Fix a redirect-on-login issue by allowing logged-out users to view 404 pages

Summary:
See T2102 and inline for discussion. This seems like the least-bad approach until we have something better.

The utility of next_uri seems much greater than the minor exposure of routable URIs.

Note that attackers can //not// detect if routable URIs are //valid// (e.g., "/D999" will always hit the login page whether it exists or not), just that they're routable. So you can only really tell if apps are installed or not.

Test Plan: Hit `/alsdknlkasnbla` while logged out, got 404 instead of login.

Reviewers: vrana, codeblock, btrahan

Reviewed By: codeblock

CC: aran

Maniphest Tasks: T2102

Differential Revision: https://secure.phabricator.com/D4012

src/applications/base/controller/Phabricator404Controller.php

@@ -2,6 +2,35 @@
 
 final class Phabricator404Controller extends PhabricatorController {
 
+  public function shouldRequireLogin() {
+
+    // NOTE: See T2102 for discussion. When a logged-out user requests a page,
+    // we give them a login form and set a `next_uri` cookie so we send them
+    // back to the page they requested after they login. However, some browsers
+    // or extensions request resources which may not exist (like
+    // "apple-touch-icon.png" and "humans.txt") and these requests may overwrite
+    // the stored "next_uri" after the login page loads. Our options for dealing
+    // with this are all bad:
+    //
+    //  1. We can't put the URI in the form because some login methods (OAuth2)
+    //     issue redirects to third-party sites. After T1536 we might be able
+    //     to.
+    //  2. We could set the cookie only if it doesn't exist, but then a user who
+    //     declines to login will end up in the wrong place if they later do
+    //     login.
+    //  3. We can blacklist all the resources browsers request, but this is a
+    //     mess.
+    //  4. We can just allow users to access the 404 page without login, so
+    //     requesting bad URIs doesn't set the cookie.
+    //
+    // This implements (4). The main downside is that an attacker can now detect
+    // if a URI is routable (i.e., some application is installed) by testing for
+    // 404 vs login. If possible, we should implement T1536 in such a way that
+    // we can pass the next URI through the login process.
+
+    return false;
+  }
+
   public function processRequest() {
     return new Aphront404Response();
   }