https://groups.google.com/forum/#!topic/seastar-dev/RuK-OajeqHk

https://www.google.com/search?ei=gTH-Wtr5O4WQsAW-io7wCQ&q=dl_iterate_phdr+gcc+exception&oq=dl_iterate_phdr+gcc+exception&gs_l=psy-ab.3...5055.7797.0.8041.10.8.0.0.0.0.382.997.3-3.3.0....0...1.1.64.psy-ab..7.1.306...33i160k1.0.dDZSKQ-pAA4

g++/glibc combination provides a c++ runtime with non scaleable 
c++ exception. This is dues to global locks in stack unwinding code. 
    
To summarize all the locks we have now and their purpose: 
 1. There is a lock in _Unwind_Find_FDE (libgcc) that protects 
    list of FDEs registered with __register_frame* functions. 
    The catch is that dynamically linked binary do not do that, 
    so all that it protects is checking that a certain list is empty. 
    This lock no longer relevant in gcc7 since there is a path there 
    checks that list is empty outside of the lock. 
 2. The lock in dl_iterate_phdr (glibc) that protects loaded object 
    list against runtime object loading/unloading.

To get rid of the first lock one has to use gcc7.

To get rid of the second one we can use the fact that we do not 
load/unload objects dynamically (at least for now). To do that we 
can mirror all elf header information in seastar and provide our 
own dl_iterate_phdr symbol which uses this mirror without locking. 
    
Unfortunately there is another gotcha in this approach: dl_iterate_phdr 
supplied by glibc never calls more then one callback simultaneously as an 
unintended consequences of the lock there, but unfortunately libgcc relies 
on that to maintain small cache of translations. The access to the cache is 
not protected by any lock since up until now only one callback could have 
run at a time. But luckily libgcc cannot use the cache if older version 
of dl_phdr_info is provided to the callback because the older version 
did not have an indication that loaded object list may have changed, 
so libgcc does not know when cache should be invalidated and disables it 
entirely. By calling the callback with old version of dl_phdr_info from 
our dl_iterate_phdr we can effectively make libgcc callback thread safe.

diff --git a/configure.py b/configure.py 
index facdd8f..33b310b 100755 
--- a/configure.py 
+++ b/configure.py 
@@ -297,6 +297,8 @@ arg_parser.add_argument('--static-boost', dest = 'staticboost', action = 'store_ 
 add_tristate(arg_parser, name = 'hwloc', dest = 'hwloc', help = 'hwloc support') 
 arg_parser.add_argument('--enable-gcc6-concepts', dest='gcc6_concepts', action='store_true', default=False, 
                         help='enable experimental support for C++ Concepts as implemented in GCC 6') 
+arg_parser.add_argument('--disable-exception-scalability-workaround', dest='exception_workaround', action='store_true', default=False, 
+        help='disabling override of dl_iterate_phdr symbol to workaround C++ exception scalability issues') 
 args = arg_parser.parse_args() 
  
 libnet = [ 
@@ -337,6 +339,7 @@ core = [ 
     'net/inet_address.cc', 
     'rpc/rpc.cc', 
     'rpc/lz4_compressor.cc', 
+    'core/exception_hacks.cc', 
     ] 
  
 protobuf = [ 
@@ -392,6 +395,9 @@ if args.gcc6_concepts: 
     defines.append('HAVE_GCC6_CONCEPTS') 
     args.user_cflags += ' -fconcepts' 
  
+if args.exception_workaround: 
+    defines.append('NO_EXCEPTION_HACK') 

 if args.staticcxx: 
     libs = libs.replace('-lstdc++', '') 
     libs += ' -static-libgcc -static-libstdc++' 
diff --git a/core/exception_hacks.cc b/core/exception_hacks.cc 
new file mode 100644 
index 0000000..6f04630 
--- /dev/null 
+++ b/core/exception_hacks.cc 
@@ -0,0 +1,108 @@ 
+/* 
+ * This file is open source software, licensed to you under the terms 
+ * of the Apache License, Version 2.0 (the "License").  See the NOTICE file 
+ * distributed with this work for additional information regarding copyright 
+ * ownership.  You may not use this file except in compliance with the License. 
+ * 
+ * You may obtain a copy of the License at 
+ * 
+ *   http://www.apache.org/licenses/LICENSE-2.0 
+ * 
+ * Unless required by applicable law or agreed to in writing, 
+ * software distributed under the License is distributed on an 
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY 
+ * KIND, either express or implied.  See the License for the 
+ * specific language governing permissions and limitations 
+ * under the License. 
+ */ 
+/* 
+ * Copyright (C) 2017 ScyllaDB 
+ */ 

