Your CI caught this: the check passed on SQLite and never fired on the
stack where the database container runs on a different timezone from the
app, so a stale token stayed valid. The comparison was between a token's
`iat`, which is UTC seconds from Node, and `auth.modified_on`, which the
driver hands back interpreted in the app's timezone. With the app on
Australia/Brisbane and the database on UTC, that column comes back ten
hours in the past and the token always looks newer than the change.
Record the moment in `auth.meta.password_changed_at` instead, written by
`setPassword` with the same `Date.now()` clock that mints `iat`. Same
unit on both sides, one clock, and no timestamp parsing: the Date and
local-string branch is gone, and so is the whole-second flooring that
Postgres microseconds made necessary.
Rows written before this have no marker and revoke nothing until their
next password change, which is the safe direction to be wrong in.
Tokens are stateless JWTs, so changing a password left every session that
the old one had opened working until its own expiry, up to a day later.
That is the case the password change is meant to close: an administrator
resetting a compromised account did not evict whoever was already in it.
The auth row already records when the password last changed, so no
migration is needed: `Access.init()` reads it alongside the user it
already loads and refuses a token whose `iat` is older. Both sides are
compared as whole seconds, which is all `iat` carries, so a token minted
in the same second as the change is kept. Postgres stores that column to
the microsecond, which is why the comparison is not done in milliseconds.
It is reported as 401 rather than the usual 403 because that is what the
frontend clears the session on, so the browser holding the dead token
lands on the login page instead of a page full of errors, and `can()`
lets that one error through unwrapped for the same reason.
Only the password does this. A user row changing (a rename, an avatar,
permissions) does not, and a user with no password auth row, which is
what a login through an external provider looks like, is not affected.
The catch block in `access.can()` constructed `errs.PermissionError`
without `new`. The error constructors in `backend/lib/error.js` are plain
constructor functions that assign to `this` and return nothing, so calling
one without `new` evaluates to `undefined`.
The statement therefore did `throw undefined`, the express error handler
in `backend/app.js` received undefined (and could not read `.status` or
`.public` off it), and the request fell through to the catch-all 404
handler in `backend/routes/main.js`.
Net effect: every authorization failure raised by `access.can(...)` was
reported to clients as `404 Not Found` instead of `403 Forbidden`.
Line 45 of the same file already used `new` correctly, which shows this
was an oversight rather than deliberate resource-existence hiding. A grep
over `backend/` confirms this was the only error constructor invoked
without `new`.
* Fix wrapping when too many hosts are shown (#207)
* Update npm packages, fixes CVE-2019-10757
* Revert some breaking packages
* Major overhaul
- Docker buildx support in CI
- Cypress API Testing in CI
- Restructured folder layout (insert clean face meme)
- Added Swagger documentation and validate API against that (to be completed)
- Use common base image for all supported archs, which includes updated nginx with ipv6 support
- Updated certbot and changes required for it
- Large amount of Hosts names will wrap in UI
- Updated packages for frontend
- Version bump 2.1.0
* Updated documentation
* Fix JWT expire time going crazy. Now set to 1day
* Backend JS formatting rules
* Remove v1 importer, I doubt anyone is using v1 anymore
* Added backend formatting rules and enforce them
in Jenkins builds
* Fix CI, doesn't need a tty
* Thanks bcrypt. Why can't you just be normal.
* Cleanup after syntax check
Co-authored-by: Marcelo Castagna <margaale@users.noreply.github.com>
José M. Requena Plens
Timothé Stoifl
Jamie Curnow
Timothé Stoifl
