Re: [PATCH] powerpc/64s/radix: Fix unmapping huge vmaps when CONFIG_HUGETLB_PAGE=n

Excerpts from Daniel Axtens's message of November 26, 2021 4:09 pm:

Hi,

quoted

pmd_huge is defined out to false when HUGETLB_PAGE is not configured,
but the vmap code still installs huge PMDs. This leads to errors
encountering bad PMDs when vunmapping because it is not seen as a
huge PTE, and the bad PMD check catches it. The end result may not
be much more serious than some bad pmd warning messages, because the
pmd_none_or_clear_bad() does what we wanted and clears the huge PTE
anyway.

Huh. So vmap seems to key off arch_vmap_p?d_supported which checks for
radix and HAVE_ARCH_HUGE_VMAP.

quoted

Fix this by checking pmd_is_leaf(), which checks for a PTE regardless
of config options. The whole huge/large/leaf stuff is a tangled mess
but that's kernel-wide and not something we can improve much in
arch/powerpc code.

I guess I'm a bit late to the party here because p?d_is_leaf was added
in 2019 in commit d6eacedd1f0e ("powerpc/book3s: Use config independent
helpers for page table walk") but why wouldn't we just make pmd_huge()
not config dependent?

I guess so it constant folds code if hugetlbfs is not configured 
(and maybe so !huge kernels would correctly print a bad PMD warning if
they got huge PMD in user mappings).

Also, looking at that commit, there are a few places that might still
throw warnings, e.g. find_linux_pte, find_current_mm_pte, pud_page which
seem like they might still throw warnings if they were to encounter a
huge vmap page:

struct page *pud_page(pud_t pud)
{
	if (pud_is_leaf(pud)) {
		VM_WARN_ON(!pud_huge(pud));

Oh, hmm. That is used in vmalloc.c so maybe that warning should be
removed as a false positive. Good catch.

Do these functions need special treatment for huge vmappings()?

find_linux_pte etc could be called for vmaps. I'm not sure I see a
problem in that function.

Thanks,
Nick

Apart from those questions, the patch itself makes sense to me and I can
follow how it would fix a problem.

Reviewed-by: Daniel Axtens <redacted>

Kind regards,
Daniel