sdl2_frt/src/video/kmsdrm/SDL_kmsdrmopengles.c

347 lines
12 KiB
C
Raw Normal View History

/*
Simple DirectMedia Layer
2020-01-17 05:49:25 +01:00
Copyright (C) 1997-2020 Sam Lantinga <slouken@libsdl.org>
This software is provided 'as-is', without any express or implied
warranty. In no event will the authors be held liable for any damages
arising from the use of this software.
Permission is granted to anyone to use this software for any purpose,
including commercial applications, and to alter it and redistribute it
freely, subject to the following restrictions:
1. The origin of this software must not be misrepresented; you must not
claim that you wrote the original software. If you use this software
in a product, an acknowledgment in the product documentation would be
appreciated but is not required.
2. Altered source versions must be plainly marked as such, and must not be
misrepresented as being the original software.
3. This notice may not be removed or altered from any source distribution.
*/
#include "../../SDL_internal.h"
#if SDL_VIDEO_DRIVER_KMSDRM && SDL_VIDEO_OPENGL_EGL
#include "SDL_kmsdrmvideo.h"
#include "SDL_kmsdrmopengles.h"
#include "SDL_kmsdrmdyn.h"
#ifndef EGL_PLATFORM_GBM_MESA
#define EGL_PLATFORM_GBM_MESA 0x31D7
#endif
/* EGL implementation of SDL OpenGL support */
int
KMSDRM_GLES_LoadLibrary(_THIS, const char *path) {
NativeDisplayType display = (NativeDisplayType)((SDL_VideoData *)_this->driverdata)->gbm_dev;
return SDL_EGL_LoadLibrary(_this, path, display, EGL_PLATFORM_GBM_MESA);
}
SDL_EGL_CreateContext_impl(KMSDRM)
int KMSDRM_GLES_SetSwapInterval(_THIS, int interval) {
if (!_this->egl_data) {
return SDL_SetError("EGL not initialized");
}
if (interval == 0 || interval == 1) {
_this->egl_data->egl_swapinterval = interval;
} else {
return SDL_SetError("Only swap intervals of 0 or 1 are supported");
}
return 0;
}
/*********************************/
/* Atomic functions block */
/*********************************/
#define VOID2U64(x) ((uint64_t)(unsigned long)(x))
static EGLSyncKHR create_fence(int fd, _THIS)
{
EGLint attrib_list[] = {
EGL_SYNC_NATIVE_FENCE_FD_ANDROID, fd,
EGL_NONE,
};
EGLSyncKHR fence = _this->egl_data->eglCreateSyncKHR(_this->egl_data->egl_display,
EGL_SYNC_NATIVE_FENCE_ANDROID, attrib_list);
assert(fence);
return fence;
}
/*
-We have to stop the GPU from executing the cmdstream again because the front butter
has been marked as back buffer so it can be selected by EGL to draw on it BUT
the pageflip has not completed, so it's still on screen, so letting GPU render on it
would cause tearing. (kms_fence)
-We have to stop the DISPLAY from doing the changes we requested in the atomic ioctl,
like the pageflip, until the GPU has completed the cmdstream execution,
because we don't want the pageflip to be done in the middle of a frame rendering.
(gpu_fence).
-We have to stop the program from doing a new atomic ioctl until the previous one
has been finished. (kms_fence)
*/
int
KMSDRM_GLES_SwapWindowDOUBLE(_THIS, SDL_Window * window) {
Fixed bug 4966 - KMSDRM: Add dynamic modeset support Anthony Pesch * Remove triple buffering support. As far as I can tell, this goes against the libdrm API; the EGL implementations themselves control the buffering. Removing it isn't absolutely necessary as it seemingly works on the Pi at least, but I noticed this while doing my work and explained my reasoning in the commit. * Replace the crtc_ready logic which allocates an extra bo to perform the initial CRTC configuration (which is required before calling drmModePageFlip) with a call to drmModeSetCrtc after the front and back buffers are allocated, avoiding this allocation. * Standardized the SDL_*Data variable names and null checks to improve readability. Given that there were duplicate fields in each SDL_*Data structure, having generic names such as "data" at times was very confusing. * Removed unused fields from the SDL_*Data structures and moves all display related fields out of SDL_VideoData and into SDL_DisplayData. Not required since the code only supports a single display right now, but this was helpful in reading and understanding the code initially. * Implement KMSDRM_GetDisplayModes / KMSDRM_SetDisplayMode to provide dynamic modeset support. These changes have been tested on a Raspberry Pi 4 and a Dell XPS laptop with an HD 520. As an update, I went back over the triple buffer changes and left them in. I didn't entirely get the code originally, I had just seen it calling KMSDRM_gbm_surface_lock_front_buffer twice for a single swap and had removed it because I was paranoid of bugs stemming from it while working on the modeset changes. I've made a few small changes to the logic that had thrown me off originally and rebased the changes: * The condition wrapping the call to release buffer was incorrect. * The first call to KMSDRM_gbm_surface_lock_front_buffer has been removed. I don't understand why it existed. * Added additional comments describing what was going on in the code (as it does fix the buffer release pattern of the original code before it).
