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 static struct pmu_event pme_test[] = { 17 { 18 .metric_expr = "inst_retired.any / cpu_clk_unhalted.thread", 19 .metric_name = "IPC", 20 .metric_group = "group1", 21 }, 22 { 23 .metric_expr = "idq_uops_not_delivered.core / (4 * (( ( cpu_clk_unhalted.thread / 2 ) * " 24 "( 1 + cpu_clk_unhalted.one_thread_active / cpu_clk_unhalted.ref_xclk ) )))", 25 .metric_name = "Frontend_Bound_SMT", 26 }, 27 { 28 .metric_expr = "l1d\\-loads\\-misses / inst_retired.any", 29 .metric_name = "dcache_miss_cpi", 30 }, 31 { 32 .metric_expr = "l1i\\-loads\\-misses / inst_retired.any", 33 .metric_name = "icache_miss_cycles", 34 }, 35 { 36 .metric_expr = "(dcache_miss_cpi + icache_miss_cycles)", 37 .metric_name = "cache_miss_cycles", 38 .metric_group = "group1", 39 }, 40 { 41 .metric_expr = "l2_rqsts.demand_data_rd_hit + l2_rqsts.pf_hit + l2_rqsts.rfo_hit", 42 .metric_name = "DCache_L2_All_Hits", 43 }, 44 { 45 .metric_expr = "max(l2_rqsts.all_demand_data_rd - l2_rqsts.demand_data_rd_hit, 0) + " 46 "l2_rqsts.pf_miss + l2_rqsts.rfo_miss", 47 .metric_name = "DCache_L2_All_Miss", 48 }, 49 { 50 .metric_expr = "dcache_l2_all_hits + dcache_l2_all_miss", 51 .metric_name = "DCache_L2_All", 52 }, 53 { 54 .metric_expr = "d_ratio(dcache_l2_all_hits, dcache_l2_all)", 55 .metric_name = "DCache_L2_Hits", 56 }, 57 { 58 .metric_expr = "d_ratio(dcache_l2_all_miss, dcache_l2_all)", 59 .metric_name = "DCache_L2_Misses", 60 }, 61 { 62 .metric_expr = "ipc + m2", 63 .metric_name = "M1", 64 }, 65 { 66 .metric_expr = "ipc + m1", 67 .metric_name = "M2", 68 }, 69 { 70 .metric_expr = "1/m3", 71 .metric_name = "M3", 72 }, 73 { 74 .metric_expr = "64 * l1d.replacement / 1000000000 / duration_time", 75 .metric_name = "L1D_Cache_Fill_BW", 76 }, 77 { 78 .name = NULL, 79 } 80 }; 81 82 static const struct pmu_events_map map = { 83 .cpuid = "test", 84 .version = "1", 85 .type = "core", 86 .table = pme_test, 87 }; 88 89 struct value { 90 const char *event; 91 u64 val; 92 }; 93 94 static u64 find_value(const char *name, struct value *values) 95 { 96 struct value *v = values; 97 98 while (v->event) { 99 if (!strcmp(name, v->event)) 100 return v->val; 101 v++; 102 } 103 return 0; 104 } 105 106 static void load_runtime_stat(struct runtime_stat *st, struct evlist *evlist, 107 struct value *vals) 108 { 109 struct evsel *evsel; 110 u64 count; 111 112 perf_stat__reset_shadow_stats(); 113 evlist__for_each_entry(evlist, evsel) { 114 count = find_value(evsel->name, vals); 115 perf_stat__update_shadow_stats(evsel, count, 0, st); 116 if (!strcmp(evsel->name, "duration_time")) 117 update_stats(&walltime_nsecs_stats, count); 118 } 119 } 120 121 static double compute_single(struct rblist *metric_events, struct evlist *evlist, 122 struct runtime_stat *st, const char *name) 123 { 124 struct metric_expr *mexp; 125 struct metric_event *me; 126 struct evsel *evsel; 127 128 evlist__for_each_entry(evlist, evsel) { 129 me = metricgroup__lookup(metric_events, evsel, false); 130 if (me != NULL) { 131 list_for_each_entry (mexp, &me->head, nd) { 132 if (strcmp(mexp->metric_name, name)) 133 continue; 134 return test_generic_metric(mexp, 0, st); 135 } 136 } 137 } 138 return 0.; 139 } 140 141 static int __compute_metric(const char *name, struct value *vals, 142 const char *name1, double *ratio1, 143 const char *name2, double *ratio2) 144 { 145 struct rblist metric_events = { 146 .nr_entries = 0, 147 }; 148 struct perf_cpu_map *cpus; 149 struct runtime_stat st; 150 struct evlist *evlist; 151 int err; 152 153 /* 154 * We need to prepare evlist for stat mode running on CPU 0 155 * because that's where all the stats are going to be created. 