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/mm.h>
17#include <linux/notifier.h>
18#include <linux/percpu.h>
19#include <linux/cpu.h>
20#include <linux/freezer.h>
21#include <linux/kthread.h>
22#include <linux/rcupdate.h>
23#include <linux/ftrace.h>
24#include <linux/smp.h>
25#include <linux/smpboot.h>
26#include <linux/tick.h>
27#include <linux/irq.h>
28
29#define CREATE_TRACE_POINTS
30#include <trace/events/irq.h>
31
32/*
33 - No shared variables, all the data are CPU local.
34 - If a softirq needs serialization, let it serialize itself
35 by its own spinlocks.
36 - Even if softirq is serialized, only local cpu is marked for
37 execution. Hence, we get something sort of weak cpu binding.
38 Though it is still not clear, will it result in better locality
39 or will not.
40
41 Examples:
42 - NET RX softirq. It is multithreaded and does not require
43 any global serialization.
44 - NET TX softirq. It kicks software netdevice queues, hence
45 it is logically serialized per device, but this serialization
46 is invisible to common code.
47 - Tasklets: serialized wrt itself.
48 */
49
50#ifndef __ARCH_IRQ_STAT
51DEFINE_PER_CPU_ALIGNED(irq_cpustat_t, irq_stat);
52EXPORT_PER_CPU_SYMBOL(irq_stat);
53#endif
54
55static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
56
57DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
58
59const char * const softirq_to_name[NR_SOFTIRQS] = {
60 "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "IRQ_POLL",
61 "TASKLET", "SCHED", "HRTIMER", "RCU"
62};
63
64/*
65 * we cannot loop indefinitely here to avoid userspace starvation,
66 * but we also don't want to introduce a worst case 1/HZ latency
67 * to the pending events, so lets the scheduler to balance
68 * the softirq load for us.
69 */
70static void wakeup_softirqd(void)
71{
72 /* Interrupts are disabled: no need to stop preemption */
73 struct task_struct *tsk = __this_cpu_read(ksoftirqd);
74
75 if (tsk && tsk->state != TASK_RUNNING)
76 wake_up_process(tsk);
77}
78
79/*
80 * If ksoftirqd is scheduled, we do not want to process pending softirqs
81 * right now. Let ksoftirqd handle this at its own rate, to get fairness,
82 * unless we're doing some of the synchronous softirqs.
83 */
84#define SOFTIRQ_NOW_MASK ((1 << HI_SOFTIRQ) | (1 << TASKLET_SOFTIRQ))
85static bool ksoftirqd_running(unsigned long pending)
86{
87 struct task_struct *tsk = __this_cpu_read(ksoftirqd);
88
89 if (pending & SOFTIRQ_NOW_MASK)
90 return false;
91 return tsk && (tsk->state == TASK_RUNNING) &&
92 !__kthread_should_park(tsk);
93}
94
95/*
96 * preempt_count and SOFTIRQ_OFFSET usage:
97 * - preempt_count is changed by SOFTIRQ_OFFSET on entering or leaving
98 * softirq processing.
99 * - preempt_count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
100 * on local_bh_disable or local_bh_enable.
101 * This lets us distinguish between whether we are currently processing
102 * softirq and whether we just have bh disabled.
103 */
104
105/*
106 * This one is for softirq.c-internal use,
107 * where hardirqs are disabled legitimately:
108 */
109#ifdef CONFIG_TRACE_IRQFLAGS
110void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
111{
112 unsigned long flags;
113
114 WARN_ON_ONCE(in_irq());
115
116 raw_local_irq_save(flags);
117 /*
118 * The preempt tracer hooks into preempt_count_add and will break
119 * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
120 * is set and before current->softirq_enabled is cleared.
121 * We must manually increment preempt_count here and manually
122 * call the trace_preempt_off later.
123 */
124 __preempt_count_add(cnt);
125 /*
126 * Were softirqs turned off above:
127 */
128 if (softirq_count() == (cnt & SOFTIRQ_MASK))
129 trace_softirqs_off(ip);
130 raw_local_irq_restore(flags);
131
132 if (preempt_count() == cnt) {
133#ifdef CONFIG_DEBUG_PREEMPT
134 current->preempt_disable_ip = get_lock_parent_ip();
135#endif
136 trace_preempt_off(CALLER_ADDR0, get_lock_parent_ip());
137 }
138}
139EXPORT_SYMBOL(__local_bh_disable_ip);
140#endif /* CONFIG_TRACE_IRQFLAGS */
141
142static void __local_bh_enable(unsigned int cnt)
143{
144 lockdep_assert_irqs_disabled();
145
146 if (preempt_count() == cnt)
147 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
148
149 if (softirq_count() == (cnt & SOFTIRQ_MASK))
150 trace_softirqs_on(_RET_IP_);
151
152 __preempt_count_sub(cnt);
153}
154
155/*
156 * Special-case - softirqs can safely be enabled by __do_softirq(),
157 * without processing still-pending softirqs:
158 */
159void _local_bh_enable(void)
160{
161 WARN_ON_ONCE(in_irq());
162 __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
163}
164EXPORT_SYMBOL(_local_bh_enable);
165
166void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
167{
168 WARN_ON_ONCE(in_irq());
169 lockdep_assert_irqs_enabled();
170#ifdef CONFIG_TRACE_IRQFLAGS
171 local_irq_disable();
172#endif
173 /*
174 * Are softirqs going to be turned on now:
175 */
176 if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
177 trace_softirqs_on(ip);
178 /*
179 * Keep preemption disabled until we are done with
180 * softirq processing:
181 */
182 preempt_count_sub(cnt - 1);
183
184 if (unlikely(!in_interrupt() && local_softirq_pending())) {
185 /*
186 * Run softirq if any pending. And do it in its own stack
187 * as we may be calling this deep in a task call stack already.