+// The purpose of the hacks here is to workaround C++ exception scalability problem 
+// with gcc and glibc. For the best result gcc-7 is required. 
+// 
+// To summarize all the locks we have now and their purpose: 
+// 1. There is a lock in _Unwind_Find_FDE (libgcc) that protects 
+//    list of FDEs registered with __register_frame* functions. 
+//    The catch is that dynamically linked binary do not do that, 
+//    so all that it protects is checking that a certain list is empty. 
+//    This lock no longer relevant in gcc-7 since there is a path there 
+//    that checks that list is empty outside of the lock and it will be 
+//    always true for us. 
+// 2. The lock in dl_iterate_phdr (glibc) that protects loaded object 
+//    list against runtime object loading/unloading. 
+// 
+// To get rid of the first lock using gcc-7 is required. 
+// 
+// To get rid of the second one we can use the fact that we do not 
+// load/unload objects dynamically (at least for now). To do that we 
+// can mirror all elf header information in seastar and provide our 
+// own dl_iterate_phdr symbol which uses this mirror without locking. 
+// 
+// Unfortunately there is another gotcha in this approach: dl_iterate_phdr 
+// supplied by glibc never calls more then one callback simultaneously as an 
+// unintended consequences of the lock there, but unfortunately libgcc relies 
+// on that to maintain small cache of translations. The access to the cache is 
+// not protected by any lock since up until now only one callback could have 
+// run at a time. But luckily libgcc cannot use the cache if older version 
+// of dl_phdr_info is provided to the callback because the older version 
+// did not have an indication that loaded object list may have changed, 
+// so libgcc does not know when cache should be invalidated and disables it 
+// entirely. By calling the callback with old version of dl_phdr_info from 
+// our dl_iterate_phdr we can effectively make libgcc callback thread safe. 

+#ifndef NO_EXCEPTION_HACK 
+#include <link.h> 
+#include <dlfcn.h> 
+#include <assert.h> 
+#include <vector> 
+#include <cstddef> 
+#include "exception_hacks.hh" 

+namespace seastar { 
+using dl_iterate_fn = int (*) (int (*callback) (struct dl_phdr_info *info, size_t size, void *data), void *data); 

+static dl_iterate_fn dl_iterate_phdr_org() { 
+    static dl_iterate_fn org = [] { 
+        auto org = (dl_iterate_fn)dlsym (RTLD_NEXT, "dl_iterate_phdr"); 
+        assert(org); 
+        return org; 
+    }(); 
+    return org; 
+} 

+static std::vector<dl_phdr_info> phdrs_cache; 

+void init_phdr_cache() { 
+    // Fill out elf header cache for access without locking. 
+    // This assumes no dynamic object loading/unloading after this point 
+    dl_iterate_phdr_org()([] (struct dl_phdr_info *info, size_t size, void *data) { 
+        phdrs_cache.push_back(*info); 
+        return 0; 
+    }, nullptr); 
+} 
+} 

