kernel/trace/ring_buffer_benchmark.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * ring buffer tester and benchmark
   4  *
   5  * Copyright (C) 2009 Steven Rostedt <srostedt@redhat.com>
   6  */
   7 #include <linux/ring_buffer.h>
   8 #include <linux/completion.h>
   9 #include <linux/kthread.h>
  10 #include <uapi/linux/sched/types.h>
  11 #include <linux/module.h>
  12 #include <linux/ktime.h>
  13 #include <asm/local.h>
  14
  15 struct rb_page {
  16         u64             ts;
  17         local_t         commit;
  18         char            data[4080];
  19 };
  20
  21 /* run time and sleep time in seconds */
  22 #define RUN_TIME        10ULL
  23 #define SLEEP_TIME      10
  24
  25 /* number of events for writer to wake up the reader */
  26 static int wakeup_interval = 100;
  27
  28 static int reader_finish;
  29 static DECLARE_COMPLETION(read_start);
  30 static DECLARE_COMPLETION(read_done);
  31
  32 static struct trace_buffer *buffer;
  33 static struct task_struct *producer;
  34 static struct task_struct *consumer;
  35 static unsigned long read;
  36
  37 static unsigned int disable_reader;
  38 module_param(disable_reader, uint, 0644);
  39 MODULE_PARM_DESC(disable_reader, "only run producer");
  40
  41 static unsigned int write_iteration = 50;
  42 module_param(write_iteration, uint, 0644);
  43 MODULE_PARM_DESC(write_iteration, "# of writes between timestamp readings");
  44
  45 static int producer_nice = MAX_NICE;
  46 static int consumer_nice = MAX_NICE;
  47
  48 static int producer_fifo;
  49 static int consumer_fifo;
  50
  51 module_param(producer_nice, int, 0644);
  52 MODULE_PARM_DESC(producer_nice, "nice prio for producer");
  53
  54 module_param(consumer_nice, int, 0644);
  55 MODULE_PARM_DESC(consumer_nice, "nice prio for consumer");
  56
  57 module_param(producer_fifo, int, 0644);
  58 MODULE_PARM_DESC(producer_fifo, "use fifo for producer: 0 - disabled, 1 - low prio, 2 - fifo");
  59
  60 module_param(consumer_fifo, int, 0644);
  61 MODULE_PARM_DESC(consumer_fifo, "use fifo for consumer: 0 - disabled, 1 - low prio, 2 - fifo");
  62
  63 static int read_events;
  64
  65 static int test_error;
  66
  67 #define TEST_ERROR()                            \
  68         do {                                    \
  69                 if (!test_error) {              \
  70                         test_error = 1;         \
  71                         WARN_ON(1);             \
  72                 }                               \
  73         } while (0)
  74
  75 enum event_status {
  76         EVENT_FOUND,
  77         EVENT_DROPPED,
  78 };
  79
  80 static bool break_test(void)
  81 {
  82         return test_error || kthread_should_stop();
  83 }
  84
  85 static enum event_status read_event(int cpu)
  86 {
  87         struct ring_buffer_event *event;
  88         int *entry;
  89         u64 ts;
  90
  91         event = ring_buffer_consume(buffer, cpu, &ts, NULL);
  92         if (!event)
  93                 return EVENT_DROPPED;
  94
  95         entry = ring_buffer_event_data(event);
  96         if (*entry != cpu) {
  97                 TEST_ERROR();
  98                 return EVENT_DROPPED;
  99         }
 100
 101         read++;
 102         return EVENT_FOUND;
 103 }
 104
 105 static enum event_status read_page(int cpu)
 106 {
 107         struct ring_buffer_event *event;
 108         struct rb_page *rpage;
 109         unsigned long commit;
 110         void *bpage;
 111         int *entry;
 112         int ret;
 113         int inc;
 114         int i;
 115
 116         bpage = ring_buffer_alloc_read_page(buffer, cpu);
 117         if (IS_ERR(bpage))
 118                 return EVENT_DROPPED;
 119
 120         ret = ring_buffer_read_page(buffer, &bpage, PAGE_SIZE, cpu, 1);
 121         if (ret >= 0) {
 122                 rpage = bpage;
 123                 /* The commit may have missed event flags set, clear them */
 124                 commit = local_read(&rpage->commit) & 0xfffff;
 125                 for (i = 0; i < commit && !test_error ; i += inc) {
 126
 127                         if (i >= (PAGE_SIZE - offsetof(struct rb_page, data))) {
 128                                 TEST_ERROR();
 129                                 break;
 130                         }
 131
 132                         inc = -1;
 133                         event = (void *)&rpage->data[i];
 134                         switch (event->type_len) {
 135                         case RINGBUF_TYPE_PADDING:
 136                                 /* failed writes may be discarded events */
 137                                 if (!event->time_delta)
 138                                         TEST_ERROR();
 139                                 inc = event->array[0] + 4;
 140                                 break;
 141                         case RINGBUF_TYPE_TIME_EXTEND:
 142                                 inc = 8;
 143                                 break;
 144                         case 0:
 145                                 entry = ring_buffer_event_data(event);
