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
11#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
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/smpboot.h>
27#include <linux/tick.h>
28#include <linux/irq.h>
29
30#define CREATE_TRACE_POINTS
31#include <trace/events/irq.h>
32
33/*
34 - No shared variables, all the data are CPU local.
35 - If a softirq needs serialization, let it serialize itself
36 by its own spinlocks.
37 - Even if softirq is serialized, only local cpu is marked for
38 execution. Hence, we get something sort of weak cpu binding.
39 Though it is still not clear, will it result in better locality
40 or will not.
41
42 Examples:
43 - NET RX softirq. It is multithreaded and does not require
44 any global serialization.
45 - NET TX softirq. It kicks software netdevice queues, hence
46 it is logically serialized per device, but this serialization
47 is invisible to common code.
48 - Tasklets: serialized wrt itself.
49 */
50
51#ifndef __ARCH_IRQ_STAT
52irq_cpustat_t irq_stat[NR_CPUS] ____cacheline_aligned;
53EXPORT_SYMBOL(irq_stat);
54#endif
55
56static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
57
58DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
59
60const char * const softirq_to_name[NR_SOFTIRQS] = {
61 "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "BLOCK_IOPOLL",
62 "TASKLET", "SCHED", "HRTIMER", "RCU"
63};
64
65/*
66 * we cannot loop indefinitely here to avoid userspace starvation,
67 * but we also don't want to introduce a worst case 1/HZ latency
68 * to the pending events, so lets the scheduler to balance
69 * the softirq load for us.
70 */
71static void wakeup_softirqd(void)
72{
73 /* Interrupts are disabled: no need to stop preemption */
74 struct task_struct *tsk = __this_cpu_read(ksoftirqd);
75
76 if (tsk && tsk->state != TASK_RUNNING)
77 wake_up_process(tsk);
78}
79
80/*
81 * preempt_count and SOFTIRQ_OFFSET usage:
82 * - preempt_count is changed by SOFTIRQ_OFFSET on entering or leaving
83 * softirq processing.
84 * - preempt_count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
85 * on local_bh_disable or local_bh_enable.
86 * This lets us distinguish between whether we are currently processing
87 * softirq and whether we just have bh disabled.
88 */
89
90/*
91 * This one is for softirq.c-internal use,
92 * where hardirqs are disabled legitimately:
93 */
94#ifdef CONFIG_TRACE_IRQFLAGS
95void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
96{
97 unsigned long flags;
98
99 WARN_ON_ONCE(in_irq());
100
101 raw_local_irq_save(flags);
102 /*
103 * The preempt tracer hooks into preempt_count_add and will break
104 * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
105 * is set and before current->softirq_enabled is cleared.
106 * We must manually increment preempt_count here and manually
107 * call the trace_preempt_off later.
108 */
109 __preempt_count_add(cnt);
110 /*
111 * Were softirqs turned off above:
112 */
113 if (softirq_count() == (cnt & SOFTIRQ_MASK))
114 trace_softirqs_off(ip);
115 raw_local_irq_restore(flags);
116
117 if (preempt_count() == cnt)
118 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
119}
120EXPORT_SYMBOL(__local_bh_disable_ip);
121#endif /* CONFIG_TRACE_IRQFLAGS */
122
123static void __local_bh_enable(unsigned int cnt)
124{
125 WARN_ON_ONCE(!irqs_disabled());
126
127 if (softirq_count() == (cnt & SOFTIRQ_MASK))
128 trace_softirqs_on(_RET_IP_);
129 preempt_count_sub(cnt);
130}
131
132/*
133 * Special-case - softirqs can safely be enabled in
134 * cond_resched_softirq(), or by __do_softirq(),
135 * without processing still-pending softirqs:
136 */
137void _local_bh_enable(void)
138{
139 WARN_ON_ONCE(in_irq());
140 __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
141}
142EXPORT_SYMBOL(_local_bh_enable);
143
144void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
145{
146 WARN_ON_ONCE(in_irq() || irqs_disabled());
147#ifdef CONFIG_TRACE_IRQFLAGS
148 local_irq_disable();
149#endif
150 /*
151 * Are softirqs going to be turned on now:
152 */
153 if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
154 trace_softirqs_on(ip);
155 /*
156 * Keep preemption disabled until we are done with
157 * softirq processing:
158 */
159 preempt_count_sub(cnt - 1);
160
161 if (unlikely(!in_interrupt() && local_softirq_pending())) {
162 /*
163 * Run softirq if any pending. And do it in its own stack
164 * as we may be calling this deep in a task call stack already.
165 */
166 do_softirq();
167 }
168
169 preempt_count_dec();
170#ifdef CONFIG_TRACE_IRQFLAGS
171 local_irq_enable();
172#endif
173 preempt_check_resched();
174}
175EXPORT_SYMBOL(__local_bh_enable_ip);
176
177/*
178 * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times,
179 * but break the loop if need_resched() is set or after 2 ms.
180 * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in
181 * certain cases, such as stop_machine(), jiffies may cease to
182 * increment and so we need the MAX_SOFTIRQ_RESTART limit as
183 * well to make sure we eventually return from this method.
184 *
185 * These limits have been established via experimentation.
186 * The two things to balance is latency against fairness -
187 * we want to handle softirqs as soon as possible, but they
188 * should not be able to lock up the box.
189 */
190#define MAX_SOFTIRQ_TIME msecs_to_jiffies(2)
191#define MAX_SOFTIRQ_RESTART 10
192
193#ifdef CONFIG_TRACE_IRQFLAGS
194/*
195 * When we run softirqs from irq_exit() and thus on the hardirq stack we need
196 * to keep the lockdep irq context tracking as tight as possible in order to
197 * not miss-qualify lock contexts and miss possible deadlocks.
