1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (C) 2021 Red Hat Inc, Daniel Bristot de Oliveira <bristot@kernel.org> 4 */ 5 6 #define _GNU_SOURCE 7 #ifdef HAVE_LIBCPUPOWER_SUPPORT 8 #include <cpuidle.h> 9 #endif /* HAVE_LIBCPUPOWER_SUPPORT */ 10 #include <dirent.h> 11 #include <stdarg.h> 12 #include <stdlib.h> 13 #include <string.h> 14 #include <unistd.h> 15 #include <ctype.h> 16 #include <errno.h> 17 #include <fcntl.h> 18 #include <sched.h> 19 #include <stdio.h> 20 #include <limits.h> 21 22 #include "common.h" 23 24 #define MAX_MSG_LENGTH 1024 25 int config_debug; 26 27 /* 28 * err_msg - print an error message to the stderr 29 */ 30 void err_msg(const char *fmt, ...) 31 { 32 char message[MAX_MSG_LENGTH]; 33 va_list ap; 34 35 va_start(ap, fmt); 36 vsnprintf(message, sizeof(message), fmt, ap); 37 va_end(ap); 38 39 fprintf(stderr, "%s", message); 40 } 41 42 /* 43 * debug_msg - print a debug message to stderr if debug is set 44 */ 45 void debug_msg(const char *fmt, ...) 46 { 47 char message[MAX_MSG_LENGTH]; 48 va_list ap; 49 50 if (!config_debug) 51 return; 52 53 va_start(ap, fmt); 54 vsnprintf(message, sizeof(message), fmt, ap); 55 va_end(ap); 56 57 fprintf(stderr, "%s", message); 58 } 59 60 /* 61 * fatal - print an error message and EOL to stderr and exit with ERROR 62 */ 63 void fatal(const char *fmt, ...) 64 { 65 va_list ap; 66 67 va_start(ap, fmt); 68 vfprintf(stderr, fmt, ap); 69 va_end(ap); 70 fprintf(stderr, "\n"); 71 72 exit(ERROR); 73 } 74 75 /* 76 * get_llong_from_str - get a long long int from a string 77 */ 78 long long get_llong_from_str(char *start) 79 { 80 long long value; 81 char *end; 82 83 errno = 0; 84 value = strtoll(start, &end, 10); 85 if (errno || start == end) 86 return -1; 87 88 return value; 89 } 90 91 /* 92 * get_duration - fill output with a human readable duration since start_time 93 */ 94 void get_duration(time_t start_time, char *output, int output_size) 95 { 96 time_t now = time(NULL); 97 struct tm *tm_info; 98 time_t duration; 99 100 duration = difftime(now, start_time); 101 tm_info = gmtime(&duration); 102 103 snprintf(output, output_size, "%3d %02d:%02d:%02d", 104 tm_info->tm_yday, 105 tm_info->tm_hour, 106 tm_info->tm_min, 107 tm_info->tm_sec); 108 } 109 110 /* 111 * parse_cpu_set - parse a cpu_list filling cpu_set_t argument 112 * 113 * Receives a cpu list, like 1-3,5 (cpus 1, 2, 3, 5), and then set 114 * filling cpu_set_t argument. 115 * 116 * Returns 0 on success, 1 otherwise. 117 */ 118 int parse_cpu_set(char *cpu_list, cpu_set_t *set) 119 { 120 const char *p; 121 int end_cpu; 122 int cpu; 123 int i; 124 125 CPU_ZERO(set); 126 127 for (p = cpu_list; *p; ) { 128 cpu = atoi(p); 129 if (cpu < 0 || (!cpu && *p != '0') || cpu >= nr_cpus) 130 goto err; 131 132 while (isdigit(*p)) 133 p++; 134 if (*p == '-') { 135 p++; 136 end_cpu = atoi(p); 137 if (end_cpu < cpu || (!end_cpu && *p != '0') || end_cpu >= nr_cpus) 138 goto err; 139 while (isdigit(*p)) 140 p++; 141 } else 142 end_cpu = cpu; 143 144 if (cpu == end_cpu) { 145 debug_msg("cpu_set: adding cpu %d\n", cpu); 146 CPU_SET(cpu, set); 147 } else { 148 for (i = cpu; i <= end_cpu; i++) { 149 debug_msg("cpu_set: adding cpu %d\n", i); 150 CPU_SET(i, set); 151 } 152 } 153 154 if (*p == ',') 155 p++; 156 } 157 158 return 0; 159 err: 160 debug_msg("Error parsing the cpu set %s\n", cpu_list); 161 return 1; 162 } 163 164 /* 165 * parse_stack_format - parse the stack format 166 * 167 * Return: the stack format on success, -1 otherwise. 