overlay: Track requests per-process

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

overlay/gpu-perf.c

@@ -30,6 +30,8 @@ struct sample_event {
 	uint64_t time;
 	uint64_t id;
 	uint32_t raw_size;
+	uint32_t raw_hdr0;
+	uint32_t raw_hdr1;
 	uint32_t raw[0];
 };
 
@@ -117,7 +119,6 @@ static int perf_tracepoint_open(struct gpu_perf *gp,
 
 		if (gp->nr_events)
 			ioctl(fd[n], PERF_EVENT_IOC_SET_OUTPUT, gp->fd[n]);
-
 	}
 
 	gp->nr_events++;
@@ -149,8 +150,53 @@ err:
 	return EINVAL;
 }
 
+static char *get_comm(pid_t pid, char *comm, int len)
+{
+	char filename[1024];
+	int fd;
+
+	*comm = '\0';
+	snprintf(filename, sizeof(filename), "/proc/%d/comm", pid);
+
+	fd = open(filename, 0);
+	if (fd >= 0) {
+		len = read(fd, comm, len-1);
+		if (len >= 0)
+			comm[len-1] = '\0';
+		close(fd);
+	}
+
+	return comm;
+}
+
+static int request_add(struct gpu_perf *gp, const void *event)
+{
+	const struct sample_event *sample = event;
+	struct gpu_perf_comm *comm;
+
+	for (comm = gp->comm; comm != NULL; comm = comm->next) {
+		if (comm->pid == sample->pid)
+			break;
+	}
+	if (comm == NULL) {
+		comm = malloc(sizeof(*comm));
+		if (comm == NULL)
+			return 0;
+
+		comm->next = gp->comm;
+		gp->comm = comm;
+		get_comm(sample->pid, comm->name, sizeof(comm->name));
+		comm->pid = sample->pid;
+		memset(comm->nr_requests, 0, sizeof(comm->nr_requests));
+	}
+
+	comm->nr_requests[sample->raw[1]]++;
+	return 1;
+}
+
 static int seqno_start(struct gpu_perf *gp, const void *event)
 {
+	printf ("seqno_start\n");
 	return 0;
 }
 
@@ -171,9 +217,9 @@ void gpu_perf_init(struct gpu_perf *gp, unsigned flags)
 	gp->nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 	gp->page_size = getpagesize();
 
+	perf_tracepoint_open(gp, "i915", "i915_gem_request_add", request_add);
 	if (perf_tracepoint_open(gp, "i915", "i915_gem_ring_complete", seqno_end) == 0)
 		perf_tracepoint_open(gp, "i915", "i915_gem_ring_dispatch", seqno_start);
-
 	perf_tracepoint_open(gp, "i915", "i915_flip_complete", flip_complete);
 
 	if (gp->nr_events == 0)
@@ -218,13 +264,18 @@ int gpu_perf_update(struct gpu_perf *gp)
 		struct perf_event_mmap_page *mmap = gp->map[n];
 		const uint8_t *data;
 		uint64_t head, tail;
+		int wrap = 0;
 
 		tail = mmap->data_tail;
 		head = mmap->data_head;
 		rmb();
 
-		if (head < tail)
+		if (head < tail) {
+			wrap = 1;
+			tail &= mask;
+			head &= mask;
 			head += size;
+		}
 
 		data = (uint8_t *)mmap + gp->page_size;
 		while (head - tail >= sizeof (struct perf_event_header)) {
@@ -259,7 +310,9 @@ int gpu_perf_update(struct gpu_perf *gp)
 			tail += header->size;
 		}
 
-		mmap->data_tail = tail & mask;
+		if (wrap)
+			tail &= mask;
+		mmap->data_tail = tail;
 	}
 
 	free(buffer);

overlay/gpu-perf.h

@@ -12,6 +12,12 @@ struct gpu_perf {
 	} *sample;
 
 	int flip_complete;
+	struct gpu_perf_comm {
+		struct gpu_perf_comm *next;
+		char name[256];
+		pid_t pid;
+		int nr_requests[4];
+	} *comm;
 };
 
 void gpu_perf_init(struct gpu_perf *gp, unsigned flags);

overlay/overlay.c

@@ -1,4 +1,3 @@
-#include <X11/Xlib.h>
 #include <sys/types.h>
 #include <sys/mman.h>
 #include <cairo.h>
@@ -7,6 +6,7 @@
 #include <stdbool.h>
 #include <string.h>
 #include <unistd.h>
+#include <fcntl.h>
 #include <errno.h>
 