198 */
199
200static inline bool lockdep_softirq_start(void)
201{
202 bool in_hardirq = false;
203
204 if (trace_hardirq_context(current)) {
205 in_hardirq = true;
206 trace_hardirq_exit();
207 }
208
209 lockdep_softirq_enter();
210
211 return in_hardirq;
212}
213
214static inline void lockdep_softirq_end(bool in_hardirq)
215{
216 lockdep_softirq_exit();
217
218 if (in_hardirq)
219 trace_hardirq_enter();
220}
221#else
222static inline bool lockdep_softirq_start(void) { return false; }
223static inline void lockdep_softirq_end(bool in_hardirq) { }
224#endif
225
226asmlinkage __visible void __do_softirq(void)
227{
228 unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
229 unsigned long old_flags = current->flags;
230 int max_restart = MAX_SOFTIRQ_RESTART;
231 struct softirq_action *h;
232 bool in_hardirq;
233 __u32 pending;
234 int softirq_bit;
235 int cpu;
236
237 /*
238 * Mask out PF_MEMALLOC s current task context is borrowed for the
239 * softirq. A softirq handled such as network RX might set PF_MEMALLOC
240 * again if the socket is related to swap
241 */
242 current->flags &= ~PF_MEMALLOC;
243
244 pending = local_softirq_pending();
245 account_irq_enter_time(current);
246
247 __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
248 in_hardirq = lockdep_softirq_start();
249
250 cpu = smp_processor_id();
251restart:
252 /* Reset the pending bitmask before enabling irqs */
253 set_softirq_pending(0);
254
255 local_irq_enable();
256
257 h = softirq_vec;
258
259 while ((softirq_bit = ffs(pending))) {
260 unsigned int vec_nr;
261 int prev_count;
262
263 h += softirq_bit - 1;
264
265 vec_nr = h - softirq_vec;
266 prev_count = preempt_count();
267
268 kstat_incr_softirqs_this_cpu(vec_nr);
269
270 trace_softirq_entry(vec_nr);
271 h->action(h);
272 trace_softirq_exit(vec_nr);
273 if (unlikely(prev_count != preempt_count())) {
274 pr_err("huh, entered softirq %u %s %p with preempt_count %08x, exited with %08x?\n",
275 vec_nr, softirq_to_name[vec_nr], h->action,
276 prev_count, preempt_count());
277 preempt_count_set(prev_count);
278 }
279 rcu_bh_qs(cpu);
280 h++;
281 pending >>= softirq_bit;
282 }
283
284 local_irq_disable();
285
286 pending = local_softirq_pending();
287 if (pending) {
288 if (time_before(jiffies, end) && !need_resched() &&
289 --max_restart)
290 goto restart;
291
292 wakeup_softirqd();
293 }
294
295 lockdep_softirq_end(in_hardirq);
296 account_irq_exit_time(current);
297 __local_bh_enable(SOFTIRQ_OFFSET);
298 WARN_ON_ONCE(in_interrupt());
299 tsk_restore_flags(current, old_flags, PF_MEMALLOC);
300}
301
302asmlinkage __visible void do_softirq(void)
303{
304 __u32 pending;
305 unsigned long flags;
306
307 if (in_interrupt())
308 return;
309
310 local_irq_save(flags);
311
312 pending = local_softirq_pending();
313
314 if (pending)
315 do_softirq_own_stack();
316
317 local_irq_restore(flags);
318}
319
320/*
321 * Enter an interrupt context.
322 */
323void irq_enter(void)
324{
325 rcu_irq_enter();
326 if (is_idle_task(current) && !in_interrupt()) {
327 /*
328 * Prevent raise_softirq from needlessly waking up ksoftirqd
329 * here, as softirq will be serviced on return from interrupt.
330 */
331 local_bh_disable();
332 tick_irq_enter();
333 _local_bh_enable();
334 }
335
336 __irq_enter();
337}
338
339static inline void invoke_softirq(void)
340{
341 if (!force_irqthreads) {
342#ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
343 /*
344 * We can safely execute softirq on the current stack if
345 * it is the irq stack, because it should be near empty
346 * at this stage.
347 */
348 __do_softirq();
349#else
350 /*
351 * Otherwise, irq_exit() is called on the task stack that can
352 * be potentially deep already. So call softirq in its own stack
353 * to prevent from any overrun.
354 */
355 do_softirq_own_stack();
356#endif
357 } else {
358 wakeup_softirqd();
359 }
360}
361
362static inline void tick_irq_exit(void)
363{
364#ifdef CONFIG_NO_HZ_COMMON
365 int cpu = smp_processor_id();
366
367 /* Make sure that timer wheel updates are propagated */
368 if ((idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
369 if (!in_interrupt())
370 tick_nohz_irq_exit();
371 }
372#endif
373}
374
375/*
376 * Exit an interrupt context. Process softirqs if needed and possible:
377 */
378void irq_exit(void)
379{
380#ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
381 local_irq_disable();
382#else
383 WARN_ON_ONCE(!irqs_disabled());
384#endif
385
386 account_irq_exit_time(current);
387 preempt_count_sub(HARDIRQ_OFFSET);
388 if (!in_interrupt() && local_softirq_pending())
389 invoke_softirq();
390
391 tick_irq_exit();
392 rcu_irq_exit();
393 trace_hardirq_exit(); /* must be last! */
394}
395
396/*
397 * This function must run with irqs disabled!
398 */
399inline void raise_softirq_irqoff(unsigned int nr)
400{
401 __raise_softirq_irqoff(nr);
402
403 /*
404 * If we're in an interrupt or softirq, we're done
405 * (this also catches softirq-disabled code). We will
406 * actually run the softirq once we return from
407 * the irq or softirq.
408 *
409 * Otherwise we wake up ksoftirqd to make sure we
410 * schedule the softirq soon.
411 */
412 if (!in_interrupt())
413 wakeup_softirqd();
414}
415
416void raise_softirq(unsigned int nr)
417{
418 unsigned long flags;
419
420 local_irq_save(flags);
421 raise_softirq_irqoff(nr);
422 local_irq_restore(flags);
423}
424
425void __raise_softirq_irqoff(unsigned int nr)
426{
427 trace_softirq_raise(nr);
428 or_softirq_pending(1UL << nr);
429}
430
431void open_softirq(int nr, void (*action)(struct softirq_action *))
432{
433 softirq_vec[nr].action = action;
434}
435
436/*
437 * Tasklets
438 */
439struct tasklet_head {
440 struct tasklet_struct *head;
441 struct tasklet_struct **tail;
442};
443
444static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
445static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
446
447void __tasklet_schedule(struct tasklet_struct *t)
448{
449 unsigned long flags;
450
451 local_irq_save(flags);
452 t->next = NULL;
453 *__this_cpu_read(tasklet_vec.tail) = t;
454 __this_cpu_write(tasklet_vec.tail, &(t->next));
