boost.asio的又一個bug.

最近在設計一個多線程分塊支持續傳的http的異步客戶端時, 測試部門常常發現http下載模
塊退出時偶爾會卡住, 在win7系統上由爲明顯. 反覆檢查代碼, 並未明顯問題, 因而專門寫
了一個反覆退出的單元測試, 當即發現問題, 並定位在io_service的析構函數中, 奇怪的是,
個人投遞io的全部socket都早已經關閉, run線程也已經退出, 按理說, 這時的io_service的
outstanding_work_應該爲0纔是, 可我一看它倒是1, 因而始終在win_iocp_io_service.hpp的
shutdown_service裏一直循環調用GetQueuedCompletionStatus, 從而致使沒法正常退出...
很明顯, 這是asio對outstanding_work_計數維護的有問題, 爲了解決
問題, 因而我很快想到不使用iocp, 添加宏BOOST_ASIO_DISABLE_IOCP一切就正常了...
因爲本身使用的是boost.1.45版本, 因而換了個boost.1.46.1再試試, 結果同樣. 難道這麼嚴
重的bug跨在了這兩個很是重要的發行版本上而沒人知道?
在官方的郵件列表中細節檢查, 終於看到了某人的bug報告和我描述的狀況差很少, 並且
在那我的報告了bug的次日, asio做者就發佈了補丁, 但這個補丁並未更新到boost.1.45
和boost.1.46中, 唉, 這兩個版本但是大版本啊, 估計受害人很多...
不過幸運的是, 我在boost的主分枝中看到了修正的代碼.
下面是這個補丁內容:

From 81a6a51c0cb66de6bc77e1fa5dcd46b2794995e4 Mon Sep 17 00:00:00 2001
From: Christopher Kohlhoff <chris@kohlhoff.com>
Date: Wed, 23 Mar 2011 15:03:56 +1100
Subject: [PATCH] On Windows, ensure the count of outstanding work is decremented for
 abandoned operations (i.e. operations that are being cleaned up within
 the io_service destructor).

---
 asio/include/asio/detail/impl/dev_poll_reactor.ipp |    2 ++
 asio/include/asio/detail/impl/epoll_reactor.ipp    |    2 ++
 asio/include/asio/detail/impl/kqueue_reactor.ipp   |    2 ++
 asio/include/asio/detail/impl/select_reactor.ipp   |    2 ++
 .../asio/detail/impl/signal_set_service.ipp        |    2 ++
 asio/include/asio/detail/impl/task_io_service.ipp  |    7 +++++++
 .../asio/detail/impl/win_iocp_io_service.ipp       |   11 +++++++++++
 asio/include/asio/detail/task_io_service.hpp       |    4 ++++
 asio/include/asio/detail/win_iocp_io_service.hpp   |    4 ++++
 9 files changed, 36 insertions(+), 0 deletions(-)

diff --git a/asio/include/asio/detail/impl/dev_poll_reactor.ipp b/asio/include/asio/detail/impl/dev_poll_reactor.ipp
index 06d89ea..2a01993 100644
--- a/asio/include/asio/detail/impl/dev_poll_reactor.ipp
+++ b/asio/include/asio/detail/impl/dev_poll_reactor.ipp
@@ -63,6 +63,8 @@ void dev_poll_reactor::shutdown_service()
     op_queue_[i].get_all_operations(ops);
 
   timer_queues_.get_all_timers(ops);
+
+  io_service_.abandon_operations(ops);
 } 
 
 // Helper class to re-register all descriptors with /dev/poll.
diff --git a/asio/include/asio/detail/impl/epoll_reactor.ipp b/asio/include/asio/detail/impl/epoll_reactor.ipp
index 22f567a..d08dedb 100644
--- a/asio/include/asio/detail/impl/epoll_reactor.ipp
+++ b/asio/include/asio/detail/impl/epoll_reactor.ipp
@@ -84,6 +84,8 @@ void epoll_reactor::shutdown_service()
   }
 
   timer_queues_.get_all_timers(ops);
+
+  io_service_.abandon_operations(ops);
 }
 
 void epoll_reactor::fork_service(asio::io_service::fork_event fork_ev)
diff --git a/asio/include/asio/detail/impl/kqueue_reactor.ipp b/asio/include/asio/detail/impl/kqueue_reactor.ipp
index f0cdf73..45aff60 100644
--- a/asio/include/asio/detail/impl/kqueue_reactor.ipp
+++ b/asio/include/asio/detail/impl/kqueue_reactor.ipp
@@ -74,6 +74,8 @@ void kqueue_reactor::shutdown_service()
   }
 
