Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/kernel/softirq.c
4 *
5 * Copyright (C) 1992 Linus Torvalds
6 *
7 * Rewritten. Old one was good in 2.2, but in 2.3 it was immoral. --ANK (990903)
8 */
9
10#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
12#include <linux/export.h>
13#include <linux/kernel_stat.h>
14#include <linux/interrupt.h>
15#include <linux/init.h>
16#include <linux/local_lock.h>
17#include <linux/mm.h>
18#include <linux/notifier.h>
19#include <linux/percpu.h>
20#include <linux/cpu.h>
21#include <linux/freezer.h>
22#include <linux/kthread.h>
23#include <linux/rcupdate.h>
24#include <linux/ftrace.h>
25#include <linux/smp.h>
26#include <linux/smpboot.h>
27#include <linux/tick.h>
28#include <linux/irq.h>
29#include <linux/wait_bit.h>
30#include <linux/workqueue.h>
31
32#include <asm/softirq_stack.h>
33
34#define CREATE_TRACE_POINTS
35#include <trace/events/irq.h>
36
37/*
38 - No shared variables, all the data are CPU local.
39 - If a softirq needs serialization, let it serialize itself
40 by its own spinlocks.
41 - Even if softirq is serialized, only local cpu is marked for
42 execution. Hence, we get something sort of weak cpu binding.
43 Though it is still not clear, will it result in better locality
44 or will not.
45
46 Examples:
47 - NET RX softirq. It is multithreaded and does not require
48 any global serialization.
49 - NET TX softirq. It kicks software netdevice queues, hence
50 it is logically serialized per device, but this serialization
51 is invisible to common code.
52 - Tasklets: serialized wrt itself.
53 */
54
55#ifndef __ARCH_IRQ_STAT
56DEFINE_PER_CPU_ALIGNED(irq_cpustat_t, irq_stat);
57EXPORT_PER_CPU_SYMBOL(irq_stat);
58#endif
59
60static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp;
61
62DEFINE_PER_CPU(struct task_struct *, ksoftirqd);
63
64const char * const softirq_to_name[NR_SOFTIRQS] = {
65 "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "IRQ_POLL",
66 "TASKLET", "SCHED", "HRTIMER", "RCU"
67};
68
69/*
70 * we cannot loop indefinitely here to avoid userspace starvation,
71 * but we also don't want to introduce a worst case 1/HZ latency
72 * to the pending events, so lets the scheduler to balance
73 * the softirq load for us.
74 */
75static void wakeup_softirqd(void)
76{
77 /* Interrupts are disabled: no need to stop preemption */
78 struct task_struct *tsk = __this_cpu_read(ksoftirqd);
79
80 if (tsk)
81 wake_up_process(tsk);
82}
83
84#ifdef CONFIG_TRACE_IRQFLAGS
85DEFINE_PER_CPU(int, hardirqs_enabled);
86DEFINE_PER_CPU(int, hardirq_context);
87EXPORT_PER_CPU_SYMBOL_GPL(hardirqs_enabled);
88EXPORT_PER_CPU_SYMBOL_GPL(hardirq_context);
89#endif
90
91/*
92 * SOFTIRQ_OFFSET usage:
93 *
94 * On !RT kernels 'count' is the preempt counter, on RT kernels this applies
95 * to a per CPU counter and to task::softirqs_disabled_cnt.
96 *
97 * - count is changed by SOFTIRQ_OFFSET on entering or leaving softirq
98 * processing.
99 *
100 * - count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
101 * on local_bh_disable or local_bh_enable.
102 *
103 * This lets us distinguish between whether we are currently processing
104 * softirq and whether we just have bh disabled.
105 */
106#ifdef CONFIG_PREEMPT_RT
107
108/*
109 * RT accounts for BH disabled sections in task::softirqs_disabled_cnt and
110 * also in per CPU softirq_ctrl::cnt. This is necessary to allow tasks in a
111 * softirq disabled section to be preempted.
112 *
113 * The per task counter is used for softirq_count(), in_softirq() and
114 * in_serving_softirqs() because these counts are only valid when the task
115 * holding softirq_ctrl::lock is running.
116 *
117 * The per CPU counter prevents pointless wakeups of ksoftirqd in case that
118 * the task which is in a softirq disabled section is preempted or blocks.
119 */
120struct softirq_ctrl {
121 local_lock_t lock;
122 int cnt;
123};
124
125static DEFINE_PER_CPU(struct softirq_ctrl, softirq_ctrl) = {
126 .lock = INIT_LOCAL_LOCK(softirq_ctrl.lock),
127};
128
129/**
130 * local_bh_blocked() - Check for idle whether BH processing is blocked
131 *
132 * Returns false if the per CPU softirq::cnt is 0 otherwise true.
133 *
134 * This is invoked from the idle task to guard against false positive
135 * softirq pending warnings, which would happen when the task which holds
136 * softirq_ctrl::lock was the only running task on the CPU and blocks on
137 * some other lock.
138 */
139bool local_bh_blocked(void)
140{
141 return __this_cpu_read(softirq_ctrl.cnt) != 0;
142}
143
144void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
145{
146 unsigned long flags;
147 int newcnt;
148
149 WARN_ON_ONCE(in_hardirq());
150
151 /* First entry of a task into a BH disabled section? */
152 if (!current->softirq_disable_cnt) {
153 if (preemptible()) {
154 local_lock(&softirq_ctrl.lock);
155 /* Required to meet the RCU bottomhalf requirements. */
156 rcu_read_lock();
157 } else {
158 DEBUG_LOCKS_WARN_ON(this_cpu_read(softirq_ctrl.cnt));
159 }
160 }
161
162 /*
163 * Track the per CPU softirq disabled state. On RT this is per CPU
164 * state to allow preemption of bottom half disabled sections.
165 */
166 newcnt = __this_cpu_add_return(softirq_ctrl.cnt, cnt);
167 /*
168 * Reflect the result in the task state to prevent recursion on the
169 * local lock and to make softirq_count() & al work.
170 */
171 current->softirq_disable_cnt = newcnt;
172
173 if (IS_ENABLED(CONFIG_TRACE_IRQFLAGS) && newcnt == cnt) {
174 raw_local_irq_save(flags);
175 lockdep_softirqs_off(ip);
176 raw_local_irq_restore(flags);
177 }
178}
179EXPORT_SYMBOL(__local_bh_disable_ip);
180
181static void __local_bh_enable(unsigned int cnt, bool unlock)
182{
183 unsigned long flags;
184 int newcnt;
185
186 DEBUG_LOCKS_WARN_ON(current->softirq_disable_cnt !=
187 this_cpu_read(softirq_ctrl.cnt));
188
189 if (IS_ENABLED(CONFIG_TRACE_IRQFLAGS) && softirq_count() == cnt) {
190 raw_local_irq_save(flags);
191 lockdep_softirqs_on(_RET_IP_);
192 raw_local_irq_restore(flags);
193 }
194
195 newcnt = __this_cpu_sub_return(softirq_ctrl.cnt, cnt);
196 current->softirq_disable_cnt = newcnt;
197
198 if (!newcnt && unlock) {
199 rcu_read_unlock();
200 local_unlock(&softirq_ctrl.lock);
201 }
202}
203
204void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
205{
206 bool preempt_on = preemptible();
207 unsigned long flags;
208 u32 pending;
209 int curcnt;
210
211 WARN_ON_ONCE(in_hardirq());
212 lockdep_assert_irqs_enabled();
213
214 local_irq_save(flags);
215 curcnt = __this_cpu_read(softirq_ctrl.cnt);
216
217 /*
218 * If this is not reenabling soft interrupts, no point in trying to
219 * run pending ones.
