Re: [PATCH 2/4] msi vector targeting abstractions

[Mark Maule]

On Wed, Dec 21, 2005 at 12:05:58PM -0700, Matthew Wilcox wrote:
> On Wed, Dec 21, 2005 at 12:42:41PM -0600, Mark Maule wrote:
> 
> > @@ -108,28 +125,38 @@
> >     		if (!(pos = pci_find_capability(entry->dev, PCI_CAP_ID_MSI)))
> >  			return;
> >  
> > +		pci_read_config_dword(entry->dev, msi_upper_address_reg(pos),
> > +			&address_hi);
> >  		pci_read_config_dword(entry->dev, msi_lower_address_reg(pos),
> > -			&address.lo_address.value);
> > -		address.lo_address.value &= MSI_ADDRESS_DEST_ID_MASK;
> > -		address.lo_address.value |= (cpu_physical_id(dest_cpu) <<
> > -									MSI_TARGET_CPU_SHIFT);
> > -		entry->msi_attrib.current_cpu = cpu_physical_id(dest_cpu);
> > +			&address_lo);
> > +
> > +		msi_callouts.msi_target(vector, dest_cpu,
> > +					&address_hi, &address_lo);
> > +
> > +		pci_write_config_dword(entry->dev, msi_upper_address_reg(pos),
> > +			address_hi);
> >  		pci_write_config_dword(entry->dev, msi_lower_address_reg(pos),
> > -			address.lo_address.value);
> > +			address_lo);
> 
> But actually, I don't understand why you don't just pass a msg_address
> pointer to msi_target instead.

Mainly I did it this way 'cause msg_address seems to be geared toward specific
hw (apic?).  In the case of altix interrupt hw, we don't know about
dest_mode et. al., but only care about the raw address.

I think this style makes it clearer that the core code should only be
using opaque data when interacting with the platform hooks and the MSI
registers.

> 
> (last two points apply throughtout this patch)
> 
> >  
> > +	(*msi_callouts.msi_teardown)(vector);
> > +
> 
> Yuck.  There's a reason C allows you to call through function pointers as if
> they were functions.

My bad ... I used the alternate style elsewhere, just botched this one up.

> 
> > +int
> > +msi_register_callouts(struct msi_callouts *co)
> > +{
> > +	msi_callouts = *co;	/* structure copy */
> > +	return 0;
> 
> Why do it this way instead of having a pointer to a struct?

Are you suggesting just have:

struct msi_callouts *msi_callouts = (some default value or NULL)

and then having each platform just assign msi_callouts in their msi_arch_init?

Doesn't matter to me either way ... I thought having an interface to set
the callouts was cleaner.

> 
> > -struct msg_data {
> > +union msg_data {
> > +	struct {
> 
> How about leaving struct msg_data alone and adding
> 
> union good_name {
> 	struct msg_data;
> 	u32 value;
> }
> 
> Or possibly struct msg_data should just be deleted and we should use
> shift/mask to access the contents of it.  ISTR GCC handled that much
> better.

Christoph had similiar comments.  Will put some thought into it.

Mark

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/