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