[PATCH i-g-t 11/16] tools/xe-perf: Add xe_perf_recorder
Ashutosh Dixit
ashutosh.dixit at intel.com
Fri Jun 28 20:43:43 UTC 2024
Add xe_perf_recorder. The recorder exercises the Xe OA uapi to capture OAG
data in a specific format, similar to i915_perf_recorder.
Signed-off-by: Ashutosh Dixit <ashutosh.dixit at intel.com>
Reviewed-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa at intel.com>
---
tools/meson.build | 1 +
tools/xe-perf/meson.build | 5 +
tools/xe-perf/xe_perf_recorder.c | 1227 +++++++++++++++++++++
tools/xe-perf/xe_perf_recorder_commands.h | 31 +
4 files changed, 1264 insertions(+)
create mode 100644 tools/xe-perf/meson.build
create mode 100644 tools/xe-perf/xe_perf_recorder.c
create mode 100644 tools/xe-perf/xe_perf_recorder_commands.h
diff --git a/tools/meson.build b/tools/meson.build
index 1656355eef..c02c020d0b 100644
--- a/tools/meson.build
+++ b/tools/meson.build
@@ -113,4 +113,5 @@ if libudev.found()
endif
subdir('i915-perf')
+subdir('xe-perf')
subdir('null_state_gen')
diff --git a/tools/xe-perf/meson.build b/tools/xe-perf/meson.build
new file mode 100644
index 0000000000..a5f59a67fe
--- /dev/null
+++ b/tools/xe-perf/meson.build
@@ -0,0 +1,5 @@
+executable('xe-perf-recorder',
+ [ 'xe_perf_recorder.c' ],
+ include_directories: inc,
+ dependencies: [lib_igt, lib_igt_xe_oa],
+ install: true)
diff --git a/tools/xe-perf/xe_perf_recorder.c b/tools/xe-perf/xe_perf_recorder.c
new file mode 100644
index 0000000000..bd2d645fa1
--- /dev/null
+++ b/tools/xe-perf/xe_perf_recorder.c
@@ -0,0 +1,1227 @@
+// SPDX-License-Identifier: MIT
+/*
+ * Copyright © 2024 Intel Corporation
+ */
+
+#include <assert.h>
+#include <dirent.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <getopt.h>
+#include <inttypes.h>
+#include <limits.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 "igt_core.h"
+#include "intel_chipset.h"
+#include "ioctl_wrappers.h"
+#include "linux_scaffold.h"
+#include "xe/xe_oa.h"
+#include "xe/xe_oa_data.h"
+#include "xe/xe_query.h"
+
+#include "xe_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 intel_xe_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 void
+print_intel_devices(void)
+{
+ fprintf(stdout, "Available devices:\n");
+ for (int i = 0; i < 128; i++) {
+ if (read_device_param("card", i, "vendor") == 0x8086) {
+ uint32_t devid = read_device_param("card", i, "device");
+ const struct intel_device_info *devinfo =
+ intel_get_device_info(devid);
+ fprintf(stdout, " %i: %s (0x%04hx)\n", i,
+ devinfo ? devinfo->codename : "unknwon",
+ devid);
+ }
+ }
+}
+
+static int
+open_render_node(uint32_t *devid, int card)
+{
+ char *name;
+ int ret;
+ int fd;
+ int render;
+
+ if (card < 0) {
+ render = find_intel_render_node();
+ if (render < 0)
+ return -1;
+ } else {
+ render = 128 + card;
+ }
+
+ 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_cs_timestamp_frequency(int drm_fd)
+{
+ return xe_gt_list(drm_fd)->gt_list[0].reference_clock;
+}
+
+static uint64_t
+get_device_oa_timestamp_frequency(int drm_fd)
+{
+ struct drm_xe_query_oa_units *qoa = xe_oa_units(drm_fd);
+ struct drm_xe_oa_unit *oau = (struct drm_xe_oa_unit *)&qoa->oa_units[0];
