Linux Audio

Check our new training course

Embedded Linux training

Mar 10-20, 2025, special US time zones
Register
Loading...
v5.4
  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/mm.h>
 17#include <linux/notifier.h>
 18#include <linux/percpu.h>
 19#include <linux/cpu.h>
 20#include <linux/freezer.h>
 21#include <linux/kthread.h>
 22#include <linux/rcupdate.h>
 23#include <linux/ftrace.h>
 24#include <linux/smp.h>
 25#include <linux/smpboot.h>
 26#include <linux/tick.h>
 27#include <linux/irq.h>
 
 
 
 
 28
 29#define CREATE_TRACE_POINTS
 30#include <trace/events/irq.h>
 31
 32/*
 33   - No shared variables, all the data are CPU local.
 34   - If a softirq needs serialization, let it serialize itself
 35     by its own spinlocks.
 36   - Even if softirq is serialized, only local cpu is marked for
 37     execution. Hence, we get something sort of weak cpu binding.
 38     Though it is still not clear, will it result in better locality
 39     or will not.
 40
 41   Examples:
 42   - NET RX softirq. It is multithreaded and does not require
 43     any global serialization.
 44   - NET TX softirq. It kicks software netdevice queues, hence
 45     it is logically serialized per device, but this serialization
 46     is invisible to common code.
 47   - Tasklets: serialized wrt itself.
 48 */
 49
 50#ifndef __ARCH_IRQ_STAT
 51DEFINE_PER_CPU_ALIGNED(irq_cpustat_t, irq_stat);
 52EXPORT_PER_CPU_SYMBOL(irq_stat);
 53#endif
 54
 55static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
 56
 57DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
 58
 59const char * const softirq_to_name[NR_SOFTIRQS] = {
 60	"HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "IRQ_POLL",
 61	"TASKLET", "SCHED", "HRTIMER", "RCU"
 62};
 63
 64/*
 65 * we cannot loop indefinitely here to avoid userspace starvation,
 66 * but we also don't want to introduce a worst case 1/HZ latency
 67 * to the pending events, so lets the scheduler to balance
 68 * the softirq load for us.
 69 */
 70static void wakeup_softirqd(void)
 71{
 72	/* Interrupts are disabled: no need to stop preemption */
 73	struct task_struct *tsk = __this_cpu_read(ksoftirqd);
 74
 75	if (tsk && tsk->state != TASK_RUNNING)
 76		wake_up_process(tsk);
 77}
 78
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 79/*
 80 * If ksoftirqd is scheduled, we do not want to process pending softirqs
 81 * right now. Let ksoftirqd handle this at its own rate, to get fairness,
 82 * unless we're doing some of the synchronous softirqs.
 
 
 
 
 
 
 
 83 */
 84#define SOFTIRQ_NOW_MASK ((1 << HI_SOFTIRQ) | (1 << TASKLET_SOFTIRQ))
 85static bool ksoftirqd_running(unsigned long pending)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 86{
 87	struct task_struct *tsk = __this_cpu_read(ksoftirqd);
 
 88
 89	if (pending & SOFTIRQ_NOW_MASK)
 90		return false;
 91	return tsk && (tsk->state == TASK_RUNNING) &&
 92		!__kthread_should_park(tsk);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 93}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 94
 95/*
 96 * preempt_count and SOFTIRQ_OFFSET usage:
 97 * - preempt_count is changed by SOFTIRQ_OFFSET on entering or leaving
 98 *   softirq processing.
 99 * - preempt_count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
100 *   on local_bh_disable or local_bh_enable.
101 * This lets us distinguish between whether we are currently processing
102 * softirq and whether we just have bh disabled.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
103 */
 
 
 
 
 
 
 
104
105/*
106 * This one is for softirq.c-internal use,
107 * where hardirqs are disabled legitimately:
108 */
109#ifdef CONFIG_TRACE_IRQFLAGS
110void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
111{
112	unsigned long flags;
113
114	WARN_ON_ONCE(in_irq());
115
116	raw_local_irq_save(flags);
117	/*
118	 * The preempt tracer hooks into preempt_count_add and will break
119	 * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
120	 * is set and before current->softirq_enabled is cleared.
121	 * We must manually increment preempt_count here and manually
122	 * call the trace_preempt_off later.
123	 */
124	__preempt_count_add(cnt);
125	/*
126	 * Were softirqs turned off above:
127	 */
128	if (softirq_count() == (cnt & SOFTIRQ_MASK))
129		trace_softirqs_off(ip);
130	raw_local_irq_restore(flags);
131
132	if (preempt_count() == cnt) {
133#ifdef CONFIG_DEBUG_PREEMPT
134		current->preempt_disable_ip = get_lock_parent_ip();
135#endif
136		trace_preempt_off(CALLER_ADDR0, get_lock_parent_ip());
137	}
138}
139EXPORT_SYMBOL(__local_bh_disable_ip);
140#endif /* CONFIG_TRACE_IRQFLAGS */
141
142static void __local_bh_enable(unsigned int cnt)
143{
144	lockdep_assert_irqs_disabled();
145
146	if (preempt_count() == cnt)
147		trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
148
149	if (softirq_count() == (cnt & SOFTIRQ_MASK))
150		trace_softirqs_on(_RET_IP_);
151
152	__preempt_count_sub(cnt);
153}
154
155/*
156 * Special-case - softirqs can safely be enabled by __do_softirq(),
157 * without processing still-pending softirqs:
158 */
159void _local_bh_enable(void)
160{
161	WARN_ON_ONCE(in_irq());
162	__local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
163}
164EXPORT_SYMBOL(_local_bh_enable);
165
166void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
167{
168	WARN_ON_ONCE(in_irq());
169	lockdep_assert_irqs_enabled();
170#ifdef CONFIG_TRACE_IRQFLAGS
171	local_irq_disable();
172#endif
173	/*
174	 * Are softirqs going to be turned on now:
175	 */
176	if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
177		trace_softirqs_on(ip);
178	/*
179	 * Keep preemption disabled until we are done with
180	 * softirq processing:
181	 */
182	preempt_count_sub(cnt - 1);
183
184	if (unlikely(!in_interrupt() && local_softirq_pending())) {
185		/*
186		 * Run softirq if any pending. And do it in its own stack
187		 * as we may be calling this deep in a task call stack already.
188		 */
189		do_softirq();
190	}
191
192	preempt_count_dec();
193#ifdef CONFIG_TRACE_IRQFLAGS
194	local_irq_enable();
195#endif
196	preempt_check_resched();
197}
198EXPORT_SYMBOL(__local_bh_enable_ip);
199
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
200/*
201 * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times,
202 * but break the loop if need_resched() is set or after 2 ms.
203 * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in
204 * certain cases, such as stop_machine(), jiffies may cease to
205 * increment and so we need the MAX_SOFTIRQ_RESTART limit as
206 * well to make sure we eventually return from this method.
207 *
208 * These limits have been established via experimentation.
209 * The two things to balance is latency against fairness -
210 * we want to handle softirqs as soon as possible, but they
211 * should not be able to lock up the box.
212 */
213#define MAX_SOFTIRQ_TIME  msecs_to_jiffies(2)
214#define MAX_SOFTIRQ_RESTART 10
215
216#ifdef CONFIG_TRACE_IRQFLAGS
217/*
218 * When we run softirqs from irq_exit() and thus on the hardirq stack we need
219 * to keep the lockdep irq context tracking as tight as possible in order to
220 * not miss-qualify lock contexts and miss possible deadlocks.
221 */
222
223static inline bool lockdep_softirq_start(void)
224{
225	bool in_hardirq = false;
226
227	if (trace_hardirq_context(current)) {
228		in_hardirq = true;
229		trace_hardirq_exit();
230	}
231
232	lockdep_softirq_enter();
233
234	return in_hardirq;
235}
236
237static inline void lockdep_softirq_end(bool in_hardirq)
238{
239	lockdep_softirq_exit();
240
241	if (in_hardirq)
242		trace_hardirq_enter();
243}
244#else
245static inline bool lockdep_softirq_start(void) { return false; }
246static inline void lockdep_softirq_end(bool in_hardirq) { }
247#endif
248
249asmlinkage __visible void __softirq_entry __do_softirq(void)
250{
251	unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
252	unsigned long old_flags = current->flags;
253	int max_restart = MAX_SOFTIRQ_RESTART;
254	struct softirq_action *h;
255	bool in_hardirq;
256	__u32 pending;
257	int softirq_bit;
258
259	/*
260	 * Mask out PF_MEMALLOC as the current task context is borrowed for the
261	 * softirq. A softirq handled, such as network RX, might set PF_MEMALLOC
262	 * again if the socket is related to swapping.
263	 */
264	current->flags &= ~PF_MEMALLOC;
265
266	pending = local_softirq_pending();
267	account_irq_enter_time(current);
268
269	__local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
270	in_hardirq = lockdep_softirq_start();
 
