1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (C) 1992, 1998-2004 Linus Torvalds, Ingo Molnar
4 *
5 * This file contains spurious interrupt handling.
6 */
7
8 #include <linux/jiffies.h>
9 #include <linux/irq.h>
10 #include <linux/module.h>
11 #include <linux/interrupt.h>
12 #include <linux/moduleparam.h>
13 #include <linux/timer.h>
14
15 #include "internals.h"
16
17 static int irqfixup __read_mostly;
18
19 #define POLL_SPURIOUS_IRQ_INTERVAL (HZ/10)
20 static void poll_spurious_irqs(struct timer_list *unused);
21 static DEFINE_TIMER(poll_spurious_irq_timer, poll_spurious_irqs);
22 static int irq_poll_cpu;
23 static atomic_t irq_poll_active;
24
25 /*
26 * We wait here for a poller to finish.
27 *
28 * If the poll runs on this CPU, then we yell loudly and return
29 * false. That will leave the interrupt line disabled in the worst
30 * case, but it should never happen.
31 *
32 * We wait until the poller is done and then recheck disabled and
33 * action (about to be disabled). Only if it's still active, we return
34 * true and let the handler run.
35 */
36 bool irq_wait_for_poll(struct irq_desc *desc)
37 {
38 if (WARN_ONCE(irq_poll_cpu == smp_processor_id(),
39 "irq poll in progress on cpu %d for irq %d\n",
40 smp_processor_id(), desc->irq_data.irq))
41 return false;
42
43 #ifdef CONFIG_SMP
44 do {
45 raw_spin_unlock(&desc->lock);
46 while (irqd_irq_inprogress(&desc->irq_data))
47 cpu_relax();
48 raw_spin_lock(&desc->lock);
49 } while (irqd_irq_inprogress(&desc->irq_data));
50 /* Might have been disabled in meantime */
51 return !irqd_irq_disabled(&desc->irq_data) && desc->action;
52 #else
53 return false;
54 #endif
55 }
56
57
58 /*
59 * Recovery handler for misrouted interrupts.
60 */
61 static int try_one_irq(struct irq_desc *desc, bool force)
62 {
63 irqreturn_t ret = IRQ_NONE;
64 struct irqaction *action;
65
66 raw_spin_lock(&desc->lock);
67
68 /*
69 * PER_CPU, nested thread interrupts and interrupts explicitly
70 * marked polled are excluded from polling.
71 */
72 if (irq_settings_is_per_cpu(desc) ||
73 irq_settings_is_nested_thread(desc) ||
74 irq_settings_is_polled(desc))
75 goto out;
76
77 /*
78 * Do not poll disabled interrupts unless the spurious
79 * disabled poller asks explicitly.
80 */
81 if (irqd_irq_disabled(&desc->irq_data) && !force)
82 goto out;
83
84 /*
85 * All handlers must agree on IRQF_SHARED, so we test just the
86 * first.
