Android minigbm框架普法

   Android minigbm框架普法



引言

假设存在这么一个场景，我的GPU的上层实现走的不是标准的Mesa接口，且GPU也没有提专门配套的gralloc和hwcompoer实现。那么我们的Android要怎么使用到EGL和GLES库呢，并且此GPU驱动是支持drm实现的，也有video overlay层。这个就是我们这个博客准备或者是探索要解决的。这里无意中发现mingbm + hwcompoer的组合可能解决，这里我就花费一定的时间来分析这个模块！



一.minigbm结构


1.1 minigbm目录结构

minigbm├── amdgpu.c├── Android.bp├── Android.gralloc.mk├── common.mk├── cros_gralloc├── dri.c├── dri.h├── drv.c├── drv.h├── drv_priv.h├── evdi.c├── exynos.c├── gbm.c├── gbm.h├── gbm_helpers.c├── gbm_helpers.h├── gbm.pc├── gbm_priv.h├── helpers_array.c├── helpers_array.h├── helpers.c├── helpers.h├── i915.c├── LICENSE├── Makefile├── marvell.c├── mediatek.c├── meson.c├── METADATA├── MODULE_LICENSE_BSD├── msm.c├── nouveau.c├── OWNERS├── OWNERS.android├── PRESUBMIT.cfg├── presubmit.sh├── radeon.c├── rockchip.c├── synaptics.c├── tegra.c├── udl.c├── util.h├── vc4.c├── vgem.c├── virgl_hw.h├── virtgpu_drm.h└── virtio_gpu.c1 directory, 46 files

1.2 minigbm框架结构

Minigbm的整个组织结构，可以用下面的框图来表示！

在这里插入图片描述


在这里插入图片描述

Generic Android Buffer Allocator - Linaro



二. minigbm源码分析

minig的核心主要是提供gralloc的HAL实现，但是我看它也提供了私有的HIDL service实现，关于HIDL的实现这个暂时不在我们考虑的范围之内。我们重点分析minigbm作为gralloc hal的实现的代码！

注意我这里分析的minigbm的版本如下:

commit aa65a6a67cd41390d4e8a0f87e20d3191db2994d (HEAD, tag: android-11.0.0_r46, tag: android-11.0.0_r43, tag: android-11.0.0_r40, tag: android-11.0.0_r39, tag: android-11.0.0_r38, tag: android-11.0.0_r37, tag: android-11.0.0_r36, tag: android-11.0.0_r35, tag: android-11.0.0_r34, tag: android-11.0.0_r33, tag: android-11.0.0_r32, tag: android-11.0.0_r30, tag: m/android-11.0.0_r37, aosp/android11-qpr3-s1-release, aosp/android11-qpr3-release, aosp/android11-qpr2-release, aosp/android11-qpr1-c-release)
Author: Jason Macnak <natsu@google.com>
Date:   Sat Jul 25 06:02:22 2020 -0700cros_gralloc: map custom drm fourcc back to standard fourccMinigbm uses a custom fourcc DRM_FORMAT_YVU420_ANDROID tospecify a DRM_FORMAT_YVU420 format with the extra Androidspecific alignment requirement. Mapper should map thiscustom code back into the standard one for metadata get().Bug: b/146515640Test: launch_cvd, open youtube video, observe no hwc warningChange-Id: Id9ac2bb233837b6c7aa093eddbea81da0bdf3c1b

2.1 minigbm注册gralloc hal

肯定mingbm作为gralloc的hal的实现，我们肯定得重点分析了。地球人都知道HAL的实现的入口HAL_MODULE_INFO_SYM，我们来看看minigbm是怎么实现的:

//cros_gralloc/gralloc0/gralloc0.cc
struct gralloc0_module {gralloc_module_t base;std::unique_ptr<alloc_device_t> alloc;std::unique_ptr<cros_gralloc_driver> driver;bool initialized;std::mutex initialization_mutex;
};// clang-format off
static struct hw_module_methods_t gralloc0_module_methods = { .open = gralloc0_open };
// clang-format onstruct gralloc0_module HAL_MODULE_INFO_SYM = {.base ={.common ={.tag = HARDWARE_MODULE_TAG,.module_api_version = GRALLOC_MODULE_API_VERSION_0_3,.hal_api_version = 0,.id = GRALLOC_HARDWARE_MODULE_ID,.name = "CrOS Gralloc",.author = "Chrome OS",.methods = &gralloc0_module_methods,},/*** @brief * 当其他进程分配的GraphicBuffer传递到当前进程后，需要通过该方法映射到当前进程，* 为后续的lock做好准备*/.registerBuffer = gralloc0_register_buffer,/*** @brief * 取消GraphicBuffer在当前进程的映射，后续不能调用lock了*/.unregisterBuffer = gralloc0_unregister_buffer,/*** @brief * 调用lock后，才能访问图形buffer，假如usage指定了GRALLOC_USAGE_SW_* flag，* vaddr将被填充成图形Buffer在虚拟内存中的地址*/.lock = gralloc0_lock,.unlock = gralloc0_unlock,.perform = gralloc0_perform,.lock_ycbcr = gralloc0_lock_ycbcr,.lockAsync = gralloc0_lock_async,.unlockAsync = gralloc0_unlock_async,.lockAsync_ycbcr = gralloc0_lock_async_ycbcr,.validateBufferSize = NULL,.getTransportSize = NULL,},/*** @brief * 是不是这个地方感觉有点奇怪，咋alloc的实现为null* 好戏在后头*/.alloc = nullptr,.driver = nullptr,.initialized = false,
};

