Feature: SslBump using Bump-Server-First method
Goal: Allow bumping of intercepted SSL connections. Prep for mimicking server certificates details.
Developer: AlexRousskov and Christos Tsantilas
This feature was replaced in Squid-3.5 by peek-n-splice
The first SslBump implementation works well for HTTP CONNECT requests naming the host that Squid must establish a TCP tunnel with. Such requests are sent by some browsers when they are explicitly configured to use a proxy. When dealing with such requests in SslBump mode, Squid can use the supplied host name to dynamically generate a server certificate and then impersonate the named server. This allows Squid to get the real HTTP request from the client (e.g., HTTP GET or POST), decrypt it, and, eventually, connect to the real server and forward the request.
The above scheme fails when SSL connections are intercepted because intercepted connections start with an SSL handshake and not an HTTP CONNECT request. Thus, Squid does not receive the origin server host name from the client. Squid knows the destination IP address of the intercepted connection, but an IP address is not usable for SSL certificate generation. This makes it impossible to generate a matching server certificate. Without such a certificate, Squid cannot impersonate the server.
A very similar failure happens when certain clients (e.g., Rekonq browser v0.7.x) send CONNECT requests that use an IP address instead of a host name to specify the tunnel destination.
Another problem with the older "bump-client-first" approach is that whenever the server sends a partially defective or an outright invalid SSL certificate, it is too late to propagate that problem to the client and let the client deal with it. This is unfortunate both because the final decision should be, ideally, done by the user, not Squid and because browsers already have rather sophisticated tools for warning the user about the problem, examining invalid certificates, ignoring problems, caching user decision, etc. (and we do not really want to duplicate that). While this project will not forward certificate problems to the client, it is a required step towards supporting that frequently requested functionality in the future.
To bump intercepted SSL connections, this project completely changes the order of bumped connection processing events in Squid. When an intercepted connection is received, Squid first connects to the server using SSL and receives the server certificate. Squid then uses the host name inside the true server certificate to generate a fake one and impersonates the server while still using the already established secure connection to the server.
Reversing the order of connection processing events is a complex task, affecting several areas of core Squid code, and violating some of the basic Squid assumptions spread throughout the code (e.g., that every server connection is backed by an HTTP request). The following caveats were discovered during implementation and initial testing:
Bumped connection goes through several stages. Each stage affects what information is available to various squid.conf ACLs.
For intercepted connections:
- When Squid makes the connection to the server to peek the certificate, there is no HTTP request and no server name. ACLs using source and destination IP addresses/ports should work during this stage.
- After Squid receives the server certificate, the actual server name becomes available (from the CN field of the certificate). Squid starts using that name when reporting certificate details on error pages if needed, but does not assume that the future request will be directed to the same server. Thus, destination domain ACLs will not work at this stage.
- After Squid receives the first HTTP request, all HTTP request-specific ACLs should be available. For each request, Squid verifies that the requested host matches the certificate CN retrieved earlier. A SQUID_X509_V_ERR_DOMAIN_MISMATCH error is triggered and the connection with the client is terminated if there is no match.
For bumped CONNECT requests:
- When Squid makes the connection to the server to peek the certificate, there is only CONNECT HTTP request. That request may have a server name but some browsers CONNECT using IP address instead (e.g., Konqueror). ACLs using source and destination IP addresses/ports should work during this stage.
- After Squid receives the server certificate, the server name becomes available (from the CN field of the certificate) even if it was not available before. ACLs using server domain name should now work. TODO: Squid does not check whether the CONNECT host name matches the subsequent embedded HTTP request Host value. Correctly adding such checks is difficult because one of the two names may be a location/CDN-specific IP address and because the tunnel ends may be designed to use multiple host names (e.g., the server end of the tunnel could be a proxy).
- After Squid receives the first bumped HTTP request, all HTTP request-specific ACLs should be available. For each request, Squid verifies that the requested host matches the certificate CN retrieved earlier. A SQUID_X509_V_ERR_DOMAIN_MISMATCH error is triggered and the connection with the client is terminated if there is no match.
Please note that sslproxy_cert_error ACLs always check the true server certificate and not the generated fake one.
Without bumping, a client opens a secure connection to the origin server and sends/receives a few messages on that single connection. It is reasonable to assume that at least some clients depend on the destination of that connection to remain the same even if Squid bumps their connection. We have encountered such assumption when dealing with NTLM authentication, for example. However, Squid has to deal with two separate connections (one with the client and the other one with the unsuspecting server) so it is possible that Squid will have to close the server connection at any time. Squid tries to minimize the chances that the server connection will change during the client connection lifetime by using the following approach:
When establishing a server connection to peek at the server certificate, Squid pins the server connection to the client connection. Subsequent client requests will all go to that server connection as if Squid was not there. Squid also remembers the peeked server certificate.
- If server closes the connection but the client keeps sending more requests, Squid opens a new connection to the server and pins it to the client connection again. This reopening is necessary to minimize compatibility problems where the client did not expect the server to close the connection because Squid-to-client connection signaling is different from server-to-Squid connection signaling. TODO: In the future, we may send "Connection: close" to the client if the origin server says so.
- When reopening a server connection, Squid verifies that the server SSL certificate has not changed much. If server certificate has changed, Squid responds with a SQUID_X509_V_ERR_DOMAIN_MISMATCH error which was added during this project. This feature minimizes the probability that another an attacker can inject itself into the post-Squid message stream after Squid already sent a fake server certificate to the client and the client approved that fake certificate.
Why bump the server first when dealing with CONNECT requests?
Bumping server first is essentially required for handling intercepted HTTPS connections but the same scheme should be used for most HTTP CONNECT requests because it offers a few advantages compared to the old bump-client-first approach:
- When Squid knows valid server certificate details, it can generate its fake server certificate with those details. With the bump-client-first scheme, all those details are lost. In general, browsers do not care about those details but there may be HTTP clients (or even human users) that require or could benefit from knowing them.
When a server sends a bad certificate, Squid may be able to replicate that brokenness in its own fake certificate, giving the HTTP client control whether to ignore the problem or terminate the transaction. With bump-client-furst, it is difficult to support similar dynamic, user-directed opt out; Squid itself has to decide what to do when the server certificate cannot be validated.
When a server asks for a client certificate, Squid may be able to ask the client and then forward the client certificate to the server. Such client certificate handling may not be possible with the bump-client-first scheme because it would have to be done after the SSL handshake.
- Some clients (e.g., Rekonq browser v0.7.x) do not send host names in CONNECT requests. Such clients require bump-server-first even in forward proxying mode. Unfortunately, there are other problems with fully supporting such clients (i.e., Squid does not know whether the IP address in the CONNECT request is what the user have typed into the address bar) so not all features will work well for them until more specialized detection code is added.
The code being tested uses bump-server-first for CONNECT requests, but that one-for-all decision is debatable, and the choice may become configurable using a squid.conf ACL (further complicating the already rather convoluted code). Feedback is welcome.
Why not just use Server Name Indication (SNI) instead?
Instead of bumping the server first, it is possible to get the intended server name during SSL or TLS handshake using a SNI feature. We have not taken that shortcut because:
- There is no SNI support in Internet Explorer running on Windows XP.
It is not possible to mimic the server certificate so that the user can (a) decide whether to ignore any certificate problems and (b) cache that decision (see server certificate mimicking).
This project will not support forwarding of SSL Server Name Indication (SNI) information to the origin server and will make such support a little more difficult. However, SNI forwarding has its own serious challenges (beyond the scope of this document) that far outweigh the added forwarding difficulties.