On 13/08/14 15:22, Russell King - ARM Linux wrote:
> On Wed, Aug 13, 2014 at 06:57:18AM -0700, Stephen Boyd wrote:
>> Commit 1a6b69b6548c (ARM: gic: add CPU migration support,
>> 2012-04-12) introduced an acquisition of the irq_controller_lock
>> in gic_raise_softirq() which can lead to a spinlock recursion if
>> the gic_arch_extn hooks call into the scheduler (via complete()
>> or wake_up(), etc.). This happens because gic_arch_extn hooks are
>> normally called with the irq_controller_lock held and calling
>> into the scheduler may cause us to call smp_send_reschedule()
>> which will grab the irq_controller_lock again. Here's an example
>> from a vendor kernel (note that the gic_arch_extn hook code here
>> isn't actually in mainline):
>
> Here's a question: why would you want to call into the scheduler from
> the gic_arch_extn code?
>
> Oh. My. God. Thomas, what have you done to the generic IRQ layer?
> This is /totally/ unsafe:
>
> void disable_irq(unsigned int irq)
> {
> if (!__disable_irq_nosync(irq))
> synchronize_irq(irq);
> }
>
> static int __disable_irq_nosync(unsigned int irq)
> {
> unsigned long flags;
> struct irq_desc *desc = irq_get_desc_buslock(irq, &flags, IRQ_GET_DESC_CHECK_GLOBAL);
irq_get_desc_buslock() results in us owning the descriptor's lock
(raw_spinlock_t).
>
> if (!desc)
> return -EINVAL;
> __disable_irq(desc, irq, false);
> irq_put_desc_busunlock(desc, flags);
> return 0;
> }
>
> void __disable_irq(struct irq_desc *desc, unsigned int irq, bool suspend)
> {
> if (suspend) {
> if (!desc->action || (desc->action->flags & IRQF_NO_SUSPEND))
> return;
> desc->istate |= IRQS_SUSPENDED;
> }
>
> if (!desc->depth++)
> irq_disable(desc);
> }
>
> You realise that disable_irq() and enable_irq() can be called by
> concurrently by different drivers for the /same/ interrupt. For
> starters, that post-increment there is completely unprotected against
> races. Secondly, the above is completely racy against a concurrent
> enable_irq() - what if we're in disable_irq(), we've incremented
> depth, but have yet to call irq_disable(). The count now has a
> value of 1.
>
> We then preempt, and run another thread which calls enable_irq()
> on it. This results in the depth being decremented, and the IRQ
> is now enabled.
We shouldn't get that far due to the spinlock taken during the disable.
> We resume the original thread, and continue to call irq_disable(),
> resulting in the interrupt being disabled.
>
> That's not nice (the right answer is that it's strictly an unbalanced
> enable_irq(), but that's no excuse here.)
>