Fwd: Re: Universal parsing library in the stdlib to alleviate security issues
Forward to the list because Abusix had blocked google.com initially. Nam ---------- Forwarded message --------- From: Nam Nguyen <bitsink@gmail.com> Date: Sun, Jul 28, 2019 at 10:18 AM Subject: Re: [Python-ideas] Re: Universal parsing library in the stdlib to alleviate security issues To: Sebastian Kreft <skreft@gmail.com> Cc: Paul Moore <p.f.moore@gmail.com>, python-ideas <python-ideas@python.org> Let's circle back to the beginning one last time ;). On Thu, Jul 25, 2019 at 8:15 AM Sebastian Kreft <skreft@gmail.com> wrote:
Nam, I think it'd be better to frame the proposal as a security enhancement. Stating some of the common bugs/gotchas found when manually implementing parsers, and the impact this has had on python over the years. Seeing a full list of security issues (CVEs) by module would give us a sense of how widespread the problem is.
Since my final exam was done this weekend, I gathered some more info into this spreadsheet. https://docs.google.com/spreadsheets/d/1TlWSf8iM7eIzEPXanJAP8Ztyzt4ZD28xFvUK... I think a strict parser can help with the majority of those problems. They are in HTTP headers, emails, cookies, URLs, and even low level socket code (inet_atoi).
Then survey the stdlib for what kind of grammars are currently being parsed, what ad-hoc parsing strategy are implemented and provide examples of whether having a general purpose parser would have prevented the security issues you have previously cited.
Most grammars I have seen here come straight from RFCs, which are in ABNF and thus context-free. Current implementations are based on regexes or string splitting. My previous example showed that at least 30500, 36216, 36742 were non-issues if we started out with a strict parser.
Right now, it is not clear what the impact of such refactor would be, nor the worth of such attempt.
Exactly the kind of response I'm looking for. It is okay to suggest that the benefits aren't clear or that there are requirements X and Y that a general parser won't be able to meet, but it's not convincing to brush aside this because there is "existing, working code." Many of the bugs in that sheet are still open. It's not comfortable to say the code is working with a straight face as I have experienced with my own fix for 30500. I just couldn't tell if it was doing the right thing.
What others have said earlier is that you are the one that needs to provide some of the requirements for the proposed private parsing library. And from what I read from your emails you do have some ideas. For example, you want it to be easy to write and review (I guess here you would eventually like it to be a close translation from whatever is specified in the RFC or grammar specification).
Yes, that's the most important point because "readability counts." It's hard to reason about correctness when there are many transformations between the authoritative spec and the implementation. I definitely don't want to touch the regexes, string splits, and custom logic that I don't understand "why" they are that way in the beginning. How do I, for example, know what this regex is about ^(([^:/?#]+):)?(//([^/?#]*))?([^?#]*)(\?([^#]*))?(#(.*))? (It's from RFC 3986.) But you also need to take into consideration some of the list's concerns,
the parser library has to be performant, as a performance regression is likely not to be tolerable.
Absolutely. That's where I need inputs from the list. I have provided my own set of requirements for such a parser library. I'm sure most of us have different needs too. So if a parser library can help you, let's hear what you want from it. If you think it can't, please let me understand why. Thanks, Nam
On Thu, Jul 25, 2019 at 10:54 AM Nam Nguyen <bitsink@gmail.com> wrote:
On Thu, Jul 25, 2019 at 2:32 AM Paul Moore <p.f.moore@gmail.com> wrote:
On Thu, 25 Jul 2019 at 02:16, Nam Nguyen <bitsink@gmail.com> wrote:
Back to my original requests to the list: 1) Whether we want to have a (possibly private) parsing library in the stdlib
In the abstract, no. Propose a specific library, and that answer would change to "maybe".
I have no specific library to propose. I'm looking for a list of features such a library should have.
and 2) What features it should have.
That question only makes sense if you get agreement to the abstract proposal that "we should add a parsing library. And as I said, I don't agree to that so I can't answer the second question.
As Chris summarized it correctly, I am advocating for a general solution to individual problems (which have the same nature). We can certainly solve the problems when they are reported, or we can take a proactive approach to make them less likely to occur. I am talking about a class of input validation issues here and I thought parsing would be a very natural solution to that. This is quite similar to a context-sensitive templating library that prevents cross-site-scripting on the output side. So I don't know why (or what it takes) to convince people that it's a good thing(tm).
