tjh.dev
Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
selftests: Add test of PMU instruction counting on powerpc
This commit adds a test of instruction counting using the PMU on powerpc.
Although the bulk of the code is architecture agnostic, the code needs to
run a precisely sized loop which is implemented in assembler.
Signed-off-by: Michael Ellerman <michael@ellerman.id.au>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
authored by
Michael Ellerman
and committed by
Benjamin Herrenschmidt
cb96143d
2fae0d7c
+349
-1
+1
-1
tools/testing/selftests/powerpc/Makefile
+1
-1
tools/testing/selftests/powerpc/Makefile
+23
tools/testing/selftests/powerpc/pmu/Makefile
+23
tools/testing/selftests/powerpc/pmu/Makefile
···
1
+
noarg:
2
+
$(MAKE) -C ../
3
+
4
+
PROGS := count_instructions
5
+
EXTRA_SOURCES := ../harness.c event.c
6
+
7
+
all: $(PROGS)
8
+
9
+
$(PROGS): $(EXTRA_SOURCES)
10
+
11
+
# loop.S can only be built 64-bit
12
+
count_instructions: loop.S count_instructions.c $(EXTRA_SOURCES)
13
+
$(CC) $(CFLAGS) -m64 -o $@ $^
14
+
15
+
run_tests: all
16
+
@-for PROG in $(PROGS); do \
17
+
./$$PROG; \
18
+
done;
19
+
20
+
clean:
21
+
rm -f $(PROGS) loop.o
22
+
23
+
.PHONY: all run_tests clean
+135
tools/testing/selftests/powerpc/pmu/count_instructions.c
+135
tools/testing/selftests/powerpc/pmu/count_instructions.c
···
1
+
/*
2
+
* Copyright 2013, Michael Ellerman, IBM Corp.
3
+
* Licensed under GPLv2.
4
+
*/
5
+
6
+
#define _GNU_SOURCE
7
+
8
+
#include <stdio.h>
9
+
#include <stdbool.h>
10
+
#include <string.h>
11
+
#include <sys/prctl.h>
12
+
13
+
#include "event.h"
14
+
#include "utils.h"
15
+
16
+
extern void thirty_two_instruction_loop(u64 loops);
17
+
18
+
static void setup_event(struct event *e, u64 config, char *name)
19
+
{
20
+
event_init_opts(e, config, PERF_TYPE_HARDWARE, name);
21
+
22
+
e->attr.disabled = 1;
23
+
e->attr.exclude_kernel = 1;
24
+
e->attr.exclude_hv = 1;
25
+
e->attr.exclude_idle = 1;
26
+
}
27
+
28
+
static int do_count_loop(struct event *events, u64 instructions,
29
+
u64 overhead, bool report)
30
+
{
31
+
s64 difference, expected;
32
+
double percentage;
33
+
34
+
prctl(PR_TASK_PERF_EVENTS_ENABLE);
35
+
36
+
/* Run for 1M instructions */
37
+
thirty_two_instruction_loop(instructions >> 5);
38
+
39
+
prctl(PR_TASK_PERF_EVENTS_DISABLE);
40
+
41
+
event_read(&events[0]);
42
+
event_read(&events[1]);
43
+
44
+
expected = instructions + overhead;
45
+
difference = events[0].result.value - expected;
46
+
percentage = (double)difference / events[0].result.value * 100;
47
+
48
+
if (report) {
49
+
event_report(&events[0]);
50
+
event_report(&events[1]);
51
+
52
+
printf("Looped for %llu instructions, overhead %llu\n", instructions, overhead);
53
+
printf("Expected %llu\n", expected);
54
+
printf("Actual %llu\n", events[0].result.value);
55
+
printf("Delta %lld, %f%%\n", difference, percentage);
56
+
}
57
+
58
+
event_reset(&events[0]);
59
+
event_reset(&events[1]);
60
+
61
+
if (difference < 0)
62
+
difference = -difference;
63
+
64
+
/* Tolerate a difference below 0.0001 % */
65
+
difference *= 10000 * 100;
66
+
if (difference / events[0].result.value)
67
+
return -1;
68
+
69
+
return 0;
70
+
}
71
+
72
+
/* Count how many instructions it takes to do a null loop */
73
+
static u64 determine_overhead(struct event *events)
74
+
{
75
+
u64 current, overhead;
76
+
int i;
77
+
78
+
do_count_loop(events, 0, 0, false);
79
+
overhead = events[0].result.value;
80
+
81
+
for (i = 0; i < 100; i++) {
82
+
do_count_loop(events, 0, 0, false);
83
+
current = events[0].result.value;
84
+
if (current < overhead) {
85
+
printf("Replacing overhead %llu with %llu\n", overhead, current);
86
+
overhead = current;
87
+
}
88
+
}
89
+
90
+
return overhead;
91
+
}
92
+
93
+
static int count_instructions(void)
94
+
{
95
+
struct event events[2];
96
+
u64 overhead;
97
+
98
+
setup_event(&events[0], PERF_COUNT_HW_INSTRUCTIONS, "instructions");
99
+
setup_event(&events[1], PERF_COUNT_HW_CPU_CYCLES, "cycles");
