Pardon the interruption: Loop preemption in Go 1.14

迭代点:

• 1.5

  • concurrent garbage collector

• 1.6

  • concurrent finalizer scans

  • dirty stack tracking

  • global write barriers

• 1.7

  • concurrent stack shrinking

• 1.8

  • hybrid write barrier

抢占:

• 1.0

  • voluntary preemption only

• 1.2

  • function call preemption

    • function entry -> grow stack/preempt -> Y -> call runtime

• 1.4

  • asynchronous preemption

    • +0.3% binary size

    • 0 performance overhead

    • ~25us typical preemption bound

一些技术尝试:

• preempt back-edges

  • adding instructions to loops isn’t viable

    • 每次loop判断preempt

      • 8% slowdown

• “fowward simulation”

• use signals, like an OS kernal

  • zero overhead, but doesn’t work with GC

    • 一次切换寄存器地址变了

• signals + gc stack maps at every instruction

  • +10% binary size

  • stack maps everywhere bloats binaries and doesn’t “keep it simple”

• signal-based preemption with conservative innermost frame scanning

  • 1.14

  • 实现

    • send a signal to the goroutine

      • 23 SIGURG

    • examine where it stopped/conservative scanning/precisely scanning

      • runtime.asyncPreempt

      • requires non-moving GC

        • https://en.wikipedia.org/wiki/Tracing_garbage_collection#Moving_vs._non-moving