linux/kernel/trace/ring_buffer_benchmark.c

420 lines
8.5 KiB
C
Raw Normal View History

ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
/*
* ring buffer tester and benchmark
*
* Copyright (C) 2009 Steven Rostedt <srostedt@redhat.com>
*/
#include <linux/ring_buffer.h>
#include <linux/completion.h>
#include <linux/kthread.h>
#include <linux/module.h>
#include <linux/time.h>
struct rb_page {
u64 ts;
local_t commit;
char data[4080];
};
/* run time and sleep time in seconds */
#define RUN_TIME 10
#define SLEEP_TIME 10
/* number of events for writer to wake up the reader */
static int wakeup_interval = 100;
static int reader_finish;
static struct completion read_start;
static struct completion read_done;
static struct ring_buffer *buffer;
static struct task_struct *producer;
static struct task_struct *consumer;
static unsigned long read;
static int disable_reader;
module_param(disable_reader, uint, 0644);
MODULE_PARM_DESC(disable_reader, "only run producer");
static int read_events;
static int kill_test;
#define KILL_TEST() \
do { \
if (!kill_test) { \
kill_test = 1; \
WARN_ON(1); \
} \
} while (0)
enum event_status {
EVENT_FOUND,
EVENT_DROPPED,
};
static enum event_status read_event(int cpu)
{
struct ring_buffer_event *event;
int *entry;
u64 ts;
event = ring_buffer_consume(buffer, cpu, &ts);
if (!event)
return EVENT_DROPPED;
entry = ring_buffer_event_data(event);
if (*entry != cpu) {
KILL_TEST();
return EVENT_DROPPED;
}
read++;
return EVENT_FOUND;
}
static enum event_status read_page(int cpu)
{
struct ring_buffer_event *event;
struct rb_page *rpage;
unsigned long commit;
void *bpage;
int *entry;
int ret;
int inc;
int i;
bpage = ring_buffer_alloc_read_page(buffer);
if (!bpage)
return EVENT_DROPPED;
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
ret = ring_buffer_read_page(buffer, &bpage, PAGE_SIZE, cpu, 1);
if (ret >= 0) {
rpage = bpage;
commit = local_read(&rpage->commit);
for (i = 0; i < commit && !kill_test; i += inc) {
if (i >= (PAGE_SIZE - offsetof(struct rb_page, data))) {
KILL_TEST();
break;
}
inc = -1;
event = (void *)&rpage->data[i];
switch (event->type_len) {
case RINGBUF_TYPE_PADDING:
/* failed writes may be discarded events */
if (!event->time_delta)
KILL_TEST();
inc = event->array[0] + 4;
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
break;
case RINGBUF_TYPE_TIME_EXTEND:
inc = 8;
break;
case 0:
entry = ring_buffer_event_data(event);
if (*entry != cpu) {
KILL_TEST();
break;
}
read++;
if (!event->array[0]) {
KILL_TEST();
break;
}
inc = event->array[0] + 4;
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
break;
default:
entry = ring_buffer_event_data(event);
if (*entry != cpu) {
KILL_TEST();
break;
}
read++;
inc = ((event->type_len + 1) * 4);
}
if (kill_test)
break;
if (inc <= 0) {
KILL_TEST();
break;
}
}
}
ring_buffer_free_read_page(buffer, bpage);
if (ret < 0)
return EVENT_DROPPED;
return EVENT_FOUND;
}
static void ring_buffer_consumer(void)
{
/* toggle between reading pages and events */
read_events ^= 1;
read = 0;
while (!reader_finish && !kill_test) {
int found;
do {
int cpu;
found = 0;
for_each_online_cpu(cpu) {
enum event_status stat;
if (read_events)
stat = read_event(cpu);
else
stat = read_page(cpu);
if (kill_test)
break;
if (stat == EVENT_FOUND)
found = 1;
}
} while (found && !kill_test);
set_current_state(TASK_INTERRUPTIBLE);
if (reader_finish)
break;
schedule();
__set_current_state(TASK_RUNNING);
}
reader_finish = 0;
complete(&read_done);
}
static void ring_buffer_producer(void)
{
struct timeval start_tv;
struct timeval end_tv;
unsigned long long time;
unsigned long long entries;
unsigned long long overruns;
unsigned long missed = 0;
unsigned long hit = 0;
