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// 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
110
111DEFINE_PER_CPU(int, hardirqs_enabled);
112DEFINE_PER_CPU(int, hardirq_context);
113EXPORT_PER_CPU_SYMBOL_GPL(hardirqs_enabled);
114EXPORT_PER_CPU_SYMBOL_GPL(hardirq_context);
115
116void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
117{
118 unsigned long flags;
119
120 WARN_ON_ONCE(in_irq());
121
122 raw_local_irq_save(flags);
123 /*
124 * The preempt tracer hooks into preempt_count_add and will break
125 * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
126 * is set and before current->softirq_enabled is cleared.
127 * We must manually increment preempt_count here and manually
128 * call the trace_preempt_off later.
129 */
130 __preempt_count_add(cnt);
131 /*
132 * Were softirqs turned off above:
133 */
134 if (softirq_count() == (cnt & SOFTIRQ_MASK))
135 lockdep_softirqs_off(ip);
136 raw_local_irq_restore(flags);
137
138 if (preempt_count() == cnt) {
139#ifdef CONFIG_DEBUG_PREEMPT
140 current->preempt_disable_ip = get_lock_parent_ip();
141#endif
142 trace_preempt_off(CALLER_ADDR0, get_lock_parent_ip());
143 }
144}
145EXPORT_SYMBOL(__local_bh_disable_ip);
146#endif /* CONFIG_TRACE_IRQFLAGS */
147
148static void __local_bh_enable(unsigned int cnt)
149{
150 lockdep_assert_irqs_disabled();
151
152 if (preempt_count() == cnt)
153 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
154
155 if (softirq_count() == (cnt & SOFTIRQ_MASK))
156 lockdep_softirqs_on(_RET_IP_);
157
158 __preempt_count_sub(cnt);
159}
160
161/*
162 * Special-case - softirqs can safely be enabled by __do_softirq(),
163 * without processing still-pending softirqs:
164 */
165void _local_bh_enable(void)
166{
167 WARN_ON_ONCE(in_irq());
168 __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
169}
170EXPORT_SYMBOL(_local_bh_enable);
171
172void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
173{
174 WARN_ON_ONCE(in_irq());
175 lockdep_assert_irqs_enabled();
176#ifdef CONFIG_TRACE_IRQFLAGS
177 local_irq_disable();
178#endif
179 /*
180 * Are softirqs going to be turned on now:
181 */
182 if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
183 lockdep_softirqs_on(ip);
184 /*
185 * Keep preemption disabled until we are done with
186 * softirq processing:
187 */
188 preempt_count_sub(cnt - 1);
189
190 if (unlikely(!in_interrupt() && local_softirq_pending())) {
191 /*
192 * Run softirq if any pending. And do it in its own stack
193 * as we may be calling this deep in a task call stack already.
194 */
195 do_softirq();
196 }
197
198 preempt_count_dec();
199#ifdef CONFIG_TRACE_IRQFLAGS
200 local_irq_enable();
201#endif
202 preempt_check_resched();
203}
204EXPORT_SYMBOL(__local_bh_enable_ip);
205
206/*
207 * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times,
208 * but break the loop if need_resched() is set or after 2 ms.
209 * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in
210 * certain cases, such as stop_machine(), jiffies may cease to
211 * increment and so we need the MAX_SOFTIRQ_RESTART limit as
212 * well to make sure we eventually return from this method.
213 *
214 * These limits have been established via experimentation.
215 * The two things to balance is latency against fairness -
216 * we want to handle softirqs as soon as possible, but they
217 * should not be able to lock up the box.
218 */
219#define MAX_SOFTIRQ_TIME msecs_to_jiffies(2)
220#define MAX_SOFTIRQ_RESTART 10
221
222#ifdef CONFIG_TRACE_IRQFLAGS
223/*
224 * When we run softirqs from irq_exit() and thus on the hardirq stack we need
225 * to keep the lockdep irq context tracking as tight as possible in order to
226 * not miss-qualify lock contexts and miss possible deadlocks.
227 */
228
229static inline bool lockdep_softirq_start(void)
230{
231 bool in_hardirq = false;
232
233 if (lockdep_hardirq_context()) {
234 in_hardirq = true;
235 lockdep_hardirq_exit();
236 }
237
238 lockdep_softirq_enter();
239
240 return in_hardirq;
241}
242
243static inline void lockdep_softirq_end(bool in_hardirq)
244{
245 lockdep_softirq_exit();
246
247 if (in_hardirq)
248 lockdep_hardirq_enter();
249}
250#else
251static inline bool lockdep_softirq_start(void) { return false; }
252static inline void lockdep_softirq_end(bool in_hardirq) { }
253#endif
254
255asmlinkage __visible void __softirq_entry __do_softirq(void)
256{
257 unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
258 unsigned long old_flags = current->flags;
259 int max_restart = MAX_SOFTIRQ_RESTART;
260 struct softirq_action *h;
261 bool in_hardirq;
262 __u32 pending;
263 int softirq_bit;
264
265 /*
266 * Mask out PF_MEMALLOC as the current task context is borrowed for the
267 * softirq. A softirq handled, such as network RX, might set PF_MEMALLOC
268 * again if the socket is related to swapping.