 146                                 if (*entry != cpu) {
 147                                         TEST_ERROR();
 148                                         break;
 149                                 }
 150                                 read++;
 151                                 if (!event->array[0]) {
 152                                         TEST_ERROR();
 153                                         break;
 154                                 }
 155                                 inc = event->array[0] + 4;
 156                                 break;
 157                         default:
 158                                 entry = ring_buffer_event_data(event);
 159                                 if (*entry != cpu) {
 160                                         TEST_ERROR();
 161                                         break;
 162                                 }
 163                                 read++;
 164                                 inc = ((event->type_len + 1) * 4);
 165                         }
 166                         if (test_error)
 167                                 break;
 168
 169                         if (inc <= 0) {
 170                                 TEST_ERROR();
 171                                 break;
 172                         }
 173                 }
 174         }
 175         ring_buffer_free_read_page(buffer, cpu, bpage);
 176
 177         if (ret < 0)
 178                 return EVENT_DROPPED;
 179         return EVENT_FOUND;
 180 }
 181
 182 static void ring_buffer_consumer(void)
 183 {
 184         /* toggle between reading pages and events */
 185         read_events ^= 1;
 186
 187         read = 0;
 188         /*
 189          * Continue running until the producer specifically asks to stop
 190          * and is ready for the completion.
 191          */
 192         while (!READ_ONCE(reader_finish)) {
 193                 int found = 1;
 194
 195                 while (found && !test_error) {
 196                         int cpu;
 197
 198                         found = 0;
 199                         for_each_online_cpu(cpu) {
 200                                 enum event_status stat;
 201
 202                                 if (read_events)
 203                                         stat = read_event(cpu);
 204                                 else
 205                                         stat = read_page(cpu);
 206
 207                                 if (test_error)
 208                                         break;
 209
 210                                 if (stat == EVENT_FOUND)
 211                                         found = 1;
 212
 213                         }
 214                 }
 215
 216                 /* Wait till the producer wakes us up when there is more data
 217                  * available or when the producer wants us to finish reading.
 218                  */
 219                 set_current_state(TASK_INTERRUPTIBLE);
 220                 if (reader_finish)
 221                         break;
 222
 223                 schedule();
 224         }
 225         __set_current_state(TASK_RUNNING);
 226         reader_finish = 0;
 227         complete(&read_done);
 228 }
 229
 230 static void ring_buffer_producer(void)
 231 {
 232         ktime_t start_time, end_time, timeout;
 233         unsigned long long time;
 234         unsigned long long entries;
 235         unsigned long long overruns;
 236         unsigned long missed = 0;
 237         unsigned long hit = 0;
 238         unsigned long avg;
 239         int cnt = 0;
 240
 241         /*
 242          * Hammer the buffer for 10 secs (this may
 243          * make the system stall)
 244          */
 245         trace_printk("Starting ring buffer hammer\n");
 246         start_time = ktime_get();
 247         timeout = ktime_add_ns(start_time, RUN_TIME * NSEC_PER_SEC);
 248         do {
 249                 struct ring_buffer_event *event;
 250                 int *entry;
 251                 int i;
 252
 253                 for (i = 0; i < write_iteration; i++) {
 254                         event = ring_buffer_lock_reserve(buffer, 10);
 255                         if (!event) {
 256                                 missed++;
 257                         } else {
 258                                 hit++;
 259                                 entry = ring_buffer_event_data(event);
 260                                 *entry = smp_processor_id();
 261                                 ring_buffer_unlock_commit(buffer, event);
 262                         }
 263                 }
 264                 end_time = ktime_get();
 265
 266                 cnt++;
 267                 if (consumer && !(cnt % wakeup_interval))
 268                         wake_up_process(consumer);
 269
 270 #ifndef CONFIG_PREEMPTION
 271                 /*
 272                  * If we are a non preempt kernel, the 10 seconds run will
 273                  * stop everything while it runs. Instead, we will call
 274                  * cond_resched and also add any time that was lost by a
 275                  * reschedule.