220 */
221 if (curcnt != cnt)
222 goto out;
223
224 pending = local_softirq_pending();
225 if (!pending)
226 goto out;
227
228 /*
229 * If this was called from non preemptible context, wake up the
230 * softirq daemon.
231 */
232 if (!preempt_on) {
233 wakeup_softirqd();
234 goto out;
235 }
236
237 /*
238 * Adjust softirq count to SOFTIRQ_OFFSET which makes
239 * in_serving_softirq() become true.
240 */
241 cnt = SOFTIRQ_OFFSET;
242 __local_bh_enable(cnt, false);
243 __do_softirq();
244
245out:
246 __local_bh_enable(cnt, preempt_on);
247 local_irq_restore(flags);
248}
249EXPORT_SYMBOL(__local_bh_enable_ip);
250
251/*
252 * Invoked from ksoftirqd_run() outside of the interrupt disabled section
253 * to acquire the per CPU local lock for reentrancy protection.
254 */
255static inline void ksoftirqd_run_begin(void)
256{
257 __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
258 local_irq_disable();
259}
260
261/* Counterpart to ksoftirqd_run_begin() */
262static inline void ksoftirqd_run_end(void)
263{
264 __local_bh_enable(SOFTIRQ_OFFSET, true);
265 WARN_ON_ONCE(in_interrupt());
266 local_irq_enable();
267}
268
269static inline void softirq_handle_begin(void) { }
270static inline void softirq_handle_end(void) { }
271
272static inline bool should_wake_ksoftirqd(void)
273{
274 return !this_cpu_read(softirq_ctrl.cnt);
275}
276
277static inline void invoke_softirq(void)
278{
279 if (should_wake_ksoftirqd())
280 wakeup_softirqd();
281}
282
283#define SCHED_SOFTIRQ_MASK BIT(SCHED_SOFTIRQ)
284
285/*
286 * flush_smp_call_function_queue() can raise a soft interrupt in a function
287 * call. On RT kernels this is undesired and the only known functionalities
288 * are in the block layer which is disabled on RT, and in the scheduler for
289 * idle load balancing. If soft interrupts get raised which haven't been
290 * raised before the flush, warn if it is not a SCHED_SOFTIRQ so it can be
291 * investigated.
292 */
293void do_softirq_post_smp_call_flush(unsigned int was_pending)
294{
295 unsigned int is_pending = local_softirq_pending();
296
297 if (unlikely(was_pending != is_pending)) {
298 WARN_ON_ONCE(was_pending != (is_pending & ~SCHED_SOFTIRQ_MASK));
299 invoke_softirq();
300 }
301}
302
303#else /* CONFIG_PREEMPT_RT */
304
305/*
306 * This one is for softirq.c-internal use, where hardirqs are disabled
307 * legitimately:
308 */
309#ifdef CONFIG_TRACE_IRQFLAGS
310void __local_bh_disable_ip(unsigned long ip, unsigned int cnt)
311{
312 unsigned long flags;
313
314 WARN_ON_ONCE(in_hardirq());
315
316 raw_local_irq_save(flags);
317 /*
318 * The preempt tracer hooks into preempt_count_add and will break
319 * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
320 * is set and before current->softirq_enabled is cleared.
321 * We must manually increment preempt_count here and manually
322 * call the trace_preempt_off later.
323 */
324 __preempt_count_add(cnt);
325 /*
326 * Were softirqs turned off above:
327 */
328 if (softirq_count() == (cnt & SOFTIRQ_MASK))
329 lockdep_softirqs_off(ip);
330 raw_local_irq_restore(flags);
331
332 if (preempt_count() == cnt) {
333#ifdef CONFIG_DEBUG_PREEMPT
334 current->preempt_disable_ip = get_lock_parent_ip();
335#endif
336 trace_preempt_off(CALLER_ADDR0, get_lock_parent_ip());
337 }
338}
339EXPORT_SYMBOL(__local_bh_disable_ip);
340#endif /* CONFIG_TRACE_IRQFLAGS */
341
342static void __local_bh_enable(unsigned int cnt)
343{
344 lockdep_assert_irqs_disabled();
345
346 if (preempt_count() == cnt)
347 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
348
349 if (softirq_count() == (cnt & SOFTIRQ_MASK))
350 lockdep_softirqs_on(_RET_IP_);
351
352 __preempt_count_sub(cnt);
353}
354
355/*
356 * Special-case - softirqs can safely be enabled by __do_softirq(),
357 * without processing still-pending softirqs:
358 */
359void _local_bh_enable(void)
360{
361 WARN_ON_ONCE(in_hardirq());
362 __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
363}
364EXPORT_SYMBOL(_local_bh_enable);
365
366void __local_bh_enable_ip(unsigned long ip, unsigned int cnt)
367{
368 WARN_ON_ONCE(in_hardirq());
369 lockdep_assert_irqs_enabled();
370#ifdef CONFIG_TRACE_IRQFLAGS
371 local_irq_disable();
372#endif
373 /*
374 * Are softirqs going to be turned on now:
375 */
376 if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
377 lockdep_softirqs_on(ip);
378 /*
379 * Keep preemption disabled until we are done with
380 * softirq processing:
381 */
382 __preempt_count_sub(cnt - 1);
383
384 if (unlikely(!in_interrupt() && local_softirq_pending())) {
385 /*
386 * Run softirq if any pending. And do it in its own stack
387 * as we may be calling this deep in a task call stack already.
388 */
389 do_softirq();
390 }
391
392 preempt_count_dec();
393#ifdef CONFIG_TRACE_IRQFLAGS
394 local_irq_enable();
395#endif
396 preempt_check_resched();
397}
398EXPORT_SYMBOL(__local_bh_enable_ip);
399
400static inline void softirq_handle_begin(void)
401{
402 __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
403}
404
405static inline void softirq_handle_end(void)
406{
407 __local_bh_enable(SOFTIRQ_OFFSET);
408 WARN_ON_ONCE(in_interrupt());
409}
410
411static inline void ksoftirqd_run_begin(void)
412{
413 local_irq_disable();
414}
415
416static inline void ksoftirqd_run_end(void)
417{
418 local_irq_enable();
419}
420
421static inline bool should_wake_ksoftirqd(void)
422{
423 return true;
424}
425
426static inline void invoke_softirq(void)
427{
428 if (!force_irqthreads() || !__this_cpu_read(ksoftirqd)) {
429#ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK
430 /*
431 * We can safely execute softirq on the current stack if
432 * it is the irq stack, because it should be near empty
433 * at this stage.
434 */
435 __do_softirq();
436#else
437 /*
438 * Otherwise, irq_exit() is called on the task stack that can
439 * be potentially deep already. So call softirq in its own stack
440 * to prevent from any overrun.
441 */
442 do_softirq_own_stack();
443#endif
444 } else {
445 wakeup_softirqd();
446 }
447}
448
449asmlinkage __visible void do_softirq(void)
450{
451 __u32 pending;
452 unsigned long flags;
453
454 if (in_interrupt())
455 return;
456
457 local_irq_save(flags);
458
459 pending = local_softirq_pending();
460
461 if (pending)
462 do_softirq_own_stack();
463
464 local_irq_restore(flags);
465}
466
467#endif /* !CONFIG_PREEMPT_RT */
468
469/*
470 * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times,
471 * but break the loop if need_resched() is set or after 2 ms.
