Re: [PATCH v2] mm/gup: fix try_grab_compound_head() race with split_huge_page()

From: Andrew Morton
Date: Mon Jun 14 2021 - 22:36:39 EST

On Tue, 15 Jun 2021 03:20:14 +0200 Jann Horn <jannh@xxxxxxxxxx> wrote:

> try_grab_compound_head() is used to grab a reference to a page from
> get_user_pages_fast(), which is only protected against concurrent
> freeing of page tables (via local_irq_save()), but not against
> concurrent TLB flushes, freeing of data pages, or splitting of compound
> pages.
> Because no reference is held to the page when try_grab_compound_head()
> is called, the page may have been freed and reallocated by the time its
> refcount has been elevated; therefore, once we're holding a stable
> reference to the page, the caller re-checks whether the PTE still points
> to the same page (with the same access rights).
> The problem is that try_grab_compound_head() has to grab a reference on
> the head page; but between the time we look up what the head page is and
> the time we actually grab a reference on the head page, the compound
> page may have been split up (either explicitly through split_huge_page()
> or by freeing the compound page to the buddy allocator and then
> allocating its individual order-0 pages).
> If that happens, get_user_pages_fast() may end up returning the right
> page but lifting the refcount on a now-unrelated page, leading to
> use-after-free of pages.
> To fix it:
> Re-check whether the pages still belong together after lifting the
> refcount on the head page.
> Move anything else that checks compound_head(page) below the refcount
> increment.
> This can't actually happen on bare-metal x86 (because there, disabling
> IRQs locks out remote TLB flushes), but it can happen on virtualized x86
> (e.g. under KVM) and probably also on arm64. The race window is pretty
> narrow, and constantly allocating and shattering hugepages isn't exactly
> fast; for now I've only managed to reproduce this in an x86 KVM guest with
> an artificially widened timing window (by adding a loop that repeatedly
> calls `inl(0x3f8 + 5)` in `try_get_compound_head()` to force VM exits,
> so that PV TLB flushes are used instead of IPIs).
> As requested on the list, also replace the existing VM_BUG_ON_PAGE()
> with a warning and bailout. Since the existing code only performed the
> BUG_ON check on DEBUG_VM kernels, ensure that the new code also only
> performs the check under that configuration - I don't want to mix two
> logically separate changes together too much.
> The macro VM_WARN_ON_ONCE_PAGE() doesn't return a value on !DEBUG_VM,
> so wrap the whole check in an #ifdef block.
> An alternative would be to change the VM_WARN_ON_ONCE_PAGE() definition
> for !DEBUG_VM such that it always returns false, but since that would
> differ from the behavior of the normal WARN macros, it might be too
> confusing for readers.
> ...
> --- a/mm/gup.c
> +++ b/mm/gup.c
> @@ -43,8 +43,25 @@ static void hpage_pincount_sub(struct page *page, int refs)
> atomic_sub(refs, compound_pincount_ptr(page));
> }
> +/* Equivalent to calling put_page() @refs times. */
> +static void put_page_refs(struct page *page, int refs)
> +{
> + if (VM_WARN_ON_ONCE_PAGE(page_ref_count(page) < refs, page))
> + return;
> +#endif

Well dang those ifdefs.

With CONFIG_DEBUG_VM=n, this expands to

if (((void)(sizeof((__force long)(page_ref_count(page) < refs))))

which will fail with "void value not ignored as it ought to be".
Because VM_WARN_ON_ONCE_PAGE() is an rval with CONFIG_DEBUG_VM=y and is
not an rval with CONFIG_DEBUG_VM=n. So the ifdefs are needed.

I know we've been around this loop before, but it still sucks! Someone
please remind me of the reasoning?

Can we do

#define VM_WARN_ON_ONCE_PAGE(cond, page) {