output: add max_render_time

include/sway/commands.h

@@ -271,6 +271,7 @@ sway_cmd output_cmd_background;
 sway_cmd output_cmd_disable;
 sway_cmd output_cmd_dpms;
 sway_cmd output_cmd_enable;
+sway_cmd output_cmd_max_render_time;
 sway_cmd output_cmd_mode;
 sway_cmd output_cmd_position;
 sway_cmd output_cmd_scale;

include/sway/config.h

@@ -212,6 +212,7 @@ struct output_config {
 	float scale;
 	int32_t transform;
 	enum wl_output_subpixel subpixel;
+	int max_render_time; // In milliseconds
 
 	char *background;
 	char *background_option;

include/sway/output.h

@@ -51,6 +51,11 @@ struct sway_output {
 	struct {
 		struct wl_signal destroy;
 	} events;
+
+	struct timespec last_presentation;
+	uint32_t refresh_nsec;
+	int max_render_time; // In milliseconds
+	struct wl_event_source *repaint_timer;
 };
 
 struct sway_output *output_create(struct wlr_output *wlr_output);
@@ -71,6 +76,8 @@ typedef void (*sway_surface_iterator_func_t)(struct sway_output *output,
 	struct wlr_surface *surface, struct wlr_box *box, float rotation,
 	void *user_data);
 
+int output_repaint_timer_handler(void *data);
+
 void output_damage_whole(struct sway_output *output);
 
 void output_damage_surface(struct sway_output *output, double ox, double oy,

sway/commands/output.c

@@ -12,6 +12,7 @@ static struct cmd_handler output_handlers[] = {
 	{ "disable", output_cmd_disable },
 	{ "dpms", output_cmd_dpms },
 	{ "enable", output_cmd_enable },
+	{ "max_render_time", output_cmd_max_render_time },
 	{ "mode", output_cmd_mode },
 	{ "pos", output_cmd_position },
 	{ "position", output_cmd_position },

sway/commands/output/max_render_time.c

@@ -0,0 +1,28 @@
+#include <strings.h>
+#include "sway/commands.h"
+#include "sway/config.h"
+
+struct cmd_results *output_cmd_max_render_time(int argc, char **argv) {
+	if (!config->handler_context.output_config) {
+		return cmd_results_new(CMD_FAILURE, "Missing output config");
+	}
+	if (!argc) {
+		return cmd_results_new(CMD_INVALID, "Missing max render time argument.");
+	}
+
+	int max_render_time;
+	if (!strcmp(*argv, "off")) {
+		max_render_time = 0;
+	} else {
+		char *end;
+		max_render_time = strtol(*argv, &end, 10);
+		if (*end || max_render_time <= 0) {
+			return cmd_results_new(CMD_INVALID, "Invalid max render time.");
+		}
+	}
+	config->handler_context.output_config->max_render_time = max_render_time;
+
+	config->handler_context.leftovers.argc = argc - 1;
+	config->handler_context.leftovers.argv = argv + 1;
+	return NULL;
+}

sway/config/output.c

@@ -47,6 +47,7 @@ struct output_config *new_output_config(const char *name) {
 	oc->scale = -1;
 	oc->transform = -1;
 	oc->subpixel = WL_OUTPUT_SUBPIXEL_UNKNOWN;
+	oc->max_render_time = -1;
 	return oc;
 }
 
@@ -81,6 +82,9 @@ void merge_output_config(struct output_config *dst, struct output_config *src) {
 	if (src->transform != -1) {
 		dst->transform = src->transform;
 	}
+	if (src->max_render_time != -1) {
+		dst->max_render_time = src->max_render_time;
+	}
 	if (src->background) {
 		free(dst->background);
 		dst->background = strdup(src->background);
@@ -153,11 +157,12 @@ static void merge_id_on_name(struct output_config *oc) {
 			list_add(config->output_configs, ion_oc);
 			sway_log(SWAY_DEBUG, "Generated id on name output config \"%s\""
 				" (enabled: %d) (%dx%d@%fHz position %d,%d scale %f "
-				"transform %d) (bg %s %s) (dpms %d)", ion_oc->name,
+				"transform %d) (bg %s %s) (dpms %d) (max render time: %d)",
-				ion_oc->enabled, ion_oc->width, ion_oc->height,
+				ion_oc->name, ion_oc->enabled, ion_oc->width, ion_oc->height,
 				ion_oc->refresh_rate, ion_oc->x, ion_oc->y, ion_oc->scale,
 				ion_oc->transform, ion_oc->background,
-				ion_oc->background_option, ion_oc->dpms_state);
+				ion_oc->background_option, ion_oc->dpms_state,
+				ion_oc->max_render_time);
 		}
 	}
 	free(id_on_name);
@@ -197,10 +202,12 @@ struct output_config *store_output_config(struct output_config *oc) {
 	}
 
