httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Rodent of Unusual Size <>
Subject [STATUS] (httpd-trunk) Wed Nov 15 23:49:15 2006
Date Thu, 16 Nov 2006 04:49:15 GMT
APACHE 2.3 STATUS:                                              -*-text-*-
Last modified at [$Date: 2006-08-22 16:41:03 -0400 (Tue, 22 Aug 2006) $]

The current version of this file can be found at:


Documentation status is maintained seperately and can be found at:

  * docs/STATUS in this source tree, or

Consult the following STATUS files for information on related projects:


Patches considered for backport are noted in their branches' STATUS:


Release history:
    [NOTE that x.{odd}.z versions are strictly Alpha/Beta releases,
          while x.{even}.z versions are Stable/GA releases.]

    2.3.0   : in development

Contributors looking for a mission:

    * Just do an egrep on "TODO" or "XXX" in the source.

    * Review the bug database at:

    * Review the "PatchAvailable" bugs in the bug database:

      After testing, you can append a comment saying "Reviewed and tested".

    * Open bugs in the bug database.



    * Handling of non-trailing / config by non-default handler is broken
      jerenkrantz asks: Why should this block a release?
      wsanchez agrees: this may be a change in behavior, but isn't
        clearly wrong, and even if so, it doesn't seem like a

    * the edge connection filter cannot be removed

      jerenkrantz asks: Why should this block a release?

      stas replies: because it requires a rewrite of the filters stack
            implementation (you have suggested that) and once 2.2 is
            released you can't do that anymore. 


    * If the parent process dies, should the remaining child processes
      "gracefully" self-terminate. Or maybe we should make it a runtime
      option, or have a concept of 2 parent processes (one being a 
      "hot spare").
      See: Message-ID: <3C58232C.FE91F19F@Golux.Com>

      Self-destruct: Ken, Martin, Lars
      Not self-destruct: BrianP, Ian, Cliff, BillS
      Make it runtime configurable: Aaron, jim, Justin, wrowe, rederpj, nd

      /* The below was a concept on *how* to handle the problem */
      Have 2 parents: +1: jim
                      -1: Justin, wrowe, rederpj, nd
                      +0: Lars, Martin (while standing by, could it do
                                        something useful?)

    * Make the worker MPM the default MPM for threaded Unix boxes.
      +1:   Justin, Ian, Cliff, BillS, striker, wrowe, nd
      +0:   BrianP, Aaron (mutex contention is looking better with the
            latest code, let's continue tuning and testing), rederpj, jim
      -0:   Lars

      pquerna: Do we want to change this for 2.2?


    * Patches submitted to the bug database:

    * Filter stacks and subrequests, redirects and fast redirects.
      There's at least one PR that suffers from the current unclean behaviour
      (which lets the server send garbage): PR 17629
      nd says: Every subrequest should get its own filter stack with the
               subreq_core filter as bottom-most. That filter does two things:
                 - swallow EOS buckets
                 - redirect the data stream to the upper request's (rr->main)
                   filter chain directly after the subrequest's starting
               Once we have a clean solution, we can try to optimize
               it, so that the server won't be slow down too much.

    * RFC 2616 violations.
      Closed PRs: 15857.
      Open PRs: 15852, 15859, 15861, 15864, 15865, 15866, 15868, 15869,
                15870, 16120, 16125, 16126, 16133, 16135, 16136, 16137,
                16138, 16139, 16140, 16142, 16518, 16520, 16521, 
      jerenkrantz says: need to decide how many we need to backport and/or
                        if these rise to showstopper status.
      wrowe suggests: it would be nice to see "MUST" v.s. "SHOULD" v.s. "MAY"
                      out of this list, without reviewing them individually.

    * There is a bug in how we sort some hooks, at least the pre-config
      hook.  The first time we call the hooks, they are in the correct 
      order, but the second time, we don't sort them correctly.  Currently,
      the modules/http/config.m4 file has been renamed to 
      modules/http/config2.m4 to work around this problem, it should moved
      back when this is fixed.

        OtherBill offers that this is a SERIOUS problem.  We do not sort
        correctly by the ordering arguments passed to the register hook
        functions.  This was proven when I reordered the open_logs hook
        to attempt to open the error logs prior to the access logs.  Possibly
        the entire sorting code needs to be refactored.