2.2 minigbm下HAL hw_module open函数实现

没有啥文字可以描述的，我们直接上源码！

static int gralloc0_open(const struct hw_module_t *mod, const char *name, struct hw_device_t **dev)
{auto const_module = reinterpret_cast<const struct gralloc0_module *>(mod);auto module = const_cast<struct gralloc0_module *>(const_module);if (module->initialized) {*dev = &module->alloc->common;return 0;}/*** @brief Construct a new if object* libhardware/include/hardware/gralloc.h* #define GRALLOC_HARDWARE_GPU0 "gpu0"*/if (strcmp(name, GRALLOC_HARDWARE_GPU0)) {drv_log("Incorrect device name - %s.\n", name);return -EINVAL;}if (gralloc0_init(module, true))return -ENODEV;*dev = &module->alloc->common;return 0;
}

我们接着看下核心的函数gralloc0_init的实现，不能停，一口气炫下去:

//cros_gralloc/cros_gralloc_driver.h
class cros_gralloc_driver
{public:cros_gralloc_driver();~cros_gralloc_driver();int32_t init();bool is_supported(const struct cros_gralloc_buffer_descriptor *descriptor);int32_t allocate(const struct cros_gralloc_buffer_descriptor *descriptor,buffer_handle_t *out_handle);int32_t retain(buffer_handle_t handle);int32_t release(buffer_handle_t handle);int32_t lock(buffer_handle_t handle, int32_t acquire_fence, bool close_acquire_fence,const struct rectangle *rect, uint32_t map_flags,uint8_t *addr[DRV_MAX_PLANES]);int32_t unlock(buffer_handle_t handle, int32_t *release_fence);int32_t invalidate(buffer_handle_t handle);int32_t flush(buffer_handle_t handle, int32_t *release_fence);int32_t get_backing_store(buffer_handle_t handle, uint64_t *out_store);int32_t resource_info(buffer_handle_t handle, uint32_t strides[DRV_MAX_PLANES],uint32_t offsets[DRV_MAX_PLANES]);int32_t get_reserved_region(buffer_handle_t handle, void **reserved_region_addr,uint64_t *reserved_region_size);uint32_t get_resolved_drm_format(uint32_t drm_format, uint64_t usage);void for_each_handle(const std::function<void(cros_gralloc_handle_t)> &function);private:cros_gralloc_driver(cros_gralloc_driver const &);cros_gralloc_driver operator=(cros_gralloc_driver const &);cros_gralloc_buffer *get_buffer(cros_gralloc_handle_t hnd);struct driver *drv_;std::mutex mutex_;std::unordered_map<uint32_t, cros_gralloc_buffer *> buffers_;std::unordered_map<cros_gralloc_handle_t, std::pair<cros_gralloc_buffer *, int32_t>>handles_;
};//cros_gralloc/gralloc0/gralloc0.cc
static int gralloc0_init(struct gralloc0_module *mod, bool initialize_alloc)
{std::lock_guard<std::mutex> lock(mod->initialization_mutex);if (mod->initialized)return 0;mod->driver = std::make_unique<cros_gralloc_driver>();/*** @brief * 根据vendor，初始化不同驱动模型，实现在cros_gralloc_driver.cc里面* 如果要扩展的话，就得实现phy自己的逻辑*/if (mod->driver->init()) {drv_log("Failed to initialize driver.\n");return -ENODEV;}if (initialize_alloc) {mod->alloc = std::make_unique<alloc_device_t>();/*** @brief * alloc在此处被赋值**/mod->alloc->alloc = gralloc0_alloc;mod->alloc->free = gralloc0_free;mod->alloc->common.tag = HARDWARE_DEVICE_TAG;mod->alloc->common.version = 0;mod->alloc->common.module = (hw_module_t *)mod;mod->alloc->common.close = gralloc0_close;}mod->initialized = true;return 0;
}

2.3 cros_gralloc_driver的init实现

我们接着看下cros_gralloc_driver的init是如何被初始的:

//cros_gralloc/cros_gralloc_driver.cc
int32_t cros_gralloc_driver::init()
{/** Create a driver from rendernode while filtering out* the specified undesired driver.** TODO(gsingh): Enable render nodes on udl/evdi.*/int fd;drmVersionPtr version;char const *str = "%s/renderD%d";const char *undesired[2] = { "vgem", nullptr };uint32_t num_nodes = 63;uint32_t min_node = 128;uint32_t max_node = (min_node + num_nodes);for (uint32_t i = 0; i < ARRAY_SIZE(undesired); i++) {for (uint32_t j = min_node; j < max_node; j++) {char *node;/*** @brief * 打开/dev/dri/renderD0-renderD128*/if (asprintf(&node, str, DRM_DIR_NAME, j) < 0)continue;fd = open(node, O_RDWR, 0);free(node);if (fd < 0)continue;version = drmGetVersion(fd);if (!version) {close(fd);continue;}if (undesired[i] && !strcmp(version->name, undesired[i])) {close(fd);drmFreeVersion(version);continue;}drmFreeVersion(version);/*** @brief * 加载不同的driver*/drv_ = drv_create(fd);if (drv_)return 0;close(fd);}}return -ENODEV;
}

我们接着玩下看drv_create的实现，这个是根绝实际驱动情况，决定backend后端:

//drv.c
struct driver *drv_create(int fd)
{struct driver *drv;int ret;drv = (struct driver *)calloc(1, sizeof(*drv));if (!drv)return NULL;drv->fd = fd;/*** @brief * 获取不同的后端，譬如i915，amdgpu，msm等* 或者阔以构建一个通用的backend，譬如xxx后端，但是是根据drm的驱动名称进行匹配的，* 所以那怕是通用的，也需要进行相关的扩展*/drv->backend = drv_get_backend(fd);if (!drv->backend)goto free_driver;if (pthread_mutex_init(&drv->driver_lock, NULL))goto free_driver;//drmHashCreate这个API不知道是用来干啥的drv->buffer_table = drmHashCreate();if (!drv->buffer_table)goto free_lock;drv->mappings = drv_array_init(sizeof(struct mapping));if (!drv->mappings)goto free_buffer_table;drv->combos = drv_array_init(sizeof(struct combination));if (!drv->combos)goto free_mappings;if (drv->backend->init) {ret = drv->backend->init(drv);if (ret) {drv_array_destroy(drv->combos);goto free_mappings;}}return drv;free_mappings:drv_array_destroy(drv->mappings);
free_buffer_table:drmHashDestroy(drv->buffer_table);
free_lock:pthread_mutex_destroy(&drv->driver_lock);
free_driver:free(drv);return NULL;
}