+
+ return oau->oa_timestamp_freq;
+}
+
+struct recording_context {
+ int drm_fd;
+ int perf_fd;
+
+ uint32_t devid;
+ uint64_t oa_timestamp_frequency;
+ uint64_t cs_timestamp_frequency;
+
+ const struct intel_device_info *devinfo;
+
+ struct intel_xe_topology_info *topology;
+ uint32_t topology_size;
+
+ struct intel_xe_perf *perf;
+ struct intel_xe_perf_metric_set *metric_set;
+
+ uint32_t oa_exponent;
+
+ struct circular_buffer circular_buffer;
+ FILE *output_stream;
+
+ const char *command_fifo;
+ int command_fifo_fd;
+
+ int gt;
+ struct drm_xe_engine_class_instance eci;
+ struct drm_xe_engine_class_instance *hwe;
+ struct drm_xe_oa_unit *oa_unit;
+};
+
+static void set_fd_flags(int fd, int flags)
+{
+ int old = fcntl(fd, F_GETFL, 0);
+
+ igt_assert_lte(0, old);
+ igt_assert_eq(0, fcntl(fd, F_SETFL, old | flags));
+}
+
+enum xe_oa_report_header {
+ HDR_32_BIT = 0,
+ HDR_64_BIT,
+};
+
+struct xe_oa_format {
+ uint32_t counter_select;
+ int size;
+ int oa_type;
+ enum xe_oa_report_header header;
+ uint16_t counter_size;
+ uint16_t bc_report;
+};
+
+#define DRM_FMT(x) DRM_XE_OA_FMT_TYPE_##x
+
+static const struct xe_oa_format oa_formats[] = {
+ [XE_OA_FORMAT_C4_B8] = { 7, 64, DRM_FMT(OAG) },
+ [XE_OA_FORMAT_A12] = { 0, 64, DRM_FMT(OAG) },
+ [XE_OA_FORMAT_A12_B8_C8] = { 2, 128, DRM_FMT(OAG) },
+ [XE_OA_FORMAT_A32u40_A4u32_B8_C8] = { 5, 256, DRM_FMT(OAG) },
+ [XE_OAR_FORMAT_A32u40_A4u32_B8_C8] = { 5, 256, DRM_FMT(OAR) },
+ [XE_OA_FORMAT_A24u40_A14u32_B8_C8] = { 5, 256, DRM_FMT(OAG) },
+ [XE_OAC_FORMAT_A24u64_B8_C8] = { 1, 320, DRM_FMT(OAC), HDR_64_BIT },
+ [XE_OAC_FORMAT_A22u32_R2u32_B8_C8] = { 2, 192, DRM_FMT(OAC), HDR_64_BIT },
+ [XE_OAM_FORMAT_MPEC8u64_B8_C8] = { 1, 192, DRM_FMT(OAM_MPEC), HDR_64_BIT },
+ [XE_OAM_FORMAT_MPEC8u32_B8_C8] = { 2, 128, DRM_FMT(OAM_MPEC), HDR_64_BIT },
+ [XE_OA_FORMAT_PEC64u64] = { 1, 576, DRM_FMT(PEC), HDR_64_BIT, 1, 0 },
+ [XE_OA_FORMAT_PEC64u64_B8_C8] = { 1, 640, DRM_FMT(PEC), HDR_64_BIT, 1, 1 },
+ [XE_OA_FORMAT_PEC64u32] = { 1, 320, DRM_FMT(PEC), HDR_64_BIT },
+ [XE_OA_FORMAT_PEC32u64_G1] = { 5, 320, DRM_FMT(PEC), HDR_64_BIT, 1, 0 },
+ [XE_OA_FORMAT_PEC32u32_G1] = { 5, 192, DRM_FMT(PEC), HDR_64_BIT },
+ [XE_OA_FORMAT_PEC32u64_G2] = { 6, 320, DRM_FMT(PEC), HDR_64_BIT, 1, 0 },
+ [XE_OA_FORMAT_PEC32u32_G2] = { 6, 192, DRM_FMT(PEC), HDR_64_BIT },
+ [XE_OA_FORMAT_PEC36u64_G1_32_G2_4] = { 3, 320, DRM_FMT(PEC), HDR_64_BIT, 1, 0 },
+ [XE_OA_FORMAT_PEC36u64_G1_4_G2_32] = { 4, 320, DRM_FMT(PEC), HDR_64_BIT, 1, 0 },
+};
+
+static uint64_t oa_format_fields(uint64_t name)
+{
+#define FIELD_PREP_ULL(_mask, _val) \
+ (((_val) << (__builtin_ffsll(_mask) - 1)) & (_mask))
+
+ struct xe_oa_format f = oa_formats[name];
+
+ /* 0 format name is invalid */
+ if (!name)
+ memset(&f, 0xff, sizeof(f));
+
+ return FIELD_PREP_ULL(DRM_XE_OA_FORMAT_MASK_FMT_TYPE, (u64)f.oa_type) |
+ FIELD_PREP_ULL(DRM_XE_OA_FORMAT_MASK_COUNTER_SEL, (u64)f.counter_select) |
+ FIELD_PREP_ULL(DRM_XE_OA_FORMAT_MASK_COUNTER_SIZE, (u64)f.counter_size) |
+ FIELD_PREP_ULL(DRM_XE_OA_FORMAT_MASK_BC_REPORT, (u64)f.bc_report);
+#undef FIELD_PREP_ULL
+}
+#define __ff oa_format_fields
+
