Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/kernel/softirq.c
4 *
5 * Copyright (C) 1992 Linus Torvalds
6 *
7 * Rewritten. Old one was good in 2.2, but in 2.3 it was immoral. --ANK (990903)
8 */
9
10#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
12#include <linux/export.h>
13#include <linux/kernel_stat.h>
14#include <linux/interrupt.h>
15#include <linux/init.h>
16#include <linux/local_lock.h>
17#include <linux/mm.h>
18#include <linux/notifier.h>
19#include <linux/percpu.h>
20#include <linux/cpu.h>
21#include <linux/freezer.h>
22#include <linux/kthread.h>
23#include <linux/rcupdate.h>
24#include <linux/ftrace.h>
25#include <linux/smp.h>
26#include <linux/smpboot.h>
27#include <linux/tick.h>
28#include <linux/irq.h>
29#include <linux/wait_bit.h>
30#include <linux/workqueue.h>
31
32#include <asm/softirq_stack.h>
33
34#define CREATE_TRACE_POINTS
35#include <trace/events/irq.h>
36
37/*
38 - No shared variables, all the data are CPU local.
39 - If a softirq needs serialization, let it serialize itself
40 by its own spinlocks.
41 - Even if softirq is serialized, only local cpu is marked for
42 execution. Hence, we get something sort of weak cpu binding.
43 Though it is still not clear, will it result in better locality
44 or will not.
45
46 Examples:
47 - NET RX softirq. It is multithreaded and does not require
48 any global serialization.
49 - NET TX softirq. It kicks software netdevice queues, hence
50 it is logically serialized per device, but this serialization
51 is invisible to common code.
52 - Tasklets: serialized wrt itself.
53 */
54
55#ifndef __ARCH_IRQ_STAT
56DEFINE_PER_CPU_ALIGNED(irq_cpustat_t, irq_stat);
57EXPORT_PER_CPU_SYMBOL(irq_stat);
58#endif
59
60static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
61
62DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
63
64const char * const softirq_to_name[NR_SOFTIRQS] = {
65 "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "IRQ_POLL",
66 "TASKLET", "SCHED", "HRTIMER", "RCU"
67};
68
69/*
70 * we cannot loop indefinitely here to avoid userspace starvation,
71 * but we also don't want to introduce a worst case 1/HZ latency
72 * to the pending events, so lets the scheduler to balance
73 * the softirq load for us.
74 */
75static void wakeup_softirqd(void)
76{
77 /* Interrupts are disabled: no need to stop preemption */
78 struct task_struct *tsk = __this_cpu_read(ksoftirqd);
79
80 if (tsk)
81 wake_up_process(tsk);
82}
83
84#ifdef CONFIG_TRACE_IRQFLAGS
85DEFINE_PER_CPU(int, hardirqs_enabled);
86DEFINE_PER_CPU(int, hardirq_context);
87EXPORT_PER_CPU_SYMBOL_GPL(hardirqs_enabled);
88EXPORT_PER_CPU_SYMBOL_GPL(hardirq_context);
89#endif
90
91/*
92 * SOFTIRQ_OFFSET usage:
93 *
94 * On !RT kernels 'count' is the preempt counter, on RT kernels this applies
95 * to a per CPU counter and to task::softirqs_disabled_cnt.
96 *
97 * - count is changed by SOFTIRQ_OFFSET on entering or leaving softirq
98 * processing.
99 *
100 * - count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
101 * on local_bh_disable or local_bh_enable.
102 *
103 * This lets us distinguish between whether we are currently processing
104 * softirq and whether we just have bh disabled.
105 */
106#ifdef CONFIG_PREEMPT_RT
107
108/*
109 * RT accounts for BH disabled sections in task::softirqs_disabled_cnt and
110 * also in per CPU softirq_ctrl::cnt. This is necessary to allow tasks in a
111 * softirq disabled section to be preempted.
112 *
113 * The per task counter is used for softirq_count(), in_softirq() and
114 * in_serving_softirqs() because these counts are only valid when the task
115 * holding softirq_ctrl::lock is running.
116 *
117 * The per CPU counter prevents pointless wakeups of ksoftirqd in case that
118 * the task which is in a softirq disabled section is preempted or blocks.
