nptl/nptl-init.c

   1 /* Copyright (C) 2002-2016 Free Software Foundation, Inc.
   2    This file is part of the GNU C Library.
   3    Contributed by Ulrich Drepper <drepper@redhat.com>, 2002.
   4
   5    The GNU C Library is free software; you can redistribute it and/or
   6    modify it under the terms of the GNU Lesser General Public
   7    License as published by the Free Software Foundation; either
   8    version 2.1 of the License, or (at your option) any later version.
   9
  10    The GNU C Library 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 GNU
  13    Lesser General Public License for more details.
  14
  15    You should have received a copy of the GNU Lesser General Public
  16    License along with the GNU C Library; if not, see
  17    <http://www.gnu.org/licenses/>.  */
  18
  19 #include <assert.h>
  20 #include <errno.h>
  21 #include <limits.h>
  22 #include <signal.h>
  23 #include <stdlib.h>
  24 #include <unistd.h>
  25 #include <sys/param.h>
  26 #include <sys/resource.h>
  27 #include <pthreadP.h>
  28 #include <atomic.h>
  29 #include <ldsodefs.h>
  30 #include <tls.h>
  31 #include <list.h>
  32 #include <fork.h>
  33 #include <version.h>
  34 #include <shlib-compat.h>
  35 #include <smp.h>
  36 #include <lowlevellock.h>
  37 #include <futex-internal.h>
  38 #include <kernel-features.h>
  39 #include <libc-internal.h>
  40 #include <pthread-pids.h>
  41
  42 #ifndef TLS_MULTIPLE_THREADS_IN_TCB
  43 /* Pointer to the corresponding variable in libc.  */
  44 int *__libc_multiple_threads_ptr attribute_hidden;
  45 #endif
  46
  47 /* Size and alignment of static TLS block.  */
  48 size_t __static_tls_size;
  49 size_t __static_tls_align_m1;
  50
  51 #ifndef __ASSUME_SET_ROBUST_LIST
  52 /* Negative if we do not have the system call and we can use it.  */
  53 int __set_robust_list_avail;
  54 # define set_robust_list_not_avail() \
  55   __set_robust_list_avail = -1
  56 #else
  57 # define set_robust_list_not_avail() do { } while (0)
  58 #endif
  59
  60 #ifndef __ASSUME_FUTEX_CLOCK_REALTIME
  61 /* Nonzero if we do not have FUTEX_CLOCK_REALTIME.  */
  62 int __have_futex_clock_realtime;
  63 # define __set_futex_clock_realtime() \
  64   __have_futex_clock_realtime = 1
  65 #else
  66 #define __set_futex_clock_realtime() do { } while (0)
  67 #endif
  68
  69 /* Version of the library, used in libthread_db to detect mismatches.  */
  70 static const char nptl_version[] __attribute_used__ = VERSION;
  71
  72
  73 #ifndef SHARED
  74 extern void __libc_setup_tls (size_t tcbsize, size_t tcbalign);
  75 #endif
  76
  77 #ifdef SHARED
  78 static
  79 #else
  80 extern
  81 #endif
  82 void __nptl_set_robust (struct pthread *);
  83
  84 #ifdef SHARED
  85 static void nptl_freeres (void);
  86
  87
  88 static const struct pthread_functions pthread_functions =
  89   {
  90     .ptr_pthread_attr_destroy = __pthread_attr_destroy,
  91 # if SHLIB_COMPAT(libpthread, GLIBC_2_0, GLIBC_2_1)
  92     .ptr___pthread_attr_init_2_0 = __pthread_attr_init_2_0,
  93 # endif
  94     .ptr___pthread_attr_init_2_1 = __pthread_attr_init_2_1,
  95     .ptr_pthread_attr_getdetachstate = __pthread_attr_getdetachstate,
  96     .ptr_pthread_attr_setdetachstate = __pthread_attr_setdetachstate,
  97     .ptr_pthread_attr_getinheritsched = __pthread_attr_getinheritsched,
  98     .ptr_pthread_attr_setinheritsched = __pthread_attr_setinheritsched,
  99     .ptr_pthread_attr_getschedparam = __pthread_attr_getschedparam,
 100     .ptr_pthread_attr_setschedparam = __pthread_attr_setschedparam,
