Re: [Fwd: Re: [Fwd: Re: gdb/725: Crash using debug target and regcaches (in 5.3 branch?)]]

[Daniel Jacobowitz <drow@mvista.com>]

On Tue, Dec 03, 2002 at 02:20:17PM -0500, Andrew Cagney wrote:
> 
> >>Er, no I wont :-(
> >>
> >>The attached is the refind patch.  I added the comment:
> >>
> >>+  /* NOTE: cagney/2002-12-02: This assumes that the target code can
> >>+     directly transfer the register values into the register cache.
> >>+     This works fine when there is a 1:1 mapping between light weight
> >>+     process (LWP) (a.k.a. process on GNU/Linux) and the thread.  On
> >>+     an N:1 (user-land threads), or N:M (combination of user-land and
> >>+     LWP threading), this does not work.  An LWP can be sitting in the
> >>+     thread context switch code and hence, the LWP's registers belong
> >>+     to no thread.  */
> >
> >
> >First of all, this comment is wrong.
> 
> Why?
> 
> The code is assuming that the LWP registers belong to the currently 
> selected thread's regcache.  That is a pretty scary assumption.
> 
> [I'll use that wording]

It's not an assumption at this point.  proc-service.c:230 to
thread_db_fetch_registers is the only path into lin_lwp_fetch_registers.
And that does:
  inferior_ptid = BUILD_LWP (lwpid, ph->pid);

So at this point we _know_ that the thread we're querying has its
registers in the LWP.  That's the whole point.

> 
> > I think we're miscommunicating
> >on what the patch does.  At this point the fetch_inferior_registers
> >code has an inferior_ptid which looks like this:
> >  PID = pid, LWPID = 0, TID = 0
> >or
> >  PID = pid, LWPID = otherpid, TID = 0
> 
> >Don't get confused by the use of TIDGET.  Look at the definition of
> >TIDGET; it gets the _LWP_ id.  This's a search and destroy candidate if
> >I ever saw one.
> 
> I'll add that.
> 
> >Some upper layer has already taken the TID, mapped it to an LWP id, and
> >is asking for that LWP's registers by the time we get here.  So the LWP
> >is known to belong to the thread we are querying.
> 
> 
> >>however, with the patch applied, I see (and consistently, well 2 out of 
> >>2, which is pretty amasing for the thread testsuite) the new failure:
> >>
> >>
> >>gdb.threads/killed.exp: GDB exits after multi-threaded program exits 
> >>messily
> >>
> >>looking at the log file:
> >>
> >>(gdb) run
> >>Starting program: /home/cagney/gdb/native/gdb/testsuite/gdb.threads/killed
> >>[New Thread 1024 (LWP 6831)]
> >>[New Thread 2049 (LWP 6832)]
> >>[New Thread 1026 (LWP 6833)]
> >>Cannot find user-level thread for LWP 6833: generic error
> >>(gdb) PASS: gdb.threads/killed.exp: run program to completion
> >>quit
> >>The program is running.  Exit anyway? (y or n) y
> >>Cannot find thread 2049: generic error
> >>(gdb) FAIL: gdb.threads/killed.exp: GDB exits after multi-threaded 
> >>program exits
> >> messily (gdb/568)
> >>
> >>Which doesn't occure when the patch isn't applied.
> >
> >
> >Are you sure about this last bit?  I see this failure even without the
> >patch, on an i386 SMP system.  I just checked it moments ago.
> 
> Yes.  Not on an SMP machine though.

According to Michael it already shows up in all of his configurations
in current CVS... I see the same thing here.  It's a little timing
sensitive, I don't know why it didn't show up beforehand for you.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer