Fix Win32 named pipe handling

[mod_fastcgi.git] / CHANGES

Changes with mod_fastcgi 2.2.5

 *) Fix Win32 named pipe handling (problems with large responses).

Changes with mod_fastcgi 2.2.4

 *) Beta WinNT support.  David Allen [djallen@raleigh.ibm.com] and 
    Rob Saccoccio [robs@chelsea.net]

 *) Remove request type restriction (GET and POST only) in order to support
    Web DAV requests.  Scott Robertson [sroberts@codeit.com] 

 *) Allow requests for URLs such as /server/some/other/qualifier to match a
    FastCGI server defined as /server.  This was done primarily for Java
        Servlets, but is generically useful.

 *) Change the comm between the PM and the request handlers from a regular
    file to a pipe and an assortment of other dynamic fixes.

 *) Log dynamic process termination scheduling and the resulting 
    process exit notification.

 *) Change the default singleThreshhold from 10 to 0 to prevent the last dynamic
    application process from being killed off due to low demand.

 *) Clean up FastCGI application pathnames (e.g. remove duplicate slashes). 
 
 *) Fix bugs that prevented dynamic processes from being shutdown when
    the load subsided.  Lars Heete [heete@do.isst.fhg.de]

 *) Prevent dynamic processes from being scheduled to be started before
    the init-start-delay or restart-delay period expires.  This prevents
        the queuing of process start requests while an application which has
        a long initialization period starts up.

 *) When suexec is enabled, allow the process manager to become root again in
    order to signal applications it spawned.

 *) Change the "server started" log message level from INFO to WARN.

 *) Fix dynamic server "Connection Refused" handling.

 *) Fix demand-based dynamic process spawing (uses a more portable approach).

 *) Add -idle-timeout arg to FastCgiServer, FastCgiExternalServer, and
    FastCgiConfig.  mod_fastcgi will now abort a connection if inactive for
        longer than this period.  It applies to the initiation of connections as
        well (and thus is similar to appConnTimeout).  Default is 30 seconds.

 *) appConnTimeout is 0 by default now resulting in blocking connect()s.  This
    is more platform portable/predicable.

 *) Leave STDOUT and STDERR open to the main server error_log.  This should
    eliminate a pile of protocol errors and having to track down 3rd party
    libs writing to stderr inadvertantly.  This doesn't mean these shouldn't
    be fixed in the application to use the FastCGI I/O, it just allows these
    applications to run when previously they'd crash and burn.

 *) Add the pass-header arg to server directives.

 *) Allow supplementary groups to be passed to spawned FastCGI applications (as
    is done for CGI) - at least until PR2580 is resolved.

 *) Initialize the default (empty) environment in a more socially acceptable
    way.

 *) Miscellaneous doc improvements based on notes I've collected up over the
    last few months.

 *) Fix a bug in the stderr handling introduced in 2.2.2.  There were conditions
    that resulted in not or improperly terminated strings.

 *) Fix the call to setsockopt() to disable Nagle.  [based on a bug report by
    Johannes Plassmann <johannes.plassmann@pdb.siemens.de>]

Changes with mod_fastcgi 2.2.2

 *) Added support for blocking connect()s by setting appConnTimeout to 0.
    Non-blocking connect()s (the default) can be troublesome on some platforms.
    Its expected that blocking connect()s will become the default (and
    non-blocking connect()s will become optional) in the next release.

 *) Dump a compile time error if the version of Apache is too old.

 *) Wrap the SIGPIPE handler manipulation code such that it is only applicable
    to Apache releases prior to 1.3.6.

 *) Minor tweaks for RUSSIAN_APACHE.  [Sergey Gershtein <sg@mplik.ru>]

 *) Dynamic updates:

    Always restart a failed dynamic application if it is the last instance.
    This means there once started there will always be at least one process
    instance of a dynamic application.

    Send PLEASE_START to the PM when a connect() results in an ECONNREFUSED.
    ECONNREFUSED means the listen queue is full (or there isn't one).  Asking
    the PM to start (another) application instance may help empty it faster.

    Change two sleep() calls to select() based snoozes because alarm() is in
    effect and sleep() and alarm() don't always play nice together.

    Fixed a couple of error messages.

 *) Fix -listen-queue-depth arg on FastCgiConfig (dynamic).  Previously it
    was ignored and the default was always used.

 *) Allow the -initial-env argument to be used to pass variables from the
    Apache process environment to the FastCGI server (by specifying a
    variable name without the "=" or a value).  [suggested by
    Martin Lichtin <lichtin@oanda.com>]

 *) Cleanup some debug macros.

