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 */
irq_wait_for_poll(struct irq_desc * desc)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 */
try_one_irq(int irq,struct irq_desc * desc,bool force)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.
84 */
85 action = desc->action;
86 if (!action || !(action->flags & IRQF_SHARED) ||
87 (action->flags & __IRQF_TIMER))
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 /* Make sure that there is still a valid action */
106 action = desc->action;
107 } while ((desc->istate & IRQS_PENDING) && action);
108 desc->istate &= ~IRQS_POLL_INPROGRESS;
109 out:
110 raw_spin_unlock(&desc->lock);
111 return ret == IRQ_HANDLED;
112 }
113
misrouted_irq(int irq)114 static int misrouted_irq(int irq)
115 {
116 struct irq_desc *desc;
117 int i, ok = 0;
118
119 if (atomic_inc_return(&irq_poll_active) != 1)
120 goto out;
121
122 irq_poll_cpu = smp_processor_id();
123
124 for_each_irq_desc(i, desc) {
125 if (!i)
126 continue;
127
128 if (i == irq) /* Already tried */
129 continue;
130
131 if (try_one_irq(i, desc, false))
132 ok = 1;
133 }
134 out:
135 atomic_dec(&irq_poll_active);
136 /* So the caller can adjust the irq error counts */
137 return ok;
138 }
139
poll_spurious_irqs(unsigned long dummy)140 static void poll_spurious_irqs(unsigned long dummy)
141 {
142 struct irq_desc *desc;
143 int i;
144
145 if (atomic_inc_return(&irq_poll_active) != 1)
146 goto out;
147 irq_poll_cpu = smp_processor_id();
148
149 for_each_irq_desc(i, desc) {
150 unsigned int state;
151
152 if (!i)
153 continue;
154
155 /* Racy but it doesn't matter */
156 state = desc->istate;
157 barrier();
158 if (!(state & IRQS_SPURIOUS_DISABLED))
159 continue;
160
161 local_irq_disable();
162 try_one_irq(i, desc, true);
163 local_irq_enable();
164 }
165 out:
166 atomic_dec(&irq_poll_active);
167 mod_timer(&poll_spurious_irq_timer,
168 jiffies + POLL_SPURIOUS_IRQ_INTERVAL);
169 }
170
bad_action_ret(irqreturn_t action_ret)171 static inline int bad_action_ret(irqreturn_t action_ret)
172 {
173 if (likely(action_ret <= (IRQ_HANDLED | IRQ_WAKE_THREAD)))
174 return 0;
175 return 1;
176 }
177
178 /*
179 * If 99,900 of the previous 100,000 interrupts have not been handled
180 * then assume that the IRQ is stuck in some manner. Drop a diagnostic
181 * and try to turn the IRQ off.
182 *
183 * (The other 100-of-100,000 interrupts may have been a correctly
184 * functioning device sharing an IRQ with the failing one)
185 */
186 static void
__report_bad_irq(unsigned int irq,struct irq_desc * desc,irqreturn_t action_ret)187 __report_bad_irq(unsigned int irq, struct irq_desc *desc,
188 irqreturn_t action_ret)
189 {
190 struct irqaction *action;
191 unsigned long flags;
192
193 if (bad_action_ret(action_ret)) {
194 printk(KERN_ERR "irq event %d: bogus return value %x\n",
195 irq, action_ret);
196 } else {
197 printk(KERN_ERR "irq %d: nobody cared (try booting with "
198 "the \"irqpoll\" option)\n", irq);
199 }
200 dump_stack();
201 printk(KERN_ERR "handlers:\n");
202
203 /*
204 * We need to take desc->lock here. note_interrupt() is called
205 * w/o desc->lock held, but IRQ_PROGRESS set. We might race
206 * with something else removing an action. It's ok to take
207 * desc->lock here. See synchronize_irq().
