Refactor: Replace DRM swap buffer implementation with asynchronous page-flipping and framebuffer caching
The original implementation created/destroyed framebuffers (FBs) per-frame, leading to kernel overhead and screen tearing. This commit replaces it with a different approach using: - Asynchronous `drmModePageFlip()` with vblank sync - Framebuffer caching to reduce repeated FB creation/removal operations - Proper resource management through BO callbacks and buffer release synchronization - Added error handling for busy displays, cache overflows, and flip failures - Event-driven cleanup via page_flip_handler to prevent GPU/scanout conflicts Co-authored-by: rob-bits
This commit is contained in:
katonar
parent
5f497d0687
commit
060bd787b1
1 changed files with 220 additions and 25 deletions
|
|
@ -71,6 +71,21 @@
|
|||
#include "EGL/egl.h" // Native platform windowing system interface
|
||||
#include "EGL/eglext.h" // EGL extensions
|
||||
|
||||
#include <poll.h> // for drmHandleEvent poll
|
||||
#include <errno.h> //for EBUSY, EAGAIN
|
||||
|
||||
#define MAX_CACHED_BOS 3
|
||||
|
||||
typedef struct {
|
||||
struct gbm_bo *bo;
|
||||
uint32_t fb_id; // DRM framebuffer ID
|
||||
} FramebufferCache;
|
||||
|
||||
static FramebufferCache fbCache[MAX_CACHED_BOS];
|
||||
static volatile int fbCacheCount = 0;
|
||||
static volatile bool pendingFlip = false;
|
||||
static bool crtcSet = false;
|
||||
|
||||
#ifndef EGL_OPENGL_ES3_BIT
|
||||
#define EGL_OPENGL_ES3_BIT 0x40
|
||||
#endif
|
||||
|
|
@ -217,6 +232,7 @@ static const short linuxToRaylibMap[KEYMAP_SIZE] = {
|
|||
//----------------------------------------------------------------------------------
|
||||
// Module Internal Functions Declaration
|
||||
//----------------------------------------------------------------------------------
|
||||
int InitSwapScreenBuffer(void);
|
||||
int InitPlatform(void); // Initialize platform (graphics, inputs and more)
|
||||
void ClosePlatform(void); // Close platform
|
||||
|
||||
|
|
@ -551,34 +567,207 @@ void DisableCursor(void)
|
|||
CORE.Input.Mouse.cursorHidden = true;
|
||||
}
|
||||
|
||||
// Swap back buffer with front buffer (screen drawing)
|
||||
void SwapScreenBuffer(void)
|
||||
{
|
||||
eglSwapBuffers(platform.device, platform.surface);
|
||||
static void drm_fb_destroy_callback(struct gbm_bo *bo, void *data) {
|
||||
uint32_t fb_id = (uintptr_t)data;
|
||||
// Remove from cache
|
||||
for (int i = 0; i < fbCacheCount; i++) {
|
||||
if (fbCache[i].bo == bo) {
|
||||
TRACELOG(LOG_INFO, "DRM: fb removed %u", (uintptr_t)fb_id);
|
||||
drmModeRmFB(platform.fd, fbCache[i].fb_id); // Release DRM FB
|
||||
// Shift remaining entries
|
||||
for (int j = i; j < fbCacheCount - 1; j++) {
|
||||
fbCache[j] = fbCache[j + 1];
|
||||
}
|
||||
fbCacheCount--;
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (!platform.gbmSurface || (-1 == platform.fd) || !platform.connector || !platform.crtc) TRACELOG(LOG_ERROR, "DISPLAY: DRM initialization failed to swap");
|
||||
|
||||
struct gbm_bo *bo = gbm_surface_lock_front_buffer(platform.gbmSurface);
|
||||
if (!bo) TRACELOG(LOG_ERROR, "DISPLAY: Failed GBM to lock front buffer");
|
||||
|
||||
uint32_t fb = 0;
