complete the implementation of std.mutex on all platforms

[shawnl]

As threads can be suspended at any time (even kernel threads!), we need not just spinlocks, but blocking locks implemented with futexes and their equivalent. This is necessary to fix #1363

https://www.kernel.org/doc/Documentation/robust-futexes.txt
http://man7.org/linux/man-pages/man2/futex.2.html
http://man7.org/linux/man-pages/man7/futex.7.html
https://www.akkadia.org/drepper/futex.pdf

[shawnl]

Lock must be the same struct on all architectures so they can be shared between heterogeneous processes. Use-case: https://keithp.com/blogs/Shared_Memory_Fences/

Because musl and glibc do not do this correctly we can't be compatible with them. (although they could become compatible with us in the future)

[kristate]

WebKit guys have their ParkingLot implementation.

[shawnl]

in regards to "Concurrent Hashtable of Synchronized Queues" of the ParkingLot blog post: I am wondering how much spin-lock contention a concurrent red-black tree has, because that would eliminate out-of-memory and dynamic memory allocation in ParkingLot as the the rb-node could be stored on the stack. https://xuezhaokun.github.io/150-algorithm/

Also, it would allow WTF::Lock-type locks to be shared between processes. (through multiple instances of WTF::ParkingLot)

Wouldn't perform as well on NUMA systems however...

[andrewrk]

Lock must be the same struct on all architectures so they can be shared between heterogeneous processes

When a user wants a lock that is the same struct on all architectures or whatever, they should explicitly request it by using a type that makes this guarantee. The default std lib mutex should not make this guarantee.

[andrewrk]

I believe this is done now, thanks to @kprotty's work. If not, please open separate issues for each OS that is not supported.