| FazBrowse GitHub Viewer | Trending | | Home |
| Tools: [Download Repo ZIP] [Original HTTPS Page] |
path.win32.normalize() checked names like CONx or NULs against the Windows reserved device list (CON, NUL, PRN, LPT1, ...) by slicing off the last character whenever the path had no colon, so any file name that happened to be a reserved name plus one extra character got wrongly treated as a device and prefixed with `.\`. A previous attempt at this fix (b0a4f16, reverted in c4429c8) guarded the check with `colonIndex !== -1`, but that also suppressed the colon-less case Windows itself treats as reserved: a name followed immediately by a single dot, e.g. NUL. and COM9. (per Microsoft's own docs, "avoid these names followed immediately by an extension; NUL.txt and NUL.tar.gz are both equivalent to NUL"), which is why it broke Windows CI. This version keeps that dot case working while still rejecting an arbitrary trailing character. Signed-off-by: Alexander Kireev <ak.chatman.media@gmail.com>
|
Review requested:
|
Sorry, something went wrong.
| Back | FazBrowse Home | New Git URL |
path.win32.normalize() treats a bare reserved device name (CON, NUL, PRN, LPT1, ...) as a device path when it isn't followed by a colon, by slicing off path's last character and comparing that against the reserved-name list. So 'CONx'.slice(0, -1) is 'CON', which matches, and the file name gets wrongly rewritten to .\CONx. Same for NULs, LPT1x, etc. — any ordinary name that happens to be reservedName + 1 char gets this treatment even though it has nothing to do with the device.
#64159 fixed exactly this by requiring a colon before running the check at all, but that also killed the colon-less case Windows genuinely treats as reserved: a trailing single dot. Per Microsoft's own docs, "avoid these names followed immediately by an extension; NUL.txt and NUL.tar.gz are both equivalent to NUL," and there's already a test (COM9. → .\COM9.) baking that in. That's presumably what broke Windows CI and got it reverted in c4429c8 — I couldn't get at the Jenkins logs to confirm, but I hit the exact regression myself while re-testing the fix, so I'm fairly confident that's the mechanism.
This version narrows the no-colon check instead of dropping it: it only treats the name as reserved when the character being sliced off is specifically a ., so NUL./COM9. still match but CONx/NULs no longer do. Added both back to test-path-normalize.js, checked the CVE-2024-36139 tests and the full test-path-win32-normalize-device-names.js suite still pass (ran the latter directly since it's Windows-gated and I'm on macOS).