119 */
120struct softirq_ctrl {
121 local_lock_t lock;
122 int cnt;
123};
124
125static DEFINE_PER_CPU(struct softirq_ctrl, softirq_ctrl) = {
126 .lock = INIT_LOCAL_LOCK(softirq_ctrl.lock),
127};
128
129/**
130 * local_bh_blocked() - Check for idle whether BH processing is blocked
131 *
132 * Returns false if the per CPU softirq::cnt is 0 otherwise true.
133 *
134 * This is invoked from the idle task to guard against false positive
135 * softirq pending warnings, which would happen when the task which holds
136 * softirq_ctrl::lock was the only running task on the CPU and blocks on
137 * some other lock.
138 */
139bool local_bh_blocked(void)
140{
141 return __this_cpu_read(softirq_ctrl.cnt) != 0;
142}
143
144void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
145{
146 unsigned long flags;
147 int newcnt;
148
149 WARN_ON_ONCE(in_hardirq());
150
151 /* First entry of a task into a BH disabled section? */
152 if (!current->softirq_disable_cnt) {
153 if (preemptible()) {
154 local_lock(&softirq_ctrl.lock);
155 /* Required to meet the RCU bottomhalf requirements. */
156 rcu_read_lock();
157 } else {
158 DEBUG_LOCKS_WARN_ON(this_cpu_read(softirq_ctrl.cnt));
159 }
160 }
161
162 /*
163 * Track the per CPU softirq disabled state. On RT this is per CPU
164 * state to allow preemption of bottom half disabled sections.
165 */
166 newcnt = __this_cpu_add_return(softirq_ctrl.cnt, cnt);
167 /*
168 * Reflect the result in the task state to prevent recursion on the
169 * local lock and to make softirq_count() & al work.
170 */
171 current->softirq_disable_cnt = newcnt;
172
173 if (IS_ENABLED(CONFIG_TRACE_IRQFLAGS) && newcnt == cnt) {
174 raw_local_irq_save(flags);
175 lockdep_softirqs_off(ip);
176 raw_local_irq_restore(flags);
177 }
178}
179EXPORT_SYMBOL(__local_bh_disable_ip);
180
181static void __local_bh_enable(unsigned int cnt, bool unlock)
182{
183 unsigned long flags;
184 int newcnt;
185
186 DEBUG_LOCKS_WARN_ON(current->softirq_disable_cnt !=
187 this_cpu_read(softirq_ctrl.cnt));
188
189 if (IS_ENABLED(CONFIG_TRACE_IRQFLAGS) && softirq_count() == cnt) {
190 raw_local_irq_save(flags);
191 lockdep_softirqs_on(_RET_IP_);
192 raw_local_irq_restore(flags);
193 }
194
195 newcnt = __this_cpu_sub_return(softirq_ctrl.cnt, cnt);
196 current->softirq_disable_cnt = newcnt;
197
198 if (!newcnt && unlock) {
199 rcu_read_unlock();
200 local_unlock(&softirq_ctrl.lock);
201 }
202}
203
204void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
205{
206 bool preempt_on = preemptible();
207 unsigned long flags;
208 u32 pending;
209 int curcnt;
210
211 WARN_ON_ONCE(in_hardirq());
212 lockdep_assert_irqs_enabled();
213
214 local_irq_save(flags);
215 curcnt = __this_cpu_read(softirq_ctrl.cnt);
216
217 /*
218 * If this is not reenabling soft interrupts, no point in trying to
219 * run pending ones.
220 */
221 if (curcnt != cnt)
222 goto out;
223
224 pending = local_softirq_pending();
225 if (!pending)
226 goto out;
227
228 /*
229 * If this was called from non preemptible context, wake up the
230 * softirq daemon.
231 */
232 if (!preempt_on) {
233 wakeup_softirqd();
234 goto out;
235 }
236
237 /*
238 * Adjust softirq count to SOFTIRQ_OFFSET which makes
239 * in_serving_softirq() become true.
