1// SPDX-License-Identifier: GPL-2.0 2/* 3 * Cache Allocation Technology (CAT) test 4 * 5 * Copyright (C) 2018 Intel Corporation 6 * 7 * Authors: 8 * Sai Praneeth Prakhya <sai.praneeth.prakhya@intel.com>, 9 * Fenghua Yu <fenghua.yu@intel.com> 10 */ 11#include "resctrl.h" 12#include <unistd.h> 13 14#define RESULT_FILE_NAME "result_cat" 15#define NUM_OF_RUNS 5 16 17/* 18 * Minimum difference in LLC misses between a test with n+1 bits CBM to the 19 * test with n bits is MIN_DIFF_PERCENT_PER_BIT * (n - 1). With e.g. 5 vs 4 20 * bits in the CBM mask, the minimum difference must be at least 21 * MIN_DIFF_PERCENT_PER_BIT * (4 - 1) = 3 percent. 22 * 23 * The relationship between number of used CBM bits and difference in LLC 24 * misses is not expected to be linear. With a small number of bits, the 25 * margin is smaller than with larger number of bits. For selftest purposes, 26 * however, linear approach is enough because ultimately only pass/fail 27 * decision has to be made and distinction between strong and stronger 28 * signal is irrelevant. 29 */ 30#define MIN_DIFF_PERCENT_PER_BIT 1UL 31 32static int show_results_info(__u64 sum_llc_val, int no_of_bits, 33 unsigned long cache_span, 34 unsigned long min_diff_percent, 35 unsigned long num_of_runs, bool platform, 36 __s64 *prev_avg_llc_val) 37{ 38 __u64 avg_llc_val = 0; 39 float avg_diff; 40 int ret = 0; 41 42 avg_llc_val = sum_llc_val / num_of_runs; 43 if (*prev_avg_llc_val) { 44 float delta = (__s64)(avg_llc_val - *prev_avg_llc_val); 45 46 avg_diff = delta / *prev_avg_llc_val; 47 ret = platform && (avg_diff * 100) < (float)min_diff_percent; 48 49 ksft_print_msg("%s Check cache miss rate changed more than %.1f%%\n", 50 ret ? "Fail:" : "Pass:", (float)min_diff_percent); 51 52 ksft_print_msg("Percent diff=%.1f\n", avg_diff * 100); 53 } 54 *prev_avg_llc_val = avg_llc_val; 55 56 show_cache_info(no_of_bits, avg_llc_val, cache_span, true); 57 58 return ret; 59} 60 61/* Remove the highest bit from CBM */ 62static unsigned long next_mask(unsigned long current_mask) 63{ 64 return current_mask & (current_mask >> 1); 65} 66 67static int check_results(struct resctrl_val_param *param, const char *cache_type, 68 unsigned long cache_total_size, unsigned long full_cache_mask, 69 unsigned long current_mask) 70{ 71 char *token_array[8], temp[512]; 72 __u64 sum_llc_perf_miss = 0; 73 __s64 prev_avg_llc_val = 0; 74 unsigned long alloc_size; 75 int runs = 0; 76 int fail = 0; 77 int ret; 78 FILE *fp; 79 80 ksft_print_msg("Checking for pass/fail\n"); 81 fp = fopen(param->filename, "r"); 82 if (!fp) { 83 ksft_perror("Cannot open file"); 84 85 return -1; 86 } 87 88 while (fgets(temp, sizeof(temp), fp)) { 89 char *token = strtok(temp, ":\t"); 90 int fields = 0; 91 int bits; 92 93 while (token) { 94 token_array[fields++] = token; 95 token = strtok(NULL, ":\t"); 96 } 97 98 sum_llc_perf_miss += strtoull(token_array[3], NULL, 0); 99 runs++; 100 101 if (runs < NUM_OF_RUNS) 102 continue; 103 104 if (!current_mask) { 105 ksft_print_msg("Unexpected empty cache mask\n"); 106 break; 107 } 108 109 alloc_size = cache_portion_size(cache_total_size, current_mask, full_cache_mask); 110 111 bits = count_bits(current_mask); 112 113 ret = show_results_info(sum_llc_perf_miss, bits, 114 alloc_size / 64, 115 MIN_DIFF_PERCENT_PER_BIT * (bits - 1), 116 runs, get_vendor() == ARCH_INTEL, 117 &prev_avg_llc_val); 118 if (ret) 119 fail = 1; 120 121 runs = 0; 122 sum_llc_perf_miss = 0; 123 current_mask = next_mask(current_mask); 124 } 125 126 fclose(fp); 127 128 return fail; 129} 130 131void cat_test_cleanup(void) 132{ 133 remove(RESULT_FILE_NAME); 134} 135 136/* 137 * cat_test - Execute CAT benchmark and measure cache misses 138 * @test: Test information structure 139 * @uparams: User supplied parameters 140 * @param: Parameters passed to cat_test() 141 * @span: Buffer size for the benchmark 142 * @current_mask Start mask for the first iteration 143 * 144 * Run CAT selftest by varying the allocated cache portion and comparing the 145 * impact on cache misses (the result analysis is done in check_results() 146 * and show_results_info(), not in this function). 