168 */ 169 int parse_stack_format(char *arg) 170 { 171 if (!strcmp(arg, "truncate")) 172 return STACK_FORMAT_TRUNCATE; 173 if (!strcmp(arg, "skip")) 174 return STACK_FORMAT_SKIP; 175 if (!strcmp(arg, "full")) 176 return STACK_FORMAT_FULL; 177 178 debug_msg("Error parsing the stack format %s\n", arg); 179 return -1; 180 } 181 182 /* 183 * parse_duration - parse duration with s/m/h/d suffix converting it to seconds 184 */ 185 long parse_seconds_duration(char *val) 186 { 187 char *end; 188 long t; 189 190 t = strtol(val, &end, 10); 191 192 if (end) { 193 switch (*end) { 194 case 's': 195 case 'S': 196 break; 197 case 'm': 198 case 'M': 199 t *= 60; 200 break; 201 case 'h': 202 case 'H': 203 t *= 60 * 60; 204 break; 205 206 case 'd': 207 case 'D': 208 t *= 24 * 60 * 60; 209 break; 210 } 211 } 212 213 return t; 214 } 215 216 /* 217 * match_time_unit - check if str starts with unit followed by end-of-string or ':' 218 * 219 * This allows the time unit parser to work both in standalone duration strings 220 * like "100ms" and in colon-delimited SCHED_DEADLINE specifications like 221 * "d:10ms:100ms", while still rejecting malformed input like "100msx". 222 */ 223 static bool match_time_unit(const char *str, const char *unit) 224 { 225 size_t len = strlen(unit); 226 227 return strncmp(str, unit, len) == 0 && 228 (str[len] == '\0' || str[len] == ':'); 229 } 230 231 /* 232 * parse_ns_duration - parse duration with ns/us/ms/s converting it to nanoseconds 233 */ 234 long parse_ns_duration(char *val) 235 { 236 char *end; 237 long t; 238 239 t = strtol(val, &end, 10); 240 241 if (end) { 242 if (match_time_unit(end, "ns")) { 243 return t; 244 } else if (match_time_unit(end, "us")) { 245 t *= 1000; 246 return t; 247 } else if (match_time_unit(end, "ms")) { 248 t *= 1000 * 1000; 249 return t; 250 } else if (match_time_unit(end, "s")) { 251 t *= 1000 * 1000 * 1000; 252 return t; 253 } 254 return -1; 255 } 256 257 return t; 258 } 259 260 /* 261 * This is a set of helper functions to use SCHED_DEADLINE. 262 */ 263 #ifndef __NR_sched_setattr 264 # ifdef __x86_64__ 265 # define __NR_sched_setattr 314 266 # elif __i386__ 267 # define __NR_sched_setattr 351 268 # elif __arm__ 269 # define __NR_sched_setattr 380 270 # elif __aarch64__ || __riscv 271 # define __NR_sched_setattr 274 272 # elif __powerpc__ 273 # define __NR_sched_setattr 355 274 # elif __s390x__ 275 # define __NR_sched_setattr 345 276 # elif __loongarch__ 277 # define __NR_sched_setattr 274 278 # endif 279 #endif 280 281 #define SCHED_DEADLINE 6 282 283 static inline int syscall_sched_setattr(pid_t pid, const struct sched_attr *attr, 284 unsigned int flags) { 285 return syscall(__NR_sched_setattr, pid, attr, flags); 286 } 287 288 int __set_sched_attr(int pid, struct sched_attr *attr) 289 { 290 int flags = 0; 291 int retval; 292 293 retval = syscall_sched_setattr(pid, attr, flags); 294 if (retval < 0) { 295 err_msg("Failed to set sched attributes to the pid %d: %s\n", 296 pid, strerror(errno)); 297 return 1; 298 } 299 300 return 0; 301 } 302 303 /* 304 * procfs_is_workload_pid - check if a procfs entry contains a comm_prefix* comm 305 * 306 * Check if the procfs entry is a directory of a process, and then check if the 307 * process has a comm with the prefix set in char *comm_prefix. As the 308 * current users of this function only check for kernel threads, there is no 309 * need to check for the threads for the process. 