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 <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 (!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, ¶m); 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, ¶m); 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"); 499