Fix assorted issues with legacy systems #644

Some systems (like CentOS 7) still run their C compilers in some C89 compatibility mode, so any C99 code will fail to build there. Since we do not require C99, declare variables at the start of blocks. Note that we do use some other features like C++ comments so we cannot enable strict C89 compatibility mode. CentOS 7 uses "-std=gnu90" which includes various GNU extensions such as line comments, but not variable declarations in arbitrary places.

It turns out that not all Emscripten toolchains support "--no-undefined" flag and complain about that by breaking builds, in particular on our Buildbot infrastructure. We use "--no-undefined" to get warnings about missing dependencies which are usually caused by OpenSSL header files that do not match the library binaries actually linked. We use BoringSSL with Emscripten by default so we don't need those warnings that much here.

struct secure_session_user_callbacks_type and others include architecture-specific types such as pointers and size_t values. Commit 85af22d (Drop "bindgen" from libthemis-sys dependencies, 2020-04-25) has replaced on-the-fly invocation of Bindgen with using pregenerated files. Unfortunately, we cannot use layout tests in these files since they end up architecture-specific. On 64-bit machines -- such as the one used to generate the files -- secure_session_user_callbacks_type uses 40 bytes of storage while on 32-bit machiens is requires only 20 bytes. We cannot include a test that simply compares the size with a constant "40". I don't want to complicate things too much just because of those tests so let's simply drop them if *they* are the only platform-dependent part here.

This test triggers a panic in "Vec::reserve" call in Token Protect implementation on 32-bit machines. Corrupted input data is not detected right away by Themis and we try to allocate more than 2 GB of memory. It fails on 32-bit systems and Rust panics. There is a nightly-only experimental API -- Vec::try_reserve -- but we need to support stable Rust. We can't do much about this issue other than assume a safe allocation interval which is only an assumption. So for now let's panic, and disable this test on 32-bit machines so that the test suite is still useful.

Similar to Rust, GoThemis panics on allocation when trying to allocate slice of negative size (a side effect of casting big C.size_t to int on 32-bit systems). This time this is a genuine overflow that we can and should avoid, but right now we don't have time to audit and update all C.size_t casts. Leave this issue as is and skip the faulty test for now.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix assorted issues with legacy systems #644

Fix assorted issues with legacy systems #644

Commits on May 25, 2020