+extern "C" 
+[[gnu::visibility("default")]] 
+[[gnu::externally_visible]] 
+int dl_iterate_phdr(int (*callback) (struct dl_phdr_info *info, size_t size, void *data), void *data) { 
+    if (!seastar::phdrs_cache.size()) { 
+        // Cache is not yet populated, pass through to original function 
+        return seastar::dl_iterate_phdr_org()(callback, data); 
+    } 
+    int r = 0; 
+    for (auto h : seastar::phdrs_cache) { 
+        // Pass dl_phdr_info size that does not include dlpi_adds and dlpi_subs. 
+        // This forces libgcc to disable caching which is not thread safe and 
+        // requires dl_iterate_phdr to serialize calls to callback. Since we do 
+        // not serialize here we have to disable caching. 
+        r = callback(&h, offsetof(dl_phdr_info, dlpi_adds), data); 
+        if (r) { 
+            break; 
+        } 
+    } 
+    return r; 
+} 
+#endif 
diff --git a/core/exception_hacks.hh b/core/exception_hacks.hh 
new file mode 100644 
index 0000000..f5ff51f 
--- /dev/null 
+++ b/core/exception_hacks.hh 
@@ -0,0 +1,24 @@ 
+/* 
+ * This file is open source software, licensed to you under the terms 
+ * of the Apache License, Version 2.0 (the "License").  See the NOTICE file 
+ * distributed with this work for additional information regarding copyright 
+ * ownership.  You may not use this file except in compliance with the License. 
+ * 
+ * You may obtain a copy of the License at 
+ * 
+ *   http://www.apache.org/licenses/LICENSE-2.0 
+ * 
+ * Unless required by applicable law or agreed to in writing, 
+ * software distributed under the License is distributed on an 
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY 
+ * KIND, either express or implied.  See the License for the 
+ * specific language governing permissions and limitations 
+ * under the License. 
+ */ 
+/* 
+ * Copyright (C) 2017 ScyllaDB 
+ */ 

+namespace seastar { 
+void init_phdr_cache(); 
+} 
diff --git a/core/reactor.cc b/core/reactor.cc 
index 2beef59..b4b4e8d 100644 
--- a/core/reactor.cc 
+++ b/core/reactor.cc 
@@ -93,6 +93,7 @@ 
 #include "util/defer.hh" 
 #include "core/metrics.hh" 
 #include "execution_stage.hh" 
+#include "exception_hacks.hh" 
  
 namespace seastar { 
  
@@ -3369,6 +3370,9 @@ smp::get_options_description() 
         ("max-io-requests", bpo::value<unsigned>(), "Maximum amount of concurrent requests to be sent to the disk. Defaults to 128 times the number of processors") 
 #endif 
         ("mbind", bpo::value<bool>()->default_value(true), "enable mbind") 
+#ifndef NO_EXCEPTION_HACK 
+        ("enable-glibc-exception-scaling-workaround", bpo::value<bool>()->default_value(true), "enable workaround for glibc/gcc c++ exception scalablity problem") 
+#endif 
         ; 
     return opts; 
 } 
@@ -3514,6 +3518,12 @@ static void sigabrt_action() noexcept { 
  
 void smp::configure(boost::program_options::variables_map configuration) 
 { 
+#ifndef NO_EXCEPTION_HACK 
+    if (configuration["enable-glibc-exception-scaling-workaround"].as<bool>()) { 
+        init_phdr_cache(); 
+    } 
+#endif 

     // Mask most, to prevent threads (esp. dpdk helper threads) 
     // from servicing a signal.  Individual reactors will unmask signals 
     // as they become prepared to handle them. 
-- 
                        Gleb.

C++ 的 runtime exception是没有扩展性的的更多相关文章

  1. 优秀开源项目之三:高性能、高并发、高扩展性和可读性的网络服务器架构State Threads

    译文在后面. State Threads for Internet Applications Introduction State Threads is an application library ...

  2. Atitit.兼具兼容性和扩展性的配置方案attilax总结

    Atitit.兼具兼容性和扩展性的配置方案attilax总结 文件配置法1 Jdbc多数据源文件配置发1 Bat文件配置法1 改进的文件配置法(采用类似i18n技术) 推荐1 使用自动化pc_id的方 ...

  3. Atitit.软件架构高扩展性and兼容性原理与概论实践attilax总结

    Atitit.软件架构高扩展性and兼容性原理与概论实践attilax总结 1. 什么是可扩展的应用程序?1 2. 松耦合(ioc)2 3. 接口的思考 2 4. 单一用途&模块化,小粒度化2 ...

  4. OpenStack 企业私有云的若干需求(6):大规模扩展性支持

    本系列会介绍OpenStack 企业私有云的几个需求: 自动扩展(Auto-scaling)支持 多租户和租户隔离 (multi-tenancy and tenancy isolation) 混合云( ...

