netCDF与HDF5库冲突引发段错误的成因及解决方法问询

问题说明 #

经过数小时的程序调试，确认netCDF与HDF5库之间存在冲突（程序需要同时读写两种格式的文件），已找到部分临时解决方案，详见本文末尾。

最小复现代码 #

如下精简代码运行后会直接触发段错误，甚至不需要调用包含HDF5对象定义的函数：

#include <iostream>
#include <string>

#include "H5Cpp.h"

#include <netcdf>

using namespace std;

void stupidfunction() // 注意：该函数从未被调用
{
    H5::Group grp1; // 仅该对象的潜在定义就会导致netCDF出现段错误！
}

int main(int argn, char ** args)
{
    std::string outputFilename = "/tmp/test.nc";

    try
    {   
        std::cout << "Now opening " << outputFilename << std::endl;
        netCDF::NcFile sfc;
        sfc.open(outputFilename, netCDF::NcFile::replace);

        std::cout << "closing file" << std::endl;
        sfc.close();

        return true;
    }
    catch(netCDF::exceptions::NcException& e)
    {
        std::cout << "EX: " << e.what() << std::endl;
        return false;
    }
    
    return 0;
}

编译命令 #

h5c++ test.cpp -std=gnu++11 -O0 -g3 -lnetcdf_c++4 -lnetcdf  -o test

相关依赖版本 #

libnetcdf-c++4-1                       4.3.1-2build1                         amd64        C++ interface for scientific data access to large binary data
libnetcdf-c++4-dev                     4.3.1-2build1                         amd64        creation, access, and sharing of scientific data in C++
libnetcdf-dev                          1:4.7.3-1                             amd64        creation, access, and sharing of scientific data
libnetcdf15:amd64                      1:4.7.3-1                             amd64        Interface for scientific data access to large binary data
netcdf-bin                             1:4.7.3-1                             amd64        Programs for reading and writing NetCDF files
netcdf-doc                             1:4.7.3-1                             all          Documentation for NetCDF

hdf5-helpers                           1.10.4+repack-11ubuntu1               amd64        Hierarchical Data Format 5 (HDF5) - Helper tools
hdf5-tools                             1.10.4+repack-11ubuntu1               amd64        Hierarchical Data Format 5 (HDF5) - Runtime tools
libhdf4-0                              4.2.14-1ubuntu1                       amd64        Hierarchical Data Format library (embedded NetCDF)
libhdf5-103:amd64                      1.10.4+repack-11ubuntu1               amd64        Hierarchical Data Format 5 (HDF5) - runtime files - serial version
libhdf5-cpp-103:amd64                  1.10.4+repack-11ubuntu1               amd64        Hierarchical Data Format 5 (HDF5) - C++ libraries
libhdf5-dev                            1.10.4+repack-11ubuntu1               amd64        Hierarchical Data Format 5 (HDF5) - development files - serial versio

错误信息 #

GDB运行报错 #

Program received signal SIGSEGV, Segmentation fault.
0x00007ffff7a47a01 in __vfprintf_internal (s=s@entry=0x7fffff7ff480, 
    format=format@entry=0x7ffff77e13a8 "can't locate ID", 
    ap=ap@entry=0x7fffff7ff5e0, mode_flags=mode_flags@entry=2)
    at vfprintf-internal.c:1289
1289    vfprintf-internal.c: No such file or directory.

