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/time.h>
11#include <asm/local.h>
12
13struct 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	10
21#define SLEEP_TIME	10
22
23/* number of events for writer to wake up the reader */
24static int wakeup_interval = 100;
25
26static int reader_finish;
27static struct completion read_start;
28static struct completion read_done;
29
30static struct ring_buffer *buffer;
31static struct task_struct *producer;
32static struct task_struct *consumer;
33static unsigned long read;
34
35static int disable_reader;
36module_param(disable_reader, uint, 0644);
37MODULE_PARM_DESC(disable_reader, "only run producer");
38
39static int write_iteration = 50;
40module_param(write_iteration, uint, 0644);
41MODULE_PARM_DESC(write_iteration, "# of writes between timestamp readings");
42
43static int producer_nice = MAX_NICE;
44static int consumer_nice = MAX_NICE;
45
46static int producer_fifo = -1;
47static int consumer_fifo = -1;
48
49module_param(producer_nice, uint, 0644);
50MODULE_PARM_DESC(producer_nice, "nice prio for producer");
51
52module_param(consumer_nice, uint, 0644);
53MODULE_PARM_DESC(consumer_nice, "nice prio for consumer");
54
55module_param(producer_fifo, uint, 0644);
56MODULE_PARM_DESC(producer_fifo, "fifo prio for producer");
57
58module_param(consumer_fifo, uint, 0644);
59MODULE_PARM_DESC(consumer_fifo, "fifo prio for consumer");
60
61static int read_events;
62
63static 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
73enum event_status {
74	EVENT_FOUND,
75	EVENT_DROPPED,
76};
77
78static 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
98static 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
175static void ring_buffer_consumer(void)
176{
177	/* toggle between reading pages and events */
178	read_events ^= 1;
179
180	read = 0;
181	while (!reader_finish && !kill_test) {
182		int found;
183
184		do {
185			int cpu;
186
187			found = 0;
188			for_each_online_cpu(cpu) {
189				enum event_status stat;
190
191				if (read_events)
192					stat = read_event(cpu);
193				else
194					stat = read_page(cpu);
195
196				if (kill_test)
197					break;
198				if (stat == EVENT_FOUND)
199					found = 1;
200			}
201		} while (found && !kill_test);
202
203		set_current_state(TASK_INTERRUPTIBLE);
204		if (reader_finish)
205			break;
206
207		schedule();
208	}
209	reader_finish = 0;
210	complete(&read_done);
211}
212
213static void ring_buffer_producer(void)
214{
215	struct timeval start_tv;
216	struct timeval end_tv;
217	unsigned long long time;
218	unsigned long long entries;
219	unsigned long long overruns;
220	unsigned long missed = 0;
221	unsigned long hit = 0;
222	unsigned long avg;
223	int cnt = 0;
224
225	/*
226	 * Hammer the buffer for 10 secs (this may
227	 * make the system stall)
228	 */
229	trace_printk("Starting ring buffer hammer\n");
230	do_gettimeofday(&start_tv);
231	do {
232		struct ring_buffer_event *event;
233		int *entry;
234		int i;
235
236		for (i = 0; i < write_iteration; i++) {
237			event = ring_buffer_lock_reserve(buffer, 10);
238			if (!event) {
239				missed++;
240			} else {
241				hit++;
242				entry = ring_buffer_event_data(event);
243				*entry = smp_processor_id();
244				ring_buffer_unlock_commit(buffer, event);
245			}
246		}
247		do_gettimeofday(&end_tv);
248
249		cnt++;
250		if (consumer && !(cnt % wakeup_interval))
251			wake_up_process(consumer);
252
253#ifndef CONFIG_PREEMPT
254		/*
255		 * If we are a non preempt kernel, the 10 second run will
256		 * stop everything while it runs. Instead, we will call
257		 * cond_resched and also add any time that was lost by a
258		 * rescedule.
259		 *
260		 * Do a cond resched at the same frequency we would wake up
261		 * the reader.
