libglvnd: the GL Vendor-Neutral Dispatch library ================================================ Introduction ------------ This is a work-in-progress implementation of the vendor-neutral dispatch layer for arbitrating OpenGL API calls between multiple vendors on a per-screen basis, as described by Andy Ritger's OpenGL ABI proposal [1]. Currently, only the GLX window-system API and OpenGL are supported, but in the future this library may support EGL and OpenGL ES as well. Building the library ---------------------- libglvnd build-depends on xorg-server, libx11, glproto and libxext. On Debian and derivatives, run: sudo apt-get install xserver-xorg-dev 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. - GLdispatch/ contains code for libGLdispatch, which is really just a thin wrapper around Mesa's glapi that tries to hide most of the complexity of managing dispatch tables. Its interface is defined in GLdispatch.h. This implements the guts of the core GL API libraries. - EGL/ and GLESv{1,2}/ are placeholders for now. GLESv{1,2}/ implement static GL entrypoints which use the context defined in libGLdispatch, while EGL/ will contain libEGL, which may be implemented similarly to libGLX. - GL/ and OpenGL/ respectively contain code to generate libGL and libOpenGL, which are both merely wrapper libraries for libGLX and 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. - util/ contains generic utility code, and arch/ contains architecture-specific defines. 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. At the same time, look at x11glvnd{client,server}.c to see how the "x11glvnd" extension which retrieves the appropriate mappings is implemented. The tests/ directory contains several unit tests which verify that dispatching to different vendors actually works. Run `make check` to run these unit tests. Note some of the unit tests require a special X server configuration and are skipped by default. To include these tests (and X server initialization/teardown), run `make check DO_X11_TESTS=1`. Architecture ------------ The library organization differs slightly from that of Andy's original proposal. See the diagram below:
┌──────────────────────────────────┐ │ │ ┌─────┤ Application │ │ │ │ │ └─────┬───────────────────┬────────┘ │ │ │ │ ┌─────▾─────┐ │ ┌──────────────┐ │ │ │ │ │ │ │ │ libOpenGL │ │ │ │ │ │ │ │ │ X server │ │ └─────┬─────┘ │ │ │ │ DT_FILTER │ │ │ │ ┌─────▾──────────┐ ┌──────▾────────┐ │ ┌──────────┐ │ │ │ │ │ │ └─│x11glvnd │─┘ │ │ [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. x11glvnd is a simple X 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 current dispatch tables will eventually be shared between GLX and EGL, similarly to how glapi operates when Mesa is compiled with the --shared-glapi option. libOpenGL is a wrapper library to libGLdispatch which exposes OpenGL 4.x core and compatibility entry points. Eventually, there will be a libGLESv{1,2} which will also be wrapper libraries to libGLdispatch that expose GLES entry points. libGL is a wrapper library on libGLdispatch and libGLX which is provided for backwards-compatibility with applications which link against the old ABI. NOTE: Logically, libGL should be a wrapper library to libOpenGL rather than libGLdispatch, as libGLdispatch is an implementation detail of libglvnd. However, we have this current arrangement for performance reasons since ELF symbol filtering is disabled by default (see Issues). ### 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 correspending `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. But XIDs could be created by another process. See the Issues section below. * To minimize code complexity from error checking, define a noop GLX dispatch table. This is returned by `__glXGet{,Current}Dispatch()` in case no other dispatch table can be found. * Similarly, `__glXScreenFrom{Context,FBConfig,Drawable}()` may fail to find a screen matching the specified GLX object. In this case, the returned screen number is -1, but the caller should just pass the screen number through to `__glXGetDispatch()` or `__glX{Add,Remove}Screen{Context,FBConfig,Drawable}Mapping()`. Those functions are expected to deal gracefully with the invalid screen number. 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 is an internal detail of libGLX. The ABI provided to the vendor library exposes a mapping from drawables to (screen, vendor) pairs. The interface does not make any assumptions about how screens and vendors correspond to each other. * 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? * Should we map XIDs directly to vendors, rather than to screens? * The current libGLX implementation stores the mapping between screen and all objects of the same type in one hash table. I.e., all pointer types (GLXContext and GLXFBConfig) in one table, and all XID types (GLXDrawable, GLXPixmap, GLXWindow, and GLXPbuffer) in another table. Should there instead be more finer-grained hash tables? There probably couldn't be finer-grained tables for XIDs, because GLXDrawable is used interchangably with the other XID-based types. * The issue above applies to XIDs in the x11glvnd extension as well: we currently don't make any distinction between window, GLX pixmap, GLX window, or GLX pbuffer XIDs. * 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? * Currently the library does not attempt to clean up allocations and unload vendor libraries if the application unloads it. This will need to be implemented eventually for the library to be usable in a production environment. What will the sequencing of this look like? Should we also hook into XCloseDisplay()? * Should x11glvnd be an extension on top of GLX 1.4, or a "GLX Next" feature? 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