Re: [RFC PATCH] percpu system call: fast userspace percpu critical sections

[Josh Triplett <josh@joshtriplett.org>]

On Thu, May 21, 2015 at 10:44:47AM -0400, Mathieu Desnoyers wrote:
> Expose a new system call allowing userspace threads to register
> a TLS area used as an ABI between the kernel and userspace to
> share information required to create efficient per-cpu critical
> sections in user-space.
> 
> This ABI consists of a thread-local structure containing:
> 
> - a nesting count surrounding the critical section,
> - a signal number to be sent to the thread when preempting a thread
>   with non-zero nesting count,
> - a flag indicating whether the signal has been sent within the
>   critical section,
> - an integer where to store the current CPU number, updated whenever
>   the thread is preempted. This CPU number cache is not strictly
>   needed, but performs better than getcpu vdso.
> 
> This approach is inspired by Paul Turner and Andrew Hunter's work
> on percpu atomics, which lets the kernel handle restart of critical
> sections, ref. http://www.linuxplumbersconf.org/2013/ocw/system/presentations/1695/original/LPC%20-%20PerCpu%20Atomics.pdf
> 
> What is done differently here compared to percpu atomics: we track
> a single nesting counter per thread rather than many ranges of
> instruction pointer values. We deliver a signal to user-space and
> let the logic of restart be handled in user-space, thus moving
> the complexity out of the kernel. The nesting counter approach
> allows us to skip the complexity of interacting with signals that
> would be otherwise needed with the percpu atomics approach, which
> needs to know which instruction pointers are preempted, including
> when preemption occurs on a signal handler nested over an instruction
> pointer of interest.
> 
> Advantages of this approach over percpu atomics:
> - kernel code is relatively simple: complexity of restart sections
>   is in user-space,
> - easy to port to other architectures: just need to reserve a new
>   system call,
> - for threads which have registered a TLS structure, the fast-path
>   at preemption is only a nesting counter check, along with the
>   optional store of the current CPU number, rather than comparing
>   instruction pointer with possibly many registered ranges,
> 
> Caveats of this approach compared to the percpu atomics:
> - We need a signal number for this, so it cannot be done without
>   designing the application accordingly,
> - Handling restart in user-space is currently performed with page
>   protection, for which we install a SIGSEGV signal handler. Again,
>   this requires designing the application accordingly, especially
>   if the application installs its own segmentation fault handler,
> - It cannot be used for tracing of processes by injection of code
>   into their address space, due to interactions with application
>   signal handlers.
> 
> The user-space proof of concept code implementing the restart section
> can be found here: https://github.com/compudj/percpu-dev
> 
> Benchmarking sched_getcpu() vs tls cache approach. Getting the
> current CPU number:
> 
> - With Linux vdso:            12.7 ns
> - With TLS-cached cpu number:  0.3 ns
> 
> We will use the TLS-cached cpu number for the following
> benchmarks.
> 
> On an Intel(R) Xeon(R) CPU E5-2630 v3 @ 2.40GHz, comparison
> with a baseline running very few load/stores (no locking,
> no getcpu, assuming one thread per CPU with affinity),
> against locking scheme based on "lock; cmpxchg", "cmpxchg"
> (using restart signal), load-store (using restart signal).
> This is performed with 32 threads on a 16-core, hyperthread
> system:
> 
>                  ns/loop      overhead (ns)
> Baseline:          3.7           0.0
> lock; cmpxchg:    22.0          18.3
> cmpxchg:          11.1           7.4
> load-store:        9.4           5.7
> 
> Therefore, the load-store scheme has a speedup of 3.2x over the
> "lock; cmpxchg" scheme if both are using the tls-cache for the
> CPU number. If we use Linux sched_getcpu() for "lock; cmpxchg"
> we reach of speedup of 5.4x for load-store+tls-cache vs
> "lock; cmpxchg"+vdso-getcpu.
> 
> I'm sending this out to trigger discussion, and hopefully to see
> Paul and Andrew's patches being posted publicly at some point, so
> we can compare our approaches.

The idea seems sensible.  One quick comment: as with any new syscall,
please include a flags argument.

- Josh Triplett