[Lldb-commits] [lldb][PATCH] Clear thread unwinder when stack is reset
jmolenda at apple.com
Mon Oct 1 23:38:32 PDT 2012
Thanks for submitting the patches Andy. This part of the unwinder is a little fragile and hard to see the dependencies correctly - I'm not yet sure what I think here.
Normally a thread has an Unwind object (usually an UnwindLLDB object). The UnwindLLDB object provides the Thread with RegisterContexts up the stack -- it provides register values for all the stack frames on a thread. So if you want to get rip on frame 4, the Unwind will give you a RegisterContext for frame 4 that provides the value, assuming it is retrievable (and it will usually do this by seeing where frame 3 stored the value, going down the stack frames until it finds someone who saved it.)
The problem is this: The Thread has an Unwind object in the m_unwinder_ap, and the Unwind object has an array of RegisterContexts that it is responsible for. But IIRC the StackFrames also have pointers into these RegisterContexts. If we're throwing away all known StackFrames for a Thread, it should be safe to reset the Unwind object. But if we're *restoring* stack frames that were saved aside earlier, I worry that they may have pointers to RegisterContexts that have since been freed or something.
(and to further complicate things, I think the RegisterContextLLBB's refer to their "owning" UnwindLLDB object to retrieve values).
This is probably all done with refcounted shared pointers and will "just work" but it's a complicated enough interdependency that I need to look at it a bit more closely before I'm confident about what the right answer is here. I haven't looked at this part of the code in probably a year or so.
On Oct 1, 2012, at 1:42 PM, "Kaylor, Andrew" <andrew.kaylor at intel.com> wrote:
> In debugging a problem with expression evaluation after an inferior function call, I tracked it back to the Thread object’s m_unwinder_ap member not being cleared when the Thread state was restored from a checkpoint. I was able to fix the problem by inserting a call to Unwind::Clear in the Thread::ResetFrameZeroRegisters function. I thought that it might make sense to clear the unwinder every time the thread’s stack was cleared (i.e. from Thread::ClearStackFrames), but I wasn’t sure how that is used and whether it always indicates an actual stack change that the unwinder needs to know about.
> The problem in question showed up in the BasicExprCommandsTestCase.test_many_expr_commands test on x86_64 Linux.
> I’m attaching two patches, one with each implementation, though only one is needed. Can someone offer some guidance as to which approach is preferred?
> lldb-commits mailing list
> lldb-commits at cs.uiuc.edu
More information about the lldb-commits