Is test-and-set (or other atomic RMW operation) a privileged instruction on any architecture?

Question

Hardware provides atomic instructions like test-and-set, compare-and-swap, load-linked-store-conditional. Are these privileged instructions? That is, can only the OS execute them (and thus requiring a system call)?

I thought they are not privileged and can be called in user space. But http://faculty.salina.k-state.edu/tim/ossg/IPC_sync/ts.html seems to suggest otherwise. But then, futex(7), under certain conditions, can achieve locking without a system call, which means it must execute the instruction (like test-and-set) without privilege.

Contradiction? If so, which is right?

Answer 1

That page is wrong. It seems to be claiming that lock-free atomic operations are privileged on ISAs in general, but that is not the case. I've never heard of one where atomic test-and-set or any other lock-free operation required kernel mode.

If that was the case, it would require C++11 lock-free atomic read-modify-write operations to compile to system calls, but they don't on x86, ARM, AArch64, MIPS, PowerPC, or any other normal CPU. (try it on https://godbolt.org/).

It would also make "light-weight" locking (which tries to take the lock without a system call) impossible. (http://preshing.com/20111124/always-use-a-lightweight-mutex/)

Normal ISAs allow user-space to do atomic RMW operations, on memory shared between threads or even between separate processes. I'm not aware of a mechanism for disabling atomic RMW for user-space on x86. Even if there is such a thing on any ISA, it's not the normal mode of operation.

Read-only or write-only accesses are usually naturally atomic on aligned locations on all ISAs, up to a certain width (Why is integer assignment on a naturally aligned variable atomic on x86?), but atomic RMW does need hardware support.

---

On x86, TAS is lock bts, which is unprivileged. (Documentation for the lock prefix). x86 has a wide selection of other atomic operations, like lock add [mem], reg/immediate, lock cmpxchg [mem], reg, and even lock xadd [mem], reg which implements fetch_add when the return value is needed. (Can num++ be atomic for 'int num'?)

Most RISCs have LL/SC, including ARM, MIPS, and PowerPC, as well as all the older no longer common RISC ISAs.

---

futex(2) is a system call. If you call it, everything it does is in kernel mode.

It's the fallback mechanism used by light-weight locking in case there is contention, which provides OS-assisted sleep/wake. So it's not futex itself that does anything in user-space, but rather lock implementations built around futex can avoid making system calls in the uncontended or low-contention case.

(e.g. spin in user-space a few times in case the lock becomes available.)

This is what the futex(7) man page is describing. But I find it a bit weird to call it a "futex operation" if you don't actually make the system call. I guess it's operating on memory that kernel code might look at on behalf of other threads that are waiting, so the necessary semantics for modifying memory locations in user-space code depends on futex.