src/intel/perf/gen_perf.c

   1 /*
   2  * Copyright © 2018 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include <dirent.h>
  25
  26 #include <sys/types.h>
  27 #include <sys/stat.h>
  28 #include <fcntl.h>
  29 #include <unistd.h>
  30 #include <errno.h>
  31
  32 #include <drm-uapi/i915_drm.h>
  33
  34 #include "gen_perf.h"
  35 #include "perf/gen_perf_metrics.h"
  36
  37 #include "dev/gen_debug.h"
  38 #include "dev/gen_device_info.h"
  39 #include "util/bitscan.h"
  40
  41 #define FILE_DEBUG_FLAG DEBUG_PERFMON
  42
  43 static bool
  44 get_sysfs_dev_dir(struct gen_perf_config *perf, int fd)
  45 {
  46    struct stat sb;
  47    int min, maj;
  48    DIR *drmdir;
  49    struct dirent *drm_entry;
  50    int len;
  51
  52    perf->sysfs_dev_dir[0] = '\0';
  53
  54    if (fstat(fd, &sb)) {
  55       DBG("Failed to stat DRM fd\n");
  56       return false;
  57    }
  58
  59    maj = major(sb.st_rdev);
  60    min = minor(sb.st_rdev);
  61
  62    if (!S_ISCHR(sb.st_mode)) {
  63       DBG("DRM fd is not a character device as expected\n");
  64       return false;
  65    }
  66
  67    len = snprintf(perf->sysfs_dev_dir,
  68                   sizeof(perf->sysfs_dev_dir),
  69                   "/sys/dev/char/%d:%d/device/drm", maj, min);
  70    if (len < 0 || len >= sizeof(perf->sysfs_dev_dir)) {
  71       DBG("Failed to concatenate sysfs path to drm device\n");
  72       return false;
  73    }
  74
  75    drmdir = opendir(perf->sysfs_dev_dir);
  76    if (!drmdir) {
  77       DBG("Failed to open %s: %m\n", perf->sysfs_dev_dir);
  78       return false;
  79    }
  80
  81    while ((drm_entry = readdir(drmdir))) {
  82       if ((drm_entry->d_type == DT_DIR ||
  83            drm_entry->d_type == DT_LNK) &&
  84           strncmp(drm_entry->d_name, "card", 4) == 0)
  85       {
  86          len = snprintf(perf->sysfs_dev_dir,
  87                         sizeof(perf->sysfs_dev_dir),
  88                         "/sys/dev/char/%d:%d/device/drm/%s",
  89                         maj, min, drm_entry->d_name);
  90          closedir(drmdir);
  91          if (len < 0 || len >= sizeof(perf->sysfs_dev_dir))
  92             return false;
  93          else
  94             return true;
  95       }
  96    }
  97
  98    closedir(drmdir);
  99
 100    DBG("Failed to find cardX directory under /sys/dev/char/%d:%d/device/drm\n",
 101        maj, min);
 102
 103    return false;
 104 }
 105
 106 static bool
 107 read_file_uint64(const char *file, uint64_t *val)
 108 {
 109     char buf[32];
 110     int fd, n;
 111
 112     fd = open(file, 0);
 113     if (fd < 0)
 114        return false;
 115     while ((n = read(fd, buf, sizeof (buf) - 1)) < 0 &&
 116            errno == EINTR);
 117     close(fd);
 118     if (n < 0)
 119        return false;
 120
 121     buf[n] = '\0';
 122     *val = strtoull(buf, NULL, 0);
 123
 124     return true;
 125 }
 126
 127 static bool
 128 read_sysfs_drm_device_file_uint64(struct gen_perf_config *perf,
 129                                   const char *file,
 130                                   uint64_t *value)
 131 {
 132    char buf[512];
 133    int len;
 134
 135    len = snprintf(buf, sizeof(buf), "%s/%s", perf->sysfs_dev_dir, file);
 136    if (len < 0 || len >= sizeof(buf)) {
 137       DBG("Failed to concatenate sys filename to read u64 from\n");
 138       return false;
 139    }
 140
 141    return read_file_uint64(buf, value);
 142 }
 143
 144 static void
 145 register_oa_config(struct gen_perf_config *perf,
 146                    const struct gen_perf_query_info *query,
 147                    uint64_t config_id)
 148 {
 149    struct gen_perf_query_info *registred_query =
 150       gen_perf_query_append_query_info(perf, 0);
 151
 152    *registred_query = *query;
 153    registred_query->oa_metrics_set_id = config_id;
 154    DBG("metric set registred: id = %" PRIu64", guid = %s\n",
 155        registred_query->oa_metrics_set_id, query->guid);
 156 }
 157
 158 static void
 159 enumerate_sysfs_metrics(struct gen_perf_config *perf)
 160 {
 161    DIR *metricsdir = NULL;
 162    struct dirent *metric_entry;
 163    char buf[256];
 164    int len;
 165
 166    len = snprintf(buf, sizeof(buf), "%s/metrics", perf->sysfs_dev_dir);
 167    if (len < 0 || len >= sizeof(buf)) {
 168       DBG("Failed to concatenate path to sysfs metrics/ directory\n");
 169       return;
 170    }
 171
 172    metricsdir = opendir(buf);
 173    if (!metricsdir) {
 174       DBG("Failed to open %s: %m\n", buf);
 175       return;
 176    }
 177
 178    while ((metric_entry = readdir(metricsdir))) {
 179       struct hash_entry *entry;
 180
 181       if ((metric_entry->d_type != DT_DIR &&
 182            metric_entry->d_type != DT_LNK) ||
 183           metric_entry->d_name[0] == '.')
