Upstream | cc2ee17 | 1970-01-12 13:46:40 +0000 | [diff] [blame] | 1 | /** |
| 2 | * @file opd_perfmon.c |
| 3 | * perfmonctl() handling |
| 4 | * |
| 5 | * @remark Copyright 2003 OProfile authors |
| 6 | * @remark Read the file COPYING |
| 7 | * |
| 8 | * @author John Levon |
| 9 | */ |
| 10 | |
| 11 | #ifdef __ia64__ |
| 12 | |
| 13 | /* need this for sched_setaffinity() in <sched.h> */ |
| 14 | #define _GNU_SOURCE |
| 15 | |
| 16 | #include "oprofiled.h" |
| 17 | #include "opd_perfmon.h" |
| 18 | #include "opd_events.h" |
| 19 | |
| 20 | #include "op_cpu_type.h" |
| 21 | #include "op_libiberty.h" |
| 22 | #include "op_hw_config.h" |
| 23 | |
| 24 | #include <sys/syscall.h> |
| 25 | #include <sys/wait.h> |
| 26 | #include <unistd.h> |
| 27 | #include <limits.h> |
| 28 | #include <signal.h> |
| 29 | #include <stdio.h> |
| 30 | #include <stdlib.h> |
| 31 | #include <string.h> |
| 32 | #include <errno.h> |
| 33 | #ifdef HAVE_SCHED_SETAFFINITY |
| 34 | #include <sched.h> |
| 35 | #endif |
| 36 | |
| 37 | extern op_cpu cpu_type; |
| 38 | |
| 39 | #ifndef HAVE_SCHED_SETAFFINITY |
| 40 | |
| 41 | /* many glibc's are not yet up to date */ |
| 42 | #ifndef __NR_sched_setaffinity |
| 43 | #define __NR_sched_setaffinity 1231 |
| 44 | #endif |
| 45 | |
| 46 | /* Copied from glibc's <sched.h> and <bits/sched.h> and munged */ |
| 47 | #define CPU_SETSIZE 1024 |
| 48 | #define __NCPUBITS (8 * sizeof (unsigned long)) |
| 49 | typedef struct |
| 50 | { |
| 51 | unsigned long __bits[CPU_SETSIZE / __NCPUBITS]; |
| 52 | } cpu_set_t; |
| 53 | |
| 54 | #define CPU_SET(cpu, cpusetp) \ |
| 55 | ((cpusetp)->__bits[(cpu)/__NCPUBITS] |= (1UL << ((cpu) % __NCPUBITS))) |
| 56 | #define CPU_ZERO(cpusetp) \ |
| 57 | memset((cpusetp), 0, sizeof(cpu_set_t)) |
| 58 | |
| 59 | static int |
| 60 | sched_setaffinity(pid_t pid, size_t len, cpu_set_t const * cpusetp) |
| 61 | { |
| 62 | return syscall(__NR_sched_setaffinity, pid, len, cpusetp); |
| 63 | } |
| 64 | #endif |
| 65 | |
| 66 | |
| 67 | #ifndef HAVE_PERFMONCTL |
| 68 | #ifndef __NR_perfmonctl |
| 69 | #define __NR_perfmonctl 1175 |
| 70 | #endif |
| 71 | |
| 72 | static int perfmonctl(int fd, int cmd, void * arg, int narg) |
| 73 | { |
| 74 | return syscall(__NR_perfmonctl, fd, cmd, arg, narg); |
| 75 | } |
| 76 | #endif |
| 77 | |
| 78 | |
| 79 | static unsigned char uuid[16] = { |
| 80 | 0x77, 0x7a, 0x6e, 0x61, 0x20, 0x65, 0x73, 0x69, |
| 81 | 0x74, 0x6e, 0x72, 0x20, 0x61, 0x65, 0x0a, 0x6c |
| 82 | }; |
| 83 | |
| 84 | |
| 85 | static size_t nr_cpus; |
| 86 | |
| 87 | struct child { |
| 88 | pid_t pid; |
| 89 | int up_pipe[2]; |
| 90 | int ctx_fd; |
| 91 | sig_atomic_t sigusr1; |
