Re: 2 problems with shutdown

On Fri, Jun 29, 2001 at 02:09:44PM +0300, Zeev Suraski wrote: > At 18:45 28/6/2001, Thies C. Arntzen wrote: > > > hi, > > > > we have two problems in the current code when it comes to > > shutting down apache processes: > > > > 1) > > in main.c php_module_shutdown(): > > we call php_config_ini_shutdown() before we call the modules > > MSHUTDOWN functions - which basically means that php is > > already "half-dead" when MSHUTDOWN is called - so modules can > > can not rely on certain things (like error_log settings) > > during MSHUDOWN. is there any known reason not to move > > php_config_ini_shutdown() further down in > > php_module_shutdown()? > > The problem is that (if I remember correctly) the INI mechanism may call > callbacks in modules that have already been unloaded. It can probably be > fixed, but would require some time to play with. agreed - but we should do it, right? > > > > 2) (more serious) > > if you do an "apachectl restart" the master httpd will signal > > the worker httpd's. the workers will than call the > > child_exit hook in all modules no matter where in the code > > they just happen to be. the problem is that we "try" > > to clean up the best we can -but- this might cause recursive > > calls into 3th party library code (which is not supported in > > most cases!) > > > > sample: > > > > script calls ociexecute($stmt) (which might take a few > > seconds) now the admin does "apachectl restart" - and we > > get interrupted deep down in the oci library. now the worker > > httpd tries to clean up and calls the child_exit hooks. php > > will now free all resources and it will also try to do a > > rollback on $stmt and after that it'll free $stmt. the > > problem here is that the oracle-server _doesn't_ like > > recursive calls at all and might crash just 'cause of that. > > > > OK - i agree - oracle should fix their code, -but- i don't > > think its smart to try a real-cleanup if we we're > > interrunpted during a request. i think we need to protect the > > engine agains recursive calls. so if child_exit is called > > while we're still in execution mode simply do nothing > > (bacause all we could do might be harmful) and leave the > > clean up to unix. > > I don't know if that qualifies as a recursive call - it's really doing what > it's supposed to do. Perhaps the Oracle module should protect itself, but > not running cleanup is wrong IMHO - consider semaphores or the likes which > won't clean up automatically. Or SQL servers that won't immediately detect > that their client is gone. I think that cleaning up is generally an > important thing. i agree - in theory. but "apachectl restart" will sometimes SEGFAULT some httpd processes because of php. just create a trivial page that will take some time: test.php <?php for ($i = 0; $i < 1000000; $i++) ; ?> run "ab -n1000 http://0/test.php" and then do an apachectl restart (a few times) and you'll see zend_execute.c(376) : Freeing 0x08162BDC (12 bytes), script=/var/www/html/t.php zend_hash.c(287) : Freeing 0x08162B84 (37 bytes), script=/var/www/html/t.php Last leak repeated 160 times zend_opcode.c(303) : Freeing 0x08161AEC (720 bytes), script=/var/www/html/t.php zend_language_scanner.c(4153) : Freeing 0x08162B4C (2 bytes), script=/var/www/html/t.php zend_opcode.c(237) : Freeing 0x08162B0C (12 bytes), script=/var/www/html/t.php zend_language_scanner.c(4238) : Freeing 0x08162AD4 (2 bytes), script=/var/www/html/t.php Last leak repeated 2 times zend_opcode.c(71) : Freeing 0x08161AB4 (4 bytes), script=/var/www/html/t.php zend_compile.c(118) : Freeing 0x081619C4 (20 bytes), ... in the error_log - and sometimes it'll even SEGFAULT. is this good? tc