APACHE PORTABLE RUNTIME (APR) LIBRARY STATUS: -*-text-*- Last modified at [$Date$] Releases: Standalone 1.0.0 rc2 : tagged June 30th 2004 (APR_1_0_RC2) 1.0.0 rc1 : tagged June 18th 2004 (APR_1_0_RC1) 0.9.3 : tagged March 30, 2003 0.9.2 : released March 22, 2003 0.9.1 : released September 11, 2002 0.9.0 : released August 28, 2002 Bundled with httpd 2.0a9 : released December 12, 2000 2.0a8 : released November 20, 2000 2.0a7 : released October 8, 2000 2.0a6 : released August 18, 2000 2.0a5 : released August 4, 2000 2.0a4 : released June 7, 2000 2.0a3 : released April 28, 2000 2.0a2 : released March 31, 2000 2.0a1 : released March 10, 2000 RELEASE 1.0.1 SHOWSTOPPERS: CURRENT VOTES: CURRENT test/testall -v EXCEPTIONS: Please add any platform anomilies to the following exception list. * various tests fail on Unix in VPATH builds. * 'testipsub' will tickle an Solaris 8 getaddrinfo() IPv6 bug, causing the test to hang. Configure with --disable-ipv6 if using an unpatched Solaris 8 installation. * The 'testdso' tests will not work if configured with --disable-shared since the loadable modules cannot be built. * 'testdso' fails on older versions of OpenBSD due to dlsym(NULL, ...) segfaulting. * BUG: Win32 fails test in File Info: test_stat_eq_finfo apr_stat and apr_getfileinfo differ in protection ... wrowe guesses that we are checking the handle objects' permissions rather than the filesystem objects' permissions. * Win32 Not Implemented tests Pipes: set_timeout/read_write; can't timeout blocking pipes Socket Creation: tcp6_socket and udp6_socket (at least by default) Socket Options: corkable: TCP isn't corkable Users: username: Groups from apr_uid_get not implemented RELEASE NON-SHOWSTOPPERS BUT WOULD BE REAL NICE TO WRAP THESE UP: * Someone needs to port testucs to Unix. Right now it only works on Windows. * The return type of a thread function (void *) is inconsistent with the type used in apr_thread_exit()/apr_thread_join() (apr_status_t). The thread function's return type should be changed to apr_status_t so that a return from the thread main function has the same effect as apr_thread_exit(). See Message-Id: for thread discussing this. +1: BrianH, Aaron, david, jerenkrantz Status: Will Rowe was working on this. * Need some architecture/OS specific versions of the atomic operations. progress: generic, solaris Sparc, FreeBSD5, linux, and OS/390 done need: AIX, AS400, HPUX * The new lock API is a full replacement for the old API, but is not yet complete on all platforms. Components that are incomplete or missing include: Netware: apr_proc_mutex_*() (Is proc_mutex unnecessary on Netware?) * proc_mutex is not necessary on NetWare since the OS does not support processes. The proc_mutex APIs actually redirect to the thread_mutex APIs. (bnicholes) OS/2: apr_thread_cond_*(), apr_proc_mutex_*() Less critical components that we may wish to add at some point: Beos: apr_thread_rwlock_try*lock() apr_proc_mutex_trylock() Unix: apr_thread_rwlock_*() for platforms w/o rwlocks in pthread Win32: apr_thread_cond_timedwait(), apr_proc_mutex_*() (Is proc_mutex unnecessary on Win32?) * Need to contemplate apr_strftime... platforms vary. OtherBill suggested this solution (but has no time to implement): Document our list of 'supported' escapes. Run some autoconf/m4 magic against the complete list we support. Move the strftime re-implementation from time/win32 to time/unix. Add some APR_HAVE_STRFTIME magic to use the system fn, or fail over to time/unix/strftime.c. Message-ID: <025e01c1a891$bf41f660$94c0b0d0@v505> * Using reentrant libraries with non-threaded APR - Anecdotal evidence exists that suggests it is bad to mix reentrant and non-reentrant libraries and therefore we should always use the reentrant versions. - Unfortunately, on some platforms (AIX 4.2.1) defining the reentrant flag (-D_THREAD_SAFE) causes builds to fail, and so one would expect --disable-threads to fix this. Although this has been fixed for that particular version of AIX, it may be useful to only enable the reentrant versions when threads are enabled. How will we deal with this issue once APR becomes a standalone library? It is perfectly legitimate to have apps needing both versions (threaded/reentrant and non-threaded/non-reentrant) on the same machine. * Pools debugging - Find a way to do check if a pool is used in multiple threads, while the creation flags say it isn't. IOW, when the pool was created with APR_POOL_FNEWALLOCATOR, but without APR_POOL_FLOCK. Currently, no matter what the