Generally, things go into the stdlib when they have been developed externally and proved their value. The bar for designing a whole library from scratch, "specifically" targeted at stdlib inclusion, is very high, and you're nowhere near reaching it IMO.
This is a misunderstanding. I have not proposed any from-scratch, or existing library to be used. And on this note, please allow me to make it clear once more time that I am not asking for a publicly-facing library either.
These are good points to set as targets! What does it take for me to get the list to agree on one such set of criteria?
You need to start by getting agreement on the premise that adding a newly-written parser to the stdlib is a good idea. And so far your *only* argument seems to be that "it will avoid a class of security bugs" which I find extremely unconvincing (and I get the impression others do, too).
Why? What is unconvincing about a parsing library being able... parse (and therefore, validate) inputs?
But even if "using a real parser" was useful in that context, there's *still* no argument for writing one from scratch, rather than using an existing, proven library.
Never a goal.
At the most basic level, what if there's a bug in your new parsing library? If we're using it in security-critical code, such a bug would be a vulnerability just like the ones you're suggesting your parser would avoid. Are you asking us to believe that your code will be robust enough to trust over code that's been used in production systems for years?
I think you need to stop getting distracted by details, and focus on your stated initial request "Whether we want to have a (possibly private) parsing library in the stdlib". You don't seem to me to have persuaded anyone of this basic suggestion yet,
Good observation. How do I convince you that complex input validation tasks should be left to a parser?
Thanks! Nam
_______________________________________________ Python-ideas mailing list -- python-ideas@python.org To unsubscribe send an email to python-ideas-leave@python.org https://mail.python.org/mailman3/lists/python-ideas.python.org/ Message archived at https://mail.python.org/archives/list/python-ideas@python.org/message/FCPU4Z... Code of Conduct: http://python.org/psf/codeofconduct/
-- Sebastian Kreft
Nam Nguyen writes:
Since my final exam was done this weekend, I gathered some more info into this spreadsheet.
https://docs.google.com/spreadsheets/d/1TlWSf8iM7eIzEPXanJAP8Ztyzt4ZD28xFvUK...
This is useful!
Most grammars I have seen here come straight from RFCs,
Grammars are only truly relevant if you have a parser-compiler for them. Otherwise, you're still translating by hand. See following discussion of urlsplit, which has issues #30500 and #36216 etc.
which are in ABNF and thus context-free. Current implementations are based on regexes or string splitting. My previous example showed that at least 30500, 36216, 36742 were non-issues if we started out with a strict parser.
For #30500, I don't recall you dealing with the claim that the parse is not for RFC 3986 URIs, but for a sort of "raw URI" which might not be percent-encoded. (AFAICT, RFC 3986 taken strictly restricts proper URIs to a subset of ASCII.) In that case, the translation would not be straightforward. This understanding is supported by #36216, which refers to IDNA. It's possible that urlsplit is intended to deal with RFC 3987 IRIs, but I tend to the view that it's just unclear. ;-)
that sheet are still open. It's not comfortable to say the code is working with a straight face as I have experienced with my own fix for 30500. I just couldn't tell if it was doing the right thing.
If it's just a straight implementation of RFC 3986, that shouldn't be too hard.
How do I, for example, know what this regex is about
^(([^:/?#]+):)?(//([^/?#]*))?([^?#]*)(\?([^#]*))?(#(.*))?
from RFC 3986?
By rewriting it as uri = r"""^(?: ([^:/?#]+): # group 1, optional scheme )?(?: //([^/?#]*) # group 2, optional authority )? ([^?#]*) # group 3, optional path (?: \?([^#]*) # group 4, optional query )?(?: #(.*) # group 5, optional fragment )?$""" (where I've revised it to use non-capturing groups and explicitly continue to end-of-string) and compiling with re.VERBOSE, or some similar device. An experienced Pythonista will likely do a double-take at group 3, then realize that "path" is a component that isn't delimited by a reserved character. Note that this regexp assumes a string containing no characters outside of the ASCII subset defined in the RFC! I'm not sure which is easier to read, this or the rather long grammar of RFC 3986 (or the even longer grammar of RFC 3897).
Absolutely. That's where I need inputs from the list. I have provided my own set of requirements for such a parser library. I'm sure most of us have different needs too. So if a parser library can help you, let's hear what you want from it.
I believe you've already heard from the people on this list who care. Its members are active participants, mostly. Its lurkers are frequently core devs who figure if it gets traction they will give their input on -dev. I think a better place to take a poll like this would be python-list@python.org. Steve
participants (2)
-
Nam Nguyen -
Stephen J. Turnbull