public inbox for igt-dev@lists.freedesktop.org
 help / color / mirror / Atom feed
* [igt-dev] [PATCH i-g-t 0/4] Add perf OA tools for GPUvis
@ 2020-02-15  1:11 Umesh Nerlige Ramappa
  2020-02-15  1:11 ` [igt-dev] [PATCH i-g-t 2/4] lib/i915/perf: Add support for loading perf configurations Umesh Nerlige Ramappa
                   ` (3 more replies)
  0 siblings, 4 replies; 5+ messages in thread
From: Umesh Nerlige Ramappa @ 2020-02-15  1:11 UTC (permalink / raw)
  To: igt-dev, Joonas Lahtinen, Ashutosh Dixit, Lionel G Landwerlin

The tools provided here enable capturing performance metrics from the i915
driver and are used in conjunction with the GPUvis software here - 

https://github.com/mikesart/gpuvis

The changes required in GPUvis are wip and will be posted following the merge of
these tools.

For more information, view tools/i915-perf/README in this patch series

Lionel Landwerlin (4):
  lib/i915/perf: Add i915_perf library
  lib/i915/perf: Add support for loading perf configurations
  tools/i915/perf: Add i915 perf recorder tool
  lib/i915/perf: Add i915 perf data reader

 lib/i915-perf.pc.in                           |    10 +
 lib/i915/perf-configs/README.md               |   115 +
 lib/i915/perf-configs/codegen.py              |    33 +
 lib/i915/perf-configs/guids.xml               |   282 +
 lib/i915/perf-configs/mdapi-xml-convert.py    |  1000 +
 lib/i915/perf-configs/oa-bdw.xml              | 15653 ++++++++++++++++
 lib/i915/perf-configs/oa-bxt.xml              |  9595 ++++++++++
 lib/i915/perf-configs/oa-cflgt2.xml           | 10866 +++++++++++
 lib/i915/perf-configs/oa-cflgt3.xml           | 10933 +++++++++++
 lib/i915/perf-configs/oa-chv.xml              |  9757 ++++++++++
 lib/i915/perf-configs/oa-cnl.xml              | 10411 ++++++++++
 lib/i915/perf-configs/oa-glk.xml              |  9346 +++++++++
 lib/i915/perf-configs/oa-hsw.xml              |  4615 +++++
 lib/i915/perf-configs/oa-icl.xml              | 11899 ++++++++++++
 lib/i915/perf-configs/oa-kblgt2.xml           | 10866 +++++++++++
 lib/i915/perf-configs/oa-kblgt3.xml           | 10933 +++++++++++
 lib/i915/perf-configs/oa-sklgt2.xml           | 11895 ++++++++++++
 lib/i915/perf-configs/oa-sklgt3.xml           | 10933 +++++++++++
 lib/i915/perf-configs/oa-sklgt4.xml           | 10956 +++++++++++
 lib/i915/perf-configs/perf-codegen.py         |   854 +
 lib/i915/perf-configs/update-guids.py         |   230 +
 lib/i915/perf.c                               |   332 +
 lib/i915/perf.h                               |   240 +
 lib/i915/perf_data.h                          |    88 +
 lib/i915/perf_data_reader.c                   |   330 +
 lib/i915/perf_data_reader.h                   |   103 +
 lib/meson.build                               |    67 +
 tools/i915-perf/README                        |    70 +
 tools/i915-perf/i915_perf_configs.c           |   277 +
 tools/i915-perf/i915_perf_control.c           |   133 +
 tools/i915-perf/i915_perf_recorder.c          |   931 +
 tools/i915-perf/i915_perf_recorder_commands.h |    39 +
 tools/i915-perf/meson.build                   |    17 +
 tools/meson.build                             |     1 +
 34 files changed, 153810 insertions(+)
 create mode 100644 lib/i915-perf.pc.in
 create mode 100644 lib/i915/perf-configs/README.md
 create mode 100644 lib/i915/perf-configs/codegen.py
 create mode 100644 lib/i915/perf-configs/guids.xml
 create mode 100755 lib/i915/perf-configs/mdapi-xml-convert.py
 create mode 100644 lib/i915/perf-configs/oa-bdw.xml
 create mode 100644 lib/i915/perf-configs/oa-bxt.xml
 create mode 100644 lib/i915/perf-configs/oa-cflgt2.xml
 create mode 100644 lib/i915/perf-configs/oa-cflgt3.xml
 create mode 100644 lib/i915/perf-configs/oa-chv.xml
 create mode 100644 lib/i915/perf-configs/oa-cnl.xml
 create mode 100644 lib/i915/perf-configs/oa-glk.xml
 create mode 100644 lib/i915/perf-configs/oa-hsw.xml
 create mode 100644 lib/i915/perf-configs/oa-icl.xml
 create mode 100644 lib/i915/perf-configs/oa-kblgt2.xml
 create mode 100644 lib/i915/perf-configs/oa-kblgt3.xml
 create mode 100644 lib/i915/perf-configs/oa-sklgt2.xml
 create mode 100644 lib/i915/perf-configs/oa-sklgt3.xml
 create mode 100644 lib/i915/perf-configs/oa-sklgt4.xml
 create mode 100755 lib/i915/perf-configs/perf-codegen.py
 create mode 100755 lib/i915/perf-configs/update-guids.py
 create mode 100644 lib/i915/perf.c
 create mode 100644 lib/i915/perf.h
 create mode 100644 lib/i915/perf_data.h
 create mode 100644 lib/i915/perf_data_reader.c
 create mode 100644 lib/i915/perf_data_reader.h
 create mode 100644 tools/i915-perf/README
 create mode 100644 tools/i915-perf/i915_perf_configs.c
 create mode 100644 tools/i915-perf/i915_perf_control.c
 create mode 100644 tools/i915-perf/i915_perf_recorder.c
 create mode 100644 tools/i915-perf/i915_perf_recorder_commands.h
 create mode 100644 tools/i915-perf/meson.build

-- 
2.20.1

_______________________________________________
igt-dev mailing list
igt-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/igt-dev

^ permalink raw reply	[flat|nested] 5+ messages in thread

* [igt-dev] [PATCH i-g-t 2/4] lib/i915/perf: Add support for loading perf configurations
  2020-02-15  1:11 [igt-dev] [PATCH i-g-t 0/4] Add perf OA tools for GPUvis Umesh Nerlige Ramappa
@ 2020-02-15  1:11 ` Umesh Nerlige Ramappa
  2020-02-15  1:11 ` [igt-dev] [PATCH i-g-t 3/4] tools/i915/perf: Add i915 perf recorder tool Umesh Nerlige Ramappa
                   ` (2 subsequent siblings)
  3 siblings, 0 replies; 5+ messages in thread
From: Umesh Nerlige Ramappa @ 2020-02-15  1:11 UTC (permalink / raw)
  To: igt-dev, Joonas Lahtinen, Ashutosh Dixit, Lionel G Landwerlin

From: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

Add support for loading perf configurations used by gpuvis.

v2: rebase fixes for igt list (Umesh)

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com>
---
 lib/i915/perf.c | 138 ++++++++++++++++++++++++++++++++++++++++++++++++
 lib/i915/perf.h |   2 +
 2 files changed, 140 insertions(+)

diff --git a/lib/i915/perf.c b/lib/i915/perf.c
index 1627f102..ae786701 100644
--- a/lib/i915/perf.c
+++ b/lib/i915/perf.c
@@ -20,8 +20,18 @@
  * SOFTWARE.
  */
 
+#include <assert.h>
+#include <errno.h>
+#include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#include <dirent.h>
+#include <fcntl.h>
+#include <sys/ioctl.h>
+#include <sys/stat.h>
+#include <sys/sysmacros.h>
+#include <sys/types.h>
+#include <unistd.h>
 
 #include "intel_chipset.h"
 #include "perf.h"
@@ -192,3 +202,131 @@ intel_perf_add_metric_set(struct intel_perf *perf,
 {
 	igt_list_add_tail(&metric_set->link, &perf->metric_sets);
 }
+
+static bool
+read_file_uint64(const char *file, uint64_t *value)
+{
+	char buf[32];
+	int fd, n;
+
+	fd = open(file, 0);
+	if (fd < 0)
+		return false;
+	n = read(fd, buf, sizeof (buf) - 1);
+	close(fd);
+	if (n < 0)
+		return false;
+
+	buf[n] = '\0';
+	*value = strtoull(buf, 0, 0);
+
+	return true;
+}
+
+static int
+get_card_for_fd(int fd)
+{
+	struct stat sb;
+	int mjr, mnr;
+	char buffer[128];
+	DIR *drm_dir;
+	struct dirent *entry;
+	int retval = -1;
+
+	if (fstat(fd, &sb))
+		return -1;
+
+	mjr = major(sb.st_rdev);
+	mnr = minor(sb.st_rdev);
+
+	snprintf(buffer, sizeof(buffer), "/sys/dev/char/%d:%d/device/drm", mjr, mnr);
+
+	drm_dir = opendir(buffer);
+	assert(drm_dir != NULL);
+
+	while ((entry = readdir(drm_dir))) {
+		if (entry->d_type == DT_DIR && strncmp(entry->d_name, "card", 4) == 0) {
+			retval = strtoull(entry->d_name + 4, NULL, 10);
+			break;
+		}
+	}
+
+	closedir(drm_dir);
+
+	return retval;
+}
+
+static void
+load_metric_set_config(struct intel_perf_metric_set *metric_set, int drm_fd)
+{
+	struct drm_i915_perf_oa_config config;
+	uint64_t config_id = 0;
+
+	memset(&config, 0, sizeof(config));
+
+	memcpy(config.uuid, metric_set->hw_config_guid, sizeof(config.uuid));
+
+	config.n_mux_regs = metric_set->n_mux_regs;
+	config.mux_regs_ptr = (uintptr_t) metric_set->mux_regs;
+
+	config.n_boolean_regs = metric_set->n_b_counter_regs;
+	config.boolean_regs_ptr = (uintptr_t) metric_set->b_counter_regs;
+
+	config.n_flex_regs = metric_set->n_flex_regs;
+	config.flex_regs_ptr = (uintptr_t) metric_set->flex_regs;
+
+	while (ioctl(drm_fd, DRM_IOCTL_I915_PERF_ADD_CONFIG, &config) < 0 &&
+	       (errno == EAGAIN || errno == EINTR));
+
+	metric_set->perf_oa_metrics_set = config_id;
+}
+
+void
+intel_perf_load_perf_configs(struct intel_perf *perf, int drm_fd)
+{
+	int drm_card = get_card_for_fd(drm_fd);
+	struct dirent *entry;
+	char metrics_path[128];
+	DIR *metrics_dir;
+	struct intel_perf_metric_set *metric_set;
+
+	snprintf(metrics_path, sizeof(metrics_path),
+		 "/sys/class/drm/card%d/metrics", drm_card);
+	metrics_dir = opendir(metrics_path);
+	if (!metrics_dir)
+		return;
+
+	while ((entry = readdir(metrics_dir))) {
+		char *metric_id_path;
+		uint64_t metric_id;
+
+		if (entry->d_type != DT_DIR)
+			continue;
+
+		asprintf(&metric_id_path, "%s/%s/id",
+			 metrics_path, entry->d_name);
+
+		if (!read_file_uint64(metric_id_path, &metric_id)) {
+			free(metric_id_path);
+			continue;
+		}
+
+		free(metric_id_path);
+
+		igt_list_for_each_entry(metric_set, &perf->metric_sets, link) {
+			if (!strcmp(metric_set->hw_config_guid, entry->d_name)) {
+				metric_set->perf_oa_metrics_set = metric_id;
+				break;
+			}
+		}
+	}
+
+	closedir(metrics_dir);
+
+	igt_list_for_each_entry(metric_set, &perf->metric_sets, link) {
+		if (metric_set->perf_oa_metrics_set)
+			continue;
+
+		load_metric_set_config(metric_set, drm_fd);
+	}
+}
diff --git a/lib/i915/perf.h b/lib/i915/perf.h
index 5a091c46..0b66efe1 100644
--- a/lib/i915/perf.h
+++ b/lib/i915/perf.h
@@ -231,6 +231,8 @@ void intel_perf_add_logical_counter(struct intel_perf *perf,
 void intel_perf_add_metric_set(struct intel_perf *perf,
 			       struct intel_perf_metric_set *metric_set);
 