creation flags say, we always create a lock. Without it integrity_check() and apr_pool_num_bytes() blow up (because they traverse pools child lists that possibly belong to another thread, in combination with the pool having no lock). However, this might actually hide problems like creating a child pool of a pool belonging to another thread. Maybe a debug function apr_pool_set_owner(apr_thread_t *) in combination with extra checks in integrity_check() will point out these problems. apr_pool_set_owner() would need to be called everytime the owner(the thread the pool is being used in) of the pool changes. - Implement apr_pool_join and apr_pool_lock. Those functions are noops at the moment. - Add stats to the pools code. We already have basic stats in debug mode. Stats that tell us about wasted memory in the production code require more thought. Status: Sander Striker is looking into this (low priority) David says this is a 1.1 issue. * Get OTHER_CHILD support into Win32 Status: Bill S. is looking into this * Win32 apr_proc_create fails to create 16 bit apps detached (a win32 bug.) The question - test in advance (slow) or recover gracefully from failure and try again? Only the test method will work on Win9x, since it will appear to work, only to encounter mangled pipes. Win2K (NT?) simply fails. * SysV semaphore support isn't usable by Apache when started as root because we don't have a way to allow the semaphore to be used by the configured User and Group. Current work-around: change the initial permissions to 0666. Needed code: See 1.3's http_main.c, SysV sem flavor of accept_mutex_init(). Status: Jim will look into this Update: Apache deals with this itself, though it might be nice if APR could do something. * Build scripts do not recognise AIX 4.2.1 pthreads Justin says: "Is this still true?" * FirstBill says we need a new procattr, APR_CREATE_SUSPENDED (or something similar) to direct ap_create_process to create the process suspended. We also need a call to wake up the suspended process. This may not be able to be implemented everywhere though. Status: OtherBill asks, why? What is the benefit, how is it portably implemented? Unless this creates some tangible that mirrors another platform, then I'm -1. * Replace tables with a proper opaque ADT that has pluggable implementations (including something like the existing data type, plus hash tables for speed, with options for more later). Status: fanf is working on this. * add a version number to apr_initialize() as an extra failsafe against (APR) library version skew. MsgID: Status: Greg -1, Jeff +1, Ryan +1, Tony -0(?), david +1 * add apr_crypt() and APR_HAS_CRYPT for apps to determine whether the crypt() function is available, and a way to call it (whether it is located in libc, libcrypt, or libufc) Justin says: Should apr_crypt() be in apr-util? Status: Greg +1 (volunteers) * configure.in does post-processing on the AC_OUTPUT files (for VPATH support). This means that config.status doesn't do the right thing when you re-run it. We ought to revamp the makefiles to do the right AC_SUBST stuff rather than depend upon rewriting. Sascha: As the rewriter is a crude hack, I would not worry too much about it. It is designed to go away once we have a proper build system in place. One of the perceived deficiencies of automake is that it uses AC_SUBST too often, thereby slowing down the task of generating Makefiles significantly, because it applies dozens of substitutions to each Makefile. And why? Make's built-in macro processing is much more powerful, and combined with the include facility, generating Makefiles becomes simpler and faster. Justin says: "I think this got fixed with Roy's build changes." * use os_(un)cork in network_io/unix/sendrecv.c for FreeBSD's sendfile implementation. david: The socket options stuff is now in and using it should reduce the number of syscalls that are required for os_cork and uncork, so the code should be reviewed to make use of the new calls. If no-one beats me to it I'll get around to it soonish... * toss the per-Makefile setup of INCLUDES; shift to rules.mk.in rbb: This is a bad thing IMHO. If we do this, then we can't use these makefiles for anything else. For example, apr-util * add the rest of the pool accessor declare/impl macros. Justin says: Both thread and file have the accessors now. Any others? Status: Greg volunteers * I think apr_open_stderr() and friends *should* dup() the descriptor. That would allow the new/returned file to be closed (via pool