irq: Handle spurios irq detection for threaded irqs
[linux-3.10.git] / kernel / irq / spurious.c
1 /*
2  * linux/kernel/irq/spurious.c
3  *
4  * Copyright (C) 1992, 1998-2004 Linus Torvalds, Ingo Molnar
5  *
6  * This file contains spurious interrupt handling.
7  */
8
9 #include <linux/jiffies.h>
10 #include <linux/irq.h>
11 #include <linux/module.h>
12 #include <linux/kallsyms.h>
13 #include <linux/interrupt.h>
14 #include <linux/moduleparam.h>
15 #include <linux/timer.h>
16
17 #include "internals.h"
18
19 static int irqfixup __read_mostly;
20
21 #define POLL_SPURIOUS_IRQ_INTERVAL (HZ/10)
22 static void poll_spurious_irqs(unsigned long dummy);
23 static DEFINE_TIMER(poll_spurious_irq_timer, poll_spurious_irqs, 0, 0);
24 static int irq_poll_cpu;
25 static atomic_t irq_poll_active;
26
27 /*
28  * We wait here for a poller to finish.
29  *
30  * If the poll runs on this CPU, then we yell loudly and return
31  * false. That will leave the interrupt line disabled in the worst
32  * case, but it should never happen.
33  *
34  * We wait until the poller is done and then recheck disabled and
35  * action (about to be disabled). Only if it's still active, we return
36  * true and let the handler run.
37  */
38 bool irq_wait_for_poll(struct irq_desc *desc)
39 {
40         if (WARN_ONCE(irq_poll_cpu == smp_processor_id(),
41                       "irq poll in progress on cpu %d for irq %d\n",
42                       smp_processor_id(), desc->irq_data.irq))
43                 return false;
44
45 #ifdef CONFIG_SMP
46         do {
47                 raw_spin_unlock(&desc->lock);
48                 while (irqd_irq_inprogress(&desc->irq_data))
49                         cpu_relax();
50                 raw_spin_lock(&desc->lock);
51         } while (irqd_irq_inprogress(&desc->irq_data));
52         /* Might have been disabled in meantime */
53         return !irqd_irq_disabled(&desc->irq_data) && desc->action;
54 #else
55         return false;
56 #endif
57 }
58
59
60 /*
61  * Recovery handler for misrouted interrupts.
62  */
63 static int try_one_irq(int irq, struct irq_desc *desc, bool force)
64 {
65         irqreturn_t ret = IRQ_NONE;
66         struct irqaction *action;
67
68         raw_spin_lock(&desc->lock);
69
70         /* PER_CPU and nested thread interrupts are never polled */
71         if (irq_settings_is_per_cpu(desc) || irq_settings_is_nested_thread(desc))
72                 goto out;
73
74         /*
75          * Do not poll disabled interrupts unless the spurious
76          * disabled poller asks explicitely.
77          */
78         if (irqd_irq_disabled(&desc->irq_data) && !force)
79                 goto out;
80
81         /*
82          * All handlers must agree on IRQF_SHARED, so we test just the
83          * first. Check for action->next as well.