269 */
270 current->flags &= ~PF_MEMALLOC;
271
272 pending = local_softirq_pending();
273 account_irq_enter_time(current);
274
275 __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
276 in_hardirq = lockdep_softirq_start();
277
278restart:
279 /* Reset the pending bitmask before enabling irqs */
280 set_softirq_pending(0);
281
282 local_irq_enable();
283
284 h = softirq_vec;
285
286 while ((softirq_bit = ffs(pending))) {
287 unsigned int vec_nr;
288 int prev_count;
289
290 h += softirq_bit - 1;
291
292 vec_nr = h - softirq_vec;
293 prev_count = preempt_count();
294
295 kstat_incr_softirqs_this_cpu(vec_nr);
296
297 trace_softirq_entry(vec_nr);
298 h->action(h);
299 trace_softirq_exit(vec_nr);
300 if (unlikely(prev_count != preempt_count())) {
301 pr_err("huh, entered softirq %u %s %p with preempt_count %08x, exited with %08x?\n",
302 vec_nr, softirq_to_name[vec_nr], h->action,
303 prev_count, preempt_count());
304 preempt_count_set(prev_count);
305 }
306 h++;
307 pending >>= softirq_bit;
308 }
309
310 if (__this_cpu_read(ksoftirqd) == current)
311 rcu_softirq_qs();
312 local_irq_disable();
313
314 pending = local_softirq_pending();
315 if (pending) {
316 if (time_before(jiffies, end) && !need_resched() &&
317 --max_restart)
318 goto restart;
319
320 wakeup_softirqd();
321 }
322
323 lockdep_softirq_end(in_hardirq);
324 account_irq_exit_time(current);
325 __local_bh_enable(SOFTIRQ_OFFSET);
326 WARN_ON_ONCE(in_interrupt());
327 current_restore_flags(old_flags, PF_MEMALLOC);
328}
329
330asmlinkage __visible void do_softirq(void)
331{
332 __u32 pending;
333 unsigned long flags;
334
335 if (in_interrupt())
336 return;
337
338 local_irq_save(flags);
339
340 pending = local_softirq_pending();
341
342 if (pending && !ksoftirqd_running(pending))
343 do_softirq_own_stack();
344
345 local_irq_restore(flags);
346}
347
348/**
349 * irq_enter_rcu - Enter an interrupt context with RCU watching
350 */
351void irq_enter_rcu(void)
352{
353 if (is_idle_task(current) && !in_interrupt()) {
354 /*
355 * Prevent raise_softirq from needlessly waking up ksoftirqd
356 * here, as softirq will be serviced on return from interrupt.
357 */
358 local_bh_disable();
359 tick_irq_enter();
360 _local_bh_enable();
361 }
362 __irq_enter();
363}
364
365/**
366 * irq_enter - Enter an interrupt context including RCU update
367 */
368void irq_enter(void)
369{
370 rcu_irq_enter();
371 irq_enter_rcu();
372}
373
374static inline void invoke_softirq(void)
375{
376 if (ksoftirqd_running(local_softirq_pending()))
377 return;
378
379 if (!force_irqthreads) {
380#ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
381 /*
382 * We can safely execute softirq on the current stack if
383 * it is the irq stack, because it should be near empty
384 * at this stage.
385 */
386 __do_softirq();
387#else
388 /*
389 * Otherwise, irq_exit() is called on the task stack that can
390 * be potentially deep already. So call softirq in its own stack
391 * to prevent from any overrun.
392 */
393 do_softirq_own_stack();
394#endif
395 } else {
396 wakeup_softirqd();
397 }
398}
399
400static inline void tick_irq_exit(void)
401{
402#ifdef CONFIG_NO_HZ_COMMON
403 int cpu = smp_processor_id();
404
405 /* Make sure that timer wheel updates are propagated */
406 if ((idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
407 if (!in_irq())
408 tick_nohz_irq_exit();
409 }
410#endif
411}
412
413static inline void __irq_exit_rcu(void)
414{
415#ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
416 local_irq_disable();
417#else
418 lockdep_assert_irqs_disabled();
419#endif
420 account_irq_exit_time(current);
421 preempt_count_sub(HARDIRQ_OFFSET);
422 if (!in_interrupt() && local_softirq_pending())
423 invoke_softirq();
424
425 tick_irq_exit();
426}
427
428/**
429 * irq_exit_rcu() - Exit an interrupt context without updating RCU
430 *
431 * Also processes softirqs if needed and possible.
