[PATCH v3] ARM: early fixmap support for earlycon

[Mark Salter]

On Sat, 2015-06-06 at 10:29 -0500, Rob Herring wrote:
> On Sat, Jun 6, 2015 at 8:29 AM, Stefan Agner <stefan at agner.ch> wrote:
> > On 2015-06-06 14:48, Russell King - ARM Linux wrote:
> >> On Sat, Jun 06, 2015 at 02:31:28PM +0200, Stefan Agner wrote:
> > <snip>
> >>> --- a/arch/arm/include/asm/fixmap.h
> >>> +++ b/arch/arm/include/asm/fixmap.h
> >>> @@ -6,9 +6,13 @@
> >>>  #define FIXADDR_TOP         (FIXADDR_END - PAGE_SIZE)
> >>>
> >>>  #include <asm/kmap_types.h>
> >>> +#include <asm/pgtable.h>
> >>>
> >>>  enum fixed_addresses {
> >>> -    FIX_KMAP_BEGIN,
> >>> +    FIX_EARLYCON_MEM_BASE,
> >>> +    __end_of_permanent_fixed_addresses,
> >>> +
> >>> +    FIX_KMAP_BEGIN = __end_of_permanent_fixed_addresses,
> >>>      FIX_KMAP_END = FIX_KMAP_BEGIN + (KM_TYPE_NR * NR_CPUS) - 1,
> >>>
> >>>      /* Support writing RO kernel text via kprobes, jump labels, etc. */
> >>> @@ -18,7 +22,14 @@ enum fixed_addresses {
> >>>      __end_of_fixed_addresses
> >>>  };
> >>>
> >>> +#define FIXMAP_PAGE_COMMON  (L_PTE_YOUNG | L_PTE_PRESENT | L_PTE_XN | L_PTE_DIRTY)
> >>> +
> >>> +#define FIXMAP_PAGE_NORMAL  (FIXMAP_PAGE_COMMON | L_PTE_MT_WRITEBACK)
> >>> +#define FIXMAP_PAGE_IO              (FIXMAP_PAGE_COMMON | L_PTE_MT_DEV_SHARED | L_PTE_SHARED)
> >>> +#define FIXMAP_PAGE_NOCACHE FIXMAP_PAGE_IO
> >>
> >> I'm really not happy with this.  What are the expected semantics of the
> >> set_fixmap_nocache() and set_fixmap_offset_nocache() ?  Are they there
> >> for mapping a device, or are they there for mapping _memory_ ?
> >>
> >> I would prefer that FIXMAP_PAGE_NOCACHE is _not_ provided until the
> >> semantics of that can be clarified, as there is a difference between the
> >> two on ARM.
> >
> > set_fixmap_nocache is used for PCIe and IO APIC stuff, so actually IO
> > devices. However, as far as I understand the x86 page table definitions
> > in pgtable_types.h, IO devices are not explicitly uncached on x86. The
> > define set_fixmap_nocache make sure IO devices are mapped uncached...
> > Maybe set_fixmap_io_nocache instead of set_fixmap_nocache would be more
> > appropriate...?
> >
> > Maybe the initial authors of the common fixmap.h have a more clear
> > understanding of the expected semantics, Mark?
> 
> They should be the same as ioremap and ioremap_nocache which are
> aliases of each other on ARM. If you want RAM then you should be using
> FIXMAP_PAGE_NORMAL. Perhaps we can come up with some error checking
> here.
> 
> > Since the focus currently is on set_fixmap_io, we could define
> > FIXMAP_PAGE_IO only for the time being...?
> 
> We could, but I think it is better to sort this out now.
> 

I agree with Rob here. AFAICT, x86 is the only arch which actually uses
set_fixmap_nocache. Other arches created definitions for it, but all
uses appear to depend on CONFIG_X86.