[PATCH][BZ #19329] Fix race between tls allocation at thread creation and dlopen

[Torvald Riegel]

On Mon, 2016-01-18 at 16:00 -0200, Adhemerval Zanella wrote:
> I would say we attack one problem at time, even when they are related.

I think we should at least have something of a high-level plan for how
to fix related problems.  One will need that eventually unless the
different parts are indeed orthogonal, so it's not wasted work.  But if
trying to fix one part of a group of related problems until the end
before looking at others, there's a risk of wasting time on details of a
solution (eg, reviews) that might not work if considering the whole
problem space.

So, I think the best approach here is to first understand what we
actually want to solve and provide in terms of behavior, then decide on
a high-level scheme for how to do this (ie, the major atomicity /
ordering pieces), and then implementing parts of this scheme one at a
time (or fixing/adapting current code for this).