 *) Improved script stderr handling.  [based on suggestions by
    David Birnbaum <davidb@chelsea.net>]

 *) Added IRIX and FreeBSD to the list of supported platforms and other minor
    updates to the INSTALL doc.

 *) Changed the default listen-queue-depth (FCGI_DEFAULT_LISTEN_Q) from
    5 to 100.  Its still configurable with the -listen-queue-depth option.
    This should help eliminate the FAQ - Why do I see "Connection Refused"
    messages in the log?

 *) Fix a bug in FastCgiExternalServer that broke support for external servers
    on other hosts.  [Dave Neuer <dneuer@futuristics.net>]

Changes with mod_fastcgi 2.2.1

 *) Updates to the INSTALL doc to describe building as a DSO.

 *) If the FastCgiIpcDir directive was in use and "httpd -t" was issued while
    Apache was running ("apachectl restart" does this), the contents of the
    dynamic directory (dynamic sockets and the mbox) would be blown away
    (a 2.1b1 bug).

 *) Add an extern declaration for ap_sys_siglist (its exposed but not in any
    Apache header file) to prevent a compile error on systems without
    SYS_SIGLIST defined.

 *) During auth requests, the subprocess_env table was left holding variables
    sent to the auth FastCGI server (including REMOTE_PASSWD!) which means they
    were passed to other processing phases (such as CGI/FastCGI).  The
    subprocess_env is now restored to its pre-auth condition.

 *) Added a FastCGI Authorizer Role compatibility mode which implements the
    specification to the tee.  Use -compat arg with any of the Auth server
    directives.  This is intended for new and existing FastCGI authorizer
    applications that require compatibility with other server implementations.

 *) Fix logging from the process manager.  Previously, all calls to the
    logging routines used NULL for the server_rec.  This works fine in a dev
    environment, but results in most of these messages being tossed by
    log_error_core() because DEFAULT_LOGLEVEL is too low.  Odd logic.

 *) Always setup fr->header so that in the event the FastCGI server
    doesn't get a valid FastCGI protocol header over the wire, we
    can still print our error message to the log from do_work().

 *) Fix a compile error in open_connection_to_fs() on systems with TCP_NODELAY
    defined (Don Locrasto [locrasto@iceminer.com]).

Changes with mod_fastcgi 2.2.0

 *) Serious rewrite of mod_fastcgi.html.  An example conf will have
    to wait until another time - I'm burned out on docs updates.

 *) Add to the mod_fastcgi to the server version string.

 *) SCO doesn't like the const in the arg to inet_addr() and
    gethostbyname().

 *) Use Apache's NET_SIZE_T to get the correct arg type for getsockopt().

 *) Deleted redundant header files from fcgi.h.

 *) Full path names are no longer required for directives which take
    file names as arguments.  Like other Apache directives you can
    specify paths from server_root (don't start with a "/").

 *) The fcgiKillMgr is gone.  This intermediate process sat between
    the Apache parent and the FastCGI Process Manager under Apache
    1.2.         The process manager tries to name itself fcgi_pm (it used
    to be fcgiProcMgr).

 *) Authentication, Authorization, and Access phases are now
    supported. I've bent the FastCgi spec a bit: as many of the standard
    environment variables (that were easy) are sent (the spec says don't
    send some by name), all headers (except Status) sent by the auth
    FastCgi server are passed to subprocesses (CGI/FastCGI invocations)
    as environment variables rather than just those prefixed by
    "Variable-".  Custom responses for auth failures aren't supported
    (yet.. we'll see what the demand for this is like).  In addition to
    the FastCGI protocol defined environment variable "FCGI_ROLE" being
    set to "AUTHORIZER" an environment variable "FCGI_APACHE_ROLE" is
    set indicating which of the three phases is being processed.  See
    the docs for info on the new per-directory directives:

        FastCgiAuthenticator, FastCgiAuthenticatorAuthoritative,
        FastCgiAuthorizer, FastCgiAuthorizerAuthoritative,
        FastCgiAccessChecker, FastCgiAccessCheckerAuthoritative

 *) The code has been broken into logical chunks making figuring it
    out much easier.  All of the #defines a user may want to change are
    now in mod_fastcgi.h.

 *) All the logging has been converted to the "supported" Apache
    logging routines, ap_log_error() and ap_log_rerror().  Log entries
    for request specific errors should now be directed to the correct
    server errorlog (if your running more than one).  Some attempt has
    been made at setting appropriate log levels (comments of course are
    welcome) and printing errno info when its of value.  All of log
    messages now have "FastCGI" in them (nice for grep).