+void intel_perf_load_perf_configs(struct intel_perf *perf, int drm_fd);
+
 #ifdef __cplusplus
 };
 #endif
-- 
2.20.1

_______________________________________________
igt-dev mailing list
igt-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/igt-dev

^ permalink raw reply related	[flat|nested] 5+ messages in thread

* [igt-dev] [PATCH i-g-t 3/4] tools/i915/perf: Add i915 perf recorder tool
  2020-02-15  1:11 [igt-dev] [PATCH i-g-t 0/4] Add perf OA tools for GPUvis Umesh Nerlige Ramappa
  2020-02-15  1:11 ` [igt-dev] [PATCH i-g-t 2/4] lib/i915/perf: Add support for loading perf configurations Umesh Nerlige Ramappa
@ 2020-02-15  1:11 ` Umesh Nerlige Ramappa
  2020-02-15  1:11 ` [igt-dev] [PATCH i-g-t 4/4] lib/i915/perf: Add i915 perf data reader Umesh Nerlige Ramappa
  2020-02-17 13:42 ` [igt-dev] [PATCH i-g-t 0/4] Add perf OA tools for GPUvis Lionel Landwerlin
  3 siblings, 0 replies; 5+ messages in thread
From: Umesh Nerlige Ramappa @ 2020-02-15  1:11 UTC (permalink / raw)
  To: igt-dev, Joonas Lahtinen, Ashutosh Dixit, Lionel G Landwerlin

From: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

i915 perf recorder tool captures OA perf data for a specific metric set
in a circular buffer of specified size. The i915 perf control tool is
used to dump the data captured in the circular buffer to a trace file.
The data captured is used to view relevant events in gpuvis.

v2: (Umesh)
- rebase fixes for igt_list apis
- memset circular_buffer to 0 to initialize size, beginpos and enpos
- _FORTIFY_SOURCE=2 caused snprintf to go through __snprintf_chk that
  falsely flagged a buffer overflow and sent an abort signal to
  i915_perf_control when capturing traces. undef the _FORTIFY_SOURCE
  selectively for i915 control tool.

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com>
---
 lib/i915/perf_data.h                          |  88 ++
 lib/meson.build                               |   1 +
 tools/i915-perf/i915_perf_control.c           | 133 +++
 tools/i915-perf/i915_perf_recorder.c          | 931 ++++++++++++++++++
 tools/i915-perf/i915_perf_recorder_commands.h |  39 +
 tools/i915-perf/meson.build                   |  12 +
 6 files changed, 1204 insertions(+)
 create mode 100644 lib/i915/perf_data.h
 create mode 100644 tools/i915-perf/i915_perf_control.c
 create mode 100644 tools/i915-perf/i915_perf_recorder.c
 create mode 100644 tools/i915-perf/i915_perf_recorder_commands.h

diff --git a/lib/i915/perf_data.h b/lib/i915/perf_data.h
new file mode 100644
index 00000000..13791187
--- /dev/null
+++ b/lib/i915/perf_data.h
@@ -0,0 +1,88 @@
+/*
+ * Copyright (C) 2019 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef PERF_DATA_H
+#define PERF_DATA_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* The structures below are embedded in the i915-perf stream so as to
+ * provide metadata. The types used in the
+ * drm_i915_perf_record_header.type are defined in
+ * intel_perf_record_type.
+ *
+ * Once defined, those structures cannot change. If you need to add
+ * new data, just define a new structure & record_type.
+ */
+
+#include <stdint.h>
+
+enum intel_perf_record_type {
+	/* Start at 65536, which is pretty safe since after 3years the
+	 * kernel hasn't defined more than 3 entries.
+	 */
+
+	/* intel_perf_record_device_info */
+	INTEL_PERF_RECORD_TYPE_DEVICE_INFO = 1 << 16,
+
+	/* intel_perf_record_device_topology */
+	INTEL_PERF_RECORD_TYPE_DEVICE_TOPOLOGY,
+
+	/* intel_perf_record_timestamp_correlation */
+	INTEL_PERF_RECORD_TYPE_TIMESTAMP_CORRELATION,
+};
+
+struct intel_perf_record_device_info {
+	/* Frequency of the timestamps in the records. */
+	uint64_t timestamp_frequency;
+
+	/* PCI ID */
+	uint32_t device_id;
+
+	/* enum drm_i915_oa_format */
+	uint32_t oa_format;
+
+	/* Configuration identifier */
+	char uuid[40];
+};
+
+/* Topology as reported by i915. */
+struct intel_perf_record_device_topology {
+	struct drm_i915_query_topology_info topology;
+};
+
+/* Timestamp correlation between CPU/GPU. */
+struct intel_perf_record_timestamp_correlation {
+	/* In CLOCK_MONOTONIC */
+	uint64_t cpu_timestamp;
+
+	/* Engine timestamp associated with the OA unit */
+	uint64_t gpu_timestamp;
+};
+
+#ifdef __cplusplus
+};
+#endif
+
+#endif /* PERF_DATA_H */
diff --git a/lib/meson.build b/lib/meson.build
index edff8a67..6e935d45 100644
--- a/lib/meson.build
+++ b/lib/meson.build
@@ -217,6 +217,7 @@ install_headers(
   'igt_list.h',
   'intel_chipset.h',
   'i915/perf.h',
+  'i915/perf_data.h',
   subdir : 'i915-perf'
 )
 