472 * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in
473 * certain cases, such as stop_machine(), jiffies may cease to
474 * increment and so we need the MAX_SOFTIRQ_RESTART limit as
475 * well to make sure we eventually return from this method.
476 *
477 * These limits have been established via experimentation.
478 * The two things to balance is latency against fairness -
479 * we want to handle softirqs as soon as possible, but they
480 * should not be able to lock up the box.
481 */
482#define MAX_SOFTIRQ_TIME msecs_to_jiffies(2)
483#define MAX_SOFTIRQ_RESTART 10
484
485#ifdef CONFIG_TRACE_IRQFLAGS
486/*
487 * When we run softirqs from irq_exit() and thus on the hardirq stack we need
488 * to keep the lockdep irq context tracking as tight as possible in order to
489 * not miss-qualify lock contexts and miss possible deadlocks.
490 */
491
492static inline bool lockdep_softirq_start(void)
493{
494 bool in_hardirq = false;
495
496 if (lockdep_hardirq_context()) {
497 in_hardirq = true;
498 lockdep_hardirq_exit();
499 }
500
501 lockdep_softirq_enter();
502
503 return in_hardirq;
504}
505
506static inline void lockdep_softirq_end(bool in_hardirq)
507{
508 lockdep_softirq_exit();
509
510 if (in_hardirq)
511 lockdep_hardirq_enter();
512}
513#else
514static inline bool lockdep_softirq_start(void) { return false; }
515static inline void lockdep_softirq_end(bool in_hardirq) { }
516#endif
517
518static void handle_softirqs(bool ksirqd)
519{
520 unsigned long end = jiffies + MAX_SOFTIRQ_TIME;
521 unsigned long old_flags = current->flags;
522 int max_restart = MAX_SOFTIRQ_RESTART;
523 struct softirq_action *h;
524 bool in_hardirq;
525 __u32 pending;
526 int softirq_bit;
527
528 /*
529 * Mask out PF_MEMALLOC as the current task context is borrowed for the
530 * softirq. A softirq handled, such as network RX, might set PF_MEMALLOC
531 * again if the socket is related to swapping.
532 */
533 current->flags &= ~PF_MEMALLOC;
534
535 pending = local_softirq_pending();
536
537 softirq_handle_begin();
538 in_hardirq = lockdep_softirq_start();
539 account_softirq_enter(current);
540
541restart:
542 /* Reset the pending bitmask before enabling irqs */
543 set_softirq_pending(0);
544
545 local_irq_enable();
546
547 h = softirq_vec;
548
549 while ((softirq_bit = ffs(pending))) {
550 unsigned int vec_nr;
551 int prev_count;
552
553 h += softirq_bit - 1;
554
555 vec_nr = h - softirq_vec;
556 prev_count = preempt_count();
557
558 kstat_incr_softirqs_this_cpu(vec_nr);
559
560 trace_softirq_entry(vec_nr);
561 h->action();
562 trace_softirq_exit(vec_nr);
563 if (unlikely(prev_count != preempt_count())) {
564 pr_err("huh, entered softirq %u %s %p with preempt_count %08x, exited with %08x?\n",
565 vec_nr, softirq_to_name[vec_nr], h->action,
566 prev_count, preempt_count());
567 preempt_count_set(prev_count);
568 }
569 h++;
570 pending >>= softirq_bit;
571 }
572
573 if (!IS_ENABLED(CONFIG_PREEMPT_RT) && ksirqd)
574 rcu_softirq_qs();
575
576 local_irq_disable();
577
578 pending = local_softirq_pending();
579 if (pending) {
580 if (time_before(jiffies, end) && !need_resched() &&
581 --max_restart)
582 goto restart;
583
584 wakeup_softirqd();
585 }
586
587 account_softirq_exit(current);
588 lockdep_softirq_end(in_hardirq);
589 softirq_handle_end();
590 current_restore_flags(old_flags, PF_MEMALLOC);
591}
592
593asmlinkage __visible void __softirq_entry __do_softirq(void)
594{
595 handle_softirqs(false);
596}
597
598/**
599 * irq_enter_rcu - Enter an interrupt context with RCU watching
600 */
601void irq_enter_rcu(void)
602{
603 __irq_enter_raw();
604
605 if (tick_nohz_full_cpu(smp_processor_id()) ||
606 (is_idle_task(current) && (irq_count() == HARDIRQ_OFFSET)))
607 tick_irq_enter();
608
609 account_hardirq_enter(current);
610}
611
612/**
613 * irq_enter - Enter an interrupt context including RCU update
614 */
615void irq_enter(void)
616{
617 ct_irq_enter();
618 irq_enter_rcu();
619}
620
621static inline void tick_irq_exit(void)
622{
623#ifdef CONFIG_NO_HZ_COMMON
624 int cpu = smp_processor_id();
625
626 /* Make sure that timer wheel updates are propagated */
627 if ((sched_core_idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) {
628 if (!in_hardirq())
629 tick_nohz_irq_exit();
630 }
631#endif
632}
633
634#ifdef CONFIG_IRQ_FORCED_THREADING
635DEFINE_PER_CPU(struct task_struct *, ktimerd);
636DEFINE_PER_CPU(unsigned long, pending_timer_softirq);
637
638static void wake_timersd(void)
639{
640 struct task_struct *tsk = __this_cpu_read(ktimerd);
641
642 if (tsk)
643 wake_up_process(tsk);
644}
645
646#else
647
648static inline void wake_timersd(void) { }
649
650#endif
651
652static inline void __irq_exit_rcu(void)
653{
654#ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
655 local_irq_disable();
656#else
657 lockdep_assert_irqs_disabled();
658#endif
659 account_hardirq_exit(current);
660 preempt_count_sub(HARDIRQ_OFFSET);
661 if (!in_interrupt() && local_softirq_pending())
662 invoke_softirq();
663
664 if (IS_ENABLED(CONFIG_IRQ_FORCED_THREADING) && force_irqthreads() &&
665 local_timers_pending_force_th() && !(in_nmi() | in_hardirq()))
666 wake_timersd();
667
668 tick_irq_exit();
669}
670
671/**
672 * irq_exit_rcu() - Exit an interrupt context without updating RCU
673 *
674 * Also processes softirqs if needed and possible.
675 */
676void irq_exit_rcu(void)
677{
678 __irq_exit_rcu();
679 /* must be last! */
680 lockdep_hardirq_exit();
681}
682
683/**
684 * irq_exit - Exit an interrupt context, update RCU and lockdep
685 *
686 * Also processes softirqs if needed and possible.
687 */
688void irq_exit(void)
689{
690 __irq_exit_rcu();
691 ct_irq_exit();
692 /* must be last! */
693 lockdep_hardirq_exit();
694}
695
696/*
697 * This function must run with irqs disabled!
698 */
699inline void raise_softirq_irqoff(unsigned int nr)
700{
701 __raise_softirq_irqoff(nr);
702
703 /*
704 * If we're in an interrupt or softirq, we're done
705 * (this also catches softirq-disabled code). We will
706 * actually run the softirq once we return from
707 * the irq or softirq.
708 *
709 * Otherwise we wake up ksoftirqd to make sure we
710 * schedule the softirq soon.
