httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From (Robert S. Thau)
Subject Yet *another* new Shambhala spin...
Date Tue, 04 Jul 1995 11:49:19 GMT
I'm calling this one 0.5.1; it's in the usual place,

Possible problem (which hit me in retrospect) --- this has code for
mutual exclusion around the accept(), based on fcntl() locks on the
error log (not flock()), though it's only #ifdef'ed in on systems
where I believe it to matter.  On some systems, that means that if the
error log is located on another server, and is NFS-mounted, the remote
server's *notorious* lockf daemon will come into play; that may cost
performance.  If the error log is local, there won't be any problem.

Aside from that, I think it's generally an improvement on, well, the
last thing.

Changes with 0.5.1: [Hopefully complete]

  *) Generalized cleanup interface in alloc.c --- any function can be
     registered with alloc.c as a cleanup for a resource pool;
     tracking of files and file descriptors has been reimplemented in
     terms of this interface, so I can give it some sort of a test.

  *) More changes in alloc.c --- new cleanup_for_exec() function,
     which tracks down and closes all file descriptors which have been
     registered with the alloc.c machinery before the server exec()s a
     child process for CGI or <!--#exec-->.  CGI children now get
     started with exactly three file descriptors open.  Hopefully,
     this cures the problem Rob H. was having with overly persistent
     CGI connections.

  *) Mutual exclusion around the accept() in child_main() --- this is
     required on at least SGI, Solaris and Linux, and is #ifdef'ed in
     by default on those systems only (-DFCNTL_SERIALIZED_ACCEPT).
     This uses fcntl(F_SETLK,...) on the error log descriptor because
     flock() on that descriptor won't work on systems which have BSD
     flock() semantics, including (I think) Linux 1.3 and Solaris.

     This does work on SunOS (when the server is idle, only one
     process in the pool is waiting on accept()); it *ought* to work
     on the other systems.

  *) FreeBSD and BSDI portability tweaks from Chuck Murcko.

  *) sizeof(*sa_client) bugfix from Rob Hartill.

  *) pstrdup(..., NULL) returns NULL, pace Randy Terbush.

  *) block_alarms() to avoid leaking the DBM* in dbm auth (this should
     be unnecessary if I go to the revised timeout-handling scheme).

  *) For NCSA bug-compatibility, set QUERY_STRING env var (to a null
     string) even if none came in with the request.

  *) CHANGES file added to distribution ;-).

View raw message