diff --git a/tools/i915-perf/i915_perf_control.c b/tools/i915-perf/i915_perf_control.c
new file mode 100644
index 00000000..a8d0d30f
--- /dev/null
+++ b/tools/i915-perf/i915_perf_control.c
@@ -0,0 +1,133 @@
+/*
+ * Copyright (C) 2019 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <getopt.h>
+#include <stdbool.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+
+#include "i915_perf_recorder_commands.h"
+
+static void
+usage(const char *name)
+{
+	fprintf(stdout,
+		"Usage: %s [options]\n"
+		"\n"
+		"     --help,               -h         Print this screen\n"
+		"     --command-fifo,       -f <path>  Path to a command fifo\n"
+		"     --dump,               -d <path>  Write a content of circular buffer to path\n",
+		name);
+}
+
+int
+main(int argc, char *argv[])
+{
+	const struct option long_options[] = {
+		{"help",                       no_argument, 0, 'h'},
+		{"dump",                 required_argument, 0, 'd'},
+		{"command-fifo",         required_argument, 0, 'f'},
+		{"quit",                       no_argument, 0, 'q'},
+		{0, 0, 0, 0}
+	};
+	const char *command_fifo = I915_PERF_RECORD_FIFO_PATH, *dump_file = NULL;
+	FILE *command_fifo_file;
+	int opt;
+	bool quit = false;
+
+	while ((opt = getopt_long(argc, argv, "hd:f:q", long_options, NULL)) != -1) {
+		switch (opt) {
+		case 'h':
+			usage(argv[0]);
+			return EXIT_SUCCESS;
+		case 'd':
+			dump_file = optarg;
+			break;
+		case 'f':
+			command_fifo = optarg;
+			break;
+		case 'q':
+			quit = true;
+			break;
+		default:
+			fprintf(stderr, "Internal error: "
+				"unexpected getopt value: %d\n", opt);
+			usage(argv[0]);
+			return EXIT_FAILURE;
+		}
+	}
+
+	if (!command_fifo)
+		return EXIT_FAILURE;
+
+	command_fifo_file = fopen(command_fifo, "r+");
+	if (!command_fifo_file) {
+		fprintf(stderr, "Unable to open command file\n");
+		return EXIT_FAILURE;
+	}
+
+	if (dump_file) {
+		if (dump_file[0] == '/') {
+			uint32_t total_len =
+				sizeof(struct recorder_command_base) + strlen(dump_file) + 1;
+			struct {
+				struct recorder_command_base base;
+				struct recorder_command_dump dump;
+			} *data = malloc(total_len);
+
+			data->base.command = RECORDER_COMMAND_DUMP;
+			data->base.size = total_len;
+			snprintf((char *) data->dump.path, strlen(dump_file) + 1, "%s", dump_file);
+
+			fwrite(data, total_len, 1, command_fifo_file);
+		} else {
+			char *cwd = get_current_dir_name();
+			uint32_t path_len = strlen(cwd) + 1 + strlen(dump_file) + 1;
+			uint32_t total_len = sizeof(struct recorder_command_base) + path_len;
+			struct {
+				struct recorder_command_base base;
+				struct recorder_command_dump dump;
+			} *data = malloc(total_len);
+
+			data->base.command = RECORDER_COMMAND_DUMP;
+			data->base.size = total_len;
+			snprintf((char *) data->dump.path, path_len, "%s/%s", cwd, dump_file);
+
+			fwrite(data, total_len, 1, command_fifo_file);
+		}
+	}
+
+	if (quit) {
+		struct recorder_command_base base = {
+			.command = RECORDER_COMMAND_QUIT,
+			.size = sizeof(base),
+		};
+
+		fwrite(&base, sizeof(base), 1, command_fifo_file);
+	}
+
+	fclose(command_fifo_file);
+
+	return EXIT_SUCCESS;
+}
diff --git a/tools/i915-perf/i915_perf_recorder.c b/tools/i915-perf/i915_perf_recorder.c
new file mode 100644
index 00000000..61bde5ba
--- /dev/null
+++ b/tools/i915-perf/i915_perf_recorder.c
@@ -0,0 +1,931 @@
+/*
+ * Copyright (C) 2019 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <assert.h>
+#include <dirent.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <getopt.h>
+#include <inttypes.h>
+#include <poll.h>
+#include <signal.h>
+#include <stdbool.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/ioctl.h>
+#include <sys/stat.h>
+#include <sys/sysmacros.h>
+#include <sys/time.h>
+#include <sys/types.h>
+#include <time.h>
+#include <unistd.h>
+
+#include <i915_drm.h>
+
+#include "igt_core.h"
+#include "intel_chipset.h"
+#include "i915/perf.h"
+#include "i915/perf_data.h"
+
+#include "i915_perf_recorder_commands.h"
+
+#define ALIGN(v, a) (((v) + (a)-1) & ~((a)-1))
+#define ARRAY_SIZE(arr) (sizeof(arr)/sizeof((arr)[0]))
+#define MAX(a,b) ((a) > (b) ? (a) : (b))
+#define MIN(a,b) ((a) < (b) ? (a) : (b))
+
+struct circular_buffer {
+	char   *data;
+	size_t  allocated_size;
+	size_t  size;
+	size_t  beginpos;
+	size_t  endpos;
+};
+
+struct chunk {
+	char *data;
+	size_t len;
+};
+
+static size_t
+circular_available_size(const struct circular_buffer *buffer)
+{
+	assert(buffer->size <= buffer->allocated_size);
+	return buffer->allocated_size - buffer->size;
+}
+
+static void
+get_chunks(struct chunk *chunks, struct circular_buffer *buffer, bool write, size_t len)
+{
+	size_t offset = write ? buffer->endpos : buffer->beginpos;
+
+	if (write)
+		assert(circular_available_size(buffer) >= len);
+	else
+		assert(buffer->size >= len);
+
+	chunks[0].data = &buffer->data[offset];
+
+	if ((offset + len) > buffer->allocated_size) {
+		chunks[0].len = buffer->allocated_size - offset;
+		chunks[1].data = buffer->data;
+		chunks[1].len = len - (buffer->allocated_size - offset);
+	} else {
+		chunks[0].len = len;
+		chunks[1].data = NULL;
+		chunks[1].len = 0;
+	}
+}
+
+static ssize_t
+circular_buffer_read(void *c, char *buf, size_t size)
+{
+	struct circular_buffer *buffer = c;
+	struct chunk chunks[2];
+
+	if (buffer->size < size)
+		return -1;
+
+	get_chunks(chunks, buffer, false, size);
+
+	memcpy(buf, chunks[0].data, chunks[0].len);
+	memcpy(buf + chunks[0].len, chunks[1].data, chunks[1].len);
+	buffer->beginpos = (buffer->beginpos + size) % buffer->allocated_size;
+	buffer->size -= size;
+
+	return size;
+}
+
+static size_t
+peek_item_size(struct circular_buffer *buffer)
+{
+	struct drm_i915_perf_record_header header;
+	struct chunk chunks[2];
+
+	if (!buffer->size)
+		return 0;
+
+	assert(buffer->size >= sizeof(header));
+
+	get_chunks(chunks, buffer, false, sizeof(header));
+	memcpy(&header, chunks[0].data, chunks[0].len);
+	memcpy((char *) &header + chunks[0].len, chunks[1].data, chunks[1].len);
+
+	return header.size;
+}
+
+static void
+circular_shrink(struct circular_buffer *buffer, size_t size)
+{
+	size_t shrank = 0, item_size;
+
+	assert(size <= buffer->allocated_size);
+
+	while (shrank < size && buffer->size > (item_size = peek_item_size(buffer))) {
+		assert(item_size > 0 && item_size <= buffer->allocated_size);
+
+		buffer->beginpos = (buffer->beginpos + item_size) % buffer->allocated_size;
+		buffer->size -= item_size;
+
+		shrank += item_size;
+	}
+}
+
+static ssize_t
+circular_buffer_write(void *c, const char *buf, size_t _size)
+{
+	struct circular_buffer *buffer = c;
+	size_t size = _size;
+
+	while (size) {
+		size_t avail = circular_available_size(buffer), item_size;
+		struct chunk chunks[2];
+
+		/* Make space in the buffer if there is too much data. */
+		if (avail < size)
+			circular_shrink(buffer, size - avail);
+
+		item_size = MIN(circular_available_size(buffer), size);
+
+		get_chunks(chunks, buffer, true, item_size);
+
+		memcpy(chunks[0].data, buf, chunks[0].len);
+		memcpy(chunks[1].data, buf + chunks[0].len, chunks[1].len);
+
+		buf += item_size;
+		size -= item_size;
+
+		buffer->endpos = (buffer->endpos + item_size) % buffer->allocated_size;
+		buffer->size += item_size;
+	}
+
+	return _size;
+}
+
+static int
+circular_buffer_seek(void *c, off64_t *offset, int whence)
+{
+	return -1;
+}
+
+static int
+circular_buffer_close(void *c)
+{
+	return 0;
+}
+
+cookie_io_functions_t circular_buffer_functions = {
+	.read  = circular_buffer_read,
+	.write = circular_buffer_write,
+	.seek  = circular_buffer_seek,
+	.close = circular_buffer_close,
+};
+
+
+static bool
+read_file_uint64(const char *file, uint64_t *value)
+{
+	char buf[32];
+	int fd, n;
+
+	fd = open(file, 0);
+	if (fd < 0)
+		return false;
+	n = read(fd, buf, sizeof (buf) - 1);
+	close(fd);
+	if (n < 0)
+		return false;
+
+	buf[n] = '\0';
+	*value = strtoull(buf, 0, 0);
+
+	return true;
+}
+
+static uint32_t
+read_device_param(const char *stem, int id, const char *param)
+{
+	char *name;
+	int ret = asprintf(&name, "/sys/class/drm/%s%u/device/%s", stem, id, param);
+	uint64_t value;
+	bool success;
+
+	assert(ret != -1);
+
+	success = read_file_uint64(name, &value);
+	free(name);
+
+	return success ? value : 0;
+}
+
+static int
+find_intel_render_node(void)
+{
+	for (int i = 128; i < (128 + 16); i++) {
+		if (read_device_param("renderD", i, "vendor") == 0x8086)
+			return i;
+	}
+
+	return -1;
+}
+
+static int
+open_render_node(uint32_t *devid)
+{
+	char *name;
+	int ret;
+	int fd;
+
+	int render = find_intel_render_node();
+	if (render < 0)
+		return -1;
+
+	ret = asprintf(&name, "/dev/dri/renderD%u", render);
+	assert(ret != -1);
+
+	*devid = read_device_param("renderD", render, "device");
+
+	fd = open(name, O_RDWR);
+	free(name);
+
+	return fd;
+}
+
+static uint32_t
+oa_exponent_for_period(uint64_t device_timestamp_frequency, double period)
+{
+	uint64_t period_ns = 1000 * 1000 * 1000 * period;
+	uint64_t device_periods[32];
+
+	for (uint32_t i = 0; i < ARRAY_SIZE(device_periods); i++)
+		device_periods[i] = 1000000000ull * (1u << i) / device_timestamp_frequency;
+
+	for (uint32_t i = 1; i < ARRAY_SIZE(device_periods); i++) {
+		if (period_ns >= device_periods[i - 1] &&
+		    period_ns < device_periods[i]) {
+			if ((device_periods[i] - period_ns) >
+			    (period_ns - device_periods[i - 1]))
+				return i - 1;
+			return i;
+		}
+	}
+
+	return -1;
+}
+
+static int
+perf_ioctl(int fd, unsigned long request, void *arg)
+{
+	int ret;
+
+	do {
+		ret = ioctl(fd, request, arg);
+	} while (ret == -1 && (errno == EINTR || errno == EAGAIN));