240 */
241 cnt = SOFTIRQ_OFFSET;
242 __local_bh_enable(cnt, false);
243 __do_softirq();
244
245out:
246 __local_bh_enable(cnt, preempt_on);
247 local_irq_restore(flags);
248}
249EXPORT_SYMBOL(__local_bh_enable_ip);
250
251/*
252 * Invoked from ksoftirqd_run() outside of the interrupt disabled section
253 * to acquire the per CPU local lock for reentrancy protection.
254 */
255static inline void ksoftirqd_run_begin(void)
256{
257 __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
258 local_irq_disable();
259}
260
261/* Counterpart to ksoftirqd_run_begin() */
262static inline void ksoftirqd_run_end(void)
263{
264 __local_bh_enable(SOFTIRQ_OFFSET, true);
265 WARN_ON_ONCE(in_interrupt());
266 local_irq_enable();
267}
268
269static inline void softirq_handle_begin(void) { }
270static inline void softirq_handle_end(void) { }
271
272static inline bool should_wake_ksoftirqd(void)
273{
274 return !this_cpu_read(softirq_ctrl.cnt);
275}
276
277static inline void invoke_softirq(void)
278{
279 if (should_wake_ksoftirqd())
280 wakeup_softirqd();
281}
282
283#define SCHED_SOFTIRQ_MASK BIT(SCHED_SOFTIRQ)
284
285/*
286 * flush_smp_call_function_queue() can raise a soft interrupt in a function
287 * call. On RT kernels this is undesired and the only known functionalities
288 * are in the block layer which is disabled on RT, and in the scheduler for
289 * idle load balancing. If soft interrupts get raised which haven't been
290 * raised before the flush, warn if it is not a SCHED_SOFTIRQ so it can be
291 * investigated.
292 */
293void do_softirq_post_smp_call_flush(unsigned int was_pending)
294{
295 unsigned int is_pending = local_softirq_pending();
296
297 if (unlikely(was_pending != is_pending)) {
298 WARN_ON_ONCE(was_pending != (is_pending & ~SCHED_SOFTIRQ_MASK));
299 invoke_softirq();
300 }
301}
302
303#else /* CONFIG_PREEMPT_RT */
304
305/*
306 * This one is for softirq.c-internal use, where hardirqs are disabled
307 * legitimately:
308 */
309#ifdef CONFIG_TRACE_IRQFLAGS
310void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
311{
312 unsigned long flags;
313
314 WARN_ON_ONCE(in_hardirq());
315
316 raw_local_irq_save(flags);
317 /*
318 * The preempt tracer hooks into preempt_count_add and will break
319 * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
320 * is set and before current->softirq_enabled is cleared.
321 * We must manually increment preempt_count here and manually
322 * call the trace_preempt_off later.
323 */
324 __preempt_count_add(cnt);
325 /*
326 * Were softirqs turned off above:
327 */
328 if (softirq_count() == (cnt & SOFTIRQ_MASK))
329 lockdep_softirqs_off(ip);
330 raw_local_irq_restore(flags);
331
332 if (preempt_count() == cnt) {
333#ifdef CONFIG_DEBUG_PREEMPT
334 current->preempt_disable_ip = get_lock_parent_ip();
335#endif
336 trace_preempt_off(CALLER_ADDR0, get_lock_parent_ip());
337 }
338}
339EXPORT_SYMBOL(__local_bh_disable_ip);
340#endif /* CONFIG_TRACE_IRQFLAGS */
341
342static void __local_bh_enable(unsigned int cnt)
343{
344 lockdep_assert_irqs_disabled();
345
346 if (preempt_count() == cnt)
347 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
348
349 if (softirq_count() == (cnt & SOFTIRQ_MASK))
350 lockdep_softirqs_on(_RET_IP_);
351
352 __preempt_count_sub(cnt);
353}
354
355/*
356 * Special-case - softirqs can safely be enabled by __do_softirq(),
357 * without processing still-pending softirqs:
358 */
359void _local_bh_enable(void)
360{
361 WARN_ON_ONCE(in_hardirq());
362 __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
363}
364EXPORT_SYMBOL(_local_bh_enable);
365
366void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
367{
368 WARN_ON_ONCE(in_hardirq());
369 lockdep_assert_irqs_enabled();
370#ifdef CONFIG_TRACE_IRQFLAGS
371 local_irq_disable();
372#endif
373 /*
374 * Are softirqs going to be turned on now:
375 */
376 if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
377 lockdep_softirqs_on(ip);
378 /*
379 * Keep preemption disabled until we are done with
380 * softirq processing:
381 */
382 __preempt_count_sub(cnt - 1);
383
384 if (unlikely(!in_interrupt() && local_softirq_pending())) {
385 /*
386 * Run softirq if any pending. And do it in its own stack
387 * as we may be calling this deep in a task call stack already.
