Why glibc 2.34 removed libpthread

Why glibc 2.34 removed libpthread Skip to main content

The recent 2.34 release of the GNU C library, glibc, removes libpthread as a separate library. This article explains the motivation behind this change and some consequences for developers and system administrators.

For a long time, glibc was split into multiple, separate, shared objects. For example, the threading library libpthread was contained in a shared object libpthread.so.0, and the application interface for the dynamic linker, libdl, in the file libdl.so.2. There was even a time, some twenty years ago, when there were two separate implementations of libpthread, the LinuxThreads implementation for Linux 2.4 and earlier and the Native POSIX Threads Library (NPTL) implementation for Linux 2.6 and later.

In the glibc 2.34 release, we have integrated most components that used to be in separate shared objects into the main libc object, libc.so.6. These changes have been implemented in a backward-compatible fashion, so even though libpthread is gone as a separate object, all the public functions it used to provide (such as pthread_create) are still available. In this consolidation effort, glibc follows the pioneering work of the musl C library, which provides absolutely everything (including the dynamic linker) in a single shared object.

The developer view

The developer view

A textbook "Hello, world!" example using C++ threads look like this:

#include <iostream>
#include <thread>

int
main()
{
  std::thread thr{[]() {
    std::cout << "Hello, world!\n";
  }};
  thr.join();
}

Building the program with g++ on a system that uses glibc 2.33 or earlier results in an unexpected error:

$ g++ -o hello hello.cpp
/usr/bin/ld: /tmp/ccJckARF.o: in function `std::thread::thread<:>(main::{lambda()#1}&&)':
hello.cpp:(.text+0x9b): undefined reference to `pthread_create'
collect2: error: ld returned 1 exit status
{lambda()#1},>

For a beginner, this error message is very confusing. The programmer did not write pthread_create, so it is not clear why the linker would complain about its absence.

The fix is to link with libpthread, the separate thread library implementation:

$ g++ -o hello hello.cpp -lpthread
$ ./hello
Hello, world!

But with glibc 2.34, the command works without -lpthread:

$ g++ -o hello hello.cpp
$ ./hello
Hello, world!

The -lpthread option still works because glibc provides an empty libpthread.a file. This file replaces the libpthread.so symbolic link to the shared object file libpthread.so.0. The shared object still exists so that existing applications that link dynamically against it can still launch. dlopen also continues to work, but the file is empty apart from a few placeholder symbols. The presence of this file helps distribution dependency generators provide the correct set of dependencies.

The reorganization of files can seem like a trivial change, but integrating glibc components into the main library also helps more advanced use cases. For example, a programmer might want to link statically against the Gio library but dynamically against glibc. A typical way to mix linking strategies is to use pkg-config with the --push-state/--pop-state linker bracket:

$ gcc -o application main.o -Wl,--push-state,-Bstatic \
  $(pkg-config --static --libs gio-2.0)  -Wl,--pop-state

The command pkg-config --static --libs gio-2.0 prints the static libraries required by gio-2.0, and only those are marked for static linking with -Bstatic. But the command does not work as expected in glibc versions before 2.34, because the pkg-config output includes -ldl, a glibc component, and the static libdl.a library provided is incompatible with dynamic linking. Thus, the command leads to the following error:

/usr/bin/ld: /usr/lib64/libdl.a(dlopen.o): in function `dlopen':
(.text+0x9): undefined reference to `__dlopen'

This error is not very hard to fix: Just filter out -ldl from the pkg-config output and include it after the -Wl,--pop-state option instead. But in glibc 2.34, libdl has also been integrated, so libdl.a is now empty, and linking against it works in both static and dynamic linker invocations.

A downside of these changes is that we had to add many new GLIBC_2.34 symbol versions for existing functions. However, we had to add a new __libc_start_main@@GLIBC_2.34 symbol version to implement a long-requested feature, startup code hardening. __libc_start_main is called by all applications during startup. This new symbol version prevents applications that have been built against glibc 2.34 from launching on systems that have installed glibc 2.33 and earlier. Further GLIBC_2.34 symbols added for integrating libpthread and the other components did not seem much of an additional burden because of that.

The system administrator view

The system administrator view

Splitting glibc into multiple components means that some components are loaded on process start (certainly the dynamic linker and the main libc.so.6 library), whereas other components might be loaded later. Such components can be loaded indirectly, for example, if the Name Service Switch (NSS) is used to look up user information using the getpwnam function. In this case, NSS modules such as nss_files or nss_systemd are loaded behind the scenes. Some of these modules are part of glibc itself (nss_files). However, others are part of other software (e.g., systemd), and those could depend on glibc components that are not initially loaded. In both cases, it is not always possible during glibc upgrades to preserve the internal application binary interface (ABI) between the initially-loaded glibc components and the components loaded later.

If the system administrator performs a glibc upgrade and neglects to restart all services (typically with a reboot of the system), late loading of glibc components might pull updated versions of the components described in the previous paragraph into a process that uses parts of the old glibc installation. The resulting ABI mismatches can result in hard-to-diagnose failures, including crashes. One common example is that systemd can no longer launch services that use a User= directive (although systemctl daemon-reexec can usually be used to work around this).

The use of incompatible dependencies is particular problematic for libpthread, due to its tight integration with the rest of glibc. But also, when backporting changes to nss_files, distributions had to attempt to preserve the internal ABI with custom downstream patches, which is somewhat cumbersome.

Loading as much as possible of glibc at process startup makes these issues go away. Long-running processes keep using the old glibc version.

Performance considerations

Performance considerations

Most processes on a typical GNU/Linux system are already dynamically linking to libpthread, even before its integration. Loading these processes is now marginally faster, because the dynamic linker has to process fewer symbol lookups and relocations. Processes that did not load libpthread before invoke one additional system call (set_robust_list) that has been avoided before. This system call is required to make process-shared robust mutexes work even if pthread_create is never called. All the integrated components had few relative relocations, which means that that they do not contribute significantly to the overall glibc relocation overhead.

Historically, some applications interpret the absence of the pthread_create symbol as an indicator to switch from thread-safe algorithms to single-threaded algorithms, as an optimization. This optimization no longer works because the pthread_create symbol is now always present. Instead, applications should enable such optimizations based on the __libc_single_threaded variable, which was introduced in glibc 2.32, partly in preparation for the libpthread integration changes.

Remaining issues

Remaining issues

Currently, the dynamic linker still lives in a separate shared object (/lib64/ld-linux-x86-64.so.2 on x86-64, for example). As musl shows, it is theoretically possible to provide the entire C library through the dynamic linker. For glibc, this would require additional (non-mechanical) changes. Without further work, it would no longer be possible to load an optimized libc.so.6 implementation file based on CPU characteristics (e.g., a libc.so.6 version that uses PCREL instructions on POWER10, something that cannot be achieved through IFUNC-based optimizations).

We have also been unable to complete the transition of the libm, libmvec, and libresolv components in time for the glibc 2.34 release. This means that some linking and upgrade hazards remain. We hope to complete these transitions in a future release.

These updates should make it easier for distributions to backport bug fixes and other changes. Some distributions might even want to experiment with seamless upgrades across major glibc releases without requiring reboots.