271
272restart:
273	/* Reset the pending bitmask before enabling irqs */
274	set_softirq_pending(0);
275
276	local_irq_enable();
277
278	h = softirq_vec;
279
280	while ((softirq_bit = ffs(pending))) {
281		unsigned int vec_nr;
282		int prev_count;
283
284		h += softirq_bit - 1;
285
286		vec_nr = h - softirq_vec;
287		prev_count = preempt_count();
288
289		kstat_incr_softirqs_this_cpu(vec_nr);
290
291		trace_softirq_entry(vec_nr);
292		h->action(h);
293		trace_softirq_exit(vec_nr);
294		if (unlikely(prev_count != preempt_count())) {
295			pr_err("huh, entered softirq %u %s %p with preempt_count %08x, exited with %08x?\n",
296			       vec_nr, softirq_to_name[vec_nr], h->action,
297			       prev_count, preempt_count());
298			preempt_count_set(prev_count);
299		}
300		h++;
301		pending >>= softirq_bit;
302	}
303
304	if (__this_cpu_read(ksoftirqd) == current)
305		rcu_softirq_qs();
 
306	local_irq_disable();
307
308	pending = local_softirq_pending();
309	if (pending) {
310		if (time_before(jiffies, end) && !need_resched() &&
311		    --max_restart)
312			goto restart;
313
314		wakeup_softirqd();
315	}
316
 
317	lockdep_softirq_end(in_hardirq);
318	account_irq_exit_time(current);
319	__local_bh_enable(SOFTIRQ_OFFSET);
320	WARN_ON_ONCE(in_interrupt());
321	current_restore_flags(old_flags, PF_MEMALLOC);
322}
323
324asmlinkage __visible void do_softirq(void)
325{
326	__u32 pending;
327	unsigned long flags;
328
329	if (in_interrupt())
330		return;
331
332	local_irq_save(flags);
333
334	pending = local_softirq_pending();
335
336	if (pending && !ksoftirqd_running(pending))
337		do_softirq_own_stack();
338
339	local_irq_restore(flags);
340}
341
342/*
343 * Enter an interrupt context.
344 */
345void irq_enter(void)
346{
347	rcu_irq_enter();
348	if (is_idle_task(current) && !in_interrupt()) {
349		/*
350		 * Prevent raise_softirq from needlessly waking up ksoftirqd
351		 * here, as softirq will be serviced on return from interrupt.
352		 */
353		local_bh_disable();
354		tick_irq_enter();
355		_local_bh_enable();
356	}
357
358	__irq_enter();
359}
360
361static inline void invoke_softirq(void)
 
 
 
362{
363	if (ksoftirqd_running(local_softirq_pending()))
364		return;
365
366	if (!force_irqthreads) {
367#ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
368		/*
369		 * We can safely execute softirq on the current stack if
370		 * it is the irq stack, because it should be near empty
371		 * at this stage.
372		 */
373		__do_softirq();
374#else
375		/*
376		 * Otherwise, irq_exit() is called on the task stack that can
377		 * be potentially deep already. So call softirq in its own stack
378		 * to prevent from any overrun.
379		 */
380		do_softirq_own_stack();
381#endif
382	} else {
383		wakeup_softirqd();
384	}
385}
386
387static inline void tick_irq_exit(void)
388{
389#ifdef CONFIG_NO_HZ_COMMON
390	int cpu = smp_processor_id();
391
392	/* Make sure that timer wheel updates are propagated */
393	if ((idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
394		if (!in_irq())
395			tick_nohz_irq_exit();
396	}
397#endif
398}
399
400/*
401 * Exit an interrupt context. Process softirqs if needed and possible:
402 */
403void irq_exit(void)
404{
405#ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
406	local_irq_disable();
407#else
408	lockdep_assert_irqs_disabled();
409#endif
410	account_irq_exit_time(current);
411	preempt_count_sub(HARDIRQ_OFFSET);
412	if (!in_interrupt() && local_softirq_pending())
413		invoke_softirq();
414
415	tick_irq_exit();
416	rcu_irq_exit();
417	trace_hardirq_exit(); /* must be last! */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
418}
419
420/*
421 * This function must run with irqs disabled!
422 */
423inline void raise_softirq_irqoff(unsigned int nr)
424{
425	__raise_softirq_irqoff(nr);
426
427	/*
428	 * If we're in an interrupt or softirq, we're done
429	 * (this also catches softirq-disabled code). We will
430	 * actually run the softirq once we return from
431	 * the irq or softirq.
432	 *
433	 * Otherwise we wake up ksoftirqd to make sure we
434	 * schedule the softirq soon.
435	 */
436	if (!in_interrupt())
437		wakeup_softirqd();
438}
439
440void raise_softirq(unsigned int nr)
441{
442	unsigned long flags;
443
444	local_irq_save(flags);
445	raise_softirq_irqoff(nr);
446	local_irq_restore(flags);
447}
448
449void __raise_softirq_irqoff(unsigned int nr)
450{
 
