1 // SPDX-License-Identifier: GPL-2.0 2 #include <linux/compiler.h> 3 #include <string.h> 4 #include <perf/cpumap.h> 5 #include <perf/evlist.h> 6 #include "metricgroup.h" 7 #include "tests.h" 8 #include "pmu-events/pmu-events.h" 9 #include "evlist.h" 10 #include "rblist.h" 11 #include "debug.h" 12 #include "expr.h" 13 #include "stat.h" 14 #include "pmu.h" 15 16 struct value { 17 const char *event; 18 u64 val; 19 }; 20 21 static u64 find_value(const char *name, struct value *values) 22 { 23 struct value *v = values; 24 25 while (v->event) { 26 if (!strcmp(name, v->event)) 27 return v->val; 28 v++; 29 } 30 return 0; 31 } 32 33 static void load_runtime_stat(struct evlist *evlist, struct value *vals) 34 { 35 struct evsel *evsel; 36 u64 count; 37 38 evlist__alloc_aggr_stats(evlist, 1); 39 evlist__for_each_entry(evlist, evsel) { 40 count = find_value(evsel->name, vals); 41 evsel->stats->aggr->counts.val = count; 42 if (!strcmp(evsel->name, "duration_time")) 43 update_stats(&walltime_nsecs_stats, count); 44 } 45 } 46 47 static double compute_single(struct rblist *metric_events, struct evlist *evlist, 48 const char *name) 49 { 50 struct metric_expr *mexp; 51 struct metric_event *me; 52 struct evsel *evsel; 53 54 evlist__for_each_entry(evlist, evsel) { 55 me = metricgroup__lookup(metric_events, evsel, false); 56 if (me != NULL) { 57 list_for_each_entry (mexp, &me->head, nd) { 58 if (strcmp(mexp->metric_name, name)) 59 continue; 60 return test_generic_metric(mexp, 0); 61 } 62 } 63 } 64 return 0.; 65 } 66 67 static int __compute_metric(const char *name, struct value *vals, 68 const char *name1, double *ratio1, 69 const char *name2, double *ratio2) 70 { 71 struct rblist metric_events = { 72 .nr_entries = 0, 73 }; 74 const struct pmu_metrics_table *pme_test; 75 struct perf_cpu_map *cpus; 76 struct evlist *evlist; 77 int err; 78 79 /* 80 * We need to prepare evlist for stat mode running on CPU 0 81 * because that's where all the stats are going to be created. 82 */ 83 evlist = evlist__new(); 84 if (!evlist) 85 return -ENOMEM; 86 87 cpus = perf_cpu_map__new("0"); 88 if (!cpus) { 89 evlist__delete(evlist); 90 return -ENOMEM; 91 } 92 93 perf_evlist__set_maps(&evlist->core, cpus, NULL); 94 95 /* Parse the metric into metric_events list. */ 96 pme_test = find_core_metrics_table("testarch", "testcpu"); 97 err = metricgroup__parse_groups_test(evlist, pme_test, name, 98 &metric_events); 99 if (err) 100 goto out; 101 102 err = evlist__alloc_stats(/*config=*/NULL, evlist, /*alloc_raw=*/false); 103 if (err) 104 goto out; 105 106 /* Load the runtime stats with given numbers for events. */ 107 load_runtime_stat(evlist, vals); 108 109 /* And execute the metric */ 110 if (name1 && ratio1) 111 *ratio1 = compute_single(&metric_events, evlist, name1); 112 if (name2 && ratio2) 113 *ratio2 = compute_single(&metric_events, evlist, name2); 114 115 out: 116 /* ... cleanup. */ 117 metricgroup__rblist_exit(&metric_events); 118 evlist__free_stats(evlist); 119 