I made `FileSystemWriteAccess` be a subclass of `FileSystemAccess` (like in [JS](64001cc02c/javascript/ql/src/semmle/javascript/Concepts.qll (L68-L74))), but then I started wondering about how I could give a good result for `getAPathArgument`, and what would a good result even be? The argument to the `open` call, or the object that the `write` method is called on? I can't see how doing either of these enables us to do anything useful...
So I looked closer at how JS uses `FileSystemWriteAccess`:
1. as sink for zip-slip: 7c51dff0f7/javascript/ql/src/semmle/javascript/security/dataflow/ZipSlipCustomizations.qll (L121)
2. as sink for downloading unsafe files (identified through their extension) through non-secure connections: 89ef6ea4eb/javascript/ql/src/semmle/javascript/security/dataflow/InsecureDownloadCustomizations.qll (L134-L150)
3. as sink for writing untrusted data to a local file 93b1e59d62/javascript/ql/src/semmle/javascript/security/dataflow/HttpToFileAccessCustomizations.qll (L43-L46)
for the 2 first sinks, it's important that `getAPathArgument` has a proper result... so that solves the problem, and highlights that it _can_ be important to give proper results for `getAPathArgument` (if possible).
So I'm trying to do best effort for `f = open(...); f.write(...)`, but with this current code we won't always be able to give a result (as highlighted by the tests). It will also be the case that there are multiple `FileSystemAccess` with the same path-argument, which could be a little strange.
overall, I'm not super confident about the way this new concept and implementation turned out, but it also seems like the best I could come up with right now...
The obvious alternative solution is to NOT make `FileSystemWriteAccess` a subclass of `FileSystemAccess`, but I'm not very tempted to go down this path, given the examples of this being useful above, and just the general notion that we should be able to model writes as being a specialized kind of `FileSystemAccess`.
Going all the way to the AST layer seemed excessive to me, so I rewrote
it to do most of the logic at the data-flow layer. In principle this
_could_ result in more names being computed (due to splitting), but in
practice I don't expect this make a big difference.
Problematic part is
```codeql
/** A escape from string format with `markupsafe.Markup` as the format string. */
private class MarkupEscapeFromStringFormat extends MarkupSafeEscape, Markup::StringFormat {
override DataFlow::Node getAnInput() {
result in [this.getArg(_), this.getArgByName(_)] and
not result = Markup::instance()
}
override DataFlow::Node getOutput() { result = this }
}
```
since the char-pred still holds even if `getAnInput` has no results...
I will say that doing it this way feels kinda dirty, and we _could_ fix
this by including the logic in `getAnInput` in the char-pred as well.
But as I see it, that would just lead to a lot of code duplication,
which isn't very nice.
Since expectation tests had so many changes from ConceptsTest, I'm going
to do the changes for that on in a separate commit. The important part
is the changes to taint-tracking, which is highlighted in this commit.
The other approach felt a bit too much like specifying magic strings
that you had to get right. (crossing your fingers that no-one writes
`HTML` instead of `html`)
According to the official documentation, the purpose of `__main__.py`
files is that their presence in a package (say, `foo`) means one can
execute the package directly using `python -m foo` (which will run the
aforementioned `foo/__main__.py` file).
In principle this means that adding `if __name__ == "__main__"` in these
files is superfluous, as they are only intended to be executed (and not
imported by some other file).
However, in practice people often _do_ include the above construct.
Here are some instances of this on LGTM.com:
https://lgtm.com/query/7521266095072095777/
In particular, 10 out of 33 files in `cpython` have this construct.
This causes some confusion in our module naming, as we usually see the
presence of `__name__ == "__main__"` as an indication that a file may
be run directly (and hence with "absolute import" semantics). However,
when run with `python -m`, the interpreter uses the usual package
semantics, and this leads to modules getting multiple names.
For this reason, I think it makes sense to simply exclude `__main__.py`
files from consideration. Note that if there is a `#!` line mentioning
the Python interpreter, then they will still be included as entry
points.