2.4 minigbm的backend以及backend的初始化

这里的backend，我把它理解它为minigbm对各种gpu抽象出来的一种能力的统称。因为前端就是gralloc hal的接口定义，都是一致的，然后后端backend抽象出来，然后根据实际的情况，调用具体的gpu用户层接口实现！我们来看backend是如何抽象和初始化的。

#ifdef DRV_AMDGPU
extern const struct backend backend_amdgpu;
#endif
/*** @brief * 可扩展虚拟显示接口（EVDI）*/
extern const struct backend backend_evdi;
#ifdef DRV_EXYNOS
extern const struct backend backend_exynos;
#endif
#ifdef DRV_I915
extern const struct backend backend_i915;
#endif
#ifdef DRV_MARVELL
extern const struct backend backend_marvell;
#endif
#ifdef DRV_MEDIATEK
extern const struct backend backend_mediatek;
#endif
#ifdef DRV_MESON
extern const struct backend backend_meson;
#endif
#ifdef DRV_MSM
extern const struct backend backend_msm;
#endif
extern const struct backend backend_nouveau;
#ifdef DRV_RADEON
extern const struct backend backend_radeon;
#endif
#ifdef DRV_ROCKCHIP
extern const struct backend backend_rockchip;
#endif
#ifdef DRV_SYNAPTICS
extern const struct backend backend_synaptics;
#endif
#ifdef DRV_TEGRA
extern const struct backend backend_tegra;
#endif
extern const struct backend backend_udl;
#ifdef DRV_VC4
extern const struct backend backend_vc4;
#endif
extern const struct backend backend_vgem;
extern const struct backend backend_virtio_gpu;static const struct backend *drv_get_backend(int fd)
{drmVersionPtr drm_version;unsigned int i;drm_version = drmGetVersion(fd);if (!drm_version)return NULL;/*** @brief * 后面如果要扩展，则是在backend_list中* 添加backend_phy*/const struct backend *backend_list[] = {
#ifdef DRV_AMDGPU&backend_amdgpu,
#endif&backend_evdi,
#ifdef DRV_EXYNOS&backend_exynos,
#endif
#ifdef DRV_I915&backend_i915,
#endif
#ifdef DRV_MARVELL&backend_marvell,
#endif
#ifdef DRV_MEDIATEK&backend_mediatek,
#endif
#ifdef DRV_MESON&backend_meson,
#endif
#ifdef DRV_MSM&backend_msm,
#endif&backend_nouveau,
#ifdef DRV_RADEON&backend_radeon,
#endif
#ifdef DRV_ROCKCHIP&backend_rockchip,
#endif
#ifdef DRV_SYNAPTICS&backend_synaptics,
#endif
#ifdef DRV_TEGRA&backend_tegra,
#endif&backend_udl,
#ifdef DRV_VC4&backend_vc4,
#endif&backend_vgem,	    &backend_virtio_gpu,};for (i = 0; i < ARRAY_SIZE(backend_list); i++) {const struct backend *b = backend_list[i];// Exactly one of the main create functions must be defined.assert((b->bo_create != NULL) ^ (b->bo_create_from_metadata != NULL));// Either both or neither must be implemented.assert((b->bo_compute_metadata != NULL) == (b->bo_create_from_metadata != NULL));// Both can't be defined, but it's okay for neither to be (i.e. only bo_create).assert((b->bo_create_with_modifiers == NULL) ||(b->bo_create_from_metadata == NULL));if (!strcmp(drm_version->name, b->name)) {drmFreeVersion(drm_version);return b;}}drmFreeVersion(drm_version);return NULL;
}

这里我们可以看到这里抽象出来了几种典型的GPU的backend，然后再根据前面获取到的drm的驱动名称进行匹配，匹配到了则返回合适的backend，没有则返回NULL。这里我们以evdi为例子来进行相关的分析。我么来看evdi backend的init。

//helpers.c
int drv_modify_linear_combinations(struct driver *drv)
{/** All current drivers can scanout linear XRGB8888/ARGB8888 as a primary* plane and as a cursor.*/drv_modify_combination(drv, DRM_FORMAT_XRGB8888, &LINEAR_METADATA,BO_USE_CURSOR | BO_USE_SCANOUT);drv_modify_combination(drv, DRM_FORMAT_ARGB8888, &LINEAR_METADATA,BO_USE_CURSOR | BO_USE_SCANOUT);return 0;
}static const uint32_t render_target_formats[] = { DRM_FORMAT_ARGB8888, DRM_FORMAT_XRGB8888 };static int evdi_init(struct driver *drv)
{/*** @brief Construct a new drv add combinations object* add default  render_target_formats*/drv_add_combinations(drv, render_target_formats, ARRAY_SIZE(render_target_formats),&LINEAR_METADATA, BO_USE_RENDER_MASK);return drv_modify_linear_combinations(drv);
}const struct backend backend_evdi = {.name = "evdi",.init = evdi_init,.bo_create = drv_dumb_bo_create,.bo_destroy = drv_dumb_bo_destroy,.bo_import = drv_prime_bo_import,.bo_map = drv_dumb_bo_map,.bo_unmap = drv_bo_munmap,
};

到这里gralloc hal也注册OK了，cross_gralloc_driver和backend也构建初始化OK了，后面就是上层通过grallc api来进行相关调用了。我们接着继续往下看。


