Valgrind sometimes takes a very long time on some tests

[brson]

One of the bots has currently been spending 30 minutes valgrinding task-perf-word-count. This makes me think something is going terribly wrong with memory management. Possibly related to native stacks.

[brson]

This could be related to #1527, since I seem to be able to reproduce this behavior with optimizations on, but not off. We could be getting into a scenario where we repeatedly reallocate hunks of stack, pissing off valgrind.

[brson]

It could even be just the act of preparing the stack segments for valgrind that makes it so slow - it's a fairly expensive operation that involves setting a status bit for every byte of the stack.

[brson]

Actually, I can't reproduce this locally. I had a build take 12 hours last night on one of the bots.

[brson]

Some investigation leads me to believe this could be a result of some pathological locking behavior in rust_port_detach.

[brson]

I'm confident now that #1924 is the reason the linux bots randomly take ridiculous amounts of time.

[brson]

Fixed.