Apache breaks down request handling into a series of steps, more or less the same way the Netscape server API does (although this API has a few more stages than NetSite does, as hooks for stuff I thought might be useful in the future). These are:

• URI -> Filename translation
• Auth ID checking [is the user who they say they are?]
• Auth access checking [is the user authorized here?]
• Access checking other than auth
• Determining MIME type of the object requested
• `Fixups' -- there aren't any of these yet, but the phase is intended as a hook for possible extensions like SetEnv, which don't really fit well elsewhere.
• Actually sending a response back to the client.
• Logging the request

These phases are handled by looking at each of a succession of modules, looking to see if each of them has a handler for the phase, and attempting invoking it if so. The handler can typically do one of three things:

• Handle the request, and indicate that it has done so by returning the magic constant OK.
• Decline to handle the request, by returning the magic integer constant DECLINED. In this case, the server behaves in all respects as if the handler simply hadn't been there.
• Signal an error, by returning one of the HTTP error codes. This terminates normal handling of the request, although an ErrorDocument may be invoked to try to mop up, and it will be logged in any case.

Most phases are terminated by the first module that handles them; however, for logging, `fixups', and non-access authentication checking, all handlers always run (barring an error). Also, the response phase is unique in that modules may declare multiple handlers for it, via a dispatch table keyed on the MIME type of the requested object. Modules may declare a response-phase handler which can handle any request, by giving it the key */* (i.e., a wildcard MIME type specification). However, wildcard handlers are only invoked if the server has already tried and failed to find a more specific response handler for the MIME type of the requested object (either none existed, or they all declined).

The handlers themselves are functions of one argument (a request_rec structure. vide infra), which returns an integer, as above.

At this point, we need to explain the structure of a module. Our candidate will be one of the messier ones, the CGI module -- this handles both CGI scripts and the ScriptAlias config file command. It's actually a great deal more complicated than most modules, but if we're going to have only one example, it might as well be the one with its fingers in every place.

Let's begin with handlers. In order to handle the CGI scripts, the module declares a response handler for them. Because of ScriptAlias, it also has handlers for the name translation phase (to recognize ScriptAliased URIs), the type-checking phase (any ScriptAliased request is typed as a CGI script).

The module needs to maintain some per (virtual) server information, namely, the ScriptAliases in effect; the module structure therefore contains pointers to a functions which builds these structures, and to another which combines two of them (in case the main server and a virtual server both have ScriptAliases declared).

Finally, this module contains code to handle the ScriptAlias command itself. This particular module only declares one command, but there could be more, so modules have command tables which declare their commands, and describe where they are permitted, and how they are to be invoked.

A final note on the declared types of the arguments of some of these commands: a pool is a pointer to a resource pool structure; these are used by the server to keep track of the memory which has been allocated, files opened, etc., either to service a particular request, or to handle the process of configuring itself. That way, when the request is over (or, for the configuration pool, when the server is restarting), the memory can be freed, and the files closed, en masse, without anyone having to write explicit code to track them all down and dispose of them. Also, a cmd_parms structure contains various information about the config file being read, and other status information, which is sometimes of use to the function which processes a config-file command (such as ScriptAlias). With no further ado, the module itself:

  STANDARD_MODULE_STUFF,
  NULL,                     /* initializer */
  NULL,                     /* dir config creator */
  NULL,                     /* dir merger */
  make_cgi_server_config,   /* server config */
  merge_cgi_server_config,  /* merge server config */
  cgi_cmds,                 /* command table */
  cgi_handlers,             /* handlers */
  translate_scriptalias,    /* filename translation */
  NULL,                     /* check_user_id */
  NULL,                     /* check auth */
  NULL,                     /* check access */
  type_scriptalias,         /* type_checker */
  NULL,                     /* fixups */
  NULL,                     /* logger */
  NULL                      /* header parser */
};

The request_rec contains pointers to a resource pool which will be cleared when the server is finished handling the request; to structures containing per-server and per-connection information, and most importantly, information on the request itself.

The most important such information is a small set of character strings describing attributes of the object being requested, including its URI, filename, content-type and content-encoding (these being filled in by the translation and type-check handlers which handle the request, respectively).

Other commonly used data items are tables giving the MIME headers on the client's original request, MIME headers to be sent back with the response (which modules can add to at will), and environment variables for any subprocesses which are spawned off in the course of servicing the request. These tables are manipulated using the ap_table_get and ap_table_set routines.

