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 15 static struct pmu_event pme_test[] = { 16 { 17 .metric_expr = "inst_retired.any / cpu_clk_unhalted.thread", 18 .metric_name = "IPC", 19 .metric_group = "group1", 20 }, 21 { 22 .metric_expr = "idq_uops_not_delivered.core / (4 * (( ( cpu_clk_unhalted.thread / 2 ) * " 23 "( 1 + cpu_clk_unhalted.one_thread_active / cpu_clk_unhalted.ref_xclk ) )))", 24 .metric_name = "Frontend_Bound_SMT", 25 }, 26 { 27 .metric_expr = "l1d\\-loads\\-misses / inst_retired.any", 28 .metric_name = "dcache_miss_cpi", 29 }, 30 { 31 .metric_expr = "l1i\\-loads\\-misses / inst_retired.any", 32 .metric_name = "icache_miss_cycles", 33 }, 34 { 35 .metric_expr = "(dcache_miss_cpi + icache_miss_cycles)", 36 .metric_name = "cache_miss_cycles", 37 .metric_group = "group1", 38 }, 39 { 40 .metric_expr = "l2_rqsts.demand_data_rd_hit + l2_rqsts.pf_hit + l2_rqsts.rfo_hit", 41 .metric_name = "DCache_L2_All_Hits", 42 }, 43 { 44 .metric_expr = "max(l2_rqsts.all_demand_data_rd - l2_rqsts.demand_data_rd_hit, 0) + " 45 "l2_rqsts.pf_miss + l2_rqsts.rfo_miss", 46 .metric_name = "DCache_L2_All_Miss", 47 }, 48 { 49 .metric_expr = "dcache_l2_all_hits + dcache_l2_all_miss", 50 .metric_name = "DCache_L2_All", 51 }, 52 { 53 .metric_expr = "d_ratio(dcache_l2_all_hits, dcache_l2_all)", 54 .metric_name = "DCache_L2_Hits", 55 }, 56 { 57 .metric_expr = "d_ratio(dcache_l2_all_miss, dcache_l2_all)", 58 .metric_name = "DCache_L2_Misses", 59 }, 60 { 61 .metric_expr = "ipc + m2", 62 .metric_name = "M1", 63 }, 64 { 65 .metric_expr = "ipc + m1", 66 .metric_name = "M2", 67 }, 68 { 69 .metric_expr = "1/m3", 70 .metric_name = "M3", 71 }, 72 { 73 .name = NULL, 74 } 75 }; 76 77 static struct pmu_events_map map = { 78 .cpuid = "test", 79 .version = "1", 80 .type = "core", 81 .table = pme_test, 82 }; 83 84 struct value { 85 const char *event; 86 u64 val; 87 }; 88 89 static u64 find_value(const char *name, struct value *values) 90 { 91 struct value *v = values; 92 93 while (v->event) { 94 if (!strcmp(name, v->event)) 95 return v->val; 96 v++; 97 }; 98 return 0; 99 } 100 101 static void load_runtime_stat(struct runtime_stat *st, struct evlist *evlist, 102 struct value *vals) 103 { 104 struct evsel *evsel; 105 u64 count; 106 107 evlist__for_each_entry(evlist, evsel) { 108 count = find_value(evsel->name, vals); 109 perf_stat__update_shadow_stats(evsel, count, 0, st); 110 } 111 } 112 113 static double compute_single(struct rblist *metric_events, struct evlist *evlist, 114 struct runtime_stat *st, const char *name) 115 { 116 struct metric_expr *mexp; 117 struct metric_event *me; 118 struct evsel *evsel; 119 120 evlist__for_each_entry(evlist, evsel) { 121 me = metricgroup__lookup(metric_events, evsel, false); 122 if (me != NULL) { 123 list_for_each_entry (mexp, &me->head, nd) { 124 if (strcmp(mexp->metric_name, name)) 125 continue; 126 return test_generic_metric(mexp, 0, st); 127 } 128 } 129 } 130 return 0.; 131 } 132 133 static int __compute_metric(const char *name, struct value *vals, 134 const char *name1, double *ratio1, 135 const char *name2, double *ratio2) 136 { 137 struct rblist metric_events = { 138 .nr_entries = 0, 139 }; 140 struct perf_cpu_map *cpus; 141 struct runtime_stat st; 142 struct evlist *evlist; 143 int err; 144 145 /* 146 * We need to prepare evlist for stat mode running on CPU 0 147 * because that's where all the stats are going to be created. 