 184          continue;
 185
 186       DBG("metric set: %s\n", metric_entry->d_name);
 187       entry = _mesa_hash_table_search(perf->oa_metrics_table,
 188                                       metric_entry->d_name);
 189       if (entry) {
 190          uint64_t id;
 191
 192          len = snprintf(buf, sizeof(buf), "%s/metrics/%s/id",
 193                         perf->sysfs_dev_dir, metric_entry->d_name);
 194          if (len < 0 || len >= sizeof(buf)) {
 195             DBG("Failed to concatenate path to sysfs metric id file\n");
 196             continue;
 197          }
 198
 199          if (!read_file_uint64(buf, &id)) {
 200             DBG("Failed to read metric set id from %s: %m", buf);
 201             continue;
 202          }
 203
 204          register_oa_config(perf, (const struct gen_perf_query_info *)entry->data, id);
 205       } else
 206          DBG("metric set not known by mesa (skipping)\n");
 207    }
 208
 209    closedir(metricsdir);
 210 }
 211
 212 static bool
 213 kernel_has_dynamic_config_support(struct gen_perf_config *perf, int fd)
 214 {
 215    uint64_t invalid_config_id = UINT64_MAX;
 216
 217    return perf->ioctl(fd, DRM_IOCTL_I915_PERF_REMOVE_CONFIG,
 218                       &invalid_config_id) < 0 && errno == ENOENT;
 219 }
 220
 221 bool
 222 gen_perf_load_metric_id(struct gen_perf_config *perf, const char *guid,
 223                         uint64_t *metric_id)
 224 {
 225    char config_path[280];
 226
 227    snprintf(config_path, sizeof(config_path), "%s/metrics/%s/id",
 228             perf->sysfs_dev_dir, guid);
 229
 230    /* Don't recreate already loaded configs. */
 231    return read_file_uint64(config_path, metric_id);
 232 }
 233
 234 static void
 235 init_oa_configs(struct gen_perf_config *perf, int fd)
 236 {
 237    hash_table_foreach(perf->oa_metrics_table, entry) {
 238       const struct gen_perf_query_info *query = entry->data;
 239       struct drm_i915_perf_oa_config config;
 240       uint64_t config_id;
 241       int ret;
 242
 243       if (gen_perf_load_metric_id(perf, query->guid, &config_id)) {
 244          DBG("metric set: %s (already loaded)\n", query->guid);
 245          register_oa_config(perf, query, config_id);
 246          continue;
 247       }
 248
 249       memset(&config, 0, sizeof(config));
 250
 251       memcpy(config.uuid, query->guid, sizeof(config.uuid));
 252
 253       config.n_mux_regs = query->n_mux_regs;
 254       config.mux_regs_ptr = (uintptr_t) query->mux_regs;
 255
 256       config.n_boolean_regs = query->n_b_counter_regs;
 257       config.boolean_regs_ptr = (uintptr_t) query->b_counter_regs;
 258
 259       config.n_flex_regs = query->n_flex_regs;
 260       config.flex_regs_ptr = (uintptr_t) query->flex_regs;
 261
 262       ret = perf->ioctl(fd, DRM_IOCTL_I915_PERF_ADD_CONFIG, &config);
 263       if (ret < 0) {
 264          DBG("Failed to load \"%s\" (%s) metrics set in kernel: %s\n",
 265              query->name, query->guid, strerror(errno));
 266          continue;
 267       }
 268
 269       register_oa_config(perf, query, ret);
 270       DBG("metric set: %s (added)\n", query->guid);
 271    }
 272 }
 273
 274 static void
 275 compute_topology_builtins(struct gen_perf_config *perf,
 276                           const struct gen_device_info *devinfo)
 277 {
 278    perf->sys_vars.slice_mask = devinfo->slice_masks;
 279    perf->sys_vars.n_eu_slices = devinfo->num_slices;
 280
 281    for (int i = 0; i < sizeof(devinfo->subslice_masks[i]); i++) {
 282       perf->sys_vars.n_eu_sub_slices +=
 283          __builtin_popcount(devinfo->subslice_masks[i]);
 284    }
 285
 286    for (int i = 0; i < sizeof(devinfo->eu_masks); i++)
