Re: [RFA/testsuite] attach.exp: Add small delay in busy loop...

[Michael Snyder <msnyder@redhat.com>]

Daniel Jacobowitz wrote:
> On Tue, Nov 18, 2003 at 04:26:07PM -0800, Michael Snyder wrote:
> 
>>Joel Brobecker wrote:
>>
>>>Hello,
>>>
>>>The attach.exp sometimes fails on certain platforms (eg mips-irix),
>>>and causes an attach process to be left behind. Since it is doing a busy
>>>loop, this runaway process left behind consumes 99.9% of the CPU,
>>>and considerably slows down the execution of the rest of the testsuite.
>>>
>>>I suggest the following change to add a small delay at each iteration
>>>of the busy loop. I had to make some adjustments to attach.exp:
>>>
>>> a. Line number 19 became line 32.
>>>    Just like Elena recently upgraded a test to avoid hard-coded
>>>    line number, we should probably clean this up, someday. This can
>>>    be a separate patch, however.
>>>
>>> b. The program was attached to while inside the busy loop, so the
>>>    test was expecting the debugger to report that the inferior was
>>>    inside function main() after the attach command was performed.
>>>    This is no longer the case, since the inferior is most likely
>>>    inside a system call, doing the delay. I felt that it was not
>>>    a necessity to checke where the debugger thought the inferior
>>>    was stopped, so removed that part of the expected output. What
>>>    I can do is add an extra test that does a backtrace and verifies
>>>    that it contains a frame for function main().
>>>
>>>2003-11-18  J. Brobecker  <brobecker@gnat.com>
>>>
>>>       * gdb.base/attach.c: Add small delay in busy loop.
>>>       * gdb.base/attach.exp: Make some associated adjustments.
>>>
>>>OK to apply?
>>
>>Seems to work on Linux.  I'd sure like to see that backtrace test,
>>though, to confirm that we are able to build a meaningful machine
>>state after we attach.
> 
> 
> Seems reasonable to me.  Warning: this will be yet another place we
> backtrace from syscalls, and sometimes we just can't do that.  We
> already have a couple of configurations where GDB can't reasonably be
> expected to backtrace out of nanosleep.

Well, we could run to a known breakpoint and then backtrace --
but that wouldn't test the state immediately after attach.
Can you think of a good test for a valid state, other than
a backtrace?