Simple C++ Symbol Visibility Demo

This post gives a demonstration/test of the dangers of using default visibility for symbols in C++ shared libraries. It briefly walks through why and how to use the -fvisibility=hidden compilation flag along with __attribute__((__visibility__("default"))) in symbol declarations.

On Linux, symbols have a default visibility that allows them to be loaded at runtime. If multiple libraries export the same symbol, this can lead to undefined behavior. In real terms, this means segmentation faults, deadlocks, or other quirky behavior can occur. While multiple solutions to this problem exist, this document discusses how to prevent this behavior by specifying symbol visibility during shared library compilation.

Prerequisites

To run the demo, gcc and Make are required.

An Example

Example code: Simple C++ Symbol Visibility Demo

To run the demonstration, run demonstrate.sh. This will build and run the demo twice — building once with default visibility and building once with hidden visibility. On Linux, default visibility causes libget_three.so and libget_seven.so to effectively share the same definition for internal_do_calculation(), which produces incorrect results as far as the author of PublicGetSeven() is concerned:

PublicGetThree returned 3
PublicGetSeven returned 3

Hidden visibility causes the libraries to use their own definitions of internal_do_calculation(), so running the .sos with hidden visibility outputs correct results:

PublicGetThree returned 3
PublicGetSeven returned 7

In Detail

get_three.cpp and get_seven.cpp are the implementations of two shared libraries, respectively named libget_three.so and libget_seven.so. get_three.cpp and get_seven.cpp each define a function with the signature int internal_do_calculation(). Meanwhile, test.cpp links to both of these libraries and calls functions from their public APIs, respectively named PublicGetThree() and PublicGetSeven() — both of these functions call internal_do_calculation() to get the number to be returned:

  • internal_do_calculation() in get_three.cpp returns 3
  • internal_do_calculation() in get_seven.cpp returns 7

Naively, it seems like this should work, since each shared library doesn't know about the other shared library's version of internal_do_calculation(). However, on Linux, when the test program is executed (without mitigating factors) the following events occur at runtime:

  • the dynamic loader looks for PublicGetThree and finds it in libget_three.so
  • the dynamic loader looks for PublicGetSeven and finds it in libget_seven.so
  • the dynamic loader looks for internal_do_calculation and in this case finds it in libget_three.so

For the purposes of this demonstration, the loader is done at this point and doesn't care that libget_seven.so will call internal_do_calculation() as defined in libget_three.so instead of its own internal_do_calculation():

  • main() in test calls:
  • PublicGetSeven() in libget_seven.so, which calls:
  • internal_do_calculation() in libget_three.so

In real-world situations, this can be problematic, especially since common libraries may be statically compiled into multiple shared libraries, such as via .a archive files or header-only include files. Conflicting versions of such common libraries can be loaded at runtime, resulting in segfaults, deadlocks, and other undefined behavior. For example, see the symbol conflict between LJM and ROS (fixed as of LJM 1.19) below.

Solution: -fvisibility=hidden and __attribute__((__visibility__("default")))

To solve the above problem using hidden visibility:

  1. Use the -fvisibility=hidden compilation flag to compile all symbols with hidden visibility (unless a symbol is explicitly specified to be public).
  2. Use the __attribute__((__visibility__("default"))) attribute to explicitly define the public API as public.

Part 1: -fvisibility=hidden

Hidden visibility for a given symbol causes a shared library to never load that hidden symbol from any other component's symbol definitions, but instead load that symbol from the shared library's own symbol definitions. Additionally, symbols with hidden visibility are not exported — other components will not be able to use them. (More on this below.)

In the second build of demonstrate.sh, the libraries are compiled with -fvisibility=hidden to make all symbols hidden (except for symbols that have been explicitly marked with a different visibility). This means that at runtime:

  • libget_three.so will load its own internal_do_calculation() and
  • libget_seven.so will load its own internal_do_calculation().

Part 2: __attribute__((__visibility__("default")))

Outside a component, trying to explicitly call a function with hidden visibility results in undefined reference errors. To get around this and allow a shared library's public API functions to be called, the visibility for those chosen public functions should be set to the default visibility.

("Default" is an unfortunate name for this visibility type since the -fvisibility compiler flag sort of sets a default visibility type — it sets the visibility type for all symbols that don't have a more explicit visibility type. In other words, it may be helpful to think of the "default" visibility type as the visible or exported visibility type, while the -fvisibility flag sets the default visibility type.)

