Re: [PATCH] ubsan: disable unsigned-integer-overflow sanitizer with clang

From: Arnd Bergmann
Date: Wed Jan 06 2021 - 04:13:52 EST


On Tue, Jan 5, 2021 at 10:25 AM Arnd Bergmann <arnd@xxxxxxxxxx> wrote:
>
> On Mon, Jan 4, 2021 at 11:33 PM Nathan Chancellor
> <natechancellor@xxxxxxxxx> wrote:
> > On Wed, Dec 30, 2020 at 05:13:03PM +0100, Marco Elver wrote:
> > > On Wed, 30 Dec 2020 at 16:47, Arnd Bergmann <arnd@xxxxxxxxxx> wrote:
> > > >
> > > > From: Arnd Bergmann <arnd@xxxxxxxx>
> > > >
> > > > Building ubsan kernels even for compile-testing introduced these
> > > > warnings in my randconfig environment:
> > > >
> > > > crypto/blake2b_generic.c:98:13: error: stack frame size of 9636 bytes in function 'blake2b_compress' [-Werror,-Wframe-larger-than=]
> > > > static void blake2b_compress(struct blake2b_state *S,
> > > > crypto/sha512_generic.c:151:13: error: stack frame size of 1292 bytes in function 'sha512_generic_block_fn' [-Werror,-Wframe-larger-than=]
> > > > static void sha512_generic_block_fn(struct sha512_state *sst, u8 const *src,
> > > > lib/crypto/curve25519-fiat32.c:312:22: error: stack frame size of 2180 bytes in function 'fe_mul_impl' [-Werror,-Wframe-larger-than=]
> > > > static noinline void fe_mul_impl(u32 out[10], const u32 in1[10], const u32 in2[10])
> > > > lib/crypto/curve25519-fiat32.c:444:22: error: stack frame size of 1588 bytes in function 'fe_sqr_impl' [-Werror,-Wframe-larger-than=]
> > > > static noinline void fe_sqr_impl(u32 out[10], const u32 in1[10])
> > > >
> > > > Further testing showed that this is caused by
> > > > -fsanitize=unsigned-integer-overflow.
> > > >
> > > > The one in blake2b immediately overflows the 8KB stack area on 32-bit
> > > > architectures, so better ensure this never happens by making this
> > > > option gcc-only.
> >
> > This patch also fixes the failed BUILD_BUG issue in mm/mremap.c that you
> > sent a patch for [1], along with a couple of other issues I see such as:
>
> I'm fairly sure I still saw that BUILD_BUG() even after I had applied this
> patch, I would guess that one just depends on inlining decisions that
> are influenced by all kinds of compiler options including
> -fsanitize=unsigned-integer-overflow, so it becomes less likely.
>
> I'll revert my other patch in the randconfig tree to see if it comes back.

The qcom/gpi.c failure still happens with this patch applied:

In file included from /git/arm-soc/drivers/dma/qcom/gpi.c:8:
In function 'field_multiplier',
inlined from 'gpi_update_reg' at
/git/arm-soc/include/linux/bitfield.h:124:17:
/git/arm-soc/include/linux/bitfield.h:119:3: error: call to
'__bad_mask' declared with attribute error: bad bitfield mask
119 | __bad_mask();
| ^~~~~~~~~~~~
In function 'field_multiplier',
inlined from 'gpi_update_reg' at
/git/arm-soc/include/linux/bitfield.h:154:1:
/git/arm-soc/include/linux/bitfield.h:119:3: error: call to
'__bad_mask' declared with attribute error: bad bitfield mask
119 | __bad_mask();
| ^~~~~~~~~~~~

See https://pastebin.com/8UH6x4A2 for the .config

Arnd