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