See also bug 1046165. The plugin process is being killed for syscalls that aren't seen on non-LSan runs. I don't know if this is due to *Sanitizer instrumentation itself, or some other config option used with it, or simply a timing dependency (some other thread doing things that aren't ordered with respect to sandbox startup). No clues from the buildbot runs because LSan disables crash reporting, so I'll have to repro locally.

Turns out it's just the __sanitizer thread creation hook — it wants to get the stack bounds (which uses pthread_getattr_np, which calls sched_getaffinity but that part of the pthread_attr isn't used) and set an alternate signal stack. So this will be simple.

Going to fold this into bug 1012951, because it's not a larger issue as I'd initially thought, and it just needs a small modification to the (now backed-out) GMP sandbox policy.