perf_cpu_map__put(cpus); 120 evlist__delete(evlist); 121 return err; 122 } 123 124 static int compute_metric(const char *name, struct value *vals, double *ratio) 125 { 126 return __compute_metric(name, vals, name, ratio, NULL, NULL); 127 } 128 129 static int compute_metric_group(const char *name, struct value *vals, 130 const char *name1, double *ratio1, 131 const char *name2, double *ratio2) 132 { 133 return __compute_metric(name, vals, name1, ratio1, name2, ratio2); 134 } 135 136 static int test_ipc(void) 137 { 138 double ratio; 139 struct value vals[] = { 140 { .event = "inst_retired.any", .val = 300 }, 141 { .event = "cpu_clk_unhalted.thread", .val = 200 }, 142 { .event = NULL, }, 143 }; 144 145 TEST_ASSERT_VAL("failed to compute metric", 146 compute_metric("IPC", vals, &ratio) == 0); 147 148 TEST_ASSERT_VAL("IPC failed, wrong ratio", 149 ratio == 1.5); 150 return 0; 151 } 152 153 static int test_frontend(void) 154 { 155 double ratio; 156 struct value vals[] = { 157 { .event = "idq_uops_not_delivered.core", .val = 300 }, 158 { .event = "cpu_clk_unhalted.thread", .val = 200 }, 159 { .event = "cpu_clk_unhalted.one_thread_active", .val = 400 }, 160 { .event = "cpu_clk_unhalted.ref_xclk", .val = 600 }, 161 { .event = NULL, }, 162 }; 163 164 TEST_ASSERT_VAL("failed to compute metric", 165 compute_metric("Frontend_Bound_SMT", vals, &ratio) == 0); 166 167 TEST_ASSERT_VAL("Frontend_Bound_SMT failed, wrong ratio", 168 ratio == 0.45); 169 return 0; 170 } 171 172 static int test_cache_miss_cycles(void) 173 { 174 double ratio; 175 struct value vals[] = { 176 { .event = "l1d-loads-misses", .val = 300 }, 177 { .event = "l1i-loads-misses", .val = 200 }, 178 { .event = "inst_retired.any", .val = 400 }, 179 { .event = NULL, }, 180 }; 181 182 TEST_ASSERT_VAL("failed to compute metric", 183 compute_metric("cache_miss_cycles", vals, &ratio) == 0); 184 185 TEST_ASSERT_VAL("cache_miss_cycles failed, wrong ratio", 186 ratio == 1.25); 187 return 0; 188 } 189 190 191 /* 192 * DCache_L2_All_Hits = l2_rqsts.demand_data_rd_hit + l2_rqsts.pf_hit + l2_rqsts.rfo_hi 193 * DCache_L2_All_Miss = max(l2_rqsts.all_demand_data_rd - l2_rqsts.demand_data_rd_hit, 0) + 194 * l2_rqsts.pf_miss + l2_rqsts.rfo_miss 195 * DCache_L2_All = dcache_l2_all_hits + dcache_l2_all_miss 196 * DCache_L2_Hits = d_ratio(dcache_l2_all_hits, dcache_l2_all) 197 * DCache_L2_Misses = d_ratio(dcache_l2_all_miss, dcache_l2_all) 198 * 199 * l2_rqsts.demand_data_rd_hit = 100 200 * l2_rqsts.pf_hit = 200 201 * l2_rqsts.rfo_hi = 300 202 * l2_rqsts.all_demand_data_rd = 400 203 * l2_rqsts.pf_miss = 500 204 * l2_rqsts.rfo_miss = 600 205 * 206 * DCache_L2_All_Hits = 600 207 * DCache_L2_All_Miss = MAX(400 - 100, 0) + 500 + 600 = 1400 208 * DCache_L2_All = 600 + 1400 = 2000 209 * DCache_L2_Hits = 600 / 2000 = 0.3 210 * DCache_L2_Misses = 1400 / 2000 = 0.7 211 */ 212 static int test_dcache_l2(void) 213 { 214 double ratio; 215 struct value vals[] = { 216 { .event = "l2_rqsts.demand_data_rd_hit", .val = 100 }, 217 { .event = "l2_rqsts.pf_hit", .val = 200 }, 218 { .event = "l2_rqsts.rfo_hit", .val = 300 }, 219 { .event = "l2_rqsts.all_demand_data_rd", .val = 400 }, 220 { .event = "l2_rqsts.pf_miss", .val = 500 }, 221 { .event = "l2_rqsts.rfo_miss", .val = 600 }, 222 { .event = NULL, }, 223 }; 224 225 TEST_ASSERT_VAL("failed to compute metric", 226 compute_metric("DCache_L2_Hits", vals, &ratio) == 0); 227 228 TEST_ASSERT_VAL("DCache_L2_Hits failed, wrong ratio", 229 ratio == 0.3); 230 231 TEST_ASSERT_VAL("failed to compute metric", 232 compute_metric("DCache_L2_Misses", vals, &ratio) == 0); 233 234 TEST_ASSERT_VAL("DCache_L2_Misses failed, wrong ratio", 235 ratio == 0.7); 236 return 0; 237 } 238 239 static int test_recursion_fail(void) 240 { 241 double ratio; 242 struct value vals[] = { 243 { .event = "inst_retired.any", .val = 300 }, 244 { .event = "cpu_clk_unhalted.thread", .val = 200 }, 245 { .event = NULL, }, 246 }; 247 248 TEST_ASSERT_VAL("failed to find recursion", 249 compute_metric("M1", vals, &ratio) == -1); 250 251 TEST_ASSERT_VAL("failed to find recursion", 252 compute_metric("M3", vals, &ratio) == -1); 253 return 0; 254 } 255 256 static int test_memory_bandwidth(void) 257 { 258 double ratio; 259 struct value vals[] = { 260 { .event = "l1d.replacement", .val = 4000000 }, 261 { .event = "duration_time", .val = 200000000 }, 262 { .event = NULL, }, 263 }; 264 265 TEST_ASSERT_VAL("failed to compute metric", 266 compute_metric("L1D_Cache_Fill_BW", vals, &ratio) == 0); 267 TEST_ASSERT_VAL("L1D_Cache_Fill_BW, wrong ratio", 268 1.28 == ratio); 269 270 return 0; 271 } 272 273 static int test_metric_group(void) 274 { 275 double ratio1, ratio2; 276 struct value vals[] = { 277 { .event = "cpu_clk_unhalted.thread", .val = 200 }, 278 { .event = "l1d-loads-misses", .val = 300 }, 279 { .event = "l1i-loads-misses", .val = 200 }, 280 { .event = "inst_retired.any", .val = 400 }, 281 { .event = NULL, }, 282 }; 283 284 TEST_ASSERT_VAL("failed to find recursion", 285 compute_metric_group("group1", vals, 286 "IPC", &ratio1, 287 "cache_miss_cycles", &ratio2) == 0); 288 289 TEST_ASSERT_VAL("group IPC failed, wrong ratio", 290 ratio1 == 2.0); 291 292 TEST_ASSERT_VAL("group cache_miss_cycles failed, wrong ratio", 293 ratio2 == 1.25); 294 return 0; 295 } 296 297 static int test__parse_metric(struct test_suite *test __maybe_unused, int subtest __maybe_unused) 298 { 299 TEST_ASSERT_VAL("IPC failed", test_ipc() == 0); 300 TEST_ASSERT_VAL("frontend failed", test_frontend() == 0); 301 TEST_ASSERT_VAL("DCache_L2 failed", test_dcache_l2() == 0); 302 TEST_ASSERT_VAL("recursion fail failed", test_recursion_fail() == 0); 303 TEST_ASSERT_VAL("Memory bandwidth", test_memory_bandwidth() == 0); 304 305 if (!perf_pmu__has_hybrid()) { 306 TEST_ASSERT_VAL("cache_miss_cycles failed", test_cache_miss_cycles() == 0); 307 TEST_ASSERT_VAL("test metric group", test_metric_group() == 0); 308 } 309 return 0; 310 } 311 312 DEFINE_SUITE("Parse and process metrics", parse_metric); 313