tools/tracing/rtla/src/timerlat_hist.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2021 Red Hat Inc, Daniel Bristot de Oliveira <bristot@kernel.org>
   4  */
   5
   6 #include <getopt.h>
   7 #include <stdlib.h>
   8 #include <string.h>
   9 #include <signal.h>
  10 #include <unistd.h>
  11 #include <stdio.h>
  12 #include <time.h>
  13
  14 #include "utils.h"
  15 #include "osnoise.h"
  16 #include "timerlat.h"
  17
  18 struct timerlat_hist_params {
  19         char                    *cpus;
  20         char                    *monitored_cpus;
  21         char                    *trace_output;
  22         unsigned long long      runtime;
  23         long long               stop_us;
  24         long long               stop_total_us;
  25         long long               timerlat_period_us;
  26         long long               print_stack;
  27         int                     sleep_time;
  28         int                     output_divisor;
  29         int                     duration;
  30         int                     set_sched;
  31         int                     dma_latency;
  32         struct sched_attr       sched_param;
  33         struct trace_events     *events;
  34
  35         char                    no_irq;
  36         char                    no_thread;
  37         char                    no_header;
  38         char                    no_summary;
  39         char                    no_index;
  40         char                    with_zeros;
  41         int                     bucket_size;
  42         int                     entries;
  43 };
  44
  45 struct timerlat_hist_cpu {
  46         int                     *irq;
  47         int                     *thread;
  48
  49         int                     irq_count;
  50         int                     thread_count;
  51
  52         unsigned long long      min_irq;
  53         unsigned long long      sum_irq;
  54         unsigned long long      max_irq;
  55
  56         unsigned long long      min_thread;
  57         unsigned long long      sum_thread;
  58         unsigned long long      max_thread;
  59 };
  60
  61 struct timerlat_hist_data {
  62         struct timerlat_hist_cpu        *hist;
  63         int                             entries;
  64         int                             bucket_size;
  65         int                             nr_cpus;
  66 };
  67
  68 /*
  69  * timerlat_free_histogram - free runtime data
  70  */
  71 static void
  72 timerlat_free_histogram(struct timerlat_hist_data *data)
  73 {
  74         int cpu;
  75
  76         /* one histogram for IRQ and one for thread, per CPU */
  77         for (cpu = 0; cpu < data->nr_cpus; cpu++) {
  78                 if (data->hist[cpu].irq)
  79                         free(data->hist[cpu].irq);
  80
  81                 if (data->hist[cpu].thread)
  82                         free(data->hist[cpu].thread);
  83         }
  84
  85         /* one set of histograms per CPU */
  86         if (data->hist)
  87                 free(data->hist);
  88
  89         free(data);
  90 }
  91
  92 /*
  93  * timerlat_alloc_histogram - alloc runtime data
  94  */
  95 static struct timerlat_hist_data
  96 *timerlat_alloc_histogram(int nr_cpus, int entries, int bucket_size)
  97 {
  98         struct timerlat_hist_data *data;
  99         int cpu;
 100
 101         data = calloc(1, sizeof(*data));
 102         if (!data)
 103                 return NULL;
 104
 105         data->entries = entries;
 106         data->bucket_size = bucket_size;
 107         data->nr_cpus = nr_cpus;
 108
 109         /* one set of histograms per CPU */
 110         data->hist = calloc(1, sizeof(*data->hist) * nr_cpus);
 111         if (!data->hist)
 112                 goto cleanup;
 113
 114         /* one histogram for IRQ and one for thread, per cpu */
 115         for (cpu = 0; cpu < nr_cpus; cpu++) {
 116                 data->hist[cpu].irq = calloc(1, sizeof(*data->hist->irq) * (entries + 1));
 117                 if (!data->hist[cpu].irq)
 118                         goto cleanup;
 119                 data->hist[cpu].thread = calloc(1, sizeof(*data->hist->thread) * (entries + 1));
 120                 if (!data->hist[cpu].thread)
 121                         goto cleanup;
 122         }
 123
 124         /* set the min to max */
 125         for (cpu = 0; cpu < nr_cpus; cpu++) {
 126                 data->hist[cpu].min_irq = ~0;
 127                 data->hist[cpu].min_thread = ~0;
 128         }
 129
 130         return data;
 131
 132 cleanup:
 133         timerlat_free_histogram(data);
 134         return NULL;
 135 }
 136
 137 /*
 138  * timerlat_hist_update - record a new timerlat occurent on cpu, updating data
 139  */
 140 static void