 287       perf->sys_vars.n_eus += __builtin_popcount(devinfo->eu_masks[i]);
 288
 289    perf->sys_vars.eu_threads_count = devinfo->num_thread_per_eu;
 290
 291    /* The subslice mask builtin contains bits for all slices. Prior to Gen11
 292     * it had groups of 3bits for each slice, on Gen11 it's 8bits for each
 293     * slice.
 294     *
 295     * Ideally equations would be updated to have a slice/subslice query
 296     * function/operator.
 297     */
 298    perf->sys_vars.subslice_mask = 0;
 299
 300    int bits_per_subslice = devinfo->gen == 11 ? 8 : 3;
 301
 302    for (int s = 0; s < util_last_bit(devinfo->slice_masks); s++) {
 303       for (int ss = 0; ss < (devinfo->subslice_slice_stride * 8); ss++) {
 304          if (gen_device_info_subslice_available(devinfo, s, ss))
 305             perf->sys_vars.subslice_mask |= 1ULL << (s * bits_per_subslice + ss);
 306       }
 307    }
 308 }
 309
 310 static bool
 311 init_oa_sys_vars(struct gen_perf_config *perf, const struct gen_device_info *devinfo)
 312 {
 313    uint64_t min_freq_mhz = 0, max_freq_mhz = 0;
 314
 315    if (!read_sysfs_drm_device_file_uint64(perf, "gt_min_freq_mhz", &min_freq_mhz))
 316       return false;
 317
 318    if (!read_sysfs_drm_device_file_uint64(perf,  "gt_max_freq_mhz", &max_freq_mhz))
 319       return false;
 320
 321    memset(&perf->sys_vars, 0, sizeof(perf->sys_vars));
 322    perf->sys_vars.gt_min_freq = min_freq_mhz * 1000000;
 323    perf->sys_vars.gt_max_freq = max_freq_mhz * 1000000;
 324    perf->sys_vars.timestamp_frequency = devinfo->timestamp_frequency;
 325    perf->sys_vars.revision = devinfo->revision;
 326    compute_topology_builtins(perf, devinfo);
 327
 328    return true;
 329 }
 330
 331 typedef void (*perf_register_oa_queries_t)(struct gen_perf_config *);
 332
 333 static perf_register_oa_queries_t
 334 get_register_queries_function(const struct gen_device_info *devinfo)
 335 {
 336    if (devinfo->is_haswell)
 337       return gen_oa_register_queries_hsw;
 338    if (devinfo->is_cherryview)
 339       return gen_oa_register_queries_chv;
 340    if (devinfo->is_broadwell)
 341       return gen_oa_register_queries_bdw;
 342    if (devinfo->is_broxton)
 343       return gen_oa_register_queries_bxt;
 344    if (devinfo->is_skylake) {
 345       if (devinfo->gt == 2)
 346          return gen_oa_register_queries_sklgt2;
 347       if (devinfo->gt == 3)
 348          return gen_oa_register_queries_sklgt3;
 349       if (devinfo->gt == 4)
 350          return gen_oa_register_queries_sklgt4;
 351    }
 352    if (devinfo->is_kabylake) {
 353       if (devinfo->gt == 2)
 354          return gen_oa_register_queries_kblgt2;
 355       if (devinfo->gt == 3)
 356          return gen_oa_register_queries_kblgt3;
 357    }
 358    if (devinfo->is_geminilake)
 359       return gen_oa_register_queries_glk;
 360    if (devinfo->is_coffeelake) {
 361       if (devinfo->gt == 2)
 362          return gen_oa_register_queries_cflgt2;
 363       if (devinfo->gt == 3)
 364          return gen_oa_register_queries_cflgt3;
 365    }
 366    if (devinfo->is_cannonlake)
 367       return gen_oa_register_queries_cnl;
 368    if (devinfo->gen == 11)
 369       return gen_oa_register_queries_icl;
 370
 371    return NULL;
 372 }
 373
 374 bool
 375 gen_perf_load_oa_metrics(struct gen_perf_config *perf, int fd,
 376                          const struct gen_device_info *devinfo)
 377 {
 378    perf_register_oa_queries_t oa_register = get_register_queries_function(devinfo);
 379    bool i915_perf_oa_available = false;
 380    struct stat sb;
 381
 382    /* The existence of this sysctl parameter implies the kernel supports
 383     * the i915 perf interface.