+
+	return ret;
+}
+
+static uint64_t
+get_device_timestamp_frequency(const struct intel_device_info *devinfo, int drm_fd)
+{
+	drm_i915_getparam_t gp;
+	int timestamp_frequency;
+
+	gp.param = I915_PARAM_CS_TIMESTAMP_FREQUENCY;
+	gp.value = &timestamp_frequency;
+	if (perf_ioctl(drm_fd, DRM_IOCTL_I915_GETPARAM, &gp) == 0)
+		return timestamp_frequency;
+
+	if (devinfo->gen > 9) {
+		fprintf(stderr, "Unable to query timestamp frequency from i915, please update kernel.\n");
+		return 0;
+	}
+
+	fprintf(stderr, "Warning: unable to query timestamp frequency from i915, guessing values...\n");
+
+	if (devinfo->gen <= 8)
+		return 12500000;
+	if (devinfo->is_broxton)
+		return 19200000;
+	return 12000000;
+}
+
+static int
+perf_open(int drm_fd,
+	  const struct intel_device_info *devinfo,
+	  uint32_t oa_exponent,
+	  const struct intel_perf_metric_set *metric_set)
+{
+	uint64_t properties[DRM_I915_PERF_PROP_MAX * 2];
+	struct drm_i915_perf_open_param param;
+	int p = 0, stream_fd;
+
+	properties[p++] = DRM_I915_PERF_PROP_SAMPLE_OA;
+	properties[p++] = true;
+
+	properties[p++] = DRM_I915_PERF_PROP_OA_METRICS_SET;
+	properties[p++] = metric_set->perf_oa_metrics_set;
+
+	properties[p++] = DRM_I915_PERF_PROP_OA_FORMAT;
+	properties[p++] = metric_set->perf_oa_format;
+
+	properties[p++] = DRM_I915_PERF_PROP_OA_EXPONENT;
+	properties[p++] = oa_exponent;
+
+	memset(&param, 0, sizeof(param));
+	param.flags = 0;
+	param.flags |= I915_PERF_FLAG_FD_CLOEXEC | I915_PERF_FLAG_FD_NONBLOCK;
+	param.properties_ptr = (uintptr_t)properties;
+	param.num_properties = p / 2;
+
+	stream_fd = perf_ioctl(drm_fd, DRM_IOCTL_I915_PERF_OPEN, &param);
+	return stream_fd;
+}
+
+static bool quit = false;
+
+static void
+sigint_handler(int val)
+{
+	quit = true;
+}
+
+static bool
+write_header(FILE *output,
+	     uint32_t device_id,
+	     uint64_t timestamp_frequency,
+	     const struct intel_perf_metric_set *metric_set)
+{
+	struct intel_perf_record_device_info info = {
+		.timestamp_frequency = timestamp_frequency,
+		.device_id = device_id,
+		.oa_format = metric_set->perf_oa_format,
+	};
+	struct drm_i915_perf_record_header header = {
+		.type = INTEL_PERF_RECORD_TYPE_DEVICE_INFO,
+		.size = sizeof(header) + sizeof(info),
+	};
+
+	snprintf(info.uuid, sizeof(info.uuid), "%s", metric_set->hw_config_guid);
+
+	if (fwrite(&header, sizeof(header), 1, output) != 1)
+		return false;
+
+	if (fwrite(&info, sizeof(info), 1, output) != 1)
+		return false;
+
+	return true;
+}
+
+static bool
+write_topology(FILE *output, int drm_fd)
+{
+	struct drm_i915_perf_record_header header = {
+		.type = INTEL_PERF_RECORD_TYPE_DEVICE_TOPOLOGY,
+	};
+	struct drm_i915_query query = {};
+	struct drm_i915_query_topology_info *topo_info;
+	struct drm_i915_query_item item = {
+		.query_id = DRM_I915_QUERY_TOPOLOGY_INFO,
+	};
+	int ret;
+
+	query.num_items = 1;
+	query.items_ptr = (uintptr_t) &item;
+
+	/* Maybe not be available on older kernels. */
+	ret = perf_ioctl(drm_fd, DRM_IOCTL_I915_QUERY, &query);
+	if (ret < 0)
+		return true;
+
+	assert(item.length > 0);
+	topo_info = malloc(item.length);
+	item.data_ptr = (uintptr_t) topo_info;
+
+	ret = perf_ioctl(drm_fd, DRM_IOCTL_I915_QUERY, &query);
+	assert(ret == 0);
+
+	header.size = sizeof(header) + item.length;
+	if (fwrite(&header, sizeof(header), 1, output) != 1)
+		return false;
+
+	if (fwrite(topo_info, item.length, 1, output) != 1)
+		return false;
+
+	return true;
+}
+
+static bool
+write_i915_perf_data(FILE *output, int perf_fd)
+{
+	ssize_t ret;
+	char data[4096];
+
+	while ((ret = read(perf_fd, data, sizeof(data))) > 0 ||
+	       errno == EINTR) {
+		if (fwrite(data, ret, 1, output) != 1)
+			return false;
+	}
+
+	return true;
+}
+
+static uint64_t timespec_diff(struct timespec *begin,
+			      struct timespec *end)
+{
+	return 1000000000ull * (end->tv_sec - begin->tv_sec) + end->tv_nsec - begin->tv_nsec;
+}
+
+static clock_t correlation_clock_id = CLOCK_MONOTONIC;
+
+static bool
+get_correlation_timestamps(struct intel_perf_record_timestamp_correlation *corr, int drm_fd)
+{
+	struct drm_i915_reg_read reg_read;
+	struct {
+		struct timespec cpu_ts_begin;
+		struct timespec cpu_ts_end;
+		uint64_t gpu_ts;
+	} attempts[3];
+	uint32_t best = 0;
+
+#define RENDER_RING_TIMESTAMP 0x2358
+
+        reg_read.offset = RENDER_RING_TIMESTAMP | I915_REG_READ_8B_WA;
+
+	/* Gather 3 correlations. */
+	for (uint32_t i = 0; i < ARRAY_SIZE(attempts); i++) {
+		clock_gettime(correlation_clock_id, &attempts[i].cpu_ts_begin);
+		if (perf_ioctl(drm_fd, DRM_IOCTL_I915_REG_READ, &reg_read) < 0)
+			return false;
+		clock_gettime(correlation_clock_id, &attempts[i].cpu_ts_end);
+
+		attempts[i].gpu_ts = reg_read.val;
+	}
+
+	/* Now select the best. */
+	for (uint32_t i = 1; i < ARRAY_SIZE(attempts); i++) {
+		if (timespec_diff(&attempts[i].cpu_ts_begin,
+				  &attempts[i].cpu_ts_end) <
+		    timespec_diff(&attempts[best].cpu_ts_begin,
+				  &attempts[best].cpu_ts_end))
+			best = i;
+	}
+
+	corr->cpu_timestamp =
+		(attempts[best].cpu_ts_begin.tv_sec * 1000000000ull +
+		 attempts[best].cpu_ts_begin.tv_nsec) +
+		timespec_diff(&attempts[best].cpu_ts_begin,
+			      &attempts[best].cpu_ts_end) / 2;
+	corr->gpu_timestamp = attempts[best].gpu_ts;
+
+	return true;
+}
+
+static bool
+write_saved_correlation_timestamps(FILE *output,
+				   const struct intel_perf_record_timestamp_correlation *corr)
+{
+	struct drm_i915_perf_record_header header = {
+		.type = INTEL_PERF_RECORD_TYPE_TIMESTAMP_CORRELATION,
+		.size = sizeof(header) + sizeof(*corr),
+	};
+
+	if (fwrite(&header, sizeof(header), 1, output) != 1)
+		return false;
+
+	if (fwrite(corr, sizeof(*corr), 1, output) != 1)
+		return false;
+
+	return true;
+}
+
+static bool
+write_correlation_timestamps(FILE *output, int drm_fd)
+{
+	struct intel_perf_record_timestamp_correlation corr;
+
+	if (!get_correlation_timestamps(&corr, drm_fd))
+		return false;
+
+	return write_saved_correlation_timestamps(output, &corr);
+}
+
+static void
+read_command_file(int command_fd, FILE *output_stream, struct circular_buffer *buffer,
+		  int drm_fd, uint32_t devid, uint64_t timestamp_frequency,
+		  struct intel_perf_metric_set *metric_set)
+{
+	struct recorder_command_base header;
+	ssize_t ret = read(command_fd, &header, sizeof(header));
+
+	if (ret < 0)
+		return;
+
+	switch (header.command) {
+	case RECORDER_COMMAND_DUMP: {
+		uint32_t len = header.size - sizeof(header), offset = 0;
+		struct recorder_command_dump *dump = malloc(len);
+		FILE *file;
+
+		while (offset < len &&
+		       ((ret = read(command_fd, (void *) dump + offset, len - offset)) > 0
+			|| errno == EAGAIN)) {
+			if (ret > 0)
+				offset += ret;
+		}
+
+		fprintf(stdout, "Writing circular buffer to %s\n", dump->path);
+
+		file = fopen((const char *) dump->path, "w+");
+		if (file) {
+			struct chunk chunks[2];
+
+			fflush(output_stream);
+			get_chunks(chunks, buffer, false, buffer->size);
+
+			if (!write_header(file, devid, timestamp_frequency, metric_set) ||
+			    !write_topology(file, drm_fd) ||
+			    fwrite(chunks[0].data, chunks[0].len, 1, file) != 1 ||
+			    (chunks[1].len > 0 &&
+			     fwrite(chunks[1].data, chunks[1].len, 1, file) != 1) ||
+			    !write_correlation_timestamps(file, drm_fd)) {
+				fprintf(stderr, "Unable to write circular buffer data in file '%s'\n",
+					dump->path);
+			}
+			fclose(file);
+		} else
+			fprintf(stderr, "Unable to write dump file '%s'\n", dump->path);
+
+		free(dump);
+		break;
+	}
+	case RECORDER_COMMAND_QUIT:
+		quit = true;
+		break;
+	default:
+		fprintf(stderr, "Unknown command 0x%x\n", header.command);
+		break;
+	}
+}
+
+static void
+print_metric_sets(const struct intel_perf *perf)
+{
+	struct intel_perf_metric_set *metric_set;
+	uint32_t longest_name = 0;
+
+	igt_list_for_each_entry(metric_set, &perf->metric_sets, link) {
+		longest_name = MAX(longest_name, strlen(metric_set->symbol_name));
+	}
+
+	igt_list_for_each_entry(metric_set, &perf->metric_sets, link) {
+		fprintf(stdout, "%s:%*s%s\t\n",
+			metric_set->symbol_name,
+			(int) (longest_name - strlen(metric_set->symbol_name) + 1), " ",
+			metric_set->name);
+	}
+}
+
+static void
+print_metric_set_counters(const struct intel_perf_metric_set *metric_set)
+{
+	uint32_t longest_name = 0;
+	for (uint32_t i = 0; i < metric_set->n_counters; i++) {
+		longest_name = MAX(longest_name, strlen(metric_set->counters[i].name));
+	}
+
+	fprintf(stdout, "Metric set %s:\n", metric_set->name);
+	for (uint32_t i = 0; i < metric_set->n_counters; i++) {
+		struct intel_perf_logical_counter *counter = &metric_set->counters[i];
+
+		fprintf(stdout, "%s:%*s%s\n",
+			counter->name,
+			(int)(longest_name - strlen(counter->name) + 1), " ",
+			counter->desc);
+	}
+}
+
+static void
+usage(const char *name)
+{
+	fprintf(stdout,
+		"Usage: %s [options]\n"
+		"\n"
+		"     --help,               -h          Print this screen\n"
+		"     --correlation-period, -c <value>  Time period of timestamp correlation in seconds\n"
+		"                                       (default = 1.0)\n"
+		"     --perf-period,        -p <value>  Time period of i915-perf reports in seconds\n"
+		"                                       (default = 0.001)\n"
+		"     --metric,             -m <value>  i915 metric to sample with\n"
+		"     --counters,           -C          List counters for a given metric and exit\n"
+		"     --size,               -s <value>  Size of circular buffer to use in kilobytes\n"
+		"                                       If specified, a maximum amount of <value> data will\n"