100
+
101
+
if (event_open(&events[0])) {
102
+
perror("perf_event_open");
103
+
return -1;
104
+
}
105
+
106
+
if (event_open_with_group(&events[1], events[0].fd)) {
107
+
perror("perf_event_open");
108
+
return -1;
109
+
}
110
+
111
+
overhead = determine_overhead(events);
112
+
printf("Overhead of null loop: %llu instructions\n", overhead);
113
+
114
+
/* Run for 1M instructions */
115
+
FAIL_IF(do_count_loop(events, 0x100000, overhead, true));
116
+
117
+
/* Run for 10M instructions */
118
+
FAIL_IF(do_count_loop(events, 0xa00000, overhead, true));
119
+
120
+
/* Run for 100M instructions */
121
+
FAIL_IF(do_count_loop(events, 0x6400000, overhead, true));
122
+
123
+
/* Run for 1G instructions */
124
+
FAIL_IF(do_count_loop(events, 0x40000000, overhead, true));
125
+
126
+
event_close(&events[0]);
127
+
event_close(&events[1]);
128
+
129
+
return 0;
130
+
}
131
+
132
+
int main(void)
133
+
{
134
+
return test_harness(count_instructions, "count_instructions");
135
+
}
+105
tools/testing/selftests/powerpc/pmu/event.c
+105
tools/testing/selftests/powerpc/pmu/event.c
···
1
+
/*
2
+
* Copyright 2013, Michael Ellerman, IBM Corp.
3
+
* Licensed under GPLv2.
4
+
*/
5
+
6
+
#define _GNU_SOURCE
7
+
#include <unistd.h>
8
+
#include <sys/syscall.h>
9
+
#include <string.h>
10
+
#include <stdio.h>
11
+
#include <sys/ioctl.h>
12
+
13
+
#include "event.h"
14
+
15
+
16
+
int perf_event_open(struct perf_event_attr *attr, pid_t pid, int cpu,
17
+
int group_fd, unsigned long flags)
18
+
{
19
+
return syscall(__NR_perf_event_open, attr, pid, cpu,
20
+
group_fd, flags);
21
+
}
22
+
23
+
void event_init_opts(struct event *e, u64 config, int type, char *name)
24
+
{
25
+
memset(e, 0, sizeof(*e));
26
+
27
+
e->name = name;
28
+
29
+
e->attr.type = type;
30
+
e->attr.config = config;
31
+
e->attr.size = sizeof(e->attr);
32
+
/* This has to match the structure layout in the header */
33
+
e->attr.read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | \
34
+
PERF_FORMAT_TOTAL_TIME_RUNNING;
35
+
}
36
+
37
+
void event_init_named(struct event *e, u64 config, char *name)
38
+
{
39
+
event_init_opts(e, config, PERF_TYPE_RAW, name);
40
+
}
41
+
42
+
#define PERF_CURRENT_PID 0
43
+
#define PERF_NO_CPU -1
44
+
#define PERF_NO_GROUP -1
45
+
46
+
int event_open_with_options(struct event *e, pid_t pid, int cpu, int group_fd)
47
+
{
48
+
e->fd = perf_event_open(&e->attr, pid, cpu, group_fd, 0);
49
+
if (e->fd == -1) {
50
+
perror("perf_event_open");
51
+
return -1;
52
+
}
53
+
54
+
return 0;
55
+
}
56
+
57
+
int event_open_with_group(struct event *e, int group_fd)
58
+
{
59
+
return event_open_with_options(e, PERF_CURRENT_PID, PERF_NO_CPU, group_fd);
60
+
}
61
+
62
+
int event_open(struct event *e)
63
+
{
64
+
return event_open_with_options(e, PERF_CURRENT_PID, PERF_NO_CPU, PERF_NO_GROUP);
65
+
}
66
+
67
+
void event_close(struct event *e)
68
+
{
69
+
close(e->fd);
70
+
}
71
+
72
+
int event_reset(struct event *e)
73
+
{
74
+
return ioctl(e->fd, PERF_EVENT_IOC_RESET);
75
+
}
76
+
77
+
int event_read(struct event *e)
78
+
{
79
+
int rc;
80
+
81
+
rc = read(e->fd, &e->result, sizeof(e->result));
82
+
if (rc != sizeof(e->result)) {
83
+
fprintf(stderr, "read error on event %p!\n", e);
84
+
return -1;
85
+
}
86
+
87
+
return 0;
88
+
}
89
+
90
+
void event_report_justified(struct event *e, int name_width, int result_width)
91
+
{
92
+
printf("%*s: result %*llu ", name_width, e->name, result_width,
93
+
e->result.value);
94
+
95
+
if (e->result.running == e->result.enabled)
96
+
printf("running/enabled %llu\n", e->result.running);
97
+
else
98
+
printf("running %llu enabled %llu\n", e->result.running,
99
+
e->result.enabled);
100
+
}
101
+
102
+
void event_report(struct event *e)
103
+
{
104
+
event_report_justified(e, 0, 0);
105
+
}
+39
tools/testing/selftests/powerpc/pmu/event.h
+39
tools/testing/selftests/powerpc/pmu/event.h
···
1
+
/*
2
+
* Copyright 2013, Michael Ellerman, IBM Corp.