2.5 minigbm的graloc hal下alloc的实现

Android通过alloc开构建GraphicBuffer，我们来看看minigbm是如何实现的。

static int gralloc0_alloc(alloc_device_t *dev, int w, int h, int format, int usage,buffer_handle_t *handle, int *stride)
{int32_t ret;bool supported;struct cros_gralloc_buffer_descriptor descriptor;auto mod = (struct gralloc0_module const *)dev->common.module;descriptor.width = w;descriptor.height = h;descriptor.droid_format = format;descriptor.droid_usage = usage;descriptor.drm_format = cros_gralloc_convert_format(format);descriptor.use_flags = gralloc0_convert_usage(usage);descriptor.reserved_region_size = 0;/*** @brief * judge driver is support descriptor or not*/supported = mod->driver->is_supported(&descriptor);if (!supported && (usage & GRALLOC_USAGE_HW_COMPOSER)) {descriptor.use_flags &= ~BO_USE_SCANOUT;supported = mod->driver->is_supported(&descriptor);}if (!supported && (usage & GRALLOC_USAGE_HW_VIDEO_ENCODER) &&!gralloc0_droid_yuv_format(format)) {// Unmask BO_USE_HW_VIDEO_ENCODER in the case of non-yuv formats// because they are not input to a hw encoder but used as an// intermediate format (e.g. camera).descriptor.use_flags &= ~BO_USE_HW_VIDEO_ENCODER;supported = mod->driver->is_supported(&descriptor);}if (!supported) {drv_log("Unsupported combination -- HAL format: %u, HAL usage: %u, ""drv_format: %4.4s, use_flags: %llu\n",format, usage, reinterpret_cast<char *>(&descriptor.drm_format),static_cast<unsigned long long>(descriptor.use_flags));return -EINVAL;}ret = mod->driver->allocate(&descriptor, handle);if (ret)return ret;auto hnd = cros_gralloc_convert_handle(*handle);*stride = hnd->pixel_stride;return 0;
}

这块我们先从整体概括下该部分的核心逻辑:

  • 判断是否支持申请buffer的format
  • 如支持则通过driver的allocate申请buffer
  • 然后将申请的buffer返回

我们接着往下看driver是如何allocate的：

//cros_gralloc/cros_gralloc_driver.cc
int32_t cros_gralloc_driver::allocate(const struct cros_gralloc_buffer_descriptor *descriptor,buffer_handle_t *out_handle)
{uint32_t id;size_t num_planes;size_t num_fds;size_t num_ints;size_t num_bytes;uint32_t resolved_format;uint32_t bytes_per_pixel;uint64_t use_flags;int32_t reserved_region_fd;char *name;struct bo *bo;struct cros_gralloc_handle *hnd;resolved_format = drv_resolve_format(drv_, descriptor->drm_format, descriptor->use_flags);use_flags = descriptor->use_flags;/** TODO(b/79682290): ARC++ assumes NV12 is always linear and doesn't* send modifiers across Wayland protocol, so we or in the* BO_USE_LINEAR flag here. We need to fix ARC++ to allocate and work* with tiled buffers.*/if (resolved_format == DRM_FORMAT_NV12)use_flags |= BO_USE_LINEAR;/** This unmask is a backup in the case DRM_FORMAT_FLEX_IMPLEMENTATION_DEFINED is resolved* to non-YUV formats.*/if (descriptor->drm_format == DRM_FORMAT_FLEX_IMPLEMENTATION_DEFINED &&(resolved_format == DRM_FORMAT_XBGR8888 || resolved_format == DRM_FORMAT_ABGR8888)) {use_flags &= ~BO_USE_HW_VIDEO_ENCODER;}/*** @brief * construct struct bo*/bo = drv_bo_create(drv_, descriptor->width, descriptor->height, resolved_format, use_flags);if (!bo) {drv_log("Failed to create bo.\n");return -ENOMEM;}/** If there is a desire for more than one kernel buffer, this can be* removed once the ArcCodec and Wayland service have the ability to* send more than one fd. GL/Vulkan drivers may also have to modified.*/if (drv_num_buffers_per_bo(bo) != 1) {drv_bo_destroy(bo);drv_log("Can only support one buffer per bo.\n");return -EINVAL;}num_planes = drv_bo_get_num_planes(bo);num_fds = num_planes;if (descriptor->reserved_region_size > 0) {reserved_region_fd =create_reserved_region(descriptor->name, descriptor->reserved_region_size);if (reserved_region_fd < 0) {drv_bo_destroy(bo);return reserved_region_fd;}num_fds += 1;} else {reserved_region_fd = -1;}num_bytes = sizeof(struct cros_gralloc_handle);num_bytes += (descriptor->name.size() + 1);/** Ensure that the total number of bytes is a multiple of sizeof(int) as* native_handle_clone() copies data based on hnd->base.numInts.*/num_bytes = ALIGN(num_bytes, sizeof(int));num_ints = num_bytes - sizeof(native_handle_t) - num_fds;/** malloc is used as handles are ultimetly destroyed via free in* native_handle_delete().*/hnd = static_cast<struct cros_gralloc_handle *>(malloc(num_bytes));hnd->base.version = sizeof(hnd->base);hnd->base.numFds = num_fds;hnd->base.numInts = num_ints;hnd->num_planes = num_planes;for (size_t plane = 0; plane < num_planes; plane++) {/*** @brief * use drmPrimeHandleToFD get fd, so can binder  transation* 将handle转换fd，方便后续通过binder跨进程传输fd*/hnd->fds[plane] = drv_bo_get_plane_fd(bo, plane);hnd->strides[plane] = drv_bo_get_plane_stride(bo, plane);hnd->offsets[plane] = drv_bo_get_plane_offset(bo, plane);hnd->sizes[plane] = drv_bo_get_plane_size(bo, plane);}hnd->fds[hnd->num_planes] = reserved_region_fd;hnd->reserved_region_size = descriptor->reserved_region_size;static std::atomic<uint32_t> next_buffer_id{ 1 };hnd->id = next_buffer_id++;hnd->width = drv_bo_get_width(bo);hnd->height = drv_bo_get_height(bo);hnd->format = drv_bo_get_format(bo);hnd->format_modifier = drv_bo_get_plane_format_modifier(bo, 0);hnd->use_flags = descriptor->use_flags;bytes_per_pixel = drv_bytes_per_pixel_from_format(hnd->format, 0);hnd->pixel_stride = DIV_ROUND_UP(hnd->strides[0], bytes_per_pixel);hnd->magic = cros_gralloc_magic;hnd->droid_format = descriptor->droid_format;hnd->usage = descriptor->droid_usage;hnd->total_size = descriptor->reserved_region_size + bo->meta.total_size;hnd->name_offset = handle_data_size;name = (char *)(&hnd->base.data[hnd->name_offset]);snprintf(name, descriptor->name.size() + 1, "%s", descriptor->name.c_str());id = drv_bo_get_plane_handle(bo, 0).u32;/*** @brief * new class cros_gralloc_buffer*/auto buffer = new cros_gralloc_buffer(id, bo, hnd, hnd->fds[hnd->num_planes],hnd->reserved_region_size);std::lock_guard<std::mutex> lock(mutex_);buffers_.emplace(id, buffer);handles_.emplace(hnd, std::make_pair(buffer, 1));/*** @brief * return buffer_handle_t*/*out_handle = reinterpret_cast<buffer_handle_t>(hnd);return 0;
}