 #include "overlay.h"
@@ -52,6 +52,11 @@ static void overlay_hide(cairo_surface_t *surface)
 }
 #endif
 
+struct overlay_context {
+	cairo_t *cr;
+	int last_y;
+};
+
 struct overlay_gpu_top {
 	struct gpu_top gpu_top;
 	struct chart busy[MAX_RINGS];
@@ -100,7 +105,7 @@ static void init_gpu_top(struct overlay_gpu_top *gt,
 	}
 }
 
-static void show_gpu_top(cairo_t *cr, struct overlay_gpu_top *gt)
+static void show_gpu_top(struct overlay_context *ctx, struct overlay_gpu_top *gt)
 {
 	int y, n, update;
 
@@ -109,16 +114,16 @@ static void show_gpu_top(cairo_t *cr, struct overlay_gpu_top *gt)
 		if (update)
 			chart_add_sample(&gt->wait[n],
 					 gt->gpu_top.ring[n].u.u.wait + gt->gpu_top.ring[n].u.u.sema);
-		chart_draw(&gt->wait[n], cr);
+		chart_draw(&gt->wait[n], ctx->cr);
 	}
 	for (n = 0; n < gt->gpu_top.num_rings; n++) {
 		if (update)
 			chart_add_sample(&gt->busy[n],
 					 gt->gpu_top.ring[n].u.u.busy);
-		chart_draw(&gt->busy[n], cr);
+		chart_draw(&gt->busy[n], ctx->cr);
 	}
 
-	cairo_set_source_rgb(cr, 1, 1, 1);
+	cairo_set_source_rgb(ctx->cr, 1, 1, 1);
 
 	y = 12;
 	for (n = 0; n < gt->gpu_top.num_rings; n++) {
@@ -135,10 +140,12 @@ static void show_gpu_top(cairo_t *cr, struct overlay_gpu_top *gt)
 			len += sprintf(txt + len, ", %d%% sema",
 				       gt->gpu_top.ring[n].u.u.sema);
 
-		cairo_move_to(cr, 12, y);
-		cairo_show_text(cr, txt);
+		cairo_move_to(ctx->cr, 12, y);
+		cairo_show_text(ctx->cr, txt);
 		y += 14;
 	}
+
+	ctx->last_y = 112;
 }
 
 struct overlay_gpu_perf {
@@ -146,29 +153,75 @@ struct overlay_gpu_perf {
 };
 
 static void init_gpu_perf(struct overlay_gpu_perf *gp,
-			 cairo_surface_t *surface)
+			  cairo_surface_t *surface)
 {
 	gpu_perf_init(&gp->gpu_perf, 0);
 }
 
-static void show_gpu_perf(cairo_t *cr, struct overlay_gpu_perf *gp)
+static char *get_comm(pid_t pid, char *comm, int len)
 {
+	char filename[1024];
+	int fd;
+
+	*comm = '\0';
+	snprintf(filename, sizeof(filename), "/proc/%d/comm", pid);
+
+	fd = open(filename, 0);
+	if (fd >= 0) {
+		len = read(fd, comm, len-1);
+		if (len >= 0)
+			comm[len-1] = '\0';
+		close(fd);
+	}
+
+	return comm;
+}
+
+static void show_gpu_perf(struct overlay_context *ctx, struct overlay_gpu_perf *gp)
+{
+	struct gpu_perf_comm *comm, **prev;
+	const char *ring_name[] = {
+		"render",
+		"video",
+		"blt",
+	};
 	char buf[1024];
-	int y;
+	int y, n;
 
 	gpu_perf_update(&gp->gpu_perf);
 