 141 timerlat_hist_update(struct osnoise_tool *tool, int cpu,
 142                      unsigned long long thread,
 143                      unsigned long long latency)
 144 {
 145         struct timerlat_hist_params *params = tool->params;
 146         struct timerlat_hist_data *data = tool->data;
 147         int entries = data->entries;
 148         int bucket;
 149         int *hist;
 150
 151         if (params->output_divisor)
 152                 latency = latency / params->output_divisor;
 153
 154         if (data->bucket_size)
 155                 bucket = latency / data->bucket_size;
 156
 157         if (!thread) {
 158                 hist = data->hist[cpu].irq;
 159                 data->hist[cpu].irq_count++;
 160                 update_min(&data->hist[cpu].min_irq, &latency);
 161                 update_sum(&data->hist[cpu].sum_irq, &latency);
 162                 update_max(&data->hist[cpu].max_irq, &latency);
 163         } else {
 164                 hist = data->hist[cpu].thread;
 165                 data->hist[cpu].thread_count++;
 166                 update_min(&data->hist[cpu].min_thread, &latency);
 167                 update_sum(&data->hist[cpu].sum_thread, &latency);
 168                 update_max(&data->hist[cpu].max_thread, &latency);
 169         }
 170
 171         if (bucket < entries)
 172                 hist[bucket]++;
 173         else
 174                 hist[entries]++;
 175 }
 176
 177 /*
 178  * timerlat_hist_handler - this is the handler for timerlat tracer events
 179  */
 180 static int
 181 timerlat_hist_handler(struct trace_seq *s, struct tep_record *record,
 182                      struct tep_event *event, void *data)
 183 {
 184         struct trace_instance *trace = data;
 185         unsigned long long thread, latency;
 186         struct osnoise_tool *tool;
 187         int cpu = record->cpu;
 188
 189         tool = container_of(trace, struct osnoise_tool, trace);
 190
 191         tep_get_field_val(s, event, "context", record, &thread, 1);
 192         tep_get_field_val(s, event, "timer_latency", record, &latency, 1);
 193
 194         timerlat_hist_update(tool, cpu, thread, latency);
 195
 196         return 0;
 197 }
 198
 199 /*
 200  * timerlat_hist_header - print the header of the tracer to the output
 201  */
 202 static void timerlat_hist_header(struct osnoise_tool *tool)
 203 {
 204         struct timerlat_hist_params *params = tool->params;
 205         struct timerlat_hist_data *data = tool->data;
 206         struct trace_seq *s = tool->trace.seq;
 207         char duration[26];
 208         int cpu;
 209
 210         if (params->no_header)
 211                 return;
 212
 213         get_duration(tool->start_time, duration, sizeof(duration));
 214         trace_seq_printf(s, "# RTLA timerlat histogram\n");
 215         trace_seq_printf(s, "# Time unit is %s (%s)\n",
 216                         params->output_divisor == 1 ? "nanoseconds" : "microseconds",
 217                         params->output_divisor == 1 ? "ns" : "us");
 218
 219         trace_seq_printf(s, "# Duration: %s\n", duration);
 220
 221         if (!params->no_index)
 222                 trace_seq_printf(s, "Index");
 223
 224         for (cpu = 0; cpu < data->nr_cpus; cpu++) {
 225                 if (params->cpus && !params->monitored_cpus[cpu])
 226                         continue;
 227
 228                 if (!data->hist[cpu].irq_count && !data->hist[cpu].thread_count)
 229                         continue;
 230
 231                 if (!params->no_irq)
 232                         trace_seq_printf(s, "   IRQ-%03d", cpu);
 233
 234                 if (!params->no_thread)
 235                         trace_seq_printf(s, "   Thr-%03d", cpu);
 236         }
 237         trace_seq_printf(s, "\n");
 238
 239
 240         trace_seq_do_printf(s);
 241         trace_seq_reset(s);
 242 }
 243
 244 /*
 245  * timerlat_print_summary - print the summary of the hist data to the output
 246  */
 247 static void
 248 timerlat_print_summary(struct timerlat_hist_params *params,
 249                        struct trace_instance *trace,
 250                        struct timerlat_hist_data *data)
 251 {
 252         int cpu;
 253
 254         if (params->no_summary)
 255                 return;
 256
 257         if (!params->no_index)
 258                 trace_seq_printf(trace->seq, "count:");
 259
 260         for (cpu = 0; cpu < data->nr_cpus; cpu++) {
 261                 if (params->cpus && !params->monitored_cpus[cpu])
 262                         continue;
 263
 264                 if (!data->hist[cpu].irq_count && !data->hist[cpu].thread_count)
 265                         continue;
 266
 267                 if (!params->no_irq)
 268                         trace_seq_printf(trace->seq, "%9d ",