310 * 311 * Return: True if the proc_entry contains a comm file with comm_prefix*. 312 * Otherwise returns false. 313 */ 314 static int procfs_is_workload_pid(const char *comm_prefix, struct dirent *proc_entry) 315 { 316 char buffer[MAX_PATH]; 317 int comm_fd, retval; 318 char *t_name; 319 320 if (proc_entry->d_type != DT_DIR) 321 return 0; 322 323 if (*proc_entry->d_name == '.') 324 return 0; 325 326 /* check if the string is a pid */ 327 for (t_name = proc_entry->d_name; t_name; t_name++) { 328 if (!isdigit(*t_name)) 329 break; 330 } 331 332 if (*t_name != '\0') 333 return 0; 334 335 snprintf(buffer, MAX_PATH, "/proc/%s/comm", proc_entry->d_name); 336 comm_fd = open(buffer, O_RDONLY); 337 if (comm_fd < 0) 338 return 0; 339 340 memset(buffer, 0, MAX_PATH); 341 retval = read(comm_fd, buffer, MAX_PATH); 342 343 close(comm_fd); 344 345 if (retval <= 0) 346 return 0; 347 348 buffer[MAX_PATH-1] = '\0'; 349 if (!str_has_prefix(buffer, comm_prefix)) 350 return 0; 351 352 /* comm already have \n */ 353 debug_msg("Found workload pid:%s comm:%s", proc_entry->d_name, buffer); 354 355 return 1; 356 } 357 358 /* 359 * set_comm_sched_attr - set sched params to threads starting with char *comm_prefix 360 * 361 * This function uses procfs to list the currently running threads and then set the 362 * sched_attr *attr to the threads that start with char *comm_prefix. It is 363 * mainly used to set the priority to the kernel threads created by the 364 * tracers. 365 */ 366 int set_comm_sched_attr(const char *comm_prefix, struct sched_attr *attr) 367 { 368 struct dirent *proc_entry; 369 DIR *procfs; 370 int retval; 371 int pid; 372 373 if (strlen(comm_prefix) >= MAX_PATH) { 374 err_msg("Command prefix is too long: %d < strlen(%s)\n", 375 MAX_PATH, comm_prefix); 376 return 1; 377 } 378 379 procfs = opendir("/proc"); 380 if (!procfs) { 381 err_msg("Could not open procfs\n"); 382 return 1; 383 } 384 385 while ((proc_entry = readdir(procfs))) { 386 387 retval = procfs_is_workload_pid(comm_prefix, proc_entry); 388 if (!retval) 389 continue; 390 391 if (strtoi(proc_entry->d_name, &pid)) { 392 err_msg("'%s' is not a valid pid", proc_entry->d_name); 393 goto out_err; 394 } 395 /* procfs_is_workload_pid confirmed it is a pid */ 396 retval = __set_sched_attr(pid, attr); 397 if (retval) { 398 err_msg("Error setting sched attributes for pid:%s\n", proc_entry->d_name); 399 goto out_err; 400 } 401 402 debug_msg("Set sched attributes for pid:%s\n", proc_entry->d_name); 403 } 404 return 0; 405 406 out_err: 407 closedir(procfs); 408 return 1; 409 } 410 411 #define INVALID_VAL (~0L) 412 static long get_long_ns_after_colon(char *start) 413 { 414 long val = INVALID_VAL; 415 416 /* find the ":" */ 417 start = strstr(start, ":"); 418 if (!start) 419 return -1; 420 421 /* skip ":" */ 422 start++; 423 val = parse_ns_duration(start); 424 425 return val; 426 } 427 428 static long get_long_after_colon(char *start) 429 { 430 long val = INVALID_VAL; 431 432 /* find the ":" */ 433 start = strstr(start, ":"); 434 if (!start) 435 return -1; 436 437 /* skip ":" */ 438 start++; 439 val = get_llong_from_str(start); 440 441 return val; 442 } 443 444 /* 445 * parse priority in the format: 446 * SCHED_OTHER: 447 * o:<prio> 448 * O:<prio> 449 * SCHED_RR: 450 * r:<prio> 451 * R:<prio> 452 * SCHED_FIFO: 453 * f:<prio> 454 * F:<prio> 455 * SCHED_DEADLINE: 456 * d:runtime:period 457 * D:runtime:period 458 */ 459 int parse_prio(char *arg, struct sched_attr *sched_param) 460 { 461 long prio; 462 long runtime; 463 long period; 464 465 memset(sched_param, 0, sizeof(*sched_param)); 466 sched_param->size = sizeof(*sched_param); 467 468 switch (arg[0]) { 469 case 'd': 470 case 'D': 471 /* d:runtime:period */ 472 if (strlen(arg) < 4) 473 return -1; 474 475 runtime = get_long_ns_after_colon(arg); 476 if (runtime == INVALID_VAL) 477 return -1; 478 479 period = get_long_ns_after_colon(&arg[2]); 480 if (period == INVALID_VAL) 481 return -1; 482 483 if (runtime > period) 484 return -1; 485 486 sched_param->sched_policy = SCHED_DEADLINE; 487 sched_param->sched_runtime = runtime; 488 sched_param->sched_deadline = period; 489 sched_param->sched_period = period; 490 break; 491 case 'f': 492 case 'F': 493 /* f:prio */ 494 prio = get_long_after_colon(arg); 495 if (prio == INVALID_VAL) 496 return -1; 497 498 if (prio < sched_get_priority_min(SCHED_FIFO)) 499 return -1; 500 if (prio > sched_get_priority_max(SCHED_FIFO)) 501 return -1; 502 503 sched_param->sched_policy = SCHED_FIFO; 504 sched_param->sched_priority = prio; 505 break; 506 case 'r': 507 case 'R': 508 /* r:prio */ 509 prio = get_long_after_colon(arg); 510 if (prio == INVALID_VAL) 511 return -1; 512 513 if (prio < sched_get_priority_min(SCHED_RR)) 514 return -1; 515 if (prio > sched_get_priority_max(SCHED_RR)) 516 return -1; 517 518 sched_param->sched_policy = SCHED_RR; 519 sched_param->sched_priority = prio; 520 break; 521 case 'o': 522 case 'O': 523 /* o:prio */ 524 prio = get_long_after_colon(arg); 525 if (prio == INVALID_VAL) 526 return -1; 527 528 if (prio < MIN_NICE) 529 return -1; 530 if (prio > MAX_NICE) 531 return -1; 532 533 sched_param->sched_policy = SCHED_OTHER; 534 sched_param->sched_nice = prio; 535 break; 536 default: 537 return -1; 538 } 539 return 0; 540 } 541 542 /* 543 * set_cpu_dma_latency - set the /dev/cpu_dma_latecy 544 * 545 * This is used to reduce the exit from idle latency. The value 546 * will be reset once the file descriptor of /dev/cpu_dma_latecy 547 * is closed. 548 * 549 * Return: the /dev/cpu_dma_latecy file descriptor 550 */ 551 int set_cpu_dma_latency(int32_t latency) 552 { 553 int retval; 554 int fd; 555 556 fd = open("/dev/cpu_dma_latency", O_RDWR); 557 if (fd < 0) { 558 err_msg("Error opening /dev/cpu_dma_latency\n"); 559 return -1; 560 } 561 562 retval = write(fd, &latency, 4); 563 if (retval < 1) { 564 err_msg("Error setting /dev/cpu_dma_latency\n"); 565 close(fd); 566 return -1; 567 } 568 569 debug_msg("Set /dev/cpu_dma_latency to %d\n", latency); 570 571 return fd; 572 } 573 574 #ifdef HAVE_LIBCPUPOWER_SUPPORT 575 static unsigned int **saved_cpu_idle_disable_state; 576 static size_t saved_cpu_idle_disable_state_alloc_ctr; 577 578 /* 579 * save_cpu_idle_state_disable - save disable for all idle states of a cpu 580 * 581 * Saves the current disable of all idle states of a cpu, to be subsequently 582 * restored via restore_cpu_idle_disable_state. 583 * 584 * Return: idle state count on success, negative on error 585 */ 586 int save_cpu_idle_disable_state(unsigned int cpu) 587 { 588 unsigned int nr_states; 589 unsigned int state; 590 int disabled; 591 592 nr_states = cpuidle_state_count(cpu); 593 594 if (nr_states == 0) 595 return 0; 596 597 if (saved_cpu_idle_disable_state == NULL) { 598 saved_cpu_idle_disable_state = calloc(nr_cpus, sizeof(unsigned int *)); 599 if (!saved_cpu_idle_disable_state) 600 return -1; 601 } 602 603 saved_cpu_idle_disable_state[cpu] = calloc(nr_states, sizeof(unsigned int)); 604 if (!saved_cpu_idle_disable_state[cpu]) 605 return -1; 606 saved_cpu_idle_disable_state_alloc_ctr++; 607 608 for (state = 0; state < nr_states; state++) { 609 disabled = cpuidle_is_state_disabled(cpu, state); 610 if (disabled < 0) 611 return disabled; 612 saved_cpu_idle_disable_state[cpu][state] = disabled; 613 } 614 615 return nr_states; 616 } 617 618 /* 619 * restore_cpu_idle_disable_state - restore disable for all idle states of a cpu 620 * 621 * Restores the current disable state of all idle states of a cpu that was 622 * previously saved by save_cpu_idle_disable_state. 623 * 624 * Return: idle state count on success, negative on error 625 */ 626 int restore_cpu_idle_disable_state(unsigned int cpu) 627 { 628 unsigned int nr_states; 629 unsigned int state; 630 int disabled; 631 int result; 632 633 nr_states = cpuidle_state_count(cpu); 634 635 if (nr_states == 0) 636 return 0; 637 638 if (!saved_cpu_idle_disable_state) 639 return -1; 640 641 for (state = 0; state < nr_states; state++) { 642 if (!saved_cpu_idle_disable_state[cpu]) 643 return -1; 644 disabled = saved_cpu_idle_disable_state[cpu][state]; 645 result = cpuidle_state_disable(cpu, state, disabled); 646 if (result < 0) 647 return result; 648 } 649 650 free(saved_cpu_idle_disable_state[cpu]); 651 saved_cpu_idle_disable_state[cpu] = NULL; 652 saved_cpu_idle_disable_state_alloc_ctr--; 653 if (saved_cpu_idle_disable_state_alloc_ctr == 0) { 654 free(saved_cpu_idle_disable_state); 655 saved_cpu_idle_disable_state = NULL; 656 } 657 658 return nr_states; 659 } 660 661 /* 662 * free_cpu_idle_disable_states - free saved idle state disable for all cpus 663 * 664 * Frees the memory used for storing cpu idle state disable for all cpus 665 * and states. 666 * 667 * Normally, the memory is freed automatically in 668 * restore_cpu_idle_disable_state; this is mostly for cleaning up after an 669 * error. 670 */ 671 void free_cpu_idle_disable_states(void) 672 { 673 int cpu; 674 675 if (!saved_cpu_idle_disable_state) 676 return; 677 678 for (cpu = 0; cpu < nr_cpus; cpu++) { 679 free(saved_cpu_idle_disable_state[cpu]); 680 saved_cpu_idle_disable_state[cpu] = NULL; 681 } 682 683 free(saved_cpu_idle_disable_state); 684 saved_cpu_idle_disable_state = NULL; 685 } 686 687 /* 688 * set_deepest_cpu_idle_state - limit idle state of cpu 689 * 690 * Disables all idle states deeper than the one given in 691 * deepest_state (assuming states with higher number are deeper). 692 * 693 * This is used to reduce the exit from idle latency. Unlike 694 * set_cpu_dma_latency, it can disable idle states per cpu. 695 * 696 * Return: idle state count on success, negative on error 697 */ 698 int set_deepest_cpu_idle_state(unsigned int cpu, unsigned int deepest_state) 699 { 700 unsigned int nr_states; 701 unsigned int state; 702 int result; 703 704 nr_states = cpuidle_state_count(cpu); 705 706 for (state = deepest_state + 1; state < nr_states; state++) { 707 result = cpuidle_state_disable(cpu, state, 1); 708 if (result < 0) 709 return result; 710 } 711 712 return nr_states; 713 } 714 #endif /* HAVE_LIBCPUPOWER_SUPPORT */ 715 716 #define _STR(x) #x 717 #define STR(x) _STR(x) 718 719 /* 720 * find_mount - find a the mount point of a given fs 721 * 722 * Returns 0 if mount is not found, otherwise return 1 and fill mp 723 * with the mount point. 724 */ 725 static const int find_mount(const char *fs, char *mp, int sizeof_mp) 726 { 727 char mount_point[MAX_PATH+1]; 728 char type[100]; 729 int found = 0; 730 FILE *fp; 731 732 fp = fopen("/proc/mounts", "r"); 733 if (!fp) 734 return 0; 735 736 while (fscanf(fp, "%*s %" STR(MAX_PATH) "s %99s %*s %*d %*d\n", mount_point, type) == 2) { 737 if (strcmp(type, fs) == 0) { 738 found = 1; 739 break; 740 } 741 } 742 fclose(fp); 743 744 if (!found) 745 return 0; 746 747 memset(mp, 0, sizeof_mp); 748 strncpy(mp, mount_point, sizeof_mp - 1); 749 750 debug_msg("Fs %s found at %s\n", fs, mp); 751 return 1; 752 } 753 754 /* 755 * get_self_cgroup - get the current thread cgroup path 756 * 757 * Parse /proc/$$/cgroup file to get the thread's cgroup. As an example of line to parse: 758 * 759 * 0::/user.slice/user-0.slice/session-3.scope'\n' 760 * 761 * This function is interested in the content after the second : and before the '\n'. 762 * 763 * Returns 1 if a string was found, 0 otherwise. 764 */ 765 static int get_self_cgroup(char *self_cg, int sizeof_self_cg) 766 { 767 char path[MAX_PATH], *start; 768 int fd, retval; 769 770 snprintf(path, MAX_PATH, "/proc/%d/cgroup", getpid()); 771 772 fd = open(path, O_RDONLY); 773 if (fd < 0) 774 return 0; 775 776 memset(path, 0, sizeof(path)); 777 retval = read(fd, path, MAX_PATH); 778 779 close(fd); 780 781 if (retval <= 0) 782 return 0; 783 784 path[MAX_PATH-1] = '\0'; 785 start = path; 786 787 start = strstr(start, ":"); 788 if (!start) 789 return 0; 790 791 /* skip ":" */ 792 start++; 793 794 start = strstr(start, ":"); 795 if (!start) 796 return 0; 797 798 /* skip ":" */ 799 start++; 800 801 if (strlen(start) >= sizeof_self_cg) 802 return 0; 803 804 snprintf(self_cg, sizeof_self_cg, "%s", start); 805 806 /* Swap '\n' with '\0' */ 807 start = strstr(self_cg, "\n"); 808 809 /* there must be '\n' */ 810 if (!start) 811 return 0; 812 813 /* ok, it found a string after the second : and before the \n */ 814 *start = '\0'; 815 816 return 1; 817 } 818 819 /* 820 * open_cgroup_procs - Open the cgroup.procs file for the given cgroup 821 * 822 * If cgroup argument is not NULL, the cgroup.procs file for that cgroup 823 * will be opened. Otherwise, the cgroup of the calling, i.e., rtla, thread 824 * will be used. 825 * 826 * Supports cgroup v2. 827 * 828 * Returns the file descriptor on success, -1 otherwise. 