Several customers are experiencing a similar problem.  When shutting down a database engine, not all of shared memory is released when hugepages are in use.  This is occurring on systems running Oracle and Sybase as well as on RHEL 5.3 and 5.4.

The amount lost varies from system to system.  But the loss is consistent on each system.  In other words, if the loss is 1 gigabyte, it will always be 1 gigabyte on that system.  Each time the database is started and later stopped, more memory is leaked until there is not enough available to start the database.

The ipcs command shows the shared memory segment as having been released after the database shutdown is complete, but the HugePages_Free field in /proc/meminfo does not show all the memory having been released.

Attempting to manually force the pages to be released by echoing 0 into /proc/sys/vm/nr_hugepages does not seem to have any effect on this issue.

This is the same problem as bz593131 but on 5.4.