2020-02-09 20:44:22 +01:00
SDL_WindowData *windata = ((SDL_WindowData *) window->driverdata);
SDL_DisplayData *dispdata = (SDL_DisplayData *) SDL_GetDisplayForWindow(window)->driverdata;
KMSDRM_FBInfo *fb;
Fixed bug 5140 - KMSDRM: Dynamic vsync toggle does not work Manuel Alfayate Corchete The KMSDRM backend was doing things wrong because of some small (but important) misconceptions on how KMS/DRM works: to implement a largely broken non-vsync refresh mechanism, the SwapWindow() function was issuing new pageflips before previous ones had completed, thus causing EBUSY returns, buffer mismanagement, etc... resulting in general breakage on vsync disabling from apps, that would not allow vsync to work again without KMSDRM video re-initialization. To further clarify, on most DRM drivers async pageflips are NOT working nowadays, so all issued pageflips will complete on next VBLANK, NOT ASAP (calling drmModePageFlip() with the DRM_MODE_PAGE_FLIP_ASYNC flag will return error). The old code was assuming that can just issue a synchronous (=on VBLANK) pageflip and then pass a 0 timeout to the pull() function so we do not wait for the pageflip event, thinking that this will lead to correct non-vsynced screen updates from the program: That is plain wrong. Each pageflip has to be waite before issuing a new one, ALWAYS. And if we do not support ASYNC pageflips on the DRM driver level, then we are forced to wait for the next VBLANK. There is no way around it. I have also added many comments on the KMSDRM code. This is needed for future reference for me or others who may need to look at this code: KMS/DRM terminology regarding what SYNC and ASYNC mean in pageflip terms, and where to do certain things and why, is not trivial. It is not desirable or possible to invest time on researching the same concepts every time there is need to dive into this code. So please leave all these comments in the patch.
2020-05-27 01:27:00 +02:00
EGLint status;
int ret;
uint32_t flags = DRM_MODE_ATOMIC_NONBLOCK;
Fixed bug 5140 - KMSDRM: Dynamic vsync toggle does not work Manuel Alfayate Corchete The KMSDRM backend was doing things wrong because of some small (but important) misconceptions on how KMS/DRM works: to implement a largely broken non-vsync refresh mechanism, the SwapWindow() function was issuing new pageflips before previous ones had completed, thus causing EBUSY returns, buffer mismanagement, etc... resulting in general breakage on vsync disabling from apps, that would not allow vsync to work again without KMSDRM video re-initialization. To further clarify, on most DRM drivers async pageflips are NOT working nowadays, so all issued pageflips will complete on next VBLANK, NOT ASAP (calling drmModePageFlip() with the DRM_MODE_PAGE_FLIP_ASYNC flag will return error). The old code was assuming that can just issue a synchronous (=on VBLANK) pageflip and then pass a 0 timeout to the pull() function so we do not wait for the pageflip event, thinking that this will lead to correct non-vsynced screen updates from the program: That is plain wrong. Each pageflip has to be waite before issuing a new one, ALWAYS. And if we do not support ASYNC pageflips on the DRM driver level, then we are forced to wait for the next VBLANK. There is no way around it. I have also added many comments on the KMSDRM code. This is needed for future reference for me or others who may need to look at this code: KMS/DRM terminology regarding what SYNC and ASYNC mean in pageflip terms, and where to do certain things and why, is not trivial. It is not desirable or possible to invest time on researching the same concepts every time there is need to dive into this code. So please leave all these comments in the patch.
2020-05-27 01:27:00 +02:00
EGLSyncKHR kms_in_fence = NULL;
EGLSyncKHR kms_out_fence = NULL;
/* Create the display in-fence, and export it as a fence fd to pass into the kernel. */
kms_in_fence = create_fence(EGL_NO_NATIVE_FENCE_FD_ANDROID, _this);
assert(kms_in_fence);
dispdata->kms_in_fence_fd = _this->egl_data->eglDupNativeFenceFDANDROID(_this->egl_data->egl_display, kms_in_fence);
/* Mark the back buffer as the next front buffer, and the current front buffer as elegible
by EGL as back buffer to draw into.
This will not really happen until we post the pageflip though KMS with the atomic ioctl
and it completes on next vsync. */
_this->egl_data->eglSwapBuffers(_this->egl_data->egl_display, windata->egl_surface);
/* Lock the buffer that is marked by eglSwapBuffers() to become the next front buffer (so it can not
be chosen by EGL as back buffer to draw on), and get a handle to it, so we can use that handle
to request the KMS pageflip that will set it as front buffer. */
windata->next_bo = KMSDRM_gbm_surface_lock_front_buffer(windata->gs);
if (!windata->next_bo) {
printf("Failed to lock frontbuffer\n");
return -1;
}
fb = KMSDRM_FBFromBO(_this, windata->next_bo);
if (!fb) {
printf("Failed to get a new framebuffer BO\n");
return -1;
}
/*
* Issue atomic pageflip to post the pageflip thru KMS. Returns immediately.
* Never issue an atomic ioctl while the previuos one is pending: it will be rejected.
*/
ret = drm_atomic_commit(_this, fb->fb_id, flags);
if (ret) {
printf("failed to do atomic commit\n");
return -1;
}
/* Get the display out fence, returned by the atomic ioctl. */
kms_out_fence = create_fence(dispdata->kms_out_fence_fd, _this);
assert(kms_out_fence);
/* Wait on the CPU side for the _previous_ commit to complete before we post the flip through KMS,
* because atomic will reject the commit if we post a new one while the previous one is still pending. */
do {
status = _this->egl_data->eglClientWaitSyncKHR(_this->egl_data->egl_display, kms_out_fence, 0, EGL_FOREVER_KHR);
} while (status != EGL_CONDITION_SATISFIED_KHR);
/* Destroy both in and out display fences. */
_this->egl_data->eglDestroySyncKHR(_this->egl_data->egl_display, kms_out_fence);
_this->egl_data->eglDestroySyncKHR(_this->egl_data->egl_display, kms_in_fence);
/* Now that the pageflip is complete, release last front buffer so EGL can chose it
* as back buffer and render on it again: */
if (windata->curr_bo) {
KMSDRM_gbm_surface_release_buffer(windata->gs, windata->curr_bo);
windata->curr_bo = NULL;
}
/* Take note of the current front buffer, so it can be freed next time this function is called. */
windata->curr_bo = windata->next_bo;
return ret;
}
int
KMSDRM_GLES_SwapWindow(_THIS, SDL_Window * window) {
SDL_WindowData *windata = ((SDL_WindowData *) window->driverdata);
SDL_DisplayData *dispdata = (SDL_DisplayData *) SDL_GetDisplayForWindow(window)->driverdata;
KMSDRM_FBInfo *fb;
int ret;
uint32_t flags = DRM_MODE_ATOMIC_NONBLOCK;
/* Create the fence that will be inserted in the cmdstream exactly at the end
of the gl commands that form a frame. KMS will have to wait on it before doing a pageflip. */
dispdata->gpu_fence = create_fence(EGL_NO_NATIVE_FENCE_FD_ANDROID, _this);
assert(dispdata->gpu_fence);
_this->egl_data->eglSwapBuffers(_this->egl_data->egl_display, windata->egl_surface);
/* It's safe to get the gpu_fence FD now, because eglSwapBuffers flushes it
down the cmdstream, so it's now in place in the cmdstream.
Atomic ioctl will pass the in-fence fd into the kernel. */
dispdata->kms_in_fence_fd = _this->egl_data->eglDupNativeFenceFDANDROID(_this->egl_data->egl_display, dispdata->gpu_fence);
_this->egl_data->eglDestroySyncKHR(_this->egl_data->egl_display, dispdata->gpu_fence);
assert(dispdata->kms_in_fence_fd != -1);
/* Lock the buffer that is marked by eglSwapBuffers() to become the next front buffer (so it can not
be chosen by EGL as back buffer to draw on), and get a handle to it to request the pageflip on it. */
windata->next_bo = KMSDRM_gbm_surface_lock_front_buffer(windata->gs);
if (!windata->next_bo) {
printf("Failed to lock frontbuffer\n");
return -1;
}
fb = KMSDRM_FBFromBO(_this, windata->next_bo);
if (!fb) {
printf("Failed to get a new framebuffer BO\n");
return -1;
}
/* Don't issue another atomic ioctl until previous one has completed. */
if (dispdata->kms_fence) {
EGLint status;
do {
status = _this->egl_data->eglClientWaitSyncKHR(_this->egl_data->egl_display, dispdata->kms_fence, 0, EGL_FOREVER_KHR);
