tools/perf/arch/arm/util/cs-etm.c

   1 /*
   2  * Copyright(C) 2015 Linaro Limited. All rights reserved.
   3  * Author: Mathieu Poirier <mathieu.poirier@linaro.org>
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 as published by
   7  * the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  *
  14  * You should have received a copy of the GNU General Public License along with
  15  * this program.  If not, see <http://www.gnu.org/licenses/>.
  16  */
  17
  18 #include <api/fs/fs.h>
  19 #include <linux/bitops.h>
  20 #include <linux/coresight-pmu.h>
  21 #include <linux/kernel.h>
  22 #include <linux/log2.h>
  23 #include <linux/types.h>
  24
  25 #include "cs-etm.h"
  26 #include "../../perf.h"
  27 #include "../../util/auxtrace.h"
  28 #include "../../util/cpumap.h"
  29 #include "../../util/evlist.h"
  30 #include "../../util/pmu.h"
  31 #include "../../util/thread_map.h"
  32 #include "../../util/cs-etm.h"
  33
  34 #include <stdlib.h>
  35
  36 struct cs_etm_recording {
  37         struct auxtrace_record  itr;
  38         struct perf_pmu         *cs_etm_pmu;
  39         struct perf_evlist      *evlist;
  40         bool                    snapshot_mode;
  41         size_t                  snapshot_size;
  42 };
  43
  44 static bool cs_etm_is_etmv4(struct auxtrace_record *itr, int cpu);
  45
  46 static int cs_etm_parse_snapshot_options(struct auxtrace_record *itr,
  47                                          struct record_opts *opts,
  48                                          const char *str)
  49 {
  50         struct cs_etm_recording *ptr =
  51                                 container_of(itr, struct cs_etm_recording, itr);
  52         unsigned long long snapshot_size = 0;
  53         char *endptr;
  54
  55         if (str) {
  56                 snapshot_size = strtoull(str, &endptr, 0);
  57                 if (*endptr || snapshot_size > SIZE_MAX)
  58                         return -1;
  59         }
  60
  61         opts->auxtrace_snapshot_mode = true;
  62         opts->auxtrace_snapshot_size = snapshot_size;
  63         ptr->snapshot_size = snapshot_size;
  64
  65         return 0;
  66 }
  67
  68 static int cs_etm_recording_options(struct auxtrace_record *itr,
  69                                     struct perf_evlist *evlist,
  70                                     struct record_opts *opts)
  71 {
  72         struct cs_etm_recording *ptr =
  73                                 container_of(itr, struct cs_etm_recording, itr);
  74         struct perf_pmu *cs_etm_pmu = ptr->cs_etm_pmu;
  75         struct perf_evsel *evsel, *cs_etm_evsel = NULL;
  76         const struct cpu_map *cpus = evlist->cpus;
  77         bool privileged = (geteuid() == 0 || perf_event_paranoid() < 0);
  78
  79         ptr->evlist = evlist;
  80         ptr->snapshot_mode = opts->auxtrace_snapshot_mode;
  81
  82         evlist__for_each(evlist, evsel) {
  83                 if (evsel->attr.type == cs_etm_pmu->type) {
  84                         if (cs_etm_evsel) {
  85                                 pr_err("There may be only one %s event\n",
  86                                        CORESIGHT_ETM_PMU_NAME);
  87                                 return -EINVAL;
  88                         }
  89                         evsel->attr.freq = 0;
  90                         evsel->attr.sample_period = 1;
  91                         cs_etm_evsel = evsel;
  92                         opts->full_auxtrace = true;
  93                 }
  94         }
  95
  96         /* no need to continue if at least one event of interest was found */
  97         if (!cs_etm_evsel)
  98                 return 0;
  99
 100         if (opts->use_clockid) {
 101                 pr_err("Cannot use clockid (-k option) with %s\n",
 102                        CORESIGHT_ETM_PMU_NAME);
 103                 return -EINVAL;
 104         }
 105
 106         /* we are in snapshot mode */
 107         if (opts->auxtrace_snapshot_mode) {
 108                 /*
 109                  * No size were given to '-S' or '-m,', so go with
 110                  * the default
 111                  */
 112                 if (!opts->auxtrace_snapshot_size &&
 113                     !opts->auxtrace_mmap_pages) {
 114                         if (privileged) {
 115                                 opts->auxtrace_mmap_pages = MiB(4) / page_size;
 116                         } else {
 117                                 opts->auxtrace_mmap_pages =
 118                                                         KiB(128) / page_size;
 119                                 if (opts->mmap_pages == UINT_MAX)
 120                                         opts->mmap_pages = KiB(256) / page_size;
 121                         }
 122                 } else if (!opts->auxtrace_mmap_pages && !privileged &&
 123                                                 opts->mmap_pages == UINT_MAX) {
 124                         opts->mmap_pages = KiB(256) / page_size;
 125                 }
 126
 127                 /*
 128                  * '-m,xyz' was specified but no snapshot size, so make the
 129                  * snapshot size as big as the auxtrace mmap area.
