Django hangs when Content-Length has incorrect value

[Alexander Charykov]

When sending incorrect content-length header to application, django hangs on reading request.body.

Steps to reproduce:

curl -v -X POST -H "Accept: */*" -H "Content-Length: 22" -d "1" ​http://127.0.0.1:8000

with following handler:

from django.http import HttpResponse
from django.views.decorators.csrf import csrf_exempt

@csrf_exempt
def index(request):
    return HttpResponse(content=request.body)

[Tim Graham]

I can reproduce this, although I had to add @csrf_exempt to the view to bypass the CSRF error page.

[Tim Graham]

The security team discussed this issue to confirm that we shouldn't treat it as a security issue.

From Markus Holtermann:

The issue can also be reproduced with:

$ gunicorn foo.wsgi --access-logfile - --worker-connections 5

From Simon Charette:

I cannot reproduce when Django is behind a reverse proxy such as NGINX with proxy_request_buffering on which is the default.

Given the development server should never be used in production and gunicorn documentation strongly recommends to use it behind a proxy server this doesn't sound severe to me. We should add a timeout on body read to the development server but it's unlikely to be an attack vector if you already follow best practice.

I mean, both Django's development server and gunicorn have been known to be vulnerable to slow loris attacks and friends for a while at this point at it feels to me that serving Django apps behind a reverse proxy is the norm nowadays.

From Collin Anderson:

I this has come up before (or maybe it's a related issue): ​https://groups.google.com/d/topic/django-developers/bYzHczKNoqM/discussion

From Claude Paroz:

Might be worth reaching to Graham Dumpleton about this issue. The WSGI specs talk about a CONTENT_LENGTH shorter than the payload, but I'm not sure it addresses the reverse issue.

[Patrik Sletmo]

I skimmed through the WSGI specification and it contains a description on how to handle Content-Length headers shorter than the payload, but as Claude Paroz guessed it does not give any instructions for handling the opposite case.

As the recommended way of running Django in production is behind a reverse proxy I don't think that there should be any fix implemented that attempts to mitigate the effect of a Slowloris attack. I do however think that from a usability standpoint it could make sense not to hang the entire application in case the request given above is received.

Would there be any practical implications that could motivate against using a timeout period for receiving the data? If nginx is used as a reverse proxy with the default option proxy_request_buffering set, the timeout would limit the time required for transmitting the request from the reverese proxy to Django, and this really shouldn't take any substantial amount of time. I suppose that implementing a timeout option in Django in some situations could create the need to configure a timeout in both the reverse proxy and in Django, but I think that these cases are rather slim.

[Patrik Sletmo]

I've been thinking some more on how to handle this issue and right now I'm considering the possibility of not fixing it at all. The two alternatives would in that case be the following:

1. Mark the issue as "wontfix".

or

2. Document the issue so that users are aware of its existence, and refer to the use of a reverse proxy. My suggestion is that it could be added here somehow ​https://docs.djangoproject.com/en/2.1/topics/security/#additional-security-topics.

The reasoning behind not providing any fix beyond what has been suggested above is simple. The types of HTTP requests triggering the hang are typically only crafted by a malicious party with the intention to trigger this sort of bug. I can not imagine any case in which the behaviour documented in this issue would come as a surprise to anyone performing this request on their development server. As this bug is nearly never triggered unintentionally when developing, and also protected against when deployed properly in production, I think that it seems strange to introduce additional functionality that must be maintained. I would say that an introduced timeout comes with more complications than it solves.

Do you have any opinion on the above mentioned proposals, Alexander Charykov and Tim Graham?

[Patrik Sletmo]

Since I lack the required insight in the Django project and thus cannot decide on what solution would be the best one on my own, I've chosen to deassign myself from this issue. I hope that my reasoning above may become useful in the future.

[Jannik Schürg]

I think this is what it boils down to:

from wsgiref.simple_server import make_server

def app(environ, start_response):
    # does not timeout for malformd 'Content-Length'
    environ['wsgi.input'].read()

make_server('', 8000, app).serve_forever()

[Zach Garwood]

Replying to Patrik Sletmo:

2. Document the issue so that users are aware of its existence, and refer to the use of a reverse proxy. My suggestion is that it could be added here somehow ​https://docs.djangoproject.com/en/2.1/topics/security/#additional-security-topics.

I think this is the best course of action.

[Manan]

Added a csrf_exempt to bypass the default 403 CSRF error page