【图形图像】Android SurfaceFlinger之OpenGL库加载过程

1、egl_init_drivers

Android中OpenGL库加载从egl_init_drivers函数开始，源码位置在frameworks/native/opengl/libs/EGL/egl.cpp。

static pthread_mutex_t sInitDriverMutex = PTHREAD_MUTEX_INITIALIZER;

EGLBoolean egl_init_drivers() {
    EGLBoolean res;
    pthread_mutex_lock(&sInitDriverMutex);
    res = egl_init_drivers_locked();
    pthread_mutex_unlock(&sInitDriverMutex);
    return res;
}

初始化过程中使用了mutex进行同步保护，下面分析egl_init_drivers_locked。

2、egl_init_drivers_locked

// this mutex protects:
//    d->disp[]
//    egl_init_drivers_locked()
//
static EGLBoolean egl_init_drivers_locked() {
    if (sEarlyInitState) {
        // initialized by static ctor. should be set here.
        return EGL_FALSE;
    }

    // get our driver loader
    Loader& loader(Loader::getInstance());

    // dynamically load our EGL implementation
    egl_connection_t* cnx = &gEGLImpl;
    if (cnx->dso == 0) {
        cnx->hooks[egl_connection_t::GLESv1_INDEX] =
                &gHooks[egl_connection_t::GLESv1_INDEX];
        cnx->hooks[egl_connection_t::GLESv2_INDEX] =
                &gHooks[egl_connection_t::GLESv2_INDEX];
        cnx->dso = loader.open(cnx);
    }

    return cnx->dso ? EGL_TRUE : EGL_FALSE;
}

从上面的egl_init_drivers_locked实现中可以看出OpenGL库最终是通过loader对象完成的，首先看一下sEarlyInitState是何方神圣。

3、sEarlyInitState

static void early_egl_init(void)
{
    int numHooks = sizeof(gHooksNoContext) / sizeof(EGLFuncPointer);
    EGLFuncPointer *iter = reinterpret_cast<EGLFuncPointer*>(&gHooksNoContext);
    for (int hook = 0; hook < numHooks; ++hook) {
        *(iter++) = reinterpret_cast<EGLFuncPointer>(gl_no_context);
    }

    setGLHooksThreadSpecific(&gHooksNoContext);
}

static pthread_once_t once_control = PTHREAD_ONCE_INIT;
static int sEarlyInitState = pthread_once(&once_control, &early_egl_init);

sEarlyInitState为pthread_once的返回值，成功时返回0。pthread_once保证多线程时early_egl_init只执行一次，early_egl_init用来设置线程特有的数据（pthread_setspecific/pthread_getspecific），这也是一种数据共享方式。EGLFuncPointer为函数指针，如下所示：

typedef void (*__eglMustCastToProperFunctionPointerType)(void);
typedef __eglMustCastToProperFunctionPointerType EGLFuncPointer;

gHooksNoContext是个有趣的数据结构，如下所示：

// maximum number of GL extensions that can be used simultaneously in
// a given process. this limitation exists because we need to have
// a static function for each extension and currently these static functions
// are generated at compile time.
#define MAX_NUMBER_OF_GL_EXTENSIONS 256

struct gl_hooks_t {
    struct gl_t {
        #include "entries.in"
    } gl;
    struct gl_ext_t {
        __eglMustCastToProperFunctionPointerType extensions[MAX_NUMBER_OF_GL_EXTENSIONS];
    } ext;
};

extern gl_hooks_t gHooksNoContext;

上面的gl_t结构体实际上包含了同OpenGL函数一致的函数指针，通过“entries.in”和GL_ENTRY进行扩展，如下所示：

#define GL_ENTRY(_r, _api, ...) _r (*(_api))(__VA_ARGS__);

// entries.in
GL_ENTRY(void, glActiveShaderProgram, GLuint pipeline, GLuint program)
GL_ENTRY(void, glActiveShaderProgramEXT, GLuint pipeline, GLuint program)
...
...