451	trace_softirq_raise(nr);
452	or_softirq_pending(1UL << nr);
453}
454
455void open_softirq(int nr, void (*action)(struct softirq_action *))
456{
457	softirq_vec[nr].action = action;
458}
459
460/*
461 * Tasklets
462 */
463struct tasklet_head {
464	struct tasklet_struct *head;
465	struct tasklet_struct **tail;
466};
467
468static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
469static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
470
471static void __tasklet_schedule_common(struct tasklet_struct *t,
472				      struct tasklet_head __percpu *headp,
473				      unsigned int softirq_nr)
474{
475	struct tasklet_head *head;
476	unsigned long flags;
477
478	local_irq_save(flags);
479	head = this_cpu_ptr(headp);
480	t->next = NULL;
481	*head->tail = t;
482	head->tail = &(t->next);
483	raise_softirq_irqoff(softirq_nr);
484	local_irq_restore(flags);
485}
486
487void __tasklet_schedule(struct tasklet_struct *t)
488{
489	__tasklet_schedule_common(t, &tasklet_vec,
490				  TASKLET_SOFTIRQ);
491}
492EXPORT_SYMBOL(__tasklet_schedule);
493
494void __tasklet_hi_schedule(struct tasklet_struct *t)
495{
496	__tasklet_schedule_common(t, &tasklet_hi_vec,
497				  HI_SOFTIRQ);
498}
499EXPORT_SYMBOL(__tasklet_hi_schedule);
500
 
 
 
 
 
 
 
 
 
 
 
 
 
 
501static void tasklet_action_common(struct softirq_action *a,
502				  struct tasklet_head *tl_head,
503				  unsigned int softirq_nr)
504{
505	struct tasklet_struct *list;
506
507	local_irq_disable();
508	list = tl_head->head;
509	tl_head->head = NULL;
510	tl_head->tail = &tl_head->head;
511	local_irq_enable();
512
513	while (list) {
514		struct tasklet_struct *t = list;
515
516		list = list->next;
517
518		if (tasklet_trylock(t)) {
519			if (!atomic_read(&t->count)) {
520				if (!test_and_clear_bit(TASKLET_STATE_SCHED,
521							&t->state))
522					BUG();
523				t->func(t->data);
 
 
 
 
 
 
 
524				tasklet_unlock(t);
525				continue;
526			}
527			tasklet_unlock(t);
528		}
529
530		local_irq_disable();
531		t->next = NULL;
532		*tl_head->tail = t;
533		tl_head->tail = &t->next;
534		__raise_softirq_irqoff(softirq_nr);
535		local_irq_enable();
536	}
537}
538
539static __latent_entropy void tasklet_action(struct softirq_action *a)
540{
 
541	tasklet_action_common(a, this_cpu_ptr(&tasklet_vec), TASKLET_SOFTIRQ);
542}
543
544static __latent_entropy void tasklet_hi_action(struct softirq_action *a)
545{
 
546	tasklet_action_common(a, this_cpu_ptr(&tasklet_hi_vec), HI_SOFTIRQ);
547}
548
 
 
 
 
 
 
 
 
 
 
 
 
549void tasklet_init(struct tasklet_struct *t,
550		  void (*func)(unsigned long), unsigned long data)
551{
552	t->next = NULL;
553	t->state = 0;
554	atomic_set(&t->count, 0);
555	t->func = func;
 
556	t->data = data;
557}
558EXPORT_SYMBOL(tasklet_init);
559
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
560void tasklet_kill(struct tasklet_struct *t)
561{
562	if (in_interrupt())
563		pr_notice("Attempt to kill tasklet from interrupt\n");
564
565	while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
566		do {
567			yield();
568		} while (test_bit(TASKLET_STATE_SCHED, &t->state));
569	}
570	tasklet_unlock_wait(t);
571	clear_bit(TASKLET_STATE_SCHED, &t->state);
572}
573EXPORT_SYMBOL(tasklet_kill);
574
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
575void __init softirq_init(void)
576{
577	int cpu;
578
579	for_each_possible_cpu(cpu) {
580		per_cpu(tasklet_vec, cpu).tail =
581			&per_cpu(tasklet_vec, cpu).head;
582		per_cpu(tasklet_hi_vec, cpu).tail =
583			&per_cpu(tasklet_hi_vec, cpu).head;
584	}
585
586	open_softirq(TASKLET_SOFTIRQ, tasklet_action);
587	open_softirq(HI_SOFTIRQ, tasklet_hi_action);
588}
589
590static int ksoftirqd_should_run(unsigned int cpu)
591{
592	return local_softirq_pending();
593}
594
595static void run_ksoftirqd(unsigned int cpu)
596{
597	local_irq_disable();
598	if (local_softirq_pending()) {
599		/*
600		 * We can safely run softirq on inline stack, as we are not deep
601		 * in the task stack here.
602		 */
603		__do_softirq();
604		local_irq_enable();
605		cond_resched();
606		return;
607	}
608	local_irq_enable();
609}
610
611#ifdef CONFIG_HOTPLUG_CPU
612/*
613 * tasklet_kill_immediate is called to remove a tasklet which can already be
614 * scheduled for execution on @cpu.
615 *
616 * Unlike tasklet_kill, this function removes the tasklet
617 * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
618 *
619 * When this function is called, @cpu must be in the CPU_DEAD state.
620 */
621void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
622{
623	struct tasklet_struct **i;
624
625	BUG_ON(cpu_online(cpu));
626	BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
627
628	if (!test_bit(TASKLET_STATE_SCHED, &t->state))
629		return;
630
631	/* CPU is dead, so no lock needed. */
632	for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
633		if (*i == t) {
634			*i = t->next;
635			/* If this was the tail element, move the tail ptr */
636			if (*i == NULL)
637				per_cpu(tasklet_vec, cpu).tail = i;
638			return;
639		}
640	}
641	BUG();
642}
643
644static int takeover_tasklets(unsigned int cpu)
645{
 