455 raise_softirq_irqoff(TASKLET_SOFTIRQ);
456 local_irq_restore(flags);
457}
458EXPORT_SYMBOL(__tasklet_schedule);
459
460void __tasklet_hi_schedule(struct tasklet_struct *t)
461{
462 unsigned long flags;
463
464 local_irq_save(flags);
465 t->next = NULL;
466 *__this_cpu_read(tasklet_hi_vec.tail) = t;
467 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
468 raise_softirq_irqoff(HI_SOFTIRQ);
469 local_irq_restore(flags);
470}
471EXPORT_SYMBOL(__tasklet_hi_schedule);
472
473void __tasklet_hi_schedule_first(struct tasklet_struct *t)
474{
475 BUG_ON(!irqs_disabled());
476
477 t->next = __this_cpu_read(tasklet_hi_vec.head);
478 __this_cpu_write(tasklet_hi_vec.head, t);
479 __raise_softirq_irqoff(HI_SOFTIRQ);
480}
481EXPORT_SYMBOL(__tasklet_hi_schedule_first);
482
483static void tasklet_action(struct softirq_action *a)
484{
485 struct tasklet_struct *list;
486
487 local_irq_disable();
488 list = __this_cpu_read(tasklet_vec.head);
489 __this_cpu_write(tasklet_vec.head, NULL);
490 __this_cpu_write(tasklet_vec.tail, &__get_cpu_var(tasklet_vec).head);
491 local_irq_enable();
492
493 while (list) {
494 struct tasklet_struct *t = list;
495
496 list = list->next;
497
498 if (tasklet_trylock(t)) {
499 if (!atomic_read(&t->count)) {
500 if (!test_and_clear_bit(TASKLET_STATE_SCHED,
501 &t->state))
502 BUG();
503 t->func(t->data);
504 tasklet_unlock(t);
505 continue;
506 }
507 tasklet_unlock(t);
508 }
509
510 local_irq_disable();
511 t->next = NULL;
512 *__this_cpu_read(tasklet_vec.tail) = t;
513 __this_cpu_write(tasklet_vec.tail, &(t->next));
514 __raise_softirq_irqoff(TASKLET_SOFTIRQ);
515 local_irq_enable();
516 }
517}
518
519static void tasklet_hi_action(struct softirq_action *a)
520{
521 struct tasklet_struct *list;
522
523 local_irq_disable();
524 list = __this_cpu_read(tasklet_hi_vec.head);
525 __this_cpu_write(tasklet_hi_vec.head, NULL);
526 __this_cpu_write(tasklet_hi_vec.tail, &__get_cpu_var(tasklet_hi_vec).head);
527 local_irq_enable();
528
529 while (list) {
530 struct tasklet_struct *t = list;
531
532 list = list->next;
533
534 if (tasklet_trylock(t)) {
535 if (!atomic_read(&t->count)) {
536 if (!test_and_clear_bit(TASKLET_STATE_SCHED,
537 &t->state))
538 BUG();
539 t->func(t->data);
540 tasklet_unlock(t);
541 continue;
542 }
543 tasklet_unlock(t);
544 }
545
546 local_irq_disable();
547 t->next = NULL;
548 *__this_cpu_read(tasklet_hi_vec.tail) = t;
549 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
550 __raise_softirq_irqoff(HI_SOFTIRQ);
551 local_irq_enable();
552 }
553}
554
555void tasklet_init(struct tasklet_struct *t,
556 void (*func)(unsigned long), unsigned long data)
557{
558 t->next = NULL;
559 t->state = 0;
560 atomic_set(&t->count, 0);
561 t->func = func;
562 t->data = data;
563}
564EXPORT_SYMBOL(tasklet_init);
565
566void tasklet_kill(struct tasklet_struct *t)
567{
568 if (in_interrupt())
569 pr_notice("Attempt to kill tasklet from interrupt\n");
570
571 while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
572 do {
573 yield();
574 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
575 }
576 tasklet_unlock_wait(t);
577 clear_bit(TASKLET_STATE_SCHED, &t->state);
578}
579EXPORT_SYMBOL(tasklet_kill);
580
581/*
582 * tasklet_hrtimer
583 */
584
585/*
586 * The trampoline is called when the hrtimer expires. It schedules a tasklet
587 * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
588 * hrtimer callback, but from softirq context.
589 */
590static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer)
591{
592 struct tasklet_hrtimer *ttimer =
593 container_of(timer, struct tasklet_hrtimer, timer);
594
595 tasklet_hi_schedule(&ttimer->tasklet);
596 return HRTIMER_NORESTART;
597}
598
599/*
600 * Helper function which calls the hrtimer callback from
601 * tasklet/softirq context
602 */
603static void __tasklet_hrtimer_trampoline(unsigned long data)
604{
605 struct tasklet_hrtimer *ttimer = (void *)data;
606 enum hrtimer_restart restart;
607
608 restart = ttimer->function(&ttimer->timer);
609 if (restart != HRTIMER_NORESTART)
610 hrtimer_restart(&ttimer->timer);
611}
612
613/**
614 * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
615 * @ttimer: tasklet_hrtimer which is initialized
616 * @function: hrtimer callback function which gets called from softirq context
617 * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
618 * @mode: hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
619 */
620void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer,
621 enum hrtimer_restart (*function)(struct hrtimer *),
622 clockid_t which_clock, enum hrtimer_mode mode)
623{
624 hrtimer_init(&ttimer->timer, which_clock, mode);
625 ttimer->timer.function = __hrtimer_tasklet_trampoline;
626 tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline,
627 (unsigned long)ttimer);
628 ttimer->function = function;
629}
630EXPORT_SYMBOL_GPL(tasklet_hrtimer_init);
631
632void __init softirq_init(void)
633{
634 int cpu;
635
636 for_each_possible_cpu(cpu) {
637 per_cpu(tasklet_vec, cpu).tail =
638 &per_cpu(tasklet_vec, cpu).head;
639 per_cpu(tasklet_hi_vec, cpu).tail =
640 &per_cpu(tasklet_hi_vec, cpu).head;
641 }
642
643 open_softirq(TASKLET_SOFTIRQ, tasklet_action);
644 open_softirq(HI_SOFTIRQ, tasklet_hi_action);
645}
646
647static int ksoftirqd_should_run(unsigned int cpu)
648{
649 return local_softirq_pending();
650}
651
652static void run_ksoftirqd(unsigned int cpu)
653{
654 local_irq_disable();
655 if (local_softirq_pending()) {
656 /*
657 * We can safely run softirq on inline stack, as we are not deep
658 * in the task stack here.
659 */
660 __do_softirq();
661 rcu_note_context_switch(cpu);
662 local_irq_enable();
663 cond_resched();
664 return;
665 }
666 local_irq_enable();
667}
668
669#ifdef CONFIG_HOTPLUG_CPU
670/*
671 * tasklet_kill_immediate is called to remove a tasklet which can already be
672 * scheduled for execution on @cpu.
673 *
674 * Unlike tasklet_kill, this function removes the tasklet
675 * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
676 *
677 * When this function is called, @cpu must be in the CPU_DEAD state.
