Crash in eclipse (JIT bug)

[Ramon Garcia]

The symptom is that eclipse crashes silently (no log, no message, no hs_.....log file in home directory). Only a core file in home directory. It happens when the m2eclipse plugin from sonatype is installed, and it fetches indexes from maven servers. A workaround is to disable the Internet connection while eclipse starts.

This is a Ubuntu 10.04 x64 system completely up to date.

Looking at the core file, this looks like a code generation bug. Perhaps the latest update to openjdk broke something.

gdb -c ~/core
(gdb) print $rsp
$9 = (void *) 0x1

Obviously the stack pointer is corrupted. Probably this is the reason why no hs_.....log is produced. Calling the crash handler is imposible.

If we disassemble a few instructions above:

(gdb) x /20i $rip - 52
   0x7f12a956bd3b: je 0x7f12a956bd46
   0x7f12a956bd41: jmpq 0x7f12a955d330
   0x7f12a956bd46: mov -0x38(%rbp),%r13
   0x7f12a956bd4a: mov -0x30(%rbp),%r14
   0x7f12a956bd4e: retq
   0x7f12a956bd4f: mov -0x40(%rbp),%rsp
   0x7f12a956bd53: movq $0x0,-0x10(%rbp)
   0x7f12a956bd5b: movabs $0x7f12ae146988,%rsi
   0x7f12a956bd65: movabs $0x0,%rdx
=> 0x7f12a956bd6f: callq 0x7f12a956bd79
   0x7f12a956bd74: jmpq 0x7f12a956bdfa

We see that the instruction mov -0x40(%rbp),%rsp overwrites the stack pointer. We confirm this by looking at the assigned value:
(gdb) print {long} ($rbp - 0x40)
$13 = 1

[Ramon Garcia (ramon-garcia-f+ubuntu)]

I am keeping the core file, in case futher details are required.

[Robin Stocker (nibor)]

I also had this on Fedora 13 and just now on an Ubuntu 10.04 x86_64 machine. It happens just when M2Eclipse is doing "Updating indexes".

[Vreixo Formoso (metalpain2002)]

A simple workaround is to disable respository index updates on startup. Just go to Window->Preferences->Maven and uncheck that option. Everything seems to work after it, but I have just find that option, so I have not tested it too much...

[Matthias Klose (doko)]

is this seen with the eclipse version distributed in Ubuntu?

[wdesmet (kromagg)]

I just tested it. It is indeed also a problem with eclipse as distributed in ubuntu (not that there was any reason to think otherwise).

[Ramon Garcia (ramon-garcia-f+ubuntu)]

The recently issued openjdk update, 6b18-1.8-4ubuntu3, no longer shows the problem.

(By the way, the discussion about the eclipse version used is pointless. A Java virtual machine cannot crash. The code causing the crash was JIT code. It was a code generation bug)