 269                                         data->hist[cpu].irq_count);
 270
 271                 if (!params->no_thread)
 272                         trace_seq_printf(trace->seq, "%9d ",
 273                                         data->hist[cpu].thread_count);
 274         }
 275         trace_seq_printf(trace->seq, "\n");
 276
 277         if (!params->no_index)
 278                 trace_seq_printf(trace->seq, "min:  ");
 279
 280         for (cpu = 0; cpu < data->nr_cpus; cpu++) {
 281                 if (params->cpus && !params->monitored_cpus[cpu])
 282                         continue;
 283
 284                 if (!data->hist[cpu].irq_count && !data->hist[cpu].thread_count)
 285                         continue;
 286
 287                 if (!params->no_irq)
 288                         trace_seq_printf(trace->seq, "%9llu ",
 289                                         data->hist[cpu].min_irq);
 290
 291                 if (!params->no_thread)
 292                         trace_seq_printf(trace->seq, "%9llu ",
 293                                         data->hist[cpu].min_thread);
 294         }
 295         trace_seq_printf(trace->seq, "\n");
 296
 297         if (!params->no_index)
 298                 trace_seq_printf(trace->seq, "avg:  ");
 299
 300         for (cpu = 0; cpu < data->nr_cpus; cpu++) {
 301                 if (params->cpus && !params->monitored_cpus[cpu])
 302                         continue;
 303
 304                 if (!data->hist[cpu].irq_count && !data->hist[cpu].thread_count)
 305                         continue;
 306
 307                 if (!params->no_irq) {
 308                         if (data->hist[cpu].irq_count)
 309                                 trace_seq_printf(trace->seq, "%9llu ",
 310                                                  data->hist[cpu].sum_irq / data->hist[cpu].irq_count);
 311                         else
 312                                 trace_seq_printf(trace->seq, "        - ");
 313                 }
 314
 315                 if (!params->no_thread) {
 316                         if (data->hist[cpu].thread_count)
 317                                 trace_seq_printf(trace->seq, "%9llu ",
 318                                                 data->hist[cpu].sum_thread / data->hist[cpu].thread_count);
 319                         else
 320                                 trace_seq_printf(trace->seq, "        - ");
 321                 }
 322         }
 323         trace_seq_printf(trace->seq, "\n");
 324
 325         if (!params->no_index)
 326                 trace_seq_printf(trace->seq, "max:  ");
 327
 328         for (cpu = 0; cpu < data->nr_cpus; cpu++) {
 329                 if (params->cpus && !params->monitored_cpus[cpu])
 330                         continue;
 331
 332                 if (!data->hist[cpu].irq_count && !data->hist[cpu].thread_count)
 333                         continue;
 334
 335                 if (!params->no_irq)
 336                         trace_seq_printf(trace->seq, "%9llu ",
 337                                         data->hist[cpu].max_irq);
 338
 339                 if (!params->no_thread)
 340                         trace_seq_printf(trace->seq, "%9llu ",
 341                                         data->hist[cpu].max_thread);
 342         }
 343         trace_seq_printf(trace->seq, "\n");
 344         trace_seq_do_printf(trace->seq);
 345         trace_seq_reset(trace->seq);
 346 }
 347
 348 /*
 349  * timerlat_print_stats - print data for all CPUs
 350  */
 351 static void
 352 timerlat_print_stats(struct timerlat_hist_params *params, struct osnoise_tool *tool)
 353 {
 354         struct timerlat_hist_data *data = tool->data;
 355         struct trace_instance *trace = &tool->trace;
 356         int bucket, cpu;
 357         int total;
 358
 359         timerlat_hist_header(tool);
 360
 361         for (bucket = 0; bucket < data->entries; bucket++) {
 362                 total = 0;
 363
 364                 if (!params->no_index)
 365                         trace_seq_printf(trace->seq, "%-6d",
 366                                          bucket * data->bucket_size);
 367
 368                 for (cpu = 0; cpu < data->nr_cpus; cpu++) {
 369                         if (params->cpus && !params->monitored_cpus[cpu])
 370                                 continue;
 371
 372                         if (!data->hist[cpu].irq_count && !data->hist[cpu].thread_count)
 373                                 continue;
 374
 375                         if (!params->no_irq) {
 376                                 total += data->hist[cpu].irq[bucket];
 377                                 trace_seq_printf(trace->seq, "%9d ",
 378                                                 data->hist[cpu].irq[bucket]);
