kernel/trace/ring_buffer_benchmark.c

   1 /*
   2  * ring buffer tester and benchmark
   3  *
   4  * Copyright (C) 2009 Steven Rostedt <[email protected]>
   5  */
   6 #include <linux/ring_buffer.h>
   7 #include <linux/completion.h>
   8 #include <linux/kthread.h>
   9 #include <uapi/linux/sched/types.h>
  10 #include <linux/module.h>
  11 #include <linux/ktime.h>
  12 #include <asm/local.h>
  13
  14 struct rb_page {
  15         u64             ts;
  16         local_t         commit;
  17         char            data[4080];
  18 };
  19
  20 /* run time and sleep time in seconds */
  21 #define RUN_TIME        10ULL
  22 #define SLEEP_TIME      10
  23
  24 /* number of events for writer to wake up the reader */
  25 static int wakeup_interval = 100;
  26
  27 static int reader_finish;
  28 static DECLARE_COMPLETION(read_start);
  29 static DECLARE_COMPLETION(read_done);
  30
  31 static struct ring_buffer *buffer;
  32 static struct task_struct *producer;
  33 static struct task_struct *consumer;
  34 static unsigned long read;
  35
  36 static unsigned int disable_reader;
  37 module_param(disable_reader, uint, 0644);
  38 MODULE_PARM_DESC(disable_reader, "only run producer");
  39
  40 static unsigned int write_iteration = 50;
  41 module_param(write_iteration, uint, 0644);
  42 MODULE_PARM_DESC(write_iteration, "# of writes between timestamp readings");
  43
  44 static int producer_nice = MAX_NICE;
  45 static int consumer_nice = MAX_NICE;
  46
  47 static int producer_fifo = -1;
  48 static int consumer_fifo = -1;
  49
  50 module_param(producer_nice, int, 0644);
  51 MODULE_PARM_DESC(producer_nice, "nice prio for producer");
  52
  53 module_param(consumer_nice, int, 0644);
  54 MODULE_PARM_DESC(consumer_nice, "nice prio for consumer");
  55
  56 module_param(producer_fifo, int, 0644);
  57 MODULE_PARM_DESC(producer_fifo, "fifo prio for producer");
  58
  59 module_param(consumer_fifo, int, 0644);
  60 MODULE_PARM_DESC(consumer_fifo, "fifo prio for consumer");
  61
  62 static int read_events;
  63
  64 static int test_error;
  65
  66 #define TEST_ERROR()                            \
  67         do {                                    \
  68                 if (!test_error) {              \
  69                         test_error = 1;         \
  70                         WARN_ON(1);             \
  71                 }                               \
  72         } while (0)
  73
  74 enum event_status {
  75         EVENT_FOUND,
  76         EVENT_DROPPED,
  77 };
  78
  79 static bool break_test(void)
  80 {
  81         return test_error || kthread_should_stop();
  82 }
  83
  84 static enum event_status read_event(int cpu)
  85 {
  86         struct ring_buffer_event *event;
  87         int *entry;
  88         u64 ts;
  89
  90         event = ring_buffer_consume(buffer, cpu, &ts, NULL);
  91         if (!event)
  92                 return EVENT_DROPPED;
  93
  94         entry = ring_buffer_event_data(event);
  95         if (*entry != cpu) {
  96                 TEST_ERROR();
  97                 return EVENT_DROPPED;
  98         }
  99
 100         read++;
 101         return EVENT_FOUND;
 102 }
 103
 104 static enum event_status read_page(int cpu)
 105 {
 106         struct ring_buffer_event *event;
 107         struct rb_page *rpage;
 108         unsigned long commit;
 109         void *bpage;
 110         int *entry;
 111         int ret;
 112         int inc;
 113         int i;
 114
 115         bpage = ring_buffer_alloc_read_page(buffer, cpu);
 116         if (IS_ERR(bpage))
 117                 return EVENT_DROPPED;
 118
 119         ret = ring_buffer_read_page(buffer, &bpage, PAGE_SIZE, cpu, 1);
 120         if (ret >= 0) {
 121                 rpage = bpage;
 122                 /* The commit may have missed event flags set, clear them */
 123                 commit = local_read(&rpage->commit) & 0xfffff;
 124                 for (i = 0; i < commit && !test_error ; i += inc) {
 125
 126                         if (i >= (PAGE_SIZE - offsetof(struct rb_page, data))) {
 127                                 TEST_ERROR();
 128                                 break;
 129                         }
 130
 131                         inc = -1;
 132                         event = (void *)&rpage->data[i];
 133                         switch (event->type_len) {
 134                         case RINGBUF_TYPE_PADDING:
 135                                 /* failed writes may be discarded events */
 136                                 if (!event->time_delta)
 137                                         TEST_ERROR();
 138                                 inc = event->array[0] + 4;
 139                                 break;
 140                         case RINGBUF_TYPE_TIME_EXTEND:
 141                                 inc = 8;
 142                                 break;
 143                         case 0:
 144                                 entry = ring_buffer_event_data(event);
 145                                 if (*entry != cpu) {