 
646	/* CPU is dead, so no lock needed. */
647	local_irq_disable();
648
649	/* Find end, append list for that CPU. */
650	if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
651		*__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
652		__this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
653		per_cpu(tasklet_vec, cpu).head = NULL;
654		per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
655	}
656	raise_softirq_irqoff(TASKLET_SOFTIRQ);
657
658	if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
659		*__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
660		__this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
661		per_cpu(tasklet_hi_vec, cpu).head = NULL;
662		per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
663	}
664	raise_softirq_irqoff(HI_SOFTIRQ);
665
666	local_irq_enable();
667	return 0;
668}
669#else
670#define takeover_tasklets	NULL
671#endif /* CONFIG_HOTPLUG_CPU */
672
673static struct smp_hotplug_thread softirq_threads = {
674	.store			= &ksoftirqd,
675	.thread_should_run	= ksoftirqd_should_run,
676	.thread_fn		= run_ksoftirqd,
677	.thread_comm		= "ksoftirqd/%u",
678};
679
680static __init int spawn_ksoftirqd(void)
681{
682	cpuhp_setup_state_nocalls(CPUHP_SOFTIRQ_DEAD, "softirq:dead", NULL,
683				  takeover_tasklets);
684	BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
685
686	return 0;
687}
688early_initcall(spawn_ksoftirqd);
689
690/*
691 * [ These __weak aliases are kept in a separate compilation unit, so that
692 *   GCC does not inline them incorrectly. ]
693 */
694
695int __init __weak early_irq_init(void)
696{
697	return 0;
698}
699
700int __init __weak arch_probe_nr_irqs(void)
701{
702	return NR_IRQS_LEGACY;
703}
704
705int __init __weak arch_early_irq_init(void)
706{
707	return 0;
708}
709
710unsigned int __weak arch_dynirq_lower_bound(unsigned int from)
711{
712	return from;
713}
v6.9.4
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 *	linux/kernel/softirq.c
   4 *
   5 *	Copyright (C) 1992 Linus Torvalds
   6 *
   7 *	Rewritten. Old one was good in 2.2, but in 2.3 it was immoral. --ANK (990903)
   8 */
   9
  10#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12#include <linux/export.h>
  13#include <linux/kernel_stat.h>
  14#include <linux/interrupt.h>
  15#include <linux/init.h>
  16#include <linux/local_lock.h>
  17#include <linux/mm.h>
  18#include <linux/notifier.h>
  19#include <linux/percpu.h>
  20#include <linux/cpu.h>
  21#include <linux/freezer.h>
  22#include <linux/kthread.h>
  23#include <linux/rcupdate.h>
  24#include <linux/ftrace.h>
  25#include <linux/smp.h>
  26#include <linux/smpboot.h>
  27#include <linux/tick.h>
  28#include <linux/irq.h>
  29#include <linux/wait_bit.h>
  30#include <linux/workqueue.h>
  31
  32#include <asm/softirq_stack.h>
  33
  34#define CREATE_TRACE_POINTS
  35#include <trace/events/irq.h>
  36
  37/*
  38   - No shared variables, all the data are CPU local.
  39   - If a softirq needs serialization, let it serialize itself
  40     by its own spinlocks.
  41   - Even if softirq is serialized, only local cpu is marked for
  42     execution. Hence, we get something sort of weak cpu binding.
  43     Though it is still not clear, will it result in better locality
  44     or will not.
  45
  46   Examples:
  47   - NET RX softirq. It is multithreaded and does not require
  48     any global serialization.
  49   - NET TX softirq. It kicks software netdevice queues, hence
  50     it is logically serialized per device, but this serialization
  51     is invisible to common code.
  52   - Tasklets: serialized wrt itself.
  53 */
  54
  55#ifndef __ARCH_IRQ_STAT
  56DEFINE_PER_CPU_ALIGNED(irq_cpustat_t, irq_stat);
  57EXPORT_PER_CPU_SYMBOL(irq_stat);
  58#endif
  59
  60static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
  61
  62DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
  63
  64const char * const softirq_to_name[NR_SOFTIRQS] = {
  65	"HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "IRQ_POLL",
  66	"TASKLET", "SCHED", "HRTIMER", "RCU"
  67};
  68
  69/*
  70 * we cannot loop indefinitely here to avoid userspace starvation,
  71 * but we also don't want to introduce a worst case 1/HZ latency
  72 * to the pending events, so lets the scheduler to balance
  73 * the softirq load for us.
  74 */
  75static void wakeup_softirqd(void)
  76{
  77	/* Interrupts are disabled: no need to stop preemption */
  78	struct task_struct *tsk = __this_cpu_read(ksoftirqd);
  79
  80	if (tsk)
  81		wake_up_process(tsk);
  82}
  83
  84#ifdef CONFIG_TRACE_IRQFLAGS
  85DEFINE_PER_CPU(int, hardirqs_enabled);
  86DEFINE_PER_CPU(int, hardirq_context);
  87EXPORT_PER_CPU_SYMBOL_GPL(hardirqs_enabled);
  88EXPORT_PER_CPU_SYMBOL_GPL(hardirq_context);
  89#endif
  90
  91/*
  92 * SOFTIRQ_OFFSET usage:
  93 *
  94 * On !RT kernels 'count' is the preempt counter, on RT kernels this applies
  95 * to a per CPU counter and to task::softirqs_disabled_cnt.
  96 *
  97 * - count is changed by SOFTIRQ_OFFSET on entering or leaving softirq
  98 *   processing.
  99 *
 100 * - count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
 101 *   on local_bh_disable or local_bh_enable.
 102 *
 103 * This lets us distinguish between whether we are currently processing
 104 * softirq and whether we just have bh disabled.
 105 */
 106#ifdef CONFIG_PREEMPT_RT
 107
 108/*
 109 * RT accounts for BH disabled sections in task::softirqs_disabled_cnt and
 110 * also in per CPU softirq_ctrl::cnt. This is necessary to allow tasks in a
 111 * softirq disabled section to be preempted.
 112 *
 113 * The per task counter is used for softirq_count(), in_softirq() and
 114 * in_serving_softirqs() because these counts are only valid when the task
 115 * holding softirq_ctrl::lock is running.
 116 *
 117 * The per CPU counter prevents pointless wakeups of ksoftirqd in case that
 118 * the task which is in a softirq disabled section is preempted or blocks.
 119 */
 120struct softirq_ctrl {
 121	local_lock_t	lock;
 122	int		cnt;
 123};
 124
 125static DEFINE_PER_CPU(struct softirq_ctrl, softirq_ctrl) = {
 126	.lock	= INIT_LOCAL_LOCK(softirq_ctrl.lock),
 127};
 128
 129/**
 130 * local_bh_blocked() - Check for idle whether BH processing is blocked
 131 *
 132 * Returns false if the per CPU softirq::cnt is 0 otherwise true.
 133 *
 134 * This is invoked from the idle task to guard against false positive
 135 * softirq pending warnings, which would happen when the task which holds
 136 * softirq_ctrl::lock was the only running task on the CPU and blocks on
 137 * some other lock.
 138 */
 139bool local_bh_blocked(void)
 140{
 141	return __this_cpu_read(softirq_ctrl.cnt) != 0;
 142}
 143
 144void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
 145{
 146	unsigned long flags;
 147	int newcnt;
 148
 149	WARN_ON_ONCE(in_hardirq());
 150
 151	/* First entry of a task into a BH disabled section? */
 152	if (!current->softirq_disable_cnt) {
 153		if (preemptible()) {
 154			local_lock(&softirq_ctrl.lock);
 155			/* Required to meet the RCU bottomhalf requirements. */
 156			rcu_read_lock();
 157		} else {
 158			DEBUG_LOCKS_WARN_ON(this_cpu_read(softirq_ctrl.cnt));
 159		}
 160	}
 161
 162	/*
 163	 * Track the per CPU softirq disabled state. On RT this is per CPU
 164	 * state to allow preemption of bottom half disabled sections.
 165	 */
 166	newcnt = __this_cpu_add_return(softirq_ctrl.cnt, cnt);
 167	/*
 168	 * Reflect the result in the task state to prevent recursion on the
 169	 * local lock and to make softirq_count() & al work.
 170	 */
 171	current->softirq_disable_cnt = newcnt;
 172
 173	if (IS_ENABLED(CONFIG_TRACE_IRQFLAGS) && newcnt == cnt) {
 174		raw_local_irq_save(flags);
 175		lockdep_softirqs_off(ip);
 176		raw_local_irq_restore(flags);
 177	}
 178}
 179EXPORT_SYMBOL(__local_bh_disable_ip);
 180
 181static void __local_bh_enable(unsigned int cnt, bool unlock)
 182{
 183	unsigned long flags;
 184	int newcnt;
 185
 186	DEBUG_LOCKS_WARN_ON(current->softirq_disable_cnt !=
 187			    this_cpu_read(softirq_ctrl.cnt));
 188
 189	if (IS_ENABLED(CONFIG_TRACE_IRQFLAGS) && softirq_count() == cnt) {
 190		raw_local_irq_save(flags);
 191		lockdep_softirqs_on(_RET_IP_);
 192		raw_local_irq_restore(flags);
 193	}
 194
 195	newcnt = __this_cpu_sub_return(softirq_ctrl.cnt, cnt);
 196	current->softirq_disable_cnt = newcnt;
 197
 198	if (!newcnt && unlock) {
 199		rcu_read_unlock();
 200		local_unlock(&softirq_ctrl.lock);
 201	}
 202}
 203
 204void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
 205{
 206	bool preempt_on = preemptible();
 207	unsigned long flags;
 208	u32 pending;
 209	int curcnt;
 210
 211	WARN_ON_ONCE(in_hardirq());
 212	lockdep_assert_irqs_enabled();
 213
 214	local_irq_save(flags);