 379                         }
 380
 381                         if (!params->no_thread) {
 382                                 total += data->hist[cpu].thread[bucket];
 383                                 trace_seq_printf(trace->seq, "%9d ",
 384                                                 data->hist[cpu].thread[bucket]);
 385                         }
 386
 387                 }
 388
 389                 if (total == 0 && !params->with_zeros) {
 390                         trace_seq_reset(trace->seq);
 391                         continue;
 392                 }
 393
 394                 trace_seq_printf(trace->seq, "\n");
 395                 trace_seq_do_printf(trace->seq);
 396                 trace_seq_reset(trace->seq);
 397         }
 398
 399         if (!params->no_index)
 400                 trace_seq_printf(trace->seq, "over: ");
 401
 402         for (cpu = 0; cpu < data->nr_cpus; cpu++) {
 403                 if (params->cpus && !params->monitored_cpus[cpu])
 404                         continue;
 405
 406                 if (!data->hist[cpu].irq_count && !data->hist[cpu].thread_count)
 407                         continue;
 408
 409                 if (!params->no_irq)
 410                         trace_seq_printf(trace->seq, "%9d ",
 411                                          data->hist[cpu].irq[data->entries]);
 412
 413                 if (!params->no_thread)
 414                         trace_seq_printf(trace->seq, "%9d ",
 415                                          data->hist[cpu].thread[data->entries]);
 416         }
 417         trace_seq_printf(trace->seq, "\n");
 418         trace_seq_do_printf(trace->seq);
 419         trace_seq_reset(trace->seq);
 420
 421         timerlat_print_summary(params, trace, data);
 422 }
 423
 424 /*
 425  * timerlat_hist_usage - prints timerlat top usage message
 426  */
 427 static void timerlat_hist_usage(char *usage)
 428 {
 429         int i;
 430
 431         char *msg[] = {
 432                 "",
 433                 "  usage: [rtla] timerlat hist [-h] [-q] [-d s] [-D] [-n] [-a us] [-p us] [-i us] [-T us] [-s us] \\",
 434                 "         [-t[=file]] [-e sys[:event]] [--filter <filter>] [--trigger <trigger>] [-c cpu-list] \\",
 435                 "         [-P priority] [-E N] [-b N] [--no-irq] [--no-thread] [--no-header] [--no-summary] \\",
 436                 "         [--no-index] [--with-zeros] [--dma-latency us]",
 437                 "",
 438                 "         -h/--help: print this menu",
 439                 "         -a/--auto: set automatic trace mode, stopping the session if argument in us latency is hit",
 440                 "         -p/--period us: timerlat period in us",
 441                 "         -i/--irq us: stop trace if the irq latency is higher than the argument in us",
 442                 "         -T/--thread us: stop trace if the thread latency is higher than the argument in us",
 443                 "         -s/--stack us: save the stack trace at the IRQ if a thread latency is higher than the argument in us",
 444                 "         -c/--cpus cpus: run the tracer only on the given cpus",
 445                 "         -d/--duration time[m|h|d]: duration of the session in seconds",
 446                 "         -D/--debug: print debug info",
 447                 "         -t/--trace[=file]: save the stopped trace to [file|timerlat_trace.txt]",
 448                 "         -e/--event <sys:event>: enable the <sys:event> in the trace instance, multiple -e are allowed",
 449                 "            --filter <filter>: enable a trace event filter to the previous -e event",
 450                 "            --trigger <trigger>: enable a trace event trigger to the previous -e event",
 451                 "         -n/--nano: display data in nanoseconds",
 452                 "         -b/--bucket-size N: set the histogram bucket size (default 1)",
 453                 "         -E/--entries N: set the number of entries of the histogram (default 256)",
 454                 "            --no-irq: ignore IRQ latencies",
 455                 "            --no-thread: ignore thread latencies",
 456                 "            --no-header: do not print header",
 457                 "            --no-summary: do not print summary",
 458                 "            --no-index: do not print index",
 459                 "            --with-zeros: print zero only entries",
 460                 "            --dma-latency us: set /dev/cpu_dma_latency latency <us> to reduce exit from idle latency",
 461                 "         -P/--priority o:prio|r:prio|f:prio|d:runtime:period : set scheduling parameters",
 462                 "               o:prio - use SCHED_OTHER with prio",