711 */
712 if (!in_interrupt() && should_wake_ksoftirqd())
713 wakeup_softirqd();
714}
715
716void raise_softirq(unsigned int nr)
717{
718 unsigned long flags;
719
720 local_irq_save(flags);
721 raise_softirq_irqoff(nr);
722 local_irq_restore(flags);
723}
724
725void __raise_softirq_irqoff(unsigned int nr)
726{
727 lockdep_assert_irqs_disabled();
728 trace_softirq_raise(nr);
729 or_softirq_pending(1UL << nr);
730}
731
732void open_softirq(int nr, void (*action)(void))
733{
734 softirq_vec[nr].action = action;
735}
736
737/*
738 * Tasklets
739 */
740struct tasklet_head {
741 struct tasklet_struct *head;
742 struct tasklet_struct **tail;
743};
744
745static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
746static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
747
748static void __tasklet_schedule_common(struct tasklet_struct *t,
749 struct tasklet_head __percpu *headp,
750 unsigned int softirq_nr)
751{
752 struct tasklet_head *head;
753 unsigned long flags;
754
755 local_irq_save(flags);
756 head = this_cpu_ptr(headp);
757 t->next = NULL;
758 *head->tail = t;
759 head->tail = &(t->next);
760 raise_softirq_irqoff(softirq_nr);
761 local_irq_restore(flags);
762}
763
764void __tasklet_schedule(struct tasklet_struct *t)
765{
766 __tasklet_schedule_common(t, &tasklet_vec,
767 TASKLET_SOFTIRQ);
768}
769EXPORT_SYMBOL(__tasklet_schedule);
770
771void __tasklet_hi_schedule(struct tasklet_struct *t)
772{
773 __tasklet_schedule_common(t, &tasklet_hi_vec,
774 HI_SOFTIRQ);
775}
776EXPORT_SYMBOL(__tasklet_hi_schedule);
777
778static bool tasklet_clear_sched(struct tasklet_struct *t)
779{
780 if (test_and_clear_wake_up_bit(TASKLET_STATE_SCHED, &t->state))
781 return true;
782
783 WARN_ONCE(1, "tasklet SCHED state not set: %s %pS\n",
784 t->use_callback ? "callback" : "func",
785 t->use_callback ? (void *)t->callback : (void *)t->func);
786
787 return false;
788}
789
790static void tasklet_action_common(struct tasklet_head *tl_head,
791 unsigned int softirq_nr)
792{
793 struct tasklet_struct *list;
794
795 local_irq_disable();
796 list = tl_head->head;
797 tl_head->head = NULL;
798 tl_head->tail = &tl_head->head;
799 local_irq_enable();
800
801 while (list) {
802 struct tasklet_struct *t = list;
803
804 list = list->next;
805
806 if (tasklet_trylock(t)) {
807 if (!atomic_read(&t->count)) {
808 if (tasklet_clear_sched(t)) {
809 if (t->use_callback) {
810 trace_tasklet_entry(t, t->callback);
811 t->callback(t);
812 trace_tasklet_exit(t, t->callback);
813 } else {
814 trace_tasklet_entry(t, t->func);
815 t->func(t->data);
816 trace_tasklet_exit(t, t->func);
817 }
818 }
819 tasklet_unlock(t);
820 continue;
821 }
822 tasklet_unlock(t);
823 }
824
825 local_irq_disable();
826 t->next = NULL;
827 *tl_head->tail = t;
828 tl_head->tail = &t->next;
829 __raise_softirq_irqoff(softirq_nr);
830 local_irq_enable();
831 }
832}
833
834static __latent_entropy void tasklet_action(void)
835{
836 workqueue_softirq_action(false);
837 tasklet_action_common(this_cpu_ptr(&tasklet_vec), TASKLET_SOFTIRQ);
838}
839
840static __latent_entropy void tasklet_hi_action(void)
841{
842 workqueue_softirq_action(true);
843 tasklet_action_common(this_cpu_ptr(&tasklet_hi_vec), HI_SOFTIRQ);
844}
845
846void tasklet_setup(struct tasklet_struct *t,
847 void (*callback)(struct tasklet_struct *))
848{
849 t->next = NULL;
850 t->state = 0;
851 atomic_set(&t->count, 0);
852 t->callback = callback;
853 t->use_callback = true;
854 t->data = 0;
855}
856EXPORT_SYMBOL(tasklet_setup);
857
858void tasklet_init(struct tasklet_struct *t,
859 void (*func)(unsigned long), unsigned long data)
860{
861 t->next = NULL;
862 t->state = 0;
863 atomic_set(&t->count, 0);
864 t->func = func;
865 t->use_callback = false;
866 t->data = data;
867}
868EXPORT_SYMBOL(tasklet_init);
869
870#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
871/*
872 * Do not use in new code. Waiting for tasklets from atomic contexts is
873 * error prone and should be avoided.
874 */
875void tasklet_unlock_spin_wait(struct tasklet_struct *t)
876{
877 while (test_bit(TASKLET_STATE_RUN, &(t)->state)) {
878 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
879 /*
880 * Prevent a live lock when current preempted soft
881 * interrupt processing or prevents ksoftirqd from
882 * running. If the tasklet runs on a different CPU
883 * then this has no effect other than doing the BH
884 * disable/enable dance for nothing.
885 */
886 local_bh_disable();
887 local_bh_enable();
888 } else {
889 cpu_relax();
890 }
891 }
892}
893EXPORT_SYMBOL(tasklet_unlock_spin_wait);
894#endif
895
896void tasklet_kill(struct tasklet_struct *t)
897{
898 if (in_interrupt())
899 pr_notice("Attempt to kill tasklet from interrupt\n");
900
901 wait_on_bit_lock(&t->state, TASKLET_STATE_SCHED, TASK_UNINTERRUPTIBLE);
902
903 tasklet_unlock_wait(t);
904 tasklet_clear_sched(t);
905}
906EXPORT_SYMBOL(tasklet_kill);
907
908#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
909void tasklet_unlock(struct tasklet_struct *t)
910{
911 clear_and_wake_up_bit(TASKLET_STATE_RUN, &t->state);
912}
913EXPORT_SYMBOL_GPL(tasklet_unlock);
914
915void tasklet_unlock_wait(struct tasklet_struct *t)
916{
917 wait_on_bit(&t->state, TASKLET_STATE_RUN, TASK_UNINTERRUPTIBLE);
918}
919EXPORT_SYMBOL_GPL(tasklet_unlock_wait);
920#endif
921
922void __init softirq_init(void)
923{
924 int cpu;
925
926 for_each_possible_cpu(cpu) {
927 per_cpu(tasklet_vec, cpu).tail =
928 &per_cpu(tasklet_vec, cpu).head;
929 per_cpu(tasklet_hi_vec, cpu).tail =
930 &per_cpu(tasklet_hi_vec, cpu).head;
931 }
932
933 open_softirq(TASKLET_SOFTIRQ, tasklet_action);
934 open_softirq(HI_SOFTIRQ, tasklet_hi_action);
935}
936
937static int ksoftirqd_should_run(unsigned int cpu)
938{
939 return local_softirq_pending();
940}
941
942static void run_ksoftirqd(unsigned int cpu)
943{
944 ksoftirqd_run_begin();
945 if (local_softirq_pending()) {
946 /*
947 * We can safely run softirq on inline stack, as we are not deep
948 * in the task stack here.
