pthread stack cache and valgrind helgrind false positive race error on tls variables

[Roland McGrath]

Tuning the stack cache size is one of the poster children for our medium-
to long-term need ("plan" would be overly generous at this stage) to figure
out a general tuning mechanism.  When we have such a mechanism, then this
particular parameter will be a trivial instance of it.  But we won't have
that very soon.

Separate from the implementation details of the workaround you propose, I
think we would like to understand more clearly what is going on to confuse
helgrind.  Though it's all unclear to me so far, I tend to suspect both
that this workaround is not what we'd consider the ideal fix and that the
underlying problem is likely to have other symptoms (for other parts of the
code, other usage patterns, and/or other code outside of libc that rolls
its own synchronization primitives).


Thanks,
Roland