Re: lin-thread cannot handle thread exit

[Michael Snyder <msnyder at cygnus dot com>]

Mark Kettenis wrote:
> 
> Hi Michael,
> 
> The thread_db assisted debugging code doesn't handle exiting threads
> properly, at least in combination with glibc 2.1.3.  There are at
> least two problems that prevent this from working:
> 
> 1. In lin-thread.c:enable_thread_event_reporting(), GDB requests
>    TD_DEATH events to be reported, and sets a breakpoint at the
>    appropriate location.  The problem is that the LinuxThreads
>    implementation included with glibc 2.1.3 triggers that breakpoint
>    after it has flagged the thread as terminated.  As a conseuqence
>    when GDB tries to fetch the registers for that thread it doesn't
>    succeed, and GDB complains about a breakpoint at location 0x0 that
>    it doesn't know of.
> 
> 2. If I disable the reporting of TD_DEATH events, things still don't
>    work.  The problem is that when the thread really exits, a
>    TARGET_WAITKIND_EXITED event is reported to GDB in.  So the answer to
>    the question ``Can I get this event mistakenly from a thread?'' in
>    lin-thread.c:thread_db_wait() is ``Yes''.  Since the
>    TARGET_WAITKIND_EXITED event is passed on to other code in GDB, GDB
>    thinks that the entire process has died.
> 
> Ignoring those "spurious" TARGET_WAITKIND_EXITED events doesn't help
> since some parts of GDB still think that the thread is still alive
> then.
> 
> Looking through the code, it seems that there quite a few "loose
> ends".  Is this still "work in progress"?

Hi Mark, 

The lin-thread / thread_db module was completed under time 
pressure and took much longer to get working than expected; 
plus of course it is a brand new native thread "port" 
(although I had access to Eric Paire's work as an example).
So no, it isn't formally regarded as a "work in progress", 
except in the sense that GDB itself is a work in progress, 
but it isn't exactly mature code either.

I appreciate your helping to find problems with it, and
I'd like to  know what else in the code you regard as a
"loose end".  One of my problems is that I'm not really
an experienced writer of multi-threaded apps -- I'm just
the person on the GDB team with the most experience with
GDB multi-thread debugging.

				Thanks, 
				Michael