Fix retire/release race when resource close yields.

Co-authored-by: Claude <noreply@anthropic.com>

Author: wtn

lib/async/pool/controller.rb

@@ -12,6 +12,7 @@
 require "async/semaphore"
 
 require "thread"
+require "set"
 
 module Async
 	module Pool
@@ -50,6 +51,9 @@ def initialize(constructor, limit: nil, concurrency: (limit || 1), policy: nil,
 				# Used to signal when a resource has been released:
 				@mutex = Thread::Mutex.new
 				@condition = Thread::ConditionVariable.new
+				
+				# Resources currently being retired (close may yield):
+				@retiring = Set.new
 			end
 
 			# @attribute [Proc] The constructor used to create new resources.
@@ -234,9 +238,11 @@ def prune(retain = 0)
 			def retire(resource)
 				Console.debug(self){"Retire #{resource}"}
 
-				@resources.delete(resource)
+				return false unless @resources.delete(resource)
 
+				@retiring.add(resource)
 				resource.close
+				@retiring.delete(resource)
 
 				@mutex.synchronize{@condition.broadcast}
 
@@ -286,6 +292,9 @@ def reuse(resource)
 
 				usage = @resources[resource]
 
+				# Already retired (e.g. connection reset during retire close):
+				return false if usage.nil? && @retiring.include?(resource)
+				
 				if usage.nil? || usage.zero?
 					raise "Trying to reuse unacquired resource: #{resource}!"
 				end

test/async/pool/retire_race.rb

@@ -0,0 +1,78 @@
+# frozen_string_literal: true
+
+# Verifies that retire + release on the same resource does not raise
+# "Trying to reuse unacquired resource".
+#
+# In practice this happens with async-http when an HTTP/2 connection is reset:
+# the background reader retires the connection, but retire's resource.close
+# yields (e.g. sending GOAWAY), allowing another fiber to call release while
+# the resource is deleted from @resources but still reports reusable? = true.
+
+require "async/pool/controller"
+require "async/pool/resource"
+require "sus/fixtures/async/reactor_context"
+
+# A resource whose close yields (simulating async IO like sending GOAWAY),
+# but whose reusable? check is synchronous (no yield).
+class SlowCloseResource < Async::Pool::Resource
+	def close
+		Async::Task.current.yield
+		super
+	end
+end
+
+describe Async::Pool::Controller do
+	include Sus::Fixtures::Async::ReactorContext
+
+	with "retire/release race on slow-close resource" do
+		let(:pool) {subject.new(SlowCloseResource)}
+
+		it "gracefully handles release after retire begins closing" do
+			resource = pool.acquire
+
+			# Start retire in a child task. It runs synchronously up to the
+			# yield inside SlowCloseResource#close, then pauses. At that point
+			# @resources[resource] has been deleted but resource.close hasn't
+			# finished, so reusable? still returns true.
+			retire_task = Async do
+				pool.retire(resource)
+			end
+
+			# No yield needed — the child already ran to its yield point.
+			# Verify the race window exists:
+			expect(resource).to be(:reusable?)
+			expect(pool.resources).not.to be(:key?, resource)
+
+			# The client's error handler now tries to release the same resource.
+			# This should not raise — retire already claimed ownership.
+			pool.release(resource)
+
+			retire_task.wait
+		end
+
+		it "gracefully handles multiplexed release after retire begins closing" do
+			constructor = lambda { SlowCloseResource.new(128) }
+			pool = subject.new(constructor)
+
+			# Two streams on the same HTTP/2 connection:
+			r1 = pool.acquire
+			r2 = pool.acquire
+			expect(r1).to be_equal(r2)
+
+			# Background reader retires (yields during close):
+			retire_task = Async do
+				pool.retire(r1)
+			end
+
+			# The race window is open: resource deleted from pool but not
+			# yet closed. First stream's error handler hits the race:
+			expect(r1).to be(:reusable?)
+			expect(pool.resources).not.to be(:key?, r1)
+
+			# Should not raise:
+			pool.release(r1)
+
+			retire_task.wait
+		end
+	end
+end