87 */
88 action = desc->action;
89 if (!action || !(action->flags & IRQF_SHARED) ||
90 (action->flags & __IRQF_TIMER))
91 goto out;
92
93 /* Already running on another processor */
94 if (irqd_irq_inprogress(&desc->irq_data)) {
95 /*
96 * Already running: If it is shared get the other
97 * CPU to go looking for our mystery interrupt too
98 */
99 desc->istate |= IRQS_PENDING;
100 goto out;
101 }
102
103 /* Mark it poll in progress */
104 desc->istate |= IRQS_POLL_INPROGRESS;
105 do {
106 if (handle_irq_event(desc) == IRQ_HANDLED)
107 ret = IRQ_HANDLED;
108 /* Make sure that there is still a valid action */
109 action = desc->action;
110 } while ((desc->istate & IRQS_PENDING) && action);
111 desc->istate &= ~IRQS_POLL_INPROGRESS;
112 out:
113 raw_spin_unlock(&desc->lock);
114 return ret == IRQ_HANDLED;
115 }
116
117 static int misrouted_irq(int irq)
118 {
119 struct irq_desc *desc;
120 int i, ok = 0;
121
122 if (atomic_inc_return(&irq_poll_active) != 1)
123 goto out;
124
125 irq_poll_cpu = smp_processor_id();
126
127 for_each_irq_desc(i, desc) {
128 if (!i)
129 continue;
130
131 if (i == irq) /* Already tried */
132 continue;
133
134 if (try_one_irq(desc, false))
135 ok = 1;
136 }
137 out:
138 atomic_dec(&irq_poll_active);
139 /* So the caller can adjust the irq error counts */
140 return ok;
141 }
142
143 static void poll_spurious_irqs(struct timer_list *unused)
144 {
145 struct irq_desc *desc;
146 int i;
147
148 if (atomic_inc_return(&irq_poll_active) != 1)
149 goto out;
150 irq_poll_cpu = smp_processor_id();
151
152 for_each_irq_desc(i, desc) {
153 unsigned int state;
154
155 if (!i)
156 continue;
157
158 /* Racy but it doesn't matter */
159 state = desc->istate;
160 barrier();
161 if (!(state & IRQS_SPURIOUS_DISABLED))
162 continue;
163
164 local_irq_disable();
165 try_one_irq(desc, true);
166 local_irq_enable();
167 }
168 out:
169 atomic_dec(&irq_poll_active);
170 mod_timer(&poll_spurious_irq_timer,
171 jiffies + POLL_SPURIOUS_IRQ_INTERVAL);
172 }
173
174 static inline int bad_action_ret(irqreturn_t action_ret)
175 {
176 unsigned int r = action_ret;
177
178 if (likely(r <= (IRQ_HANDLED | IRQ_WAKE_THREAD)))
179 return 0;
180 return 1;
181 }
182
183 /*
184 * If 99,900 of the previous 100,000 interrupts have not been handled
185 * then assume that the IRQ is stuck in some manner. Drop a diagnostic
186 * and try to turn the IRQ off.
187 *
188 * (The other 100-of-100,000 interrupts may have been a correctly
189 * functioning device sharing an IRQ with the failing one)
190 */
191 static void __report_bad_irq(struct irq_desc *desc, irqreturn_t action_ret)
192 {
193 unsigned int irq = irq_desc_get_irq(desc);
194 struct irqaction *action;
195 unsigned long flags;
196
197 if (bad_action_ret(action_ret)) {
198 printk(KERN_ERR "irq event %d: bogus return value %x\n",
199 irq, action_ret);
200 } else {
201 printk(KERN_ERR "irq %d: nobody cared (try booting with "
202 "the \"irqpoll\" option)\n", irq);
203 }
204 dump_stack();
205 printk(KERN_ERR "handlers:\n");
206
207 /*
208 * We need to take desc->lock here. note_interrupt() is called
209 * w/o desc->lock held, but IRQ_PROGRESS set. We might race
210 * with something else removing an action. It's ok to take
211 * desc->lock here. See synchronize_irq().