Finally, there are pointers to two data structures which, in turn, point to per-module configuration structures. Specifically, these hold pointers to the data structures which the module has built to describe the way it has been configured to operate in a given directory (via .htaccess files or Directory sections), for private data it has built in the course of servicing the request (so modules' handlers for one phase can pass `notes' to their handlers for other phases). There is another such configuration vector in the server_rec data structure pointed to by the request_rec, which contains per (virtual) server configuration data.

Here is an abridged declaration, giving the fields most commonly used:

char *args;           /* QUERY_ARGS, if any */
struct stat finfo;    /* Set by server core;
                       * st_mode set to zero if no such file */

int header_only;     /* HEAD request, as opposed to GET */
char *protocol;      /* Protocol, as given to us, or HTTP/0.9 */
char *method;        /* GET, HEAD, POST, etc. */
int method_number;   /* M_GET, M_POST, etc. */

void *per_dir_config;   /* Options set in config files, etc. */
void *request_config;   /* Notes on *this* request */

Most request_rec structures are built by reading an HTTP request from a client, and filling in the fields. However, there are a few exceptions:

• If the request is to an imagemap, a type map (i.e., a *.var file), or a CGI script which returned a local `Location:', then the resource which the user requested is going to be ultimately located by some URI other than what the client originally supplied. In this case, the server does an internal redirect, constructing a new request_rec for the new URI, and processing it almost exactly as if the client had requested the new URI directly.
• If some handler signaled an error, and an ErrorDocument is in scope, the same internal redirect machinery comes into play.

• Finally, a handler occasionally needs to investigate `what would happen if' some other request were run. For instance, the directory indexing module needs to know what MIME type would be assigned to a request for each directory entry, in order to figure out what icon to use.

  Such handlers can construct a sub-request, using the functions ap_sub_req_lookup_file, ap_sub_req_lookup_uri, and ap_sub_req_method_uri; these construct a new request_rec structure and processes it as you would expect, up to but not including the point of actually sending a response. (These functions skip over the access checks if the sub-request is for a file in the same directory as the original request).

  (Server-side includes work by building sub-requests and then actually invoking the response handler for them, via the function ap_run_sub_req).

As discussed above, each handler, when invoked to handle a particular request_rec, has to return an int to indicate what happened. That can either be

• OK -- the request was handled successfully. This may or may not terminate the phase.
• DECLINED -- no erroneous condition exists, but the module declines to handle the phase; the server tries to find another.
• an HTTP error code, which aborts handling of the request.

Note that if the error code returned is REDIRECT, then the module should put a Location in the request's headers_out, to indicate where the client should be redirected to.

Handlers for most phases do their work by simply setting a few fields in the request_rec structure (or, in the case of access checkers, simply by returning the correct error code). However, response handlers have to actually send a request back to the client.

They should begin by sending an HTTP response header, using the function ap_send_http_header. (You don't have to do anything special to skip sending the header for HTTP/0.9 requests; the function figures out on its own that it shouldn't do anything). If the request is marked header_only, that's all they should do; they should return after that, without attempting any further output.

Otherwise, they should produce a request body which responds to the client as appropriate. The primitives for this are ap_rputc and ap_rprintf, for internally generated output, and ap_send_fd, to copy the contents of some FILE * straight to the client.

At this point, you should more or less understand the following piece of code, which is the handler which handles GET requests which have no more specific handler; it also shows how conditional GETs can be handled, if it's desirable to do so in a particular response handler -- ap_set_last_modified checks against the If-modified-since value supplied by the client, if any, and returns an appropriate code (which will, if nonzero, be USE_LOCAL_COPY). No similar considerations apply for ap_set_content_length, but it returns an error code for symmetry.

Finally, if all of this is too much of a challenge, there are a few ways out of it. First off, as shown above, a response handler which has not yet produced any output can simply return an error code, in which case the server will automatically produce an error response. Secondly, it can punt to some other handler by invoking ap_internal_redirect, which is how the internal redirection machinery discussed above is invoked. A response handler which has internally redirected should always return OK.

(Invoking ap_internal_redirect from handlers which are not response handlers will lead to serious confusion).

Stuff that should be discussed here in detail:

• Authentication-phase handlers not invoked unless auth is configured for the directory.
• Common auth configuration stored in the core per-dir configuration; it has accessors ap_auth_type, ap_auth_name, and ap_requires.
• Common routines, to handle the protocol end of things, at least for HTTP basic authentication (ap_get_basic_auth_pw, which sets the connection->user structure field automatically, and ap_note_basic_auth_failure, which arranges for the proper WWW-Authenticate: header to be sent back).

