The main use of spinlocks that I'm aware of is dealing with interrupt handlers in driver code. In this situation you generally can't go to sleep (sleeping with interrupts disabled is generally a good way of never waking up) so calling "mutex_lock" is simply out of the question. That's probably a niche use but that's literally the only situation I've ever used actual spin locks instead of mutexes, mainly for the reasons outlined by TFA.
Most of the time I would say if lock performance impacts your application you are likely doing something wrong. There are many ways of solving typical synchronization problems without locks and if you need locks there are many ways to amortize costs.
Mutexes are not faster than spin locks the same way slower cars are not faster than fast cars. You might still crash in a fast car and be slower to the end of the race (to the supermarket) but that is just your failure to use the power you were given responsibly.
The feature article ; the freaking article ; the friendly article ; the f$%^&*( article, c.f. RTFM.
And yeah, I believe that it derives from "RTFA" (read the fucking article) which would be what you told people who obviously commented without reading the article.
the F is usually something else though ...
Also, the disruptor design is based on a system where there is a 1:1 correspondance between threads and cores. If this is not the case it will likely still interact somewhat poorly with the OS's scheduler without any blocking at all, for the same reason as spinlocks (in fact, if you use the library you can customize this behaviour).
Random Google search seems to validate that.
https://stackoverflow.com/questions/1383363/is-my-spin-lock-...
The highest performance design is to use a ringbuffer. To write to the ringbuffer, you atomic-increment the "claim" counter, giving you an index. Now take index modulo ringbuffer size. That's the slot to write to. After you're done writing to it, set your "publish" counter to the index.
Each writer has a "publish" counter, and the "claim" counter isn't allowed to move beyond the lowest "publish" counter modulo ringbuffer size.
Each reader uses a similar strategy: the reader has a current counter. You find the minimum of all publish counter values, then process each slot up to that value, and set your counter to that value. The "claim" counter isn't allowed to move past the minimum of all reader counters.
Hence, everyone is reading from / writing to separate cache lines, and there are no locks at all. The sole primitive is atomic increment, which can never fail. (The only failure condition is "one of the slots hasn't been processed yet" (i.e. the publish counter is <= claim counter minus ringbuffer size) at which point everybody waits for the slot to be processed.
You can wait using multiple strategies: a spin loop (which isn't the same as a spinlock because you're only reading a value in memory, not calling any atomic primitives), yielding CPU time via sched_yield(), or calling sleep. All strategies have tradeoffs. Calling sleep is the least CPU intensive, but highest latency. Vice-versa for spin loop.
Takeaway: there are no locks. Just increments.
I'd be interested to see benchmarks showing that spinlocks with CAS have similar throughput to a ringbuffer with atomic increment.
Note that with the ringbuffer approach, each reader can process many slots at once, since you're taking the minimum of all published slot numbers. If you last processed slot 3, and the minimum publish count is 9, you can process slot 4 through 9 without doing any atomic operations at all. The design guarantees safety with no extra work.
It's not a minor trick; it's one of the main reasons throughput is orders of magnitude higher.
Benchmarks: https://github.com/LMAX-Exchange/disruptor/wiki/Performance-...
Beyond that, the ringbuffer approach also solves the problem you'll always run into: if you use a queue, you have to decide the max size of the queue. It's very hard to use all available resources without allocating too much and swapping to disk, which murders performance. Real systems with queues have memory usage patterns that tend to fluctuate by tens of GB, whereas a fixed-size ringbuffer avoids the problem.
Also as far as I'm aware, at least on intel all atomic operations take pretty much exactly the same number of clock cyles (except for CAS and DCAS which are ~10% and 50% more expensive, IIRC)
For context: before Pentium with its glueless 2x2 SMP/redundancy support there were various approaches to shared memory x86 multiprocessors with wildly different memory coherence models. (And some of the “lets design a board with eight 80386” are the reason why Intel had designed i586 to be glueless and such systems are probably still used to this day, althought unsupported)
As forward progress is a pretty basic requirements, in practice even LL/SC platforms in practice do that, but is instead of having a single instruction with guaranteed forward progress you have to use a few special (but sometimes underspecified) sequences of instructions between the ll/sc pairs.
FWIW RISC-V guarantees forward progress for reasonable uses:
> We mandate that LR/SC sequences of bounded length (16 consecutive static instructions) will eventually succeed, provided they contain only base ISA instructions other than loads, stores, and taken branches.
what happens if those 16 instructions touch 16 different cache lines? I'm not an hardware expert (and even less on coherency protocols), but I think it would be extremely hard to make sure livelocking is avoided in all cases, short of having some extremely expensive and heavy handed global 'bus' lock fallback.
That's quite the corner case.
What if your work queues are running on a multitasking operating system that runs services? And what about a hypervisor?
If you’re not sharing cores between VMs it’s typical to do the same at the hypervisor layer.
I wrote some software at an automated trading firm that would do the subset of "parsing ticks and maintaining order books" sufficient to map ticks to symbols, and no more work than that. My code was pinned to one core and spun waiting for an expensive network card to DMA some ticks into memory, then when ticks we cared about were found it would load them onto one or more queues. Each queue corresponded to a different thread spin waiting on its own pinned core.
So one use case is "transforming ticks into trading decisions slightly faster than you otherwise could"
Whether there is a use case for spinlocks outside low-latency trading, i have no idea!
The main use of spinlocks that i'm aware of is minimising latency in inter-processor communication.
That is, if you have a worker task which is waiting for a supervisor task to tell it to do something, then to minimise the time between the supervisor giving the order and the worker getting to work, use a spinlock.
For this to really work, you need to implement both tasks as threads pinned to dedicated cores, so they won't be preeempted. You will burn a huge amount of CPU doing this, and so it won't be "faster" from a throughput point of view. But the latency will be as low as it's possible to go.