KUnit's deferred action API accepts a void(*)(void *) function pointer which is called when the test is exited. However, we very frequently want to use existing functions which accept a single pointer, but which may not be of type void*. While this is probably dodgy enough to be on the wrong side of the C standard, it's been often used for similar callbacks, and gcc's -Wcast-function-type seems to ignore cases where the only difference is the type of the argument, assuming it's compatible (i.e., they're both pointers to data).
However, clang 16 has introduced -Wcast-function-type-strict, which no longer permits any deviation in function pointer type. This seems to be because it'd break CFI, which validates the type of function calls.
This rather ruins our attempts to cast functions to defer them, and leaves us with a few options: 1. Stick our fingers in our ears an ignore the warning. (It's worked so far, but probably isn't the right thing to do.) 2. Find some horrible way of casting which fools the compiler into letting us do the cast. (It'd still break CFI, though.) 3. Disable the warning, and CFI for this function. This isn't optimal, but may make sense for test-only code. However, I think we'd have to do this for every function called, not just the caller, so maybe it's not practical. 4. Manually write wrappers around any such functions. This is ugly (do we really want two copies of each function, one of which has no type info and just forwards to the other). It could get repetitive. 5. Generate these wrappers with a macro. That's what this patch does.
I'm broadly okay with any of the options above, though whatever we go with will no doubt require some bikeshedding of details (should these wrappers be public, do we dedupe them, etc).
Thoughts?
Link: https://github.com/ClangBuiltLinux/linux/issues/1750 Signed-off-by: David Gow davidgow@google.com ---
I finally got around to setting up clang 16 to look into these warnings:
lib/kunit/test.c:764:38: warning: cast from 'void (*)(const void *)' to 'kunit_action_t *' (aka 'void (*)(void *)') converts to incompatible function type [-Wcast-function-type-strict] if (kunit_add_action_or_reset(test, (kunit_action_t *)kfree, data) != 0) ^~~~~~~~~~~~~~~~~~~~~~~ lib/kunit/test.c:776:29: warning: cast from 'void (*)(const void *)' to 'kunit_action_t *' (aka 'void (*)(void *)') converts to incompatible function type [-Wcast-function-type-strict] kunit_release_action(test, (kunit_action_t *)kfree, (void *)ptr); ^~~~~~~~~~~~~~~~~~~~~~~ 2 warnings generated.
It's probably something which needs fixing with wrappers, not with the "just keep casting things until the compiler forgets" strategy.
There are few enough uses of kunit_add_action() that now's the time to change things if we want to fix these warnings (and, I guess, work with CFI). This patch uses an ugly macro, but we're definitely still at the point where doing this by hand might make more sense.
Don't take this exact patch too seriously: it's mostly a discussion starter so we can decide on a plan.
Cheers, -- David
--- include/kunit/resource.h | 9 +++++++++ lib/kunit/executor_test.c | 7 +++---- lib/kunit/test.c | 6 ++++-- 3 files changed, 16 insertions(+), 6 deletions(-)
diff --git a/include/kunit/resource.h b/include/kunit/resource.h index c7383e90f5c9..4110e13970dc 100644 --- a/include/kunit/resource.h +++ b/include/kunit/resource.h @@ -390,6 +390,15 @@ void kunit_remove_resource(struct kunit *test, struct kunit_resource *res); /* A 'deferred action' function to be used with kunit_add_action. */ typedef void (kunit_action_t)(void *);
+/* We can't cast function pointers to kunit_action_t if CFI is enabled. */ +#define KUNIT_DEFINE_ACTION_WRAPPER(wrapper, orig, arg_type) \ + static void wrapper(void *in) \ + { \ + arg_type arg = (arg_type)in; \ + orig(arg); \ + } + + /** * kunit_add_action() - Call a function when the test ends. * @test: Test case to associate the action with. diff --git a/lib/kunit/executor_test.c b/lib/kunit/executor_test.c index b4f6f96b2844..14ac64f4f71b 100644 --- a/lib/kunit/executor_test.c +++ b/lib/kunit/executor_test.c @@ -256,9 +256,8 @@ kunit_test_suites(&executor_test_suite);
/* Test helpers */
-/* Use the resource API to register a call to kfree(to_free). - * Since we never actually use the resource, it's safe to use on const data. - */ +KUNIT_DEFINE_ACTION_WRAPPER(kfree_action_wrapper, kfree, const void *) +/* Use the resource API to register a call to kfree(to_free). */ static void kfree_at_end(struct kunit *test, const void *to_free) { /* kfree() handles NULL already, but avoid allocating a no-op cleanup. */ @@ -266,7 +265,7 @@ static void kfree_at_end(struct kunit *test, const void *to_free) return;
kunit_add_action(test, - (kunit_action_t *)kfree, + kfree_action_wrapper, (void *)to_free); }