 384     */
 385    if (stat("/proc/sys/dev/i915/perf_stream_paranoid", &sb) == 0) {
 386
 387       /* If _paranoid == 1 then on Gen8+ we won't be able to access OA
 388        * metrics unless running as root.
 389        */
 390       if (devinfo->is_haswell)
 391          i915_perf_oa_available = true;
 392       else {
 393          uint64_t paranoid = 1;
 394
 395          read_file_uint64("/proc/sys/dev/i915/perf_stream_paranoid", &paranoid);
 396
 397          if (paranoid == 0 || geteuid() == 0)
 398             i915_perf_oa_available = true;
 399       }
 400    }
 401
 402    if (!i915_perf_oa_available ||
 403        !oa_register ||
 404        !get_sysfs_dev_dir(perf, fd) ||
 405        !init_oa_sys_vars(perf, devinfo))
 406       return false;
 407
 408    perf->oa_metrics_table =
 409       _mesa_hash_table_create(perf, _mesa_key_hash_string,
 410                               _mesa_key_string_equal);
 411
 412    /* Index all the metric sets mesa knows about before looking to see what
 413     * the kernel is advertising.
 414     */
 415    oa_register(perf);
 416
 417    if (likely((INTEL_DEBUG & DEBUG_NO_OACONFIG) == 0) &&
 418        kernel_has_dynamic_config_support(perf, fd))
 419       init_oa_configs(perf, fd);
 420    else
 421       enumerate_sysfs_metrics(perf);
 422
 423    return true;
 424 }
 425
 426 /* Accumulate 32bits OA counters */
 427 static inline void
 428 accumulate_uint32(const uint32_t *report0,
 429                   const uint32_t *report1,
 430                   uint64_t *accumulator)
 431 {
 432    *accumulator += (uint32_t)(*report1 - *report0);
 433 }
 434
 435 /* Accumulate 40bits OA counters */
 436 static inline void
 437 accumulate_uint40(int a_index,
 438                   const uint32_t *report0,
 439                   const uint32_t *report1,
 440                   uint64_t *accumulator)
 441 {
 442    const uint8_t *high_bytes0 = (uint8_t *)(report0 + 40);
 443    const uint8_t *high_bytes1 = (uint8_t *)(report1 + 40);
 444    uint64_t high0 = (uint64_t)(high_bytes0[a_index]) << 32;
 445    uint64_t high1 = (uint64_t)(high_bytes1[a_index]) << 32;
 446    uint64_t value0 = report0[a_index + 4] | high0;
 447    uint64_t value1 = report1[a_index + 4] | high1;
 448    uint64_t delta;
 449
 450    if (value0 > value1)
 451       delta = (1ULL << 40) + value1 - value0;
 452    else
 453       delta = value1 - value0;
 454
 455    *accumulator += delta;
 456 }
 457
 458 static void
 459 gen8_read_report_clock_ratios(const uint32_t *report,
 460                               uint64_t *slice_freq_hz,
 461                               uint64_t *unslice_freq_hz)
 462 {
 463    /* The lower 16bits of the RPT_ID field of the OA reports contains a
 464     * snapshot of the bits coming from the RP_FREQ_NORMAL register and is
 465     * divided this way :
 466     *
 467     * RPT_ID[31:25]: RP_FREQ_NORMAL[20:14] (low squashed_slice_clock_frequency)
 468     * RPT_ID[10:9]:  RP_FREQ_NORMAL[22:21] (high squashed_slice_clock_frequency)
 469     * RPT_ID[8:0]:   RP_FREQ_NORMAL[31:23] (squashed_unslice_clock_frequency)
 470     *
 471     * RP_FREQ_NORMAL[31:23]: Software Unslice Ratio Request
 472     *                        Multiple of 33.33MHz 2xclk (16 MHz 1xclk)
 473     *
 474     * RP_FREQ_NORMAL[22:14]: Software Slice Ratio Request
 475     *                        Multiple of 33.33MHz 2xclk (16 MHz 1xclk)
