Re: [PATCH 9/9] gdb: use reopen_exec_file from reread_symbols

[Andrew Burgess <aburgess@redhat.com>]

Tom Tromey <tromey@adacore.com> writes:

>>>>>> "Andrew" == Andrew Burgess <aburgess@redhat.com> writes:
>
> Andrew> I didn't fully understand all the discussion w.r.t. gnulib stat vs
> Andrew> system stat, but I'm hoping the change above, which is less that your
> Andrew> originally proposed full change, might be mergable.
>
> FTR (and IIUC) the issue is that gnulib's stat yields different times on
> Windows hosts, because it tries to handle the timezone -- on Unix, stat
> reports in UTC, but on Windows it may report in local time.
>
> The result is that BFD will report times that are offset from the times
> reported by gnulib.
>
> gnulib doesn't provide a way to change this behavior, so one proposal in
> the thread was to hack gnulib.
>
>>> I don't think gdb has a very clear idea of when bfd_cache_close_all
>>> ought to be called.
>
> Andrew> I agree this stuff doesn't seem well defined at all, but I don't think I
> Andrew> made anything particularly worse in this respect, so I'm leaving that as
> Andrew> a problem for the future.
>
> One thing I wonder is whether bfd_cache_close_all is even needed.
> gdbserver doesn't do this and AFAIK, gdb just keeps all those remote fds
> open.
>
> Maybe we're missing some testing scenario, though, like gdbserver
> extended-remote with "gdb --write".

So, I touched a bfd_cache_close_all recently[1], and dug into the history
of these calls a little at the time.  They were first added to GDB in
this commit:

  commit ce7d45220e4ed342d4a77fcd2f312e85e1100971
  Date:   Fri Jul 30 12:05:45 2004 +0000

Before this there were no bfd_cache_close_all calls in the GDB tree.

The mailing list thread associated with this commit is:

  https://sourceware.org/pipermail/gdb-patches/2004-July/036407.html

Though this thread seems to reference some older thread that I've not
been able to track down.  The above thread is titled:

  [RFA] win32: bfd_cache_close after kill

Here is the text from the older (unfound) thread that is quoted:

  > When the inferior is killed, it is safe the release the different file
  > handles that BFD keeps open. It is particularly useful on Win32 (and
  > presumably on HP UX) to be able to recompile and restart a new debugging
  > session without quitting GDB...

Now I've never done any development work on Win32, I very occasionally
boot a Windows machine to do a test build of GDB, but beyond that my
knowledge of Windows is like 20+ years old :-/  However, I do have a
vague memory that Windows didn't like you writing to a file that was
opened by some other application?  Maybe this memory is wrong, but that
would seem to align with the above text.

If that is the case, then it would seem (to me) that we want to ensure
BFD is not holding open any files at a time when GDB itself is not
actively doing anything that might read from a file, this would be:

  1. When GDB is sat idle at a prompt, and

  2. When GDB is blocked waiting for an inferior event.

Having a bfd_cache_close_all call in e.g. reopen_exec_file, seems pretty
pointless really, as it's not impossible that GDB will then immediately
re-open the BFD from some other function.

I wonder if we should remove all bfd_cache_close_all calls, and just
have two calls associated with the two states above?

[1] https://inbox.sourceware.org/gdb-patches/6219a1ae148c0b9796f7153b9a6c3b7173fb8f5a.1694706545.git.aburgess@redhat.com/

Thanks,
Andrew