188 */
189 do_softirq();
190 }
191
192 preempt_count_dec();
193#ifdef CONFIG_TRACE_IRQFLAGS
194 local_irq_enable();
195#endif
196 preempt_check_resched();
197}
198EXPORT_SYMBOL(__local_bh_enable_ip);
199
200/*
201 * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times,
202 * but break the loop if need_resched() is set or after 2 ms.
203 * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in
204 * certain cases, such as stop_machine(), jiffies may cease to
205 * increment and so we need the MAX_SOFTIRQ_RESTART limit as
206 * well to make sure we eventually return from this method.
207 *
208 * These limits have been established via experimentation.
209 * The two things to balance is latency against fairness -
210 * we want to handle softirqs as soon as possible, but they
211 * should not be able to lock up the box.
212 */
213#define MAX_SOFTIRQ_TIME msecs_to_jiffies(2)
214#define MAX_SOFTIRQ_RESTART 10
215
216#ifdef CONFIG_TRACE_IRQFLAGS
217/*
218 * When we run softirqs from irq_exit() and thus on the hardirq stack we need
219 * to keep the lockdep irq context tracking as tight as possible in order to
220 * not miss-qualify lock contexts and miss possible deadlocks.
221 */
222
223static inline bool lockdep_softirq_start(void)
224{
225 bool in_hardirq = false;
226
227 if (trace_hardirq_context(current)) {
228 in_hardirq = true;
229 trace_hardirq_exit();
230 }
231
232 lockdep_softirq_enter();
233
234 return in_hardirq;
235}
236
237static inline void lockdep_softirq_end(bool in_hardirq)
238{
239 lockdep_softirq_exit();
240
241 if (in_hardirq)
242 trace_hardirq_enter();
243}
244#else
245static inline bool lockdep_softirq_start(void) { return false; }
246static inline void lockdep_softirq_end(bool in_hardirq) { }
247#endif
248
249asmlinkage __visible void __softirq_entry __do_softirq(void)
250{
251 unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
252 unsigned long old_flags = current->flags;
253 int max_restart = MAX_SOFTIRQ_RESTART;
254 struct softirq_action *h;
255 bool in_hardirq;
256 __u32 pending;
257 int softirq_bit;
258
259 /*
260 * Mask out PF_MEMALLOC as the current task context is borrowed for the
261 * softirq. A softirq handled, such as network RX, might set PF_MEMALLOC
262 * again if the socket is related to swapping.
263 */
264 current->flags &= ~PF_MEMALLOC;
265
266 pending = local_softirq_pending();
267 account_irq_enter_time(current);
268
269 __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
270 in_hardirq = lockdep_softirq_start();
271
272restart:
273 /* Reset the pending bitmask before enabling irqs */
274 set_softirq_pending(0);
275
276 local_irq_enable();
277
278 h = softirq_vec;
279
280 while ((softirq_bit = ffs(pending))) {
281 unsigned int vec_nr;
282 int prev_count;
283
284 h += softirq_bit - 1;
285
286 vec_nr = h - softirq_vec;
287 prev_count = preempt_count();
288
289 kstat_incr_softirqs_this_cpu(vec_nr);
290
291 trace_softirq_entry(vec_nr);
292 h->action(h);
293 trace_softirq_exit(vec_nr);
294 if (unlikely(prev_count != preempt_count())) {
295 pr_err("huh, entered softirq %u %s %p with preempt_count %08x, exited with %08x?\n",
296 vec_nr, softirq_to_name[vec_nr], h->action,
297 prev_count, preempt_count());
298 preempt_count_set(prev_count);
299 }
300 h++;
301 pending >>= softirq_bit;
302 }
303
304 if (__this_cpu_read(ksoftirqd) == current)
305 rcu_softirq_qs();
306 local_irq_disable();
307
308 pending = local_softirq_pending();
309 if (pending) {
310 if (time_before(jiffies, end) && !need_resched() &&
311 --max_restart)
312 goto restart;
313
314 wakeup_softirqd();
315 }
316
317 lockdep_softirq_end(in_hardirq);
318 account_irq_exit_time(current);
319 __local_bh_enable(SOFTIRQ_OFFSET);
320 WARN_ON_ONCE(in_interrupt());
321 current_restore_flags(old_flags, PF_MEMALLOC);
322}
323
324asmlinkage __visible void do_softirq(void)
325{
326 __u32 pending;
327 unsigned long flags;
328
329 if (in_interrupt())
330 return;
331
332 local_irq_save(flags);
333
334 pending = local_softirq_pending();
335
336 if (pending && !ksoftirqd_running(pending))
337 do_softirq_own_stack();
338
339 local_irq_restore(flags);
340}
341
342/*
343 * Enter an interrupt context.
344 */
345void irq_enter(void)
346{
347 rcu_irq_enter();
348 if (is_idle_task(current) && !in_interrupt()) {
349 /*
350 * Prevent raise_softirq from needlessly waking up ksoftirqd
351 * here, as softirq will be serviced on return from interrupt.
352 */
353 local_bh_disable();
354 tick_irq_enter();
355 _local_bh_enable();
356 }
357
358 __irq_enter();
359}
360
361static inline void invoke_softirq(void)
362{
363 if (ksoftirqd_running(local_softirq_pending()))
364 return;
365
366 if (!force_irqthreads) {
367#ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
368 /*
369 * We can safely execute softirq on the current stack if
370 * it is the irq stack, because it should be near empty
371 * at this stage.
372 */
373 __do_softirq();
374#else
375 /*
376 * Otherwise, irq_exit() is called on the task stack that can
377 * be potentially deep already. So call softirq in its own stack
378 * to prevent from any overrun.
379 */
380 do_softirq_own_stack();
381#endif
382 } else {
383 wakeup_softirqd();
384 }
385}
386
387static inline void tick_irq_exit(void)
388{
389#ifdef CONFIG_NO_HZ_COMMON
390 int cpu = smp_processor_id();
391
392 /* Make sure that timer wheel updates are propagated */
393 if ((idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
394 if (!in_irq())
395 tick_nohz_irq_exit();
396 }
397#endif
398}
399
400/*
401 * Exit an interrupt context. Process softirqs if needed and possible:
402 */
403void irq_exit(void)
404{
405#ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
406 local_irq_disable();
407#else
408 lockdep_assert_irqs_disabled();
409#endif
410 account_irq_exit_time(current);
411 preempt_count_sub(HARDIRQ_OFFSET);
412 if (!in_interrupt() && local_softirq_pending())
413 invoke_softirq();
414
415 tick_irq_exit();
416 rcu_irq_exit();
417 trace_hardirq_exit(); /* must be last! */
418}
419
420/*
421 * This function must run with irqs disabled!
422 */
423inline void raise_softirq_irqoff(unsigned int nr)
424{
425 __raise_softirq_irqoff(nr);
426
427 /*
428 * If we're in an interrupt or softirq, we're done
429 * (this also catches softirq-disabled code). We will
430 * actually run the softirq once we return from
431 * the irq or softirq.
432 *
433 * Otherwise we wake up ksoftirqd to make sure we
434 * schedule the softirq soon.
435 */
436 if (!in_interrupt())
437 wakeup_softirqd();
438}
439
440void raise_softirq(unsigned int nr)
441{
442 unsigned long flags;
443
444 local_irq_save(flags);
445 raise_softirq_irqoff(nr);
446 local_irq_restore(flags);
447}
448
449void __raise_softirq_irqoff(unsigned int nr)
450{
451 trace_softirq_raise(nr);
452 or_softirq_pending(1UL << nr);
453}
454
455void open_softirq(int nr, void (*action)(struct softirq_action *))
456{
457 softirq_vec[nr].action = action;
458}
459
460/*
461 * Tasklets
462 */
463struct tasklet_head {
464 struct tasklet_struct *head;
465 struct tasklet_struct **tail;
466};
467
468static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
469static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
470
471static void __tasklet_schedule_common(struct tasklet_struct *t,
472 struct tasklet_head __percpu *headp,
473 unsigned int softirq_nr)
474{
475 struct tasklet_head *head;
476 unsigned long flags;
477
478 local_irq_save(flags);
479 head = this_cpu_ptr(headp);
480 t->next = NULL;
481 *head->tail = t;
482 head->tail = &(t->next);
483 raise_softirq_irqoff(softirq_nr);
484 local_irq_restore(flags);
485}
486
487void __tasklet_schedule(struct tasklet_struct *t)
488{
489 __tasklet_schedule_common(t, &tasklet_vec,
490 TASKLET_SOFTIRQ);
491}
492EXPORT_SYMBOL(__tasklet_schedule);
493
494void __tasklet_hi_schedule(struct tasklet_struct *t)
495{
496 __tasklet_schedule_common(t, &tasklet_hi_vec,
497 HI_SOFTIRQ);
498}
499EXPORT_SYMBOL(__tasklet_hi_schedule);
500
501static void tasklet_action_common(struct softirq_action *a,
502 struct tasklet_head *tl_head,
503 unsigned int softirq_nr)
504{
505 struct tasklet_struct *list;
506
507 local_irq_disable();
508 list = tl_head->head;
509 tl_head->head = NULL;
510 tl_head->tail = &tl_head->head;
511 local_irq_enable();
512
513 while (list) {
514 struct tasklet_struct *t = list;
515
516 list = list->next;
517
518 if (tasklet_trylock(t)) {
519 if (!atomic_read(&t->count)) {
520 if (!test_and_clear_bit(TASKLET_STATE_SCHED,
521 &t->state))
522 BUG();
523 t->func(t->data);
524 tasklet_unlock(t);
525 continue;
526 }
527 tasklet_unlock(t);
528 }
529
530 local_irq_disable();
531 t->next = NULL;
532 *tl_head->tail = t;
533 tl_head->tail = &t->next;
534 __raise_softirq_irqoff(softirq_nr);
535 local_irq_enable();
536 }
537}
538
539static __latent_entropy void tasklet_action(struct softirq_action *a)
540{
541 tasklet_action_common(a, this_cpu_ptr(&tasklet_vec), TASKLET_SOFTIRQ);
542}
543
544static __latent_entropy void tasklet_hi_action(struct softirq_action *a)
545{
546 tasklet_action_common(a, this_cpu_ptr(&tasklet_hi_vec), HI_SOFTIRQ);
547}
548
549void tasklet_init(struct tasklet_struct *t,
550 void (*func)(unsigned long), unsigned long data)
551{
552 t->next = NULL;
553 t->state = 0;
554 atomic_set(&t->count, 0);
555 t->func = func;
556 t->data = data;
557}