 215	curcnt = __this_cpu_read(softirq_ctrl.cnt);
 216
 217	/*
 218	 * If this is not reenabling soft interrupts, no point in trying to
 219	 * run pending ones.
 220	 */
 221	if (curcnt != cnt)
 222		goto out;
 223
 224	pending = local_softirq_pending();
 225	if (!pending)
 226		goto out;
 227
 228	/*
 229	 * If this was called from non preemptible context, wake up the
 230	 * softirq daemon.
 231	 */
 232	if (!preempt_on) {
 233		wakeup_softirqd();
 234		goto out;
 235	}
 236
 237	/*
 238	 * Adjust softirq count to SOFTIRQ_OFFSET which makes
 239	 * in_serving_softirq() become true.
 240	 */
 241	cnt = SOFTIRQ_OFFSET;
 242	__local_bh_enable(cnt, false);
 243	__do_softirq();
 244
 245out:
 246	__local_bh_enable(cnt, preempt_on);
 247	local_irq_restore(flags);
 248}
 249EXPORT_SYMBOL(__local_bh_enable_ip);
 250
 251/*
 252 * Invoked from ksoftirqd_run() outside of the interrupt disabled section
 253 * to acquire the per CPU local lock for reentrancy protection.
 254 */
 255static inline void ksoftirqd_run_begin(void)
 256{
 257	__local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
 258	local_irq_disable();
 259}
 260
 261/* Counterpart to ksoftirqd_run_begin() */
 262static inline void ksoftirqd_run_end(void)
 263{
 264	__local_bh_enable(SOFTIRQ_OFFSET, true);
 265	WARN_ON_ONCE(in_interrupt());
 266	local_irq_enable();
 267}
 268
 269static inline void softirq_handle_begin(void) { }
 270static inline void softirq_handle_end(void) { }
 271
 272static inline bool should_wake_ksoftirqd(void)
 273{
 274	return !this_cpu_read(softirq_ctrl.cnt);
 275}
 276
 277static inline void invoke_softirq(void)
 278{
 279	if (should_wake_ksoftirqd())
 280		wakeup_softirqd();
 281}
 282
 283/*
 284 * flush_smp_call_function_queue() can raise a soft interrupt in a function
 285 * call. On RT kernels this is undesired and the only known functionality
 286 * in the block layer which does this is disabled on RT. If soft interrupts
 287 * get raised which haven't been raised before the flush, warn so it can be
 288 * investigated.
 289 */
 290void do_softirq_post_smp_call_flush(unsigned int was_pending)
 291{
 292	if (WARN_ON_ONCE(was_pending != local_softirq_pending()))
 293		invoke_softirq();
 294}
 295
 296#else /* CONFIG_PREEMPT_RT */
 297
 298/*
 299 * This one is for softirq.c-internal use, where hardirqs are disabled
 300 * legitimately:
 301 */
 302#ifdef CONFIG_TRACE_IRQFLAGS
 303void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
 304{
 305	unsigned long flags;
 306
 307	WARN_ON_ONCE(in_hardirq());
 308
 309	raw_local_irq_save(flags);
 310	/*
 311	 * The preempt tracer hooks into preempt_count_add and will break
 312	 * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
 313	 * is set and before current->softirq_enabled is cleared.
 314	 * We must manually increment preempt_count here and manually
 315	 * call the trace_preempt_off later.
 316	 */
 317	__preempt_count_add(cnt);
 318	/*
 319	 * Were softirqs turned off above:
 320	 */
 321	if (softirq_count() == (cnt & SOFTIRQ_MASK))
 322		lockdep_softirqs_off(ip);
 323	raw_local_irq_restore(flags);
 324
 325	if (preempt_count() == cnt) {
 326#ifdef CONFIG_DEBUG_PREEMPT
 327		current->preempt_disable_ip = get_lock_parent_ip();
 328#endif
 329		trace_preempt_off(CALLER_ADDR0, get_lock_parent_ip());
 330	}
 331}
 332EXPORT_SYMBOL(__local_bh_disable_ip);
 333#endif /* CONFIG_TRACE_IRQFLAGS */
 334
 335static void __local_bh_enable(unsigned int cnt)
 336{
 337	lockdep_assert_irqs_disabled();
 338
 339	if (preempt_count() == cnt)
 340		trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
 341
 342	if (softirq_count() == (cnt & SOFTIRQ_MASK))
 343		lockdep_softirqs_on(_RET_IP_);
 344
 345	__preempt_count_sub(cnt);
 346}
 347
 348/*
 349 * Special-case - softirqs can safely be enabled by __do_softirq(),
 350 * without processing still-pending softirqs:
 351 */
 352void _local_bh_enable(void)
 353{
 354	WARN_ON_ONCE(in_hardirq());
 355	__local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
 356}
 357EXPORT_SYMBOL(_local_bh_enable);
 358
 359void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
 360{
 361	WARN_ON_ONCE(in_hardirq());
 362	lockdep_assert_irqs_enabled();
 363#ifdef CONFIG_TRACE_IRQFLAGS
 364	local_irq_disable();
 365#endif
 366	/*
 367	 * Are softirqs going to be turned on now:
 368	 */
 369	if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
 370		lockdep_softirqs_on(ip);
 371	/*
 372	 * Keep preemption disabled until we are done with
 373	 * softirq processing:
 374	 */
 375	__preempt_count_sub(cnt - 1);
 376
 377	if (unlikely(!in_interrupt() && local_softirq_pending())) {
 378		/*
 379		 * Run softirq if any pending. And do it in its own stack
 380		 * as we may be calling this deep in a task call stack already.
 381		 */
 382		do_softirq();
 383	}
 384
 385	preempt_count_dec();
 386#ifdef CONFIG_TRACE_IRQFLAGS
 387	local_irq_enable();
 388#endif
 389	preempt_check_resched();
 390}
 391EXPORT_SYMBOL(__local_bh_enable_ip);
 392
 393static inline void softirq_handle_begin(void)
 394{
 395	__local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
 396}
 397
 398static inline void softirq_handle_end(void)
 399{
 400	__local_bh_enable(SOFTIRQ_OFFSET);
 401	WARN_ON_ONCE(in_interrupt());
 402}
 403
 404static inline void ksoftirqd_run_begin(void)
 405{
 406	local_irq_disable();
 407}
 408
 409static inline void ksoftirqd_run_end(void)
 410{
 411	local_irq_enable();
 412}
 413
 414static inline bool should_wake_ksoftirqd(void)
 415{
 416	return true;
 417}
 418
 419static inline void invoke_softirq(void)
 420{
 421	if (!force_irqthreads() || !__this_cpu_read(ksoftirqd)) {
 422#ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
 423		/*
 424		 * We can safely execute softirq on the current stack if
 425		 * it is the irq stack, because it should be near empty
 426		 * at this stage.
 427		 */
 428		__do_softirq();
 429#else
 430		/*
 431		 * Otherwise, irq_exit() is called on the task stack that can
 432		 * be potentially deep already. So call softirq in its own stack
 433		 * to prevent from any overrun.
 434		 */
 435		do_softirq_own_stack();
 436#endif
 437	} else {
 438		wakeup_softirqd();
 439	}
 440}
 441
 442asmlinkage __visible void do_softirq(void)
 443{
 444	__u32 pending;
 445	unsigned long flags;
 446
 447	if (in_interrupt())
 448		return;
 449
 450	local_irq_save(flags);
 451
 452	pending = local_softirq_pending();
 453
 454	if (pending)
 455		do_softirq_own_stack();
 456
 457	local_irq_restore(flags);
 458}
 459
 460#endif /* !CONFIG_PREEMPT_RT */
 461
 462/*
 463 * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times,
 464 * but break the loop if need_resched() is set or after 2 ms.
 465 * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in
 466 * certain cases, such as stop_machine(), jiffies may cease to
 467 * increment and so we need the MAX_SOFTIRQ_RESTART limit as
 468 * well to make sure we eventually return from this method.
 469 *
 470 * These limits have been established via experimentation.
 471 * The two things to balance is latency against fairness -
 472 * we want to handle softirqs as soon as possible, but they
 473 * should not be able to lock up the box.
 474 */
 475#define MAX_SOFTIRQ_TIME  msecs_to_jiffies(2)
 476#define MAX_SOFTIRQ_RESTART 10
 477
 478#ifdef CONFIG_TRACE_IRQFLAGS
 479/*
 480 * When we run softirqs from irq_exit() and thus on the hardirq stack we need
 481 * to keep the lockdep irq context tracking as tight as possible in order to
 482 * not miss-qualify lock contexts and miss possible deadlocks.
 483 */
 484
 485static inline bool lockdep_softirq_start(void)
 486{
 487	bool in_hardirq = false;
 488
 489	if (lockdep_hardirq_context()) {
 490		in_hardirq = true;
 491		lockdep_hardirq_exit();
 492	}
 493
 494	lockdep_softirq_enter();
 495
 496	return in_hardirq;
 497}
 498
 499static inline void lockdep_softirq_end(bool in_hardirq)
 500{
 501	lockdep_softirq_exit();
 502
 503	if (in_hardirq)
 504		lockdep_hardirq_enter();
 505}
 506#else
 507static inline bool lockdep_softirq_start(void) { return false; }
 508static inline void lockdep_softirq_end(bool in_hardirq) { }
 509#endif
 510
 511static void handle_softirqs(bool ksirqd)
 512{
 513	unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
 514	unsigned long old_flags = current->flags;
 515	int max_restart = MAX_SOFTIRQ_RESTART;
 516	struct softirq_action *h;
 517	bool in_hardirq;
 518	__u32 pending;
 519	int softirq_bit;
 520
 521	/*
 522	 * Mask out PF_MEMALLOC as the current task context is borrowed for the
 523	 * softirq. A softirq handled, such as network RX, might set PF_MEMALLOC
 524	 * again if the socket is related to swapping.
 525	 */
 526	current->flags &= ~PF_MEMALLOC;
 527
 528	pending = local_softirq_pending();
 