-	y = 130;
+	y = ctx->last_y + 18;
+
+	for (prev = &gp->gpu_perf.comm; (comm = *prev) != NULL; ) {
+		cairo_move_to(ctx->cr, 12, y);
+		sprintf(buf, "%s:", comm->name);
+		cairo_show_text(ctx->cr, buf);
+		for (n = 0; n < 3; n++) {
+			if (comm->nr_requests[n] == 0)
+				continue;
+			sprintf(buf, " %d %s", comm->nr_requests[n], ring_name[n]);
+			cairo_show_text(ctx->cr, buf);
+		}
+		y += 14;
+
+		memset(comm->nr_requests, 0, sizeof(comm->nr_requests));
+		if (strcmp(comm->name, get_comm(comm->pid, buf, sizeof(buf)))) {
+			*prev = comm->next;
+			free(comm);
+		} else
+			prev = &comm->next;
+	}
 
 	sprintf(buf, "Flips: %d", gp->gpu_perf.flip_complete);
-	cairo_move_to(cr, 12, y);
-	cairo_show_text(cr, buf);
+	gp->gpu_perf.flip_complete = 0;
+	cairo_move_to(ctx->cr, 12, y);
+	cairo_show_text(ctx->cr, buf);
 	y += 14;
 
-	gp->gpu_perf.flip_complete = 0;
+	ctx->last_y = y;
 }
 
-static void show_gem_objects(cairo_t *cr)
+static void show_gem_objects(struct overlay_context *ctx)
 {
 	char gem_objects[1024], *s, *t, *end;
 	int len, y;
@@ -177,7 +230,7 @@ static void show_gem_objects(cairo_t *cr)
 	if (len <= 0)
 		return;
 
-	y = 150;
+	y = ctx->last_y + 18;
 
 	s = gem_objects;
 	end = s + len - 1;
@@ -187,12 +240,14 @@ static void show_gem_objects(cairo_t *cr)
 			t = end;
 		*t = '\0';
 
-		cairo_move_to(cr, 12, y);
-		cairo_show_text(cr, s);
+		cairo_move_to(ctx->cr, 12, y);
+		cairo_show_text(ctx->cr, s);
 		y += 14;
 
 		s = t+1;
 	}
+
+	ctx->last_y = y;
 }
 
 int main(int argc, char **argv)
@@ -215,32 +270,33 @@ int main(int argc, char **argv)
 	init_gpu_perf(&gpu_perf, surface);
 
 	while (1) {
-		cairo_t *cr;
+		struct overlay_context ctx;
 
 		usleep(500*1000);
 
-		cr = cairo_create(surface);
-		cairo_set_operator(cr, CAIRO_OPERATOR_CLEAR);
-		cairo_paint(cr);
-		cairo_set_operator(cr, CAIRO_OPERATOR_OVER);
+		ctx.cr = cairo_create(surface);
+		cairo_set_operator(ctx.cr, CAIRO_OPERATOR_CLEAR);
+		cairo_paint(ctx.cr);
+		cairo_set_operator(ctx.cr, CAIRO_OPERATOR_OVER);
+		ctx.last_y = 6;
 
 		{
 			char buf[80];
 			cairo_text_extents_t extents;
 			sprintf(buf, "%d", i++);
-			cairo_set_source_rgb(cr, .5, .5, .5);
-			cairo_text_extents(cr, buf, &extents);
-			cairo_move_to(cr,
+			cairo_set_source_rgb(ctx.cr, .5, .5, .5);
+			cairo_text_extents(ctx.cr, buf, &extents);
+			cairo_move_to(ctx.cr,
 				      cairo_image_surface_get_width(surface)-extents.width-6,
 				      6+extents.height);
-			cairo_show_text(cr, buf);
+			cairo_show_text(ctx.cr, buf);
 		}
 
-		show_gpu_top(cr, &gpu_top);
-		show_gpu_perf(cr, &gpu_perf);
-		show_gem_objects(cr);
+		show_gpu_top(&ctx, &gpu_top);
+		show_gpu_perf(&ctx, &gpu_perf);
+		show_gem_objects(&ctx);
 
-		cairo_destroy(cr);
+		cairo_destroy(ctx.cr);
 
 		overlay_show(surface);
 	}