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