+		"                                       be recorded.\n"
+		"     --command-fifo,       -f <path>   Path to a command fifo, implies circular buffer\n"
+		"                                       (To use with i915-perf-control)\n"
+		"     --output,             -o <path>   Output file (default = i915_perf.record)\n"
+		"     --cpu-clock,          -k <path>   Cpu clock to use for correlations\n"
+		"                                       Values: boot, mono, mono_raw (default = mono)\n",
+		name);
+}
+
+int
+main(int argc, char *argv[])
+{
+	const struct option long_options[] = {
+		{"help",                       no_argument, 0, 'h'},
+		{"correlation-period",   required_argument, 0, 'c'},
+		{"perf-period",          required_argument, 0, 'p'},
+		{"metric",               required_argument, 0, 'm'},
+		{"counters",                   no_argument, 0, 'C'},
+		{"output",               required_argument, 0, 'o'},
+		{"size",                 required_argument, 0, 's'},
+		{"command-fifo",         required_argument, 0, 'f'},
+		{"cpu-clock",            required_argument, 0, 'k'},
+		{0, 0, 0, 0}
+	};
+	const struct {
+		clock_t id;
+		const char *name;
+	} clock_names[] = {
+		{ CLOCK_BOOTTIME,      "boot" },
+		{ CLOCK_MONOTONIC,     "mono" },
+		{ CLOCK_MONOTONIC_RAW, "mono_raw" },
+	};
+	const struct intel_device_info *devinfo;
+	double corr_period = 1.0, perf_period = 0.001;
+	const char *metric_name = NULL, *output_file = "i915_perf.record", *command_fifo = I915_PERF_RECORD_FIFO_PATH;
+	struct intel_perf *perf;
+	struct intel_perf_metric_set *metric_set, *selected_metric_set = NULL;
+	struct intel_perf_record_timestamp_correlation initial_correlation;
+	struct circular_buffer circular_buffer;
+	struct timespec now;
+	uint64_t corr_period_ns, poll_time_ns, timestamp_frequency;
+	uint32_t devid = 0, oa_exponent;
+	uint32_t circular_size = 0;
+	int drm_fd, perf_fd, command_fifo_fd = -1;
+	int opt;
+	bool list_counters = false;
+	FILE *output = NULL, *output_stream;
+
+	while ((opt = getopt_long(argc, argv, "hc:p:m:Co:s:f:k:", long_options, NULL)) != -1) {
+		switch (opt) {
+		case 'h':
+			usage(argv[0]);
+			return EXIT_SUCCESS;
+		case 'c':
+			corr_period = atof(optarg);
+			break;
+		case 'p':
+			perf_period = atof(optarg);
+			break;
+		case 'm':
+			metric_name = optarg;
+			break;
+		case 'C':
+			list_counters = true;
+			break;
+		case 'o':
+			output_file = optarg;
+			break;
+		case 's':
+			circular_size = MAX(8, atoi(optarg)) * 1024;
+			break;
+		case 'f':
+			command_fifo = optarg;
+			circular_size = 8 * 1024 * 1024;
+			break;
+		case 'k': {
+			bool found = false;
+			for (uint32_t i = 0; i < ARRAY_SIZE(clock_names); i++) {
+				if (!strcmp(clock_names[i].name, optarg)) {
+					correlation_clock_id = clock_names[i].id;
+					found = true;
+					break;
+				}
+			}
+			if (!found) {
+				fprintf(stderr, "Unknown clock name '%s'\n", optarg);
+				return EXIT_FAILURE;
+			}
+			break;
+		}
+		default:
+			fprintf(stderr, "Internal error: "
+				"unexpected getopt value: %d\n", opt);
+			usage(argv[0]);
+			return EXIT_FAILURE;
+		}
+	}
+
+	drm_fd = open_render_node(&devid);
+
+	devinfo = intel_get_device_info(devid);
+	if (!devinfo) {
+		fprintf(stderr, "No device info found.\n");
+		return EXIT_FAILURE;
+	}
+
+	fprintf(stdout, "Device name=%s gen=%i gt=%i id=0x%x\n",
+		devinfo->codename, devinfo->gen, devinfo->gt, devid);
+
+	perf = intel_perf_for_devinfo(devinfo);
+	if (!perf) {
+		fprintf(stderr, "No perf data found.\n");
+		return EXIT_FAILURE;
+	}
+
+	if (!metric_name) {
+		print_metric_sets(perf);
+		return EXIT_FAILURE;
+	}
+
+	igt_list_for_each_entry(metric_set, &perf->metric_sets, link) {
+		if (!strcasecmp(metric_set->symbol_name, metric_name)) {
+			selected_metric_set = metric_set;
+			break;
+		}
+	}
+
+	if (!selected_metric_set) {
+		fprintf(stderr, "Unknown metric set '%s'\n", metric_name);
+		print_metric_sets(perf);
+		return EXIT_FAILURE;
+	}
+
+	if (list_counters) {
+		print_metric_set_counters(selected_metric_set);
+		return EXIT_SUCCESS;
+	}
+
+	intel_perf_load_perf_configs(perf, drm_fd);
+
+	timestamp_frequency = get_device_timestamp_frequency(devinfo, drm_fd);
+
+	signal(SIGINT, sigint_handler);
+
+	if (command_fifo) {
+		if (mkfifo(command_fifo, S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP | S_IROTH | S_IWOTH) != 0) {
+			fprintf(stderr, "Unable to create command fifo '%s': %s\n",
+				command_fifo, strerror(errno));
+			return EXIT_FAILURE;
+		}
+
+		command_fifo_fd = open(command_fifo, O_RDWR);
+		if (command_fifo_fd < 0) {
+			fprintf(stderr, "Unable to open command fifo '%s': %s\n",
+				command_fifo, strerror(errno));
+			return EXIT_FAILURE;
+		}
+	} else {
+		output = fopen(output_file, "w+");
+		if (!output) {
+			fprintf(stderr, "Unable to open output file '%s'\n",
+				output_file);
+			return EXIT_FAILURE;
+		}
+	}
+
+	if (circular_size) {
+		memset(&circular_buffer, 0, sizeof(circular_buffer));
+		circular_buffer.allocated_size = circular_size;
+		circular_buffer.data = malloc(circular_size);
+		if (!circular_buffer.data) {
+			fprintf(stderr, "Unable to allocate circular buffer\n");
+			return EXIT_FAILURE;
+		}
+
+		output_stream = fopencookie(&circular_buffer, "w+",
+					    circular_buffer_functions);
+		if (!output_stream) {
+			fprintf(stderr, "Unable to create circular buffer\n");
+			return EXIT_FAILURE;
+		}
+
+		if (!get_correlation_timestamps(&initial_correlation, drm_fd)) {
+			fprintf(stderr, "Unable to correlation timestamps\n");
+			return EXIT_FAILURE;
+		}
+
+		write_correlation_timestamps(output_stream, drm_fd);
+	} else {
+		if (!write_header(output, devid, timestamp_frequency, selected_metric_set) ||
+		    !write_topology(output, drm_fd) ||
+		    !write_correlation_timestamps(output, drm_fd)) {
+			fprintf(stderr, "Unable to write header in file '%s'\n",
+				output_file);
+			return EXIT_FAILURE;
+		}
+
+		output_stream = output;
+	}
+
+	if (selected_metric_set->perf_oa_metrics_set == 0) {
+		fprintf(stderr,
+			"Unable to load performance configuration, consider running:\n"
+			"   sysctl dev.i915.perf_stream_paranoid=0\n");
+		return EXIT_FAILURE;
+	}
+
+	oa_exponent = oa_exponent_for_period(timestamp_frequency, perf_period);
+	fprintf(stdout, "Opening perf stream with metric_id=%lu oa_exponent=%u\n",
+		selected_metric_set->perf_oa_metrics_set, oa_exponent);
+
+	perf_fd = perf_open(drm_fd, devinfo, oa_exponent, selected_metric_set);
+	if (perf_fd < 0) {
+		fprintf(stderr, "Unable to open i915 perf stream: %s\n",
+			strerror(errno));
+		return EXIT_FAILURE;
+	}
+
+	corr_period_ns = corr_period * 1000000000ul;
+	poll_time_ns = corr_period_ns;
+
+	while (!quit) {
+		struct pollfd pollfd[2] = {
+			{         perf_fd, POLLIN, 0 },
+			{ command_fifo_fd, POLLIN, 0 },
+		};
+		uint64_t elapsed_ns;
+		int ret;
+
+		igt_gettime(&now);
+		ret = poll(pollfd, command_fifo_fd != -1 ? 2 : 1, poll_time_ns / 1000000);
+		if (ret < 0 && errno != EINTR) {
+			fprintf(stderr, "Failed to poll i915-perf stream: %s\n",
+				strerror(errno));
+			break;
+		}
+
+		if (ret > 0) {
+			if (pollfd[0].revents & POLLIN) {
+				if (!write_i915_perf_data(output_stream, perf_fd)) {
+					fprintf(stderr, "Failed to write i915-perf data: %s\n",
+						strerror(errno));
+					break;
+				}
+			}
+
+			if (pollfd[1].revents & POLLIN) {
+				read_command_file(command_fifo_fd, output_stream,
+						  &circular_buffer,
+						  drm_fd, devid, timestamp_frequency,
+						  selected_metric_set);
+			}
+		}
+
+		elapsed_ns = igt_nsec_elapsed(&now);
+		if (elapsed_ns > poll_time_ns) {
+			poll_time_ns = corr_period_ns;
+			if (!write_correlation_timestamps(output_stream, drm_fd)) {
+				fprintf(stderr,
+					"Failed to write i915 timestamp correlation data: %s\n",
+					strerror(errno));
+				break;
+			}
+		} else {
+			poll_time_ns -= elapsed_ns;
+		}
+	}
+
+	fprintf(stdout, "Exiting...\n");
+
+	if (!write_correlation_timestamps(output_stream, drm_fd)) {
+		fprintf(stderr,
+			"Failed to write final i915 timestamp correlation data: %s\n",
+			strerror(errno));
+	}
+
+	fclose(output_stream);
+
+	if (command_fifo)
+		unlink(command_fifo);
+
+	free(circular_buffer.data);
+
+	close(drm_fd);
+
+	return EXIT_SUCCESS;
+}
diff --git a/tools/i915-perf/i915_perf_recorder_commands.h b/tools/i915-perf/i915_perf_recorder_commands.h
new file mode 100644
index 00000000..4855d80f
--- /dev/null
+++ b/tools/i915-perf/i915_perf_recorder_commands.h
@@ -0,0 +1,39 @@
+/*
+ * Copyright (C) 2019 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <stdint.h>
+
+#define I915_PERF_RECORD_FIFO_PATH "/tmp/.i915-perf-record"
+
+enum recorder_command {
+	RECORDER_COMMAND_DUMP = 1,
+	RECORDER_COMMAND_QUIT,
+};
+
+struct recorder_command_base {
+	uint32_t command;
+	uint32_t size;
+};
+
+struct recorder_command_dump {
+	uint8_t path[0];
+};
diff --git a/tools/i915-perf/meson.build b/tools/i915-perf/meson.build
index 0ebdd185..1be3ab22 100644
--- a/tools/i915-perf/meson.build
+++ b/tools/i915-perf/meson.build
@@ -3,3 +3,15 @@ executable('i915-perf-configs',
            include_directories: inc,
            dependencies: [lib_igt_chipset, lib_igt_i915_perf],
            install: true)
+
+executable('i915-perf-recorder',
+           [ 'i915_perf_recorder.c' ],
+           include_directories: inc,
+           dependencies: [lib_igt, lib_igt_i915_perf],
+           install: true)
+
+executable('i915-perf-control',
+           [ 'i915_perf_control.c' ],
+           c_args: '-U_FORTIFY_SOURCE',
+           include_directories: inc,
+           install: true)
-- 
2.20.1