gl_no_context用于检查OpenGL函数调用是否设置了Context，其中用到了设置线程键的egl_tls_t类和收集函数调用栈帧的CallStack类。

static int gl_no_context() {
    if (egl_tls_t::logNoContextCall()) {
        char const* const error = "call to OpenGL ES API with "
                "no current context (logged once per thread)";
        if (LOG_NDEBUG) {
            ALOGE(error);
        } else {
            LOG_ALWAYS_FATAL(error);
        }
        char value[PROPERTY_VALUE_MAX];
        property_get("debug.egl.callstack", value, "0");
        if (atoi(value)) {
            CallStack stack(LOG_TAG);
        }
    }
    return 0;
}

随后，将修改过的gHooksNoContext通过setGLHooksThreadSpecific设置线程键，而__get_tls则是通过汇编实现的（不同的硬件有不同的汇编代码），如下所示：

void setGLHooksThreadSpecific(gl_hooks_t const *value) {
    setGlThreadSpecific(value);
}

void setGlThreadSpecific(gl_hooks_t const *value) {
    gl_hooks_t const * volatile * tls_hooks = get_tls_hooks();
    tls_hooks[TLS_SLOT_OPENGL_API] = value;
}

inline gl_hooks_t const * volatile * get_tls_hooks() {
    volatile void *tls_base = __get_tls();
    gl_hooks_t const * volatile * tls_hooks =
            reinterpret_cast<gl_hooks_t const * volatile *>(tls_base);
    return tls_hooks;
}

// Well-known TLS slots. What data goes in which slot is arbitrary unless otherwise noted.
enum {
  TLS_SLOT_SELF = 0, // The kernel requires this specific slot for x86.
  TLS_SLOT_THREAD_ID,
  TLS_SLOT_ERRNO,

  // These two aren't used by bionic itself, but allow the graphics code to
  // access TLS directly rather than using the pthread API.
  TLS_SLOT_OPENGL_API = 3,
  TLS_SLOT_OPENGL = 4,

  // This slot is only used to pass information from the dynamic linker to
  // libc.so when the C library is loaded in to memory. The C runtime init
  // function will then clear it. Since its use is extremely temporary,
  // we reuse an existing location that isn't needed during libc startup.
  TLS_SLOT_BIONIC_PREINIT = TLS_SLOT_OPENGL_API,

  TLS_SLOT_STACK_GUARD = 5, // GCC requires this specific slot for x86.
  TLS_SLOT_DLERROR,

  // Fast storage for Thread::Current() in ART.
  TLS_SLOT_ART_THREAD_SELF,

  // Lets TSAN avoid using pthread_getspecific for finding the current thread
  // state.
  TLS_SLOT_TSAN,

  BIONIC_TLS_SLOTS // Must come last!
};

4、egl_connection_t

在分析Loader前先来看一下egl_connection_t，其结构如下所示：

extern egl_connection_t gEGLImpl;
extern gl_hooks_t gHooks[2];

struct egl_connection_t {
    enum {
        GLESv1_INDEX = 0,
        GLESv2_INDEX = 1
    };

    inline egl_connection_t() : dso(0) { }
    void *              dso;
    gl_hooks_t *        hooks[2];
    EGLint              major;
    EGLint              minor;
    egl_t               egl;

    void*               libEgl;
    void*               libGles1;
    void*               libGles2;
};

从上面的egl_connection_t结构体及前面的egl_init_drivers_locked函数定义可以看出，包括两个hook，GLESv1和GLESv2，而dso则实际上指向了libEGL、libGLESv1和libGLESv2（从后面的Loader分析中可以看出），egl_t类似于gl_t，包含了一系列egl函数指针。下面着重分析Loader。

