libglvnd: the GL Vendor-Neutral Dispatch library ================================================ Introduction ------------ libglvnd is a vendor-neutral dispatch layer for arbitrating OpenGL API calls between multiple vendors. It allows multiple drivers from different vendors to coexist on the same filesystem, and determines which vendor to dispatch each API call to at runtime. Both GLX and EGL are supported, in any combination with OpenGL and OpenGL ES. libglvnd was originally described in Andy Ritger's OpenGL ABI proposal [1]. Building the library ---------------------- libglvnd build-depends on libx11, glproto and libxext. On Debian and derivatives, run: sudo apt-get install libxext-dev libx11-dev x11proto-gl-dev Run `./autogen.sh`, then run `./configure` and `make`. Code overview ------------- The code in the src/ directory is organized as follows: - GLX/ contains code for libGLX, the GLX window-system API library. - EGL/ contains code for libEGL, the EGL window-system API library. - GLdispatch/ contains code for libGLdispatch, which is responsible for dispatching OpenGL functions to the correct vendor library based on the current EGL or GLX rendering context. This implements the guts of the GL API libraries. Most of the dispatch code is based on Mesa's glapi. - OpenGL/, GLESv1/, and GLESv2/ contain code to generate libOpenGL.so, libGLESv1\_CM.so, and libGLESv2.so, respectively. All three are merely wrapper libraries for libGLdispatch. Ideally, these could be implemented via ELF symbol filtering, but in practice they need to be implemented manually. See the Issues section for details on why this is the case. - GL/ contains code for libGL. This is a wrapper around libGLdispatch and libGLX. - util/ contains generic utility code. In addition, libglvnd uses a GLX extension, [GLX\_EXT\_libglvnd](https://khronos.org/registry/OpenGL/extensions/EXT/GLX_EXT_libglvnd.txt), to determine which vendor library to use for a screen or XID. There are a few good starting points for familiarizing oneself with the code: - Look at the vendor-library to GLX ABI defined in `libglxabi.h`. - Follow the flow of `glXGetProcAddress() -> __glDispatchGetProcAddress() -> _glapi_get_proc_address()` to see how the dispatch table is updated as new GL stubs are generated, and how GLX looks for vendor-library-implemented dispatchers for GLX extension functions. - Follow the flow of `glXMakeContextCurrent() -> __glDispatchMakeCurrent() -> _glapi_set_current()` to see how the current dispatch table and state is updated by the API library. - Look at `libglxmapping.c:__glXLookupVendorBy{Name,Screen}()` to see how vendor library names are queried. - For EGL, follow the flow of `eglGetPlatformDisplay()` to see how EGL selects a vendor library. The tests/ directory contains several unit tests which verify that dispatching to different vendors actually works. Run `make check` to run these unit tests. Architecture ------------ The library organization differs slightly from that of Andy's original proposal. See the diagram below:
┌──────────────────────────────────┐ │ │ ┌─────┤ Application │ │ │ │ │ └─────┬───────────────────┬────────┘ │ │ │ │ ┌─────▾─────┐ │ ┌──────────────────────┐ │ │ │ │ │ │ │ │ libOpenGL │ │ │ │ │ │ │ │ │ X server │ │ └─────┬─────┘ │ │ │ │ DT_FILTER │ │ │ │ ┌─────▾──────────┐ ┌──────▾────────┐ │ ┌──────────────────┐ │ │ │ │ │ │ └─│GLX_EXT_libglvnd │─┘ │ │ [mapi/glapi] ◂─▸ │ │extension │ │ │ libGLdispatch │ │ libGLX ├─────────────▸──────────────────┘ │ │ │ │ ◂──────────┬─────────────────┐ │ └───────▴────────┘ └──────▴────────┘ │ │ │ DT_FILTER DT_FILTER ┌─▾─────────┐ ┌───▾────────┐ │ ┌───────┴─────────────────┴────────┐ │ │ │ │ │ │ │ │ │ │ │ └─────▸ libGL │ │ GLX_vendor│ │ GLX_vendor2│ └──────────────────────────────────┘ │ │ │ │ └───────────┘ └────────────┘In this diagram, * `A ───▸ B` indicates that module A calls into module B. * `A ── DT_FILTER ──▸ B` indicates that DSO A is (logically) a filter library on DSO B. If ELF symbol filtering is enabled, symbols exported by A are resolved to entrypoints in B. libGLX manages loading GLX vendor libraries and dispatching GLX core and extension functions to the right vendor. GLX\_EXT\_libglvnd is a simple GLX extension which allows libGLX to determine the number of the screen belonging to an arbitrary drawable XID, and also the GL vendor to use for a given screen. libGLdispatch implements core GL dispatching and TLS. It acts as a thin wrapper around glapi which provides some higher-level functionality for managing dispatch tables, requesting vendor proc addresses, and making current to a given context + dispatch