+static int
+perf_open(struct recording_context *ctx)
+{
+ int stream_fd;
+
+ uint64_t properties[] = {
+ DRM_XE_OA_PROPERTY_OA_UNIT_ID, ctx->oa_unit->oa_unit_id,
+
+ /* Include OA reports in samples */
+ DRM_XE_OA_PROPERTY_SAMPLE_OA, true,
+
+ /* OA unit configuration */
+ DRM_XE_OA_PROPERTY_OA_METRIC_SET, ctx->metric_set->perf_oa_metrics_set,
+ DRM_XE_OA_PROPERTY_OA_FORMAT, __ff(ctx->metric_set->perf_oa_format),
+ DRM_XE_OA_PROPERTY_OA_PERIOD_EXPONENT, ctx->oa_exponent,
+ };
+ struct intel_xe_oa_open_prop param = {
+ .num_properties = ARRAY_SIZE(properties) / 2,
+ .properties_ptr = to_user_pointer(properties),
+ };
+
+ stream_fd = intel_xe_perf_ioctl(ctx->drm_fd, DRM_XE_PERF_OP_STREAM_OPEN, ¶m);
+ if (stream_fd < 0) {
+ errno = 0;
+ goto exit;
+ }
+
+ set_fd_flags(stream_fd, O_CLOEXEC | O_NONBLOCK);
+exit:
+ return stream_fd;
+}
+
+static bool quit = false;
+
+static void
+sigint_handler(int val)
+{
+ quit = true;
+}
+
+static bool
+write_version(FILE *output, struct recording_context *ctx)
+{
+ struct intel_xe_perf_record_version version = {
+ .version = INTEL_XE_PERF_RECORD_VERSION,
+ };
+ struct intel_xe_perf_record_header header = {
+ .type = INTEL_XE_PERF_RECORD_TYPE_VERSION,
+ .size = sizeof(header) + sizeof(version),
+ };
+
+ if (fwrite(&header, sizeof(header), 1, output) != 1)
+ return false;
+
+ if (fwrite(&version, sizeof(version), 1, output) != 1)
+ return false;
+
+ return true;
+}
+
+static bool
+write_header(FILE *output, struct recording_context *ctx)
+{
+ struct intel_xe_perf_record_device_info info = {
+ .timestamp_frequency = ctx->oa_timestamp_frequency,
+ .device_id = ctx->perf->devinfo.devid,
+ .device_revision = ctx->perf->devinfo.revision,
+ .gt_min_frequency = ctx->perf->devinfo.gt_min_freq,
+ .gt_max_frequency = ctx->perf->devinfo.gt_max_freq,
+ .oa_format = ctx->metric_set->perf_oa_format,
+ .engine_class = ctx->hwe->engine_class,
+ .engine_instance = ctx->hwe->engine_instance,
+ };
+ struct intel_xe_perf_record_header header = {
+ .type = INTEL_XE_PERF_RECORD_TYPE_DEVICE_INFO,
+ .size = sizeof(header) + sizeof(info),
+ };
+
+ snprintf(info.metric_set_name, sizeof(info.metric_set_name),
+ "%s", ctx->metric_set->symbol_name);
+ snprintf(info.metric_set_uuid, sizeof(info.metric_set_uuid),
+ "%s", ctx->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 struct intel_xe_topology_info *get_topology(struct recording_context *ctx)
+{
+ return xe_fill_topology_info(ctx->drm_fd, ctx->devid, &ctx->topology_size);
+}
+
+static bool
+write_topology(FILE *output, struct recording_context *ctx)
+{
+ struct intel_xe_perf_record_header header = {
+ .type = INTEL_XE_PERF_RECORD_TYPE_DEVICE_TOPOLOGY,
+ };
+
+ header.size = sizeof(header) + ctx->topology_size;
+ if (fwrite(&header, sizeof(header), 1, output) != 1)
+ return false;
+
+ if (fwrite(ctx->topology, ctx->topology_size, 1, output) != 1)
+ return false;
+
+ return true;
+}
+
+static int get_stream_status(int perf_fd, u32 *oa_status)
+{
+ struct drm_xe_oa_stream_status status;
+ int ret;
+
+ ret = perf_ioctl(perf_fd, DRM_XE_PERF_IOCTL_STATUS, &status);
+ if (ret)
+ return ret;
+
+ *oa_status = status.oa_status;