388 */
389 do_softirq();
390 }
391
392 preempt_count_dec();
393#ifdef CONFIG_TRACE_IRQFLAGS
394 local_irq_enable();
395#endif
396 preempt_check_resched();
397}
398EXPORT_SYMBOL(__local_bh_enable_ip);
399
400static inline void softirq_handle_begin(void)
401{
402 __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
403}
404
405static inline void softirq_handle_end(void)
406{
407 __local_bh_enable(SOFTIRQ_OFFSET);
408 WARN_ON_ONCE(in_interrupt());
409}
410
411static inline void ksoftirqd_run_begin(void)
412{
413 local_irq_disable();
414}
415
416static inline void ksoftirqd_run_end(void)
417{
418 local_irq_enable();
419}
420
421static inline bool should_wake_ksoftirqd(void)
422{
423 return true;
424}
425
426static inline void invoke_softirq(void)
427{
428 if (!force_irqthreads() || !__this_cpu_read(ksoftirqd)) {
429#ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
430 /*
431 * We can safely execute softirq on the current stack if
432 * it is the irq stack, because it should be near empty
433 * at this stage.
434 */
435 __do_softirq();
436#else
437 /*
438 * Otherwise, irq_exit() is called on the task stack that can
439 * be potentially deep already. So call softirq in its own stack
440 * to prevent from any overrun.
441 */
442 do_softirq_own_stack();
443#endif
444 } else {
445 wakeup_softirqd();
446 }
447}
448
449asmlinkage __visible void do_softirq(void)
450{
451 __u32 pending;
452 unsigned long flags;
453
454 if (in_interrupt())
455 return;
456
457 local_irq_save(flags);
458
459 pending = local_softirq_pending();
460
461 if (pending)
462 do_softirq_own_stack();
463
464 local_irq_restore(flags);
465}
466
467#endif /* !CONFIG_PREEMPT_RT */
468
469/*
470 * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times,
471 * but break the loop if need_resched() is set or after 2 ms.
472 * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in
473 * certain cases, such as stop_machine(), jiffies may cease to
474 * increment and so we need the MAX_SOFTIRQ_RESTART limit as
475 * well to make sure we eventually return from this method.
476 *
477 * These limits have been established via experimentation.
478 * The two things to balance is latency against fairness -
479 * we want to handle softirqs as soon as possible, but they
480 * should not be able to lock up the box.
481 */
482#define MAX_SOFTIRQ_TIME msecs_to_jiffies(2)
483#define MAX_SOFTIRQ_RESTART 10
484
485#ifdef CONFIG_TRACE_IRQFLAGS
486/*
487 * When we run softirqs from irq_exit() and thus on the hardirq stack we need
488 * to keep the lockdep irq context tracking as tight as possible in order to
489 * not miss-qualify lock contexts and miss possible deadlocks.
490 */
491
492static inline bool lockdep_softirq_start(void)
493{
494 bool in_hardirq = false;
495
496 if (lockdep_hardirq_context()) {
497 in_hardirq = true;
498 lockdep_hardirq_exit();
499 }
500
501 lockdep_softirq_enter();
502
503 return in_hardirq;
504}
505
506static inline void lockdep_softirq_end(bool in_hardirq)
507{
508 lockdep_softirq_exit();
509
510 if (in_hardirq)
511 lockdep_hardirq_enter();
512}
513#else
514static inline bool lockdep_softirq_start(void) { return false; }
515static inline void lockdep_softirq_end(bool in_hardirq) { }
516#endif
517
518static void handle_softirqs(bool ksirqd)
519{
520 unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
521 unsigned long old_flags = current->flags;
522 int max_restart = MAX_SOFTIRQ_RESTART;
523 struct softirq_action *h;
524 bool in_hardirq;
525 __u32 pending;
526 int softirq_bit;
527
528 /*
529 * Mask out PF_MEMALLOC as the current task context is borrowed for the
530 * softirq. A softirq handled, such as network RX, might set PF_MEMALLOC
531 * again if the socket is related to swapping.