 *) All of the calls to Apache routines have been updated to the ap_
    convention, eliminating the need for the compat header.

 *) By default we no longer flush every piece of data we get from the
    FastCGI server to the client.  This allows the FastCGI server to
    release without having to wait for the client flush to complete.
    The old behaviour can revived with the -flush argument to AppClass.

 *) I renamed some of the directives for consistency (the old names
    still work).

        AppClass -> FastCgiServer,
        FCGIConfig -> FastCgiConfig,
        ExternalAppClass -> FastCgiExternalServer

 *) Directive arguments are no longer case sensitive.

 *) The module now uses hard_timeout() rather than soft_timeout.  This
    means that the module will longjump out of the request if the
    drop-dead timeout expires (set with Apache's Timeout directive, the
    default is 5min) or if the client closes the connection (SIGPIPE).
    This is more typical for a Apache modules and I'm not convinced we
    properly handle all of the error cases well enough to use
    soft_timeout (I can think of one place we don't anyway). This means
    your FastCGI application can see SIGPIPE.

 *) SIGPIPE is now ignored by default in FastCGI servers spawned by
    Apache.  Without this the default behaviour is exit().

 *) It should (I think) run now under DSO now.

 *) The bug which prevented sending of binary data in 2.1b1 is fixed.

 *) The broken -initial-env bug ins 2.1b1 is fixed.

 *) Maybe some other stuff.



*** Well there's a big gap here.  Maybe I'll go back and extract the CVS
commit notes and stick 'em in here.. when I get some free time. ;)

    --robs  (8 Feb 99)


*** Originally from docs/README..

Apache[X1.03.01]/mod_fastcgi[02.00.05] 19970909 unsupported
From: Stanley Gambarin <stanleyg@cs.bu.edu>

 *) Yet more changes to the source distribution.  Separated out
    the Tcl dynamic string and buffer libraries into separate files.
    Did the same for the OS library, however, it is not really an
    abstraction... more like a bunch of wrappers.  Added to DEVNOTES
    a note about the problem of Apache not allowing including header
    file more than once.

Apache[X1.03.01]/mod_fastcgi[02.00.04] 19970908 unsupported
From: Stanley Gambarin <stanleyg@cs.bu.edu>

 *) More changes to the installation scripts.  Theoretically we
    should be able to to use the same Makefile/installation script
    for both Apache 1.2.x and 1.3.x.  Need some sort of the abstraction
    layer to use the same source code for both 1.2.x and 1.3.x sources.

Apache[X1.03.01]/mod_fastcgi[02.00.03] 19970905 unsupported
From: Stanley Gambarin <stanleyg@cs.bu.edu>

 *) Created Makefile to compile the module into the archive library,
    just like the proxy module.  Also initial draft of the installation
    script, which should work for both 1.2 and 1.3 sources of Apache.

Apache[01.02.04]/mod_fastcgi[02.00.02] 19970903 unsupported
From: Stanley Gambarin <stanleyg@cs.bu.edu>

 *) More source reorg.  Separated out some header files.  Need much
    more work on this.  Updated the TODO file.  Added fcgivers.h and
    CHANGES     file to track the history.

Apache[01.02.04]/mod_fastcgi[02.00.01] 19970902 unsupported
From: Stanley Gambarin <stanleyg@cs.bu.edu>

 *) Source reorganization.  This is done to provide a basis for later
    functionality implementation.  See TODO file for more information
    of what possible future enhancements are possible/desired.

Apache[01.02.04]/mod_fastcgi[02.00.00] 19970902 unsupported
From:  David MacKenzie <djm@va.pubnix.com>

 *) Create the "dynamic" dir and "mbox" as the user specified in the
    User and Group directives instead of as root.  They are created with
    restrictive permissions, and then the module checks to see whether the
    user specified in the User and Group directives has read, write, and
    execute permissions on them.  Those permissions had been explicitly
    denied by creating them as "root".

 *) Some documentation was garbled or incomplete.

 *) When a FastCGI application can't be execl'd, the code used a
    value of errno that may have been stomped on by intervening system
    calls.  That happened on BSD/OS 2.1, where the error_log was
    reporting errno=25 (NOTTY) instead of the correct errno=13 (EACCES).
    An intervening stdio call was setting errno as a side-effect, which
    it has the right to do.