 146                                         TEST_ERROR();
 147                                         break;
 148                                 }
 149                                 read++;
 150                                 if (!event->array[0]) {
 151                                         TEST_ERROR();
 152                                         break;
 153                                 }
 154                                 inc = event->array[0] + 4;
 155                                 break;
 156                         default:
 157                                 entry = ring_buffer_event_data(event);
 158                                 if (*entry != cpu) {
 159                                         TEST_ERROR();
 160                                         break;
 161                                 }
 162                                 read++;
 163                                 inc = ((event->type_len + 1) * 4);
 164                         }
 165                         if (test_error)
 166                                 break;
 167
 168                         if (inc <= 0) {
 169                                 TEST_ERROR();
 170                                 break;
 171                         }
 172                 }
 173         }
 174         ring_buffer_free_read_page(buffer, cpu, bpage);
 175
 176         if (ret < 0)
 177                 return EVENT_DROPPED;
 178         return EVENT_FOUND;
 179 }
 180
 181 static void ring_buffer_consumer(void)
 182 {
 183         /* toggle between reading pages and events */
 184         read_events ^= 1;
 185
 186         read = 0;
 187         /*
 188          * Continue running until the producer specifically asks to stop
 189          * and is ready for the completion.
 190          */
 191         while (!READ_ONCE(reader_finish)) {
 192                 int found = 1;
 193
 194                 while (found && !test_error) {
 195                         int cpu;
 196
 197                         found = 0;
 198                         for_each_online_cpu(cpu) {
 199                                 enum event_status stat;
 200
 201                                 if (read_events)
 202                                         stat = read_event(cpu);
 203                                 else
 204                                         stat = read_page(cpu);
 205
 206                                 if (test_error)
 207                                         break;
 208
 209                                 if (stat == EVENT_FOUND)
 210                                         found = 1;
 211
 212                         }
 213                 }
 214
 215                 /* Wait till the producer wakes us up when there is more data
 216                  * available or when the producer wants us to finish reading.
 217                  */
 218                 set_current_state(TASK_INTERRUPTIBLE);
 219                 if (reader_finish)
 220                         break;
 221
 222                 schedule();
 223         }
 224         __set_current_state(TASK_RUNNING);
 225         reader_finish = 0;
 226         complete(&read_done);
 227 }
 228
 229 static void ring_buffer_producer(void)
 230 {
 231         ktime_t start_time, end_time, timeout;
 232         unsigned long long time;
 233         unsigned long long entries;
 234         unsigned long long overruns;
 235         unsigned long missed = 0;
 236         unsigned long hit = 0;
 237         unsigned long avg;
 238         int cnt = 0;
 239
 240         /*
 241          * Hammer the buffer for 10 secs (this may
 242          * make the system stall)
 243          */
 244         trace_printk("Starting ring buffer hammer\n");
 245         start_time = ktime_get();
 246         timeout = ktime_add_ns(start_time, RUN_TIME * NSEC_PER_SEC);
 247         do {
 248                 struct ring_buffer_event *event;
 249                 int *entry;
 250                 int i;
 251
 252                 for (i = 0; i < write_iteration; i++) {
 253                         event = ring_buffer_lock_reserve(buffer, 10);
 254                         if (!event) {
 255                                 missed++;
 256                         } else {
 257                                 hit++;
 258                                 entry = ring_buffer_event_data(event);
 259                                 *entry = smp_processor_id();
 260                                 ring_buffer_unlock_commit(buffer, event);
 261                         }
 262                 }
 263                 end_time = ktime_get();
 264
 265                 cnt++;
 266                 if (consumer && !(cnt % wakeup_interval))
 267                         wake_up_process(consumer);
 268
 269 #ifndef CONFIG_PREEMPT
 270                 /*
 271                  * If we are a non preempt kernel, the 10 second run will
 272                  * stop everything while it runs. Instead, we will call
 273                  * cond_resched and also add any time that was lost by a
 274                  * rescedule.