678 */
679void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
680{
681 struct tasklet_struct **i;
682
683 BUG_ON(cpu_online(cpu));
684 BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
685
686 if (!test_bit(TASKLET_STATE_SCHED, &t->state))
687 return;
688
689 /* CPU is dead, so no lock needed. */
690 for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
691 if (*i == t) {
692 *i = t->next;
693 /* If this was the tail element, move the tail ptr */
694 if (*i == NULL)
695 per_cpu(tasklet_vec, cpu).tail = i;
696 return;
697 }
698 }
699 BUG();
700}
701
702static void takeover_tasklets(unsigned int cpu)
703{
704 /* CPU is dead, so no lock needed. */
705 local_irq_disable();
706
707 /* Find end, append list for that CPU. */
708 if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
709 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
710 this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
711 per_cpu(tasklet_vec, cpu).head = NULL;
712 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
713 }
714 raise_softirq_irqoff(TASKLET_SOFTIRQ);
715
716 if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
717 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
718 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
719 per_cpu(tasklet_hi_vec, cpu).head = NULL;
720 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
721 }
722 raise_softirq_irqoff(HI_SOFTIRQ);
723
724 local_irq_enable();
725}
726#endif /* CONFIG_HOTPLUG_CPU */
727
728static int cpu_callback(struct notifier_block *nfb, unsigned long action,
729 void *hcpu)
730{
731 switch (action) {
732#ifdef CONFIG_HOTPLUG_CPU
733 case CPU_DEAD:
734 case CPU_DEAD_FROZEN:
735 takeover_tasklets((unsigned long)hcpu);
736 break;
737#endif /* CONFIG_HOTPLUG_CPU */
738 }
739 return NOTIFY_OK;
740}
741
742static struct notifier_block cpu_nfb = {
743 .notifier_call = cpu_callback
744};
745
746static struct smp_hotplug_thread softirq_threads = {
747 .store = &ksoftirqd,
748 .thread_should_run = ksoftirqd_should_run,
749 .thread_fn = run_ksoftirqd,
750 .thread_comm = "ksoftirqd/%u",
751};
752
753static __init int spawn_ksoftirqd(void)
754{
755 register_cpu_notifier(&cpu_nfb);
756
757 BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
758
759 return 0;
760}
761early_initcall(spawn_ksoftirqd);
762
763/*
764 * [ These __weak aliases are kept in a separate compilation unit, so that
765 * GCC does not inline them incorrectly. ]
766 */
767
768int __init __weak early_irq_init(void)
769{
770 return 0;
771}
772
773int __init __weak arch_probe_nr_irqs(void)
774{
775 return NR_IRQS_LEGACY;
776}
777
778int __init __weak arch_early_irq_init(void)
779{
780 return 0;
781}
782
783unsigned int __weak arch_dynirq_lower_bound(unsigned int from)
784{
785 return from;
786}
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#include <linux/workqueue.h>
31
32#include <asm/softirq_stack.h>
33
34#define CREATE_TRACE_POINTS
35#include <trace/events/irq.h>
36
37/*
38 - No shared variables, all the data are CPU local.
39 - If a softirq needs serialization, let it serialize itself
40 by its own spinlocks.
41 - Even if softirq is serialized, only local cpu is marked for
42 execution. Hence, we get something sort of weak cpu binding.
43 Though it is still not clear, will it result in better locality
44 or will not.
45
46 Examples:
47 - NET RX softirq. It is multithreaded and does not require
48 any global serialization.
49 - NET TX softirq. It kicks software netdevice queues, hence
50 it is logically serialized per device, but this serialization
51 is invisible to common code.
52 - Tasklets: serialized wrt itself.
53 */
54
55#ifndef __ARCH_IRQ_STAT
56DEFINE_PER_CPU_ALIGNED(irq_cpustat_t, irq_stat);
57EXPORT_PER_CPU_SYMBOL(irq_stat);
58#endif
59
60static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
61
62DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
63
64const char * const softirq_to_name[NR_SOFTIRQS] = {
65 "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "IRQ_POLL",
66 "TASKLET", "SCHED", "HRTIMER", "RCU"
67};
68
69/*
70 * we cannot loop indefinitely here to avoid userspace starvation,
71 * but we also don't want to introduce a worst case 1/HZ latency
72 * to the pending events, so lets the scheduler to balance
73 * the softirq load for us.
74 */
75static void wakeup_softirqd(void)
76{
77 /* Interrupts are disabled: no need to stop preemption */
78 struct task_struct *tsk = __this_cpu_read(ksoftirqd);
79
80 if (tsk)
81 wake_up_process(tsk);
82}
83
84#ifdef CONFIG_TRACE_IRQFLAGS
85DEFINE_PER_CPU(int, hardirqs_enabled);
86DEFINE_PER_CPU(int, hardirq_context);
87EXPORT_PER_CPU_SYMBOL_GPL(hardirqs_enabled);
88EXPORT_PER_CPU_SYMBOL_GPL(hardirq_context);
89#endif
90
91/*
92 * SOFTIRQ_OFFSET usage:
93 *
94 * On !RT kernels 'count' is the preempt counter, on RT kernels this applies
95 * to a per CPU counter and to task::softirqs_disabled_cnt.
96 *
97 * - count is changed by SOFTIRQ_OFFSET on entering or leaving softirq
98 * processing.
99 *
100 * - count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
101 * on local_bh_disable or local_bh_enable.
102 *
103 * This lets us distinguish between whether we are currently processing
104 * softirq and whether we just have bh disabled.
105 */
106#ifdef CONFIG_PREEMPT_RT
107
108/*
109 * RT accounts for BH disabled sections in task::softirqs_disabled_cnt and
110 * also in per CPU softirq_ctrl::cnt. This is necessary to allow tasks in a
111 * softirq disabled section to be preempted.
112 *
113 * The per task counter is used for softirq_count(), in_softirq() and
114 * in_serving_softirqs() because these counts are only valid when the task
115 * holding softirq_ctrl::lock is running.
116 *
117 * The per CPU counter prevents pointless wakeups of ksoftirqd in case that
118 * the task which is in a softirq disabled section is preempted or blocks.
119 */
120struct softirq_ctrl {
121 local_lock_t lock;
122 int cnt;
123};
124
125static DEFINE_PER_CPU(struct softirq_ctrl, softirq_ctrl) = {
126 .lock = INIT_LOCAL_LOCK(softirq_ctrl.lock),
127};
128
129/**
130 * local_bh_blocked() - Check for idle whether BH processing is blocked
131 *
132 * Returns false if the per CPU softirq::cnt is 0 otherwise true.
133 *
134 * This is invoked from the idle task to guard against false positive
135 * softirq pending warnings, which would happen when the task which holds
136 * softirq_ctrl::lock was the only running task on the CPU and blocks on
137 * some other lock.
138 */
139bool local_bh_blocked(void)
140{
141 return __this_cpu_read(softirq_ctrl.cnt) != 0;
142}
143
144void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
145{
146 unsigned long flags;
147 int newcnt;
148
149 WARN_ON_ONCE(in_hardirq());
150
151 /* First entry of a task into a BH disabled section? */
152 if (!current->softirq_disable_cnt) {
153 if (preemptible()) {
154 local_lock(&softirq_ctrl.lock);
155 /* Required to meet the RCU bottomhalf requirements. */
156 rcu_read_lock();
157 } else {
158 DEBUG_LOCKS_WARN_ON(this_cpu_read(softirq_ctrl.cnt));
159 }
160 }
161
162 /*
163 * Track the per CPU softirq disabled state. On RT this is per CPU
164 * state to allow preemption of bottom half disabled sections.
