kernel/irq/spurious.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * linux/kernel/irq/spurious.c
   4  *
   5  * Copyright (C) 1992, 1998-2004 Linus Torvalds, Ingo Molnar
   6  *
   7  * This file contains spurious interrupt handling.
   8  */
   9
  10 #include <linux/jiffies.h>
  11 #include <linux/irq.h>
  12 #include <linux/module.h>
  13 #include <linux/kallsyms.h>
  14 #include <linux/interrupt.h>
  15 #include <linux/moduleparam.h>
  16 #include <linux/timer.h>
  17
  18 #include "internals.h"
  19
  20 static int irqfixup __read_mostly;
  21
  22 #define POLL_SPURIOUS_IRQ_INTERVAL (HZ/10)
  23 static void poll_spurious_irqs(unsigned long dummy);
  24 static DEFINE_TIMER(poll_spurious_irq_timer, poll_spurious_irqs);
  25 static int irq_poll_cpu;
  26 static atomic_t irq_poll_active;
  27
  28 /*
  29  * We wait here for a poller to finish.
  30  *
  31  * If the poll runs on this CPU, then we yell loudly and return
  32  * false. That will leave the interrupt line disabled in the worst
  33  * case, but it should never happen.
  34  *
  35  * We wait until the poller is done and then recheck disabled and
  36  * action (about to be disabled). Only if it's still active, we return
  37  * true and let the handler run.
  38  */
  39 bool irq_wait_for_poll(struct irq_desc *desc)
  40 {
  41         if (WARN_ONCE(irq_poll_cpu == smp_processor_id(),
  42                       "irq poll in progress on cpu %d for irq %d\n",
  43                       smp_processor_id(), desc->irq_data.irq))
  44                 return false;
  45
  46 #ifdef CONFIG_SMP
  47         do {
  48                 raw_spin_unlock(&desc->lock);
  49                 while (irqd_irq_inprogress(&desc->irq_data))
  50                         cpu_relax();
  51                 raw_spin_lock(&desc->lock);
  52         } while (irqd_irq_inprogress(&desc->irq_data));
  53         /* Might have been disabled in meantime */
  54         return !irqd_irq_disabled(&desc->irq_data) && desc->action;
  55 #else
  56         return false;
  57 #endif
  58 }
  59
  60
  61 /*
  62  * Recovery handler for misrouted interrupts.
  63  */
  64 static int try_one_irq(struct irq_desc *desc, bool force)
  65 {
  66         irqreturn_t ret = IRQ_NONE;
  67         struct irqaction *action;
  68
  69         raw_spin_lock(&desc->lock);
  70
  71         /*
  72          * PER_CPU, nested thread interrupts and interrupts explicitely
  73          * marked polled are excluded from polling.
  74          */
  75         if (irq_settings_is_per_cpu(desc) ||
  76             irq_settings_is_nested_thread(desc) ||
  77             irq_settings_is_polled(desc))
  78                 goto out;
  79
  80         /*
  81          * Do not poll disabled interrupts unless the spurious
  82          * disabled poller asks explicitely.
  83          */
  84         if (irqd_irq_disabled(&desc->irq_data) && !force)
  85                 goto out;
  86
  87         /*
  88          * All handlers must agree on IRQF_SHARED, so we test just the
  89          * first.