532 */
533 current->flags &= ~PF_MEMALLOC;
534
535 pending = local_softirq_pending();
536
537 softirq_handle_begin();
538 in_hardirq = lockdep_softirq_start();
539 account_softirq_enter(current);
540
541restart:
542 /* Reset the pending bitmask before enabling irqs */
543 set_softirq_pending(0);
544
545 local_irq_enable();
546
547 h = softirq_vec;
548
549 while ((softirq_bit = ffs(pending))) {
550 unsigned int vec_nr;
551 int prev_count;
552
553 h += softirq_bit - 1;
554
555 vec_nr = h - softirq_vec;
556 prev_count = preempt_count();
557
558 kstat_incr_softirqs_this_cpu(vec_nr);
559
560 trace_softirq_entry(vec_nr);
561 h->action();
562 trace_softirq_exit(vec_nr);
563 if (unlikely(prev_count != preempt_count())) {
564 pr_err("huh, entered softirq %u %s %p with preempt_count %08x, exited with %08x?\n",
565 vec_nr, softirq_to_name[vec_nr], h->action,
566 prev_count, preempt_count());
567 preempt_count_set(prev_count);
568 }
569 h++;
570 pending >>= softirq_bit;
571 }
572
573 if (!IS_ENABLED(CONFIG_PREEMPT_RT) && ksirqd)
574 rcu_softirq_qs();
575
576 local_irq_disable();
577
578 pending = local_softirq_pending();
579 if (pending) {
580 if (time_before(jiffies, end) && !need_resched() &&
581 --max_restart)
582 goto restart;
583
584 wakeup_softirqd();
585 }
586
587 account_softirq_exit(current);
588 lockdep_softirq_end(in_hardirq);
589 softirq_handle_end();
590 current_restore_flags(old_flags, PF_MEMALLOC);
591}
592
593asmlinkage __visible void __softirq_entry __do_softirq(void)
594{
595 handle_softirqs(false);
596}
597
598/**
599 * irq_enter_rcu - Enter an interrupt context with RCU watching
600 */
601void irq_enter_rcu(void)
602{
603 __irq_enter_raw();
604
605 if (tick_nohz_full_cpu(smp_processor_id()) ||
606 (is_idle_task(current) && (irq_count() == HARDIRQ_OFFSET)))
607 tick_irq_enter();
608
609 account_hardirq_enter(current);
610}
611
612/**
613 * irq_enter - Enter an interrupt context including RCU update
614 */
615void irq_enter(void)
616{
617 ct_irq_enter();
618 irq_enter_rcu();
619}
620
621static inline void tick_irq_exit(void)
622{
623#ifdef CONFIG_NO_HZ_COMMON
624 int cpu = smp_processor_id();
625
626 /* Make sure that timer wheel updates are propagated */
627 if ((sched_core_idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
628 if (!in_hardirq())
629 tick_nohz_irq_exit();
630 }
631#endif
632}
633
634#ifdef CONFIG_IRQ_FORCED_THREADING
635DEFINE_PER_CPU(struct task_struct *, ktimerd);
636DEFINE_PER_CPU(unsigned long, pending_timer_softirq);
637
638static void wake_timersd(void)
639{
640 struct task_struct *tsk = __this_cpu_read(ktimerd);
641
642 if (tsk)
643 wake_up_process(tsk);
644}
645
646#else
647
648static inline void wake_timersd(void) { }
649
650#endif
651
652static inline void __irq_exit_rcu(void)
653{
654#ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
655 local_irq_disable();
656#else
657 lockdep_assert_irqs_disabled();
658#endif
659 account_hardirq_exit(current);
660 preempt_count_sub(HARDIRQ_OFFSET);
661 if (!in_interrupt() && local_softirq_pending())
662 invoke_softirq();
663
664 if (IS_ENABLED(CONFIG_IRQ_FORCED_THREADING) && force_irqthreads() &&
665 local_timers_pending_force_th() && !(in_nmi() | in_hardirq()))
666 wake_timersd();
667
668 tick_irq_exit();
669}
670
671/**
672 * irq_exit_rcu() - Exit an interrupt context without updating RCU
673 *
674 * Also processes softirqs if needed and possible.
