[PATCH] arm64: support ACPI tables outside of kernel RAM

[Mark Salter]

On Fri, 2015-05-22 at 11:34 +0100, Catalin Marinas wrote:
> On Mon, May 18, 2015 at 06:49:28PM +0200, Ard Biesheuvel wrote:
> > On 18 May 2015 at 18:41, Catalin Marinas <catalin.marinas at arm.com> wrote:
> > > On Mon, May 18, 2015 at 09:58:45AM -0400, Mark Salter wrote:
> > >> On Mon, 2015-05-18 at 12:11 +0100, Catalin Marinas wrote:
> > >> > On Thu, May 14, 2015 at 10:22:53AM -0400, Mark Salter wrote:
> > >> > > There is no guarantee that ACPI tables will be located in RAM linearly
> > >> > > mapped by the kernel. This could be because UEFI placed them below the
> > >> > > kernel image or because mem= places them beyond the reach of the linear
> > >> > > kernel mapping. Even though these tables are outside the linear mapped
> > >> > > RAM, they still need to be accessed as normal memory in order to support
> > >> > > unaligned accesses from ACPI code. In this case, the page_is_ram() test
> > >> > > in acpi_os_ioremap() is not sufficient.
> > >> >
> > >> > And can we not simply add the rest of the RAM to the resource list as
> > >> > "System RAM" without being part of memblock?
> > >>
> > >> If it is in "System RAM", then it needs a valid pfn and struct page.
> > >> Parts of the kernel expect that (page_is_ram(), memory hotplug, etc).
> > >
> > > OK, I had the impression that we could get away with this.
> > >
> > >> > > Additionally, if the table spans multiple pages, it may fall partially
> > >> > > within the linear map and partially without. If the table overlaps the
> > >> > > end of the linear map, the test for whether or not to use the existing
> > >> > > mapping in ioremap_cache() could lead to a panic when ACPI code tries
> > >> > > to access the part beyond the end of the linear map. This patch
> > >> > > attempts to address these problems.
> > >> >
> > >> > That's a problem with ioremap_cache() that should be fixed independently.
> > >>
> > >> I can submit that separately if you prefer.
> > >
> > > Yes, please.
> > >
> > >> > Ideally, I'd like to see the ACPI code use different APIs to distinguish
> > >> > between table access in RAM and device access, so that we don't have to
> > >> > guess whether the page is RAM or not.
> > >>
> > >> I don't think the ACPI code has enough info to make that decision, but
> > >> I'm not sure honestly.
> > >
> > > Do we have a guarantee that UEFI tells the kernel about the whole RAM?
> > 
> > Yes, the UEFI memory map must describe all of RAM, no matter how it is
> > used. I may also describe some MMIO regions, but typically only
> > regions that it needs itself to implement the UEFI Runtime Services
> > (e.g., RTC base address, NOR flash for the variable store)
> > 
> > So we could potentially query the UEFI memory map directly to find out
> > whether some otherwise unqualified region is backed by RAM or not,
> > although I'd prefer some intermediate data structure (such as the
> > physmem memblock table) if we go that route.
> 
> OK, so my preferred options, in this order:
> 
> 1. Change the core ACPI kernel code to distinguish between mapping I/O
>    or RAM (could be as simple as acpi_map not using acpi_os_ioremap but
>    another API). I guess the code knows when it plans to map tables or
>    I/O registers
> 
> 2. If the above is not possible, add the extra checks as per Mark's
>    patch but I would rather call this resource "UEFI RAM" than "ACPI",
>    it's not really ACPI specific.
> 
Actually, it is ACPI specific. The patch only registers resources for
EfiACPIReclaimMemory and EfiACPIMemoryNVS regions which are also
marked as cacheable. On x86 these show up in /proc/iomem as
"ACPI Tables" and "ACPI Non-volatile Storage". I used "ACPI RAM" to
avoid having to search for two strings.