代码量比较多，但是我们是有原则的，不能啥都啃，不然咋下嘴不是。上面的代码核心逻辑主要为:

  • 构建struct bo对象
  • 然后通过博对象，填充cros_gralloc_handle
  • 然后继续将前面构建出来的相关对象填充cros_gralloc_buffer
  • 然后将buffer和handle以一定的结构保存

我们对前面的分开展开，继续分析，先看drv_bo_create的实现:

//drv.c
struct bo *drv_bo_create(struct driver *drv, uint32_t width, uint32_t height, uint32_t format,uint64_t use_flags)
{int ret;size_t plane;struct bo *bo;bool is_test_alloc;is_test_alloc = use_flags & BO_USE_TEST_ALLOC;use_flags &= ~BO_USE_TEST_ALLOC;/*** @brief * new struct bo*/bo = drv_bo_new(drv, width, height, format, use_flags, is_test_alloc);if (!bo)return NULL;ret = -EINVAL;if (drv->backend->bo_compute_metadata) {ret = drv->backend->bo_compute_metadata(bo, width, height, format, use_flags, NULL,0);if (!is_test_alloc && ret == 0)ret = drv->backend->bo_create_from_metadata(bo);} else if (!is_test_alloc) {/*** @brief * use backend bo_create*/ret = drv->backend->bo_create(bo, width, height, format, use_flags);}if (ret) {free(bo);return NULL;}pthread_mutex_lock(&drv->driver_lock);for (plane = 0; plane < bo->meta.num_planes; plane++) {if (plane > 0)assert(bo->meta.offsets[plane] >= bo->meta.offsets[plane - 1]);drv_increment_reference_count(drv, bo, plane);}pthread_mutex_unlock(&drv->driver_lock);return bo;
}

上面代码的核心逻辑主要是通过backend的bo_create函数继续填充struct bo，战斗还没有完，我们接着继续往下看:

bo_createdrv_dumb_bo_create（）drv_dumb_bo_create_ex()// helpers.c
int drv_dumb_bo_create_ex(struct bo *bo, uint32_t width, uint32_t height, uint32_t format,uint64_t use_flags, uint64_t quirks)
{int ret;size_t plane;uint32_t aligned_width, aligned_height;struct drm_mode_create_dumb create_dumb;aligned_width = width;aligned_height = height;switch (format) {case DRM_FORMAT_R16:/* HAL_PIXEL_FORMAT_Y16 requires that the buffer's width be 16 pixel* aligned. See hardware/interfaces/graphics/common/1.0/types.hal. */aligned_width = ALIGN(width, 16);break;case DRM_FORMAT_YVU420_ANDROID:/* HAL_PIXEL_FORMAT_YV12 requires that the buffer's height not* be aligned. Update 'height' so that drv_bo_from_format below* uses the non-aligned height. */height = bo->meta.height;/* Align width to 32 pixels, so chroma strides are 16 bytes as* Android requires. */aligned_width = ALIGN(width, 32);/* Adjust the height to include room for chroma planes. */aligned_height = 3 * DIV_ROUND_UP(height, 2);break;case DRM_FORMAT_YVU420:case DRM_FORMAT_NV12:case DRM_FORMAT_NV21:/* Adjust the height to include room for chroma planes */aligned_height = 3 * DIV_ROUND_UP(height, 2);break;default:break;}memset(&create_dumb, 0, sizeof(create_dumb));if (quirks & BO_QUIRK_DUMB32BPP) {aligned_width =DIV_ROUND_UP(aligned_width * layout_from_format(format)->bytes_per_pixel[0], 4);create_dumb.bpp = 32;} else {create_dumb.bpp = layout_from_format(format)->bytes_per_pixel[0] * 8;}create_dumb.width = aligned_width;create_dumb.height = aligned_height;create_dumb.flags = 0;ret = drmIoctl(bo->drv->fd, DRM_IOCTL_MODE_CREATE_DUMB, &create_dumb);if (ret) {drv_log("DRM_IOCTL_MODE_CREATE_DUMB failed (%d, %d)\n", bo->drv->fd, errno);return -errno;}drv_bo_from_format(bo, create_dumb.pitch, height, format);/*** @brief * store obj_gem handle*/for (plane = 0; plane < bo->meta.num_planes; plane++)bo->handles[plane].u32 = create_dumb.handle;bo->meta.total_size = create_dumb.size;return 0;
}

对于libdrm比较熟悉的小伙伴来说，上面就比较简单了通过ioctl的DRM_IOCTL_MODE_CREATE_DUMB指令，申请构建GEM buffer。我们知道gem_obj的handle是不同跨进程传输的，必须转换成fd才行，这个是在那个地方执行的呢？
在这里插入图片描述

