src/thread/pthread_create.c

   1 #define _GNU_SOURCE
   2 #include "pthread_impl.h"
   3 #include "stdio_impl.h"
   4 #include "libc.h"
   5 #include "lock.h"
   6 #include <sys/mman.h>
   7 #include <string.h>
   8 #include <stddef.h>
   9
  10 static void dummy_0()
  11 {
  12 }
  13 weak_alias(dummy_0, __acquire_ptc);
  14 weak_alias(dummy_0, __release_ptc);
  15 weak_alias(dummy_0, __pthread_tsd_run_dtors);
  16 weak_alias(dummy_0, __do_orphaned_stdio_locks);
  17 weak_alias(dummy_0, __dl_thread_cleanup);
  18 weak_alias(dummy_0, __membarrier_init);
  19
  20 static int tl_lock_count;
  21 static int tl_lock_waiters;
  22
  23 void __tl_lock(void)
  24 {
  25         int tid = __pthread_self()->tid;
  26         int val = __thread_list_lock;
  27         if (val == tid) {
  28                 tl_lock_count++;
  29                 return;
  30         }
  31         while ((val = a_cas(&__thread_list_lock, 0, tid)))
  32                 __wait(&__thread_list_lock, &tl_lock_waiters, val, 0);
  33 }
  34
  35 void __tl_unlock(void)
  36 {
  37         if (tl_lock_count) {
  38                 tl_lock_count--;
  39                 return;
  40         }
  41         a_store(&__thread_list_lock, 0);
  42         if (tl_lock_waiters) __wake(&__thread_list_lock, 1, 0);
  43 }
  44
  45 void __tl_sync(pthread_t td)
  46 {
  47         a_barrier();
  48         int val = __thread_list_lock;
  49         if (!val) return;
  50         __wait(&__thread_list_lock, &tl_lock_waiters, val, 0);
  51         if (tl_lock_waiters) __wake(&__thread_list_lock, 1, 0);
  52 }
  53
  54 _Noreturn void __pthread_exit(void *result)
  55 {
  56         pthread_t self = __pthread_self();
  57         sigset_t set;
  58
  59         self->canceldisable = 1;
  60         self->cancelasync = 0;
  61         self->result = result;
  62
  63         while (self->cancelbuf) {
  64                 void (*f)(void *) = self->cancelbuf->__f;
  65                 void *x = self->cancelbuf->__x;
  66                 self->cancelbuf = self->cancelbuf->__next;
  67                 f(x);
  68         }
  69
  70         __pthread_tsd_run_dtors();
  71
  72         __block_app_sigs(&set);
  73
  74         /* This atomic potentially competes with a concurrent pthread_detach
  75          * call; the loser is responsible for freeing thread resources. */
  76         int state = a_cas(&self->detach_state, DT_JOINABLE, DT_EXITING);
  77
  78         if (state==DT_DETACHED && self->map_base) {
  79                 /* Since __unmapself bypasses the normal munmap code path,
  80                  * explicitly wait for vmlock holders first. This must be
  81                  * done before any locks are taken, to avoid lock ordering
  82                  * issues that could lead to deadlock. */
  83                 __vm_wait();
  84         }
  85
  86         /* Access to target the exiting thread with syscalls that use
  87          * its kernel tid is controlled by killlock. For detached threads,
  88          * any use past this point would have undefined behavior, but for
  89          * joinable threads it's a valid usage that must be handled.
  90          * Signals must be blocked since pthread_kill must be AS-safe. */
  91         LOCK(self->killlock);
  92
  93         /* The thread list lock must be AS-safe, and thus depends on
  94          * application signals being blocked above. */
  95         __tl_lock();
  96
  97         /* If this is the only thread in the list, don't proceed with
  98          * termination of the thread, but restore the previous lock and
  99          * signal state to prepare for exit to call atexit handlers. */
 100         if (self->next == self) {
 101                 __tl_unlock();
 102                 UNLOCK(self->killlock);
 103                 self->detach_state = state;
 104                 __restore_sigs(&set);
 105                 exit(0);
 106         }
 107
 108         /* At this point we are committed to thread termination. */
 109
 110         /* Process robust list in userspace to handle non-pshared mutexes
 111          * and the detached thread case where the robust list head will
 112          * be invalid when the kernel would process it. */
 113         __vm_lock();
 114         volatile void *volatile *rp;
 115         while ((rp=self->robust_list.head) && rp != &self->robust_list.head) {
 116                 pthread_mutex_t *m = (void *)((char *)rp
 117                         - offsetof(pthread_mutex_t, _m_next));
 118                 int waiters = m->_m_waiters;
 119                 int priv = (m->_m_type & 128) ^ 128;
 120                 self->robust_list.pending = rp;
 121                 self->robust_list.head = *rp;
 122                 int cont = a_swap(&m->_m_lock, 0x40000000);
 123                 self->robust_list.pending = 0;
 124                 if (cont < 0 || waiters)
 125                         __wake(&m->_m_lock, 1, priv);
 126         }
 127         __vm_unlock();
 128
 129         __do_orphaned_stdio_locks();
 130         __dl_thread_cleanup();
 131
 132         /* Last, unlink thread from the list. This change will not be visible
 133          * until the lock is released, which only happens after SYS_exit
 134          * has been called, via the exit futex address pointing at the lock.
 135          * This needs to happen after any possible calls to LOCK() that might
 136          * skip locking if process appears single-threaded. */
 137         if (!--libc.threads_minus_1) libc.need_locks = -1;
 138         self->next->prev = self->prev;
 139         self->prev->next = self->next;
 140         self->prev = self->next = self;
 141
 142         if (state==DT_DETACHED && self->map_base) {
 143                 /* Detached threads must block even implementation-internal
 144                  * signals, since they will not have a stack in their last
 145                  * moments of existence. */
 146                 __block_all_sigs(&set);
 147
 148                 /* Robust list will no longer be valid, and was already
 149                  * processed above, so unregister it with the kernel. */
 150                 if (self->robust_list.off)
 151                         __syscall(SYS_set_robust_list, 0, 3*sizeof(long));
 152
 153                 /* The following call unmaps the thread's stack mapping
 154                  * and then exits without touching the stack. */
 155                 __unmapself(self->map_base, self->map_size);
 156         }
 157
 158         /* Wake any joiner. */
 159         a_store(&self->detach_state, DT_EXITED);
 160         __wake(&self->detach_state, 1, 1);
 161
 162         /* After the kernel thread exits, its tid may be reused. Clear it
 163          * to prevent inadvertent use and inform functions that would use
 164          * it that it's no longer available. */
 165         self->tid = 0;
 166         UNLOCK(self->killlock);
 167
 168         for (;;) __syscall(SYS_exit, 0);
 169 }
 170
 171 void __do_cleanup_push(struct __ptcb *cb)
 172 {
 173         struct pthread *self = __pthread_self();
 174         cb->__next = self->cancelbuf;
 175         self->cancelbuf = cb;
 176 }
 177
 178 void __do_cleanup_pop(struct __ptcb *cb)
 179 {
 180         __pthread_self()->cancelbuf = cb->__next;
 181 }
 182
 183 struct start_args {
 184         void *(*start_func)(void *);
 185         void *start_arg;
 186         volatile int control;
 187         unsigned long sig_mask[_NSIG/8/sizeof(long)];
 188 };