212 */
213 raw_spin_lock_irqsave(&desc->lock, flags);
214 for_each_action_of_desc(desc, action) {
215 printk(KERN_ERR "[<%p>] %ps", action->handler, action->handler);
216 if (action->thread_fn)
217 printk(KERN_CONT " threaded [<%p>] %ps",
218 action->thread_fn, action->thread_fn);
219 printk(KERN_CONT "\n");
220 }
221 raw_spin_unlock_irqrestore(&desc->lock, flags);
222 }
223
224 static void report_bad_irq(struct irq_desc *desc, irqreturn_t action_ret)
225 {
226 static int count = 100;
227
228 if (count > 0) {
229 count--;
230 __report_bad_irq(desc, action_ret);
231 }
232 }
233
234 static inline int
235 try_misrouted_irq(unsigned int irq, struct irq_desc *desc,
236 irqreturn_t action_ret)
237 {
238 struct irqaction *action;
239
240 if (!irqfixup)
241 return 0;
242
243 /* We didn't actually handle the IRQ - see if it was misrouted? */
244 if (action_ret == IRQ_NONE)
245 return 1;
246
247 /*
248 * But for 'irqfixup == 2' we also do it for handled interrupts if
249 * they are marked as IRQF_IRQPOLL (or for irq zero, which is the
250 * traditional PC timer interrupt.. Legacy)
251 */
252 if (irqfixup < 2)
253 return 0;
254
255 if (!irq)
256 return 1;
257
258 /*
259 * Since we don't get the descriptor lock, "action" can
260 * change under us. We don't really care, but we don't
261 * want to follow a NULL pointer. So tell the compiler to
262 * just load it once by using a barrier.
263 */
264 action = desc->action;
265 barrier();
266 return action && (action->flags & IRQF_IRQPOLL);
267 }
268
269 #define SPURIOUS_DEFERRED 0x80000000
270
271 void note_interrupt(struct irq_desc *desc, irqreturn_t action_ret)
272 {
273 unsigned int irq;
274
275 if (desc->istate & IRQS_POLL_INPROGRESS ||
276 irq_settings_is_polled(desc))
277 return;
278
279 if (bad_action_ret(action_ret)) {
280 report_bad_irq(desc, action_ret);
281 return;
282 }
283
284 /*
285 * We cannot call note_interrupt from the threaded handler
286 * because we need to look at the compound of all handlers
287 * (primary and threaded). Aside of that in the threaded
288 * shared case we have no serialization against an incoming
289 * hardware interrupt while we are dealing with a threaded
290 * result.
291 *
292 * So in case a thread is woken, we just note the fact and
293 * defer the analysis to the next hardware interrupt.
294 *
295 * The threaded handlers store whether they successfully
296 * handled an interrupt and we check whether that number
297 * changed versus the last invocation.
298 *
299 * We could handle all interrupts with the delayed by one
300 * mechanism, but for the non forced threaded case we'd just
301 * add pointless overhead to the straight hardirq interrupts
302 * for the sake of a few lines less code.
303 */
304 if (action_ret & IRQ_WAKE_THREAD) {
305 /*
306 * There is a thread woken. Check whether one of the
307 * shared primary handlers returned IRQ_HANDLED. If
308 * not we defer the spurious detection to the next
309 * interrupt.
310 */
311 if (action_ret == IRQ_WAKE_THREAD) {
312 int handled;
313 /*
314 * We use bit 31 of thread_handled_last to
315 * denote the deferred spurious detection
316 * active. No locking necessary as
317 * thread_handled_last is only accessed here
318 * and we have the guarantee that hard
319 * interrupts are not reentrant.
320 */
321 if (!(desc->threads_handled_last & SPURIOUS_DEFERRED)) {
322 desc->threads_handled_last |= SPURIOUS_DEFERRED;
323 return;
324 }
325 /*
326 * Check whether one of the threaded handlers
327 * returned IRQ_HANDLED since the last
328 * interrupt happened.
329 *
330 * For simplicity we just set bit 31, as it is
331 * set in threads_handled_last as well. So we
332 * avoid extra masking. And we really do not
333 * care about the high bits of the handled
334 * count. We just care about the count being
335 * different than the one we saw before.
336 */
337 handled = atomic_read(&desc->threads_handled);
338 handled |= SPURIOUS_DEFERRED;
339 if (handled != desc->threads_handled_last) {
340 action_ret = IRQ_HANDLED;
341 /*
342 * Note: We keep the SPURIOUS_DEFERRED
343 * bit set. We are handling the
344 * previous invocation right now.
345 * Keep it for the current one, so the
346 * next hardware interrupt will
347 * account for it.
348 */
349 desc->threads_handled_last = handled;
350 } else {
351 /*
352 * None of the threaded handlers felt
353 * responsible for the last interrupt
354 *
355 * We keep the SPURIOUS_DEFERRED bit
356 * set in threads_handled_last as we
357 * need to account for the current
358 * interrupt as well.
359 */
360 action_ret = IRQ_NONE;
361 }
362 } else {
363 /*
364 * One of the primary handlers returned
365 * IRQ_HANDLED. So we don't care about the
366 * threaded handlers on the same line. Clear
367 * the deferred detection bit.
368 *
369 * In theory we could/should check whether the
370 * deferred bit is set and take the result of
371 * the previous run into account here as
372 * well. But it's really not worth the
373 * trouble. If every other interrupt is
374 * handled we never trigger the spurious
375 * detector. And if this is just the one out
376 * of 100k unhandled ones which is handled
377 * then we merily delay the spurious detection
378 * by one hard interrupt. Not a real problem.
379 */
380 desc->threads_handled_last &= ~SPURIOUS_DEFERRED;
381 }
382 }
383
384 if (unlikely(action_ret == IRQ_NONE)) {
385 /*
386 * If we are seeing only the odd spurious IRQ caused by
387 * bus asynchronicity then don't eventually trigger an error,
388 * otherwise the counter becomes a doomsday timer for otherwise
389 * working systems
390 */
391 if (time_after(jiffies, desc->last_unhandled + HZ/10))
392 desc->irqs_unhandled = 1;
393 else
394 desc->irqs_unhandled++;
395 desc->last_unhandled = jiffies;
396 }
397
398 irq = irq_desc_get_irq(desc);
399 if (unlikely(try_misrouted_irq(irq, desc, action_ret))) {
400 int ok = misrouted_irq(irq);
401 if (action_ret == IRQ_NONE)
402 desc->irqs_unhandled -= ok;
403 }
404
405 desc->irq_count++;
406 if (likely(desc->irq_count < 100000))
407 return;
408
409 desc->irq_count = 0;
410 if (unlikely(desc->irqs_unhandled > 99900)) {
411 /*
412 * The interrupt is stuck
413 */
414 __report_bad_irq(desc, action_ret);
415 /*
416 * Now kill the IRQ
417 */
418 printk(KERN_EMERG "Disabling IRQ #%d\n", irq);
419 desc->istate |= IRQS_SPURIOUS_DISABLED;
420 desc->depth++;
421 irq_disable(desc);
422
423 mod_timer(&poll_spurious_irq_timer,
424 jiffies + POLL_SPURIOUS_IRQ_INTERVAL);
425 }
426 desc->irqs_unhandled = 0;
427 }
428
429 bool noirqdebug __read_mostly;
430
431 int noirqdebug_setup(char *str)
432 {
433 noirqdebug = 1;
434 printk(KERN_INFO "IRQ lockup detection disabled\n");
435
436 return 1;
437 }
438
439 __setup("noirqdebug", noirqdebug_setup);
440 module_param(noirqdebug, bool, 0644);
441 MODULE_PARM_DESC(noirqdebug, "Disable irq lockup detection when true");
442
443 static int __init irqfixup_setup(char *str)
444 {
445 irqfixup = 1;
446 printk(KERN_WARNING "Misrouted IRQ fixup support enabled.\n");
447 printk(KERN_WARNING "This may impact system performance.\n");
448
449 return 1;
450 }
451
452 __setup("irqfixup", irqfixup_setup);
453 module_param(irqfixup, int, 0644);
454
455 static int __init irqpoll_setup(char *str)
456 {
457 irqfixup = 2;
458 printk(KERN_WARNING "Misrouted IRQ fixup and polling support "
459 "enabled\n");
460 printk(KERN_WARNING "This may significantly impact system "
461 "performance\n");
462 return 1;
463 }
464
465 __setup("irqpoll", irqpoll_setup);