949 */
950 handle_softirqs(true);
951 ksoftirqd_run_end();
952 cond_resched();
953 return;
954 }
955 ksoftirqd_run_end();
956}
957
958#ifdef CONFIG_HOTPLUG_CPU
959static int takeover_tasklets(unsigned int cpu)
960{
961 workqueue_softirq_dead(cpu);
962
963 /* CPU is dead, so no lock needed. */
964 local_irq_disable();
965
966 /* Find end, append list for that CPU. */
967 if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
968 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
969 __this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
970 per_cpu(tasklet_vec, cpu).head = NULL;
971 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
972 }
973 raise_softirq_irqoff(TASKLET_SOFTIRQ);
974
975 if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
976 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
977 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
978 per_cpu(tasklet_hi_vec, cpu).head = NULL;
979 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
980 }
981 raise_softirq_irqoff(HI_SOFTIRQ);
982
983 local_irq_enable();
984 return 0;
985}
986#else
987#define takeover_tasklets NULL
988#endif /* CONFIG_HOTPLUG_CPU */
989
990static struct smp_hotplug_thread softirq_threads = {
991 .store = &ksoftirqd,
992 .thread_should_run = ksoftirqd_should_run,
993 .thread_fn = run_ksoftirqd,
994 .thread_comm = "ksoftirqd/%u",
995};
996
997#ifdef CONFIG_IRQ_FORCED_THREADING
998static void ktimerd_setup(unsigned int cpu)
999{
1000 /* Above SCHED_NORMAL to handle timers before regular tasks. */
1001 sched_set_fifo_low(current);
1002}
1003
1004static int ktimerd_should_run(unsigned int cpu)
1005{
1006 return local_timers_pending_force_th();
1007}
1008
1009void raise_ktimers_thread(unsigned int nr)
1010{
1011 trace_softirq_raise(nr);
1012 __this_cpu_or(pending_timer_softirq, BIT(nr));
1013}
1014
1015static void run_ktimerd(unsigned int cpu)
1016{
1017 unsigned int timer_si;
1018
1019 ksoftirqd_run_begin();
1020
1021 timer_si = local_timers_pending_force_th();
1022 __this_cpu_write(pending_timer_softirq, 0);
1023 or_softirq_pending(timer_si);
1024
1025 __do_softirq();
1026
1027 ksoftirqd_run_end();
1028}
1029
1030static struct smp_hotplug_thread timer_thread = {
1031 .store = &ktimerd,
1032 .setup = ktimerd_setup,
1033 .thread_should_run = ktimerd_should_run,
1034 .thread_fn = run_ktimerd,
1035 .thread_comm = "ktimers/%u",
1036};
1037#endif
1038
1039static __init int spawn_ksoftirqd(void)
1040{
1041 cpuhp_setup_state_nocalls(CPUHP_SOFTIRQ_DEAD, "softirq:dead", NULL,
1042 takeover_tasklets);
1043 BUG_ON(smpboot_register_percpu_thread(&softirq_threads));
1044#ifdef CONFIG_IRQ_FORCED_THREADING
1045 if (force_irqthreads())
1046 BUG_ON(smpboot_register_percpu_thread(&timer_thread));
1047#endif
1048 return 0;
1049}
1050early_initcall(spawn_ksoftirqd);
1051
1052/*
1053 * [ These __weak aliases are kept in a separate compilation unit, so that
1054 * GCC does not inline them incorrectly. ]
1055 */
1056
1057int __init __weak early_irq_init(void)
1058{
1059 return 0;
1060}
1061
1062int __init __weak arch_probe_nr_irqs(void)
1063{
1064 return NR_IRQS_LEGACY;
1065}
1066
1067int __init __weak arch_early_irq_init(void)
1068{
1069 return 0;
1070}
1071
1072unsigned int __weak arch_dynirq_lower_bound(unsigned int from)
1073{
1074 return from;
1075}
1/*
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 * Remote softirq infrastructure is by Jens Axboe.
11 */
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/tick.h>
27
28#define CREATE_TRACE_POINTS
29#include <trace/events/irq.h>
30
31#include <asm/irq.h>
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
51irq_cpustat_t irq_stat[NR_CPUS] ____cacheline_aligned;
52EXPORT_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
59char *softirq_to_name[NR_SOFTIRQS] = {
60 "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "BLOCK_IOPOLL",
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 * preempt_count and SOFTIRQ_OFFSET usage:
81 * - preempt_count is changed by SOFTIRQ_OFFSET on entering or leaving
82 * softirq processing.
83 * - preempt_count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET)
84 * on local_bh_disable or local_bh_enable.
85 * This lets us distinguish between whether we are currently processing
86 * softirq and whether we just have bh disabled.
87 */
88
89/*
90 * This one is for softirq.c-internal use,
91 * where hardirqs are disabled legitimately:
92 */
93#ifdef CONFIG_TRACE_IRQFLAGS
94static void __local_bh_disable(unsigned long ip, unsigned int cnt)
95{
96 unsigned long flags;
97
98 WARN_ON_ONCE(in_irq());
99
100 raw_local_irq_save(flags);
101 /*
102 * The preempt tracer hooks into add_preempt_count and will break
103 * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET
104 * is set and before current->softirq_enabled is cleared.
105 * We must manually increment preempt_count here and manually
106 * call the trace_preempt_off later.
107 */
108 preempt_count() += cnt;
109 /*
110 * Were softirqs turned off above:
111 */
112 if (softirq_count() == cnt)
113 trace_softirqs_off(ip);
114 raw_local_irq_restore(flags);
115
116 if (preempt_count() == cnt)
117 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
118}
119#else /* !CONFIG_TRACE_IRQFLAGS */
120static inline void __local_bh_disable(unsigned long ip, unsigned int cnt)
121{
122 add_preempt_count(cnt);
123 barrier();
124}
125#endif /* CONFIG_TRACE_IRQFLAGS */
126
127void local_bh_disable(void)
128{
129 __local_bh_disable((unsigned long)__builtin_return_address(0),
130 SOFTIRQ_DISABLE_OFFSET);
131}
132
133EXPORT_SYMBOL(local_bh_disable);
134
135static void __local_bh_enable(unsigned int cnt)
136{
137 WARN_ON_ONCE(in_irq());
138 WARN_ON_ONCE(!irqs_disabled());
139
140 if (softirq_count() == cnt)
141 trace_softirqs_on((unsigned long)__builtin_return_address(0));
142 sub_preempt_count(cnt);
143}
144
145/*
146 * Special-case - softirqs can safely be enabled in
147 * cond_resched_softirq(), or by __do_softirq(),
148 * without processing still-pending softirqs:
149 */
150void _local_bh_enable(void)
151{
152 __local_bh_enable(SOFTIRQ_DISABLE_OFFSET);
153}
154
155EXPORT_SYMBOL(_local_bh_enable);
156
157static inline void _local_bh_enable_ip(unsigned long ip)
158{
159 WARN_ON_ONCE(in_irq() || irqs_disabled());
160#ifdef CONFIG_TRACE_IRQFLAGS
161 local_irq_disable();
162#endif
163 /*
164 * Are softirqs going to be turned on now:
165 */
166 if (softirq_count() == SOFTIRQ_DISABLE_OFFSET)
167 trace_softirqs_on(ip);
168 /*
169 * Keep preemption disabled until we are done with
170 * softirq processing:
171 */
172 sub_preempt_count(SOFTIRQ_DISABLE_OFFSET - 1);
173
174 if (unlikely(!in_interrupt() && local_softirq_pending()))
175 do_softirq();
176
177 dec_preempt_count();
178#ifdef CONFIG_TRACE_IRQFLAGS
179 local_irq_enable();
180#endif
181 preempt_check_resched();
182}
183
184void local_bh_enable(void)
185{
186 _local_bh_enable_ip((unsigned long)__builtin_return_address(0));
187}
188EXPORT_SYMBOL(local_bh_enable);
189
190void local_bh_enable_ip(unsigned long ip)
191{
192 _local_bh_enable_ip(ip);
193}
194EXPORT_SYMBOL(local_bh_enable_ip);
195
196/*
197 * We restart softirq processing MAX_SOFTIRQ_RESTART times,
198 * and we fall back to softirqd after that.