147 * 148 * One bit is removed from the CAT allocation bit mask (in current_mask) for 149 * each subsequent test which keeps reducing the size of the allocated cache 150 * portion. A single test flushes the buffer, reads it to warm up the cache, 151 * and reads the buffer again. The cache misses are measured during the last 152 * read pass. 153 * 154 * Return: 0 when the test was run, < 0 on error. 155 */ 156static int cat_test(const struct resctrl_test *test, 157 const struct user_params *uparams, 158 struct resctrl_val_param *param, 159 size_t span, unsigned long current_mask) 160{ 161 char *resctrl_val = param->resctrl_val; 162 struct perf_event_read pe_read; 163 struct perf_event_attr pea; 164 cpu_set_t old_affinity; 165 unsigned char *buf; 166 char schemata[64]; 167 int ret, i, pe_fd; 168 pid_t bm_pid; 169 170 if (strcmp(param->filename, "") == 0) 171 sprintf(param->filename, "stdio"); 172 173 bm_pid = getpid(); 174 175 /* Taskset benchmark to specified cpu */ 176 ret = taskset_benchmark(bm_pid, uparams->cpu, &old_affinity); 177 if (ret) 178 return ret; 179 180 /* Write benchmark to specified con_mon grp, mon_grp in resctrl FS*/ 181 ret = write_bm_pid_to_resctrl(bm_pid, param->ctrlgrp, param->mongrp, 182 resctrl_val); 183 if (ret) 184 goto reset_affinity; 185 186 perf_event_attr_initialize(&pea, PERF_COUNT_HW_CACHE_MISSES); 187 perf_event_initialize_read_format(&pe_read); 188 pe_fd = perf_open(&pea, bm_pid, uparams->cpu); 189 if (pe_fd < 0) { 190 ret = -1; 191 goto reset_affinity; 192 } 193 194 buf = alloc_buffer(span, 1); 195 if (!buf) { 196 ret = -1; 197 goto pe_close; 198 } 199 200 while (current_mask) { 201 snprintf(schemata, sizeof(schemata), "%lx", param->mask & ~current_mask); 202 ret = write_schemata("", schemata, uparams->cpu, test->resource); 203 if (ret) 204 goto free_buf; 205 snprintf(schemata, sizeof(schemata), "%lx", current_mask); 206 ret = write_schemata(param->ctrlgrp, schemata, uparams->cpu, test->resource); 207 if (ret) 208 goto free_buf; 209 210 for (i = 0; i < NUM_OF_RUNS; i++) { 211 mem_flush(buf, span); 212 fill_cache_read(buf, span, true); 213 214 ret = perf_event_reset_enable(pe_fd); 215 if (ret) 216 goto free_buf; 217 218 fill_cache_read(buf, span, true); 219 220 ret = perf_event_measure(pe_fd, &pe_read, param->filename, bm_pid); 221 if (ret) 222 goto free_buf; 223 } 224 current_mask = next_mask(current_mask); 225 } 226 227free_buf: 228 free(buf); 229pe_close: 230 close(pe_fd); 231reset_affinity: 232 taskset_restore(bm_pid, &old_affinity); 233 234 return ret; 235} 236 237static int cat_run_test(const struct resctrl_test *test, const struct user_params *uparams) 238{ 239 unsigned long long_mask, start_mask, full_cache_mask; 240 unsigned long cache_total_size = 0; 241 int n = uparams->bits; 242 unsigned int start; 243 int count_of_bits; 244 size_t span; 245 int ret; 246 247 ret = get_full_cbm(test->resource, &full_cache_mask); 248 if (ret) 249 return ret; 250 /* Get the largest contiguous exclusive portion of the cache */ 251 ret = get_mask_no_shareable(test->resource, &long_mask); 252 if (ret) 253 return ret; 254 255 /* Get L3/L2 cache size */ 256 ret = get_cache_size(uparams->cpu, test->resource, &cache_total_size); 257 if (ret) 258 return ret; 259 ksft_print_msg("Cache size :%lu\n", cache_total_size); 260 261 count_of_bits = count_contiguous_bits(long_mask, &start); 262 263 if (!n) 264 n = count_of_bits / 2; 265 266 if (n > count_of_bits - 1) { 267 ksft_print_msg("Invalid input value for no_of_bits n!