 189
 190 static int start(void *p)
 191 {
 192         struct start_args *args = p;
 193         int state = args->control;
 194         if (state) {
 195                 if (a_cas(&args->control, 1, 2)==1)
 196                         __wait(&args->control, 0, 2, 1);
 197                 if (args->control) {
 198                         __syscall(SYS_set_tid_address, &args->control);
 199                         for (;;) __syscall(SYS_exit, 0);
 200                 }
 201         }
 202         __syscall(SYS_rt_sigprocmask, SIG_SETMASK, &args->sig_mask, 0, _NSIG/8);
 203         __pthread_exit(args->start_func(args->start_arg));
 204         return 0;
 205 }
 206
 207 static int start_c11(void *p)
 208 {
 209         struct start_args *args = p;
 210         int (*start)(void*) = (int(*)(void*)) args->start_func;
 211         __pthread_exit((void *)(uintptr_t)start(args->start_arg));
 212         return 0;
 213 }
 214
 215 #define ROUND(x) (((x)+PAGE_SIZE-1)&-PAGE_SIZE)
 216
 217 /* pthread_key_create.c overrides this */
 218 static volatile size_t dummy = 0;
 219 weak_alias(dummy, __pthread_tsd_size);
 220 static void *dummy_tsd[1] = { 0 };
 221 weak_alias(dummy_tsd, __pthread_tsd_main);
 222
 223 static FILE *volatile dummy_file = 0;
 224 weak_alias(dummy_file, __stdin_used);
 225 weak_alias(dummy_file, __stdout_used);
 226 weak_alias(dummy_file, __stderr_used);
 227
 228 static void init_file_lock(FILE *f)
 229 {
 230         if (f && f->lock<0) f->lock = 0;
 231 }
 232
 233 int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict attrp, void *(*entry)(void *), void *restrict arg)
 234 {
 235         int ret, c11 = (attrp == __ATTRP_C11_THREAD);
 236         size_t size, guard;
 237         struct pthread *self, *new;
 238         unsigned char *map = 0, *stack = 0, *tsd = 0, *stack_limit;
 239         unsigned flags = CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND
 240                 | CLONE_THREAD | CLONE_SYSVSEM | CLONE_SETTLS
 241                 | CLONE_PARENT_SETTID | CLONE_CHILD_CLEARTID | CLONE_DETACHED;
 242         pthread_attr_t attr = { 0 };
 243         sigset_t set;
 244
 245         if (!libc.can_do_threads) return ENOSYS;
 246         self = __pthread_self();
 247         if (!libc.threaded) {
 248                 for (FILE *f=*__ofl_lock(); f; f=f->next)
 249                         init_file_lock(f);
 250                 __ofl_unlock();
 251                 init_file_lock(__stdin_used);
 252                 init_file_lock(__stdout_used);
 253                 init_file_lock(__stderr_used);
 254                 __syscall(SYS_rt_sigprocmask, SIG_UNBLOCK, SIGPT_SET, 0, _NSIG/8);
 255                 self->tsd = (void **)__pthread_tsd_main;
 256                 __membarrier_init();
 257                 libc.threaded = 1;
 258         }
 259         if (attrp && !c11) attr = *attrp;
 260
 261         __acquire_ptc();
 262         if (!attrp || c11) {
 263                 attr._a_stacksize = __default_stacksize;
 264                 attr._a_guardsize = __default_guardsize;
 265         }
 266
 267         if (attr._a_stackaddr) {
 268                 size_t need = libc.tls_size + __pthread_tsd_size;
 269                 size = attr._a_stacksize;
 270                 stack = (void *)(attr._a_stackaddr & -16);
 271                 stack_limit = (void *)(attr._a_stackaddr - size);
 272                 /* Use application-provided stack for TLS only when
 273                  * it does not take more than ~12% or 2k of the
 274                  * application's stack space. */
 275                 if (need < size/8 && need < 2048) {
 276                         tsd = stack - __pthread_tsd_size;
 277                         stack = tsd - libc.tls_size;
 278                         memset(stack, 0, need);
 279                 } else {
 280                         size = ROUND(need);
 281                 }
 282                 guard = 0;
 283         } else {
 284                 guard = ROUND(attr._a_guardsize);
 285                 size = guard + ROUND(attr._a_stacksize
 286                         + libc.tls_size +  __pthread_tsd_size);
 287         }
 288
 289         if (!tsd) {
 290                 if (guard) {
 291                         map = __mmap(0, size, PROT_NONE, MAP_PRIVATE|MAP_ANON, -1, 0);
 292                         if (map == MAP_FAILED) goto fail;