84          */
85         action = desc->action;
86         if (!action || !(action->flags & IRQF_SHARED) ||
87             (action->flags & __IRQF_TIMER) || !action->next)
88                 goto out;
89
90         /* Already running on another processor */
91         if (irqd_irq_inprogress(&desc->irq_data)) {
92                 /*
93                  * Already running: If it is shared get the other
94                  * CPU to go looking for our mystery interrupt too
95                  */
96                 desc->istate |= IRQS_PENDING;
97                 goto out;
98         }
99
100         /* Mark it poll in progress */
101         desc->istate |= IRQS_POLL_INPROGRESS;
102         do {
103                 if (handle_irq_event(desc) == IRQ_HANDLED)
104                         ret = IRQ_HANDLED;
105                 action = desc->action;
106         } while ((desc->istate & IRQS_PENDING) && action);
107         desc->istate &= ~IRQS_POLL_INPROGRESS;
108 out:
109         raw_spin_unlock(&desc->lock);
110         return ret == IRQ_HANDLED;
111 }
112
113 static int misrouted_irq(int irq)
114 {
115         struct irq_desc *desc;
116         int i, ok = 0;
117
118         if (atomic_inc_return(&irq_poll_active) == 1)
119                 goto out;
120
121         irq_poll_cpu = smp_processor_id();
122
123         for_each_irq_desc(i, desc) {
124                 if (!i)
125                          continue;
126
127                 if (i == irq)   /* Already tried */
128                         continue;
129
130                 if (try_one_irq(i, desc, false))
131                         ok = 1;
132         }
133 out:
134         atomic_dec(&irq_poll_active);
135         /* So the caller can adjust the irq error counts */
136         return ok;
137 }
138
139 static void poll_spurious_irqs(unsigned long dummy)
140 {
141         struct irq_desc *desc;
142         int i;
143
144         if (atomic_inc_return(&irq_poll_active) != 1)
145                 goto out;
146         irq_poll_cpu = smp_processor_id();
147
148         for_each_irq_desc(i, desc) {
149                 unsigned int state;
150
151                 if (!i)
152                          continue;
153
154                 /* Racy but it doesn't matter */
155                 state = desc->istate;
156                 barrier();
157                 if (!(state & IRQS_SPURIOUS_DISABLED))
158                         continue;
159
160                 local_irq_disable();
161                 try_one_irq(i, desc, true);
162                 local_irq_enable();
163         }
164 out:
165         atomic_dec(&irq_poll_active);
166         mod_timer(&poll_spurious_irq_timer,
167                   jiffies + POLL_SPURIOUS_IRQ_INTERVAL);
168 }
169
170 static inline int bad_action_ret(irqreturn_t action_ret)
171 {
172         if (likely(action_ret <= (IRQ_HANDLED | IRQ_WAKE_THREAD)))
173                 return 0;
174         return 1;
175 }
176
177 /*
178  * If 99,900 of the previous 100,000 interrupts have not been handled
179  * then assume that the IRQ is stuck in some manner. Drop a diagnostic
180  * and try to turn the IRQ off.
181  *
182  * (The other 100-of-100,000 interrupts may have been a correctly
183  *  functioning device sharing an IRQ with the failing one)
184  */
185 static void
186 __report_bad_irq(unsigned int irq, struct irq_desc *desc,
187                  irqreturn_t action_ret)
188 {
189         struct irqaction *action;
190         unsigned long flags;
191
192         if (bad_action_ret(action_ret)) {
193                 printk(KERN_ERR "irq event %d: bogus return value %x\n",
194                                 irq, action_ret);
195         } else {
196                 printk(KERN_ERR "irq %d: nobody cared (try booting with "
197                                 "the \"irqpoll\" option)\n", irq);
198         }
199         dump_stack();
200         printk(KERN_ERR "handlers:\n");
201
202         /*
203          * We need to take desc->lock here. note_interrupt() is called
204          * w/o desc->lock held, but IRQ_PROGRESS set. We might race
205          * with something else removing an action. It's ok to take
206          * desc->lock here. See synchronize_irq().
207          */
208         raw_spin_lock_irqsave(&desc->lock, flags);
209         action = desc->action;
210         while (action) {
211                 printk(KERN_ERR "[<%p>] %pf", action->handler, action->handler);
212                 if (action->thread_fn)
213                         printk(KERN_CONT " threaded [<%p>] %pf",
214                                         action->thread_fn, action->thread_fn);
215                 printk(KERN_CONT "\n");
216                 action = action->next;
217         }
218         raw_spin_unlock_irqrestore(&desc->lock, flags);
219 }
220
221 static void
222 report_bad_irq(unsigned int irq, struct irq_desc *desc, irqreturn_t action_ret)
223 {
224         static int count = 100;
225
226         if (count > 0) {
227                 count--;
228                 __report_bad_irq(irq, desc, action_ret);
229         }
230 }
231
232 static inline int
233 try_misrouted_irq(unsigned int irq, struct irq_desc *desc,
234                   irqreturn_t action_ret)
235 {
236         struct irqaction *action;
237
238         if (!irqfixup)
239                 return 0;
240
241         /* We didn't actually handle the IRQ - see if it was misrouted? */
242         if (action_ret == IRQ_NONE)
243                 return 1;
244
245         /*
246          * But for 'irqfixup == 2' we also do it for handled interrupts if
247          * they are marked as IRQF_IRQPOLL (or for irq zero, which is the
248          * traditional PC timer interrupt.. Legacy)
249          */
250         if (irqfixup < 2)
251                 return 0;
252
253         if (!irq)
254                 return 1;
255
256         /*
257          * Since we don't get the descriptor lock, "action" can
258          * change under us.  We don't really care, but we don't
259          * want to follow a NULL pointer. So tell the compiler to
260          * just load it once by using a barrier.