table. This is a separate library rather than statically linked into libGLX, since the same dispatching code is used by both GLX and EGL. libOpenGL is a wrapper library to libGLdispatch which exposes OpenGL 4.5 core and compatibility entry points. libGLESv{1,2} are wrapper libraries to libGLdispatch which expose OpenGL ES entrypoints. libGL is a wrapper library to libGLdispatch and libGLX which is provided for backwards-compatibility with applications which link against the old ABI. Note that since all OpenGL functions are dispatched through the same table in libGLdispatch, it doesn't matter which library is used to find the entrypoint. The same OpenGL function in libGL, libOpenGL, libGLES, and the function pointer returned by glXGetProcAddress are all interchangeable. ### OpenGL dispatching ### By definition, all OpenGL functions are dispatched based on the current context. OpenGL dispatching is handled in libGLdispatch, which is used by both EGL and GLX. libGLdispatch uses a per-thread dispatch table to look up the correct vendor library function for every OpenGL function. When an application calls eglMakeCurrent or glXMakeCurrent, the EGL or GLX library finds the correct dispatch table and then calls into libGLdispatch to set that table for the current thread. Since they're all dispatched through the common libGLdispatch layer, that also means that all OpenGL entrypoints will work correctly, regardless of whether the current context is from EGL or GLX. ### GLX dispatching ### Unlike core OpenGL functions, whose vendor can be determined from the current context, many GLX functions are context-independent. In order to successfully map GLX API calls to the right vendor, we use the following strategy: * Most GLX entry points specify (either explicitly, or implicitly) an X screen. * On a per-entry point basis, dispatch the call to the `libGLX_VENDOR.so` for that screen. * The first time `libGLX.so` gets called with a unique combination of X Display + screen, do the following: * Use the Display connection to query the X server for the GLX vendor of that X screen. * Load the corresponding `libGLX_VENDOR.so`. * Read the vendor's GLX dispatch table from the `libGLX_VENDOR.so`. * Cache that Display + screen <=> vendor dispatch table mapping, for use in subsequent dispatching. * Some GLX entry points imply an X screen by a GLX object they specify. Such GLX objects are: * GLXContext (an opaque pointer) * GLXFBConfig (an opaque pointer) * GLXPixmap (an XID) * GLXDrawable (an XID) * GLXWindow (an XID) * GLXPbuffer (an XID) To map from object to screen, record the corresponding screen when the object is created. This means the current process needs to see a GLX call to create the object. In the case of the opaque pointers, this is reasonable, since the pointer is only meaningful within the current process. XIDs, however, can be created by another process, so libGLX may not know in advance which screen they belong to. To deal with that, libGLX queries the server using the GLX extension GLX\_EXT\_libglvnd. ### EGL dispatching ### EGL dispatching works similarly to GLX, but there are fewer object types to deal with. Almost all EGL functions are dispatched based on an EGLDisplay or EGLDeviceEXT parameter. EGL can't rely on asking an X server for a vendor name like GLX can, so instead, it enumerates and loads every available vendor library. Loading every vendor is also needed to support extensions such as EGL\_EXT\_device\_enumeration. In order to find the available vendor libraries, each vendor provides a JSON file in a well-known directory, similar to how Vulkan ICD's are loaded. When the application calls eglGetPlatformDisplay, EGL will simply call into each vendor library until it finds one that succeeds. After that, whichever vendor succeeded owns that display. As with GLX, vendor libraries must provide dispatch stubs for any display or device extensions that they support, so that they can add new extensions without having to modify libglvnd. Since libglvnd passes eglGetPlatformDisplay calls through to each vendor, a vendor can also add a new platform extension (e.g., EGL\_KHR\_platform\_x11) without changing libglvnd. Other EGL client extensions, by definition, do require modifying libglvnd. Those are handled on a case-by-case basis. Issues ------ * Ideally, several components of libglvnd (namely, the `libGL` wrapper library and the `libOpenGL, libGLES{v1_CM,v2}` interface libraries) could be implemented via ELF symbol filtering (see [2] for a demonstration of this). However, a loader bug (tracked in [3]) makes this mechanism unreliable: dlopen(3)ing a