 101     .ptr_pthread_attr_getschedpolicy = __pthread_attr_getschedpolicy,
 102     .ptr_pthread_attr_setschedpolicy = __pthread_attr_setschedpolicy,
 103     .ptr_pthread_attr_getscope = __pthread_attr_getscope,
 104     .ptr_pthread_attr_setscope = __pthread_attr_setscope,
 105     .ptr_pthread_condattr_destroy = __pthread_condattr_destroy,
 106     .ptr_pthread_condattr_init = __pthread_condattr_init,
 107     .ptr___pthread_cond_broadcast = __pthread_cond_broadcast,
 108     .ptr___pthread_cond_destroy = __pthread_cond_destroy,
 109     .ptr___pthread_cond_init = __pthread_cond_init,
 110     .ptr___pthread_cond_signal = __pthread_cond_signal,
 111     .ptr___pthread_cond_wait = __pthread_cond_wait,
 112     .ptr___pthread_cond_timedwait = __pthread_cond_timedwait,
 113 # if SHLIB_COMPAT(libpthread, GLIBC_2_0, GLIBC_2_3_2)
 114     .ptr___pthread_cond_broadcast_2_0 = __pthread_cond_broadcast_2_0,
 115     .ptr___pthread_cond_destroy_2_0 = __pthread_cond_destroy_2_0,
 116     .ptr___pthread_cond_init_2_0 = __pthread_cond_init_2_0,
 117     .ptr___pthread_cond_signal_2_0 = __pthread_cond_signal_2_0,
 118     .ptr___pthread_cond_wait_2_0 = __pthread_cond_wait_2_0,
 119     .ptr___pthread_cond_timedwait_2_0 = __pthread_cond_timedwait_2_0,
 120 # endif
 121     .ptr_pthread_equal = __pthread_equal,
 122     .ptr___pthread_exit = __pthread_exit,
 123     .ptr_pthread_getschedparam = __pthread_getschedparam,
 124     .ptr_pthread_setschedparam = __pthread_setschedparam,
 125     .ptr_pthread_mutex_destroy = __pthread_mutex_destroy,
 126     .ptr_pthread_mutex_init = __pthread_mutex_init,
 127     .ptr_pthread_mutex_lock = __pthread_mutex_lock,
 128     .ptr_pthread_mutex_unlock = __pthread_mutex_unlock,
 129     .ptr_pthread_self = __pthread_self,
 130     .ptr___pthread_setcancelstate = __pthread_setcancelstate,
 131     .ptr_pthread_setcanceltype = __pthread_setcanceltype,
 132     .ptr___pthread_cleanup_upto = __pthread_cleanup_upto,
 133     .ptr___pthread_once = __pthread_once,
 134     .ptr___pthread_rwlock_rdlock = __pthread_rwlock_rdlock,
 135     .ptr___pthread_rwlock_wrlock = __pthread_rwlock_wrlock,
 136     .ptr___pthread_rwlock_unlock = __pthread_rwlock_unlock,
 137     .ptr___pthread_key_create = __pthread_key_create,
 138     .ptr___pthread_getspecific = __pthread_getspecific,
 139     .ptr___pthread_setspecific = __pthread_setspecific,
 140     .ptr__pthread_cleanup_push_defer = __pthread_cleanup_push_defer,
 141     .ptr__pthread_cleanup_pop_restore = __pthread_cleanup_pop_restore,
 142     .ptr_nthreads = &__nptl_nthreads,
 143     .ptr___pthread_unwind = &__pthread_unwind,
 144     .ptr__nptl_deallocate_tsd = __nptl_deallocate_tsd,
 145 # ifdef SIGSETXID
 146     .ptr__nptl_setxid = __nptl_setxid,
 147 # endif
 148     /* For now only the stack cache needs to be freed.  */
 149     .ptr_freeres = nptl_freeres,
 150     .ptr_set_robust = __nptl_set_robust
 151   };
 152 # define ptr_pthread_functions &pthread_functions
 153 #else
 154 # define ptr_pthread_functions NULL
 155 #endif
 156
 157
 158 #ifdef SHARED
 159 /* This function is called indirectly from the freeres code in libc.  */
 160 static void
 161 __libc_freeres_fn_section
 162 nptl_freeres (void)
 163 {
 164   __unwind_freeres ();
 165   __free_stacks (0);
 166 }
 167
 168
 169 static
 170 #endif
 171 void
 172 __nptl_set_robust (struct pthread *self)
 173 {
 174 #ifdef __NR_set_robust_list
 175   INTERNAL_SYSCALL_DECL (err);
 176   INTERNAL_SYSCALL (set_robust_list, err, 2, &self->robust_head,