When a request has internally redirected, there is the question of what to log. Apache handles this by bundling the entire chain of redirects into a list of request_rec structures which are threaded through the r->prev and r->next pointers. The request_rec which is passed to the logging handlers in such cases is the one which was originally built for the initial request from the client; note that the bytes_sent field will only be correct in the last request in the chain (the one for which a response was actually sent).

Memory is allocated to pools by calling the function ap_palloc, which takes two arguments, one being a pointer to a resource pool structure, and the other being the amount of memory to allocate (in chars). Within handlers for handling requests, the most common way of getting a resource pool structure is by looking at the pool slot of the relevant request_rec; hence the repeated appearance of the following idiom in module code:

Note that there is no ap_pfree -- ap_palloced memory is freed only when the associated resource pool is cleared. This means that ap_palloc does not have to do as much accounting as malloc(); all it does in the typical case is to round up the size, bump a pointer, and do a range check.

(It also raises the possibility that heavy use of ap_palloc could cause a server process to grow excessively large. There are two ways to deal with this, which are dealt with below; briefly, you can use malloc, and try to be sure that all of the memory gets explicitly freed, or you can allocate a sub-pool of the main pool, allocate your memory in the sub-pool, and clear it out periodically. The latter technique is discussed in the section on sub-pools below, and is used in the directory-indexing code, in order to avoid excessive storage allocation when listing directories with thousands of files).

There are functions which allocate initialized memory, and are frequently useful. The function ap_pcalloc has the same interface as ap_palloc, but clears out the memory it allocates before it returns it. The function ap_pstrdup takes a resource pool and a char * as arguments, and allocates memory for a copy of the string the pointer points to, returning a pointer to the copy. Finally ap_pstrcat is a varargs-style function, which takes a pointer to a resource pool, and at least two char * arguments, the last of which must be NULL. It allocates enough memory to fit copies of each of the strings, as a unit; for instance:

returns a pointer to 8 bytes worth of memory, initialized to "foo/bar".

A pool is really defined by its lifetime more than anything else. There are some static pools in http_main which are passed to various non-http_main functions as arguments at opportune times. Here they are:

permanent_pool

never passed to anything else, this is the ancestor of all pools

pconf

• subpool of permanent_pool
• created at the beginning of a config "cycle"; exists until the server is terminated or restarts; passed to all config-time routines, either via cmd->pool, or as the "pool *p" argument on those which don't take pools
• passed to the module init() functions

ptemp

• sorry I lie, this pool isn't called this currently in 1.3, I renamed it this in my pthreads development. I'm referring to the use of ptrans in the parent... contrast this with the later definition of ptrans in the child.
• subpool of permanent_pool
• created at the beginning of a config "cycle"; exists until the end of config parsing; passed to config-time routines via cmd->temp_pool. Somewhat of a "bastard child" because it isn't available everywhere. Used for temporary scratch space which may be needed by some config routines but which is deleted at the end of config.

pchild

• subpool of permanent_pool
• created when a child is spawned (or a thread is created); lives until that child (thread) is destroyed
• passed to the module child_init functions
• destruction happens right after the child_exit functions are called... (which may explain why I think child_exit is redundant and unneeded)

ptrans

• should be a subpool of pchild, but currently is a subpool of permanent_pool, see above
• cleared by the child before going into the accept() loop to receive a connection
• used as connection->pool

r->pool

• for the main request this is a subpool of connection->pool; for subrequests it is a subpool of the parent request's pool.
• exists until the end of the request (i.e., ap_destroy_sub_req, or in child_main after process_request has finished)
• note that r itself is allocated from r->pool; i.e., r->pool is first created and then r is the first thing palloc()d from it

For almost everything folks do, r->pool is the pool to use. But you can see how other lifetimes, such as pchild, are useful to some modules... such as modules that need to open a database connection once per child, and wish to clean it up when the child dies.

You can also see how some bugs have manifested themself, such as setting connection->user to a value from r->pool -- in this case connection exists for the lifetime of ptrans, which is longer than r->pool (especially if r->pool is a subrequest!). So the correct thing to do is to allocate from connection->pool.

