Re: [PATCH v2 09/11] s390: mm: Convert to GENERIC_IOREMAP

From: Niklas Schnelle
Date: Mon Aug 22 2022 - 11:51:07 EST


On Sun, 2022-08-21 at 00:05 -0700, Christoph Hellwig wrote:
> > +void __iomem *
> > +arch_ioremap(phys_addr_t *paddr, size_t size, unsigned long *prot_val)
> > {
> > if (!static_branch_unlikely(&have_mio))
> > + return (void __iomem *) *paddr;
> > + return NULL;
>
> This logic isn't new in the patch, but it could really use a comment
> as it is rather non-obvious.

Yes, makes sense. Basically we fake MMIO addresses because the s390
architecture doesn't have MMIO as a concept. That is until the PCI MIO
instructions introduced pseudo-MMIO though only for specific PCI
load/store instructions. Without those PCI BAR spaces as well as config
space is accessed with so called function handles. As these are a bad
fit for Linux' MMIO based APIs we create fake MMIO addresses (called
address cookies) that encode an index into the zpci_iomap_start[] which
can be decoded by our implementation of ioread*/iowrite*().

I don't think this is the right place to describe this overall scheme
in detail but maybe we can leave a a good bread crumb. Maybe something
like below?

/*
* When PCI MIO instructions are unavailable the "physical" address encodes
* a hint for accessing the PCI memory space it represents. Just pass it
* unchanged such that ioread/iowrite can decode it.
*/