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