[Intel-gfx] [PATCH i-g-t] overlay: parse tracepoints from sysfs to figure out fields' location

Lionel Landwerlin lionel.g.landwerlin at intel.com
Tue Dec 19 00:36:19 UTC 2017


With changes going to drm-tip, the tracepoints field locations are
going to change. This change introduces a tracepoint parser (using a
peg parser) which lets us figure out field positions on the fly.

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin at intel.com>
---
 configure.ac                  |   5 ++
 overlay/gpu-perf.c            | 164 ++++++++++++++++++++++++++++++++++--------
 overlay/meson.build           |  11 ++-
 overlay/tracepoint_format.leg |  34 +++++++++
 4 files changed, 184 insertions(+), 30 deletions(-)
 create mode 100644 overlay/tracepoint_format.leg

diff --git a/configure.ac b/configure.ac
index 8740f7a4..87ca9f91 100644
--- a/configure.ac
+++ b/configure.ac
@@ -151,6 +151,11 @@ else
 	enable_overlay_xlib="no"
 	enable_assembler="no"
 	BUILD_SHADER_DEBUGGER="no"
+
+	AC_CHECK_TOOL([LEG], [leg])
+	if test "x$LEG" != "xleg"; then
+		AC_MSG_ERROR([leg command missing, try : apt-get install peg])
+	fi
 fi
 AM_CONDITIONAL(BUILD_X86, [test "x$build_x86" = xyes])
 
diff --git a/overlay/gpu-perf.c b/overlay/gpu-perf.c
index 3d4a9be9..7fa561e3 100644
--- a/overlay/gpu-perf.c
+++ b/overlay/gpu-perf.c
@@ -57,25 +57,120 @@ 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];
+	uint8_t tracepoint_data[0];
 };
 
 enum {
-	DEVICE = 0,
-	CTX,
-	ENGINE,
-	CTX_SEQNO,
-	GLOBAL_SEQNO
+	TP_GEM_REQUEST_ADD,
+	TP_GEM_REQUEST_WAIT_BEGIN,
+	TP_GEM_REQUEST_WAIT_END,
+	TP_FLIP_COMPLETE,
+	TP_GEM_RING_SYNC_TO,
+	TP_GEM_RING_SWITCH_CONTEXT,
+
+	TP_NB
 };
 
-static uint64_t tracepoint_id(const char *sys, const char *name)
+struct tracepoint {
+	struct {
+		char name[128];
+		int offset;
+		int size;
+		int is_signed;
+	} fields[20];
+	const char *name;
+	int n_fields;
+
+	int device_field;
+	int ctx_field;
+	int ring_field;
+	int seqno_field;
+	int global_seqno_field;
+	int plane_field;
+} tracepoints[TP_NB] = {
+	[TP_GEM_REQUEST_ADD]         = { .name = "i915/i915_gem_request_add", },
+	[TP_GEM_REQUEST_WAIT_BEGIN]  = { .name = "i915/i915_gem_request_wait_begin", },
+	[TP_GEM_REQUEST_WAIT_END]    = { .name = "i915/i915_gem_request_wait_end", },
+	[TP_FLIP_COMPLETE]           = { .name = "i915/flip_complete", },
+	[TP_GEM_RING_SYNC_TO]        = { .name = "i915/gem_ring_sync_to", },
+	[TP_GEM_RING_SWITCH_CONTEXT] = { .name = "i915/gem_ring_switch_context", },
+};
+
+union parser_value {
+    char *string;
+    int integer;
+};
+
+struct parser_ctx {
+	struct tracepoint *tp;
+	FILE *fp;
+};
+
+#define YY_CTX_LOCAL
+#define YY_CTX_MEMBERS struct parser_ctx ctx;
+#define YYSTYPE union parser_value
+#define YY_PARSE(T) static T
+#define YY_INPUT(yy, buf, result, max)				\
+	{							\
+		int yyc = getc(yy->ctx.fp);			\
+		result = (EOF == yyc) ? 0 : (*(buf)= yyc, 1);	\
+	}
+
+#include "tracepoint_format.h"
+
+static int
+parse_tracepoint(int tp_id)
+{
+	struct tracepoint *tp = &tracepoints[tp_id];
+	yycontext ctx;
+	char buf[1024];
+
+	/* Already parsed? */
+	if (tp->n_fields != 0)
+		return 0;
+
+	snprintf(buf, sizeof(buf), "%s/tracing/events/%s/format",
+		 debugfs_path, tp->name);
+
+	memset(&ctx, 0, sizeof(ctx));
+	ctx.ctx.tp = tp;
+	ctx.ctx.fp = fopen(buf, "r");
+
+	if (ctx.ctx.fp == NULL)
+		return 0;
+
+	if (yyparse(&ctx)) {
+		for (int f = 0; f < tp->n_fields; f++) {
+			if (!strcmp(tp->fields[f].name, "device")) {
+				tp->device_field = f;
+			} else if (!strcmp(tp->fields[f].name, "ctx")) {
+				tp->ctx_field = f;
+			} else if (!strcmp(tp->fields[f].name, "ring")) {
+				tp->ring_field = f;
+			} else if (!strcmp(tp->fields[f].name, "seqno")) {
+				tp->seqno_field = f;
+			} else if (!strcmp(tp->fields[f].name, "global_seqno")) {
+				tp->global_seqno_field = f;
+			} else if (!strcmp(tp->fields[f].name, "plane")) {
+				tp->plane_field = f;
+			}
+		}
+        } else
+		tp->n_fields = 0;
+
+	yyrelease(&ctx);
+	fclose(ctx.ctx.fp);
+
+	return tp->n_fields;
+}
+
+static uint64_t tracepoint_id(int tp_id)
 {
 	char buf[1024];
 	int fd, n;
 
