lh | 9ed821d | 2023-04-07 01:36:19 -0700 | [diff] [blame] | 1 | /* Linuxthreads - a simple clone()-based implementation of Posix */ |
| 2 | /* threads for Linux. */ |
| 3 | /* Copyright (C) 1996 Xavier Leroy (Xavier.Leroy@inria.fr) */ |
| 4 | /* */ |
| 5 | /* This program is free software; you can redistribute it and/or */ |
| 6 | /* modify it under the terms of the GNU Library General Public License */ |
| 7 | /* as published by the Free Software Foundation; either version 2 */ |
| 8 | /* of the License, or (at your option) any later version. */ |
| 9 | /* */ |
| 10 | /* This program is distributed in the hope that it will be useful, */ |
| 11 | /* but WITHOUT ANY WARRANTY; without even the implied warranty of */ |
| 12 | /* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */ |
| 13 | /* GNU Library General Public License for more details. */ |
| 14 | |
| 15 | /* The "thread manager" thread: manages creation and termination of threads */ |
| 16 | |
| 17 | #include <features.h> |
| 18 | #include <errno.h> |
| 19 | #include <sched.h> |
| 20 | #include <stddef.h> |
| 21 | #include <stdio.h> |
| 22 | #include <stdlib.h> |
| 23 | #include <string.h> |
| 24 | #include <unistd.h> |
| 25 | #include <sys/poll.h> /* for poll */ |
| 26 | #include <sys/mman.h> /* for mmap */ |
| 27 | #include <sys/param.h> |
| 28 | #include <sys/time.h> |
| 29 | #include <sys/wait.h> /* for waitpid macros */ |
| 30 | |
| 31 | #include "pthread.h" |
| 32 | #include "internals.h" |
| 33 | #include "spinlock.h" |
| 34 | #include "restart.h" |
| 35 | #include "semaphore.h" |
| 36 | #include "debug.h" /* PDEBUG, added by StS */ |
| 37 | |
| 38 | #ifndef THREAD_STACK_OFFSET |
| 39 | #define THREAD_STACK_OFFSET 0 |
| 40 | #endif |
| 41 | |
| 42 | /* poll() is not supported in kernel <= 2.0, therefore is __NR_poll is |
| 43 | * not available, we assume an old Linux kernel is in use and we will |
| 44 | * use select() instead. */ |
| 45 | #include <sys/syscall.h> |
| 46 | #ifndef __NR_poll |
| 47 | # define USE_SELECT |
| 48 | #endif |
| 49 | |
| 50 | libpthread_hidden_proto(waitpid) |
| 51 | libpthread_hidden_proto(raise) |
| 52 | |
| 53 | /* Array of active threads. Entry 0 is reserved for the initial thread. */ |
| 54 | struct pthread_handle_struct __pthread_handles[PTHREAD_THREADS_MAX] = |
| 55 | { { __LOCK_INITIALIZER, &__pthread_initial_thread, 0}, |
| 56 | { __LOCK_INITIALIZER, &__pthread_manager_thread, 0}, /* All NULLs */ }; |
| 57 | |
| 58 | /* For debugging purposes put the maximum number of threads in a variable. */ |
| 59 | const int __linuxthreads_pthread_threads_max = PTHREAD_THREADS_MAX; |
| 60 | |
| 61 | /* Indicate whether at least one thread has a user-defined stack (if 1), |
| 62 | or if all threads have stacks supplied by LinuxThreads (if 0). */ |
| 63 | int __pthread_nonstandard_stacks; |
| 64 | |
| 65 | /* Number of active entries in __pthread_handles (used by gdb) */ |
| 66 | volatile int __pthread_handles_num = 2; |
| 67 | |
| 68 | /* Whether to use debugger additional actions for thread creation |
| 69 | (set to 1 by gdb) */ |
| 70 | volatile int __pthread_threads_debug; |
| 71 | |
| 72 | /* Globally enabled events. */ |
| 73 | volatile td_thr_events_t __pthread_threads_events; |
| 74 | |
| 75 | /* Pointer to thread descriptor with last event. */ |
| 76 | volatile pthread_descr __pthread_last_event; |
| 77 | |
| 78 | /* Mapping from stack segment to thread descriptor. */ |
| 79 | /* Stack segment numbers are also indices into the __pthread_handles array. */ |
| 80 | /* Stack segment number 0 is reserved for the initial thread. */ |
| 81 | |
| 82 | static __inline__ pthread_descr thread_segment(int seg) |
| 83 | { |
| 84 | return (pthread_descr)(THREAD_STACK_START_ADDRESS - (seg - 1) * STACK_SIZE) |
| 85 | - 1; |
| 86 | } |
| 87 | |
| 88 | /* Flag set in signal handler to record child termination */ |
| 89 | |
| 90 | static volatile int terminated_children = 0; |
| 91 | |
| 92 | /* Flag set when the initial thread is blocked on pthread_exit waiting |
| 93 | for all other threads to terminate */ |
| 94 | |
| 95 | static int main_thread_exiting = 0; |
| 96 | |
| 97 | /* Counter used to generate unique thread identifier. |
| 98 | Thread identifier is pthread_threads_counter + segment. */ |
| 99 | |
| 100 | static pthread_t pthread_threads_counter = 0; |
| 101 | |
| 102 | /* Forward declarations */ |
| 103 | |
| 104 | static int pthread_handle_create(pthread_t *thread, const pthread_attr_t *attr, |
| 105 | void * (*start_routine)(void *), void *arg, |
| 106 | sigset_t *mask, int father_pid, |
| 107 | int report_events, |
| 108 | td_thr_events_t *event_maskp); |
| 109 | static void pthread_handle_free(pthread_t th_id); |
| 110 | static void pthread_handle_exit(pthread_descr issuing_thread, int exitcode) attribute_noreturn; |
| 111 | static void pthread_reap_children(void); |
