Any futex operation starts in userspace, but it may necessary to communicate with the kernel using the futex(2) system call.
To up a futex, execute the proper assembler instructions that will cause the host CPU to atomically increment the integer. Afterwards, check if it has in fact changed from 0 to 1, in which case there were no waiters and the operation is done. This is the non-contended case which is fast and should be common.
In the contended case, the atomic increment changed the counter from -1 (or some other negative number). If this is detected, there are waiters. Userspace should now set the counter to 1 and instruct the kernel to wake up any waiters using the FUTEX_WAKE operation.
Waiting on a futex, to down it, is the reverse operation. Atomically decrement the counter and check if it changed to 0, in which case the operation is done and the futex was uncontended. In all other circumstances, the process should set the counter to -1 and request that the kernel wait for another process to up the futex. This is done using the FUTEX_WAIT operation.
The futex system call can optionally be passed a timeout specifying how long the kernel should wait for the futex to be upped. In this case, semantics are more complex and the programmer is referred to futex(2) for more details. The same holds for asynchronous futex waiting.
To reiterate, bare futexes are not intended as an easy to use abstraction for end-users. Implementors are expected to be assembly literate and to have read the sources of the futex userspace library referenced below.
This man page illustrates the most common use of the futex(2) primitives: it is by no means the only one.
Futexes were designed and worked on by Hubertus Franke (IBM Thomas J. Watson Research Center), Matthew Kirkwood, Ingo Molnar (Red Hat) and Rusty Russell (IBM Linux Technology Center). This page written by bert hubert.
Initial futex support was merged in Linux 2.5.7 but with different semantics from those described above. Current semantics are available from Linux 2.5.40 onwards.