\n"); 268 ksft_print_msg("Please enter value in range 1 to %d\n", 269 count_of_bits - 1); 270 return -1; 271 } 272 start_mask = create_bit_mask(start, n); 273 274 struct resctrl_val_param param = { 275 .resctrl_val = CAT_STR, 276 .ctrlgrp = "c1", 277 .filename = RESULT_FILE_NAME, 278 .num_of_runs = 0, 279 }; 280 param.mask = long_mask; 281 span = cache_portion_size(cache_total_size, start_mask, full_cache_mask); 282 283 remove(param.filename); 284 285 ret = cat_test(test, uparams, ¶m, span, start_mask); 286 if (ret) 287 goto out; 288 289 ret = check_results(¶m, test->resource, 290 cache_total_size, full_cache_mask, start_mask); 291out: 292 cat_test_cleanup(); 293 294 return ret; 295} 296 297static int noncont_cat_run_test(const struct resctrl_test *test, 298 const struct user_params *uparams) 299{ 300 unsigned long full_cache_mask, cont_mask, noncont_mask; 301 unsigned int eax, ebx, ecx, edx, sparse_masks; 302 int bit_center, ret; 303 char schemata[64]; 304 305 /* Check to compare sparse_masks content to CPUID output. */ 306 ret = resource_info_unsigned_get(test->resource, "sparse_masks", &sparse_masks); 307 if (ret) 308 return ret; 309 310 if (!strcmp(test->resource, "L3")) 311 __cpuid_count(0x10, 1, eax, ebx, ecx, edx); 312 else if (!strcmp(test->resource, "L2")) 313 __cpuid_count(0x10, 2, eax, ebx, ecx, edx); 314 else 315 return -EINVAL; 316 317 if (sparse_masks != ((ecx >> 3) & 1)) { 318 ksft_print_msg("CPUID output doesn't match 'sparse_masks' file content!\n"); 319 return 1; 320 } 321 322 /* Write checks initialization. */ 323 ret = get_full_cbm(test->resource, &full_cache_mask); 324 if (ret < 0) 325 return ret; 326 bit_center = count_bits(full_cache_mask) / 2; 327 328 /* 329 * The bit_center needs to be at least 3 to properly calculate the CBM 330 * hole in the noncont_mask. If it's smaller return an error since the 331 * cache mask is too short and that shouldn't happen. 332 */ 333 if (bit_center < 3) 334 return -EINVAL; 335 cont_mask = full_cache_mask >> bit_center; 336 337 /* Contiguous mask write check. */ 338 snprintf(schemata, sizeof(schemata), "%lx", cont_mask); 339 ret = write_schemata("", schemata, uparams->cpu, test->resource); 340 if (ret) { 341 ksft_print_msg("Write of contiguous CBM failed\n"); 342 return 1; 343 } 344 345 /* 346 * Non-contiguous mask write check. CBM has a 0xf hole approximately in the middle. 347 * Output is compared with support information to catch any edge case errors. 348 */ 349 noncont_mask = ~(0xfUL << (bit_center - 2)) & full_cache_mask; 350 snprintf(schemata, sizeof(schemata), "%lx", noncont_mask); 351 ret = write_schemata("", schemata, uparams->cpu, test->resource); 352 if (ret && sparse_masks) 353 ksft_print_msg("Non-contiguous CBMs supported but write of non-contiguous CBM failed\n"); 354 else if (ret && !sparse_masks) 355 ksft_print_msg("Non-contiguous CBMs not supported and write of non-contiguous CBM failed as expected\n"); 356 else if (!ret && !sparse_masks) 357 ksft_print_msg("Non-contiguous CBMs not supported but write of non-contiguous CBM succeeded\n"); 358 359 return !ret == !sparse_masks; 360} 361 362static bool noncont_cat_feature_check(const struct resctrl_test *test) 363{ 364 if (!resctrl_resource_exists(test->resource)) 365 return false; 366 367 return resource_info_file_exists(test->resource, "sparse_masks"); 368} 369 370struct resctrl_test l3_cat_test = { 371 .name = "L3_CAT", 372 .group = "CAT", 373 .resource = "L3", 374 .feature_check = test_resource_feature_check, 375 .run_test = cat_run_test, 376}; 377 378struct resctrl_test l3_noncont_cat_test = { 379 .name = "L3_NONCONT_CAT", 380 .group = "CAT", 381 .resource = "L3", 382 .feature_check = noncont_cat_feature_check, 383 .run_test = noncont_cat_run_test, 384}; 385 386struct resctrl_test l2_noncont_cat_test = { 387 .name = "L2_NONCONT_CAT", 388 .group = "CAT", 389 .resource = "L2", 390 .feature_check = noncont_cat_feature_check, 391 .run_test = noncont_cat_run_test, 392}; 393