[libcxx-commits] [PATCH] D92110: [SystemZ][ZOS] Provide PATH_MAX macro for libcxx

[Hubert Tong via Phabricator via libcxx-commits]

hubert.reinterpretcast added inline comments.


================
Comment at: libcxx/src/filesystem/operations.cpp:545
+  #if defined(__MVS__)
+    char buff[1024 +1];
+  #else
----------------
ldionne wrote:
> hubert.reinterpretcast wrote:
> > Minor nit: style.
> Hardcoding buffer sizes like that is a great way to trigger a buffer overflow when assumptions are not respected. I would much rather we don't introduce this sort of code in libc++, to be honest. Is there a reason why zOS can't just provide `PATH_MAX`?
> 
> Or, even much better, why doesn't zOS satisfy `defined(_POSIX_VERSION) && _POSIX_VERSION >= 200112`? Then, we could use the safe and non-buggy version of `realpath`. The buffer-taking version is documented in various places, including `realpath`'s manpage itself, as being buggy.
> Hardcoding buffer sizes like that is a great way to trigger a buffer overflow when assumptions are not respected.

Precisely because violating the assumptions mean that buffer overflow would be introduced into compiled programs, changes that would cause the assumption to not be respected cannot realistically be made.

> I would much rather we don't introduce this sort of code in libc++, to be honest. Is there a reason why zOS can't just provide `PATH_MAX`?

The POSIX specification provides rationale for not defining PATH_MAX in <limits.h>.

> Or, even much better, why doesn't zOS satisfy `defined(_POSIX_VERSION) && _POSIX_VERSION >= 200112`? Then, we could use the safe and non-buggy version of `realpath`. The buffer-taking version is documented in various places, including `realpath`'s manpage itself, as being buggy.

Deployed systems have stability considerations that may be incompatible with such a change. The IBM documentation for released versions clearly state that passing NULL to the second parameter produces an error. Indeed the bugginess of the buffer-taking version is why the hardcoded 1024 that the implementation of `realpath` on z/OS uses is preferable to an implementation that does not implement a fixed upper bound on the buffer-taking form of `realpath`.



Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D92110/new/

https://reviews.llvm.org/D92110