Re: [GIT PULL] arm64 fix for 5.14

[Linus Torvalds <torvalds@linux-foundation.org>]

On Fri, Aug 27, 2021 at 12:40 AM Christoph Hellwig <hch@lst.de> wrote:
>
> > In a PC world that would be (for example) the legacy PCI space at
> > 0xa0000-0xfffff, but I could easily imagine other platforms having
> > other situations.
>
> So what would be the correct check for "this is not actually page backed
> normal RAM"?

It would probably be interesting to have the arm people explain the
call chain for the warning that caused that revert, so we'd have a
very concrete example of the situation that goes wrong, but taking a
wild stab at it, the code might be something like

            /* Don't allow RAM to be mapped */
            if (WARN_ON_ONCE(phys_addr_is_ram(phys_addr)))
                    return DMA_MAPPING_ERROR;

and then having something like

  static inline bool phys_addr_is_ram(phys_addr_t phys_addr)
  {
        unsigned long pfn = PHYS_PFN(phys_addr);

        if (!pfn_valid(pfn))
                return false;
        return is_zero_pfn(pfn) || !PageReserved(pfn_to_page(pfn));
  }

might be close to right.

The ARM code actually uses that complex pfn_to_section_nr() and
memblock_is_memory() etc. That seems a bit of an overkill, since the
memblock code should have translated all that into being reserved.

But again, I don't actually know exactly what triggered the issue on
ARM, so the above is just my "this seems to be a more proper check"
suggestion.

Will?

                   Linus