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