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
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}