[APACHE DOCUMENTATION]

Apache HTTP Server Version 1.3

An In-Depth Discussion of VirtualHost Matching

This is a very rough document that was probably out of date the moment it was written. It attempts to explain exactly what the code does when deciding what virtual host to serve a hit from. It's provided on the assumption that something is better than nothing. The server version under discussion is Apache 1.2.

If you just want to "make it work" without understanding how, there's a What Works section at the bottom.

Config File Parsing

There is a main_server which consists of all the definitions appearing outside of VirtualHost sections. There are virtual servers, called vhosts, which are defined by VirtualHost sections.

The directives Port, ServerName, ServerPath, and ServerAlias can appear anywhere within the definition of a server. However, each appearance overrides the previous appearance (within that server).

The default value of the Port field for main_server is 80. The main_server has no default ServerName, ServerPath, or ServerAlias.

In the absence of any Listen directives, the (final if there are multiple) Port directive in the main_server indicates which port httpd will listen on.

The Port and ServerName directives for any server main or virtual are used when generating URLs such as during redirects.

Each address appearing in the VirtualHost directive can have an optional port. If the port is unspecified it defaults to the value of the main_server's most recent Port statement. The special port * indicates a wildcard that matches any port. Collectively the entire set of addresses (including multiple A record results from DNS lookups) are called the vhost's address set.

The magic _default_ address has significance during the matching algorithm. It essentially matches any unspecified address.

After parsing the VirtualHost directive, the vhost server is given a default Port equal to the port assigned to the first name in its VirtualHost directive. The complete list of names in the VirtualHost directive are treated just like a ServerAlias (but are not overridden by any ServerAlias statement). Note that subsequent Port statements for this vhost will not affect the ports assigned in the address set.

All vhosts are stored in a list which is in the reverse order that they appeared in the config file. For example, if the config file is:

    <VirtualHost A>
    ...
    </VirtualHost>

    <VirtualHost B>
    ...
    </VirtualHost>

    <VirtualHost C>
    ...
    </VirtualHost>
Then the list will be ordered: main_server, C, B, A. Keep this in mind.

After parsing has completed, the list of servers is scanned, and various merges and default values are set. In particular:

  1. If a vhost has no ServerAdmin, ResourceConfig, AccessConfig, Timeout, KeepAliveTimeout, KeepAlive, MaxKeepAliveRequests, or SendBufferSize directive then the respective value is inherited from the main_server. (That is, inherited from whatever the final setting of that value is in the main_server.)
  2. The "lookup defaults" that define the default directory permissions for a vhost are merged with those of the main server. This includes any per-directory configuration information for any module.
  3. The per-server configs for each module from the main_server are merged into the vhost server.
Essentially, the main_server is treated as "defaults" or a "base" on which to build each vhost. But the positioning of these main_server definitions in the config file is largely irrelevant -- the entire config of the main_server has been parsed when this final merging occurs. So even if a main_server definition appears after a vhost definition it might affect the vhost definition.

If the main_server has no ServerName at this point, then the hostname of the machine that httpd is running on is used instead. We will call the main_server address set those IP addresses returned by a DNS lookup on the ServerName of the main_server.

Now a pass is made through the vhosts to fill in any missing ServerName fields and to classify the vhost as either an IP-based vhost or a name-based vhost. A vhost is considered a name-based vhost if any of its address set overlaps the main_server (the port associated with each address must match the main_server's Port). Otherwise it is considered an IP-based vhost.

For any undefined ServerName fields, a name-based vhost defaults to the address given first in the VirtualHost statement defining the vhost. Any vhost that includes the magic _default_ wildcard is given the same ServerName as the main_server. Otherwise the vhost (which is necessarily an IP-based vhost) is given a ServerName based on the result of a reverse DNS lookup on the first address given in the VirtualHost statement.

Vhost Matching

Apache 1.3 differs from what is documented here, and documentation still has to be written.

The server determines which vhost to use for a request as follows:

find_virtual_server: When the connection is first made by the client, the local IP address (the IP address to which the client connected) is looked up in the server list. A vhost is matched if it is an IP-based vhost, the IP address matches and the port matches (taking into account wildcards).

If no vhosts are matched then the last occurrence, if it appears, of a _default_ address (which if you recall the ordering of the server list mentioned above means that this would be the first occurrence of _default_ in the config file) is matched.

In any event, if nothing above has matched, then the main_server is matched.

The vhost resulting from the above search is stored with data about the connection. We'll call this the connection vhost. The connection vhost is constant over all requests in a particular TCP/IP session -- that is, over all requests in a KeepAlive/persistent session.