208 */
209 raw_spin_lock_irqsave(&desc->lock, flags);
210 action = desc->action;
211 while (action) {
212 printk(KERN_ERR "[<%p>] %pf", action->handler, action->handler);
213 if (action->thread_fn)
214 printk(KERN_CONT " threaded [<%p>] %pf",
215 action->thread_fn, action->thread_fn);
216 printk(KERN_CONT "\n");
217 action = action->next;
218 }
219 raw_spin_unlock_irqrestore(&desc->lock, flags);
220 }
221
222 static void
report_bad_irq(unsigned int irq,struct irq_desc * desc,irqreturn_t action_ret)223 report_bad_irq(unsigned int irq, struct irq_desc *desc, irqreturn_t action_ret)
224 {
225 static int count = 100;
226
227 if (count > 0) {
228 count--;
229 __report_bad_irq(irq, desc, action_ret);
230 }
231 }
232
233 static inline int
try_misrouted_irq(unsigned int irq,struct irq_desc * desc,irqreturn_t action_ret)234 try_misrouted_irq(unsigned int irq, struct irq_desc *desc,
235 irqreturn_t action_ret)
236 {
237 struct irqaction *action;
238
239 if (!irqfixup)
240 return 0;
241
242 /* We didn't actually handle the IRQ - see if it was misrouted? */
243 if (action_ret == IRQ_NONE)
244 return 1;
245
246 /*
247 * But for 'irqfixup == 2' we also do it for handled interrupts if
248 * they are marked as IRQF_IRQPOLL (or for irq zero, which is the
249 * traditional PC timer interrupt.. Legacy)
250 */
251 if (irqfixup < 2)
252 return 0;
253
254 if (!irq)
255 return 1;
256
257 /*
258 * Since we don't get the descriptor lock, "action" can
259 * change under us. We don't really care, but we don't
260 * want to follow a NULL pointer. So tell the compiler to
261 * just load it once by using a barrier.
262 */
263 action = desc->action;
264 barrier();
265 return action && (action->flags & IRQF_IRQPOLL);
266 }
267
268 #define SPURIOUS_DEFERRED 0x80000000
269
note_interrupt(unsigned int irq,struct irq_desc * desc,irqreturn_t action_ret)270 void note_interrupt(unsigned int irq, struct irq_desc *desc,
271 irqreturn_t action_ret)
272 {
273 if (desc->istate & IRQS_POLL_INPROGRESS)
274 return;
275
276 if (bad_action_ret(action_ret)) {
277 report_bad_irq(irq, desc, action_ret);
278 return;
279 }
280
281 /*
282 * We cannot call note_interrupt from the threaded handler
283 * because we need to look at the compound of all handlers
284 * (primary and threaded). Aside of that in the threaded
285 * shared case we have no serialization against an incoming
286 * hardware interrupt while we are dealing with a threaded
287 * result.
288 *
289 * So in case a thread is woken, we just note the fact and
290 * defer the analysis to the next hardware interrupt.
291 *
292 * The threaded handlers store whether they sucessfully
293 * handled an interrupt and we check whether that number
294 * changed versus the last invocation.
295 *
296 * We could handle all interrupts with the delayed by one
297 * mechanism, but for the non forced threaded case we'd just
298 * add pointless overhead to the straight hardirq interrupts
299 * for the sake of a few lines less code.
300 */
301 if (action_ret & IRQ_WAKE_THREAD) {
302 /*
303 * There is a thread woken. Check whether one of the
304 * shared primary handlers returned IRQ_HANDLED. If
305 * not we defer the spurious detection to the next
306 * interrupt.
307 */
308 if (action_ret == IRQ_WAKE_THREAD) {
309 int handled;
310 /*
311 * We use bit 31 of thread_handled_last to
312 * denote the deferred spurious detection
313 * active. No locking necessary as
314 * thread_handled_last is only accessed here
315 * and we have the guarantee that hard
316 * interrupts are not reentrant.
317 */
318 if (!(desc->threads_handled_last & SPURIOUS_DEFERRED)) {
319 desc->threads_handled_last |= SPURIOUS_DEFERRED;
320 return;
321 }
322 /*
323 * Check whether one of the threaded handlers
324 * returned IRQ_HANDLED since the last
325 * interrupt happened.
326 *
327 * For simplicity we just set bit 31, as it is
328 * set in threads_handled_last as well. So we
329 * avoid extra masking. And we really do not
330 * care about the high bits of the handled
331 * count. We just care about the count being
332 * different than the one we saw before.