165 */
166 newcnt = __this_cpu_add_return(softirq_ctrl.cnt, cnt);
167 /*
168 * Reflect the result in the task state to prevent recursion on the
169 * local lock and to make softirq_count() & al work.
170 */
171 current->softirq_disable_cnt = newcnt;
172
173 if (IS_ENABLED(CONFIG_TRACE_IRQFLAGS) && newcnt == cnt) {
174 raw_local_irq_save(flags);
175 lockdep_softirqs_off(ip);
176 raw_local_irq_restore(flags);
177 }
178}
179EXPORT_SYMBOL(__local_bh_disable_ip);
180
181static void __local_bh_enable(unsigned int cnt, bool unlock)
182{
183 unsigned long flags;
184 int newcnt;
185
186 DEBUG_LOCKS_WARN_ON(current->softirq_disable_cnt !=
187 this_cpu_read(softirq_ctrl.cnt));
188
189 if (IS_ENABLED(CONFIG_TRACE_IRQFLAGS) && softirq_count() == cnt) {
190 raw_local_irq_save(flags);
191 lockdep_softirqs_on(_RET_IP_);
192 raw_local_irq_restore(flags);
193 }
194
195 newcnt = __this_cpu_sub_return(softirq_ctrl.cnt, cnt);
196 current->softirq_disable_cnt = newcnt;
197
198 if (!newcnt && unlock) {
199 rcu_read_unlock();
200 local_unlock(&softirq_ctrl.lock);
201 }
202}
203
204void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
205{
206 bool preempt_on = preemptible();
207 unsigned long flags;
208 u32 pending;
209 int curcnt;
210
211 WARN_ON_ONCE(in_hardirq());
212 lockdep_assert_irqs_enabled();
213
214 local_irq_save(flags);
215 curcnt = __this_cpu_read(softirq_ctrl.cnt);
216
217 /*
218 * If this is not reenabling soft interrupts, no point in trying to
219 * run pending ones.
220 */
221 if (curcnt != cnt)
222 goto out;
223
224 pending = local_softirq_pending();
225 if (!pending)
226 goto out;
227
228 /*
229 * If this was called from non preemptible context, wake up the
230 * softirq daemon.
231 */
232 if (!preempt_on) {
233 wakeup_softirqd();
234 goto out;
235 }
236
237 /*
238 * Adjust softirq count to SOFTIRQ_OFFSET which makes
239 * in_serving_softirq() become true.
240 */
241 cnt = SOFTIRQ_OFFSET;
242 __local_bh_enable(cnt, false);
243 __do_softirq();
244
245out:
246 __local_bh_enable(cnt, preempt_on);
247 local_irq_restore(flags);
248}
249EXPORT_SYMBOL(__local_bh_enable_ip);
250
251/*
252 * Invoked from ksoftirqd_run() outside of the interrupt disabled section
253 * to acquire the per CPU local lock for reentrancy protection.
254 */
255static inline void ksoftirqd_run_begin(void)
256{
257 __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
258 local_irq_disable();
259}
260
261/* Counterpart to ksoftirqd_run_begin() */
262static inline void ksoftirqd_run_end(void)
263{
264 __local_bh_enable(SOFTIRQ_OFFSET, true);
265 WARN_ON_ONCE(in_interrupt());
266 local_irq_enable();
267}
268
269static inline void softirq_handle_begin(void) { }
270static inline void softirq_handle_end(void) { }
271
272static inline bool should_wake_ksoftirqd(void)
273{
274 return !this_cpu_read(softirq_ctrl.cnt);
275}
276
277static inline void invoke_softirq(void)
278{
279 if (should_wake_ksoftirqd())
280 wakeup_softirqd();
281}
282
283#define SCHED_SOFTIRQ_MASK BIT(SCHED_SOFTIRQ)
284
285/*
286 * flush_smp_call_function_queue() can raise a soft interrupt in a function
287 * call. On RT kernels this is undesired and the only known functionalities
288 * are in the block layer which is disabled on RT, and in the scheduler for
289 * idle load balancing. If soft interrupts get raised which haven't been
290 * raised before the flush, warn if it is not a SCHED_SOFTIRQ so it can be
291 * investigated.
292 */
293void do_softirq_post_smp_call_flush(unsigned int was_pending)
294{
295 unsigned int is_pending = local_softirq_pending();
296
297 if (unlikely(was_pending != is_pending)) {
298 WARN_ON_ONCE(was_pending != (is_pending & ~SCHED_SOFTIRQ_MASK));
299 invoke_softirq();
300 }
301}
302
303#else /* CONFIG_PREEMPT_RT */
304
305/*
306 * This one is for softirq.c-internal use, where hardirqs are disabled
307 * legitimately:
308 */
309#ifdef CONFIG_TRACE_IRQFLAGS
310void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
311{
312 unsigned long flags;
313
314 WARN_ON_ONCE(in_hardirq());
315
316 raw_local_irq_save(flags);
317 /*
318 * The preempt tracer hooks into preempt_count_add and will break
319 * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
320 * is set and before current->softirq_enabled is cleared.
321 * We must manually increment preempt_count here and manually
322 * call the trace_preempt_off later.
323 */
324 __preempt_count_add(cnt);
325 /*
326 * Were softirqs turned off above:
327 */
328 if (softirq_count() == (cnt & SOFTIRQ_MASK))
329 lockdep_softirqs_off(ip);
330 raw_local_irq_restore(flags);
331
332 if (preempt_count() == cnt) {
333#ifdef CONFIG_DEBUG_PREEMPT
334 current->preempt_disable_ip = get_lock_parent_ip();
335#endif
336 trace_preempt_off(CALLER_ADDR0, get_lock_parent_ip());
337 }
338}
339EXPORT_SYMBOL(__local_bh_disable_ip);
340#endif /* CONFIG_TRACE_IRQFLAGS */
341
342static void __local_bh_enable(unsigned int cnt)
343{
344 lockdep_assert_irqs_disabled();
345
346 if (preempt_count() == cnt)
347 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
348
349 if (softirq_count() == (cnt & SOFTIRQ_MASK))
350 lockdep_softirqs_on(_RET_IP_);
351
352 __preempt_count_sub(cnt);
353}
354
355/*
356 * Special-case - softirqs can safely be enabled by __do_softirq(),
357 * without processing still-pending softirqs:
358 */
359void _local_bh_enable(void)
360{
361 WARN_ON_ONCE(in_hardirq());
362 __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
363}
364EXPORT_SYMBOL(_local_bh_enable);
365
366void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
367{
368 WARN_ON_ONCE(in_hardirq());
369 lockdep_assert_irqs_enabled();
370#ifdef CONFIG_TRACE_IRQFLAGS
371 local_irq_disable();
372#endif
373 /*
374 * Are softirqs going to be turned on now:
375 */
376 if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
377 lockdep_softirqs_on(ip);
378 /*
379 * Keep preemption disabled until we are done with
380 * softirq processing:
381 */
382 __preempt_count_sub(cnt - 1);
383
384 if (unlikely(!in_interrupt() && local_softirq_pending())) {
385 /*
386 * Run softirq if any pending. And do it in its own stack
387 * as we may be calling this deep in a task call stack already.