 177                     sizeof (struct robust_list_head));
 178 #endif
 179 }
 180
 181
 182 #ifdef SIGCANCEL
 183 /* For asynchronous cancellation we use a signal.  This is the handler.  */
 184 static void
 185 sigcancel_handler (int sig, siginfo_t *si, void *ctx)
 186 {
 187   /* Safety check.  It would be possible to call this function for
 188      other signals and send a signal from another process.  This is not
 189      correct and might even be a security problem.  Try to catch as
 190      many incorrect invocations as possible.  */
 191   if (sig != SIGCANCEL
 192       || si->si_pid != __getpid()
 193       || si->si_code != SI_TKILL)
 194     return;
 195
 196   struct pthread *self = THREAD_SELF;
 197
 198   int oldval = THREAD_GETMEM (self, cancelhandling);
 199   while (1)
 200     {
 201       /* We are canceled now.  When canceled by another thread this flag
 202          is already set but if the signal is directly send (internally or
 203          from another process) is has to be done here.  */
 204       int newval = oldval | CANCELING_BITMASK | CANCELED_BITMASK;
 205
 206       if (oldval == newval || (oldval & EXITING_BITMASK) != 0)
 207         /* Already canceled or exiting.  */
 208         break;
 209
 210       int curval = THREAD_ATOMIC_CMPXCHG_VAL (self, cancelhandling, newval,
 211                                               oldval);
 212       if (curval == oldval)
 213         {
 214           /* Set the return value.  */
 215           THREAD_SETMEM (self, result, PTHREAD_CANCELED);
 216
 217           /* Make sure asynchronous cancellation is still enabled.  */
 218           if ((newval & CANCELTYPE_BITMASK) != 0)
 219             /* Run the registered destructors and terminate the thread.  */
 220             __do_cancel ();
 221
 222           break;
 223         }
 224
 225       oldval = curval;
 226     }
 227 }
 228 #endif
 229
 230
 231 #ifdef SIGSETXID
 232 struct xid_command *__xidcmd attribute_hidden;
 233
 234 /* We use the SIGSETXID signal in the setuid, setgid, etc. implementations to
 235    tell each thread to call the respective setxid syscall on itself.  This is
 236    the handler.  */
 237 static void
 238 sighandler_setxid (int sig, siginfo_t *si, void *ctx)
 239 {
 240   int result;
 241
 242   /* Safety check.  It would be possible to call this function for
 243      other signals and send a signal from another process.  This is not
 244      correct and might even be a security problem.  Try to catch as
 245      many incorrect invocations as possible.  */
 246   if (sig != SIGSETXID
 247       || si->si_pid != __getpid ()
 248       || si->si_code != SI_TKILL)
 249     return;
 250
 251   INTERNAL_SYSCALL_DECL (err);
 252   result = INTERNAL_SYSCALL_NCS (__xidcmd->syscall_no, err, 3, __xidcmd->id[0],
 253                                  __xidcmd->id[1], __xidcmd->id[2]);
 254   int error = 0;
 255   if (__glibc_unlikely (INTERNAL_SYSCALL_ERROR_P (result, err)))
 256     error = INTERNAL_SYSCALL_ERRNO (result, err);
 257   __nptl_setxid_error (__xidcmd, error);
 258
 259   /* Reset the SETXID flag.  */
 260   struct pthread *self = THREAD_SELF;
 261   int flags, newval;
 262   do
 263     {
 264       flags = THREAD_GETMEM (self, cancelhandling);
 265       newval = THREAD_ATOMIC_CMPXCHG_VAL (self, cancelhandling,
 266                                           flags & ~SETXID_BITMASK, flags);
 267     }
 268   while (flags != newval);
 269
 270   /* And release the futex.  */
 271   self->setxid_futex = 1;
 272   futex_wake (&self->setxid_futex, 1, FUTEX_PRIVATE);
