[Bug] Failing kunit test on ARCH=arm and LLVM=1

[Christian Schrefl]

Greetings,

when trying to port Rust to ARM I noticed that the DEFINE_FLEX_test
kunit test in lib/overflow_kunit.c:1188 fails when combining LLVM=1
and ARCH=arm.

I have reproduced this on v6.10-rc1 and next-20240606.

Here is the clang/llvm version I'm using:
clang version 18.1.6 (Fedora 18.1.6-3.fc40)
Target: x86_64-redhat-linux-gnu
Thread model: posix
InstalledDir: /usr/bin
Configuration file: /etc/clang/x86_64-redhat-linux-gnu-clang.cfg

I have not looked closer at the failure so I'm unsure if this is a
problem with LLVM or if the test case is to speciffic.

Let me know if I should open a issue at
https://github.com/ClangBuiltLinux/linux/issues instead.

Cheers,
Christian

[Nathan Chancellor]

Hi Christian,

On Fri, Jun 07, 2024 at 02:37:19PM +0200, Christian Schrefl wrote:
> Greetings,
>
> when trying to port Rust to ARM I noticed that the DEFINE_FLEX_test
> kunit test in lib/overflow_kunit.c:1188 fails when combining LLVM=1
> and ARCH=arm.
>
> I have reproduced this on v6.10-rc1 and next-20240606.
>
> Here is the clang/llvm version I'm using:
> clang version 18.1.6 (Fedora 18.1.6-3.fc40)
> Target: x86_64-redhat-linux-gnu
> Thread model: posix
> InstalledDir: /usr/bin
> Configuration file: /etc/clang/x86_64-redhat-linux-gnu-clang.cfg
>
> I have not looked closer at the failure so I'm unsure if this is a
> problem with LLVM or if the test case is to speciffic.

Thanks a lot for the report! I can reproduce this with tip of tree LLVM
as well.

$ echo 'CONFIG_KUNIT=y
CONFIG_OVERFLOW_KUNIT_TEST=y
CONFIG_RUNTIME_KERNEL_TESTING_MENU=y' >kernel/configs/repro.config

$ make -skj"$(nproc)" ARCH=arm LLVM=1 {def,repro.}config zImage

$ boot-qemu.py -a arm -k .
...
[ 0.000000] Linux version 6.10.0-rc2-00235-g8a92980606e3 (nathan@thelio-3990X) (ClangBuiltLinux clang version 19.0.0git (https://github.com/llvm/llvm-project e635520be888335dd59874038d33e60cca3a7143), ClangBuiltLinux LLD 19.0.0) #1 SMP Fri Jun 7 06:12:02 MST 2024
...
[ 1.832472] # castable_to_type_test: 75 castable_to_type() tests finished
[ 1.833483] ok 21 castable_to_type_test
[ 1.834122] # DEFINE_FLEX_test: EXPECTATION FAILED at lib/overflow_kunit.c:1188
[ 1.834122] Expected __builtin_dynamic_object_size(two, 0) == sizeof(struct foo) + sizeof(s16) + sizeof(s16), but
[ 1.834122] __builtin_dynamic_object_size(two, 0) == 8 (0x8)
[ 1.834122] sizeof(struct foo) + sizeof(s16) + sizeof(s16) == 12 (0xc)
[ 1.834746] not ok 22 DEFINE_FLEX_test
...

I don't see the same failure with GCC 13.2.0. This test fails when
building for arm64 and x86_64 as well, so it does not appear to be
architecture specific.

I think I see what is going on here. Looking at the documentation for
DEFINE_RAW_FLEX(), it states "Define an on-stack instance of structure
with a trailing flexible array member, when it does not have a
__counted_by annotation." but commit d8e45f2929b9 ("overflow: Change
DEFINE_FLEX to take __counted_by member") defined 'struct foo' with
__counted_by on it. __counted_by informs __builtin_dynamic_object_size()
about the size of the flexible array. With DEFINE_FLEX_RAW, the counter
is zero, so the size of array in 'struct foo' is zero, meaning this test
is incorrect when built with a compiler that supports __counted_by,
which is just Clang 18+ right now (it should land in GCC 15 if I
understand correctly).

I see two potential solutions that work for me.

One would be to stop using DEFINE_RAW_FLEX() and match the other uses
(but I assume testing DEFINE_RAW_FLEX() was intentional):

diff --git a/lib/overflow_kunit.c b/lib/overflow_kunit.c
index 4ef31b0bb74d..883670adf0cc 100644
--- a/lib/overflow_kunit.c
+++ b/lib/overflow_kunit.c
@@ -1180,7 +1180,7 @@ struct foo {

static void DEFINE_FLEX_test(struct kunit *test)
{
- DEFINE_RAW_FLEX(struct foo, two, array, 2);
+ DEFINE_FLEX(struct foo, two, array, counter, 2);
DEFINE_FLEX(struct foo, eight, array, counter, 8);
DEFINE_FLEX(struct foo, empty, array, counter, 0);


The other would be making the size of the array conditional on not
having __counted_by support (which is admittedly ugly):

diff --git a/lib/overflow_kunit.c b/lib/overflow_kunit.c
index 4ef31b0bb74d..7eed0890e25f 100644
--- a/lib/overflow_kunit.c
+++ b/lib/overflow_kunit.c
@@ -1185,7 +1185,11 @@ static void DEFINE_FLEX_test(struct kunit *test)
DEFINE_FLEX(struct foo, empty, array, counter, 0);

KUNIT_EXPECT_EQ(test, __struct_size(two),
- sizeof(struct foo) + sizeof(s16) + sizeof(s16));
+ sizeof(struct foo)
+#if !__has_attribute(__counted_by__)
+ + sizeof(s16) + sizeof(s16)
+#endif
+ );
KUNIT_EXPECT_EQ(test, __struct_size(eight), 24);
KUNIT_EXPECT_EQ(test, __struct_size(empty), sizeof(struct foo));
}

Kees, am I missing anything here?

Cheers,
Nathan

[Kees Cook]

Thanks for analyzing this! I've sent a patch for this now. It's similar
to what you've suggested here, but I wanted to break out the
non-counted_by usage as well, which is how DEFINE_RAW_FLEX() is supposed
to be used, but it's good to capture the expected behavior of RAW with
counted_by too.

--
Kees Cook