For each request made on the connection the following sequence of events further determines the actual vhost that will be used to serve the request.

check_fulluri: If the requestURI is an absoluteURI, that is it includes http://hostname/, then an attempt is made to determine if the hostname's address (and optional port) match that of the connection vhost. If it does then the hostname portion of the URI is saved as the request_hostname. If it does not match, then the URI remains untouched. Note: to achieve this address comparison, the hostname supplied goes through a DNS lookup unless it matches the ServerName or the local IP address of the client's socket.

parse_uri: If the URI begins with a protocol (i.e., http:, ftp:) then the request is considered a proxy request. Note that even though we may have stripped an http://hostname/ in the previous step, this could still be a proxy request.

read_request: If the request does not have a hostname from the earlier step, then any Host: header sent by the client is used as the request hostname.

check_hostalias: If the request now has a hostname, then an attempt is made to match for this hostname. The first step of this match is to compare any port, if one was given in the request, against the Port field of the connection vhost. If there's a mismatch then the vhost used for the request is the connection vhost. (This is a bug, see observations.)

If the port matches, then httpd scans the list of vhosts starting with the next server after the connection vhost. This scan does not stop if there are any matches, it goes through all possible vhosts, and in the end uses the last match it found. The comparisons performed are as follows:

  • Compare the request hostname:port with the vhost ServerName and Port.
  • Compare the request hostname against any and all addresses given in the VirtualHost directive for this vhost.
  • Compare the request hostname against the ServerAlias given for the vhost.

check_serverpath: If the request has no hostname (back up a few paragraphs) then a scan similar to the one in check_hostalias is performed to match any ServerPath directives given in the vhosts. Note that the last match is used regardless (again consider the ordering of the virtual hosts).

Observations

  • It is difficult to define an IP-based vhost for the machine's "main IP address". You essentially have to create a bogus ServerName for the main_server that does not match the machine's IPs.
  • During the scans in both check_hostalias and check_serverpath no check is made that the vhost being scanned is actually a name-based vhost. This means, for example, that it's possible to match an IP-based vhost through another address. But because the scan starts in the vhost list at the first vhost that matched the local IP address of the connection, not all IP-based vhosts can be matched.

    Consider the config file above with three vhosts A, B, C. Suppose that B is a named-based vhost, and A and C are IP-based vhosts. If a request comes in on B or C's address containing a header "Host: A" then it will be served from A's config. If a request comes in on A's address then it will always be served from A's config regardless of any Host: header.

  • Unless you have a _default_ vhost, it doesn't matter if you mix name-based vhosts in amongst IP-based vhosts. During the find_virtual_server phase above no named-based vhost will be matched, so the main_server will remain the connection vhost. Then scans will cover all vhosts in the vhost list.

    If you do have a _default_ vhost, then you cannot place named-based vhosts after it in the config. This is because on any connection to the main server IPs the connection vhost will always be the _default_ vhost since none of the name-based are considered during find_virtual_server.

  • You should never specify DNS names in VirtualHost directives because it will force your server to rely on DNS to boot. Furthermore it poses a security threat if you do not control the DNS for all the domains listed. There's more information available on this and the next two topics.
  • ServerName should always be set for each vhost. Otherwise A DNS lookup is required for each vhost.
  • A DNS lookup is always required for the main_server's ServerName (or to generate that if it isn't specified in the config).
  • If a ServerPath directive exists which is a prefix of another ServerPath directive that appears later in the configuration file, then the former will always be matched and the latter will never be matched. (That is assuming that no Host header was available to disambiguate the two.)
  • If a vhost that would otherwise be a name-vhost includes a Port statement that doesn't match the main_server Port then it will be considered an IP-based vhost. Then find_virtual_server will match it (because the ports associated with each address in the address set default to the port of the main_server) as the connection vhost. Then check_hostalias will refuse to check any other name-based vhost because of the port mismatch. The result is that the vhost will steal all hits going to the main_server address.
  • If two IP-based vhosts have an address in common, the vhost appearing later in the file is always matched. Such a thing might happen inadvertently. If the config has name-based vhosts and for some reason the main_server ServerName resolves to the wrong address then all the name-based vhosts will be parsed as ip-based vhosts. Then the last of them will steal all the hits.
  • The last name-based vhost in the config is always matched for any hit which doesn't match one of the other name-based vhosts.

What Works

In addition to the tips on the DNS Issues page, here are some further tips:

  • Place all main_server definitions before any VirtualHost definitions. (This is to aid the readability of the configuration -- the post-config merging process makes it non-obvious that definitions mixed in around virtualhosts might affect all virtualhosts.)
  • Arrange your VirtualHosts such that all name-based virtual hosts come first, followed by IP-based virtual hosts, followed by any _default_ virtual host
  • Avoid ServerPaths which are prefixes of other ServerPaths. If you cannot avoid this then you have to ensure that the longer (more specific) prefix vhost appears earlier in the configuration file than the shorter (less specific) prefix (i.e., "ServerPath /abc" should appear after "ServerPath /abcdef").
  • Do not use port-based vhosts in the same server as name-based vhosts. A loose definition for port-based is a vhost which is determined by the port on the server (i.e., one server with ports 8000, 8080, and 80 - all of which have different configurations).

Apache HTTP Server Version 1.3

Index Home