 463                 "               r:prio - use SCHED_RR with prio",
 464                 "               f:prio - use SCHED_FIFO with prio",
 465                 "               d:runtime[us|ms|s]:period[us|ms|s] - use SCHED_DEADLINE with runtime and period",
 466                 "                                                      in nanoseconds",
 467                 NULL,
 468         };
 469
 470         if (usage)
 471                 fprintf(stderr, "%s\n", usage);
 472
 473         fprintf(stderr, "rtla timerlat hist: a per-cpu histogram of the timer latency (version %s)\n",
 474                         VERSION);
 475
 476         for (i = 0; msg[i]; i++)
 477                 fprintf(stderr, "%s\n", msg[i]);
 478         exit(1);
 479 }
 480
 481 /*
 482  * timerlat_hist_parse_args - allocs, parse and fill the cmd line parameters
 483  */
 484 static struct timerlat_hist_params
 485 *timerlat_hist_parse_args(int argc, char *argv[])
 486 {
 487         struct timerlat_hist_params *params;
 488         struct trace_events *tevent;
 489         int auto_thresh;
 490         int retval;
 491         int c;
 492
 493         params = calloc(1, sizeof(*params));
 494         if (!params)
 495                 exit(1);
 496
 497         /* disabled by default */
 498         params->dma_latency = -1;
 499
 500         /* display data in microseconds */
 501         params->output_divisor = 1000;
 502         params->bucket_size = 1;
 503         params->entries = 256;
 504
 505         while (1) {
 506                 static struct option long_options[] = {
 507                         {"auto",                required_argument,      0, 'a'},
 508                         {"cpus",                required_argument,      0, 'c'},
 509                         {"bucket-size",         required_argument,      0, 'b'},
 510                         {"debug",               no_argument,            0, 'D'},
 511                         {"entries",             required_argument,      0, 'E'},
 512                         {"duration",            required_argument,      0, 'd'},
 513                         {"help",                no_argument,            0, 'h'},
 514                         {"irq",                 required_argument,      0, 'i'},
 515                         {"nano",                no_argument,            0, 'n'},
 516                         {"period",              required_argument,      0, 'p'},
 517                         {"priority",            required_argument,      0, 'P'},
 518                         {"stack",               required_argument,      0, 's'},
 519                         {"thread",              required_argument,      0, 'T'},
 520                         {"trace",               optional_argument,      0, 't'},
 521                         {"event",               required_argument,      0, 'e'},
 522                         {"no-irq",              no_argument,            0, '0'},
 523                         {"no-thread",           no_argument,            0, '1'},
 524                         {"no-header",           no_argument,            0, '2'},
 525                         {"no-summary",          no_argument,            0, '3'},
 526                         {"no-index",            no_argument,            0, '4'},
 527                         {"with-zeros",          no_argument,            0, '5'},
 528                         {"trigger",             required_argument,      0, '6'},
 529                         {"filter",              required_argument,      0, '7'},
 530                         {"dma-latency",         required_argument,      0, '8'},
 531                         {0, 0, 0, 0}
 532                 };
 533
 534                 /* getopt_long stores the option index here. */
 535                 int option_index = 0;
 536
 537                 c = getopt_long(argc, argv, "a:c:b:d:e:E:Dhi:np:P:s:t::T:0123456:7:8:",
 538                                  long_options, &option_index);
 539
 540                 /* detect the end of the options. */
 541                 if (c == -1)
 542                         break;
 543
 544                 switch (c) {
 545                 case 'a':
 546                         auto_thresh = get_llong_from_str(optarg);
 547
 548                         /* set thread stop to auto_thresh */
 549                         params->stop_total_us = auto_thresh;
 550
 551                         /* get stack trace */
 552                         params->print_stack = auto_thresh;
 553
 554                         /* set trace */
 555                         params->trace_output = "timerlat_trace.txt";
 556
 557                         break;
 558                 case 'c':
 559                         retval = parse_cpu_list(optarg, &params->monitored_cpus);
