While C allows function pointer casts, it is UB to call a function with
a different type than its actual type signature. That is, even though
`void f(int *)` and `void g(void *)` have the same ABI, it is UB to
cast `f` to a `void(*)(void *)` and then call it through that pointer.
Clang CFI will try to enforce this rule.
The recent CL to call X509_print in tests revealed that all the i2? and
?2i callbacks in X509V3_EXT_METHODs were implemented with functions of
the wrong type, out of some combination of missing consts and void*
turned into T*.
This CL fixes this. Where the function wasn't exported, or had no
callers, I just fixed the function itself. Where it had extension
callers, I added a wrapper function with a void* type.
I'm not positive whether the wrappers are the right call. On the one
hand, keeping the exported functions as-is is more type-safe and more
OpenSSL-compatible. However, most (but not all) uses of these are in
other code defining X509V3_EXT_METHODs themselves, so the void*
signature is more correct for them too. And the functions have a type
signature meant for X509V3_EXT_METHOD, complete with method pointer.
I've gone with leaving the exported ones as-is for now. Probably the
right answer anyway is to migrate the external callers, of either type
signature.
Change-Id: Ib8f2995cbd890221eaa9ac864a7e553cb6711901
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/52686
Commit-Queue: Bob Beck <bbe@google.com>
Reviewed-by: Bob Beck <bbe@google.com>
At one point in the SSLeay days, all the ASN1_STRING typedefs were
separate structs (but only in debug builds) and the M_ASN1_* macros
included type casts to handle this.
This is long gone, but we still have the M_ASN1_* macros. Remove the
casts and switch code within the library to call the macros. Some
subtleties:
- The "MSTRING" types (what OpenSSL calls its built-in CHOICEs
containing some set of string types) are weird because the M_FOO_new()
macro and the tasn_new.c FOO_new() function behave differently. I've
split those into a separate CL.
- ASN1_STRING_type, etc., call into the macro, which accesses the field
directly. This CL inverts the dependency.
- ASN1_INTEGER_new and ASN1_INTEGER_free, etc., are generated via
IMPLEMENT_ASN1_STRING_FUNCTIONS in tasn_typ.c. I've pointed
M_ASN1_INTEGER_new and M_ASN1_INTEGER_free to these fields. (The free
function is a no-op, but consistent.)
- The other macros like M_ASN1_BIT_STRING_dup largely do not have
corresponding functions. I've aligned with OpenSSL in just using the
generic ASN1_STRING_dup function. But some others, like
M_ASN1_OCTET_STRING_dup have a corresponding ASN1_OCTET_STRING_dup
function. OpenSSL retained these, so I have too.
Update-Note: Some external code uses the M_ASN1_* macros. This should
remain compatible, but some type errors may have gotten through
unnoticed. This CL restores type-checking.
Change-Id: I8656abc7d0f179192e05a852c97483c021ad9b20
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/44045
Reviewed-by: Adam Langley <agl@google.com>
David Benjamin
Adam Langley