overlay: Monitor per-ring context switch rate

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

overlay/gpu-perf.c

@@ -40,10 +40,12 @@
 
 #if defined(__i386__)
 #define rmb()           asm volatile("lock; addl $0,0(%%esp)" ::: "memory")
+#define wmb()           asm volatile("lock; addl $0,0(%%esp)" ::: "memory")
 #endif
 
 #if defined(__x86_64__)
 #define rmb()           asm volatile("lfence" ::: "memory")
+#define wmb()           asm volatile("sfence" ::: "memory")
 #endif
 
 #define N_PAGES 32
@@ -228,6 +230,14 @@ static int flip_complete(struct gpu_perf *gp, const void *event)
 	return 1;
 }
 
+static int ctx_switch(struct gpu_perf *gp, const void *event)
+{
+	const struct sample_event *sample = event;
+
+	gp->ctx_switch[sample->raw[1]]++;
+	return 1;
+}
+
 static int ring_sync(struct gpu_perf *gp, const void *event)
 {
 	const struct sample_event *sample = event;
@@ -293,6 +303,7 @@ void gpu_perf_init(struct gpu_perf *gp, unsigned flags)
 		perf_tracepoint_open(gp, "i915", "i915_gem_request_wait_end", wait_end);
 	perf_tracepoint_open(gp, "i915", "i915_flip_complete", flip_complete);
 	perf_tracepoint_open(gp, "i915", "i915_gem_ring_sync_to", ring_sync);
+	perf_tracepoint_open(gp, "i915", "i915_gem_ring_switch_context", ctx_switch);
 
 	if (gp->nr_events == 0) {
 		gp->error = "i915.ko tracepoints not available";
@@ -303,20 +314,19 @@ void gpu_perf_init(struct gpu_perf *gp, unsigned flags)
 		return;
 }
 
-static int process_sample(struct gpu_perf *gp,
+static int process_sample(struct gpu_perf *gp, int cpu,
 			  const struct perf_event_header *header)
 {
 	const struct sample_event *sample = (const struct sample_event *)header;
 	int n, update = 0;
 
 	/* hash me! */
-	for (n = 0; n < gp->nr_cpus * gp->nr_events; n++) {
+	for (n = 0; n < gp->nr_events; n++) {
-		if (gp->sample[n].id != sample->id)
+		int m = n * gp->nr_cpus + cpu;
+		if (gp->sample[m].id != sample->id)
 			continue;
 
-		update = 1;
+		update = gp->sample[m].func(gp, sample);
-		if (gp->sample[n].func)
-			update = gp->sample[n].func(gp, sample);
 		break;
 	}
 
@@ -380,13 +390,14 @@ int gpu_perf_update(struct gpu_perf *gp)
 			}
 
 			if (header->type == PERF_RECORD_SAMPLE)
-				update += process_sample(gp, header);
+				update += process_sample(gp, n, header);
 			tail += header->size;
 		}
 
 		if (wrap)
 			tail &= mask;
 		mmap->data_tail = tail;
+		wmb();
 	}
 
 	free(buffer);

overlay/gpu-perf.h

@@ -41,7 +41,9 @@ struct gpu_perf {
 		int (*func)(struct gpu_perf *, const void *);
 	} *sample;
 
-	int flip_complete[4];
+	unsigned flip_complete[MAX_RINGS];
+	unsigned ctx_switch[MAX_RINGS];
+
 	struct gpu_perf_comm {
 		struct gpu_perf_comm *next;
 		char name[256];

overlay/overlay.c

@@ -102,6 +102,7 @@ struct overlay_gpu_top {
 
 struct overlay_gpu_perf {
 	struct gpu_perf gpu_perf;
+	time_t show_ctx;
 };
 
 struct overlay_gpu_freq {
@@ -127,6 +128,8 @@ struct overlay_context {
 	cairo_t *cr;
 	int width, height;
 
+	time_t time;
+
 	struct overlay_gpu_top gpu_top;
 	struct overlay_gpu_perf gpu_perf;
 	struct overlay_gpu_freq gpu_freq;
@@ -270,6 +273,8 @@ static void init_gpu_perf(struct overlay_context *ctx,
 			  struct overlay_gpu_perf *gp)
 {
 	gpu_perf_init(&gp->gpu_perf, 0);
+
+	gp->show_ctx = 0;
 }
 
 static char *get_comm(pid_t pid, char *comm, int len)
@@ -310,6 +315,16 @@ static void show_gpu_perf(struct overlay_context *ctx, struct overlay_gpu_perf *
 	char buf[1024];
 	cairo_pattern_t *linear;
 	int x, y, y1, y2, n;
+	int has_ctx = 0;
+
+	gpu_perf_update(&gp->gpu_perf);
+
+	for (n = 4; n > 0; n--) {
+		if (gp->gpu_perf.ctx_switch[n-1]) {
+			has_ctx = n;
+			break;
+		}
+	}
 
 	cairo_rectangle(ctx->cr, ctx->width/2+HALF_PAD-.5, PAD-.5, ctx->width/2-SIZE_PAD+1, ctx->height/2-SIZE_PAD+1);
 	cairo_set_source_rgb(ctx->cr, .15, .15, .15);
@@ -326,12 +341,9 @@ static void show_gpu_perf(struct overlay_context *ctx, struct overlay_gpu_perf *
 		return;
 	}
 
-	gpu_perf_update(&gp->gpu_perf);
-
 	y = PAD + 12 - 2;
 	x = ctx->width/2 + HALF_PAD;
 
-
 	for (comm = gp->gpu_perf.comm; comm; comm = comm->next) {
 		int total;
 
@@ -369,6 +381,8 @@ static void show_gpu_perf(struct overlay_context *ctx, struct overlay_gpu_perf *
 		chart_draw(comm->user_data, ctx->cr);
 		y2 += 14;
 	}
+	if (has_ctx || gp->show_ctx)
+		y2 += 14;
 	y1 += -12 - 2;
 	y2 += 14 - 14 + 4;
 
@@ -465,6 +479,27 @@ skip_comm:
 	cairo_move_to(ctx->cr, x, y);
 	cairo_show_text(ctx->cr, buf);
 	y += 14;
+
+	cairo_set_source_rgba(ctx->cr, 1, 1, 1, 1);
+	cairo_move_to(ctx->cr, x, y);
+	if (has_ctx) {
+		int len = sprintf(buf, "Contexts:");
+		for (n = 0; n < has_ctx; n++)
+			len += sprintf(buf + len, "%s %d",
+				       n ? "," : "",
+				       gp->gpu_perf.ctx_switch[n]);
+
+		memset(gp->gpu_perf.ctx_switch, 0, sizeof(gp->gpu_perf.ctx_switch));
+		gp->show_ctx = ctx->time;
+
+		cairo_show_text(ctx->cr, buf);
+		y += 14;
+	} else if (gp->show_ctx) {
+		cairo_show_text(ctx->cr, "Contexts: 0");
+		y += 14;
+		if (ctx->time - gp->show_ctx > 10)
+			gp->show_ctx = 0;
+	}
 }
 
 static void init_gpu_freq(struct overlay_context *ctx,
@@ -841,6 +876,8 @@ int main(int argc, char **argv)
 
 	i = 0;
 	while (1) {
+		ctx.time = time(NULL);
+
 		ctx.cr = cairo_create(ctx.surface);
 		cairo_set_operator(ctx.cr, CAIRO_OPERATOR_CLEAR);
 		cairo_paint(ctx.cr);