linux/Documentation/arm/interrupts.rst

==========
Interrupts
==========

2.5.2-rmk5:
  This is the first kernel that contains a major shake up of some of the
  major architecture-specific subsystems.

Firstly, it contains some pretty major changes to the way we handle the
MMU TLB.  Each MMU TLB variant is now handled completely separately -
we have TLB v3, TLB v4 (without write buffer), TLB v4 (with write buffer),
and finally TLB v4 (with write buffer, with I TLB invalidate entry).
There is more assembly code inside each of these functions, mainly to
allow more flexible TLB handling for the future.

Secondly, the IRQ subsystem.

The 2.5 kernels will be having major changes to the way IRQs are handled.
Unfortunately, this means that machine types that touch the irq_desc[]
array (basically all machine types) will break, and this means every
machine type that we currently have.

Lets take an example.  On the Assabet with Neponset, we have::

                  GPIO25                 IRR:2
        SA1100 ------------> Neponset -----------> SA1111
                                         IIR:1
                                      -----------> USAR
                                         IIR:0
                                      -----------> SMC9196

The way stuff currently works, all SA1111 interrupts are mutually
exclusive of each other - if you're processing one interrupt from the
SA1111 and another comes in, you have to wait for that interrupt to
finish processing before you can service the new interrupt.  Eg, an
IDE PIO-based interrupt on the SA1111 excludes all other SA1111 and
SMC9196 interrupts until it has finished transferring its multi-sector
data, which can be a long time.  Note also that since we loop in the
SA1111 IRQ handler, SA1111 IRQs can hold off SMC9196 IRQs indefinitely.


The new approach brings several new ideas...

We introduce the concept of a "parent" and a "child".  For example,
to the Neponset handler, the "parent" is GPIO25, and the "children"d
are SA1111, SMC9196 and USAR.

We also bring the idea of an IRQ "chip" (mainly to reduce the size of
the irqdesc array).  This doesn't have to be a real "IC"; indeed the
SA11x0 IRQs are handled by two separate "chip" structures, one for
GPIO0-10, and another for all the rest.  It is just a container for
the various operations (maybe this'll change to a better name).
This structure has the following operations::

  struct irqchip {
          /*
           * Acknowledge the IRQ.
           * If this is a level-based IRQ, then it is expected to mask the IRQ
           * as well.
           */
          void (*ack)(unsigned int irq);
          /*
           * Mask the IRQ in hardware.
           */
          void (*mask)(unsigned int irq);
          /*
           * Unmask the IRQ in hardware.
           */
          void (*unmask)(unsigned int irq);
          /*
           * Re-run the IRQ
           */
          void (*rerun)(unsigned int irq);
          /*
           * Set the type of the IRQ.
           */
          int (*type)(unsigned int irq, unsigned int, type);
  };

ack
       - required.  May be the same function as mask for IRQs
         handled by do_level_IRQ.
mask
       - required.
unmask
       - required.
rerun
       - optional.  Not required if you're using do_level_IRQ for all
         IRQs that use this 'irqchip'.  Generally expected to re-trigger
         the hardware IRQ if possible.  If not, may call the handler
	 directly.
type
       - optional.  If you don't support changing the type of an IRQ,
         it should be null so people can detect if they are unable to
         set the IRQ type.

For each IRQ, we keep the following information:

        - "disable" depth (number of disable_irq()s without enable_irq()s)
        - flags indicating what we can do with this IRQ (valid, probe,
          noautounmask) as before
        - status of the IRQ (probing, enable, etc)
        - chip
        - per-IRQ handler
        - irqaction structure list

The handler can be one of the 3 standard handlers - "level", "edge" and
"simple", or your own specific handler if you need to do something special.

The "level" handler is what we currently have - its pretty simple.
"edge" knows about the brokenness of such IRQ implementations - that you
need to leave the hardware IRQ enabled while processing it, and queueing
further IRQ events should the IRQ happen again while processing.  The
"simple" handler is very basic, and does not perform any hardware
manipulation, nor state tracking.  This is useful for things like the
SMC9196 and USAR above.

So, what's changed?
===================

1. Machine implementations must not write to the irqdesc array.

2. New functions to manipulate the irqdesc array.  The first 4 are expected
   to be useful only to machine specific code.  The last is recommended to
   only be used by machine specific code, but may be used in drivers if
   absolutely necessary.

        set_irq_chip(irq,chip)
                Set the mask/unmask methods for handling this IRQ

        set_irq_handler(irq,handler)
                Set the handler for this IRQ (level, edge, simple)

        set_irq_chained_handler(irq,handler)
                Set a "chained" handler for this IRQ - automatically
                enables this IRQ (eg, Neponset and SA1111 handlers).

        set_irq_flags(irq,flags)
                Set the valid/probe/noautoenable flags.

        set_irq_type(irq,type)
                Set active the IRQ edge(s)/level.  This replaces the
                SA1111 INTPOL manipulation, and the set_GPIO_IRQ_edge()
                function.  Type should be one of IRQ_TYPE_xxx defined in
		<linux/irq.h>

3. set_GPIO_IRQ_edge() is obsolete, and should be replaced by set_irq_type.

4. Direct access to SA1111 INTPOL is deprecated.  Use set_irq_type instead.

5. A handler is expected to perform any necessary acknowledgement of the
   parent IRQ via the correct chip specific function.  For instance, if
   the SA1111 is directly connected to a SA1110 GPIO, then you should
   acknowledge the SA1110 IRQ each time you re-read the SA1111 IRQ status.

6. For any child which doesn't have its own IRQ enable/disable controls
   (eg, SMC9196), the handler must mask or acknowledge the parent IRQ
   while the child handler is called, and the child handler should be the
   "simple" handler (not "edge" nor "level").  After the handler completes,
   the parent IRQ should be unmasked, and the status of all children must
   be re-checked for pending events.  (see the Neponset IRQ handler for
   details).

7. fixup_irq() is gone, as is `arch/arm/mach-*/include/mach/irq.h`

Please note that this will not solve all problems - some of them are
hardware based.  Mixing level-based and edge-based IRQs on the same
parent signal (eg neponset) is one such area where a software based
solution can't provide the full answer to low IRQ latency.