Re: [GIT PULL] arm64 fix for 5.14

[Catalin Marinas]

On Fri, Aug 27, 2021 at 10:16:27AM -0700, Linus Torvalds wrote:
> On Fri, Aug 27, 2021 at 10:10 AM Christoph Hellwig <hch@xxxxxx> wrote:
> > They CCed me on their earlier discussion, but I did not catch up on it
> > until you responded to the pull request  If I understood it correct it
> > was about a platform device mapping a MMIO region (like a PCI bar),
> > but something about section alignment cause pfn_valid to mistrigger.
> 
> Yeah, so I can easily see the maxpfn numbers can easily end up being
> rounded up to a whole memory section etc.
> 
> I think my suggested solution should JustWork(tm) - exactly because if
> the area is then in that "this pfn is valid" area, it will
> double-check the actual underlying page.
> 
> That said, I think x86 avoids the problem another way - by just making
> sure max_pfn is exact. That works too, as long as there are no holes
> in the RAM map that might be used for PCI BAR's.
> 
> So I think arm could fix it that way too, depending on their memory layout.

The suggested solution in the original thread was to change the generic
DMA code to use memblock_is_memory() instead of pfn_valid():

https://lore.kernel.org/lkml/b720e7c8-ca44-0a25-480b-05bf49d03c35@xxxxxxxxxx/

Given how late we discovered this in the -rc cycle, the decision was to
revert the pfn_valid() patch. We'll re-instate it at some point but
someone needs to sanity check the other pfn_valid() call sites and the
expected semantics.

-- 
Catalin