To expose a given symbol, you can use __attribute__((__visibility__("default"))) to decorate a symbol, e.g.:

__attribute__((__visibility__("default"))) int PublicGetSeven();

Other methods to specify visibility exist, such as using #pragma directives. See below for more details.

Benefit: Compatibility

Compiling with hidden visibility means that:

  • Symbol definitions are assumed to be located within the component using those symbols
  • Symbols are not exported

Therefore, if a shared library is compiled with hidden visibility:

  • It will not accidentally use symbols from another component
  • It will not let its own symbols be accidentally used by another component

This is an advantage over -Bsymbolic or -Bsymbolic-functions because it doesn't matter as much if other shared libraries adhere to correct symbol visibility.

Benefit: Optimizations

Another benefit to hidden symbols is that additional optimizations can be made. In short:

  • Library load time is improved
  • Runtime speed can be improved because the compiler knows that it can optimize via devirtualization and inlining of functions
  • Shared library binary size can be reduced because the compiler can omit hidden symbols from the exported symbols table

Downside: Unit Testing is Harder

Since symbols are not exported when using hidden visibility, unit testing of internal functions requires building the library with default visibility. This may not be much of a problem, however, since there are different compilation flags that you'd want to apply to debug and release builds.

Example debug build flags:

  • -g - add debug info
  • -O0 - no optimization (improves experience using debuggers)

Example release build flags:

  • -fvisibility=hidden
  • -O2 - moderate size and speed optimizations (many other optimization options exist as well)

Alternative to building for both debug and release, source code could be compiled into a single component with unit test code, so that compiled unit test code has access to compiled source code symbols.

Downside: Problems with C++ Exceptions

To use hidden visibility, any defined type that is passed out of a shared library must have exported symbols (i.e. default visibility), including exceptions. This is because when binary code in a given component catches an exception, the exception requires a typeinfo lookup. However, typeinfo symbols for symbols with hidden visibility are also hidden.

In short, hidden visibility can still be used with C++ exceptions, but if the shared library's API conceptually does not include exception throwing, it's even simpler because there's no danger of throwing an exception of a hidden type.

On macOS

-fvisibility=hidden is not needed to fix this on macOS. From the GNU documentation:

On Darwin, default visibility means that the declaration is visible to other modules.

This is in contrast to ELF format (i.e. Linux), which allows declarations to also be overridden:

On ELF, default visibility means that the declaration is visible to other modules and, in shared libraries, means that the declared entity may be overridden.

Running demonstrate.sh on macOS produces the following output for both runs of test, with and without hidden visibility:

PublicGetThree returned 3
PublicGetSeven returned 7

LD_PRELOAD

The environment variable LD_PRELOAD can be used to force the runtime loader to load symbols from one library instead of others. However:

  • LD_PRELOAD is ignored for symbols that were compiled with hidden visibility.
  • LD_PRELOAD is ignored for libraries that were linked with -Bsymbolic/-Bsymbolic-functions.

Other Solutions

Other solutions include:

  • ld Version scripts
  • The -Bsymbolic linker flag(s)
  • -fvisibility=protected
  • Namespaces
  • Dynamic linking

ld Version Scripts

LJM 1.20 and onward will use version scripts to hide all symbols except for the public API in LabJackM.h.

The ld linker has the capability to hide all symbols except for a set of specified symbols. This works similar to how Windows .def files work. However, ld version scripts can specify patterns so that listing all functions explicitly is not required. For example, LJM's version script looks as follows.

libLabJackM.version:

LJM_1 {
  global: LJM_*;
  local: *;
};

This exposes all symbols that start with LJM_, while hiding all others. (* is a wildcard character.) The LJM_1 is semantic.

LJM is then linked with the following link flag:

-Wl,--version-script=libLabJackM.version

(The -Wl, is used because gcc is used as the linker runner.)

It's important to note that with this version script setup, static libraries, such as libprotobuf.a are both looked up internally and also not exported. This makes LJM compatible with other binaries that include a different version of libprotobuf.a. Hidden visibility would work for this situation, except that libprotobuf.a has its own build scripts that don't use hidden visibility. It's easier to let those build scripts work by default, and it's less error prone, because all LJM functions that are meant to be externally available will begin with the LJM_ prefix anyway.

