Merge pull request #8632 from apolcyn/fix_ruby_deadlock

remove wait queue from ruby thread pool to avoid possibility of deadlock

src/ruby/lib/grpc/generic/rpc_server.rb

@@ -48,6 +48,10 @@ module GRPC
       @stop_cond = ConditionVariable.new
       @workers = []
       @keep_alive = keep_alive
+
+      # Each worker thread has its own queue to push and pull jobs
+      # these queues are put into @ready_queues when that worker is idle
+      @ready_workers = Queue.new
     end
 
     # Returns the number of jobs waiting
@@ -55,6 +59,13 @@ module GRPC
       @jobs.size
     end
 
+    def ready_for_work?
+      # Busy worker threads are either doing work, or have a single job
+      # waiting on them. Workers that are idle with no jobs waiting
+      # have their "queues" in @ready_workers
+      !@ready_workers.empty?
+    end
+
     # Runs the given block on the queue with the provided args.
     #
     # @param args the args passed blk when it is called
@@ -67,7 +78,11 @@ module GRPC
           return
         end
         GRPC.logger.info('schedule another job')
-        @jobs << [blk, args]
+        fail 'No worker threads available' if @ready_workers.empty?
+        worker_queue = @ready_workers.pop
+
+        fail 'worker already has a task waiting' unless worker_queue.empty?
+        worker_queue << [blk, args]
       end
     end
 
@@ -77,9 +92,11 @@ module GRPC
         fail 'already stopped' if @stopped
       end
       until @workers.size == @size.to_i
-        next_thread = Thread.new do
+        new_worker_queue = Queue.new
+        @ready_workers << new_worker_queue
+        next_thread = Thread.new(new_worker_queue) do |jobs|
           catch(:exit) do  # allows { throw :exit } to kill a thread
-            loop_execute_jobs
+            loop_execute_jobs(jobs)
           end
           remove_current_thread
         end
@@ -90,7 +107,7 @@ module GRPC
     # Stops the jobs in the pool
     def stop
       GRPC.logger.info('stopping, will wait for all the workers to exit')
-      @workers.size.times { schedule { throw :exit } }
+      schedule { throw :exit } while ready_for_work?
       @stop_mutex.synchronize do  # wait @keep_alive for works to stop
         @stopped = true
         @stop_cond.wait(@stop_mutex, @keep_alive) if @workers.size > 0
@@ -125,15 +142,18 @@ module GRPC
       end
     end
 
-    def loop_execute_jobs
+    def loop_execute_jobs(worker_queue)
       loop do
         begin
-          blk, args = @jobs.pop
+          blk, args = worker_queue.pop
           blk.call(*args)
         rescue StandardError => e
           GRPC.logger.warn('Error in worker thread')
           GRPC.logger.warn(e)
         end
+        # there shouldn't be any work given to this thread while its busy
+        fail('received a task while busy') unless worker_queue.empty?
+        @ready_workers << worker_queue
       end
     end
   end
@@ -147,10 +167,10 @@ module GRPC
 
     def_delegators :@server, :add_http2_port
 
-    # Default thread pool size is 3
-    DEFAULT_POOL_SIZE = 3
+    # Default thread pool size is 30
+    DEFAULT_POOL_SIZE = 30
 
-    # Default max_waiting_requests size is 20
+    # Deprecated due to internal changes to the thread pool
     DEFAULT_MAX_WAITING_REQUESTS = 20
 
     # Default poll period is 1s
@@ -175,11 +195,11 @@ module GRPC
     # instance.
     #
     # * pool_size: the size of the thread pool the server uses to run its
-    # threads
+    # threads. No more concurrent requests can be made than the size
+    # of the thread pool
     #
-    # * max_waiting_requests: the maximum number of requests that are not
-    # being handled to allow. When this limit is exceeded, the server responds
-    # with not available to new requests
+    # * max_waiting_requests: Deprecated due to internal changes to the thread
+    # pool. This is still an argument for compatibility but is ignored.
     #
     # * poll_period: when present, the server polls for new events with this
     # period
@@ -330,10 +350,8 @@ module GRPC
 
