以前只掌握gdb一些基础知识,还没有真正"实战"过。刚好最近同事一个进程coredump了,原因比较深,正好利用这个机会来分析下
// @ 运行:gdb [可执行程序] -c [coredump文件] gdb edu_info_recommend_svr -c core_edu_info_recomm // @ 查看堆栈信息:bt (backtrace) (gdb) bt 10 #0 0x00007fa0809b6144 in __strcmp_sse42 () from /lib64/libc.so.6 #1 0x00000000004aa8f7 in Json::Value::CZString::operator< (this=<value optimized out>, other=<value optimized out>) at /data/home/yaaf_proj/workspace/yaaf-ng/deps/libjsoncpp/json_value.cpp:191 #2 0x00000000004abf1a in operator() (this=<value optimized out>, key=<value optimized out>) at /usr/lib/gcc/x86_64-redhat-linux/4.4.7/../../../../include/c++/4.4.7/bits/stl_function.h:230 #3 find (this=<value optimized out>, key=<value optimized out>) at /usr/lib/gcc/x86_64-redhat-linux/4.4.7/../../../../include/c++/4.4.7/bits/stl_tree.h:1424 #4 find (this=<value optimized out>, key=<value optimized out>) at /usr/lib/gcc/x86_64-redhat-linux/4.4.7/../../../../include/c++/4.4.7/bits/stl_map.h:659 #5 Json::Value::operator[] (this=<value optimized out>, key=<value optimized out>) at /data/home/yaaf_proj/workspace/yaaf-ng/deps/libjsoncpp/json_value.cpp:1195 #6 0x0000000000438448 in Cmp6 (CourseInfo1=..., CourseInfo2=...) at /data/home/tinshuang/work/edu_proj/edu_info_recommend_svr/worker.cpp:226 #7 0x0000000000448a73 in __unguarded_partition<__gnu_cxx::__normal_iterator<Json::Value*, std::vector<Json::Value, std::allocator<Json::Value> > >, Json::Value, bool (*)(Json::Value const&, Json::Value const&)> (__first=..., __last=..., __depth_limit=10, __comp=0x438430 <Cmp6(Json::Value const&, Json::Value const&)>) at /usr/lib/gcc/x86_64-redhat-linux/4.4.7/../../../../include/c++/4.4.7/bits/stl_algo.h:2230 #8 std::__introsort_loop<__gnu_cxx::__normal_iterator<Json::Value*, std::vector<Json::Value, std::allocator<Json::Value> > >, long, bool (*)(Json::Value const&, Json::Value const&)> (__first=..., __last=..., __depth_limit=10, __comp=0x438430 <Cmp6(Json::Value const&, Json::Value const&)>) at /usr/lib/gcc/x86_64-redhat-linux/4.4.7/../../../../include/c++/4.4.7/bits/stl_algo.h:2301 #9 0x0000000000448aea in std::__introsort_loop<__gnu_cxx::__normal_iterator<Json::Value*, std::vector<Json::Value, std::allocator<Json::Value> > >, long, bool (*)(Json::Value const&, Json::Value const&)> (__first=..., __last=..., __depth_limit=10, __comp=0x438430 <Cmp6(Json::Value const&, Json::Value const&)>) at /usr/lib/gcc/x86_64-redhat-linux/4.4.7/../../../../include/c++/4.4.7/bits/stl_algo.h:2302 (More stack frames follow...)
从总的堆栈信息看,很多和STL与Json.Cpp相关,最终core在了Json.Cpp调用strcmp()上面。STL和Json.Cpp都是久经考验的第三方库,怎么可能会core?所以开始时很疑惑
// @ 切换当前栈:f (frame) (gdb) f 1 #1 0x00000000004aa8f7 in Json::Value::CZString::operator< (this=<value optimized out>, other=<value optimized out>) at /data/home/yaaf_proj/workspace/yaaf-ng/deps/libjsoncpp/json_value.cpp:191 191 return strcmp( cstr_, other.cstr_ ) < 0; // @ 打印相关的源代码:l (list) (gdb) l 187 bool 188 Value::CZString::operator<( const CZString &other ) const 189 { 190 if ( cstr_ ) 191 return strcmp( cstr_, other.cstr_ ) < 0; 192 return index_ < other.index_; 193 } // @ 查看当前栈|函数的入参:info args (gdb) info args this = <value optimized out> other = <value optimized out>
可以看到,堆栈#1core在了json_value.cpp内的strcmp(cstr_, other.cstr_)里,但是打印参数this和other都是<value optimized out>。这是因为程序在编译时加了-O2优化,一些变量会存储在CPU的寄存器中,没有放入内存。所以这里我再往上层的堆栈查看
