This may seem a bit backwards: we take nice C++ list<string> and do the
C dance.  It's a bit defensive: this class is used in the segv handler
(in the backtrace() and backtrace_symbol() path), so we want to minimize
the work we do on the heap in that case.  (For the list<string> path,
we can do whatever we like.)

Signed-off-by: Sage Weil <sage@newdream.net>

Signed-off-by: Sage Weil <sage@newdream.net>

Extend handle_pyerror() to generate a crash dump.  Pass some additional
context through from the callers (including the ability to not generate
a crash dump in the CLI handler case).

Extra crash dump fields look like so:

    "backtrace": [
        "  File \"/home/sage/src/ceph/src/pybind/mgr/balancer/module.py\", line 652, in serve\n    self.ifail()",
        "  File \"/home/sage/src/ceph/src/pybind/mgr/balancer/module.py\", line 648, in ifail\n    raise RuntimeError('test')",
    ],
    "mgr_module": "balancer",
    "mgr_module_caller": "PyModuleRunner::serve",
    "mgr_python_exception": "RuntimeError",

Notably, the backtrace deliberately excludes the 'value' of the exception,
as that may leak identifying information about the system.  Instead, we
only include the exception *type* and the portion of the traceback that
identifies the call path (where in the code we crashed).

Also note: a side-effect of this change is that module exceptions will
trigger cluster health warnings about daemon crashes.

Signed-off-by: Sage Weil <sage@newdream.net>

Generate a different warning for crashes in mgr module python code, as
they do not mean that the entire mgr daemon crashed.  Document.

Signed-off-by: Sage Weil <sage@newdream.net>

Broken by 8c009e2

Signed-off-by: Sage Weil <sage@newdream.net>

Include the exception value in teh crash dump, but redact it in telemetry.
That way the operator can see it (it's useful info!) but we don't risk
sharing identifying data via telemetry.

Signed-off-by: Sage Weil <sage@newdream.net>