    * pipes deadlock on all platforms with limited pipe buffers (e.g. both
      Linux and Win32, as opposed to only Win32 on 1.3).  The right solution
      is either GStein's proposal for a "CGI Brigade", or OtherBill's proposal
      for "Poll Buckets" for "Polling Filter Chains".  Or maybe both :-)

    * All handlers should always send content down even if r->header_only
      is set.  If not, it means that the HEAD requests don't generate the
      same headers as a GET which is wrong.

    * exec cmd and suexec arg-passing enhancements
      Status: Patches proposed
      Message-ID: <20020526041748.A29148@prodigy.Redbrick.DCU.IE>
      (see the "proc.patch" and "suexec-shell.patch" links in this message)

    * The 2.0.36 worker MPM graceless shutdown changes work but are
      a bit clunky on some platforms; eg, on Linux, the loop to
      join each worker thread seems to hang, and the parent ends up
      killing off the child with SIGKILL.  But at least it shuts down.

        chrisd: Has this been fixed by the changes for PR 38737?

    * --enable-mods-shared="foo1 foo2" is busted on Darwin.  Pier
        posted a patch (Message-ID: <>).

    * We do not properly substitute the prefix-variables in the configuration
      scripts or generated-configs.  (i.e. if sysconfdir is etc,
      httpd-std.conf points to conf.)

    * If any request gets through ap_process_request_internal() and is
      scheduled to be served by the core handler, without a flag that this 
      r->filename was tested by dir/file_walk, we need to 500 at the very 
      end of the ap_process_request_internal() processing so sub_req-esters
      know this request cannot be run.  This provides authors of older 
      modules better compatibility, while still improving the security and 
      robustness of 2.0. 

        Status: still need to decide where this goes, OtherBill comments...
        Message-ID: <065701c14526$495203b0$>
        [Deleted comments regarding the ap_run_handler phase, as irrelevant
            as BillS points out that "common case will be caught in
  	    default_handler already (with the r->finfo.filetype == 0 check)"
            and the issue is detecting this -before- we try to run the req.]

	gregames says: can this happen somehow without a broken module
            being involved?  If not, why waste cycles trying to defend against
            potential broken modules?  It seems futile.
        wrowe counters: no, it shouldn't happen unless the module is broken.
            But the right answer is to fail the request up-front in dir/file
            walk if the path was entirely invalid; and we can't do that either
            UNTIL 2.1 or we break modules that haven't hooked map_to_storage.

    * With AP_MODE_EXHAUSTIVE in the core, it is finally clear to me
      how the Perchild MPM should be re-written.  It hasn't worked
      correctly since filters were added because it wasn't possible to
      get the content that had already been written and the socket at
      the same time.  This mode lets us do that, so the MPM can be

    * Can a static httpd be built reliably?
        Message-ID: <>

    * Usage of APR_BRIGADE_NORMALIZE in core_input_filter should be
      removed if possible.
        Message-ID: <>
        Jeff wonders if we still care about this.  It is no longer an
        API issue but simply an extra trip through the brigade.

    * Get perchild to work on platforms other than Linux. This
      will require a portable mechanism to pass data and file/socket
      descriptors between vhost child groups. An API was proposed
      on dev@apr:
        Message-ID: <>

    * Try to get libtool inter-library dependency code working on AIX.
        Message-ID: <>

      Justin says: If we get it working on AIX, we can enable this
                   on all platforms and clean up our build system
      Jeff says:   I thought I tested a patch for you sometime in
                   January that you were going to commit within a few

    * Handling of %2f in URIs.  Currently both 1.3 and 2.0
      completely disallow %2f in the request URI path (see
      ap_unescape_url() in util.c).  It's permitted and passed
      through in the query string, however.  Roy says the
      original reason for disallowing it, from five years ago,
      was to protect CGI scripts that applied PATH_INFO to
      a filesystem location and which might be tricked by
      ..%2f..%2f(...).  We *should* allow path-info of the
      form ''.
      Since we've revamped a lot of our processing of path
      segments, it would be nice to allow this, or at least
      allow it conditionally with a directive.

        OtherBill adds that %2f as the SECOND character of a multibyte
        sequence causes the request to fail!  This happens notably in
        the ja-jis encoding.

    * FreeBSD, threads, and worker MPM.  All seems to work fine 
      if you only have one worker process with many threads.  Add 
      a second worker process and the accept lock seems to be
      lost.  This might be an APR issue with how it deals with
      the child_init hook (i.e. the fcntl lock needs to be resynced).
      More examination and analysis is required.
        Status: Works with FreeBSD 5.3. Does not work in previous versions.
                This has also been reported on Cygwin.