 130                  */
 131                 if (!opts->auxtrace_snapshot_size) {
 132                         opts->auxtrace_snapshot_size =
 133                                 opts->auxtrace_mmap_pages * (size_t)page_size;
 134                 }
 135
 136                 /*
 137                  * -Sxyz was specified but no auxtrace mmap area, so make the
 138                  * auxtrace mmap area big enough to fit the requested snapshot
 139                  * size.
 140                  */
 141                 if (!opts->auxtrace_mmap_pages) {
 142                         size_t sz = opts->auxtrace_snapshot_size;
 143
 144                         sz = round_up(sz, page_size) / page_size;
 145                         opts->auxtrace_mmap_pages = roundup_pow_of_two(sz);
 146                 }
 147
 148                 /* Snapshost size can't be bigger than the auxtrace area */
 149                 if (opts->auxtrace_snapshot_size >
 150                                 opts->auxtrace_mmap_pages * (size_t)page_size) {
 151                         pr_err("Snapshot size %zu must not be greater than AUX area tracing mmap size %zu\n",
 152                                opts->auxtrace_snapshot_size,
 153                                opts->auxtrace_mmap_pages * (size_t)page_size);
 154                         return -EINVAL;
 155                 }
 156
 157                 /* Something went wrong somewhere - this shouldn't happen */
 158                 if (!opts->auxtrace_snapshot_size ||
 159                     !opts->auxtrace_mmap_pages) {
 160                         pr_err("Failed to calculate default snapshot size and/or AUX area tracing mmap pages\n");
 161                         return -EINVAL;
 162                 }
 163         }
 164
 165         /* We are in full trace mode but '-m,xyz' wasn't specified */
 166         if (opts->full_auxtrace && !opts->auxtrace_mmap_pages) {
 167                 if (privileged) {
 168                         opts->auxtrace_mmap_pages = MiB(4) / page_size;
 169                 } else {
 170                         opts->auxtrace_mmap_pages = KiB(128) / page_size;
 171                         if (opts->mmap_pages == UINT_MAX)
 172                                 opts->mmap_pages = KiB(256) / page_size;
 173                 }
 174
 175         }
 176
 177         /* Validate auxtrace_mmap_pages provided by user */
 178         if (opts->auxtrace_mmap_pages) {
 179                 unsigned int max_page = (KiB(128) / page_size);
 180                 size_t sz = opts->auxtrace_mmap_pages * (size_t)page_size;
 181
 182                 if (!privileged &&
 183                     opts->auxtrace_mmap_pages > max_page) {
 184                         opts->auxtrace_mmap_pages = max_page;
 185                         pr_err("auxtrace too big, truncating to %d\n",
 186                                max_page);
 187                 }
 188
 189                 if (!is_power_of_2(sz)) {
 190                         pr_err("Invalid mmap size for %s: must be a power of 2\n",
 191                                CORESIGHT_ETM_PMU_NAME);
 192                         return -EINVAL;
 193                 }
 194         }
 195
 196         if (opts->auxtrace_snapshot_mode)
 197                 pr_debug2("%s snapshot size: %zu\n", CORESIGHT_ETM_PMU_NAME,
 198                           opts->auxtrace_snapshot_size);
 199
 200         if (cs_etm_evsel) {
 201                 /*
 202                  * To obtain the auxtrace buffer file descriptor, the auxtrace
 203                  * event must come first.