 *) Erroneous fprintf arguments, a missing return value, an unused
    function, missing declarations, and other problems detected by gcc
    -Wall.

 *) Duplicated code to create the lock file name merged into a single
    function, closing several memory leaks.

 *) FCGIConfig -minProcesses didn't allow a value of 0, so the last
    FCGI app in a given dynamic class would live forever, even if it
    hadn't been requested for weeks.

 *) The test for whether to keep looking for dynamic app victims to
    kill was backwards, so no dynamic apps were ever selected as
    victims.

 *) The killInterval and updateInterval were being ignored; the
    sigsuspend() forced recalculations only when a child died,
    instead of at the intervals specified.

 *) When a dynamic app couldn't be started, the program SEGV'd when
    trying to free the ipcAddrPtr twice.

 *) Remove the dead lock file and socket when cleaning up after a
        server whose last child has been killed, so the FastCGIHandler
    isn't fooled into thinking there is still a process serving
    that app.

 *) Make file locking robust in the presence of signals.

 *) Rename some badly named variables.

 *) Fix typos in many comments.

 *) Fix some memory and file descriptor leaks.

 *) Make the blocking kill of a server closer to working.

 *) More fixes to calculations in dynamic application management.


*** Originally from docs/README.OMI..

What's New: Version 2.0b1, 16 Apr 1997

 *) Implemented a mechanism by which FastCGI applications are started by
    the web server on the first request and continue running.  Also,  a
    heuristics have been implemented to allow for dynamic killing of the
    running FastCGI apps.  The configuration options are supplied via
        the new FCGIConfig directive, that is described in the
        mod_fastcgi.html

 *) When performing internal redirect, since the original request's body
    has already been read, do not allow the redirected request to think
    that is has one [body].

 *) More conditional compilation for OS/2.

 *) Fixed occassional  "Assertion failed: len > 0, file mod_fastcgi.c"

 *) Fixed "failed assertion `count >= 0 && count <= bufPtr->length'"

 *) Added bflush() call after bwrite() in DrainReqOutBuf function to
    force some output to be written in the period of inactivity. This
    circumvents the Apache's buffering during the server-push.

 *) Added a header-parser function placeholder for compatibility with an
    Apache source code.

 *) Additional checks have been placed with regard to permissions of the
    FastCGI processes.  These include the existence of the  file,
    ExecCGI and IncludesNOEXEC checks, disallowing of nph- scripts, etc.

 *) Implemented a restart cleanup, so that no parentlesses FastCGI
        applications are left after Apache is restarted and/or terminated.
        This was accomplished via two independent mechanisms, where the
        first one prevents a start of the process manager on the first
        reading of the configuration files.  The second mechanism deals away
        from Apache's  implementation of fork() (via spawn_child()) and
        implements its own forking, thus removing any dependencies on the
        Apache to cleanup  processes during termination phase.

 *) Removed the definition of the Sigfunc, as Apache 2.0b10 defines it
    itself in the file conf.h

What's New: Version 1.4.3, 15 Jan 1997

 *) Fixed compilation warnings for various platforms, as well as
        conditional compilation for OS/2.

What's New: Version 1.4.2, 12 Dec 1996

 *) mod_fastcgi.c is ported to Apache 1.2b.  Any further development
        will proceed under this version of Apache web server.

 *) As the result of porting, the "include virtual" construct of SSI
        will now work correctly using either <Location> or <Directory>
        directive.

What's New: Version 1.4.1, 4 Dec 1996

 *) Checks have been removed from the ScanCGIHeaders that provided
    for the presence of both Status and Location headers as being
    an error.  Contradictory to CGI/1.1 Internet Draft, both of
    these headers are used by the current CGI applications.

What's New: Version 1.4, 22 Nov 1996

 *) Added the -port option to AppClass, allowing TCP/IP communication.
    Added the -socket option to AppClass, allowing Unix domain
    communication via a configurable pathname.

 *) Added the ExternalAppClass directive, allowing TCP/IP
    communication with remote FastCGI applications.

 *) The handler had its own code for generating HTTP response headers;
    now it uses Apache's.  This reduces the size of the module.
    More importantly, it fixes the bug in which "include virtual"
    sees the HTTP response headers.

 *) The response header parser performed very little checking.
    Now the parser enforces the guidelines in the CGI/1.1
    Internet-Draft: Status and Location are mutually exclusive,
    Location can only be a response to GET or HEAD,
    CGI response headers can't be repeated, etc.  (The CGI response
    headers are the ones the handler interprets: Status, Location,
    and Content-type.)