  90          */
  91         action = desc->action;
  92         if (!action || !(action->flags & IRQF_SHARED) ||
  93             (action->flags & __IRQF_TIMER))
  94                 goto out;
  95
  96         /* Already running on another processor */
  97         if (irqd_irq_inprogress(&desc->irq_data)) {
  98                 /*
  99                  * Already running: If it is shared get the other
 100                  * CPU to go looking for our mystery interrupt too
 101                  */
 102                 desc->istate |= IRQS_PENDING;
 103                 goto out;
 104         }
 105
 106         /* Mark it poll in progress */
 107         desc->istate |= IRQS_POLL_INPROGRESS;
 108         do {
 109                 if (handle_irq_event(desc) == IRQ_HANDLED)
 110                         ret = IRQ_HANDLED;
 111                 /* Make sure that there is still a valid action */
 112                 action = desc->action;
 113         } while ((desc->istate & IRQS_PENDING) && action);
 114         desc->istate &= ~IRQS_POLL_INPROGRESS;
 115 out:
 116         raw_spin_unlock(&desc->lock);
 117         return ret == IRQ_HANDLED;
 118 }
 119
 120 static int misrouted_irq(int irq)
 121 {
 122         struct irq_desc *desc;
 123         int i, ok = 0;
 124
 125         if (atomic_inc_return(&irq_poll_active) != 1)
 126                 goto out;
 127
 128         irq_poll_cpu = smp_processor_id();
 129
 130         for_each_irq_desc(i, desc) {
 131                 if (!i)
 132                          continue;
 133
 134                 if (i == irq)   /* Already tried */
 135                         continue;
 136
 137                 if (try_one_irq(desc, false))
 138                         ok = 1;
 139         }
 140 out:
 141         atomic_dec(&irq_poll_active);
 142         /* So the caller can adjust the irq error counts */
 143         return ok;
 144 }
 145
 146 static void poll_spurious_irqs(unsigned long dummy)
 147 {
 148         struct irq_desc *desc;
 149         int i;
 150
 151         if (atomic_inc_return(&irq_poll_active) != 1)
 152                 goto out;
 153         irq_poll_cpu = smp_processor_id();
 154
 155         for_each_irq_desc(i, desc) {
 156                 unsigned int state;
 157
 158                 if (!i)
 159                          continue;
 160
 161                 /* Racy but it doesn't matter */
 162                 state = desc->istate;
 163                 barrier();
 164                 if (!(state & IRQS_SPURIOUS_DISABLED))
 165                         continue;
 166
 167                 local_irq_disable();
 168                 try_one_irq(desc, true);
 169                 local_irq_enable();
 170         }
 171 out:
 172         atomic_dec(&irq_poll_active);
 173         mod_timer(&poll_spurious_irq_timer,
 174                   jiffies + POLL_SPURIOUS_IRQ_INTERVAL);
 175 }
 176
 177 static inline int bad_action_ret(irqreturn_t action_ret)
 178 {
 179         unsigned int r = action_ret;
 180
 181         if (likely(r <= (IRQ_HANDLED | IRQ_WAKE_THREAD)))
 182                 return 0;
 183         return 1;
 184 }
 185
 186 /*
 187  * If 99,900 of the previous 100,000 interrupts have not been handled
 188  * then assume that the IRQ is stuck in some manner. Drop a diagnostic
 189  * and try to turn the IRQ off.
 190  *
 191  * (The other 100-of-100,000 interrupts may have been a correctly
 192  *  functioning device sharing an IRQ with the failing one)
 193  */
 194 static void __report_bad_irq(struct irq_desc *desc, irqreturn_t action_ret)
 195 {
 196         unsigned int irq = irq_desc_get_irq(desc);
 197         struct irqaction *action;
 198         unsigned long flags;
 199
 200         if (bad_action_ret(action_ret)) {
 201                 printk(KERN_ERR "irq event %d: bogus return value %x\n",
 202                                 irq, action_ret);
 203         } else {
 204                 printk(KERN_ERR "irq %d: nobody cared (try booting with "
 205                                 "the \"irqpoll\" option)\n", irq);
 206         }
 207         dump_stack();
 208         printk(KERN_ERR "handlers:\n");
 209
 210         /*
 211          * We need to take desc->lock here. note_interrupt() is called
 212          * w/o desc->lock held, but IRQ_PROGRESS set. We might race
 213          * with something else removing an action. It's ok to take
 214          * desc->lock here. See synchronize_irq().