//drv.c
int drv_bo_get_plane_fd(struct bo *bo, size_t plane)
{int ret, fd;assert(plane < bo->meta.num_planes);if (bo->is_test_buffer) {return -EINVAL;}/*** @brief * handle是每个gem context拥有的，不能跨进程传输。要跨进程传输可以将handle和fd联系起来。* fd可以跨进程传输，android大部分buffer都是通过binder传递fd，所以这里调用drmPrimeHandleToFD* 将gem_obj的句柄handle转换成fd以便于传输*/ret = drmPrimeHandleToFD(bo->drv->fd, bo->handles[plane].u32, DRM_CLOEXEC | DRM_RDWR, &fd);// Older DRM implementations blocked DRM_RDWR, but gave a read/write mapping anywaysif (ret)ret = drmPrimeHandleToFD(bo->drv->fd, bo->handles[plane].u32, DRM_CLOEXEC, &fd);return (ret) ? ret : fd;
}

到这里基allocate的流程就基本就结尾了，其核心逻辑用伪代码表示就是

drmIoctl(bo->drv->fd, DRM_IOCTL_MODE_CREATE_DUMB, &create_dumb)
drmPrimeHandleToFD

2.6 minigbm的graloc hal下import的实现

我们知道当进程通过sf申请到GraphicBuffer之后，并不能立马使用必须通过import映射到申请的进程才行，这里我们看看底下import是如何实现的。我们通过前面gralloc的hal注册可知，registerBuffer指向了gralloc0_register_buffer，所以我们只需要看它的实现逻辑即可:

//cros_gralloc/gralloc0/gralloc0.cc
gralloc0_register_buffer()mod->driver->retain(handle)int32_t cros_gralloc_driver::retain(buffer_handle_t handle)
{uint32_t id;std::lock_guard<std::mutex> lock(mutex_);auto hnd = cros_gralloc_convert_handle(handle);if (!hnd) {drv_log("Invalid handle.\n");return -EINVAL;}auto buffer = get_buffer(hnd);if (buffer) {handles_[hnd].second++;buffer->increase_refcount();return 0;}/*** @brief Construct a new if object* https://juejin.cn/post/7184685220158210107* 从fd导入bo* gem都是通过handle去查找一个bo，要通过fd转换成一个bo，首先要将fd转换成handle。drm提供了* drmPrimeFDToHandle 接口，获得fd 对应的handle。* 对应的handle 转fd函数是drmPrimeHandleToFD*/if (drmPrimeFDToHandle(drv_get_fd(drv_), hnd->fds[0], &id)) {drv_log("drmPrimeFDToHandle failed.\n");return -errno;}/*** @brief * */if (buffers_.count(id)) {buffer = buffers_[id];buffer->increase_refcount();} else {struct bo *bo;struct drv_import_fd_data data;data.format = hnd->format;data.width = hnd->width;data.height = hnd->height;data.use_flags = hnd->use_flags;memcpy(data.fds, hnd->fds, sizeof(data.fds));memcpy(data.strides, hnd->strides, sizeof(data.strides));memcpy(data.offsets, hnd->offsets, sizeof(data.offsets));for (uint32_t plane = 0; plane < DRV_MAX_PLANES; plane++) {data.format_modifiers[plane] = hnd->format_modifier;}/*** @brief * It's not going to be this way.* DRM_IOCTL_PRIME_FD_TO_HANDLE*/bo = drv_bo_import(drv_, &data);if (!bo)return -EFAULT;id = drv_bo_get_plane_handle(bo, 0).u32;buffer = new cros_gralloc_buffer(id, bo, nullptr, hnd->fds[hnd->num_planes],hnd->reserved_region_size);buffers_.emplace(id, buffer);}handles_.emplace(hnd, std::make_pair(buffer, 1));return 0;
}

这块的逻辑比较简单，就是取出前面aallocate中构建的buffer，看是否存在存在取出，没有则继续后续逻辑。


2.7 minigbm的graloc hal下lock的实现

当我们上层创建出来的GraphicBuffer被创建出来后，在正式被使用前还必须要lock锁定这块，才能被正式使用，我们看看mingi是如何实现的。

.lockAsync = gralloc0_lock_async
lock = gralloc0_lockmodule->lockAsyncgralloc0_lock_asyncauto hnd = cros_gralloc_convert_handle(handle)map_flags = gralloc0_convert_map_usage(usage)mod->driver->lock()cros_gralloc_driver::lockhnd = cros_gralloc_convert_handle(handle)auto buffer = get_buffer(hnd)buffer->lock()//cros_gralloc_buffer::lock

这里我们展开来看下cros_gralloc_buffer::lock的实现:

//cros_gralloc/cros_gralloc_buffer.cc
int32_t cros_gralloc_buffer::lock(const struct rectangle *rect, uint32_t map_flags,uint8_t *addr[DRV_MAX_PLANES])
{void *vaddr = nullptr;memset(addr, 0, DRV_MAX_PLANES * sizeof(*addr));/** Gralloc consumers don't support more than one kernel buffer per buffer object yet, so* just use the first kernel buffer.*/if (drv_num_buffers_per_bo(bo_) != 1) {drv_log("Can only support one buffer per bo.\n");return -EINVAL;}if (map_flags) {if (lock_data_[0]) {drv_bo_invalidate(bo_, lock_data_[0]);vaddr = lock_data_[0]->vma->addr;} else {struct rectangle r = *rect;if (!r.width && !r.height && !r.x && !r.y) {/** Android IMapper.hal: An accessRegion of all-zeros means the* entire buffer.*/r.width = drv_bo_get_width(bo_);r.height = drv_bo_get_height(bo_);}vaddr = drv_bo_map(bo_, &r, map_flags, &lock_data_[0], 0);}if (vaddr == MAP_FAILED) {drv_log("Mapping failed.\n");return -EFAULT;}}for (uint32_t plane = 0; plane < num_planes_; plane++)addr[plane] = static_cast<uint8_t *>(vaddr) + drv_bo_get_plane_offset(bo_, plane);lockcount_++;return 0;
}