 560                         if (retval)
 561                                 timerlat_hist_usage("\nInvalid -c cpu list\n");
 562                         params->cpus = optarg;
 563                         break;
 564                 case 'b':
 565                         params->bucket_size = get_llong_from_str(optarg);
 566                         if ((params->bucket_size == 0) || (params->bucket_size >= 1000000))
 567                                 timerlat_hist_usage("Bucket size needs to be > 0 and <= 1000000\n");
 568                         break;
 569                 case 'D':
 570                         config_debug = 1;
 571                         break;
 572                 case 'd':
 573                         params->duration = parse_seconds_duration(optarg);
 574                         if (!params->duration)
 575                                 timerlat_hist_usage("Invalid -D duration\n");
 576                         break;
 577                 case 'e':
 578                         tevent = trace_event_alloc(optarg);
 579                         if (!tevent) {
 580                                 err_msg("Error alloc trace event");
 581                                 exit(EXIT_FAILURE);
 582                         }
 583
 584                         if (params->events)
 585                                 tevent->next = params->events;
 586
 587                         params->events = tevent;
 588                         break;
 589                 case 'E':
 590                         params->entries = get_llong_from_str(optarg);
 591                         if ((params->entries < 10) || (params->entries > 9999999))
 592                                         timerlat_hist_usage("Entries must be > 10 and < 9999999\n");
 593                         break;
 594                 case 'h':
 595                 case '?':
 596                         timerlat_hist_usage(NULL);
 597                         break;
 598                 case 'i':
 599                         params->stop_us = get_llong_from_str(optarg);
 600                         break;
 601                 case 'n':
 602                         params->output_divisor = 1;
 603                         break;
 604                 case 'p':
 605                         params->timerlat_period_us = get_llong_from_str(optarg);
 606                         if (params->timerlat_period_us > 1000000)
 607                                 timerlat_hist_usage("Period longer than 1 s\n");
 608                         break;
 609                 case 'P':
 610                         retval = parse_prio(optarg, &params->sched_param);
 611                         if (retval == -1)
 612                                 timerlat_hist_usage("Invalid -P priority");
 613                         params->set_sched = 1;
 614                         break;
 615                 case 's':
 616                         params->print_stack = get_llong_from_str(optarg);
 617                         break;
 618                 case 'T':
 619                         params->stop_total_us = get_llong_from_str(optarg);
 620                         break;
 621                 case 't':
 622                         if (optarg)
 623                                 /* skip = */
 624                                 params->trace_output = &optarg[1];
 625                         else
 626                                 params->trace_output = "timerlat_trace.txt";
 627                         break;
 628                 case '0': /* no irq */
 629                         params->no_irq = 1;
 630                         break;
 631                 case '1': /* no thread */
 632                         params->no_thread = 1;
 633                         break;
 634                 case '2': /* no header */
 635                         params->no_header = 1;
 636                         break;
 637                 case '3': /* no summary */
 638                         params->no_summary = 1;
 639                         break;
 640                 case '4': /* no index */
 641                         params->no_index = 1;
 642                         break;
 643                 case '5': /* with zeros */
 644                         params->with_zeros = 1;
 645                         break;
 646                 case '6': /* trigger */
 647                         if (params->events) {
 648                                 retval = trace_event_add_trigger(params->events, optarg);
 649                                 if (retval) {
 650                                         err_msg("Error adding trigger %s\n", optarg);
 651                                         exit(EXIT_FAILURE);
 652                                 }
 653                         } else {
