Salikh Zakirov wrote:
Hi harmonists,The latest combination of DRLVM and Classlib is highly instable on my machine (Gentoo Linux/i686 Pentium 4 HT), causing smoke tests fail frequently and in unstable manner. Fails are caused by either incorrect memory deallocation (caught by glibc), or by segmentation fault. Tests worked much better just a few days ago, so I tried to bisect commits to find the failure cause. Bisection pointed to the commit Author: Oliver Deakin <[EMAIL PROTECTED]> Date: Mon Oct 15 15:20:32 2007 +0000 [r584804] Adding DetachCurrentThread() call to main.c after discussion on dev list: http://mail-archives.apache.org/mod_mbox/harmony-dev/200710.mbox/[EMAIL PROTECTED] Is anyone else seeing this instability? Does anyone know what needs to be fixed in DRLVM to prevent it?
The change was supposed to affect only debugging scenarios since there was a bug that some JVMTI events were attempted to be sent in a thread that is no longer attached to VM.
I've recently discovered some worse scenarios on shutdown related also to glibc complaining about corrupted allocation list and various segfaults, but all of them are related to shutdown with JDWP agent (see HARMONY-5019 - it took me a night to debug assembly code in glibc and libgcc_s to find out why pthread_cancel crashes). The problem existed before Oliver's commit and isn't connected to it in any way.
Could you give some details, typical stack traces? I could at least say if I saw them already or not (BTW smoke tests work for me). I am also running Gentoo on x86 (stable) and amd64 (stable) but I have only 1 CPU so I don't catch all of the race conditions related to SMP.
-- Gregory