| 92 | sig_atomic_t sigusr2; |
| 93 | sig_atomic_t sigterm; |
| 94 | }; |
| 95 | |
| 96 | static struct child * children; |
| 97 | |
| 98 | static void perfmon_start_child(int ctx_fd) |
| 99 | { |
| 100 | if (perfmonctl(ctx_fd, PFM_START, 0, 0) == -1) { |
| 101 | perror("Couldn't start perfmon: "); |
| 102 | exit(EXIT_FAILURE); |
| 103 | } |
| 104 | } |
| 105 | |
| 106 | |
| 107 | static void perfmon_stop_child(int ctx_fd) |
| 108 | { |
| 109 | if (perfmonctl(ctx_fd, PFM_STOP, 0, 0) == -1) { |
| 110 | perror("Couldn't stop perfmon: "); |
| 111 | exit(EXIT_FAILURE); |
| 112 | } |
| 113 | } |
| 114 | |
| 115 | |
| 116 | static void child_sigusr1(int val __attribute__((unused))) |
| 117 | { |
| 118 | size_t i; |
| 119 | |
| 120 | for (i = 0; i < nr_cpus; ++i) { |
| 121 | if (children[i].pid == getpid()) { |
| 122 | children[i].sigusr1 = 1; |
| 123 | return; |
| 124 | } |
| 125 | } |
| 126 | } |
| 127 | |
| 128 | |
| 129 | static void child_sigusr2(int val __attribute__((unused))) |
| 130 | { |
| 131 | size_t i; |
| 132 | |
| 133 | for (i = 0; i < nr_cpus; ++i) { |
| 134 | if (children[i].pid == getpid()) { |
| 135 | children[i].sigusr2 = 1; |
| 136 | return; |
| 137 | } |
| 138 | } |
| 139 | } |
| 140 | |
| 141 | |
| 142 | static void child_sigterm(int val __attribute__((unused))) |
| 143 | { |
| 144 | printf("Child received SIGTERM, killing parent.\n"); |
| 145 | kill(getppid(), SIGTERM); |
| 146 | } |
| 147 | |
| 148 | |
| 149 | static void set_affinity(size_t cpu) |
| 150 | { |
| 151 | cpu_set_t set; |
| 152 | |
| 153 | CPU_ZERO(&set); |
| 154 | CPU_SET(cpu, &set); |
| 155 | |
| 156 | int err = sched_setaffinity(getpid(), sizeof(set), &set); |
| 157 | |
| 158 | if (err == -1) { |
| 159 | fprintf(stderr, "Failed to set affinity: %s\n", |
| 160 | strerror(errno)); |
| 161 | exit(EXIT_FAILURE); |
| 162 | } |
| 163 | } |
| 164 | |
| 165 | |
| 166 | static void setup_signals(void) |
| 167 | { |
| 168 | struct sigaction act; |
| 169 | sigset_t mask; |
| 170 | |
| 171 | sigemptyset(&mask); |
| 172 | sigaddset(&mask, SIGUSR1); |
| 173 | sigaddset(&mask, SIGUSR2); |
| 174 | sigprocmask(SIG_BLOCK, &mask, NULL); |
| 175 | |
| 176 | act.sa_handler = child_sigusr1; |
| 177 | act.sa_flags = 0; |
| 178 | sigemptyset(&act.sa_mask); |
| 179 | |
| 180 | if (sigaction(SIGUSR1, &act, NULL)) { |
| 181 | perror("oprofiled: install of SIGUSR1 handler failed: "); |
| 182 | exit(EXIT_FAILURE); |
| 183 | } |
| 184 | |
| 185 | act.sa_handler = child_sigusr2; |
| 186 | act.sa_flags = 0; |
| 187 | sigemptyset(&act.sa_mask); |
| 188 | |
| 189 | if (sigaction(SIGUSR2, &act, NULL)) { |
| 190 | perror("oprofiled: install of SIGUSR2 handler failed: "); |