And there was another interesting bug in mod_include / mod_cgi. You'll see in those that they do this test to decide if they should use r->pool or r->main->pool. In this case the resource that they are registering for cleanup is a child process. If it were registered in r->pool, then the code would wait() for the child when the subrequest finishes. With mod_include this could be any old #include, and the delay can be up to 3 seconds... and happened quite frequently. Instead the subprocess is registered in r->main->pool which causes it to be cleaned up when the entire request is done -- i.e., after the output has been sent to the client and logging has happened.

As indicated above, resource pools are also used to track other sorts of resources besides memory. The most common are open files. The routine which is typically used for this is ap_pfopen, which takes a resource pool and two strings as arguments; the strings are the same as the typical arguments to fopen, e.g.,

There is also a ap_popenf routine, which parallels the lower-level open system call. Both of these routines arrange for the file to be closed when the resource pool in question is cleared.

Unlike the case for memory, there are functions to close files allocated with ap_pfopen, and ap_popenf, namely ap_pfclose and ap_pclosef. (This is because, on many systems, the number of files which a single process can have open is quite limited). It is important to use these functions to close files allocated with ap_pfopen and ap_popenf, since to do otherwise could cause fatal errors on systems such as Linux, which react badly if the same FILE* is closed more than once.

(Using the close functions is not mandatory, since the file will eventually be closed regardless, but you should consider it in cases where your module is opening, or could open, a lot of files).

More text goes here. Describe the the cleanup primitives in terms of which the file stuff is implemented; also, spawn_process.

Pool cleanups live until clear_pool() is called: clear_pool(a) recursively calls destroy_pool() on all subpools of a; then calls all the cleanups for a; then releases all the memory for a. destroy_pool(a) calls clear_pool(a) and then releases the pool structure itself. i.e., clear_pool(a) doesn't delete a, it just frees up all the resources and you can start using it again immediately.

On rare occasions, too-free use of ap_palloc() and the associated primitives may result in undesirably profligate resource allocation. You can deal with such a case by creating a sub-pool, allocating within the sub-pool rather than the main pool, and clearing or destroying the sub-pool, which releases the resources which were associated with it. (This really is a rare situation; the only case in which it comes up in the standard module set is in case of listing directories, and then only with very large directories. Unnecessary use of the primitives discussed here can hair up your code quite a bit, with very little gain).

The primitive for creating a sub-pool is ap_make_sub_pool, which takes another pool (the parent pool) as an argument. When the main pool is cleared, the sub-pool will be destroyed. The sub-pool may also be cleared or destroyed at any time, by calling the functions ap_clear_pool and ap_destroy_pool, respectively. (The difference is that ap_clear_pool frees resources associated with the pool, while ap_destroy_pool also deallocates the pool itself. In the former case, you can allocate new resources within the pool, and clear it again, and so forth; in the latter case, it is simply gone).

One final note -- sub-requests have their own resource pools, which are sub-pools of the resource pool for the main request. The polite way to reclaim the resources associated with a sub request which you have allocated (using the ap_sub_req_... functions) is ap_destroy_sub_req, which frees the resource pool. Before calling this function, be sure to copy anything that you care about which might be allocated in the sub-request's resource pool into someplace a little less volatile (for instance, the filename in its request_rec structure).

(Again, under most circumstances, you shouldn't feel obliged to call this function; only 2K of memory or so are allocated for a typical sub request, and it will be freed anyway when the main request pool is cleared. It is only when you are allocating many, many sub-requests for a single main request that you should seriously consider the ap_destroy_... functions).

Let's look out how all of this plays out in mod_mime.c, which defines the file typing handler which emulates the NCSA server's behavior of determining file types from suffixes. What we'll be looking at, here, is the code which implements the AddType and AddEncoding commands. These commands can appear in .htaccess files, so they must be handled in the module's private per-directory data, which in fact, consists of two separate tables for MIME types and encoding information, and is declared as follows:

typedef struct {
    table *forced_types;      /* Additional AddTyped stuff */
    table *encoding_types;    /* Added with AddEncoding... */
} mime_dir_config;

When the server is reading a configuration file, or Directory section, which includes one of the MIME module's commands, it needs to create a mime_dir_config structure, so those commands have something to act on. It does this by invoking the function it finds in the module's `create per-dir config slot', with two arguments: the name of the directory to which this configuration information applies (or NULL for srm.conf), and a pointer to a resource pool in which the allocation should happen.

(If we are reading a .htaccess file, that resource pool is the per-request resource pool for the request; otherwise it is a resource pool which is used for configuration data, and cleared on restarts. Either way, it is important for the structure being created to vanish when the pool is cleared, by registering a cleanup on the pool if necessary).