unsigned long avg;
int cnt = 0;
/*
* Hammer the buffer for 10 secs (this may
* make the system stall)
*/
trace_printk("Starting ring buffer hammer\n");
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
do_gettimeofday(&start_tv);
do {
struct ring_buffer_event *event;
int *entry;
event = ring_buffer_lock_reserve(buffer, 10);
if (!event) {
missed++;
} else {
hit++;
entry = ring_buffer_event_data(event);
*entry = smp_processor_id();
ring_buffer_unlock_commit(buffer, event);
}
do_gettimeofday(&end_tv);
cnt++;
if (consumer && !(cnt % wakeup_interval))
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
wake_up_process(consumer);
#ifndef CONFIG_PREEMPT
/*
* If we are a non preempt kernel, the 10 second run will
* stop everything while it runs. Instead, we will call
* cond_resched and also add any time that was lost by a
* rescedule.
*
* Do a cond resched at the same frequency we would wake up
* the reader.
*/
if (cnt % wakeup_interval)
cond_resched();
#endif
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
} while (end_tv.tv_sec < (start_tv.tv_sec + RUN_TIME) && !kill_test);
trace_printk("End ring buffer hammer\n");
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
if (consumer) {
/* Init both completions here to avoid races */
init_completion(&read_start);
init_completion(&read_done);
/* the completions must be visible before the finish var */
smp_wmb();
reader_finish = 1;
/* finish var visible before waking up the consumer */
smp_wmb();
wake_up_process(consumer);
wait_for_completion(&read_done);
}
time = end_tv.tv_sec - start_tv.tv_sec;
time *= USEC_PER_SEC;
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
time += (long long)((long)end_tv.tv_usec - (long)start_tv.tv_usec);
entries = ring_buffer_entries(buffer);
overruns = ring_buffer_overruns(buffer);
if (kill_test)
trace_printk("ERROR!\n");
trace_printk("Time: %lld (usecs)\n", time);
trace_printk("Overruns: %lld\n", overruns);
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
if (disable_reader)
trace_printk("Read: (reader disabled)\n");
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
else
trace_printk("Read: %ld (by %s)\n", read,
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
read_events ? "events" : "pages");
trace_printk("Entries: %lld\n", entries);
trace_printk("Total: %lld\n", entries + overruns + read);
trace_printk("Missed: %ld\n", missed);
trace_printk("Hit: %ld\n", hit);
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
/* Convert time from usecs to millisecs */
do_div(time, USEC_PER_MSEC);
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
if (time)
hit /= (long)time;
else
trace_printk("TIME IS ZERO??\n");
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
trace_printk("Entries per millisec: %ld\n", hit);
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
if (hit) {
/* Calculate the average time in nanosecs */
avg = NSEC_PER_MSEC / hit;
trace_printk("%ld ns per entry\n", avg);
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
}
if (missed) {
if (time)
missed /= (long)time;
trace_printk("Total iterations per millisec: %ld\n",
hit + missed);
/* it is possible that hit + missed will overflow and be zero */
if (!(hit + missed)) {
trace_printk("hit + missed overflowed and totalled zero!\n");
hit--; /* make it non zero */
}
/* Caculate the average time in nanosecs */
avg = NSEC_PER_MSEC / (hit + missed);
trace_printk("%ld ns per entry\n", avg);
}
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
}
static void wait_to_die(void)
{
set_current_state(TASK_INTERRUPTIBLE);
while (!kthread_should_stop()) {
schedule();
set_current_state(TASK_INTERRUPTIBLE);
}
__set_current_state(TASK_RUNNING);
}