| 191 | exit(EXIT_FAILURE); |
| 192 | } |
| 193 | |
| 194 | act.sa_handler = child_sigterm; |
| 195 | act.sa_flags = 0; |
| 196 | sigemptyset(&act.sa_mask); |
| 197 | |
| 198 | if (sigaction(SIGTERM, &act, NULL)) { |
| 199 | perror("oprofiled: install of SIGTERM handler failed: "); |
| 200 | exit(EXIT_FAILURE); |
| 201 | } |
| 202 | } |
| 203 | |
| 204 | |
| 205 | /** create the per-cpu context */ |
| 206 | static void create_context(struct child * self) |
| 207 | { |
| 208 | pfarg_context_t ctx; |
| 209 | int err; |
| 210 | |
| 211 | memset(&ctx, 0, sizeof(pfarg_context_t)); |
| 212 | memcpy(&ctx.ctx_smpl_buf_id, &uuid, 16); |
| 213 | ctx.ctx_flags = PFM_FL_SYSTEM_WIDE; |
| 214 | |
| 215 | err = perfmonctl(0, PFM_CREATE_CONTEXT, &ctx, 1); |
| 216 | if (err == -1) { |
| 217 | fprintf(stderr, "CREATE_CONTEXT failed: %s\n", |
| 218 | strerror(errno)); |
| 219 | exit(EXIT_FAILURE); |
| 220 | } |
| 221 | |
| 222 | self->ctx_fd = ctx.ctx_fd; |
| 223 | } |
| 224 | |
| 225 | |
| 226 | /** program the perfmon counters */ |
| 227 | static void write_pmu(struct child * self) |
| 228 | { |
| 229 | pfarg_reg_t pc[OP_MAX_COUNTERS]; |
| 230 | pfarg_reg_t pd[OP_MAX_COUNTERS]; |
| 231 | int err; |
| 232 | size_t i; |
| 233 | |
| 234 | memset(pc, 0, sizeof(pc)); |
| 235 | memset(pd, 0, sizeof(pd)); |
| 236 | |
| 237 | #define PMC_GEN_INTERRUPT (1UL << 5) |
| 238 | #define PMC_PRIV_MONITOR (1UL << 6) |
| 239 | /* McKinley requires pmc4 to have bit 23 set (enable PMU). |
| 240 | * It is supposedly ignored in other pmc registers. |
| 241 | */ |
| 242 | #define PMC_MANDATORY (1UL << 23) |
| 243 | #define PMC_USER (1UL << 3) |
| 244 | #define PMC_KERNEL (1UL << 0) |
| 245 | for (i = 0; i < op_nr_counters && opd_events[i].name; ++i) { |
| 246 | struct opd_event * event = &opd_events[i]; |
| 247 | pc[i].reg_num = event->counter + 4; |
| 248 | pc[i].reg_value = PMC_GEN_INTERRUPT; |
| 249 | pc[i].reg_value |= PMC_PRIV_MONITOR; |
| 250 | pc[i].reg_value |= PMC_MANDATORY; |
| 251 | (event->user) ? (pc[i].reg_value |= PMC_USER) |
| 252 | : (pc[i].reg_value &= ~PMC_USER); |
| 253 | (event->kernel) ? (pc[i].reg_value |= PMC_KERNEL) |
| 254 | : (pc[i].reg_value &= ~PMC_KERNEL); |
| 255 | pc[i].reg_value &= ~(0xff << 8); |
| 256 | pc[i].reg_value |= ((event->value & 0xff) << 8); |
| 257 | pc[i].reg_value &= ~(0xf << 16); |
| 258 | pc[i].reg_value |= ((event->um & 0xf) << 16); |
| 259 | pc[i].reg_smpl_eventid = event->counter; |
| 260 | } |
| 261 | |
| 262 | for (i = 0; i < op_nr_counters && opd_events[i].name; ++i) { |
| 263 | struct opd_event * event = &opd_events[i]; |
| 264 | pd[i].reg_value = ~0UL - event->count + 1; |
