Fixes#8093.
This PR modifies and audits all uses of `subprocess` and `Popen` outside of tests, `certbot-ci/`, `certbot-compatibility-test/`, `letsencrypt-auto-source/`, `tools/`, and `windows-installer/`. Calls to outside programs have their `env` modified to remove the `SNAP` components of paths, if they exist. This includes any calls made from hooks, calls to `apachectl` and `nginx`, and to `openssl` from `ocsp.py`.
For testing manually, rsync flags will look something like:
```
rsync -avzhe ssh root@focal.domain:/home/certbot/certbot/certbot_*_amd64.snap .
rsync -avzhe ssh certbot_*_amd64.snap root@centos7.domain:/root/certbot/
```
With these modifications, `certbot plugins --prepare` now passes on Centos 7.
If I'm wrong and we package the `openssl` binary, the modifications should be removed from `ocsp.py`, and `env` should be passed into `run_script` rather than set internally in its calls from nginx and apache.
One caveat with this approach is the disconnect between why it's a problem (packaging) and where it's solved (internal to Certbot). I considered a wrapping approach, but we'd still have to audit specific calls. I think the best way to address this is robust testing; specifically, running the snap on other systems.
For hooks, all calls will remove the snap paths if they exist. This is probably fine, because even if the hook intends to call back into certbot, it can do that, it'll just create a new snap.
I'm not sure if we need these modifications for the Mac OS X/ Darwin calls, but they can't hurt.
* Add method to plugins util to get env without snap paths
* Use modified environment in Nginx plugin
* Pass through env to certbot.util.run_script
* Use modified environment in Apache plugin
* move env_no_snap_for_external_calls to certbot.util
* Set env internally to run_script, since we use that only to call out
* Add env to mac subprocess calls in certbot.util
* Add env to openssl call in ocsp.py
* Add env for hooks calls in certbot.compat.misc.
* Pass env into execute_command to avoid circular dependency
* Update hook test to assert called with env
* Fix mypy type hint to account for new param
* Change signature to include Optional
* go back to using CERTBOT_PLUGIN_PATH
* no need to modify PYTHONPATH in env
* robustly detect when we're in a snap
* Improve env util fxn docstring
* Update changelog
* Add unit tests for env_no_snap_for_external_calls
* Import compat.os
According to `distutils/version.py`, StrictVersion is pretty strict in
what version numbers to accept:
> A version number consists of two or three dot-separated numeric
> components, with an optional "pre-release" tag on the end. The
> pre-release tag consists of the letter 'a' or 'b' followed by a number.
This assumption already fails for some pretty basic python libraries
itself, like setuptools, also available in `46.1.3.post20200610`, a
completely valid version number according to
https://www.python.org/dev/peps/pep-0440/#post-releases.
There doesn't seem to be a particular reason on why StrictVersion has
been used here, so let's use LooseVersion, to be compatible with these
versions.
Co-authored-by: Adrien Ferrand <adferrand@users.noreply.github.com>
This PR gets its root from an observation I did on current version of Certbot (1.3.0): the `renewal-hooks` directory in Certbot configuration directory is created on Windows with write permissions to everybody.
I thought it was a critical bug since this directory contains hooks that are executed by Certbot, and you certainly do not want this folder to be open to any malicious hook that could be inserted by everyone, then executed with administrator privileges by Certbot.
Turns out for this specific problem that the bug is not critical for the hooks, because the scripts are expected to be in subdirectories of `renewal-hooks` (namely `pre`, `post` and `deploy`), and these subdirectories have proper permissions because we set them explicitly when Certbot is starting.
Still, there is a divergence here between Linux and Windows: on Linux all Certbot directories without explicit permissions have at maximum `0o755` permissions by default, while on Windows it is a `0o777` equivalent. It is not an immediate security risk, but it is definitly error-prone, not expected, and so a potential breach in the future if we forget about it.
Root cause is that umask is not existing in Windows. Indeed under Linux the umask defines the default permissions when you create a file or a directory. Python takes that into account, with an API for `os.open` and `os.mkdir` that expose a `mode` parameter with default value of `0o777`. In practice it is never `0o777` (either you the the `mode` explictly or left the default one) because the effective mode is masked by the current umask value in the system: on Linux it is `0o022`, so files/directories have a maximum mode of `0o755` if you did not set the umask explicitly, and it is what it is observed for Certbot.
