Spin locks are discouraged in FreeBSD kernel code too, except for very low-level areas (serial console, pmap, and interrupt handling come to mind) where you can't safely ask the scheduler to do anything for you.

That said, our sleep locks are actually adaptive locks, so on SMP systems they'll spin for a bit before giving up and asking to be descheduled.

They are encouraged in the Linux kernel in places where it is nessesary to lock for a very short time. Using mutexes takes a bit of overhead, so they can't give you very short sleep times; that is when spinlocks have to be used.

http://kernel.org/doc/Documentation/spinlocks.txt

In FreeBSD the vast majority of sleep lock acquisitions are uncontended and end up being a single compare-and-exchange -- very fast. Spin locks are slightly more work, since we disable interrupts during them.

Exactly

Spinlocks in linux are the "last resource locks", usually in kernel code only in situations where you can't have the scheduler kicking in but can have a race condition (in SMP systems usually)

So it's basically interrupt handlers.

It sounds smart but it occurred to me that that I can only think of two situations where a spinlock needs to be promoted. 1. Too many cores access that tiny portion of the system at the same time or 2) a bug(like using a spinlock where something else should be used and a more old-fashioned bug).

Will 1) really happen often enough to not use simple spin locks?

As an example of where you might use a user space spinlock, consider:

If you have:

… a multithreaded process with shared resources that need locks

… and for the uncontested case your spinlock is much faster than your next fastest lock

… and you have resources for which the probability of a contested access is very close to zero

… and the locks are only held briefly (relative to compute quantum) (especially if they are held shorter than thread context switch)

You may find that the speed gained by using the spinlock is a win even if once in a while you end up burning your compute quantum because a thread got suspended holding a spinlock.

I agree with this assessment. Where the probability of collision is low, and spins are in that case of short duration anyway, and especially if test and set intrinsics are available, spin locks in userspace can be very valuable. They also, besides saving scheduler cycles, require only a common word access, even while that implicates barriers and cache lines.

You need to be careful though.

Indeed spinlocks should be used carefully. But there are some places where you can't avoid them.