| 112 | static void pthread_kill_all_threads(int sig, int main_thread_also); |
| 113 | |
| 114 | /* The server thread managing requests for thread creation and termination */ |
| 115 | |
| 116 | int attribute_noreturn __pthread_manager(void *arg) |
| 117 | { |
| 118 | int reqfd = (int) (long int) arg; |
| 119 | #ifdef USE_SELECT |
| 120 | struct timeval tv; |
| 121 | fd_set fd; |
| 122 | #else |
| 123 | struct pollfd ufd; |
| 124 | #endif |
| 125 | sigset_t manager_mask; |
| 126 | int n; |
| 127 | struct pthread_request request; |
| 128 | |
| 129 | /* If we have special thread_self processing, initialize it. */ |
| 130 | #ifdef INIT_THREAD_SELF |
| 131 | INIT_THREAD_SELF(&__pthread_manager_thread, 1); |
| 132 | #endif |
| 133 | /* Set the error variable. */ |
| 134 | __pthread_manager_thread.p_errnop = &__pthread_manager_thread.p_errno; |
| 135 | __pthread_manager_thread.p_h_errnop = &__pthread_manager_thread.p_h_errno; |
| 136 | |
| 137 | #ifdef __UCLIBC_HAS_XLOCALE__ |
| 138 | /* Initialize thread's locale to the global locale. */ |
| 139 | __pthread_manager_thread.locale = __global_locale; |
| 140 | #endif /* __UCLIBC_HAS_XLOCALE__ */ |
| 141 | |
| 142 | /* Block all signals except __pthread_sig_cancel and SIGTRAP */ |
| 143 | __sigfillset(&manager_mask); |
| 144 | sigdelset(&manager_mask, __pthread_sig_cancel); /* for thread termination */ |
| 145 | sigdelset(&manager_mask, SIGTRAP); /* for debugging purposes */ |
| 146 | if (__pthread_threads_debug && __pthread_sig_debug > 0) |
| 147 | sigdelset(&manager_mask, __pthread_sig_debug); |
| 148 | sigprocmask(SIG_SETMASK, &manager_mask, NULL); |
| 149 | /* Raise our priority to match that of main thread */ |
| 150 | __pthread_manager_adjust_prio(__pthread_main_thread->p_priority); |
| 151 | /* Synchronize debugging of the thread manager */ |
| 152 | n = TEMP_FAILURE_RETRY(read(reqfd, (char *)&request, |
| 153 | sizeof(request))); |
| 154 | #ifndef USE_SELECT |
| 155 | ufd.fd = reqfd; |
| 156 | ufd.events = POLLIN; |
| 157 | #endif |
| 158 | /* Enter server loop */ |
| 159 | while(1) { |
| 160 | #ifdef USE_SELECT |
| 161 | tv.tv_sec = 2; |
| 162 | tv.tv_usec = 0; |
| 163 | FD_ZERO (&fd); |
| 164 | FD_SET (reqfd, &fd); |
| 165 | n = select (reqfd + 1, &fd, NULL, NULL, &tv); |
| 166 | #else |
| 167 | PDEBUG("before poll\n"); |
| 168 | n = poll(&ufd, 1, 2000); |
| 169 | PDEBUG("after poll\n"); |
| 170 | #endif |
| 171 | /* Check for termination of the main thread */ |
| 172 | if (getppid() == 1) { |
| 173 | pthread_kill_all_threads(SIGKILL, 0); |
| 174 | _exit(0); |
| 175 | } |
| 176 | /* Check for dead children */ |
| 177 | if (terminated_children) { |
| 178 | terminated_children = 0; |
| 179 | pthread_reap_children(); |
| 180 | } |
| 181 | /* Read and execute request */ |
| 182 | #ifdef USE_SELECT |
| 183 | if (n == 1) |
| 184 | #else |
| 185 | if (n == 1 && (ufd.revents & POLLIN)) |
| 186 | #endif |
| 187 | { |
| 188 | |
| 189 | PDEBUG("before read\n"); |
| 190 | n = read(reqfd, (char *)&request, sizeof(request)); |
| 191 | PDEBUG("after read, n=%d\n", n); |
| 192 | switch(request.req_kind) { |
| 193 | case REQ_CREATE: |
| 194 | PDEBUG("got REQ_CREATE\n"); |
| 195 | request.req_thread->p_retcode = |
| 196 | pthread_handle_create((pthread_t *) &request.req_thread->p_retval, |
| 197 | request.req_args.create.attr, |
| 198 | request.req_args.create.fn, |
| 199 | request.req_args.create.arg, |
| 200 | &request.req_args.create.mask, |
| 201 | request.req_thread->p_pid, |
| 202 | request.req_thread->p_report_events, |
| 203 | &request.req_thread->p_eventbuf.eventmask); |
| 204 | PDEBUG("restarting %p\n", request.req_thread); |
| 205 | restart(request.req_thread); |
| 206 | break; |
| 207 | case REQ_FREE: |
| 208 | PDEBUG("got REQ_FREE\n"); |
| 209 | pthread_handle_free(request.req_args.free.thread_id); |
| 210 | break; |
| 211 | case REQ_PROCESS_EXIT: |
| 212 | PDEBUG("got REQ_PROCESS_EXIT from %p, exit code = %d\n", |
| 213 | request.req_thread, request.req_args.exit.code); |
| 214 | pthread_handle_exit(request.req_thread, |
| 215 | request.req_args.exit.code); |
| 216 | break; |
| 217 | case REQ_MAIN_THREAD_EXIT: |
| 218 | PDEBUG("got REQ_MAIN_THREAD_EXIT\n"); |
| 219 | main_thread_exiting = 1; |
| 220 | /* Reap children in case all other threads died and the signal handler |
| 221 | went off before we set main_thread_exiting to 1, and therefore did |
| 222 | not do REQ_KICK. */ |
| 223 | pthread_reap_children(); |
| 224 | |
| 225 | if (__pthread_main_thread->p_nextlive == __pthread_main_thread) { |
| 226 | restart(__pthread_main_thread); |
| 227 | /* The main thread will now call exit() which will trigger an |
| 228 | __on_exit handler, which in turn will send REQ_PROCESS_EXIT |
| 229 | to the thread manager. In case you are wondering how the |
| 230 | manager terminates from its loop here. */ |
| 231 | } |
| 232 | break; |