GDB栈回溯信息 #

(gdb) bt
#0  0x00007ffff7a47a01 in __vfprintf_internal (s=s@entry=0x7fffff7ff480, format=format@entry=0x7ffff77e13a8 "can't locate ID", 
    ap=ap@entry=0x7fffff7ff5e0, mode_flags=mode_flags@entry=2) at vfprintf-internal.c:1289
#1  0x00007ffff7a5cd4a in __vasprintf_internal (result_ptr=0x7fffff7ff5d8, format=0x7ffff77e13a8 "can't locate ID", args=0x7fffff7ff5e0, mode_flags=2)
    at vasprintf.c:57
#2  0x00007ffff75c7e56 in H5E_printf_stack () from /usr/lib/x86_64-linux-gnu/libhdf5_serial.so.103
#3  0x00007ffff76553b9 in H5I_inc_ref () from /usr/lib/x86_64-linux-gnu/libhdf5_serial.so.103
...
many many lines repeating H5E_printf_stack, H5E__push_stack and H5I_inc_ref
...
#56139 0x00007ffff76553b9 in H5I_inc_ref () from /usr/lib/x86_64-linux-gnu/libhdf5_serial.so.103
#56140 0x00007ffff75c7c2f in H5E__push_stack () from /usr/lib/x86_64-linux-gnu/libhdf5_serial.so.103
#56141 0x00007ffff75c7e7e in H5E_printf_stack () from /usr/lib/x86_64-linux-gnu/libhdf5_serial.so.103
#56142 0x00007ffff761fc85 in H5G_loc () from /usr/lib/x86_64-linux-gnu/libhdf5_serial.so.103
#56143 0x00007ffff7546903 in H5Acreate1 () from /usr/lib/x86_64-linux-gnu/libhdf5_serial.so.103
#56144 0x00007ffff790b11b in NC4_write_provenance () from /usr/lib/x86_64-linux-gnu/libnetcdf.so.15
#56145 0x00007ffff790b5a8 in ?? () from /usr/lib/x86_64-linux-gnu/libnetcdf.so.15
#56146 0x00007ffff790b7b0 in nc4_close_hdf5_file () from /usr/lib/x86_64-linux-gnu/libnetcdf.so.15
#56147 0x00007ffff790b9ea in NC4_close () from /usr/lib/x86_64-linux-gnu/libnetcdf.so.15
#56148 0x00007ffff78ca579 in nc_close () from /usr/lib/x86_64-linux-gnu/libnetcdf.so.15
#56149 0x00007ffff7f82270 in netCDF::NcFile::close() () from /usr/lib/x86_64-linux-gnu/libnetcdf_c++4.so.1
#56150 0x00005555555a7959 in main (argn=1, args=0x7fffffffe5b8) at test.cpp:29
(gdb) 

问题根因 #

该错误是典型的库符号冲突导致的：

1. netCDF4格式的底层本身依赖HDF5实现读写，你手动链接的独立HDF5库和netCDF依赖的HDF5库存在大量重复符号
2. HDF5的C++接口会注册全局的构造/析构逻辑，即使你没有显式调用相关代码，这些全局逻辑也会污染netCDF内部调用的HDF5上下文，导致HDF5内部ID查找出现死循环，最终栈溢出触发段错误
3. 指定classic格式后，netCDF不会调用底层HDF5相关逻辑，所以不会触发冲突

解决方案 #

临时规避方案 #

将netCDF文件格式指定为classic或classic64即可避免错误，修改示例如下：

sfc.open(outputFilename, netCDF::NcFile::replace, netCDF::NcFile::classic);

或者

sfc.open(outputFilename, netCDF::NcFile::replace, netCDF::NcFile::classic64);

永久修复方案 #

1. 调整链接顺序，把netCDF相关库放在HDF5库前面，让链接器优先使用netCDF依赖的HDF5符号，编译命令修改为：

   h5c++ test.cpp -std=gnu++11 -O0 -g3 -lnetcdf_c++4 -lnetcdf -lhdf5_cpp -lhdf5 -o test
   

2. 编译netCDF的时候使用和项目完全一致的HDF5版本，保证两者的HDF5依赖完全对齐
3. 如果不需要在同一个编译单元同时调用两种库的接口，可以把HDF5相关代码和netCDF相关代码拆分到不同的动态库中，避免符号直接冲突

内容的提问来源于stack exchange，提问作者Knut Stanley Jacobsen