+ return 0;
+}
+
+static bool write_stream_status(struct recording_context *ctx, FILE *output)
+{
+ u32 oa_status;
+
+ if (!get_stream_status(ctx->perf_fd, &oa_status)) {
+ struct intel_xe_perf_record_header header = { .size = sizeof(header) };
+
+ if (oa_status & DRM_XE_OASTATUS_REPORT_LOST)
+ header.type = INTEL_XE_PERF_RECORD_OA_TYPE_REPORT_LOST;
+ else if (oa_status & DRM_XE_OASTATUS_BUFFER_OVERFLOW)
+ header.type = INTEL_XE_PERF_RECORD_OA_TYPE_BUFFER_LOST;
+ else
+ return true;
+
+ if (fwrite(&header, sizeof(header), 1, output) != 1)
+ return false;
+ }
+
+ return true;
+}
+
+static bool write_stream_data(struct recording_context *ctx,
+ char *data, ssize_t size, FILE *output)
+{
+ ssize_t format_size = oa_formats[ctx->metric_set->perf_oa_format].size;
+
+ assert(!(size % format_size));
+
+ for (int i = 0; i < size / format_size; i++) {
+ struct intel_xe_perf_record_header header = {
+ .type = INTEL_XE_PERF_RECORD_TYPE_SAMPLE,
+ .size = sizeof(header) + format_size,
+ };
+
+ if (fwrite(&header, sizeof(header), 1, output) != 1)
+ return false;
+
+ if (fwrite(data, format_size, 1, output) != 1)
+ return false;
+ }
+
+ return true;
+}
+
+static bool write_perf_data(FILE *output, struct recording_context *ctx)
+{
+ char data[4096];
+ ssize_t len;
+ bool ret;
+
+ while (1) {
+ len = read(ctx->perf_fd, data, sizeof(data));
+
+ if (len < 0) {
+ switch (errno) {
+ case EIO:
+ ret = write_stream_status(ctx, output);
+ if (!ret)
+ return ret;
+ break;
+ case EAGAIN:
+ case EINTR:
+ return true;
+ default:
+ /* Not expecting -EFAULT, -ENOSPC, -EINVAL */
+ assert(0);
+ }
+ } else {
+ ret = write_stream_data(ctx, data, len, output);
+ if (!ret)
+ return ret;
+ }
+ }
+
+ /* Should not reach here */
+ return false;
+}
+
+static clock_t correlation_clock_id = CLOCK_MONOTONIC;
+
+static const char *
+get_correlation_clock_name(clock_t clock_id)
+{
+ switch (clock_id) {
+ case CLOCK_BOOTTIME: return "bootime";
+ case CLOCK_MONOTONIC: return "monotonic";
+ case CLOCK_MONOTONIC_RAW: return "monotonic_raw";
+ default: return "*unknown*";
+ }
+}
+
+static int query_engine_cycles(int fd, struct drm_xe_query_engine_cycles *ts)
+{
+ struct drm_xe_device_query query = {
+ .extensions = 0,
+ .query = DRM_XE_DEVICE_QUERY_ENGINE_CYCLES,
+ .size = sizeof(*ts),
+ .data = (uintptr_t)ts,
+ };
+
+ return perf_ioctl(fd, DRM_IOCTL_XE_DEVICE_QUERY, &query);
+}
+
+static bool get_correlation_timestamps(struct recording_context *ctx,
+ struct intel_xe_perf_record_timestamp_correlation *corr)
+{
+ struct drm_xe_query_engine_cycles ts = {};
+
+ ts.eci = *ctx->hwe;
+ ts.clockid = correlation_clock_id;
+
+ if (query_engine_cycles(ctx->drm_fd, &ts))
+ return false;
+
+ corr->cpu_timestamp = ts.cpu_timestamp + ts.cpu_delta / 2;
+ corr->gpu_timestamp = ts.engine_cycles;
+
+ return true;
+}
+
+static bool
+write_saved_correlation_timestamps(FILE *output,
+ const struct intel_xe_perf_record_timestamp_correlation *corr)
+{
+ struct intel_xe_perf_record_header header = {
+ .type = INTEL_XE_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(struct recording_context *ctx, FILE *output)
+{