 273
 274   if (atomic_decrement_val (&__xidcmd->cntr) == 0)
 275     futex_wake ((unsigned int *) &__xidcmd->cntr, 1, FUTEX_PRIVATE);
 276 }
 277 #endif
 278
 279
 280 /* When using __thread for this, we do it in libc so as not
 281    to give libpthread its own TLS segment just for this.  */
 282 extern void **__libc_dl_error_tsd (void) __attribute__ ((const));
 283
 284
 285 /* This can be set by the debugger before initialization is complete.  */
 286 static bool __nptl_initial_report_events __attribute_used__;
 287
 288 void
 289 __pthread_initialize_minimal_internal (void)
 290 {
 291 #ifndef SHARED
 292   /* Unlike in the dynamically linked case the dynamic linker has not
 293      taken care of initializing the TLS data structures.  */
 294   __libc_setup_tls (TLS_TCB_SIZE, TLS_TCB_ALIGN);
 295
 296   /* We must prevent gcc from being clever and move any of the
 297      following code ahead of the __libc_setup_tls call.  This function
 298      will initialize the thread register which is subsequently
 299      used.  */
 300   __asm __volatile ("");
 301 #endif
 302
 303   /* Minimal initialization of the thread descriptor.  */
 304   struct pthread *pd = THREAD_SELF;
 305   __pthread_initialize_pids (pd);
 306   THREAD_SETMEM (pd, specific[0], &pd->specific_1stblock[0]);
 307   THREAD_SETMEM (pd, user_stack, true);
 308   if (LLL_LOCK_INITIALIZER != 0)
 309     THREAD_SETMEM (pd, lock, LLL_LOCK_INITIALIZER);
 310 #if HP_TIMING_AVAIL
 311   THREAD_SETMEM (pd, cpuclock_offset, GL(dl_cpuclock_offset));
 312 #endif
 313
 314   /* Initialize the robust mutex data.  */
 315   {
 316 #ifdef __PTHREAD_MUTEX_HAVE_PREV
 317     pd->robust_prev = &pd->robust_head;
 318 #endif
 319     pd->robust_head.list = &pd->robust_head;
 320 #ifdef __NR_set_robust_list
 321     pd->robust_head.futex_offset = (offsetof (pthread_mutex_t, __data.__lock)
 322                                     - offsetof (pthread_mutex_t,
 323                                                 __data.__list.__next));
 324     INTERNAL_SYSCALL_DECL (err);
 325     int res = INTERNAL_SYSCALL (set_robust_list, err, 2, &pd->robust_head,
 326                                 sizeof (struct robust_list_head));
 327     if (INTERNAL_SYSCALL_ERROR_P (res, err))
 328 #endif
 329       set_robust_list_not_avail ();
 330   }
 331
 332 #ifdef __NR_futex
 333 # ifndef __ASSUME_PRIVATE_FUTEX
 334   /* Private futexes are always used (at least internally) so that
 335      doing the test once this early is beneficial.  */
 336   {
 337     int word = 0;
 338     INTERNAL_SYSCALL_DECL (err);
 339     word = INTERNAL_SYSCALL (futex, err, 3, &word,
 340                             FUTEX_WAKE | FUTEX_PRIVATE_FLAG, 1);
 341     if (!INTERNAL_SYSCALL_ERROR_P (word, err))
 342       THREAD_SETMEM (pd, header.private_futex, FUTEX_PRIVATE_FLAG);
 343   }
 344
 345   /* Private futexes have been introduced earlier than the
 346      FUTEX_CLOCK_REALTIME flag.  We don't have to run the test if we
 347      know the former are not supported.  This also means we know the
 348      kernel will return ENOSYS for unknown operations.  */
 349   if (THREAD_GETMEM (pd, header.private_futex) != 0)
 350 # endif
 351 # ifndef __ASSUME_FUTEX_CLOCK_REALTIME
 352     {
 353       int word = 0;
 354       /* NB: the syscall actually takes six parameters.  The last is the
 355          bit mask.  But since we will not actually wait at all the value
 356          is irrelevant.  Given that passing six parameters is difficult
 357          on some architectures we just pass whatever random value the
