Print scheduler details (ExecutionContext)

[ysbaddaden]

Prints runtime information about a program, either at regular interval, programmatically or on demand (through a signal), with or without the list of fibers (that can be huge). Stops the world to be able to print an accurate map of all execution contexts,

For example (idle program with a couple additional execution contexts):

sched.details time=571371878
ExecutionContext::SingleThreaded name=DEFAULT global_queue.size=0
  Scheduler name=DEFAULT thread=0x7f22d27a9740 local_queue.size=0 status=event-loop
ExecutionContext::SingleThreaded name=ST1 global_queue.size=0
  Scheduler name=ST1 thread=0x7f22cd750700 local_queue.size=0 status=parked
ExecutionContext::SingleThreaded name=ST2 global_queue.size=0
  Scheduler name=ST2 thread=0x7f22ccf4f700 local_queue.size=0 status=parked

Requirements:

• crystal with Thread.stop_world (merged for 1.14.0);
• execution_context (latest main)

Usage:

require "perf-tools"
require "execution_context"

# print on demand (UNIX only):
PerfTools::SchedulerTrace.on(Signal::USR1, details: true)

# print at regular intervals:
PerfTools::SchedulerTrace.every(5.seconds, details: false)

# print manually:
PerfTools::SchedulerTrace.print_runtime_status(details: true)

Specify details: true to also print the list of fibers in each execution context.

[ysbaddaden]

This line doesn't compile: we can't pass a variable (here details) to a C closure and we can't pass custom data to the sigaction handler either. We need a global variable (meh) or keep a Hash(Int32, Bool) as a class property (sounds better).

[straight-shoota]

Since we only need to cover two cases, maybe we could create two different procs?

if details 
  LibC::SigactionHandlerT.new do |_, _, _|
    print_runtime_status(true)
  end
else
  LibC::SigactionHandlerT.new do |_, _, _|
    print_runtime_status(false)
  end
end