+ struct intel_xe_perf_record_timestamp_correlation corr;
+
+ if (!get_correlation_timestamps(ctx, &corr))
+ return false;
+
+ return write_saved_correlation_timestamps(output, &corr);
+}
+
+static void
+read_command_file(struct recording_context *ctx)
+{
+ struct recorder_command_base header;
+ ssize_t ret = read(ctx->command_fifo_fd, &header, sizeof(header));
+
+ if (ret < 0)
+ return;
+
+ switch (header.command) {
+ case RECORDER_COMMAND_DUMP: {
+ uint32_t len = header.size - sizeof(header), offset = 0;
+ uint8_t *dump = malloc(len);
+ FILE *file;
+
+ while (offset < len &&
+ ((ret = read(ctx->command_fifo_fd,
+ (void *) dump + offset, len - offset)) > 0
+ || errno == EAGAIN)) {
+ if (ret > 0)
+ offset += ret;
+ }
+
+ fprintf(stdout, "Writing circular buffer to %s\n", dump);
+
+ file = fopen((const char *) dump, "w+");
+ if (file) {
+ struct chunk chunks[2];
+
+ fflush(ctx->output_stream);
+ get_chunks(chunks, &ctx->circular_buffer,
+ false, ctx->circular_buffer.size);
+
+ if (!write_version(file, ctx) ||
+ !write_header(file, ctx) ||
+ !write_topology(file, ctx) ||
+ 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(ctx, file)) {
+ fprintf(stderr, "Unable to write circular buffer data in file '%s'\n",
+ dump);
+ }
+ fclose(file);
+ } else
+ fprintf(stderr, "Unable to write dump file '%s'\n", dump);
+
+ 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_xe_perf *perf)
+{
+ struct intel_xe_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\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_xe_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, "%s (%s):\n", metric_set->symbol_name, metric_set->name);
+ for (uint32_t i = 0; i < metric_set->n_counters; i++) {
+ struct intel_xe_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
+print_metric_sets_counters(struct intel_xe_perf *perf)
+{
+ struct intel_xe_perf_metric_set *metric_set;
+
+ igt_list_for_each_entry(metric_set, &perf->metric_sets, link)
+ print_metric_set_counters(metric_set);
+}
+
+static void
+usage(const char *name)
+{
+ fprintf(stdout,
+ "Usage: %s [options]\n"
+ "Recording tool for xe-oa\n"
+ "\n"
+ " --help, -h Print this screen\n"
+ " --device, -d <value> Device to use\n"
+ " (value=list to list devices\n"
+ " value=1 to use /dev/dri/card1)\n"
+ " --correlation-period, -c <value> Time period of timestamp correlation in seconds\n"
+ " (default = 1.0)\n"
+ " --perf-period, -p <value> Time period of xe-oa reports in seconds\n"
+ " (default = 0.001)\n"
+ " --metric, -m <value> xe-oa metric to sample with (use value=list to list all metrics)\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 xe-perf-control)\n"
+ " --output, -o <path> Output file (default = xe_perf.record)\n"
+ " --cpu-clock, -k <path> Cpu clock to use for correlations\n"
+ " Values: boot, mono, mono_raw (default = mono)\n"
+ " --engine-class -e <value> Engine class used for the OA capture.\n"
+ " --engine-instance -i <value> Engine instance used for the OA capture.\n",
+ name);
+}
+
+static void
+teardown_recording_context(struct recording_context *ctx)
+{
+ if (ctx->topology)
+ free(ctx->topology);
+
+ if (ctx->perf)
+ intel_xe_perf_free(ctx->perf);
