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