558EXPORT_SYMBOL(tasklet_init);
559
560void tasklet_kill(struct tasklet_struct *t)
561{
562 if (in_interrupt())
563 pr_notice("Attempt to kill tasklet from interrupt\n");
564
565 while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
566 do {
567 yield();
568 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
569 }
570 tasklet_unlock_wait(t);
571 clear_bit(TASKLET_STATE_SCHED, &t->state);
572}
573EXPORT_SYMBOL(tasklet_kill);
574
575void __init softirq_init(void)
576{
577 int cpu;
578
579 for_each_possible_cpu(cpu) {
580 per_cpu(tasklet_vec, cpu).tail =
581 &per_cpu(tasklet_vec, cpu).head;
582 per_cpu(tasklet_hi_vec, cpu).tail =
583 &per_cpu(tasklet_hi_vec, cpu).head;
584 }
585
586 open_softirq(TASKLET_SOFTIRQ, tasklet_action);
587 open_softirq(HI_SOFTIRQ, tasklet_hi_action);
588}
589
590static int ksoftirqd_should_run(unsigned int cpu)
591{
592 return local_softirq_pending();
593}
594
595static void run_ksoftirqd(unsigned int cpu)
596{
597 local_irq_disable();
598 if (local_softirq_pending()) {
599 /*
600 * We can safely run softirq on inline stack, as we are not deep
601 * in the task stack here.
602 */
603 __do_softirq();
604 local_irq_enable();
605 cond_resched();
606 return;
607 }
608 local_irq_enable();
609}
610
611#ifdef CONFIG_HOTPLUG_CPU
612/*
613 * tasklet_kill_immediate is called to remove a tasklet which can already be
614 * scheduled for execution on @cpu.
615 *
616 * Unlike tasklet_kill, this function removes the tasklet
617 * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
618 *
619 * When this function is called, @cpu must be in the CPU_DEAD state.
620 */
621void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
622{
623 struct tasklet_struct **i;
624
625 BUG_ON(cpu_online(cpu));
626 BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
627
628 if (!test_bit(TASKLET_STATE_SCHED, &t->state))
629 return;
630
631 /* CPU is dead, so no lock needed. */
632 for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
633 if (*i == t) {
634 *i = t->next;
635 /* If this was the tail element, move the tail ptr */
636 if (*i == NULL)
637 per_cpu(tasklet_vec, cpu).tail = i;
638 return;
639 }
640 }
641 BUG();
642}
643
644static int takeover_tasklets(unsigned int cpu)
645{
646 /* CPU is dead, so no lock needed. */
647 local_irq_disable();
648
649 /* Find end, append list for that CPU. */
650 if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
651 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
652 __this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
653 per_cpu(tasklet_vec, cpu).head = NULL;
654 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
655 }
656 raise_softirq_irqoff(TASKLET_SOFTIRQ);
657
658 if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
659 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
660 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
661 per_cpu(tasklet_hi_vec, cpu).head = NULL;
662 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
663 }
664 raise_softirq_irqoff(HI_SOFTIRQ);
665
666 local_irq_enable();
667 return 0;
668}
669#else
670#define takeover_tasklets NULL
671#endif /* CONFIG_HOTPLUG_CPU */
672
673static struct smp_hotplug_thread softirq_threads = {
674 .store = &ksoftirqd,
675 .thread_should_run = ksoftirqd_should_run,
676 .thread_fn = run_ksoftirqd,
677 .thread_comm = "ksoftirqd/%u",
678};
679
680static __init int spawn_ksoftirqd(void)
681{
682 cpuhp_setup_state_nocalls(CPUHP_SOFTIRQ_DEAD, "softirq:dead", NULL,
683 takeover_tasklets);
684 BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
685
686 return 0;
687}
688early_initcall(spawn_ksoftirqd);
689
690/*
691 * [ These __weak aliases are kept in a separate compilation unit, so that
692 * GCC does not inline them incorrectly. ]
693 */
694
695int __init __weak early_irq_init(void)
696{
697 return 0;
698}
699
700int __init __weak arch_probe_nr_irqs(void)
701{
702 return NR_IRQS_LEGACY;
703}
704
705int __init __weak arch_early_irq_init(void)
706{
707 return 0;
708}
709
710unsigned int __weak arch_dynirq_lower_bound(unsigned int from)
711{
712 return from;
713}
1/*
2 * linux/kernel/softirq.c
3 *
4 * Copyright (C) 1992 Linus Torvalds
5 *
6 * Distribute under GPLv2.
7 *
8 * Rewritten. Old one was good in 2.2, but in 2.3 it was immoral. --ANK (990903)
9 */
10
11#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12
13#include <linux/export.h>
14#include <linux/kernel_stat.h>
15#include <linux/interrupt.h>
16#include <linux/init.h>
17#include <linux/mm.h>
18#include <linux/notifier.h>
19#include <linux/percpu.h>
20#include <linux/cpu.h>
21#include <linux/freezer.h>
22#include <linux/kthread.h>
23#include <linux/rcupdate.h>
24#include <linux/ftrace.h>
25#include <linux/smp.h>
26#include <linux/smpboot.h>
27#include <linux/tick.h>
28#include <linux/irq.h>
29
30#define CREATE_TRACE_POINTS
31#include <trace/events/irq.h>
32
33/*
34 - No shared variables, all the data are CPU local.