261          */
262         action = desc->action;
263         barrier();
264         return action && (action->flags & IRQF_IRQPOLL);
265 }
266
267 void note_interrupt(unsigned int irq, struct irq_desc *desc,
268                     irqreturn_t action_ret)
269 {
270         if (desc->istate & IRQS_POLL_INPROGRESS)
271                 return;
272
273         /* we get here again via the threaded handler */
274         if (action_ret == IRQ_WAKE_THREAD)
275                 return;
276
277         if (bad_action_ret(action_ret)) {
278                 report_bad_irq(irq, desc, action_ret);
279                 return;
280         }
281
282         if (unlikely(action_ret == IRQ_NONE)) {
283                 /*
284                  * If we are seeing only the odd spurious IRQ caused by
285                  * bus asynchronicity then don't eventually trigger an error,
286                  * otherwise the counter becomes a doomsday timer for otherwise
287                  * working systems
288                  */
289                 if (time_after(jiffies, desc->last_unhandled + HZ/10))
290                         desc->irqs_unhandled = 1;
291                 else
292                         desc->irqs_unhandled++;
293                 desc->last_unhandled = jiffies;
294         }
295
296         if (unlikely(try_misrouted_irq(irq, desc, action_ret))) {
297                 int ok = misrouted_irq(irq);
298                 if (action_ret == IRQ_NONE)
299                         desc->irqs_unhandled -= ok;
300         }
301
302         desc->irq_count++;
303         if (likely(desc->irq_count < 100000))
304                 return;
305
306         desc->irq_count = 0;
307         if (unlikely(desc->irqs_unhandled > 99900)) {
308                 /*
309                  * The interrupt is stuck
310                  */
311                 __report_bad_irq(irq, desc, action_ret);
312                 /*
313                  * Now kill the IRQ
314                  */
315                 printk(KERN_EMERG "Disabling IRQ #%d\n", irq);
316                 desc->istate |= IRQS_SPURIOUS_DISABLED;
317                 desc->depth++;
318                 irq_disable(desc);
319
320                 mod_timer(&poll_spurious_irq_timer,
321                           jiffies + POLL_SPURIOUS_IRQ_INTERVAL);
322         }
323         desc->irqs_unhandled = 0;
324 }
325
326 int noirqdebug __read_mostly;
327
328 int noirqdebug_setup(char *str)
329 {
330         noirqdebug = 1;
331         printk(KERN_INFO "IRQ lockup detection disabled\n");
332
333         return 1;
334 }
335
336 __setup("noirqdebug", noirqdebug_setup);
337 module_param(noirqdebug, bool, 0644);
338 MODULE_PARM_DESC(noirqdebug, "Disable irq lockup detection when true");
339
340 static int __init irqfixup_setup(char *str)
341 {
342         irqfixup = 1;
343         printk(KERN_WARNING "Misrouted IRQ fixup support enabled.\n");
344         printk(KERN_WARNING "This may impact system performance.\n");
345
346         return 1;
347 }
348
349 __setup("irqfixup", irqfixup_setup);
350 module_param(irqfixup, int, 0644);
351
352 static int __init irqpoll_setup(char *str)
353 {
354         irqfixup = 2;
355         printk(KERN_WARNING "Misrouted IRQ fixup and polling support "
356                                 "enabled\n");
357         printk(KERN_WARNING "This may significantly impact system "
358                                 "performance\n");
359         return 1;
360 }
361
362 __setup("irqpoll", irqpoll_setup);