GCJ Thread Dump

Bryce McKinlay mckinlay@redhat.com
Thu Feb 10 18:04:00 GMT 2005


David Daney wrote:

> As I said earlier It would probably have to be done from something 
> similar to the GC's stop-the-world.  You would print a stack trace for 
> each thread.
>
> The problem I see is that for targets that have no frame pointer.  
> These have to use the DWARF .eh_frame information for generating 
> stacktraces.    The stop-the-world uses asynchronous signals and these 
> are not compatible with DWARF exception handling.


On most platforms, we can unwind through signal handlers thanks to 
MD_FALLBACK_FRAME_STATE_FOR - this is how NullPointerException is 
implemented.

>   It might be possible to generate enough eh_frame data to allow 
> getting a stack trace from any instruction, but without adding 
> everything needed for full exception handling (I don't really know not 
> being an expert in this area).
>
> We could compile all GC/libgcj/etc libraries with 
> -fnon-call-exceptions, but what would happen for threads blocked in 
> libc or system calls?  I think those might be difficult to backtrace 
> through.


Ah, good point - presumably this is what Andrew was referring to. 
MD_FALLBACK_FRAME_STATE_FOR would fail if, say, we signaled a thread 
during a glibc leaf function call which had no unwind info. Non-Linux 
systems may not have unwind info for their system calls at all.

Bryce



More information about the Java mailing list