156 */ 157 evlist = evlist__new(); 158 if (!evlist) 159 return -ENOMEM; 160 161 cpus = perf_cpu_map__new("0"); 162 if (!cpus) { 163 evlist__delete(evlist); 164 return -ENOMEM; 165 } 166 167 perf_evlist__set_maps(&evlist->core, cpus, NULL); 168 runtime_stat__init(&st); 169 170 /* Parse the metric into metric_events list. */ 171 err = metricgroup__parse_groups_test(evlist, &map, name, 172 false, false, 173 &metric_events); 174 if (err) 175 goto out; 176 177 err = evlist__alloc_stats(evlist, false); 178 if (err) 179 goto out; 180 181 /* Load the runtime stats with given numbers for events. */ 182 load_runtime_stat(&st, evlist, vals); 183 184 /* And execute the metric */ 185 if (name1 && ratio1) 186 *ratio1 = compute_single(&metric_events, evlist, &st, name1); 187 if (name2 && ratio2) 188 *ratio2 = compute_single(&metric_events, evlist, &st, name2); 189 190 out: 191 /* ... cleanup. */ 192 metricgroup__rblist_exit(&metric_events); 193 runtime_stat__exit(&st); 194 evlist__free_stats(evlist); 195 perf_cpu_map__put(cpus); 196 evlist__delete(evlist); 197 return err; 198 } 199 200 static int compute_metric(const char *name, struct value *vals, double *ratio) 201 { 202 return __compute_metric(name, vals, name, ratio, NULL, NULL); 203 } 204 205 static int compute_metric_group(const char *name, struct value *vals, 206 const char *name1, double *ratio1, 207 const char *name2, double *ratio2) 208 { 209 return __compute_metric(name, vals, name1, ratio1, name2, ratio2); 210 } 211 212 static int test_ipc(void) 213 { 214 double ratio; 215 struct value vals[] = { 216 { .event = "inst_retired.any", .val = 300 }, 217 { .event = "cpu_clk_unhalted.thread", .val = 200 }, 218 { .event = NULL, }, 219 }; 220 221 TEST_ASSERT_VAL("failed to compute metric", 222 compute_metric("IPC", vals, &ratio) == 0); 223 224 TEST_ASSERT_VAL("IPC failed, wrong ratio", 225 ratio == 1.5); 226 return 0; 227 } 228 229 static int test_frontend(void) 230 { 231 double ratio; 232 struct value vals[] = { 233 { .event = "idq_uops_not_delivered.core", .val = 300 }, 234 { .event = "cpu_clk_unhalted.thread", .val = 200 }, 235 { .event = "cpu_clk_unhalted.one_thread_active", .val = 400 }, 236 { .event = "cpu_clk_unhalted.ref_xclk", .val = 600 }, 237 { .event = NULL, }, 238 }; 239 240 TEST_ASSERT_VAL("failed to compute metric", 241 compute_metric("Frontend_Bound_SMT", vals, &ratio) == 0); 242 243 TEST_ASSERT_VAL("Frontend_Bound_SMT failed, wrong ratio", 244 ratio == 0.45); 245 return 0; 246 } 247 248 static int test_cache_miss_cycles(void) 249 { 250 double ratio; 251 struct value vals[] = { 252 { .event = "l1d-loads-misses", .val = 300 }, 253 { .event = "l1i-loads-misses", .val = 200 }, 254 { .event = "inst_retired.any", .val = 400 }, 255 { .event = NULL, }, 256 }; 257 258 TEST_ASSERT_VAL("failed to compute metric", 259 compute_metric("cache_miss_cycles", vals, &ratio) == 0); 260 261 TEST_ASSERT_VAL("cache_miss_cycles failed, wrong ratio", 262 ratio == 1.25); 263 return 0; 264 } 265 266 267 /* 268 * DCache_L2_All_Hits = l2_rqsts.demand_data_rd_hit + l2_rqsts.pf_hit + l2_rqsts.rfo_hi 269 * DCache_L2_All_Miss = max(l2_rqsts.all_demand_data_rd - l2_rqsts.demand_data_rd_hit, 0) + 270 * l2_rqsts.pf_miss + l2_rqsts.rfo_miss 271 * DCache_L2_All = dcache_l2_all_hits + dcache_l2_all_miss 272 * DCache_L2_Hits = d_ratio(dcache_l2_all_hits, dcache_l2_all) 273 * DCache_L2_Misses = d_ratio(dcache_l2_all_miss, dcache_l2_all) 274 * 275 * l2_rqsts.demand_data_rd_hit = 100 276 * l2_rqsts.pf_hit = 200 277 * l2_rqsts.rfo_hi = 300 278 * l2_rqsts.all_demand_data_rd = 400 279 * l2_rqsts.pf_miss = 500 280 * l2_rqsts.rfo_miss = 600 281 * 282 * DCache_L2_All_Hits = 600 283 * DCache_L2_All_Miss = MAX(400 - 100, 0) + 500 + 600 = 1400 284 * DCache_L2_All = 600 + 1400 = 2000 285 * DCache_L2_Hits = 600 / 2000 = 0.3 286 * DCache_L2_Misses = 1400 / 2000 = 0.7 287 */ 288 static int test_dcache_l2(void) 289 { 290 double ratio; 291 struct value vals[] = { 292 { .event = "l2_rqsts.demand_data_rd_hit", .val = 100 }, 293 { .event = "l2_rqsts.pf_hit", .val = 200 }, 294 { .event = "l2_rqsts.rfo_hit", .val = 300 }, 295 { .event = "l2_rqsts.all_demand_data_rd", .val = 400 }, 296 { .event = "l2_rqsts.pf_miss", .val = 500 }, 297 { .event = "l2_rqsts.rfo_miss", .val = 600 }, 298 { .event = NULL, }, 299 }; 300 301 TEST_ASSERT_VAL("failed to compute metric", 302 compute_metric("DCache_L2_Hits", vals, &ratio) == 0); 303 304 TEST_ASSERT_VAL("DCache_L2_Hits failed, wrong ratio", 305 ratio == 0.3); 306 307 TEST_ASSERT_VAL("failed to compute metric", 308 compute_metric("DCache_L2_Misses", vals, &ratio) == 0); 309 310 TEST_ASSERT_VAL("DCache_L2_Misses failed, wrong ratio", 311 ratio == 0.7); 312 return 0; 313 } 314 315 static int test_recursion_fail(void) 316 { 317 double ratio; 318 struct value vals[] = { 319 { .event = "inst_retired.any", .val = 300 }, 320 { .event = "cpu_clk_unhalted.thread", .val = 200 }, 321 { .event = NULL, }, 322 }; 323 324 TEST_ASSERT_VAL("failed to find recursion", 325 compute_metric("M1", vals, &ratio) == -1); 326 327 TEST_ASSERT_VAL("failed to find recursion", 328 compute_metric("M3", vals, &ratio) == -1); 329 return 0; 330 } 331 332 static int test_memory_bandwidth(void) 333 { 334 double ratio; 335 struct value vals[] = { 336 { .event = "l1d.replacement", .val = 4000000 }, 337 { .event = "duration_time", .val = 200000000 }, 338 { .event = NULL, }, 339 }; 340 341 TEST_ASSERT_VAL("failed to compute metric", 342 compute_metric("L1D_Cache_Fill_BW", vals, &ratio) == 0); 343 TEST_ASSERT_VAL("L1D_Cache_Fill_BW, wrong ratio", 344 1.28 == ratio); 345 346 return 0; 347 } 348 349 static int test_metric_group(void) 350 { 351 double ratio1, ratio2; 352 struct value vals[] = { 353 { .event = "cpu_clk_unhalted.thread", .val = 200 }, 354 { .event = "l1d-loads-misses", .val = 300 }, 355 { .event = "l1i-loads-misses", .val = 200 }, 356 { .event = "inst_retired.any", .val = 400 }, 357 { .event = NULL, }, 358 }; 359 360 TEST_ASSERT_VAL("failed to find recursion", 361 compute_metric_group("group1", vals, 362 "IPC", &ratio1, 363 "cache_miss_cycles", &ratio2) == 0); 364 365 TEST_ASSERT_VAL("group IPC failed, wrong ratio", 366 ratio1 == 2.0); 367 368 TEST_ASSERT_VAL("group cache_miss_cycles failed, wrong ratio", 369 ratio2 == 1.25); 370 return 0; 371 } 372 373 static int test__parse_metric(struct test_suite *test __maybe_unused, int subtest __maybe_unused) 374 { 375 TEST_ASSERT_VAL("IPC failed", test_ipc() == 0); 376 TEST_ASSERT_VAL("frontend failed", test_frontend() == 0); 377 TEST_ASSERT_VAL("DCache_L2 failed", test_dcache_l2() == 0); 378 TEST_ASSERT_VAL("recursion fail failed", test_recursion_fail() == 0); 379 TEST_ASSERT_VAL("Memory bandwidth", test_memory_bandwidth() == 0); 380 381 if (!perf_pmu__has_hybrid()) { 382 TEST_ASSERT_VAL("cache_miss_cycles failed", test_cache_miss_cycles() == 0); 383 TEST_ASSERT_VAL("test metric group", test_metric_group() == 0); 384 } 385 return 0; 386 } 387 388 DEFINE_SUITE("Parse and process metrics", parse_metric); 389