 293                         if (__mprotect(map+guard, size-guard, PROT_READ|PROT_WRITE)
 294                             && errno != ENOSYS) {
 295                                 __munmap(map, size);
 296                                 goto fail;
 297                         }
 298                 } else {
 299                         map = __mmap(0, size, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANON, -1, 0);
 300                         if (map == MAP_FAILED) goto fail;
 301                 }
 302                 tsd = map + size - __pthread_tsd_size;
 303                 if (!stack) {
 304                         stack = tsd - libc.tls_size;
 305                         stack_limit = map + guard;
 306                 }
 307         }
 308
 309         new = __copy_tls(tsd - libc.tls_size);
 310         new->map_base = map;
 311         new->map_size = size;
 312         new->stack = stack;
 313         new->stack_size = stack - stack_limit;
 314         new->guard_size = guard;
 315         new->self = new;
 316         new->tsd = (void *)tsd;
 317         new->locale = &libc.global_locale;
 318         if (attr._a_detach) {
 319                 new->detach_state = DT_DETACHED;
 320         } else {
 321                 new->detach_state = DT_JOINABLE;
 322         }
 323         new->robust_list.head = &new->robust_list.head;
 324         new->canary = self->canary;
 325         new->sysinfo = self->sysinfo;
 326
 327         /* Setup argument structure for the new thread on its stack.
 328          * It's safe to access from the caller only until the thread
 329          * list is unlocked. */
 330         stack -= (uintptr_t)stack % sizeof(uintptr_t);
 331         stack -= sizeof(struct start_args);
 332         struct start_args *args = (void *)stack;
 333         args->start_func = entry;
 334         args->start_arg = arg;
 335         args->control = attr._a_sched ? 1 : 0;
 336
 337         /* Application signals (but not the synccall signal) must be
 338          * blocked before the thread list lock can be taken, to ensure
 339          * that the lock is AS-safe. */
 340         __block_app_sigs(&set);
 341
 342         /* Ensure SIGCANCEL is unblocked in new thread. This requires
 343          * working with a copy of the set so we can restore the
 344          * original mask in the calling thread. */
 345         memcpy(&args->sig_mask, &set, sizeof args->sig_mask);
 346         args->sig_mask[(SIGCANCEL-1)/8/sizeof(long)] &=
 347                 ~(1UL<<((SIGCANCEL-1)%(8*sizeof(long))));
 348
 349         __tl_lock();
 350         if (!libc.threads_minus_1++) libc.need_locks = 1;
 351         ret = __clone((c11 ? start_c11 : start), stack, flags, args, &new->tid, TP_ADJ(new), &__thread_list_lock);
 352
 353         /* All clone failures translate to EAGAIN. If explicit scheduling
 354          * was requested, attempt it before unlocking the thread list so
 355          * that the failed thread is never exposed and so that we can
 356          * clean up all transient resource usage before returning. */
 357         if (ret < 0) {
 358                 ret = -EAGAIN;
 359         } else if (attr._a_sched) {
 360                 ret = __syscall(SYS_sched_setscheduler,
 361                         new->tid, attr._a_policy, &attr._a_prio);
 362                 if (a_swap(&args->control, ret ? 3 : 0)==2)
 363                         __wake(&args->control, 1, 1);
 364                 if (ret)
 365                         __wait(&args->control, 0, 3, 0);
 366         }
 367
 368         if (ret >= 0) {
 369                 new->next = self->next;
 370                 new->prev = self;
 371                 new->next->prev = new;
 372                 new->prev->next = new;
 373         } else {
 374                 if (!--libc.threads_minus_1) libc.need_locks = 0;
 375         }
 376         __tl_unlock();
 377         __restore_sigs(&set);
 378         __release_ptc();
 379
 380         if (ret < 0) {
 381                 if (map) __munmap(map, size);
 382                 return -ret;
 383         }
 384
 385         *res = new;
 386         return 0;
 387 fail:
 388         __release_ptc();
 389         return EAGAIN;
 390 }
 391
 392 weak_alias(__pthread_exit, pthread_exit);
 393 weak_alias(__pthread_create, pthread_create);