 476     */
 477
 478    uint32_t unslice_freq = report[0] & 0x1ff;
 479    uint32_t slice_freq_low = (report[0] >> 25) & 0x7f;
 480    uint32_t slice_freq_high = (report[0] >> 9) & 0x3;
 481    uint32_t slice_freq = slice_freq_low | (slice_freq_high << 7);
 482
 483    *slice_freq_hz = slice_freq * 16666667ULL;
 484    *unslice_freq_hz = unslice_freq * 16666667ULL;
 485 }
 486
 487 void
 488 gen_perf_query_result_read_frequencies(struct gen_perf_query_result *result,
 489                                        const struct gen_device_info *devinfo,
 490                                        const uint32_t *start,
 491                                        const uint32_t *end)
 492 {
 493    /* Slice/Unslice frequency is only available in the OA reports when the
 494     * "Disable OA reports due to clock ratio change" field in
 495     * OA_DEBUG_REGISTER is set to 1. This is how the kernel programs this
 496     * global register (see drivers/gpu/drm/i915/i915_perf.c)
 497     *
 498     * Documentation says this should be available on Gen9+ but experimentation
 499     * shows that Gen8 reports similar values, so we enable it there too.
 500     */
 501    if (devinfo->gen < 8)
 502       return;
 503
 504    gen8_read_report_clock_ratios(start,
 505                                  &result->slice_frequency[0],
 506                                  &result->unslice_frequency[0]);
 507    gen8_read_report_clock_ratios(end,
 508                                  &result->slice_frequency[1],
 509                                  &result->unslice_frequency[1]);
 510 }
 511
 512 void
 513 gen_perf_query_result_accumulate(struct gen_perf_query_result *result,
 514                                  const struct gen_perf_query_info *query,
 515                                  const uint32_t *start,
 516                                  const uint32_t *end)
 517 {
 518    int i, idx = 0;
 519
 520    result->hw_id = start[2];
 521    result->reports_accumulated++;
 522
 523    switch (query->oa_format) {
 524    case I915_OA_FORMAT_A32u40_A4u32_B8_C8:
 525       accumulate_uint32(start + 1, end + 1, result->accumulator + idx++); /* timestamp */
 526       accumulate_uint32(start + 3, end + 3, result->accumulator + idx++); /* clock */
 527
 528       /* 32x 40bit A counters... */
 529       for (i = 0; i < 32; i++)
 530          accumulate_uint40(i, start, end, result->accumulator + idx++);
 531
 532       /* 4x 32bit A counters... */
 533       for (i = 0; i < 4; i++)
 534          accumulate_uint32(start + 36 + i, end + 36 + i, result->accumulator + idx++);
 535
 536       /* 8x 32bit B counters + 8x 32bit C counters... */
 537       for (i = 0; i < 16; i++)
 538          accumulate_uint32(start + 48 + i, end + 48 + i, result->accumulator + idx++);
 539       break;
 540
 541    case I915_OA_FORMAT_A45_B8_C8:
 542       accumulate_uint32(start + 1, end + 1, result->accumulator); /* timestamp */
 543
 544       for (i = 0; i < 61; i++)
 545          accumulate_uint32(start + 3 + i, end + 3 + i, result->accumulator + 1 + i);
 546       break;
 547
 548    default:
 549       unreachable("Can't accumulate OA counters in unknown format");
 550    }
 551
 552 }
 553
 554 void
 555 gen_perf_query_result_clear(struct gen_perf_query_result *result)
 556 {
 557    memset(result, 0, sizeof(*result));
 558    result->hw_id = 0xffffffff; /* invalid */
 559 }