Re: [Qemu-devel] [PATCH 07/15] ac97: convert to new PCI API

[malc]

On Tue, 9 Feb 2010, Anthony Liguori wrote:

> On 02/09/2010 05:36 PM, malc wrote:
> > Let's see:
> > 
> > Currently we have this
> > 
> > 
> > readb(...):
> >    dostuff
> >    return stuff
> > 
> > readw(...):
> >    dostuff
> >    return stuff
> >    
> 
> And this is completely wrong.

It's completely right, until some future C standard implements proper
algebraic data types.

> 
> For the most part, device models don't consistently handle access to registers
> via their non native sizes.  Some devices return the lower part of a dword
> register when accessed with a word accessor.  Some devices only return the
> register when there's a dword access.

That's up to device to register an accessor and return what's appropriate.

> 
> What's worse, is that if a device only registers a byte accessor, because of
> the default accessors, a word access returns two registers but a dword access
> returns ~0U.  Some device models actually rely on this behaviour.
> 
> > You are replacing it with
> > 
> > read(size...):
> >    if (size == 1): do1
> >    elif (size == 2): do2
> >    else: # and here your code assumes that everything is handy dandy
> >          # and size is 4
> >      do4
> >    
> 
> This is ugly because it's a programmatic conversion.  Once we have this API,
> we can switch to:
> 
> read(addr, size...):
>   switch(addr):
>   case REG0:
>     return s->reg0;
>   case REG1:
>     return s->reg1;

This is exactly equivalent to your original if, due to the C language
being what it is.

> Along with having a table like:
> 
> pci_regs = { {REG0, DWORD}, {REG1, BYTE}, {REG2, BYTE}, {REG3, WORD} }
> 
> That allows the PCI layer to invoke the callback and do the right operations
> to transparently handle the access size without the device model having to
> deal with it.  The reason I've left size in the callback at all is that I
> suspect there will always be a device that behaves weirdly and needs to know
> about the accessor size.  But for the most part, I think this model will make
> things much more consistent and eliminate a huge class of emulation bugs.

Eh? Care to name one emulation bug?

> 
> We can even take it further, and do something like:
> 
> pci_regs = {
>   PCI_REG_DWORD(REG0, DeviceState, reg0),
>   PCI_REG_BYTE(REG1, DeviceState, reg1),
>   ...
> }
> 
> In which case, we only need to handle the case in switch() if we need to
> implement a side effect of the register operation.
> 
> But none of this is possible if we completely rely on every device to
> implement non-dword access in it's own broken way.
> 

Lovely, we are heading full speed into fanatsy land, where PCI BUS itself
accesses device specific stuff.

-- 
mailto:address@hidden