(gdb) f 2 #2 0x00000000004abf1a in operator() (this=<value optimized out>, key=<value optimized out>) at /usr/lib/gcc/x86_64-redhat-linux/4.4.7/../../../../include/c++/4.4.7/bits/stl_function.h:230 230 { return __x < __y; } (gdb) l 225 template<typename _Tp> 226 struct less : public binary_function<_Tp, _Tp, bool> 227 { 228 bool 229 operator()(const _Tp& __x, const _Tp& __y) const 230 { return __x < __y; } 231 }; (gdb) info args this = 0x3277aa0 __x = @0x7f9fbfeb2250 __y = @0x31c26b0 // @ 打印变量:p (print) (gdb) p __x $1 = (const Json::Value::CZString &) @0x7f9fbfeb2250: { cstr_ = 0x50477b "sub_bgtime", index_ = 0 } (gdb) p __y $2 = (const Json::Value::CZString &) @0x31c26b0: { cstr_ = 0x0, index_ = 0 }
可以看出,这一层堆栈core在了STL比较2个Json::Value::CZString类型变量的大小上。翻了Json.cpp的源代码,其实Json::Value::CZString只有2个变量:
{ const char * cstr_ ; int index_ }
其中cstr_表示CZString的具体字符串内容,但是可以发现,__y的cstr_=0x0,即指向NULL!(不是指向空串,空串是类似cstr_=0x50477b ""的形式)
再结合#0~#5和Json.Cpp的源代码,core的原因就比较清楚了:Json.Cpp调用[]在找"sub_bgtime"这个Key,其内部是用到std::map::find()来查找,std::map内部维护了一棵红黑树,红黑树在拿"sub_bgtime"和原有的Key(可简单看成const char*类型)做比较时,有一个Key=NULL,所以在调用strcmp("sub_bgtime",NULL)时,strcmp内部访问NULL引发了Segmentation fault段错误
所以现在问题就是为什么这个Json会有一个Key=NULL呢?从代码上看,是core在了对vector<Json::value> vec_course_list做std::sort(),以为是vec_course_list里面的Json数据有问题,所以这里从vec_course_list的堆栈内继续分析下去
(gdb) f 13 #13 0x0000000000442132 in InfoWorker::GetWillCourseListOnClient (this=0x15fb1e0, req_body=<value optimized out>) at /data/home/tinshuang/work/edu_proj/edu_info_recommend_svr/worker.cpp:894 // @ 设置打印选项:set print [pretty|union|static-members...] [on|off] (这里为了显示方便) (gdb) set print union on (gdb) set print static-members off // @ 打印STL::vector:pvec (原生gdb对打印STL容器支持不是很好,网上这个叫dbinit_stl_views的小插件蛮好用的) // 可以看到这个vector拥有193个元素(Json),第1个Json值存储的内存地址为value_.map_,即0x13fe450 (gdb) pvec vec_course_list 0 elem[0]: $416 = { value_ = { int_ = 20964432, uint_ = 20964432, real_ = 1.0357805635774896e-316, bool_ = 80, string_ = 0x13fe450 "\020w\r\002", map_ = 0x13fe450 }, type_ = Json::objectValue, allocated_ = -1, comments_ = 0x0 } Vector size = 193 Vector capacity = 256 Element type = Json::Value * // 按照其数据类型,把这个Json::Value内容打印出来。这里就可以看到这个Json::Value的各个Key了 (gdb) p *(Json::Value::ValueHolder::ObjectValues *) 0x13fe450 $419 = std::map with 18 elements = { [{ cstr_ = 0x19598f0 "apply_num", index_ = 1 }] = { value_ = { int_ = 1401, uint_ = 1401, real_ = 6.9218596982358641e-321, bool_ = 121, string_ = 0x579 <Address 0x579 out of bounds>, map_ = 0x579 }, type_ = Json::uintValue, allocated_ = -1, comments_ = 0x0 }, ....
当然这里把vector里面193个Json::Value的每个Key都搞出来工作量很大,现在还没什么好方法。所以先结合代码逻辑和程序日志,发现vector内各个Json::Value其实都是正常的,不可能存在Key=NULL的情况。
到这里,就开始怀疑是std::sort有问题了,或者是我们使用不当。看了STL这块的源代码,比较晦涩,所以Google了下"stl sort core",发现这个问题在网上已经被很多人提及到:std::sort()在排序时,如果比较函数对相等的元素返回true,会导致程序coredump
原因分析:std::sort()的排序分2种,当元素个数>16(_S_threshold)时选择快速排序,<=16个则选择插入排序(对象少时快排性能不理想)。按照快排原理,每次都是遍历所有值和一个中间值比较,小的放左边,大的放右边。从下面STL源代码可看出,std::sort()在遍历比较时,是没有边界保护的。如果比较相等的元素返回真,则在极端情况下(如所有元素相等,__pivot为最小|最大值时)会出现访问越界,导致coredump:
/// This is a helper function... template<typename _RandomAccessIterator, typename _Tp, typename _Compare> _RandomAccessIterator __unguarded_partition(_RandomAccessIterator __first, _RandomAccessIterator __last, _Tp __pivot, _Compare __comp) { while (true) { while (__comp(*__first, __pivot)) ++__first; --__last; while (__comp(__pivot, *__last)) --__last; if (!(__first < __last)) return __first; std::iter_swap(__first, __last); ++__first; } }