_______________________________________________
igt-dev mailing list
igt-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/igt-dev

^ permalink raw reply related	[flat|nested] 5+ messages in thread

* [igt-dev] [PATCH i-g-t 4/4] lib/i915/perf: Add i915 perf data reader
  2020-02-15  1:11 [igt-dev] [PATCH i-g-t 0/4] Add perf OA tools for GPUvis Umesh Nerlige Ramappa
  2020-02-15  1:11 ` [igt-dev] [PATCH i-g-t 2/4] lib/i915/perf: Add support for loading perf configurations Umesh Nerlige Ramappa
  2020-02-15  1:11 ` [igt-dev] [PATCH i-g-t 3/4] tools/i915/perf: Add i915 perf recorder tool Umesh Nerlige Ramappa
@ 2020-02-15  1:11 ` Umesh Nerlige Ramappa
  2020-02-17 13:42 ` [igt-dev] [PATCH i-g-t 0/4] Add perf OA tools for GPUvis Lionel Landwerlin
  3 siblings, 0 replies; 5+ messages in thread
From: Umesh Nerlige Ramappa @ 2020-02-15  1:11 UTC (permalink / raw)
  To: igt-dev, Joonas Lahtinen, Ashutosh Dixit, Lionel G Landwerlin

From: Lionel Landwerlin <lionel.g.landwerlin@intel.com>

Read perf OA records and correlate timestamps between the GPU and CPU.

v2: (Umesh)
- Add README on usage
- rebase fixes for igt_list

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com>
---
 lib/i915/perf_data_reader.c | 330 ++++++++++++++++++++++++++++++++++++
 lib/i915/perf_data_reader.h | 103 +++++++++++
 lib/meson.build             |   2 +
 tools/i915-perf/README      |  70 ++++++++
 4 files changed, 505 insertions(+)
 create mode 100644 lib/i915/perf_data_reader.c
 create mode 100644 lib/i915/perf_data_reader.h
 create mode 100644 tools/i915-perf/README