432 */
433void irq_exit_rcu(void)
434{
435 __irq_exit_rcu();
436 /* must be last! */
437 lockdep_hardirq_exit();
438}
439
440/**
441 * irq_exit - Exit an interrupt context, update RCU and lockdep
442 *
443 * Also processes softirqs if needed and possible.
444 */
445void irq_exit(void)
446{
447 __irq_exit_rcu();
448 rcu_irq_exit();
449 /* must be last! */
450 lockdep_hardirq_exit();
451}
452
453/*
454 * This function must run with irqs disabled!
455 */
456inline void raise_softirq_irqoff(unsigned int nr)
457{
458 __raise_softirq_irqoff(nr);
459
460 /*
461 * If we're in an interrupt or softirq, we're done
462 * (this also catches softirq-disabled code). We will
463 * actually run the softirq once we return from
464 * the irq or softirq.
465 *
466 * Otherwise we wake up ksoftirqd to make sure we
467 * schedule the softirq soon.
468 */
469 if (!in_interrupt())
470 wakeup_softirqd();
471}
472
473void raise_softirq(unsigned int nr)
474{
475 unsigned long flags;
476
477 local_irq_save(flags);
478 raise_softirq_irqoff(nr);
479 local_irq_restore(flags);
480}
481
482void __raise_softirq_irqoff(unsigned int nr)
483{
484 trace_softirq_raise(nr);
485 or_softirq_pending(1UL << nr);
486}
487
488void open_softirq(int nr, void (*action)(struct softirq_action *))
489{
490 softirq_vec[nr].action = action;
491}
492
493/*
494 * Tasklets
495 */
496struct tasklet_head {
497 struct tasklet_struct *head;
498 struct tasklet_struct **tail;
499};
500
501static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
502static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
503
504static void __tasklet_schedule_common(struct tasklet_struct *t,
505 struct tasklet_head __percpu *headp,
506 unsigned int softirq_nr)
507{
508 struct tasklet_head *head;
509 unsigned long flags;
510
511 local_irq_save(flags);
512 head = this_cpu_ptr(headp);
513 t->next = NULL;
514 *head->tail = t;
515 head->tail = &(t->next);
516 raise_softirq_irqoff(softirq_nr);
517 local_irq_restore(flags);
518}
519
520void __tasklet_schedule(struct tasklet_struct *t)
521{
522 __tasklet_schedule_common(t, &tasklet_vec,
523 TASKLET_SOFTIRQ);
524}
525EXPORT_SYMBOL(__tasklet_schedule);
526
527void __tasklet_hi_schedule(struct tasklet_struct *t)
528{
529 __tasklet_schedule_common(t, &tasklet_hi_vec,
530 HI_SOFTIRQ);
531}
532EXPORT_SYMBOL(__tasklet_hi_schedule);
533
534static void tasklet_action_common(struct softirq_action *a,
535 struct tasklet_head *tl_head,
536 unsigned int softirq_nr)
537{
538 struct tasklet_struct *list;
539
540 local_irq_disable();
541 list = tl_head->head;
542 tl_head->head = NULL;
543 tl_head->tail = &tl_head->head;
544 local_irq_enable();
545
546 while (list) {
547 struct tasklet_struct *t = list;
548
549 list = list->next;
550
551 if (tasklet_trylock(t)) {
552 if (!atomic_read(&t->count)) {
553 if (!test_and_clear_bit(TASKLET_STATE_SCHED,
554 &t->state))
555 BUG();
556 if (t->use_callback)
557 t->callback(t);
558 else
559 t->func(t->data);
560 tasklet_unlock(t);
561 continue;
562 }
563 tasklet_unlock(t);
564 }
565
566 local_irq_disable();
567 t->next = NULL;
568 *tl_head->tail = t;
569 tl_head->tail = &t->next;
570 __raise_softirq_irqoff(softirq_nr);
571 local_irq_enable();
572 }
573}
574
575static __latent_entropy void tasklet_action(struct softirq_action *a)
576{
577 tasklet_action_common(a, this_cpu_ptr(&tasklet_vec), TASKLET_SOFTIRQ);
578}
579
580static __latent_entropy void tasklet_hi_action(struct softirq_action *a)
581{
582 tasklet_action_common(a, this_cpu_ptr(&tasklet_hi_vec), HI_SOFTIRQ);
583}
584
585void tasklet_setup(struct tasklet_struct *t,
586 void (*callback)(struct tasklet_struct *))
587{
588 t->next = NULL;
589 t->state = 0;
590 atomic_set(&t->count, 0);
591 t->callback = callback;
592 t->use_callback = true;
593 t->data = 0;
594}
595EXPORT_SYMBOL(tasklet_setup);
596
597void tasklet_init(struct tasklet_struct *t,
598 void (*func)(unsigned long), unsigned long data)
599{
600 t->next = NULL;
601 t->state = 0;
602 atomic_set(&t->count, 0);
603 t->func = func;
604 t->use_callback = false;
605 t->data = data;
606}
