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