#36158 [Asn]: SIGTERM is not handled correctly when running as a FastCGI server

ID: 36158 User updated by: chris at mysociety dot org Reported By: chris at mysociety dot org Status: Assigned Bug Type: CGI related Operating System: * PHP Version: 5.1.2, 4.4.2 Assigned To: dmitry New Comment: Yeah, that's a fair point. Matters would be improved by adding the FCGI_FAIL_ON_INTR flag to the call to FCGX_InitRequest; that's not a complete solution, since that might cause the server process to quit on receipt of another signal in the call to FCGX_Accept_r, but that's not a very serious problem. Here's a patch that includes that change: http://bitter.ukcod.org.uk/~chris/tmp/20060201/php-4.3.10-fastcgi-sigterm-fix.patch Previous Comments: ------------------------------------------------------------------------ [2006-02-01 18:20:43] stefan at luli dot de Hi there, I am struggeling with the described problem for quite some time now. I have traced the problem now in the code back to the SIGTERM play between fcgi_pm and php. The question is if fcgi_pm should send a SIGTERM to the script even though it processes a request at this momment. One would think that fcgi_pm should be smart enough to deal with this situation (wait until request is done/send no more requests/send SIGTERM). But apparently it isn't. The patch seams to do the trick, but is also prevents php to terminate in case it's just idel. So php will sit there until the next request comes in and terminate once the request is done before waiting for the next request to come in. This should be solved. ------------------------------------------------------------------------ [2006-01-26 08:58:25] sniper@php.net Dmitry, does the new code fix this problem..? ------------------------------------------------------------------------ [2006-01-26 01:55:15] chris at mysociety dot org That snapshot (a) doesn't fix the problem -- the FastCGI code is essentially unchanged since 4.x, though with a bunch of whitespace noise; and (b) wouldn't help us in any case since it's PHP5 and our applications are written for PHP4. ------------------------------------------------------------------------ [2006-01-25 23:51:05] sniper@php.net Please try using this CVS snapshot: http://snaps.php.net/php5.1-latest.tar.gz For Windows: http://snaps.php.net/win32/php5.1-win32-latest.zip ------------------------------------------------------------------------ [2006-01-25 21:00:06] chris at mysociety dot org Description: ------------ When php is running as a FastCGI server, it should handle SIGTERM and execute a graceful shutdown, because the mod_fastcgi process manager uses this signal to shut down excess dynamic servers. "Graceful" in this context means that the server must finish handling any outstanding request and then terminate its connection to the web server. Presently, PHP does not handle SIGTERM at all in the FastCGI request loop, so that on receipt of SIGTERM, php aborts, even if half-way through a request. This typically results in the web server logging "Incomplete headers (0 bytes) received", and returning "500 Internal Server Error" to the client. I have prepared a patch for this issue, here: http://bitter.ukcod.org.uk/~chris/tmp/20060125/php-4.3.10-fastcgi-sigterm-fix.patch -- this is against 4.3.10, but it applies to 4.4.2 fine. Reproduce code: --------------- Put the following in a .php file: <? sleep(1); print "fish"; ?> and set it up to run under FastCGI. Set up a script to request the page in a loop ("while : ; do wget -O/dev/null -q http://path/to/script ; done") Then kill the PHP process (pkill php will do this under Linux). If you catch the script in the sleep, this will result in the client getting a 500 and the web server logging a message about php failing, as describe above. Expected result: ---------------- No errors. Actual result: -------------- 500 error. ------------------------------------------------------------------------ -- Edit this bug report at http://bugs.php.net/?id=36158&edit=1