| 233 | case REQ_POST: |
| 234 | PDEBUG("got REQ_POST\n"); |
| 235 | __new_sem_post(request.req_args.post); |
| 236 | break; |
| 237 | case REQ_DEBUG: |
| 238 | PDEBUG("got REQ_DEBUG\n"); |
| 239 | /* Make gdb aware of new thread and gdb will restart the |
| 240 | new thread when it is ready to handle the new thread. */ |
| 241 | if (__pthread_threads_debug && __pthread_sig_debug > 0) { |
| 242 | PDEBUG("about to call raise(__pthread_sig_debug)\n"); |
| 243 | raise(__pthread_sig_debug); |
| 244 | } |
| 245 | case REQ_KICK: |
| 246 | /* This is just a prod to get the manager to reap some |
| 247 | threads right away, avoiding a potential delay at shutdown. */ |
| 248 | break; |
| 249 | } |
| 250 | } |
| 251 | } |
| 252 | } |
| 253 | |
| 254 | int attribute_noreturn __pthread_manager_event(void *arg) |
| 255 | { |
| 256 | /* If we have special thread_self processing, initialize it. */ |
| 257 | #ifdef INIT_THREAD_SELF |
| 258 | INIT_THREAD_SELF(&__pthread_manager_thread, 1); |
| 259 | #endif |
| 260 | |
| 261 | /* Get the lock the manager will free once all is correctly set up. */ |
| 262 | __pthread_lock (THREAD_GETMEM((&__pthread_manager_thread), p_lock), NULL); |
| 263 | /* Free it immediately. */ |
| 264 | __pthread_unlock (THREAD_GETMEM((&__pthread_manager_thread), p_lock)); |
| 265 | |
| 266 | __pthread_manager(arg); |
| 267 | } |
| 268 | |
| 269 | /* Process creation */ |
| 270 | static int |
| 271 | attribute_noreturn |
| 272 | pthread_start_thread(void *arg) |
| 273 | { |
| 274 | pthread_descr self = (pthread_descr) arg; |
| 275 | struct pthread_request request; |
| 276 | void * outcome; |
| 277 | /* Initialize special thread_self processing, if any. */ |
| 278 | #ifdef INIT_THREAD_SELF |
| 279 | INIT_THREAD_SELF(self, self->p_nr); |
| 280 | #endif |
| 281 | PDEBUG("\n"); |
| 282 | /* Make sure our pid field is initialized, just in case we get there |
| 283 | before our father has initialized it. */ |
| 284 | THREAD_SETMEM(self, p_pid, getpid()); |
| 285 | /* Initial signal mask is that of the creating thread. (Otherwise, |
| 286 | we'd just inherit the mask of the thread manager.) */ |
| 287 | sigprocmask(SIG_SETMASK, &self->p_start_args.mask, NULL); |
| 288 | /* Set the scheduling policy and priority for the new thread, if needed */ |
| 289 | if (THREAD_GETMEM(self, p_start_args.schedpolicy) >= 0) |
| 290 | /* Explicit scheduling attributes were provided: apply them */ |
| 291 | sched_setscheduler(THREAD_GETMEM(self, p_pid), |
| 292 | THREAD_GETMEM(self, p_start_args.schedpolicy), |
| 293 | &self->p_start_args.schedparam); |
| 294 | else if (__pthread_manager_thread.p_priority > 0) |
| 295 | /* Default scheduling required, but thread manager runs in realtime |
| 296 | scheduling: switch new thread to SCHED_OTHER policy */ |
| 297 | { |
| 298 | struct sched_param default_params; |
| 299 | default_params.sched_priority = 0; |
| 300 | sched_setscheduler(THREAD_GETMEM(self, p_pid), |
| 301 | SCHED_OTHER, &default_params); |
| 302 | } |
| 303 | /* Make gdb aware of new thread */ |
| 304 | if (__pthread_threads_debug && __pthread_sig_debug > 0) { |
| 305 | request.req_thread = self; |
| 306 | request.req_kind = REQ_DEBUG; |
| 307 | TEMP_FAILURE_RETRY(write(__pthread_manager_request, |
| 308 | (char *) &request, sizeof(request))); |
| 309 | suspend(self); |
| 310 | } |
| 311 | /* Run the thread code */ |
| 312 | outcome = self->p_start_args.start_routine(THREAD_GETMEM(self, |
| 313 | p_start_args.arg)); |
| 314 | /* Exit with the given return value */ |
| 315 | __pthread_do_exit(outcome, CURRENT_STACK_FRAME); |
| 316 | } |
| 317 | |
| 318 | static int |
| 319 | attribute_noreturn |
| 320 | pthread_start_thread_event(void *arg) |
| 321 | { |
| 322 | pthread_descr self = (pthread_descr) arg; |
| 323 | |
| 324 | #ifdef INIT_THREAD_SELF |
| 325 | INIT_THREAD_SELF(self, self->p_nr); |
| 326 | #endif |
| 327 | /* Make sure our pid field is initialized, just in case we get there |
| 328 | before our father has initialized it. */ |
| 329 | THREAD_SETMEM(self, p_pid, getpid()); |
| 330 | /* Get the lock the manager will free once all is correctly set up. */ |
| 331 | __pthread_lock (THREAD_GETMEM(self, p_lock), NULL); |
| 332 | /* Free it immediately. */ |
| 333 | __pthread_unlock (THREAD_GETMEM(self, p_lock)); |
| 334 | |
| 335 | /* Continue with the real function. */ |
| 336 | pthread_start_thread (arg); |
| 337 | } |
| 338 | |
| 339 | static int pthread_allocate_stack(const pthread_attr_t *attr, |
| 340 | pthread_descr default_new_thread, |
| 341 | int pagesize, |
| 342 | pthread_descr * out_new_thread, |
| 343 | char ** out_new_thread_bottom, |
| 344 | char ** out_guardaddr, |
| 345 | size_t * out_guardsize) |
| 346 | { |
| 347 | pthread_descr new_thread; |
| 348 | char * new_thread_bottom; |
| 349 | char * guardaddr; |
| 350 | size_t stacksize, guardsize; |
| 351 | |
