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 19static int irqfixup __read_mostly; 20 21#define POLL_SPURIOUS_IRQ_INTERVAL (HZ/10) 22static void poll_spurious_irqs(unsigned long dummy); 23static DEFINE_TIMER(poll_spurious_irq_timer, poll_spurious_irqs, 0, 0); 24static int irq_poll_cpu; 25static 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 */ 38bool 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 */ 63static int try_one_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 /* 71 * PER_CPU, nested thread interrupts and interrupts explicitely 72 * marked polled are excluded from polling. 73 */ 74 if (irq_settings_is_per_cpu(desc) || 75 irq_settings_is_nested_thread(desc) || 76 irq_settings_is_polled(desc)) 77 goto out; 78 79 /* 80 * Do not poll disabled interrupts unless the spurious 81 * disabled poller asks explicitely. 82 */ 83 if (irqd_irq_disabled(&desc->irq_data) && !force) 84 goto out; 85 86 /* 87 * All handlers must agree on IRQF_SHARED, so we test just the 88 * first. 89 */ 90 action = desc->action; 91 if (!action || !(action->flags & IRQF_SHARED) || 92 (action->flags & __IRQF_TIMER)) 93 goto out; 94 95 /* Already running on another processor */ 96 if (irqd_irq_inprogress(&desc->irq_data)) { 97 /* 98 * Already running: If it is shared get the other 99 * CPU to go looking for our mystery interrupt too 100 */ 101 desc->istate |= IRQS_PENDING; 102 goto out; 103 } 104 105 /* Mark it poll in progress */ 106 desc->istate |= IRQS_POLL_INPROGRESS; 107 do { 108 if (handle_irq_event(desc) == IRQ_HANDLED) 109 ret = IRQ_HANDLED; 110 /* Make sure that there is still a valid action */ 111 action = desc->action; 112 } while ((desc->istate & IRQS_PENDING) && action); 113 desc->istate &= ~IRQS_POLL_INPROGRESS; 114out: 115 raw_spin_unlock(&desc->lock); 116 return ret == IRQ_HANDLED; 117} 118 119static int misrouted_irq(int irq) 120{ 121 struct irq_desc *desc; 122 int i, ok = 0; 123 124 if (atomic_inc_return(&irq_poll_active) != 1) 125 goto out; 126 127 irq_poll_cpu = smp_processor_id(); 128 129 for_each_irq_desc(i, desc) { 130 if (!i) 131 continue; 132 133 if (i == irq) /* Already tried */ 134 continue; 135 136 if (try_one_irq(desc, false)) 137 ok = 1; 138 } 139out: 140 atomic_dec(&irq_poll_active); 141 /* So the caller can adjust the irq error counts */ 142 return ok; 143} 144 145static void poll_spurious_irqs(unsigned long dummy) 146{ 147 struct irq_desc *desc; 148 int i; 149 150 if (atomic_inc_return(&irq_poll_active) != 1) 151 goto out; 152 irq_poll_cpu = smp_processor_id(); 153 154 for_each_irq_desc(i, desc) { 155 unsigned int state; 156 157 if (!i) 158 continue; 159 160 /* Racy but it doesn't matter */ 161 state = desc->istate; 162 barrier(); 163 if (!(state & IRQS_SPURIOUS_DISABLED)) 164 continue; 165 166 local_irq_disable(); 167 try_one_irq(desc, true); 168 local_irq_enable(); 169 } 170out: 171 atomic_dec(&irq_poll_active); 172 mod_timer(&poll_spurious_irq_timer, 173 jiffies + POLL_SPURIOUS_IRQ_INTERVAL); 174} 175 176static inline int bad_action_ret(irqreturn_t action_ret) 177{ 178 if (likely(action_ret <= (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 */ 191static 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 action = desc->action; 215 while (action) { 216 printk(KERN_ERR "[<%p>] %pf", action->handler, action->handler); 217 if (action->thread_fn) 218 printk(KERN_CONT " threaded [<%p>] %pf", 219 action->thread_fn, action->thread_fn); 220 printk(KERN_CONT "\n"); 221 action = action->next; 222 } 223 raw_spin_unlock_irqrestore(&desc->lock, flags); 