 215          */
 216         raw_spin_lock_irqsave(&desc->lock, flags);
 217         for_each_action_of_desc(desc, action) {
 218                 printk(KERN_ERR "[<%p>] %pf", action->handler, action->handler);
 219                 if (action->thread_fn)
 220                         printk(KERN_CONT " threaded [<%p>] %pf",
 221                                         action->thread_fn, action->thread_fn);
 222                 printk(KERN_CONT "\n");
 223         }
 224         raw_spin_unlock_irqrestore(&desc->lock, flags);
 225 }
 226
 227 static void report_bad_irq(struct irq_desc *desc, irqreturn_t action_ret)
 228 {
 229         static int count = 100;
 230
 231         if (count > 0) {
 232                 count--;
 233                 __report_bad_irq(desc, action_ret);
 234         }
 235 }
 236
 237 static inline int
 238 try_misrouted_irq(unsigned int irq, struct irq_desc *desc,
 239                   irqreturn_t action_ret)
 240 {
 241         struct irqaction *action;
 242
 243         if (!irqfixup)
 244                 return 0;
 245
 246         /* We didn't actually handle the IRQ - see if it was misrouted? */
 247         if (action_ret == IRQ_NONE)
 248                 return 1;
 249
 250         /*
 251          * But for 'irqfixup == 2' we also do it for handled interrupts if
 252          * they are marked as IRQF_IRQPOLL (or for irq zero, which is the
 253          * traditional PC timer interrupt.. Legacy)
 254          */
 255         if (irqfixup < 2)
 256                 return 0;
 257
 258         if (!irq)
 259                 return 1;
 260
 261         /*
 262          * Since we don't get the descriptor lock, "action" can
 263          * change under us.  We don't really care, but we don't
 264          * want to follow a NULL pointer. So tell the compiler to
 265          * just load it once by using a barrier.
 266          */
 267         action = desc->action;
 268         barrier();
 269         return action && (action->flags & IRQF_IRQPOLL);
 270 }
 271
 272 #define SPURIOUS_DEFERRED       0x80000000
 273
 274 void note_interrupt(struct irq_desc *desc, irqreturn_t action_ret)
 275 {
 276         unsigned int irq;
 277
 278         if (desc->istate & IRQS_POLL_INPROGRESS ||
 279             irq_settings_is_polled(desc))
 280                 return;
 281
 282         if (bad_action_ret(action_ret)) {
 283                 report_bad_irq(desc, action_ret);
 284                 return;
 285         }
 286
 287         /*
 288          * We cannot call note_interrupt from the threaded handler
 289          * because we need to look at the compound of all handlers
 290          * (primary and threaded). Aside of that in the threaded
 291          * shared case we have no serialization against an incoming
 292          * hardware interrupt while we are dealing with a threaded
 293          * result.
 294          *
 295          * So in case a thread is woken, we just note the fact and
 296          * defer the analysis to the next hardware interrupt.
 297          *
 298          * The threaded handlers store whether they sucessfully
 299          * handled an interrupt and we check whether that number
 300          * changed versus the last invocation.
 301          *
 302          * We could handle all interrupts with the delayed by one
 303          * mechanism, but for the non forced threaded case we'd just
 304          * add pointless overhead to the straight hardirq interrupts
 305          * for the sake of a few lines less code.
 306          */
 307         if (action_ret & IRQ_WAKE_THREAD) {
 308                 /*
 309                  * There is a thread woken. Check whether one of the
 310                  * shared primary handlers returned IRQ_HANDLED. If
 311                  * not we defer the spurious detection to the next
 312                  * interrupt.
 313                  */
 314                 if (action_ret == IRQ_WAKE_THREAD) {
 315                         int handled;
 316                         /*
 317                          * We use bit 31 of thread_handled_last to
 318                          * denote the deferred spurious detection
 319                          * active. No locking necessary as
 320                          * thread_handled_last is only accessed here
 321                          * and we have the guarantee that hard
 322                          * interrupts are not reentrant.
 323                          */
 324                         if (!(desc->threads_handled_last & SPURIOUS_DEFERRED)) {
 325                                 desc->threads_handled_last |= SPURIOUS_DEFERRED;
 326                                 return;
 327                         }
 328                         /*
 329                          * Check whether one of the threaded handlers
 330                          * returned IRQ_HANDLED since the last
 331                          * interrupt happened.
 332                          *
 333                          * For simplicity we just set bit 31, as it is
 334                          * set in threads_handled_last as well. So we
 335                          * avoid extra masking. And we really do not
 336                          * care about the high bits of the handled
 337                          * count. We just care about the count being
 338                          * different than the one we saw before.
 339                          */
 340                         handled = atomic_read(&desc->threads_handled);
 341                         handled |= SPURIOUS_DEFERRED;
 342                         if (handled != desc->threads_handled_last) {
 343                                 action_ret = IRQ_HANDLED;
 344                                 /*
 345                                  * Note: We keep the SPURIOUS_DEFERRED
 346                                  * bit set. We are handling the
 347                                  * previous invocation right now.
 348                                  * Keep it for the current one, so the
 349                                  * next hardware interrupt will
 350                                  * account for it.