35 - If a softirq needs serialization, let it serialize itself
36 by its own spinlocks.
37 - Even if softirq is serialized, only local cpu is marked for
38 execution. Hence, we get something sort of weak cpu binding.
39 Though it is still not clear, will it result in better locality
40 or will not.
41
42 Examples:
43 - NET RX softirq. It is multithreaded and does not require
44 any global serialization.
45 - NET TX softirq. It kicks software netdevice queues, hence
46 it is logically serialized per device, but this serialization
47 is invisible to common code.
48 - Tasklets: serialized wrt itself.
49 */
50
51#ifndef __ARCH_IRQ_STAT
52irq_cpustat_t irq_stat[NR_CPUS] ____cacheline_aligned;
53EXPORT_SYMBOL(irq_stat);
54#endif
55
56static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
57
58DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
59
60const char * const softirq_to_name[NR_SOFTIRQS] = {
61 "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "IRQ_POLL",
62 "TASKLET", "SCHED", "HRTIMER", "RCU"
63};
64
65/*
66 * we cannot loop indefinitely here to avoid userspace starvation,
67 * but we also don't want to introduce a worst case 1/HZ latency
68 * to the pending events, so lets the scheduler to balance
69 * the softirq load for us.
70 */
71static void wakeup_softirqd(void)
72{
73 /* Interrupts are disabled: no need to stop preemption */
74 struct task_struct *tsk = __this_cpu_read(ksoftirqd);
75
76 if (tsk && tsk->state != TASK_RUNNING)
77 wake_up_process(tsk);
78}
79
80/*
81 * If ksoftirqd is scheduled, we do not want to process pending softirqs
82 * right now. Let ksoftirqd handle this at its own rate, to get fairness.
83 */
84static bool ksoftirqd_running(void)
85{
86 struct task_struct *tsk = __this_cpu_read(ksoftirqd);
87
88 return tsk && (tsk->state == TASK_RUNNING);
89}
90
91/*
92 * preempt_count and SOFTIRQ_OFFSET usage:
93 * - preempt_count is changed by SOFTIRQ_OFFSET on entering or leaving
94 * softirq processing.
95 * - preempt_count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
96 * on local_bh_disable or local_bh_enable.
97 * This lets us distinguish between whether we are currently processing
98 * softirq and whether we just have bh disabled.
99 */
100
101/*
102 * This one is for softirq.c-internal use,
103 * where hardirqs are disabled legitimately:
104 */
105#ifdef CONFIG_TRACE_IRQFLAGS
106void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
107{
108 unsigned long flags;
109
110 WARN_ON_ONCE(in_irq());
111
112 raw_local_irq_save(flags);
113 /*
114 * The preempt tracer hooks into preempt_count_add and will break
115 * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
116 * is set and before current->softirq_enabled is cleared.
117 * We must manually increment preempt_count here and manually
118 * call the trace_preempt_off later.
119 */
120 __preempt_count_add(cnt);
121 /*
122 * Were softirqs turned off above:
123 */
124 if (softirq_count() == (cnt & SOFTIRQ_MASK))
125 trace_softirqs_off(ip);
126 raw_local_irq_restore(flags);
127
128 if (preempt_count() == cnt) {
129#ifdef CONFIG_DEBUG_PREEMPT
130 current->preempt_disable_ip = get_lock_parent_ip();
131#endif
132 trace_preempt_off(CALLER_ADDR0, get_lock_parent_ip());
133 }
134}
135EXPORT_SYMBOL(__local_bh_disable_ip);
136#endif /* CONFIG_TRACE_IRQFLAGS */
137
138static void __local_bh_enable(unsigned int cnt)
139{
140 lockdep_assert_irqs_disabled();
141
142 if (softirq_count() == (cnt & SOFTIRQ_MASK))
143 trace_softirqs_on(_RET_IP_);
144 preempt_count_sub(cnt);
145}
146
147/*
148 * Special-case - softirqs can safely be enabled in
149 * cond_resched_softirq(), or by __do_softirq(),
150 * without processing still-pending softirqs:
151 */
152void _local_bh_enable(void)
153{
154 WARN_ON_ONCE(in_irq());
155 __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
156}
157EXPORT_SYMBOL(_local_bh_enable);
158
159void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
160{
161 WARN_ON_ONCE(in_irq());
162 lockdep_assert_irqs_enabled();
163#ifdef CONFIG_TRACE_IRQFLAGS
164 local_irq_disable();
165#endif
166 /*
167 * Are softirqs going to be turned on now:
168 */
169 if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
170 trace_softirqs_on(ip);
171 /*
172 * Keep preemption disabled until we are done with
173 * softirq processing:
174 */
175 preempt_count_sub(cnt - 1);
176
177 if (unlikely(!in_interrupt() && local_softirq_pending())) {
178 /*
179 * Run softirq if any pending. And do it in its own stack
180 * as we may be calling this deep in a task call stack already.
181 */
182 do_softirq();
183 }
184
185 preempt_count_dec();
186#ifdef CONFIG_TRACE_IRQFLAGS
187 local_irq_enable();
188#endif
189 preempt_check_resched();
190}
191EXPORT_SYMBOL(__local_bh_enable_ip);
192
193/*
194 * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times,
195 * but break the loop if need_resched() is set or after 2 ms.
