[lldb-dev] Can Process hold a TargetSP instead of a Target&?
Zachary Turner via lldb-dev
lldb-dev at lists.llvm.org
Fri Aug 28 14:18:24 PDT 2015
We already do this in DoDestroy(), but it looks like for whatever reason
DoDestroy is not getting called on us even though the Target is being
destroyed. Or maybe it is and our DoDestroy is getting into some edge
condition that doesn't cleanup correctly. But it's hard to debug because
it only happens from the test suite, and only when the system is under
heavy load (i.e. you run the entire test suite, and not just one test).
In the future I had planned to make an option for dotest that would allow
lldb to write full logs of every run during the test suite, so we could see
the sequence of events that are happening, but it's a bigger task.
A weak_ptr would work just as well and avoid the problem you describe, so
I'll wait and see if anyone has an issue with that.
On Fri, Aug 28, 2015 at 2:01 PM Pavel Labath <labath at google.com> wrote:
> I think it should be a weak_ptr if anything. Target already holds a
> shared_ptr of the process, and you will get pointer loops otherwise.
> Couldn't you make sure the debug thread exits (and processes all
> events) before the Target gets deleted (e.g. shut it down in
> Process::Finalize() or somewhere...)? If there is currently an
> invariant that Process should never outlive its Target (which would
> seem to be implied by the fact it holds a Target&), I would prefer if
> it can be preserved.
> On 28 August 2015 at 19:34, Zachary Turner via lldb-dev
> <lldb-dev at lists.llvm.org> wrote:
> > We've been seeing races during shutdown of inferiors for months, and I
> > finally tracked it down to the fact that Process holds a Target&. When
> > inferior is exiting on Windows, we will get a notification of this and we
> > try to do various cleanup related with the target. But there are times
> > where the Target gets deleted even when the Process is still around, due
> > some interactions between our debug loop and the timing of when certain
> > debug events that get sent by the operating system.
> > As a result, the race leads to us getting one of the notifications from
> > OS and us trying to access the target, which is stored by reference
> > to a crash.
> > It seems like a purely mechanical change to make Process hold a TargetSP
> > instead of a Target&. I've already started down this patch locally, but
> > want to make sure there are no objections or concerns before I continue
> > this path, since it's kind of mundane work.
> > _______________________________________________
> > lldb-dev mailing list
> > lldb-dev at lists.llvm.org
> > http://lists.llvm.org/cgi-bin/mailman/listinfo/lldb-dev
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the lldb-dev