On Tue, Oct 15, 2024 at 01:12:36PM +0200, David Hildenbrand wrote:
> pmd_leaf()/pud_leaf() only implies a pmd_present()/pud_present() check on
> some architectures.
Should we clarify what behaviour we actually want from arch code?
> We really should check for
> pmd_present()/pud_present() first.
>
> This should explain the report we got on ppc64 (which has
> CONFIG_PGTABLE_HAS_HUGE_LEAVES set in the config) that triggered:
> VM_WARN_ON_ONCE(pmd_leaf(pmdp_get_lockless(pmdp)));
>
> Likely we had a PMD migration entry for which pmd_leaf() did not
> trigger. We raced with restoring the PMD migration entry, and suddenly
> saw a pmd_leaf(). In this case, pte_offset_map_lock() saved us from more
> trouble, because it rechecks the PMD value, but we would not have processed
> the migration entry -- which is not too bad because the only user of
> FW_MIGRATION is KSM for unsharing, and KSM only applies to small folios.
>
> Further, we shouldn't re-read the PMD/PUD value for our warning, the
> primary purpose of the VM_WARN_ON_ONCE() is to find spurious use of
> pmd_leaf()/pud_leaf() without CONFIG_PGTABLE_HAS_HUGE_LEAVES.
>
> As a side note, we are currently not implementing FW_MIGRATION support
> for PUD migration entries, which likely should exist due to hugetlb. Add
> a TODO so this won't fall through the cracks if more FW_MIGRATION users
> get added.
>
> Fixes: aa39ca6940f1 ("mm/pagewalk: introduce folio_walk_start() + folio_walk_end()")
> Reported-by: syzbot+7d917f67c05066cec295@xxxxxxxxxxxxxxxxxxxxxxxxx
> Closes: https://lkml.kernel.org/r/670d3248.050a0220.3e960.0064.GAE@xxxxxxxxxx
> Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> Cc: Jann Horn <jannh@xxxxxxxxxx>
> Signed-off-by: David Hildenbrand <david@xxxxxxxxxx>
Acked-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
--
Kiryl Shutsemau / Kirill A. Shutemov
