Logo Questions Linux Laravel Mysql Ubuntu Git Menu
 

c++ linux double destruction of static variable. linking symbols overlap

Tags:

Environment: linux x64, compiler gcc 4.x

Project has following structure:

static library "slib"
-- inside this library, there is static object "sobj"

dynamic library "dlib"
-- links statically "slib"

executable "exe":
-- links "slib" statically
-- links "dlib" dynamically

at end of the program, "sobj" is destructed twice. That behaviour is expected, BUT it is destructed twice at same memory address, i.e. same "this" in destructor - as the result there is double destruction problem. I think it is due some symbol overlapping.

What the solution for that conflict? Maybe some linking option?


Here is test case:


main_exe.cpp

#include <cstdlib>

#include "static_lib.h"
#include "dynamic_lib.h"

int main(int argc, char *argv[])
{
    stat_useStatic();
    din_useStatic();
    return EXIT_SUCCESS;
}

static_lib.h

#ifndef STATIC_LIB_H
#define STATIC_LIB_H

#include <cstdio>

void stat_useStatic();
struct CTest
{
    CTest(): status(isAlive)
    {
        printf("CTest() this=%d\n",this);
    }
    ~CTest()
    {
        printf("~CTest() this=%d, %s\n",this,status==isAlive?"is Alive":"is Dead");
        status=isDead;
    }
    void use()
    {
        printf("use\n");
    }
    static const int isAlive=12385423;
    static const int isDead=6543421;
    int status;

    static CTest test;
};

#endif

static_lib.cpp

#include "static_lib.h"

CTest CTest::test;

void stat_useStatic()
{
    CTest::test.use();
}

dynamic_lib.h

#ifndef DYNAMIC_LIB_H
#define DYNAMIC_LIB_H

#include "static_lib.h"

#ifdef WIN32
#define DLLExport __declspec(dllexport)
#else
#define DLLExport 
#endif
DLLExport void din_useStatic();


#endif

dynamic_lib.cpp

#include "dynamic_lib.h"

DLLExport void din_useStatic()
{
    CTest::test.use();
}

CMakeLists.txt

project( StaticProblem )
cmake_minimum_required(VERSION 2.6)
if(WIN32)
else(WIN32)
    ADD_DEFINITIONS(-fPIC)
endif(WIN32)

ADD_LIBRARY( static_lib  STATIC static_lib.cpp static_lib.h)

ADD_LIBRARY( dynamic_lib SHARED dynamic_lib.cpp dynamic_lib.h)
TARGET_LINK_LIBRARIES( dynamic_lib static_lib )

ADD_EXECUTABLE( main_exe main_exe.cpp )
TARGET_LINK_LIBRARIES( main_exe static_lib dynamic_lib )

That example works OK, on windows, but on linux - there is problem. As it works ok on windows, solution should be like change some linking option or something like that, but not change project structure or not use static vars.

Output:

Windows

CTest() this=268472624
CTest() this=4231488
use
use
~CTest() this=4231488, is Alive
~CTest() this=268472624, is Alive

Linux

CTest() this=6296204
CTest() this=6296204
use
use
~CTest() this=6296204, is Alive
~CTest() this=6296204, is Dead
like image 212
John Avatar asked Jul 15 '11 23:07

John


2 Answers

TL;DR: you should not link a library once as a static dependency and once as a dynamic dependency.


How are the destructors of static variables executed in the Itanium ABI (used by clang, gcc, icc...)?

The C++ Standard Library offers a standard facility to schedule the execution of a function during the program shut-down (after main has ended) in the format of atexit.

The behavior is relatively simple, atexit basically builds a stack of callbacks and will thus execute them in the reverse order of their scheduling.

Whenever a static variable is constructed, immediately after its construction ends, a callback is registered in the atexit stack to destroy it during shutdown.


What happens when a static variable exists both in a statically linked library and a dynamically linked library?

It attempts to exist twice.

Each library will have:

  • a memory area reserved for the variable, pointed to by the corresponding symbol (the mangled name for the variable),
  • an entry in the load section to build the variable, and schedule its destruction.

The surprise comes from the way symbol resolution works in the loader. Essentially, the loader builds up a mapping between symbol and location (pointer), in a first come first serve basis.

However, the load/unload sections are nameless, and therefore each of them is executed in full.

Therefore:

  • the static variable is constructed a first time,
  • the static variable is constructed a second time over the first one (which is leaked),
  • the static variable is destructed a first time,
  • the static variable is destructed a second time; which is generally where the problem is detected.

So what?

The solution is simple: NOT linking against both a static library A (directly) and a dynamic library B also linking against A (dynamically or statically).

Depending on the use case, you may either:

  • link statically against B,
  • link dynamically against both A and B.

As it works OK on windows, solution should be like change some linking option or something like that, but not change project structure or not use static vars.

In the unlikely event where you really need two independent instances of the static variable, apart from refactoring your code, it is possible to instead hide the symbols in your dynamic library.

This Windows' default behavior, which is why the DLLExport attribute is required there, and why since it was forgotten for CTest::test the behavior on Windows is different.

Do note, however, that any future maintainer of this project will curse you loudly if you opt for this behavior. Nobody expects a static variable to have multiple instances.

like image 156
Matthieu M. Avatar answered Oct 15 '22 06:10

Matthieu M.


OK, I have found solution:

http://gcc.gnu.org/wiki/Visibility

For example if change

static CTest test;

to

__attribute__ ((visibility ("hidden"))) static CTest test;

problem will gone. Linux:

CTest() this=-1646158468
CTest() this=6296196
use
use
~CTest() this=6296196, is Alive
~CTest() this=-1646158468, is Alive

nm output before fix was:

0000000000200dd4 B _ZN5CTest4testE

after fix:

0000000000200d7c b _ZN5CTest4testE

Difference is changed global symbol "B" to local symbol "b".

Instead of adding "attribute ((visibility ("hidden")))" to symbols, it is possible to use compiler option "-fvisibility=hidden". That option makes gcc to behave much more like Windows env.

like image 43
John Avatar answered Oct 15 '22 06:10

John