kernel/irq/spurious.c

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