 276                  *
 277                  * Do a cond resched at the same frequency we would wake up
 278                  * the reader.
 279                  */
 280                 if (cnt % wakeup_interval)
 281                         cond_resched();
 282 #endif
 283         } while (ktime_before(end_time, timeout) && !break_test());
 284         trace_printk("End ring buffer hammer\n");
 285
 286         if (consumer) {
 287                 /* Init both completions here to avoid races */
 288                 init_completion(&read_start);
 289                 init_completion(&read_done);
 290                 /* the completions must be visible before the finish var */
 291                 smp_wmb();
 292                 reader_finish = 1;
 293                 wake_up_process(consumer);
 294                 wait_for_completion(&read_done);
 295         }
 296
 297         time = ktime_us_delta(end_time, start_time);
 298
 299         entries = ring_buffer_entries(buffer);
 300         overruns = ring_buffer_overruns(buffer);
 301
 302         if (test_error)
 303                 trace_printk("ERROR!\n");
 304
 305         if (!disable_reader) {
 306                 if (consumer_fifo)
 307                         trace_printk("Running Consumer at SCHED_FIFO %s\n",
 308                                      consumer_fifo == 1 ? "low" : "high");
 309                 else
 310                         trace_printk("Running Consumer at nice: %d\n",
 311                                      consumer_nice);
 312         }
 313         if (producer_fifo)
 314                 trace_printk("Running Producer at SCHED_FIFO %s\n",
 315                              producer_fifo == 1 ? "low" : "high");
 316         else
 317                 trace_printk("Running Producer at nice: %d\n",
 318                              producer_nice);
 319
 320         /* Let the user know that the test is running at low priority */
 321         if (!producer_fifo && !consumer_fifo &&
 322             producer_nice == MAX_NICE && consumer_nice == MAX_NICE)
 323                 trace_printk("WARNING!!! This test is running at lowest priority.\n");
 324
 325         trace_printk("Time:     %lld (usecs)\n", time);
 326         trace_printk("Overruns: %lld\n", overruns);
 327         if (disable_reader)
 328                 trace_printk("Read:     (reader disabled)\n");
 329         else
 330                 trace_printk("Read:     %ld  (by %s)\n", read,
 331                         read_events ? "events" : "pages");
 332         trace_printk("Entries:  %lld\n", entries);
 333         trace_printk("Total:    %lld\n", entries + overruns + read);
 334         trace_printk("Missed:   %ld\n", missed);
 335         trace_printk("Hit:      %ld\n", hit);
 336
 337         /* Convert time from usecs to millisecs */
 338         do_div(time, USEC_PER_MSEC);
 339         if (time)
 340                 hit /= (long)time;
 341         else
 342                 trace_printk("TIME IS ZERO??\n");
 343
 344         trace_printk("Entries per millisec: %ld\n", hit);
 345
 346         if (hit) {
 347                 /* Calculate the average time in nanosecs */
 348                 avg = NSEC_PER_MSEC / hit;
 349                 trace_printk("%ld ns per entry\n", avg);
 350         }
 351
 352         if (missed) {
 353                 if (time)
 354                         missed /= (long)time;
 355
 356                 trace_printk("Total iterations per millisec: %ld\n",
 357                              hit + missed);
 358
 359                 /* it is possible that hit + missed will overflow and be zero */
 360                 if (!(hit + missed)) {
 361                         trace_printk("hit + missed overflowed and totalled zero!\n");
 362                         hit--; /* make it non zero */
 363                 }
 364
 365                 /* Calculate the average time in nanosecs */
 366                 avg = NSEC_PER_MSEC / (hit + missed);
 367                 trace_printk("%ld ns per entry\n", avg);
