| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * Benchmark scanning sysfs files for PMU information. |
| 4 | * |
| 5 | * Copyright 2023 Google LLC. |
| 6 | */ |
| 7 | #include <errno.h> |
| 8 | #include <stdio.h> |
| 9 | #include "bench.h" |
| 10 | #include "util/debug.h" |
| 11 | #include "util/pmu.h" |
| 12 | #include "util/pmus.h" |
| 13 | #include "util/stat.h" |
| 14 | #include <linux/atomic.h> |
| 15 | #include <linux/err.h> |
| 16 | #include <linux/time64.h> |
| 17 | #include <subcmd/parse-options.h> |
| 18 | |
| 19 | static unsigned int iterations = 100; |
| 20 | |
| 21 | struct pmu_scan_result { |
| 22 | char *name; |
| 23 | int nr_aliases; |
| 24 | int nr_formats; |
| 25 | int nr_caps; |
| 26 | bool is_core; |
| 27 | }; |
| 28 | |
| 29 | static const struct option options[] = { |
| 30 | OPT_UINTEGER('i', "iterations" , &iterations, |
| 31 | "Number of iterations used to compute average" ), |
| 32 | OPT_END() |
| 33 | }; |
| 34 | |
| 35 | static const char *const bench_usage[] = { |
| 36 | "perf bench internals pmu-scan <options>" , |
| 37 | NULL |
| 38 | }; |
| 39 | |
| 40 | static int nr_pmus; |
| 41 | static struct pmu_scan_result *results; |
| 42 | |
| 43 | static int save_result(void) |
| 44 | { |
| 45 | struct perf_pmu *pmu = NULL; |
| 46 | struct list_head *list; |
| 47 | struct pmu_scan_result *r; |
| 48 | |
| 49 | while ((pmu = perf_pmus__scan(pmu)) != NULL) { |
| 50 | r = realloc(results, (nr_pmus + 1) * sizeof(*r)); |
| 51 | if (r == NULL) |
| 52 | return -ENOMEM; |
| 53 | |
| 54 | results = r; |
| 55 | r = results + nr_pmus; |
| 56 | |
| 57 | r->name = strdup(pmu->name); |
| 58 | r->is_core = pmu->is_core; |
| 59 | r->nr_caps = pmu->nr_caps; |
| 60 | |
| 61 | r->nr_aliases = perf_pmu__num_events(pmu); |
| 62 | |
| 63 | r->nr_formats = 0; |
| 64 | list_for_each(list, &pmu->format) |
| 65 | r->nr_formats++; |
| 66 | |
| 67 | pr_debug("pmu[%d] name=%s, nr_caps=%d, nr_aliases=%d, nr_formats=%d\n" , |
| 68 | nr_pmus, r->name, r->nr_caps, r->nr_aliases, r->nr_formats); |
| 69 | nr_pmus++; |
| 70 | } |
| 71 | |
| 72 | perf_pmus__destroy(); |
| 73 | return 0; |
| 74 | } |
| 75 | |
| 76 | static int check_result(bool core_only) |
| 77 | { |
| 78 | struct pmu_scan_result *r; |
| 79 | struct perf_pmu *pmu; |
| 80 | struct list_head *list; |
| 81 | int nr; |
| 82 | |
| 83 | for (int i = 0; i < nr_pmus; i++) { |
| 84 | r = &results[i]; |
| 85 | if (core_only && !r->is_core) |
| 86 | continue; |
| 87 | |
| 88 | pmu = perf_pmus__find(r->name); |
| 89 | if (pmu == NULL) { |
| 90 | pr_err("Cannot find PMU %s\n" , r->name); |
| 91 | return -1; |
| 92 | } |
| 93 | |
| 94 | if (pmu->nr_caps != (u32)r->nr_caps) { |
| 95 | pr_err("Unmatched number of event caps in %s: expect %d vs got %d\n" , |
| 96 | pmu->name, r->nr_caps, pmu->nr_caps); |
| 97 | return -1; |
| 98 | } |
| 99 | |
| 100 | nr = perf_pmu__num_events(pmu); |
| 101 | if (nr != r->nr_aliases) { |
| 102 | pr_err("Unmatched number of event aliases in %s: expect %d vs got %d\n" , |
| 103 | pmu->name, r->nr_aliases, nr); |
| 104 | return -1; |
| 105 | } |
| 106 | |
| 107 | nr = 0; |
| 108 | list_for_each(list, &pmu->format) |
| 109 | nr++; |
| 110 | if (nr != r->nr_formats) { |
| 111 | pr_err("Unmatched number of event formats in %s: expect %d vs got %d\n" , |
| 112 | pmu->name, r->nr_formats, nr); |
| 113 | return -1; |
| 114 | } |
| 115 | } |
| 116 | return 0; |
| 117 | } |
| 118 | |
| 119 | static void delete_result(void) |
| 120 | { |
| 121 | for (int i = 0; i < nr_pmus; i++) |
| 122 | free(results[i].name); |
| 123 | free(results); |
| 124 | |
| 125 | results = NULL; |
| 126 | nr_pmus = 0; |
| 127 | } |
| 128 | |
| 129 | static int run_pmu_scan(void) |
| 130 | { |
| 131 | struct stats stats; |
| 132 | struct timeval start, end, diff; |
| 133 | double time_average, time_stddev; |
| 134 | u64 runtime_us; |
| 135 | int ret; |
| 136 | |
| 137 | init_stats(&stats); |
| 138 | pr_info("Computing performance of sysfs PMU event scan for %u times\n" , |
| 139 | iterations); |
| 140 | |
| 141 | if (save_result() < 0) { |
| 142 | pr_err("Failed to initialize PMU scan result\n" ); |
| 143 | return -1; |
| 144 | } |
| 145 | |
| 146 | for (int j = 0; j < 2; j++) { |
| 147 | bool core_only = (j == 0); |
| 148 | |
| 149 | for (unsigned int i = 0; i < iterations; i++) { |
| 150 | gettimeofday(&start, NULL); |
| 151 | if (core_only) |
| 152 | perf_pmus__scan_core(NULL); |
| 153 | else |
| 154 | perf_pmus__scan(NULL); |
| 155 | gettimeofday(&end, NULL); |
| 156 | timersub(&end, &start, &diff); |
| 157 | runtime_us = diff.tv_sec * USEC_PER_SEC + diff.tv_usec; |
| 158 | update_stats(&stats, runtime_us); |
| 159 | |
| 160 | ret = check_result(core_only); |
| 161 | perf_pmus__destroy(); |
| 162 | if (ret < 0) |
| 163 | break; |
| 164 | } |
| 165 | time_average = avg_stats(&stats); |
| 166 | time_stddev = stddev_stats(&stats); |
| 167 | pr_info(" Average%s PMU scanning took: %.3f usec (+- %.3f usec)\n" , |
| 168 | core_only ? " core" : "" , time_average, time_stddev); |
| 169 | } |
| 170 | delete_result(); |
| 171 | return 0; |
| 172 | } |
| 173 | |
| 174 | int bench_pmu_scan(int argc, const char **argv) |
| 175 | { |
| 176 | int err = 0; |
| 177 | |
| 178 | argc = parse_options(argc, argv, options, bench_usage, 0); |
| 179 | if (argc) { |
| 180 | usage_with_options(bench_usage, options); |
| 181 | exit(EXIT_FAILURE); |
| 182 | } |
| 183 | |
| 184 | err = run_pmu_scan(); |
| 185 | |
| 186 | return err; |
| 187 | } |
| 188 | |