The -Bsymbolic Linker Flag(s)

The -Bsymbolic or -Bsymbolic-functions linker flags (run as -Wl,-Bsymbolic or -Wl,-Bsymbolic-functions when with gcc) is another way to solve this issue. For example, you could compile and link this demo as follows:

g++ -g -Wall -o get_three.o -c -fPIC get_three.cpp
g++ -Wl,-Bsymbolic -o libget_three.so -shared get_three.o
g++ -g -Wall -o get_seven.o -c -fPIC get_seven.cpp
g++ -Wl,-Bsymbolic -o libget_seven.so -shared get_seven.o
g++ -g -Wall -Wl,-rpath=. -o test test.cpp -L. -lget_three -lget_seven

This would produce the following (desired) output when running ./test:

PublicGetThree returned 3
PublicGetSeven returned 7

One advantage hidden visibility has over -Bsymbolic or -Bsymbolic-functions is that -Bsymbolic or -Bsymbolic-functions still export their symbols, so another library can accidentally use them. For example, while compiling libget_seven.so with -Bsymbolic may give the correct results:

...
g++ -o libget_three.so -shared get_three.o
g++ -o libget_seven.so -shared get_seven.o -Wl,-Bsymbolic
g++ -g -Wall -Wl,-rpath=. -o test test.cpp -L. -lget_three -lget_seven
./test
PublicGetThree returned 3
PublicGetSeven returned 7

But if the linking order while linking test is swapped so that libget_seven.so is seen be the loader first, ./test gives incorrect results:

g++ -g -Wall -Wl,-rpath=. -o test test.cpp -L. -lget_seven -lget_three
./test
PublicGetThree returned 7
PublicGetSeven returned 7

This is not a problem when compiling with hidden visibility — either get_three.o or get_seven.o could be compiled with -fvisibility=hidden and the correct results are still achieved regardless.

Another advantage hidden visibility has over -Bsymbolic or -Bsymbolic-functions is that hidden visibility helps enable additional size and speed optimizations

An advantage -Bsymbolic or -Bsymbolic-functions has over hidden visibility is that unit testing internal functions can be done without building the library twice.

-fvisibility=protected

Protected visibility is similar to hidden visibility in that it binds symbol references within a component so that a shared library compiled with -fvisibility=protected will not have its symbols overridden. However, protected visibility is similar to default visibility in that it can still "pollute" the symbol table so that unaware libraries might still get their symbols overridden.

For example this link order is successful:

g++ -g -Wall -o get_three.o -c -fPIC get_three.cpp
g++ -g -Wall -o get_seven.o -c -fPIC get_seven.cpp -fvisibility=protected
...
g++ -g -Wall -Wl,-rpath=. -o test test.cpp -L. -lget_three -lget_seven
./test
PublicGetThree returned 3
PublicGetSeven returned 7

But this link order is not:

g++ -g -Wall -o get_three.o -c -fPIC get_three.cpp -fvisibility=protected
g++ -g -Wall -o get_seven.o -c -fPIC get_seven.cpp
...
g++ -g -Wall -Wl,-rpath=. -o test test.cpp -L. -lget_three -lget_seven
./test
PublicGetThree returned 3
PublicGetSeven returned 3

Namespaces

Namespaces can solve this issue if you have control of the source code being compiled. However, if you're writing a shared library that statically links a popular dependent library, your library's version of a function defined in that popular library may conflict with another shared library's version of that popular library. For an example, see below for the problem with libLabJackM.so (pre 1.18.6) and libroscpp.so both statically linking Boost headers.

In such a case, hidden visibility circumvents the need to modify third-party code or third-party binaries.

Dynamic Linking

If libraries A and B both use a dependent library C, it can be a problem if A and B use different versions of C, because the internal implementation in C may have changed such that problems occur. Imagine an old version of library C defines a given data structure one way and a newer version of library C defines it in a different way. It could be that the size of a given data structure changes or the order of data fields in that data structure. In this case, if an instance of that data structure is initialized in the older version of library C and used in the newer version of library C (or vice versa), problems will occur.

However, if A and B are both compatible with a given major version of library C, dynamic linking solves this issue. All data structures will be initialized and used with the same version of library C.

In some cases, dynamic linking can present a disadvantage. One reason to statically link is to have minimal dependencies for the sake of simple distribution. Another reason is that it's difficult to test all versions of a dependency, so statically linking to a particular version of a dependency allows for consistent behavior of that dependency.

Real World Example: Boost Thread Conflict in LJM and ROS

First of all, thanks to Jeremy Xu and Mike Purvis from Clearpath Robotics for their patience in helping LabJack uncover and fix this issue!

libLabJackM.so (version 1.18.5 and prior) and libroscpp.so (tested with version Kinetic Kame) both contain external symbol references related to the Boost C++ libraries. Specifically, boost::thread_group caused conflict at runtime, causing deadlocks to occur. (libLabJackM.so version 1.19.0 and later is compiled with hidden visibility to solve this issue.)

