1 | // SPDX-License-Identifier: GPL-2.0-only |
2 | /* |
3 | * Copyright 2013, Michael Ellerman, IBM Corp. |
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 | #include "lib.h" |
16 | |
17 | extern void thirty_two_instruction_loop(u64 loops); |
18 | |
19 | static void setup_event(struct event *e, u64 config, char *name) |
20 | { |
21 | event_init_opts(e, config, type: PERF_TYPE_HARDWARE, name); |
22 | |
23 | e->attr.disabled = 1; |
24 | e->attr.exclude_kernel = 1; |
25 | e->attr.exclude_hv = 1; |
26 | e->attr.exclude_idle = 1; |
27 | } |
28 | |
29 | static int do_count_loop(struct event *events, u64 instructions, |
30 | u64 overhead, bool report) |
31 | { |
32 | s64 difference, expected; |
33 | double percentage; |
34 | |
35 | prctl(PR_TASK_PERF_EVENTS_ENABLE); |
36 | |
37 | /* Run for 1M instructions */ |
38 | thirty_two_instruction_loop(loops: instructions >> 5); |
39 | |
40 | prctl(PR_TASK_PERF_EVENTS_DISABLE); |
41 | |
42 | event_read(e: &events[0]); |
43 | event_read(e: &events[1]); |
44 | |
45 | expected = instructions + overhead; |
46 | difference = events[0].result.value - expected; |
47 | percentage = (double)difference / events[0].result.value * 100; |
48 | |
49 | if (report) { |
50 | event_report(e: &events[0]); |
51 | event_report(e: &events[1]); |
52 | |
53 | printf("Looped for %llu instructions, overhead %llu\n" , instructions, overhead); |
54 | printf("Expected %llu\n" , expected); |
55 | printf("Actual %llu\n" , events[0].result.value); |
56 | printf("Delta %lld, %f%%\n" , difference, percentage); |
57 | } |
58 | |
59 | event_reset(e: &events[0]); |
60 | event_reset(e: &events[1]); |
61 | |
62 | if (difference < 0) |
63 | difference = -difference; |
64 | |
65 | /* Tolerate a difference below 0.0001 % */ |
66 | difference *= 10000 * 100; |
67 | if (difference / events[0].result.value) |
68 | return -1; |
69 | |
70 | return 0; |
71 | } |
72 | |
73 | /* Count how many instructions it takes to do a null loop */ |
74 | static u64 determine_overhead(struct event *events) |
75 | { |
76 | u64 current, overhead; |
77 | int i; |
78 | |
79 | do_count_loop(events, instructions: 0, overhead: 0, report: false); |
80 | overhead = events[0].result.value; |
81 | |
82 | for (i = 0; i < 100; i++) { |
83 | do_count_loop(events, instructions: 0, overhead: 0, report: false); |
84 | current = events[0].result.value; |
85 | if (current < overhead) { |
86 | printf("Replacing overhead %llu with %llu\n" , overhead, current); |
87 | overhead = current; |
88 | } |
89 | } |
90 | |
91 | return overhead; |
92 | } |
93 | |
94 | static int test_body(void) |
95 | { |
96 | struct event events[2]; |
97 | u64 overhead; |
98 | |
99 | setup_event(e: &events[0], config: PERF_COUNT_HW_INSTRUCTIONS, name: "instructions" ); |
100 | setup_event(e: &events[1], config: PERF_COUNT_HW_CPU_CYCLES, name: "cycles" ); |
101 | |
102 | if (event_open(e: &events[0])) { |
103 | perror("perf_event_open" ); |
104 | return -1; |
105 | } |
106 | |
107 | if (event_open_with_group(e: &events[1], group_fd: events[0].fd)) { |
108 | perror("perf_event_open" ); |
109 | return -1; |
110 | } |
111 | |
112 | overhead = determine_overhead(events); |
113 | printf("Overhead of null loop: %llu instructions\n" , overhead); |
114 | |
115 | /* Run for 1Mi instructions */ |
116 | FAIL_IF(do_count_loop(events, instructions: 1000000, overhead, report: true)); |
117 | |
118 | /* Run for 10Mi instructions */ |
119 | FAIL_IF(do_count_loop(events, instructions: 10000000, overhead, report: true)); |
120 | |
121 | /* Run for 100Mi instructions */ |
122 | FAIL_IF(do_count_loop(events, instructions: 100000000, overhead, report: true)); |
123 | |
124 | /* Run for 1Bi instructions */ |
125 | FAIL_IF(do_count_loop(events, instructions: 1000000000, overhead, report: true)); |
126 | |
127 | /* Run for 16Bi instructions */ |
128 | FAIL_IF(do_count_loop(events, instructions: 16000000000, overhead, report: true)); |
129 | |
130 | /* Run for 64Bi instructions */ |
131 | FAIL_IF(do_count_loop(events, instructions: 64000000000, overhead, report: true)); |
132 | |
133 | event_close(e: &events[0]); |
134 | event_close(e: &events[1]); |
135 | |
136 | return 0; |
137 | } |
138 | |
139 | static int count_instructions(void) |
140 | { |
141 | return eat_cpu(test_function: test_body); |
142 | } |
143 | |
144 | int main(void) |
145 | { |
146 | return test_harness(count_instructions, "count_instructions" ); |
147 | } |
148 | |