 204                  */
 205                 perf_evlist__to_front(evlist, cs_etm_evsel);
 206                 /*
 207                  * In the case of per-cpu mmaps, we need the CPU on the
 208                  * AUX event.
 209                  */
 210                 if (!cpu_map__empty(cpus))
 211                         perf_evsel__set_sample_bit(cs_etm_evsel, CPU);
 212         }
 213
 214         /* Add dummy event to keep tracking */
 215         if (opts->full_auxtrace) {
 216                 struct perf_evsel *tracking_evsel;
 217                 int err;
 218
 219                 err = parse_events(evlist, "dummy:u", NULL);
 220                 if (err)
 221                         return err;
 222
 223                 tracking_evsel = perf_evlist__last(evlist);
 224                 perf_evlist__set_tracking_event(evlist, tracking_evsel);
 225
 226                 tracking_evsel->attr.freq = 0;
 227                 tracking_evsel->attr.sample_period = 1;
 228
 229                 /* In per-cpu case, always need the time of mmap events etc */
 230                 if (!cpu_map__empty(cpus))
 231                         perf_evsel__set_sample_bit(tracking_evsel, TIME);
 232         }
 233
 234         return 0;
 235 }
 236
 237 static u64 cs_etm_get_config(struct auxtrace_record *itr)
 238 {
 239         u64 config = 0;
 240         struct cs_etm_recording *ptr =
 241                         container_of(itr, struct cs_etm_recording, itr);
 242         struct perf_pmu *cs_etm_pmu = ptr->cs_etm_pmu;
 243         struct perf_evlist *evlist = ptr->evlist;
 244         struct perf_evsel *evsel;
 245
 246         evlist__for_each(evlist, evsel) {
 247                 if (evsel->attr.type == cs_etm_pmu->type) {
 248                         /*
 249                          * Variable perf_event_attr::config is assigned to
 250                          * ETMv3/PTM.  The bit fields have been made to match
 251                          * the ETMv3.5 ETRMCR register specification.  See the
 252                          * PMU_FORMAT_ATTR() declarations in
 253                          * drivers/hwtracing/coresight/coresight-perf.c for
 254                          * details.
 255                          */
 256                         config = evsel->attr.config;
 257                         break;
 258                 }
 259         }
 260
 261         return config;
 262 }
 263
 264 static size_t
 265 cs_etm_info_priv_size(struct auxtrace_record *itr __maybe_unused,
 266                       struct perf_evlist *evlist __maybe_unused)
 267 {
 268         int i;
 269         int etmv3 = 0, etmv4 = 0;
 270         const struct cpu_map *cpus = evlist->cpus;
 271
 272         /* cpu map is not empty, we have specific CPUs to work with */
 273         if (!cpu_map__empty(cpus)) {
 274                 for (i = 0; i < cpu_map__nr(cpus); i++) {
 275                         if (cs_etm_is_etmv4(itr, cpus->map[i]))
 276                                 etmv4++;
 277                         else
 278                                 etmv3++;
 279                 }
 280         } else {
 281                 /* get configuration for all CPUs in the system */
 282                 for (i = 0; i < cpu__max_cpu(); i++) {
 283                         if (cs_etm_is_etmv4(itr, i))
 284                                 etmv4++;
 285                         else
 286                                 etmv3++;
 287                 }
 288         }
 289
 290         return (CS_ETM_HEADER_SIZE +
 291                (etmv4 * CS_ETMV4_PRIV_SIZE) +
 292                (etmv3 * CS_ETMV3_PRIV_SIZE));
 293 }
 294
 295 static const char *metadata_etmv3_ro[CS_ETM_PRIV_MAX] = {
 296         [CS_ETM_ETMCCER]        = "mgmt/etmccer",
 297         [CS_ETM_ETMIDR]         = "mgmt/etmidr",