 529
 530	softirq_handle_begin();
 531	in_hardirq = lockdep_softirq_start();
 532	account_softirq_enter(current);
 533
 534restart:
 535	/* Reset the pending bitmask before enabling irqs */
 536	set_softirq_pending(0);
 537
 538	local_irq_enable();
 539
 540	h = softirq_vec;
 541
 542	while ((softirq_bit = ffs(pending))) {
 543		unsigned int vec_nr;
 544		int prev_count;
 545
 546		h += softirq_bit - 1;
 547
 548		vec_nr = h - softirq_vec;
 549		prev_count = preempt_count();
 550
 551		kstat_incr_softirqs_this_cpu(vec_nr);
 552
 553		trace_softirq_entry(vec_nr);
 554		h->action(h);
 555		trace_softirq_exit(vec_nr);
 556		if (unlikely(prev_count != preempt_count())) {
 557			pr_err("huh, entered softirq %u %s %p with preempt_count %08x, exited with %08x?\n",
 558			       vec_nr, softirq_to_name[vec_nr], h->action,
 559			       prev_count, preempt_count());
 560			preempt_count_set(prev_count);
 561		}
 562		h++;
 563		pending >>= softirq_bit;
 564	}
 565
 566	if (!IS_ENABLED(CONFIG_PREEMPT_RT) && ksirqd)
 567		rcu_softirq_qs();
 568
 569	local_irq_disable();
 570
 571	pending = local_softirq_pending();
 572	if (pending) {
 573		if (time_before(jiffies, end) && !need_resched() &&
 574		    --max_restart)
 575			goto restart;
 576
 577		wakeup_softirqd();
 578	}
 579
 580	account_softirq_exit(current);
 581	lockdep_softirq_end(in_hardirq);
 582	softirq_handle_end();
 
