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