148 */ 149 evlist = evlist__new(); 150 if (!evlist) 151 return -ENOMEM; 152 153 cpus = perf_cpu_map__new("0"); 154 if (!cpus) { 155 evlist__delete(evlist); 156 return -ENOMEM; 157 } 158 159 perf_evlist__set_maps(&evlist->core, cpus, NULL); 160 runtime_stat__init(&st); 161 162 /* Parse the metric into metric_events list. */ 163 err = metricgroup__parse_groups_test(evlist, &map, name, 164 false, false, 165 &metric_events); 166 if (err) 167 goto out; 168 169 err = perf_evlist__alloc_stats(evlist, false); 170 if (err) 171 goto out; 172 173 /* Load the runtime stats with given numbers for events. */ 174 load_runtime_stat(&st, evlist, vals); 175 176 /* And execute the metric */ 177 if (name1 && ratio1) 178 *ratio1 = compute_single(&metric_events, evlist, &st, name1); 179 if (name2 && ratio2) 180 *ratio2 = compute_single(&metric_events, evlist, &st, name2); 181 182 out: 183 /* ... clenup. */ 184 metricgroup__rblist_exit(&metric_events); 185 runtime_stat__exit(&st); 186 perf_evlist__free_stats(evlist); 187 perf_cpu_map__put(cpus); 188 evlist__delete(evlist); 189 return err; 190 } 191 192 static int compute_metric(const char *name, struct value *vals, double *ratio) 193 { 194 return __compute_metric(name, vals, name, ratio, NULL, NULL); 195 } 196 197 static int compute_metric_group(const char *name, struct value *vals, 198 const char *name1, double *ratio1, 199 const char *name2, double *ratio2) 200 { 201 return __compute_metric(name, vals, name1, ratio1, name2, ratio2); 202 } 203 204 static int test_ipc(void) 205 { 206 double ratio; 207 struct value vals[] = { 208 { .event = "inst_retired.any", .val = 300 }, 209 { .event = "cpu_clk_unhalted.thread", .val = 200 }, 210 { .event = NULL, }, 211 }; 212 213 TEST_ASSERT_VAL("failed to compute metric", 214 compute_metric("IPC", vals, &ratio) == 0); 215 216 TEST_ASSERT_VAL("IPC failed, wrong ratio", 217 ratio == 1.5); 218 return 0; 219 } 220 221 static int test_frontend(void) 222 { 223 double ratio; 224 struct value vals[] = { 225 { .event = "idq_uops_not_delivered.core", .val = 300 }, 226 { .event = "cpu_clk_unhalted.thread", .val = 200 }, 227 { .event = "cpu_clk_unhalted.one_thread_active", .val = 400 }, 228 { .event = "cpu_clk_unhalted.ref_xclk", .val = 600 }, 229 { .event = NULL, }, 230 }; 231 232 TEST_ASSERT_VAL("failed to compute metric", 233 compute_metric("Frontend_Bound_SMT", vals, &ratio) == 0); 234 235 TEST_ASSERT_VAL("Frontend_Bound_SMT failed, wrong ratio", 236 ratio == 0.45); 237 return 0; 238 } 239 240 static int test_cache_miss_cycles(void) 241 { 242 double ratio; 243 struct value vals[] = { 244 { .event = "l1d-loads-misses", .val = 300 }, 245 { .event = "l1i-loads-misses", .val = 200 }, 246 { .event = "inst_retired.any", .val = 400 }, 247 { .event = NULL, }, 248 }; 249 250 TEST_ASSERT_VAL("failed to compute metric", 251 compute_metric("cache_miss_cycles", vals, &ratio) == 0); 252 253 TEST_ASSERT_VAL("cache_miss_cycles failed, wrong ratio", 254 ratio == 1.25); 255 return 0; 256 } 257 258 259 /* 260 * DCache_L2_All_Hits = l2_rqsts.demand_data_rd_hit + l2_rqsts.pf_hit + l2_rqsts.rfo_hi 261 * DCache_L2_All_Miss = max(l2_rqsts.all_demand_data_rd - l2_rqsts.demand_data_rd_hit, 0) + 262 * l2_rqsts.pf_miss + l2_rqsts.rfo_miss 263 * DCache_L2_All = dcache_l2_all_hits + dcache_l2_all_miss 264 * DCache_L2_Hits = d_ratio(dcache_l2_all_hits, dcache_l2_all) 265 * DCache_L2_Misses = d_ratio(dcache_l2_all_miss, dcache_l2_all) 266 * 267 * l2_rqsts.demand_data_rd_hit = 100 268 * l2_rqsts.pf_hit = 200 269 * l2_rqsts.rfo_hi = 300 270 * l2_rqsts.all_demand_data_rd = 400 271 * l2_rqsts.pf_miss = 500 272 * l2_rqsts.rfo_miss = 600 273 * 274 * DCache_L2_All_Hits = 600 275 * DCache_L2_All_Miss = MAX(400 - 100, 0) + 500 + 600 = 1400 276 * DCache_L2_All = 600 + 1400 = 2000 277 * DCache_L2_Hits = 600 / 2000 = 0.3 278 * DCache_L2_Misses = 1400 / 2000 = 0.7 279 */ 280 static int test_dcache_l2(void) 281 { 282 double ratio; 283 struct value vals[] = { 284 { .event = "l2_rqsts.demand_data_rd_hit", .val = 100 }, 285 { .event = "l2_rqsts.pf_hit", .val = 200 }, 286 { .event = "l2_rqsts.rfo_hit", .val = 300 }, 287 { .event = "l2_rqsts.all_demand_data_rd", .val = 400 }, 288 { .event = "l2_rqsts.pf_miss", .val = 500 }, 289 { .event = "l2_rqsts.rfo_miss", .val = 600 }, 290 { .event = NULL, }, 291 }; 292 293 TEST_ASSERT_VAL("failed to compute metric", 294 compute_metric("DCache_L2_Hits", vals, &ratio) == 0); 295 296 TEST_ASSERT_VAL("DCache_L2_Hits failed, wrong ratio", 297 ratio == 0.3); 298 299 TEST_ASSERT_VAL("failed to compute metric", 300 compute_metric("DCache_L2_Misses", vals, &ratio) == 0); 301 302 TEST_ASSERT_VAL("DCache_L2_Misses failed, wrong ratio", 303 ratio == 0.7); 304 return 0; 305 } 306 307 static int test_recursion_fail(void) 308 { 309 double ratio; 310 struct value vals[] = { 311 { .event = "inst_retired.any", .val = 300 }, 312 { .event = "cpu_clk_unhalted.thread", .val = 200 }, 313 { .event = NULL, }, 314 }; 315 316 TEST_ASSERT_VAL("failed to find recursion", 317 compute_metric("M1", vals, &ratio) == -1); 318 319 TEST_ASSERT_VAL("failed to find recursion", 320 compute_metric("M3", vals, &ratio) == -1); 321 return 0; 322 } 323 324 static int test_metric_group(void) 325 { 326 double ratio1, ratio2; 327 struct value vals[] = { 328 { .event = "cpu_clk_unhalted.thread", .val = 200 }, 329 { .event = "l1d-loads-misses", .val = 300 }, 330 { .event = "l1i-loads-misses", .val = 200 }, 331 { .event = "inst_retired.any", .val = 400 }, 332 { .event = NULL, }, 333 }; 334 335 TEST_ASSERT_VAL("failed to find recursion", 336 compute_metric_group("group1", vals, 337 "IPC", &ratio1, 338 "cache_miss_cycles", &ratio2) == 0); 339 340 TEST_ASSERT_VAL("group IPC failed, wrong ratio", 341 ratio1 == 2.0); 342 343 TEST_ASSERT_VAL("group cache_miss_cycles failed, wrong ratio", 344 ratio2 == 1.25); 345 return 0; 346 } 347 348 int test__parse_metric(struct test *test __maybe_unused, int subtest __maybe_unused) 349 { 350 TEST_ASSERT_VAL("IPC failed", test_ipc() == 0); 351 TEST_ASSERT_VAL("frontend failed", test_frontend() == 0); 352 TEST_ASSERT_VAL("cache_miss_cycles failed", test_cache_miss_cycles() == 0); 353 TEST_ASSERT_VAL("DCache_L2 failed", test_dcache_l2() == 0); 354 TEST_ASSERT_VAL("recursion fail failed", test_recursion_fail() == 0); 355 TEST_ASSERT_VAL("test metric group", test_metric_group() == 0); 356 return 0; 357 } 358