 
 583	current_restore_flags(old_flags, PF_MEMALLOC);
 584}
 585
 586asmlinkage __visible void __softirq_entry __do_softirq(void)
 587{
 588	handle_softirqs(false);
 
 
 
 
 
 
 
 
 
 
 
 
 
 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 (tick_nohz_full_cpu(smp_processor_id()) ||
 599	    (is_idle_task(current) && (irq_count() == HARDIRQ_OFFSET)))
 
 
 
 600		tick_irq_enter();
 
 
 601
 602	account_hardirq_enter(current);
 603}
 604
 605/**
 606 * irq_enter - Enter an interrupt context including RCU update
 607 */
 608void irq_enter(void)
 609{
 610	ct_irq_enter();
 611	irq_enter_rcu();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 612}
 613
 614static inline void tick_irq_exit(void)
 615{
 616#ifdef CONFIG_NO_HZ_COMMON
 617	int cpu = smp_processor_id();
 618
 619	/* Make sure that timer wheel updates are propagated */
 620	if ((sched_core_idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
 621		if (!in_hardirq())
 622			tick_nohz_irq_exit();
 623	}
 624#endif
 625}
 626
 627static inline void __irq_exit_rcu(void)
 
 
 
 628{
 629#ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
 630	local_irq_disable();
 631#else
 632	lockdep_assert_irqs_disabled();
 633#endif
 634	account_hardirq_exit(current);
 635	preempt_count_sub(HARDIRQ_OFFSET);
 636	if (!in_interrupt() && local_softirq_pending())
 637		invoke_softirq();
 638
 639	tick_irq_exit();
 640}
 641
 642/**
 643 * irq_exit_rcu() - Exit an interrupt context without updating RCU
 644 *
 645 * Also processes softirqs if needed and possible.
 646 */
 647void irq_exit_rcu(void)
 648{
 649	__irq_exit_rcu();
 650	 /* must be last! */
 651	lockdep_hardirq_exit();
 652}
 653
 654/**
 655 * irq_exit - Exit an interrupt context, update RCU and lockdep
 656 *
 657 * Also processes softirqs if needed and possible.
 658 */
 659void irq_exit(void)
 660{
 661	__irq_exit_rcu();
 662	ct_irq_exit();
 663	 /* must be last! */
 664	lockdep_hardirq_exit();
 665}
 666
 667/*
 668 * This function must run with irqs disabled!
 669 */
 670inline void raise_softirq_irqoff(unsigned int nr)
 671{
 672	__raise_softirq_irqoff(nr);
 673
 674	/*
 675	 * If we're in an interrupt or softirq, we're done
 676	 * (this also catches softirq-disabled code). We will
 677	 * actually run the softirq once we return from
 678	 * the irq or softirq.
 679	 *
 680	 * Otherwise we wake up ksoftirqd to make sure we
 681	 * schedule the softirq soon.
 682	 */
 683	if (!in_interrupt() && should_wake_ksoftirqd())
 684		wakeup_softirqd();
 685}
 686
 687void raise_softirq(unsigned int nr)
 688{
 689	unsigned long flags;
 690
 691	local_irq_save(flags);
 692	raise_softirq_irqoff(nr);
 693	local_irq_restore(flags);
 694}
 695
 696void __raise_softirq_irqoff(unsigned int nr)
 697{
 698	lockdep_assert_irqs_disabled();
 699	trace_softirq_raise(nr);
 700	or_softirq_pending(1UL << nr);
 701}
 702
 703void open_softirq(int nr, void (*action)(struct softirq_action *))
 704{
 705	softirq_vec[nr].action = action;
 706}
 707
 708/*
 709 * Tasklets
 710 */
 711struct tasklet_head {
 712	struct tasklet_struct *head;
 713	struct tasklet_struct **tail;
 714};
 715
 716static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
 717static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
 718
 719static void __tasklet_schedule_common(struct tasklet_struct *t,
 720				      struct tasklet_head __percpu *headp,
 721				      unsigned int softirq_nr)
 722{
 723	struct tasklet_head *head;
 724	unsigned long flags;
 725
 726	local_irq_save(flags);
 727	head = this_cpu_ptr(headp);
 728	t->next = NULL;
 729	*head->tail = t;
 730	head->tail = &(t->next);
 731	raise_softirq_irqoff(softirq_nr);
 732	local_irq_restore(flags);
 733}
 734
 735void __tasklet_schedule(struct tasklet_struct *t)
 736{
 737	__tasklet_schedule_common(t, &tasklet_vec,
 738				  TASKLET_SOFTIRQ);
 739}
 740EXPORT_SYMBOL(__tasklet_schedule);
 741
 742void __tasklet_hi_schedule(struct tasklet_struct *t)
 743{
 744	__tasklet_schedule_common(t, &tasklet_hi_vec,
 745				  HI_SOFTIRQ);
 746}
 747EXPORT_SYMBOL(__tasklet_hi_schedule);
 748
 749static bool tasklet_clear_sched(struct tasklet_struct *t)
 750{
 751	if (test_and_clear_bit(TASKLET_STATE_SCHED, &t->state)) {
 752		wake_up_var(&t->state);
 753		return true;
 754	}
 755
 756	WARN_ONCE(1, "tasklet SCHED state not set: %s %pS\n",
 757		  t->use_callback ? "callback" : "func",
 758		  t->use_callback ? (void *)t->callback : (void *)t->func);
 759
 760	return false;
 761}
 762
 763static void tasklet_action_common(struct softirq_action *a,
 764				  struct tasklet_head *tl_head,
 765				  unsigned int softirq_nr)
 766{
 767	struct tasklet_struct *list;
 768
 769	local_irq_disable();
 770	list = tl_head->head;
 771	tl_head->head = NULL;
 772	tl_head->tail = &tl_head->head;
 773	local_irq_enable();
 774
 775	while (list) {
 776		struct tasklet_struct *t = list;
 777
 778		list = list->next;
 779
 780		if (tasklet_trylock(t)) {
 781			if (!atomic_read(&t->count)) {
 782				if (tasklet_clear_sched(t)) {
 783					if (t->use_callback) {
 784						trace_tasklet_entry(t, t->callback);
 785						t->callback(t);
 786						trace_tasklet_exit(t, t->callback);
 787					} else {
 788						trace_tasklet_entry(t, t->func);
 789						t->func(t->data);
 790						trace_tasklet_exit(t, t->func);
 791					}
 792				}
 793				tasklet_unlock(t);
 794				continue;
 795			}
 796			tasklet_unlock(t);
 797		}
 798
 799		local_irq_disable();
 800		t->next = NULL;
 801		*tl_head->tail = t;
 802		tl_head->tail = &t->next;
 803		__raise_softirq_irqoff(softirq_nr);
 804		local_irq_enable();
 805	}
 806}
 807
 808static __latent_entropy void tasklet_action(struct softirq_action *a)
 809{
 810	workqueue_softirq_action(false);
 811	tasklet_action_common(a, this_cpu_ptr(&tasklet_vec), TASKLET_SOFTIRQ);
 812}
 813
 814static __latent_entropy void tasklet_hi_action(struct softirq_action *a)
 815{
 816	workqueue_softirq_action(true);
 817	tasklet_action_common(a, this_cpu_ptr(&tasklet_hi_vec), HI_SOFTIRQ);
 818}
 819
 820void tasklet_setup(struct tasklet_struct *t,
 821		   void (*callback)(struct tasklet_struct *))
 822{
 823	t->next = NULL;
 824	t->state = 0;
 825	atomic_set(&t->count, 0);
 826	t->callback = callback;
 827	t->use_callback = true;
 828	t->data = 0;
 829}
 830EXPORT_SYMBOL(tasklet_setup);
 831
 832void tasklet_init(struct tasklet_struct *t,
 833		  void (*func)(unsigned long), unsigned long data)
 834{
 835	t->next = NULL;
 836	t->state = 0;
 837	atomic_set(&t->count, 0);
 838	t->func = func;
 839	t->use_callback = false;
 840	t->data = data;
 841}
 842EXPORT_SYMBOL(tasklet_init);
 843
 844#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
 845/*
 846 * Do not use in new code. Waiting for tasklets from atomic contexts is
 847 * error prone and should be avoided.
 848 */
 849void tasklet_unlock_spin_wait(struct tasklet_struct *t)
 850{
 851	while (test_bit(TASKLET_STATE_RUN, &(t)->state)) {
 852		if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
 853			/*
 854			 * Prevent a live lock when current preempted soft
 855			 * interrupt processing or prevents ksoftirqd from
 856			 * running. If the tasklet runs on a different CPU
 857			 * then this has no effect other than doing the BH
 858			 * disable/enable dance for nothing.
 859			 */
 860			local_bh_disable();
 861			local_bh_enable();
 862		} else {
 863			cpu_relax();
 864		}
 865	}
 866}
 867EXPORT_SYMBOL(tasklet_unlock_spin_wait);
 868#endif
 869
 870void tasklet_kill(struct tasklet_struct *t)
 871{
 872	if (in_interrupt())
 873		pr_notice("Attempt to kill tasklet from interrupt\n");
 874
 875	while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state))
 876		wait_var_event(&t->state, !test_bit(TASKLET_STATE_SCHED, &t->state));
 877
 