| 352 | if (attr != NULL && attr->__stackaddr_set) |
| 353 | { |
| 354 | /* The user provided a stack. */ |
| 355 | new_thread = (pthread_descr) ((long)(attr->__stackaddr) & -sizeof(void *)) - 1; |
| 356 | new_thread_bottom = (char *) attr->__stackaddr - attr->__stacksize; |
| 357 | guardaddr = NULL; |
| 358 | guardsize = 0; |
| 359 | __pthread_nonstandard_stacks = 1; |
| 360 | #ifndef __ARCH_USE_MMU__ |
| 361 | /* check the initial thread stack boundaries so they don't overlap */ |
| 362 | NOMMU_INITIAL_THREAD_BOUNDS((char *) new_thread, (char *) new_thread_bottom); |
| 363 | |
| 364 | PDEBUG("initial stack: bos=%p, tos=%p\n", __pthread_initial_thread_bos, |
| 365 | __pthread_initial_thread_tos); |
| 366 | #endif |
| 367 | } |
| 368 | else |
| 369 | { |
| 370 | #ifdef __ARCH_USE_MMU__ |
| 371 | stacksize = STACK_SIZE - pagesize; |
| 372 | if (attr != NULL) |
| 373 | stacksize = MIN(stacksize, roundup(attr->__stacksize, pagesize)); |
| 374 | /* Allocate space for stack and thread descriptor at default address */ |
| 375 | new_thread = default_new_thread; |
| 376 | new_thread_bottom = (char *) (new_thread + 1) - stacksize; |
| 377 | if (mmap((caddr_t)((char *)(new_thread + 1) - INITIAL_STACK_SIZE), |
| 378 | INITIAL_STACK_SIZE, PROT_READ | PROT_WRITE | PROT_EXEC, |
| 379 | MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED | MAP_GROWSDOWN, |
| 380 | -1, 0) == MAP_FAILED) |
| 381 | /* Bad luck, this segment is already mapped. */ |
| 382 | return -1; |
| 383 | /* We manage to get a stack. Now see whether we need a guard |
| 384 | and allocate it if necessary. Notice that the default |
| 385 | attributes (stack_size = STACK_SIZE - pagesize) do not need |
| 386 | a guard page, since the RLIMIT_STACK soft limit prevents stacks |
| 387 | from running into one another. */ |
| 388 | if (stacksize == (size_t) (STACK_SIZE - pagesize)) |
| 389 | { |
| 390 | /* We don't need a guard page. */ |
| 391 | guardaddr = NULL; |
| 392 | guardsize = 0; |
| 393 | } |
| 394 | else |
| 395 | { |
| 396 | /* Put a bad page at the bottom of the stack */ |
| 397 | guardsize = attr->__guardsize; |
| 398 | guardaddr = (void *)new_thread_bottom - guardsize; |
| 399 | if (mmap((caddr_t) guardaddr, guardsize, 0, MAP_FIXED, -1, 0) |
| 400 | == MAP_FAILED) |
| 401 | { |
| 402 | /* We don't make this an error. */ |
| 403 | guardaddr = NULL; |
| 404 | guardsize = 0; |
| 405 | } |
| 406 | } |
| 407 | #else |
| 408 | /* We cannot mmap to this huge chunk of stack space when we don't have |
| 409 | * an MMU. Pretend we are using a user provided stack even if there was |
| 410 | * none provided by the user. Thus, we get around the mmap and reservation |
| 411 | * of a huge stack segment. -StS */ |
| 412 | |
| 413 | stacksize = INITIAL_STACK_SIZE; |
| 414 | /* The user may want to use a non-default stacksize */ |
| 415 | if (attr != NULL) |
| 416 | { |
| 417 | stacksize = attr->__stacksize; |
| 418 | } |
| 419 | |
| 420 | /* malloc a stack - memory from the bottom up */ |
| 421 | if ((new_thread_bottom = malloc(stacksize)) == NULL) |
| 422 | { |
| 423 | /* bad luck, we cannot malloc any more */ |
| 424 | return -1 ; |
| 425 | } |
| 426 | PDEBUG("malloced chunk: base=%p, size=0x%04x\n", new_thread_bottom, stacksize); |
| 427 | |
| 428 | /* Set up the pointers. new_thread marks the TOP of the stack frame and |
| 429 | * the address of the pthread_descr struct at the same time. Therefore we |
| 430 | * must account for its size and fit it in the malloc()'ed block. The |
| 431 | * value of `new_thread' is then passed to clone() as the stack argument. |
| 432 | * |
| 433 | * ^ +------------------------+ |
| 434 | * | | pthread_descr struct | |
| 435 | * | +------------------------+ <- new_thread |
| 436 | * malloc block | | | |
| 437 | * | | thread stack | |
| 438 | * | | | |
| 439 | * v +------------------------+ <- new_thread_bottom |
| 440 | * |
| 441 | * Note: The calculated value of new_thread must be word aligned otherwise |
| 442 | * the kernel chokes on a non-aligned stack frame. Choose the lower |
| 443 | * available word boundary. |
| 444 | */ |
| 445 | new_thread = ((pthread_descr) ((int)(new_thread_bottom + stacksize) & -sizeof(void*))) - 1; |
| 446 | guardaddr = NULL; |
| 447 | guardsize = 0; |
| 448 | |
| 449 | PDEBUG("thread stack: bos=%p, tos=%p\n", new_thread_bottom, new_thread); |
| 450 | |
| 451 | /* check the initial thread stack boundaries so they don't overlap */ |
| 452 | NOMMU_INITIAL_THREAD_BOUNDS((char *) new_thread, (char *) new_thread_bottom); |
| 453 | |
| 454 | PDEBUG("initial stack: bos=%p, tos=%p\n", __pthread_initial_thread_bos, |
| 455 | __pthread_initial_thread_tos); |
| 456 | |
| 457 | /* on non-MMU systems we always have non-standard stack frames */ |
| 458 | __pthread_nonstandard_stacks = 1; |
| 459 | |
| 460 | #endif /* __ARCH_USE_MMU__ */ |
| 461 | } |
| 462 | |