|
||||
int result = drmModeAddFB(platform.fd, platform.connector->modes[platform.modeIndex].hdisplay, platform.connector->modes[platform.modeIndex].vdisplay, 24, 32, gbm_bo_get_stride(bo), gbm_bo_get_handle(bo).u32, &fb);
|
||||
if (result != 0) TRACELOG(LOG_ERROR, "DISPLAY: drmModeAddFB() failed with result: %d", result);
|
||||
|
||||
result = drmModeSetCrtc(platform.fd, platform.crtc->crtc_id, fb, 0, 0, &platform.connector->connector_id, 1, &platform.connector->modes[platform.modeIndex]);
|
||||
if (result != 0) TRACELOG(LOG_ERROR, "DISPLAY: drmModeSetCrtc() failed with result: %d", result);
|
||||
|
||||
if (platform.prevFB)
|
||||
{
|
||||
result = drmModeRmFB(platform.fd, platform.prevFB);
|
||||
if (result != 0) TRACELOG(LOG_ERROR, "DISPLAY: drmModeRmFB() failed with result: %d", result);
|
||||
// Create or retrieve cached DRM FB for BO
|
||||
static uint32_t get_or_create_fb_for_bo(struct gbm_bo *bo) {
|
||||
// Try to find existing cache entry
|
||||
for (int i = 0; i < fbCacheCount; i++) {
|
||||
if (fbCache[i].bo == bo) {
|
||||
return fbCache[i].fb_id;
|
||||
}
|
||||
}
|
||||
|
||||
platform.prevFB = fb;
|
||||
// Create new entry if cache not full
|
||||
if (fbCacheCount >= MAX_CACHED_BOS) {
|
||||
//FB cache full!
|
||||
return 0;
|
||||
}
|
||||
|
||||
if (platform.prevBO) gbm_surface_release_buffer(platform.gbmSurface, platform.prevBO);
|
||||
uint32_t handle = gbm_bo_get_handle(bo).u32;
|
||||
uint32_t stride = gbm_bo_get_stride(bo);
|
||||
uint32_t width = gbm_bo_get_width(bo);
|
||||
uint32_t height = gbm_bo_get_height(bo);
|
||||
|
||||
uint32_t fb_id;
|
||||
if (drmModeAddFB(platform.fd, width, height, 24, 32, stride, handle, &fb_id)) {
|
||||
//rmModeAddFB failed
|
||||
return 0;
|
||||
}
|
||||
|
||||
// Store in cache
|
||||
fbCache[fbCacheCount] = (FramebufferCache){ .bo = bo, .fb_id = fb_id };
|
||||
fbCacheCount++;
|
||||
|
||||
// Set destroy callback to auto-cleanup
|
||||
gbm_bo_set_user_data(bo, (void*)(uintptr_t)fb_id, drm_fb_destroy_callback);
|
||||
|
||||
TRACELOG(LOG_INFO, "DRM: added new bo %u" , (uintptr_t)fb_id);
|
||||
return fb_id;
|
||||
}
|
||||
|
||||
// Renders a blank frame to allocate initial buffers
|
||||
void RenderBlankFrame() {
|
||||
glClearColor(0, 0, 0, 1);
|
||||
glClear(GL_COLOR_BUFFER_BIT);
|
||||
eglSwapBuffers(platform.device, platform.surface);
|
||||
|
||||
// Ensure the buffer is processed
|
||||
glFinish();
|
||||
}
|
||||
|
||||
// Initialize with first buffer only
|
||||
int InitSwapScreenBuffer() {
|
||||
if (!platform.gbmSurface || platform.fd < 0) {
|
||||
TRACELOG(LOG_ERROR, "DRM not initialized");
|
||||
return -1;
|
||||
}
|
||||
|
||||
// Render a blank frame to allocate buffers
|
||||
RenderBlankFrame();
|
||||
|
||||
// Get first buffer
|
||||
struct gbm_bo *bo = gbm_surface_lock_front_buffer(platform.gbmSurface);
|
||||
if (!bo) {
|
||||
TRACELOG(LOG_ERROR, "Failed to lock initial buffer");
|
||||
return -1;
|
||||
}
|
||||
|
||||
// Create FB for first buffer
|
||||
uint32_t fb_id = get_or_create_fb_for_bo(bo);
|
||||
if (!fb_id) {
|
||||
gbm_surface_release_buffer(platform.gbmSurface, bo);
|
||||
return -1;
|
||||
}
|
||||
|
||||
// Initial CRTC setup
|
||||
