Fix task release in session disconnect

queue/abstract.lua

@@ -151,18 +151,22 @@ function tube.ack(self, id)
     return result
 end
 
-function tube.release(self, id, opts)
+local function tube_release_internal(self, id, opts, session_id)
     opts = opts or {}
-    local _taken = box.space._queue_taken:get{session.id(), self.tube_id, id}
+    local _taken = box.space._queue_taken:get{session_id, self.tube_id, id}
     if _taken == nil then
         error("Task was not taken in the session")
     end
 
-    box.space._queue_taken:delete{session.id(), self.tube_id, id}
+    box.space._queue_taken:delete{session_id, self.tube_id, id}
     self:peek(id)
     return self.raw:normalize_task(self.raw:release(id, opts))
 end
 
+function tube.release(self, id, opts)
+    return tube_release_internal(self, id, opts, session.id())
+end
+
 function tube.peek(self, id)
     local task = self.raw:peek(id)
     if task == nil then
@@ -371,7 +375,7 @@ function method._on_consumer_disconnect()
             log.warn("Consumer %s disconnected, release task %s(%s)",
                 id, task[3], tube[1])
 
-            queue.tube[tube[1]]:release(task[3])
+            tube_release_internal(queue.tube[tube[1]], task[3], nil, id)
         end
     end
 end

t/110-disconnect-trigger-check.t

@@ -0,0 +1,98 @@
+#!/usr/bin/env tarantool
+
+local fiber = require('fiber')
+local netbox = require('net.box')
+local os = require('os')
+local queue = require('queue')
+local tap = require('tap')
+local tnt = require('t.tnt')
+
+local tube
+local test = tap.test('lost a session id after yield')
+
+-- The test cases are in check_result().
+test:plan(2)
+
+-- Verify that _queue_taken space is empty.
+local function check_result()
+    if tube == nil then
+        os.exit(1)
+    end
+
+    -- tube:drop() is most simple way to check that _queue_taken
+    -- is empty. It give an error if it is not so.
+    local ok, res = pcall(tube.drop, tube)
+    test:is(ok, true, 'drop empty queue')
+    test:is(res, true, 'tube:drop() result is true')
+
+    tnt.finish()
+    os.exit(test:check() and 0 or 1)
+end
+
+-- Yield in queue's on_disconnect trigger (which handles a client
+-- disconnection) may lead to a situation when _queue_taken
+-- temporary space is not cleaned and becomes inconsistent with
+-- 'status' field in <tube_name> space. This appears only on
+-- tarantool versions affected by gh-4627.
+--
+-- See https://github.com/tarantool/queue/issues/103
+-- See https://github.com/tarantool/tarantool/issues/4627
+local function test_lost_session_id_after_yield()
+    -- We must check the results of a test after
+    -- the queue._on_consumer_disconnect trigger
+    -- has been done.
+    --
+    -- Triggers are run in LIFO order.
+    box.session.on_disconnect(check_result)
+
+    local listen = 'localhost:1918'
+    tnt.cfg{listen = listen}
+
+    local driver = 'fifottl'
+    tube = queue.create_tube('test_tube', driver, {if_not_exists = true})
+
+    rawset(_G, 'queue', require('queue'))
+    tube:grant('guest', {call = true})
+
+    -- We need at least two tasks to trigger box.session.id()
+    -- call after a yield in the queue._on_consumer_disconnect
+    -- trigger (in the version of queue before the fix). See
+    -- more below.
+    queue.tube.test_tube:put('1')
+    queue.tube.test_tube:put('2')
+    local connection = netbox.connect(listen)
+    connection:call('queue.tube.test_tube:take')
+    connection:call('queue.tube.test_tube:take')
+
+    -- After disconnection of a client the _on_consumer_disconnect
+    -- trigger is run. It changes 'status' field for tuples in
+    -- <tube_name> space in a loop and removes the corresponding
+    -- tuples from _queue_taken space. The version before the fix
+    -- operates in this way:
+    --
+    --  | <_on_consumer_disconnect>
+    --  | for task in tasks of the client:
+    --  |     call <task:release>
+    --  |
+    --  | <task:release>
+    --  |     delete _queue_taken tuple using box.session.id()
+    --  |     update <tube_name> space using task_id -- !! yield
+    --
+    -- So the deletion from _queue_taken may be unable to delete
+    -- right tuples for second and following tasks, because
+    -- box.session.id() may give a garbage.
+    connection:close()
+
+    -- Wait for check_result() trigger, which will ensure that
+    -- _queue_taken space is cleaned and will exit successfully
+    -- in the case (or exit abnormally otherwise).
+    fiber.sleep(5)
+
+    -- Wrong session id may lead to 'Task was not taken in the
+    -- session' error in the _on_consumer_disconnect and so the
+    -- second on_disconnect trigger (check_result) will not be
+    -- fired.
+    os.exit(1)
+end
+
+test_lost_session_id_after_yield()