Linux Kernel Architecture

Chapter 14: Kernel Activities

When stacks with 8-KiB size, that is, two page frames, are used, IRQ handling is simplified because a

potential stack switch does not need to be taken into account andirq_desc[irq]->handle_irqcan be

called immediately in any case.

Old-Style Processing

In the discussion of how AMD64 calls the flow control handler, it was mentioned that the code ends up in

generic_handle_irq, which selects and activates the properhandle_irqfunction from the IRQ database

irq_desc. However,generic_handle_irqis a little more complicated in practice:

<irq.h>

static inline void generic_handle_irq(unsigned int irq)

{

struct irq_desc *desc = irq_desc + irq;

#ifdef CONFIG_GENERIC_HARDIRQS_NO__DO_IRQ

desc->handle_irq(irq, desc);

#else

if (likely(desc->handle_irq))

desc->handle_irq(irq, desc);

else

__do_IRQ(irq);

#endif

}

Before the generic IRQ rework, the kernel used a colorful mixture of architecture-dependentapproaches

to IRQ handling. Most important, there was no separation between flow handling and ISR han-

dling: Both tasks were performed simultaneously in a single architecture-specific routine usually

called__do_IRQ.

Modern code should activate the configuration optionGENRIC_HARDIRQS_NO__DO_IRQand implement

flow handling as shown in the preceding discussions. In this case,generic_handle_irqreally boils

down to just callingirq_desc[irq]->handle_irq.

What if this option is not set? The kernel provides a default implementation of__do_IRQthat com-

bines flow handling for all interrupt types, and also calls the required ISRs.^15 Basically, there are three

possibilities of how to use this function and implement flow handling:

1. Use generic flow handlers for some IRQs, and leave the handlers for others undefined. For
   these,__do_IRQis employed to handle both flow and high-level processing. It is required to
   callgeneric_handle_IRQfromdo_IRQin this case.


2. Call__do_IRQdirectly fromdo_IRQ. This bypasses the flow separation completely. Some
   off-mainstream architectures like M32R, H8300, SuperH, and Cris still use this approach.


3. Handle IRQs in a completely architecture-dependent way without reusing any of the exist-
   ing frameworks. Clearly, this is not the brightest idea — to say the least.

Since it is needless to say that the long-term goal forall architectures is to convert to the generic IRQ

framework,__do_IRQis not discussed in detail.

(^15) The implementation is based on the version used on IA-32 systems before the generic IRQ framework was introduced.