     # Sends RESOURCE_EXHAUSTED if there are too many unprocessed jobs
     def available?(an_rpc)
-      jobs_count, max = @pool.jobs_waiting, @max_waiting_requests
-      GRPC.logger.info("waiting: #{jobs_count}, max: #{max}")
-      return an_rpc if @pool.jobs_waiting <= @max_waiting_requests
-      GRPC.logger.warn("NOT AVAILABLE: too many jobs_waiting: #{an_rpc}")
+      return an_rpc if @pool.ready_for_work?
+      GRPC.logger.warn('no free worker threads currently')
       noop = proc { |x| x }
       c = ActiveCall.new(an_rpc.call, noop, noop, an_rpc.deadline,
                          metadata_received: true)

src/ruby/spec/generic/rpc_server_pool_spec.rb

@@ -29,6 +29,8 @@
 
 require 'grpc'
 
+Thread.abort_on_exception = true
+
 describe GRPC::Pool do
   Pool = GRPC::Pool
 
@@ -44,32 +46,34 @@ describe GRPC::Pool do
     end
   end
 
-  describe '#jobs_waiting' do
-    it 'at start, it is zero' do
+  describe '#ready_for_work?' do
+    it 'before start it is not ready' do
       p = Pool.new(1)
-      expect(p.jobs_waiting).to be(0)
+      expect(p.ready_for_work?).to be(false)
     end
 
-    it 'it increases, with each scheduled job if the pool is not running' do
-      p = Pool.new(1)
-      job = proc {}
-      expect(p.jobs_waiting).to be(0)
-      5.times do |i|
+    it 'it stops being ready after all workers jobs waiting or running' do
+      p = Pool.new(5)
+      p.start
+      job = proc { sleep(3) } # sleep so workers busy when done scheduling
+      5.times do
+        expect(p.ready_for_work?).to be(true)
         p.schedule(&job)
-        expect(p.jobs_waiting).to be(i + 1)
       end
+      expect(p.ready_for_work?).to be(false)
     end
 
-    it 'it decreases as jobs are run' do
-      p = Pool.new(1)
+    it 'it becomes ready again after jobs complete' do
+      p = Pool.new(5)
+      p.start
       job = proc {}
-      expect(p.jobs_waiting).to be(0)
-      3.times do
+      5.times do
+        expect(p.ready_for_work?).to be(true)
         p.schedule(&job)
       end
-      p.start
-      sleep 2
-      expect(p.jobs_waiting).to be(0)
+      expect(p.ready_for_work?).to be(false)
+      sleep 5 # give the pool time do get at least one task done
+      expect(p.ready_for_work?).to be(true)
     end
   end
 
@@ -90,6 +94,18 @@ describe GRPC::Pool do
       expect(q.pop).to be(o)
       p.stop
     end
+
+    it 'it throws an error if all of the workers have tasks to do' do
+      p = Pool.new(5)
+      p.start
+      job = proc {}
+      5.times do
+        expect(p.ready_for_work?).to be(true)
+        p.schedule(&job)
+      end
+      expect { p.schedule(&job) }.to raise_error
+      expect { p.schedule(&job) }.to raise_error
+    end
   end
 
   describe '#stop' do
@@ -113,18 +129,8 @@ describe GRPC::Pool do
   end
 
   describe '#start' do
-    it 'runs pre-scheduled jobs' do
-      p = Pool.new(2)
-      o, q = Object.new, Queue.new
-      n = 5  # arbitrary
-      n.times { p.schedule(o, &q.method(:push)) }
-      p.start
-      n.times { expect(q.pop).to be(o) }
-      p.stop
-    end
-
     it 'runs jobs as they are scheduled' do
-      p = Pool.new(2)
+      p = Pool.new(5)
       o, q = Object.new, Queue.new
       p.start
       n = 5  # arbitrary