However on Windows, the `mode` value passed (and got from default) to the `open` and `mkdir` of `certbot.compat.filesystem` module is taken verbatim, since umask does not exit, and then is used to calculate the DACL of the newly created file/directory. So if the mode is not set explicitly, we end up with files and directories with `0o777` permissions.
This PR fixes this problem by implementing a umask behavior in the `certbot.compat.filesystem` module, that will be applied to any file or directory created by Certbot since we forbid to use the `os` module directly.
The implementation is quite straight-forward. For Linux the behavior is not changed. On Windows a `mask` parameter is added to the function that calculates the DACL, to be invoked appropriately when file or directory are created. The actual value of the mask is taken from an internal class of the `filesystem` module: its default value is `0o755` to match default umasks on Linux, and can be changed with the new method `umask` that have the same behavior than the original `os.umask`. Of course `os.umask` becomes a forbidden function and `filesystem.umask` must be used instead.
Existing code that is impacted have been updated, and new unit tests are created for this new function.
* Implement umask for Windows
* Set umask at the beginning of tests
* Fix lint, update local oldest requirements
* Update certbot-apache/setup.py
Co-authored-by: Brad Warren <bmw@users.noreply.github.com>
* Improve tests
* Adapt filesystem.makedirs for Windows
* Fix
* Update certbot-apache/setup.py
Co-authored-by: Brad Warren <bmw@users.noreply.github.com>
* Changelog entries
* Fix lint
* Update certbot/CHANGELOG.md
Co-authored-by: Brad Warren <bmw@users.noreply.github.com>
Co-authored-by: Brad Warren <bmw@users.noreply.github.com>
In #7771, the Apache configurator gained the ability to identify what
version of OpenSSL Apache's ssl_module is linked against. However, the
detection was only functional if the module was built as a DSO (which is
almost always the case).
This commit covers the case where the ssl_module is statically linked
within the Apache binary. It requires the user to specify the path to
the binary (with --apache-bin) and emits a warning if static linking is
detected but no path has been provided.
* Error out in apache installer when mod_ssl is not available
* Update to MisconfigurationError and add/fix tests
* Remove error cases we no longer hit and associated test
* mock out function to have consistent error across machines
* improve changelog message
* only check key in modules list, not value
Part of #7886.
This PR conditionally installs mock in `apache/setup.py` based on setuptools version and python version, when possible. It then updates `apache` tests to use `unittest.mock` when `mock` isn't available.
* Conditionally install mock in apache
* error out on newer python and older setuptools
* error when trying to build wheels with old setuptools
* use unittest.mock when third-party mock isn't available in apache, with no cover and type ignore
* Fix dangerous default argument
* Remove unused imports
* Remove unnecessary comprehension
* Use literal syntax to create data structure
* Use literal syntax instead of function calls to create data structure
Co-authored-by: deepsource-autofix[bot] <62050782+deepsource-autofix[bot]@users.noreply.github.com>
* Load apacheconfig dependency, gate behind flag
* Bump apacheconfig dependency to latest version and install dev version of apache for coverage tests
* Move augeasnode_test tests to more generic parsernode_test
* Revert "Move augeasnode_test tests to more generic parsernode_test"
This reverts commit 6bb986ef78.
* Mock AugeasNode into DualNode's place, and run augeasnode tests exclusively on AugeasNode
* Don't calculate coverage for skeleton functions
* clean up helper function in augeasnode_test
Fixes#7350.
This PR changes the parsed modules from a `set` to a `dict`, with the filepath argument as the value. Accordingly, after calling `enable_mod` to enable `ssl_module`, modules now need to be re-parsed, so call `reset_modules`.
* Add mechanism for selecting apache config file, based on work done in #7191.
* Check OpenSSL version
* Remove os imports
* debian override still needs os
* Reformat remaining apache tests with modules dict syntax
* Clean up more apache tests
* Switch from property to method for openssl and add tests for coverage.