675 */
676void irq_exit_rcu(void)
677{
678 __irq_exit_rcu();
679 /* must be last! */
680 lockdep_hardirq_exit();
681}
682
683/**
684 * irq_exit - Exit an interrupt context, update RCU and lockdep
685 *
686 * Also processes softirqs if needed and possible.
687 */
688void irq_exit(void)
689{
690 __irq_exit_rcu();
691 ct_irq_exit();
692 /* must be last! */
693 lockdep_hardirq_exit();
694}
695
696/*
697 * This function must run with irqs disabled!
698 */
699inline void raise_softirq_irqoff(unsigned int nr)
700{
701 __raise_softirq_irqoff(nr);
702
703 /*
704 * If we're in an interrupt or softirq, we're done
705 * (this also catches softirq-disabled code). We will
706 * actually run the softirq once we return from
707 * the irq or softirq.
708 *
709 * Otherwise we wake up ksoftirqd to make sure we
710 * schedule the softirq soon.
711 */
712 if (!in_interrupt() && should_wake_ksoftirqd())
713 wakeup_softirqd();
714}
715
716void raise_softirq(unsigned int nr)
717{
718 unsigned long flags;
719
720 local_irq_save(flags);
721 raise_softirq_irqoff(nr);
722 local_irq_restore(flags);
723}
724
725void __raise_softirq_irqoff(unsigned int nr)
726{
727 lockdep_assert_irqs_disabled();
728 trace_softirq_raise(nr);
729 or_softirq_pending(1UL << nr);
730}
731
732void open_softirq(int nr, void (*action)(void))
733{
734 softirq_vec[nr].action = action;
735}
736
737/*
738 * Tasklets
739 */
740struct tasklet_head {
741 struct tasklet_struct *head;
742 struct tasklet_struct **tail;
743};
744
745static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
746static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
747
748static void __tasklet_schedule_common(struct tasklet_struct *t,
749 struct tasklet_head __percpu *headp,
750 unsigned int softirq_nr)
751{
752 struct tasklet_head *head;
753 unsigned long flags;
754
755 local_irq_save(flags);
756 head = this_cpu_ptr(headp);
757 t->next = NULL;
758 *head->tail = t;
759 head->tail = &(t->next);
760 raise_softirq_irqoff(softirq_nr);
761 local_irq_restore(flags);
762}
763
764void __tasklet_schedule(struct tasklet_struct *t)
765{
766 __tasklet_schedule_common(t, &tasklet_vec,
767 TASKLET_SOFTIRQ);
768}
769EXPORT_SYMBOL(__tasklet_schedule);
770
771void __tasklet_hi_schedule(struct tasklet_struct *t)
772{
773 __tasklet_schedule_common(t, &tasklet_hi_vec,
774 HI_SOFTIRQ);
775}
776EXPORT_SYMBOL(__tasklet_hi_schedule);
777
778static bool tasklet_clear_sched(struct tasklet_struct *t)
779{
780 if (test_and_clear_wake_up_bit(TASKLET_STATE_SCHED, &t->state))
781 return true;
782
783 WARN_ONCE(1, "tasklet SCHED state not set: %s %pS\n",
784 t->use_callback ? "callback" : "func",
785 t->use_callback ? (void *)t->callback : (void *)t->func);
786
787 return false;
788}
789
790static void tasklet_action_common(struct tasklet_head *tl_head,
791 unsigned int softirq_nr)
792{
793 struct tasklet_struct *list;
794
795 local_irq_disable();
796 list = tl_head->head;
797 tl_head->head = NULL;
798 tl_head->tail = &tl_head->head;
799 local_irq_enable();
800
801 while (list) {
802 struct tasklet_struct *t = list;
803
804 list = list->next;
805
806 if (tasklet_trylock(t)) {
807 if (!atomic_read(&t->count)) {
808 if (tasklet_clear_sched(t)) {
809 if (t->use_callback) {
810 trace_tasklet_entry(t, t->callback);
811 t->callback(t);
812 trace_tasklet_exit(t, t->callback);
813 } else {
814 trace_tasklet_entry(t, t->func);
815 t->func(t->data);