199 *
200 * This number has been established via experimentation.
201 * The two things to balance is latency against fairness -
202 * we want to handle softirqs as soon as possible, but they
203 * should not be able to lock up the box.
204 */
205#define MAX_SOFTIRQ_RESTART 10
206
207asmlinkage void __do_softirq(void)
208{
209 struct softirq_action *h;
210 __u32 pending;
211 int max_restart = MAX_SOFTIRQ_RESTART;
212 int cpu;
213
214 pending = local_softirq_pending();
215 account_system_vtime(current);
216
217 __local_bh_disable((unsigned long)__builtin_return_address(0),
218 SOFTIRQ_OFFSET);
219 lockdep_softirq_enter();
220
221 cpu = smp_processor_id();
222restart:
223 /* Reset the pending bitmask before enabling irqs */
224 set_softirq_pending(0);
225
226 local_irq_enable();
227
228 h = softirq_vec;
229
230 do {
231 if (pending & 1) {
232 unsigned int vec_nr = h - softirq_vec;
233 int prev_count = preempt_count();
234
235 kstat_incr_softirqs_this_cpu(vec_nr);
236
237 trace_softirq_entry(vec_nr);
238 h->action(h);
239 trace_softirq_exit(vec_nr);
240 if (unlikely(prev_count != preempt_count())) {
241 printk(KERN_ERR "huh, entered softirq %u %s %p"
242 "with preempt_count %08x,"
243 " exited with %08x?\n", vec_nr,
244 softirq_to_name[vec_nr], h->action,
245 prev_count, preempt_count());
246 preempt_count() = prev_count;
247 }
248
249 rcu_bh_qs(cpu);
250 }
251 h++;
252 pending >>= 1;
253 } while (pending);
254
255 local_irq_disable();
256
257 pending = local_softirq_pending();
258 if (pending && --max_restart)
259 goto restart;
260
261 if (pending)
262 wakeup_softirqd();
263
264 lockdep_softirq_exit();
265
266 account_system_vtime(current);
267 __local_bh_enable(SOFTIRQ_OFFSET);
268}
269
270#ifndef __ARCH_HAS_DO_SOFTIRQ
271
272asmlinkage void do_softirq(void)
273{
274 __u32 pending;
275 unsigned long flags;
276
277 if (in_interrupt())
278 return;
279
280 local_irq_save(flags);
281
282 pending = local_softirq_pending();
283
284 if (pending)
285 __do_softirq();
286
287 local_irq_restore(flags);
288}
289
290#endif
291
292/*
293 * Enter an interrupt context.
294 */
295void irq_enter(void)
296{
297 int cpu = smp_processor_id();
298
299 rcu_irq_enter();
300 if (is_idle_task(current) && !in_interrupt()) {
301 /*
302 * Prevent raise_softirq from needlessly waking up ksoftirqd
303 * here, as softirq will be serviced on return from interrupt.
304 */
305 local_bh_disable();
306 tick_check_idle(cpu);
307 _local_bh_enable();
308 }
309
310 __irq_enter();
311}
312
313static inline void invoke_softirq(void)
314{
315 if (!force_irqthreads) {
316#ifdef __ARCH_IRQ_EXIT_IRQS_DISABLED
317 __do_softirq();
318#else
319 do_softirq();
320#endif
321 } else {
322 __local_bh_disable((unsigned long)__builtin_return_address(0),
323 SOFTIRQ_OFFSET);
324 wakeup_softirqd();
325 __local_bh_enable(SOFTIRQ_OFFSET);
326 }
327}
328
329/*
330 * Exit an interrupt context. Process softirqs if needed and possible:
331 */
332void irq_exit(void)
333{
334 account_system_vtime(current);
335 trace_hardirq_exit();
336 sub_preempt_count(IRQ_EXIT_OFFSET);
337 if (!in_interrupt() && local_softirq_pending())
338 invoke_softirq();
339
340#ifdef CONFIG_NO_HZ
341 /* Make sure that timer wheel updates are propagated */
342 if (idle_cpu(smp_processor_id()) && !in_interrupt() && !need_resched())
343 tick_nohz_irq_exit();
344#endif
345 rcu_irq_exit();
346 sched_preempt_enable_no_resched();
347}
348
349/*
350 * This function must run with irqs disabled!
351 */
352inline void raise_softirq_irqoff(unsigned int nr)
353{
354 __raise_softirq_irqoff(nr);
355
356 /*
357 * If we're in an interrupt or softirq, we're done
358 * (this also catches softirq-disabled code). We will
359 * actually run the softirq once we return from
360 * the irq or softirq.
361 *
362 * Otherwise we wake up ksoftirqd to make sure we
363 * schedule the softirq soon.
364 */
365 if (!in_interrupt())
366 wakeup_softirqd();
367}
368
369void raise_softirq(unsigned int nr)
370{
371 unsigned long flags;
372
373 local_irq_save(flags);
374 raise_softirq_irqoff(nr);
375 local_irq_restore(flags);
376}
377
378void __raise_softirq_irqoff(unsigned int nr)
379{
380 trace_softirq_raise(nr);
381 or_softirq_pending(1UL << nr);
382}
383
384void open_softirq(int nr, void (*action)(struct softirq_action *))
385{
386 softirq_vec[nr].action = action;
387}
388
389/*
390 * Tasklets
391 */
392struct tasklet_head
393{
394 struct tasklet_struct *head;
395 struct tasklet_struct **tail;
396};
397
398static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec);
399static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec);
400
401void __tasklet_schedule(struct tasklet_struct *t)
402{
403 unsigned long flags;
404
405 local_irq_save(flags);
406 t->next = NULL;
407 *__this_cpu_read(tasklet_vec.tail) = t;
408 __this_cpu_write(tasklet_vec.tail, &(t->next));
409 raise_softirq_irqoff(TASKLET_SOFTIRQ);
410 local_irq_restore(flags);
411}
412
413EXPORT_SYMBOL(__tasklet_schedule);
414
415void __tasklet_hi_schedule(struct tasklet_struct *t)
416{
417 unsigned long flags;
418
419 local_irq_save(flags);
420 t->next = NULL;
421 *__this_cpu_read(tasklet_hi_vec.tail) = t;
422 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
423 raise_softirq_irqoff(HI_SOFTIRQ);
424 local_irq_restore(flags);
425}
426
427EXPORT_SYMBOL(__tasklet_hi_schedule);
428
429void __tasklet_hi_schedule_first(struct tasklet_struct *t)
430{
431 BUG_ON(!irqs_disabled());
432
433 t->next = __this_cpu_read(tasklet_hi_vec.head);
434 __this_cpu_write(tasklet_hi_vec.head, t);
435 __raise_softirq_irqoff(HI_SOFTIRQ);
436}
437
438EXPORT_SYMBOL(__tasklet_hi_schedule_first);
439
440static void tasklet_action(struct softirq_action *a)
441{
442 struct tasklet_struct *list;
443
444 local_irq_disable();
445 list = __this_cpu_read(tasklet_vec.head);
446 __this_cpu_write(tasklet_vec.head, NULL);
447 __this_cpu_write(tasklet_vec.tail, &__get_cpu_var(tasklet_vec).head);
448 local_irq_enable();
449
450 while (list) {
451 struct tasklet_struct *t = list;
452