    * There is increasing demand from module writers for an API
      that will allow them to control the server  la apachectl.
      Reasons include sole-function servers that need to die if
      an external dependency (e.g., a database) fails, et cetera.
      Perhaps something in the (ever more abused) scoreboard?
             On the other hand, we already have a pipe that goes between parent
             and child for graceful shutdown events, along with an API that
             can be used to send a message down that pipe.  In threaded MPMs,
             it is easy enough to make that one pipe be used for graceful
             and graceless events, and it is also easy to open that pipe
             to both parent and child for writing.  Then we just need to
             figure out how to do graceless on non-threaded MPMs.

    * Allow the DocumentRoot directive within <Location > scopes?  This
      allows the beloved (crusty) Alias /foo/ /somepath/foo/ followed
      by a <Directory /somepath/foo> to become simply 
      <Location /foo/> DocumentRoot /somefile/foo (IMHO a bit more legible
      and in-your-face.)  DocumentRoot unset would be accepted [and would
      not permit content to be served, only virtual resources such as
      server-info or server-status.
      This proposed change would _not_ depricate Alias.
        striker: See the thread starting with Message-ID:

    * Win32: Rotatelogs sometimes is not terminated when Apache
      goes down hard.  FirstBill was looking at possibly tracking the 
      child's-child processes in the parent process.
        stoddard: Shared scoreboard might offer a good way for the parent 
        to keep track of 'other child' processes and whack them if the child 
        goes down.
        Other thoughts on walking the process chain using the NT kernel
        have also been proposed on APR.

    * Eliminate unnecessary creation of pipes in mod_cgid

    * Combine log_child and piped_log_spawn. Clean up http_log.c.
      Common logging API.

    * Platforms that do not support fork (primarily Win32 and AS/400)
      Architect start-up code that avoids initializing all the modules 
      in the parent process on platforms that do not support fork.

    * There are still a number of places in the code where we are
      losing error status (i.e. throwing away the error returned by a
      system call and replacing it with a generic error code)

    * Mass vhosting version of suEXEC.

    * All DBMs suffer from confusion in support/dbmmanage (perl script) since 
      the dbmmanage employs the first-matched dbm format.  This is not
      necessarily the library that Apache was built with.  Aught to
      rewrite dbmmanage upon installation to bin/ with the proper library 
      for predictable mod_auth_dbm administration.
        Questions; htdbm exists, time to kill dbmmanage, or does it remain
                   useful as a perl dbm management example?  If we keep it,
                   do we address the issue above?

    * Integrate mod_dav.
        Some additional items remaining:
        - case_preserved_filename stuff
            (use the new canonical name stuff?)
        - find a new home for ap_text(_header)
        - is it possible to remove the DAV: namespace stuff from util_xml?

    * ap_core_translate() and its use by mod_mmap_static and mod_file_cache
      are a bit wonky.  The function should probably be exposed as a utility 
      function (such as ap_translate_url2fs() or ap_validate_fs_url() or 
      something).  Another approach would be a new hook phase after
      "translate" which would allow the module to munge what the
      translation has decided to do.
        Status: Greg +1 (volunteers)

    * Explore use of a post-config hook for the code in http_main.c which
      calls ap_fixup_virutal_hosts(), ap_fini_vhost_config(), and
      ap_sort_hooks()  [to reduce the logic in main()]

    * read the config tree just once, and process N times (as necessary)

    * (possibly) use UUIDs in mod_unique_id and/or mod_usertrack

    * (possibly) port the bug fix for PR 6942 (segv when LoadModule is put
      into a VirtualHost container) to 2.0.

    * shift stuff to mod_core.h

    * callers of ap_run_create_request() should check the return value
      for failure (Doug volunteers)

    * Fix the worker MPM to use POD to kill child processes instead
      of ap_os_killpg, regardless of how they should die.

        chrisd: Is this done, by any chance?  See r92598 and r93358.