816 trace_tasklet_exit(t, t->func);
817 }
818 }
819 tasklet_unlock(t);
820 continue;
821 }
822 tasklet_unlock(t);
823 }
824
825 local_irq_disable();
826 t->next = NULL;
827 *tl_head->tail = t;
828 tl_head->tail = &t->next;
829 __raise_softirq_irqoff(softirq_nr);
830 local_irq_enable();
831 }
832}
833
834static __latent_entropy void tasklet_action(void)
835{
836 workqueue_softirq_action(false);
837 tasklet_action_common(this_cpu_ptr(&tasklet_vec), TASKLET_SOFTIRQ);
838}
839
840static __latent_entropy void tasklet_hi_action(void)
841{
842 workqueue_softirq_action(true);
843 tasklet_action_common(this_cpu_ptr(&tasklet_hi_vec), HI_SOFTIRQ);
844}
845
846void tasklet_setup(struct tasklet_struct *t,
847 void (*callback)(struct tasklet_struct *))
848{
849 t->next = NULL;
850 t->state = 0;
851 atomic_set(&t->count, 0);
852 t->callback = callback;
853 t->use_callback = true;
854 t->data = 0;
855}
856EXPORT_SYMBOL(tasklet_setup);
857
858void tasklet_init(struct tasklet_struct *t,
859 void (*func)(unsigned long), unsigned long data)
860{
861 t->next = NULL;
862 t->state = 0;
863 atomic_set(&t->count, 0);
864 t->func = func;
865 t->use_callback = false;
866 t->data = data;
867}
868EXPORT_SYMBOL(tasklet_init);
869
870#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
871/*
872 * Do not use in new code. Waiting for tasklets from atomic contexts is
873 * error prone and should be avoided.
874 */
875void tasklet_unlock_spin_wait(struct tasklet_struct *t)
876{
877 while (test_bit(TASKLET_STATE_RUN, &(t)->state)) {
878 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
879 /*
880 * Prevent a live lock when current preempted soft
881 * interrupt processing or prevents ksoftirqd from
882 * running. If the tasklet runs on a different CPU
883 * then this has no effect other than doing the BH
884 * disable/enable dance for nothing.
885 */
886 local_bh_disable();
887 local_bh_enable();
888 } else {
889 cpu_relax();
890 }
891 }
892}
893EXPORT_SYMBOL(tasklet_unlock_spin_wait);
894#endif
895
896void tasklet_kill(struct tasklet_struct *t)
897{
898 if (in_interrupt())
899 pr_notice("Attempt to kill tasklet from interrupt\n");
900
901 wait_on_bit_lock(&t->state, TASKLET_STATE_SCHED, TASK_UNINTERRUPTIBLE);
902
903 tasklet_unlock_wait(t);
904 tasklet_clear_sched(t);
905}
906EXPORT_SYMBOL(tasklet_kill);
907
908#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
909void tasklet_unlock(struct tasklet_struct *t)
910{
911 clear_and_wake_up_bit(TASKLET_STATE_RUN, &t->state);
912}
913EXPORT_SYMBOL_GPL(tasklet_unlock);
914
915void tasklet_unlock_wait(struct tasklet_struct *t)
916{
917 wait_on_bit(&t->state, TASKLET_STATE_RUN, TASK_UNINTERRUPTIBLE);
918}
919EXPORT_SYMBOL_GPL(tasklet_unlock_wait);
920#endif
921
922void __init softirq_init(void)
923{
924 int cpu;
925
926 for_each_possible_cpu(cpu) {
927 per_cpu(tasklet_vec, cpu).tail =
928 &per_cpu(tasklet_vec, cpu).head;
929 per_cpu(tasklet_hi_vec, cpu).tail =
930 &per_cpu(tasklet_hi_vec, cpu).head;
931 }
932
933 open_softirq(TASKLET_SOFTIRQ, tasklet_action);
934 open_softirq(HI_SOFTIRQ, tasklet_hi_action);
935}
936
937static int ksoftirqd_should_run(unsigned int cpu)
938{
939 return local_softirq_pending();
940}
941
942static void run_ksoftirqd(unsigned int cpu)
943{
944 ksoftirqd_run_begin();
945 if (local_softirq_pending()) {
946 /*
947 * We can safely run softirq on inline stack, as we are not deep
948 * in the task stack here.