 358          calling convention calls for to the kernel.  It causes no harm.  */
 359       INTERNAL_SYSCALL_DECL (err);
 360       word = INTERNAL_SYSCALL (futex, err, 5, &word,
 361                                FUTEX_WAIT_BITSET | FUTEX_CLOCK_REALTIME
 362                                | FUTEX_PRIVATE_FLAG, 1, NULL, 0);
 363       assert (INTERNAL_SYSCALL_ERROR_P (word, err));
 364       if (INTERNAL_SYSCALL_ERRNO (word, err) != ENOSYS)
 365         __set_futex_clock_realtime ();
 366     }
 367 # endif
 368 #endif
 369
 370   /* Set initial thread's stack block from 0 up to __libc_stack_end.
 371      It will be bigger than it actually is, but for unwind.c/pt-longjmp.c
 372      purposes this is good enough.  */
 373   THREAD_SETMEM (pd, stackblock_size, (size_t) __libc_stack_end);
 374
 375   /* Initialize the list of all running threads with the main thread.  */
 376   INIT_LIST_HEAD (&__stack_user);
 377   list_add (&pd->list, &__stack_user);
 378
 379   /* Before initializing __stack_user, the debugger could not find us and
 380      had to set __nptl_initial_report_events.  Propagate its setting.  */
 381   THREAD_SETMEM (pd, report_events, __nptl_initial_report_events);
 382
 383 #if defined SIGCANCEL || defined SIGSETXID
 384   struct sigaction sa;
 385   __sigemptyset (&sa.sa_mask);
 386
 387 # ifdef SIGCANCEL
 388   /* Install the cancellation signal handler.  If for some reason we
 389      cannot install the handler we do not abort.  Maybe we should, but
 390      it is only asynchronous cancellation which is affected.  */
 391   sa.sa_sigaction = sigcancel_handler;
 392   sa.sa_flags = SA_SIGINFO;
 393   (void) __libc_sigaction (SIGCANCEL, &sa, NULL);
 394 # endif
 395
 396 # ifdef SIGSETXID
 397   /* Install the handle to change the threads' uid/gid.  */
 398   sa.sa_sigaction = sighandler_setxid;
 399   sa.sa_flags = SA_SIGINFO | SA_RESTART;
 400   (void) __libc_sigaction (SIGSETXID, &sa, NULL);
 401 # endif
 402
 403   /* The parent process might have left the signals blocked.  Just in
 404      case, unblock it.  We reuse the signal mask in the sigaction
 405      structure.  It is already cleared.  */
 406 # ifdef SIGCANCEL
 407   __sigaddset (&sa.sa_mask, SIGCANCEL);
 408 # endif
 409 # ifdef SIGSETXID
 410   __sigaddset (&sa.sa_mask, SIGSETXID);
 411 # endif
 412   {
 413     INTERNAL_SYSCALL_DECL (err);
 414     (void) INTERNAL_SYSCALL (rt_sigprocmask, err, 4, SIG_UNBLOCK, &sa.sa_mask,
 415                              NULL, _NSIG / 8);
 416   }
 417 #endif
 418
 419   /* Get the size of the static and alignment requirements for the TLS
 420      block.  */
 421   size_t static_tls_align;
 422   _dl_get_tls_static_info (&__static_tls_size, &static_tls_align);
 423
 424   /* Make sure the size takes all the alignments into account.  */
 425   if (STACK_ALIGN > static_tls_align)
 426     static_tls_align = STACK_ALIGN;
 427   __static_tls_align_m1 = static_tls_align - 1;
 428
 429   __static_tls_size = roundup (__static_tls_size, static_tls_align);
 430
 431   /* Determine the default allowed stack size.  This is the size used
 432      in case the user does not specify one.  */
 433   struct rlimit limit;
 434   if (__getrlimit (RLIMIT_STACK, &limit) != 0
 435       || limit.rlim_cur == RLIM_INFINITY)
 436     /* The system limit is not usable.  Use an architecture-specific
 437        default.  */
 438     limit.rlim_cur = ARCH_STACK_DEFAULT_SIZE;
 439   else if (limit.rlim_cur < PTHREAD_STACK_MIN)
 440     /* The system limit is unusably small.