static int ring_buffer_consumer_thread(void *arg)
{
while (!kthread_should_stop() && !kill_test) {
complete(&read_start);
ring_buffer_consumer();
set_current_state(TASK_INTERRUPTIBLE);
if (kthread_should_stop() || kill_test)
break;
schedule();
__set_current_state(TASK_RUNNING);
}
__set_current_state(TASK_RUNNING);
if (kill_test)
wait_to_die();
return 0;
}
static int ring_buffer_producer_thread(void *arg)
{
init_completion(&read_start);
while (!kthread_should_stop() && !kill_test) {
ring_buffer_reset(buffer);
if (consumer) {
smp_wmb();
wake_up_process(consumer);
wait_for_completion(&read_start);
}
ring_buffer_producer();
trace_printk("Sleeping for 10 secs\n");
ring-buffer: add benchmark and tester This patch adds code that can benchmark the ring buffer as well as test it. This code can be compiled into the kernel (not recommended) or as a module. A separate ring buffer is used to not interfer with other users, like ftrace. It creates a producer and a consumer (option to disable creation of the consumer) and will run for 10 seconds, then sleep for 10 seconds and then repeat. While running, the producer will write 10 byte loads into the ring buffer with just putting in the current CPU number. The reader will continually try to read the buffer. The reader will alternate from reading the buffer via event by event, or by full pages. The output is a pr_info, thus it will fill up the syslogs. Starting ring buffer hammer End ring buffer hammer Time: 9000349 (usecs) Overruns: 12578640 Read: 5358440 (by events) Entries: 0 Total: 17937080 Missed: 0 Hit: 17937080 Entries per millisec: 1993 501 ns per entry Sleeping for 10 secs Starting ring buffer hammer End ring buffer hammer Time: 9936350 (usecs) Overruns: 0 Read: 28146644 (by pages) Entries: 74 Total: 28146718 Missed: 0 Hit: 28146718 Entries per millisec: 2832 353 ns per entry Sleeping for 10 secs Time: is the time the test ran Overruns: the number of events that were overwritten and not read Read: the number of events read (either by pages or events) Entries: the number of entries left in the buffer (the by pages will only read full pages) Total: Entries + Read + Overruns Missed: the number of entries that failed to write Hit: the number of entries that were written The above example shows that it takes ~353 nanosecs per entry when there is a reader, reading by pages (and no overruns) The event by event reader slowed the producer down to 501 nanosecs. [ Impact: see how changes to the ring buffer affect stability and performance ] Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
2009-05-06 02:47:18 +00:00
set_current_state(TASK_INTERRUPTIBLE);
schedule_timeout(HZ * SLEEP_TIME);
__set_current_state(TASK_RUNNING);
}
if (kill_test)
wait_to_die();
return 0;
}
static int __init ring_buffer_benchmark_init(void)
{
int ret;
/* make a one meg buffer in overwite mode */
buffer = ring_buffer_alloc(1000000, RB_FL_OVERWRITE);
if (!buffer)
return -ENOMEM;
if (!disable_reader) {
consumer = kthread_create(ring_buffer_consumer_thread,
NULL, "rb_consumer");
ret = PTR_ERR(consumer);
if (IS_ERR(consumer))
goto out_fail;
}
producer = kthread_run(ring_buffer_producer_thread,
NULL, "rb_producer");
ret = PTR_ERR(producer);
if (IS_ERR(producer))
goto out_kill;
return 0;
out_kill:
if (consumer)
kthread_stop(consumer);
out_fail:
ring_buffer_free(buffer);
return ret;
}
static void __exit ring_buffer_benchmark_exit(void)
{
kthread_stop(producer);
if (consumer)
kthread_stop(consumer);
ring_buffer_free(buffer);
}
module_init(ring_buffer_benchmark_init);
module_exit(ring_buffer_benchmark_exit);
MODULE_AUTHOR("Steven Rostedt");
MODULE_DESCRIPTION("ring_buffer_benchmark");
MODULE_LICENSE("GPL");