Re: [PATCH v3] mm: pgtable: protect lockless kernel page table walks with RCU

[David CARLIER]

> It's not teardown is it? The only caller is vmap_try_huge_pmd().

  Right, it's the huge-pmd install collapsing the pte table, not teardown.

  > And if !CONFIG_ASYNC_KERNEL_PGTABLE_FREE then ... everything's
broken again isn't it?

  Yeah, that's the real bug. Non-async pagetable_free_kernel() frees
  immediately, so the rcu_read_lock() does nothing. v3 only ever helped the
  async config.

  > now we're unconditionally waiting a grace period ... this seems... silly?

  It only fires for kernel pte tables freed by the huge collapse and by
  hot-remove, not every page table free. User pte tables already rcu-free,
  this just does the same for the kernel ones.

  > why the arm64 page table split path is safe

  It only walks the range it's actively splitting, so nothing frees a table
  under it. Same for the other walk_kernel_page_table_range() users. ptdump
  walks the whole kernel range for the W+X check and can't take exclusive
  access, so it's the one that races.

  > I'd rather you add walk_kernel_page_table_range_rcu()

  Will do.

  Resend standalone in a day or two.