453 list = list->next;
454
455 if (tasklet_trylock(t)) {
456 if (!atomic_read(&t->count)) {
457 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
458 BUG();
459 t->func(t->data);
460 tasklet_unlock(t);
461 continue;
462 }
463 tasklet_unlock(t);
464 }
465
466 local_irq_disable();
467 t->next = NULL;
468 *__this_cpu_read(tasklet_vec.tail) = t;
469 __this_cpu_write(tasklet_vec.tail, &(t->next));
470 __raise_softirq_irqoff(TASKLET_SOFTIRQ);
471 local_irq_enable();
472 }
473}
474
475static void tasklet_hi_action(struct softirq_action *a)
476{
477 struct tasklet_struct *list;
478
479 local_irq_disable();
480 list = __this_cpu_read(tasklet_hi_vec.head);
481 __this_cpu_write(tasklet_hi_vec.head, NULL);
482 __this_cpu_write(tasklet_hi_vec.tail, &__get_cpu_var(tasklet_hi_vec).head);
483 local_irq_enable();
484
485 while (list) {
486 struct tasklet_struct *t = list;
487
488 list = list->next;
489
490 if (tasklet_trylock(t)) {
491 if (!atomic_read(&t->count)) {
492 if (!test_and_clear_bit(TASKLET_STATE_SCHED, &t->state))
493 BUG();
494 t->func(t->data);
495 tasklet_unlock(t);
496 continue;
497 }
498 tasklet_unlock(t);
499 }
500
501 local_irq_disable();
502 t->next = NULL;
503 *__this_cpu_read(tasklet_hi_vec.tail) = t;
504 __this_cpu_write(tasklet_hi_vec.tail, &(t->next));
505 __raise_softirq_irqoff(HI_SOFTIRQ);
506 local_irq_enable();
507 }
508}
509
510
511void tasklet_init(struct tasklet_struct *t,
512 void (*func)(unsigned long), unsigned long data)
513{
514 t->next = NULL;
515 t->state = 0;
516 atomic_set(&t->count, 0);
517 t->func = func;
518 t->data = data;
519}
520
521EXPORT_SYMBOL(tasklet_init);
522
523void tasklet_kill(struct tasklet_struct *t)
524{
525 if (in_interrupt())
526 printk("Attempt to kill tasklet from interrupt\n");
527
528 while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) {
529 do {
530 yield();
531 } while (test_bit(TASKLET_STATE_SCHED, &t->state));
532 }
533 tasklet_unlock_wait(t);
534 clear_bit(TASKLET_STATE_SCHED, &t->state);
535}
536
537EXPORT_SYMBOL(tasklet_kill);
538
539/*
540 * tasklet_hrtimer
541 */
542
543/*
544 * The trampoline is called when the hrtimer expires. It schedules a tasklet
545 * to run __tasklet_hrtimer_trampoline() which in turn will call the intended
546 * hrtimer callback, but from softirq context.
547 */
548static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer)
549{
550 struct tasklet_hrtimer *ttimer =
551 container_of(timer, struct tasklet_hrtimer, timer);
552
553 tasklet_hi_schedule(&ttimer->tasklet);
554 return HRTIMER_NORESTART;
555}
556
557/*
558 * Helper function which calls the hrtimer callback from
559 * tasklet/softirq context
560 */
561static void __tasklet_hrtimer_trampoline(unsigned long data)
562{
563 struct tasklet_hrtimer *ttimer = (void *)data;
564 enum hrtimer_restart restart;
565
566 restart = ttimer->function(&ttimer->timer);
567 if (restart != HRTIMER_NORESTART)
568 hrtimer_restart(&ttimer->timer);
569}
570
571/**
572 * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks
573 * @ttimer: tasklet_hrtimer which is initialized
574 * @function: hrtimer callback function which gets called from softirq context
575 * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME)
576 * @mode: hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL)
577 */
578void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer,
579 enum hrtimer_restart (*function)(struct hrtimer *),
580 clockid_t which_clock, enum hrtimer_mode mode)
581{
582 hrtimer_init(&ttimer->timer, which_clock, mode);
583 ttimer->timer.function = __hrtimer_tasklet_trampoline;
584 tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline,
585 (unsigned long)ttimer);
586 ttimer->function = function;
587}
588EXPORT_SYMBOL_GPL(tasklet_hrtimer_init);
589
590/*
591 * Remote softirq bits
592 */
593
594DEFINE_PER_CPU(struct list_head [NR_SOFTIRQS], softirq_work_list);
595EXPORT_PER_CPU_SYMBOL(softirq_work_list);
596
597static void __local_trigger(struct call_single_data *cp, int softirq)
598{
599 struct list_head *head = &__get_cpu_var(softirq_work_list[softirq]);
600
601 list_add_tail(&cp->list, head);
602
603 /* Trigger the softirq only if the list was previously empty. */
604 if (head->next == &cp->list)
605 raise_softirq_irqoff(softirq);
606}
607
608#ifdef CONFIG_USE_GENERIC_SMP_HELPERS
609static void remote_softirq_receive(void *data)
610{
611 struct call_single_data *cp = data;
612 unsigned long flags;
613 int softirq;
614
615 softirq = cp->priv;
616
617 local_irq_save(flags);
618 __local_trigger(cp, softirq);
619 local_irq_restore(flags);
620}
621
622static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
623{
624 if (cpu_online(cpu)) {
625 cp->func = remote_softirq_receive;
626 cp->info = cp;
627 cp->flags = 0;
628 cp->priv = softirq;
629
630 __smp_call_function_single(cpu, cp, 0);
631 return 0;
632 }
633 return 1;
634}
635#else /* CONFIG_USE_GENERIC_SMP_HELPERS */
636static int __try_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
637{
638 return 1;
639}
640#endif
641
642/**
643 * __send_remote_softirq - try to schedule softirq work on a remote cpu
644 * @cp: private SMP call function data area
645 * @cpu: the remote cpu
646 * @this_cpu: the currently executing cpu
647 * @softirq: the softirq for the work
648 *
649 * Attempt to schedule softirq work on a remote cpu. If this cannot be
650 * done, the work is instead queued up on the local cpu.
651 *
652 * Interrupts must be disabled.
653 */
654void __send_remote_softirq(struct call_single_data *cp, int cpu, int this_cpu, int softirq)
655{
656 if (cpu == this_cpu || __try_remote_softirq(cp, cpu, softirq))
657 __local_trigger(cp, softirq);
658}
659EXPORT_SYMBOL(__send_remote_softirq);
660
661/**
662 * send_remote_softirq - try to schedule softirq work on a remote cpu
663 * @cp: private SMP call function data area
664 * @cpu: the remote cpu
665 * @softirq: the softirq for the work
666 *
667 * Like __send_remote_softirq except that disabling interrupts and
668 * computing the current cpu is done for the caller.
