(unsigned int cpu)
908{
909 return local_softirq_pending();
910}
911
912static void run_ksoftirqd(unsigned int cpu)
913{
914 ksoftirqd_run_begin();
915 if (local_softirq_pending()) {
916 /*
917 * We can safely run softirq on inline stack, as we are not deep
918 * in the task stack here.
919 */
920 __do_softirq();
921 ksoftirqd_run_end();
922 cond_resched();
923 return;
924 }
925 ksoftirqd_run_end();
926}
927
928#ifdef CONFIG_HOTPLUG_CPU
929static int takeover_tasklets(unsigned int cpu)
930{
931 /* CPU is dead, so no lock needed. */
932 local_irq_disable();
933
934 /* Find end, append list for that CPU. */
935 if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
936 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
937 __this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
938 per_cpu(tasklet_vec, cpu).head = NULL;
939 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
940 }
941 raise_softirq_irqoff(TASKLET_SOFTIRQ);
942
943 if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
944 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
945 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
946 per_cpu(tasklet_hi_vec, cpu).head = NULL;
947 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
948 }
949 raise_softirq_irqoff(HI_SOFTIRQ);
950
951 local_irq_enable();
952 return 0;
953}
954#else
955#define takeover_tasklets NULL
956#endif /* CONFIG_HOTPLUG_CPU */
957
958static struct smp_hotplug_thread softirq_threads = {
959 .store = &ksoftirqd,
960 .thread_should_run = ksoftirqd_should_run,
961 .thread_fn = run_ksoftirqd,
962 .thread_comm = "ksoftirqd/%u",
963};
964
965static __init int spawn_ksoftirqd(void)
966{
967 cpuhp_setup_state_nocalls(CPUHP_SOFTIRQ_DEAD, "softirq:dead", NULL,
968 takeover_tasklets);
969 BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
970
971 return 0;
972}
973early_initcall(spawn_ksoftirqd);
974
975/*
976 * [ These __weak aliases are kept in a separate compilation unit, so that
977 * GCC does not inline them incorrectly. ]
978 */
979
980int __init __weak early_irq_init(void)
981{
982 return 0;
983}
984
985int __init __weak arch_probe_nr_irqs(void)
986{
987 return NR_IRQS_LEGACY;
988}
989
990int __init __weak arch_early_irq_init(void)
991{
992 return 0;
993}
994
995unsigned int __weak arch_dynirq_lower_bound(unsigned int from)
996{
997 return from;
998}