The problem was that a reference to:

  • boost::thread_group::~thread_group() in libLabJackM's DeviceFactory.cpp

...was jumping to:

  • boost::thread_group::~thread_group() in /opt/ros/kinetic/lib/libroscpp.so:

For example, while debugging a program that linked to both LJM 1.18.5 and ROS Kinetic Kame, a backtrace during deadlock shows the issue.

(gdb) bt
#0  __pthread_cond_destroy (cond=0x7fffdebebb30) at pthread_cond_destroy.c:76
#1  0x00007f8587b419e0 in boost::thread_group::~thread_group() () from /opt/ros/kinetic/lib/libroscpp.so
#2  0x00007f8586d9d9d6 in DeviceFactory::OpenAllMatchingConnections ([email protected]=0x22b47d0, request=..., cTypes=...,
    [email protected]=0x7fffdebebcc0) at source_files/os_indep/DeviceFactory.cpp:581
#3  0x00007f8586d9dd4c in DeviceFactory::OpenRequestAfterLocks ([email protected]=0x22b47d0, request=..., cTypes=...,
    [email protected]=0x7fffdebebcc0) at source_files/os_indep/DeviceFactory.cpp:432
#4  0x00007f8586d9df58 in DeviceFactory::OpenRequest ([email protected]=0x22b47d0, request=..., [email protected]=0x7fffdebebcc0)
    at source_files/os_indep/DeviceFactory.cpp:411
#5  0x00007f8586d9e18d in DeviceFactory::OpenDevice (this=0x22b47d0, adapter=0x22b46d0) at source_files/os_indep/DeviceFactory.cpp:193
#6  0x00007f8586d6fa11 in LJM_Open (DeviceType=4, ConnectionType=1, Identifier=0x415abb "LJM_idANY", Handle=0x7fffdebec050)
    at source_files/api/ExposedAPI.cpp:286
#7  0x000000000040e74d in LabJackT4::LabJackT4() ()
#8  0x000000000041450d in main ()

boost::thread_group is compiled into LabJackM and roscpp as a .hpp file, meaning that LabJackM and roscpp contain boost::thread_group symbols, while libboost_thread.a does not:

$ readelf -a /usr/lib/x86_64-linux-gnu/libboost_thread.a | grep thread_group
$ readelf -a /usr/lib/x86_64-linux-gnu/libboost_thread.so.1.58.0 | grep thread_group
$ readelf -a /opt/ros/kinetic/lib/libroscpp.so | grep thread_group
  1471: 000000000014c920   368 FUNC    WEAK   DEFAULT   12 _ZN5boost12thread_groupD1
  3163: 000000000014c920   368 FUNC    WEAK   DEFAULT   12 _ZN5boost12thread_groupD2
$ readelf -a /usr/local/lib/libLabJackM.so.1.18.5 | grep thread_group
  1438: 00000000001c8d20   796 FUNC    WEAK   DEFAULT   11 _ZN5boost12thread_group13
  1657: 00000000001c6120   170 FUNC    WEAK   DEFAULT   11 _ZN5boost12thread_groupD2
  3373: 00000000001c8b60   446 FUNC    WEAK   DEFAULT   11 _ZN5boost12thread_group8j
  6808: 00000000001c7f20   265 FUNC    WEAK   DEFAULT   11 _ZN5boost12thread_groupC1
  8215: 00000000001c7f20   265 FUNC    WEAK   DEFAULT   11 _ZN5boost12thread_groupC2
 10770: 00000000001c6120   170 FUNC    WEAK   DEFAULT   11 _ZN5boost12thread_groupD1

The header for that section as output by readelf is:

   Num:    Value          Size Type    Bind   Vis      Ndx Name

After compiling LabJackM with -fvisibility=hidden, the boost::thread_group symbols bindings were LOCAL instead of WEAK:

$ readelf -a build/libLabJackM.so.1.18.5  | grep thread_group
  1982: 000000000014c560   784 FUNC    LOCAL  DEFAULT   11 _ZN5boost12thread_group13
  3312: 000000000014c3a0   446 FUNC    LOCAL  DEFAULT   11 _ZN5boost12thread_group8j
  7193: 00000000001498e0   170 FUNC    LOCAL  DEFAULT   11 _ZN5boost12thread_groupD1
  8355: 00000000001498e0   170 FUNC    LOCAL  DEFAULT   11 _ZN5boost12thread_groupD2

More Details