  5. 使用Lua脚本语言开发出高扩展性的系统,AgileEAS.NET SOA中间件Lua脚本引擎介绍

    一.前言 AgileEAS.NET SOA 中间件平台是一款基于基于敏捷并行开发思想和Microsoft .Net构件(组件)开发技术而构建的一个快速开发应用平台.用于帮助中小型软件企业建立一条适合市 ...

  6. jetbrick,新一代 Java 模板引擎,具有高性能和高扩展性

    新一代 Java 模板引擎,具有高性能和高扩展性. <!-- Jetbrick Template Engineer --> <dependency> <groupId&g ...

  7. Android ImageCache图片缓存,使用简单,支持预取,支持多种缓存算法,支持不同网络类型,扩展性强

    本文主要介绍一个支持图片自动预取.支持多种缓存算法的图片缓存的使用及功能.图片较大需要SD卡保存情况推荐使用ImageSDCardCache. 与Android LruCache相比主要特性:(1). ...

  8. Checked Exception与Runtime Exception 的区别

    Java里有个很重要的特色是Exception ,也就是说允许程序产生例外状况.而在学Java 的时候,我们也只知道Exception 的写法,却未必真能了解不同种类的Exception 的区别. 首 ...

  9. 深入NGINX:我们如何设计它的性能和扩展性

    为了更好地理解设计,你需要了解NGINX是如何工作的.NGINX之所以能在性能上如此优越,是由于其背后的设计.许多web服务器和应用服务器使用简单的线程的(threaded).或基于流程的 (proc ...

随机推荐

  1. .NET零基础入门06:面向对象入门

    一:前言 在本系列课程的第一部分,我们说明为了要选择C#作为你成为程序员的第一门语言: • 首先,C#是一门非常优秀的面向对象编程的语言: 凡是对编码感兴趣的同学一定听说过"面向对象编程&q ...

  2. 部署maven的一些要点、遇到的问题和心得体会

    maven的部署.遇到的问题和心得体会 2013-10-24 | 阅:  转:  |  分享         部署maven的一些要点.遇到的问题和心得体会 (图片看不了,可以下载doc文件) 一.  ...

  3. FastJson、Jackson、Gson进行Java对象转换Json的细节处理

    前言 Java对象在转json的时候,如果对象里面有属性值为null的话,那么在json序列化的时候要不要序列出来呢?对比以下json转换方式 一.fastJson 1.fastJson在转换java ...

  4. 使用强大的 Mockito 测试框架来测试你的代码

    原文链接 : Unit tests with Mockito - Tutorial 译文出自 : 掘金翻译计划 译者 : edvardhua 校对者: hackerkevin, futureshine ...

  5. MECE分析法(Mutually Exclusive Collectively Exhaustive)

    什么是MECE分析法? MECE,是Mutually Exclusive Collectively Exhaustive,中文意思是“相互独立,完全穷尽”. 也就是对于一个重大的议题,能够做到不重叠. ...

  6. 不越狱安装破解软件,iResign重签名方法

    http://www.baidu.com/s?wd=iresign%E8%BD%AF%E4%BB%B6&rsv_spt=1&issp=1&rsv_bp=0&ie=utf ...

  7. [转]MySQL单列索引和组合索引的区别介绍

    FROM : http://database.ctocio.com.cn/353/11664853.shtml MySQL单列索引是我们使用MySQL数据库中经常会见到的,MySQL单列索引和组合索引 ...

  8. C++二维数组 取地址 复制给 二维指针

    本来应该是个简单的问题,但是我就不明白了,为什么会段错误了... #include<stdio.h> #define UINT32 unsigned int UINT32 NType1_X ...

  9. 第十三章 redis-cluster原理

    一.基本定义 虚拟槽slot分区算法,优点是扩容缩容简单:直接把slot及每个slot上的数据进行缩放即可 redis定义了0-16383(总共为16384个slot,即214个slot) slot会 ...

  10. ios开发第三方库--cocoapods安装

    1. ruby -e "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/master/install)&quo ...