} while (status != EGL_CONDITION_SATISFIED_KHR);
_this->egl_data->eglDestroySyncKHR(_this->egl_data->egl_display, dispdata->kms_fence);
}
/* Issue atomic commit. */
ret = drm_atomic_commit(_this, fb->fb_id, flags);
if (ret) {
printf("failed to do atomic commit\n");
return -1;
}
/* release last front buffer so EGL can chose it as back buffer and render on it again: */
if (windata->curr_bo) {
KMSDRM_gbm_surface_release_buffer(windata->gs, windata->curr_bo);
windata->curr_bo = NULL;
}
/* Take note of the current front buffer, so it can be freed next time this function is called. */
windata->curr_bo = windata->next_bo;
/* Import out fence from the out fence fd and tell the GPU to wait on it
until the requested pageflip has completed. */
dispdata->kms_fence = create_fence(dispdata->kms_out_fence_fd, _this);
assert(dispdata->kms_fence);
dispdata->kms_out_fence_fd = -1;
_this->egl_data->eglWaitSyncKHR(_this->egl_data->egl_display, dispdata->kms_fence, 0);
return ret;
}
int
KMSDRM_GLES_SwapWindowOLD(_THIS, SDL_Window * window) {
SDL_WindowData *windata = ((SDL_WindowData *) window->driverdata);
SDL_DisplayData *dispdata = (SDL_DisplayData *) SDL_GetDisplayForWindow(window)->driverdata;
KMSDRM_FBInfo *fb;
int ret;
uint32_t flags = DRM_MODE_ATOMIC_NONBLOCK;
dispdata->kms_fence = NULL; /* in-fence to gpu, out-fence from kms */
dispdata->gpu_fence = NULL; /* in-fence to kms, out-fence from gpu, */
/* Allow modeset (which is done inside atomic_commit). */
flags |= DRM_MODE_ATOMIC_ALLOW_MODESET;
/* Import out fence from the out fence fd and tell the GPU to wait on it
until the requested pageflip has completed. */
dispdata->kms_fence = create_fence(dispdata->kms_out_fence_fd, _this);
assert(dispdata->kms_fence);
dispdata->kms_out_fence_fd = -1;
_this->egl_data->eglWaitSyncKHR(_this->egl_data->egl_display, dispdata->kms_fence, 0);
/* GL DRAW */
/* Create the gpu fence here so it's inserted in the cmdstream exactly
at the end of the gl commands that form a frame. */
dispdata->gpu_fence = create_fence(EGL_NO_NATIVE_FENCE_FD_ANDROID, _this);
assert(dispdata->gpu_fence);
_this->egl_data->eglSwapBuffers(_this->egl_data->egl_display, windata->egl_surface);
/* It's safe to get the gpu_fence fd now, because eglSwapBuffers() flushes the fd. */
dispdata->kms_in_fence_fd = _this->egl_data->eglDupNativeFenceFDANDROID(_this->egl_data->egl_display, dispdata->gpu_fence);
_this->egl_data->eglDestroySyncKHR(_this->egl_data->egl_display, dispdata->gpu_fence);
assert(dispdata->kms_in_fence_fd != -1);
windata->next_bo = KMSDRM_gbm_surface_lock_front_buffer(windata->gs);
if (!windata->next_bo) {
printf("Failed to lock frontbuffer\n");
return -1;
}
fb = KMSDRM_FBFromBO(_this, windata->next_bo);
if (!fb) {
printf("Failed to get a new framebuffer BO\n");
return -1;
}
if (dispdata->kms_fence) {
EGLint status;
do {
status = _this->egl_data->eglClientWaitSyncKHR(_this->egl_data->egl_display, dispdata->kms_fence, 0, EGL_FOREVER_KHR);
} while (status != EGL_CONDITION_SATISFIED_KHR);
_this->egl_data->eglDestroySyncKHR(_this->egl_data->egl_display, dispdata->kms_fence);
}
/* Issue atomic commit. */
ret = drm_atomic_commit(_this, fb->fb_id, flags);
if (ret) {
printf("failed to do atomic commit\n");
return -1;
}
/* release last front buffer so EGL can chose it as back buffer and render on it again: */
if (windata->curr_bo) {
KMSDRM_gbm_surface_release_buffer(windata->gs, windata->curr_bo);
windata->curr_bo = NULL;
}
/* Take note of the current front buffer, so it can be freed next time this function is called. */
windata->curr_bo = windata->next_bo;
/* Allow a modeset change for the first commit only. */
flags &= ~(DRM_MODE_ATOMIC_ALLOW_MODESET);
return ret;
}
/***************************************/
/* End of Atomic functions block */
/***************************************/
SDL_EGL_MakeCurrent_impl(KMSDRM)
#endif /* SDL_VIDEO_DRIVER_KMSDRM && SDL_VIDEO_OPENGL_EGL */
/* vi: set ts=4 sw=4 expandtab: */