For the MIME module, the per-dir config creation function just ap_pallocs the structure above, and a creates a couple of tables to fill it. That looks like this:

Now, suppose we've just read in a .htaccess file. We already have the per-directory configuration structure for the next directory up in the hierarchy. If the .htaccess file we just read in didn't have any AddType or AddEncoding commands, its per-directory config structure for the MIME module is still valid, and we can just use it. Otherwise, we need to merge the two structures somehow.

To do that, the server invokes the module's per-directory config merge function, if one is present. That function takes three arguments: the two structures being merged, and a resource pool in which to allocate the result. For the MIME module, all that needs to be done is overlay the tables from the new per-directory config structure with those from the parent:

As a note -- if there is no per-directory merge function present, the server will just use the subdirectory's configuration info, and ignore the parent's. For some modules, that works just fine (e.g., for the includes module, whose per-directory configuration information consists solely of the state of the XBITHACK), and for those modules, you can just not declare one, and leave the corresponding structure slot in the module itself NULL.

Now that we have these structures, we need to be able to figure out how to fill them. That involves processing the actual AddType and AddEncoding commands. To find commands, the server looks in the module's command table. That table contains information on how many arguments the commands take, and in what formats, where it is permitted, and so forth. That information is sufficient to allow the server to invoke most command-handling functions with pre-parsed arguments. Without further ado, let's look at the AddType command handler, which looks like this (the AddEncoding command looks basically the same, and won't be shown here):

This command handler is unusually simple. As you can see, it takes four arguments, two of which are pre-parsed arguments, the third being the per-directory configuration structure for the module in question, and the fourth being a pointer to a cmd_parms structure. That structure contains a bunch of arguments which are frequently of use to some, but not all, commands, including a resource pool (from which memory can be allocated, and to which cleanups should be tied), and the (virtual) server being configured, from which the module's per-server configuration data can be obtained if required.

Another way in which this particular command handler is unusually simple is that there are no error conditions which it can encounter. If there were, it could return an error message instead of NULL; this causes an error to be printed out on the server's stderr, followed by a quick exit, if it is in the main config files; for a .htaccess file, the syntax error is logged in the server error log (along with an indication of where it came from), and the request is bounced with a server error response (HTTP error status, code 500).

The MIME module's command table has entries for these commands, which look like this:

The entries in these tables are:

• The name of the command
• The function which handles it
• a (void *) pointer, which is passed in the cmd_parms structure to the command handler --- this is useful in case many similar commands are handled by the same function.
• A bit mask indicating where the command may appear. There are mask bits corresponding to each AllowOverride option, and an additional mask bit, RSRC_CONF, indicating that the command may appear in the server's own config files, but not in any .htaccess file.
• A flag indicating how many arguments the command handler wants pre-parsed, and how they should be passed in. TAKE2 indicates two pre-parsed arguments. Other options are TAKE1, which indicates one pre-parsed argument, FLAG, which indicates that the argument should be On or Off, and is passed in as a boolean flag, RAW_ARGS, which causes the server to give the command the raw, unparsed arguments (everything but the command name itself). There is also ITERATE, which means that the handler looks the same as TAKE1, but that if multiple arguments are present, it should be called multiple times, and finally ITERATE2, which indicates that the command handler looks like a TAKE2, but if more arguments are present, then it should be called multiple times, holding the first argument constant.
• Finally, we have a string which describes the arguments that should be present. If the arguments in the actual config file are not as required, this string will be used to help give a more specific error message. (You can safely leave this NULL).

Finally, having set this all up, we have to use it. This is ultimately done in the module's handlers, specifically for its file-typing handler, which looks more or less like this; note that the per-directory configuration structure is extracted from the request_rec's per-directory configuration vector by using the ap_get_module_config function.

The basic ideas behind per-server module configuration are basically the same as those for per-directory configuration; there is a creation function and a merge function, the latter being invoked where a virtual server has partially overridden the base server configuration, and a combined structure must be computed. (As with per-directory configuration, the default if no merge function is specified, and a module is configured in some virtual server, is that the base configuration is simply ignored).

The only substantial difference is that when a command needs to configure the per-server private module data, it needs to go to the cmd_parms data to get at it. Here's an example, from the alias module, which also indicates how a syntax error can be returned (note that the per-directory configuration argument to the command handler is declared as a dummy, since the module doesn't actually have per-directory config data):