Logged out sessions are resurrected by concurrent requests

[Jonas Borgström]

1. User logs in
2. User loads a slow page in separate tab or as an ajax request, which modifies the session
3. User logs out before request in step 2 completes. This will delete the session from the db

Expected behavior

User/session stays logged out since the user explicitly logged out and the session row was delete in step 3.

Actual behavior

The previously deleted session is re-inserted into the database when the request from step 2 completes. So the previously logged out user is now logged in again.

[Jonas Borgström]

Proposed fix (against django 1.4)

[Jonas Borgström]

bump

[Russell Keith-Magee]

Seems like a reasonable request, and the patch looks like a decent start -- but it needs tests.

[nikl@…]

Finalized on the train and airport: ​https://github.com/django/django/pull/2678

Thanks to everybody at DjangoIsland who helped me tackle this - looking forward to your feedback!

• Nikl

[Tim Graham]

Please don't mark your own patch as RFC. Someone who reviews the patch should do that.

[Sasha Romijn]

I'm not entirely getting this. When a user logs out, ​the session is flushed. Flushing the session ​clears it and deletes it. The database session store performs this deletion by actually ​deleting the record from the DB. The cached_db backend ​deletes it from the DB and the cache. So basically, all records of this session should be deleted. If you would post a new request with the now deleted session ID, ​Django will reject it, and assign you a new session with a new session ID.

The reporter says that django re-inserts the session when a request arrives with the old session ID, and will re-insert it with the old session data. But I don't see that anywhere in the code. As far as I can see, Django would reject the session ID, as loading would fail as the session object has been deleted, and the user would be assigned a new session. Even if there were a flaw in that logic: once the session data has been deleted, how would any code know how to recreate the session? The request doesn't contain any hint on what user should be logged in.

The only explanation I can come up with is that we're talking about cookie backed sessions, for which this is a documented limitation: you can't guarantee deletion of a cookie backed session or it's data, no matter what we do in Django: its the nature of cookies.

[Tim Graham]

I couldn't reproduce this using steps 1-3 in the description (SQLite). After logging out in a separate tag, the slow page loaded, but subsequent requests redirected to the admin login page. There also seem to be some concerns from Nick's review on the PR.

[Jonas Borgström]

I think one key detail missing from the initial reproduction steps is that the "slow page" needs to modify the session to make it dirty. Otherwise the session will not be resurrected.

Anyway, I've now create a complete reproduction test case here:
​https://github.com/jborg/django-21608

See README.txt for details.

[Collin Anderson]

would session.save(force_update=True) fix this issue?

[Tore Lundqvist]

I have reproduced this bug with the test site (​https://github.com/jborg/django-21608) on both Django 1.8.7 and 1.9 (using SQLite).

These were the steps I used:

• Open the admin page in a tab and log in. ​http://localhost:8000/admin/
• Switch to a new tab and open the slow page. ​http://localhost:8000/slow/ Act fast after this step to complete the next three steps before the slow page has finished loaded, <10 sec
• Switch back to tab with the admin page.
• Click the "Logout" link on the top right corner of the page. -> Now you are on the logout page
• Reload page -> Now you are on the login page.
• Wait for slow page to finish loading.
• Reload the tab with the login page.
• Logged in again without entering credentials!

This is a security issue, not critical though, as someone might think that they have logged out but is actually still logged in.
If you logout and leave a public computer while a page is still loading in another tab there is a risk that the next person using that computer can get access to your account.
It would be nice to have this fixed 1.8 and up.

[Tore Lundqvist]

Run the unit test without the fix to verify the bug. This fix should also be back ported to 1.8 as it is a security fix.

[Tim Graham]

Are you able to submit the patch as a pull request?

[Tore Lundqvist]

Absolutely, which branch should I create a PR against?

[Tim Graham]

Master please.

[Tore Lundqvist]

Hi again, pull request is up.

​https://github.com/django/django/pull/5950

[Esbjörn Ekberg]

I've gone through the patch review checklist and everything is looking good.

[Tim Graham <timograham@…>]

In 3389c5e:

Fixed #21608 -- Prevented logged out sessions being resurrected by concurrent requests.

Thanks Simon Charette for the review.

[Tim Graham <timograham@…>]

In 5faf745:

Refs #21608 -- Fixed incorrect cache key in cache session backend's save().

The bug was introduced commit 3389c5ea229884a1943873fe7e7ffc2800cefc22.

[Dan Tao]

I commented ​here, but just to raise visibility: I'm concerned that the change to fix this bug resulted in a logical error (or at least unintuitive behavior) in SessionStore.save(). Namely: now must_create=False implies must_update=True, which I would argue is wrong. must_create=False should probably mean that either creating or updating is acceptable.

Has this already been discussed elsewhere, and perhaps there's something I'm missing?