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. SVG.js 图案使用和use引用

    一.SVG.Pattern 图案设置 var draw = SVG('svg1').size(300, 300); //SVG.Pattern 图案设置 var pattern = draw.patt ...

  2. KOL:Key Opinion Leader

    什么是关键意见领袖 在营销学上,为各厂家宣传的专家或权威被称为“关键意见领袖(Key Opinion Leader),通常被定义为:拥有更多.更准确的产品信息,且为相关群体所接受或信任,并对该群体的购 ...

  3. HttpAsyncClient的连接池使用

    代码示例 public static void main(String[] args) throws Exception { ConnectingIOReactor ioReactor = new D ...

  4. 【BZOJ】【4146】 【AMPPZ2014】Divisors

    暴力 由于值的范围很小($ \leq 2*10^6$),所以用一个cnt数组统计每个值有多少个数,然后从小到大,统计每个数的倍数即可. 根据调和数?的神奇性质= =这样是$O(nlogn)$的…… / ...

  5. 深入JVM系列(二)之GC机制、收集器与GC调优

    一.回想JVM内存分配 须要了解很多其它内存模式与内存分配的,请看 深入JVM系列(一)之内存模型与内存分配 1.1.内存分配: 1.对象优先在EDEN分配 2.大对象直接进入老年代  3.长期存活的 ...

  6. hyper-v 尝试更改 状态时 应用程序遇到错误 无法初始化

    刚还原了一个问题,现在 又来一个: 让我崩溃的微软 hyper-v.这次错误 提示也没了. http://social.technet.microsoft.com/Forums/de-DE/751b2 ...

  7. [leetcode]Median of Two Sorted Arrays @ Python

    原题地址:https://oj.leetcode.com/problems/median-of-two-sorted-arrays/ 题意:There are two sorted arrays A ...

  8. 【Javascript设计模式1】-单例模式

    <parctical common lisp>的作者曾说,如果你需要一种模式,那一定是哪里出了问题.他所说的问题是指因为语言的天生缺陷,不得不去寻求和总结一种通用的解决方案. 不管是弱类型 ...

  9. CSS命名规范和规则

    一.命名规则 ).尽量不缩写,除非一看就明白的单词 二.class的命名 (1).red { color: red; } .f60 {color: #f60; } .ff8600{ color: #f ...

  10. Hadoop:开发机运行spark程序,抛出异常:ERROR Shell: Failed to locate the winutils binary in the hadoop binary path

    问题: windows开发机运行spark程序,抛出异常:ERROR Shell: Failed to locate the winutils binary in the hadoop binary ...