Fix make resource failure when cc parallel is enabled

[NobodyXu]

Fixed #962

[NobodyXu]

Verified locally by building cargo-near, that this PR indeed fixed #962

[weihanglo]

Thank you for your hard work chasing this down!

Just fly-by. Could we have one or two simple paragraphs describing what was wrong and how this PR gets it correct?

[NobodyXu]

Yes, so the problem is that make < 4.3 pass jobserver using annoymous pipe fd, and thus they share the file description between parent and child processes, e.g. flags O_NONBLOCK are shared among make, cargo, build.rs, etc.

When cc::Build::compile in one build.rs sets O_NONBLOCK on the pipe, all other processes, including make, will also affected by this change.

Since make < 4.3 is not able to handle jobserver pipe with O_NONBLOCK set, it will fail.

make >= 4.3 uses O_NONBLOCK flag itself and it supports the named fifo mode, which is more robust than annoymous
pipe, because they don't share the file description.

Edit:

This PR fixed this, by reverting to use crate jobserver, which does not set the O_NONBLOCK and spawns a thread to avoid blocking instead.

[NobodyXu]

Note that jobserver currently has a problem: It doesn't clone the anonymous pipe passed via env.

So if someone calls Build::compile, creates a jobserver::Client and drop it, then calling Build::compile again will use the now closed pipe fd.

Since fd is reused, this end up creates some really strange bug (this happened in opencv before twistedfall/opencv-rust#480 , they fixed it by using jobslot, which does clone the annoymous pipe fd)

I've created rust-lang/jobserver-rs#57 but it has not been merged yet

Plus jobserver now has msrv 1.63, and cc is still on 1.56

[n-eq]

Confirmed this fixes the issue on our side, both for MacOS and iOS targets. Thanks @NobodyXu ! Looking forward to the next release