Re: Threaded interrupt handlers broken?

[Michael Buesch <mb@bu3sch.de>]

On Sunday 16 August 2009 11:53:13 Michael Buesch wrote:
> Hi,
> 
> I was trying to use threaded interrupt handlers, but the code always crashes
> within irq_thread() with a "BUG: spinlock bad magic 00000000".
> The spinlock that's not properly initialized is from the wait_for_threads waitqueue.
> It crashes on line 526 (see below).
> The initialization of the waitqueue struct seems to depend on whether the IRQ is
> shared or not. I don't know if that's correct, but I patched it to unconditionally
> initialize the struct. That did not help.
> 
> Any ideas?
> 
> 
> 490 static int irq_thread(void *data)
> 491 {
> 492         struct sched_param param = { .sched_priority = MAX_USER_RT_PRIO/2, };
> 493         struct irqaction *action = data;
> 494         struct irq_desc *desc = irq_to_desc(action->irq);
> 495         int wake;
> 496 
> 497         sched_setscheduler(current, SCHED_FIFO, &param);
> 498         current->irqaction = action;
> 499 
> 500         while (!irq_wait_for_interrupt(action)) {
> 501 
> 502                 irq_thread_check_affinity(desc, action);
> 503 
> 504                 atomic_inc(&desc->threads_active);
> 505 
> 506                 spin_lock_irq(&desc->lock);
> 507                 if (unlikely(desc->status & IRQ_DISABLED)) {
> 508                         /*
> 509                          * CHECKME: We might need a dedicated
> 510                          * IRQ_THREAD_PENDING flag here, which
> 511                          * retriggers the thread in check_irq_resend()
> 512                          * but AFAICT IRQ_PENDING should be fine as it
> 513                          * retriggers the interrupt itself --- tglx
> 514                          */
> 515                         desc->status |= IRQ_PENDING;
> 516                         spin_unlock_irq(&desc->lock);
> 517                 } else {
> 518                         spin_unlock_irq(&desc->lock);
> 519 
> 520                         action->thread_fn(action->irq, action->dev_id);
> 521                 }
> 522 
> 523                 wake = atomic_dec_and_test(&desc->threads_active);

Is this test logic inverted? atomic_dec_and_test() means
(threads_active - 1) == 0
Shouldn't it be like this?
(threads_active - 1) != 0

> 524 
> 525                 if (wake && waitqueue_active(&desc->wait_for_threads))
> 526                         wake_up(&desc->wait_for_threads);                <<<<<<<<<<<<<<<<<<<<<<<<
> 527         }
> 528 
> 529         /*
> 530          * Clear irqaction. Otherwise exit_irq_thread() would make
> 531          * fuzz about an active irq thread going into nirvana.
> 532          */
> 533         current->irqaction = NULL;
> 534         return 0;
> 535 }
> 



-- 
Greetings, Michael.