 298 };
 299
 300 static const char *metadata_etmv4_ro[CS_ETMV4_PRIV_MAX] = {
 301         [CS_ETMV4_TRCIDR0]              = "trcidr/trcidr0",
 302         [CS_ETMV4_TRCIDR1]              = "trcidr/trcidr1",
 303         [CS_ETMV4_TRCIDR2]              = "trcidr/trcidr2",
 304         [CS_ETMV4_TRCIDR8]              = "trcidr/trcidr8",
 305         [CS_ETMV4_TRCAUTHSTATUS]        = "mgmt/trcauthstatus",
 306 };
 307
 308 static bool cs_etm_is_etmv4(struct auxtrace_record *itr, int cpu)
 309 {
 310         bool ret = false;
 311         char path[PATH_MAX];
 312         int scan;
 313         unsigned int val;
 314         struct cs_etm_recording *ptr =
 315                         container_of(itr, struct cs_etm_recording, itr);
 316         struct perf_pmu *cs_etm_pmu = ptr->cs_etm_pmu;
 317
 318         /* Take any of the RO files for ETMv4 and see if it present */
 319         snprintf(path, PATH_MAX, "cpu%d/%s",
 320                  cpu, metadata_etmv4_ro[CS_ETMV4_TRCIDR0]);
 321         scan = perf_pmu__scan_file(cs_etm_pmu, path, "%x", &val);
 322
 323         /* The file was read successfully, we have a winner */
 324         if (scan == 1)
 325                 ret = true;
 326
 327         return ret;
 328 }
 329
 330 static int cs_etm_get_ro(struct perf_pmu *pmu, int cpu, const char *path)
 331 {
 332         char pmu_path[PATH_MAX];
 333         int scan;
 334         unsigned int val = 0;
 335
 336         /* Get RO metadata from sysfs */
 337         snprintf(pmu_path, PATH_MAX, "cpu%d/%s", cpu, path);
 338
 339         scan = perf_pmu__scan_file(pmu, pmu_path, "%x", &val);
 340         if (scan != 1)
 341                 pr_err("%s: error reading: %s\n", __func__, pmu_path);
 342
 343         return val;
 344 }
 345
 346 static void cs_etm_get_metadata(int cpu, u32 *offset,
 347                                 struct auxtrace_record *itr,
 348                                 struct auxtrace_info_event *info)
 349 {
 350         u32 increment;
 351         u64 magic;
 352         struct cs_etm_recording *ptr =
 353                         container_of(itr, struct cs_etm_recording, itr);
 354         struct perf_pmu *cs_etm_pmu = ptr->cs_etm_pmu;
 355
 356         /* first see what kind of tracer this cpu is affined to */
 357         if (cs_etm_is_etmv4(itr, cpu)) {
 358                 magic = __perf_cs_etmv4_magic;
 359                 /* Get trace configuration register */
 360                 info->priv[*offset + CS_ETMV4_TRCCONFIGR] =
 361                                                 cs_etm_get_config(itr);
 362                 /* Get traceID from the framework */
 363                 info->priv[*offset + CS_ETMV4_TRCTRACEIDR] =
 364                                                 coresight_get_trace_id(cpu);
 365                 /* Get read-only information from sysFS */
 366                 info->priv[*offset + CS_ETMV4_TRCIDR0] =
 367                         cs_etm_get_ro(cs_etm_pmu, cpu,
 368                                       metadata_etmv4_ro[CS_ETMV4_TRCIDR0]);
 369                 info->priv[*offset + CS_ETMV4_TRCIDR1] =
 370                         cs_etm_get_ro(cs_etm_pmu, cpu,
 371                                       metadata_etmv4_ro[CS_ETMV4_TRCIDR1]);
 372                 info->priv[*offset + CS_ETMV4_TRCIDR2] =
 373                         cs_etm_get_ro(cs_etm_pmu, cpu,
 374                                       metadata_etmv4_ro[CS_ETMV4_TRCIDR2]);
 375                 info->priv[*offset + CS_ETMV4_TRCIDR8] =
 376                         cs_etm_get_ro(cs_etm_pmu, cpu,
 377                                       metadata_etmv4_ro[CS_ETMV4_TRCIDR8]);
 378                 info->priv[*offset + CS_ETMV4_TRCAUTHSTATUS] =
 379                         cs_etm_get_ro(cs_etm_pmu, cpu,
