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