cleanup or manually) without accidentally closing stderr/out. * need to export (in code, not just build scripts) the shared library extension (e.g. ".so") for the platform. clients need to use this to construct filenames to pass to apr_dso_load() -- note on Win32 we distinguish 'apache module' names from other 'loadable module' names, so be careful with Apache's directive. * Possible gmtime_r replacement in explode_time On Solaris (and possibly others), the gmtime_r libc function obtains a mutex. We have seen 21/25 threads being blocked in this mutex on a threaded httpd MPM when requesting static pages. It may be worth it to hand optimize this since there is no real need for a mutex at the system level (straight arithmetic from what I can tell). If you have access to the Solaris source code: osnet_volume/usr/src/lib/libc/port/gen/time_comm.c. * Add a way to query APR for what features it has at runtime (i.e. threads). Justin says: I'm not completely sold on this, but it has been mentioned before and at least added to STATUS. * apr_xlate.h generates a bunch of compiler warnings. Jeff asks: which platform? Justin says: Solaris with Forte 6.1. * fcntl() oddness on Solaris. Under high loads, fcntl() decides to return error code 46 (ENOLCK). httpd (prefork MPM) error log says (predictably): (46)No record locks available: couldn't grab the accept mutex All of the children report this and subsequently exits. httpd is now hosed. AFAICT, this does not look to be an out-of-fds error. Solaris's man page says: ENOLCK The cmd argument is F_SETLK, F_SETLK64, F_SETLKW, or F_SETLKW64 and satisfying the lock or unlock request would result in the number of locked regions in the system exceeding a system-imposed limit. Justin says: What is this system-imposed limit and how do we change it? This gives me more rationale for switching the default interprocess lock mechanism to pthread (if available). Explanation (from Kristofer Spinka ): ============ The system imposed default limit of outstanding lock requests is 512. You can verify this by, in a contemporary version of Solaris: # mdb -k > tune_t_flckrec/D tune_t_flckrec: tune_t_flckrec: 512 This can be increased by adding the following to /etc/system: set tune_t_flckrec=1024 and rebooting. Of course "1024" can be any reasonable limit, although we do not know what "reasonable" should be, so be conservative, only increase this as necessary. * Generate a good bug report to send to the FreeBSD hackers that details the problems we have seen with threads and system calls (specifically sendfile data is corrupted). From our analysis so far, we don't think that this is an APR issue, but rather a FreeBSD kernel issue. Our current solution is to just disable threads across the board on FreeBSD. MsgID: <20010828091959.D17570@ebuilt.com> Status: Fixed in -CURRENT. MFC in about a week. Continuing testing with threads on FreeBSD. FreeBSD PR kern/32684: http://www.freebsd.org/cgi/query-pr.cgi?pr=kern/32684 * There are some optimizations that can be done to the new apr_proc_*() functions (on UNIX). One that may reduce pointer indirection would be to make the apr_proc_mutex_unix_lock_methods_t first-class members of the apr_proc_mutex_t structure. * Condition variables are tricky enough to use, and even trickier to implement properly. We could really use a better test case for those subtle quirks that sometimes creep into CV implementations. * Once we are fully satisfied with the new lock API, we can begin to migrate the old API to be implemented on top of the new one, or just decide to get rid of it altogether. * FreeBSD returns 45 (EOPNOTSUPP) when the lockfile is on a NFS partition when you call fcntl(F_SETLKW). It may be good if we can somehow detect this and error out when creating the lock rather than waiting for the error to occur when acquiring lock. * Fix autoconf tests for strerror_r on BeOS and remove the hack in misc/unix/errorcodes.c to get error reporting working. Committed as the solution is elusive at present. * implement APR_PROGRAM_ENV and APR_PROGRAM_PATH on BeOS, OS/2, Netware, and Win32. * stat() on a few platforms (notably Solaris and AIX) succeeds for a non-directory even if a trailing '/' was specified in the name. APR should perhaps simulate the normal -1/ENOTDIR behavior in APR routines which retrieve information about the file. Note: Win2K fails GetFileAttributesEx in this scenario. See OtherBill's comments in this message to dev@httpd.apache.org: Message-Id: <5.1.0.14.2.20020315080852.00bce168@localhost> Documentation