* Sometimes the dict location will be None in which case we should in fact return None
* warn thoroughly and consistently in openssl_version function
* update tests for new warnings
* read file as bytes, and factor out the open for testing
* normalize ssl_module_location path to account for being relative to server root
* Use byte literals in a python 2 and 3 compatible way
* string does need to be a literal
* patch builtins open
* add debug, remove space
* Add test to check if OpenSSL detection is working on different systems
* fix relative test location for cwd
* put </IfModule> on its own line in test case
* Revert test file to status in master.
* Call augeas load before reparsing modules to pick up the changes
* fix grep, tail, and mod_ssl location on centos
* strip the trailing whitespace from fedora
* just use LooseVersion in test
* call apache2ctl on debian systems
* Use sudo for apache2ctl command
* add check to make sure we're getting a version
* Add boolean so we don't warn on debian/ubuntu before trying to enable mod_ssl
* Reduce warnings while testing by setting mock _openssl_version.
* Make sure we're not throwing away any unwritten changes to the config
* test last warning case for coverage
* text changes for clarity
This PR builds on #7657 and cleans up additional unnecessary pylint comments and some stray comments referring to pylint: disable comments that have been deleted that I didn't notice in my review of that PR.
* Remove stray pylint link.
* Cleanup more pylint comments
* Cleanup magic_typing imports
* Remove unneeded pylint: enable comments
As pylint is evolving, it improves its accuracy, and several pylint error suppression (`# pylint: disable=ERROR) added in certbot codebase months or years ago are not needed anymore to make it happy.
There is a (disabled by default) pylint error to detect the useless suppressions (pylint-ception: `useless-suppression`). It is not working perfectly (it has also false-positives ...) but it is a good start to clean the codebase.
This PR removes several of these useless suppressions as detected by the current pylint version we use.
* Remove useless suppress
* Remove useless lines
Part of #7204.
Makes the smaller changes described at https://github.com/certbot/certbot/issues/7204#issuecomment-571838185 to disable many old ciphersuites and TLS versions < 1.2. Does not add checks for OpenSSL version or modify session tickets.
Since Apache uses TLS protocol blacklisting instead of whitelisting (as in NGINX), we additionally may not need to determine if the server supports TLS1.3 and turn it on or off based on Apache version.
* Update SSL versions and ciphersuites based on Mozilla intermediate recommendations for apache
* Update constants with hashes of new config files
* Update changelog
Part of #7550
This PR makes appropriate corrections to run pylint on Python 3.
Why not keeping the dependencies unchanged and just run pylint on Python 3?
Because the old version of pylint breaks horribly on Python 3 because of unsupported version of astroid.
Why updating pylint + astroid to the latest version ?
Because this version only fixes some internal errors occuring during the lint of Certbot code, and is also ready to run gracefully on Python 3.8.
Why upgrading mypy ?
Because the old version does not support the new version of astroid required to run pylint correctly.
Why not upgrading mypy to its latest version ?
Because this latest version includes a new typshed version, that adds a lot of new type definitions, and brings dozens of new errors on the Certbot codebase. I would like to fix that in a future PR.
That said so, the work has been to find the correct set of new dependency versions, then configure pylint for sane configuration errors in our situation, disable irrelevant lintings errors, then fixing (or ignoring for good reason) the remaining mypy errors.
I also made PyLint and MyPy checks run correctly on Windows.
* Start configuration
* Reconfigure travis
* Suspend a check specific to python 3. Start fixing code.
* Repair call_args
* Fix return + elif lints
* Reconfigure development to run mainly on python3
* Remove incompatible Python 3.4 jobs
* Suspend pylint in some assertions
* Remove pylint in dev
* Take first mypy that supports typed-ast>=1.4.0 to limit the migration path
* Various return + else lint errors
* Find a set of deps that is working with current mypy version
* Update local oldest requirements
* Remove all current pylint errors
* Rebuild letsencrypt-auto
* Update mypy to fix pylint with new astroid version, and fix mypy issues
* Explain type: ignore
* Reconfigure tox, fix none path
* Simplify pinning
* Remove useless directive
* Remove debugging code
* Remove continue
* Update requirements
* Disable unsubscriptable-object check
* Disable one check, enabling two more
* Plug certbot dev version for oldest requirements
* Remove useless disable directives
* Remove useless no-member disable
* Remove no-else-* checks. Use elif in symetric branches.
* Add back assertion
* Add new line
* Remove unused pylint disable
* Remove other pylint disable