| 265 | pd[i].reg_short_reset = ~0UL - event->count + 1; |
| 266 | pd[i].reg_num = event->counter + 4; |
| 267 | } |
| 268 | |
| 269 | err = perfmonctl(self->ctx_fd, PFM_WRITE_PMCS, pc, i); |
| 270 | if (err == -1) { |
| 271 | perror("Couldn't write PMCs: "); |
| 272 | exit(EXIT_FAILURE); |
| 273 | } |
| 274 | |
| 275 | err = perfmonctl(self->ctx_fd, PFM_WRITE_PMDS, pd, i); |
| 276 | if (err == -1) { |
| 277 | perror("Couldn't write PMDs: "); |
| 278 | exit(EXIT_FAILURE); |
| 279 | } |
| 280 | } |
| 281 | |
| 282 | |
| 283 | static void load_context(struct child * self) |
| 284 | { |
| 285 | pfarg_load_t load_args; |
| 286 | int err; |
| 287 | |
| 288 | memset(&load_args, 0, sizeof(load_args)); |
| 289 | load_args.load_pid = self->pid; |
| 290 | |
| 291 | err = perfmonctl(self->ctx_fd, PFM_LOAD_CONTEXT, &load_args, 1); |
| 292 | if (err == -1) { |
| 293 | perror("Couldn't load context: "); |
| 294 | exit(EXIT_FAILURE); |
| 295 | } |
| 296 | } |
| 297 | |
| 298 | |
| 299 | static void notify_parent(struct child * self, size_t cpu) |
| 300 | { |
| 301 | for (;;) { |
| 302 | ssize_t ret; |
| 303 | ret = write(self->up_pipe[1], &cpu, sizeof(size_t)); |
| 304 | if (ret == sizeof(size_t)) |
| 305 | break; |
| 306 | if (ret < 0 && errno != EINTR) { |
| 307 | fprintf(stderr, "Failed to write child pipe with %s\n", |
| 308 | strerror(errno)); |
| 309 | exit(EXIT_FAILURE); |
| 310 | } |
| 311 | } |
| 312 | } |
| 313 | |
| 314 | |
| 315 | static void run_child(size_t cpu) |
| 316 | { |
| 317 | struct child * self = &children[cpu]; |
| 318 | |
| 319 | self->pid = getpid(); |
| 320 | self->sigusr1 = 0; |
| 321 | self->sigusr2 = 0; |
| 322 | self->sigterm = 0; |
| 323 | |
| 324 | setup_signals(); |
| 325 | |
| 326 | set_affinity(cpu); |
| 327 | |
| 328 | create_context(self); |
| 329 | |
| 330 | write_pmu(self); |
| 331 | |
| 332 | load_context(self); |
| 333 | |
| 334 | notify_parent(self, cpu); |
| 335 | |
| 336 | for (;;) { |
| 337 | sigset_t sigmask; |
| 338 | sigfillset(&sigmask); |
| 339 | sigdelset(&sigmask, SIGUSR1); |
| 340 | sigdelset(&sigmask, SIGUSR2); |
| 341 | sigdelset(&sigmask, SIGTERM); |
| 342 | |
| 343 | if (self->sigusr1) { |
| 344 | printf("PFM_START on CPU%d\n", (int)cpu); |
| 345 | fflush(stdout); |
| 346 | perfmon_start_child(self->ctx_fd); |
| 347 | self->sigusr1 = 0; |
| 348 | } |
| 349 | |
| 350 | if (self->sigusr2) { |
| 351 | printf("PFM_STOP on CPU%d\n", (int)cpu); |
| 352 | fflush(stdout); |
| 353 | perfmon_stop_child(self->ctx_fd); |
| 354 | self->sigusr2 = 0; |
| 355 | } |
| 356 | |
| 357 | sigsuspend(&sigmask); |
| 358 | } |
| 359 | } |
| 360 | |
| 361 | |