-	snprintf(buf, sizeof(buf), "%s/tracing/events/%s/%s/id", debugfs_path, sys, name);
+	snprintf(buf, sizeof(buf), "%s/tracing/events/%s/id", debugfs_path,
+		 tracepoints[tp_id].name);
 	fd = open(buf, 0);
 	if (fd < 0)
 		return 0;
@@ -88,8 +183,12 @@ static uint64_t tracepoint_id(const char *sys, const char *name)
 	return strtoull(buf, 0, 0);
 }
 
-static int perf_tracepoint_open(struct gpu_perf *gp,
-				const char *sys, const char *name,
+#define READ_TP_FIELD_U32(sample, tp_id, field_name)			\
+	(*(const uint32_t *)((sample)->tracepoint_data +		\
+			     tracepoints[tp_id].fields[			\
+				     tracepoints[tp_id].field_name##_field].offset))
+
+static int perf_tracepoint_open(struct gpu_perf *gp, int tp_id,
 				int (*func)(struct gpu_perf *, const void *))
 {
 	struct perf_event_attr attr;
@@ -99,10 +198,13 @@ static int perf_tracepoint_open(struct gpu_perf *gp,
 	memset(&attr, 0, sizeof (attr));
 
 	attr.type = PERF_TYPE_TRACEPOINT;
-	attr.config = tracepoint_id(sys, name);
+	attr.config = tracepoint_id(tp_id);
 	if (attr.config == 0)
 		return ENOENT;
 
+	if (parse_tracepoint(tp_id) == 0)
+		return ENOENT;
+
 	attr.sample_period = 1;
 	attr.sample_type = (PERF_SAMPLE_TIME | PERF_SAMPLE_STREAM_ID | PERF_SAMPLE_TID | PERF_SAMPLE_RAW);
 	attr.read_format = PERF_FORMAT_ID;
@@ -227,7 +329,7 @@ static int request_add(struct gpu_perf *gp, const void *event)
 	if (comm == NULL)
 		return 0;
 
-	comm->nr_requests[sample->raw[ENGINE]]++;
+	comm->nr_requests[READ_TP_FIELD_U32(sample, TP_GEM_REQUEST_ADD, ring)]++;
 	return 1;
 }
 
@@ -235,7 +337,7 @@ static int flip_complete(struct gpu_perf *gp, const void *event)
 {
 	const struct sample_event *sample = event;
 
-	gp->flip_complete[sample->raw[0]]++;
+	gp->flip_complete[READ_TP_FIELD_U32(sample, TP_FLIP_COMPLETE, plane)]++;
 	return 1;
 }
 
@@ -243,7 +345,7 @@ static int ctx_switch(struct gpu_perf *gp, const void *event)
 {
 	const struct sample_event *sample = event;
 
-	gp->ctx_switch[sample->raw[ENGINE]]++;
+	gp->ctx_switch[READ_TP_FIELD_U32(sample, TP_GEM_RING_SWITCH_CONTEXT, ring)]++;
 	return 1;
 }
 
@@ -278,11 +380,11 @@ static int wait_begin(struct gpu_perf *gp, const void *event)
 
 	wait->comm = comm;
 	wait->comm->active = true;
-	wait->context = sample->raw[ENGINE];
-	wait->seqno = sample->raw[CTX_SEQNO];
+	wait->context = READ_TP_FIELD_U32(sample, TP_GEM_REQUEST_WAIT_BEGIN, ctx);
+	wait->seqno = READ_TP_FIELD_U32(sample, TP_GEM_REQUEST_WAIT_BEGIN, seqno);
 	wait->time = sample->time;
-	wait->next = gp->wait[sample->raw[CTX]];
-	gp->wait[sample->raw[CTX]] = wait;
+	wait->next = gp->wait[READ_TP_FIELD_U32(sample, TP_GEM_REQUEST_WAIT_BEGIN, ring)];
+	gp->wait[READ_TP_FIELD_U32(sample, TP_GEM_REQUEST_WAIT_BEGIN, ring)] = wait;
 
 	return 0;
 }
@@ -292,9 +394,13 @@ static int wait_end(struct gpu_perf *gp, const void *event)
 	const struct sample_event *sample = event;
 	struct gpu_perf_time *wait, **prev;
 