 275                  *
 276                  * Do a cond resched at the same frequency we would wake up
 277                  * the reader.
 278                  */
 279                 if (cnt % wakeup_interval)
 280                         cond_resched();
 281 #endif
 282         } while (ktime_before(end_time, timeout) && !break_test());
 283         trace_printk("End ring buffer hammer\n");
 284
 285         if (consumer) {
 286                 /* Init both completions here to avoid races */
 287                 init_completion(&read_start);
 288                 init_completion(&read_done);
 289                 /* the completions must be visible before the finish var */
 290                 smp_wmb();
 291                 reader_finish = 1;
 292                 wake_up_process(consumer);
 293                 wait_for_completion(&read_done);
 294         }
 295
 296         time = ktime_us_delta(end_time, start_time);
 297
 298         entries = ring_buffer_entries(buffer);
 299         overruns = ring_buffer_overruns(buffer);
 300
 301         if (test_error)
 302                 trace_printk("ERROR!\n");
 303
 304         if (!disable_reader) {
 305                 if (consumer_fifo < 0)
 306                         trace_printk("Running Consumer at nice: %d\n",
 307                                      consumer_nice);
 308                 else
 309                         trace_printk("Running Consumer at SCHED_FIFO %d\n",
 310                                      consumer_fifo);
 311         }
 312         if (producer_fifo < 0)
 313                 trace_printk("Running Producer at nice: %d\n",
 314                              producer_nice);
 315         else
 316                 trace_printk("Running Producer at SCHED_FIFO %d\n",
 317                              producer_fifo);
 318
 319         /* Let the user know that the test is running at low priority */
 320         if (producer_fifo < 0 && consumer_fifo < 0 &&
 321             producer_nice == MAX_NICE && consumer_nice == MAX_NICE)
 322                 trace_printk("WARNING!!! This test is running at lowest priority.\n");
 323
 324         trace_printk("Time:     %lld (usecs)\n", time);
 325         trace_printk("Overruns: %lld\n", overruns);
 326         if (disable_reader)
 327                 trace_printk("Read:     (reader disabled)\n");
 328         else
 329                 trace_printk("Read:     %ld  (by %s)\n", read,
 330                         read_events ? "events" : "pages");
 331         trace_printk("Entries:  %lld\n", entries);
 332         trace_printk("Total:    %lld\n", entries + overruns + read);
 333         trace_printk("Missed:   %ld\n", missed);
 334         trace_printk("Hit:      %ld\n", hit);
 335
 336         /* Convert time from usecs to millisecs */
 337         do_div(time, USEC_PER_MSEC);
 338         if (time)
 339                 hit /= (long)time;
 340         else
 341                 trace_printk("TIME IS ZERO??\n");
 342
 343         trace_printk("Entries per millisec: %ld\n", hit);
 344
 345         if (hit) {
 346                 /* Calculate the average time in nanosecs */
 347                 avg = NSEC_PER_MSEC / hit;
 348                 trace_printk("%ld ns per entry\n", avg);
 349         }
 350
 351         if (missed) {
 352                 if (time)
 353                         missed /= (long)time;
 354
 355                 trace_printk("Total iterations per millisec: %ld\n",
 356                              hit + missed);
 357
 358                 /* it is possible that hit + missed will overflow and be zero */
 359                 if (!(hit + missed)) {
 360                         trace_printk("hit + missed overflowed and totalled zero!\n");
 361                         hit--; /* make it non zero */
 362                 }
 363
 364                 /* Caculate the average time in nanosecs */
 365                 avg = NSEC_PER_MSEC / (hit + missed);
 366                 trace_printk("%ld ns per entry\n", avg);
 367         }
 368 }
 369
 370 static void wait_to_die(void)
 371 {
