dri: make KGSL presentation waits opt-in on native Wayland

docs/envvars.rst

@@ -1081,6 +1081,21 @@ Hasvk driver environment variables
 DRI environment variables
 -------------------------
 
+.. envvar:: MESA_DRI3_DISABLE_PRESENT_WAIT_FENCE
+
+   Controls the X11 Present wait-fence bridge and the shared KGSL
+   presentation-completion wait. For native Wayland EGL contexts the
+   default is ``true``, restoring the earlier nonblocking presentation
+   behavior. For X11 contexts, including XWayland and Termux:X11, the
+   default is ``false``, preserving the Present wait-fence bridge.
+
+   Set ``1`` to disable both presentation mechanisms, or ``0`` to keep
+   them enabled, including the conservative KGSL completion wait on
+   native Wayland. This does not disable resource submission, OpenGL
+   synchronization objects, or EGL native fences. Disabling the wait
+   does not add an asynchronous fence handoff to the compositor; visual
+   correctness still depends on the existing buffer synchronization.
+
 .. envvar:: DRI_NO_MSAA
 
    disable MSAA for GLX/EGL MSAA visuals

src/egl/drivers/dri2/egl_dri2.c

@@ -1330,6 +1330,15 @@ dri2_create_context(_EGLDisplay *disp, _EGLConfig *conf,
    if (!dri2_ctx->dri_context)
       goto cleanup;
 
+   if (disp->Platform == _EGL_PLATFORM_WAYLAND) {
+      /* Native Wayland has no X11 Present wait-fence bridge. Keep its
+       * pre-bridge default, with the same environment override as DRI3.
+       * This is a context policy, so later attachment flushes agree too.
+       */
+      dri_disable_present_wait_fence(dri2_ctx->dri_context,
+         debug_get_bool_option("MESA_DRI3_DISABLE_PRESENT_WAIT_FENCE", true));
+   }
+
    mtx_unlock(&dri2_dpy->lock);
 
    return &dri2_ctx->base;

src/gallium/drivers/freedreno/freedreno_context.c

@@ -147,6 +147,9 @@ fd_set_context_param(struct pipe_context *pctx, enum pipe_context_param param,
    case PIPE_CONTEXT_PARAM_EXPLICIT_PRESENT_FENCE:
       ctx->explicit_present_fence = value;
       break;
+   case PIPE_CONTEXT_PARAM_DISABLE_PRESENT_WAIT_FENCE:
+      ctx->disable_present_wait_fence = value;
+      break;
    default:
       break;
    }

src/gallium/drivers/freedreno/freedreno_context.h

@@ -362,6 +362,7 @@ struct fd_context {
     * the window-system presentation path.
     */
    bool explicit_present_fence;
+   bool disable_present_wait_fence;
 
    /*
     * Counter to keep track of batch's most recent update.  Ie. the batch with

src/gallium/drivers/freedreno/freedreno_resource.c

@@ -770,9 +770,11 @@ fd_flush_resource(struct pipe_context *pctx, struct pipe_resource *prsc)
     * If the window-system path is not carrying the next submission's native
     * fence explicitly, wait here before handing the resource to an
     * implicit-sync consumer.  The explicit Present-fence path avoids this
-    * per-frame CPU stall.
+    * per-frame CPU stall. The compatibility opt-out restores the previous
+    * window-system behavior without changing resource submission or API fences.
     */
-   if (ctx->screen->is_kgsl && !ctx->explicit_present_fence)
+   if (ctx->screen->is_kgsl && !ctx->explicit_present_fence &&
+       !ctx->disable_present_wait_fence)
       fd_resource_wait(ctx, rsc, FD_BO_PREP_READ);
 }
 

src/gallium/frontends/dri/dri_context.c

@@ -32,6 +32,7 @@
 #include "dri_screen.h"
 #include "dri_drawable.h"
 #include "dri_context.h"
+#include "dri_util.h"
 #include "frontend/drm_driver.h"
 
 #include "pipe/p_context.h"
@@ -42,6 +43,16 @@
 #include "util/u_memory.h"
 #include "util/u_debug.h"
 
+void
+dri_disable_present_wait_fence(struct dri_context *ctx, bool disable)
+{
+   struct pipe_context *pipe = ctx->st->pipe;
+   if (pipe->set_context_param)
+      pipe->set_context_param(pipe,
+                              PIPE_CONTEXT_PARAM_DISABLE_PRESENT_WAIT_FENCE,
+                              disable);
+}
+
 struct dri_context *
 dri_create_context(struct dri_screen *screen,
                    struct st_context_attribs *attribs,
@@ -136,6 +147,9 @@ dri_create_context(struct dri_screen *screen,
    if (!thread_safe)
       enable_glthread = false;
 
+   dri_disable_present_wait_fence(ctx, debug_get_bool_option(
+      "MESA_DRI3_DISABLE_PRESENT_WAIT_FENCE", false));
+
    /* Do this last. */
    if (enable_glthread)
       _mesa_glthread_init(ctx->st->ctx);

src/gallium/frontends/dri/dri_util.h

@@ -139,6 +139,9 @@ dri_query_renderer_integer(struct dri_screen *_screen, int param,
 
 PUBLIC void
 dri_flush_drawable(struct dri_drawable *dPriv);
+/* Set the window-system policy before the context is made current. */
+PUBLIC void
+dri_disable_present_wait_fence(struct dri_context *ctx, bool disable);
 PUBLIC void
 dri_flush(struct dri_context *cPriv,
           struct dri_drawable *dPriv,

src/gallium/include/pipe/p_defines.h

@@ -1269,6 +1269,11 @@ enum pipe_context_param
     */
    PIPE_CONTEXT_PARAM_EXPLICIT_PRESENT_FENCE,
 
+   /* Compatibility policy for window-system presentation. Disable only the
+    * driver's additional CPU completion wait, not submission or API fences.
+    */
+   PIPE_CONTEXT_PARAM_DISABLE_PRESENT_WAIT_FENCE,
+
    /* The minification/magnification switch-over point, the constant c of the
     * GL and ES specs. Value 1 selects c = 0.5, value 0 selects c = 0.
     *

src/gallium/targets/dri/dri.sym.in

@@ -27,6 +27,7 @@
 		dri_query_renderer_string;
 		dri_query_renderer_integer;
 		dri_flush_drawable;
+		dri_disable_present_wait_fence;
 		dri_flush;
 		dri_invalidate_drawable;
 		dri2GalliumConfigQuery*;