SimpleCookies are not correctly serialized with the file or database cache backends

[Raphael Kubo da Costa]

As discussed in the ​django-developers mailing list, it appears that SimpleCookies in HttpResponses are not being correctly serialized when one uses either the file or the database cache backends.

The following steps are enough to trigger the incorrect behaviour:

1. Enable UpdateCacheMiddleware and FetchFromCacheMiddleware in settings.py, and set CACHE_BACKEND accordingly
2. Enable SessionMiddleware and CsrfViewMiddleware
3. Have a view with a simple form and no specific cache decorators. Since the session application is being used, the Vary: Cookie header will be added anyway.
4. In the template used by the view, include the csrf_token tag, as usual.
5. Access the view, either via curl or a web browser.
6. The first time the view is accessed, the csrf token is both set in the header as a cookie and displayed as a hidden form element, as expected. The header has the format Set-Cookie: csrftoken=XX; Max-Age: YY; Path=/.
7. The next times the view is accessed, the cookie header has the format Set-Cookie: csrftoken="Set-Cookie: csrftoken=XX Max-Age: YY; Path=/", and so has the csrf form element, which causes the submitted form to be invalid when the csrf checks are made.

It turns out that UpdateCacheMiddleware serializes the returned HttpResponse in process_response, and both the file and the database cache backends use pickle.dumps with protocol=pickle.HIGHEST_PROTOCOL. It is known that ​SimpleCookies are incompatible with pickle.HIGHEST_PROTOCOL. FetchFromCacheMiddleware later retrieves this same HttpResponse and returns it, however the cookies have invalid values.

The attached testcase triggers the problem in the unit tests.

[Raphael Kubo da Costa]

Testcase referenced in the bug description.

[Paul McMillan]

Marking this as accepted, per the mailing list discussion.

The attached testcase triggers a similar error, and is the reason that this is a problem, but it is not a valid test for this bug (since properly fixing this bug will not cause the test to pass).

I'll also note here that the local memory cache backend should use the highest pickle protocol so that it is similar to the other backends.

[Raphael Kubo da Costa]

Proposed addition to Django's test suite

[Raphael Kubo da Costa]

Hey there. I've finally found some time to go back to this problem. The patch I've just attached is suits better in Django's unit tests -- it uses middlewares in the base test class, but by being in the base class it is easier to test all existing backends.

As for changing the locmem cache backend to use the highest pickle protocol, would you like me to create a new ticket with a patch for that?

[teolicy]

I ran into a similar snag (caching breaks cookies, because pylibmc uses the highest protocol version).

Now, perhaps I misunderstood something along the way, but as far as I can see the solution should be to make SimpleCookie a properly serializable object.

I mean, what's so special about SimpleCookie that makes it unreasonable to ask for this code to simply work?

>>> from django.http import HttpResponse
>>> resp = HttpResponse('')
>>> resp.set_cookie('xx', 'yy')
>>> import cPickle
>>> cPickle.loads(cPickle.dumps(resp)).cookies # this is fine
<SimpleCookie: xx='yy'>
>>> cPickle.loads(cPickle.dumps(resp, 2)).cookies # this is broken
<SimpleCookie: xx='Set-Cookie: xx=yy; Path=/'>
>>>

Also, I humbly think that the ticket's Component should be HTTP Handling, not Cache, and that this is something that should be slated for 1.3.1, not 1.4.

[Luke Plant]

@teolicy:

From the Python bug report, it looks like they are going to WONTFIX it. Ideally, our SimpleCookie is exactly the same as the stdlib SimpleCookie (currently it isn't with any released version of Python, but as fixes get in to Pythno stdlib it should become that). Since we don't to diverge from stdlib if we can avoid it, we'd prefer to put the fix elsewhere.

Regarding Django 1.3.1, our current policy is that we do not backport bugfixes like this to the stable branch.

[Paul McMillan]

I split off the issue of locmem pickle protocol into #16378.

[Jacob]

Milestone 1.4 deleted

[Paul McMillan]

I'm marking this as a release blocker. It shouldn't be possible to mix and match the built in functionality we provide to silently return incorrect results.

I'm leaving it assigned to myself since I've looked at the issue already, but if anyone else wants to fix it before I get to it, please feel free to take it over.

The fix here is to serialize the SimpleCookie to a string before pickling the request for the cache.

[Luke Plant]

Patch (incorporating tests)

[Luke Plant]

In [17200]:

Fixed #15863 - SimpleCookies are not correctly serialized with the file or database cache backends

Thanks to rakuco for the report and for the tests.

[Claude Paroz <claude@…>]

In 4d817b38875c900d70793acd528afc9e954bbcb7:

Fixed #19262 -- Support cookie pickling in SimpleTemplateResponse

Refs #15863.

[Claude Paroz <claude@…>]

In 6554137eebe4bd10bdf3f1be21f63f0a9cffd7ff:

[1.5.x] Fixed #19262 -- Support cookie pickling in SimpleTemplateResponse

Refs #15863.
Backport of 4d817b3887 from master.