-	for (prev = &gp->wait[sample->raw[ENGINE]]; (wait = *prev) != NULL; prev = &wait->next) {
-		if (wait->context != sample->raw[CTX] ||
-		    wait->seqno != sample->raw[CTX_SEQNO])
+	for (prev = &gp->wait[READ_TP_FIELD_U32(sample, TP_GEM_REQUEST_WAIT_END, ring)];
+	     (wait = *prev) != NULL;
+	     prev = &wait->next) {
+		if (wait->context != READ_TP_FIELD_U32(sample,
+							      TP_GEM_REQUEST_WAIT_END, ctx) ||
+		    wait->seqno != READ_TP_FIELD_U32(sample,
+							    TP_GEM_REQUEST_WAIT_END, seqno))
 			continue;
 
 		wait->comm->wait_time += sample->time - wait->time;
@@ -314,12 +420,12 @@ 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_request_wait_begin", wait_begin) == 0)
-		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);
+	perf_tracepoint_open(gp, TP_GEM_REQUEST_ADD, request_add);
+	if (perf_tracepoint_open(gp, TP_GEM_REQUEST_WAIT_BEGIN, wait_begin) == 0)
+		perf_tracepoint_open(gp, TP_GEM_REQUEST_WAIT_END, wait_end);
+	perf_tracepoint_open(gp, TP_FLIP_COMPLETE, flip_complete);
+	perf_tracepoint_open(gp, TP_GEM_RING_SYNC_TO, ring_sync);
+	perf_tracepoint_open(gp, TP_GEM_RING_SWITCH_CONTEXT, ctx_switch);
 
 	if (gp->nr_events == 0) {
 		gp->error = "i915.ko tracepoints not available";
diff --git a/overlay/meson.build b/overlay/meson.build
index afacff5e..6a5691dc 100644
--- a/overlay/meson.build
+++ b/overlay/meson.build
@@ -14,6 +14,14 @@ gpu_overlay_src = [
 	'rc6.c',
 ]
 
+leg = find_program('leg', required : true)
+leg_gen = generator(leg,
+		    output : '@BASENAME at .h',
+		    arguments : ['-o', '@OUTPUT@', '@INPUT@'])
+leg_file = leg_gen.process('tracepoint_format.leg')
+leg_lib = static_library('leg', leg_file)
+leg_inc = leg_lib.private_dir_include()
+
 xv = dependency('xv', required : false)
 x11 = dependency('x11', required : false)
 xext = dependency('xext', required : false)
@@ -53,8 +61,9 @@ gpu_overlay_src += 'kms/kms-overlay.c'
 
 if xrandr.found() and cairo.found()
 	executable('intel-gpu-overlay', gpu_overlay_src,
-			include_directories : inc,
+			include_directories : [inc, leg_inc],
 			c_args : gpu_overlay_cflags,
 			dependencies : gpu_overlay_deps,
+                        link_with : leg_lib,
 			install : true)
 endif
diff --git a/overlay/tracepoint_format.leg b/overlay/tracepoint_format.leg
new file mode 100644
index 00000000..ea83db91
--- /dev/null
+++ b/overlay/tracepoint_format.leg
@@ -0,0 +1,34 @@
+TracepointFormat =
+    'name' ':' Space PropertyName EndLine
+    'ID' ':' Space Number EndLine
+    'format' ':' EndLine
+    Field+
+    'print fmt' ':' .*
+    !.
+
+Field         = Space (Property ';' Space)+ EndLine
+        { yy->ctx.tp->n_fields++; }
+              | EndLine
+
+Property      = 'offset' ':' v:Number
+        { yy->ctx.tp->fields[yy->ctx.tp->n_fields].offset = v.integer; }
+              | 'size' ':' v:Number
+        { yy->ctx.tp->fields[yy->ctx.tp->n_fields].size = v.integer; }
+              | 'signed' ':' v:Number
+        { yy->ctx.tp->fields[yy->ctx.tp->n_fields].is_signed = v.integer != 0; }
+              | 'field' ':' v:PropertyValue
+        { snprintf(yy->ctx.tp->fields[yy->ctx.tp->n_fields].name,
+                   sizeof(yy->ctx.tp->fields[yy->ctx.tp->n_fields].name),
+                   "%s", strrchr(v.string, ' ') + 1); free(v.string); }
+              | n:PropertyName ':' v:PropertyValue
+        { free(n.string); free(v.string); }
+
+PropertyName  = < [A-Za-z0-9_]+ >
+        { $$.string = strdup(yytext); }
+PropertyValue = < [^;]+ >
+        { $$.string = strdup(yytext); }
+Number        = < [0-9]+ >
+        { $$.integer = atoi(yytext); }
+
+EndLine       = [\n]
+Space         = [ \t]*
-- 
2.15.1



More information about the Intel-gfx mailing list