196 * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in
197 * certain cases, such as stop_machine(), jiffies may cease to
198 * increment and so we need the MAX_SOFTIRQ_RESTART limit as
199 * well to make sure we eventually return from this method.
200 *
201 * These limits have been established via experimentation.
202 * The two things to balance is latency against fairness -
203 * we want to handle softirqs as soon as possible, but they
204 * should not be able to lock up the box.
205 */
206#define MAX_SOFTIRQ_TIME msecs_to_jiffies(2)
207#define MAX_SOFTIRQ_RESTART 10
208
209#ifdef CONFIG_TRACE_IRQFLAGS
210/*
211 * When we run softirqs from irq_exit() and thus on the hardirq stack we need
212 * to keep the lockdep irq context tracking as tight as possible in order to
213 * not miss-qualify lock contexts and miss possible deadlocks.
214 */
215
216static inline bool lockdep_softirq_start(void)
217{
218 bool in_hardirq = false;
219
220 if (trace_hardirq_context(current)) {
221 in_hardirq = true;
222 trace_hardirq_exit();
223 }
224
225 lockdep_softirq_enter();
226
227 return in_hardirq;
228}
229
230static inline void lockdep_softirq_end(bool in_hardirq)
231{
232 lockdep_softirq_exit();
233
234 if (in_hardirq)
235 trace_hardirq_enter();
236}
237#else
238static inline bool lockdep_softirq_start(void) { return false; }
239static inline void lockdep_softirq_end(bool in_hardirq) { }
240#endif
241
242asmlinkage __visible void __softirq_entry __do_softirq(void)
243{
244 unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
245 unsigned long old_flags = current->flags;
246 int max_restart = MAX_SOFTIRQ_RESTART;
247 struct softirq_action *h;
248 bool in_hardirq;
249 __u32 pending;
250 int softirq_bit;
251
252 /*
253 * Mask out PF_MEMALLOC s current task context is borrowed for the
254 * softirq. A softirq handled such as network RX might set PF_MEMALLOC
255 * again if the socket is related to swap
256 */
257 current->flags &= ~PF_MEMALLOC;
258
259 pending = local_softirq_pending();
260 account_irq_enter_time(current);
261
262 __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
263 in_hardirq = lockdep_softirq_start();
264
265restart:
266 /* Reset the pending bitmask before enabling irqs */
267 set_softirq_pending(0);
268
269 local_irq_enable();
270
271 h = softirq_vec;
272
273 while ((softirq_bit = ffs(pending))) {
274 unsigned int vec_nr;
275 int prev_count;
276
277 h += softirq_bit - 1;
278
279 vec_nr = h - softirq_vec;
280 prev_count = preempt_count();
281
282 kstat_incr_softirqs_this_cpu(vec_nr);
283
284 trace_softirq_entry(vec_nr);
285 h->action(h);
286 trace_softirq_exit(vec_nr);
287 if (unlikely(prev_count != preempt_count())) {
288 pr_err("huh, entered softirq %u %s %p with preempt_count %08x, exited with %08x?\n",
289 vec_nr, softirq_to_name[vec_nr], h->action,
290 prev_count, preempt_count());
291 preempt_count_set(prev_count);
292 }
293 h++;
294 pending >>= softirq_bit;
295 }
296
297 rcu_bh_qs();
298 local_irq_disable();
299
300 pending = local_softirq_pending();
301 if (pending) {
302 if (time_before(jiffies, end) && !need_resched() &&
303 --max_restart)
304 goto restart;
305
306 wakeup_softirqd();
307 }
308
309 lockdep_softirq_end(in_hardirq);
310 account_irq_exit_time(current);
311 __local_bh_enable(SOFTIRQ_OFFSET);
312 WARN_ON_ONCE(in_interrupt());
313 current_restore_flags(old_flags, PF_MEMALLOC);
314}
315
316asmlinkage __visible void do_softirq(void)
317{
318 __u32 pending;
319 unsigned long flags;
320
321 if (in_interrupt())
322 return;
323
324 local_irq_save(flags);
325
326 pending = local_softirq_pending();
327
328 if (pending && !ksoftirqd_running())
329 do_softirq_own_stack();
330
331 local_irq_restore(flags);
332}
333
334/*
335 * Enter an interrupt context.
336 */
337void irq_enter(void)
338{
339 rcu_irq_enter();
340 if (is_idle_task(current) && !in_interrupt()) {
341 /*
342 * Prevent raise_softirq from needlessly waking up ksoftirqd
343 * here, as softirq will be serviced on return from interrupt.
344 */
345 local_bh_disable();
346 tick_irq_enter();
347 _local_bh_enable();
348 }
349
350 __irq_enter();
351}
352
353static inline void invoke_softirq(void)
354{
355 if (ksoftirqd_running())
356 return;
357
358 if (!force_irqthreads) {
359#ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
360 /*
361 * We can safely execute softirq on the current stack if
362 * it is the irq stack, because it should be near empty
363 * at this stage.
364 */
365 __do_softirq();
366#else
367 /*
368 * Otherwise, irq_exit() is called on the task stack that can
369 * be potentially deep already. So call softirq in its own stack
370 * to prevent from any overrun.