3
+
* Licensed under GPLv2.
4
+
*/
5
+
6
+
#ifndef _SELFTESTS_POWERPC_PMU_EVENT_H
7
+
#define _SELFTESTS_POWERPC_PMU_EVENT_H
8
+
9
+
#include <unistd.h>
10
+
#include <linux/perf_event.h>
11
+
12
+
#include "utils.h"
13
+
14
+
15
+
struct event {
16
+
struct perf_event_attr attr;
17
+
char *name;
18
+
int fd;
19
+
/* This must match the read_format we use */
20
+
struct {
21
+
u64 value;
22
+
u64 running;
23
+
u64 enabled;
24
+
} result;
25
+
};
26
+
27
+
void event_init(struct event *e, u64 config);
28
+
void event_init_named(struct event *e, u64 config, char *name);
29
+
void event_init_opts(struct event *e, u64 config, int type, char *name);
30
+
int event_open_with_options(struct event *e, pid_t pid, int cpu, int group_fd);
31
+
int event_open_with_group(struct event *e, int group_fd);
32
+
int event_open(struct event *e);
33
+
void event_close(struct event *e);
34
+
int event_reset(struct event *e);
35
+
int event_read(struct event *e);
36
+
void event_report_justified(struct event *e, int name_width, int result_width);
37
+
void event_report(struct event *e);
38
+
39
+
#endif /* _SELFTESTS_POWERPC_PMU_EVENT_H */
+46
tools/testing/selftests/powerpc/pmu/loop.S
+46
tools/testing/selftests/powerpc/pmu/loop.S
···
1
+
/*
2
+
* Copyright 2013, Michael Ellerman, IBM Corp.
3
+
* Licensed under GPLv2.
4
+
*/
5
+
6
+
.text
7
+
8
+
.global thirty_two_instruction_loop
9
+
.type .thirty_two_instruction_loop,@function
10
+
.section ".opd","aw",@progbits
11
+
thirty_two_instruction_loop:
12
+
.quad .thirty_two_instruction_loop, .TOC.@tocbase, 0
13
+
.previous
14
+
.thirty_two_instruction_loop:
15
+
cmpwi %r3,0
16
+
beqlr
17
+
addi %r4,%r3,1
18
+
addi %r4,%r4,1
19
+
addi %r4,%r4,1
20
+
addi %r4,%r4,1
21
+
addi %r4,%r4,1
22
+
addi %r4,%r4,1
23
+
addi %r4,%r4,1
24
+
addi %r4,%r4,1
25
+
addi %r4,%r4,1
26
+
addi %r4,%r4,1
27
+
addi %r4,%r4,1
28
+
addi %r4,%r4,1
29
+
addi %r4,%r4,1
30
+
addi %r4,%r4,1
31
+
addi %r4,%r4,1
32
+
addi %r4,%r4,1
33
+
addi %r4,%r4,1
34
+
addi %r4,%r4,1
35
+
addi %r4,%r4,1
36
+
addi %r4,%r4,1
37
+
addi %r4,%r4,1
38
+
addi %r4,%r4,1
39
+
addi %r4,%r4,1
40
+
addi %r4,%r4,1
41
+
addi %r4,%r4,1
42
+
addi %r4,%r4,1
43
+
addi %r4,%r4,1
44
+
addi %r4,%r4,1 # 28 addi's
45
+
subi %r3,%r3,1
46
+
b .thirty_two_instruction_loop