| 463 | /* Clear the thread data structure. */ |
| 464 | memset (new_thread, '\0', sizeof (*new_thread)); |
| 465 | *out_new_thread = new_thread; |
| 466 | *out_new_thread_bottom = new_thread_bottom; |
| 467 | *out_guardaddr = guardaddr; |
| 468 | *out_guardsize = guardsize; |
| 469 | return 0; |
| 470 | } |
| 471 | |
| 472 | static int pthread_handle_create(pthread_t *thread, const pthread_attr_t *attr, |
| 473 | void * (*start_routine)(void *), void *arg, |
| 474 | sigset_t * mask, int father_pid, |
| 475 | int report_events, |
| 476 | td_thr_events_t *event_maskp) |
| 477 | { |
| 478 | size_t sseg; |
| 479 | int pid; |
| 480 | pthread_descr new_thread; |
| 481 | char * new_thread_bottom; |
| 482 | char * new_thread_top; |
| 483 | pthread_t new_thread_id; |
| 484 | char *guardaddr = NULL; |
| 485 | size_t guardsize = 0; |
| 486 | int pagesize = getpagesize(); |
| 487 | int saved_errno = 0; |
| 488 | |
| 489 | /* First check whether we have to change the policy and if yes, whether |
| 490 | we can do this. Normally this should be done by examining the |
| 491 | return value of the sched_setscheduler call in pthread_start_thread |
| 492 | but this is hard to implement. FIXME */ |
| 493 | if (attr != NULL && attr->__schedpolicy != SCHED_OTHER && geteuid () != 0) |
| 494 | return EPERM; |
| 495 | /* Find a free segment for the thread, and allocate a stack if needed */ |
| 496 | for (sseg = 2; ; sseg++) |
| 497 | { |
| 498 | if (sseg >= PTHREAD_THREADS_MAX) |
| 499 | return EAGAIN; |
| 500 | if (__pthread_handles[sseg].h_descr != NULL) |
| 501 | continue; |
| 502 | if (pthread_allocate_stack(attr, thread_segment(sseg), pagesize, |
| 503 | &new_thread, &new_thread_bottom, |
| 504 | &guardaddr, &guardsize) == 0) |
| 505 | break; |
| 506 | #ifndef __ARCH_USE_MMU__ |
| 507 | else |
| 508 | /* When there is MMU, mmap () is used to allocate the stack. If one |
| 509 | * segment is already mapped, we should continue to see if we can |
| 510 | * use the next one. However, when there is no MMU, malloc () is used. |
| 511 | * It's waste of CPU cycles to continue to try if it fails. */ |
| 512 | return EAGAIN; |
| 513 | #endif |
| 514 | } |
| 515 | __pthread_handles_num++; |
| 516 | /* Allocate new thread identifier */ |
| 517 | pthread_threads_counter += PTHREAD_THREADS_MAX; |
| 518 | new_thread_id = sseg + pthread_threads_counter; |
| 519 | /* Initialize the thread descriptor. Elements which have to be |
| 520 | initialized to zero already have this value. */ |
| 521 | new_thread->p_tid = new_thread_id; |
| 522 | new_thread->p_lock = &(__pthread_handles[sseg].h_lock); |
| 523 | new_thread->p_cancelstate = PTHREAD_CANCEL_ENABLE; |
| 524 | new_thread->p_canceltype = PTHREAD_CANCEL_DEFERRED; |
| 525 | new_thread->p_errnop = &new_thread->p_errno; |
| 526 | new_thread->p_h_errnop = &new_thread->p_h_errno; |
| 527 | #ifdef __UCLIBC_HAS_XLOCALE__ |
| 528 | /* Initialize thread's locale to the global locale. */ |
| 529 | new_thread->locale = __global_locale; |
| 530 | #endif /* __UCLIBC_HAS_XLOCALE__ */ |
| 531 | new_thread->p_guardaddr = guardaddr; |
| 532 | new_thread->p_guardsize = guardsize; |
| 533 | new_thread->p_self = new_thread; |
| 534 | new_thread->p_nr = sseg; |
| 535 | /* Initialize the thread handle */ |
| 536 | __pthread_init_lock(&__pthread_handles[sseg].h_lock); |
| 537 | __pthread_handles[sseg].h_descr = new_thread; |
| 538 | __pthread_handles[sseg].h_bottom = new_thread_bottom; |
| 539 | /* Determine scheduling parameters for the thread */ |
| 540 | new_thread->p_start_args.schedpolicy = -1; |
| 541 | if (attr != NULL) { |
| 542 | new_thread->p_detached = attr->__detachstate; |
| 543 | new_thread->p_userstack = attr->__stackaddr_set; |
| 544 | |
| 545 | switch(attr->__inheritsched) { |
| 546 | case PTHREAD_EXPLICIT_SCHED: |
| 547 | new_thread->p_start_args.schedpolicy = attr->__schedpolicy; |
| 548 | memcpy (&new_thread->p_start_args.schedparam, &attr->__schedparam, |
| 549 | sizeof (struct sched_param)); |
| 550 | break; |
| 551 | case PTHREAD_INHERIT_SCHED: |
| 552 | new_thread->p_start_args.schedpolicy = sched_getscheduler(father_pid); |
| 553 | sched_getparam(father_pid, &new_thread->p_start_args.schedparam); |
| 554 | break; |
| 555 | } |
| 556 | new_thread->p_priority = |
| 557 | new_thread->p_start_args.schedparam.sched_priority; |
| 558 | } |
| 559 | /* Finish setting up arguments to pthread_start_thread */ |
| 560 | new_thread->p_start_args.start_routine = start_routine; |
| 561 | new_thread->p_start_args.arg = arg; |
| 562 | new_thread->p_start_args.mask = *mask; |
| 563 | /* Raise priority of thread manager if needed */ |
| 564 | __pthread_manager_adjust_prio(new_thread->p_priority); |
| 565 | /* Do the cloning. We have to use two different functions depending |
| 566 | on whether we are debugging or not. */ |
| 567 | pid = 0; /* Note that the thread never can have PID zero. */ |
