Bug #68486 [Com]: PHP SegFault zend_hash_find

Edit report at https://bugs.php.net/bug.php?id=68486&edit=1 ID: 68486 Comment by: php at bof dot de Reported by: wattwood at tcstire dot com Summary: PHP SegFault zend_hash_find Status: Open Type: Bug Package: Apache2 related Operating System: Ubuntu 14.04.1 LTS PHP Version: 5.5.19 Block user comment: N Private report: N New Comment: Production test of the one-liner patch, using Apache 2.4, ran flawlessly, now for five hours straight. Just to be sure it's good, I'll run a second build under Apache 2.2 + put that on one of my not-yet-upgraded production servers tomorrow. But I think this patch is good to go in. I added it now as a patch to this bug report, for that reason. Previous Comments: ------------------------------------------------------------------------ [2015-03-18 09:39:52] php at bof dot de Re: [2015-03-17 23:22 UTC] gmoniker at gmail dot com Jep, I think your one line fix to the end of php_handler(), fixes the immediate problem here. Initial testing confirms. I'll test in production later today. I didn't realize before (but now confirmed in apache sources), that apr_pool_cleanup_run() _cancels_ an exact same previously set cleanup callback. Reentry by virtual() still dumps core (even with ap_rflush removed there), as probably will the other funny reentry cases. But as far as I can see, all those can only be triggered by suitable PHP code, and not at will by external requests - so they are better handled separately in new bug reports. ------------------------------------------------------------------------ [2015-03-18 04:45:33] phpdev at ehrhardt dot nl I applied http://bei.bof.de/php/sapi_apache2.noreentry.v3.patch on my Centos6 development server with Apache 2.4.12 with PHP 5.6.6 as mod_php. Result: the segfaults did not happen anymore, so the patch seems to be OK. Limesurvey, Drupal7 and Piwik ran fine after applying the patch. Especially Piwik is worth mentioning because it seems to stress PHP to the max. After patch v3 it produced an occasional 'zend_mm_heap corrupted' (like it always does), but no segfault. ------------------------------------------------------------------------ [2015-03-17 23:22:47] gmoniker at gmail dot com It seems to me that it will be hard to revalidate this handler if it is changed extensively. Just did some testing with a one line adjustment to the code of "sapi/apache2handler/sapi_apache2.c". Towards the end of php_handler(request_rec *r) I included the line: apr_pool_cleanup_run(r->pool, (void *)&SG(server_context), php_server_context_cleanup); So it becomes: if (!parent_req) { php_apache_request_dtor(r TSRMLS_CC); ctx->request_processed = 1; bucket = apr_bucket_eos_create(r->connection->bucket_alloc); APR_BRIGADE_INSERT_TAIL(brigade, bucket); rv = ap_pass_brigade(r->output_filters, brigade); if (rv != APR_SUCCESS || r->connection->aborted) { zend_first_try { php_handle_aborted_connection(); } zend_end_try(); } apr_brigade_cleanup(brigade); apr_pool_cleanup_run(r->pool, (void *)&SG(server_context), php_server_context_cleanup); } else { ctx->r = parent_req; } It seems to me this solves the missing automatic request pool destruction after a request in Apache 2.4 leading the php_handler routine to consider independent follow-on requests a subrequest of the first request when they are not. The request pool can continue on, but without the php_struct in it. I am afraid the Apache 2.4 handling of these pipelines can cause memory bloat of the server process, but that belongs in Apaches corner. ------------------------------------------------------------------------ [2015-03-17 16:42:24] php at bof dot de ab@php.net : regarding reproducing the issue you can take any PHP script, a simple one-line echo is sufficient, and then request it twice using the echo|netcat command as shown in several of the comments here. ------------------------------------------------------------------------ [2015-03-17 16:38:32] php at bof dot de Sorry for the broken gist patches. Apparently gists do funny things with whitespace... I should have tested that, but didn't. To make things easy I put the patches up on my personal webspace. Please don't link to that in any blog posts or stuff like that :) Same patch as in the last gist, without and with added debug/trace calls: http://bei.bof.de/php/sapi_apache2.noreentry.v2.patch http://bei.bof.de/php/sapi_apache2.noreentry.debug.v2.patch Since then I also added something alluded to on internals in a new thread "PHP apache2handler virtual() function" - now implemented with a new PHP function apache_tail_request(). This can be found, on top of the changes of the v2 patch, in these patches: http://bei.bof.de/php/sapi_apache2.noreentry.v3.patch http://bei.bof.de/php/sapi_apache2.noreentry.debug.v3.patch All patches were made on top of PHP 5.6.7RC1, checked out by tag from git. For about 2 hours now, I have the v3 patch actively running on one of my production servers, under apache 2.4 now, and without any coredumps / segvs in the logs or any kind of performance regression. This is with some 80 requests per second, so it gets a good workout. The patch appears to work. ------------------------------------------------------------------------ The remainder of the comments for this report are too long. To view the rest of the comments, please view the bug report online at https://bugs.php.net/bug.php?id=68486 -- Edit this bug report at https://bugs.php.net/bug.php?id=68486&edit=1