669 */
670void send_remote_softirq(struct call_single_data *cp, int cpu, int softirq)
671{
672 unsigned long flags;
673 int this_cpu;
674
675 local_irq_save(flags);
676 this_cpu = smp_processor_id();
677 __send_remote_softirq(cp, cpu, this_cpu, softirq);
678 local_irq_restore(flags);
679}
680EXPORT_SYMBOL(send_remote_softirq);
681
682static int __cpuinit remote_softirq_cpu_notify(struct notifier_block *self,
683 unsigned long action, void *hcpu)
684{
685 /*
686 * If a CPU goes away, splice its entries to the current CPU
687 * and trigger a run of the softirq
688 */
689 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
690 int cpu = (unsigned long) hcpu;
691 int i;
692
693 local_irq_disable();
694 for (i = 0; i < NR_SOFTIRQS; i++) {
695 struct list_head *head = &per_cpu(softirq_work_list[i], cpu);
696 struct list_head *local_head;
697
698 if (list_empty(head))
699 continue;
700
701 local_head = &__get_cpu_var(softirq_work_list[i]);
702 list_splice_init(head, local_head);
703 raise_softirq_irqoff(i);
704 }
705 local_irq_enable();
706 }
707
708 return NOTIFY_OK;
709}
710
711static struct notifier_block __cpuinitdata remote_softirq_cpu_notifier = {
712 .notifier_call = remote_softirq_cpu_notify,
713};
714
715void __init softirq_init(void)
716{
717 int cpu;
718
719 for_each_possible_cpu(cpu) {
720 int i;
721
722 per_cpu(tasklet_vec, cpu).tail =
723 &per_cpu(tasklet_vec, cpu).head;
724 per_cpu(tasklet_hi_vec, cpu).tail =
725 &per_cpu(tasklet_hi_vec, cpu).head;
726 for (i = 0; i < NR_SOFTIRQS; i++)
727 INIT_LIST_HEAD(&per_cpu(softirq_work_list[i], cpu));
728 }
729
730 register_hotcpu_notifier(&remote_softirq_cpu_notifier);
731
732 open_softirq(TASKLET_SOFTIRQ, tasklet_action);
733 open_softirq(HI_SOFTIRQ, tasklet_hi_action);
734}
735
736static int run_ksoftirqd(void * __bind_cpu)
737{
738 set_current_state(TASK_INTERRUPTIBLE);
739
740 while (!kthread_should_stop()) {
741 preempt_disable();
742 if (!local_softirq_pending()) {
743 schedule_preempt_disabled();
744 }
745
746 __set_current_state(TASK_RUNNING);
747
748 while (local_softirq_pending()) {
749 /* Preempt disable stops cpu going offline.
750 If already offline, we'll be on wrong CPU:
751 don't process */
752 if (cpu_is_offline((long)__bind_cpu))
753 goto wait_to_die;
754 local_irq_disable();
755 if (local_softirq_pending())
756 __do_softirq();
757 local_irq_enable();
758 sched_preempt_enable_no_resched();
759 cond_resched();
760 preempt_disable();
761 rcu_note_context_switch((long)__bind_cpu);
762 }
763 preempt_enable();
764 set_current_state(TASK_INTERRUPTIBLE);
765 }
766 __set_current_state(TASK_RUNNING);
767 return 0;
768
769wait_to_die:
770 preempt_enable();
771 /* Wait for kthread_stop */
772 set_current_state(TASK_INTERRUPTIBLE);
773 while (!kthread_should_stop()) {
774 schedule();
775 set_current_state(TASK_INTERRUPTIBLE);
776 }
777 __set_current_state(TASK_RUNNING);
778 return 0;
779}
780
781#ifdef CONFIG_HOTPLUG_CPU
782/*
783 * tasklet_kill_immediate is called to remove a tasklet which can already be
784 * scheduled for execution on @cpu.
785 *
786 * Unlike tasklet_kill, this function removes the tasklet
787 * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state.
788 *
789 * When this function is called, @cpu must be in the CPU_DEAD state.
790 */
791void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu)
792{
793 struct tasklet_struct **i;
794
795 BUG_ON(cpu_online(cpu));
796 BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state));
797
798 if (!test_bit(TASKLET_STATE_SCHED, &t->state))
799 return;
800
801 /* CPU is dead, so no lock needed. */
802 for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) {
803 if (*i == t) {
804 *i = t->next;
805 /* If this was the tail element, move the tail ptr */
806 if (*i == NULL)
807 per_cpu(tasklet_vec, cpu).tail = i;
808 return;
809 }
810 }
811 BUG();
812}
813
814static void takeover_tasklets(unsigned int cpu)
815{
816 /* CPU is dead, so no lock needed. */
817 local_irq_disable();
818
819 /* Find end, append list for that CPU. */
820 if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
821 *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head;
822 this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail);
823 per_cpu(tasklet_vec, cpu).head = NULL;
824 per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
825 }
826 raise_softirq_irqoff(TASKLET_SOFTIRQ);
827
828 if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
829 *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head;
830 __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail);
831 per_cpu(tasklet_hi_vec, cpu).head = NULL;
832 per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
833 }
834 raise_softirq_irqoff(HI_SOFTIRQ);
835
836 local_irq_enable();
837}
838#endif /* CONFIG_HOTPLUG_CPU */
839
840static int __cpuinit cpu_callback(struct notifier_block *nfb,
841 unsigned long action,
842 void *hcpu)
843{
844 int hotcpu = (unsigned long)hcpu;
845 struct task_struct *p;
846
847 switch (action) {
848 case CPU_UP_PREPARE:
849 case CPU_UP_PREPARE_FROZEN:
850 p = kthread_create_on_node(run_ksoftirqd,
851 hcpu,
852 cpu_to_node(hotcpu),
853 "ksoftirqd/%d", hotcpu);
854 if (IS_ERR(p)) {
855 printk("ksoftirqd for %i failed\n", hotcpu);
856 return notifier_from_errno(PTR_ERR(p));
857 }
858 kthread_bind(p, hotcpu);
859 per_cpu(ksoftirqd, hotcpu) = p;
860 break;
861 case CPU_ONLINE:
862 case CPU_ONLINE_FROZEN:
863 wake_up_process(per_cpu(ksoftirqd, hotcpu));
864 break;
865#ifdef CONFIG_HOTPLUG_CPU
866 case CPU_UP_CANCELED:
867 case CPU_UP_CANCELED_FROZEN:
868 if (!per_cpu(ksoftirqd, hotcpu))
869 break;
870 /* Unbind so it can run. Fall thru. */
871 kthread_bind(per_cpu(ksoftirqd, hotcpu),
872 cpumask_any(cpu_online_mask));
873 case CPU_DEAD:
874 case CPU_DEAD_FROZEN: {
875 static const struct sched_param param = {
876 .sched_priority = MAX_RT_PRIO-1
877 };
878
879 p = per_cpu(ksoftirqd, hotcpu);
880 per_cpu(ksoftirqd, hotcpu) = NULL;
881 sched_setscheduler_nocheck(p, SCHED_FIFO, ¶m);
882 kthread_stop(p);
883 takeover_tasklets(hotcpu);
884 break;
885 }
886#endif /* CONFIG_HOTPLUG_CPU */
887 }
888 return NOTIFY_OK;
889}
890
891static struct notifier_block __cpuinitdata cpu_nfb = {
892 .notifier_call = cpu_callback
893};
894
895static __init int spawn_ksoftirqd(void)
896{
897 void *cpu = (void *)(long)smp_processor_id();
898 int err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
899
900 BUG_ON(err != NOTIFY_OK);
901 cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
902 register_cpu_notifier(&cpu_nfb);
903 return 0;
904}
905early_initcall(spawn_ksoftirqd);
906
907/*
908 * [ These __weak aliases are kept in a separate compilation unit, so that
909 * GCC does not inline them incorrectly. ]
910 */
911
912int __init __weak early_irq_init(void)
913{
914 return 0;
915}
916
917#ifdef CONFIG_GENERIC_HARDIRQS
918int __init __weak arch_probe_nr_irqs(void)
919{
920 return NR_IRQS_LEGACY;
921}
922
923int __init __weak arch_early_irq_init(void)
924{
925 return 0;
926}
927#endif