nsz Git - musl/blob - 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 <sys/mman.h>
   6 #include <string.h>
   7 #include <stddef.h>
   8
   9 static void dummy_0()
  10 {
  11 }
  12 weak_alias(dummy_0, __acquire_ptc);
  13 weak_alias(dummy_0, __release_ptc);
  14 weak_alias(dummy_0, __pthread_tsd_run_dtors);
  15 weak_alias(dummy_0, __do_orphaned_stdio_locks);
  16 weak_alias(dummy_0, __dl_thread_cleanup);
  17
  18 static void *dummy_1(void *p)
  19 {
  20         return 0;
  21 }
  22 weak_alias(dummy_1, __start_sched);
  23
  24 _Noreturn void __pthread_exit(void *result)
  25 {
  26         pthread_t self = __pthread_self();
  27         sigset_t set;
  28
  29         self->canceldisable = 1;
  30         self->cancelasync = 0;
  31         self->result = result;
  32
  33         while (self->cancelbuf) {
  34                 void (*f)(void *) = self->cancelbuf->__f;
  35                 void *x = self->cancelbuf->__x;
  36                 self->cancelbuf = self->cancelbuf->__next;
  37                 f(x);
  38         }
  39
  40         __pthread_tsd_run_dtors();
  41
  42         /* Access to target the exiting thread with syscalls that use
  43          * its kernel tid is controlled by killlock. For detached threads,
  44          * any use past this point would have undefined behavior, but for
  45          * joinable threads it's a valid usage that must be handled. */
  46         LOCK(self->killlock);
  47
  48         /* Block all signals before decrementing the live thread count.
  49          * This is important to ensure that dynamically allocated TLS
  50          * is not under-allocated/over-committed, and possibly for other
  51          * reasons as well. */
  52         __block_all_sigs(&set);
  53
  54         /* It's impossible to determine whether this is "the last thread"
  55          * until performing the atomic decrement, since multiple threads
  56          * could exit at the same time. For the last thread, revert the
  57          * decrement, restore the tid, and unblock signals to give the
  58          * atexit handlers and stdio cleanup code a consistent state. */
  59         if (a_fetch_add(&libc.threads_minus_1, -1)==0) {
  60                 libc.threads_minus_1 = 0;
  61                 UNLOCK(self->killlock);
  62                 __restore_sigs(&set);
  63                 exit(0);
  64         }
  65
  66         /* Process robust list in userspace to handle non-pshared mutexes
  67          * and the detached thread case where the robust list head will
  68          * be invalid when the kernel would process it. */
  69         __vm_lock();
  70         volatile void *volatile *rp;
  71         while ((rp=self->robust_list.head) && rp != &self->robust_list.head) {
  72                 pthread_mutex_t *m = (void *)((char *)rp
  73                         - offsetof(pthread_mutex_t, _m_next));
  74                 int waiters = m->_m_waiters;
  75                 int priv = (m->_m_type & 128) ^ 128;
  76                 self->robust_list.pending = rp;
  77                 self->robust_list.head = *rp;
  78                 int cont = a_swap(&m->_m_lock, 0x40000000);
  79                 self->robust_list.pending = 0;
  80                 if (cont < 0 || waiters)
  81                         __wake(&m->_m_lock, 1, priv);
  82         }
  83         __vm_unlock();
  84
  85         __do_orphaned_stdio_locks();
  86         __dl_thread_cleanup();
  87
  88         /* This atomic potentially competes with a concurrent pthread_detach
  89          * call; the loser is responsible for freeing thread resources. */
  90         int state = a_cas(&self->detach_state, DT_JOINABLE, DT_EXITING);
  91
  92         if (state>=DT_DETACHED && self->map_base) {
  93                 /* Detached threads must avoid the kernel clear_child_tid
  94                  * feature, since the virtual address will have been
  95                  * unmapped and possibly already reused by a new mapping
  96                  * at the time the kernel would perform the write. In
  97                  * the case of threads that started out detached, the
  98                  * initial clone flags are correct, but if the thread was
  99                  * detached later, we need to clear it here. */