 654                                 timerlat_hist_usage("--trigger requires a previous -e\n");
 655                         }
 656                         break;
 657                 case '7': /* filter */
 658                         if (params->events) {
 659                                 retval = trace_event_add_filter(params->events, optarg);
 660                                 if (retval) {
 661                                         err_msg("Error adding filter %s\n", optarg);
 662                                         exit(EXIT_FAILURE);
 663                                 }
 664                         } else {
 665                                 timerlat_hist_usage("--filter requires a previous -e\n");
 666                         }
 667                         break;
 668                 case '8':
 669                         params->dma_latency = get_llong_from_str(optarg);
 670                         if (params->dma_latency < 0 || params->dma_latency > 10000) {
 671                                 err_msg("--dma-latency needs to be >= 0 and < 10000");
 672                                 exit(EXIT_FAILURE);
 673                         }
 674                         break;
 675                 default:
 676                         timerlat_hist_usage("Invalid option");
 677                 }
 678         }
 679
 680         if (geteuid()) {
 681                 err_msg("rtla needs root permission\n");
 682                 exit(EXIT_FAILURE);
 683         }
 684
 685         if (params->no_irq && params->no_thread)
 686                 timerlat_hist_usage("no-irq and no-thread set, there is nothing to do here");
 687
 688         if (params->no_index && !params->with_zeros)
 689                 timerlat_hist_usage("no-index set with with-zeros is not set - it does not make sense");
 690
 691         return params;
 692 }
 693
 694 /*
 695  * timerlat_hist_apply_config - apply the hist configs to the initialized tool
 696  */
 697 static int
 698 timerlat_hist_apply_config(struct osnoise_tool *tool, struct timerlat_hist_params *params)
 699 {
 700         int retval;
 701
 702         if (!params->sleep_time)
 703                 params->sleep_time = 1;
 704
 705         if (params->cpus) {
 706                 retval = osnoise_set_cpus(tool->context, params->cpus);
 707                 if (retval) {
 708                         err_msg("Failed to apply CPUs config\n");
 709                         goto out_err;
 710                 }
 711         }
 712
 713         if (params->stop_us) {
 714                 retval = osnoise_set_stop_us(tool->context, params->stop_us);
 715                 if (retval) {
 716                         err_msg("Failed to set stop us\n");
 717                         goto out_err;
 718                 }
 719         }
 720
 721         if (params->stop_total_us) {
 722                 retval = osnoise_set_stop_total_us(tool->context, params->stop_total_us);
 723                 if (retval) {
 724                         err_msg("Failed to set stop total us\n");
 725                         goto out_err;
 726                 }
 727         }
 728
 729         if (params->timerlat_period_us) {
 730                 retval = osnoise_set_timerlat_period_us(tool->context, params->timerlat_period_us);
 731                 if (retval) {
 732                         err_msg("Failed to set timerlat period\n");
 733                         goto out_err;
 734                 }
 735         }
 736
 737         if (params->print_stack) {
 738                 retval = osnoise_set_print_stack(tool->context, params->print_stack);
 739                 if (retval) {
 740                         err_msg("Failed to set print stack\n");
 741                         goto out_err;
 742                 }
 743         }
 744
 745         return 0;
 746
 747 out_err:
 748         return -1;
 749 }
 750
 751 /*
 752  * timerlat_init_hist - initialize a timerlat hist tool with parameters
 753  */
 754 static struct osnoise_tool
 755 *timerlat_init_hist(struct timerlat_hist_params *params)
 756 {
 757         struct osnoise_tool *tool;
 758         int nr_cpus;
 759
 760         nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
 761
 762         tool = osnoise_init_tool("timerlat_hist");
 763         if (!tool)
 764                 return NULL;
 765
 766         tool->data = timerlat_alloc_histogram(nr_cpus, params->entries, params->bucket_size);
 767         if (!tool->data)
 768                 goto out_err;
 769
 770         tool->params = params;
 771
 772         tep_register_event_handler(tool->trace.tep, -1, "ftrace", "timerlat",
 773                                    timerlat_hist_handler, tool);
 774
 775         return tool;
 776
 777 out_err:
 778         osnoise_destroy_tool(tool);