shared library with `DT_FILTER` fields can crash the application. Instead, for now, ELF symbol filtering is disabled by default, and an alternate approach is used to implement these libraries. * The library currently indirectly associates a drawable with a vendor, by first mapping a drawable to its screen, then mapping the screen to its vendor. However, it may make sense in render offload scenarios to allow direct mapping from drawables to vendors, so multiple vendors could potentially operate on drawables in the same screen. The problem with this is that several GLX functions, such as glXChooseFBConfig(), explicitly refer to screens, and so it becomes a gray area which vendor the call should be dispatched to. Given this issue, does it still make more sense to use a direct drawable to vendor mapping? How would this be implemented? Should we add new API calls to "GLX Next"? * Note that the (drawable -> screen -> vendor) mapping mainly exists in the GLX_EXT_libglvnd extension. libGLX itself keeps a simple (drawable -> vendor) mapping, and exposes that mapping to the vendor libraries. * Along the same lines, would it be useful to include a "glXGetProcAddressFromVendor()" or "glXGetProcAddressFromScreen()" entrypoint in a new GLX version to obviate the need for this library in future applications? * Global state is required by both libGLX.so and libGLdispatch.so for various purposes, and needs to be protected by locks in multithreaded environments. Is it reasonable for the vendor-neutral library to depend on pthreads for implementing these locks? While there is no harm in having the API libraries link against pthreads even if the application does not, we would like to avoid pthread locking overhead if the application is single-threaded. Hence, this library uses a `glvnd_pthread` wrapper library which provides single-threaded fallbacks for applications which are not linked against pthreads. It is expected that multi-threaded applications will either statically link against pthreads, or load pthreads prior to loading libGL. * Is using a hash table to store GLX extension entrypoints performant enough for dispatching? Should we be using a flat array instead? * How should malloc(3) failures be handled? * How should forking be handled? * The current libGLX implementation stores all GLXContext and GLXFBConfig handles in global hashtables, which means that GLXContext and GLXFBConfig handles must be unique between vendors. That is, two vendor libraries must not come up with the same handle value for a GLXContext or GLXFBConfig. To that end, GLXContext and GLXFBConfig handles must be pointers to memory addresses that the vendor library somehow controls. The values are otherwise opaque. * Querying an XID <=> screen mapping without somehow "locking" the XID is inherently racy, since a different process may destroy the drawable, and X may recycle the XID, after the mapping is saved client-side. Is there a mechanism we could use to notify the API library when a mapping is no longer valid? References ---------- [1] https://github.com/aritger/linux-opengl-abi-proposal/blob/master/linux-opengl-abi-proposal.txt [2] https://github.com/aritger/libgl-elf-tricks-demo [3] https://sourceware.org/bugzilla/show_bug.cgi?id=16272 Acknowledgements ------- Thanks to Andy Ritger for the original libGLX implementation and README documentation. ### libglvnd ### libglvnd itself (excluding components listed below) is licensed as follows: Copyright (c) 2013, NVIDIA CORPORATION. Permission is hereby granted, free of charge, to any person obtaining a copy of this software and/or associated documentation files (the "Materials"), to deal in the Materials without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Materials, and to permit persons to whom the Materials are furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included unaltered in all copies or substantial portions of the Materials. Any additions, deletions, or changes to the original source files must be clearly indicated in accompanying documentation. If only executable code is distributed, then the accompanying documentation must state that "this software is based in part on the work of the Khronos Group." THE MATERIALS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE MATERIALS OR THE USE OR OTHER DEALINGS IN THE MATERIALS. ### X.Org ### libglvnd contains list.h, a linked list implementation from the X.Org project. Source code from the X.Org project is available from: http://cgit.freedesktop.org/xorg/xserver list.h carries the following license: Copyright © 2010 Intel Corporation Copyright © 2010 Francisco Jerez