5、Loader-open

Loader是个单实例，如下所示：

class Loader : public Singleton<Loader>

从Loader::open开始，包括两部分，load_driver和load_wrapper，如下所示：

void* Loader::open(egl_connection_t* cnx)
{
    void* dso;
    driver_t* hnd = 0;

    setEmulatorGlesValue();

    dso = load_driver("GLES", cnx, EGL | GLESv1_CM | GLESv2);
    if (dso) {
        hnd = new driver_t(dso);
    } else {
        // Always load EGL first
        dso = load_driver("EGL", cnx, EGL);
        if (dso) {
            hnd = new driver_t(dso);
            hnd->set( load_driver("GLESv1_CM", cnx, GLESv1_CM), GLESv1_CM );
            hnd->set( load_driver("GLESv2",    cnx, GLESv2),    GLESv2 );
        }
    }

    LOG_ALWAYS_FATAL_IF(!hnd, "couldn't find an OpenGL ES implementation");

    cnx->libEgl   = load_wrapper(EGL_WRAPPER_DIR "/libEGL.so");
    cnx->libGles2 = load_wrapper(EGL_WRAPPER_DIR "/libGLESv2.so");
    cnx->libGles1 = load_wrapper(EGL_WRAPPER_DIR "/libGLESv1_CM.so");

    LOG_ALWAYS_FATAL_IF(!cnx->libEgl,
            "couldn't load system EGL wrapper libraries");

    LOG_ALWAYS_FATAL_IF(!cnx->libGles2 || !cnx->libGles1,
            "couldn't load system OpenGL ES wrapper libraries");

    return (void*)hnd;
}

在分析load_driver和load_wrapper之前，先来看一下driver_t，这是个Loader类的内部结构体，如下所示：

struct driver_t {
        explicit driver_t(void* gles);
        ~driver_t();
        status_t set(void* hnd, int32_t api);
        void* dso[3];
    };

driver_t的dso分别保存了EGL、GLESv1_CM和GLESv2三个库。open函数中还调用了setEmulatorGlesValue，这个函数就是检查、设置一些模拟器属性（是否在模拟器中运行？在模拟中运行时是否有GPU支持？），比较简单，代码如下：

static void setEmulatorGlesValue(void) {
    char prop[PROPERTY_VALUE_MAX];
    property_get("ro.kernel.qemu", prop, "0");
    if (atoi(prop) != 1) return;

    property_get("ro.kernel.qemu.gles",prop,"0");
    if (atoi(prop) == 1) {
        ALOGD("Emulator has host GPU support, qemu.gles is set to 1.");
        property_set("qemu.gles", "1");
        return;
    }

    // for now, checking the following
    // directory is good enough for emulator system images
    const char* vendor_lib_path =
#if defined(__LP64__)
        "/vendor/lib64/egl";
#else
        "/vendor/lib/egl";
#endif

    const bool has_vendor_lib = (access(vendor_lib_path, R_OK) == 0);
    if (has_vendor_lib) {
        ALOGD("Emulator has vendor provided software renderer, qemu.gles is set to 2.");
        property_set("qemu.gles", "2");
    } else {
        ALOGD("Emulator without GPU support detected. "
              "Fallback to legacy software renderer, qemu.gles is set to 0.");
        property_set("qemu.gles", "0");
    }
}

6、Loader-load_driver

void *Loader::load_driver(const char* kind, egl_connection_t* cnx, uint32_t mask)
{
class MatchFile {};
...
}

EGL驱动需要提供一个单一的库libGLES.so，或者是三个分开的库libEGL.so、libGLESv1_CM.so和libGLESv2.so，模拟器中的软描画则需要提供单一的库libGLES_android.so，为了兼容旧版本的“egl.cfg”文件配置，这几个库后面还可能加一个名称如_emulation。在load_driver函数中还有个内部类MatchFile，用于在某个lib目录中查找特定的OpenGL库是否存在，存在的话进而通过dlopen打开，dlopen成功后再通过dlsym逐个获取OpenGL所有API的地址并将它们保存下来。

7、load_wrapper

load_wrapper相对load_driver来说较简单，就是dlopen一个so，如下所示：

static void* load_wrapper(const char* path) {
    void* so = dlopen(path, RTLD_NOW | RTLD_LOCAL);
    ALOGE_IF(!so, "dlopen(\"%s\") failed: %s", path, dlerror());
    return so;
}