Issue #20693 has been updated by kjtsanaktsidis (KJ Tsanaktsidis). I have a PR for doing this: https://github.com/ruby/ruby/pull/11403 But actually it breaks Bundler's tests, because they stub `File.writable?` in a test. I opened a fix for Bundler which needs to land first: https://github.com/rubygems/rubygems/pull/7961 ---------------------------------------- Bug #20693: Dir.tmpdir should perform a real access check before warning about writability https://bugs.ruby-lang.org/issues/20693#change-109505 * Author: kjtsanaktsidis (KJ Tsanaktsidis) * Status: Open * Backport: 3.1: UNKNOWN, 3.2: UNKNOWN, 3.3: UNKNOWN ---------------------------------------- The code in `Dir.tmpdir` attempts to warn the user if their temp directory is deficient for some reason: ```ruby case when !stat.directory? warn "#{name} is not a directory: #{dir}" when !stat.writable? warn "#{name} is not writable: #{dir}" when stat.world_writable? && !stat.sticky? warn "#{name} is world-writable: #{dir}" else break dir end ``` This check for writability is looking at the user/group/world access bits on the stat output, and determining if the user running Ruby is allowed to write to the temp directory based on that. However, modern operating systems contain other mechanisms apart from the user/group/world bits which can grant access to a directory that would otherwise be denied, or vice versa. Things like: * Posix ACL's * Linux's capabilities like CAP_DAC_OVERRIDE * Linux Security Modules like SELinux or AppArmor * Syscall filters like Linux's seccomp * Granular capability systems like FreeBSD's Capsicum * OpenBSD's pledge and unveil * Windows too has a rich ACL system for controlling filesystem access To address this, we should call `File.writable?` instead of `stat.writable?`, which asks the system whether the file is writable using the `euidaccess()` function if available. On Linux/glibc, at least, this will issue an `access(2)` syscall, and the Kernel can take all of the above into account. n.b. if Ruby is running as suid, then glibc currently will NOT ask the kernel to perform the access check in `euidaccess()`, and instead does a similar thing to what `Stat#writable?` does (https://github.com/bminor/glibc/blob/7f04bb4e49413bd57ac3215f3480b09ae713196...). This is because of the relatively new `faccessat2(2)` syscall is required to do this properly, and there is some ecosystem issues with leveraging this by default (e.g. https://bugzilla.redhat.com/show_bug.cgi?id=1900021). Since running Ruby as suid is probably a very bad idea anyway, and the glibc implementation isn't any worse than the `Stat#writable?` one, this seems OK though. -- https://bugs.ruby-lang.org/