Shared library

Library code may be shared in memory by multiple processes, and on disk. If virtual memory is used, processes would execute the same physical page of RAM that is mapped into the different address spaces of the processes. This has advantages. For instance, on the OpenStep system, applications were often only a few hundred kilobytes in size and loaded quickly; most of their code was located in libraries that had already been loaded for other purposes by the operating system.

Programs can accomplish RAM sharing by using position-independent code, as in Unix, which leads to a complex but flexible architecture, or by using common virtual addresses, as in Windows and OS/2. These systems ensure, by various means, like pre-mapping the address space and reserving slots for each shared library, that code has a high probability of being shared. A third alternative is single-level store, as used by the IBM System/38 and its successors. This allows position-dependent code, but places no significant restrictions on where code can be placed or how it can be shared.

In some cases, different versions of shared libraries can cause problems, especially when libraries of different versions have the same file name, and different applications installed on a system each require a specific version. Such a scenario is known as DLL hell, named after the Windows and OS/2 DLL file. Most modern operating systems after 2001 have clean-up methods to eliminate such situations or use application-specific "private" libraries.[1]

Dynamic linking or late binding is linking performed while a program is being loaded (load time) or executed (runtime), rather than when the executable file is created. A dynamically linked library (dynamic-link library, or DLL, under Windows and OS/2; shareable image under OpenVMS;[2] dynamic shared object, or DSO, under Unix-like systems) is a library intended for dynamic linking. Only a minimal amount of work is done by the linker when the executable file is created; it only records what library routines the program needs and the index names or numbers of the routines in the library. The majority of the work of linking is done at the time the application is loaded (load time) or during execution (runtime). Usually, the necessary linking program, called a "dynamic linker" or "linking loader", is actually part of the underlying operating system. (However, it is possible, and not exceedingly difficult, to write a program that uses dynamic linking and includes its own dynamic linker, even for an operating system that itself provides no support for dynamic linking.)

Programmers originally developed dynamic linking in the Multics operating system, starting in 1964, and the MTS (Michigan Terminal System), built in the late 1960s.[3]

Since shared libraries on most systems do not change often, systems can compute a likely load address for each shared library on the system before it is needed and store that information in the libraries and executables. If every shared library that is loaded has undergone this process, then each will load at its predetermined address, which speeds up the process of dynamic linking. This optimization is known as prebinding or prelinking on macOS and Linux, respectively. IBM z/VM uses a similar technique, called "Discontinuous Saved Segments" (DCSS).[4] Disadvantages of this technique include the time required to precompute these addresses every time the shared libraries change, the inability to use address space layout randomization, and the requirement of sufficient virtual address space for use (a problem that will be alleviated by the adoption of 64-bit architectures, at least for the time being).

Loaders for shared libraries vary widely in functionality. Some depend on the executable storing explicit paths to the libraries. Any change to the library naming or layout of the file system will cause these systems to fail. More commonly, only the name of the library (and not the path) is stored in the executable, with the operating system supplying a method to find the library on disk, based on some algorithm.

If a shared library that an executable depends on is deleted, moved, or renamed, or if an incompatible version of the library is copied to a place that is earlier in the search, the executable would fail to load. This is called dependency hell, existing on many platforms. The (infamous) Windows variant is commonly known as DLL hell. This problem cannot occur if each version of each library is uniquely identified and each program references libraries only by their full unique identifiers. The "DLL hell" problems with earlier Windows versions arose from using only the names of libraries, which were not guaranteed to be unique, to resolve dynamic links in programs. (To avoid "DLL hell", later versions of Windows rely largely on options for programs to install private DLLs—essentially a partial retreat from the use of shared libraries—along with mechanisms to prevent replacement of shared system DLLs with earlier versions of them.)

Dynamic loading, a subset of dynamic linking, involves a dynamically linked library loading and unloading at runtime on request. Such a request may be made implicitly or explicitly. Implicit requests are made when a compiler or static linker adds library references that include file paths or simply file names. Explicit requests are made when applications make direct calls to an operating system's API.

Most operating systems that support dynamically linked libraries also support dynamically loading such libraries via a run-time linker API. For instance, Microsoft Windows uses the API functions LoadLibrary, LoadLibraryEx, FreeLibrary and GetProcAddress with Microsoft Dynamic Link Libraries; POSIX-based systems, including most UNIX and UNIX-like systems, use dlopen, dlclose and dlsym. Some development systems automate this process.

• How To Write Shared Libraries by Ulrich Drepper (with much background info)