 351                                  */
 352                                 desc->threads_handled_last = handled;
 353                         } else {
 354                                 /*
 355                                  * None of the threaded handlers felt
 356                                  * responsible for the last interrupt
 357                                  *
 358                                  * We keep the SPURIOUS_DEFERRED bit
 359                                  * set in threads_handled_last as we
 360                                  * need to account for the current
 361                                  * interrupt as well.
 362                                  */
 363                                 action_ret = IRQ_NONE;
 364                         }
 365                 } else {
 366                         /*
 367                          * One of the primary handlers returned
 368                          * IRQ_HANDLED. So we don't care about the
 369                          * threaded handlers on the same line. Clear
 370                          * the deferred detection bit.
 371                          *
 372                          * In theory we could/should check whether the
 373                          * deferred bit is set and take the result of
 374                          * the previous run into account here as
 375                          * well. But it's really not worth the
 376                          * trouble. If every other interrupt is
 377                          * handled we never trigger the spurious
 378                          * detector. And if this is just the one out
 379                          * of 100k unhandled ones which is handled
 380                          * then we merily delay the spurious detection
 381                          * by one hard interrupt. Not a real problem.
 382                          */
 383                         desc->threads_handled_last &= ~SPURIOUS_DEFERRED;
 384                 }
 385         }
 386
 387         if (unlikely(action_ret == IRQ_NONE)) {
 388                 /*
 389                  * If we are seeing only the odd spurious IRQ caused by
 390                  * bus asynchronicity then don't eventually trigger an error,
 391                  * otherwise the counter becomes a doomsday timer for otherwise
 392                  * working systems
 393                  */
 394                 if (time_after(jiffies, desc->last_unhandled + HZ/10))
 395                         desc->irqs_unhandled = 1;
 396                 else
 397                         desc->irqs_unhandled++;
 398                 desc->last_unhandled = jiffies;
 399         }
 400
 401         irq = irq_desc_get_irq(desc);
 402         if (unlikely(try_misrouted_irq(irq, desc, action_ret))) {
 403                 int ok = misrouted_irq(irq);
 404                 if (action_ret == IRQ_NONE)
 405                         desc->irqs_unhandled -= ok;
 406         }
 407
 408         desc->irq_count++;
 409         if (likely(desc->irq_count < 100000))
 410                 return;
 411
 412         desc->irq_count = 0;
 413         if (unlikely(desc->irqs_unhandled > 99900)) {
 414                 /*
 415                  * The interrupt is stuck
 416                  */
 417                 __report_bad_irq(desc, action_ret);
 418                 /*
 419                  * Now kill the IRQ
 420                  */
 421                 printk(KERN_EMERG "Disabling IRQ #%d\n", irq);
 422                 desc->istate |= IRQS_SPURIOUS_DISABLED;
 423                 desc->depth++;
 424                 irq_disable(desc);
 425
 426                 mod_timer(&poll_spurious_irq_timer,
 427                           jiffies + POLL_SPURIOUS_IRQ_INTERVAL);
 428         }
 429         desc->irqs_unhandled = 0;
 430 }
 431
 432 bool noirqdebug __read_mostly;
 433
 434 int noirqdebug_setup(char *str)
 435 {
 436         noirqdebug = 1;
 437         printk(KERN_INFO "IRQ lockup detection disabled\n");
 438
 439         return 1;
 440 }
 441
 442 __setup("noirqdebug", noirqdebug_setup);
 443 module_param(noirqdebug, bool, 0644);
 444 MODULE_PARM_DESC(noirqdebug, "Disable irq lockup detection when true");
 445
 446 static int __init irqfixup_setup(char *str)
 447 {
 448         irqfixup = 1;
 449         printk(KERN_WARNING "Misrouted IRQ fixup support enabled.\n");
 450         printk(KERN_WARNING "This may impact system performance.\n");
 451
 452         return 1;
 453 }
 454
 455 __setup("irqfixup", irqfixup_setup);
 456 module_param(irqfixup, int, 0644);
 457
 458 static int __init irqpoll_setup(char *str)
 459 {
 460         irqfixup = 2;
 461         printk(KERN_WARNING "Misrouted IRQ fixup and polling support "
 462                                 "enabled\n");
 463         printk(KERN_WARNING "This may significantly impact system "
 464                                 "performance\n");
 465         return 1;
 466 }
 467
 468 __setup("irqpoll", irqpoll_setup);