SJSAS 9.0 (Sun Application Server) uses NIO for its HTTP Engine. Removing the workaround for 4953599 and running performance test on Solaris literally kill the VM, without any core/log. The problem seems to be that the C heap is corrupted, so lots of threads stack up trying to get native memory from the C heap. Meanwhile, the compiler thread detects the corruption in the C heap and tries to throw a signal, and the signal handler dies -- so either the entire process dies without a core, or the process hangs.

The attachement contains a pstack log showing the signal thread hanging, and the corresponding jstack output for that.

We've tested with 1.5.0_0X and the problem happens in all version.
Posted Date : 2005-09-29 19:41:39.0

N/A

The submitter provided access to the machine and we located the fatal error logs and core file. The current crash is because java.io.File.list is used the non-MT safe readdir (it should be using readdir_r). This is tracked by 4621211. With this fixed we need to do further testing to see if that are other issues - the submitter has accumulated a varied selection of fatal error logs which suggest problems in other areas too.
Posted Date : 2005-09-30 21:55:45.0

--

Indeed there is a second issue. With a preliminary fix for 4621211 there is still random memory corruption. I've tracked it down to a timing bug in the DevPollSelectorProvider. With a preliminary for that the tests with the app server run to completion.
Posted Date : 2005-09-30 21:55:45.0