    * Scoreboard structures could be changed in the future such that
      proper alignment is not maintained, leading to segfaults on 
      some systems.  Cliff posted a patch to deal with this issue but
      later recanted. See this message to
      Message-ID: <Pine.LNX.4.44.0203011354090.16457-200000@deepthought

    * APXS either needs to be fixed completely for use when apr is out of tree,
      or it should drop query mode altogether, and we just grow an 
      httpd-config or similar arrangement. 
      To quote a discussion in STATUS earlier:

          thommay: this doesn't fix all the problems with apxs and out of
                   tree apr/apr-util, but it's a good start. There's still the 
                   query cases; but I'm beginning to think that in these cases 
                   the app should be querying ap{r,u}-config directly
          gstein: agreed. apxs should deprecate the -q flag
          pquerna: I vote for a httpd-config, and to deprecate the -q flag.
          minfrin: +1 for httpd-config, and to deprecate -q.


    * Do we need SSL_set_read_ahead()?

    * the ssl_expr api is NOT THREAD SAFE.  race conditions exist:
       -in ssl_expr_comp() if SSLRequire is used in .htaccess
        (ssl_expr_info is global)
       -is ssl_expr_eval() if there is an error
        (ssl_expr_error is global)

    * SSLRequire directive (parsing of) leaks memory

    * Diffie-Hellman-Parameters for temporary keys are hardcoded in
      ssl_engine_dh.c, while the comment in ssl_engine_kernel.c says:
      "it is suggested that keys be changed daily or every 500
      transactions, and more often if possible."

    * ssl_var_lookup could be rewritten to be MUCH faster

    * CRL callback should be pluggable

    * session cache store should be pluggable

    * init functions should return status code rather than ssl_die()

    * ssl_engine_pphrase.c needs to be reworked so it is generic enough
      to also decrypt proxy keys

    * mod_proxy: Ability to run SSL over proxy gateway connections,
      encrypting (or reencrypting) at the proxy.

    * mod_cache: Handle ESI tags.

    * mod_cache: Resolve issue of how to cache page fragements (or perhaps
      -if- we want to cache page fragements). Today, mod_cache/mod_mem_cache
      will cache #include 'virtual' requests (but not #include 'file'
      requests). This was accomplished by making CACHE_IN a
      CONTENT_SET-1 filter to force it to run before the SUBREQ_CORE
      filter.  But now responses cannot be cached that include the
      effects of having been run through CONTENT_SET filters
      (mod_deflate, mod_expires, etc).  We could rerun all the
      CONTENT_SET filters on the cached response, but this will not
      work in all cases. For example, mod_expires relies on installing
      the EXPIRATION filter during fixups. Contents served out of
      mod_cache (out of the quick_handler) bypass -all- the request
      line server hooks (Ryan really hated this. It is great for
      performance, but bad because of the complications listed above).


    * mod_mem_cache: Consider adding a RevalidateTimeout directive to
      specify time at which local cached content is to be revalidated
      (ie, underlying file stat'ed to see if it has changed).

    * mod_cache: CacheEnable/CacheDisable should accept regular expressions.
      jerenkrantz says: Too slow.  Get regexs away from speedy caches by
                        default.  Introduce a new CacheEnableRegex if you want.

    * mod_mem_cache/mod_disk_cache: Need to be able to query cache
      status (num of entries, cache object properties, etc.).
      mod_status could be extended to query optional hooks defined
      by modules for the purpose of reporting module status.
      mod_cache (et. al.) could define optional hooks that are called
      to collect status.  Status should be queryable by
      HTTP or SNMP?
      jerenkrantz says: Yawn.  Who cares.

    * MaxRequestsPerChild measures connections, not requests.
        Until someone has a better way, we'll probably just rename it
    * Regex containers don't work in an intutive way
        Status: No one has come up with an efficient way to fix this
        behavior. Dean has suggested getting rid of regex containers
        OtherBill suggests: We at least seem to agree on eliminating
                            the <Container ~ foo> forms, and using only
                            <ContainerMatch foo> semantics.

    * orig_ct in the byterange/multipart handling may not be
      needed. Apache 1.3 just never stashed "multipart" into
      r->content_type. We should probably follow suit since the
      byterange stuff doesn't want the rest of the code to see the
      multipart content-type; the other code should still think it is
      dealing with the <orig_ct> stuff.
        Status: Greg volunteers to investigate (esp. since he was most 
                likely the one to break it :-)


    Experimental modules should eventually be be promoted to fully supported
    status or removed from the repository entirely (ie, the
    'experiment' failed). This section tracks what needs to happen to 
    get the modules promoted to fully supported status.

View raw message