 	sway_log(SWAY_DEBUG, "Config stored for output %s (enabled: %d) (%dx%d@%fHz "
-		"position %d,%d scale %f subpixel %s transform %d) (bg %s %s) (dpms %d)",
+		"position %d,%d scale %f subpixel %s transform %d) (bg %s %s) (dpms %d) "
+		"(max render time: %d)",
 		oc->name, oc->enabled, oc->width, oc->height, oc->refresh_rate,
 		oc->x, oc->y, oc->scale, sway_wl_output_subpixel_to_string(oc->subpixel),
-		oc->transform, oc->background, oc->background_option, oc->dpms_state);
+		oc->transform, oc->background, oc->background_option, oc->dpms_state,
+		oc->max_render_time);
 
 	return oc;
 }
@@ -325,6 +332,12 @@ bool apply_output_config(struct output_config *oc, struct sway_output *output) {
 		wlr_output_enable(wlr_output, false);
 	}
 
+	if (oc && oc->max_render_time >= 0) {
+		sway_log(SWAY_DEBUG, "Set %s max render time to %d",
+			oc->name, oc->max_render_time);
+		output->max_render_time = oc->max_render_time;
+	}
+
 	return true;
 }
 
@@ -343,6 +356,7 @@ static void default_output_config(struct output_config *oc,
 	oc->subpixel = output->detected_subpixel;
 	oc->transform = WL_OUTPUT_TRANSFORM_NORMAL;
 	oc->dpms_state = DPMS_ON;
+	oc->max_render_time = 0;
 }
 
 static struct output_config *get_output_config(char *identifier,
@@ -396,10 +410,11 @@ static struct output_config *get_output_config(char *identifier,
 
 		sway_log(SWAY_DEBUG, "Generated output config \"%s\" (enabled: %d)"
 			" (%dx%d@%fHz position %d,%d scale %f transform %d) (bg %s %s)"
-			" (dpms %d)", result->name, result->enabled, result->width,
+			" (dpms %d) (max render time: %d)", result->name, result->enabled,
-			result->height, result->refresh_rate, result->x, result->y,
+			result->width, result->height, result->refresh_rate,
-			result->scale, result->transform, result->background,
+			result->x, result->y, result->scale, result->transform,
-			result->background_option, result->dpms_state);
+			result->background, result->background_option, result->dpms_state,
+			result->max_render_time);
 	} else if (oc_name) {
 		// No identifier config, just return a copy of the name config
 		free(result->name);

sway/desktop/output.c

@@ -480,19 +480,13 @@ static bool scan_out_fullscreen_view(struct sway_output *output,
 	return wlr_output_commit(wlr_output);
 }
 
-static void damage_handle_frame(struct wl_listener *listener, void *data) {
+int output_repaint_timer_handler(void *data) {
-	struct sway_output *output =
+	struct sway_output *output = data;
-		wl_container_of(listener, output, damage_frame);
+	output->wlr_output->block_idle_frame = false;
-	if (!output->enabled || !output->wlr_output->enabled) {
-		return;
-	}
-
-	struct timespec now;
-	clock_gettime(CLOCK_MONOTONIC, &now);
 
 	struct sway_workspace *workspace = output->current.active_workspace;
 	if (workspace == NULL) {
-		return;
+		return 0;
 	}
 
 	struct sway_container *fullscreen_con = root->fullscreen_global;
@@ -515,7 +509,7 @@ static void damage_handle_frame(struct wl_listener *listener, void *data) {
 		last_scanned_out = scanned_out;
 
 		if (scanned_out) {
-			goto frame_done;
+			return 0;
 		}
 	}
 
@@ -524,17 +518,82 @@ static void damage_handle_frame(struct wl_listener *listener, void *data) {
 	pixman_region32_init(&damage);
 	if (!wlr_output_damage_attach_render(output->damage,
 			&needs_frame, &damage)) {
-		return;
+		return 0;
 	}
 
 	if (needs_frame) {
+		struct timespec now;
+		clock_gettime(CLOCK_MONOTONIC, &now);
+
 		output_render(output, &now, &damage);
 	}
 
 	pixman_region32_fini(&damage);
 