+
+ if (ctx->command_fifo)
+ unlink(ctx->command_fifo);
+ if (ctx->command_fifo_fd != -1)
+ close(ctx->command_fifo_fd);
+
+ if (ctx->output_stream)
+ fclose(ctx->output_stream);
+
+ free(ctx->circular_buffer.data);
+
+ if (ctx->perf_fd != -1)
+ close(ctx->perf_fd);
+ if (ctx->drm_fd != -1)
+ close(ctx->drm_fd);
+}
+
+static int assign_oa_unit(int fd, struct recording_context *ctx)
+{
+ struct drm_xe_query_oa_units *qoa = xe_oa_units(fd);
+ struct drm_xe_oa_unit *oau;
+ uint8_t *poau;
+
+ poau = (uint8_t *)&qoa->oa_units[0];
+ for (int i = 0; i < qoa->num_oa_units; i++) {
+ oau = (struct drm_xe_oa_unit *)poau;
+
+ for (int j = 0; j < oau->num_engines; j++) {
+ if (oau->eci[j].engine_class == ctx->eci.engine_class &&
+ oau->eci[j].engine_instance == ctx->eci.engine_instance) {
+ ctx->hwe = &oau->eci[j];
+ ctx->oa_unit = oau;
+ return 0;
+ }
+ }
+
+ poau += sizeof(*oau) + oau->num_engines * sizeof(oau->eci[0]);
+ }
+
+ return -1;
+}
+
+int
+main(int argc, char *argv[])
+{
+ const struct option long_options[] = {
+ {"help", no_argument, 0, 'h'},
+ {"device", required_argument, 0, 'd'},
+ {"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'},
+ {"engine-class", required_argument, 0, 'e'},
+ {"engine-instance", required_argument, 0, 'i'},
+ {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" },
+ };
+ double corr_period = 1.0, perf_period = 0.001;
+ const char *metric_name = NULL, *output_file = "xe_perf.record";
+ struct intel_xe_perf_metric_set *metric_set;
+ struct intel_xe_perf_record_timestamp_correlation initial_correlation;
+ struct timespec now;
+ uint64_t corr_period_ns, poll_time_ns;
+ uint32_t circular_size = 0;
+ int opt, dev_node_id = -1;
+ bool list_counters = false;
+ FILE *output = NULL;
+ struct recording_context ctx = {
+ .drm_fd = -1,
+ .perf_fd = -1,
+
+ .command_fifo = XE_PERF_RECORD_FIFO_PATH,
+ .command_fifo_fd = -1,
+
+ .eci = { DRM_XE_ENGINE_CLASS_RENDER, 0 },
+ };
+
+ while ((opt = getopt_long(argc, argv, "hc:d:p:m:Co:s:f:k:P:e:i:", long_options, NULL)) != -1) {
+ switch (opt) {
+ case 'h':
+ usage(argv[0]);
+ return EXIT_SUCCESS;
+ case 'c':
+ corr_period = atof(optarg);
+ break;
+ case 'd':
+ if (!strcmp(optarg, "list"))
+ dev_node_id = -2;
+ else
+ dev_node_id = atoi(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':
+ ctx.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;
+ }
+ case 'e':
+ ctx.eci.engine_class = atoi(optarg);
+ break;
+ case 'i':
+ ctx.eci.engine_instance = atoi(optarg);
+ break;
+ default:
+ fprintf(stderr, "Internal error: "
+ "unexpected getopt value: %d\n", opt);
+ usage(argv[0]);
+ return EXIT_FAILURE;
+ }
+ }
+
+ if (dev_node_id == -2) {
+ print_intel_devices();
+ return EXIT_SUCCESS;
+ }
+
+ ctx.drm_fd = open_render_node(&ctx.devid, dev_node_id);
+ if (ctx.drm_fd < 0) {
+ fprintf(stderr, "Unable to open device.\n");
+ return EXIT_FAILURE;
+ }
+
+ xe_device_get(ctx.drm_fd);
+
+ ctx.devinfo = intel_get_device_info(ctx.devid);
+ if (!ctx.devinfo) {
+ fprintf(stderr, "No device info found.\n");
+ goto fail;
+ }
+
+ if (assign_oa_unit(ctx.drm_fd, &ctx) < 0) {
+ fprintf(stderr, "assign_oa_unit failed\n");