262		 */
263		if (cnt % wakeup_interval)
264			cond_resched();
265#endif
266
267	} while (end_tv.tv_sec < (start_tv.tv_sec + RUN_TIME) && !kill_test);
268	trace_printk("End ring buffer hammer\n");
269
270	if (consumer) {
271		/* Init both completions here to avoid races */
272		init_completion(&read_start);
273		init_completion(&read_done);
274		/* the completions must be visible before the finish var */
275		smp_wmb();
276		reader_finish = 1;
277		/* finish var visible before waking up the consumer */
278		smp_wmb();
279		wake_up_process(consumer);
280		wait_for_completion(&read_done);
281	}
282
283	time = end_tv.tv_sec - start_tv.tv_sec;
284	time *= USEC_PER_SEC;
285	time += (long long)((long)end_tv.tv_usec - (long)start_tv.tv_usec);
286
287	entries = ring_buffer_entries(buffer);
288	overruns = ring_buffer_overruns(buffer);
289
290	if (kill_test)
291		trace_printk("ERROR!\n");
292
293	if (!disable_reader) {
294		if (consumer_fifo < 0)
295			trace_printk("Running Consumer at nice: %d\n",
296				     consumer_nice);
297		else
298			trace_printk("Running Consumer at SCHED_FIFO %d\n",
299				     consumer_fifo);
300	}
301	if (producer_fifo < 0)
302		trace_printk("Running Producer at nice: %d\n",
303			     producer_nice);
304	else
305		trace_printk("Running Producer at SCHED_FIFO %d\n",
306			     producer_fifo);
307
308	/* Let the user know that the test is running at low priority */
309	if (producer_fifo < 0 && consumer_fifo < 0 &&
310	    producer_nice == MAX_NICE && consumer_nice == MAX_NICE)
311		trace_printk("WARNING!!! This test is running at lowest priority.\n");
312
313	trace_printk("Time:     %lld (usecs)\n", time);
314	trace_printk("Overruns: %lld\n", overruns);
315	if (disable_reader)
316		trace_printk("Read:     (reader disabled)\n");
317	else
318		trace_printk("Read:     %ld  (by %s)\n", read,
319			read_events ? "events" : "pages");
320	trace_printk("Entries:  %lld\n", entries);
321	trace_printk("Total:    %lld\n", entries + overruns + read);
322	trace_printk("Missed:   %ld\n", missed);
323	trace_printk("Hit:      %ld\n", hit);
324
325	/* Convert time from usecs to millisecs */
326	do_div(time, USEC_PER_MSEC);
327	if (time)
328		hit /= (long)time;
329	else
330		trace_printk("TIME IS ZERO??\n");
331
332	trace_printk("Entries per millisec: %ld\n", hit);
333
334	if (hit) {
335		/* Calculate the average time in nanosecs */
336		avg = NSEC_PER_MSEC / hit;
337		trace_printk("%ld ns per entry\n", avg);
338	}
339
340	if (missed) {
341		if (time)
342			missed /= (long)time;
343
344		trace_printk("Total iterations per millisec: %ld\n",
345			     hit + missed);
346
347		/* it is possible that hit + missed will overflow and be zero */
348		if (!(hit + missed)) {
349			trace_printk("hit + missed overflowed and totalled zero!\n");
350			hit--; /* make it non zero */
351		}
352
353		/* Caculate the average time in nanosecs */
354		avg = NSEC_PER_MSEC / (hit + missed);
355		trace_printk("%ld ns per entry\n", avg);
356	}
357}
358
359static void wait_to_die(void)
360{
361	set_current_state(TASK_INTERRUPTIBLE);
362	while (!kthread_should_stop()) {
363		schedule();
364		set_current_state(TASK_INTERRUPTIBLE);
365	}
366	__set_current_state(TASK_RUNNING);
367}
368
369static int ring_buffer_consumer_thread(void *arg)
370{
371	while (!kthread_should_stop() && !kill_test) {
372		complete(&read_start);
373
374		ring_buffer_consumer();
375
376		set_current_state(TASK_INTERRUPTIBLE);
377		if (kthread_should_stop() || kill_test)
378			break;
379
380		schedule();
381	}
382	__set_current_state(TASK_RUNNING);
383
384	if (kill_test)
385		wait_to_die();
386
387	return 0;
388}
389
390static int ring_buffer_producer_thread(void *arg)
391{
392	init_completion(&read_start);
393
394	while (!kthread_should_stop() && !kill_test) {
395		ring_buffer_reset(buffer);
396
397		if (consumer) {
398			smp_wmb();
399			wake_up_process(consumer);
400			wait_for_completion(&read_start);
401		}
402
403		ring_buffer_producer();
404
405		trace_printk("Sleeping for 10 secs\n");
406		set_current_state(TASK_INTERRUPTIBLE);
407		schedule_timeout(HZ * SLEEP_TIME);
408	}
409
410	if (kill_test)
411		wait_to_die();
412
413	return 0;
414}
415
416static int __init ring_buffer_benchmark_init(void)
417{
418	int ret;
419
420	/* make a one meg buffer in overwite mode */
421	buffer = ring_buffer_alloc(1000000, RB_FL_OVERWRITE);
422	if (!buffer)
423		return -ENOMEM;
424
425	if (!disable_reader) {
426		consumer = kthread_create(ring_buffer_consumer_thread,
427					  NULL, "rb_consumer");
428		ret = PTR_ERR(consumer);
429		if (IS_ERR(consumer))
430			goto out_fail;
431	}
432
433	producer = kthread_run(ring_buffer_producer_thread,
434			       NULL, "rb_producer");
435	ret = PTR_ERR(producer);
436
437	if (IS_ERR(producer))
438		goto out_kill;
439
440	/*
441	 * Run them as low-prio background tasks by default:
442	 */
443	if (!disable_reader) {
444		if (consumer_fifo >= 0) {
445			struct sched_param param = {
446				.sched_priority = consumer_fifo
447			};
448			sched_setscheduler(consumer, SCHED_FIFO, &param);
449		} else
450			set_user_nice(consumer, consumer_nice);
451	}
452
453	if (producer_fifo >= 0) {
454		struct sched_param param = {
455			.sched_priority = consumer_fifo
456		};
457		sched_setscheduler(producer, SCHED_FIFO, &param);
458	} else
459		set_user_nice(producer, producer_nice);
460
461	return 0;
462
463 out_kill:
464	if (consumer)
465		kthread_stop(consumer);
466
467 out_fail:
468	ring_buffer_free(buffer);
469	return ret;
470}
471
472static void __exit ring_buffer_benchmark_exit(void)
473{
474	kthread_stop(producer);
475	if (consumer)
476		kthread_stop(consumer);
477	ring_buffer_free(buffer);
478}
479
480module_init(ring_buffer_benchmark_init);
481module_exit(ring_buffer_benchmark_exit);
482
483MODULE_AUTHOR("Steven Rostedt");
484MODULE_DESCRIPTION("ring_buffer_benchmark");
485MODULE_LICENSE("GPL");
486