Rethinking the futex API

The library implementation doesn't get to decide this: Application code can either use pthread_cond_signal to wake exactly on thread or pthread_cond_broadcast to wake all threads. With the naive implementation, the kernel goes through a linked list and wakes everything while tearing the list apart. All the awoken threads exit the kernel and compete for ownership of the same mutex. One of these threads will end up getting the mutex, the others all do system calls in order to enter the kernel again where kernel code than builds a new linked list of them which is attached to a different futex.

That's an insanely complicated way to transfer a linked list from a head pointer a in the kernel to another head pointer b in the kernel. It's much simpler if the kernel just wakes one thread and transfers the existing list to the other futex.