 368         }
 369 }
 370
 371 static void wait_to_die(void)
 372 {
 373         set_current_state(TASK_INTERRUPTIBLE);
 374         while (!kthread_should_stop()) {
 375                 schedule();
 376                 set_current_state(TASK_INTERRUPTIBLE);
 377         }
 378         __set_current_state(TASK_RUNNING);
 379 }
 380
 381 static int ring_buffer_consumer_thread(void *arg)
 382 {
 383         while (!break_test()) {
 384                 complete(&read_start);
 385
 386                 ring_buffer_consumer();
 387
 388                 set_current_state(TASK_INTERRUPTIBLE);
 389                 if (break_test())
 390                         break;
 391                 schedule();
 392         }
 393         __set_current_state(TASK_RUNNING);
 394
 395         if (!kthread_should_stop())
 396                 wait_to_die();
 397
 398         return 0;
 399 }
 400
 401 static int ring_buffer_producer_thread(void *arg)
 402 {
 403         while (!break_test()) {
 404                 ring_buffer_reset(buffer);
 405
 406                 if (consumer) {
 407                         wake_up_process(consumer);
 408                         wait_for_completion(&read_start);
 409                 }
 410
 411                 ring_buffer_producer();
 412                 if (break_test())
 413                         goto out_kill;
 414
 415                 trace_printk("Sleeping for 10 secs\n");
 416                 set_current_state(TASK_INTERRUPTIBLE);
 417                 if (break_test())
 418                         goto out_kill;
 419                 schedule_timeout(HZ * SLEEP_TIME);
 420         }
 421
 422 out_kill:
 423         __set_current_state(TASK_RUNNING);
 424         if (!kthread_should_stop())
 425                 wait_to_die();
 426
 427         return 0;
 428 }
 429
 430 static int __init ring_buffer_benchmark_init(void)
 431 {
 432         int ret;
 433
 434         /* make a one meg buffer in overwite mode */
 435         buffer = ring_buffer_alloc(1000000, RB_FL_OVERWRITE);
 436         if (!buffer)
 437                 return -ENOMEM;
 438
 439         if (!disable_reader) {
 440                 consumer = kthread_create(ring_buffer_consumer_thread,
 441                                           NULL, "rb_consumer");
 442                 ret = PTR_ERR(consumer);
 443                 if (IS_ERR(consumer))
 444                         goto out_fail;
 445         }
 446
 447         producer = kthread_run(ring_buffer_producer_thread,
 448                                NULL, "rb_producer");
 449         ret = PTR_ERR(producer);
 450
 451         if (IS_ERR(producer))
 452                 goto out_kill;
 453
 454         /*
 455          * Run them as low-prio background tasks by default:
 456          */
 457         if (!disable_reader) {
 458                 if (consumer_fifo >= 2)
 459                         sched_set_fifo(consumer);
 460                 else if (consumer_fifo == 1)
 461                         sched_set_fifo_low(consumer);
 462                 else
 463                         set_user_nice(consumer, consumer_nice);
 464         }
 465
 466         if (producer_fifo >= 2)
 467                 sched_set_fifo(producer);
 468         else if (producer_fifo == 1)
 469                 sched_set_fifo_low(producer);
 470         else
 471                 set_user_nice(producer, producer_nice);
 472
 473         return 0;
 474
 475  out_kill:
 476         if (consumer)
 477                 kthread_stop(consumer);
 478
 479  out_fail:
 480         ring_buffer_free(buffer);
 481         return ret;
 482 }
 483
 484 static void __exit ring_buffer_benchmark_exit(void)
 485 {
 486         kthread_stop(producer);
 487         if (consumer)
 488                 kthread_stop(consumer);
 489         ring_buffer_free(buffer);
 490 }
 491
 492 module_init(ring_buffer_benchmark_init);
 493 module_exit(ring_buffer_benchmark_exit);
 494
 495 MODULE_AUTHOR("Steven Rostedt");
 496 MODULE_DESCRIPTION("ring_buffer_benchmark");
 497 MODULE_LICENSE("GPL");