+ goto fail;
+ }
+
+ fprintf(stdout, "Device name=%s gen=%i id=0x%x oa_unit=%i gt=%i\n",
+ ctx.devinfo->codename, ctx.devinfo->graphics_ver, ctx.devid,
+ ctx.oa_unit->oa_unit_id, ctx.hwe->gt_id);
+
+ ctx.topology = get_topology(&ctx);
+ if (!ctx.topology) {
+ fprintf(stderr, "Unable to retrieve GPU topology\n");
+ goto fail;
+ }
+
+ ctx.perf = intel_xe_perf_for_fd(ctx.drm_fd, ctx.hwe->gt_id);
+ if (!ctx.perf) {
+ fprintf(stderr, "No perf data found.\n");
+ goto fail;
+ }
+
+ intel_xe_perf_load_perf_configs(ctx.perf, ctx.drm_fd);
+
+ if (metric_name) {
+ if (!strcmp(metric_name, "list")) {
+ print_metric_sets(ctx.perf);
+ return EXIT_SUCCESS;
+ }
+
+ igt_list_for_each_entry(metric_set, &ctx.perf->metric_sets, link) {
+ if (!strcasecmp(metric_set->symbol_name, metric_name)) {
+ ctx.metric_set = metric_set;
+ break;
+ }
+ }
+ }
+
+ if (list_counters) {
+ if (!ctx.metric_set)
+ print_metric_sets_counters(ctx.perf);
+ else
+ print_metric_set_counters(ctx.metric_set);
+ teardown_recording_context(&ctx);
+ return EXIT_SUCCESS;
+ }
+
+ if (!ctx.metric_set) {
+ if (!metric_name)
+ fprintf(stderr, "No metric set specified.\n");
+ else
+ fprintf(stderr, "Unknown metric set '%s'.\n", metric_name);
+ print_metric_sets(ctx.perf);
+ goto fail;
+ }
+
+ intel_xe_perf_load_perf_configs(ctx.perf, ctx.drm_fd);
+
+ ctx.oa_timestamp_frequency = get_device_oa_timestamp_frequency(ctx.drm_fd);
+ ctx.cs_timestamp_frequency = get_device_cs_timestamp_frequency(ctx.drm_fd);
+
+ signal(SIGINT, sigint_handler);
+
+ if (ctx.command_fifo) {
+ if (mkfifo(ctx.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",
+ ctx.command_fifo, strerror(errno));
+ goto fail;
+ }
+
+ ctx.command_fifo_fd = open(ctx.command_fifo, O_RDWR);
+ if (ctx.command_fifo_fd < 0) {
+ fprintf(stderr, "Unable to open command fifo '%s': %s\n",
+ ctx.command_fifo, strerror(errno));
+ goto fail;
+ }
+ }
+
+ if (circular_size) {
+ ctx.circular_buffer.allocated_size = circular_size;
+ ctx.circular_buffer.data = malloc(circular_size);
+ if (!ctx.circular_buffer.data) {
+ fprintf(stderr, "Unable to allocate circular buffer\n");
+ goto fail;
+ }
+
+ ctx.output_stream = fopencookie(&ctx.circular_buffer, "w+",
+ circular_buffer_functions);
+ if (!ctx.output_stream) {
+ fprintf(stderr, "Unable to create circular buffer\n");
+ goto fail;
+ }
+
+ if (!get_correlation_timestamps(&ctx, &initial_correlation)) {
+ fprintf(stderr, "Unable to correlation timestamps\n");
+ goto fail;
+ }
+
+ write_correlation_timestamps(&ctx, ctx.output_stream);
+ fprintf(stdout,
+ "Recoding in internal circular buffer.\n"
+ "Use xe-perf-control to snapshot into file.\n");
+ } else {
+ output = fopen(output_file, "w+");
+ if (!output) {
+ fprintf(stderr, "Unable to open output file '%s'\n",
+ output_file);
+ goto fail;
+ }
+
+ if (!write_version(output, &ctx) ||
+ !write_header(output, &ctx) ||
+ !write_topology(output, &ctx) ||
+ !write_correlation_timestamps(&ctx, output)) {
+ fprintf(stderr, "Unable to write header in file '%s'\n",
+ output_file);
+ goto fail;
+ }
+
+ ctx.output_stream = output;
+ fprintf(stdout, "Writing recoding to %s\n", output_file);
+ }
+