224} 225 226static void report_bad_irq(struct irq_desc *desc, irqreturn_t action_ret) 227{ 228 static int count = 100; 229 230 if (count > 0) { 231 count--; 232 __report_bad_irq(desc, action_ret); 233 } 234} 235 236static inline int 237try_misrouted_irq(unsigned int irq, struct irq_desc *desc, 238 irqreturn_t action_ret) 239{ 240 struct irqaction *action; 241 242 if (!irqfixup) 243 return 0; 244 245 /* We didn't actually handle the IRQ - see if it was misrouted? */ 246 if (action_ret == IRQ_NONE) 247 return 1; 248 249 /* 250 * But for 'irqfixup == 2' we also do it for handled interrupts if 251 * they are marked as IRQF_IRQPOLL (or for irq zero, which is the 252 * traditional PC timer interrupt.. Legacy) 253 */ 254 if (irqfixup < 2) 255 return 0; 256 257 if (!irq) 258 return 1; 259 260 /* 261 * Since we don't get the descriptor lock, "action" can 262 * change under us. We don't really care, but we don't 263 * want to follow a NULL pointer. So tell the compiler to 264 * just load it once by using a barrier. 265 */ 266 action = desc->action; 267 barrier(); 268 return action && (action->flags & IRQF_IRQPOLL); 269} 270 271#define SPURIOUS_DEFERRED 0x80000000 272 273void note_interrupt(struct irq_desc *desc, irqreturn_t action_ret) 274{ 275 unsigned int irq; 276 277 if (desc->istate & IRQS_POLL_INPROGRESS || 278 irq_settings_is_polled(desc)) 279 return; 280 281 if (bad_action_ret(action_ret)) { 282 report_bad_irq(desc, action_ret); 283 return; 284 } 285 286 /* 287 * We cannot call note_interrupt from the threaded handler 288 * because we need to look at the compound of all handlers 289 * (primary and threaded). Aside of that in the threaded 290 * shared case we have no serialization against an incoming 291 * hardware interrupt while we are dealing with a threaded 292 * result. 293 * 294 * So in case a thread is woken, we just note the fact and 295 * defer the analysis to the next hardware interrupt. 296 * 297 * The threaded handlers store whether they sucessfully 298 * handled an interrupt and we check whether that number 299 * changed versus the last invocation. 300 * 301 * We could handle all interrupts with the delayed by one 302 * mechanism, but for the non forced threaded case we'd just 303 * add pointless overhead to the straight hardirq interrupts 304 * for the sake of a few lines less code. 305 */ 306 if (action_ret & IRQ_WAKE_THREAD) { 307 /* 308 * There is a thread woken. Check whether one of the 309 * shared primary handlers returned IRQ_HANDLED. If 310 * not we defer the spurious detection to the next 311 * interrupt. 312 */ 313 if (action_ret == IRQ_WAKE_THREAD) { 314 int handled; 315 /* 316 * We use bit 31 of thread_handled_last to 317 * denote the deferred spurious detection 318 * active. No locking necessary as 319 * thread_handled_last is only accessed here 320 * and we have the guarantee that hard 321 * interrupts are not reentrant. 322 */ 323 if (!(desc->threads_handled_last & SPURIOUS_DEFERRED)) { 324 desc->threads_handled_last |= SPURIOUS_DEFERRED; 325 return; 326 } 327 /* 328 * Check whether one of the threaded handlers 329 * returned IRQ_HANDLED since the last 330 * interrupt happened. 331 * 332 * For simplicity we just set bit 31, as it is 333 * set in threads_handled_last as well. So we 334 * avoid extra masking. And we really do not 335 * care about the high bits of the handled 336 * count. We just care about the count being 337 * different than the one we saw before. 338 */ 339 handled = atomic_read(&desc->threads_handled); 340 handled |= SPURIOUS_DEFERRED; 341 if (handled != desc->threads_handled_last) { 342 action_ret = IRQ_HANDLED; 343 /* 344 * Note: We keep the SPURIOUS_DEFERRED 345 * bit set. We are handling the 346 * previous invocation right now. 347 * Keep it for the current one, so the 348 * next hardware interrupt will 349 * account for it. 350 */ 351 desc->threads_handled_last = handled; 352 } else { 353 /* 354 * None of the threaded handlers felt 355 * responsible for the last interrupt 356 * 357 * We keep the SPURIOUS_DEFERRED bit 358 * set in threads_handled_last as we 359 * need to account for the current 360 * interrupt as well. 361 */ 362 action_ret = IRQ_NONE; 363 } 364 } else { 365 /* 366 * One of the primary handlers returned 367 * IRQ_HANDLED. So we don't care about the 368 * threaded handlers on the same line. Clear 369 * the deferred detection bit. 370 * 371 * In theory we could/should check whether the 372 * deferred bit is set and take the result of 373 * the previous run into account here as 374 * well. But it's really not worth the 375 * trouble. If every other interrupt is 376 * handled we never trigger the spurious 377 * detector. And if this is just the one out 378 * of 100k unhandled ones which is handled 379 * then we merily delay the spurious detection 380 * by one hard interrupt. Not a real problem. 381 */ 382 desc->threads_handled_last &= ~SPURIOUS_DEFERRED; 383 } 384 } 385 386 if (unlikely(action_ret == IRQ_NONE)) { 387 /* 388 * If we are seeing only the odd spurious IRQ caused by 389 * bus asynchronicity then don't eventually trigger an error, 390 * otherwise the counter becomes a doomsday timer for otherwise 391 * working systems 392 */ 393 if (time_after(jiffies, desc->last_unhandled + HZ/10)) 394 desc->irqs_unhandled = 1; 395 else 396 desc->irqs_unhandled++; 397 desc->last_unhandled = jiffies; 398 } 399 400 irq = irq_desc_get_irq(desc); 401 if (unlikely(try_misrouted_irq(irq, desc, action_ret))) { 402 int ok = misrouted_irq(irq); 403 if (action_ret == IRQ_NONE) 404 desc->irqs_unhandled -= ok; 405 } 406 407 desc->irq_count++; 408 if (likely(desc->irq_count < 100000)) 409 return; 410 411 desc->irq_count = 0; 412 if (unlikely(desc->irqs_unhandled > 99900)) { 413 /* 414 * The interrupt is stuck 415 */ 416 __report_bad_irq(desc, action_ret); 417 /* 418 * Now kill the IRQ 419 */ 420 printk(KERN_EMERG "Disabling IRQ #%d\n", irq); 421 desc->istate |= IRQS_SPURIOUS_DISABLED; 422 desc->depth++; 423 irq_disable(desc); 424 425 mod_timer(&poll_spurious_irq_timer, 426 jiffies + POLL_SPURIOUS_IRQ_INTERVAL); 427 } 428 desc->irqs_unhandled = 0; 429} 430 431bool noirqdebug __read_mostly; 432 433int noirqdebug_setup(char *str) 434{ 435 noirqdebug = 1; 436 printk(KERN_INFO "IRQ lockup detection disabled\n"); 437 438 return 1; 439} 440 441__setup("noirqdebug", noirqdebug_setup); 442module_param(noirqdebug, bool, 0644); 443MODULE_PARM_DESC(noirqdebug, "Disable irq lockup detection when true"); 444 445static int __init irqfixup_setup(char *str) 446{ 447 irqfixup = 1; 448 printk(KERN_WARNING "Misrouted IRQ fixup support enabled.\n"); 449 printk(KERN_WARNING "This may impact system performance.\n"); 450 451 return 1; 452} 453 454__setup("irqfixup", irqfixup_setup); 455module_param(irqfixup, int, 0644); 456 457static int __init irqpoll_setup(char *str) 458{ 459 irqfixup = 2; 460 printk(KERN_WARNING "Misrouted IRQ fixup and polling support " 461 "enabled\n"); 462 printk(KERN_WARNING "This may significantly impact system " 463 "performance\n"); 464 return 1; 465} 466 467__setup("irqpoll", irqpoll_setup); 468