里面核心的逻辑就是drv_bo_map，我们看看他的逻辑是什么:

.bo_map = drv_dumb_bo_map,
//drv.caddr = bo->drv->backend->bo_map(bo, mapping.vma, plane, map_flags)
//helpers.c
void *drv_dumb_bo_map(struct bo *bo, struct vma *vma, size_t plane, uint32_t map_flags)
{int ret;size_t i;struct drm_mode_map_dumb map_dumb;memset(&map_dumb, 0, sizeof(map_dumb));map_dumb.handle = bo->handles[plane].u32;/*** @brief * map kerne space to user space*/ret = drmIoctl(bo->drv->fd, DRM_IOCTL_MODE_MAP_DUMB, &map_dumb);if (ret) {drv_log("DRM_IOCTL_MODE_MAP_DUMB failed\n");return MAP_FAILED;}for (i = 0; i < bo->meta.num_planes; i++)if (bo->handles[i].u32 == bo->handles[plane].u32)vma->length += bo->meta.sizes[i];return mmap(0, vma->length, drv_get_prot(map_flags), MAP_SHARED, bo->drv->fd,map_dumb.offset);
}


三.总结

通过前面的分析，我们可以知道minigbm将libdrm的操作剥离出Mesa架构，然后调用libdrm的用户层接口申请dumb buffer，然后封装成GraphicBuffer给上层使用。关于DRM入门的知识重点推荐何小龙大神的DRM（Direct Rendering Manager）学习简介。好了今天的博客Android minigbm框架普法就到这里了。总之，青山不改绿水长流先到这里了。如果本博客对你有所帮助，麻烦关注或者点个赞，如果觉得很烂也可以踩一脚！谢谢各位了！！

本文来自互联网用户投稿，该文观点仅代表作者本人，不代表本站立场。本站仅提供信息存储空间服务，不拥有所有权，不承担相关法律责任。如若转载，请注明出处：http://www.mzph.cn/news/713414.shtml

如若内容造成侵权/违法违规/事实不符，请联系多彩编程网进行投诉反馈email:809451989@qq.com，一经查实，立即删除！

相关文章

Galaxy生信云平台：集合操作工具大全

Galaxy生信云平台：集合操作工具大全

Galaxy平台上的文件称为数据集&#xff08;Dataset&#xff09;&#xff0c;如果将多个文件组合在一起&#xff0c;则形成数据集合&#xff08;Dataset collection&#xff09;。 上传文件后&#xff0c;可以通过工具将文件构建成数据集合。具体操作可以参考前面介绍转录组流程…
阅读更多...
后台组件体系

后台组件体系

从今天开始进入更细粒度说明。后台微服务是由组件构成的。平台的开发理念是为甲方打造一个生态环境。安装实施时为客户安装私仓来管理组件。开发微服务时鼓励拆分为组件。开发新功能时&#xff0c;先看有没有相关组件&#xff0c;有的话就在pom.xml文件&#xff08;不要问我这个…
阅读更多...
Day10：基础入门-HTTP数据包Postman构造请求方法请求头修改状态码判断

Day10：基础入门-HTTP数据包Postman构造请求方法请求头修改状态码判断

目录 数据-方法&头部&状态码 案例-文件探针 案例-登录爆破 工具-Postman自构造使用 思维导图 章节知识点&#xff1a; 应用架构&#xff1a;Web/APP/云应用/三方服务/负载均衡等 安全产品&#xff1a;CDN/WAF/IDS/IPS/蜜罐/防火墙/杀毒等 渗透命令&#xff1a;文件…
阅读更多...
什么是端点安全以及如何保护端点

什么是端点安全以及如何保护端点

什么是端点安全 端点是指可以接收信号的任何设备&#xff0c;是员工使用的一种计算设备&#xff0c;用于保存公司数据或可以访问 Internet。端点的几个示例包括&#xff1a;服务器、工作站&#xff08;台式机和笔记本电脑&#xff09;、移动设备、虚拟机、平板电脑、物联网、可…
阅读更多...
一【初识EMC】

一【初识EMC】

在作为硬件行业相关从业者&#xff0c;经常接触到EMC相关问题&#xff0c;下面来简单介绍下EMC相关方面的知识 文章目录 前言一、生活中的EMC现象&#xff1f;二、EMC是什么三、EMC的三要素四、EMI与EMS的评估方式1.RE2.CE3.HAR4.FLICKER5.Rs6.CS7.ESD8.EFT9.DIP10.PMS11.surge…
阅读更多...
Zookeeper3：客户端命令

Zookeeper3：客户端命令

文章目录 客户端命令连接服务端Zookeeper客户端内置命令 ls - 节点信息 客户端命令 连接服务端Zookeeper //客户端连接服务端zookeeper 默认连的本机2181端口的zookeeper cd /opt/module/zookeeper-3.9.1/bin && sh zkCli.sh//客户端连接远程服务端zookeeper cd /op…
阅读更多...
【小尘送书-第十一期】编程的基石，开发的核心:《算法秘籍》

【小尘送书-第十一期】编程的基石，开发的核心:《算法秘籍》

大家好&#xff0c;我是小尘&#xff0c;欢迎你的关注&#xff01;大家可以一起交流学习&#xff01;欢迎大家在CSDN后台私信我&#xff01;一起讨论学习&#xff0c;讨论如何找到满意的工作&#xff01; &#x1f468;‍&#x1f4bb;博主主页&#xff1a;小尘要自信 &#x1…
阅读更多...
如何使用GAP-Burp-Extension扫描潜在的参数和节点

如何使用GAP-Burp-Extension扫描潜在的参数和节点

关于GAP-Burp-Extension GAP-Burp-Extension是一款功能强大的Burp扩展&#xff0c;该工具在getAllParams扩展的基础上进行了升级&#xff0c;该工具不仅可以帮助广大研究人员在安全审计过程中扫描潜在的参数&#xff0c;而且还可以搜索潜在的链接并使用这些参数进行测试&#…
阅读更多...
零基础如何快速入门伦敦金交易

