Michael Ellerman's on July 23, 2019 8:52 pm:

Nicholas Piggin [off-list ref] writes:

quoted

create_physical_mapping expects physical addresses, but creating and
splitting these mappings after boot is supplying virtual (effective)
addresses. This can be hit by booting with limited memory then probing
new physical memory sections.

Cc: Reza Arbab <redacted>
Fixes: 6cc27341b21a8 ("powerpc/mm: add radix__create_section_mapping()")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>

This is not catastrophic because create_physical_mapping() just uses
start/end to construct virtual addresses anyway, and __va(__va(x)) == __va(x) ?

A bit more subtle, it calls __map_kernel_page with the pa as well.
pfn_pte ends up masking the top 0xc bits out with PTE_RPN_MASK,
which is what saves us.

Hmm, so we really should also have a VM_BUG_ON in pfn_pte if it's
given a pfn with the top PAGE_SHIFT bit or PTE_RPN_MASK bits set.
I'll add that as a patch 5.
 

Although we do pass those through as region_start/end which then go to
memblock_alloc_try_nid(). But I guess that doesn't happen after boot,
which is the case you're talking about.

So I think looks good, change log could use a bit more detail though :)

Thanks for taking a look. I'll resend after a bit more testing and
some changelog improvement.

Thanks,
Nick