Jan Andres created MODCLUSTER-381: ------------------------------------- Summary: [Patch] Stability and portability issues in native code Key: MODCLUSTER-381 URL: https://issues.jboss.org/browse/MODCLUSTER-381 Project: mod_cluster Issue Type: Patch Affects Versions: 1.2.6.Final Reporter: Jan Andres Assignee: Jean-Frederic Clere Attachments: mod_cluster-1.2.6-20140109.diff Hi, Please find attached a patch for some issues I came across while deploying mod_cluster for Apache on Solaris/SPARC. The patch applies to the 1.2.6 release and addresses a total of 3 issues: 1. A memory leak in add_balancers_workers(). The current code will allocate new memory from the config pool for the balancer's "sticky" parameter on each update but throw away the pointer (thus losing memory) if the value generated equals the balancer's current value. The patch changes this to first do the apr_psprintf() from a temporary pool and only apr_pstrdup() this temporary string into the config pool in case it differs from the old value and actually needs to be updated. 2. Memory alignment issues. Some parts of the code use custom memory alignment logic and will generally align memory blocks to a multiple of sizeof(void*). This poses a problem on some architectures such a 64-bit SPARC CPU running a "32-bit application" which really means the process has a 32-bit memory space (i.e. sizeof(void*) is 4), but can still use native 64-bit instructions to operate on data. On SPARC and some other architectures, these instructions require 8-byte data alignment and will lead to a Bus error if this is not observed. The patch replaces the custom alignment logic by invocations of APR_ALIGN_DEFAULT() which also makes the code somewhat more concise and readable. 3. Heap corruption issues on process exit, sometimes resulting in core dumps, because mod_proxy_cluster's watchdog thread never terminates and continues to run while the main thread is already freeing up resources and destroying all the memory pools. The patch adds a mechanism to signal the watchdog thread to terminate and wait for it to exit prior to the freeing of the memory resources. This mechanism is invoked via a pre-cleanup hook on the config pool. Regards, Jan -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira