1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2022 ARM Limited. 4 */ 5 6 #define _GNU_SOURCE 7 #define _POSIX_C_SOURCE 199309L 8 9 #include <errno.h> 10 #include <getopt.h> 11 #include <poll.h> 12 #include <signal.h> 13 #include <stdbool.h> 14 #include <stddef.h> 15 #include <stdio.h> 16 #include <stdlib.h> 17 #include <string.h> 18 #include <unistd.h> 19 #include <sys/auxv.h> 20 #include <sys/epoll.h> 21 #include <sys/prctl.h> 22 #include <sys/types.h> 23 #include <sys/uio.h> 24 #include <sys/wait.h> 25 #include <asm/hwcap.h> 26 27 #include "../../kselftest.h" 28 29 #define MAX_VLS 16 30 31 struct child_data { 32 char *name, *output; 33 pid_t pid; 34 int stdout; 35 bool output_seen; 36 bool exited; 37 int exit_status; 38 }; 39 40 static int epoll_fd; 41 static struct child_data *children; 42 static struct epoll_event *evs; 43 static int tests; 44 static int num_children; 45 static bool terminate; 46 47 static int startup_pipe[2]; 48 49 static int num_processors(void) 50 { 51 long nproc = sysconf(_SC_NPROCESSORS_CONF); 52 if (nproc < 0) { 53 perror("Unable to read number of processors\n"); 54 exit(EXIT_FAILURE); 55 } 56 57 return nproc; 58 } 59 60 static void child_start(struct child_data *child, const char *program) 61 { 62 int ret, pipefd[2], i; 63 struct epoll_event ev; 64 65 ret = pipe(pipefd); 66 if (ret != 0) 67 ksft_exit_fail_msg("Failed to create stdout pipe: %s (%d)\n", 68 strerror(errno), errno); 69 70 child->pid = fork(); 71 if (child->pid == -1) 72 ksft_exit_fail_msg("fork() failed: %s (%d)\n", 73 strerror(errno), errno); 74 75 if (!child->pid) { 76 /* 77 * In child, replace stdout with the pipe, errors to 78 * stderr from here as kselftest prints to stdout. 79 */ 80 ret = dup2(pipefd[1], 1); 81 if (ret == -1) { 82 fprintf(stderr, "dup2() %d\n", errno); 83 exit(EXIT_FAILURE); 84 } 85 86 /* 87 * Duplicate the read side of the startup pipe to 88 * FD 3 so we can close everything else. 89 */ 90 ret = dup2(startup_pipe[0], 3); 91 if (ret == -1) { 92 fprintf(stderr, "dup2() %d\n", errno); 93 exit(EXIT_FAILURE); 94 } 95 96 /* 97 * Very dumb mechanism to clean open FDs other than 98 * stdio. We don't want O_CLOEXEC for the pipes... 99 */ 100 for (i = 4; i < 8192; i++) 101 close(i); 102 103 /* 104 * Read from the startup pipe, there should be no data 105 * and we should block until it is closed. We just 106 * carry on on error since this isn't super critical. 107 */ 108 ret = read(3, &i, sizeof(i)); 109 if (ret < 0) 110 fprintf(stderr, "read(startp pipe) failed: %s (%d)\n", 111 strerror(errno), errno); 112 if (ret > 0) 113 fprintf(stderr, "%d bytes of data on startup pipe\n", 114 ret); 115 close(3); 116 117 ret = execl(program, program, NULL); 118 fprintf(stderr, "execl(%s) failed: %d (%s)\n", 119 program, errno, strerror(errno)); 120 121 exit(EXIT_FAILURE); 122 } else { 123 /* 124 * In parent, remember the child and close our copy of the 125 * write side of stdout. 126 */ 127 close(pipefd[1]); 128 child->stdout = pipefd[0]; 129 child->output = NULL; 130 child->exited = false; 131 child->output_seen = false; 132 133 ev.events = EPOLLIN | EPOLLHUP; 134 ev.data.ptr = child; 135 136 ret = epoll_ctl(epoll_fd, EPOLL_CTL_ADD, child->stdout, &ev); 137 if (ret < 0) { 138 ksft_exit_fail_msg("%s EPOLL_CTL_ADD failed: %s (%d)\n", 139 child->name, strerror(errno), errno); 140 } 141 } 142 } 143 144 static bool child_output_read(struct child_data *child) 145 { 146 char read_data[1024]; 147 char work[1024]; 148 int ret, len, cur_work, cur_read; 149 150 ret = read(child->stdout, read_data, sizeof(read_data)); 151 if (ret < 0) { 152 if (errno == EINTR) 153 return true; 154 155 ksft_print_msg("%s: read() failed: %s (%d)\n", 156 child->name, strerror(errno), 157 errno); 158 return false; 159 } 160 len = ret; 161 162 child->output_seen = true; 163 164 /* Pick up any partial read */ 165 if (child->output) { 166 strncpy(work, child->output, sizeof(work) - 1); 167 cur_work = strnlen(work, sizeof(work)); 168 free(child->output); 169 child->output = NULL; 170 } else { 171 cur_work = 0; 172 } 173 174 cur_read = 0; 175 while (cur_read < len) { 176 work[cur_work] = read_data[cur_read++]; 177 178 if (work[cur_work] == '\n') { 179 work[cur_work] = '\0'; 180 ksft_print_msg("%s: %s\n", child->name, work); 181 cur_work = 0; 182 } else { 183 cur_work++; 184 } 185 } 186 187 if (cur_work) { 188 work[cur_work] = '\0'; 189 ret = asprintf(&child->output, "%s", work); 190 if (ret == -1) 191 ksft_exit_fail_msg("Out of memory\n"); 192 } 193 194 return false; 195 } 196 197 static void child_output(struct child_data *child, uint32_t events, 198 bool flush) 199 { 200 bool read_more; 201 202 if (events & EPOLLIN) { 203 do { 204 read_more = child_output_read(child); 205 } while (read_more); 206 } 207 208 if (events & EPOLLHUP) { 209 close(child->stdout); 210 child->stdout = -1; 211 flush = true; 212 } 213 214 if (flush && child->output) { 215 ksft_print_msg("%s: %s<EOF>\n", child->name, child->output); 216 free(child->output); 217 child->output = NULL; 218 } 219 } 220 221 static void child_tickle(struct child_data *child) 222 { 223 if (child->output_seen && !child->exited) 224 kill(child->pid, SIGUSR2); 225 } 226 227 static void child_stop(struct child_data *child) 228 { 229 if (!child->exited) 230 kill(child->pid, SIGTERM); 231 } 232 233 static void child_cleanup(struct child_data *child) 234 { 235 pid_t ret; 236 int status; 237 bool fail = false; 238 239 if (!child->exited) { 240 do { 241 ret = waitpid(child->pid, &status, 0); 242 if (ret == -1 && errno == EINTR) 243 continue; 244 245 if (ret == -1) { 246 ksft_print_msg("waitpid(%d) failed: %s (%d)\n", 247 child->pid, strerror(errno), 248 errno); 249 fail = true; 250 break; 251 } 252 } while (!WIFEXITED(status)); 253 child->exit_status = WEXITSTATUS(status); 254 } 255 256 if (!child->output_seen) { 257 ksft_print_msg("%s no output seen\n", child->name); 258 fail = true; 259 } 260 261 if (child->exit_status != 0) { 262 ksft_print_msg("%s exited with error code %d\n", 263 child->name, child->exit_status); 264 fail = true; 265 } 266 267 ksft_test_result(!fail, "%s\n", child->name); 268 } 269 270 static void handle_child_signal(int sig, siginfo_t *info, void *context) 271 { 272 int i; 273 bool found = false; 274 275 for (i = 0; i < num_children; i++) { 276 if (children[i].pid == info->si_pid) { 277 children[i].exited = true; 278 children[i].exit_status = info->si_status; 279 found = true; 280 break; 281 } 282 } 283 284 if (!found) 285 ksft_print_msg("SIGCHLD for unknown PID %d with status %d\n", 286 info->si_pid, info->si_status); 287 } 288 289 static void handle_exit_signal(int sig, siginfo_t *info, void *context) 290 { 291 int i; 292 293 /* If we're already exiting then don't signal again */ 294 if (terminate) 295 return; 296 297 ksft_print_msg("Got signal, exiting...\n"); 298 299 terminate = true; 300 301 /* 302 * This should be redundant, the main loop should clean up 303 * after us, but for safety stop everything we can here. 304 */ 305 for (i = 0; i < num_children; i++) 306 child_stop(&children[i]); 307 } 308 309 static void start_fpsimd(struct child_data *child, int cpu, int copy) 310 { 311 int ret; 312 313 ret = asprintf(&child->name, "FPSIMD-%d-%d", cpu, copy); 314 if (ret == -1) 315 ksft_exit_fail_msg("asprintf() failed\n"); 316 317 child_start(child, "./fpsimd-test"); 318 319 ksft_print_msg("Started %s\n", child->name); 320 } 321 322 static void start_sve(struct child_data *child, int vl, int cpu) 323 { 324 int ret; 325 326 ret = prctl(PR_SVE_SET_VL, vl | PR_SVE_VL_INHERIT); 327 if (ret < 0) 328 ksft_exit_fail_msg("Failed to set SVE VL %d\n", vl); 329 330 ret = asprintf(&child->name, "SVE-VL-%d-%d", vl, cpu); 331 if (ret == -1) 332 ksft_exit_fail_msg("asprintf() failed\n"); 333 334 child_start(child, "./sve-test"); 335 336 ksft_print_msg("Started %s\n", child->name); 337 } 338 339 static void start_ssve(struct child_data *child, int vl, int cpu) 340 { 341 int ret; 342 343 ret = asprintf(&child->name, "SSVE-VL-%d-%d", vl, cpu); 344 if (ret == -1) 345 ksft_exit_fail_msg("asprintf() failed\n"); 346 347 ret = prctl(PR_SME_SET_VL, vl | PR_SME_VL_INHERIT); 348 if (ret < 0) 349 ksft_exit_fail_msg("Failed to set SME VL %d\n", ret); 350 351 child_start(child, "./ssve-test"); 352 353 ksft_print_msg("Started %s\n", child->name); 354 } 355 356 static void start_za(struct child_data *child, int vl, int cpu) 357 { 358 int ret; 359 360 ret = prctl(PR_SME_SET_VL, vl | PR_SVE_VL_INHERIT); 361 if (ret < 0) 362 ksft_exit_fail_msg("Failed to set SME VL %d\n", ret); 363 364 ret = asprintf(&child->name, "ZA-VL-%d-%d", vl, cpu); 365 if (ret == -1) 366 ksft_exit_fail_msg("asprintf() failed\n"); 367 368 child_start(child, "./za-test"); 369 370 ksft_print_msg("Started %s\n", child->name); 371 } 372 373 static void start_zt(struct child_data *child, int cpu) 374 { 375 int ret; 376 377 ret = asprintf(&child->name, "ZT-%d", cpu); 378 if (ret == -1) 379 ksft_exit_fail_msg("asprintf() failed\n"); 380 381 child_start(child, "./zt-test"); 382 383 ksft_print_msg("Started %s\n", child->name); 384 } 385 386 static void probe_vls(int vls[], int *vl_count, int set_vl) 387 { 388 unsigned int vq; 389 int vl; 390 391 *vl_count = 0; 392 393 for (vq = SVE_VQ_MAX; vq > 0; vq /= 2) { 394 vl = prctl(set_vl, vq * 16); 395 if (vl == -1) 396 ksft_exit_fail_msg("SET_VL failed: %s (%d)\n", 397 strerror(errno), errno); 398 399 vl &= PR_SVE_VL_LEN_MASK; 400 401 if (*vl_count && (vl == vls[*vl_count - 1])) 402 break; 403 404 vq = sve_vq_from_vl(vl); 405 406 vls[*vl_count] = vl; 407 *vl_count += 1; 408 } 409 } 410 411 /* Handle any pending output without blocking */ 412 static void drain_output(bool flush) 413 { 414 int ret = 1; 415 int i; 416 417 while (ret > 0) { 418 ret = epoll_wait(epoll_fd, evs, tests, 0); 419 if (ret < 0) { 420 if (errno == EINTR) 421 continue; 422 ksft_print_msg("epoll_wait() failed: %s (%d)\n", 423 strerror(errno), errno); 424 } 425 426 for (i = 0; i < ret; i++) 427 child_output(evs[i].data.ptr, evs[i].events, flush); 428 } 429 } 430 431 static const struct option options[] = { 432 { "timeout", required_argument, NULL, 't' }, 433 { } 434 }; 435 436 int main(int argc, char **argv) 437 { 438 int ret; 439 int timeout = 10; 440 int cpus, i, j, c; 441 int sve_vl_count, sme_vl_count, fpsimd_per_cpu; 442 bool all_children_started = false; 443 int seen_children; 444 int sve_vls[MAX_VLS], sme_vls[MAX_VLS]; 445 bool have_sme2; 446 struct sigaction sa; 447 448 while ((c = getopt_long(argc, argv, "t:", options, NULL)) != -1) { 449 switch (c) { 450 case 't': 451 ret = sscanf(optarg, "%d", &timeout); 452 if (ret != 1) 453 ksft_exit_fail_msg("Failed to parse timeout %s\n", 454 optarg); 455 break; 456 default: 457 ksft_exit_fail_msg("Unknown argument\n"); 458 } 459 } 460 461 cpus = num_processors(); 462 tests = 0; 463 464 if (getauxval(AT_HWCAP) & HWCAP_SVE) { 465 probe_vls(sve_vls, &sve_vl_count, PR_SVE_SET_VL); 466 tests += sve_vl_count * cpus; 467 } else { 468 sve_vl_count = 0; 469 } 470 471 if (getauxval(AT_HWCAP2) & HWCAP2_SME) { 472 probe_vls(sme_vls, &sme_vl_count, PR_SME_SET_VL); 473 tests += sme_vl_count * cpus * 2; 474 } else { 475 sme_vl_count = 0; 476 } 477 478 if (getauxval(AT_HWCAP2) & HWCAP2_SME2) { 479 tests += cpus; 480 have_sme2 = true; 481 } else { 482 have_sme2 = false; 483 } 484 485 /* Force context switching if we only have FPSIMD */ 486 if (!sve_vl_count && !sme_vl_count) 487 fpsimd_per_cpu = 2; 488 else 489 fpsimd_per_cpu = 1; 490 tests += cpus * fpsimd_per_cpu; 491 492 ksft_print_header(); 493 ksft_set_plan(tests); 494 495 ksft_print_msg("%d CPUs, %d SVE VLs, %d SME VLs, SME2 %s\n", 496 cpus, sve_vl_count, sme_vl_count, 497 have_sme2 ? "present" : "absent"); 498 499 if (timeout > 0) 500 ksft_print_msg("Will run for %ds\n", timeout); 501 else 502 ksft_print_msg("Will run until terminated\n"); 503 504 children = calloc(sizeof(*children), tests); 505 if (!children) 506 ksft_exit_fail_msg("Unable to allocate child data\n"); 507 508 ret = epoll_create1(EPOLL_CLOEXEC); 509 if (ret < 0) 510 ksft_exit_fail_msg("epoll_create1() failed: %s (%d)\n", 511 strerror(errno), ret); 512 epoll_fd = ret; 513 514 /* Create a pipe which children will block on before execing */ 515 ret = pipe(startup_pipe); 516 if (ret != 0) 517 ksft_exit_fail_msg("Failed to create startup pipe: %s (%d)\n", 518 strerror(errno), errno); 519 520 /* Get signal handers ready before we start any children */ 521 memset(&sa, 0, sizeof(sa)); 522 sa.sa_sigaction = handle_exit_signal; 523 sa.sa_flags = SA_RESTART | SA_SIGINFO; 524 sigemptyset(&sa.sa_mask); 525 ret = sigaction(SIGINT, &sa, NULL); 526 if (ret < 0) 527 ksft_print_msg("Failed to install SIGINT handler: %s (%d)\n", 528 strerror(errno), errno); 529 ret = sigaction(SIGTERM, &sa, NULL); 530 if (ret < 0) 531 ksft_print_msg("Failed to install SIGTERM handler: %s (%d)\n", 532 strerror(errno), errno); 533 sa.sa_sigaction = handle_child_signal; 534 ret = sigaction(SIGCHLD, &sa, NULL); 535 if (ret < 0) 536 ksft_print_msg("Failed to install SIGCHLD handler: %s (%d)\n", 537 strerror(errno), errno); 538 539 evs = calloc(tests, sizeof(*evs)); 540 if (!evs) 541 ksft_exit_fail_msg("Failed to allocated %d epoll events\n", 542 tests); 543 544 for (i = 0; i < cpus; i++) { 545 for (j = 0; j < fpsimd_per_cpu; j++) 546 start_fpsimd(&children[num_children++], i, j); 547 548 for (j = 0; j < sve_vl_count; j++) 549 start_sve(&children[num_children++], sve_vls[j], i); 550 551 for (j = 0; j < sme_vl_count; j++) { 552 start_ssve(&children[num_children++], sme_vls[j], i); 553 start_za(&children[num_children++], sme_vls[j], i); 554 } 555 556 if (have_sme2) 557 start_zt(&children[num_children++], i); 558 } 559 560 /* 561 * All children started, close the startup pipe and let them 562 * run. 563 */ 564 close(startup_pipe[0]); 565 close(startup_pipe[1]); 566 567 for (;;) { 568 /* Did we get a signal asking us to exit? */ 569 if (terminate) 570 break; 571 572 /* 573 * Timeout is counted in seconds with no output, the 574 * tests print during startup then are silent when 575 * running so this should ensure they all ran enough 576 * to install the signal handler, this is especially 577 * useful in emulation where we will both be slow and 578 * likely to have a large set of VLs. 579 */ 580 ret = epoll_wait(epoll_fd, evs, tests, 1000); 581 if (ret < 0) { 582 if (errno == EINTR) 583 continue; 584 ksft_exit_fail_msg("epoll_wait() failed: %s (%d)\n", 585 strerror(errno), errno); 586 } 587 588 /* Output? */ 589 if (ret > 0) { 590 for (i = 0; i < ret; i++) { 591 child_output(evs[i].data.ptr, evs[i].events, 592 false); 593 } 594 continue; 595 } 596 597 /* Otherwise epoll_wait() timed out */ 598 599 /* 600 * If the child processes have not produced output they 601 * aren't actually running the tests yet . 602 */ 603 if (!all_children_started) { 604 seen_children = 0; 605 606 for (i = 0; i < num_children; i++) 607 if (children[i].output_seen || 608 children[i].exited) 609 seen_children++; 610 611 if (seen_children != num_children) { 612 ksft_print_msg("Waiting for %d children\n", 613 num_children - seen_children); 614 continue; 615 } 616 617 all_children_started = true; 618 } 619 620 ksft_print_msg("Sending signals, timeout remaining: %d\n", 621 timeout); 622 623 for (i = 0; i < num_children; i++) 624 child_tickle(&children[i]); 625 626 /* Negative timeout means run indefinitely */ 627 if (timeout < 0) 628 continue; 629 if (--timeout == 0) 630 break; 631 } 632 633 ksft_print_msg("Finishing up...\n"); 634 terminate = true; 635 636 for (i = 0; i < tests; i++) 637 child_stop(&children[i]); 638 639 drain_output(false); 640 641 for (i = 0; i < tests; i++) 642 child_cleanup(&children[i]); 643 644 drain_output(true); 645 646 ksft_print_cnts(); 647 648 return 0; 649 } 650