 
 878	tasklet_unlock_wait(t);
 879	tasklet_clear_sched(t);
 880}
 881EXPORT_SYMBOL(tasklet_kill);
 882
 883#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
 884void tasklet_unlock(struct tasklet_struct *t)
 885{
 886	smp_mb__before_atomic();
 887	clear_bit(TASKLET_STATE_RUN, &t->state);
 888	smp_mb__after_atomic();
 889	wake_up_var(&t->state);
 890}
 891EXPORT_SYMBOL_GPL(tasklet_unlock);
 892
 893void tasklet_unlock_wait(struct tasklet_struct *t)
 894{
 895	wait_var_event(&t->state, !test_bit(TASKLET_STATE_RUN, &t->state));
 896}
 897EXPORT_SYMBOL_GPL(tasklet_unlock_wait);
 898#endif
 899
 900void __init softirq_init(void)
 901{
 902	int cpu;
 903
 904	for_each_possible_cpu(cpu) {
 905		per_cpu(tasklet_vec, cpu).tail =
 906			&per_cpu(tasklet_vec, cpu).head;
 907		per_cpu(tasklet_hi_vec, cpu).tail =
 908			&per_cpu(tasklet_hi_vec, cpu).head;
 909	}
 910
 911	open_softirq(TASKLET_SOFTIRQ, tasklet_action);
 912	open_softirq(HI_SOFTIRQ, tasklet_hi_action);
 913}
 914
 915static int ksoftirqd_should_run(unsigned int cpu)
 916{
 917	return local_softirq_pending();
 918}
 919
 920static void run_ksoftirqd(unsigned int cpu)
 921{
 922	ksoftirqd_run_begin();
 923	if (local_softirq_pending()) {
 924		/*
 925		 * We can safely run softirq on inline stack, as we are not deep
 926		 * in the task stack here.
 927		 */
 928		handle_softirqs(true);
 929		ksoftirqd_run_end();
 930		cond_resched();
 931		return;
 932	}
 933	ksoftirqd_run_end();
 934}
 935
 936#ifdef CONFIG_HOTPLUG_CPU
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 937static int takeover_tasklets(unsigned int cpu)
 938{
 939	workqueue_softirq_dead(cpu);
 940
 941	/* CPU is dead, so no lock needed. */
 942	local_irq_disable();
 943
 944	/* Find end, append list for that CPU. */
 945	if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
 946		*__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
 947		__this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
 948		per_cpu(tasklet_vec, cpu).head = NULL;
 949		per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
 950	}
 951	raise_softirq_irqoff(TASKLET_SOFTIRQ);
 952
 953	if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
 954		*__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
 955		__this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
 956		per_cpu(tasklet_hi_vec, cpu).head = NULL;
 957		per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
 958	}
 959	raise_softirq_irqoff(HI_SOFTIRQ);
 960
 961	local_irq_enable();
 962	return 0;
 963}
 964#else
 965#define takeover_tasklets	NULL
 966#endif /* CONFIG_HOTPLUG_CPU */
 967
 968static struct smp_hotplug_thread softirq_threads = {
 969	.store			= &ksoftirqd,
 970	.thread_should_run	= ksoftirqd_should_run,
 971	.thread_fn		= run_ksoftirqd,
 972	.thread_comm		= "ksoftirqd/%u",
 973};
 974
 975static __init int spawn_ksoftirqd(void)
 976{
 977	cpuhp_setup_state_nocalls(CPUHP_SOFTIRQ_DEAD, "softirq:dead", NULL,
 978				  takeover_tasklets);
 979	BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
 980
 981	return 0;
 982}
 983early_initcall(spawn_ksoftirqd);
 984
 985/*
 986 * [ These __weak aliases are kept in a separate compilation unit, so that
 987 *   GCC does not inline them incorrectly. ]
 988 */
 989
 990int __init __weak early_irq_init(void)
 991{
 992	return 0;
 993}
 994
 995int __init __weak arch_probe_nr_irqs(void)
 996{
 997	return NR_IRQS_LEGACY;
 998}
 999
1000int __init __weak arch_early_irq_init(void)
1001{
1002	return 0;
1003}
1004
1005unsigned int __weak arch_dynirq_lower_bound(unsigned int from)
1006{
1007	return from;
1008}