371 */
372 do_softirq_own_stack();
373#endif
374 } else {
375 wakeup_softirqd();
376 }
377}
378
379static inline void tick_irq_exit(void)
380{
381#ifdef CONFIG_NO_HZ_COMMON
382 int cpu = smp_processor_id();
383
384 /* Make sure that timer wheel updates are propagated */
385 if ((idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
386 if (!in_interrupt())
387 tick_nohz_irq_exit();
388 }
389#endif
390}
391
392/*
393 * Exit an interrupt context. Process softirqs if needed and possible:
394 */
395void irq_exit(void)
396{
397#ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
398 local_irq_disable();
399#else
400 lockdep_assert_irqs_disabled();
401#endif
402 account_irq_exit_time(current);
403 preempt_count_sub(HARDIRQ_OFFSET);
404 if (!in_interrupt() && local_softirq_pending())
405 invoke_softirq();
406
407 tick_irq_exit();
408 rcu_irq_exit();
409 trace_hardirq_exit(); /* must be last! */
410}
411
412/*
413 * This function must run with irqs disabled!
414 */
415inline void raise_softirq_irqoff(unsigned int nr)
416{
417 __raise_softirq_irqoff(nr);
418
419 /*
420 * If we're in an interrupt or softirq, we're done
421 * (this also catches softirq-disabled code). We will
422 * actually run the softirq once we return from
423 * the irq or softirq.
424 *
425 * Otherwise we wake up ksoftirqd to make sure we
426 * schedule the softirq soon.
427 */
428 if (!in_interrupt())
429 wakeup_softirqd();
430}
431
432void raise_softirq(unsigned int nr)
433{
434 unsigned long flags;
435
436 local_irq_save(flags);
437 raise_softirq_irqoff(nr);
438 local_irq_restore(flags);
439}
440
441void __raise_softirq_irqoff(unsigned int nr)
442{
443 trace_softirq_raise(nr);
444 or_softirq_pending(1UL << nr);
445}
446
447void open_softirq(int nr, void (*action)(struct softirq_action *))
448{
449 softirq_vec[nr].action = action;
450}
451
452/*
453 * Tasklets
454 */
455struct tasklet_head {
456 struct tasklet_struct *head;
457 struct tasklet_struct **tail;
458};
459
460static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
461static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
462
463static void __tasklet_schedule_common(struct tasklet_struct *t,
464 struct tasklet_head __percpu *headp,
465 unsigned int softirq_nr)
466{
467 struct tasklet_head *head;
468 unsigned long flags;
469
470 local_irq_save(flags);
471 head = this_cpu_ptr(headp);
472 t->next = NULL;
473 *head->tail = t;
474 head->tail = &(t->next);
475 raise_softirq_irqoff(softirq_nr);
476 local_irq_restore(flags);
477}
478
479void __tasklet_schedule(struct tasklet_struct *t)
480{
481 __tasklet_schedule_common(t, &tasklet_vec,
482 TASKLET_SOFTIRQ);
483}
484EXPORT_SYMBOL(__tasklet_schedule);
485
486void __tasklet_hi_schedule(struct tasklet_struct *t)
487{
488 __tasklet_schedule_common(t, &tasklet_hi_vec,
489 HI_SOFTIRQ);
490}
491EXPORT_SYMBOL(__tasklet_hi_schedule);
492
493static void tasklet_action_common(struct softirq_action *a,
494 struct tasklet_head *tl_head,
495 unsigned int softirq_nr)
496{
497 struct tasklet_struct *list;
498
499 local_irq_disable();
500 list = tl_head->head;
501 tl_head->head = NULL;
502 tl_head->tail = &tl_head->head;
503 local_irq_enable();
504
505 while (list) {
506 struct tasklet_struct *t = list;
507
508 list = list->next;
509
510 if (tasklet_trylock(t)) {
511 if (!atomic_read(&t->count)) {
512 if (!test_and_clear_bit(TASKLET_STATE_SCHED,
513 &t->state))
514 BUG();
515 t->func(t->data);
516 tasklet_unlock(t);
517 continue;
518 }
519 tasklet_unlock(t);
520 }
521
522 local_irq_disable();
523 t->next = NULL;
524 *tl_head->tail = t;
525 tl_head->tail = &t->next;
526 __raise_softirq_irqoff(softirq_nr);
527 local_irq_enable();
528 }
529}
530
531static __latent_entropy void tasklet_action(struct softirq_action *a)
532{
533 tasklet_action_common(a, this_cpu_ptr(&tasklet_vec), TASKLET_SOFTIRQ);
534}
535
536static __latent_entropy void tasklet_hi_action(struct softirq_action *a)
537{
538 tasklet_action_common(a, this_cpu_ptr(&tasklet_hi_vec), HI_SOFTIRQ);
539}
540
541void tasklet_init(struct tasklet_struct *t,
542 void (*func)(unsigned long), unsigned long data)
543{
544 t->next = NULL;
545 t->state = 0;
546 atomic_set(&t->count, 0);
547 t->func = func;
548 t->data = data;
549}
550EXPORT_SYMBOL(tasklet_init);
551
552void tasklet_kill(struct tasklet_struct *t)
553{
554 if (in_interrupt())
555 pr_notice("Attempt to kill tasklet from interrupt\n");
556
557 while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
558 do {
559 yield();
560 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
561 }
562 tasklet_unlock_wait(t);
563 clear_bit(TASKLET_STATE_SCHED, &t->state);
564}
565EXPORT_SYMBOL(tasklet_kill);
566
567/*
568 * tasklet_hrtimer
569 */
570
571/*
572 * The trampoline is called when the hrtimer expires. It schedules a tasklet
573 * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
574 * hrtimer callback, but from softirq context.