333 */
334 handled = atomic_read(&desc->threads_handled);
335 handled |= SPURIOUS_DEFERRED;
336 if (handled != desc->threads_handled_last) {
337 action_ret = IRQ_HANDLED;
338 /*
339 * Note: We keep the SPURIOUS_DEFERRED
340 * bit set. We are handling the
341 * previous invocation right now.
342 * Keep it for the current one, so the
343 * next hardware interrupt will
344 * account for it.
345 */
346 desc->threads_handled_last = handled;
347 } else {
348 /*
349 * None of the threaded handlers felt
350 * responsible for the last interrupt
351 *
352 * We keep the SPURIOUS_DEFERRED bit
353 * set in threads_handled_last as we
354 * need to account for the current
355 * interrupt as well.
356 */
357 action_ret = IRQ_NONE;
358 }
359 } else {
360 /*
361 * One of the primary handlers returned
362 * IRQ_HANDLED. So we don't care about the
363 * threaded handlers on the same line. Clear
364 * the deferred detection bit.
365 *
366 * In theory we could/should check whether the
367 * deferred bit is set and take the result of
368 * the previous run into account here as
369 * well. But it's really not worth the
370 * trouble. If every other interrupt is
371 * handled we never trigger the spurious
372 * detector. And if this is just the one out
373 * of 100k unhandled ones which is handled
374 * then we merily delay the spurious detection
375 * by one hard interrupt. Not a real problem.
376 */
377 desc->threads_handled_last &= ~SPURIOUS_DEFERRED;
378 }
379 }
380
381 if (unlikely(action_ret == IRQ_NONE)) {
382 /*
383 * If we are seeing only the odd spurious IRQ caused by
384 * bus asynchronicity then don't eventually trigger an error,
385 * otherwise the counter becomes a doomsday timer for otherwise
386 * working systems
387 */
388 if (time_after(jiffies, desc->last_unhandled + HZ/10))
389 desc->irqs_unhandled = 1;
390 else
391 desc->irqs_unhandled++;
392 desc->last_unhandled = jiffies;
393 }
394
395 if (unlikely(try_misrouted_irq(irq, desc, action_ret))) {
396 int ok = misrouted_irq(irq);
397 if (action_ret == IRQ_NONE)
398 desc->irqs_unhandled -= ok;
399 }
400
401 desc->irq_count++;
402 if (likely(desc->irq_count < 100000))
403 return;
404
405 desc->irq_count = 0;
406 if (unlikely(desc->irqs_unhandled > 99900)) {
407 /*
408 * The interrupt is stuck
409 */
410 __report_bad_irq(irq, desc, action_ret);
411 /*
412 * Now kill the IRQ
413 */
414 printk(KERN_EMERG "Disabling IRQ #%d\n", irq);
415 desc->istate |= IRQS_SPURIOUS_DISABLED;
416 desc->depth++;
417 irq_disable(desc);
418
419 mod_timer(&poll_spurious_irq_timer,
420 jiffies + POLL_SPURIOUS_IRQ_INTERVAL);
421 }
422 desc->irqs_unhandled = 0;
423 }
424
425 bool noirqdebug __read_mostly;
426
noirqdebug_setup(char * str)427 int noirqdebug_setup(char *str)
428 {
429 noirqdebug = 1;
430 printk(KERN_INFO "IRQ lockup detection disabled\n");
431
432 return 1;
433 }
434
435 __setup("noirqdebug", noirqdebug_setup);
436 module_param(noirqdebug, bool, 0644);
437 MODULE_PARM_DESC(noirqdebug, "Disable irq lockup detection when true");
438
irqfixup_setup(char * str)439 static int __init irqfixup_setup(char *str)
440 {
441 irqfixup = 1;
442 printk(KERN_WARNING "Misrouted IRQ fixup support enabled.\n");
443 printk(KERN_WARNING "This may impact system performance.\n");
444
445 return 1;
446 }
447
448 __setup("irqfixup", irqfixup_setup);
449 module_param(irqfixup, int, 0644);
450
irqpoll_setup(char * str)451 static int __init irqpoll_setup(char *str)
452 {
453 irqfixup = 2;
454 printk(KERN_WARNING "Misrouted IRQ fixup and polling support "
455 "enabled\n");
456 printk(KERN_WARNING "This may significantly impact system "
457 "performance\n");
458 return 1;
459 }
460
461 __setup("irqpoll", irqpoll_setup);
462