949 */
950 handle_softirqs(true);
951 ksoftirqd_run_end();
952 cond_resched();
953 return;
954 }
955 ksoftirqd_run_end();
956}
957
958#ifdef CONFIG_HOTPLUG_CPU
959static int takeover_tasklets(unsigned int cpu)
960{
961 workqueue_softirq_dead(cpu);
962
963 /* CPU is dead, so no lock needed. */
964 local_irq_disable();
965
966 /* Find end, append list for that CPU. */
967 if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
968 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
969 __this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
970 per_cpu(tasklet_vec, cpu).head = NULL;
971 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
972 }
973 raise_softirq_irqoff(TASKLET_SOFTIRQ);
974
975 if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
976 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
977 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
978 per_cpu(tasklet_hi_vec, cpu).head = NULL;
979 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
980 }
981 raise_softirq_irqoff(HI_SOFTIRQ);
982
983 local_irq_enable();
984 return 0;
985}
986#else
987#define takeover_tasklets NULL
988#endif /* CONFIG_HOTPLUG_CPU */
989
990static struct smp_hotplug_thread softirq_threads = {
991 .store = &ksoftirqd,
992 .thread_should_run = ksoftirqd_should_run,
993 .thread_fn = run_ksoftirqd,
994 .thread_comm = "ksoftirqd/%u",
995};
996
997#ifdef CONFIG_IRQ_FORCED_THREADING
998static void ktimerd_setup(unsigned int cpu)
999{
1000 /* Above SCHED_NORMAL to handle timers before regular tasks. */
1001 sched_set_fifo_low(current);
1002}
1003
1004static int ktimerd_should_run(unsigned int cpu)
1005{
1006 return local_timers_pending_force_th();
1007}
1008
1009void raise_ktimers_thread(unsigned int nr)
1010{
1011 trace_softirq_raise(nr);
1012 __this_cpu_or(pending_timer_softirq, BIT(nr));
1013}
1014
1015static void run_ktimerd(unsigned int cpu)
1016{
1017 unsigned int timer_si;
1018
1019 ksoftirqd_run_begin();
1020
1021 timer_si = local_timers_pending_force_th();
1022 __this_cpu_write(pending_timer_softirq, 0);
1023 or_softirq_pending(timer_si);
1024
1025 __do_softirq();
1026
1027 ksoftirqd_run_end();
1028}
1029
1030static struct smp_hotplug_thread timer_thread = {
1031 .store = &ktimerd,
1032 .setup = ktimerd_setup,
1033 .thread_should_run = ktimerd_should_run,
1034 .thread_fn = run_ktimerd,
1035 .thread_comm = "ktimers/%u",
1036};
1037#endif
1038
1039static __init int spawn_ksoftirqd(void)
1040{
1041 cpuhp_setup_state_nocalls(CPUHP_SOFTIRQ_DEAD, "softirq:dead", NULL,
1042 takeover_tasklets);
1043 BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
1044#ifdef CONFIG_IRQ_FORCED_THREADING
1045 if (force_irqthreads())
1046 BUG_ON(smpboot_register_percpu_thread(&timer_thread));
1047#endif
1048 return 0;
1049}
1050early_initcall(spawn_ksoftirqd);
1051
1052/*
1053 * [ These __weak aliases are kept in a separate compilation unit, so that
1054 * GCC does not inline them incorrectly. ]
1055 */
1056
1057int __init __weak early_irq_init(void)
1058{
1059 return 0;
1060}
1061
1062int __init __weak arch_probe_nr_irqs(void)
1063{
1064 return NR_IRQS_LEGACY;
1065}
1066
1067int __init __weak arch_early_irq_init(void)
1068{
1069 return 0;
1070}
1071
1072unsigned int __weak arch_dynirq_lower_bound(unsigned int from)
1073{
1074 return from;
1075}
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}