diff --git a/lib/kunit/test.c b/lib/kunit/test.c index 421f13981412..41b7d9a090fb 100644 --- a/lib/kunit/test.c +++ b/lib/kunit/test.c @@ -804,6 +804,8 @@ static struct notifier_block kunit_mod_nb = { }; #endif
+KUNIT_DEFINE_ACTION_WRAPPER(kfree_action_wrapper, kfree, const void *) + void *kunit_kmalloc_array(struct kunit *test, size_t n, size_t size, gfp_t gfp) { void *data; @@ -813,7 +815,7 @@ void *kunit_kmalloc_array(struct kunit *test, size_t n, size_t size, gfp_t gfp) if (!data) return NULL;
- if (kunit_add_action_or_reset(test, (kunit_action_t *)kfree, data) != 0) + if (kunit_add_action_or_reset(test, kfree_action_wrapper, data) != 0) return NULL;
return data; @@ -825,7 +827,7 @@ void kunit_kfree(struct kunit *test, const void *ptr) if (!ptr) return;
- kunit_release_action(test, (kunit_action_t *)kfree, (void *)ptr); + kunit_release_action(test, kfree_action_wrapper, (void *)ptr); } EXPORT_SYMBOL_GPL(kunit_kfree);
Hi David,
On Thu, Sep 14, 2023 at 10:01 PM David Gow davidgow@google.com wrote:
KUnit's deferred action API accepts a void(*)(void *) function pointer which is called when the test is exited. However, we very frequently want to use existing functions which accept a single pointer, but which may not be of type void*. While this is probably dodgy enough to be on the wrong side of the C standard, it's been often used for similar callbacks, and gcc's -Wcast-function-type seems to ignore cases where the only difference is the type of the argument, assuming it's compatible (i.e., they're both pointers to data).
However, clang 16 has introduced -Wcast-function-type-strict, which no longer permits any deviation in function pointer type. This seems to be because it'd break CFI, which validates the type of function calls.
This rather ruins our attempts to cast functions to defer them, and leaves us with a few options:
- Stick our fingers in our ears an ignore the warning. (It's worked so far, but probably isn't the right thing to do.)
- Find some horrible way of casting which fools the compiler into letting us do the cast. (It'd still break CFI, though.)
- Disable the warning, and CFI for this function. This isn't optimal, but may make sense for test-only code. However, I think we'd have to do this for every function called, not just the caller, so maybe it's not practical.
- Manually write wrappers around any such functions. This is ugly (do we really want two copies of each function, one of which has no type info and just forwards to the other). It could get repetitive.
- Generate these wrappers with a macro. That's what this patch does.
I'm broadly okay with any of the options above, though whatever we go with will no doubt require some bikeshedding of details (should these wrappers be public, do we dedupe them, etc).
Thoughts?
Using a macro to generate a wrapper is a reasonable approach IMO, and we've used it before in the kernel to fix type mismatches in indirectly called functions (v4l2-ioctl and cfg80211 come to mind at least).
Sami
On Fri, Sep 15, 2023 at 01:01:23PM +0800, David Gow wrote:
KUnit's deferred action API accepts a void(*)(void *) function pointer which is called when the test is exited. However, we very frequently want to use existing functions which accept a single pointer, but which may not be of type void*. While this is probably dodgy enough to be on the wrong side of the C standard, it's been often used for similar callbacks, and gcc's -Wcast-function-type seems to ignore cases where the only difference is the type of the argument, assuming it's compatible (i.e., they're both pointers to data).
However, clang 16 has introduced -Wcast-function-type-strict, which no longer permits any deviation in function pointer type. This seems to be because it'd break CFI, which validates the type of function calls.
This rather ruins our attempts to cast functions to defer them, and leaves us with a few options:
- Stick our fingers in our ears an ignore the warning. (It's worked so far, but probably isn't the right thing to do.)
- Find some horrible way of casting which fools the compiler into letting us do the cast. (It'd still break CFI, though.)
- Disable the warning, and CFI for this function. This isn't optimal, but may make sense for test-only code. However, I think we'd have to do this for every function called, not just the caller, so maybe it's not practical.
- Manually write wrappers around any such functions. This is ugly (do we really want two copies of each function, one of which has no type info and just forwards to the other). It could get repetitive.
- Generate these wrappers with a macro. That's what this patch does.
I'm broadly okay with any of the options above, though whatever we go with will no doubt require some bikeshedding of details (should these wrappers be public, do we dedupe them, etc).
Thoughts?
Looks awesome :)
We ended up using a wrapper in KMS to workaround this issue and would benefit from it too.
Maxime
linux-kselftest-mirror@lists.linaro.org