Failing kunit test on ARCH=arm and LLVM=1
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
Hi Christian,
Quoted 15 lines · Christian Schrefl
Greetings,
when trying to port Rust to ARM I noticed that the
DEFINE_FLEX_testkunit test inlib/overflow_kunit.c:1188fails 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
Quoted 2 lines
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):
lib/overflow_kunit.c +1 -1
@@ -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):
lib/overflow_kunit.c +5 -1
@@ -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
Quoted 99 lines · Nathan Chancellor
Hi Christian,
Quoted 15 lines · Christian Schrefl
Greetings,
when trying to port Rust to ARM I noticed that the
DEFINE_FLEX_testkunit test inlib/overflow_kunit.c:1188fails 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.cfgQuoted 2 lines
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_64as 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_byannotation." but commitd8e45f2929b9("overflow: ChangeDEFINE_FLEXto take__counted_bymember") defined 'struct foo' with__counted_byon it.__counted_byinforms__builtin_dynamic_object_size()about the size of the flexible array. WithDEFINE_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 testingDEFINE_RAW_FLEX()was intentional):lib/overflow_kunit.c +1
-1@@ -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_bysupport (which is admittedly ugly):lib/overflow_kunit.c +5
-1@@ -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 )); }
Quoted 1 line
Kees, am I missing anything here?
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.
Signature · 1 line
Kees Cook