httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Ralf S. Engelschall" <>
Subject Re: [PATCH] Extended API (EAPI)
Date Thu, 15 Apr 1999 06:22:16 GMT

In article <> you wrote:
> Time has come... [EAPI]

Two additional notes I've forgotten:

 o There is a README.EAPI in the tarball which
   describes the intentions and ideas of EAPI together with some examples. So
   when you're impatient and don't want to read the patch itself, read this
   document. I append you a version to this mail, too.

 o When you want to review the patch itself, don't read the large
   apache-1.3-eapi.patch. Instead read the eapi.patch file in the tarball,
   because it contains _annotations_ for each patch snippet and describes what
   is actually patched and why.

                                       Ralf S. Engelschall

 Extended API (EAPI)

 What is EAPI

 Extended API (EAPI) is a comprehensive API addition which can be _OPTIONALLY_
 enabled with ``Rule EAPI=yes'' in src/Configuration or ``--enable-rule=EAPI''
 on the APACI configure command line. This then defines a -DEAPI and this way
 the EAPI code is compiled into Apache. When this define is not present _NO_
 EAPI code is compiled into Apache at all, because all(!) EAPI patches are
 encapsulated in #ifdef EAPI...#endif.

 What is provided by EAPI?

 EAPI's additions to the Apache API fall into the following categories:

    o  Context Attachment Support for Data Structures
    o  Loosly-coupled Hook Interface for Inter-Module Communication
    o  Direct and Pool-based Shared Memory Support 
    o  Additional Apache Module Hooks
    o  Specialized EAPI Goodies

 They are discussed in details now....

 Context Attachment Support for Data Structures

 Attaching private information to a request_rec, conn_rec, server_rec or even
 BUFF structure is for a lot of modules the most elegant solution to keep
 states between API phases without the need for any global variables. That's
 especially true for modules which operate on lower I/O levels (where no
 per-module configuration structure is available) or have to deal with various
 callback functions of third-party libraries (where one need to find the
 private context which can be hard without global variables).

 The EAPI way to solve this situation is: 

 1. A generic context library was written which allows one
    to create a context and later store and retrieve context variables
    identified by a unique key.

 2. The Apache kernel was extended to provide contexts for all standard data
    structures like request_rec, server_rec, conn_rec, BUFF, etc.  This way
    modules can easily attach information to all these structures with the
    help of the context API.

 Point 1 is implemented by new src/ap/ap_ctx.c and src/include/ap_ctx.h source
 files.  Point 2 is implemented by EAPI patches to various src/main/*.c and
 src/include/*.h files.

  | /* a module implements on-the-fly compression for
  |    the buffer code and for this uses a third-party library which 
  |    don't uses a filedescriptor. Instead a CLIB* is used.  The module has to
  |    attach this CLIB* to the BUFF in oder to have it available whenever a
  |    BUFF is used somewhere. */
  | BUFF *buff;
  | CLIB *comp;
  | comp = CLIB_new_from_fd(buff->fd);
  | ap_ctx_set(buff->ctx, "CLIB", comp);
  |   :
  | /* later when it deals with a BUFF, it can easily find back the
  |    CLIB* via the BUFF* */
  | comp = (CLIB *)ap_ctx_get(buff->ctx, "CLIB");
  |   :
 Possible use cases from practice are:
  o  attaching third-party structures to Apache structures
  o  replacing global module variables with clean context variables
  o  custom attachments for complex modules like mod_php, mod_php, etc.
  o  companion support for the hook interface (see below)
  o  etc. pp.

 Loosly-coupled Hook Interface for Inter-Module Communication

 Apache is structured into modules which is a nice idea.  With the Dynamic
 Shared Object (DSO) facility it gets even nicer because then modules are then
 really stand-alone objects. The drawback is that DSO restricts modules.  The
 most popular problem is that no inter-module symbol references are allowed.
 The classical problem: Module A implements some nice functions module B would
 like to use to avoid reimplementing the wheel. But B cannot call A's
 functions because this violates both the design idea of stand-alone modules
 and the DSO restrictions. Additionally a module C could exists which also
 provides a variant of the functionality of A's function.  Then B should get
 the variant (either A's or C's) which is best or available at all.
 Real Life Example:

 mod_rewrite provides %{XXXX} constructs to lookup variables.  The available
 variables are (and have to be) hard-coded into mod_rewrite. Now our mod_clib
 which does on-the-fly compression provides a variable CLIB_FACTOR which gives
 information about the shrink factor of the compression and a user wants to
 use this shrink factor to make an URL-rewriting decision (<grin>). No chance
 without EAPI.  With EAPI it's easy: Inside the if-cascade for the various
 variables in mod_rewrite one replaces:

  | char *result;
  | request_rec *r;
  |    :
  | if (strcasecmp(var, "...") == 0) {
  |    :
  | else if (strcasecmp(var, "SCRIPT_GROUP") == 0) {
  |     result = ...
  | }
  | else {
  |     if (result == NULL) {
  |         ...complain...
  |     }
  | }
  |    :


  | char *result;
  | request_rec *r;
  |    :
  | if (strcasecmp(var, "...") == 0) {
  |    :
  | else if (strcasecmp(var, "SCRIPT_GROUP") == 0) {
  |     result = ...
  | }
  | else {
  |     ap_hook_use("ap::lookup_variable",
  |                 AP_HOOK_SIG4(ptr,ptr,ptr,ctx), 
  |                 AP_HOOK_DECLINE(NULL),
  |                 &result, r, var);
  |     if (result == NULL) {
  |         ...complain...
  |     }
  | }
  |    :

 What this does is that when XXXX of %{XXXX} isn't known, a hook named
 ap::lookup_variable is called with the request_rec and the var ("XXX") and
 the result variable. When no one has registered for this hook, nothing
 happens. ap_hook_use() immediately returns and nothing was changed. 

 But now let's assume mod_clib is additionally loaded as a DSO. And without
 changing anything now magically mod_rewrite implements %{CLIB_FACTOR}. How?
 Look inside mod_clib.c:

  | /* mod_clib registeres for the ap::lookup_variable hook 
  | inside it's init phase */
  | CLIB *comp;
  | ap_hook_register("ap::lookup_variable", 
  |                  my_lookup_variable, AP_HOOK_CTX(comp));
  | /* and implements the my_lookup_variable() function */
  | char *my_lookup_variable(request_rec *r, char *name, CLIB *comp)
  | {
  |     if (strcmp(name, "CLIB_FACTOR") == 0)
  |         return ap_psrintf(r->pool, "%d", comp->factor);
  |     return NULL;
  | }

 What happens? When mod_rewrite calls the ap_hook_use() function internally
 the hook facility knows that mod_clib has registered for this hook and calls
 the equivalent of

 |     result = my_lookup_variable(r, var, <comp>);

 where <comp> is the CLIB* context variable mod_clib has registered for
 itself. Now assume a second module exists which also provides variables and
 want to allow mod_rewrite to lookup them.  It registers after mod_clib with

 |      ap_hook_register("ap::lookup_variable", 
 |                        my_lookup_variable2, AP_HOOK_CTX(whatever));

 and then the following happens: The hook facility does for mod_rewrite the
 equivalent of:

 |      result = my_lookup_variable(r, var, <comp>);
 |      if (result == NULL)
 |          result = my_lookup_variable2(r, var, <whatever>);

 As you can see the hook functions decline in this example with NULL.  That's
 the NULL from AP_HOOK_DECLINE(NULL) and can be any value of any type, of

 The same idea can be also used by mod_log_config and every other module which
 wants to lookup a variable inside Apache. Which variables are available
 depend on the available modules which implement them. And this all works
 nicely with the DSO facility, because the ap_hook_xxx() API is part of the
 Apache kernel code. And nothing has to be changed inside Apache when another
 modules wants to create a new hook, because the mechanism is totally generic.

 So when our module A wants to let other modules to use it's function it just
 has to configure a hook for this.  Then other modules call this hook. Is
 module A not there the boolean return value of the hook call will indicate
 this. When module A is there the function is called.

 Direct and Pool-based Shared Memory Support

 Since years it was annoying that Apache's pre-forked process model basically
 means that every server lives it's own life (= address space) and this way
 module authors cannot easily spread module configuration or other data
 accross the processes.  The most elegant solution is to use shared memory
 segments.  The drawback is that there is no portable API for shared memory
 handling and there is no convinient memory allocation API for working inside
 shared memory segments.

 The EAPI way to solve this situation is: 

 1. A stand-alone and resuable library was written (named MM from "memory
    mapped" and available from which
    abstracts the shared memory and memory mutex fiddling into a low-level
    API.  Internally the shared memory and mutex functionality is implemented
    in various platform-depended ways: 4.4BSD or POSIX.1 anonymous memory
    mapping, /dev/zero-based memory mapping, temporary file memory mapping, or
    SysV IPC shared memory for allocating the shared memory areas and POSIX.1
    fcntl(2), BSD flock(2) or SysV IPC semaphores for implementing mutual
    exclusion capabilities.

    Additionally MM provides a high-level malloc()-style API based on this
    abstracted shared memory low-level API. The idea is just to allocate the
    requested memory chunks from shared memory segments instead of the heap.

 2. EAPI now provides an easy method (with the EAPI_MM configuration 
    variable) to build Apache against this MM library. For this the whole MM
    API (mm_xxx() functions) is encapsulated in an Apache API subpart
    (ap_mm_xxx() functions). This way the API is fixed and always present (no
    #ifdef EAPI stuff in modules!), but useable only when EAPI was used in
    conjunction with MM. A simple ``EAPI_MM=/path/to/mm ./configure
    --enable-rule=EAPI ...'' is enough to put MM under the ap_mm_xxx() API.
    This way modules can use a consistent, powerful and abstracted ap_mm_xxx()
    API for dealing with shared memory.

 3. Because inside Apache mostly all memory handling is done via the
    pool facility, additional support for ``shared memory pools'' is provided.
    This way modules can use all ap_pxxx() functions in combination with
    shared memory.

 Point 1 is implemented inside the MM package. Point 2 is implemented by the
 new src/ap/ap_mm.c and src/include/ap_mm.h source files.  Point 3 is
 implemented by EAPI patches to src/main/alloc.c and src/include/alloc.h.


 | /* inside a module init function (before the forking!)
 |    for instance a module allocates a structure with a counter
 |    in a shared memory segment */
 | pool *p;
 | pool *sp;
 | struct mystuff { int cnt } *my;
 | sp = ap_make_shared_sub_pool(p);
 | my = (struct mystuff *)ap_palloc(sp, sizeof(struct mystuff));
 | my->cnt = 0;
 |     :
 | /* then under request processing time it's changed by one process */
 | ap_acquire_pool(sp, AP_POOL_RW);
 | my->cnt++;
 | ap_release_pool(sp);
 |     :
 | /* and at the same time read by other processes */
 | ap_acquire_pool(sp, AP_POOL_RD);
 | ap_rprintf(r, "The counter is %d\n", my->cnt);
 | ap_release_pool(sp);

 Possible use cases from practice are:

  o  assembling traffic or other accounting details
  o  establishing of high-performance inter-process caches
  o  inter-process wide keeping of session state information
  o  shared memory support for mod_perl, mod_php, etc.
  o  etc. pp.

 Additional Apache Module Hooks

 The above three EAPI additions are all very generic facilities.  But there
 were also specialized things which were missing in Apache (and needed by
 modules). Mostly additional API phases. EAPI adds the following additional
 hook pointers to the module structure:

     Called from within ap_add_module() right after the module structure
     was linked into the Apache internal module list.  It is mainly
     intended to be used to define configuration defines (<IfDefine>)
     which have to be available directly after a LoadModule/AddModule.
     Actually this is the earliest possible hook a module can use.  It's
     especially important for the modules when they use the hook facility.

     Called from within ap_remove_module() right before the module
     structure is kicked out from the Apache internal module list.
     Actually this is last possible hook a module can use and exists for
     consistency with the add_module hook.

     Called right after a configuration directive line was read and
     before it is processed. It is mainly intended to be used for
     rewriting directives in order to provide backward compatibility to
     old directive variants.

     Called from within the internal new_connection() function, right
     after the conn_rec structure for the new established connection was
     created and before Apache starts processing the request with
     ap_read_request().  It is mainly intended to be used to setup/run
     connection dependent things like sending start headers for
     on-the-fly compression, etc.

     Called from within the Apache dispatching loop just before any
     ap_bclose() is performed on the socket connection, but a long time
     before any pool cleanups are done for the connection (which can be
     too late for some applications).  It is mainly intended to be used
     to close/finalize connection dependent things like sending end
     headers for on-the-fly compression, etc.

 Specialized EAPI Goodies

 And finally EAPI now uses some of the new functionality to add a few new
 EAPI-based goodies to mod_rewrite, mod_status and mod_proxy:

     The above presented example of lookup hooks is implemented which allows
     mod_rewrite to lookup arbitrary variables provides by not known modules.

     Any module now can register to an EAPI hook of mod_status which
     allows it to put additional text on the /status webpages.

     Some EAPI hooks are provided to allow other modules to control the HTTP
     client processing inside mod_proxy.  This can be used for a lot of

View raw message