 100                 if (state == DT_DYNAMIC) __syscall(SYS_set_tid_address, 0);
 101
 102                 /* Robust list will no longer be valid, and was already
 103                  * processed above, so unregister it with the kernel. */
 104                 if (self->robust_list.off)
 105                         __syscall(SYS_set_robust_list, 0, 3*sizeof(long));
 106
 107                 /* Since __unmapself bypasses the normal munmap code path,
 108                  * explicitly wait for vmlock holders first. */
 109                 __vm_wait();
 110
 111                 /* The following call unmaps the thread's stack mapping
 112                  * and then exits without touching the stack. */
 113                 __unmapself(self->map_base, self->map_size);
 114         }
 115
 116         /* After the kernel thread exits, its tid may be reused. Clear it
 117          * to prevent inadvertent use and inform functions that would use
 118          * it that it's no longer available. */
 119         self->tid = 0;
 120         UNLOCK(self->killlock);
 121
 122         for (;;) __syscall(SYS_exit, 0);
 123 }
 124
 125 void __do_cleanup_push(struct __ptcb *cb)
 126 {
 127         struct pthread *self = __pthread_self();
 128         cb->__next = self->cancelbuf;
 129         self->cancelbuf = cb;
 130 }
 131
 132 void __do_cleanup_pop(struct __ptcb *cb)
 133 {
 134         __pthread_self()->cancelbuf = cb->__next;
 135 }
 136
 137 static int start(void *p)
 138 {
 139         pthread_t self = p;
 140         if (self->unblock_cancel)
 141                 __syscall(SYS_rt_sigprocmask, SIG_UNBLOCK,
 142                         SIGPT_SET, 0, _NSIG/8);
 143         __pthread_exit(self->start(self->start_arg));
 144         return 0;
 145 }
 146
 147 static int start_c11(void *p)
 148 {
 149         pthread_t self = p;
 150         int (*start)(void*) = (int(*)(void*)) self->start;
 151         __pthread_exit((void *)(uintptr_t)start(self->start_arg));
 152         return 0;
 153 }
 154
 155 #define ROUND(x) (((x)+PAGE_SIZE-1)&-PAGE_SIZE)
 156
 157 /* pthread_key_create.c overrides this */
 158 static volatile size_t dummy = 0;
 159 weak_alias(dummy, __pthread_tsd_size);
 160 static void *dummy_tsd[1] = { 0 };
 161 weak_alias(dummy_tsd, __pthread_tsd_main);
 162
 163 volatile int __block_new_threads = 0;
 164 extern size_t __default_stacksize;
 165 extern size_t __default_guardsize;
 166
 167 static FILE *volatile dummy_file = 0;
 168 weak_alias(dummy_file, __stdin_used);
 169 weak_alias(dummy_file, __stdout_used);
 170 weak_alias(dummy_file, __stderr_used);
 171
 172 static void init_file_lock(FILE *f)
 173 {
 174         if (f && f->lock<0) f->lock = 0;
 175 }
 176
 177 int __pthread_create(pthread_t *restrict res, const pthread_attr_t *restrict attrp, void *(*entry)(void *), void *restrict arg)
 178 {
 179         int ret, c11 = (attrp == __ATTRP_C11_THREAD);
 180         size_t size, guard;
 181         struct pthread *self, *new;
 182         unsigned char *map = 0, *stack = 0, *tsd = 0, *stack_limit;
 183         unsigned flags = CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND
 184                 | CLONE_THREAD | CLONE_SYSVSEM | CLONE_SETTLS
 185                 | CLONE_PARENT_SETTID | CLONE_CHILD_CLEARTID | CLONE_DETACHED;
 186         int do_sched = 0;
 187         pthread_attr_t attr = { 0 };
 188         struct start_sched_args ssa;
 189
 190         if (!libc.can_do_threads) return ENOSYS;
 191         self = __pthread_self();
 192         if (!libc.threaded) {
 193                 for (FILE *f=*__ofl_lock(); f; f=f->next)
 194                         init_file_lock(f);
 195                 __ofl_unlock();
 196                 init_file_lock(__stdin_used);
 197                 init_file_lock(__stdout_used);
 198                 init_file_lock(__stderr_used);
 199                 __syscall(SYS_rt_sigprocmask, SIG_UNBLOCK, SIGPT_SET, 0, _NSIG/8);
 200                 self->tsd = (void **)__pthread_tsd_main;
 201                 libc.threaded = 1;
 202         }
 203         if (attrp && !c11) attr = *attrp;
 204
 205         __acquire_ptc();
 206         if (!attrp || c11) {
 207                 attr._a_stacksize = __default_stacksize;
 208                 attr._a_guardsize = __default_guardsize;