 372         set_current_state(TASK_INTERRUPTIBLE);
 373         while (!kthread_should_stop()) {
 374                 schedule();
 375                 set_current_state(TASK_INTERRUPTIBLE);
 376         }
 377         __set_current_state(TASK_RUNNING);
 378 }
 379
 380 static int ring_buffer_consumer_thread(void *arg)
 381 {
 382         while (!break_test()) {
 383                 complete(&read_start);
 384
 385                 ring_buffer_consumer();
 386
 387                 set_current_state(TASK_INTERRUPTIBLE);
 388                 if (break_test())
 389                         break;
 390                 schedule();
 391         }
 392         __set_current_state(TASK_RUNNING);
 393
 394         if (!kthread_should_stop())
 395                 wait_to_die();
 396
 397         return 0;
 398 }
 399
 400 static int ring_buffer_producer_thread(void *arg)
 401 {
 402         while (!break_test()) {
 403                 ring_buffer_reset(buffer);
 404
 405                 if (consumer) {
 406                         wake_up_process(consumer);
 407                         wait_for_completion(&read_start);
 408                 }
 409
 410                 ring_buffer_producer();
 411                 if (break_test())
 412                         goto out_kill;
 413
 414                 trace_printk("Sleeping for 10 secs\n");
 415                 set_current_state(TASK_INTERRUPTIBLE);
 416                 if (break_test())
 417                         goto out_kill;
 418                 schedule_timeout(HZ * SLEEP_TIME);
 419         }
 420
 421 out_kill:
 422         __set_current_state(TASK_RUNNING);
 423         if (!kthread_should_stop())
 424                 wait_to_die();
 425
 426         return 0;
 427 }
 428
 429 static int __init ring_buffer_benchmark_init(void)
 430 {
 431         int ret;
 432
 433         /* make a one meg buffer in overwite mode */
 434         buffer = ring_buffer_alloc(1000000, RB_FL_OVERWRITE);
 435         if (!buffer)
 436                 return -ENOMEM;
 437
 438         if (!disable_reader) {
 439                 consumer = kthread_create(ring_buffer_consumer_thread,
 440                                           NULL, "rb_consumer");
 441                 ret = PTR_ERR(consumer);
 442                 if (IS_ERR(consumer))
 443                         goto out_fail;
 444         }
 445
 446         producer = kthread_run(ring_buffer_producer_thread,
 447                                NULL, "rb_producer");
 448         ret = PTR_ERR(producer);
 449
 450         if (IS_ERR(producer))
 451                 goto out_kill;
 452
 453         /*
 454          * Run them as low-prio background tasks by default:
 455          */
 456         if (!disable_reader) {
 457                 if (consumer_fifo >= 0) {
 458                         struct sched_param param = {
 459                                 .sched_priority = consumer_fifo
 460                         };
 461                         sched_setscheduler(consumer, SCHED_FIFO, &param);
 462                 } else
 463                         set_user_nice(consumer, consumer_nice);
 464         }
 465
 466         if (producer_fifo >= 0) {
 467                 struct sched_param param = {
 468                         .sched_priority = producer_fifo
 469                 };
 470                 sched_setscheduler(producer, SCHED_FIFO, &param);
 471         } else
 472                 set_user_nice(producer, producer_nice);
 473
 474         return 0;
 475
 476  out_kill:
 477         if (consumer)
 478                 kthread_stop(consumer);
 479
 480  out_fail:
 481         ring_buffer_free(buffer);
 482         return ret;
 483 }
 484
 485 static void __exit ring_buffer_benchmark_exit(void)
 486 {
 487         kthread_stop(producer);
 488         if (consumer)
 489                 kthread_stop(consumer);
 490         ring_buffer_free(buffer);
 491 }
 492
 493 module_init(ring_buffer_benchmark_init);
 494 module_exit(ring_buffer_benchmark_exit);
 495
 496 MODULE_AUTHOR("Steven Rostedt");
 497 MODULE_DESCRIPTION("ring_buffer_benchmark");
 498 MODULE_LICENSE("GPL");