零基础如何快速入门伦敦金交易

伦敦金交易是金融市场中备受关注的一种投资方式。对于想要学习如何炒伦敦金并快速开始交易的人来说&#xff0c;本文将为您提供一份全面而详细的指南。无论您是初学者还是有经验的交易者&#xff0c;本文都将帮助您了解伦敦金交易的基本知识&#xff0c;并提供一些实用的技巧和…
阅读更多...
协议-http协议-基础概念03-http状态码-http特点-http性能-压缩和分块传输-范围请求

协议-http协议-基础概念03-http状态码-http特点-http性能-压缩和分块传输-范围请求

参考来源&#xff1a; 极客时间-透视HTTP协议(作者&#xff1a;罗剑锋)&#xff1b; 01-状态码分类 开头的 Version 部分是 HTTP 协议的版本号&#xff0c;通常是HTTP/1.1&#xff0c;用处不是很大。后面的 Reason 部分是原因短语&#xff0c;是状态码的简短文字描述&#xff…
阅读更多...
每周一算法：双向广搜

每周一算法：双向广搜

题目链接 字符串变换 题目描述 已知有两个字串 A , B A,B A,B&#xff0c;及一组字串变换的规则&#xff08;至多 6 6 6个规则&#xff09;: A 1 → B 1 A_1→B_1 A1​→B1​ A 2 → B 2 A_2→B_2 A2​→B2​ … 规则的含义为&#xff1a;在 A A A中的子串 A 1 A_1 A1​…
阅读更多...
pytorch统计属性

pytorch统计属性

目录 1.normal2. mean, sum, min, max, prod3.argmin, argmax4. topk kthvalue5. compare 1.normal torch.normal(mean, std, *, generatorNone, outNone) → Tensor返回一个张量&#xff0c;其中的每个元素随机来自独立的标准正态分布。这些分布具有给定的均值和标准差。 参数…
阅读更多...
【cmu15445c++入门】(13)C++的std::promise

【cmu15445c++入门】(13)C++的std::promise

一、说明 std::promise 是C11并发编程中常用的一个类&#xff0c;常配合std::future使用。其作用是在一个线程t1中保存一个类型typename T的值&#xff0c;可供相绑定的std::future对象在另一线程t2中获取 二、代码 #include <chrono> #include <future> #includ…
阅读更多...
Hive SQL 开发指南（二）使用（DDL、DML，DQL）

Hive SQL 开发指南（二）使用（DDL、DML，DQL）

在大数据领域&#xff0c;Hive SQL 是一种常用的查询语言&#xff0c;用于在 Hadoop上进行数据分析和处理。为了确保代码的可读性、维护性和性能&#xff0c;制定一套规范化的 Hive SQL 开发规范至关重要。本文将介绍 Hive SQL 的基础知识&#xff0c;并提供一些规范化的开发指…
阅读更多...
如何安装ProtoBuf环境

如何安装ProtoBuf环境

1 &#x1f351;下载 ProtoBuf&#x1f351; 下载 ProtoBuf 前⼀定要安装依赖库&#xff1a;autoconf automake libtool curl make g unzip 如未安装&#xff0c;安装命令如下&#xff1a; Ubuntu ⽤⼾选择&#xff1a; sudo apt-get install autoconf automake libtool cur…
阅读更多...
Unity 向量计算、欧拉角与四元数转换、输出文本、告警、错误、修改时间、定时器、路径、

Unity 向量计算、欧拉角与四元数转换、输出文本、告警、错误、修改时间、定时器、路径、

using System.Collections; using System.Collections.Generic; using UnityEngine;public class c2 : MonoBehaviour {// 定时器float t1 0;void Start(){// 向量Vector3 v1 new Vector3(0, 0, 2);Vector3 v2 new Vector3(0, 0, 3);// 计算两个向量的夹角Debug.Log(Vector3…
阅读更多...
Java 学习和实践笔记（26）：组合（component）的含义以及与继承（extends)的关系

Java 学习和实践笔记（26）：组合（component）的含义以及与继承（extends)的关系

组合的两个作用&#xff1a; 1&#xff09;通过将父类对象作为子类的属性 2&#xff09;通过第1点的作用&#xff0c;实现了代码复用。 示例代码&#xff1a; public class TestComponent {public static void main(String[] args) {Student2 s1 new Student2("jason&…
阅读更多...
灯塔：HTML笔记

灯塔：HTML笔记

网页由哪些部分组成&#xff1f; *文字 图片 音频 视频 超链接 程序员写的代码是通过浏览器转换成网页的 五大浏览器有哪些&#xff1f; *IE浏览器 *火狐浏览器&#xff08;Firefox&#xff09; *谷歌浏览器&#xff08;Chrome&#xff09; *Safari浏览器 *欧朋浏览器&…
阅读更多...
NENU OJ算法2例题||搜索E

NENU OJ算法2例题||搜索E

NENU OJ算法2例题 合集原文指路 算法2搜索E 1281: E001 数的划分 题目描述 将整数n分成k份&#xff0c;且每份不能为空&#xff0c;任意两种分法不能相同&#xff08;不考虑顺序&#xff09;。 例如&#xff1a;n7&#xff0c;k3&#xff0c;下面三种分法被认为是相同的。…
阅读更多...
【数据结构与算法】动态规划法解题20240302

【数据结构与算法】动态规划法解题20240302

这里写目录标题 一、198. 打家劫舍1、动态规划五部曲 二、213. 打家劫舍 II 一、198. 打家劫舍 你是一个专业的小偷&#xff0c;计划偷窃沿街的房屋。每间房内都藏有一定的现金&#xff0c;影响你偷窃的唯一制约因素就是相邻的房屋装有相互连通的防盗系统&#xff0c;如果两间…
阅读更多...
推荐文章
最新文章

Copyright @ 2022~2023