diff --git a/lib/i915/perf_data_reader.c b/lib/i915/perf_data_reader.c
new file mode 100644
index 00000000..43683331
--- /dev/null
+++ b/lib/i915/perf_data_reader.c
@@ -0,0 +1,330 @@
+/*
+ * Copyright (C) 2019 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <assert.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/mman.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <unistd.h>
+
+#include "intel_chipset.h"
+#include "perf.h"
+#include "perf_data_reader.h"
+
+#define MAX(a,b) ((a) > (b) ? (a) : (b))
+#define ARRAY_SIZE(arr) (sizeof(arr) / sizeof((arr)[0]))
+
+static inline bool
+oa_report_ctx_is_valid(const struct intel_perf_devinfo *devinfo,
+		       const uint8_t *_report)
+{
+	const uint32_t *report = (const uint32_t *) _report;
+
+	if (devinfo->gen < 8) {
+		return false; /* TODO */
+	} else if (devinfo->gen == 8) {
+		return report[0] & (1ul << 25);
+	} else if (devinfo->gen > 8) {
+		return report[0] & (1ul << 16);
+	}
+
+	return false;
+}
+
+static uint32_t
+oa_report_ctx_id(const struct intel_perf_devinfo *devinfo, const uint8_t *report)
+{
+	if (!oa_report_ctx_is_valid(devinfo, report))
+		return 0xffffffff;
+	return ((const uint32_t *) report)[2];
+}
+
+static inline uint64_t
+oa_report_timestamp(const uint8_t *report)
+{
+	return ((const uint32_t *)report)[1];
+}
+
+static void
+append_record(struct intel_perf_data_reader *reader,
+	      const struct drm_i915_perf_record_header *header)
+{
+	if (reader->n_records >= reader->n_allocated_records) {
+		reader->n_allocated_records = MAX(100, 2 * reader->n_allocated_records);
+		reader->records =
+			(const struct drm_i915_perf_record_header **)
+			realloc((void *) reader->records,
+				reader->n_allocated_records *
+				sizeof(*reader->records));
+		assert(reader->records);
+	}
+
+	reader->records[reader->n_records++] = header;
+}
+
+static void
+append_timestamp_correlation(struct intel_perf_data_reader *reader,
+			     const struct intel_perf_record_timestamp_correlation *corr)
+{
+	if (reader->n_correlations >= reader->n_allocated_correlations) {
+		reader->n_allocated_correlations = MAX(100, 2 * reader->n_allocated_correlations);
+		reader->correlations =
+			(const struct intel_perf_record_timestamp_correlation **)
+			realloc((void *) reader->correlations,
+				reader->n_allocated_correlations *
+				sizeof(*reader->correlations));
+		assert(reader->correlations);
+	}
+
+	reader->correlations[reader->n_correlations++] = corr;
+}
+
+static struct intel_perf_metric_set *
+find_metric_set(struct intel_perf *perf, const char *uuid)
+{
+	struct intel_perf_metric_set *metric_set;
+
+	igt_list_for_each_entry(metric_set, &perf->metric_sets, link) {
+		if (!strcmp(uuid, metric_set->hw_config_guid))
+			return metric_set;
+	}
+
+	return NULL;
+}
+
+static void
+init_devinfo(struct intel_perf_devinfo *perf_devinfo,
+	     const struct intel_device_info *devinfo,
+	     uint32_t devid,
+	     uint64_t timestamp_frequency)
+{
+	perf_devinfo->devid = devid;
+	perf_devinfo->gen = devinfo->gen;
+	perf_devinfo->timestamp_frequency = timestamp_frequency;
+}
+
+static bool
+parse_data(struct intel_perf_data_reader *reader)
+{
+	const uint8_t *end = reader->mmap_data + reader->mmap_size;
+	const uint8_t *iter = reader->mmap_data;
+	while (iter < end) {
+		const struct drm_i915_perf_record_header *header =
+			(const struct drm_i915_perf_record_header *) iter;
+
+		switch (header->type) {
+		case DRM_I915_PERF_RECORD_SAMPLE:
+			append_record(reader, header);
+			break;
+
+		case DRM_I915_PERF_RECORD_OA_REPORT_LOST:
+		case DRM_I915_PERF_RECORD_OA_BUFFER_LOST:
+			assert(header->size == sizeof(*header));
+			break;
+
+		case INTEL_PERF_RECORD_TYPE_DEVICE_INFO: {
+			const struct intel_device_info *devinfo;
+
+			reader->record_info =
+				(const struct intel_perf_record_device_info *) (header + 1);
+			assert(header->size == (sizeof(*reader->record_info) + sizeof(*header)));
+			devinfo = intel_get_device_info(reader->record_info->device_id);
+			if (!devinfo)
+				return false;
+			init_devinfo(&reader->devinfo, devinfo,
+				     reader->record_info->device_id,
+				     reader->record_info->timestamp_frequency);
+			reader->perf = intel_perf_for_devinfo(devinfo);
+			reader->metric_set = find_metric_set(reader->perf, reader->record_info->uuid);
+			break;
+		}
+
+		case INTEL_PERF_RECORD_TYPE_TIMESTAMP_CORRELATION: {
+			append_timestamp_correlation(reader,
+						     (const struct intel_perf_record_timestamp_correlation *) (header + 1));
+			break;
+		}
+		}
+
+		iter += header->size;
+	}
+
+	return true;
+}
+
+static uint64_t
+correlate_gpu_timestamp(struct intel_perf_data_reader *reader,
+			uint64_t gpu_ts)
+{
+	/* OA reports only have the lower 32bits of the timestamp
+	 * register, while our correlation data has the whole 36bits.
+	 * Try to figure what portion of the correlation data the
+	 * 32bit timestamp belongs to.
+	 */
+	uint64_t mask = 0xffffffff;
+	int corr_idx = -1;
+
+	for (uint32_t i = 0; i < reader->n_correlation_chunks; i++) {
+		if (gpu_ts >= (reader->correlation_chunks[i].gpu_ts_begin & mask) &&
+		    gpu_ts <= (reader->correlation_chunks[i].gpu_ts_end & mask)) {
+			corr_idx = reader->correlation_chunks[i].idx;
+			break;
+		}
+	}
+
+	/* Not found? Assume prior to the first timestamp correlation.
+	 */
+	if (corr_idx < 0) {
+		return reader->correlations[0]->cpu_timestamp -
+			((reader->correlations[0]->gpu_timestamp & mask) - gpu_ts) *
+			(reader->correlations[1]->cpu_timestamp - reader->correlations[0]->cpu_timestamp) /
+			(reader->correlations[1]->gpu_timestamp - reader->correlations[0]->gpu_timestamp);
+	}
+
+	for (uint32_t i = corr_idx; i < (reader->n_correlations - 1); i++) {
+		if (gpu_ts >= (reader->correlations[i]->gpu_timestamp & mask) &&
+		    gpu_ts < (reader->correlations[i + 1]->gpu_timestamp & mask)) {
+			return reader->correlations[i]->cpu_timestamp +
+				(gpu_ts - (reader->correlations[i]->gpu_timestamp & mask)) *
+				(reader->correlations[i + 1]->cpu_timestamp - reader->correlations[i]->cpu_timestamp) /
+				(reader->correlations[i + 1]->gpu_timestamp - reader->correlations[i]->gpu_timestamp);
+		}
+	}
+
+	/* This is a bit harsh, but the recording tool should ensure we have
+	 * sampling points on either side of the bag of OA reports.
+	 */
+	assert(0);
+}
+
+static void
+append_timeline_event(struct intel_perf_data_reader *reader,
+		      uint64_t ts_start, uint64_t ts_end,
+		      uint32_t record_start, uint32_t record_end,
+		      uint32_t hw_id)
+{
+	if (reader->n_timelines >= reader->n_allocated_timelines) {
+		reader->n_allocated_timelines = MAX(100, 2 * reader->n_allocated_timelines);
+		reader->timelines =
+			(struct intel_perf_timeline_item *)
+			realloc((void *) reader->timelines,
+				reader->n_allocated_timelines *
+				sizeof(*reader->timelines));
+		assert(reader->timelines);
+	}
+
+	reader->timelines[reader->n_timelines].ts_start = ts_start;
+	reader->timelines[reader->n_timelines].ts_end = ts_end;
+	reader->timelines[reader->n_timelines].cpu_ts_start =
+		correlate_gpu_timestamp(reader, ts_start);
+	reader->timelines[reader->n_timelines].cpu_ts_end =
+		correlate_gpu_timestamp(reader, ts_end);
+	reader->timelines[reader->n_timelines].record_start = record_start;
+	reader->timelines[reader->n_timelines].record_end = record_end;
+	reader->timelines[reader->n_timelines].hw_id = hw_id;
+	reader->n_timelines++;
+}
+
+static void
+generate_cpu_events(struct intel_perf_data_reader *reader)
+{
+	uint32_t last_header_idx = 0;
+	const struct drm_i915_perf_record_header *last_header = reader->records[0];
+
+	for (uint32_t i = 1; i < reader->n_records; i++) {
+		const struct drm_i915_perf_record_header *current_header =
+			reader->records[i];
+		const uint8_t *start_report = (const uint8_t *) (last_header + 1),
+			*end_report = (const uint8_t *) (current_header + 1);
+		uint32_t last_ctx_id = oa_report_ctx_id(&reader->devinfo, start_report),
+			current_ctx_id = oa_report_ctx_id(&reader->devinfo, end_report);
+		uint64_t gpu_ts_start = oa_report_timestamp(start_report),
+			gpu_ts_end = oa_report_timestamp(end_report);
+
+		if (last_ctx_id == current_ctx_id)
+			continue;
+
+		append_timeline_event(reader, gpu_ts_start, gpu_ts_end, last_header_idx, i, last_ctx_id);
+
+		last_header = current_header;
+		last_header_idx = i;
+	}
+}
+
+static void
+compute_correlation_chunks(struct intel_perf_data_reader *reader)
+{
+	uint64_t mask = ~(0xffffffff);
+	uint32_t last_idx = 0;
+	uint64_t last_ts = reader->correlations[last_idx]->gpu_timestamp;
+
+	for (uint32_t i = 0; i < reader->n_correlations; i++) {
+		if (!reader->n_correlation_chunks ||
+		    (last_ts & mask) != (reader->correlations[i]->gpu_timestamp & mask)) {
+			assert(reader->n_correlation_chunks < ARRAY_SIZE(reader->correlation_chunks));
+			reader->correlation_chunks[reader->n_correlation_chunks].gpu_ts_begin = last_ts;
+			reader->correlation_chunks[reader->n_correlation_chunks].gpu_ts_end = last_ts | ~mask;
+			reader->correlation_chunks[reader->n_correlation_chunks].idx = last_idx;
+			last_ts = reader->correlation_chunks[reader->n_correlation_chunks].gpu_ts_end + 1;
+			last_idx = i;
+			reader->n_correlation_chunks++;
+		}
+	}
+}
+
+bool
+intel_perf_data_reader_init(struct intel_perf_data_reader *reader,
+			    int perf_file_fd)
+{
+        struct stat st;
+        if (fstat(perf_file_fd, &st) != 0)
+		return false;
+
+	memset(reader, 0, sizeof(*reader));
+
+	reader->mmap_size = st.st_size;
+	reader->mmap_data = (const uint8_t *) mmap(NULL, st.st_size,
+						   PROT_READ, MAP_PRIVATE,
+						   perf_file_fd, 0);
+	if (reader->mmap_data == MAP_FAILED)
+		return false;
+
+	if (!parse_data(reader))
+		return false;
+
+	compute_correlation_chunks(reader);
+	generate_cpu_events(reader);
+
+	return true;
+}
+
+void
+intel_perf_data_reader_fini(struct intel_perf_data_reader *reader)
+{
+	intel_perf_free(reader->perf);
+	free(reader->records);
+	free(reader->timelines);
+	free(reader->correlations);
+	munmap((void *)reader->mmap_data, reader->mmap_size);
+}
diff --git a/lib/i915/perf_data_reader.h b/lib/i915/perf_data_reader.h
new file mode 100644
index 00000000..f75e96dd
--- /dev/null
+++ b/lib/i915/perf_data_reader.h
@@ -0,0 +1,103 @@
+/*
+ * Copyright (C) 2019 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef PERF_DATA_READER_H
+#define PERF_DATA_READER_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Helper to read a i915-perf recording. */
+
+#include <stdbool.h>
+#include <stdint.h>
+
+#include <i915_drm.h>
+
+#include "perf.h"
+#include "perf_data.h"
+
+struct intel_device_info;
+
+struct intel_perf_timeline_item {
+	uint64_t ts_start;
+	uint64_t ts_end;
+	uint64_t cpu_ts_start;
+	uint64_t cpu_ts_end;
+
+	/* Offsets into intel_perf_data_reader.records */
+	uint32_t record_start;
+	uint32_t record_end;
+
+	uint32_t hw_id;
+
+	/* User associated data with a given item on the i915 perf
+	 * timeline.
+	 */
+	void *user_data;
+};
+
+struct intel_perf_data_reader {
+	/* Array of pointers into the mmapped i915 perf file. */
+	const struct drm_i915_perf_record_header **records;
+	uint32_t n_records;
+	uint32_t n_allocated_records;
+
+	/**/
+	struct intel_perf_timeline_item *timelines;
+	uint32_t n_timelines;
+	uint32_t n_allocated_timelines;
+
+	/**/
+	const struct intel_perf_record_timestamp_correlation **correlations;
+	uint32_t n_correlations;
+	uint32_t n_allocated_correlations;
+
+	struct {
+		uint64_t gpu_ts_begin;
+		uint64_t gpu_ts_end;
+		uint32_t idx;
+	} correlation_chunks[4];
+	uint32_t n_correlation_chunks;
+
+	/**/
+	const struct intel_perf_record_device_info *record_info;
+
+	struct intel_perf_devinfo devinfo;
+
+	struct intel_perf *perf;
+	struct intel_perf_metric_set *metric_set;
+
+	const uint8_t *mmap_data;
+	size_t mmap_size;
+};
+
+bool intel_perf_data_reader_init(struct intel_perf_data_reader *reader,
+				 int perf_file_fd);
+void intel_perf_data_reader_fini(struct intel_perf_data_reader *reader);
+
+#ifdef __cplusplus
+};
+#endif
+
+#endif /* PERF_DATA_READER_H */
diff --git a/lib/meson.build b/lib/meson.build
index 6e935d45..f241bff7 100644
--- a/lib/meson.build
+++ b/lib/meson.build
@@ -173,6 +173,7 @@ lib_igt_perf = declare_dependency(link_with : lib_igt_perf_build,
 
 i915_perf_files = [
   'i915/perf.c',
+  'i915/perf_data_reader.c',
 ]
 