388 */
389 do_softirq();
390 }
391
392 preempt_count_dec();
393#ifdef CONFIG_TRACE_IRQFLAGS
394 local_irq_enable();
395#endif
396 preempt_check_resched();
397}
398EXPORT_SYMBOL(__local_bh_enable_ip);
399
400static inline void softirq_handle_begin(void)
401{
402 __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
403}
404
405static inline void softirq_handle_end(void)
406{
407 __local_bh_enable(SOFTIRQ_OFFSET);
408 WARN_ON_ONCE(in_interrupt());
409}
410
411static inline void ksoftirqd_run_begin(void)
412{
413 local_irq_disable();
414}
415
416static inline void ksoftirqd_run_end(void)
417{
418 local_irq_enable();
419}
420
421static inline bool should_wake_ksoftirqd(void)
422{
423 return true;
424}
425
426static inline void invoke_softirq(void)
427{
428 if (!force_irqthreads() || !__this_cpu_read(ksoftirqd)) {
429#ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
430 /*
431 * We can safely execute softirq on the current stack if
432 * it is the irq stack, because it should be near empty
433 * at this stage.
434 */
435 __do_softirq();
436#else
437 /*
438 * Otherwise, irq_exit() is called on the task stack that can
439 * be potentially deep already. So call softirq in its own stack
440 * to prevent from any overrun.
441 */
442 do_softirq_own_stack();
443#endif
444 } else {
445 wakeup_softirqd();
446 }
447}
448
449asmlinkage __visible void do_softirq(void)
450{
451 __u32 pending;
452 unsigned long flags;
453
454 if (in_interrupt())
455 return;
456
457 local_irq_save(flags);
458
459 pending = local_softirq_pending();
460
461 if (pending)
462 do_softirq_own_stack();
463
464 local_irq_restore(flags);
465}
466
467#endif /* !CONFIG_PREEMPT_RT */
468
469/*
470 * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times,
471 * but break the loop if need_resched() is set or after 2 ms.
472 * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in
473 * certain cases, such as stop_machine(), jiffies may cease to
474 * increment and so we need the MAX_SOFTIRQ_RESTART limit as
475 * well to make sure we eventually return from this method.
476 *
477 * These limits have been established via experimentation.
478 * The two things to balance is latency against fairness -
479 * we want to handle softirqs as soon as possible, but they
480 * should not be able to lock up the box.
481 */
482#define MAX_SOFTIRQ_TIME msecs_to_jiffies(2)
483#define MAX_SOFTIRQ_RESTART 10
484
485#ifdef CONFIG_TRACE_IRQFLAGS
486/*
487 * When we run softirqs from irq_exit() and thus on the hardirq stack we need
488 * to keep the lockdep irq context tracking as tight as possible in order to
489 * not miss-qualify lock contexts and miss possible deadlocks.
490 */
491
492static inline bool lockdep_softirq_start(void)
493{
494 bool in_hardirq = false;
495
496 if (lockdep_hardirq_context()) {
497 in_hardirq = true;
498 lockdep_hardirq_exit();
499 }
500
501 lockdep_softirq_enter();
502
503 return in_hardirq;
504}
505
506static inline void lockdep_softirq_end(bool in_hardirq)
507{
508 lockdep_softirq_exit();
509
510 if (in_hardirq)
511 lockdep_hardirq_enter();
512}
513#else
514static inline bool lockdep_softirq_start(void) { return false; }
515static inline void lockdep_softirq_end(bool in_hardirq) { }
516#endif
517
518static void handle_softirqs(bool ksirqd)
519{
520 unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
521 unsigned long old_flags = current->flags;
522 int max_restart = MAX_SOFTIRQ_RESTART;
523 struct softirq_action *h;
524 bool in_hardirq;
525 __u32 pending;
526 int softirq_bit;
527
528 /*
529 * Mask out PF_MEMALLOC as the current task context is borrowed for the
530 * softirq. A softirq handled, such as network RX, might set PF_MEMALLOC
531 * again if the socket is related to swapping.
532 */
533 current->flags &= ~PF_MEMALLOC;
534
535 pending = local_softirq_pending();
536
537 softirq_handle_begin();
538 in_hardirq = lockdep_softirq_start();
539 account_softirq_enter(current);
540
541restart:
542 /* Reset the pending bitmask before enabling irqs */
543 set_softirq_pending(0);
544
545 local_irq_enable();
546
547 h = softirq_vec;
548
549 while ((softirq_bit = ffs(pending))) {
550 unsigned int vec_nr;
551 int prev_count;
552
553 h += softirq_bit - 1;
554
555 vec_nr = h - softirq_vec;
556 prev_count = preempt_count();
557
558 kstat_incr_softirqs_this_cpu(vec_nr);
559
560 trace_softirq_entry(vec_nr);
561 h->action();
562 trace_softirq_exit(vec_nr);
563 if (unlikely(prev_count != preempt_count())) {
564 pr_err("huh, entered softirq %u %s %p with preempt_count %08x, exited with %08x?\n",
565 vec_nr, softirq_to_name[vec_nr], h->action,
566 prev_count, preempt_count());
567 preempt_count_set(prev_count);
568 }
569 h++;
570 pending >>= softirq_bit;
571 }
572
573 if (!IS_ENABLED(CONFIG_PREEMPT_RT) && ksirqd)
574 rcu_softirq_qs();
575
576 local_irq_disable();
577
578 pending = local_softirq_pending();
579 if (pending) {
580 if (time_before(jiffies, end) && !need_resched() &&
581 --max_restart)
582 goto restart;
583
584 wakeup_softirqd();
585 }
586
587 account_softirq_exit(current);
588 lockdep_softirq_end(in_hardirq);
589 softirq_handle_end();
590 current_restore_flags(old_flags, PF_MEMALLOC);
591}
592
593asmlinkage __visible void __softirq_entry __do_softirq(void)
594{
595 handle_softirqs(false);
596}
597
598/**
599 * irq_enter_rcu - Enter an interrupt context with RCU watching
600 */
601void irq_enter_rcu(void)
602{
603 __irq_enter_raw();
604
605 if (tick_nohz_full_cpu(smp_processor_id()) ||
606 (is_idle_task(current) && (irq_count() == HARDIRQ_OFFSET)))
607 tick_irq_enter();
608
609 account_hardirq_enter(current);
610}
611
612/**
613 * irq_enter - Enter an interrupt context including RCU update
614 */
615void irq_enter(void)
616{
617 ct_irq_enter();
618 irq_enter_rcu();
619}
620
621static inline void tick_irq_exit(void)
622{
623#ifdef CONFIG_NO_HZ_COMMON
624 int cpu = smp_processor_id();
625
626 /* Make sure that timer wheel updates are propagated */
627 if ((sched_core_idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
628 if (!in_hardirq())
629 tick_nohz_irq_exit();
630 }
631#endif
632}
633
634#ifdef CONFIG_IRQ_FORCED_THREADING
635DEFINE_PER_CPU(struct task_struct *, ktimerd);
636DEFINE_PER_CPU(unsigned long, pending_timer_softirq);
637
638static void wake_timersd(void)
639{
640 struct task_struct *tsk = __this_cpu_read(ktimerd);
641
642 if (tsk)
643 wake_up_process(tsk);
644}
645
646#else
647
648static inline void wake_timersd(void) { }
649
650#endif
651
652static inline void __irq_exit_rcu(void)
653{
654#ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
655 local_irq_disable();
656#else
657 lockdep_assert_irqs_disabled();
658#endif
659 account_hardirq_exit(current);
660 preempt_count_sub(HARDIRQ_OFFSET);
661 if (!in_interrupt() && local_softirq_pending())
662 invoke_softirq();
663
664 if (IS_ENABLED(CONFIG_IRQ_FORCED_THREADING) && force_irqthreads() &&
665 local_timers_pending_force_th() && !(in_nmi() | in_hardirq()))
666 wake_timersd();
667
668 tick_irq_exit();
669}
670
671/**
672 * irq_exit_rcu() - Exit an interrupt context without updating RCU
673 *
674 * Also processes softirqs if needed and possible.