607EXPORT_SYMBOL(tasklet_init);
608
609void tasklet_kill(struct tasklet_struct *t)
610{
611 if (in_interrupt())
612 pr_notice("Attempt to kill tasklet from interrupt\n");
613
614 while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
615 do {
616 yield();
617 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
618 }
619 tasklet_unlock_wait(t);
620 clear_bit(TASKLET_STATE_SCHED, &t->state);
621}
622EXPORT_SYMBOL(tasklet_kill);
623
624void __init softirq_init(void)
625{
626 int cpu;
627
628 for_each_possible_cpu(cpu) {
629 per_cpu(tasklet_vec, cpu).tail =
630 &per_cpu(tasklet_vec, cpu).head;
631 per_cpu(tasklet_hi_vec, cpu).tail =
632 &per_cpu(tasklet_hi_vec, cpu).head;
633 }
634
635 open_softirq(TASKLET_SOFTIRQ, tasklet_action);
636 open_softirq(HI_SOFTIRQ, tasklet_hi_action);
637}
638
639static int ksoftirqd_should_run(unsigned int cpu)
640{
641 return local_softirq_pending();
642}
643
644static void run_ksoftirqd(unsigned int cpu)
645{
646 local_irq_disable();
647 if (local_softirq_pending()) {
648 /*
649 * We can safely run softirq on inline stack, as we are not deep
650 * in the task stack here.
651 */
652 __do_softirq();
653 local_irq_enable();
654 cond_resched();
655 return;
656 }
657 local_irq_enable();
658}
659
660#ifdef CONFIG_HOTPLUG_CPU
661/*
662 * tasklet_kill_immediate is called to remove a tasklet which can already be
663 * scheduled for execution on @cpu.
664 *
665 * Unlike tasklet_kill, this function removes the tasklet
666 * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
667 *
668 * When this function is called, @cpu must be in the CPU_DEAD state.
669 */
670void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
671{
672 struct tasklet_struct **i;
673
674 BUG_ON(cpu_online(cpu));
675 BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
676
677 if (!test_bit(TASKLET_STATE_SCHED, &t->state))
678 return;
679
680 /* CPU is dead, so no lock needed. */
681 for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
682 if (*i == t) {
683 *i = t->next;
684 /* If this was the tail element, move the tail ptr */
685 if (*i == NULL)
686 per_cpu(tasklet_vec, cpu).tail = i;
687 return;
688 }
689 }
690 BUG();
691}
692
693static int takeover_tasklets(unsigned int cpu)
694{
695 /* CPU is dead, so no lock needed. */
696 local_irq_disable();
697
698 /* Find end, append list for that CPU. */
699 if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
700 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
701 __this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
702 per_cpu(tasklet_vec, cpu).head = NULL;
703 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
704 }
705 raise_softirq_irqoff(TASKLET_SOFTIRQ);
706
707 if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
708 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
709 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
710 per_cpu(tasklet_hi_vec, cpu).head = NULL;
711 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
712 }
713 raise_softirq_irqoff(HI_SOFTIRQ);
714
715 local_irq_enable();
716 return 0;
717}
718#else
719#define takeover_tasklets NULL
720#endif /* CONFIG_HOTPLUG_CPU */
721
722static struct smp_hotplug_thread softirq_threads = {
723 .store = &ksoftirqd,
724 .thread_should_run = ksoftirqd_should_run,
725 .thread_fn = run_ksoftirqd,
726 .thread_comm = "ksoftirqd/%u",
727};
728
729static __init int spawn_ksoftirqd(void)
730{
731 cpuhp_setup_state_nocalls(CPUHP_SOFTIRQ_DEAD, "softirq:dead", NULL,
732 takeover_tasklets);
733 BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
734
735 return 0;
736}
737early_initcall(spawn_ksoftirqd);
738
739/*
740 * [ These __weak aliases are kept in a separate compilation unit, so that
741 * GCC does not inline them incorrectly. ]
742 */
743
744int __init __weak early_irq_init(void)
745{
746 return 0;
747}
748
749int __init __weak arch_probe_nr_irqs(void)
750{
751 return NR_IRQS_LEGACY;
752}
753
754int __init __weak arch_early_irq_init(void)
755{
756 return 0;
757}
758
759unsigned int __weak arch_dynirq_lower_bound(unsigned int from)
760{
761 return from;
762}