that needs writing: * API documentation Ian Says: APR Stuff in now in Doxygen format, which is the first step. David says: are we planning on doing any more? I'm tempted to remove this item. * apr-site needs to be revamped with Anakia/XHTML. * Identify and implement those protection bits that have general usefulness, perhaps hidden, generic read-only [immutable], effective current user permissions, etc. * Maybe make the following functions return void instead of apr_status_t, as they cannot ever error: apr_md5_init() apr_md5_update() apr_md5_final() apr_md5_encode() apr_md5() /* plus make the obvious code tweak in this one */ (Volunteer: Karl Fogel .) However, don't do this until after apr and apr-util offer library version numbers, and httpd uses those numbers to indicate which version it needs. Until the libraries are versioned, this api change is [somewhat] painful for httpd. Status: Still in discussion, current leanings appear to be Bill Stoddard -0.5 (?) Sander Striker +1 Greg Stein +1 Karl Fogel +1 david: This was rejected for 1.0 following Ben L's comment that should we ever start using any other form of md5 (e.g. openssl) then errors would become a distinct possibility. Stuff for post 1.0: * Almost every API in APR depends on pools, but pool semantics aren't a good match for a lot of applications. We need to find a way to support alternate allocators polymorphically without a significant performance penalty. * apr_global_mutex_child_init and apr_proc_mutex_child_init aren't portable. There are a variety of problems with the locking API when it is used with apr_create_proc instead of apr_fork. First, _child_init doesn't take a lockmech_e parameter so it causes a segfault after the apr_proc_create, because the proc_mutex field hasn't been initialized. When the lockmech_e parameter is added, it _still_ doesn't work, because some lock mechanisms expect to inherit from the parent process. For example, sys V semaphores don't have a file to open, so the child process can't reaquire the lock. jerenkrantz says: This is not a showstopper and I believe the above analysis is slightly confusing. The real problem here is that apr_*_mutex_child_init assumes a shared memory space - that is, the children processes have access to the parent apr_*_mutex_t pointer. The children just call child_init on the original, inherited apr_*_mutex_t. Unlike globalmutexchild in test, apr_*_mutex_create is *not* intended to be called from the child and subsequently call child_init. Instead, apr_create_proc is intended to exec separate processes with disjoint memory addresses. Currently, APR does not provide a cross-platform mechanism for joining an already existing lock. A simple 'apr_*_mutex_join' which is intended to be called from separate processes to an already-existing lock would solve this problem. child_init is not intended to be used this way. Even with SysV semaphores, using IPC_PRIVATE should still work due to the parent-child relationship. A strawman has been posted to dev@apr: Message-Id: <213031CF0406DE1AC426A411@[10.0.1.137]> This was listed as a showstopper for 1.0, but while the 2 patches above exist neither was able to garner enough votes to be included in 1.0. Will Rowe commented that a combination of the 2 would probably be the right approach, a view that seems to have a lot of merit. Hopefully we can solve this post 1.0. There were also enough people who felt that it wasn't a real showstopper for it to be bumped. * Must namespace protect all include/apr_foo.h headers. Jon Travis has especially observed these including apr within Apache-1.3. Message-ID: <20020128100116.A4288@covalent.net> Deprecating the symbols in 0.9, eliminating them with 1.0. (Those problems have been fixed, but it is a good example of what to look for.) Some headers with issues: apr.hnw (READDIR_IS_THREAD_SAFE, ENUM_BITFIELD, _POSIX_THREAD_SAFE_FUNCTIONS (?)) * Flush out the test suite and make sure it passes on all platforms. We currently have about 450 functions in APR and 147 tests. That means we have a large number of functions that we can't verify are actually portable. This TODO includes finishing the migration to the unified test suite, and adding more tests to make the suite comprehensive. * Eliminate the TODO's and XXX's by using the doxygen @bug feature to allow us to better track the open issues, and provide historical bug lists that help porters understand what was wrong in the old versions of APR that they would be upgrading from.