| 568 | new_thread_top = ((char *)new_thread - THREAD_STACK_OFFSET); |
| 569 | |
| 570 | /* ******************************************************** */ |
| 571 | /* This code was moved from below to cope with running threads |
| 572 | * on uClinux systems. See comment below... |
| 573 | * Insert new thread in doubly linked list of active threads */ |
| 574 | new_thread->p_prevlive = __pthread_main_thread; |
| 575 | new_thread->p_nextlive = __pthread_main_thread->p_nextlive; |
| 576 | __pthread_main_thread->p_nextlive->p_prevlive = new_thread; |
| 577 | __pthread_main_thread->p_nextlive = new_thread; |
| 578 | /* ********************************************************* */ |
| 579 | |
| 580 | if (report_events) |
| 581 | { |
| 582 | /* See whether the TD_CREATE event bit is set in any of the |
| 583 | masks. */ |
| 584 | int idx = __td_eventword (TD_CREATE); |
| 585 | uint32_t m = __td_eventmask (TD_CREATE); |
| 586 | |
| 587 | if ((m & (__pthread_threads_events.event_bits[idx] |
| 588 | | event_maskp->event_bits[idx])) != 0) |
| 589 | { |
| 590 | /* Lock the mutex the child will use now so that it will stop. */ |
| 591 | __pthread_lock(new_thread->p_lock, NULL); |
| 592 | |
| 593 | /* We have to report this event. */ |
| 594 | #ifdef __ia64__ |
| 595 | pid = __clone2(pthread_start_thread_event, new_thread_top, |
| 596 | new_thread_top - new_thread_bottom, |
| 597 | CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND | |
| 598 | __pthread_sig_cancel, new_thread); |
| 599 | #else |
| 600 | pid = clone(pthread_start_thread_event, new_thread_top, |
| 601 | CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND | |
| 602 | __pthread_sig_cancel, new_thread); |
| 603 | #endif |
| 604 | |
| 605 | saved_errno = errno; |
| 606 | if (pid != -1) |
| 607 | { |
| 608 | /* Now fill in the information about the new thread in |
| 609 | the newly created thread's data structure. We cannot let |
| 610 | the new thread do this since we don't know whether it was |
| 611 | already scheduled when we send the event. */ |
| 612 | new_thread->p_eventbuf.eventdata = new_thread; |
| 613 | new_thread->p_eventbuf.eventnum = TD_CREATE; |
| 614 | __pthread_last_event = new_thread; |
| 615 | |
| 616 | /* We have to set the PID here since the callback function |
| 617 | in the debug library will need it and we cannot guarantee |
| 618 | the child got scheduled before the debugger. */ |
| 619 | new_thread->p_pid = pid; |
| 620 | |
| 621 | /* Now call the function which signals the event. */ |
| 622 | __linuxthreads_create_event (); |
| 623 | |
| 624 | /* Now restart the thread. */ |
| 625 | __pthread_unlock(new_thread->p_lock); |
| 626 | } |
| 627 | } |
| 628 | } |
| 629 | if (pid == 0) |
| 630 | { |
| 631 | PDEBUG("cloning new_thread = %p\n", new_thread); |
| 632 | #ifdef __ia64__ |
| 633 | pid = __clone2(pthread_start_thread, new_thread_top, |
| 634 | new_thread_top - new_thread_bottom, |
| 635 | CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND | |
| 636 | __pthread_sig_cancel, new_thread); |
| 637 | #else |
| 638 | pid = clone(pthread_start_thread, new_thread_top, |
| 639 | CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND | |
| 640 | __pthread_sig_cancel, new_thread); |
| 641 | #endif |
| 642 | saved_errno = errno; |
| 643 | } |
| 644 | /* Check if cloning succeeded */ |
| 645 | if (pid == -1) { |
| 646 | /******************************************************** |
| 647 | * Code inserted to remove the thread from our list of active |
| 648 | * threads in case of failure (needed to cope with uClinux), |
| 649 | * See comment below. */ |
| 650 | new_thread->p_nextlive->p_prevlive = new_thread->p_prevlive; |
| 651 | new_thread->p_prevlive->p_nextlive = new_thread->p_nextlive; |
| 652 | /********************************************************/ |
| 653 | |
| 654 | /* Free the stack if we allocated it */ |
| 655 | if (attr == NULL || !attr->__stackaddr_set) |
| 656 | { |
| 657 | #ifdef __ARCH_USE_MMU__ |
| 658 | if (new_thread->p_guardsize != 0) |
| 659 | munmap(new_thread->p_guardaddr, new_thread->p_guardsize); |
| 660 | munmap((caddr_t)((char *)(new_thread+1) - INITIAL_STACK_SIZE), |
| 661 | INITIAL_STACK_SIZE); |
| 662 | #else |
| 663 | free(new_thread_bottom); |
| 664 | #endif /* __ARCH_USE_MMU__ */ |
| 665 | } |
| 666 | __pthread_handles[sseg].h_descr = NULL; |
| 667 | __pthread_handles[sseg].h_bottom = NULL; |
| 668 | __pthread_handles_num--; |
| 669 | return saved_errno; |
| 670 | } |
| 671 | PDEBUG("new thread pid = %d\n", pid); |
| 672 | |
| 673 | #if 0 |
| 674 | /* *********************************************************** |
| 675 | This code has been moved before the call to clone(). In uClinux, |
| 676 | the use of wait on a semaphore is dependant upon that the child so |
| 677 | the child must be in the active threads list. This list is used in |