 *) The response header parser used to miss CGI headers
    with no whitespace after the colon, e.g. "Status:200 OK".

 *) The response header parser sometimes interpreted the first line
    of content as an RFC822 continuation line.

 *) The handler implemented a nonstandard version of Location
    which never used internal redirects.  The handler
    now attempts to implement Location as specified in the
    CGI/1.1 Internet-Draft.  The module documentation explains
    the new behavior.

 *) Error log entries from the response header parser were
    pretty uninformative; they are better now.  If a header is
    malformed, the log entry includes it.  If the headers are
    unterminated rather than malformed, the log entry says that,
    and says how many bytes were received from the app.

 *) When the application manager forked a new process, and that
    process ran into trouble before executing the first instruction
    of application code, the process used to exit with status =
    errno.  This made certain configuration problems (e.g. incorrect
    file permissions when server parent is root) quite difficult
    to pin down.  Now the failing child process opens up the error log
    and writes an informative entry before exiting.

 *) The module now correctly handles a slash at the end of the
    DocumentRoot directive.  This was a one-line fix.

What's New: Version 1.3.3, 17 Oct 1996

 *) The module now registers its request handler under the name
    fastcgi-script in addition to the name application/x-httpd-fcgi.
    This was a one-line addition to the module, but had quite
    a large impact on the documentation and sample configuration.

What's New: Version 1.3.2, 27 Sept 1996

 *) On some systems (SunOS, Linux?), fopen for append has a bug
    that strikes when two processes append to the same file.
    This bug causes the process manager to corrupt the error log.
    Work around the bug by calling open, then fdopen.  A patch
    to Apache 1.1.1 is also required, as described in mod_fastcgi.html.
    (Reported by Bob Ramstad.)

What's New: Version 1.3.1, 17 Sept 1996

 *) On Linux and SunOS, the Apache default user_id (-1) is not a
    legal value of uid_t, so cast it.  (Reported by Bob Ramstad,
    Scott Langley, others.)

 *) On some systems (Linux, some Solaris?, Irix?), connect requires
    write access to a Unix Domain socket, so provide it.  (Reported
    by Scott Langley, freeform@wired.com, others.)

 *) If you hit a .fcg file, but there's no AppClass defined, the
    error should be NOT_FOUND instead of SERVER_ERROR.  (Reported
    by Michael Smith.)

What's New: Version 1.3, 4 Sept 1996

 *) Module sometimes busy-waited in FastCgiDoWork.  Fixed.

 *) Module violated the Apache buff abstraction.
    Fixed.  Might now work with SSL (not tested).
    As part of the fix, eliminated support for
    Apache 1.0x versions.

 *) Module failed to chmod the Unix domain listening sockets
    it created, so protections were set according to the current
    umask.  Fixed.

 *) Module forked too many process manager processes: One
    per AppClass.  When the Apache parent ran as root, these
    process manager processes ran as root.  New process
    manager is a single process (only started if AppClass is
    used) and runs with same privileges as other children.
    The new process manager doesn't do any polling, so
    there's less system overhead than before.

 *) AppClass insisted on getting at least two arguments.  Fixed.

 *) Module used setjmp/longjmp, causing compiler warnings
    on some platforms (e.g. Linux.)  The module no longer
    uses setjmp/longjmp.

 *) Module created listening sockets in /tmp, where they were
    sometimes wiped out by cleanup scripts.  Added FastCgiIpcDir
    directive to give control over the location of listening
    sockets.

 *) Module wrote error log entries without a timestamp.  Fixed.

 *) AppClass directive wrote error messages to stderr in addition
    to returning a char * message to the Apache core.  Fixed.

What's New: Version 1.2, 3 June 1996

 *) Ported from Apache-1.0-based code to Apache-1.1b2 internals by
    Ralf S. Engelschall
    <rse@en.muc.de>
    Add version string: APACHEVERSION macro to mod_fastcgi.c

 *) chown FastCGI socket to user_id and setuid to user_id for app
    class processes

 *) Modify GetFromStream() by having it call an OS dependent
    function GetStreamSize(FILE *) which uses FILE internal data member.
    Linux users might need to modify GetStreamSize().

What's New: Version 1.1, 10 May 1996

 *) If you specify a non-existent executable in the AppClass directive,
        or if the file exists and it does not have execute permission you
        get a constant stream of error messages telling you that "program
        terminated due to a signal".

 *) The mod_fastcgi module should use the standard Apache error logging
        facility instead of writing to stderr.

Version 1.0, 30 April 1996