 i915_perf_hardware = [
@@ -218,6 +219,7 @@ install_headers(
   'intel_chipset.h',
   'i915/perf.h',
   'i915/perf_data.h',
+  'i915/perf_data_reader.h',
   subdir : 'i915-perf'
 )
 
diff --git a/tools/i915-perf/README b/tools/i915-perf/README
new file mode 100644
index 00000000..e9822345
--- /dev/null
+++ b/tools/i915-perf/README
@@ -0,0 +1,70 @@
+======================
+i915 perf tools for OA
+======================
+
+The tools provided here enable capturing performance metrics from the i915
+driver and are used in conjunction with the GPUvis software here - 
+
+https://github.com/mikesart/gpuvis
+
+Tools in IGT
+------------
+
+The following tools are generated in build/tools/i915-perf
+
+i915-perf-configs
+i915-perf-control
+i915-perf-recorder
+
+Usage in IGT
+------------
+
+Just launching i915-perf-recorder with no argument will list all available
+metrics. Once installed, the igt recorder tool can be used to record metrics in
+a circular buffer. Example below shows capture of RenderBasic metrics with an
+8Mb circular buffer.
+
+i915-perf-recorder -m RenderBasic -s 8192
+
+The circular buffer can be dumped at a given location from another terminal
+using the i915-perf-control tool :
+
+i915-perf-control -d /tmp/recording.perf
+ 
+Integration with GPUvis
+-----------------------
+
+GPUvis provides sample scripts in gpuvis/sample directory that can be modified
+and used to capture the metrics required.
+
+1. Setup the recording by launching the following scripts from gpuvis/sample
+   directory : 
+
+        trace-cmd-setup.sh
+        trace-cmd-start-tracing.sh
+
+This will setup a recording in a circular buffer.
+ 
+2. Start using the system for a specific task you want to record.
+
+3. Once the task is completed, save the circular buffer into a capture file with
+   the following script :
+
+        trace-cmd-capture.sh
+ 
+4. Once finished, tear down the circular buffer recording with :
+
+        trace-cmd-stop-tracing.sh
+
+Inspecting data captured in GPUvis
+----------------------------------
+ 
+The capture script will generate 2 files for instance : 
+
+        trace_09-26-2019_01-22-40.dat
+        trace_09-26-2019_01-22-40.i915-dat
+
+The first one contains ftrace data, the other i915-perf data. To inspect the
+data launch gpuvis with the 2 files as arguments :
+
+        gpuvis trace_09-26-2019_01-22-40.dat trace_09-26-2019_01-22-40.i915-dat
-- 
2.20.1

_______________________________________________
igt-dev mailing list
igt-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/igt-dev

^ permalink raw reply related	[flat|nested] 5+ messages in thread

* Re: [igt-dev] [PATCH i-g-t 0/4] Add perf OA tools for GPUvis
  2020-02-15  1:11 [igt-dev] [PATCH i-g-t 0/4] Add perf OA tools for GPUvis Umesh Nerlige Ramappa
                   ` (2 preceding siblings ...)
  2020-02-15  1:11 ` [igt-dev] [PATCH i-g-t 4/4] lib/i915/perf: Add i915 perf data reader Umesh Nerlige Ramappa
@ 2020-02-17 13:42 ` Lionel Landwerlin
  3 siblings, 0 replies; 5+ messages in thread
From: Lionel Landwerlin @ 2020-02-17 13:42 UTC (permalink / raw)
  To: Umesh Nerlige Ramappa, igt-dev, Joonas Lahtinen, Ashutosh Dixit

There was a small communication hiccup. Umesh thought I did not work on 
this stuff anymore, but I actually just picked up the stuff again last week.

Sending an update with more changes/updates.

Sorry for the confusion.

-Lionel

On 15/02/2020 03:11, Umesh Nerlige Ramappa wrote:
> The tools provided here enable capturing performance metrics from the i915
> driver and are used in conjunction with the GPUvis software here -
>
> https://github.com/mikesart/gpuvis
>
> The changes required in GPUvis are wip and will be posted following the merge of
> these tools.
>
> For more information, view tools/i915-perf/README in this patch series
>
> Lionel Landwerlin (4):
>    lib/i915/perf: Add i915_perf library
>    lib/i915/perf: Add support for loading perf configurations
>    tools/i915/perf: Add i915 perf recorder tool
>    lib/i915/perf: Add i915 perf data reader
>
>   lib/i915-perf.pc.in                           |    10 +
>   lib/i915/perf-configs/README.md               |   115 +
>   lib/i915/perf-configs/codegen.py              |    33 +
>   lib/i915/perf-configs/guids.xml               |   282 +
>   lib/i915/perf-configs/mdapi-xml-convert.py    |  1000 +
>   lib/i915/perf-configs/oa-bdw.xml              | 15653 ++++++++++++++++
>   lib/i915/perf-configs/oa-bxt.xml              |  9595 ++++++++++
>   lib/i915/perf-configs/oa-cflgt2.xml           | 10866 +++++++++++
>   lib/i915/perf-configs/oa-cflgt3.xml           | 10933 +++++++++++
>   lib/i915/perf-configs/oa-chv.xml              |  9757 ++++++++++
>   lib/i915/perf-configs/oa-cnl.xml              | 10411 ++++++++++
>   lib/i915/perf-configs/oa-glk.xml              |  9346 +++++++++
>   lib/i915/perf-configs/oa-hsw.xml              |  4615 +++++
>   lib/i915/perf-configs/oa-icl.xml              | 11899 ++++++++++++
>   lib/i915/perf-configs/oa-kblgt2.xml           | 10866 +++++++++++
>   lib/i915/perf-configs/oa-kblgt3.xml           | 10933 +++++++++++
>   lib/i915/perf-configs/oa-sklgt2.xml           | 11895 ++++++++++++
>   lib/i915/perf-configs/oa-sklgt3.xml           | 10933 +++++++++++
>   lib/i915/perf-configs/oa-sklgt4.xml           | 10956 +++++++++++
>   lib/i915/perf-configs/perf-codegen.py         |   854 +
>   lib/i915/perf-configs/update-guids.py         |   230 +
>   lib/i915/perf.c                               |   332 +
>   lib/i915/perf.h                               |   240 +
>   lib/i915/perf_data.h                          |    88 +
>   lib/i915/perf_data_reader.c                   |   330 +
>   lib/i915/perf_data_reader.h                   |   103 +
>   lib/meson.build                               |    67 +
>   tools/i915-perf/README                        |    70 +
>   tools/i915-perf/i915_perf_configs.c           |   277 +
>   tools/i915-perf/i915_perf_control.c           |   133 +
>   tools/i915-perf/i915_perf_recorder.c          |   931 +
>   tools/i915-perf/i915_perf_recorder_commands.h |    39 +
>   tools/i915-perf/meson.build                   |    17 +
>   tools/meson.build                             |     1 +
>   34 files changed, 153810 insertions(+)
>   create mode 100644 lib/i915-perf.pc.in
>   create mode 100644 lib/i915/perf-configs/README.md
>   create mode 100644 lib/i915/perf-configs/codegen.py
>   create mode 100644 lib/i915/perf-configs/guids.xml
>   create mode 100755 lib/i915/perf-configs/mdapi-xml-convert.py
>   create mode 100644 lib/i915/perf-configs/oa-bdw.xml
>   create mode 100644 lib/i915/perf-configs/oa-bxt.xml
>   create mode 100644 lib/i915/perf-configs/oa-cflgt2.xml
>   create mode 100644 lib/i915/perf-configs/oa-cflgt3.xml
>   create mode 100644 lib/i915/perf-configs/oa-chv.xml
>   create mode 100644 lib/i915/perf-configs/oa-cnl.xml
>   create mode 100644 lib/i915/perf-configs/oa-glk.xml
>   create mode 100644 lib/i915/perf-configs/oa-hsw.xml
>   create mode 100644 lib/i915/perf-configs/oa-icl.xml
>   create mode 100644 lib/i915/perf-configs/oa-kblgt2.xml
>   create mode 100644 lib/i915/perf-configs/oa-kblgt3.xml
>   create mode 100644 lib/i915/perf-configs/oa-sklgt2.xml
>   create mode 100644 lib/i915/perf-configs/oa-sklgt3.xml
>   create mode 100644 lib/i915/perf-configs/oa-sklgt4.xml
>   create mode 100755 lib/i915/perf-configs/perf-codegen.py
>   create mode 100755 lib/i915/perf-configs/update-guids.py
>   create mode 100644 lib/i915/perf.c
>   create mode 100644 lib/i915/perf.h
>   create mode 100644 lib/i915/perf_data.h
>   create mode 100644 lib/i915/perf_data_reader.c
>   create mode 100644 lib/i915/perf_data_reader.h
>   create mode 100644 tools/i915-perf/README
>   create mode 100644 tools/i915-perf/i915_perf_configs.c
>   create mode 100644 tools/i915-perf/i915_perf_control.c
>   create mode 100644 tools/i915-perf/i915_perf_recorder.c
>   create mode 100644 tools/i915-perf/i915_perf_recorder_commands.h
>   create mode 100644 tools/i915-perf/meson.build
>

_______________________________________________
igt-dev mailing list
igt-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/igt-dev

^ permalink raw reply	[flat|nested] 5+ messages in thread

end of thread, other threads:[~2020-02-17 13:42 UTC | newest]

Thread overview: 5+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2020-02-15  1:11 [igt-dev] [PATCH i-g-t 0/4] Add perf OA tools for GPUvis Umesh Nerlige Ramappa
2020-02-15  1:11 ` [igt-dev] [PATCH i-g-t 2/4] lib/i915/perf: Add support for loading perf configurations Umesh Nerlige Ramappa
2020-02-15  1:11 ` [igt-dev] [PATCH i-g-t 3/4] tools/i915/perf: Add i915 perf recorder tool Umesh Nerlige Ramappa
2020-02-15  1:11 ` [igt-dev] [PATCH i-g-t 4/4] lib/i915/perf: Add i915 perf data reader Umesh Nerlige Ramappa
2020-02-17 13:42 ` [igt-dev] [PATCH i-g-t 0/4] Add perf OA tools for GPUvis Lionel Landwerlin

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox