Squid Web Cache wiki

Squid Web Cache documentation

🔗 Feature: Mimic original SSL server certificate when bumping traffic

🔗 Motivation

One of the SslBump serious drawbacks is the loss of information embedded in SSL server certificate. There are two basic cases to consider from Squid point of view:

Hiding original certificate information has never been the intent of lawful SslBump deployments. Instead, it was an undesirable side-effect of the initial SslBump implementation. Fortunately, this limitation can be removed in most cases, making SslBump less intrusive and less dangerous.

🔗 Implementation overview

OpenSSL APIs allow us to extract and use origin server certificate properties when generating a fake server certificate. In general, we want to mimic all properties, but various SSL rules make micking of some properties technically infeasible, and browser behavior makes mimicking most properties undesirable under certain conditions. We detail these exceptions below. Squid administrator can tweak mimicking algorithms using sslproxy_cert_adapt and sslproxy_cert_sign configuration options.

The ssl_crtd daemon receives matching configuration options as well as the original server certificate to mimic its properties.

A bump-server-first support is required to get the original server certificate before we have to send our fake certificate to the client.

🔗 Fake Certificate properties

This section documents how each fake certificate property is generated. The “true” adjective is applied to describe a property of the SSL certificate received from the origin server. The “intended” adjective describes a property of the request or connection received from the client (including intercepted connections).

x509 certificate property After successful bumping After failed bumping
Common Name (CN) True CN by default. Can be overwritten using sslproxy_cert_adapt setCommonName. If the CONNECT address is available, then use it, subject to CN length controls discussed separately below. Otherwise, if true CN is available, then use that. If this is an intercepted connection and no true CN is available, then the certificate will have no CN (and no Subject).
Alias True alias, if any. None.
Subject True subject by default. The CN part can be overwritten (see CN). Contains CN only (see CN).
Subject Alternative Names (subjectAltName) True names, if any, by default. None if using sslproxy_cert_adapt setCommonName (browsers reject certificates where alternative names are not related to CN). None.
Signer and signature Configured trusted Squid CA by default. To mimic an “untrusted true server certificate” error, Squid generates an untrusted certificate with the trusted certificate subject prefixed by an “Untrusted by” string (Squid signs with this untrusted certificate as needed, but does not send it to the user, preventing its caching). To mimic a self-signed certificate error, Squid makes a self-signed certificate. Configured trusted Squid CA certificate.
Issuer The subject of the signing certificate (see Signer).  
Serial Number A positive 20-byte SHA1 hash of signing certificate and fake certificate properties. Browsers reject certificates that have the same Issuer, same serial number, but different CNs. Since Squid has to use the same Issuer for nearly all CNs, we must ensure that serial numbers are virtually never the same if CNs differ, even when generated on independent Squids.  
Validity dates True dates by default. If a true validity date is missing or if sslproxy_cert_adapt setValidAfter and setValidBefore is active, then the signing certificate validity date is used. Squid trusted certificate validity dates.
Version Version 3 when any certificate extension (e.g., subjectAltName) is mimicked (per RFC 5280). Otherwise, OpenSSL sets the version (usually to 1?) Set by OpenSSL (usually to 1?)
Other Not mimicked or set (see Limitations).  

All certificates generated by Squid are signed using the configured trusted CA certificate private key. This, along with the serial number generation algorithm, allows independent but identically configured Squids (including but not limited to Squid SMP workers) to generate identical certificates under similar circumstances.

🔗 Delayed error responses

When Squid fails to negotiate a secure connection with the origin server and bump-ssl-server-first is enabled, Squid remembers the error page and serves it after establishing the secure connection with the client and receiving the first encrypted client request. The error is served securely. The same approach is used for Squid redirect messages configured via deny_info. This error delay is implemented because (a) browsers like FireFox and Chromium do not display CONNECT errors correctly and (b) intercepted SSL connections must wait for the first request to serve an error.

Furthermore, when Squid encounters an error, it uses a trusted certificate with minimal properties to encrypt the connection with the client. If we try to mimic the true broken certificate instead, the user will get a browser error dialog and then, if user allows, the Squid error page with essentially the same (and possibly more detailed/friendly) information about the problem. Using a trusted certificate avoids this “double error” effect in many cases. And, after all, the information is coming from Squid and not the origin server so it is kind of wrong to mimic broken origin server details when serving that information.

Squid closes the client connection after serving the error so that no requests are sent to the broken server.

It is important to understand that Squid can be configured to ignore or tolerate certain SSL connection establishment errors using sslproxy_cert_error. If the error is allowed, Squid forgets about the error, mimics true broken certificate properties, and continues to talk to the server. Otherwise, Squid does not mimic and terminates the server connection as discussed above. Thus, if you want users to see broken certificate properties instead of Squid error pages, you must tell Squid to ignore the error.

🔗 Long domain names

Section A.1 of RFC 5280 limits a Common Name field of an SSL certificate to 64 characters. As far as we know, that implies that secure sites with longer names must use wildcard certificates. And since wildcards cannot be applied to TLDs (e.g., browsers reject a *.com wildcard), there can be no secure site with a long second-level domain label.

If Squid receives a valid true certificate, Squid does not try to enforce CN length limit and simply mimics true certificate fields as described in the table above. However, when Squid fails to connect to the origin server or fails to receive a usable true certificate, Squid has to generate a minimal fake certificate from scratch and has to deal with long domain names of the sites a user intended to visit. To shorten the name, Squid tries to replace the lower level domain label(s) with a wild card until the CN length no longer exceeds the 64 character limit. If that replacement results in a TLD wildcard such as *.com or, worse, in a bare * wildcard, then Squid produces a certificate with no CN at all. Such certificates are usually rejected by browsers with various, often misleading, errors. For example,

Long domain name in the request Certificate CN for serving Squid errors Comments
llanfairpwllgwyngyllgogerychwyrndrobwyll-llantysiliogogogoch.com llanfairpwllgwyngyllgogerychwyrndrobwyll-llantysiliogogogoch.com This domain name is exactly 64 characters long so it is within the CN limits.
www.llanfairpwllgwyngyllgogerychwyrndrobwyll-llantysiliogogogoch.com none Squid refuses to generate a *.com wildcard and replacing just “www” with “*” would exceed the 64 character limit by 2 characters.
this-long-domain-exceeds-64-chars-but-should-not-crash-ssl-crtd.example.com *.example.com Browsers will accept this wildcard and show Squid error page.
www.this-long-domain-exceeds-64-chars-but-should-not-crash-ssl-crtd.example.com *.example.com Browsers will refuse this wildcard because they apparently do not allow a wildcard to replace more than one domain label.

Hopefully, excessively long domains are rare for secure sites. TODO: Find a public secure site with a long domain name that actually works.

🔗 URLs with IP addresses

A user may type SSL server IP address in the address bar. Some browsers (e.g., Rekonq browser v0.7.x) send IP addresses in CONNECT requests even when the user typed a host name in the address bar. Currently, Squid cannot distinguish the two cases and assumes that an IP address in the CONNECT request implies that the user typed that address in the address bar. Besides assuming user input, Squid overall behavior here is meant to mimic what would happen if Squid was not in the loop. Here are a few cases when the user enters something like instead of https://www.google.com/:

Squid configuration Comments Browser displays
No SslBump This is because the server certificate does not use an IP address for CN. Browser’s internal “Server’s certificate does not match the URL” error.
SslBump with default squid.conf This matches no-SslBump behavior. However, see “Always IP” below the table. Squid’s SQUID_X509_V_ERR_DOMAIN_MISMATCH error page, served with CN set to the IP address from the CONNECT request.
sslproxy_cert_error allow all This is correct behavior because Squid was told to ignore errors and was not told to adapt the origin server CN. The origin server set CN to www.google.com or equivalent while the browser was expecting an IP address. Browser’s internal “Server’s certificate does not match the URL” error.
sslproxy_cert_error allow all sslproxy_cert_adapt setCommonName ssl::certDomainMismatch Because Squid sets fake certificate CN to the IP address from the CONNECT request. However, see “Always IP” below. Google page without an error.
sslproxy_cert_error allow all sslproxy_cert_adapt setCommonName{} ssl::certDomainMismatch Squid sets fake certificate CN to the IP address from the CONNECT request. However, see “Always IP” below. Google page without an error.

Always IP: Configurations with this comment may not work with browsers that always use IP addresses in CONNECT requests because their second request Host header will not match the CN IP. There is nothing Squid can do here until we learn how to detect CONNECT requests from such browsers.

🔗 IPv6

IPv6 addresses in Request URIs are handled as discussed above. The only IPv6-related caveat is that Squid strips surrounding square brackets when it has to form a certificate CN field based on the IP address. Browsers such as Firefox, Chromium, and Safari prefer bare IPv6 addresses in CNs even if the URL has a bracketed IPv6 address. These browsers generate confusing errors when they see bracketed CNs. For example:

  You attempted to reach [2001:470:1:18::120], but instead you actually reached
  a server identifying itself as [2001:470:1:18::120]. Chromium can say for sure
  that you reached [2001:470:1:18::120], but cannot verify that that is the same
  site as [2001:470:1:18::120] which you intended to reach.

🔗 Limitations

Some browsers (e.g., Rekonq browser v0.7.x) send IP addresses in CONNECT requests even when the user typed a host name in the address bar. Squid cannot handle both such browsers and URLs with IP addresses instead of host names because Squid cannot distinguish one case from another. There is nothing we can do about it until somebody contributes code to reliably detect CONNECT requests from those “unusual” browsers.

SQUID_X509_V_ERR_DOMAIN_MISMATCH errors are not checked until the first encrypted request arrives from the client. It is impossible to check for those errors earlier when dealing with intercepted connections or when talking to a browser that does not use domain names in CONNECT requests. It is possible to check for such errors when dealing with CONNECT requests that contain intended domain name information, but Squid does not.

Certificate chains are not mimicked.

🔗 Certificate properties not mimicked or set

Not all true certificate properties are mimicked. Initially, we thought it is a good idea to mimic everything by default, but we quickly ran into problems with browsers rejecting fake certificates due to mismatching or otherwise invalid combination of properties (e.g., alternative names not matching CN). We now mimic only the properties that are unlikely to cause problems. However, a few other properties may still be investigated for mimicking: Certificate Policies, Subject Directory Attributes, Extended Key Usage, Freshest CRL, and Subject Information Access.

The following properties may be worth setting if configured (via the CA certificate?): Authority Key Identifier, Subject Key Identifier, Key Usage, Issuer Alternative Name, Freshest CRL, and Authority Information Access.

The following properties are probably not applicable because they deal with CA or other specialized certificates (or are too vague to be mimicked safely): Basic Constraints, Name Constraints, Policy Constraints, and Inhibit anyPolicy.

Categories: Feature

Navigation: Site Search, Site Pages, Categories, 🔼 go up