Kind fetcher locks are not fully thread-safe

[mensfeld]

In case of extensive concurrent usage, the mutex handed over to two threads under same key may differ as illustrated below:

require 'concurrent'

10000.times do
  kind_fetcher_locks = Concurrent::Hash.new { |hash, key| hash[key] = Mutex.new }

  refs = Set.new
  100.times.map do |i|
    Thread.new { refs << kind_fetcher_locks[i % 50].object_id }
  end.each(&:join)

  raise "Not 50 but #{refs.count}" unless refs.size == 50
end

this can lead to really weird issues.

Works when fixed as above:

require 'concurrent'

10000.times do
  mutex = Mutex.new
  # kind_fetcher_locks = Concurrent::Hash.new { |hash, key| hash[key] = Mutex.new }

  kind_fetcher_locks = Concurrent::Hash.new do |hash, key|
    mutex.synchronize do
      break hash[key] if hash.key?(key)
      hash[key] = Mutex.new
    end
  end

  refs = Set.new
  100.times.map do |i|
    Thread.new { refs << kind_fetcher_locks[i % 50].object_id }
  end.each(&:join)

  raise "Not 50 but #{refs.count}" unless refs.size == 50
end

ref ruby-concurrency/concurrent-ruby#970

What are you trying to accomplish with this PR?
...

How is this accomplished?
...

What could go wrong?
...

[mensfeld]

I don't have an option to trigger a re-run of CI, signed the CLA

// o nice, picked up my comment