 209         }
 210
 211         if (__block_new_threads) __wait(&__block_new_threads, 0, 1, 1);
 212
 213         if (attr._a_stackaddr) {
 214                 size_t need = libc.tls_size + __pthread_tsd_size;
 215                 size = attr._a_stacksize;
 216                 stack = (void *)(attr._a_stackaddr & -16);
 217                 stack_limit = (void *)(attr._a_stackaddr - size);
 218                 /* Use application-provided stack for TLS only when
 219                  * it does not take more than ~12% or 2k of the
 220                  * application's stack space. */
 221                 if (need < size/8 && need < 2048) {
 222                         tsd = stack - __pthread_tsd_size;
 223                         stack = tsd - libc.tls_size;
 224                         memset(stack, 0, need);
 225                 } else {
 226                         size = ROUND(need);
 227                 }
 228                 guard = 0;
 229         } else {
 230                 guard = ROUND(attr._a_guardsize);
 231                 size = guard + ROUND(attr._a_stacksize
 232                         + libc.tls_size +  __pthread_tsd_size);
 233         }
 234
 235         if (!tsd) {
 236                 if (guard) {
 237                         map = __mmap(0, size, PROT_NONE, MAP_PRIVATE|MAP_ANON, -1, 0);
 238                         if (map == MAP_FAILED) goto fail;
 239                         if (__mprotect(map+guard, size-guard, PROT_READ|PROT_WRITE)
 240                             && errno != ENOSYS) {
 241                                 __munmap(map, size);
 242                                 goto fail;
 243                         }
 244                 } else {
 245                         map = __mmap(0, size, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANON, -1, 0);
 246                         if (map == MAP_FAILED) goto fail;
 247                 }
 248                 tsd = map + size - __pthread_tsd_size;
 249                 if (!stack) {
 250                         stack = tsd - libc.tls_size;
 251                         stack_limit = map + guard;
 252                 }
 253         }
 254
 255         new = __copy_tls(tsd - libc.tls_size);
 256         new->map_base = map;
 257         new->map_size = size;
 258         new->stack = stack;
 259         new->stack_size = stack - stack_limit;
 260         new->guard_size = guard;
 261         new->start = entry;
 262         new->start_arg = arg;
 263         new->self = new;
 264         new->tsd = (void *)tsd;
 265         new->locale = &libc.global_locale;
 266         if (attr._a_detach) {
 267                 new->detach_state = DT_DETACHED;
 268                 flags -= CLONE_CHILD_CLEARTID;
 269         } else {
 270                 new->detach_state = DT_JOINABLE;
 271         }
 272         if (attr._a_sched) {
 273                 do_sched = 1;
 274                 ssa.futex = -1;
 275                 ssa.start_fn = new->start;
 276                 ssa.start_arg = new->start_arg;
 277                 ssa.attr = &attr;
 278                 new->start = __start_sched;
 279                 new->start_arg = &ssa;
 280                 __block_app_sigs(&ssa.mask);
 281         }
 282         new->robust_list.head = &new->robust_list.head;
 283         new->unblock_cancel = self->cancel;
 284         new->CANARY = self->CANARY;
 285
 286         a_inc(&libc.threads_minus_1);
 287         ret = __clone((c11 ? start_c11 : start), stack, flags, new, &new->tid, TP_ADJ(new), &new->detach_state);
 288
 289         __release_ptc();
 290
 291         if (do_sched) {
 292                 __restore_sigs(&ssa.mask);
 293         }
 294
 295         if (ret < 0) {
 296                 a_dec(&libc.threads_minus_1);
 297                 if (map) __munmap(map, size);
 298                 return EAGAIN;
 299         }
 300
 301         if (do_sched) {
 302                 __futexwait(&ssa.futex, -1, 1);
 303                 ret = ssa.futex;
 304                 if (ret) return ret;
 305         }
 306
 307         *res = new;
 308         return 0;
 309 fail:
 310         __release_ptc();
 311         return EAGAIN;
 312 }
 313
 314 weak_alias(__pthread_exit, pthread_exit);
 315 weak_alias(__pthread_create, pthread_create);