| 678 | pthread_find_self() to get the pthread_descr of self. So, if the |
| 679 | child calls sem_wait before this code is executed , it will hang |
| 680 | forever and initial_thread will instead be posted by a sem_post |
| 681 | call. */ |
| 682 | |
| 683 | /* Insert new thread in doubly linked list of active threads */ |
| 684 | new_thread->p_prevlive = __pthread_main_thread; |
| 685 | new_thread->p_nextlive = __pthread_main_thread->p_nextlive; |
| 686 | __pthread_main_thread->p_nextlive->p_prevlive = new_thread; |
| 687 | __pthread_main_thread->p_nextlive = new_thread; |
| 688 | /************************************************************/ |
| 689 | #endif |
| 690 | |
| 691 | /* Set pid field of the new thread, in case we get there before the |
| 692 | child starts. */ |
| 693 | new_thread->p_pid = pid; |
| 694 | /* We're all set */ |
| 695 | *thread = new_thread_id; |
| 696 | return 0; |
| 697 | } |
| 698 | |
| 699 | |
| 700 | /* Try to free the resources of a thread when requested by pthread_join |
| 701 | or pthread_detach on a terminated thread. */ |
| 702 | |
| 703 | static void pthread_free(pthread_descr th) |
| 704 | { |
| 705 | pthread_handle handle; |
| 706 | pthread_readlock_info *iter, *next; |
| 707 | #ifndef __ARCH_USE_MMU__ |
| 708 | char *h_bottom_save; |
| 709 | #endif |
| 710 | |
| 711 | /* Make the handle invalid */ |
| 712 | handle = thread_handle(th->p_tid); |
| 713 | __pthread_lock(&handle->h_lock, NULL); |
| 714 | #ifndef __ARCH_USE_MMU__ |
| 715 | h_bottom_save = handle->h_bottom; |
| 716 | #endif |
| 717 | handle->h_descr = NULL; |
| 718 | handle->h_bottom = (char *)(-1L); |
| 719 | __pthread_unlock(&handle->h_lock); |
| 720 | #ifdef FREE_THREAD_SELF |
| 721 | FREE_THREAD_SELF(th, th->p_nr); |
| 722 | #endif |
| 723 | /* One fewer threads in __pthread_handles */ |
| 724 | __pthread_handles_num--; |
| 725 | |
| 726 | /* Destroy read lock list, and list of free read lock structures. |
| 727 | If the former is not empty, it means the thread exited while |
| 728 | holding read locks! */ |
| 729 | |
| 730 | for (iter = th->p_readlock_list; iter != NULL; iter = next) |
| 731 | { |
| 732 | next = iter->pr_next; |
| 733 | free(iter); |
| 734 | } |
| 735 | |
| 736 | for (iter = th->p_readlock_free; iter != NULL; iter = next) |
| 737 | { |
| 738 | next = iter->pr_next; |
| 739 | free(iter); |
| 740 | } |
| 741 | |
| 742 | /* If initial thread, nothing to free */ |
| 743 | if (th == &__pthread_initial_thread) return; |
| 744 | if (!th->p_userstack) |
| 745 | { |
| 746 | #ifdef __ARCH_USE_MMU__ |
| 747 | /* Free the stack and thread descriptor area */ |
| 748 | if (th->p_guardsize != 0) |
| 749 | munmap(th->p_guardaddr, th->p_guardsize); |
| 750 | munmap((caddr_t) ((char *)(th+1) - STACK_SIZE), STACK_SIZE); |
| 751 | #else |
| 752 | /* For non-MMU systems we always malloc the stack, so free it here. -StS */ |
| 753 | free(h_bottom_save); |
| 754 | #endif /* __ARCH_USE_MMU__ */ |
| 755 | } |
| 756 | } |
| 757 | |
| 758 | /* Handle threads that have exited */ |
| 759 | |
| 760 | static void pthread_exited(pid_t pid) |
| 761 | { |
| 762 | pthread_descr th; |
| 763 | int detached; |
| 764 | /* Find thread with that pid */ |
| 765 | for (th = __pthread_main_thread->p_nextlive; |
| 766 | th != __pthread_main_thread; |
| 767 | th = th->p_nextlive) { |
| 768 | if (th->p_pid == pid) { |
| 769 | /* Remove thread from list of active threads */ |
| 770 | th->p_nextlive->p_prevlive = th->p_prevlive; |
| 771 | th->p_prevlive->p_nextlive = th->p_nextlive; |
| 772 | /* Mark thread as exited, and if detached, free its resources */ |
| 773 | __pthread_lock(th->p_lock, NULL); |
| 774 | th->p_exited = 1; |
| 775 | /* If we have to signal this event do it now. */ |
| 776 | if (th->p_report_events) |
| 777 | { |
| 778 | /* See whether TD_REAP is in any of the mask. */ |
| 779 | int idx = __td_eventword (TD_REAP); |
| 780 | uint32_t mask = __td_eventmask (TD_REAP); |
| 781 | |
| 782 | if ((mask & (__pthread_threads_events.event_bits[idx] |
| 783 | | th->p_eventbuf.eventmask.event_bits[idx])) != 0) |
| 784 | { |
| 785 | /* Yep, we have to signal the reapage. */ |
| 786 | th->p_eventbuf.eventnum = TD_REAP; |
| 787 | th->p_eventbuf.eventdata = th; |
| 788 | __pthread_last_event = th; |
| 789 | |
| 790 | /* Now call the function to signal the event. */ |
| 791 | __linuxthreads_reap_event(); |
| 792 | } |
| 793 | } |
| 794 | detached = th->p_detached; |
| 795 | __pthread_unlock(th->p_lock); |
| 796 | if (detached) |
| 797 | pthread_free(th); |
| 798 | break; |
| 799 | } |
| 800 | } |
| 801 | /* If all threads have exited and the main thread is pending on a |
| 802 | pthread_exit, wake up the main thread and terminate ourselves. */ |
| 803 | if (main_thread_exiting && |
| 804 | __pthread_main_thread->p_nextlive == __pthread_main_thread) { |
| 805 | restart(__pthread_main_thread); |
