Re: [PATCH 2.5.5] Conversion of hp100 to new PCI interface

[Jean Tourrilhes <jt@bougret.hpl.hp.com>]

On Tue, Feb 26, 2002 at 09:11:15PM -0500, Jeff Garzik wrote:
> > 
> > +       /* Conversion to new PCI API :
> > +        * The memory block we use, lp->page_vaddr, was DMA allocated via
> > +        * pci_alloc_consistent(), so we just need to "retreive" the
> > +        * original mapping to bus/phys address - Jean II */
> >         ringptr->pdl = pdlptr + 1;
> > -       ringptr->pdl_paddr = virt_to_bus(pdlptr + 1);
> > +       ringptr->pdl_paddr = virt_to_phys(pdlptr + 1);
> 
> Nope..  You need to use the mapping value obtained from
> pci_alloc_consistent...

	I don't understand the objection. The memory has been declared
as DMA and the system already manages it as such. What's the catch ?
	If I can't use virt_to_phys(), can I have a function that does
exactly the same ? The new API is heavy enough, and if drivers can't
have something like this it's just messy...
	Just define something like :
------------------------------------------
static inline dma_addr_t pci_map_alloc(struct pci_dev *hwdev, void *ptr)
{
	ret virt_to_phys(ptr);
}
------------------------------------------

> Note for ISA (and EISA and VLB) devices, you also call
> pci_alloc_consistent.  You pass NULL for the pci device.

	That's what I do. But I don't claim it's tested.
	By the way, it seems weird to prefix all those functions with
"pci_" where in fact they are only loosely related to PCI stuff.
	Also, on the inconsistent side, most functions handle pci_dev
== NULL, but not all, and it's not well documented. For example, if
you pass NULL to pci_set_dma_mask(), it will kaboom !

> 	Jeff

	Jean