Re: [PATCH v2 1/2] arm64: don't make early_*map() calls post paging_init()

[Will Deacon <will.deacon-5wv7dgnIgG8@public.gmane.org>]

On Wed, Jan 07, 2015 at 01:31:09PM +0000, Leif Lindholm wrote:
> On Wed, Jan 07, 2015 at 01:13:06PM +0000, Ard Biesheuvel wrote:
> > >> > -void __init efi_idmap_init(void)
> > >> > +void __init efi_memmap_init(void)
> > >> >  {
> > >> > +   u64 mapsize;
> > >> > +
> > >> >     if (!efi_enabled(EFI_BOOT))
> > >> >             return;
> > >> >
> > >> > +   /* replace early memmap mapping with permanent mapping */
> > >> > +   mapsize = memmap.map_end - memmap.map;
> > >> > +   memmap.map = (__force void *)ioremap_cache((phys_addr_t)memmap.phys_map,
> > >> > +                                              mapsize);
> > >>
> > >> ioremap_cache() could potententially fail here if the phys_map address
> > >> doesn't have a valid pfn (not in the kernel linear ram mapping) because
> > >> some of the underlying vm support hasn't been initialized yet.
> > >
> > > Can you be more specific about the case you have in mind, please? pfn_valid
> > > uses the memblocks on arm64, and that should all have been sorted out in
> > > paging_init(). What's the issue that you're anticipating?
> > 
> > I think Mark's concern is that it is too early to call
> > __get_free_page(), which is what happens if ioremap_cache() finds that
> > the requested address is not covered by the existing linear mapping.
> > Currently, UEFI reserved RAM regions are covered by the linear
> > mapping, but that is something we intend to change in the future.
> 
> Which shouldn't be a problem, right? Since this function will be going
> away with your "stable mappings" set, and the remap call bumped down
> to an early initcall in arm64_enter_virtual_mode() (or potential
> future name for that function).

Sounds reasonable to me... Ard?

However, I'd certainly like something in the commit log and/or a comment
in the code with our reasoning.

Will