On Fri, 02 Feb 2007 17:35:31 -0700 [EMAIL PROTECTED] (Eric W. Biederman) wrote:
> When making the interrupt vectors per cpu I failed to handle a case > during irq migration. If the same interrupt comes in while we are > servicing the irq but before we migrate it the pending bit in the > local apic IRR register will be set for that irq. > > After migrating the irq to another cpu and or vector the data > structures will no longer be setup to handle this pending irq. Then as > soon as we return from servicing the irq we just migrated we will get > a nasty: "No irq handler for vector" message. > > Since we do not disable irqs for edge triggered interrupts except > in the smallest possible window during migration I cannot avoid > migrating an irq in the unlikely event that it becomes pending. > This is because by the time the irq could no longer become pending > I have already updated all of my data structures. > > Therefore this patch introduces an intermediate state that > exists soley on the cpu where we are handling the irq during > migration. The irq number is changed to negative in the > vector_irq data structure. > > Once the migration operation is complete we know we will receive > no more interrupts on this vector so the irq pending state for > this irq will no longer be updated. If the irq is not pending and > we are in the intermediate state we immediately free the vector, > otherwise in we free the vector in do_IRQ when the pending irq > arrives. So is this a for-2.6.20 thing? The bug was present in 2.6.19, so I assume it doesn't affect many people? - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
