kernel/trace/ring_buffer_benchmark.c

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