radeonsi: avoid syncing the driver thread in si_fence_finish

It is really only required when we need to flush for deferred fences.

Reviewed-by: Marek Olšák <[email protected]>

3 files changed, 49 insertions, 37 deletions

diff --git a/src/gallium/auxiliary/util/u_threaded_context.h b/src/gallium/auxiliary/util/u_threaded_context.h
index b2d904569e1..34089561f34 100644
--- a/src/gallium/auxiliary/util/u_threaded_context.h
+++ b/src/gallium/auxiliary/util/u_threaded_context.h
@@ -407,6 +407,14 @@ threaded_transfer(struct pipe_transfer *transfer)
    return (struct threaded_transfer*)transfer;
 }
 
+static inline struct pipe_context *
+threaded_context_unwrap_unsync(struct pipe_context *pipe)
+{
+   if (!pipe || !pipe->priv)
+      return pipe;
+   return (struct pipe_context*)pipe->priv;
+}
+
 static inline void
 tc_unflushed_batch_token_reference(struct tc_unflushed_batch_token **dst,
                                    struct tc_unflushed_batch_token *src)
diff --git a/src/gallium/drivers/radeonsi/si_fence.c b/src/gallium/drivers/radeonsi/si_fence.c
index ff1800ce785..5163d652c83 100644
--- a/src/gallium/drivers/radeonsi/si_fence.c
+++ b/src/gallium/drivers/radeonsi/si_fence.c
@@ -188,12 +188,8 @@ static boolean si_fence_finish(struct pipe_screen *screen,
 {
 	struct radeon_winsys *rws = ((struct r600_common_screen*)screen)->ws;
 	struct si_multi_fence *rfence = (struct si_multi_fence *)fence;
-	struct r600_common_context *rctx;
 	int64_t abs_timeout = os_time_get_absolute_timeout(timeout);
 
-	ctx = threaded_context_unwrap_sync(ctx);
-	rctx = ctx ? (struct r600_common_context*)ctx : NULL;
-
 	if (!util_queue_fence_is_signalled(&rfence->ready)) {
 		if (!timeout)
 			return false;
@@ -245,41 +241,46 @@ static boolean si_fence_finish(struct pipe_screen *screen,
 	}
 
 	/* Flush the gfx IB if it hasn't been flushed yet. */
-	if (rctx &&
-	    rfence->gfx_unflushed.ctx == rctx &&
-	    rfence->gfx_unflushed.ib_index == rctx->num_gfx_cs_flushes) {
-		/* Section 4.1.2 (Signaling) of the OpenGL 4.6 (Core profile)
-		 * spec says:
-		 *
-		 *    "If the sync object being blocked upon will not be
-		 *     signaled in finite time (for example, by an associated
-		 *     fence command issued previously, but not yet flushed to
-		 *     the graphics pipeline), then ClientWaitSync may hang
-		 *     forever. To help prevent this behavior, if
-		 *     ClientWaitSync is called and all of the following are
-		 *     true:
-		 *
-		 *     * the SYNC_FLUSH_COMMANDS_BIT bit is set in flags,
-		 *     * sync is unsignaled when ClientWaitSync is called,
-		 *     * and the calls to ClientWaitSync and FenceSync were
-		 *       issued from the same context,
-		 *
-		 *     then the GL will behave as if the equivalent of Flush
-		 *     were inserted immediately after the creation of sync."
-		 *
-		 * This means we need to flush for such fences even when we're
-		 * not going to wait.
-		 */
-		rctx->gfx.flush(rctx, timeout ? 0 : RADEON_FLUSH_ASYNC, NULL);
-		rfence->gfx_unflushed.ctx = NULL;
+	if (ctx && rfence->gfx_unflushed.ctx) {
+		struct si_context *sctx;
+
+		sctx = (struct si_context *)threaded_context_unwrap_unsync(ctx);
+		if (rfence->gfx_unflushed.ctx == &sctx->b &&
+		    rfence->gfx_unflushed.ib_index == sctx->b.num_gfx_cs_flushes) {
+			/* Section 4.1.2 (Signaling) of the OpenGL 4.6 (Core profile)
+			 * spec says:
+			 *
+			 *    "If the sync object being blocked upon will not be
+			 *     signaled in finite time (for example, by an associated
+			 *     fence command issued previously, but not yet flushed to
+			 *     the graphics pipeline), then ClientWaitSync may hang
+			 *     forever. To help prevent this behavior, if
+			 *     ClientWaitSync is called and all of the following are
+			 *     true:
+			 *
+			 *     * the SYNC_FLUSH_COMMANDS_BIT bit is set in flags,
+			 *     * sync is unsignaled when ClientWaitSync is called,
+			 *     * and the calls to ClientWaitSync and FenceSync were
+			 *       issued from the same context,
+			 *
+			 *     then the GL will behave as if the equivalent of Flush
+			 *     were inserted immediately after the creation of sync."
+			 *
+			 * This means we need to flush for such fences even when we're
+			 * not going to wait.
+			 */
+			threaded_context_unwrap_sync(ctx);
+			sctx->b.gfx.flush(&sctx->b, timeout ? 0 : RADEON_FLUSH_ASYNC, NULL);
+			rfence->gfx_unflushed.ctx = NULL;
 
-		if (!timeout)
-			return false;
+			if (!timeout)
+				return false;
 
-		/* Recompute the timeout after all that. */
-		if (timeout && timeout != PIPE_TIMEOUT_INFINITE) {
-			int64_t time = os_time_get_nano();
-			timeout = abs_timeout > time ? abs_timeout - time : 0;
+			/* Recompute the timeout after all that. */
+			if (timeout && timeout != PIPE_TIMEOUT_INFINITE) {
+				int64_t time = os_time_get_nano();
+				timeout = abs_timeout > time ? abs_timeout - time : 0;
+			}
 		}
 	}
 
diff --git a/src/gallium/drivers/radeonsi/si_hw_context.c b/src/gallium/drivers/radeonsi/si_hw_context.c
index 15234d72725..1903cf8c978 100644
--- a/src/gallium/drivers/radeonsi/si_hw_context.c
+++ b/src/gallium/drivers/radeonsi/si_hw_context.c
@@ -129,6 +129,9 @@ void si_context_gfx_flush(void *context, unsigned flags,
 	ws->cs_flush(cs, flags, &ctx->b.last_gfx_fence);
 	if (fence)
 		ws->fence_reference(fence, ctx->b.last_gfx_fence);
+
+	/* This must be after cs_flush returns, since the context's API
+	 * thread can concurrently read this value in si_fence_finish. */
 	ctx->b.num_gfx_cs_flushes++;
 
 	/* Check VM faults if needed. */