if (drmModeSetCrtc(platform.fd, platform.crtc->crtc_id, fb_id,
|
||||
0, 0, &platform.connector->connector_id, 1,
|
||||
&platform.connector->modes[platform.modeIndex])) {
|
||||
TRACELOG(LOG_ERROR, "Initial CRTC setup failed: %s", strerror(errno));
|
||||
gbm_surface_release_buffer(platform.gbmSurface, bo);
|
||||
return -1;
|
||||
}
|
||||
|
||||
// Keep first buffer locked until flipped
|
||||
platform.prevBO = bo;
|
||||
crtcSet = true;
|
||||
return 0;
|
||||
}
|
||||
|
||||
// Static page flip handler
|
||||
// this will be called once the drmModePageFlip() finished from the drmHandleEvent(platform.fd, &evctx); context
|
||||
static void page_flip_handler(int fd, unsigned int frame,
|
||||
unsigned int sec, unsigned int usec,
|
||||
void *data) {
|
||||
(void)fd; (void)frame; (void)sec; (void)usec; // Unused
|
||||
pendingFlip = false;
|
||||
struct gbm_bo *bo_to_release = (struct gbm_bo *)data;
|
||||
//Buffers are released after the flip completes (via page_flip_handler), ensuring they're no longer in use.
|
||||
// Prevents the GPU from writing to a buffer being scanned out
|
||||
if (bo_to_release) {
|
||||
gbm_surface_release_buffer(platform.gbmSurface, bo_to_release);
|
||||
}
|
||||
}
|
||||
|
||||
// Swap implementation with proper caching
|
||||
void SwapScreenBuffer() {
|
||||
static int loopCnt = 0;
|
||||
loopCnt++;
|
||||
static int errCnt[5] = {0};
|
||||
if (!crtcSet || !platform.gbmSurface) return;
|
||||
|
||||
//call this only, if pendingFlip is not set
|
||||
eglSwapBuffers(platform.device, platform.surface);
|
||||
|
||||
// Process pending events non-blocking
|
||||
drmEventContext evctx = {
|
||||
.version = DRM_EVENT_CONTEXT_VERSION,
|
||||
.page_flip_handler = page_flip_handler
|
||||
};
|
||||
|
||||
struct pollfd pfd = { .fd = platform.fd, .events = POLLIN };
|
||||
//polling for event for 0ms
|
||||
while (poll(&pfd, 1, 0) > 0) {
|
||||
drmHandleEvent(platform.fd, &evctx);
|
||||
}
|
||||
|
||||
// Skip if previous flip pending
|
||||
if (pendingFlip) {
|
||||
//Skip frame: flip pending
|
||||
errCnt[0]++;
|
||||
return;
|
||||
}
|
||||
|
||||
// Get new front buffer
|
||||
struct gbm_bo *next_bo = gbm_surface_lock_front_buffer(platform.gbmSurface);
|
||||
if (!next_bo) {
|
||||
//Failed to lock front buffer
|
||||
errCnt[1]++;
|
||||
return;
|
||||
}
|
||||
|
||||
// Get FB ID (creates new one if needed)
|
||||
uint32_t fb_id = get_or_create_fb_for_bo(next_bo);
|
||||
if (!fb_id) {
|
||||
gbm_surface_release_buffer(platform.gbmSurface, next_bo);
|
||||
errCnt[2]++;
|
||||
return;
|
||||
}
|
||||
|
||||
// Attempt page flip
|
||||
/* rmModePageFlip() schedules a buffer-flip for the next vblank and then
|
||||
* notifies us about it. It takes a CRTC-id, fb-id and an arbitrary
|
||||
* data-pointer and then schedules the page-flip. This is fully asynchronous and
|
||||
* When the page-flip happens, the DRM-fd will become readable and we can call
|
||||
* drmHandleEvent(). This will read all vblank/page-flip events and call our
|
||||
* modeset_page_flip_event() callback with the data-pointer that we passed to
|
||||
* drmModePageFlip(). We simply call modeset_draw_dev() then so the next frame
|
||||
* is rendered..