-frame_done:
+	return 0;
+}
+
+static void damage_handle_frame(struct wl_listener *listener, void *data) {
+	struct sway_output *output =
+		wl_container_of(listener, output, damage_frame);
+	if (!output->enabled || !output->wlr_output->enabled) {
+		return;
+	}
+
+	// Compute predicted milliseconds until the next refresh. It's used for
+	// delaying both output rendering and surface frame callbacks.
+	int msec_until_refresh = 0;
+
+	if (output->max_render_time != 0) {
+		struct timespec now;
+		clockid_t presentation_clock
+			= wlr_backend_get_presentation_clock(server.backend);
+		clock_gettime(presentation_clock, &now);
+
+		const long NSEC_IN_SECONDS = 1000000000;
+		struct timespec predicted_refresh = output->last_presentation;
+		predicted_refresh.tv_nsec += output->refresh_nsec % NSEC_IN_SECONDS;
+		predicted_refresh.tv_sec += output->refresh_nsec / NSEC_IN_SECONDS;
+		if (predicted_refresh.tv_nsec >= NSEC_IN_SECONDS) {
+			predicted_refresh.tv_sec += 1;
+			predicted_refresh.tv_nsec -= NSEC_IN_SECONDS;
+		}
+
+		// If the predicted refresh time is before the current time then
+		// there's no point in delaying.
+		//
+		// We only check tv_sec because if the predicted refresh time is less
+		// than a second before the current time, then msec_until_refresh will
+		// end up slightly below zero, which will effectively disable the delay
+		// without potential disasterous negative overflows that could occur if
+		// tv_sec was not checked.
+		if (predicted_refresh.tv_sec >= now.tv_sec) {
+			long nsec_until_refresh
+				= (predicted_refresh.tv_sec - now.tv_sec) * NSEC_IN_SECONDS
+					+ (predicted_refresh.tv_nsec - now.tv_nsec);
+
+			// We want msec_until_refresh to be conservative, that is, floored.
+			// If we have 7.9 msec until refresh, we better compute the delay
+			// as if we had only 7 msec, so that we don't accidentally delay
+			// more than necessary and miss a frame.
+			msec_until_refresh = nsec_until_refresh / 1000000;
+		}
+	}
+
+	int delay = msec_until_refresh - output->max_render_time;
+
+	// If the delay is less than 1 millisecond (which is the least we can wait)
+	// then just render right away.
+	if (delay < 1) {
+		output_repaint_timer_handler(output);
+	} else {
+		output->wlr_output->block_idle_frame = true;
+		wl_event_source_timer_update(output->repaint_timer, delay);
+	}
+
 	// Send frame done to all visible surfaces
+	struct timespec now;
+	clock_gettime(CLOCK_MONOTONIC, &now);
 	send_frame_done(output, &now);
 }
 
@@ -768,6 +827,9 @@ static void handle_present(struct wl_listener *listener, void *data) {
 		return;
 	}
 
+	output->last_presentation = *output_event->when;
+	output->refresh_nsec = output_event->refresh;
+
 	struct wlr_presentation_event event = {
 		.output = output->wlr_output,
 		.tv_sec = (uint64_t)output_event->when->tv_sec,
@@ -806,6 +868,9 @@ void handle_new_output(struct wl_listener *listener, void *data) {
 	wl_signal_add(&output->damage->events.destroy, &output->damage_destroy);
 	output->damage_destroy.notify = damage_handle_destroy;
 
+	output->repaint_timer = wl_event_loop_add_timer(server->wl_event_loop,
+		output_repaint_timer_handler, output);
+
 	struct output_config *oc = find_output_config(output);
 	if (!oc || oc->enabled) {
 		output_enable(output, oc);

sway/meson.build

@@ -175,6 +175,7 @@ sway_sources = files(
 	'commands/output/disable.c',
 	'commands/output/dpms.c',
 	'commands/output/enable.c',
+	'commands/output/max_render_time.c',
 	'commands/output/mode.c',
 	'commands/output/position.c',
 	'commands/output/scale.c',

sway/sway-output.5.scd

@@ -107,6 +107,30 @@ must be separated by one space. For example:
 	Enables or disables the specified output via DPMS. To turn an output off
 	(ie. blank the screen but keep workspaces as-is), one can set DPMS to off.
 
+*output* <name> max_render_time off|<msec>
+	When set to a positive number of milliseconds, enables delaying output
+	rendering to reduce latency. The rendering is delayed in such a way as
+	to leave the specified number of milliseconds before the next
+	presentation for rendering.
+
+	The output rendering normally takes place immediately after a
+	presentation (vblank, buffer flip, etc.) and the frame callbacks are
+	sent to surfaces immediately after the rendering to give surfaces the
+	most time to draw their next frame. This results in slightly below 2
+	frames of latency between the surface rendering and committing new
+	contents, and the contents being shown on screen, on average. When the
+	output rendering is delayed, the frame callbacks are sent immediately
+	after presentation, and the surfaces have a small timespan (1 /
+	(refresh rate) - max_render_time) to render and commit new contents to
+	be shown on the next presentation, resulting in below 1 frame of
+	latency.
+
+	To set this up for optimal latency:
+	. Launch some _full-screen_ application that renders continuously, like
+	  *glxgears*.
+	. Start with *max_render_time 1*. Increment by *1* if you see frame
+	  drops.
+
 # SEE ALSO
 
 *sway*(5) *sway-input*(5)

sway/tree/output.c

@@ -242,6 +242,7 @@ void output_destroy(struct sway_output *output) {
 	}
 	list_free(output->workspaces);
 	list_free(output->current.workspaces);
+	wl_event_source_remove(output->repaint_timer);
 	free(output);
 }