   timer_queues_.get_all_timers(ops);
+
+  io_service_.abandon_operations(ops);
 }
 
 void kqueue_reactor::fork_service(asio::io_service::fork_event fork_ev)
diff --git a/asio/include/asio/detail/impl/select_reactor.ipp b/asio/include/asio/detail/impl/select_reactor.ipp
index f4e0314..00fd9fc 100644
--- a/asio/include/asio/detail/impl/select_reactor.ipp
+++ b/asio/include/asio/detail/impl/select_reactor.ipp
@@ -81,6 +81,8 @@ void select_reactor::shutdown_service()
     op_queue_[i].get_all_operations(ops);
 
   timer_queues_.get_all_timers(ops);
+
+  io_service_.abandon_operations(ops);
 }
 
 void select_reactor::fork_service(asio::io_service::fork_event fork_ev)
diff --git a/asio/include/asio/detail/impl/signal_set_service.ipp b/asio/include/asio/detail/impl/signal_set_service.ipp
index f0f0e78..4cde184 100644
--- a/asio/include/asio/detail/impl/signal_set_service.ipp
+++ b/asio/include/asio/detail/impl/signal_set_service.ipp
@@ -145,6 +145,8 @@ void signal_set_service::shutdown_service()
       reg = reg->next_in_table_;
     }
   }
+
+  io_service_.abandon_operations(ops);
 }
 
 void signal_set_service::fork_service(
diff --git a/asio/include/asio/detail/impl/task_io_service.ipp b/asio/include/asio/detail/impl/task_io_service.ipp
index cb585d5..0a2c6fa 100644
--- a/asio/include/asio/detail/impl/task_io_service.ipp
+++ b/asio/include/asio/detail/impl/task_io_service.ipp
@@ -230,6 +230,13 @@ void task_io_service::post_deferred_completions(
   }
 }
 
+void task_io_service::abandon_operations(
+    op_queue<task_io_service::operation>& ops)
+{
+  op_queue<task_io_service::operation> ops2;
+  ops2.push(ops);
+}
+
 std::size_t task_io_service::do_one(mutex::scoped_lock& lock,
     task_io_service::idle_thread_info* this_idle_thread)
 {
diff --git a/asio/include/asio/detail/impl/win_iocp_io_service.ipp b/asio/include/asio/detail/impl/win_iocp_io_service.ipp
index ca3125e..7aaa6b8 100644
--- a/asio/include/asio/detail/impl/win_iocp_io_service.ipp
+++ b/asio/include/asio/detail/impl/win_iocp_io_service.ipp
@@ -262,6 +262,17 @@ void win_iocp_io_service::post_deferred_completions(
   }
 }
 
+void win_iocp_io_service::abandon_operations(
+    op_queue<win_iocp_operation>& ops)
+{
+  while (win_iocp_operation* op = ops.front())
+  {
+    ops.pop();
+    ::InterlockedDecrement(&outstanding_work_);
+    op->destroy();
+  }
+}
+
 void win_iocp_io_service::on_pending(win_iocp_operation* op)
 {
   if (::InterlockedCompareExchange(&op->ready_, 1, 0) == 1)
diff --git a/asio/include/asio/detail/task_io_service.hpp b/asio/include/asio/detail/task_io_service.hpp
index 285d83e..654f83c 100644
--- a/asio/include/asio/detail/task_io_service.hpp
+++ b/asio/include/asio/detail/task_io_service.hpp
@@ -105,6 +105,10 @@ public:
   // that work_started() was previously called for each operation.
   ASIO_DECL void post_deferred_completions(op_queue<operation>& ops);
 
+  // Process unfinished operations as part of a shutdown_service operation.
+  // Assumes that work_started() was previously called for the operations.
+  ASIO_DECL void abandon_operations(op_queue<operation>& ops);
+
 private:
   // Structure containing information about an idle thread.
   struct idle_thread_info;
diff --git a/asio/include/asio/detail/win_iocp_io_service.hpp b/asio/include/asio/detail/win_iocp_io_service.hpp
index a562834..b5d7f0b 100644
--- a/asio/include/asio/detail/win_iocp_io_service.hpp
+++ b/asio/include/asio/detail/win_iocp_io_service.hpp
@@ -126,6 +126,10 @@ public:
   ASIO_DECL void post_deferred_completions(
       op_queue<win_iocp_operation>& ops);
 
+  // Enqueue unfinished operation as part of a shutdown_service operation.
+  // Assumes that work_started() was previously called for the operations.
+  ASIO_DECL void abandon_operations(op_queue<operation>& ops);
+
   // Called after starting an overlapped I/O operation that did not complete
   // immediately. The caller must have already called work_started() prior to
   // starting the operation.
-- 
1.7.0.1

注: 在boost.asio中, 使用這個補丁時, 須要將ASIO_DECL 改爲 BOOST_ ASIO_DECL
注: 最新發布的boost1.47版本已經修復了這個bug, 建議儘可能採用新版本的boost.

這是我第二次在使用asio的過程當中, 發現的比較嚴重的bug了, 不過幸運的是, 每一次都能在官方的論壇 
或郵件列表中找到解決方案. 

結論, 再牛的人寫的代碼也會有bug, 我的很是崇拜asio的做者.