|
||||
* returns immediately.
|
||||
*/
|
||||
if (drmModePageFlip(platform.fd, platform.crtc->crtc_id, fb_id,
|
||||
DRM_MODE_PAGE_FLIP_EVENT, platform.prevBO)) {
|
||||
if (errno == EBUSY) {
|
||||
//Display busy - skip flip
|
||||
errCnt[3]++;
|
||||
} else {
|
||||
//Page flip failed
|
||||
errCnt[4]++;
|
||||
}
|
||||
gbm_surface_release_buffer(platform.gbmSurface, next_bo);
|
||||
return;
|
||||
}
|
||||
|
||||
// Success: update state
|
||||
pendingFlip = true;
|
||||
|
||||
platform.prevBO = next_bo;
|
||||
//successful usage, do benchmarking
|
||||
//in every 10 sec, at 60FPS 60*10 -> 600
|
||||
if(loopCnt >= 600) {
|
||||
TRACELOG(LOG_INFO, "DRM err counters: %d, %d, %d, %d, %d, %d",errCnt[0],errCnt[1],errCnt[2],errCnt[3],errCnt[4], loopCnt);
|
||||
//reinit the errors
|
||||
for(int i=0;i<5;i++) {
|
||||
errCnt[i] = 0;
|
||||
}
|
||||
loopCnt = 0;
|
||||
}
|
||||
}
|
||||
|
||||
//----------------------------------------------------------------------------------
|
||||
|
|
@ -910,7 +1099,8 @@ int InitPlatform(void)
|
|||
EGL_BLUE_SIZE, 8, // BLUE color bit depth (alternative: 5)
|
||||
EGL_ALPHA_SIZE, 8, // ALPHA bit depth (required for transparent framebuffer)
|
||||
//EGL_TRANSPARENT_TYPE, EGL_NONE, // Request transparent framebuffer (EGL_TRANSPARENT_RGB does not work on RPI)
|
||||
EGL_DEPTH_SIZE, 24, // Depth buffer size (Required to use Depth testing!)
|
||||
//ToDo: verify this. In 5.5 it is 16, in master it was 24
|
||||
EGL_DEPTH_SIZE, 16, // Depth buffer size (Required to use Depth testing!)
|
||||
//EGL_STENCIL_SIZE, 8, // Stencil buffer size
|
||||
EGL_SAMPLE_BUFFERS, sampleBuffer, // Activate MSAA
|
||||
EGL_SAMPLES, samples, // 4x Antialiasing if activated (Free on MALI GPUs)
|
||||
|
|
@ -1083,9 +1273,14 @@ int InitPlatform(void)
|
|||
CORE.Storage.basePath = GetWorkingDirectory();
|
||||
//----------------------------------------------------------------------------
|
||||
|
||||
TRACELOG(LOG_INFO, "PLATFORM: DRM: Initialized successfully");
|
||||
if(InitSwapScreenBuffer() == 0) {
|
||||
TRACELOG(LOG_INFO, "PLATFORM: DRM: Initialized successfully");
|
||||
return 0;
|
||||
} else {
|
||||
TRACELOG(LOG_INFO, "PLATFORM: DRM: Initialized failed");
|
||||
return -1;
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
// Close platform
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue