feat: caml_fatal_error on misuse of deserialize functions

[MisterDA]

Mimicking #12171, this PR gets intern.c closer to extern.c. It fixes a problem with possible misuse of deserialize functions, as these should only be called from within a caml_input_* entrypoint context. Otherwise, the field intern_src of struct caml_intern_state won't have been populated, and a null pointer dereference occurs.

cc @askvortsov1 and @gasche who authored and reviewed the original PR.

[kayceesrk]

I'm ok with the first line of the error message "intern_state not initialised". I don't know whether the rest of the message is clarifying the error more "this function can only be called from a caml_input_* entrypoint.".

When will this error occur? Is it that caml_deserialise_* was called directly without calling one of the caml_input_* functions?

[MisterDA]

When will this error occur? Is it that caml_deserialise_* was called directly without calling one of the caml_input_* functions?

Exactly. In that case, before this PR, the field intern_src pointing to the source of the marshaled data won't have been populated, and would be accessed, leading to the nullptr dereference. By splitting initialization of the structure (in init_intern_state) and then access (with get_intern_state), we can check whether it has been provided a marshaled data source, and error if not (admittedly nicer than a segfault).

[kayceesrk]

May I request that you add a testcase for this under testsuite/tests/lib-marshal?

Also, the message can be more clear. Instead of

"this function can only be called from a caml_input_* entrypoint."

How about

"It is likely that caml_deserialize_* function was called without going through caml_input_*.

[MisterDA]

May I request that you add a testcase for this under testsuite/tests/lib-marshal?

Maybe easier said than done? As the functions raise a fatal error, I can only test one of them by process. A failwith would allow to catch the exception and recover.

[kayceesrk]

Maybe easier said than done?

This is doable. See .run files in the test suite for how it is done. But it is not worth the trouble here as the bytecode and native code behaviours will diverge. Bytecode will perform an intern at

ocaml/runtime/startup_byt.c

Line 566 in 225f2f3

value global_data = caml_input_val(chan);

before user code is run. This will cause any test to not cause a fatal error, which is different from native code.