 441        Use the minimal size acceptable.  */
 442     limit.rlim_cur = PTHREAD_STACK_MIN;
 443
 444   /* Make sure it meets the minimum size that allocate_stack
 445      (allocatestack.c) will demand, which depends on the page size.  */
 446   const uintptr_t pagesz = GLRO(dl_pagesize);
 447   const size_t minstack = pagesz + __static_tls_size + MINIMAL_REST_STACK;
 448   if (limit.rlim_cur < minstack)
 449     limit.rlim_cur = minstack;
 450
 451   /* Round the resource limit up to page size.  */
 452   limit.rlim_cur = ALIGN_UP (limit.rlim_cur, pagesz);
 453   lll_lock (__default_pthread_attr_lock, LLL_PRIVATE);
 454   __default_pthread_attr.stacksize = limit.rlim_cur;
 455   __default_pthread_attr.guardsize = GLRO (dl_pagesize);
 456   lll_unlock (__default_pthread_attr_lock, LLL_PRIVATE);
 457
 458 #ifdef SHARED
 459   /* Make __rtld_lock_{,un}lock_recursive use pthread_mutex_{,un}lock,
 460      keep the lock count from the ld.so implementation.  */
 461   GL(dl_rtld_lock_recursive) = (void *) __pthread_mutex_lock;
 462   GL(dl_rtld_unlock_recursive) = (void *) __pthread_mutex_unlock;
 463   unsigned int rtld_lock_count = GL(dl_load_lock).mutex.__data.__count;
 464   GL(dl_load_lock).mutex.__data.__count = 0;
 465   while (rtld_lock_count-- > 0)
 466     __pthread_mutex_lock (&GL(dl_load_lock).mutex);
 467
 468   GL(dl_make_stack_executable_hook) = &__make_stacks_executable;
 469 #endif
 470
 471   GL(dl_init_static_tls) = &__pthread_init_static_tls;
 472
 473   GL(dl_wait_lookup_done) = &__wait_lookup_done;
 474
 475   /* Register the fork generation counter with the libc.  */
 476 #ifndef TLS_MULTIPLE_THREADS_IN_TCB
 477   __libc_multiple_threads_ptr =
 478 #endif
 479     __libc_pthread_init (&__fork_generation, __reclaim_stacks,
 480                          ptr_pthread_functions);
 481
 482   /* Determine whether the machine is SMP or not.  */
 483   __is_smp = is_smp_system ();
 484 }
 485 strong_alias (__pthread_initialize_minimal_internal,
 486               __pthread_initialize_minimal)
 487
 488
 489 size_t
 490 __pthread_get_minstack (const pthread_attr_t *attr)
 491 {
 492   struct pthread_attr *iattr = (struct pthread_attr *) attr;
 493
 494   return (GLRO(dl_pagesize) + __static_tls_size + PTHREAD_STACK_MIN
 495           + iattr->guardsize);
 496 }