| 362 | static void wait_for_child(struct child * child) |
| 363 | { |
| 364 | size_t tmp; |
| 365 | for (;;) { |
| 366 | ssize_t ret; |
| 367 | ret = read(child->up_pipe[0], &tmp, sizeof(size_t)); |
| 368 | if (ret == sizeof(size_t)) |
| 369 | break; |
| 370 | if (ret < 0 && errno != EINTR) { |
| 371 | fprintf(stderr, "Failed to read child pipe with %s\n", |
| 372 | strerror(errno)); |
| 373 | exit(EXIT_FAILURE); |
| 374 | } |
| 375 | } |
| 376 | printf("Perfmon child up on CPU%d\n", (int)tmp); |
| 377 | fflush(stdout); |
| 378 | |
| 379 | close(child->up_pipe[0]); |
| 380 | close(child->up_pipe[1]); |
| 381 | } |
| 382 | |
| 383 | |
| 384 | void perfmon_init(void) |
| 385 | { |
| 386 | size_t i; |
| 387 | long nr; |
| 388 | |
| 389 | if (cpu_type == CPU_TIMER_INT) |
| 390 | return; |
| 391 | |
| 392 | nr = sysconf(_SC_NPROCESSORS_ONLN); |
| 393 | if (nr == -1) { |
| 394 | fprintf(stderr, "Couldn't determine number of CPUs.\n"); |
| 395 | exit(EXIT_FAILURE); |
| 396 | } |
| 397 | |
| 398 | nr_cpus = nr; |
| 399 | |
| 400 | children = xmalloc(sizeof(struct child) * nr_cpus); |
| 401 | |
| 402 | for (i = 0; i < nr_cpus; ++i) { |
| 403 | int ret; |
| 404 | |
| 405 | if (pipe(children[i].up_pipe)) { |
| 406 | perror("Couldn't create child pipe.\n"); |
| 407 | exit(EXIT_FAILURE); |
| 408 | } |
| 409 | |
| 410 | ret = fork(); |
| 411 | if (ret == -1) { |
| 412 | fprintf(stderr, "Couldn't fork perfmon child.\n"); |
| 413 | exit(EXIT_FAILURE); |
| 414 | } else if (ret == 0) { |
| 415 | printf("Running perfmon child on CPU%d.\n", (int)i); |
| 416 | fflush(stdout); |
| 417 | run_child(i); |
| 418 | } else { |
| 419 | children[i].pid = ret; |
| 420 | printf("Waiting on CPU%d\n", (int)i); |
| 421 | wait_for_child(&children[i]); |
| 422 | } |
| 423 | } |
| 424 | } |
| 425 | |
| 426 | |
| 427 | void perfmon_exit(void) |
| 428 | { |
| 429 | size_t i; |
| 430 | |
| 431 | if (cpu_type == CPU_TIMER_INT) |
| 432 | return; |
| 433 | |
| 434 | for (i = 0; i < nr_cpus; ++i) { |
| 435 | kill(children[i].pid, SIGKILL); |
| 436 | waitpid(children[i].pid, NULL, 0); |
| 437 | } |
| 438 | } |
| 439 | |
| 440 | |
| 441 | void perfmon_start(void) |
| 442 | { |
| 443 | size_t i; |
| 444 | |
| 445 | if (cpu_type == CPU_TIMER_INT) |
| 446 | return; |
| 447 | |
| 448 | for (i = 0; i < nr_cpus; ++i) |
| 449 | kill(children[i].pid, SIGUSR1); |
| 450 | } |
| 451 | |
| 452 | |
| 453 | void perfmon_stop(void) |
| 454 | { |
| 455 | size_t i; |
| 456 | |
| 457 | if (cpu_type == CPU_TIMER_INT) |
| 458 | return; |
| 459 | |
| 460 | for (i = 0; i < nr_cpus; ++i) |
| 461 | kill(children[i].pid, SIGUSR2); |
| 462 | } |
| 463 | |
| 464 | #endif /* __ia64__ */ |