From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Markus Fritsche <fritsche.markus@gmail.com>
Date: Fri, 8 May 2026 23:30:00 +0000
Subject: [PATCH] vo_dmabuf_wayland: explicit DMA_BUF_IOCTL_SYNC on import fds

V4L2 does not attach implicit fences (dma_resv) to CAPTURE buffers
on VIDIOC_DQBUF. When the buffer is forwarded to a Wayland compositor
that imports it via wl_dmabuf and samples in the GPU, the GPU may
read from physical memory before the producer's writes have flushed,
producing all-zero output (manifests as solid green for BT.601
limited-range YUV(0,0,0) -> RGB(0, 135, 0) on the consumer side).

Issue an explicit DMA_BUF_IOCTL_SYNC(SYNC_START|SYNC_RW) +
SYNC_END(SYNC_RW) round-trip on each unique dma_buf fd before
zwp_linux_buffer_params_v1_add(). This invokes the producer driver's
dma_buf_ops->begin_cpu_access / end_cpu_access, which on most ARM
SoCs flushes write buffers and synchronizes coherent memory before
the compositor's GPU import.

This is a userspace workaround. Root cause is the missing implicit
fence on V4L2 CAPTURE DQBUF and is being addressed upstream via
the vb2_dma_resv RFC.

Without this patch, on RK3566 (hantro VPU + Mali-G52 panfrost +
KDE Plasma 6 / KWin 6.6.4), `mpv --hwdec=vaapi --vo=dmabuf-wayland`
shows solid green frames for all hardware-decoded content. With
this patch, decoded frames are presented correctly.

Signed-off-by: Markus Fritsche <fritsche.markus@gmail.com>
---
diff --git a/video/out/vo_dmabuf_wayland.c b/video/out/vo_dmabuf_wayland.c
index 6b7c511..16e3d18 100644
--- a/video/out/vo_dmabuf_wayland.c
+++ b/video/out/vo_dmabuf_wayland.c
@@ -27,6 +27,12 @@
 #include <va/va_drmcommon.h>
 #endif
 
+/* fourier patch: explicit dma_buf cache sync workaround for missing
+ * implicit-fence on V4L2 stateless CAPTURE buffers. Applies to both
+ * VAAPI and DRMPrime import paths. */
+#include <linux/dma-buf.h>
+#include <sys/ioctl.h>
+
 #include "gpu/hwdec.h"
 #include "gpu/video.h"
 #include "mpv_talloc.h"
@@ -205,6 +211,14 @@ static void vaapi_dmabuf_importer(struct buffer *buf, struct mp_image *src,
         buf->drm_format = 0;
         goto done;
     }
+    /* fourier patch: explicit cache coherency sync on each dma_buf fd
+     * before submitting to the compositor. See top-of-file comment. */
+    for (int obj_no = 0; obj_no < desc.num_objects; obj_no++) {
+        struct dma_buf_sync sync = { .flags = DMA_BUF_SYNC_START | DMA_BUF_SYNC_RW };
+        (void)ioctl(desc.objects[obj_no].fd, DMA_BUF_IOCTL_SYNC, &sync);
+        sync.flags = DMA_BUF_SYNC_END | DMA_BUF_SYNC_RW;
+        (void)ioctl(desc.objects[obj_no].fd, DMA_BUF_IOCTL_SYNC, &sync);
+    }
     for (int plane_no = 0; plane_no < desc.layers[layer_no].num_planes; ++plane_no) {
         int object = desc.layers[layer_no].object_index[plane_no];
         uint64_t modifier = desc.objects[object].drm_format_modifier;
@@ -258,6 +272,16 @@ static void drmprime_dmabuf_importer(struct buffer *buf, struct mp_image *src,
         return;
 
     buf->id = drmprime_surface_id(src);
+
+    /* fourier patch: explicit cache coherency sync on each dma_buf fd
+     * before submitting to the compositor. See top-of-file comment. */
+    for (int obj_no = 0; obj_no < desc->nb_objects; obj_no++) {
+        struct dma_buf_sync sync = { .flags = DMA_BUF_SYNC_START | DMA_BUF_SYNC_RW };
+        (void)ioctl(desc->objects[obj_no].fd, DMA_BUF_IOCTL_SYNC, &sync);
+        sync.flags = DMA_BUF_SYNC_END | DMA_BUF_SYNC_RW;
+        (void)ioctl(desc->objects[obj_no].fd, DMA_BUF_IOCTL_SYNC, &sync);
+    }
+
     for (layer_no = 0; layer_no < desc->nb_layers; layer_no++) {
         AVDRMLayerDescriptor layer = desc->layers[layer_no];
 
-- 
2.51.0