| 806 | /* Same logic as REQ_MAIN_THREAD_EXIT. */ |
| 807 | } |
| 808 | } |
| 809 | |
| 810 | static void pthread_reap_children(void) |
| 811 | { |
| 812 | pid_t pid; |
| 813 | int status; |
| 814 | PDEBUG("\n"); |
| 815 | |
| 816 | while ((pid = waitpid(-1, &status, WNOHANG | __WCLONE)) > 0) { |
| 817 | pthread_exited(pid); |
| 818 | if (WIFSIGNALED(status)) { |
| 819 | /* If a thread died due to a signal, send the same signal to |
| 820 | all other threads, including the main thread. */ |
| 821 | pthread_kill_all_threads(WTERMSIG(status), 1); |
| 822 | _exit(0); |
| 823 | } |
| 824 | } |
| 825 | } |
| 826 | |
| 827 | /* Try to free the resources of a thread when requested by pthread_join |
| 828 | or pthread_detach on a terminated thread. */ |
| 829 | |
| 830 | static void pthread_handle_free(pthread_t th_id) |
| 831 | { |
| 832 | pthread_handle handle = thread_handle(th_id); |
| 833 | pthread_descr th; |
| 834 | |
| 835 | __pthread_lock(&handle->h_lock, NULL); |
| 836 | if (invalid_handle(handle, th_id)) { |
| 837 | /* pthread_reap_children has deallocated the thread already, |
| 838 | nothing needs to be done */ |
| 839 | __pthread_unlock(&handle->h_lock); |
| 840 | return; |
| 841 | } |
| 842 | th = handle->h_descr; |
| 843 | if (th->p_exited) { |
| 844 | __pthread_unlock(&handle->h_lock); |
| 845 | pthread_free(th); |
| 846 | } else { |
| 847 | /* The Unix process of the thread is still running. |
| 848 | Mark the thread as detached so that the thread manager will |
| 849 | deallocate its resources when the Unix process exits. */ |
| 850 | th->p_detached = 1; |
| 851 | __pthread_unlock(&handle->h_lock); |
| 852 | } |
| 853 | } |
| 854 | |
| 855 | /* Send a signal to all running threads */ |
| 856 | |
| 857 | static void pthread_kill_all_threads(int sig, int main_thread_also) |
| 858 | { |
| 859 | pthread_descr th; |
| 860 | for (th = __pthread_main_thread->p_nextlive; |
| 861 | th != __pthread_main_thread; |
| 862 | th = th->p_nextlive) { |
| 863 | kill(th->p_pid, sig); |
| 864 | } |
| 865 | if (main_thread_also) { |
| 866 | kill(__pthread_main_thread->p_pid, sig); |
| 867 | } |
| 868 | } |
| 869 | |
| 870 | /* Process-wide exit() */ |
| 871 | |
| 872 | static void pthread_handle_exit(pthread_descr issuing_thread, int exitcode) |
| 873 | { |
| 874 | pthread_descr th; |
| 875 | __pthread_exit_requested = 1; |
| 876 | __pthread_exit_code = exitcode; |
| 877 | /* Send the CANCEL signal to all running threads, including the main |
| 878 | thread, but excluding the thread from which the exit request originated |
| 879 | (that thread must complete the exit, e.g. calling atexit functions |
| 880 | and flushing stdio buffers). */ |
| 881 | for (th = issuing_thread->p_nextlive; |
| 882 | th != issuing_thread; |
| 883 | th = th->p_nextlive) { |
| 884 | kill(th->p_pid, __pthread_sig_cancel); |
| 885 | } |
| 886 | /* Now, wait for all these threads, so that they don't become zombies |
| 887 | and their times are properly added to the thread manager's times. */ |
| 888 | for (th = issuing_thread->p_nextlive; |
| 889 | th != issuing_thread; |
| 890 | th = th->p_nextlive) { |
| 891 | waitpid(th->p_pid, NULL, __WCLONE); |
| 892 | } |
| 893 | restart(issuing_thread); |
| 894 | _exit(0); |
| 895 | } |
| 896 | |
| 897 | /* Handler for __pthread_sig_cancel in thread manager thread */ |
| 898 | |
| 899 | void __pthread_manager_sighandler(int sig attribute_unused) |
| 900 | { |
| 901 | int kick_manager = terminated_children == 0 && main_thread_exiting; |
| 902 | terminated_children = 1; |
| 903 | |
| 904 | /* If the main thread is terminating, kick the thread manager loop |
| 905 | each time some threads terminate. This eliminates a two second |
| 906 | shutdown delay caused by the thread manager sleeping in the |
| 907 | call to __poll(). Instead, the thread manager is kicked into |
| 908 | action, reaps the outstanding threads and resumes the main thread |
| 909 | so that it can complete the shutdown. */ |
| 910 | |
| 911 | if (kick_manager) { |
| 912 | struct pthread_request request; |
| 913 | request.req_thread = 0; |
| 914 | request.req_kind = REQ_KICK; |
| 915 | TEMP_FAILURE_RETRY(write(__pthread_manager_request, |
| 916 | (char *) &request, sizeof(request))); |
| 917 | } |
| 918 | } |
| 919 | |
| 920 | /* Adjust priority of thread manager so that it always run at a priority |
| 921 | higher than all threads */ |
| 922 | |
| 923 | void __pthread_manager_adjust_prio(int thread_prio) |
| 924 | { |
| 925 | struct sched_param param; |
| 926 | |
| 927 | if (thread_prio <= __pthread_manager_thread.p_priority) return; |
| 928 | param.sched_priority = |
| 929 | thread_prio < sched_get_priority_max(SCHED_FIFO) |
| 930 | ? thread_prio + 1 : thread_prio; |
| 931 | sched_setscheduler(__pthread_manager_thread.p_pid, SCHED_FIFO, ¶m); |
| 932 | __pthread_manager_thread.p_priority = thread_prio; |
| 933 | } |