675 */
676void irq_exit_rcu(void)
677{
678 __irq_exit_rcu();
679 /* must be last! */
680 lockdep_hardirq_exit();
681}
682
683/**
684 * irq_exit - Exit an interrupt context, update RCU and lockdep
685 *
686 * Also processes softirqs if needed and possible.
687 */
688void irq_exit(void)
689{
690 __irq_exit_rcu();
691 ct_irq_exit();
692 /* must be last! */
693 lockdep_hardirq_exit();
694}
695
696/*
697 * This function must run with irqs disabled!
698 */
699inline void raise_softirq_irqoff(unsigned int nr)
700{
701 __raise_softirq_irqoff(nr);
702
703 /*
704 * If we're in an interrupt or softirq, we're done
705 * (this also catches softirq-disabled code). We will
706 * actually run the softirq once we return from
707 * the irq or softirq.
708 *
709 * Otherwise we wake up ksoftirqd to make sure we
710 * schedule the softirq soon.
711 */
712 if (!in_interrupt() && should_wake_ksoftirqd())
713 wakeup_softirqd();
714}
715
716void raise_softirq(unsigned int nr)
717{
718 unsigned long flags;
719
720 local_irq_save(flags);
721 raise_softirq_irqoff(nr);
722 local_irq_restore(flags);
723}
724
725void __raise_softirq_irqoff(unsigned int nr)
726{
727 lockdep_assert_irqs_disabled();
728 trace_softirq_raise(nr);
729 or_softirq_pending(1UL << nr);
730}
731
732void open_softirq(int nr, void (*action)(void))
733{
734 softirq_vec[nr].action = action;
735}
736
737/*
738 * Tasklets
739 */
740struct tasklet_head {
741 struct tasklet_struct *head;
742 struct tasklet_struct **tail;
743};
744
745static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
746static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
747
748static void __tasklet_schedule_common(struct tasklet_struct *t,
749 struct tasklet_head __percpu *headp,
750 unsigned int softirq_nr)
751{
752 struct tasklet_head *head;
753 unsigned long flags;
754
755 local_irq_save(flags);
756 head = this_cpu_ptr(headp);
757 t->next = NULL;
758 *head->tail = t;
759 head->tail = &(t->next);
760 raise_softirq_irqoff(softirq_nr);
761 local_irq_restore(flags);
762}
763
764void __tasklet_schedule(struct tasklet_struct *t)
765{
766 __tasklet_schedule_common(t, &tasklet_vec,
767 TASKLET_SOFTIRQ);
768}
769EXPORT_SYMBOL(__tasklet_schedule);
770
771void __tasklet_hi_schedule(struct tasklet_struct *t)
772{
773 __tasklet_schedule_common(t, &tasklet_hi_vec,
774 HI_SOFTIRQ);
775}
776EXPORT_SYMBOL(__tasklet_hi_schedule);
777
778static bool tasklet_clear_sched(struct tasklet_struct *t)
779{
780 if (test_and_clear_wake_up_bit(TASKLET_STATE_SCHED, &t->state))
781 return true;
782
783 WARN_ONCE(1, "tasklet SCHED state not set: %s %pS\n",
784 t->use_callback ? "callback" : "func",
785 t->use_callback ? (void *)t->callback : (void *)t->func);
786
787 return false;
788}
789
790static void tasklet_action_common(struct tasklet_head *tl_head,
791 unsigned int softirq_nr)
792{
793 struct tasklet_struct *list;
794
795 local_irq_disable();
796 list = tl_head->head;
797 tl_head->head = NULL;
798 tl_head->tail = &tl_head->head;
799 local_irq_enable();
800
801 while (list) {
802 struct tasklet_struct *t = list;
803
804 list = list->next;
805
806 if (tasklet_trylock(t)) {
807 if (!atomic_read(&t->count)) {
808 if (tasklet_clear_sched(t)) {
809 if (t->use_callback) {
810 trace_tasklet_entry(t, t->callback);
811 t->callback(t);
812 trace_tasklet_exit(t, t->callback);
813 } else {
814 trace_tasklet_entry(t, t->func);
815 t->func(t->data);
816 trace_tasklet_exit(t, t->func);
817 }
818 }
819 tasklet_unlock(t);
820 continue;
821 }
822 tasklet_unlock(t);
823 }
824
825 local_irq_disable();
826 t->next = NULL;
827 *tl_head->tail = t;
828 tl_head->tail = &t->next;
829 __raise_softirq_irqoff(softirq_nr);
830 local_irq_enable();
831 }
832}
833
834static __latent_entropy void tasklet_action(void)
835{
836 workqueue_softirq_action(false);
837 tasklet_action_common(this_cpu_ptr(&tasklet_vec), TASKLET_SOFTIRQ);
838}
839
840static __latent_entropy void tasklet_hi_action(void)
841{
842 workqueue_softirq_action(true);
843 tasklet_action_common(this_cpu_ptr(&tasklet_hi_vec), HI_SOFTIRQ);
844}
845
846void tasklet_setup(struct tasklet_struct *t,
847 void (*callback)(struct tasklet_struct *))
848{
849 t->next = NULL;
850 t->state = 0;
851 atomic_set(&t->count, 0);
852 t->callback = callback;
853 t->use_callback = true;
854 t->data = 0;
855}
856EXPORT_SYMBOL(tasklet_setup);
857
858void tasklet_init(struct tasklet_struct *t,
859 void (*func)(unsigned long), unsigned long data)
860{
861 t->next = NULL;
862 t->state = 0;
863 atomic_set(&t->count, 0);
864 t->func = func;
865 t->use_callback = false;
866 t->data = data;
867}
868EXPORT_SYMBOL(tasklet_init);
869
870#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
871/*
872 * Do not use in new code. Waiting for tasklets from atomic contexts is
873 * error prone and should be avoided.
