Re: [Qemu-devel] [PATCH 01/15] pc: merge pc_piix.c into pc.c

[Anthony Liguori <aliguori@us.ibm.com>]

On 01/27/2012 07:32 AM, Jan Kiszka wrote:
> On 2012-01-27 14:07, Anthony Liguori wrote:
>> On 01/27/2012 02:50 AM, Jan Kiszka wrote:
>>> On 2012-01-26 20:00, Anthony Liguori wrote:
>>>> A long time ago, there was a grand plan to merge q35 chipset support.  The start
>>>> of that series was a refactoring of pc.c which split a bunch of the "common"
>>>> functionality into a separate file that could be shared by the two.
>>>>
>>>> But q35 never got merged and the refactoring, in retrospect, just made things
>>>> worse.  Making things proper objects and using composition is the right way
>>>> to share common devices.
>>>>
>>>> By pulling these files back together, we can start to fix some of this mess.
>>>
>>> There are surely things to clean up and improve, but a clear NACK for
>>> the general direction.
>>
>> Hi Jan,
>>
>> I think you're missing the bigger picture here.  Once this refactoring finishes,
>> here's what we'll be left with:
>>
>> 1) pc_init creates an I440FX, any bus devices (ISA serial port, PCI vga and
>> nics, etc.), sets properties appropriately, and realizes the devices.
>>
>> 2) I440FX is-a PCIHost, has-a I440FX-PMC, has-a PIIX3
>>
>> 3) PIIX3 has-a RTC, has-a I8042, has-a DMAController, etc.
>>
>> Memory creation is done by the I440FX-PMC.
>>
>> Now it's true that a newer chipset is going to be similar.  It will likely have
>> a SuperI/O chip that looks similar to PIIX3.  The right way to share code would
>> be to move most of the PIIX3 functionality to a base class (PCSuperIO) that
>> PIIX3 inherits from.
>>
>> This is probably how to support ISAPC properly too.
>
> The ISAPC is differently composed. The board creates all those
> individual chips that are otherwise part of the SuperIO block of the
> chipset. And IRQ wiring is different.

A SuperIO device can still have irq properties for each device that determines 
which ISA bus irq is used.

The other alternative is to stop making devices ISADevice, and then having 
SuperIO expose a bunch of device specific IRQs.  Those IRQs can then be routed 
in whatever way makes sense.

This is really the Right Approach but it's most likely a compatibility breaker 
so I'm trying my best to avoid these types of refactorings right now.

> So, no, I don't think it is the
> right model. I would rather think of a pc_isa.c that does the proper
> composing.

Objects compose other objects.  Functions should not be creating devices.  So if 
you're view of pc_isa.c that there's a pc_basic_init() that takes a bunch of 
devicestate pointers, then that's definitely not the direction I'm heading.

Note that if we ever want to get to a board configuration file, we need to have 
three explicit steps in device creation:

1) device initialization where devices are allocates and the composition tree is 
filled out (which makes sure that every device has an addressable path)

2) device property setting (which requires all devices have an addressable path)

3) device realization where property settings are validated, and then any 
initialization that depends on properties is done.

The problem with the current code is that it doesn't split up these phases. 
Modeling composition really helps to get this split because it forces you to 
think about things in terms of distinct phases.

>>   Once I sort out interrupts,
>> I'll attempt to tackle that.  My guess is that a SuperIO chip could be an
>> ISADevice and that we could simply make the PIIX3 has-a SuperIO.  Then the ISAPC
>> would have a trivial ISA chipset that has-a SuperIO.
>>
>> This is fairly trivial to do once we have the right structure to the code.
>>
>> But the current code has the wrong structure which is why there's so much
>> pointer chasing and passing.
>
> Just look at your code and count the generic, PIIX3-independent
> functions.

I see zero.

"PIIX3-independent" means that some other piece of code would consume it in 
exactly the same way.  pc.c had a bunch of spaghetti in it doing tricks with if 
(pci_enabled) that ended up encoding two very separate paths into one maze of code.

> Keep them in pc.c, move the rest to pc_piix.c. You could try
> to model the ISA accordingly. I think some pc_isa.c would help to
> establish a good split-up already now, in the absence of a third chipset.

I think creating a proper ISA model is a good idea and I'll certainly get there.

Regards,

Anthony Liguori

>>
>>>
>>> It's undoubted that we need a more modern chipset than this ancient
>>> PIIX3, rather sooner than later. And it is clear that there is a good
>>> amount of generic functions in pc.c for building a PC, even a fairly
>>> modern one. So we need a common lib for PC chipsets and would only
>>> revert what you start here.
>>
>> Sorry, but I don't view this as a useful requirement.  Today we support two
>> types of PCs: an i440fx based system and an ISA-only system.  We should
>> concentrate on modeling those two systems in the most natural way sharing as
>> much code as possible.
>
> A modern chipset is the only sane way to add things like PCIe,
> hotplugging, power management, etc., and to enable standard PC
> components like AHCI or EHCI/xHCI by default.
>
> Jan
>