 380                                       metadata_etmv4_ro
 381                                       [CS_ETMV4_TRCAUTHSTATUS]);
 382
 383                 /* How much space was used */
 384                 increment = CS_ETMV4_PRIV_MAX;
 385         } else {
 386                 magic = __perf_cs_etmv3_magic;
 387                 /* Get configuration register */
 388                 info->priv[*offset + CS_ETM_ETMCR] = cs_etm_get_config(itr);
 389                 /* Get traceID from the framework */
 390                 info->priv[*offset + CS_ETM_ETMTRACEIDR] =
 391                                                 coresight_get_trace_id(cpu);
 392                 /* Get read-only information from sysFS */
 393                 info->priv[*offset + CS_ETM_ETMCCER] =
 394                         cs_etm_get_ro(cs_etm_pmu, cpu,
 395                                       metadata_etmv3_ro[CS_ETM_ETMCCER]);
 396                 info->priv[*offset + CS_ETM_ETMIDR] =
 397                         cs_etm_get_ro(cs_etm_pmu, cpu,
 398                                       metadata_etmv3_ro[CS_ETM_ETMIDR]);
 399
 400                 /* How much space was used */
 401                 increment = CS_ETM_PRIV_MAX;
 402         }
 403
 404         /* Build generic header portion */
 405         info->priv[*offset + CS_ETM_MAGIC] = magic;
 406         info->priv[*offset + CS_ETM_CPU] = cpu;
 407         /* Where the next CPU entry should start from */
 408         *offset += increment;
 409 }
 410
 411 static int cs_etm_info_fill(struct auxtrace_record *itr,
 412                             struct perf_session *session,
 413                             struct auxtrace_info_event *info,
 414                             size_t priv_size)
 415 {
 416         int i;
 417         u32 offset;
 418         u64 nr_cpu, type;
 419         const struct cpu_map *cpus = session->evlist->cpus;
 420         struct cs_etm_recording *ptr =
 421                         container_of(itr, struct cs_etm_recording, itr);
 422         struct perf_pmu *cs_etm_pmu = ptr->cs_etm_pmu;
 423
 424         if (priv_size != cs_etm_info_priv_size(itr, session->evlist))
 425                 return -EINVAL;
 426
 427         if (!session->evlist->nr_mmaps)
 428                 return -EINVAL;
 429
 430         /* If the cpu_map is empty all CPUs are involved */
 431         nr_cpu = cpu_map__empty(cpus) ? cpu__max_cpu() : cpu_map__nr(cpus);
 432         /* Get PMU type as dynamically assigned by the core */
 433         type = cs_etm_pmu->type;
 434
 435         /* First fill out the session header */
 436         info->type = PERF_AUXTRACE_CS_ETM;
 437         info->priv[CS_HEADER_VERSION_0] = 0;
 438         info->priv[CS_PMU_TYPE_CPUS] = type << 32;
 439         info->priv[CS_PMU_TYPE_CPUS] |= nr_cpu;
 440         info->priv[CS_ETM_SNAPSHOT] = ptr->snapshot_mode;
 441
 442         offset = CS_ETM_SNAPSHOT + 1;
 443
 444         /* cpu map is not empty, we have specific CPUs to work with */
 445         if (!cpu_map__empty(cpus)) {
 446                 for (i = 0; i < cpu_map__nr(cpus) && offset < priv_size; i++)
 447                         cs_etm_get_metadata(cpus->map[i], &offset, itr, info);
 448         } else {
 449                 /* get configuration for all CPUs in the system */
 450                 for (i = 0; i < cpu__max_cpu(); i++)
 451                         cs_etm_get_metadata(i, &offset, itr, info);
 452         }
 453
 454         return 0;
 455 }
 456
 457 static int cs_etm_find_snapshot(struct auxtrace_record *itr __maybe_unused,
 458                                 int idx, struct auxtrace_mmap *mm,
 459                                 unsigned char *data __maybe_unused,
 460                                 u64 *head, u64 *old)
 461 {
 462         pr_debug3("%s: mmap index %d old head %zu new head %zu size %zu\n",
 463                   __func__, idx, (size_t)*old, (size_t)*head, mm->len);
 464
 465         *old = *head;
 466         *head += mm->len;
 467
 468         return 0;
 469 }
 470
 471 static int cs_etm_snapshot_start(struct auxtrace_record *itr)
 472 {
 473         struct cs_etm_recording *ptr =
 474                         container_of(itr, struct cs_etm_recording, itr);
 475         struct perf_evsel *evsel;
 476
 477         evlist__for_each(ptr->evlist, evsel) {
 478                 if (evsel->attr.type == ptr->cs_etm_pmu->type)
 479                         return perf_evsel__disable(evsel);
 480         }
 481         return -EINVAL;
 482 }
 483
 484 static int cs_etm_snapshot_finish(struct auxtrace_record *itr)
 485 {
 486         struct cs_etm_recording *ptr =
 487                         container_of(itr, struct cs_etm_recording, itr);
 488         struct perf_evsel *evsel;
 489
 490         evlist__for_each(ptr->evlist, evsel) {
 491                 int nthreads = thread_map__nr(evsel->threads);
 492                 int ncpus = cpu_map__nr(evsel->cpus);
 493
 494                 if (evsel->attr.type == ptr->cs_etm_pmu->type) {
 495                         return perf_evsel__enable(evsel, ncpus, nthreads);
 496                 }
 497         }
 498         return -EINVAL;
 499 }
 500
 501 static u64 cs_etm_reference(struct auxtrace_record *itr __maybe_unused)
 502 {
 503         return (((u64) rand() <<  0) & 0x00000000FFFFFFFFull) |
 504                 (((u64) rand() << 32) & 0xFFFFFFFF00000000ull);
 505 }
 506
 507 static void cs_etm_recording_free(struct auxtrace_record *itr)
 508 {
 509         struct cs_etm_recording *ptr =
 510                         container_of(itr, struct cs_etm_recording, itr);
 511         free(ptr);
 512 }
 513
 514 static int cs_etm_read_finish(struct auxtrace_record *itr, int idx)
 515 {
 516         struct cs_etm_recording *ptr =
 517                         container_of(itr, struct cs_etm_recording, itr);
 518         struct perf_evsel *evsel;
 519
 520         evlist__for_each(ptr->evlist, evsel) {
 521                 if (evsel->attr.type == ptr->cs_etm_pmu->type)
 522                         return perf_evlist__enable_event_idx(ptr->evlist,
 523                                                              evsel, idx);
 524         }
 525
 526         return -EINVAL;
 527 }
 528
 529 struct auxtrace_record *cs_etm_record_init(int *err)
 530 {
 531         struct perf_pmu *cs_etm_pmu;
 532         struct cs_etm_recording *ptr;
 533
 534         cs_etm_pmu = perf_pmu__find(CORESIGHT_ETM_PMU_NAME);
 535
 536         if (!cs_etm_pmu) {
 537                 *err = -EINVAL;
 538                 goto out;
 539         }
 540
 541         ptr = zalloc(sizeof(struct cs_etm_recording));
 542         if (!ptr) {
 543                 *err = -ENOMEM;
 544                 goto out;
 545         }
 546
 547         ptr->cs_etm_pmu                 = cs_etm_pmu;
 548         ptr->itr.parse_snapshot_options = cs_etm_parse_snapshot_options;
 549         ptr->itr.recording_options      = cs_etm_recording_options;
 550         ptr->itr.info_priv_size         = cs_etm_info_priv_size;
 551         ptr->itr.info_fill              = cs_etm_info_fill;
 552         ptr->itr.find_snapshot          = cs_etm_find_snapshot;
 553         ptr->itr.snapshot_start         = cs_etm_snapshot_start;
 554         ptr->itr.snapshot_finish        = cs_etm_snapshot_finish;
 555         ptr->itr.reference              = cs_etm_reference;
 556         ptr->itr.free                   = cs_etm_recording_free;
 557         ptr->itr.read_finish            = cs_etm_read_finish;
 558
 559         *err = 0;
 560         return &ptr->itr;
 561 out:
 562         return NULL;
 563 }