Re: [PATCH] x86: efi: avoid BUILD_BUG_ON() for non-constant p4d_index

On Fri, 15 Jan 2021 at 21:27, Arvind Sankar [off-list ref] wrote:

On Fri, Jan 15, 2021 at 02:07:51PM -0500, Arvind Sankar wrote:

quoted

On Thu, Jan 07, 2021 at 11:34:15PM +0100, Arnd Bergmann wrote:

quoted

From: Arnd Bergmann <arnd@arndb.de>

When 5-level page tables are enabled, clang triggers a BUILD_BUG_ON():

x86_64-linux-ld: arch/x86/platform/efi/efi_64.o: in function `efi_sync_low_kernel_mappings':
efi_64.c:(.text+0x22c): undefined reference to `__compiletime_assert_354'

Use the same method as in commit c65e774fb3f6 ("x86/mm: Make PGDIR_SHIFT
and PTRS_PER_P4D variable") and change it to MAYBE_BUILD_BUG_ON(),
so it only triggers for constant input.

Link: https://github.com/ClangBuiltLinux/linux/issues/256
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
---
 arch/x86/platform/efi/efi_64.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/arch/x86/platform/efi/efi_64.c b/arch/x86/platform/efi/efi_64.c
index e1e8d4e3a213..62bb1616b4a5 100644
--- a/arch/x86/platform/efi/efi_64.c
+++ b/arch/x86/platform/efi/efi_64.c

@@ -137,8 +137,8 @@ void efi_sync_low_kernel_mappings(void)
     * As with PGDs, we share all P4D entries apart from the one entry
     * that covers the EFI runtime mapping space.
     */
-   BUILD_BUG_ON(p4d_index(EFI_VA_END) != p4d_index(MODULES_END));
-   BUILD_BUG_ON((EFI_VA_START & P4D_MASK) != (EFI_VA_END & P4D_MASK));
+   MAYBE_BUILD_BUG_ON(p4d_index(EFI_VA_END) != p4d_index(MODULES_END));
+   MAYBE_BUILD_BUG_ON((EFI_VA_START & P4D_MASK) != (EFI_VA_END & P4D_MASK));

    pgd_efi = efi_pgd + pgd_index(EFI_VA_END);
    pgd_k = pgd_offset_k(EFI_VA_END);
--

2.29.2

I think this needs more explanation as to why clang is triggering this.
The issue mentions clang not inline p4d_index(), and I guess not
performing inter-procedural analysis either?

For the second assertion there, everything is always constant AFAICT:
EFI_VA_START, EFI_VA_END and P4D_MASK are all constants regardless of
CONFIG_5LEVEL.

For the first assertion, it isn't technically constant, but if
p4d_index() gets inlined, the compiler should be able to see that the
two are always equal, even though ptrs_per_p4d is not constant:
      EFI_VA_END >> 39 == MODULES_END >> 39
so the masking with ptrs_per_p4d-1 doesn't matter for the comparison.

As a matter of fact, it seems like the four assertions could be combined
into:
      BUILD_BUG_ON((EFI_VA_END & P4D_MASK) != (MODULES_END & P4D_MASK));
      BUILD_BUG_ON((EFI_VA_START & P4D_MASK) != (EFI_VA_END & P4D_MASK));
instead of separately asserting they're the same PGD entry and the same
P4D entry.

Thanks.

I actually don't quite get the MODULES_END check -- Ard, do you know
what that's for?

Maybe Boris remembers? He wrote the original code for the 'new' EFI
page table layout.

What we really should be checking is that EFI_VA_START is in the top-most
PGD entry and the top-most P4D entry, since we only copy PGD/P4D entries
before EFI_VA_END, but not after EFI_VA_START. So the checks should
really be
        BUILD_BUG_ON(((EFI_VA_START - 1) & P4D_MASK) != (-1ul & P4D_MASK));
        BUILD_BUG_ON(((EFI_VA_START - 1) & P4D_MASK) != (EFI_VA_END & P4D_MASK));
imo. I guess that's what using MODULES_END is effectively checking, but
it would be clearer to check it directly.

This obviously needs a comment, but checking that everything lives in
the top 512 GB of the kernel VA space seems sufficient to me,