+ if (ctx.metric_set->perf_oa_metrics_set == 0) {
+ fprintf(stderr,
+ "Unable to load performance configuration, consider running:\n"
+ " sysctl dev.xe.perf_stream_paranoid=0\n");
+ goto fail;
+ }
+
+ fprintf(stdout, "Using correlation clock: %s\n",
+ get_correlation_clock_name(correlation_clock_id));
+
+ ctx.oa_exponent = oa_exponent_for_period(ctx.oa_timestamp_frequency, perf_period);
+ fprintf(stdout, "Opening perf stream with metric_id=%"PRIu64" oa_exponent=%u oa_format=%u\n",
+ ctx.metric_set->perf_oa_metrics_set, ctx.oa_exponent,
+ ctx.metric_set->perf_oa_format);
+
+ ctx.perf_fd = perf_open(&ctx);
+ if (ctx.perf_fd < 0) {
+ fprintf(stderr, "Unable to open xe oa stream: %s\n",
+ strerror(errno));
+ goto fail;
+ }
+
+ corr_period_ns = corr_period * 1000000000ul;
+ poll_time_ns = corr_period_ns;
+
+ while (!quit) {
+ struct pollfd pollfd[2] = {
+ { ctx.perf_fd, POLLIN, 0 },
+ { ctx.command_fifo_fd, POLLIN, 0 },
+ };
+ uint64_t elapsed_ns;
+ int ret;
+
+ igt_gettime(&now);
+ ret = poll(pollfd, ctx.command_fifo_fd != -1 ? 2 : 1, poll_time_ns / 1000000);
+ if (ret < 0 && errno != EINTR) {
+ fprintf(stderr, "Failed to poll xe-oa stream: %s\n",
+ strerror(errno));
+ break;
+ }
+
+ if (ret > 0) {
+ if (pollfd[0].revents & POLLIN) {
+ if (!write_perf_data(ctx.output_stream, &ctx)) {
+ fprintf(stderr, "Failed to write xe-oa data: %s\n",
+ strerror(errno));
+ break;
+ }
+ }
+
+ if (pollfd[1].revents & POLLIN) {
+ read_command_file(&ctx);
+ }
+ }
+
+ elapsed_ns = igt_nsec_elapsed(&now);
+ if (elapsed_ns > poll_time_ns) {
+ poll_time_ns = corr_period_ns;
+ if (!write_correlation_timestamps(&ctx, ctx.output_stream)) {
+ fprintf(stderr,
+ "Failed to write xe timestamp correlation data: %s\n",
+ strerror(errno));
+ break;
+ }
+ } else {
+ poll_time_ns -= elapsed_ns;
+ }
+ }
+
+ fprintf(stdout, "Exiting...\n");
+
+ if (!write_perf_data(ctx.output_stream, &ctx)) {
+ fprintf(stderr, "Failed to write xe-oa data: %s\n",
+ strerror(errno));
+ }
+
+ if (!write_correlation_timestamps(&ctx, ctx.output_stream)) {
+ fprintf(stderr,
+ "Failed to write final xe timestamp correlation data: %s\n",
+ strerror(errno));
+ }
+
+ teardown_recording_context(&ctx);
+
+ return EXIT_SUCCESS;
+
+ fail:
+ teardown_recording_context(&ctx);
+
+ return EXIT_FAILURE;
+}
diff --git a/tools/xe-perf/xe_perf_recorder_commands.h b/tools/xe-perf/xe_perf_recorder_commands.h
new file mode 100644
index 0000000000..528813d2dc
--- /dev/null
+++ b/tools/xe-perf/xe_perf_recorder_commands.h
@@ -0,0 +1,31 @@
+/* SPDX-License-Identifier: MIT */
+/*
+ * Copyright © 2023 Intel Corporation
+ */
+
+#ifndef XE_PERF_RECORDER_COMMANDS_H
+#define XE_PERF_RECORDER_COMMANDS_H
+
+#include <stdint.h>
+
+#define XE_PERF_RECORD_FIFO_PATH "/tmp/.xe-perf-record"
+
+enum recorder_command {
+ RECORDER_COMMAND_DUMP = 1,
+ RECORDER_COMMAND_QUIT,
+};
+
+struct recorder_command_base {
+ uint32_t command;
+ uint32_t size; /* size of recorder_command_base + dump in bytes */
+};
+
+/*
+ The dump after the recorder_command_base header:
+
+struct recorder_command_dump {
+ uint8_t path[0];
+};
+*/
+
+#endif
--
2.41.0
More information about the igt-dev
mailing list