 779         return NULL;
 780 }
 781
 782 static int stop_tracing;
 783 static void stop_hist(int sig)
 784 {
 785         stop_tracing = 1;
 786 }
 787
 788 /*
 789  * timerlat_hist_set_signals - handles the signal to stop the tool
 790  */
 791 static void
 792 timerlat_hist_set_signals(struct timerlat_hist_params *params)
 793 {
 794         signal(SIGINT, stop_hist);
 795         if (params->duration) {
 796                 signal(SIGALRM, stop_hist);
 797                 alarm(params->duration);
 798         }
 799 }
 800
 801 int timerlat_hist_main(int argc, char *argv[])
 802 {
 803         struct timerlat_hist_params *params;
 804         struct osnoise_tool *record = NULL;
 805         struct osnoise_tool *tool = NULL;
 806         struct trace_instance *trace;
 807         int dma_latency_fd = -1;
 808         int return_value = 1;
 809         int retval;
 810
 811         params = timerlat_hist_parse_args(argc, argv);
 812         if (!params)
 813                 exit(1);
 814
 815         tool = timerlat_init_hist(params);
 816         if (!tool) {
 817                 err_msg("Could not init osnoise hist\n");
 818                 goto out_exit;
 819         }
 820
 821         retval = timerlat_hist_apply_config(tool, params);
 822         if (retval) {
 823                 err_msg("Could not apply config\n");
 824                 goto out_free;
 825         }
 826
 827         trace = &tool->trace;
 828
 829         retval = enable_timerlat(trace);
 830         if (retval) {
 831                 err_msg("Failed to enable timerlat tracer\n");
 832                 goto out_free;
 833         }
 834
 835         if (params->set_sched) {
 836                 retval = set_comm_sched_attr("timerlat/", &params->sched_param);
 837                 if (retval) {
 838                         err_msg("Failed to set sched parameters\n");
 839                         goto out_free;
 840                 }
 841         }
 842
 843         if (params->dma_latency >= 0) {
 844                 dma_latency_fd = set_cpu_dma_latency(params->dma_latency);
 845                 if (dma_latency_fd < 0) {
 846                         err_msg("Could not set /dev/cpu_dma_latency.\n");
 847                         goto out_free;
 848                 }
 849         }
 850
 851         trace_instance_start(trace);
 852
 853         if (params->trace_output) {
 854                 record = osnoise_init_trace_tool("timerlat");
 855                 if (!record) {
 856                         err_msg("Failed to enable the trace instance\n");
 857                         goto out_free;
 858                 }
 859
 860                 if (params->events) {
 861                         retval = trace_events_enable(&record->trace, params->events);
 862                         if (retval)
 863                                 goto out_hist;
 864                 }
 865
 866                 trace_instance_start(&record->trace);
 867         }
 868
 869         tool->start_time = time(NULL);
 870         timerlat_hist_set_signals(params);
 871
 872         while (!stop_tracing) {
 873                 sleep(params->sleep_time);
 874
 875                 retval = tracefs_iterate_raw_events(trace->tep,
 876                                                     trace->inst,
 877                                                     NULL,
 878                                                     0,
 879                                                     collect_registered_events,
 880                                                     trace);
 881                 if (retval < 0) {
 882                         err_msg("Error iterating on events\n");
 883                         goto out_hist;
 884                 }
 885
 886                 if (trace_is_off(&tool->trace, &record->trace))
 887                         break;
 888         }
 889
 890         timerlat_print_stats(params, tool);
 891
 892         return_value = 0;
 893
 894         if (trace_is_off(&tool->trace, &record->trace)) {
 895                 printf("rtla timelat hit stop tracing\n");
 896                 if (params->trace_output) {
 897                         printf("  Saving trace to %s\n", params->trace_output);
 898                         save_trace_to_file(record->trace.inst, params->trace_output);
 899                 }
 900         }
 901
 902 out_hist:
 903         if (dma_latency_fd >= 0)
 904                 close(dma_latency_fd);
 905         trace_events_destroy(&record->trace, params->events);
 906         params->events = NULL;
 907 out_free:
 908         timerlat_free_histogram(tool->data);
 909         osnoise_destroy_tool(record);
 910         osnoise_destroy_tool(tool);
 911         free(params);
 912 out_exit:
 913         exit(return_value);
 914 }