829 */ 830 static int open_cgroup_procs(const char *cgroup) 831 { 832 char cgroup_path[MAX_PATH - strlen("/cgroup.procs")]; 833 char cgroup_procs[MAX_PATH]; 834 int retval; 835 int cg_fd; 836 size_t cg_path_len; 837 838 retval = find_mount("cgroup2", cgroup_path, sizeof(cgroup_path)); 839 if (!retval) { 840 err_msg("Did not find cgroupv2 mount point\n"); 841 return -1; 842 } 843 844 cg_path_len = strlen(cgroup_path); 845 846 if (!cgroup) { 847 retval = get_self_cgroup(&cgroup_path[cg_path_len], 848 sizeof(cgroup_path) - cg_path_len); 849 if (!retval) { 850 err_msg("Did not find self cgroup\n"); 851 return -1; 852 } 853 } else { 854 snprintf(&cgroup_path[cg_path_len], 855 sizeof(cgroup_path) - cg_path_len, "%s/", cgroup); 856 } 857 858 snprintf(cgroup_procs, MAX_PATH, "%s/cgroup.procs", cgroup_path); 859 860 debug_msg("Using cgroup path at: %s\n", cgroup_procs); 861 862 cg_fd = open(cgroup_procs, O_RDWR); 863 if (cg_fd < 0) 864 return -1; 865 866 return cg_fd; 867 } 868 869 /* 870 * set_pid_cgroup - Set cgroup to pid_t pid 871 * 872 * If cgroup argument is not NULL, the threads will move to the given cgroup. 873 * Otherwise, the cgroup of the calling, i.e., rtla, thread will be used. 874 * 875 * Supports cgroup v2. 876 * 877 * Returns 1 on success, 0 otherwise. 878 */ 879 int set_pid_cgroup(pid_t pid, const char *cgroup) 880 { 881 char pid_str[24]; 882 int retval; 883 int cg_fd; 884 885 cg_fd = open_cgroup_procs(cgroup); 886 if (cg_fd < 0) 887 return 0; 888 889 snprintf(pid_str, sizeof(pid_str), "%d\n", pid); 890 891 retval = write(cg_fd, pid_str, strlen(pid_str)); 892 if (retval < 0) 893 err_msg("Error setting cgroup attributes for pid:%s - %s\n", 894 pid_str, strerror(errno)); 895 else 896 debug_msg("Set cgroup attributes for pid:%s\n", pid_str); 897 898 close(cg_fd); 899 900 return (retval >= 0); 901 } 902 903 /** 904 * set_comm_cgroup - Set cgroup to threads starting with char *comm_prefix 905 * 906 * If cgroup argument is not NULL, the threads will move to the given cgroup. 907 * Otherwise, the cgroup of the calling, i.e., rtla, thread will be used. 908 * 909 * Supports cgroup v2. 910 * 911 * Returns 1 on success, 0 otherwise. 912 */ 913 int set_comm_cgroup(const char *comm_prefix, const char *cgroup) 914 { 915 struct dirent *proc_entry; 916 DIR *procfs; 917 int retval; 918 int cg_fd; 919 920 if (strlen(comm_prefix) >= MAX_PATH) { 921 err_msg("Command prefix is too long: %d < strlen(%s)\n", 922 MAX_PATH, comm_prefix); 923 return 0; 924 } 925 926 cg_fd = open_cgroup_procs(cgroup); 927 if (cg_fd < 0) 928 return 0; 929 930 procfs = opendir("/proc"); 931 if (!procfs) { 932 err_msg("Could not open procfs\n"); 933 goto out_cg; 934 } 935 936 while ((proc_entry = readdir(procfs))) { 937 938 retval = procfs_is_workload_pid(comm_prefix, proc_entry); 939 if (!retval) 940 continue; 941 942 retval = write(cg_fd, proc_entry->d_name, strlen(proc_entry->d_name)); 943 if (retval < 0) { 944 err_msg("Error setting cgroup attributes for pid:%s - %s\n", 945 proc_entry->d_name, strerror(errno)); 946 goto out_procfs; 947 } 948 949 debug_msg("Set cgroup attributes for pid:%s\n", proc_entry->d_name); 950 } 951 952 closedir(procfs); 953 close(cg_fd); 954 return 1; 955 956 out_procfs: 957 closedir(procfs); 958 out_cg: 959 close(cg_fd); 960 return 0; 961 } 962 963 /** 964 * auto_house_keeping - Automatically move rtla out of measurement threads 965 * 966 * Try to move rtla away from the tracer, if possible. 967 * 968 * Returns 1 on success, 0 otherwise. 969 */ 970 int auto_house_keeping(cpu_set_t *monitored_cpus) 971 { 972 cpu_set_t rtla_cpus, house_keeping_cpus; 973 int retval; 974 975 /* first get the CPUs in which rtla can actually run. */ 976 retval = sched_getaffinity(getpid(), sizeof(rtla_cpus), &rtla_cpus); 977 if (retval == -1) { 978 debug_msg("Could not get rtla affinity, rtla might run with the threads!\n"); 979 return 0; 980 } 981 982 /* then check if the existing setup is already good. */ 983 CPU_AND(&house_keeping_cpus, &rtla_cpus, monitored_cpus); 984 if (!CPU_COUNT(&house_keeping_cpus)) { 985 debug_msg("rtla and the monitored CPUs do not share CPUs."); 986 debug_msg("Skipping auto house-keeping\n"); 987 return 1; 988 } 989 990 /* remove the intersection */ 991 CPU_XOR(&house_keeping_cpus, &rtla_cpus, monitored_cpus); 992 993 /* get only those that rtla can run */ 994 CPU_AND(&house_keeping_cpus, &house_keeping_cpus, &rtla_cpus); 995 996 /* is there any cpu left? */ 997 if (!CPU_COUNT(&house_keeping_cpus)) { 998 debug_msg("Could not find any CPU for auto house-keeping\n"); 999 return 0; 1000 } 1001 1002 retval = sched_setaffinity(getpid(), sizeof(house_keeping_cpus), &house_keeping_cpus); 1003 if (retval == -1) { 1004 debug_msg("Could not set affinity for auto house-keeping\n"); 1005 return 0; 1006 } 1007 1008 debug_msg("rtla automatically moved to an auto house-keeping cpu set\n"); 1009 1010 return 1; 1011 } 1012 1013 /** 1014 * parse_optional_arg - Parse optional argument value 1015 * 1016 * Parse optional argument value, which can be in the form of: 1017 * -sarg, -s/--long=arg, -s/--long arg 1018 * 1019 * Returns arg value if found, NULL otherwise. 1020 */ 1021 char *parse_optional_arg(int argc, char **argv) 1022 { 1023 if (optarg) { 1024 if (optarg[0] == '=') { 1025 /* skip the = */ 1026 return &optarg[1]; 1027 } else { 1028 return optarg; 1029 } 1030 /* parse argument of form -s [arg] and --long [arg]*/ 1031 } else if (optind < argc && argv[optind][0] != '-') { 1032 /* consume optind */ 1033 return argv[optind++]; 1034 } else { 1035 return NULL; 1036 } 1037 } 1038 1039 /* 1040 * strtoi - convert string to integer with error checking 1041 * 1042 * Returns 0 on success, -1 if conversion fails or result is out of int range. 1043 */ 1044 int strtoi(const char *s, int *res) 1045 { 1046 char *end_ptr; 1047 long lres; 1048 1049 if (!*s) 1050 return -1; 1051 1052 errno = 0; 1053 lres = strtol(s, &end_ptr, 0); 1054 if (errno || *end_ptr || lres > INT_MAX || lres < INT_MIN) 1055 return -1; 1056 1057 *res = (int) lres; 1058 return 0; 1059 } 1060 1061 static inline void fatal_alloc(void) 1062 { 1063 fatal("Error allocating memory\n"); 1064 } 1065 1066 void *calloc_fatal(size_t n, size_t size) 1067 { 1068 void *p = calloc(n, size); 1069 1070 if (!p) 1071 fatal_alloc(); 1072 1073 return p; 1074 } 1075 1076 void *reallocarray_fatal(void *p, size_t n, size_t size) 1077 { 1078 p = reallocarray(p, n, size); 1079 1080 if (!p) 1081 fatal_alloc(); 1082 1083 return p; 1084 } 1085 1086 char *strdup_fatal(const char *s) 1087 { 1088 char *p = strdup(s); 1089 1090 if (!p) 1091 fatal_alloc(); 1092 1093 return p; 1094 } 1095