575 */
576static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer)
577{
578 struct tasklet_hrtimer *ttimer =
579 container_of(timer, struct tasklet_hrtimer, timer);
580
581 tasklet_hi_schedule(&ttimer->tasklet);
582 return HRTIMER_NORESTART;
583}
584
585/*
586 * Helper function which calls the hrtimer callback from
587 * tasklet/softirq context
588 */
589static void __tasklet_hrtimer_trampoline(unsigned long data)
590{
591 struct tasklet_hrtimer *ttimer = (void *)data;
592 enum hrtimer_restart restart;
593
594 restart = ttimer->function(&ttimer->timer);
595 if (restart != HRTIMER_NORESTART)
596 hrtimer_restart(&ttimer->timer);
597}
598
599/**
600 * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
601 * @ttimer: tasklet_hrtimer which is initialized
602 * @function: hrtimer callback function which gets called from softirq context
603 * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
604 * @mode: hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
605 */
606void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer,
607 enum hrtimer_restart (*function)(struct hrtimer *),
608 clockid_t which_clock, enum hrtimer_mode mode)
609{
610 hrtimer_init(&ttimer->timer, which_clock, mode);
611 ttimer->timer.function = __hrtimer_tasklet_trampoline;
612 tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline,
613 (unsigned long)ttimer);
614 ttimer->function = function;
615}
616EXPORT_SYMBOL_GPL(tasklet_hrtimer_init);
617
618void __init softirq_init(void)
619{
620 int cpu;
621
622 for_each_possible_cpu(cpu) {
623 per_cpu(tasklet_vec, cpu).tail =
624 &per_cpu(tasklet_vec, cpu).head;
625 per_cpu(tasklet_hi_vec, cpu).tail =
626 &per_cpu(tasklet_hi_vec, cpu).head;
627 }
628
629 open_softirq(TASKLET_SOFTIRQ, tasklet_action);
630 open_softirq(HI_SOFTIRQ, tasklet_hi_action);
631}
632
633static int ksoftirqd_should_run(unsigned int cpu)
634{
635 return local_softirq_pending();
636}
637
638static void run_ksoftirqd(unsigned int cpu)
639{
640 local_irq_disable();
641 if (local_softirq_pending()) {
642 /*
643 * We can safely run softirq on inline stack, as we are not deep
644 * in the task stack here.
645 */
646 __do_softirq();
647 local_irq_enable();
648 cond_resched();
649 return;
650 }
651 local_irq_enable();
652}
653
654#ifdef CONFIG_HOTPLUG_CPU
655/*
656 * tasklet_kill_immediate is called to remove a tasklet which can already be
657 * scheduled for execution on @cpu.
658 *
659 * Unlike tasklet_kill, this function removes the tasklet
660 * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
661 *
662 * When this function is called, @cpu must be in the CPU_DEAD state.
663 */
664void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
665{
666 struct tasklet_struct **i;
667
668 BUG_ON(cpu_online(cpu));
669 BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
670
671 if (!test_bit(TASKLET_STATE_SCHED, &t->state))
672 return;
673
674 /* CPU is dead, so no lock needed. */
675 for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
676 if (*i == t) {
677 *i = t->next;
678 /* If this was the tail element, move the tail ptr */
679 if (*i == NULL)
680 per_cpu(tasklet_vec, cpu).tail = i;
681 return;
682 }
683 }
684 BUG();
685}
686
687static int takeover_tasklets(unsigned int cpu)
688{
689 /* CPU is dead, so no lock needed. */
690 local_irq_disable();
691
692 /* Find end, append list for that CPU. */
693 if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
694 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
695 this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
696 per_cpu(tasklet_vec, cpu).head = NULL;
697 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
698 }
699 raise_softirq_irqoff(TASKLET_SOFTIRQ);
700
701 if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
702 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
703 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
704 per_cpu(tasklet_hi_vec, cpu).head = NULL;
705 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
706 }
707 raise_softirq_irqoff(HI_SOFTIRQ);
708
709 local_irq_enable();
710 return 0;
711}
712#else
713#define takeover_tasklets NULL
714#endif /* CONFIG_HOTPLUG_CPU */
715
716static struct smp_hotplug_thread softirq_threads = {
717 .store = &ksoftirqd,
718 .thread_should_run = ksoftirqd_should_run,
719 .thread_fn = run_ksoftirqd,
720 .thread_comm = "ksoftirqd/%u",
721};
722
723static __init int spawn_ksoftirqd(void)
724{
725 cpuhp_setup_state_nocalls(CPUHP_SOFTIRQ_DEAD, "softirq:dead", NULL,
726 takeover_tasklets);
727 BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
728
729 return 0;
730}
731early_initcall(spawn_ksoftirqd);
732
733/*
734 * [ These __weak aliases are kept in a separate compilation unit, so that
735 * GCC does not inline them incorrectly. ]
736 */
737
738int __init __weak early_irq_init(void)
739{
740 return 0;
741}
742
743int __init __weak arch_probe_nr_irqs(void)
744{
745 return NR_IRQS_LEGACY;
746}
747
748int __init __weak arch_early_irq_init(void)
749{
750 return 0;
751}
752
753unsigned int __weak arch_dynirq_lower_bound(unsigned int from)
754{
755 return from;
756}