874 */
875void tasklet_unlock_spin_wait(struct tasklet_struct *t)
876{
877 while (test_bit(TASKLET_STATE_RUN, &(t)->state)) {
878 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
879 /*
880 * Prevent a live lock when current preempted soft
881 * interrupt processing or prevents ksoftirqd from
882 * running. If the tasklet runs on a different CPU
883 * then this has no effect other than doing the BH
884 * disable/enable dance for nothing.
885 */
886 local_bh_disable();
887 local_bh_enable();
888 } else {
889 cpu_relax();
890 }
891 }
892}
893EXPORT_SYMBOL(tasklet_unlock_spin_wait);
894#endif
895
896void tasklet_kill(struct tasklet_struct *t)
897{
898 if (in_interrupt())
899 pr_notice("Attempt to kill tasklet from interrupt\n");
900
901 wait_on_bit_lock(&t->state, TASKLET_STATE_SCHED, TASK_UNINTERRUPTIBLE);
902
903 tasklet_unlock_wait(t);
904 tasklet_clear_sched(t);
905}
906EXPORT_SYMBOL(tasklet_kill);
907
908#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
909void tasklet_unlock(struct tasklet_struct *t)
910{
911 clear_and_wake_up_bit(TASKLET_STATE_RUN, &t->state);
912}
913EXPORT_SYMBOL_GPL(tasklet_unlock);
914
915void tasklet_unlock_wait(struct tasklet_struct *t)
916{
917 wait_on_bit(&t->state, TASKLET_STATE_RUN, TASK_UNINTERRUPTIBLE);
918}
919EXPORT_SYMBOL_GPL(tasklet_unlock_wait);
920#endif
921
922void __init softirq_init(void)
923{
924 int cpu;
925
926 for_each_possible_cpu(cpu) {
927 per_cpu(tasklet_vec, cpu).tail =
928 &per_cpu(tasklet_vec, cpu).head;
929 per_cpu(tasklet_hi_vec, cpu).tail =
930 &per_cpu(tasklet_hi_vec, cpu).head;
931 }
932
933 open_softirq(TASKLET_SOFTIRQ, tasklet_action);
934 open_softirq(HI_SOFTIRQ, tasklet_hi_action);
935}
936
937static int ksoftirqd_should_run(unsigned int cpu)
938{
939 return local_softirq_pending();
940}
941
942static void run_ksoftirqd(unsigned int cpu)
943{
944 ksoftirqd_run_begin();
945 if (local_softirq_pending()) {
946 /*
947 * We can safely run softirq on inline stack, as we are not deep
948 * in the task stack here.
949 */
950 handle_softirqs(true);
951 ksoftirqd_run_end();
952 cond_resched();
953 return;
954 }
955 ksoftirqd_run_end();
956}
957
958#ifdef CONFIG_HOTPLUG_CPU
959static int takeover_tasklets(unsigned int cpu)
960{
961 workqueue_softirq_dead(cpu);
962
963 /* CPU is dead, so no lock needed. */
964 local_irq_disable();
965
966 /* Find end, append list for that CPU. */
967 if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
968 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
969 __this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
970 per_cpu(tasklet_vec, cpu).head = NULL;
971 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
972 }
973 raise_softirq_irqoff(TASKLET_SOFTIRQ);
974
975 if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
976 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
977 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
978 per_cpu(tasklet_hi_vec, cpu).head = NULL;
979 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
980 }
981 raise_softirq_irqoff(HI_SOFTIRQ);
982
983 local_irq_enable();
984 return 0;
985}
986#else
987#define takeover_tasklets NULL
988#endif /* CONFIG_HOTPLUG_CPU */
989
990static struct smp_hotplug_thread softirq_threads = {
991 .store = &ksoftirqd,
992 .thread_should_run = ksoftirqd_should_run,
993 .thread_fn = run_ksoftirqd,
994 .thread_comm = "ksoftirqd/%u",
995};
996
997#ifdef CONFIG_IRQ_FORCED_THREADING
998static void ktimerd_setup(unsigned int cpu)
999{
1000 /* Above SCHED_NORMAL to handle timers before regular tasks. */
1001 sched_set_fifo_low(current);
1002}
1003
1004static int ktimerd_should_run(unsigned int cpu)
1005{
1006 return local_timers_pending_force_th();
1007}
1008
1009void raise_ktimers_thread(unsigned int nr)
1010{
1011 trace_softirq_raise(nr);
1012 __this_cpu_or(pending_timer_softirq, BIT(nr));
1013}
1014
1015static void run_ktimerd(unsigned int cpu)
1016{
1017 unsigned int timer_si;
1018
1019 ksoftirqd_run_begin();
1020
1021 timer_si = local_timers_pending_force_th();
1022 __this_cpu_write(pending_timer_softirq, 0);
1023 or_softirq_pending(timer_si);
1024
1025 __do_softirq();
1026
1027 ksoftirqd_run_end();
1028}
1029
1030static struct smp_hotplug_thread timer_thread = {
1031 .store = &ktimerd,
1032 .setup = ktimerd_setup,
1033 .thread_should_run = ktimerd_should_run,
1034 .thread_fn = run_ktimerd,
1035 .thread_comm = "ktimers/%u",
1036};
1037#endif
1038
1039static __init int spawn_ksoftirqd(void)
1040{
1041 cpuhp_setup_state_nocalls(CPUHP_SOFTIRQ_DEAD, "softirq:dead", NULL,
1042 takeover_tasklets);
1043 BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
1044#ifdef CONFIG_IRQ_FORCED_THREADING
1045 if (force_irqthreads())
1046 BUG_ON(smpboot_register_percpu_thread(&timer_thread));
1047#endif
1048 return 0;
1049}
1050early_initcall(spawn_ksoftirqd);
1051
1052/*
1053 * [ These __weak aliases are kept in a separate compilation unit, so that
1054 * GCC does not inline them incorrectly. ]
1055 */
1056
1057int __init __weak early_irq_init(void)
1058{
1059 return 0;
1060}
1061
1062int __init __weak arch_probe_nr_irqs(void)
1063{
1064 return NR_IRQS_LEGACY;
1065}
1066
1067int __init __weak arch_early_irq_init(void)
1068{
1069 return 0;
1070}
1071
1072unsigned int __weak arch_dynirq_lower_bound(unsigned int from)
1073{
1074 return from;
1075}